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 #define CREATE_TRACE_POINTS
23 #include <trace/events/thermal_pressure.h>
25 static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
26 static struct cpumask scale_freq_counters_mask;
27 static bool scale_freq_invariant;
28 static DEFINE_PER_CPU(u32, freq_factor) = 1;
30 static bool supports_scale_freq_counters(const struct cpumask *cpus)
32 return cpumask_subset(cpus, &scale_freq_counters_mask);
35 bool topology_scale_freq_invariant(void)
37 return cpufreq_supports_freq_invariance() ||
38 supports_scale_freq_counters(cpu_online_mask);
41 static void update_scale_freq_invariant(bool status)
43 if (scale_freq_invariant == status)
47 * Task scheduler behavior depends on frequency invariance support,
48 * either cpufreq or counter driven. If the support status changes as
49 * a result of counter initialisation and use, retrigger the build of
50 * scheduling domains to ensure the information is propagated properly.
52 if (topology_scale_freq_invariant() == status) {
53 scale_freq_invariant = status;
54 rebuild_sched_domains_energy();
58 void topology_set_scale_freq_source(struct scale_freq_data *data,
59 const struct cpumask *cpus)
61 struct scale_freq_data *sfd;
65 * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
66 * supported by cpufreq.
68 if (cpumask_empty(&scale_freq_counters_mask))
69 scale_freq_invariant = topology_scale_freq_invariant();
73 for_each_cpu(cpu, cpus) {
74 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
76 /* Use ARCH provided counters whenever possible */
77 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
78 rcu_assign_pointer(per_cpu(sft_data, cpu), data);
79 cpumask_set_cpu(cpu, &scale_freq_counters_mask);
85 update_scale_freq_invariant(true);
87 EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
89 void topology_clear_scale_freq_source(enum scale_freq_source source,
90 const struct cpumask *cpus)
92 struct scale_freq_data *sfd;
97 for_each_cpu(cpu, cpus) {
98 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
100 if (sfd && sfd->source == source) {
101 rcu_assign_pointer(per_cpu(sft_data, cpu), NULL);
102 cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
109 * Make sure all references to previous sft_data are dropped to avoid
110 * use-after-free races.
114 update_scale_freq_invariant(false);
116 EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
118 void topology_scale_freq_tick(void)
120 struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data));
123 sfd->set_freq_scale();
126 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
127 EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
129 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
130 unsigned long max_freq)
135 if (WARN_ON_ONCE(!cur_freq || !max_freq))
139 * If the use of counters for FIE is enabled, just return as we don't
140 * want to update the scale factor with information from CPUFREQ.
141 * Instead the scale factor will be updated from arch_scale_freq_tick.
143 if (supports_scale_freq_counters(cpus))
146 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
148 for_each_cpu(i, cpus)
149 per_cpu(arch_freq_scale, i) = scale;
152 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
153 EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale);
155 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
157 per_cpu(cpu_scale, cpu) = capacity;
160 DEFINE_PER_CPU(unsigned long, thermal_pressure);
163 * topology_update_thermal_pressure() - Update thermal pressure for CPUs
164 * @cpus : The related CPUs for which capacity has been reduced
165 * @capped_freq : The maximum allowed frequency that CPUs can run at
167 * Update the value of thermal pressure for all @cpus in the mask. The
168 * cpumask should include all (online+offline) affected CPUs, to avoid
169 * operating on stale data when hot-plug is used for some CPUs. The
170 * @capped_freq reflects the currently allowed max CPUs frequency due to
171 * thermal capping. It might be also a boost frequency value, which is bigger
172 * than the internal 'freq_factor' max frequency. In such case the pressure
173 * value should simply be removed, since this is an indication that there is
174 * no thermal throttling. The @capped_freq must be provided in kHz.
176 void topology_update_thermal_pressure(const struct cpumask *cpus,
177 unsigned long capped_freq)
179 unsigned long max_capacity, capacity, th_pressure;
183 cpu = cpumask_first(cpus);
184 max_capacity = arch_scale_cpu_capacity(cpu);
185 max_freq = per_cpu(freq_factor, cpu);
187 /* Convert to MHz scale which is used in 'freq_factor' */
191 * Handle properly the boost frequencies, which should simply clean
192 * the thermal pressure value.
