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> | |
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> | |
2ef7a295 | 15 | #include <linux/sched/topology.h> |
bb1fbdd3 | 16 | #include <linux/cpuset.h> |
60c1b220 AP |
17 | #include <linux/cpumask.h> |
18 | #include <linux/init.h> | |
83150f5d | 19 | #include <linux/rcupdate.h> |
60c1b220 | 20 | #include <linux/sched.h> |
2ef7a295 | 21 | |
83150f5d | 22 | static DEFINE_PER_CPU(struct scale_freq_data __rcu *, sft_data); |
01e055c1 VK |
23 | static struct cpumask scale_freq_counters_mask; |
24 | static bool scale_freq_invariant; | |
c214f124 | 25 | static DEFINE_PER_CPU(u32, freq_factor) = 1; |
01e055c1 VK |
26 | |
27 | static bool supports_scale_freq_counters(const struct cpumask *cpus) | |
28 | { | |
29 | return cpumask_subset(cpus, &scale_freq_counters_mask); | |
30 | } | |
31 | ||
15e5d5b4 VS |
32 | bool topology_scale_freq_invariant(void) |
33 | { | |
34 | return cpufreq_supports_freq_invariance() || | |
01e055c1 VK |
35 | supports_scale_freq_counters(cpu_online_mask); |
36 | } | |
37 | ||
38 | static void update_scale_freq_invariant(bool status) | |
39 | { | |
40 | if (scale_freq_invariant == status) | |
41 | return; | |
42 | ||
43 | /* | |
44 | * Task scheduler behavior depends on frequency invariance support, | |
45 | * either cpufreq or counter driven. If the support status changes as | |
46 | * a result of counter initialisation and use, retrigger the build of | |
47 | * scheduling domains to ensure the information is propagated properly. | |
48 | */ | |
49 | if (topology_scale_freq_invariant() == status) { | |
50 | scale_freq_invariant = status; | |
51 | rebuild_sched_domains_energy(); | |
52 | } | |
53 | } | |
54 | ||
55 | void topology_set_scale_freq_source(struct scale_freq_data *data, | |
56 | const struct cpumask *cpus) | |
57 | { | |
58 | struct scale_freq_data *sfd; | |
59 | int cpu; | |
60 | ||
61 | /* | |
62 | * Avoid calling rebuild_sched_domains() unnecessarily if FIE is | |
63 | * supported by cpufreq. | |
64 | */ | |
65 | if (cpumask_empty(&scale_freq_counters_mask)) | |
66 | scale_freq_invariant = topology_scale_freq_invariant(); | |
67 | ||
83150f5d VK |
68 | rcu_read_lock(); |
69 | ||
01e055c1 | 70 | for_each_cpu(cpu, cpus) { |
83150f5d | 71 | sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu)); |
01e055c1 VK |
72 | |
73 | /* Use ARCH provided counters whenever possible */ | |
74 | if (!sfd || sfd->source != SCALE_FREQ_SOURCE_ARCH) { | |
83150f5d | 75 | rcu_assign_pointer(per_cpu(sft_data, cpu), data); |
01e055c1 VK |
76 | cpumask_set_cpu(cpu, &scale_freq_counters_mask); |
77 | } | |
78 | } | |
79 | ||
83150f5d VK |
80 | rcu_read_unlock(); |
81 | ||
01e055c1 | 82 | update_scale_freq_invariant(true); |
15e5d5b4 | 83 | } |
2f533958 | 84 | EXPORT_SYMBOL_GPL(topology_set_scale_freq_source); |
15e5d5b4 | 85 | |
01e055c1 VK |
86 | void topology_clear_scale_freq_source(enum scale_freq_source source, |
87 | const struct cpumask *cpus) | |
cd0ed03a | 88 | { |
01e055c1 VK |
89 | struct scale_freq_data *sfd; |
90 | int cpu; | |
91 | ||
83150f5d VK |
92 | rcu_read_lock(); |
93 | ||
01e055c1 | 94 | for_each_cpu(cpu, cpus) { |
83150f5d | 95 | sfd = rcu_dereference(*per_cpu_ptr(&sft_data, cpu)); |
01e055c1 VK |
96 | |
97 | if (sfd && sfd->source == source) { | |
83150f5d | 98 | rcu_assign_pointer(per_cpu(sft_data, cpu), NULL); |
01e055c1 VK |
99 | cpumask_clear_cpu(cpu, &scale_freq_counters_mask); |
100 | } | |
101 | } | |
102 | ||
83150f5d VK |
