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/cpu.h>
11 #include <linux/cpufreq.h>
12 #include <linux/device.h>
13 #include <linux/of.h>
14 #include <linux/slab.h>
15 #include <linux/string.h>
16 #include <linux/sched/topology.h>
17 #include <linux/cpuset.h>
18 #include <linux/cpumask.h>
19 #include <linux/init.h>
20 #include <linux/percpu.h>
21 #include <linux/sched.h>
22 #include <linux/smp.h>
23 #include <trace/hooks/topology.h>
24
topology_scale_freq_invariant(void)25 bool topology_scale_freq_invariant(void)
26 {
27 return cpufreq_supports_freq_invariance() ||
28 arch_freq_counters_available(cpu_online_mask);
29 }
30
arch_freq_counters_available(const struct cpumask * cpus)31 __weak bool arch_freq_counters_available(const struct cpumask *cpus)
32 {
33 return false;
34 }
35 DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
36 EXPORT_PER_CPU_SYMBOL_GPL(freq_scale);
37
topology_set_freq_scale(const struct cpumask * cpus,unsigned long cur_freq,unsigned long max_freq)38 void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq,
39 unsigned long max_freq)
40 {
41 unsigned long scale;
42 int i;
43
44 if (WARN_ON_ONCE(!cur_freq || !max_freq))
45 return;
46
47 /*
48 * If the use of counters for FIE is enabled, just return as we don't
49 * want to update the scale factor with information from CPUFREQ.
50 * Instead the scale factor will be updated from arch_scale_freq_tick.
51 */
52 if (arch_freq_counters_available(cpus))
53 return;
54
55 scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq;
56
57 trace_android_vh_arch_set_freq_scale(cpus, cur_freq, max_freq, &scale);
58
59 for_each_cpu(i, cpus)
60 per_cpu(freq_scale, i) = scale;
61 }
62
63 DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
64 EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale);
65
topology_set_cpu_scale(unsigned int cpu,unsigned long capacity)66 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
67 {
68 per_cpu(cpu_scale, cpu) = capacity;
69 }
70
71 DEFINE_PER_CPU(unsigned long, thermal_pressure);
72 EXPORT_PER_CPU_SYMBOL_GPL(thermal_pressure);
73
topology_set_thermal_pressure(const struct cpumask * cpus,unsigned long th_pressure)74 void topology_set_thermal_pressure(const struct cpumask *cpus,
75 unsigned long th_pressure)
76 {
77 int cpu;
78
79 for_each_cpu(cpu, cpus)
80 WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure);
81 }
82 EXPORT_SYMBOL_GPL(topology_set_thermal_pressure);
83
cpu_capacity_show(struct device * dev,struct device_attribute * attr,char * buf)84 static ssize_t cpu_capacity_show(struct device *dev,
85 struct device_attribute *attr,
86 char *buf)
87 {
88 struct cpu *cpu = container_of(dev, struct cpu, dev);
89
90 return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id));
91 }
92
93 static void update_topology_flags_workfn(struct work_struct *work);
94 static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn);
95
96 static DEVICE_ATTR_RO(cpu_capacity);
97
register_cpu_capacity_sysctl(void)98 static int register_cpu_capacity_sysctl(void)
99 {
100 int i;
101 struct device *cpu;
102
103 for_each_possible_cpu(i) {
104 cpu = get_cpu_device(i);
105 if (!cpu) {
106 pr_err("%s: too early to get CPU%d device!\n",
107 __func__, i);
108 continue;
109 }
110 device_create_file(cpu, &dev_attr_cpu_capacity);
111 }
112
113 return 0;
114 }
115 subsys_initcall(register_cpu_capacity_sysctl);
116
117 static int update_topology;
118 bool topology_update_done;
119 EXPORT_SYMBOL_GPL(topology_update_done);
120
topology_update_cpu_topology(void)121 int topology_update_cpu_topology(void)
122 {
123 return update_topology;
124 }
125
126 /*
127 * Updating the sched_domains can't be done directly from cpufreq callbacks
128 * due to locking, so queue the work for later.
