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 static DEFINE_PER_CPU(struct scale_freq_data *, sft_data);
25 static struct cpumask scale_freq_counters_mask;
26 static bool scale_freq_invariant;
28 static bool supports_scale_freq_counters(const struct cpumask *cpus)
30 return cpumask_subset(cpus, &scale_freq_counters_mask);
33 bool topology_scale_freq_invariant(void)
35 return cpufreq_supports_freq_invariance() ||
36 supports_scale_freq_counters(cpu_online_mask);
39 static void update_scale_freq_invariant(bool status)
41 if (scale_freq_invariant == status)
45 * Task scheduler behavior depends on frequency invariance support,
46 * either cpufreq or counter driven. If the support status changes as
47 * a result of counter initialisation and use, retrigger the build of
48 * scheduling domains to ensure the information is propagated properly.
50 if (topology_scale_freq_invariant() == status) {
51 scale_freq_invariant = status;
52 rebuild_sched_domains_energy();
56 void topology_set_scale_freq_source(struct scale_freq_data *data,
57 const struct cpumask *cpus)
59 struct scale_freq_data *sfd;
63 * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
64 * supported by cpufreq.
66 if (cpumask_empty(&scale_freq_counters_mask))
67 scale_freq_invariant = topology_scale_freq_invariant();
69 for_each_cpu(cpu, cpus) {
70 sfd = per_cpu(sft_data, cpu);
72 /* Use ARCH provided counters whenever possible */
73 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
74 per_cpu(sft_data, cpu) = data;
75 cpumask_set_cpu(cpu, &scale_freq_counters_mask);
79 update_scale_freq_invariant(true);
81 EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
83 void topology_clear_scale_freq_source(enum scale_freq_source source,
84 const struct cpumask *cpus)
86 struct scale_freq_data *sfd;
89 for_each_cpu(cpu, cpus) {
90 sfd = per_cpu(sft_data, cpu);
92 if (sfd && sfd->source == source) {
93 per_cpu(sft_data, cpu) = NULL;
94 cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
98 update_scale_freq_invariant(false);
100 EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
102 void topology_scale_freq_tick(void)
104 struct scale_freq_data *sfd = *this_cpu_ptr(&sft_data);
107 sfd->set_freq_scale();
110 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
111 EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
113 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
114 unsigned long max_freq)
119 if (WARN_ON_ONCE(!cur_freq || !max_freq))
123 * If the use of counters for FIE is enabled, just return as we don't
124 * want to update the scale factor with information from CPUFREQ.
125 * Instead the scale factor will be updated from arch_scale_freq_tick.
127 if (supports_scale_freq_counters(cpus))
130 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
132 for_each_cpu(i, cpus)
133 per_cpu(arch_freq_scale, i) = scale;
136 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
138 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
140 per_cpu(cpu_scale, cpu) = capacity;
143 DEFINE_PER_CPU(unsigned long, thermal_pressure);
145 void topology_set_thermal_pressure(const struct cpumask *cpus,
146 unsigned long th_pressure)
150 for_each_cpu(cpu, cpus)
151 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
154 static ssize_t cpu_capacity_show(struct device *dev,
155 struct device_attribute *attr,
158 struct cpu *cpu = container_of(dev, struct cpu, dev);
160 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
163 static void update_topology_flags_workfn(struct work_struct *work);
164 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
166 static DEVICE_ATTR_RO(cpu_capacity);
168 static int register_cpu_capacity_sysctl(void)
173 for_each_possible_cpu(i) {
174 cpu = get_cpu_device(i);
176 pr_err("%s: too early to get CPU%d device!\n",
180 device_create_file(cpu, &dev_attr_cpu_capacity);
185 subsys_initcall(register_cpu_capacity_sysctl);
187 static int update_topology;
189 int topology_update_cpu_topology(void)
191 return update_topology;
195 * Updating the sched_domains can't be done directly from cpufreq callbacks
196 * due to locking, so queue the work for later.