103 | rcu_read_unlock(); |
104 | ||
105 | /* | |
106 | * Make sure all references to previous sft_data are dropped to avoid | |
107 | * use-after-free races. | |
108 | */ | |
109 | synchronize_rcu(); | |
110 | ||
01e055c1 | 111 | update_scale_freq_invariant(false); |
cd0ed03a | 112 | } |
2f533958 | 113 | EXPORT_SYMBOL_GPL(topology_clear_scale_freq_source); |
01e055c1 VK |
114 | |
115 | void topology_scale_freq_tick(void) | |
116 | { | |
83150f5d | 117 | struct scale_freq_data *sfd = rcu_dereference_sched(*this_cpu_ptr(&sft_data)); |
01e055c1 VK |
118 | |
119 | if (sfd) | |
120 | sfd->set_freq_scale(); | |
121 | } | |
122 | ||
eec73529 | 123 | DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE; |
2f533958 | 124 | EXPORT_PER_CPU_SYMBOL_GPL(arch_freq_scale); |
2ef7a295 | 125 | |
a20b7053 IV |
126 | void topology_set_freq_scale(const struct cpumask *cpus, unsigned long cur_freq, |
127 | unsigned long max_freq) | |
2ef7a295 | 128 | { |
0e27c567 DE |
129 | unsigned long scale; |
130 | int i; | |
131 | ||
0a10d3fe IV |
132 | if (WARN_ON_ONCE(!cur_freq || !max_freq)) |
133 | return; | |
134 | ||
cd0ed03a IV |
135 | /* |
136 | * If the use of counters for FIE is enabled, just return as we don't | |
137 | * want to update the scale factor with information from CPUFREQ. | |
138 | * Instead the scale factor will be updated from arch_scale_freq_tick. | |
139 | */ | |
01e055c1 | 140 | if (supports_scale_freq_counters(cpus)) |
cd0ed03a IV |
141 | return; |
142 | ||
0e27c567 DE |
143 | scale = (cur_freq << SCHED_CAPACITY_SHIFT) / max_freq; |
144 | ||
145 | for_each_cpu(i, cpus) | |
eec73529 | 146 | per_cpu(arch_freq_scale, i) = scale; |
2ef7a295 JL |
147 | } |
148 | ||
8216f588 | 149 | DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE; |
275157b3 | 150 | EXPORT_PER_CPU_SYMBOL_GPL(cpu_scale); |
2ef7a295 | 151 | |
4ca4f26a | 152 | void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity) |
2ef7a295 JL |
153 | { |
154 | per_cpu(cpu_scale, cpu) = capacity; | |
155 | } | |
156 | ||
25980c7a VS |
157 | DEFINE_PER_CPU(unsigned long, thermal_pressure); |
158 | ||
c214f124 LL |
159 | /** |
160 | * topology_update_thermal_pressure() - Update thermal pressure for CPUs | |
161 | * @cpus : The related CPUs for which capacity has been reduced | |
162 | * @capped_freq : The maximum allowed frequency that CPUs can run at | |
163 | * | |
164 | * Update the value of thermal pressure for all @cpus in the mask. The | |
165 | * cpumask should include all (online+offline) affected CPUs, to avoid | |
166 | * operating on stale data when hot-plug is used for some CPUs. The | |
167 | * @capped_freq reflects the currently allowed max CPUs frequency due to | |
168 | * thermal capping. It might be also a boost frequency value, which is bigger | |
169 | * than the internal 'freq_factor' max frequency. In such case the pressure | |
170 | * value should simply be removed, since this is an indication that there is | |
171 | * no thermal throttling. The @capped_freq must be provided in kHz. | |
172 | */ | |
173 | void topology_update_thermal_pressure(const struct cpumask *cpus, | |
174 | unsigned long capped_freq) | |
175 | { | |
7e97b3dc | 176 | unsigned long max_capacity, capacity, th_pressure; |
c214f124 LL |
177 | u32 max_freq; |
178 | int cpu; | |
179 | ||
180 | cpu = cpumask_first(cpus); | |
181 | max_capacity = arch_scale_cpu_capacity(cpu); | |
182 | max_freq = per_cpu(freq_factor, cpu); | |
183 | ||
184 | /* Convert to MHz scale which is used in 'freq_factor' */ | |
