Commit | Line | Data |
---|---|---|
6ee97d35 | 1 | // SPDX-License-Identifier: GPL-2.0 |
2ef7a295 JL |
2 | /* |
3 | * Arch specific cpu topology information | |
4 | * | |
5 | * Copyright (C) 2016, ARM Ltd. | |
6 | * Written by: Juri Lelli, ARM Ltd. | |
2ef7a295 JL |
7 | */ |
8 | ||
9 | #include <linux/acpi.h> | |
38db9b95 | 10 | #include <linux/cacheinfo.h> |
2ef7a295 JL |
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> | |
2ef7a295 | 16 | #include <linux/sched/topology.h> |
bb1fbdd3 | 17 | #include <linux/cpuset.h> |
60c1b220 AP |
18 | #include <linux/cpumask.h> |
19 | #include <linux/init.h> | |
83150f5d | 20 | #include <linux/rcupdate.h> |
60c1b220 | 21 | #include <linux/sched.h> |
2ef7a295 | 22 | |
c3d438ee LL |
23 | #define CREATE_TRACE_POINTS |
24 | #include <trace/events/thermal_pressure.h> | |
25 | ||
83150f5d | 26 | static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data); |
01e055c1 VK |
27 | static struct cpumask scale_freq_counters_mask; |
28 | static bool scale_freq_invariant; | |
c214f124 | 29 | static DEFINE_PER_CPU(u32, freq_factor) = 1; |
01e055c1 VK |
30 | |
31 | static bool supports_scale_freq_counters(const struct cpumask *cpus) | |
32 | { | |
33 | return cpumask_subset(cpus, &scale_freq_counters_mask); | |
34 | } | |
35 | ||
15e5d5b4 VS |
36 | bool topology_scale_freq_invariant(void) |
37 | { | |
38 | return cpufreq_supports_freq_invariance() || | |
01e055c1 VK |
39 | supports_scale_freq_counters(cpu_online_mask); |
40 | } | |
41 | ||
42 | static 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 | ||
59 | void 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 | ||
83150f5d VK |
72 | rcu_read_lock(); |
73 | ||
01e055c1 | 74 | for_each_cpu(cpu, cpus) { |
83150f5d | 75 | sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu)); |
01e055c1 VK |
76 | |
77 | /* Use ARCH provided counters whenever possible */ | |
78 | if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) { | |
83150f5d | 79 | rcu_assign_pointer(per_cpu(sft_data, cpu), data); |
01e055c1 VK |
80 | cpumask_set_cpu(cpu, &scale_freq_counters_mask); |
81 | } | |
82 | } | |
83 | ||
83150f5d VK |
84 | rcu_read_unlock(); |
85 | ||
01e055c1 | 86 | update_scale_freq_invariant(true); |
15e5d5b4 | 87 | } |
2f533958 | 88 | EXPORT_SYMBOL_GPL(topology_set_scale_freq_source); |
15e5d5b4 | 89 | |
01e055c1 VK |
90 | void topology_clear_scale_freq_source(enum scale_freq_source source, |
91 | const struct cpumask *cpus) | |
cd0ed03a | 92 | { |
01e055c1 VK |
93 | struct scale_freq_data *sfd; |
94 | int cpu; | |
95 | ||
83150f5d VK |
96 | rcu_read_lock(); |
97 | ||
01e055c1 | 98 | for_each_cpu(cpu, cpus) { |
83150f5d | 99 | sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu)); |
01e055c1 VK |
100 | |
101 | if (sfd && sfd->source == source) { | |
83150f5d | 102 | rcu_assign_pointer(per_cpu(sft_data, cpu), NULL); |
01e055c1 VK |
103 | cpumask_clear_cpu(cpu, &scale_freq_counters_mask); |
104 | } | |
105 | } | |
106 | ||
83150f5d VK |
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 | ||
01e055c1 | 115 | update_scale_freq_invariant(false); |
cd0ed03a | 116 | } |
2f533958 | 117 | EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source); |
01e055c1 VK |
118 | |
119 | void topology_scale_freq_tick(void) | |
120 | { | |
83150f5d | 121 | struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data)); |
01e055c1 VK |
122 | |
123 | if (sfd) | |
124 | sfd->set_freq_scale(); | |
125 | } | |
126 | ||
eec73529 | 127 | DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE; |
2f533958 | 128 | EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale); |
2ef7a295 | 129 | |
a20b7053 IV |
130 | void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq, |
131 | unsigned long max_freq) | |
2ef7a295 | 132 | { |
0e27c567 DE |
