xref: /openbmc/linux/arch/arm64/kernel/topology.c (revision 4e1a33b1)
1 /*
2  * arch/arm64/kernel/topology.c
3  *
4  * Copyright (C) 2011,2013,2014 Linaro Limited.
5  *
6  * Based on the arm32 version written by Vincent Guittot in turn based on
7  * arch/sh/kernel/topology.c
8  *
9  * This file is subject to the terms and conditions of the GNU General Public
10  * License.  See the file "COPYING" in the main directory of this archive
11  * for more details.
12  */
13 
14 #include <linux/acpi.h>
15 #include <linux/cpu.h>
16 #include <linux/cpumask.h>
17 #include <linux/init.h>
18 #include <linux/percpu.h>
19 #include <linux/node.h>
20 #include <linux/nodemask.h>
21 #include <linux/of.h>
22 #include <linux/sched.h>
23 #include <linux/slab.h>
24 #include <linux/string.h>
25 #include <linux/cpufreq.h>
26 
27 #include <asm/cpu.h>
28 #include <asm/cputype.h>
29 #include <asm/topology.h>
30 
31 static DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
32 static DEFINE_MUTEX(cpu_scale_mutex);
33 
34 unsigned long arch_scale_cpu_capacity(struct sched_domain *sd, int cpu)
35 {
36 	return per_cpu(cpu_scale, cpu);
37 }
38 
39 static void set_capacity_scale(unsigned int cpu, unsigned long capacity)
40 {
41 	per_cpu(cpu_scale, cpu) = capacity;
42 }
43 
44 static ssize_t cpu_capacity_show(struct device *dev,
45 				 struct device_attribute *attr,
46 				 char *buf)
47 {
48 	struct cpu *cpu = container_of(dev, struct cpu, dev);
49 
50 	return sprintf(buf, "%lu\n",
51 			arch_scale_cpu_capacity(NULL, cpu->dev.id));
52 }
53 
54 static ssize_t cpu_capacity_store(struct device *dev,
55 				  struct device_attribute *attr,
56 				  const char *buf,
57 				  size_t count)
58 {
59 	struct cpu *cpu = container_of(dev, struct cpu, dev);
60 	int this_cpu = cpu->dev.id, i;
61 	unsigned long new_capacity;
62 	ssize_t ret;
63 
64 	if (count) {
65 		ret = kstrtoul(buf, 0, &new_capacity);
66 		if (ret)
67 			return ret;
68 		if (new_capacity > SCHED_CAPACITY_SCALE)
69 			return -EINVAL;
70 
71 		mutex_lock(&cpu_scale_mutex);
72 		for_each_cpu(i, &cpu_topology[this_cpu].core_sibling)
73 			set_capacity_scale(i, new_capacity);
74 		mutex_unlock(&cpu_scale_mutex);
75 	}
76 
77 	return count;
78 }
79 
80 static DEVICE_ATTR_RW(cpu_capacity);
81 
82 static int register_cpu_capacity_sysctl(void)
83 {
84 	int i;
85 	struct device *cpu;
86 
87 	for_each_possible_cpu(i) {
88 		cpu = get_cpu_device(i);
89 		if (!cpu) {
90 			pr_err("%s: too early to get CPU%d device!\n",
91 			       __func__, i);
92 			continue;
93 		}
94 		device_create_file(cpu, &dev_attr_cpu_capacity);
95 	}
96 
97 	return 0;
98 }
99 subsys_initcall(register_cpu_capacity_sysctl);
100 
101 static u32 capacity_scale;
102 static u32 *raw_capacity;
103 static bool cap_parsing_failed;
104 
105 static void __init parse_cpu_capacity(struct device_node *cpu_node, int cpu)
106 {
107 	int ret;
108 	u32 cpu_capacity;
109 
110 	if (cap_parsing_failed)
111 		return;
112 
113 	ret = of_property_read_u32(cpu_node,
114 				   "capacity-dmips-mhz",
115 				   &cpu_capacity);
116 	if (!ret) {
117 		if (!raw_capacity) {
118 			raw_capacity = kcalloc(num_possible_cpus(),