185 | capped_freq /= 1000; | |
186 | ||
187 | /* | |
188 | * Handle properly the boost frequencies, which should simply clean | |
189 | * the thermal pressure value. | |
190 | */ | |
191 | if (max_freq <= capped_freq) | |
192 | capacity = max_capacity; | |
193 | else | |
194 | capacity = mult_frac(max_capacity, capped_freq, max_freq); | |
195 | ||
7e97b3dc LL |
196 | th_pressure = max_capacity - capacity; |
197 | ||
198 | for_each_cpu(cpu, cpus) | |
199 | WRITE_ONCE(per_cpu(thermal_pressure, cpu), th_pressure); | |
c214f124 LL |
200 | } |
201 | EXPORT_SYMBOL_GPL(topology_update_thermal_pressure); | |
202 | ||
2ef7a295 JL |
203 | static ssize_t cpu_capacity_show(struct device *dev, |
204 | struct device_attribute *attr, | |
205 | char *buf) | |
206 | { | |
207 | struct cpu *cpu = container_of(dev, struct cpu, dev); | |
208 | ||
aa838896 | 209 | return sysfs_emit(buf, "%lu\n", topology_get_cpu_scale(cpu->dev.id)); |
2ef7a295 JL |
210 | } |
211 | ||
bb1fbdd3 MR |
212 | static void update_topology_flags_workfn(struct work_struct *work); |
213 | static DECLARE_WORK(update_topology_flags_work, update_topology_flags_workfn); | |
214 | ||
5d777b18 | 215 | static DEVICE_ATTR_RO(cpu_capacity); |
2ef7a295 JL |
216 | |
217 | static int register_cpu_capacity_sysctl(void) | |
218 | { | |
219 | int i; | |
220 | struct device *cpu; | |
221 | ||
222 | for_each_possible_cpu(i) { | |
223 | cpu = get_cpu_device(i); | |
224 | if (!cpu) { | |
225 | pr_err("%s: too early to get CPU%d device!\n", | |
226 | __func__, i); | |
227 | continue; | |
228 | } | |
229 | device_create_file(cpu, &dev_attr_cpu_capacity); | |
230 | } | |
231 | ||
232 | return 0; | |
233 | } | |
234 | subsys_initcall(register_cpu_capacity_sysctl); | |
235 | ||
bb1fbdd3 MR |
236 | static int update_topology; |
237 | ||
238 | int topology_update_cpu_topology(void) | |
239 | { | |
240 | return update_topology; | |
241 | } | |
242 | ||
243 | /* | |
244 | * Updating the sched_domains can't be done directly from cpufreq callbacks | |
245 | * due to locking, so queue the work for later. | |
246 | */ | |
247 | static void update_topology_flags_workfn(struct work_struct *work) | |
248 | { | |
249 | update_topology = 1; | |
250 | rebuild_sched_domains(); | |
251 | pr_debug("sched_domain hierarchy rebuilt, flags updated\n"); | |
252 | update_topology = 0; | |
253 | } | |
254 | ||
2ef7a295 | 255 | static u32 *raw_capacity; |
62de1161 | 256 | |
82d8ba71 | 257 | static int free_raw_capacity(void) |
62de1161 VK |
258 | { |
259 | kfree(raw_capacity); | |
260 | raw_capacity = NULL; | |
261 | ||
262 | return 0; | |
263 | } | |
2ef7a295 | 264 | |
4ca4f26a | 265 | void topology_normalize_cpu_scale(void) |
2ef7a295 JL |
266 | { |
267 | u64 capacity; | |
b8fe128d | 268 | u64 capacity_scale; |
2ef7a295 JL |
269 | int cpu; |
270 | ||
62de1161 | 271 | if (!raw_capacity) |
2ef7a295 JL |
272 | return; |
273 | ||
b8fe128d | 274 | capacity_scale = 1; |
2ef7a295 | 275 | for_each_possible_cpu(cpu) { |
b8fe128d JC |
276 | capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu); |
277 | capacity_scale = max(capacity, capacity_scale); | |
278 | } | |
279 | ||
280 | pr_debug("cpu_capacity: capacity_scale=%llu\n", capacity_scale); | |
281 | for_each_possible_cpu(cpu) { | |
282 | capacity = raw_capacity[cpu] * per_cpu(freq_factor, cpu); | |
283 | capacity = div64_u64(capacity << SCHED_CAPACITY_SHIFT, | |
284 | capacity_scale); | |
4ca4f26a | 285 | topology_set_cpu_scale(cpu, capacity); |