194 if (max_freq <= capped_freq)
195 capacity = max_capacity;
197 capacity = mult_frac(max_capacity, capped_freq, max_freq);
199 th_pressure = max_capacity - capacity;
201 trace_thermal_pressure_update(cpu, th_pressure);
203 for_each_cpu(cpu, cpus)
204 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
206 EXPORT_SYMBOL_GPL(topology_update_thermal_pressure);
208 static ssize_t cpu_capacity_show(struct device *dev,
209 struct device_attribute *attr,
212 struct cpu *cpu = container_of(dev, struct cpu, dev);
214 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
217 static void update_topology_flags_workfn(struct work_struct *work);
218 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
220 static DEVICE_ATTR_RO(cpu_capacity);
222 static int register_cpu_capacity_sysctl(void)
227 for_each_possible_cpu(i) {
228 cpu = get_cpu_device(i);
230 pr_err("%s: too early to get CPU%d device!\n",
234 device_create_file(cpu, &dev_attr_cpu_capacity);
239 subsys_initcall(register_cpu_capacity_sysctl);
241 static int update_topology;
243 int topology_update_cpu_topology(void)
245 return update_topology;
249 * Updating the sched_domains can't be done directly from cpufreq callbacks
250 * due to locking, so queue the work for later.
252 static void update_topology_flags_workfn(struct work_struct *work)
255 rebuild_sched_domains();
256 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
260 static u32 *raw_capacity;
262 static int free_raw_capacity(void)
270 void topology_normalize_cpu_scale(void)
280 for_each_possible_cpu(cpu) {
281 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
282 capacity_scale = max(capacity, capacity_scale);
285 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
286 for_each_possible_cpu(cpu) {
287 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
288 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
290 topology_set_cpu_scale(cpu, capacity);
291 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
292 cpu, topology_get_cpu_scale(cpu));
296 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
299 static bool cap_parsing_failed;
303 if (cap_parsing_failed)
306 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
310 raw_capacity = kcalloc(num_possible_cpus(),
311 sizeof(*raw_capacity),
314 cap_parsing_failed = true;
318 raw_capacity[cpu] = cpu_capacity;
319 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
320 cpu_node, raw_capacity[cpu]);
323 * Update freq_factor for calculating early boot cpu capacities.
324 * For non-clk CPU DVFS mechanism, there's no way to get the
325 * frequency value now, assuming they are running at the same
326 * frequency (by keeping the initial freq_factor value).
328 cpu_clk = of_clk_get(cpu_node, 0);
329 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
330 per_cpu(freq_factor, cpu) =
331 clk_get_rate(cpu_clk) / 1000;
336 pr_err("cpu_capacity: missing %pOF raw capacity\n",
338 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
340 cap_parsing_failed = true;
347 #ifdef CONFIG_ACPI_CPPC_LIB
348 #include <acpi/cppc_acpi.h>
350 void topology_init_cpu_capacity_cppc(void)
352 struct cppc_perf_caps perf_caps;
355 if (likely(acpi_disabled || !acpi_cpc_valid()))
358 raw_capacity = kcalloc(num_possible_cpus(), sizeof(*raw_capacity),
363 for_each_possible_cpu(cpu) {
364 if (!cppc_get_perf_caps(cpu, &perf_caps) &&
365 (perf_caps.highest_perf >= perf_caps.nominal_perf) &&
366 (perf_caps.highest_perf >= perf_caps.lowest_perf)) {
367 raw_capacity[cpu] = perf_caps.highest_perf;
368 pr_debug("cpu_capacity: CPU%d cpu_capacity=%u (raw).\n",
369 cpu, raw_capacity[cpu]);
373 pr_err("cpu_capacity: CPU%d missing/invalid highest performance.\n", cpu);
374 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
378 topology_normalize_cpu_scale();
379 schedule_work(&update_topology_flags_work);
380 pr_debug("cpu_capacity: cpu_capacity initialization done\n");
387 #ifdef CONFIG_CPU_FREQ
388 static cpumask_var_t cpus_to_visit;
389 static void parsing_done_workfn(struct work_struct *work);
390 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
393 init_cpu_capacity_callback(struct notifier_block *nb,
397 struct cpufreq_policy *policy = data;
403 if (val != CPUFREQ_CREATE_POLICY)
406 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
407 cpumask_pr_args(policy->related_cpus),
408 cpumask_pr_args(cpus_to_visit));
410 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
412 for_each_cpu(cpu, policy->related_cpus)
413 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
415 if (cpumask_empty(cpus_to_visit)) {
416 topology_normalize_cpu_scale();
417 schedule_work(&update_topology_flags_work);
419 pr_debug("cpu_capacity: parsing done\n");
420 schedule_work(&parsing_done_work);
426 static struct notifier_block init_cpu_capacity_notifier = {
427 .notifier_call = init_cpu_capacity_callback,
430 static int __init register_cpufreq_notifier(void)
435 * On ACPI-based systems skip registering cpufreq notifier as cpufreq
436 * information is not needed for cpu capacity initialization.
