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/sched/topology.h>
16 #include <linux/cpuset.h>
17 #include <linux/cpumask.h>
18 #include <linux/init.h>
19 #include <linux/rcupdate.h>
20 #include <linux/sched.h>
22 static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
23 static struct cpumask scale_freq_counters_mask;
24 static bool scale_freq_invariant;
26 static bool supports_scale_freq_counters(const struct cpumask *cpus)
28 return cpumask_subset(cpus, &scale_freq_counters_mask);
31 bool topology_scale_freq_invariant(void)
33 return cpufreq_supports_freq_invariance() ||
34 supports_scale_freq_counters(cpu_online_mask);
37 static void update_scale_freq_invariant(bool status)
39 if (scale_freq_invariant == status)
43 * Task scheduler behavior depends on frequency invariance support,
44 * either cpufreq or counter driven. If the support status changes as
45 * a result of counter initialisation and use, retrigger the build of
46 * scheduling domains to ensure the information is propagated properly.
48 if (topology_scale_freq_invariant() == status) {
49 scale_freq_invariant = status;
50 rebuild_sched_domains_energy();
54 void topology_set_scale_freq_source(struct scale_freq_data *data,
55 const struct cpumask *cpus)
57 struct scale_freq_data *sfd;
61 * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
62 * supported by cpufreq.
64 if (cpumask_empty(&scale_freq_counters_mask))
65 scale_freq_invariant = topology_scale_freq_invariant();
69 for_each_cpu(cpu, cpus) {
70 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
72 /* Use ARCH provided counters whenever possible */
73 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
74 rcu_assign_pointer(per_cpu(sft_data, cpu), data);
75 cpumask_set_cpu(cpu, &scale_freq_counters_mask);
81 update_scale_freq_invariant(true);
83 EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
85 void topology_clear_scale_freq_source(enum scale_freq_source source,
86 const struct cpumask *cpus)
88 struct scale_freq_data *sfd;
93 for_each_cpu(cpu, cpus) {
94 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
96 if (sfd && sfd->source == source) {
97 rcu_assign_pointer(per_cpu(sft_data, cpu), NULL);
98 cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
105 * Make sure all references to previous sft_data are dropped to avoid
106 * use-after-free races.
110 update_scale_freq_invariant(false);
112 EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
114 void topology_scale_freq_tick(void)
116 struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data));
119 sfd->set_freq_scale();
122 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
123 EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
125 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
126 unsigned long max_freq)
131 if (WARN_ON_ONCE(!cur_freq || !max_freq))
135 * If the use of counters for FIE is enabled, just return as we don't
136 * want to update the scale factor with information from CPUFREQ.
137 * Instead the scale factor will be updated from arch_scale_freq_tick.
139 if (supports_scale_freq_counters(cpus))
142 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
144 for_each_cpu(i, cpus)
145 per_cpu(arch_freq_scale, i) = scale;
148 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
149 EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale);
151 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
153 per_cpu(cpu_scale, cpu) = capacity;
156 DEFINE_PER_CPU(unsigned long, thermal_pressure);
158 void topology_set_thermal_pressure(const struct cpumask *cpus,
159 unsigned long th_pressure)
163 for_each_cpu(cpu, cpus)
164 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
166 EXPORT_SYMBOL_GPL(topology_set_thermal_pressure);
168 static ssize_t cpu_capacity_show(struct device *dev,
169 struct device_attribute *attr,
172 struct cpu *cpu = container_of(dev, struct cpu, dev);
174 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
177 static void update_topology_flags_workfn(struct work_struct *work);
178 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
180 static DEVICE_ATTR_RO(cpu_capacity);
182 static int register_cpu_capacity_sysctl(void)
187 for_each_possible_cpu(i) {
188 cpu = get_cpu_device(i);
190 pr_err("%s: too early to get CPU%d device!\n",
194 device_create_file(cpu, &dev_attr_cpu_capacity);
199 subsys_initcall(register_cpu_capacity_sysctl);
201 static int update_topology;
203 int topology_update_cpu_topology(void)
205 return update_topology;
209 * Updating the sched_domains can't be done directly from cpufreq callbacks
210 * due to locking, so queue the work for later.
