projects/linux·blob master · refs

979 lines · 24.2 KiB

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/cpu_smt.h>
15#include <linux/device.h>
16#include <linux/of.h>
17#include <linux/slab.h>
18#include <linux/sched/topology.h>
19#include <linux/cpuset.h>
20#include <linux/cpumask.h>
21#include <linux/init.h>
22#include <linux/rcupdate.h>
23#include <linux/sched.h>
24#include <linux/units.h>
25
26#define CREATE_TRACE_POINTS
27#include <trace/events/hw_pressure.h>
28
29static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
30static struct cpumask scale_freq_counters_mask;
31static bool scale_freq_invariant;
32DEFINE_PER_CPU(unsigned long, capacity_freq_ref) = 0;
33EXPORT_PER_CPU_SYMBOL_GPL(capacity_freq_ref);
34
35static bool supports_scale_freq_counters(const struct cpumask *cpus)
36{
37 int i;
38
39 for_each_cpu(i, cpus) {
40 if (cpumask_test_cpu(i, &scale_freq_counters_mask))
41 return true;
42 }
43
44 return false;
45}
46
47bool topology_scale_freq_invariant(void)
48{
49 return cpufreq_supports_freq_invariance() ||
50 supports_scale_freq_counters(cpu_online_mask);
51}
52
53static void update_scale_freq_invariant(bool status)
54{
55 if (scale_freq_invariant == status)
56 return;
57
58 /*
59 * Task scheduler behavior depends on frequency invariance support,
60 * either cpufreq or counter driven. If the support status changes as
61 * a result of counter initialisation and use, retrigger the build of
62 * scheduling domains to ensure the information is propagated properly.
63 */
64 if (topology_scale_freq_invariant() == status) {
65 scale_freq_invariant = status;
66 rebuild_sched_domains_energy();
67 }
68}
69
70void topology_set_scale_freq_source(struct scale_freq_data *data,
71 const struct cpumask *cpus)
72{
73 struct scale_freq_data *sfd;
74 int cpu;
75
76 /*
77 * Avoid calling rebuild_sched_domains() unnecessarily if FIE is
78 * supported by cpufreq.
79 */
80 if (cpumask_empty(&scale_freq_counters_mask))
81 scale_freq_invariant = topology_scale_freq_invariant();
82
83 rcu_read_lock();
84
85 for_each_cpu(cpu, cpus) {
86 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
87
88 /* Use ARCH provided counters whenever possible */
89 if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) {
90 rcu_assign_pointer(per_cpu(sft_data, cpu), data);
91 cpumask_set_cpu(cpu, &scale_freq_counters_mask);
92 }
93 }
94
95 rcu_read_unlock();
96
97 update_scale_freq_invariant(true);
98}
99EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
100
101void topology_clear_scale_freq_source(enum scale_freq_source source,
102 const struct cpumask *cpus)
103{
104 struct scale_freq_data *sfd;
105 int cpu;
106
107 rcu_read_lock();
108
109 for_each_cpu(cpu, cpus) {
110 sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu));
111
112 if (sfd && sfd->source == source) {
113 rcu_assign_pointer(per_cpu(sft_data, cpu), NULL);
114 cpumask_clear_cpu(cpu, &scale_freq_counters_mask);
115 }
116 }
117
118 rcu_read_unlock();
119
120 /*
121 * Make sure all references to previous sft_data are dropped to avoid
122 * use-after-free races.
123 */
124 synchronize_rcu();
125
126 update_scale_freq_invariant(false);
127}
128EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
129
130void topology_scale_freq_tick(void)
131{
132 struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data));
133
134 if (sfd)
135 sfd->set_freq_scale();
136}
137
138DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
139EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
140
141void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
142 unsigned long max_freq)
143{
144 unsigned long scale;
145 int i;
146
147 if (WARN_ON_ONCE(!cur_freq || !max_freq))
148 return;
149
150 /*
151 * If the use of counters for FIE is enabled, just return as we don't
152 * want to update the scale factor with information from CPUFREQ.
153 * Instead the scale factor will be updated from arch_scale_freq_tick.
154 */
155 if (supports_scale_freq_counters(cpus))
156 return;
157
158 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
159
160 for_each_cpu(i, cpus)
161 per_cpu(arch_freq_scale, i) = scale;
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 void update_topology_flags_workfn(struct work_struct *work);
210static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
211
212static int update_topology;
213
214int topology_update_cpu_topology(void)
215{
216 return update_topology;
217}
218
219/*
220 * Updating the sched_domains can't be done directly from cpufreq callbacks
221 * due to locking, so queue the work for later.