438 if (!acpi_disabled || !raw_capacity)
441 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
444 cpumask_copy(cpus_to_visit, cpu_possible_mask);
446 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
447 CPUFREQ_POLICY_NOTIFIER);
450 free_cpumask_var(cpus_to_visit);
454 core_initcall(register_cpufreq_notifier);
456 static void parsing_done_workfn(struct work_struct *work)
458 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
459 CPUFREQ_POLICY_NOTIFIER);
460 free_cpumask_var(cpus_to_visit);
464 core_initcall(free_raw_capacity);
467 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
469 * This function returns the logic cpu number of the node.
470 * There are basically three kinds of return values:
471 * (1) logic cpu number which is > 0.
472 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
473 * there is no possible logical CPU in the kernel to match. This happens
474 * when CONFIG_NR_CPUS is configure to be smaller than the number of
475 * CPU nodes in DT. We need to just ignore this case.
476 * (3) -1 if the node does not exist in the device tree
478 static int __init get_cpu_for_node(struct device_node *node)
480 struct device_node *cpu_node;
483 cpu_node = of_parse_phandle(node, "cpu", 0);
487 cpu = of_cpu_node_to_id(cpu_node);
489 topology_parse_cpu_capacity(cpu_node, cpu);
491 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
492 cpu_node, cpumask_pr_args(cpu_possible_mask));
494 of_node_put(cpu_node);
498 static int __init parse_core(struct device_node *core, int package_id,
505 struct device_node *t;
508 snprintf(name, sizeof(name), "thread%d", i);
509 t = of_get_child_by_name(core, name);
512 cpu = get_cpu_for_node(t);
514 cpu_topology[cpu].package_id = package_id;
515 cpu_topology[cpu].core_id = core_id;
516 cpu_topology[cpu].thread_id = i;
517 } else if (cpu != -ENODEV) {
518 pr_err("%pOF: Can't get CPU for thread\n", t);
527 cpu = get_cpu_for_node(core);
530 pr_err("%pOF: Core has both threads and CPU\n",
535 cpu_topology[cpu].package_id = package_id;
536 cpu_topology[cpu].core_id = core_id;
537 } else if (leaf && cpu != -ENODEV) {
538 pr_err("%pOF: Can't get CPU for leaf core\n", core);
545 static int __init parse_cluster(struct device_node *cluster, int depth)
549 bool has_cores = false;
550 struct device_node *c;
551 static int package_id __initdata;
556 * First check for child clusters; we currently ignore any
557 * information about the nesting of clusters and present the
558 * scheduler with a flat list of them.