119 					       sizeof(*raw_capacity),
120 					       GFP_KERNEL);
121 			if (!raw_capacity) {
122 				pr_err("cpu_capacity: failed to allocate memory for raw capacities\n");
123 				cap_parsing_failed = true;
124 				return;
125 			}
126 		}
127 		capacity_scale = max(cpu_capacity, capacity_scale);
128 		raw_capacity[cpu] = cpu_capacity;
129 		pr_debug("cpu_capacity: %s cpu_capacity=%u (raw)\n",
130 			cpu_node->full_name, raw_capacity[cpu]);
131 	} else {
132 		if (raw_capacity) {
133 			pr_err("cpu_capacity: missing %s raw capacity\n",
134 				cpu_node->full_name);
135 			pr_err("cpu_capacity: partial information: fallback to 1024 for all CPUs\n");
136 		}
137 		cap_parsing_failed = true;
138 		kfree(raw_capacity);
139 	}
140 }
141 
142 static void normalize_cpu_capacity(void)
143 {
144 	u64 capacity;
145 	int cpu;
146 
147 	if (!raw_capacity || cap_parsing_failed)
148 		return;
149 
150 	pr_debug("cpu_capacity: capacity_scale=%u\n", capacity_scale);
151 	mutex_lock(&cpu_scale_mutex);
152 	for_each_possible_cpu(cpu) {
153 		pr_debug("cpu_capacity: cpu=%d raw_capacity=%u\n",
154 			 cpu, raw_capacity[cpu]);
155 		capacity = (raw_capacity[cpu] << SCHED_CAPACITY_SHIFT)
156 			/ capacity_scale;
157 		set_capacity_scale(cpu, capacity);
158 		pr_debug("cpu_capacity: CPU%d cpu_capacity=%lu\n",
159 			cpu, arch_scale_cpu_capacity(NULL, cpu));
160 	}
161 	mutex_unlock(&cpu_scale_mutex);
162 }
163 
164 #ifdef CONFIG_CPU_FREQ
165 static cpumask_var_t cpus_to_visit;
166 static bool cap_parsing_done;
167 static void parsing_done_workfn(struct work_struct *work);
168 static DECLARE_WORK(parsing_done_work, parsing_done_workfn);
169 
170 static int
171 init_cpu_capacity_callback(struct notifier_block *nb,
172 			   unsigned long val,
173 			   void *data)
174 {
175 	struct cpufreq_policy *policy = data;
176 	int cpu;
177 
178 	if (cap_parsing_failed || cap_parsing_done)
179 		return 0;
180 
181 	switch (val) {
182 	case CPUFREQ_NOTIFY:
183 		pr_debug("cpu_capacity: init cpu capacity for CPUs [%*pbl] (to_visit=%*pbl)\n",
184 				cpumask_pr_args(policy->related_cpus),
185 				cpumask_pr_args(cpus_to_visit));
186 		cpumask_andnot(cpus_to_visit,
187 			       cpus_to_visit,
188 			       policy->related_cpus);
189 		for_each_cpu(cpu, policy->related_cpus) {
190 			raw_capacity[cpu] = arch_scale_cpu_capacity(NULL, cpu) *
191 					    policy->cpuinfo.max_freq / 1000UL;
192 			capacity_scale = max(raw_capacity[cpu], capacity_scale);
193 		}
194 		if (cpumask_empty(cpus_to_visit)) {
195 			normalize_cpu_capacity();
196 			kfree(raw_capacity);
197 			pr_debug("cpu_capacity: parsing done\n");
198 			cap_parsing_done = true;
199 			schedule_work(&parsing_done_work);
200 		}
201 	}
202 	return 0;
203 }
204 
205 static struct notifier_block init_cpu_capacity_notifier = {
206 	.notifier_call = init_cpu_capacity_callback,
207 };
208 
209 static int __init register_cpufreq_notifier(void)
210 {
211 	/*
212 	 * on ACPI-based systems we need to use the default cpu capacity
213 	 * until we have the necessary code to parse the cpu capacity, so
214 	 * skip registering cpufreq notifier.