222 */
223static void update_topology_flags_workfn(struct work_struct *work)
224{
225 update_topology = 1;
226 rebuild_sched_domains();
227 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
228 update_topology = 0;
229}
230
231static u32 *raw_capacity;
232
233static int free_raw_capacity(void)
234{
235 kfree(raw_capacity);
236 raw_capacity = NULL;
237
238 return 0;
239}
240
241void topology_normalize_cpu_scale(void)
242{
243 u64 capacity;
244 u64 capacity_scale;
245 int cpu;
246
247 if (!raw_capacity)
248 return;
249
250 capacity_scale = 1;
251 for_each_possible_cpu(cpu) {
252 capacity = raw_capacity[cpu] *
253 (per_cpu(capacity_freq_ref, cpu) ?: 1);
254 capacity_scale = max(capacity, capacity_scale);
255 }
256
257 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
258 for_each_possible_cpu(cpu) {
259 capacity = raw_capacity[cpu] *
260 (per_cpu(capacity_freq_ref, cpu) ?: 1);
261 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
262 capacity_scale);
263 topology_set_cpu_scale(cpu, capacity);
264 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
265 cpu, topology_get_cpu_scale(cpu));
266 }
267}
268
269bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
270{
271 struct clk *cpu_clk;
272 static bool cap_parsing_failed;
273 int ret;
274 u32 cpu_capacity;
275
276 if (cap_parsing_failed)
277 return false;
278
279 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
280 &cpu_capacity);
281 if (!ret) {
282 if (!raw_capacity) {
283 raw_capacity = kcalloc(num_possible_cpus(),
284 sizeof(*raw_capacity),
285 GFP_KERNEL);
286 if (!raw_capacity) {
287 cap_parsing_failed = true;
288 return false;
289 }
290 }
291 raw_capacity[cpu] = cpu_capacity;
292 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
293 cpu_node, raw_capacity[cpu]);
294
295 /*
296 * Update capacity_freq_ref for calculating early boot CPU capacities.
297 * For non-clk CPU DVFS mechanism, there's no way to get the
298 * frequency value now, assuming they are running at the same
299 * frequency (by keeping the initial capacity_freq_ref value).
300 */
301 cpu_clk = of_clk_get(cpu_node, 0);
302 if (!IS_ERR_OR_NULL(cpu_clk)) {
303 per_cpu(capacity_freq_ref, cpu) =
304 clk_get_rate(cpu_clk) / HZ_PER_KHZ;
305 clk_put(cpu_clk);
306 }
307 } else {
308 if (raw_capacity) {
309 pr_err("cpu_capacity: missing %pOF raw capacity\n",
310 cpu_node);
311 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
312 }
313 cap_parsing_failed = true;
314 free_raw_capacity();
315 }
316
317 return !ret;
318}
319
320void __weak freq_inv_set_max_ratio(int cpu, u64 max_rate)
321{
322}
323
324#ifdef CONFIG_ACPI_CPPC_LIB
325#include <acpi/cppc_acpi.h>
326
327static inline void topology_init_cpu_capacity_cppc(void)
328{
329 u64 capacity, capacity_scale = 0;
330 struct cppc_perf_caps perf_caps;
331 int cpu;
332
333 if (likely(!acpi_cpc_valid()))
334 return;
335
336 raw_capacity = kcalloc(num_possible_cpus(), sizeof(*raw_capacity),
337 GFP_KERNEL);
338 if (!raw_capacity)
339 return;
340
341 for_each_possible_cpu(cpu) {
342 if (!cppc_get_perf_caps(cpu, &perf_caps) &&
343 (perf_caps.highest_perf >= perf_caps.nominal_perf) &&
344 (perf_caps.highest_perf >= perf_caps.lowest_perf)) {
345 raw_capacity[cpu] = perf_caps.highest_perf;
346 capacity_scale = max_t(u64, capacity_scale, raw_capacity[cpu]);
347
348 per_cpu(capacity_freq_ref, cpu) = cppc_perf_to_khz(&perf_caps, raw_capacity[cpu]);
349
350 pr_debug("cpu_capacity: CPU%d cpu_capacity=%u (raw).\n",
351 cpu, raw_capacity[cpu]);
352 continue;
353 }
354
