Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Arch specific cpu topology information
4 *
5 * Copyright (C) 2016, ARM Ltd.
6 * Written by: Juri Lelli, ARM Ltd.
7 */
8
9#include <linux/acpi.h>
10#include <linux/cacheinfo.h>
11#include <linux/cpu.h>
12#include <linux/cpufreq.h>
13#include <linux/device.h>
14#include <linux/of.h>
15#include <linux/slab.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/rcupdate.h>
21#include <linux/sched.h>
22
23#define CREATE_TRACE_POINTS
24#include <trace/events/thermal_pressure.h>
25
26static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
27static struct cpumask scale_freq_counters_mask;
28static bool scale_freq_invariant;
29static DEFINE_PER_CPU(u32, freq_factor) = 1;
30
31static bool supports_scale_freq_counters(const struct cpumask *cpus)
32{
33 return cpumask_subset(cpus, &scale_freq_counters_mask);
34}
35
36bool topology_scale_freq_invariant(void)
37{
38 return cpufreq_supports_freq_invariance() ||
39 supports_scale_freq_counters(cpu_online_mask);
40}
41
42static void update_scale_freq_invariant(bool status)
43{
44 if (scale_freq_invariant == status)
45 return;
46
47 /*
48 * Task scheduler behavior depends on frequency invariance support,
49 * either cpufreq or counter driven. If the support status changes as
50 * a result of counter initialisation and use, retrigger the build of
51 * scheduling domains to ensure the information is propagated properly.
52 */
53 if (topology_scale_freq_invariant() == status) {
54 scale_freq_invariant = status;
55 rebuild_sched_domains_energy();
56 }
57}
58
59void topology_set_scale_freq_source(struct scale_freq_data *data,
60 const struct cpumask *cpus)
61{
62 struct scale_freq_data *sfd;
63 int cpu;
64
65 /*
66 * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
67 * supported by cpufreq.
68 */
69 if (cpumask_empty(&scale_freq_counters_mask))
70 scale_freq_invariant = topology_scale_freq_invariant();
71
72 rcu_read_lock();
73
74 for_each_cpu(cpu, cpus) {
75 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
76
77 /* Use ARCH provided counters whenever possible */
78 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
79 rcu_assign_pointer(per_cpu(sft_data, cpu), data);
80 cpumask_set_cpu(cpu, &scale_freq_counters_mask);
81 }
82 }
83
84 rcu_read_unlock();
85
86 update_scale_freq_invariant(true);
87}
88EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
89
90void topology_clear_scale_freq_source(enum scale_freq_source source,
91 const struct cpumask *cpus)
92{
93 struct scale_freq_data *sfd;
94 int cpu;
95
96 rcu_read_lock();
97
98 for_each_cpu(cpu, cpus) {
99 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
100
101 if (sfd && sfd->source == source) {
102 rcu_assign_pointer(per_cpu(sft_data, cpu), NULL);
103 cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
104 }
105 }
106
107 rcu_read_unlock();
108
109 /*
110 * Make sure all references to previous sft_data are dropped to avoid
111 * use-after-free races.
112 */
113 synchronize_rcu();
114
115 update_scale_freq_invariant(false);
116}
117EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
118
119void topology_scale_freq_tick(void)
120{
121 struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data));
122
123 if (sfd)
124 sfd->set_freq_scale();
125}
126
127DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
128EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
129
130void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
131 unsigned long max_freq)
132{
133 unsigned long scale;
134 int i;
135
136 if (WARN_ON_ONCE(!cur_freq || !max_freq))
137 return;
138
139 /*
140 * If the use of counters for FIE is enabled, just return as we don't
141 * want to update the scale factor with information from CPUFREQ.
142 * Instead the scale factor will be updated from arch_scale_freq_tick.
143 */
144 if (supports_scale_freq_counters(cpus))
145 return;
146
147 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
148
149 for_each_cpu(i, cpus)
150 per_cpu(arch_freq_scale, i) = scale;
151}
152
153DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
154EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale);
155
156void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
157{
158 per_cpu(cpu_scale, cpu) = capacity;
159}
160
161DEFINE_PER_CPU(unsigned long, thermal_pressure);
162
163/**
164 * topology_update_thermal_pressure() - Update thermal pressure for CPUs
165 * @cpus : The related CPUs for which capacity has been reduced
166 * @capped_freq : The maximum allowed frequency that CPUs can run at
167 *
168 * Update the value of thermal pressure for all @cpus in the mask. The
169 * cpumask should include all (online+offline) affected CPUs, to avoid
170 * operating on stale data when hot-plug is used for some CPUs. The
171 * @capped_freq reflects the currently allowed max CPUs frequency due to
172 * thermal capping. It might be also a boost frequency value, which is bigger
173 * than the internal 'freq_factor' max frequency. In such case the pressure
174 * value should simply be removed, since this is an indication that there is
175 * no thermal throttling. The @capped_freq must be provided in kHz.
176 */
177void topology_update_thermal_pressure(const struct cpumask *cpus,
178 unsigned long capped_freq)
179{
180 unsigned long max_capacity, capacity, th_pressure;
181 u32 max_freq;
182 int cpu;
183
184 cpu = cpumask_first(cpus);
185 max_capacity = arch_scale_cpu_capacity(cpu);
186 max_freq = per_cpu(freq_factor, cpu);
187
188 /* Convert to MHz scale which is used in 'freq_factor' */
189 capped_freq /= 1000;
190
191 /*
192 * Handle properly the boost frequencies, which should simply clean
193 * the thermal pressure value.