198 static void update_topology_flags_workfn(struct work_struct *work)
201 rebuild_sched_domains();
202 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
206 static DEFINE_PER_CPU(u32, freq_factor) = 1;
207 static u32 *raw_capacity;
209 static int free_raw_capacity(void)
217 void topology_normalize_cpu_scale(void)
227 for_each_possible_cpu(cpu) {
228 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
229 capacity_scale = max(capacity, capacity_scale);
232 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
233 for_each_possible_cpu(cpu) {
234 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
235 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
237 topology_set_cpu_scale(cpu, capacity);
238 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
239 cpu, topology_get_cpu_scale(cpu));
243 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
246 static bool cap_parsing_failed;
250 if (cap_parsing_failed)
253 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
257 raw_capacity = kcalloc(num_possible_cpus(),
258 sizeof(*raw_capacity),
261 cap_parsing_failed = true;
265 raw_capacity[cpu] = cpu_capacity;
266 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
267 cpu_node, raw_capacity[cpu]);
270 * Update freq_factor for calculating early boot cpu capacities.
271 * For non-clk CPU DVFS mechanism, there's no way to get the
272 * frequency value now, assuming they are running at the same
273 * frequency (by keeping the initial freq_factor value).
275 cpu_clk = of_clk_get(cpu_node, 0);
276 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
277 per_cpu(freq_factor, cpu) =
278 clk_get_rate(cpu_clk) / 1000;
283 pr_err("cpu_capacity: missing %pOF raw capacity\n",
285 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
287 cap_parsing_failed = true;
294 #ifdef CONFIG_CPU_FREQ
295 static cpumask_var_t cpus_to_visit;
296 static void parsing_done_workfn(struct work_struct *work);
297 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
300 init_cpu_capacity_callback(struct notifier_block *nb,
304 struct cpufreq_policy *policy = data;
310 if (val != CPUFREQ_CREATE_POLICY)
313 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
314 cpumask_pr_args(policy->related_cpus),
315 cpumask_pr_args(cpus_to_visit));
317 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
319 for_each_cpu(cpu, policy->related_cpus)
320 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
322 if (cpumask_empty(cpus_to_visit)) {
323 topology_normalize_cpu_scale();
324 schedule_work(&update_topology_flags_work);
326 pr_debug("cpu_capacity: parsing done\n");
327 schedule_work(&parsing_done_work);
333 static struct notifier_block init_cpu_capacity_notifier = {
334 .notifier_call = init_cpu_capacity_callback,
337 static int __init register_cpufreq_notifier(void)
342 * on ACPI-based systems we need to use the default cpu capacity
343 * until we have the necessary code to parse the cpu capacity, so
344 * skip registering cpufreq notifier.
346 if (!acpi_disabled || !raw_capacity)
349 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
352 cpumask_copy(cpus_to_visit, cpu_possible_mask);
354 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
355 CPUFREQ_POLICY_NOTIFIER);
358 free_cpumask_var(cpus_to_visit);
362 core_initcall(register_cpufreq_notifier);
364 static void parsing_done_workfn(struct work_struct *work)
366 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
367 CPUFREQ_POLICY_NOTIFIER);
368 free_cpumask_var(cpus_to_visit);
372 core_initcall(free_raw_capacity);
375 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
377 * This function returns the logic cpu number of the node.
378 * There are basically three kinds of return values:
379 * (1) logic cpu number which is > 0.
380 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
381 * there is no possible logical CPU in the kernel to match. This happens
382 * when CONFIG_NR_CPUS is configure to be smaller than the number of
383 * CPU nodes in DT. We need to just ignore this case.
384 * (3) -1 if the node does not exist in the device tree
386 static int __init get_cpu_for_node(struct device_node *node)
388 struct device_node *cpu_node;
391 cpu_node = of_parse_phandle(node, "cpu", 0);
395 cpu = of_cpu_node_to_id(cpu_node);
397 topology_parse_cpu_capacity(cpu_node, cpu);
399 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
400 cpu_node, cpumask_pr_args(cpu_possible_mask));
402 of_node_put(cpu_node);
406 static int __init parse_core(struct device_node *core, int package_id,
413 struct device_node *t;
416 snprintf(name, sizeof(name), "thread%d", i);
417 t = of_get_child_by_name(core, name);
420 cpu = get_cpu_for_node(t);
422 cpu_topology[cpu].package_id = package_id;
423 cpu_topology[cpu].core_id = core_id;
424 cpu_topology[cpu].thread_id = i;
425 } else if (cpu != -ENODEV) {
426 pr_err("%pOF: Can't get CPU for thread\n", t);
435 cpu = get_cpu_for_node(core);
438 pr_err("%pOF: Core has both threads and CPU\n",
443 cpu_topology[cpu].package_id = package_id;
444 cpu_topology[cpu].core_id = core_id;
445 } else if (leaf && cpu != -ENODEV) {
446 pr_err("%pOF: Can't get CPU for leaf core\n", core);
453 static int __init parse_cluster(struct device_node *cluster, int depth)
457 bool has_cores = false;
458 struct device_node *c;
459 static int package_id __initdata;
464 * First check for child clusters; we currently ignore any
465 * information about the nesting of clusters and present the
466 * scheduler with a flat list of them.