355 pr_err("cpu_capacity: CPU%d missing/invalid highest performance.\n", cpu);
356 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
357 goto exit;
358 }
359
360 for_each_possible_cpu(cpu) {
361 freq_inv_set_max_ratio(cpu,
362 per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
363
364 capacity = raw_capacity[cpu];
365 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
366 capacity_scale);
367 topology_set_cpu_scale(cpu, capacity);
368 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
369 cpu, topology_get_cpu_scale(cpu));
370 }
371
372 schedule_work(&update_topology_flags_work);
373 pr_debug("cpu_capacity: cpu_capacity initialization done\n");
374
375exit:
376 free_raw_capacity();
377}
378void acpi_processor_init_invariance_cppc(void)
379{
380 topology_init_cpu_capacity_cppc();
381}
382#endif
383
384#ifdef CONFIG_CPU_FREQ
385static cpumask_var_t cpus_to_visit;
386static void parsing_done_workfn(struct work_struct *work);
387static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
388
389static int
390init_cpu_capacity_callback(struct notifier_block *nb,
391 unsigned long val,
392 void *data)
393{
394 struct cpufreq_policy *policy = data;
395 int cpu;
396
397 if (val != CPUFREQ_CREATE_POLICY)
398 return 0;
399
400 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
401 cpumask_pr_args(policy->related_cpus),
402 cpumask_pr_args(cpus_to_visit));
403
404 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
405
406 for_each_cpu(cpu, policy->related_cpus) {
407 per_cpu(capacity_freq_ref, cpu) = policy->cpuinfo.max_freq;
408 freq_inv_set_max_ratio(cpu,
409 per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
410 }
411
412 if (cpumask_empty(cpus_to_visit)) {
413 if (raw_capacity) {
414 topology_normalize_cpu_scale();
415 schedule_work(&update_topology_flags_work);
416 free_raw_capacity();
417 }
418 pr_debug("cpu_capacity: parsing done\n");
419 schedule_work(&parsing_done_work);
420 }
421
422 return 0;
423}
424
425static struct notifier_block init_cpu_capacity_notifier = {
426 .notifier_call = init_cpu_capacity_callback,
427};
428
429static int __init register_cpufreq_notifier(void)
430{
431 int ret;
432
433 /*
434 * On ACPI-based systems skip registering cpufreq notifier as cpufreq
435 * information is not needed for cpu capacity initialization.
436 */
437 if (!acpi_disabled)
438 return -EINVAL;
439
440 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
441 return -ENOMEM;
442
443 cpumask_copy(cpus_to_visit, cpu_possible_mask);
444
445 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
446 CPUFREQ_POLICY_NOTIFIER);
447
448 if (ret)
449 free_cpumask_var(cpus_to_visit);
450
451 return ret;
452}
453core_initcall(register_cpufreq_notifier);
454
455static void parsing_done_workfn(struct work_struct *work)
456{
457 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
458 CPUFREQ_POLICY_NOTIFIER);
459 free_cpumask_var(cpus_to_visit);
460}
461
462#else
463core_initcall(free_raw_capacity);
464#endif
465
466#if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
467
468/* Used to enable the SMT control */
469static unsigned int max_smt_thread_num = 1;
470
471/*
472 * This function returns the logic cpu number of the node.
473 * There are basically three kinds of return values:
474 * (1) logic cpu number which is > 0.
475 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
476 * there is no possible logical CPU in the kernel to match. This happens
477 * when CONFIG_NR_CPUS is configure to be smaller than the number of
478 * CPU nodes in DT. We need to just ignore this case.