194 */
195 if (max_freq <= capped_freq)
196 capacity = max_capacity;
197 else
198 capacity = mult_frac(max_capacity, capped_freq, max_freq);
199
200 th_pressure = max_capacity - capacity;
201
202 trace_thermal_pressure_update(cpu, th_pressure);
203
204 for_each_cpu(cpu, cpus)
205 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
206}
207EXPORT_SYMBOL_GPL(topology_update_thermal_pressure);
208
209static ssize_t cpu_capacity_show(struct device *dev,
210 struct device_attribute *attr,
211 char *buf)
212{
213 struct cpu *cpu = container_of(dev, struct cpu, dev);
214
215 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
216}
217
218static void update_topology_flags_workfn(struct work_struct *work);
219static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
220
221static DEVICE_ATTR_RO(cpu_capacity);
222
223static int register_cpu_capacity_sysctl(void)
224{
225 int i;
226 struct device *cpu;
227
228 for_each_possible_cpu(i) {
229 cpu = get_cpu_device(i);
230 if (!cpu) {
231 pr_err("%s: too early to get CPU%d device!\n",
232 __func__, i);
233 continue;
234 }
235 device_create_file(cpu, &dev_attr_cpu_capacity);
236 }
237
238 return 0;
239}
240subsys_initcall(register_cpu_capacity_sysctl);
241
242static int update_topology;
243
244int topology_update_cpu_topology(void)
245{
246 return update_topology;
247}
248
249/*
250 * Updating the sched_domains can't be done directly from cpufreq callbacks
251 * due to locking, so queue the work for later.
252 */
253static void update_topology_flags_workfn(struct work_struct *work)
254{
255 update_topology = 1;
256 rebuild_sched_domains();
257 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
258 update_topology = 0;
259}
260
261static u32 *raw_capacity;
262
263static int free_raw_capacity(void)
264{
265 kfree(raw_capacity);
266 raw_capacity = NULL;
267
268 return 0;
269}
270
271void topology_normalize_cpu_scale(void)
272{
273 u64 capacity;
274 u64 capacity_scale;
275 int cpu;
276
277 if (!raw_capacity)
278 return;
279
280 capacity_scale = 1;
281 for_each_possible_cpu(cpu) {
282 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
283 capacity_scale = max(capacity, capacity_scale);
284 }
285
286 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
287 for_each_possible_cpu(cpu) {
288 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
289 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
290 capacity_scale);
291 topology_set_cpu_scale(cpu, capacity);
292 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
293 cpu, topology_get_cpu_scale(cpu));
294 }
295}
296
297bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
298{
299 struct clk *cpu_clk;
300 static bool cap_parsing_failed;
301 int ret;
302 u32 cpu_capacity;
303
304 if (cap_parsing_failed)
305 return false;
306
307 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
308 &cpu_capacity);
309 if (!ret) {
310 if (!raw_capacity) {
311 raw_capacity = kcalloc(num_possible_cpus(),
312 sizeof(*raw_capacity),
313 GFP_KERNEL);
314 if (!raw_capacity) {
315 cap_parsing_failed = true;
316 return false;
317 }
318 }
319 raw_capacity[cpu] = cpu_capacity;
320 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
321 cpu_node, raw_capacity[cpu]);
322
323 /*
324 * Update freq_factor for calculating early boot cpu capacities.
325 * For non-clk CPU DVFS mechanism, there's no way to get the
326 * frequency value now, assuming they are running at the same
327 * frequency (by keeping the initial freq_factor value).
328 */
329 cpu_clk = of_clk_get(cpu_node, 0);
330 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
331 per_cpu(freq_factor, cpu) =
332 clk_get_rate(cpu_clk) / 1000;
333 clk_put(cpu_clk);
334 }
335 } else {
336 if (raw_capacity) {
337 pr_err("cpu_capacity: missing %pOF raw capacity\n",
338 cpu_node);
339 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
340 }
341 cap_parsing_failed = true;
342 free_raw_capacity();
343 }
344
345 return !ret;
346}
347
348#ifdef CONFIG_ACPI_CPPC_LIB
349#include <acpi/cppc_acpi.h>
350
351void topology_init_cpu_capacity_cppc(void)
352{
353 struct cppc_perf_caps perf_caps;
354 int cpu;
355
356 if (likely(!acpi_cpc_valid()))
357 return;
358
359 raw_capacity = kcalloc(num_possible_cpus(), sizeof(*raw_capacity),
360 GFP_KERNEL);
361 if (!raw_capacity)
362 return;
363
364 for_each_possible_cpu(cpu) {
365 if (!cppc_get_perf_caps(cpu, &perf_caps) &&
366 (perf_caps.highest_perf >= perf_caps.nominal_perf) &&
367 (perf_caps.highest_perf >= perf_caps.lowest_perf)) {
368 raw_capacity[cpu] = perf_caps.highest_perf;
369 pr_debug("cpu_capacity: CPU%d cpu_capacity=%u (raw).\n",
370 cpu, raw_capacity[cpu]);
371 continue;
372 }
373
374 pr_err("cpu_capacity: CPU%d missing/invalid highest performance.\n", cpu);
375 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
376 goto exit;
377 }
378
379 topology_normalize_cpu_scale();
380 schedule_work(&update_topology_flags_work);
381 pr_debug("cpu_capacity: cpu_capacity initialization done\n");
382
383exit:
384 free_raw_capacity();
385}
386#endif
387
388#ifdef CONFIG_CPU_FREQ
389static cpumask_var_t cpus_to_visit;
390static void parsing_done_workfn(struct work_struct *work);
391static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
392
393static int
394init_cpu_capacity_callback(struct notifier_block *nb,
395 unsigned long val,
396 void *data)
397{
398 struct cpufreq_policy *policy = data;
399 int cpu;
400
401 if (!raw_capacity)
402 return 0;
403
404 if (val != CPUFREQ_CREATE_POLICY)
405 return 0;
406
407 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
408 cpumask_pr_args(policy->related_cpus),
409 cpumask_pr_args(cpus_to_visit));
410
411 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
412
413 for_each_cpu(cpu, policy->related_cpus)
414 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
415
416 if (cpumask_empty(cpus_to_visit)) {
417 topology_normalize_cpu_scale();
418 schedule_work(&update_topology_flags_work);
419 free_raw_capacity();
420 pr_debug("cpu_capacity: parsing done\n");
421 schedule_work(&parsing_done_work);
422 }
423
424 return 0;
425}
426
427static struct notifier_block init_cpu_capacity_notifier = {
428 .notifier_call = init_cpu_capacity_callback,
429};
430
431static int __init register_cpufreq_notifier(void)
432{
433 int ret;
434
435 /*
436 * On ACPI-based systems skip registering cpufreq notifier as cpufreq
437 * information is not needed for cpu capacity initialization.