133 | unsigned long scale; |
134 | int i; | |
135 | ||
0a10d3fe IV |
136 | if (WARN_ON_ONCE(!cur_freq || !max_freq)) |
137 | return; | |
138 | ||
cd0ed03a IV |
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 | */ | |
01e055c1 | 144 | if (supports_scale_freq_counters(cpus)) |
cd0ed03a IV |
145 | return; |
146 | ||
0e27c567 DE |
147 | scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq; |
148 | ||
149 | for_each_cpu(i, cpus) | |
eec73529 | 150 | per_cpu(arch_freq_scale, i) = scale; |
2ef7a295 JL |
151 | } |
152 | ||
8216f588 | 153 | DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE; |
275157b3 | 154 | EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale); |
2ef7a295 | 155 | |
4ca4f26a | 156 | void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity) |
2ef7a295 JL |
157 | { |
158 | per_cpu(cpu_scale, cpu) = capacity; | |
159 | } | |
160 | ||
25980c7a VS |
161 | DEFINE_PER_CPU(unsigned long, thermal_pressure); |
162 | ||
c214f124 LL |
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 | */ | |
177 | void topology_update_thermal_pressure(const struct cpumask *cpus, | |
178 | unsigned long capped_freq) | |
179 | { | |
7e97b3dc | 180 | unsigned long max_capacity, capacity, th_pressure; |
c214f124 LL |
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 | ||
7e97b3dc LL |
200 | th_pressure = max_capacity - capacity; |
201 | ||
c3d438ee LL |
202 | trace_thermal_pressure_update(cpu, th_pressure); |
203 | ||
7e97b3dc LL |
204 | for_each_cpu(cpu, cpus) |
205 | WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure); | |
c214f124 LL |
206 | } |
207 | EXPORT_SYMBOL_GPL(topology_update_thermal_pressure); | |
208 | ||
2ef7a295 JL |
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 | ||
aa838896 | 215 | return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id)); |
2ef7a295 JL |
216 | } |
217 | ||
bb1fbdd3 MR |
218 | static void update_topology_flags_workfn(struct work_struct *work); |
219 | static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn); | |
220 | ||
5d777b18 | 221 | static DEVICE_ATTR_RO(cpu_capacity); |
2ef7a295 JL |
222 | |
223 | static 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 | } | |
240 | subsys_initcall(register_cpu_capacity_sysctl); | |
241 | ||
bb1fbdd3 MR |
242 | static int update_topology; |
243 | ||
244 | int 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 | */ | |
253 | static 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 | ||
2ef7a295 | 261 | static u32 *raw_capacity; |
62de1161 | 262 | |
82d8ba71 | 263 | static int free_raw_capacity(void) |
62de1161 VK |
264 | { |
265 | kfree(raw_capacity); | |
266 | raw_capacity = NULL; | |
267 | ||
268 | return 0; | |
269 | } | |
2ef7a295 | 270 | |
4ca4f26a | 271 | void topology_normalize_cpu_scale(void) |
2ef7a295 JL |
272 | { |
273 | u64 capacity; | |
b8fe128d | 274 | u64 capacity_scale; |
2ef7a295 JL |
275 | int cpu; |
276 | ||
62de1161 | 277 | if (!raw_capacity) |
2ef7a295 JL |
278 | return; |
279 | ||
b8fe128d | 280 | capacity_scale = 1; |
2ef7a295 | 281 | for_each_possible_cpu(cpu) { |
b8fe128d JC |
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); | |
4ca4f26a | 291 | topology_set_cpu_scale(cpu, capacity); |
2ef7a295 | 292 | pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n", |
8ec59c0f | 293 | cpu, topology_get_cpu_scale(cpu)); |
2ef7a295 | 294 | } |
2ef7a295 JL |
295 | } |
296 | ||
805df296 | 297 | bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu) |
2ef7a295 | 298 | { |
b8fe128d | 299 | struct clk *cpu_clk; |
62de1161 | 300 | static bool cap_parsing_failed; |
805df296 | 301 | int ret; |
2ef7a295 JL |
302 | u32 cpu_capacity; |
303 | ||
304 | if (cap_parsing_failed) | |
805df296 | 305 | return false; |
2ef7a295 | 306 | |
3eeba1a2 | 307 | ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz", |