479 * (3) -1 if the node does not exist in the device tree
480 */
481static int __init get_cpu_for_node(struct device_node *node)
482{
483 int cpu;
484 struct device_node *cpu_node __free(device_node) =
485 of_parse_phandle(node, "cpu", 0);
486
487 if (!cpu_node)
488 return -1;
489
490 cpu = of_cpu_node_to_id(cpu_node);
491 if (cpu >= 0)
492 topology_parse_cpu_capacity(cpu_node, cpu);
493 else
494 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
495 cpu_node, cpumask_pr_args(cpu_possible_mask));
496
497 return cpu;
498}
499
500static int __init parse_core(struct device_node *core, int package_id,
501 int cluster_id, int core_id)
502{
503 char name[20];
504 bool leaf = true;
505 int i = 0;
506 int cpu;
507
508 do {
509 snprintf(name, sizeof(name), "thread%d", i);
510 struct device_node *t __free(device_node) =
511 of_get_child_by_name(core, name);
512
513 if (!t)
514 break;
515
516 leaf = false;
517 cpu = get_cpu_for_node(t);
518 if (cpu >= 0) {
519 cpu_topology[cpu].package_id = package_id;
520 cpu_topology[cpu].cluster_id = cluster_id;
521 cpu_topology[cpu].core_id = core_id;
522 cpu_topology[cpu].thread_id = i;
523 } else if (cpu != -ENODEV) {
524 pr_err("%pOF: Can't get CPU for thread\n", t);
525 return -EINVAL;
526 }
527 i++;
528 } while (1);
529
530 max_smt_thread_num = max_t(unsigned int, max_smt_thread_num, i);
531
532 cpu = get_cpu_for_node(core);
533 if (cpu >= 0) {
534 if (!leaf) {
535 pr_err("%pOF: Core has both threads and CPU\n",
536 core);
537 return -EINVAL;
538 }
539
540 cpu_topology[cpu].package_id = package_id;
541 cpu_topology[cpu].cluster_id = cluster_id;
542 cpu_topology[cpu].core_id = core_id;
543 } else if (leaf && cpu != -ENODEV) {
544 pr_err("%pOF: Can't get CPU for leaf core\n", core);
545 return -EINVAL;
546 }
547
548 return 0;
549}
550
551static int __init parse_cluster(struct device_node *cluster, int package_id,
552 int cluster_id, int depth)
553{
554 char name[20];
555 bool leaf = true;
556 bool has_cores = false;
557 int core_id = 0;
558 int i, ret;
559
560 /*
561 * First check for child clusters; we currently ignore any
562 * information about the nesting of clusters and present the
563 * scheduler with a flat list of them.
564 */
565 i = 0;
566 do {
567 snprintf(name, sizeof(name), "cluster%d", i);
568 struct device_node *c __free(device_node) =
569 of_get_child_by_name(cluster, name);
570
571 if (!c)
572 break;
573
574 leaf = false;
575 ret = parse_cluster(c, package_id, i, depth + 1);
576 if (depth > 0)
577 pr_warn("Topology for clusters of clusters not yet supported\n");
578 if (ret != 0)
579 return ret;
580 i++;
581 } while (1);
582
583 /* Now check for cores */
584 i = 0;
585 do {
586 snprintf(name, sizeof(name), "core%d", i);
587 struct device_node *c __free(device_node) =
588 of_get_child_by_name(cluster, name);
589
590 if (!c)
591 break;
592
593 has_cores = true;
594
595 if (depth == 0) {
596 pr_err("%pOF: cpu-map children should be clusters\n", c);
597 return -EINVAL;
598 }
599
600 if (leaf) {
601 ret = parse_core(c, package_id, cluster_id, core_id++);
602 if (ret != 0)
603 return ret;
604 } else {
605 pr_err("%pOF: Non-leaf cluster with core %s\n",
606 cluster, name);
607 return -EINVAL;
608 }
609
610 i++;
611 } while (1);
612
613 if (leaf && !has_cores)
614 pr_warn("%pOF: empty cluster\n", cluster);
615
616 return 0;
617}
618
619static int __init parse_socket(struct device_node *socket)
620{
621 char name[20];
622 bool has_socket = false;
623 int package_id = 0, ret;
624
625 do {
626 snprintf(name, sizeof(name), "socket%d", package_id);
627 struct device_node *c __free(device_node) =
628 of_get_child_by_name(socket, name);
629
630 if (!c)
631 break;
632
633 has_socket = true;
634 ret = parse_cluster(c, package_id, -1, 0);
635 if (ret != 0)
636 return ret;
637
638 package_id++;
639 } while (1);
640
641 if (!has_socket)
642 ret = parse_cluster(socket, 0, -1, 0);
643
644 /*
645 * Reset the max_smt_thread_num to 1 on failure. Since on failure
646 * we need to notify the framework the SMT is not supported, but
647 * max_smt_thread_num can be initialized to the SMT thread number
648 * of the cores which are successfully parsed.
649 */
650 if (ret)
651 max_smt_thread_num = 1;
652
653 cpu_smt_set_num_threads(max_smt_thread_num, max_smt_thread_num);
654
655 return ret;
656}
657
658static int __init parse_dt_topology(void)
659{
660 int ret = 0;
661 int cpu;
662 struct device_node *cn __free(device_node) =
663 of_find_node_by_path("/cpus");
664
665 if (!cn) {
666 pr_err("No CPU information found in DT\n");
667 return 0;
668 }
669
670 /*
671 * When topology is provided cpu-map is essentially a root
672 * cluster with restricted subnodes.