438 */
439 if (!acpi_disabled || !raw_capacity)
440 return -EINVAL;
441
442 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
443 return -ENOMEM;
444
445 cpumask_copy(cpus_to_visit, cpu_possible_mask);
446
447 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
448 CPUFREQ_POLICY_NOTIFIER);
449
450 if (ret)
451 free_cpumask_var(cpus_to_visit);
452
453 return ret;
454}
455core_initcall(register_cpufreq_notifier);
456
457static void parsing_done_workfn(struct work_struct *work)
458{
459 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
460 CPUFREQ_POLICY_NOTIFIER);
461 free_cpumask_var(cpus_to_visit);
462}
463
464#else
465core_initcall(free_raw_capacity);
466#endif
467
468#if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
469/*
470 * This function returns the logic cpu number of the node.
471 * There are basically three kinds of return values:
472 * (1) logic cpu number which is > 0.
473 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
474 * there is no possible logical CPU in the kernel to match. This happens
475 * when CONFIG_NR_CPUS is configure to be smaller than the number of
476 * CPU nodes in DT. We need to just ignore this case.
477 * (3) -1 if the node does not exist in the device tree
478 */
479static int __init get_cpu_for_node(struct device_node *node)
480{
481 struct device_node *cpu_node;
482 int cpu;
483
484 cpu_node = of_parse_phandle(node, "cpu", 0);
485 if (!cpu_node)
486 return -1;
487
488 cpu = of_cpu_node_to_id(cpu_node);
489 if (cpu >= 0)
490 topology_parse_cpu_capacity(cpu_node, cpu);
491 else
492 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
493 cpu_node, cpumask_pr_args(cpu_possible_mask));
494
495 of_node_put(cpu_node);
496 return cpu;
497}
498
499static int __init parse_core(struct device_node *core, int package_id,
500 int cluster_id, int core_id)
501{
502 char name[20];
503 bool leaf = true;
504 int i = 0;
505 int cpu;
506 struct device_node *t;
507
508 do {
509 snprintf(name, sizeof(name), "thread%d", i);
510 t = of_get_child_by_name(core, name);
511 if (t) {
512 leaf = false;
513 cpu = get_cpu_for_node(t);
514 if (cpu >= 0) {
515 cpu_topology[cpu].package_id = package_id;
516 cpu_topology[cpu].cluster_id = cluster_id;
517 cpu_topology[cpu].core_id = core_id;
518 cpu_topology[cpu].thread_id = i;
519 } else if (cpu != -ENODEV) {
520 pr_err("%pOF: Can't get CPU for thread\n", t);
521 of_node_put(t);
522 return -EINVAL;
523 }
524 of_node_put(t);
525 }
526 i++;
527 } while (t);
528
529 cpu = get_cpu_for_node(core);
530 if (cpu >= 0) {
531 if (!leaf) {
532 pr_err("%pOF: Core has both threads and CPU\n",
533 core);
534 return -EINVAL;
535 }
536
537 cpu_topology[cpu].package_id = package_id;
538 cpu_topology[cpu].cluster_id = cluster_id;
539 cpu_topology[cpu].core_id = core_id;
540 } else if (leaf && cpu != -ENODEV) {
541 pr_err("%pOF: Can't get CPU for leaf core\n", core);
542 return -EINVAL;
543 }
544
545 return 0;
546}
547
548static int __init parse_cluster(struct device_node *cluster, int package_id,
549 int cluster_id, int depth)
550{
551 char name[20];
552 bool leaf = true;
553 bool has_cores = false;
554 struct device_node *c;
555 int core_id = 0;
556 int i, ret;
557
558 /*
559 * First check for child clusters; we currently ignore any
560 * information about the nesting of clusters and present the
561 * scheduler with a flat list of them.
562 */
563 i = 0;
564 do {
565 snprintf(name, sizeof(name), "cluster%d", i);
566 c = of_get_child_by_name(cluster, name);
567 if (c) {
568 leaf = false;
569 ret = parse_cluster(c, package_id, i, depth + 1);
570 if (depth > 0)
571 pr_warn("Topology for clusters of clusters not yet supported\n");
572 of_node_put(c);
573 if (ret != 0)
574 return ret;
575 }
576 i++;
577 } while (c);
578
579 /* Now check for cores */
580 i = 0;
581 do {
582 snprintf(name, sizeof(name), "core%d", i);
583 c = of_get_child_by_name(cluster, name);
584 if (c) {
585 has_cores = true;
586
587 if (depth == 0) {
588 pr_err("%pOF: cpu-map children should be clusters\n",
589 c);
590 of_node_put(c);
591 return -EINVAL;
592 }
593
594 if (leaf) {
595 ret = parse_core(c, package_id, cluster_id,
596 core_id++);
597 } else {
598 pr_err("%pOF: Non-leaf cluster with core %s\n",
599 cluster, name);
600 ret = -EINVAL;
601 }
602
603 of_node_put(c);
604 if (ret != 0)
605 return ret;
606 }
607 i++;
608 } while (c);
609
610 if (leaf && !has_cores)
611 pr_warn("%pOF: empty cluster\n", cluster);
612
613 return 0;
614}
615
616static int __init parse_socket(struct device_node *socket)
617{
618 char name[20];
619 struct device_node *c;
620 bool has_socket = false;
621 int package_id = 0, ret;
622
623 do {
624 snprintf(name, sizeof(name), "socket%d", package_id);
625 c = of_get_child_by_name(socket, name);
626 if (c) {
627 has_socket = true;
628 ret = parse_cluster(c, package_id, -1, 0);
629 of_node_put(c);
630 if (ret != 0)
631 return ret;
632 }
633 package_id++;
634 } while (c);
635
636 if (!has_socket)
637 ret = parse_cluster(socket, 0, -1, 0);
638
639 return ret;
640}
641
642static int __init parse_dt_topology(void)
643{
644 struct device_node *cn, *map;
645 int ret = 0;
646 int cpu;
647
648 cn = of_find_node_by_path("/cpus");
649 if (!cn) {
650 pr_err("No CPU information found in DT\n");
651 return 0;
652 }
653
654 /*
655 * When topology is provided cpu-map is essentially a root
656 * cluster with restricted subnodes.