2ef7a295 JL |
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) { | |
2ef7a295 | 315 | cap_parsing_failed = true; |
805df296 | 316 | return false; |
2ef7a295 JL |
317 | } |
318 | } | |
2ef7a295 | 319 | raw_capacity[cpu] = cpu_capacity; |
6ef2541f RH |
320 | pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n", |
321 | cpu_node, raw_capacity[cpu]); | |
b8fe128d JC |
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); | |
4dfff3d5 | 330 | if (!PTR_ERR_OR_ZERO(cpu_clk)) { |
b8fe128d JC |
331 | per_cpu(freq_factor, cpu) = |
332 | clk_get_rate(cpu_clk) / 1000; | |
4dfff3d5 JC |
333 | clk_put(cpu_clk); |
334 | } | |
2ef7a295 JL |
335 | } else { |
336 | if (raw_capacity) { | |
6ef2541f RH |
337 | pr_err("cpu_capacity: missing %pOF raw capacity\n", |
338 | cpu_node); | |
2ef7a295 JL |
339 | pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n"); |
340 | } | |
341 | cap_parsing_failed = true; | |
62de1161 | 342 | free_raw_capacity(); |
2ef7a295 JL |
343 | } |
344 | ||
345 | return !ret; | |
346 | } | |
347 | ||
9924fbb5 IV |
348 | #ifdef CONFIG_ACPI_CPPC_LIB |
349 | #include <acpi/cppc_acpi.h> | |
350 | ||
351 | void topology_init_cpu_capacity_cppc(void) | |
352 | { | |
353 | struct cppc_perf_caps perf_caps; | |
354 | int cpu; | |
355 | ||
a2a9d185 | 356 | if (likely(!acpi_cpc_valid())) |
9924fbb5 IV |
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 | ||
383 | exit: | |
384 | free_raw_capacity(); | |
385 | } | |
386 | #endif | |
387 | ||
2ef7a295 | 388 | #ifdef CONFIG_CPU_FREQ |
9de9a449 GI |
389 | static cpumask_var_t cpus_to_visit; |
390 | static void parsing_done_workfn(struct work_struct *work); | |
391 | static DECLARE_WORK(parsing_done_work, parsing_done_workfn); | |
2ef7a295 | 392 | |
9de9a449 | 393 | static int |
2ef7a295 JL |
394 | init_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 | ||
d8bcf4db | 401 | if (!raw_capacity) |
2ef7a295 JL |
402 | return 0; |
403 | ||
40f0fc2a | 404 | if (val != CPUFREQ_CREATE_POLICY) |
93a57081 VK |
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 | ||
b8fe128d JC |
413 | for_each_cpu(cpu, policy->related_cpus) |
414 | per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000; | |
93a57081 VK |
415 | |
416 | if (cpumask_empty(cpus_to_visit)) { | |
417 | topology_normalize_cpu_scale(); | |
bb1fbdd3 | 418 | schedule_work(&update_topology_flags_work); |
62de1161 | 419 | free_raw_capacity(); |
93a57081 | 420 | pr_debug("cpu_capacity: parsing done\n"); |
93a57081 VK |
421 | schedule_work(&parsing_done_work); |
422 | } | |
423 | ||
2ef7a295 JL |
424 | return 0; |
425 | } | |
426 | ||
9de9a449 | 427 | static struct notifier_block init_cpu_capacity_notifier = { |
2ef7a295 JL |
428 | .notifier_call = init_cpu_capacity_callback, |
429 | }; | |
430 | ||
431 | static int __init register_cpufreq_notifier(void) | |
432 | { | |
5408211a DE |
433 | int ret; |
434 | ||
2ef7a295 | 435 | /* |
9924fbb5 IV |
436 | * On ACPI-based systems skip registering cpufreq notifier as cpufreq |
437 | * information is not needed for cpu capacity initialization. | |
2ef7a295 | 438 | */ |
c105aa31 | 439 | if (!acpi_disabled || !raw_capacity) |
2ef7a295 JL |
440 | return -EINVAL; |
441 | ||
0fd33116 | 442 | if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL)) |
2ef7a295 | 443 | return -ENOMEM; |
2ef7a295 JL |
444 | |
445 | cpumask_copy(cpus_to_visit, cpu_possible_mask); | |
446 | ||
5408211a DE |
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; | |
2ef7a295 JL |
454 | } |
455 | core_initcall(register_cpufreq_notifier); | |
456 | ||
9de9a449 | 457 | static void parsing_done_workfn(struct work_struct *work) |
2ef7a295 JL |
458 | { |
459 | cpufreq_unregister_notifier(&init_cpu_capacity_notifier, | |
460 | CPUFREQ_POLICY_NOTIFIER); | |
5408211a | 461 | free_cpumask_var(cpus_to_visit); |
