Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * cacheinfo support - processor cache information via sysfs
4 *
5 * Based on arch/x86/kernel/cpu/intel_cacheinfo.c
6 * Author: Sudeep Holla <sudeep.holla@arm.com>
7 */
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/acpi.h>
11#include <linux/bitops.h>
12#include <linux/cacheinfo.h>
13#include <linux/compiler.h>
14#include <linux/cpu.h>
15#include <linux/device.h>
16#include <linux/init.h>
17#include <linux/of.h>
18#include <linux/sched.h>
19#include <linux/slab.h>
20#include <linux/smp.h>
21#include <linux/sysfs.h>
22
23/* pointer to per cpu cacheinfo */
24static DEFINE_PER_CPU(struct cpu_cacheinfo, ci_cpu_cacheinfo);
25#define ci_cacheinfo(cpu) (&per_cpu(ci_cpu_cacheinfo, cpu))
26#define cache_leaves(cpu) (ci_cacheinfo(cpu)->num_leaves)
27#define per_cpu_cacheinfo(cpu) (ci_cacheinfo(cpu)->info_list)
28#define per_cpu_cacheinfo_idx(cpu, idx) \
29 (per_cpu_cacheinfo(cpu) + (idx))
30
31/* Set if no cache information is found in DT/ACPI. */
32static bool use_arch_info;
33
34struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu)
35{
36 return ci_cacheinfo(cpu);
37}
38
39static inline bool cache_leaves_are_shared(struct cacheinfo *this_leaf,
40 struct cacheinfo *sib_leaf)
41{
42 /*
43 * For non DT/ACPI systems, assume unique level 1 caches,
44 * system-wide shared caches for all other levels.
45 */
46 if (!(IS_ENABLED(CONFIG_OF) || IS_ENABLED(CONFIG_ACPI)) ||
47 use_arch_info)
48 return (this_leaf->level != 1) && (sib_leaf->level != 1);
49
50 if ((sib_leaf->attributes & CACHE_ID) &&
51 (this_leaf->attributes & CACHE_ID))
52 return sib_leaf->id == this_leaf->id;
53
54 return sib_leaf->fw_token == this_leaf->fw_token;
55}
56
57bool last_level_cache_is_valid(unsigned int cpu)
58{
59 struct cacheinfo *llc;
60
61 if (!cache_leaves(cpu))
62 return false;
63
64 llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
65
66 return (llc->attributes & CACHE_ID) || !!llc->fw_token;
67
68}
69
70bool last_level_cache_is_shared(unsigned int cpu_x, unsigned int cpu_y)
71{
72 struct cacheinfo *llc_x, *llc_y;
73
74 if (!last_level_cache_is_valid(cpu_x) ||
75 !last_level_cache_is_valid(cpu_y))
76 return false;
77
78 llc_x = per_cpu_cacheinfo_idx(cpu_x, cache_leaves(cpu_x) - 1);
79 llc_y = per_cpu_cacheinfo_idx(cpu_y, cache_leaves(cpu_y) - 1);
80
81 return cache_leaves_are_shared(llc_x, llc_y);
82}
83
84#ifdef CONFIG_OF
85
86static bool of_check_cache_nodes(struct device_node *np);
87
88/* OF properties to query for a given cache type */
89struct cache_type_info {
90 const char *size_prop;
91 const char *line_size_props[2];
92 const char *nr_sets_prop;
93};
94
95static const struct cache_type_info cache_type_info[] = {
96 {
97 .size_prop = "cache-size",
98 .line_size_props = { "cache-line-size",
99 "cache-block-size", },
100 .nr_sets_prop = "cache-sets",
101 }, {
102 .size_prop = "i-cache-size",
103 .line_size_props = { "i-cache-line-size",
104 "i-cache-block-size", },
105 .nr_sets_prop = "i-cache-sets",
106 }, {
107 .size_prop = "d-cache-size",
108 .line_size_props = { "d-cache-line-size",
109 "d-cache-block-size", },
110 .nr_sets_prop = "d-cache-sets",
111 },
112};
113
114static inline int get_cacheinfo_idx(enum cache_type type)
115{
116 if (type == CACHE_TYPE_UNIFIED)
117 return 0;
118 return type;
119}
120
121static void cache_size(struct cacheinfo *this_leaf, struct device_node *np)
122{
123 const char *propname;
124 int ct_idx;
125
126 ct_idx = get_cacheinfo_idx(this_leaf->type);
127 propname = cache_type_info[ct_idx].size_prop;
128
129 of_property_read_u32(np, propname, &this_leaf->size);
130}
131
132/* not cache_line_size() because that's a macro in include/linux/cache.h */
133static void cache_get_line_size(struct cacheinfo *this_leaf,