657 */
658 map = of_get_child_by_name(cn, "cpu-map");
659 if (!map)
660 goto out;
661
662 ret = parse_socket(map);
663 if (ret != 0)
664 goto out_map;
665
666 topology_normalize_cpu_scale();
667
668 /*
669 * Check that all cores are in the topology; the SMP code will
670 * only mark cores described in the DT as possible.
671 */
672 for_each_possible_cpu(cpu)
673 if (cpu_topology[cpu].package_id < 0) {
674 ret = -EINVAL;
675 break;
676 }
677
678out_map:
679 of_node_put(map);
680out:
681 of_node_put(cn);
682 return ret;
683}
684#endif
685
686/*
687 * cpu topology table
688 */
689struct cpu_topology cpu_topology[NR_CPUS];
690EXPORT_SYMBOL_GPL(cpu_topology);
691
692const struct cpumask *cpu_coregroup_mask(int cpu)
693{
694 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
695
696 /* Find the smaller of NUMA, core or LLC siblings */
697 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
698 /* not numa in package, lets use the package siblings */
699 core_mask = &cpu_topology[cpu].core_sibling;
700 }
701
702 if (last_level_cache_is_valid(cpu)) {
703 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
704 core_mask = &cpu_topology[cpu].llc_sibling;
705 }
706
707 /*
708 * For systems with no shared cpu-side LLC but with clusters defined,
709 * extend core_mask to cluster_siblings. The sched domain builder will
710 * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled.
711 */
712 if (IS_ENABLED(CONFIG_SCHED_CLUSTER) &&
713 cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling))
714 core_mask = &cpu_topology[cpu].cluster_sibling;
715
716 return core_mask;
717}
718
719const struct cpumask *cpu_clustergroup_mask(int cpu)
720{
721 /*
722 * Forbid cpu_clustergroup_mask() to span more or the same CPUs as
723 * cpu_coregroup_mask().
724 */
725 if (cpumask_subset(cpu_coregroup_mask(cpu),
726 &cpu_topology[cpu].cluster_sibling))
727 return topology_sibling_cpumask(cpu);
728
729 return &cpu_topology[cpu].cluster_sibling;
730}
731
732void update_siblings_masks(unsigned int cpuid)
733{
734 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
735 int cpu, ret;
736
737 ret = detect_cache_attributes(cpuid);
738 if (ret && ret != -ENOENT)
739 pr_info("Early cacheinfo failed, ret = %d\n", ret);
740
741 /* update core and thread sibling masks */
742 for_each_online_cpu(cpu) {
743 cpu_topo = &cpu_topology[cpu];
744
745 if (last_level_cache_is_shared(cpu, cpuid)) {
746 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
747 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
748 }
749
750 if (cpuid_topo->package_id != cpu_topo->package_id)
751 continue;
752
753 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
754 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
755
756 if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
757 continue;
758
759 if (cpuid_topo->cluster_id >= 0) {
760 cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling);
761 cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling);
762 }
763
764 if (cpuid_topo->core_id != cpu_topo->core_id)
765 continue;
766
767 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
768 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
769 }
770}
771
772static void clear_cpu_topology(int cpu)
773{
774 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
775
776 cpumask_clear(&cpu_topo->llc_sibling);
777 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
778
779 cpumask_clear(&cpu_topo->cluster_sibling);
780 cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling);
781
782 cpumask_clear(&cpu_topo->core_sibling);
783 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
784 cpumask_clear(&cpu_topo->thread_sibling);
785 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
786}
787
788void __init reset_cpu_topology(void)
789{
790 unsigned int cpu;
791
792 for_each_possible_cpu(cpu) {
793 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
794
795 cpu_topo->thread_id = -1;
796 cpu_topo->core_id = -1;
797 cpu_topo->cluster_id = -1;
798 cpu_topo->package_id = -1;
799
800 clear_cpu_topology(cpu);
801 }
802}
803
804void remove_cpu_topology(unsigned int cpu)
805{
806 int sibling;
807
808 for_each_cpu(sibling, topology_core_cpumask(cpu))
809 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
810 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
811 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
812 for_each_cpu(sibling, topology_cluster_cpumask(cpu))
813 cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling));
814 for_each_cpu(sibling, topology_llc_cpumask(cpu))
815 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
816
817 clear_cpu_topology(cpu);
818}
819
820__weak int __init parse_acpi_topology(void)
821{
822 return 0;
823}
824
825#if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
826void __init init_cpu_topology(void)
827{
828 int ret;
829
830 reset_cpu_topology();
831 ret = parse_acpi_topology();
832 if (!ret)
833 ret = of_have_populated_dt() && parse_dt_topology();
834
835 if (ret) {
836 /*
837 * Discard anything that was parsed if we hit an error so we
838 * don't use partial information.
839 */
840 reset_cpu_topology();
841 return;
842 }
843}
844
845void store_cpu_topology(unsigned int cpuid)
846{
847 struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
848
849 if (cpuid_topo->package_id != -1)
850 goto topology_populated;
851
852 cpuid_topo->thread_id = -1;
853 cpuid_topo->core_id = cpuid;
854 cpuid_topo->package_id = cpu_to_node(cpuid);
855
856 pr_debug("CPU%u: package %d core %d thread %d\n",
857 cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
858 cpuid_topo->thread_id);
859
860topology_populated:
861 update_siblings_masks(cpuid);
862}
863#endif
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Arch specific cpu topology information
4 *
5 * Copyright (C) 2016, ARM Ltd.
6 * Written by: Juri Lelli, ARM Ltd.
7 */
8
9#include <linux/acpi.h>
10#include <linux/cacheinfo.h>
11#include <linux/cleanup.h>
12#include <linux/cpu.h>
13#include <linux/cpufreq.h>
14#include <linux/device.h>
15#include <linux/of.h>
16#include <linux/slab.h>
17#include <linux/sched/topology.h>
18#include <linux/cpuset.h>
19#include <linux/cpumask.h>
20#include <linux/init.h>
21#include <linux/rcupdate.h>
22#include <linux/sched.h>
23#include <linux/units.h>
24
25#define CREATE_TRACE_POINTS
26#include <trace/events/hw_pressure.h>
27
28static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
29static struct cpumask scale_freq_counters_mask;
30static bool scale_freq_invariant;
31DEFINE_PER_CPU(unsigned long, capacity_freq_ref) = 1;
32EXPORT_PER_CPU_SYMBOL_GPL(capacity_freq_ref);
33
34static bool supports_scale_freq_counters(const struct cpumask *cpus)
35{
36 return cpumask_subset(cpus, &scale_freq_counters_mask);
37}
38
39bool topology_scale_freq_invariant(void)
40{
41 return cpufreq_supports_freq_invariance() ||
42 supports_scale_freq_counters(cpu_online_mask);
43}
44
45static void update_scale_freq_invariant(bool status)
46{
47 if (scale_freq_invariant == status)
48 return;
49
50 /*
51 * Task scheduler behavior depends on frequency invariance support,
52 * either cpufreq or counter driven. If the support status changes as
53 * a result of counter initialisation and use, retrigger the build of
54 * scheduling domains to ensure the information is propagated properly.