673 */
674 struct device_node *map __free(device_node) =
675 of_get_child_by_name(cn, "cpu-map");
676
677 if (!map)
678 return ret;
679
680 ret = parse_socket(map);
681 if (ret != 0)
682 return ret;
683
684 topology_normalize_cpu_scale();
685
686 /*
687 * Check that all cores are in the topology; the SMP code will
688 * only mark cores described in the DT as possible.
689 */
690 for_each_possible_cpu(cpu)
691 if (cpu_topology[cpu].package_id < 0) {
692 return -EINVAL;
693 }
694
695 return ret;
696}
697#endif
698
699/*
700 * cpu topology table
701 */
702struct cpu_topology cpu_topology[NR_CPUS];
703EXPORT_SYMBOL_GPL(cpu_topology);
704
705const struct cpumask *cpu_coregroup_mask(int cpu)
706{
707 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
708
709 /* Find the smaller of NUMA, core or LLC siblings */
710 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
711 /* not numa in package, lets use the package siblings */
712 core_mask = &cpu_topology[cpu].core_sibling;
713 }
714
715 if (last_level_cache_is_valid(cpu)) {
716 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
717 core_mask = &cpu_topology[cpu].llc_sibling;
718 }
719
720 /*
721 * For systems with no shared cpu-side LLC but with clusters defined,
722 * extend core_mask to cluster_siblings. The sched domain builder will
723 * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled.
724 */
725 if (IS_ENABLED(CONFIG_SCHED_CLUSTER) &&
726 cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling))
727 core_mask = &cpu_topology[cpu].cluster_sibling;
728
729 return core_mask;
730}
731
732const struct cpumask *cpu_clustergroup_mask(int cpu)
733{
734 /*
735 * Forbid cpu_clustergroup_mask() to span more or the same CPUs as
736 * cpu_coregroup_mask().
737 */
738 if (cpumask_subset(cpu_coregroup_mask(cpu),
739 &cpu_topology[cpu].cluster_sibling))
740 return topology_sibling_cpumask(cpu);
741
742 return &cpu_topology[cpu].cluster_sibling;
743}
744
745void update_siblings_masks(unsigned int cpuid)
746{
747 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
748 int cpu, ret;
749
750 ret = detect_cache_attributes(cpuid);
751 if (ret && ret != -ENOENT)
752 pr_info("Early cacheinfo allocation failed, ret = %d\n", ret);
753
754 /* update core and thread sibling masks */
755 for_each_online_cpu(cpu) {
756 cpu_topo = &cpu_topology[cpu];
757
758 if (last_level_cache_is_shared(cpu, cpuid)) {
759 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
760 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
761 }
762
763 if (cpuid_topo->package_id != cpu_topo->package_id)
764 continue;
765
766 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
767 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
768
769 if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
770 continue;
771
772 if (cpuid_topo->cluster_id >= 0) {
773 cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling);
774 cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling);
775 }
776
777 if (cpuid_topo->core_id != cpu_topo->core_id)
778 continue;
779
780 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
781 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
782 }
783}
784
785static void clear_cpu_topology(int cpu)
786{
787 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
788
789 cpumask_clear(&cpu_topo->llc_sibling);
790 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
791
792 cpumask_clear(&cpu_topo->cluster_sibling);
793 cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling);
794
795 cpumask_clear(&cpu_topo->core_sibling);
796 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
797 cpumask_clear(&cpu_topo->thread_sibling);
798 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
799}
800
801void __init reset_cpu_topology(void)
802{
803 unsigned int cpu;
804
805 for_each_possible_cpu(cpu) {
806 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
807
808 cpu_topo->thread_id = -1;
809 cpu_topo->core_id = -1;
810 cpu_topo->cluster_id = -1;
811 cpu_topo->package_id = -1;
812
813 clear_cpu_topology(cpu);
814 }
815}
816
817void remove_cpu_topology(unsigned int cpu)
818{
819 int sibling;
820
821 for_each_cpu(sibling, topology_core_cpumask(cpu))
822 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
823 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
824 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
825 for_each_cpu(sibling, topology_cluster_cpumask(cpu))
826 cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling));
827 for_each_cpu(sibling, topology_llc_cpumask(cpu))
828 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
829
830 clear_cpu_topology(cpu);
831}
832
833#if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
834struct cpu_smt_info {
835 unsigned int thread_num;
836 int core_id;
837};
838
839static bool __init acpi_cpu_is_threaded(int cpu)
840{
841 int is_threaded = acpi_pptt_cpu_is_thread(cpu);
842
843 /*
844 * if the PPTT doesn't have thread information, check for architecture
845 * specific fallback if available
846 */
847 if (is_threaded < 0)
848 is_threaded = arch_cpu_is_threaded();
849
850 return !!is_threaded;
851}
852
853/*
854 * Propagate the topology information of the processor_topology_node tree to the
855 * cpu_topology array.