2ef7a295 | 286 | pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n", |
8ec59c0f | 287 | cpu, topology_get_cpu_scale(cpu)); |
2ef7a295 | 288 | } |
2ef7a295 JL |
289 | } |
290 | ||
805df296 | 291 | bool __init topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu) |
2ef7a295 | 292 | { |
b8fe128d | 293 | struct clk *cpu_clk; |
62de1161 | 294 | static bool cap_parsing_failed; |
805df296 | 295 | int ret; |
2ef7a295 JL |
296 | u32 cpu_capacity; |
297 | ||
298 | if (cap_parsing_failed) | |
805df296 | 299 | return false; |
2ef7a295 | 300 | |
3eeba1a2 | 301 | ret = of_property_read_u32(cpu_node, "capacity-dmips-mhz", |
2ef7a295 JL |
302 | &cpu_capacity); |
303 | if (!ret) { | |
304 | if (!raw_capacity) { | |
305 | raw_capacity = kcalloc(num_possible_cpus(), | |
306 | sizeof(*raw_capacity), | |
307 | GFP_KERNEL); | |
308 | if (!raw_capacity) { | |
2ef7a295 | 309 | cap_parsing_failed = true; |
805df296 | 310 | return false; |
2ef7a295 JL |
311 | } |
312 | } | |
2ef7a295 | 313 | raw_capacity[cpu] = cpu_capacity; |
6ef2541f RH |
314 | pr_debug("cpu_capacity: %pOF cpu_capacity=%u (raw)\n", |
315 | cpu_node, raw_capacity[cpu]); | |
b8fe128d JC |
316 | |
317 | /* | |
318 | * Update freq_factor for calculating early boot cpu capacities. | |
319 | * For non-clk CPU DVFS mechanism, there's no way to get the | |
320 | * frequency value now, assuming they are running at the same | |
321 | * frequency (by keeping the initial freq_factor value). | |
322 | */ | |
323 | cpu_clk = of_clk_get(cpu_node, 0); | |
4dfff3d5 | 324 | if (!PTR_ERR_OR_ZERO(cpu_clk)) { |
b8fe128d JC |
325 | per_cpu(freq_factor, cpu) = |
326 | clk_get_rate(cpu_clk) / 1000; | |
4dfff3d5 JC |
327 | clk_put(cpu_clk); |
328 | } | |
2ef7a295 JL |
329 | } else { |
330 | if (raw_capacity) { | |
6ef2541f RH |
331 | pr_err("cpu_capacity: missing %pOF raw capacity\n", |
332 | cpu_node); | |
2ef7a295 JL |
333 | pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n"); |
334 | } | |
335 | cap_parsing_failed = true; | |
62de1161 | 336 | free_raw_capacity(); |
2ef7a295 JL |
337 | } |
338 | ||
339 | return !ret; | |
340 | } | |
341 | ||
9924fbb5 IV |
342 | #ifdef CONFIG_ACPI_CPPC_LIB |
343 | #include <acpi/cppc_acpi.h> | |
344 | ||
345 | void topology_init_cpu_capacity_cppc(void) | |
346 | { | |
347 | struct cppc_perf_caps perf_caps; | |
348 | int cpu; | |
349 | ||
350 | if (likely(acpi_disabled || !acpi_cpc_valid())) | |
351 | return; | |
352 | ||
353 | raw_capacity = kcalloc(num_possible_cpus(), sizeof(*raw_capacity), | |
354 | GFP_KERNEL); | |
355 | if (!raw_capacity) | |
356 | return; | |
357 | ||
358 | for_each_possible_cpu(cpu) { | |
359 | if (!cppc_get_perf_caps(cpu, &perf_caps) && | |
360 | (perf_caps.highest_perf >= perf_caps.nominal_perf) && | |
361 | (perf_caps.highest_perf >= perf_caps.lowest_perf)) { | |
362 | raw_capacity[cpu] = perf_caps.highest_perf; | |
363 | pr_debug("cpu_capacity: CPU%d cpu_capacity=%u (raw).\n", | |
364 | cpu, raw_capacity[cpu]); | |
365 | continue; | |
366 | } | |
367 | ||
368 | pr_err("cpu_capacity: CPU%d missing/invalid highest performance.\n", cpu); | |
369 | pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n"); | |
370 | goto exit; | |
371 | } | |
372 | ||
373 | topology_normalize_cpu_scale(); | |
374 | schedule_work(&update_topology_flags_work); | |
375 | pr_debug("cpu_capacity: cpu_capacity initialization done\n"); | |
376 | ||
377 | exit: | |
378 | free_raw_capacity(); | |
379 | } | |
380 | #endif | |
381 | ||
2ef7a295 | 382 | #ifdef CONFIG_CPU_FREQ |