55 */
56 if (topology_scale_freq_invariant() == status) {
57 scale_freq_invariant = status;
58 rebuild_sched_domains_energy();
59 }
60}
61
62void topology_set_scale_freq_source(struct scale_freq_data *data,
63 const struct cpumask *cpus)
64{
65 struct scale_freq_data *sfd;
66 int cpu;
67
68 /*
69 * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
70 * supported by cpufreq.
71 */
72 if (cpumask_empty(&scale_freq_counters_mask))
73 scale_freq_invariant = topology_scale_freq_invariant();
74
75 rcu_read_lock();
76
77 for_each_cpu(cpu, cpus) {
78 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
79
80 /* Use ARCH provided counters whenever possible */
81 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
82 rcu_assign_pointer(per_cpu(sft_data, cpu), data);
83 cpumask_set_cpu(cpu, &scale_freq_counters_mask);
84 }
85 }
86
87 rcu_read_unlock();
88
89 update_scale_freq_invariant(true);
90}
91EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
92
93void topology_clear_scale_freq_source(enum scale_freq_source source,
94 const struct cpumask *cpus)
95{
96 struct scale_freq_data *sfd;
97 int cpu;
98
99 rcu_read_lock();
100
101 for_each_cpu(cpu, cpus) {
102 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
103
104 if (sfd && sfd->source == source) {
105 rcu_assign_pointer(per_cpu(sft_data, cpu), NULL);
106 cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
107 }
108 }
109
110 rcu_read_unlock();
111
112 /*
113 * Make sure all references to previous sft_data are dropped to avoid
114 * use-after-free races.
115 */
116 synchronize_rcu();
117
118 update_scale_freq_invariant(false);
119}
120EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
121
122void topology_scale_freq_tick(void)
123{
124 struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data));
125
126 if (sfd)
127 sfd->set_freq_scale();
128}
129
130DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
131EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
132
133void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
134 unsigned long max_freq)
135{
136 unsigned long scale;
137 int i;
138
139 if (WARN_ON_ONCE(!cur_freq || !max_freq))
140 return;
141
142 /*
143 * If the use of counters for FIE is enabled, just return as we don't
144 * want to update the scale factor with information from CPUFREQ.
145 * Instead the scale factor will be updated from arch_scale_freq_tick.
146 */
147 if (supports_scale_freq_counters(cpus))
148 return;
149
150 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
151
152 for_each_cpu(i, cpus)
153 per_cpu(arch_freq_scale, i) = scale;
154}
155
156DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
157EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale);
158
159void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
160{
161 per_cpu(cpu_scale, cpu) = capacity;
162}
163
164DEFINE_PER_CPU(unsigned long, hw_pressure);
165
166/**
167 * topology_update_hw_pressure() - Update HW pressure for CPUs
168 * @cpus : The related CPUs for which capacity has been reduced
169 * @capped_freq : The maximum allowed frequency that CPUs can run at
170 *
171 * Update the value of HW pressure for all @cpus in the mask. The
172 * cpumask should include all (online+offline) affected CPUs, to avoid
173 * operating on stale data when hot-plug is used for some CPUs. The
174 * @capped_freq reflects the currently allowed max CPUs frequency due to
175 * HW capping. It might be also a boost frequency value, which is bigger
176 * than the internal 'capacity_freq_ref' max frequency. In such case the
177 * pressure value should simply be removed, since this is an indication that
178 * there is no HW throttling. The @capped_freq must be provided in kHz.
179 */
180void topology_update_hw_pressure(const struct cpumask *cpus,
181 unsigned long capped_freq)
182{
183 unsigned long max_capacity, capacity, pressure;
184 u32 max_freq;
185 int cpu;
186
187 cpu = cpumask_first(cpus);
188 max_capacity = arch_scale_cpu_capacity(cpu);
189 max_freq = arch_scale_freq_ref(cpu);
190
191 /*
192 * Handle properly the boost frequencies, which should simply clean
193 * the HW pressure value.
194 */
195 if (max_freq <= capped_freq)
196 capacity = max_capacity;
197 else
198 capacity = mult_frac(max_capacity, capped_freq, max_freq);
199
200 pressure = max_capacity - capacity;
201
202 trace_hw_pressure_update(cpu, pressure);
203
204 for_each_cpu(cpu, cpus)
205 WRITE_ONCE(per_cpu(hw_pressure, cpu), pressure);
206}
207EXPORT_SYMBOL_GPL(topology_update_hw_pressure);
208
209static ssize_t cpu_capacity_show(struct device *dev,
210 struct device_attribute *attr,
211 char *buf)
212{
213 struct cpu *cpu = container_of(dev, struct cpu, dev);
214
215 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
216}
217
218static void update_topology_flags_workfn(struct work_struct *work);
219static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
220
221static DEVICE_ATTR_RO(cpu_capacity);
222
223static int cpu_capacity_sysctl_add(unsigned int cpu)
224{
225 struct device *cpu_dev = get_cpu_device(cpu);
226
227 if (!cpu_dev)
228 return -ENOENT;
229
230 device_create_file(cpu_dev, &dev_attr_cpu_capacity);
231
232 return 0;
233}
234
235static int cpu_capacity_sysctl_remove(unsigned int cpu)
236{
237 struct device *cpu_dev = get_cpu_device(cpu);
238
239 if (!cpu_dev)
240 return -ENOENT;
241
242 device_remove_file(cpu_dev, &dev_attr_cpu_capacity);
243
244 return 0;
245}
246
247static int register_cpu_capacity_sysctl(void)
248{
249 cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "topology/cpu-capacity",
250 cpu_capacity_sysctl_add, cpu_capacity_sysctl_remove);
251
252 return 0;
253}
254subsys_initcall(register_cpu_capacity_sysctl);
255
256static int update_topology;
257
258int topology_update_cpu_topology(void)
259{
260 return update_topology;
261}
262
263/*
264 * Updating the sched_domains can't be done directly from cpufreq callbacks
265 * due to locking, so queue the work for later.