856 */
857__weak int __init parse_acpi_topology(void)
858{
859 unsigned int max_smt_thread_num = 1;
860 struct cpu_smt_info *entry;
861 struct xarray hetero_cpu;
862 unsigned long hetero_id;
863 int cpu, topology_id;
864
865 if (acpi_disabled)
866 return 0;
867
868 xa_init(&hetero_cpu);
869
870 for_each_possible_cpu(cpu) {
871 topology_id = find_acpi_cpu_topology(cpu, 0);
872 if (topology_id < 0)
873 return topology_id;
874
875 if (acpi_cpu_is_threaded(cpu)) {
876 cpu_topology[cpu].thread_id = topology_id;
877 topology_id = find_acpi_cpu_topology(cpu, 1);
878 cpu_topology[cpu].core_id = topology_id;
879
880 /*
881 * In the PPTT, CPUs below a node with the 'identical
882 * implementation' flag have the same number of threads.
883 * Count the number of threads for only one CPU (i.e.
884 * one core_id) among those with the same hetero_id.
885 * See the comment of find_acpi_cpu_topology_hetero_id()
886 * for more details.
887 *
888 * One entry is created for each node having:
889 * - the 'identical implementation' flag
890 * - its parent not having the flag
891 */
892 hetero_id = find_acpi_cpu_topology_hetero_id(cpu);
893 entry = xa_load(&hetero_cpu, hetero_id);
894 if (!entry) {
895 entry = kzalloc_obj(*entry);
896 WARN_ON_ONCE(!entry);
897
898 if (entry) {
899 entry->core_id = topology_id;
900 entry->thread_num = 1;
901 xa_store(&hetero_cpu, hetero_id,
902 entry, GFP_KERNEL);
903 }
904 } else if (entry->core_id == topology_id) {
905 entry->thread_num++;
906 }
907 } else {
908 cpu_topology[cpu].thread_id = -1;
909 cpu_topology[cpu].core_id = topology_id;
910 }
911 topology_id = find_acpi_cpu_topology_cluster(cpu);
912 cpu_topology[cpu].cluster_id = topology_id;
913 topology_id = find_acpi_cpu_topology_package(cpu);
914 cpu_topology[cpu].package_id = topology_id;
915 }
916
917 /*
918 * This is a short loop since the number of XArray elements is the
919 * number of heterogeneous CPU clusters. On a homogeneous system
920 * there's only one entry in the XArray.
921 */
922 xa_for_each(&hetero_cpu, hetero_id, entry) {
923 max_smt_thread_num = max(max_smt_thread_num, entry->thread_num);
924 xa_erase(&hetero_cpu, hetero_id);
925 kfree(entry);
926 }
927
928 cpu_smt_set_num_threads(max_smt_thread_num, max_smt_thread_num);
929 xa_destroy(&hetero_cpu);
930 return 0;
931}
932
933void __init init_cpu_topology(void)
934{
935 int cpu, ret;
936
937 reset_cpu_topology();
938 ret = parse_acpi_topology();
939 if (!ret)
940 ret = of_have_populated_dt() && parse_dt_topology();
941
942 if (ret) {
943 /*
944 * Discard anything that was parsed if we hit an error so we
945 * don't use partial information. But do not return yet to give
946 * arch-specific early cache level detection a chance to run.
947 */
948 reset_cpu_topology();
949 }
950
951 for_each_possible_cpu(cpu) {
952 ret = fetch_cache_info(cpu);
953 if (!ret)
954 continue;
955 else if (ret != -ENOENT)
956 pr_err("Early cacheinfo failed, ret = %d\n", ret);
957 return;
958 }
959}
960
961void store_cpu_topology(unsigned int cpuid)
962{
963 struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
964
965 if (cpuid_topo->package_id != -1)
966 goto topology_populated;
967
968 cpuid_topo->thread_id = -1;
969 cpuid_topo->core_id = cpuid;
970 cpuid_topo->package_id = cpu_to_node(cpuid);
971
972 pr_debug("CPU%u: package %d core %d thread %d\n",
973 cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
974 cpuid_topo->thread_id);
975
976topology_populated:
977 update_siblings_masks(cpuid);
978}
979#endif