134 struct device_node *np)
135{
136 int i, lim, ct_idx;
137
138 ct_idx = get_cacheinfo_idx(this_leaf->type);
139 lim = ARRAY_SIZE(cache_type_info[ct_idx].line_size_props);
140
141 for (i = 0; i < lim; i++) {
142 int ret;
143 u32 line_size;
144 const char *propname;
145
146 propname = cache_type_info[ct_idx].line_size_props[i];
147 ret = of_property_read_u32(np, propname, &line_size);
148 if (!ret) {
149 this_leaf->coherency_line_size = line_size;
150 break;
151 }
152 }
153}
154
155static void cache_nr_sets(struct cacheinfo *this_leaf, struct device_node *np)
156{
157 const char *propname;
158 int ct_idx;
159
160 ct_idx = get_cacheinfo_idx(this_leaf->type);
161 propname = cache_type_info[ct_idx].nr_sets_prop;
162
163 of_property_read_u32(np, propname, &this_leaf->number_of_sets);
164}
165
166static void cache_associativity(struct cacheinfo *this_leaf)
167{
168 unsigned int line_size = this_leaf->coherency_line_size;
169 unsigned int nr_sets = this_leaf->number_of_sets;
170 unsigned int size = this_leaf->size;
171
172 /*
173 * If the cache is fully associative, there is no need to
174 * check the other properties.
175 */
176 if (!(nr_sets == 1) && (nr_sets > 0 && size > 0 && line_size > 0))
177 this_leaf->ways_of_associativity = (size / nr_sets) / line_size;
178}
179
180static bool cache_node_is_unified(struct cacheinfo *this_leaf,
181 struct device_node *np)
182{
183 return of_property_read_bool(np, "cache-unified");
184}
185
186static void cache_of_set_props(struct cacheinfo *this_leaf,
187 struct device_node *np)
188{
189 /*
190 * init_cache_level must setup the cache level correctly
191 * overriding the architecturally specified levels, so
192 * if type is NONE at this stage, it should be unified
193 */
194 if (this_leaf->type == CACHE_TYPE_NOCACHE &&
195 cache_node_is_unified(this_leaf, np))
196 this_leaf->type = CACHE_TYPE_UNIFIED;
197 cache_size(this_leaf, np);
198 cache_get_line_size(this_leaf, np);
199 cache_nr_sets(this_leaf, np);
200 cache_associativity(this_leaf);
201}
202
203static int cache_setup_of_node(unsigned int cpu)
204{
205 struct device_node *np, *prev;
206 struct cacheinfo *this_leaf;
207 unsigned int index = 0;
208
209 np = of_cpu_device_node_get(cpu);
210 if (!np) {
211 pr_err("Failed to find cpu%d device node\n", cpu);
212 return -ENOENT;
213 }
214
215 if (!of_check_cache_nodes(np)) {
216 of_node_put(np);
217 return -ENOENT;
218 }
219
220 prev = np;
221
222 while (index < cache_leaves(cpu)) {
223 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
224 if (this_leaf->level != 1) {
225 np = of_find_next_cache_node(np);
226 of_node_put(prev);
227 prev = np;
228 if (!np)
229 break;
230 }
231 cache_of_set_props(this_leaf, np);
232 this_leaf->fw_token = np;
233 index++;
234 }
235
236 of_node_put(np);
237
238 if (index != cache_leaves(cpu)) /* not all OF nodes populated */
239 return -ENOENT;
240
241 return 0;
242}
243
244static bool of_check_cache_nodes(struct device_node *np)
245{
246 struct device_node *next;
247
248 if (of_property_present(np, "cache-size") ||
249 of_property_present(np, "i-cache-size") ||
250 of_property_present(np, "d-cache-size") ||
251 of_property_present(np, "cache-unified"))
252 return true;
253
254 next = of_find_next_cache_node(np);
255 if (next) {
256 of_node_put(next);
257 return true;
258 }
259
260 return false;
261}
262
263static int of_count_cache_leaves(struct device_node *np)
264{
265 unsigned int leaves = 0;
266
267 if (of_property_read_bool(np, "cache-size"))
268 ++leaves;
269 if (of_property_read_bool(np, "i-cache-size"))
270 ++leaves;
271 if (of_property_read_bool(np, "d-cache-size"))
272 ++leaves;
273
274 if (!leaves) {
275 /* The '[i-|d-|]cache-size' property is required, but
276 * if absent, fallback on the 'cache-unified' property.