2ef7a295 JL |
462 | } |
463 | ||
464 | #else | |
2ef7a295 JL |
465 | core_initcall(free_raw_capacity); |
466 | #endif | |
60c1b220 AP |
467 | |
468 | #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV) | |
f3c19481 ZT |
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 | */ | |
60c1b220 AP |
479 | static 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 | |
f3c19481 ZT |
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)); | |
60c1b220 AP |
494 | |
495 | of_node_put(cpu_node); | |
496 | return cpu; | |
497 | } | |
498 | ||
499 | static int __init parse_core(struct device_node *core, int package_id, | |
556c9678 | 500 | int cluster_id, int core_id) |
60c1b220 | 501 | { |
4a33691c | 502 | char name[20]; |
60c1b220 AP |
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; | |
556c9678 | 516 | cpu_topology[cpu].cluster_id = cluster_id; |
60c1b220 AP |
517 | cpu_topology[cpu].core_id = core_id; |
518 | cpu_topology[cpu].thread_id = i; | |
f3c19481 ZT |
519 | } else if (cpu != -ENODEV) { |
520 | pr_err("%pOF: Can't get CPU for thread\n", t); | |
60c1b220 AP |
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; | |
556c9678 | 538 | cpu_topology[cpu].cluster_id = cluster_id; |
60c1b220 | 539 | cpu_topology[cpu].core_id = core_id; |
f3c19481 | 540 | } else if (leaf && cpu != -ENODEV) { |
60c1b220 AP |
541 | pr_err("%pOF: Can't get CPU for leaf core\n", core); |
542 | return -EINVAL; | |
543 | } | |
544 | ||
545 | return 0; | |
546 | } | |
547 | ||
dea8c0b4 SH |
548 | static int __init parse_cluster(struct device_node *cluster, int package_id, |
549 | int cluster_id, int depth) | |
60c1b220 | 550 | { |
4a33691c | 551 | char name[20]; |
60c1b220 AP |
552 | bool leaf = true; |
553 | bool has_cores = false; | |
554 | struct device_node *c; | |
60c1b220 AP |
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; | |
dea8c0b4 | 569 | ret = parse_cluster(c, package_id, i, depth + 1); |
00e66e37 SH |
570 | if (depth > 0) |
571 | pr_warn("Topology for clusters of clusters not yet supported\n"); | |
60c1b220 AP |
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) { | |
dea8c0b4 SH |
595 | ret = parse_core(c, package_id, cluster_id, |
596 | core_id++); | |
60c1b220 AP |
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 | ||
60c1b220 AP |
613 | return 0; |
614 | } | |
615 | ||
dea8c0b4 SH |
616 | static 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 | ||
60c1b220 AP |
642 | static 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 | ||
dea8c0b4 | 662 | ret = parse_socket(map); |
60c1b220 AP |
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) | |
5a01bb8e | 673 | if (cpu_topology[cpu].package_id < 0) { |
60c1b220 | 674 | ret = -EINVAL; |
5a01bb8e SH |
675 | break; |
676 | } | |
60c1b220 AP |
677 | |
678 | out_map: | |
679 | of_node_put(map); | |
680 | out: | |
681 | of_node_put(cn); | |
682 | return ret; | |
683 | } | |
ca74b316 | 684 | #endif |
60c1b220 AP |
685 | |
686 | /* | |
687 | * cpu topology table | |
688 | */ | |
689 | struct cpu_topology cpu_topology[NR_CPUS]; | |
690 | EXPORT_SYMBOL_GPL(cpu_topology); | |
691 | ||
692 | const 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 | } | |
f027db2f SH |
701 | |
702 | if (last_level_cache_is_valid(cpu)) { | |
60c1b220 AP |
703 | if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask)) |
704 | core_mask = &cpu_topology[cpu].llc_sibling; | |
705 | } | |
706 | ||
db1e5948 DH |
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 | ||
60c1b220 AP |
716 | return core_mask; |
717 | } | |
718 | ||
c5e22fef JC |
719 | const struct cpumask *cpu_clustergroup_mask(int cpu) |
720 | { | |
bfcc4397 IV |
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)) | |
5ac251c8 | 727 | return topology_sibling_cpumask(cpu); |