215 	 */
216 	if (!acpi_disabled || cap_parsing_failed)
217 		return -EINVAL;
218 
219 	if (!alloc_cpumask_var(&cpus_to_visit, GFP_KERNEL)) {
220 		pr_err("cpu_capacity: failed to allocate memory for cpus_to_visit\n");
221 		return -ENOMEM;
222 	}
223 	cpumask_copy(cpus_to_visit, cpu_possible_mask);
224 
225 	return cpufreq_register_notifier(&init_cpu_capacity_notifier,
226 					 CPUFREQ_POLICY_NOTIFIER);
227 }
228 core_initcall(register_cpufreq_notifier);
229 
230 static void parsing_done_workfn(struct work_struct *work)
231 {
232 	cpufreq_unregister_notifier(&init_cpu_capacity_notifier,
233 					 CPUFREQ_POLICY_NOTIFIER);
234 }
235 
236 #else
237 static int __init free_raw_capacity(void)
238 {
239 	kfree(raw_capacity);
240 
241 	return 0;
242 }
243 core_initcall(free_raw_capacity);
244 #endif
245 
246 static int __init get_cpu_for_node(struct device_node *node)
247 {
248 	struct device_node *cpu_node;
249 	int cpu;
250 
251 	cpu_node = of_parse_phandle(node, "cpu", 0);
252 	if (!cpu_node)
253 		return -1;
254 
255 	for_each_possible_cpu(cpu) {
256 		if (of_get_cpu_node(cpu, NULL) == cpu_node) {
257 			parse_cpu_capacity(cpu_node, cpu);
258 			of_node_put(cpu_node);
259 			return cpu;
260 		}
261 	}
262 
263 	pr_crit("Unable to find CPU node for %s\n", cpu_node->full_name);
264 
265 	of_node_put(cpu_node);
266 	return -1;
267 }
268 
269 static int __init parse_core(struct device_node *core, int cluster_id,
270 			     int core_id)
271 {
272 	char name[10];
273 	bool leaf = true;
274 	int i = 0;
275 	int cpu;
276 	struct device_node *t;
277 
278 	do {
279 		snprintf(name, sizeof(name), "thread%d", i);
280 		t = of_get_child_by_name(core, name);
281 		if (t) {
282 			leaf = false;
283 			cpu = get_cpu_for_node(t);
284 			if (cpu >= 0) {
285 				cpu_topology[cpu].cluster_id = cluster_id;
286 				cpu_topology[cpu].core_id = core_id;
287 				cpu_topology[cpu].thread_id = i;
288 			} else {
289 				pr_err("%s: Can't get CPU for thread\n",
290 				       t->full_name);
291 				of_node_put(t);
292 				return -EINVAL;
293 			}
294 			of_node_put(t);
295 		}
296 		i++;
297 	} while (t);
298 
299 	cpu = get_cpu_for_node(core);
300 	if (cpu >= 0) {
301 		if (!leaf) {
302 			pr_err("%s: Core has both threads and CPU\n",
303 			       core->full_name);
304 			return -EINVAL;
305 		}
306 
307 		cpu_topology[cpu].cluster_id = cluster_id;
308 		cpu_topology[cpu].core_id = core_id;
309 	} else if (leaf) {
310 		pr_err("%s: Can't get CPU for leaf core\n", core->full_name);
311 		return -EINVAL;
312 	}
313 
314 	return 0;
315 }
316 
317 static int __init parse_cluster(struct device_node *cluster, int depth)
318 {
319 	char name[10];
320 	bool leaf = true;
321 	bool has_cores = false;
322 	struct device_node *c;
323 	static int cluster_id __initdata;
324 	int core_id = 0;
325 	int i, ret;
326 
327 	/*
328 	 * First check for child clusters; we currently ignore any
329 	 * information about the nesting of clusters and present the
330 	 * scheduler with a flat list of them.
331 	 */
332 	i = 0;
333 	do {
334 		snprintf(name, sizeof(name), "cluster%d", i);
335 		c = of_get_child_by_name(cluster, name);
336 		if (c) {
337 			leaf = false;
338 			ret = parse_cluster(c, depth + 1);
339 			of_node_put(c);
340 			if (ret != 0)
341 				return ret;
342 		}
343 		i++;
344 	} while (c);
345 
346 	/* Now check for cores */
347 	i = 0;
348 	do {
349 		snprintf(name, sizeof(name), "core%d", i);
350 		c = of_get_child_by_name(cluster, name);
351 		if (c) {
352 			has_cores = true;
353 
354 			if (depth == 0) {
355 				pr_err("%s: cpu-map children should be clusters\n",
356 				       c->full_name);
357 				of_node_put(c);
358 				return -EINVAL;
359 			}
360 
361 			if (leaf) {
362 				ret = parse_core(c, cluster_id, core_id++);
363 			} else {
364 				pr_err("%s: Non-leaf cluster with core %s\n",
365 				       cluster->full_name, name);
366 				ret = -EINVAL;
367 			}
368 
369 			of_node_put(c);
370 			if (ret != 0)
371 				return ret;
372 		}
373 		i++;
374 	} while (c);
375 
376 	if (leaf && !has_cores)
377 		pr_warn("%s: empty cluster\n", cluster->full_name);
378 
379 	if (leaf)
380 		cluster_id++;
381 
382 	return 0;
383 }
384 
385 static int __init parse_dt_topology(void)
386 {
387 	struct device_node *cn, *map;
388 	int ret = 0;
389 	int cpu;
390 
391 	cn = of_find_node_by_path("/cpus");
392 	if (!cn) {
393 		pr_err("No CPU information found in DT\n");
394 		return 0;
395 	}
396 
397 	/*
398 	 * When topology is provided cpu-map is essentially a root
399 	 * cluster with restricted subnodes.
400 	 */
401 	map = of_get_child_by_name(cn, "cpu-map");
402 	if (!map) {
403 		cap_parsing_failed = true;
404 		goto out;
405 	}
406 
407 	ret = parse_cluster(map, 0);
408 	if (ret != 0)
409 		goto out_map;
410 
411 	normalize_cpu_capacity();
412 
413 	/*
414 	 * Check that all cores are in the topology; the SMP code will
415 	 * only mark cores described in the DT as possible.
