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
28 static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data);
29 static struct cpumask scale_freq_counters_mask;
30 static bool scale_freq_invariant;
31 DEFINE_PER_CPU(unsigned long, capacity_freq_ref) = 1;
32 EXPORT_PER_CPU_SYMBOL_GPL(capacity_freq_ref);
33
supports_scale_freq_counters(const struct cpumask * cpus)34 static bool supports_scale_freq_counters(const struct cpumask *cpus)
35 {
36 return cpumask_subset(cpus, &scale_freq_counters_mask);
37 }
38
topology_scale_freq_invariant(void)39 bool topology_scale_freq_invariant(void)
40 {
41 return cpufreq_supports_freq_invariance() ||
42 supports_scale_freq_counters(cpu_online_mask);
43 }
44
update_scale_freq_invariant(bool status)45 static 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
topology_set_scale_freq_source(struct scale_freq_data * data,const struct cpumask * cpus)62 void 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 }
91 EXPORT_SYMBOL_GPL(topology_set_scale_freq_source);
92
topology_clear_scale_freq_source(enum scale_freq_source source,const struct cpumask * cpus)93 void 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 }
120 EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source);
121
topology_scale_freq_tick(void)122 void 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
130 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
131 EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale);
132
topology_set_freq_scale(const struct cpumask * cpus,unsigned long cur_freq,unsigned long max_freq)133 void 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
156 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
157 EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale);
158
topology_set_cpu_scale(unsigned int cpu,unsigned long capacity)159 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
160 {
161 per_cpu(cpu_scale, cpu) = capacity;
162 }
163
164 DEFINE_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 */
topology_update_hw_pressure(const struct cpumask * cpus,unsigned long capped_freq)180 void 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 }
207 EXPORT_SYMBOL_GPL(topology_update_hw_pressure);
208
cpu_capacity_show(struct device * dev,struct device_attribute * attr,char * buf)209 static 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
218 static void update_topology_flags_workfn(struct work_struct *work);
219 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
220
221 static DEVICE_ATTR_RO(cpu_capacity);
222
cpu_capacity_sysctl_add(unsigned int cpu)223 static 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
cpu_capacity_sysctl_remove(unsigned int cpu)235 static 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
register_cpu_capacity_sysctl(void)247 static 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 }
254 subsys_initcall(register_cpu_capacity_sysctl);
255
256 static int update_topology;
257
topology_update_cpu_topology(void)258 int 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 */
update_topology_flags_workfn(struct work_struct * work)267 static 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
275 static u32 *raw_capacity;
276
free_raw_capacity(void)277 static int free_raw_capacity(void)
278 {
279 kfree(raw_capacity);
280 raw_capacity = NULL;
281
282 return 0;
283 }
284
topology_normalize_cpu_scale(void)285 void 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
topology_parse_cpu_capacity(struct device_node * cpu_node,int cpu)311 bool __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
freq_inv_set_max_ratio(int cpu,u64 max_rate)362 void __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
topology_init_cpu_capacity_cppc(void)369 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
417 exit:
418 free_raw_capacity();
419 }
420 #endif
421
422 #ifdef CONFIG_CPU_FREQ
423 static cpumask_var_t cpus_to_visit;
424 static void parsing_done_workfn(struct work_struct *work);
425 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
426
427 static int
init_cpu_capacity_callback(struct notifier_block * nb,unsigned long val,void * data)428 init_cpu_capacity_callback(struct notifier_block *nb,
429 unsigned long val,
430 void *data)
431 {
432 struct cpufreq_policy *policy = data;
433 int cpu;
434
435 if (val != CPUFREQ_CREATE_POLICY)
436 return 0;
437
438 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
439 cpumask_pr_args(policy->related_cpus),
440 cpumask_pr_args(cpus_to_visit));
441
442 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
443
444 for_each_cpu(cpu, policy->related_cpus) {
445 per_cpu(capacity_freq_ref, cpu) = policy->cpuinfo.max_freq;
446 freq_inv_set_max_ratio(cpu,
447 per_cpu(capacity_freq_ref, cpu) * HZ_PER_KHZ);
448 }
449
450 if (cpumask_empty(cpus_to_visit)) {
451 if (raw_capacity) {
452 topology_normalize_cpu_scale();
453 schedule_work(&update_topology_flags_work);
454 free_raw_capacity();
455 }
456 pr_debug("cpu_capacity: parsing done\n");
457 schedule_work(&parsing_done_work);
458 }
459
460 return 0;
461 }
462
463 static struct notifier_block init_cpu_capacity_notifier = {
464 .notifier_call = init_cpu_capacity_callback,
465 };
466
register_cpufreq_notifier(void)467 static int __init register_cpufreq_notifier(void)
468 {
469 int ret;
470
471 /*
472 * On ACPI-based systems skip registering cpufreq notifier as cpufreq
473 * information is not needed for cpu capacity initialization.