212 static void update_topology_flags_workfn(struct work_struct *work)
215 rebuild_sched_domains();
216 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
220 static DEFINE_PER_CPU(u32, freq_factor) = 1;
221 static u32 *raw_capacity;
223 static int free_raw_capacity(void)
231 void topology_normalize_cpu_scale(void)
241 for_each_possible_cpu(cpu) {
242 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
243 capacity_scale = max(capacity, capacity_scale);
246 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
247 for_each_possible_cpu(cpu) {
248 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
249 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
251 topology_set_cpu_scale(cpu, capacity);
252 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
253 cpu, topology_get_cpu_scale(cpu));
257 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
260 static bool cap_parsing_failed;
264 if (cap_parsing_failed)
267 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
271 raw_capacity = kcalloc(num_possible_cpus(),
272 sizeof(*raw_capacity),
275 cap_parsing_failed = true;
279 raw_capacity[cpu] = cpu_capacity;
280 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
281 cpu_node, raw_capacity[cpu]);
284 * Update freq_factor for calculating early boot cpu capacities.
285 * For non-clk CPU DVFS mechanism, there's no way to get the
286 * frequency value now, assuming they are running at the same
287 * frequency (by keeping the initial freq_factor value).
289 cpu_clk = of_clk_get(cpu_node, 0);
290 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
291 per_cpu(freq_factor, cpu) =
292 clk_get_rate(cpu_clk) / 1000;
297 pr_err("cpu_capacity: missing %pOF raw capacity\n",
299 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
301 cap_parsing_failed = true;
308 #ifdef CONFIG_CPU_FREQ
309 static cpumask_var_t cpus_to_visit;
310 static void parsing_done_workfn(struct work_struct *work);
311 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
314 init_cpu_capacity_callback(struct notifier_block *nb,
318 struct cpufreq_policy *policy = data;
324 if (val != CPUFREQ_CREATE_POLICY)
327 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
328 cpumask_pr_args(policy->related_cpus),
329 cpumask_pr_args(cpus_to_visit));
331 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
333 for_each_cpu(cpu, policy->related_cpus)
334 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
336 if (cpumask_empty(cpus_to_visit)) {
337 topology_normalize_cpu_scale();
338 schedule_work(&update_topology_flags_work);
340 pr_debug("cpu_capacity: parsing done\n");
341 schedule_work(&parsing_done_work);
347 static struct notifier_block init_cpu_capacity_notifier = {
348 .notifier_call = init_cpu_capacity_callback,
351 static int __init register_cpufreq_notifier(void)
356 * on ACPI-based systems we need to use the default cpu capacity
357 * until we have the necessary code to parse the cpu capacity, so
358 * skip registering cpufreq notifier.
360 if (!acpi_disabled || !raw_capacity)
363 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
366 cpumask_copy(cpus_to_visit, cpu_possible_mask);
368 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
369 CPUFREQ_POLICY_NOTIFIER);
372 free_cpumask_var(cpus_to_visit);
376 core_initcall(register_cpufreq_notifier);
378 static void parsing_done_workfn(struct work_struct *work)
380 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
381 CPUFREQ_POLICY_NOTIFIER);
382 free_cpumask_var(cpus_to_visit);
386 core_initcall(free_raw_capacity);
389 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
391 * This function returns the logic cpu number of the node.
392 * There are basically three kinds of return values:
393 * (1) logic cpu number which is > 0.
394 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
395 * there is no possible logical CPU in the kernel to match. This happens
396 * when CONFIG_NR_CPUS is configure to be smaller than the number of
397 * CPU nodes in DT. We need to just ignore this case.
398 * (3) -1 if the node does not exist in the device tree
400 static int __init get_cpu_for_node(struct device_node *node)
402 struct device_node *cpu_node;
405 cpu_node = of_parse_phandle(node, "cpu", 0);
409 cpu = of_cpu_node_to_id(cpu_node);
411 topology_parse_cpu_capacity(cpu_node, cpu);
413 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
414 cpu_node, cpumask_pr_args(cpu_possible_mask));
416 of_node_put(cpu_node);
420 static int __init parse_core(struct device_node *core, int package_id,
427 struct device_node *t;
430 snprintf(name, sizeof(name), "thread%d", i);
431 t = of_get_child_by_name(core, name);
434 cpu = get_cpu_for_node(t);
436 cpu_topology[cpu].package_id = package_id;
437 cpu_topology[cpu].core_id = core_id;
438 cpu_topology[cpu].thread_id = i;
439 } else if (cpu != -ENODEV) {
440 pr_err("%pOF: Can't get CPU for thread\n", t);
449 cpu = get_cpu_for_node(core);
452 pr_err("%pOF: Core has both threads and CPU\n",
457 cpu_topology[cpu].package_id = package_id;
458 cpu_topology[cpu].core_id = core_id;
459 } else if (leaf && cpu != -ENODEV) {
460 pr_err("%pOF: Can't get CPU for leaf core\n", core);
467 static int __init parse_cluster(struct device_node *cluster, int depth)
471 bool has_cores = false;
472 struct device_node *c;
473 static int package_id __initdata;
478 * First check for child clusters; we currently ignore any
479 * information about the nesting of clusters and present the
480 * scheduler with a flat list of them.