bfcc4397 | 728 | |
c5e22fef JC |
729 | return &cpu_topology[cpu].cluster_sibling; |
730 | } | |
731 | ||
60c1b220 AP |
732 | void update_siblings_masks(unsigned int cpuid) |
733 | { | |
734 | struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid]; | |
3fcbf1c7 SH |
735 | int cpu, ret; |
736 | ||
737 | ret = detect_cache_attributes(cpuid); | |
9b03e793 | 738 | if (ret && ret != -ENOENT) |
5944ce09 | 739 | pr_info("Early cacheinfo allocation failed, ret = %d\n", ret); |
60c1b220 AP |
740 | |
741 | /* update core and thread sibling masks */ | |
742 | for_each_online_cpu(cpu) { | |
743 | cpu_topo = &cpu_topology[cpu]; | |
744 | ||
f027db2f | 745 | if (last_level_cache_is_shared(cpu, cpuid)) { |
60c1b220 AP |
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 | ||
3f828329 SH |
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 | ||
9eb5e54f | 759 | if (cpuid_topo->cluster_id >= 0) { |
c5e22fef JC |
760 | cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling); |
761 | cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling); | |
762 | } | |
763 | ||
60c1b220 AP |
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 | ||
772 | static 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 | ||
c5e22fef JC |
779 | cpumask_clear(&cpu_topo->cluster_sibling); |
780 | cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling); | |
781 | ||
60c1b220 AP |
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 | ||
ca74b316 | 788 | void __init reset_cpu_topology(void) |
60c1b220 AP |
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; | |
c5e22fef | 797 | cpu_topo->cluster_id = -1; |
60c1b220 | 798 | cpu_topo->package_id = -1; |
60c1b220 AP |
799 | |
800 | clear_cpu_topology(cpu); | |
801 | } | |
802 | } | |
803 | ||
804 | void 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)); | |
4cc4cc28 WS |
812 | for_each_cpu(sibling, topology_cluster_cpumask(cpu)) |
813 | cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling)); | |
60c1b220 AP |
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 | ||
ca74b316 | 825 | #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV) |
60c1b220 AP |
826 | void __init init_cpu_topology(void) |
827 | { | |
5944ce09 | 828 | int cpu, ret; |
38db9b95 | 829 | |
60c1b220 | 830 | reset_cpu_topology(); |
38db9b95 SH |
831 | ret = parse_acpi_topology(); |
832 | if (!ret) | |
833 | ret = of_have_populated_dt() && parse_dt_topology(); | |
60c1b220 | 834 | |
38db9b95 SH |
835 | if (ret) { |
836 | /* | |
837 | * Discard anything that was parsed if we hit an error so we | |
e103d554 RR |
838 | * don't use partial information. But do not return yet to give |
839 | * arch-specific early cache level detection a chance to run. | |
38db9b95 | 840 | */ |
60c1b220 | 841 | reset_cpu_topology(); |
38db9b95 | 842 | } |
5944ce09 PG |
843 | |
844 | for_each_possible_cpu(cpu) { | |
845 | ret = fetch_cache_info(cpu); | |
35223401 PG |
846 | if (!ret) |
847 | continue; | |
848 | else if (ret != -ENOENT) | |
5944ce09 | 849 | pr_err("Early cacheinfo failed, ret = %d\n", ret); |
35223401 | 850 | return; |
5944ce09 | 851 | } |
60c1b220 | 852 | } |
456797da CD |
853 | |
854 | void store_cpu_topology(unsigned int cpuid) | |
855 | { | |
856 | struct cpu_topology *cpuid_topo = &cpu_topology[cpuid]; | |
857 | ||
858 | if (cpuid_topo->package_id != -1) | |
859 | goto topology_populated; | |
860 | ||
861 | cpuid_topo->thread_id = -1; | |
862 | cpuid_topo->core_id = cpuid; | |
863 | cpuid_topo->package_id = cpu_to_node(cpuid); | |
864 | ||
865 | pr_debug("CPU%u: package %d core %d thread %d\n", | |
866 | cpuid, cpuid_topo->package_id, cpuid_topo->core_id, | |
867 | cpuid_topo->thread_id); | |
868 | ||
869 | topology_populated: | |
870 | update_siblings_masks(cpuid); | |
871 | } | |
60c1b220 | 872 | #endif |