474 */
475 if (!acpi_disabled)
476 return -EINVAL;
477
478 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
479 return -ENOMEM;
480
481 cpumask_copy(cpus_to_visit, cpu_possible_mask);
482
483 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
484 CPUFREQ_POLICY_NOTIFIER);
485
486 if (ret)
487 free_cpumask_var(cpus_to_visit);
488
489 return ret;
490 }
491 core_initcall(register_cpufreq_notifier);
492
parsing_done_workfn(struct work_struct * work)493 static void parsing_done_workfn(struct work_struct *work)
494 {
495 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
496 CPUFREQ_POLICY_NOTIFIER);
497 free_cpumask_var(cpus_to_visit);
498 }
499
500 #else
501 core_initcall(free_raw_capacity);
502 #endif
503
504 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
505 /*
506 * This function returns the logic cpu number of the node.
507 * There are basically three kinds of return values:
508 * (1) logic cpu number which is > 0.
509 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
510 * there is no possible logical CPU in the kernel to match. This happens
511 * when CONFIG_NR_CPUS is configure to be smaller than the number of
512 * CPU nodes in DT. We need to just ignore this case.
513 * (3) -1 if the node does not exist in the device tree
514 */
get_cpu_for_node(struct device_node * node)515 static int __init get_cpu_for_node(struct device_node *node)
516 {
517 int cpu;
518 struct device_node *cpu_node __free(device_node) =
519 of_parse_phandle(node, "cpu", 0);
520
521 if (!cpu_node)
522 return -1;
523
524 cpu = of_cpu_node_to_id(cpu_node);
525 if (cpu >= 0)
526 topology_parse_cpu_capacity(cpu_node, cpu);
527 else
528 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
529 cpu_node, cpumask_pr_args(cpu_possible_mask));
530
531 return cpu;
532 }
533
parse_core(struct device_node * core,int package_id,int cluster_id,int core_id)534 static int __init parse_core(struct device_node *core, int package_id,
535 int cluster_id, int core_id)
536 {
537 char name[20];
538 bool leaf = true;
539 int i = 0;
540 int cpu;
541
542 do {
543 snprintf(name, sizeof(name), "thread%d", i);
544 struct device_node *t __free(device_node) =
545 of_get_child_by_name(core, name);
546
547 if (!t)
548 break;
549
550 leaf = false;
551 cpu = get_cpu_for_node(t);
552 if (cpu >= 0) {
553 cpu_topology[cpu].package_id = package_id;
554 cpu_topology[cpu].cluster_id = cluster_id;
555 cpu_topology[cpu].core_id = core_id;
556 cpu_topology[cpu].thread_id = i;
557 } else if (cpu != -ENODEV) {
558 pr_err("%pOF: Can't get CPU for thread\n", t);
559 return -EINVAL;
560 }
561 i++;
562 } while (1);
563
564 cpu = get_cpu_for_node(core);
565 if (cpu >= 0) {
566 if (!leaf) {
567 pr_err("%pOF: Core has both threads and CPU\n",
568 core);
569 return -EINVAL;
570 }
571
572 cpu_topology[cpu].package_id = package_id;
573 cpu_topology[cpu].cluster_id = cluster_id;
574 cpu_topology[cpu].core_id = core_id;
575 } else if (leaf && cpu != -ENODEV) {
576 pr_err("%pOF: Can't get CPU for leaf core\n", core);
577 return -EINVAL;
578 }
579
580 return 0;
581 }
582
parse_cluster(struct device_node * cluster,int package_id,int cluster_id,int depth)583 static int __init parse_cluster(struct device_node *cluster, int package_id,
584 int cluster_id, int depth)
585 {
586 char name[20];
587 bool leaf = true;
588 bool has_cores = false;
589 int core_id = 0;
590 int i, ret;
591
592 /*
593 * First check for child clusters; we currently ignore any
594 * information about the nesting of clusters and present the
595 * scheduler with a flat list of them.