129 */
update_topology_flags_workfn(struct work_struct * work)130 static void update_topology_flags_workfn(struct work_struct *work)
131 {
132 update_topology = 1;
133 rebuild_sched_domains();
134 topology_update_done = true;
135 trace_android_vh_update_topology_flags_workfn(NULL);
136 pr_debug("sched_domain hierarchy rebuilt, flags updated\n");
137 update_topology = 0;
138 }
139
140 static DEFINE_PER_CPU(u32, freq_factor) = 1;
141 static u32 *raw_capacity;
142
free_raw_capacity(void)143 static int free_raw_capacity(void)
144 {
145 kfree(raw_capacity);
146 raw_capacity = NULL;
147
148 return 0;
149 }
150
topology_normalize_cpu_scale(void)151 void topology_normalize_cpu_scale(void)
152 {
153 u64 capacity;
154 u64 capacity_scale;
155 int cpu;
156
157 if (!raw_capacity)
158 return;
159
160 capacity_scale = 1;
161 for_each_possible_cpu(cpu) {
162 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
163 capacity_scale = max(capacity, capacity_scale);
164 }
165
166 pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale);
167 for_each_possible_cpu(cpu) {
168 capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu);
169 capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT,
170 capacity_scale);
171 topology_set_cpu_scale(cpu, capacity);
172 pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
173 cpu, topology_get_cpu_scale(cpu));
174 }
175 }
176
topology_parse_cpu_capacity(struct device_node * cpu_node,int cpu)177 bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu)
178 {
179 struct clk *cpu_clk;
180 static bool cap_parsing_failed;
181 int ret;
182 u32 cpu_capacity;
183
184 if (cap_parsing_failed)
185 return false;
186
187 ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz",
188 &cpu_capacity);
189 if (!ret) {
190 if (!raw_capacity) {
191 raw_capacity = kcalloc(num_possible_cpus(),
192 sizeof(*raw_capacity),
193 GFP_KERNEL);
194 if (!raw_capacity) {
195 cap_parsing_failed = true;
196 return false;
197 }
198 }
199 raw_capacity[cpu] = cpu_capacity;
200 pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n",
201 cpu_node, raw_capacity[cpu]);
202
203 /*
204 * Update freq_factor for calculating early boot cpu capacities.
205 * For non-clk CPU DVFS mechanism, there's no way to get the
206 * frequency value now, assuming they are running at the same
207 * frequency (by keeping the initial freq_factor value).
208 */
209 cpu_clk = of_clk_get(cpu_node, 0);
210 if (!PTR_ERR_OR_ZERO(cpu_clk)) {
211 per_cpu(freq_factor, cpu) =
212 clk_get_rate(cpu_clk) / 1000;
213 clk_put(cpu_clk);
214 }
215 } else {
216 if (raw_capacity) {
217 pr_err("cpu_capacity: missing %pOF raw capacity\n",
218 cpu_node);
219 pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
220 }
221 cap_parsing_failed = true;
222 free_raw_capacity();
223 }
224
225 return !ret;
226 }
227
228 #ifdef CONFIG_CPU_FREQ
229 static cpumask_var_t cpus_to_visit;
230 static void parsing_done_workfn(struct work_struct *work);
231 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
232
233 static int
init_cpu_capacity_callback(struct notifier_block * nb,unsigned long val,void * data)234 init_cpu_capacity_callback(struct notifier_block *nb,
235 unsigned long val,
236 void *data)
237 {
238 struct cpufreq_policy *policy = data;
239 int cpu;
240
241 if (!raw_capacity)
242 return 0;
243
244 if (val != CPUFREQ_CREATE_POLICY)
245 return 0;
246
247 pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
248 cpumask_pr_args(policy->related_cpus),
249 cpumask_pr_args(cpus_to_visit));
250
251 cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus);
252
253 for_each_cpu(cpu, policy->related_cpus)
254 per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000;
255
256 if (cpumask_empty(cpus_to_visit)) {
257 topology_normalize_cpu_scale();
258 schedule_work(&update_topology_flags_work);
259 free_raw_capacity();
260 pr_debug("cpu_capacity: parsing done\n");
261 schedule_work(&parsing_done_work);
262 }
263
264 return 0;
265 }
266
267 static struct notifier_block init_cpu_capacity_notifier = {
268 .notifier_call = init_cpu_capacity_callback,
269 };
270
register_cpufreq_notifier(void)271 static int __init register_cpufreq_notifier(void)
272 {
273 int ret;
274
275 /*
276 * on ACPI-based systems we need to use the default cpu capacity
277 * until we have the necessary code to parse the cpu capacity, so
278 * skip registering cpufreq notifier.
279 */
280 if (!acpi_disabled || !raw_capacity)
281 return -EINVAL;
282
283 if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL))
284 return -ENOMEM;
285
286 cpumask_copy(cpus_to_visit, cpu_possible_mask);
287
288 ret = cpufreq_register_notifier(&init_cpu_capacity_notifier,
289 CPUFREQ_POLICY_NOTIFIER);
290
291 if (ret)
292 free_cpumask_var(cpus_to_visit);
293
294 return ret;
295 }
296 core_initcall(register_cpufreq_notifier);
297
parsing_done_workfn(struct work_struct * work)298 static void parsing_done_workfn(struct work_struct *work)
299 {
300 cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
301 CPUFREQ_POLICY_NOTIFIER);
302 free_cpumask_var(cpus_to_visit);
303 }
304
305 #else
306 core_initcall(free_raw_capacity);
307 #endif
308
309 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
310 /*
311 * This function returns the logic cpu number of the node.