266 */
267static void update_topology_flags_workfn(struct work_struct *work)
268{
269 update_topology = 1;
270 rebuild_sched_domains();
271 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
272 update_topology = 0;
273}
274
275static u32 *raw_capacity;
276
277static int free_raw_capacity(void)
278{
279 kfree(raw_capacity);
280 raw_capacity = NULL;
281
282 return 0;
283}
284
285void topology_normalize_cpu_scale(void)
286{
287 u64 capacity;
288 u64 capacity_scale;
289 int cpu;
290
291 if (!raw_capacity)
292 return;
293
294 capacity_scale = 1;
295 for_each_possible_cpu(cpu) {
296 capacity = raw_capacity[cpu] * per_cpu(capacity_freq_ref, cpu);
297 capacity_scale = max(capacity, capacity_scale);
298 }
299
300 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
301 for_each_possible_cpu(cpu) {
302 capacity = raw_capacity[cpu] * per_cpu(capacity_freq_ref, cpu);
303 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
304 capacity_scale);
305 topology_set_cpu_scale(cpu, capacity);
306 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
307 cpu, topology_get_cpu_scale(cpu));
308 }
309}
310
311bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
312{
313 struct clk *cpu_clk;
314 static bool cap_parsing_failed;
315 int ret;
316 u32 cpu_capacity;
317
318 if (cap_parsing_failed)
319 return false;
320
321 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
322 &cpu_capacity);
323 if (!ret) {
324 if (!raw_capacity) {
325 raw_capacity = kcalloc(num_possible_cpus(),
326 sizeof(*raw_capacity),
327 GFP_KERNEL);
328 if (!raw_capacity) {
329 cap_parsing_failed = true;
330 return false;
331 }
332 }
333 raw_capacity[cpu] = cpu_capacity;
334 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
335 cpu_node, raw_capacity[cpu]);
336
337 /*
338 * Update capacity_freq_ref for calculating early boot CPU capacities.
339 * For non-clk CPU DVFS mechanism, there's no way to get the
340 * frequency value now, assuming they are running at the same
341 * frequency (by keeping the initial capacity_freq_ref value).
342 */
343 cpu_clk = of_clk_get(cpu_node, 0);
344 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
345 per_cpu(capacity_freq_ref, cpu) =
346 clk_get_rate(cpu_clk) / HZ_PER_KHZ;
347 clk_put(cpu_clk);
348 }
349 } else {
350 if (raw_capacity) {
351 pr_err("cpu_capacity: missing %pOF raw capacity\n",
352 cpu_node);
353 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
354 }
355 cap_parsing_failed = true;
356 free_raw_capacity();
357 }
358
359 return !ret;
360}
361
362void __weak freq_inv_set_max_ratio(int cpu, u64 max_rate)
363{
364}
365
366#ifdef CONFIG_ACPI_CPPC_LIB
367#include <acpi/cppc_acpi.h>
368
369static inline void topology_init_cpu_capacity_cppc(void)
370{
371 u64 capacity, capacity_scale = 0;
372 struct cppc_perf_caps perf_caps;
373 int cpu;
374
375 if (likely(!acpi_cpc_valid()))
376 return;
377
378 raw_capacity = kcalloc(num_possible_cpus(), sizeof(*raw_capacity),
379 GFP_KERNEL);
380 if (!raw_capacity)
381 return;
382
383 for_each_possible_cpu(cpu) {
384 if (!cppc_get_perf_caps(cpu, &perf_caps) &&
385 (perf_caps.highest_perf >= perf_caps.nominal_perf) &&
386 (perf_caps.highest_perf >= perf_caps.lowest_perf)) {
387 raw_capacity[cpu] = perf_caps.highest_perf;
388 capacity_scale = max_t(u64, capacity_scale, raw_capacity[cpu]);
389
390 per_cpu(capacity_freq_ref, cpu) = cppc_perf_to_khz(&perf_caps, raw_capacity[cpu]);
391
392 pr_debug("cpu_capacity: CPU%d cpu_capacity=%u (raw).\n",
393 cpu, raw_capacity[cpu]);
394 continue;
395 }
396
397 pr_err("cpu_capacity: CPU%d missing/invalid highest performance.\n", cpu);
398 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
399 goto exit;
400 }
401
402 for_each_possible_cpu(cpu) {
403 freq_inv_set_max_ratio(cpu,
404 per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
405
406 capacity = raw_capacity[cpu];
407 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
408 capacity_scale);
409 topology_set_cpu_scale(cpu, capacity);
410 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
411 cpu, topology_get_cpu_scale(cpu));
412 }
413
414 schedule_work(&update_topology_flags_work);
415 pr_debug("cpu_capacity: cpu_capacity initialization done\n");
416
417exit:
418 free_raw_capacity();
419}
420void acpi_processor_init_invariance_cppc(void)
421{
422 topology_init_cpu_capacity_cppc();
423}
424#endif
425
426#ifdef CONFIG_CPU_FREQ
427static cpumask_var_t cpus_to_visit;
428static void parsing_done_workfn(struct work_struct *work);
429static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
430
431static int
432init_cpu_capacity_callback(struct notifier_block *nb,
433 unsigned long val,
434 void *data)
435{
436 struct cpufreq_policy *policy = data;
437 int cpu;
438
439 if (val != CPUFREQ_CREATE_POLICY)
440 return 0;
441
442 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
443 cpumask_pr_args(policy->related_cpus),
444 cpumask_pr_args(cpus_to_visit));
445
446 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
447
448 for_each_cpu(cpu, policy->related_cpus) {
449 per_cpu(capacity_freq_ref, cpu) = policy->cpuinfo.max_freq;
450 freq_inv_set_max_ratio(cpu,
451 per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
452 }
453
454 if (cpumask_empty(cpus_to_visit)) {
455 if (raw_capacity) {
456 topology_normalize_cpu_scale();
457 schedule_work(&update_topology_flags_work);
458 free_raw_capacity();
459 }
460 pr_debug("cpu_capacity: parsing done\n");
461 schedule_work(&parsing_done_work);
462 }
463
464 return 0;
465}
466
467static struct notifier_block init_cpu_capacity_notifier = {
468 .notifier_call = init_cpu_capacity_callback,
469};
470
471static int __init register_cpufreq_notifier(void)
472{
473 int ret;
474
475 /*
476 * On ACPI-based systems skip registering cpufreq notifier as cpufreq
477 * information is not needed for cpu capacity initialization.