596 */
597 i = 0;
598 do {
599 snprintf(name, sizeof(name), "cluster%d", i);
600 struct device_node *c __free(device_node) =
601 of_get_child_by_name(cluster, name);
602
603 if (!c)
604 break;
605
606 leaf = false;
607 ret = parse_cluster(c, package_id, i, depth + 1);
608 if (depth > 0)
609 pr_warn("Topology for clusters of clusters not yet supported\n");
610 if (ret != 0)
611 return ret;
612 i++;
613 } while (1);
614
615 /* Now check for cores */
616 i = 0;
617 do {
618 snprintf(name, sizeof(name), "core%d", i);
619 struct device_node *c __free(device_node) =
620 of_get_child_by_name(cluster, name);
621
622 if (!c)
623 break;
624
625 has_cores = true;
626
627 if (depth == 0) {
628 pr_err("%pOF: cpu-map children should be clusters\n", c);
629 return -EINVAL;
630 }
631
632 if (leaf) {
633 ret = parse_core(c, package_id, cluster_id, core_id++);
634 if (ret != 0)
635 return ret;
636 } else {
637 pr_err("%pOF: Non-leaf cluster with core %s\n",
638 cluster, name);
639 return -EINVAL;
640 }
641
642 i++;
643 } while (1);
644
645 if (leaf && !has_cores)
646 pr_warn("%pOF: empty cluster\n", cluster);
647
648 return 0;
649 }
650
parse_socket(struct device_node * socket)651 static int __init parse_socket(struct device_node *socket)
652 {
653 char name[20];
654 bool has_socket = false;
655 int package_id = 0, ret;
656
657 do {
658 snprintf(name, sizeof(name), "socket%d", package_id);
659 struct device_node *c __free(device_node) =
660 of_get_child_by_name(socket, name);
661
662 if (!c)
663 break;
664
665 has_socket = true;
666 ret = parse_cluster(c, package_id, -1, 0);
667 if (ret != 0)
668 return ret;
669
670 package_id++;
671 } while (1);
672
673 if (!has_socket)
674 ret = parse_cluster(socket, 0, -1, 0);
675
676 return ret;
677 }
678
parse_dt_topology(void)679 static int __init parse_dt_topology(void)
680 {
681 int ret = 0;
682 int cpu;
683 struct device_node *cn __free(device_node) =
684 of_find_node_by_path("/cpus");
685
686 if (!cn) {
687 pr_err("No CPU information found in DT\n");
688 return 0;
689 }
690
691 /*
692 * When topology is provided cpu-map is essentially a root
693 * cluster with restricted subnodes.
694 */
695 struct device_node *map __free(device_node) =
696 of_get_child_by_name(cn, "cpu-map");
697
698 if (!map)
699 return ret;
700
701 ret = parse_socket(map);
702 if (ret != 0)
703 return ret;
704
705 topology_normalize_cpu_scale();
706
707 /*
708 * Check that all cores are in the topology; the SMP code will
709 * only mark cores described in the DT as possible.
710 */
711 for_each_possible_cpu(cpu)
712 if (cpu_topology[cpu].package_id < 0) {
713 return -EINVAL;
714 }
715
716 return ret;
717 }
718 #endif
719
720 /*
721 * cpu topology table
722 */
723 struct cpu_topology cpu_topology[NR_CPUS];
724 EXPORT_SYMBOL_GPL(cpu_topology);
725
cpu_coregroup_mask(int cpu)726 const struct cpumask *cpu_coregroup_mask(int cpu)
727 {
728 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
729
730 /* Find the smaller of NUMA, core or LLC siblings */
731 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
732 /* not numa in package, lets use the package siblings */
733 core_mask = &cpu_topology[cpu].core_sibling;
734 }
735
736 if (last_level_cache_is_valid(cpu)) {
737 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
738 core_mask = &cpu_topology[cpu].llc_sibling;
739 }
740
741 /*
742 * For systems with no shared cpu-side LLC but with clusters defined,
743 * extend core_mask to cluster_siblings. The sched domain builder will
744 * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled.
745 */
746 if (IS_ENABLED(CONFIG_SCHED_CLUSTER) &&
747 cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling))
748 core_mask = &cpu_topology[cpu].cluster_sibling;
749
750 return core_mask;
751 }
752
cpu_clustergroup_mask(int cpu)753 const struct cpumask *cpu_clustergroup_mask(int cpu)
754 {
755 /*
756 * Forbid cpu_clustergroup_mask() to span more or the same CPUs as
757 * cpu_coregroup_mask().