416 	 */
417 	for_each_possible_cpu(cpu)
418 		if (cpu_topology[cpu].cluster_id == -1)
419 			ret = -EINVAL;
420 
421 out_map:
422 	of_node_put(map);
423 out:
424 	of_node_put(cn);
425 	return ret;
426 }
427 
428 /*
429  * cpu topology table
430  */
431 struct cpu_topology cpu_topology[NR_CPUS];
432 EXPORT_SYMBOL_GPL(cpu_topology);
433 
434 const struct cpumask *cpu_coregroup_mask(int cpu)
435 {
436 	return &cpu_topology[cpu].core_sibling;
437 }
438 
439 static void update_siblings_masks(unsigned int cpuid)
440 {
441 	struct cpu_topology *cpu_topo, *cpuid_topo = &cpu_topology[cpuid];
442 	int cpu;
443 
444 	/* update core and thread sibling masks */
445 	for_each_possible_cpu(cpu) {
446 		cpu_topo = &cpu_topology[cpu];
447 
448 		if (cpuid_topo->cluster_id != cpu_topo->cluster_id)
449 			continue;
450 
451 		cpumask_set_cpu(cpuid, &cpu_topo->core_sibling);
452 		if (cpu != cpuid)
453 			cpumask_set_cpu(cpu, &cpuid_topo->core_sibling);
454 
455 		if (cpuid_topo->core_id != cpu_topo->core_id)
456 			continue;
457 
458 		cpumask_set_cpu(cpuid, &cpu_topo->thread_sibling);
459 		if (cpu != cpuid)
460 			cpumask_set_cpu(cpu, &cpuid_topo->thread_sibling);
461 	}
462 }
463 
464 void store_cpu_topology(unsigned int cpuid)
465 {
466 	struct cpu_topology *cpuid_topo = &cpu_topology[cpuid];
467 	u64 mpidr;
468 
469 	if (cpuid_topo->cluster_id != -1)
470 		goto topology_populated;
471 
472 	mpidr = read_cpuid_mpidr();
473 
474 	/* Uniprocessor systems can rely on default topology values */
475 	if (mpidr & MPIDR_UP_BITMASK)
476 		return;
477 
478 	/* Create cpu topology mapping based on MPIDR. */
479 	if (mpidr & MPIDR_MT_BITMASK) {
480 		/* Multiprocessor system : Multi-threads per core */
481 		cpuid_topo->thread_id  = MPIDR_AFFINITY_LEVEL(mpidr, 0);
482 		cpuid_topo->core_id    = MPIDR_AFFINITY_LEVEL(mpidr, 1);
483 		cpuid_topo->cluster_id = MPIDR_AFFINITY_LEVEL(mpidr, 2) |
484 					 MPIDR_AFFINITY_LEVEL(mpidr, 3) << 8;
485 	} else {
486 		/* Multiprocessor system : Single-thread per core */
487 		cpuid_topo->thread_id  = -1;
488 		cpuid_topo->core_id    = MPIDR_AFFINITY_LEVEL(mpidr, 0);
489 		cpuid_topo->cluster_id = MPIDR_AFFINITY_LEVEL(mpidr, 1) |
490 					 MPIDR_AFFINITY_LEVEL(mpidr, 2) << 8 |
491 					 MPIDR_AFFINITY_LEVEL(mpidr, 3) << 16;
492 	}
493 
494 	pr_debug("CPU%u: cluster %d core %d thread %d mpidr %#016llx\n",
495 		 cpuid, cpuid_topo->cluster_id, cpuid_topo->core_id,
496 		 cpuid_topo->thread_id, mpidr);
497 
498 topology_populated:
499 	update_siblings_masks(cpuid);
500 }
501 
502 static void __init reset_cpu_topology(void)
503 {
504 	unsigned int cpu;
505 
506 	for_each_possible_cpu(cpu) {
507 		struct cpu_topology *cpu_topo = &cpu_topology[cpu];
508 
509 		cpu_topo->thread_id = -1;
510 		cpu_topo->core_id = 0;
511 		cpu_topo->cluster_id = -1;
512 
513 		cpumask_clear(&cpu_topo->core_sibling);
514 		cpumask_set_cpu(cpu, &cpu_topo->core_sibling);
515 		cpumask_clear(&cpu_topo->thread_sibling);
516 		cpumask_set_cpu(cpu, &cpu_topo->thread_sibling);
517 	}
518 }
519 
520 void __init init_cpu_topology(void)
521 {
522 	reset_cpu_topology();
523 
524 	/*
525 	 * Discard anything that was parsed if we hit an error so we
526 	 * don't use partial information.
527 	 */
528 	if (of_have_populated_dt() && parse_dt_topology())
529 		reset_cpu_topology();
530 }
531