470 snprintf(name, sizeof(name), "cluster%d", i);
471 c = of_get_child_by_name(cluster, name);
474 ret = parse_cluster(c, depth + 1);
482 /* Now check for cores */
485 snprintf(name, sizeof(name), "core%d", i);
486 c = of_get_child_by_name(cluster, name);
491 pr_err("%pOF: cpu-map children should be clusters\n",
498 ret = parse_core(c, package_id, core_id++);
500 pr_err("%pOF: Non-leaf cluster with core %s\n",
512 if (leaf && !has_cores)
513 pr_warn("%pOF: empty cluster\n", cluster);
521 static int __init parse_dt_topology(void)
523 struct device_node *cn, *map;
527 cn = of_find_node_by_path("/cpus");
529 pr_err("No CPU information found in DT\n");
534 * When topology is provided cpu-map is essentially a root
535 * cluster with restricted subnodes.
537 map = of_get_child_by_name(cn, "cpu-map");
541 ret = parse_cluster(map, 0);
545 topology_normalize_cpu_scale();
548 * Check that all cores are in the topology; the SMP code will
549 * only mark cores described in the DT as possible.
551 for_each_possible_cpu(cpu)
552 if (cpu_topology[cpu].package_id == -1)
566 struct cpu_topology cpu_topology[NR_CPUS];
567 EXPORT_SYMBOL_GPL(cpu_topology);
569 const struct cpumask *cpu_coregroup_mask(int cpu)
571 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
573 /* Find the smaller of NUMA, core or LLC siblings */
574 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
575 /* not numa in package, lets use the package siblings */
576 core_mask = &cpu_topology[cpu].core_sibling;
578 if (cpu_topology[cpu].llc_id != -1) {
579 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
580 core_mask = &cpu_topology[cpu].llc_sibling;
586 void update_siblings_masks(unsigned int cpuid)
588 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
591 /* update core and thread sibling masks */
592 for_each_online_cpu(cpu) {
593 cpu_topo = &cpu_topology[cpu];
595 if (cpuid_topo->llc_id == cpu_topo->llc_id) {
596 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
597 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
600 if (cpuid_topo->package_id != cpu_topo->package_id)
603 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
604 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
606 if (cpuid_topo->core_id != cpu_topo->core_id)
609 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
610 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
614 static void clear_cpu_topology(int cpu)
616 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
618 cpumask_clear(&cpu_topo->llc_sibling);
619 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
621 cpumask_clear(&cpu_topo->core_sibling);
622 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
623 cpumask_clear(&cpu_topo->thread_sibling);
624 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
627 void __init reset_cpu_topology(void)
631 for_each_possible_cpu(cpu) {
632 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
634 cpu_topo->thread_id = -1;
635 cpu_topo->core_id = -1;
636 cpu_topo->package_id = -1;
637 cpu_topo->llc_id = -1;
639 clear_cpu_topology(cpu);
643 void remove_cpu_topology(unsigned int cpu)
647 for_each_cpu(sibling, topology_core_cpumask(cpu))
648 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
649 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
650 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
651 for_each_cpu(sibling, topology_llc_cpumask(cpu))
652 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
654 clear_cpu_topology(cpu);
657 __weak int __init parse_acpi_topology(void)
662 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
663 void __init init_cpu_topology(void)
665 reset_cpu_topology();
668 * Discard anything that was parsed if we hit an error so we
669 * don't use partial information.
671 if (parse_acpi_topology())
672 reset_cpu_topology();
673 else if (of_have_populated_dt() && parse_dt_topology())
674 reset_cpu_topology();