9de9a449 GI |
383 | static cpumask_var_t cpus_to_visit; |
384 | static void parsing_done_workfn(struct work_struct *work); | |
385 | static DECLARE_WORK(parsing_done_work, parsing_done_workfn); | |
2ef7a295 | 386 | |
9de9a449 | 387 | static int |
2ef7a295 JL |
388 | init_cpu_capacity_callback(struct notifier_block *nb, |
389 | unsigned long val, | |
390 | void *data) | |
391 | { | |
392 | struct cpufreq_policy *policy = data; | |
393 | int cpu; | |
394 | ||
d8bcf4db | 395 | if (!raw_capacity) |
2ef7a295 JL |
396 | return 0; |
397 | ||
40f0fc2a | 398 | if (val != CPUFREQ_CREATE_POLICY) |
93a57081 VK |
399 | return 0; |
400 | ||
401 | pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n", | |
402 | cpumask_pr_args(policy->related_cpus), | |
403 | cpumask_pr_args(cpus_to_visit)); | |
404 | ||
405 | cpumask_andnot(cpus_to_visit, cpus_to_visit, policy->related_cpus); | |
406 | ||
b8fe128d JC |
407 | for_each_cpu(cpu, policy->related_cpus) |
408 | per_cpu(freq_factor, cpu) = policy->cpuinfo.max_freq / 1000; | |
93a57081 VK |
409 | |
410 | if (cpumask_empty(cpus_to_visit)) { | |
411 | topology_normalize_cpu_scale(); | |
bb1fbdd3 | 412 | schedule_work(&update_topology_flags_work); |
62de1161 | 413 | free_raw_capacity(); |
93a57081 | 414 | pr_debug("cpu_capacity: parsing done\n"); |
93a57081 VK |
415 | schedule_work(&parsing_done_work); |
416 | } | |
417 | ||
2ef7a295 JL |
418 | return 0; |
419 | } | |
420 | ||
9de9a449 | 421 | static struct notifier_block init_cpu_capacity_notifier = { |
2ef7a295 JL |
422 | .notifier_call = init_cpu_capacity_callback, |
423 | }; | |
424 | ||
425 | static int __init register_cpufreq_notifier(void) | |
426 | { | |
5408211a DE |
427 | int ret; |
428 | ||
2ef7a295 | 429 | /* |
9924fbb5 IV |
430 | * On ACPI-based systems skip registering cpufreq notifier as cpufreq |
431 | * information is not needed for cpu capacity initialization. | |
2ef7a295 | 432 | */ |
c105aa31 | 433 | if (!acpi_disabled || !raw_capacity) |
2ef7a295 JL |
434 | return -EINVAL; |
435 | ||
0fd33116 | 436 | if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL)) |
2ef7a295 | 437 | return -ENOMEM; |
2ef7a295 JL |
438 | |
439 | cpumask_copy(cpus_to_visit, cpu_possible_mask); | |
440 | ||
5408211a DE |
441 | ret = cpufreq_register_notifier(&init_cpu_capacity_notifier, |
442 | CPUFREQ_POLICY_NOTIFIER); | |
443 | ||
444 | if (ret) | |
445 | free_cpumask_var(cpus_to_visit); | |
446 | ||
447 | return ret; | |
2ef7a295 JL |
448 | } |
449 | core_initcall(register_cpufreq_notifier); | |
450 | ||
9de9a449 | 451 | static void parsing_done_workfn(struct work_struct *work) |
2ef7a295 JL |
452 | { |
453 | cpufreq_unregister_notifier(&init_cpu_capacity_notifier, | |
454 | CPUFREQ_POLICY_NOTIFIER); | |
5408211a | 455 | free_cpumask_var(cpus_to_visit); |
2ef7a295 JL |
456 | } |
457 | ||
458 | #else | |
2ef7a295 JL |
459 | core_initcall(free_raw_capacity); |
460 | #endif | |
60c1b220 AP |
461 | |
462 | #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV) | |
f3c19481 ZT |
463 | /* |
464 | * This function returns the logic cpu number of the node. | |
465 | * There are basically three kinds of return values: | |
466 | * (1) logic cpu number which is > 0. | |
467 | * (2) -ENODEV when the device tree(DT) node is valid and found in the DT but | |
468 | * there is no possible logical CPU in the kernel to match. This happens | |
469 | * when CONFIG_NR_CPUS is configure to be smaller than the number of | |
470 | * CPU nodes in DT. We need to just ignore this case. | |