312 * There are basically three kinds of return values:
313 * (1) logic cpu number which is > 0.
314 * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but
315 * there is no possible logical CPU in the kernel to match. This happens
316 * when CONFIG_NR_CPUS is configure to be smaller than the number of
317 * CPU nodes in DT. We need to just ignore this case.
318 * (3) -1 if the node does not exist in the device tree
319 */
get_cpu_for_node(struct device_node * node)320 static int __init get_cpu_for_node(struct device_node *node)
321 {
322 struct device_node *cpu_node;
323 int cpu;
324
325 cpu_node = of_parse_phandle(node, "cpu", 0);
326 if (!cpu_node)
327 return -1;
328
329 cpu = of_cpu_node_to_id(cpu_node);
330 if (cpu >= 0)
331 topology_parse_cpu_capacity(cpu_node, cpu);
332 else
333 pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n",
334 cpu_node, cpumask_pr_args(cpu_possible_mask));
335
336 of_node_put(cpu_node);
337 return cpu;
338 }
339
parse_core(struct device_node * core,int package_id,int core_id)340 static int __init parse_core(struct device_node *core, int package_id,
341 int core_id)
342 {
343 char name[20];
344 bool leaf = true;
345 int i = 0;
346 int cpu;
347 struct device_node *t;
348
349 do {
350 snprintf(name, sizeof(name), "thread%d", i);
351 t = of_get_child_by_name(core, name);
352 if (t) {
353 leaf = false;
354 cpu = get_cpu_for_node(t);
355 if (cpu >= 0) {
356 cpu_topology[cpu].package_id = package_id;
357 cpu_topology[cpu].core_id = core_id;
358 cpu_topology[cpu].thread_id = i;
359 } else if (cpu != -ENODEV) {
360 pr_err("%pOF: Can't get CPU for thread\n", t);
361 of_node_put(t);
362 return -EINVAL;
363 }
364 of_node_put(t);
365 }
366 i++;
367 } while (t);
368
369 cpu = get_cpu_for_node(core);
370 if (cpu >= 0) {
371 if (!leaf) {
372 pr_err("%pOF: Core has both threads and CPU\n",
373 core);
374 return -EINVAL;
375 }
376
377 cpu_topology[cpu].package_id = package_id;
378 cpu_topology[cpu].core_id = core_id;
379 } else if (leaf && cpu != -ENODEV) {
380 pr_err("%pOF: Can't get CPU for leaf core\n", core);
381 return -EINVAL;
382 }
383
384 return 0;
385 }
386
parse_cluster(struct device_node * cluster,int depth)387 static int __init parse_cluster(struct device_node *cluster, int depth)
388 {
389 char name[20];
390 bool leaf = true;
391 bool has_cores = false;
392 struct device_node *c;
393 static int package_id __initdata;
394 int core_id = 0;
395 int i, ret;
396
397 /*
398 * First check for child clusters; we currently ignore any
399 * information about the nesting of clusters and present the
400 * scheduler with a flat list of them.
401 */
402 i = 0;
403 do {
404 snprintf(name, sizeof(name), "cluster%d", i);
405 c = of_get_child_by_name(cluster, name);
406 if (c) {
407 leaf = false;
408 ret = parse_cluster(c, depth + 1);
409 of_node_put(c);
410 if (ret != 0)
411 return ret;
412 }
413 i++;
414 } while (c);
415
416 /* Now check for cores */
417 i = 0;
418 do {
419 snprintf(name, sizeof(name), "core%d", i);
420 c = of_get_child_by_name(cluster, name);
421 if (c) {
422 has_cores = true;
423
424 if (depth == 0) {
425 pr_err("%pOF: cpu-map children should be clusters\n",
426 c);
427 of_node_put(c);
428 return -EINVAL;
429 }
430
431 if (leaf) {
432 ret = parse_core(c, package_id, core_id++);
433 } else {
434 pr_err("%pOF: Non-leaf cluster with core %s\n",
435 cluster, name);
436 ret = -EINVAL;
437 }
438
439 of_node_put(c);
440 if (ret != 0)
441 return ret;
442 }
443 i++;
444 } while (c);
445
446 if (leaf && !has_cores)
447 pr_warn("%pOF: empty cluster\n", cluster);
448
449 if (leaf)
450 package_id++;
451
452 return 0;
453 }
454
parse_dt_topology(void)455 static int __init parse_dt_topology(void)
456 {
457 struct device_node *cn, *map;
458 int ret = 0;
459 int cpu;
460
461 cn = of_find_node_by_path("/cpus");
462 if (!cn) {
463 pr_err("No CPU information found in DT\n");
464 return 0;
465 }
466
467 /*
468 * When topology is provided cpu-map is essentially a root
469 * cluster with restricted subnodes.