562 snprintf(name, sizeof(name), "cluster%d", i);
563 c = of_get_child_by_name(cluster, name);
566 ret = parse_cluster(c, depth + 1);
574 /* Now check for cores */
577 snprintf(name, sizeof(name), "core%d", i);
578 c = of_get_child_by_name(cluster, name);
583 pr_err("%pOF: cpu-map children should be clusters\n",
590 ret = parse_core(c, package_id, core_id++);
592 pr_err("%pOF: Non-leaf cluster with core %s\n",
604 if (leaf && !has_cores)
605 pr_warn("%pOF: empty cluster\n", cluster);
613 static int __init parse_dt_topology(void)
615 struct device_node *cn, *map;
619 cn = of_find_node_by_path("/cpus");
621 pr_err("No CPU information found in DT\n");
626 * When topology is provided cpu-map is essentially a root
627 * cluster with restricted subnodes.
629 map = of_get_child_by_name(cn, "cpu-map");
633 ret = parse_cluster(map, 0);
637 topology_normalize_cpu_scale();
640 * Check that all cores are in the topology; the SMP code will
641 * only mark cores described in the DT as possible.
643 for_each_possible_cpu(cpu)
644 if (cpu_topology[cpu].package_id == -1)
658 struct cpu_topology cpu_topology[NR_CPUS];
659 EXPORT_SYMBOL_GPL(cpu_topology);
661 const struct cpumask *cpu_coregroup_mask(int cpu)
663 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
665 /* Find the smaller of NUMA, core or LLC siblings */
666 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
667 /* not numa in package, lets use the package siblings */
668 core_mask = &cpu_topology[cpu].core_sibling;
670 if (cpu_topology[cpu].llc_id != -1) {
671 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
672 core_mask = &cpu_topology[cpu].llc_sibling;
676 * For systems with no shared cpu-side LLC but with clusters defined,
677 * extend core_mask to cluster_siblings. The sched domain builder will
678 * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled.
680 if (IS_ENABLED(CONFIG_SCHED_CLUSTER) &&
681 cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling))
682 core_mask = &cpu_topology[cpu].cluster_sibling;
687 const struct cpumask *cpu_clustergroup_mask(int cpu)
689 return &cpu_topology[cpu].cluster_sibling;
692 void update_siblings_masks(unsigned int cpuid)
694 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
697 /* update core and thread sibling masks */
698 for_each_online_cpu(cpu) {
699 cpu_topo = &cpu_topology[cpu];
701 if (cpu_topo->llc_id != -1 && cpuid_topo->llc_id == cpu_topo->llc_id) {
702 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
703 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
706 if (cpuid_topo->package_id != cpu_topo->package_id)
709 if (cpuid_topo->cluster_id == cpu_topo->cluster_id &&
710 cpuid_topo->cluster_id != -1) {
711 cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling);
712 cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling);
715 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
716 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
718 if (cpuid_topo->core_id != cpu_topo->core_id)
721 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
722 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
726 static void clear_cpu_topology(int cpu)
728 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
730 cpumask_clear(&cpu_topo->llc_sibling);
731 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
733 cpumask_clear(&cpu_topo->cluster_sibling);
734 cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling);
736 cpumask_clear(&cpu_topo->core_sibling);
737 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
738 cpumask_clear(&cpu_topo->thread_sibling);
739 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
742 void __init reset_cpu_topology(void)
746 for_each_possible_cpu(cpu) {
747 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
749 cpu_topo->thread_id = -1;
750 cpu_topo->core_id = -1;
751 cpu_topo->cluster_id = -1;
752 cpu_topo->package_id = -1;
753 cpu_topo->llc_id = -1;
755 clear_cpu_topology(cpu);
759 void remove_cpu_topology(unsigned int cpu)
763 for_each_cpu(sibling, topology_core_cpumask(cpu))
764 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
765 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
766 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
767 for_each_cpu(sibling, topology_cluster_cpumask(cpu))
768 cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling));
769 for_each_cpu(sibling, topology_llc_cpumask(cpu))
770 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
772 clear_cpu_topology(cpu);
775 __weak int __init parse_acpi_topology(void)
780 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
781 void __init init_cpu_topology(void)
783 reset_cpu_topology();
786 * Discard anything that was parsed if we hit an error so we
787 * don't use partial information.
789 if (parse_acpi_topology())
790 reset_cpu_topology();
791 else if (of_have_populated_dt() && parse_dt_topology())
792 reset_cpu_topology();