471 | * (3) -1 if the node does not exist in the device tree | |
472 | */ | |
60c1b220 AP |
473 | static int __init get_cpu_for_node(struct device_node *node) |
474 | { | |
475 | struct device_node *cpu_node; | |
476 | int cpu; | |
477 | ||
478 | cpu_node = of_parse_phandle(node, "cpu", 0); | |
479 | if (!cpu_node) | |
480 | return -1; | |
481 | ||
482 | cpu = of_cpu_node_to_id(cpu_node); | |
483 | if (cpu >= 0) | |
484 | topology_parse_cpu_capacity(cpu_node, cpu); | |
485 | else | |
f3c19481 ZT |
486 | pr_info("CPU node for %pOF exist but the possible cpu range is :%*pbl\n", |
487 | cpu_node, cpumask_pr_args(cpu_possible_mask)); | |
60c1b220 AP |
488 | |
489 | of_node_put(cpu_node); | |
490 | return cpu; | |
491 | } | |
492 | ||
493 | static int __init parse_core(struct device_node *core, int package_id, | |
494 | int core_id) | |
495 | { | |
4a33691c | 496 | char name[20]; |
60c1b220 AP |
497 | bool leaf = true; |
498 | int i = 0; | |
499 | int cpu; | |
500 | struct device_node *t; | |
501 | ||
502 | do { | |
503 | snprintf(name, sizeof(name), "thread%d", i); | |
504 | t = of_get_child_by_name(core, name); | |
505 | if (t) { | |
506 | leaf = false; | |
507 | cpu = get_cpu_for_node(t); | |
508 | if (cpu >= 0) { | |
509 | cpu_topology[cpu].package_id = package_id; | |
510 | cpu_topology[cpu].core_id = core_id; | |
511 | cpu_topology[cpu].thread_id = i; | |
f3c19481 ZT |
512 | } else if (cpu != -ENODEV) { |
513 | pr_err("%pOF: Can't get CPU for thread\n", t); | |
60c1b220 AP |
514 | of_node_put(t); |
515 | return -EINVAL; | |
516 | } | |
517 | of_node_put(t); | |
518 | } | |
519 | i++; | |
520 | } while (t); | |
521 | ||
522 | cpu = get_cpu_for_node(core); | |
523 | if (cpu >= 0) { | |
524 | if (!leaf) { | |
525 | pr_err("%pOF: Core has both threads and CPU\n", | |
526 | core); | |
527 | return -EINVAL; | |
528 | } | |
529 | ||
530 | cpu_topology[cpu].package_id = package_id; | |
531 | cpu_topology[cpu].core_id = core_id; | |
f3c19481 | 532 | } else if (leaf && cpu != -ENODEV) { |
60c1b220 AP |
533 | pr_err("%pOF: Can't get CPU for leaf core\n", core); |
534 | return -EINVAL; | |
535 | } | |
536 | ||
537 | return 0; | |
538 | } | |
539 | ||
540 | static int __init parse_cluster(struct device_node *cluster, int depth) | |
541 | { | |
4a33691c | 542 | char name[20]; |
60c1b220 AP |
543 | bool leaf = true; |
544 | bool has_cores = false; | |
545 | struct device_node *c; | |
546 | static int package_id __initdata; | |
547 | int core_id = 0; | |
548 | int i, ret; | |
549 | ||
550 | /* | |
551 | * First check for child clusters; we currently ignore any | |
552 | * information about the nesting of clusters and present the | |
553 | * scheduler with a flat list of them. | |
554 | */ | |
555 | i = 0; | |
556 | do { | |
557 | snprintf(name, sizeof(name), "cluster%d", i); | |
558 | c = of_get_child_by_name(cluster, name); | |
559 | if (c) { | |
560 | leaf = false; | |
561 | ret = parse_cluster(c, depth + 1); | |
562 | of_node_put(c); | |
563 | if (ret != 0) | |
564 | return ret; | |
565 | } | |
566 | i++; | |
567 | } while (c); | |
568 | ||
569 | /* Now check for cores */ | |
570 | i = 0; | |
571 | do { | |
572 | snprintf(name, sizeof(name), "core%d", i); | |
573 | c = of_get_child_by_name(cluster, name); | |
574 | if (c) { | |
575 | has_cores = true; | |
576 | ||
577 | if (depth == 0) { | |
578 | pr_err("%pOF: cpu-map children should be clusters\n", | |
579 | c); | |
580 | of_node_put(c); | |
581 | return -EINVAL; | |