478 */
479 if (!acpi_disabled)
480 return -EINVAL;
481
482 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
483 return -ENOMEM;
484
485 cpumask_copy(cpus_to_visit, cpu_possible_mask);
486
487 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
488 CPUFREQ_POLICY_NOTIFIER);
489
490 if (ret)
491 free_cpumask_var(cpus_to_visit);
492
493 return ret;
494}
495core_initcall(register_cpufreq_notifier);
496
497static void parsing_done_workfn(struct work_struct *work)
498{
499 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
500 CPUFREQ_POLICY_NOTIFIER);
501 free_cpumask_var(cpus_to_visit);
502}
503
504#else
505core_initcall(free_raw_capacity);
506#endif
507
508#if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
509/*
510 * This function returns the logic cpu number of the node.
511 * There are basically three kinds of return values:
512 * (1) logic cpu number which is > 0.
513 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
514 * there is no possible logical CPU in the kernel to match. This happens
515 * when CONFIG_NR_CPUS is configure to be smaller than the number of
516 * CPU nodes in DT. We need to just ignore this case.
517 * (3) -1 if the node does not exist in the device tree
518 */
519static int __init get_cpu_for_node(struct device_node *node)
520{
521 int cpu;
522 struct device_node *cpu_node __free(device_node) =
523 of_parse_phandle(node, "cpu", 0);
524
525 if (!cpu_node)
526 return -1;
527
528 cpu = of_cpu_node_to_id(cpu_node);
529 if (cpu >= 0)
530 topology_parse_cpu_capacity(cpu_node, cpu);
531 else
532 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
533 cpu_node, cpumask_pr_args(cpu_possible_mask));
534
535 return cpu;
536}
537
538static int __init parse_core(struct device_node *core, int package_id,
539 int cluster_id, int core_id)
540{
541 char name[20];
542 bool leaf = true;
543 int i = 0;
544 int cpu;
545
546 do {
547 snprintf(name, sizeof(name), "thread%d", i);
548 struct device_node *t __free(device_node) =
549 of_get_child_by_name(core, name);
550
551 if (!t)
552 break;
553
554 leaf = false;
555 cpu = get_cpu_for_node(t);
556 if (cpu >= 0) {
557 cpu_topology[cpu].package_id = package_id;
558 cpu_topology[cpu].cluster_id = cluster_id;
559 cpu_topology[cpu].core_id = core_id;
560 cpu_topology[cpu].thread_id = i;
561 } else if (cpu != -ENODEV) {
562 pr_err("%pOF: Can't get CPU for thread\n", t);
563 return -EINVAL;
564 }
565 i++;
566 } while (1);
567
568 cpu = get_cpu_for_node(core);
569 if (cpu >= 0) {
570 if (!leaf) {
571 pr_err("%pOF: Core has both threads and CPU\n",
572 core);
573 return -EINVAL;
574 }
575
576 cpu_topology[cpu].package_id = package_id;
577 cpu_topology[cpu].cluster_id = cluster_id;
578 cpu_topology[cpu].core_id = core_id;
579 } else if (leaf && cpu != -ENODEV) {
580 pr_err("%pOF: Can't get CPU for leaf core\n", core);
581 return -EINVAL;
582 }
583
584 return 0;
585}
586
587static int __init parse_cluster(struct device_node *cluster, int package_id,
588 int cluster_id, int depth)
589{
590 char name[20];
591 bool leaf = true;
592 bool has_cores = false;
593 int core_id = 0;
594 int i, ret;
595
596 /*
597 * First check for child clusters; we currently ignore any
598 * information about the nesting of clusters and present the
599 * scheduler with a flat list of them.
600 */
601 i = 0;
602 do {
603 snprintf(name, sizeof(name), "cluster%d", i);
604 struct device_node *c __free(device_node) =
605 of_get_child_by_name(cluster, name);
606
607 if (!c)
608 break;
609
610 leaf = false;
611 ret = parse_cluster(c, package_id, i, depth + 1);
612 if (depth > 0)
613 pr_warn("Topology for clusters of clusters not yet supported\n");
614 if (ret != 0)
615 return ret;
616 i++;
617 } while (1);
618
619 /* Now check for cores */
620 i = 0;
621 do {
622 snprintf(name, sizeof(name), "core%d", i);
623 struct device_node *c __free(device_node) =
624 of_get_child_by_name(cluster, name);
625
626 if (!c)
627 break;
628
629 has_cores = true;
630
631 if (depth == 0) {
632 pr_err("%pOF: cpu-map children should be clusters\n", c);
633 return -EINVAL;
634 }
635
636 if (leaf) {
637 ret = parse_core(c, package_id, cluster_id, core_id++);
638 if (ret != 0)
639 return ret;
640 } else {
641 pr_err("%pOF: Non-leaf cluster with core %s\n",
642 cluster, name);
643 return -EINVAL;
644 }
645
646 i++;
647 } while (1);
648
649 if (leaf && !has_cores)
650 pr_warn("%pOF: empty cluster\n", cluster);
651
652 return 0;
653}
654
655static int __init parse_socket(struct device_node *socket)
656{
657 char name[20];
658 bool has_socket = false;
659 int package_id = 0, ret;
660
661 do {
662 snprintf(name, sizeof(name), "socket%d", package_id);
663 struct device_node *c __free(device_node) =
664 of_get_child_by_name(socket, name);
665
666 if (!c)
667 break;
668
669 has_socket = true;
670 ret = parse_cluster(c, package_id, -1, 0);
671 if (ret != 0)
672 return ret;
673
674 package_id++;
675 } while (1);
676
677 if (!has_socket)
678 ret = parse_cluster(socket, 0, -1, 0);
679
680 return ret;
681}
682
683static int __init parse_dt_topology(void)
684{
685 int ret = 0;
686 int cpu;
687 struct device_node *cn __free(device_node) =
688 of_find_node_by_path("/cpus");
689
690 if (!cn) {
691 pr_err("No CPU information found in DT\n");
692 return 0;
693 }
694
695 /*
696 * When topology is provided cpu-map is essentially a root
697 * cluster with restricted subnodes.