758 */
759 if (cpumask_subset(cpu_coregroup_mask(cpu),
760 &cpu_topology[cpu].cluster_sibling))
761 return topology_sibling_cpumask(cpu);
762
763 return &cpu_topology[cpu].cluster_sibling;
764 }
765
update_siblings_masks(unsigned int cpuid)766 void update_siblings_masks(unsigned int cpuid)
767 {
768 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
769 int cpu, ret;
770
771 ret = detect_cache_attributes(cpuid);
772 if (ret && ret != -ENOENT)
773 pr_info("Early cacheinfo allocation failed, ret = %d\n", ret);
774
775 /* update core and thread sibling masks */
776 for_each_online_cpu(cpu) {
777 cpu_topo = &cpu_topology[cpu];
778
779 if (last_level_cache_is_shared(cpu, cpuid)) {
780 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
781 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
782 }
783
784 if (cpuid_topo->package_id != cpu_topo->package_id)
785 continue;
786
787 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
788 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
789
790 if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
791 continue;
792
793 if (cpuid_topo->cluster_id >= 0) {
794 cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling);
795 cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling);
796 }
797
798 if (cpuid_topo->core_id != cpu_topo->core_id)
799 continue;
800
801 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
802 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
803 }
804 }
805
clear_cpu_topology(int cpu)806 static void clear_cpu_topology(int cpu)
807 {
808 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
809
810 cpumask_clear(&cpu_topo->llc_sibling);
811 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
812
813 cpumask_clear(&cpu_topo->cluster_sibling);
814 cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling);
815
816 cpumask_clear(&cpu_topo->core_sibling);
817 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
818 cpumask_clear(&cpu_topo->thread_sibling);
819 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
820 }
821
reset_cpu_topology(void)822 void __init reset_cpu_topology(void)
823 {
824 unsigned int cpu;
825
826 for_each_possible_cpu(cpu) {
827 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
828
829 cpu_topo->thread_id = -1;
830 cpu_topo->core_id = -1;
831 cpu_topo->cluster_id = -1;
832 cpu_topo->package_id = -1;
833
834 clear_cpu_topology(cpu);
835 }
836 }
837
remove_cpu_topology(unsigned int cpu)838 void remove_cpu_topology(unsigned int cpu)
839 {
840 int sibling;
841
842 for_each_cpu(sibling, topology_core_cpumask(cpu))
843 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
844 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
845 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
846 for_each_cpu(sibling, topology_cluster_cpumask(cpu))
847 cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling));
848 for_each_cpu(sibling, topology_llc_cpumask(cpu))
849 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
850
851 clear_cpu_topology(cpu);
852 }
853
parse_acpi_topology(void)854 __weak int __init parse_acpi_topology(void)
855 {
856 return 0;
857 }
858
859 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
init_cpu_topology(void)860 void __init init_cpu_topology(void)
861 {
862 int cpu, ret;
863
864 reset_cpu_topology();
865 ret = parse_acpi_topology();
866 if (!ret)
867 ret = of_have_populated_dt() && parse_dt_topology();
868
869 if (ret) {
870 /*
871 * Discard anything that was parsed if we hit an error so we
872 * don't use partial information. But do not return yet to give
873 * arch-specific early cache level detection a chance to run.
874 */
875 reset_cpu_topology();
876 }
877
878 for_each_possible_cpu(cpu) {
879 ret = fetch_cache_info(cpu);
880 if (!ret)
881 continue;
882 else if (ret != -ENOENT)
883 pr_err("Early cacheinfo failed, ret = %d\n", ret);
884 return;
885 }
886 }
887
store_cpu_topology(unsigned int cpuid)888 void store_cpu_topology(unsigned int cpuid)
889 {
890 struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
891
892 if (cpuid_topo->package_id != -1)
893 goto topology_populated;
894
895 cpuid_topo->thread_id = -1;
896 cpuid_topo->core_id = cpuid;
897 cpuid_topo->package_id = cpu_to_node(cpuid);
898
899 pr_debug("CPU%u: package %d core %d thread %d\n",
900 cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
901 cpuid_topo->thread_id);
902
903 topology_populated:
904 update_siblings_masks(cpuid);
905 }
906 #endif
907