582 | } | |
583 | ||
584 | if (leaf) { | |
585 | ret = parse_core(c, package_id, core_id++); | |
586 | } else { | |
587 | pr_err("%pOF: Non-leaf cluster with core %s\n", | |
588 | cluster, name); | |
589 | ret = -EINVAL; | |
590 | } | |
591 | ||
592 | of_node_put(c); | |
593 | if (ret != 0) | |
594 | return ret; | |
595 | } | |
596 | i++; | |
597 | } while (c); | |
598 | ||
599 | if (leaf && !has_cores) | |
600 | pr_warn("%pOF: empty cluster\n", cluster); | |
601 | ||
602 | if (leaf) | |
603 | package_id++; | |
604 | ||
605 | return 0; | |
606 | } | |
607 | ||
608 | static int __init parse_dt_topology(void) | |
609 | { | |
610 | struct device_node *cn, *map; | |
611 | int ret = 0; | |
612 | int cpu; | |
613 | ||
614 | cn = of_find_node_by_path("/cpus"); | |
615 | if (!cn) { | |
616 | pr_err("No CPU information found in DT\n"); | |
617 | return 0; | |
618 | } | |
619 | ||
620 | /* | |
621 | * When topology is provided cpu-map is essentially a root | |
622 | * cluster with restricted subnodes. | |
623 | */ | |
624 | map = of_get_child_by_name(cn, "cpu-map"); | |
625 | if (!map) | |
626 | goto out; | |
627 | ||
628 | ret = parse_cluster(map, 0); | |
629 | if (ret != 0) | |
630 | goto out_map; | |
631 | ||
632 | topology_normalize_cpu_scale(); | |
633 | ||
634 | /* | |
635 | * Check that all cores are in the topology; the SMP code will | |
636 | * only mark cores described in the DT as possible. | |
637 | */ | |
638 | for_each_possible_cpu(cpu) | |
639 | if (cpu_topology[cpu].package_id == -1) | |
640 | ret = -EINVAL; | |
641 | ||
642 | out_map: | |
643 | of_node_put(map); | |
644 | out: | |
645 | of_node_put(cn); | |
646 | return ret; | |
647 | } | |
ca74b316 | 648 | #endif |
60c1b220 AP |
649 | |
650 | /* | |
651 | * cpu topology table | |
652 | */ | |
653 | struct cpu_topology cpu_topology[NR_CPUS]; | |
654 | EXPORT_SYMBOL_GPL(cpu_topology); | |
655 | ||
656 | const struct cpumask *cpu_coregroup_mask(int cpu) | |
657 | { | |
658 | const cpumask_t *core_mask = cpumask_of_node(cpu_to_node(cpu)); | |
659 | ||
660 | /* Find the smaller of NUMA, core or LLC siblings */ | |
661 | if (cpumask_subset(&cpu_topology[cpu].core_sibling, core_mask)) { | |
662 | /* not numa in package, lets use the package siblings */ | |
663 | core_mask = &cpu_topology[cpu].core_sibling; | |
664 | } | |
665 | if (cpu_topology[cpu].llc_id != -1) { | |
666 | if (cpumask_subset(&cpu_topology[cpu].llc_sibling, core_mask)) | |
667 | core_mask = &cpu_topology[cpu].llc_sibling; | |
668 | } | |
669 | ||
db1e5948 DH |
670 | /* |
671 | * For systems with no shared cpu-side LLC but with clusters defined, | |
672 | * extend core_mask to cluster_siblings. The sched domain builder will | |
673 | * then remove MC as redundant with CLS if SCHED_CLUSTER is enabled. | |
674 | */ | |
675 | if (IS_ENABLED(CONFIG_SCHED_CLUSTER) && | |
676 | cpumask_subset(core_mask, &cpu_topology[cpu].cluster_sibling)) | |
677 | core_mask = &cpu_topology[cpu].cluster_sibling; | |
678 | ||
60c1b220 AP |
679 | return core_mask; |
680 | } | |
681 | ||
c5e22fef JC |
682 | const struct cpumask *cpu_clustergroup_mask(int cpu) |
683 | { | |
684 | return &cpu_topology[cpu].cluster_sibling; | |
685 | } | |
686 | ||
60c1b220 AP |
687 | void update_siblings_masks(unsigned int cpuid) |
688 | { | |
689 | struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid]; | |
690 | int cpu; | |
691 | ||
692 | /* update core and thread sibling masks */ | |
693 | for_each_online_cpu(cpu) { | |