698 */
699 struct device_node *map __free(device_node) =
700 of_get_child_by_name(cn, "cpu-map");
701
702 if (!map)
703 return ret;
704
705 ret = parse_socket(map);
706 if (ret != 0)
707 return ret;
708
709 topology_normalize_cpu_scale();
710
711 /*
712 * Check that all cores are in the topology; the SMP code will
713 * only mark cores described in the DT as possible.
714 */
715 for_each_possible_cpu(cpu)
716 if (cpu_topology[cpu].package_id < 0) {
717 return -EINVAL;
718 }
719
720 return ret;
721}
722#endif
723
724/*
725 * cpu topology table
726 */
727struct cpu_topology cpu_topology[NR_CPUS];
728EXPORT_SYMBOL_GPL(cpu_topology);
729
730const struct cpumask *cpu_coregroup_mask(int cpu)
731{
732 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
733
734 /* Find the smaller of NUMA, core or LLC siblings */
735 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
736 /* not numa in package, lets use the package siblings */
737 core_mask = &cpu_topology[cpu].core_sibling;
738 }
739
740 if (last_level_cache_is_valid(cpu)) {
741 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
742 core_mask = &cpu_topology[cpu].llc_sibling;
743 }
744
745 /*
746 * For systems with no shared cpu-side LLC but with clusters defined,
747 * extend core_mask to cluster_siblings. The sched domain builder will
748 * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled.
749 */
750 if (IS_ENABLED(CONFIG_SCHED_CLUSTER) &&
751 cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling))
752 core_mask = &cpu_topology[cpu].cluster_sibling;
753
754 return core_mask;
755}
756
757const struct cpumask *cpu_clustergroup_mask(int cpu)
758{
759 /*
760 * Forbid cpu_clustergroup_mask() to span more or the same CPUs as
761 * cpu_coregroup_mask().
762 */
763 if (cpumask_subset(cpu_coregroup_mask(cpu),
764 &cpu_topology[cpu].cluster_sibling))
765 return topology_sibling_cpumask(cpu);
766
767 return &cpu_topology[cpu].cluster_sibling;
768}
769
770void update_siblings_masks(unsigned int cpuid)
771{
772 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
773 int cpu, ret;
774
775 ret = detect_cache_attributes(cpuid);
776 if (ret && ret != -ENOENT)
777 pr_info("Early cacheinfo allocation failed, ret = %d\n", ret);
778
779 /* update core and thread sibling masks */
780 for_each_online_cpu(cpu) {
781 cpu_topo = &cpu_topology[cpu];
782
783 if (last_level_cache_is_shared(cpu, cpuid)) {
784 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
785 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
786 }
787
788 if (cpuid_topo->package_id != cpu_topo->package_id)
789 continue;
790
791 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
792 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
793
794 if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
795 continue;
796
797 if (cpuid_topo->cluster_id >= 0) {
798 cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling);
799 cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling);
800 }
801
802 if (cpuid_topo->core_id != cpu_topo->core_id)
803 continue;
804
805 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
806 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
807 }
808}
809
810static void clear_cpu_topology(int cpu)
811{
812 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
813
814 cpumask_clear(&cpu_topo->llc_sibling);
815 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
816
817 cpumask_clear(&cpu_topo->cluster_sibling);
818 cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling);
819
820 cpumask_clear(&cpu_topo->core_sibling);
821 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
822 cpumask_clear(&cpu_topo->thread_sibling);
823 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
824}
825
826void __init reset_cpu_topology(void)
827{
828 unsigned int cpu;
829
830 for_each_possible_cpu(cpu) {
831 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
832
833 cpu_topo->thread_id = -1;
834 cpu_topo->core_id = -1;
835 cpu_topo->cluster_id = -1;
836 cpu_topo->package_id = -1;
837
838 clear_cpu_topology(cpu);
839 }
840}
841
842void remove_cpu_topology(unsigned int cpu)
843{
844 int sibling;
845
846 for_each_cpu(sibling, topology_core_cpumask(cpu))
847 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
848 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
849 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
850 for_each_cpu(sibling, topology_cluster_cpumask(cpu))
851 cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling));
852 for_each_cpu(sibling, topology_llc_cpumask(cpu))
853 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
854
855 clear_cpu_topology(cpu);
856}
857
858__weak int __init parse_acpi_topology(void)
859{
860 return 0;
861}
862
863#if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
864void __init init_cpu_topology(void)
865{
866 int cpu, ret;
867
868 reset_cpu_topology();
869 ret = parse_acpi_topology();
870 if (!ret)
871 ret = of_have_populated_dt() && parse_dt_topology();
872
873 if (ret) {
874 /*
875 * Discard anything that was parsed if we hit an error so we
876 * don't use partial information. But do not return yet to give
877 * arch-specific early cache level detection a chance to run.
878 */
879 reset_cpu_topology();
880 }
881
882 for_each_possible_cpu(cpu) {
883 ret = fetch_cache_info(cpu);
884 if (!ret)
885 continue;
886 else if (ret != -ENOENT)
887 pr_err("Early cacheinfo failed, ret = %d\n", ret);
888 return;
889 }
890}
891
892void store_cpu_topology(unsigned int cpuid)
893{
894 struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
895
896 if (cpuid_topo->package_id != -1)
897 goto topology_populated;
898
899 cpuid_topo->thread_id = -1;
900 cpuid_topo->core_id = cpuid;
901 cpuid_topo->package_id = cpu_to_node(cpuid);
902
903 pr_debug("CPU%u: package %d core %d thread %d\n",
904 cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
905 cpuid_topo->thread_id);
906
907topology_populated:
908 update_siblings_masks(cpuid);
909}
910#endif