470 */
471 map = of_get_child_by_name(cn, "cpu-map");
472 if (!map)
473 goto out;
474
475 ret = parse_cluster(map, 0);
476 if (ret != 0)
477 goto out_map;
478
479 topology_normalize_cpu_scale();
480
481 /*
482 * Check that all cores are in the topology; the SMP code will
483 * only mark cores described in the DT as possible.
484 */
485 for_each_possible_cpu(cpu)
486 if (cpu_topology[cpu].package_id == -1)
487 ret = -EINVAL;
488
489 out_map:
490 of_node_put(map);
491 out:
492 of_node_put(cn);
493 return ret;
494 }
495 #endif
496
497 /*
498 * cpu topology table
499 */
500 struct cpu_topology cpu_topology[NR_CPUS];
501 EXPORT_SYMBOL_GPL(cpu_topology);
502
cpu_coregroup_mask(int cpu)503 const struct cpumask *cpu_coregroup_mask(int cpu)
504 {
505 const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu));
506
507 /* Find the smaller of NUMA, core or LLC siblings */
508 if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) {
509 /* not numa in package, lets use the package siblings */
510 core_mask = &cpu_topology[cpu].core_sibling;
511 }
512 if (cpu_topology[cpu].llc_id != -1) {
513 if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask))
514 core_mask = &cpu_topology[cpu].llc_sibling;
515 }
516
517 return core_mask;
518 }
519
update_siblings_masks(unsigned int cpuid)520 void update_siblings_masks(unsigned int cpuid)
521 {
522 struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
523 int cpu;
524
525 /* update core and thread sibling masks */
526 for_each_online_cpu(cpu) {
527 cpu_topo = &cpu_topology[cpu];
528
529 if (cpu_topo->llc_id != -1 && cpuid_topo->llc_id == cpu_topo->llc_id) {
530 cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling);
531 cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling);
532 }
533
534 if (cpuid_topo->package_id != cpu_topo->package_id)
535 continue;
536
537 cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
538 cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
539
540 if (cpuid_topo->core_id != cpu_topo->core_id)
541 continue;
542
543 cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
544 cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
545 }
546 }
547
clear_cpu_topology(int cpu)548 static void clear_cpu_topology(int cpu)
549 {
550 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
551
552 cpumask_clear(&cpu_topo->llc_sibling);
553 cpumask_set_cpu(cpu, &cpu_topo->llc_sibling);
554
555 cpumask_clear(&cpu_topo->core_sibling);
556 cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
557 cpumask_clear(&cpu_topo->thread_sibling);
558 cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
559 }
560
reset_cpu_topology(void)561 void __init reset_cpu_topology(void)
562 {
563 unsigned int cpu;
564
565 for_each_possible_cpu(cpu) {
566 struct cpu_topology *cpu_topo = &cpu_topology[cpu];
567
568 cpu_topo->thread_id = -1;
569 cpu_topo->core_id = -1;
570 cpu_topo->package_id = -1;
571 cpu_topo->llc_id = -1;
572
573 clear_cpu_topology(cpu);
574 }
575 }
576
remove_cpu_topology(unsigned int cpu)577 void remove_cpu_topology(unsigned int cpu)
578 {
579 int sibling;
580
581 for_each_cpu(sibling, topology_core_cpumask(cpu))
582 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
583 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
584 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
585 for_each_cpu(sibling, topology_llc_cpumask(cpu))
586 cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling));
587
588 clear_cpu_topology(cpu);
589 }
590
parse_acpi_topology(void)591 __weak int __init parse_acpi_topology(void)
592 {
593 return 0;
594 }
595
596 #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV)
init_cpu_topology(void)597 void __init init_cpu_topology(void)
598 {
599 reset_cpu_topology();
600
601 /*
602 * Discard anything that was parsed if we hit an error so we
603 * don't use partial information.
604 */
605 if (parse_acpi_topology())
606 reset_cpu_topology();
607 else if (of_have_populated_dt() && parse_dt_topology())
608 reset_cpu_topology();
609 }
610
store_cpu_topology(unsigned int cpuid)611 void store_cpu_topology(unsigned int cpuid)
612 {
613 struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
614
615 if (cpuid_topo->package_id != -1)
616 goto topology_populated;
617
618 cpuid_topo->thread_id = -1;
619 cpuid_topo->core_id = cpuid;
620 cpuid_topo->package_id = cpu_to_node(cpuid);
621
622 pr_debug("CPU%u: package %d core %d thread %d\n",
623 cpuid, cpuid_topo->package_id, cpuid_topo->core_id,
624 cpuid_topo->thread_id);
625
626 topology_populated:
627 update_siblings_masks(cpuid);
628 }
629 #endif
630