484 snprintf(name, sizeof(name), "cluster%d", i);
485 c = of_get_child_by_name(cluster, name);
488 ret = parse_cluster(c, depth + 1);
496 /* Now check for cores */
499 snprintf(name, sizeof(name), "core%d", i);
500 c = of_get_child_by_name(cluster, name);
505 pr_err("%pOF: cpu-map children should be clusters\n",
512 ret = parse_core(c, package_id, core_id++);
514 pr_err("%pOF: Non-leaf cluster with core %s\n",
526 if (leaf && !has_cores)
527 pr_warn("%pOF: empty cluster\n", cluster);
535 static int __init parse_dt_topology(void)
537 struct device_node *cn, *map;
541 cn = of_find_node_by_path("/cpus");
543 pr_err("No CPU information found in DT\n");
548 * When topology is provided cpu-map is essentially a root
549 * cluster with restricted subnodes.
551 map = of_get_child_by_name(cn, "cpu-map");
555 ret = parse_cluster(map, 0);
559 topology_normalize_cpu_scale();
562 * Check that all cores are in the topology; the SMP code will
563 * only mark cores described in the DT as possible.
565 for_each_possible_cpu(cpu)
566 if (cpu_topology[cpu].package_id == -1)
580 struct cpu_topology cpu_topology[NR_CPUS];
581 EXPORT_SYMBOL_GPL(cpu_topology);
583 const struct cpumask *cpu_coregroup_mask(int cpu)
585 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
587 /* Find the smaller of NUMA, core or LLC siblings */
588 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
589 /* not numa in package, lets use the package siblings */
590 core_mask = &cpu_topology[cpu].core_sibling;
592 if (cpu_topology[cpu].llc_id != -1) {
593 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
594 core_mask = &cpu_topology[cpu].llc_sibling;
600 const struct cpumask *cpu_clustergroup_mask(int cpu)
602 return &cpu_topology[cpu].cluster_sibling;
605 void update_siblings_masks(unsigned int cpuid)
607 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
610 /* update core and thread sibling masks */
611 for_each_online_cpu(cpu) {
612 cpu_topo = &cpu_topology[cpu];
614 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
615 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
616 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
619 if (cpuid_topo->package_id != cpu_topo->package_id)
622 if (cpuid_topo->cluster_id == cpu_topo->cluster_id &&
623 cpuid_topo->cluster_id != -1) {
624 cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling);
625 cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling);
628 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
629 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
631 if (cpuid_topo->core_id != cpu_topo->core_id)
634 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
635 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
639 static void clear_cpu_topology(int cpu)
641 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
643 cpumask_clear(&cpu_topo->llc_sibling);
644 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
646 cpumask_clear(&cpu_topo->cluster_sibling);
647 cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling);
649 cpumask_clear(&cpu_topo->core_sibling);
650 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
651 cpumask_clear(&cpu_topo->thread_sibling);
652 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
655 void __init reset_cpu_topology(void)
659 for_each_possible_cpu(cpu) {
660 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
662 cpu_topo->thread_id = -1;
663 cpu_topo->core_id = -1;
664 cpu_topo->cluster_id = -1;
665 cpu_topo->package_id = -1;
666 cpu_topo->llc_id = -1;
668 clear_cpu_topology(cpu);
672 void remove_cpu_topology(unsigned int cpu)
676 for_each_cpu(sibling, topology_core_cpumask(cpu))
677 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
678 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
679 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
680 for_each_cpu(sibling, topology_cluster_cpumask(cpu))
681 cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling));
682 for_each_cpu(sibling, topology_llc_cpumask(cpu))
683 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
685 clear_cpu_topology(cpu);
688 __weak int __init parse_acpi_topology(void)
693 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
694 void __init init_cpu_topology(void)
696 reset_cpu_topology();
699 * Discard anything that was parsed if we hit an error so we
700 * don't use partial information.
702 if (parse_acpi_topology())
703 reset_cpu_topology();
704 else if (of_have_populated_dt() && parse_dt_topology())
705 reset_cpu_topology();