694 | cpu_topo = &cpu_topology[cpu]; | |
695 | ||
1dc9f1a6 | 696 | if (cpu_topo->llc_id != -1 && cpuid_topo->llc_id == cpu_topo->llc_id) { |
60c1b220 AP |
697 | cpumask_set_cpu(cpu, &cpuid_topo->llc_sibling); |
698 | cpumask_set_cpu(cpuid, &cpu_topo->llc_sibling); | |
699 | } | |
700 | ||
701 | if (cpuid_topo->package_id != cpu_topo->package_id) | |
702 | continue; | |
703 | ||
c5e22fef JC |
704 | if (cpuid_topo->cluster_id == cpu_topo->cluster_id && |
705 | cpuid_topo->cluster_id != -1) { | |
706 | cpumask_set_cpu(cpu, &cpuid_topo->cluster_sibling); | |
707 | cpumask_set_cpu(cpuid, &cpu_topo->cluster_sibling); | |
708 | } | |
709 | ||
60c1b220 AP |
710 | cpumask_set_cpu(cpuid, &cpu_topo->core_sibling); |
711 | cpumask_set_cpu(cpu, &cpuid_topo->core_sibling); | |
712 | ||
713 | if (cpuid_topo->core_id != cpu_topo->core_id) | |
714 | continue; | |
715 | ||
716 | cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling); | |
717 | cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling); | |
718 | } | |
719 | } | |
720 | ||
721 | static void clear_cpu_topology(int cpu) | |
722 | { | |
723 | struct cpu_topology *cpu_topo = &cpu_topology[cpu]; | |
724 | ||
725 | cpumask_clear(&cpu_topo->llc_sibling); | |
726 | cpumask_set_cpu(cpu, &cpu_topo->llc_sibling); | |
727 | ||
c5e22fef JC |
728 | cpumask_clear(&cpu_topo->cluster_sibling); |
729 | cpumask_set_cpu(cpu, &cpu_topo->cluster_sibling); | |
730 | ||
60c1b220 AP |
731 | cpumask_clear(&cpu_topo->core_sibling); |
732 | cpumask_set_cpu(cpu, &cpu_topo->core_sibling); | |
733 | cpumask_clear(&cpu_topo->thread_sibling); | |
734 | cpumask_set_cpu(cpu, &cpu_topo->thread_sibling); | |
735 | } | |
736 | ||
ca74b316 | 737 | void __init reset_cpu_topology(void) |
60c1b220 AP |
738 | { |
739 | unsigned int cpu; | |
740 | ||
741 | for_each_possible_cpu(cpu) { | |
742 | struct cpu_topology *cpu_topo = &cpu_topology[cpu]; | |
743 | ||
744 | cpu_topo->thread_id = -1; | |
745 | cpu_topo->core_id = -1; | |
c5e22fef | 746 | cpu_topo->cluster_id = -1; |
60c1b220 AP |
747 | cpu_topo->package_id = -1; |
748 | cpu_topo->llc_id = -1; | |
749 | ||
750 | clear_cpu_topology(cpu); | |
751 | } | |
752 | } | |
753 | ||
754 | void remove_cpu_topology(unsigned int cpu) | |
755 | { | |
756 | int sibling; | |
757 | ||
758 | for_each_cpu(sibling, topology_core_cpumask(cpu)) | |
759 | cpumask_clear_cpu(cpu, topology_core_cpumask(sibling)); | |
760 | for_each_cpu(sibling, topology_sibling_cpumask(cpu)) | |
761 | cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling)); | |
4cc4cc28 WS |
762 | for_each_cpu(sibling, topology_cluster_cpumask(cpu)) |
763 | cpumask_clear_cpu(cpu, topology_cluster_cpumask(sibling)); | |
60c1b220 AP |
764 | for_each_cpu(sibling, topology_llc_cpumask(cpu)) |
765 | cpumask_clear_cpu(cpu, topology_llc_cpumask(sibling)); | |
766 | ||
767 | clear_cpu_topology(cpu); | |
768 | } | |
769 | ||
770 | __weak int __init parse_acpi_topology(void) | |
771 | { | |
772 | return 0; | |
773 | } | |
774 | ||
ca74b316 | 775 | #if defined(CONFIG_ARM64) || defined(CONFIG_RISCV) |
60c1b220 AP |
776 | void __init init_cpu_topology(void) |
777 | { | |
778 | reset_cpu_topology(); | |
779 | ||
780 | /* | |
781 | * Discard anything that was parsed if we hit an error so we | |
782 | * don't use partial information. | |
783 | */ | |
784 | if (parse_acpi_topology()) | |
785 | reset_cpu_topology(); | |
786 | else if (of_have_populated_dt() && parse_dt_topology()) | |
787 | reset_cpu_topology(); | |
788 | } | |
789 | #endif |