277 */
278 if (of_property_read_bool(np, "cache-unified"))
279 return 1;
280 else
281 return 2;
282 }
283
284 return leaves;
285}
286
287int init_of_cache_level(unsigned int cpu)
288{
289 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
290 struct device_node *np = of_cpu_device_node_get(cpu);
291 struct device_node *prev = NULL;
292 unsigned int levels = 0, leaves, level;
293
294 if (!of_check_cache_nodes(np)) {
295 of_node_put(np);
296 return -ENOENT;
297 }
298
299 leaves = of_count_cache_leaves(np);
300 if (leaves > 0)
301 levels = 1;
302
303 prev = np;
304 while ((np = of_find_next_cache_node(np))) {
305 of_node_put(prev);
306 prev = np;
307 if (!of_device_is_compatible(np, "cache"))
308 goto err_out;
309 if (of_property_read_u32(np, "cache-level", &level))
310 goto err_out;
311 if (level <= levels)
312 goto err_out;
313
314 leaves += of_count_cache_leaves(np);
315 levels = level;
316 }
317
318 of_node_put(np);
319 this_cpu_ci->num_levels = levels;
320 this_cpu_ci->num_leaves = leaves;
321
322 return 0;
323
324err_out:
325 of_node_put(np);
326 return -EINVAL;
327}
328
329#else
330static inline int cache_setup_of_node(unsigned int cpu) { return 0; }
331int init_of_cache_level(unsigned int cpu) { return 0; }
332#endif
333
334int __weak cache_setup_acpi(unsigned int cpu)
335{
336 return -ENOTSUPP;
337}
338
339unsigned int coherency_max_size;
340
341static int cache_setup_properties(unsigned int cpu)
342{
343 int ret = 0;
344
345 if (of_have_populated_dt())
346 ret = cache_setup_of_node(cpu);
347 else if (!acpi_disabled)
348 ret = cache_setup_acpi(cpu);
349
350 // Assume there is no cache information available in DT/ACPI from now.
351 if (ret && use_arch_cache_info())
352 use_arch_info = true;
353
354 return ret;
355}
356
357static int cache_shared_cpu_map_setup(unsigned int cpu)
358{
359 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
360 struct cacheinfo *this_leaf, *sib_leaf;
361 unsigned int index, sib_index;
362 int ret = 0;
363
364 if (this_cpu_ci->cpu_map_populated)
365 return 0;
366
367 /*
368 * skip setting up cache properties if LLC is valid, just need
369 * to update the shared cpu_map if the cache attributes were
370 * populated early before all the cpus are brought online
371 */
372 if (!last_level_cache_is_valid(cpu) && !use_arch_info) {
373 ret = cache_setup_properties(cpu);
374 if (ret)
375 return ret;
376 }
377
378 for (index = 0; index < cache_leaves(cpu); index++) {
379 unsigned int i;
380
381 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
382
383 cpumask_set_cpu(cpu, &this_leaf->shared_cpu_map);
384 for_each_online_cpu(i) {
385 struct cpu_cacheinfo *sib_cpu_ci = get_cpu_cacheinfo(i);
386
387 if (i == cpu || !sib_cpu_ci->info_list)
388 continue;/* skip if itself or no cacheinfo */
389 for (sib_index = 0; sib_index < cache_leaves(i); sib_index++) {
390 sib_leaf = per_cpu_cacheinfo_idx(i, sib_index);
391
392 /*
393 * Comparing cache IDs only makes sense if the leaves
394 * belong to the same cache level of same type. Skip
395 * the check if level and type do not match.
396 */
397 if (sib_leaf->level != this_leaf->level ||
398 sib_leaf->type != this_leaf->type)
399 continue;
400
401 if (cache_leaves_are_shared(this_leaf, sib_leaf)) {
402 cpumask_set_cpu(cpu, &sib_leaf->shared_cpu_map);
403 cpumask_set_cpu(i, &this_leaf->shared_cpu_map);
404 break;
405 }
406 }
407 }
408 /* record the maximum cache line size */
409 if (this_leaf->coherency_line_size > coherency_max_size)
410 coherency_max_size = this_leaf->coherency_line_size;
411 }
412
413 /* shared_cpu_map is now populated for the cpu */
414 this_cpu_ci->cpu_map_populated = true;
415 return 0;
416}
417
418static void cache_shared_cpu_map_remove(unsigned int cpu)
419{
420 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
421 struct cacheinfo *this_leaf, *sib_leaf;
422 unsigned int sibling, index, sib_index;
423
424 for (index = 0; index < cache_leaves(cpu); index++) {
425 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
426 for_each_cpu(sibling, &this_leaf->shared_cpu_map) {
427 struct cpu_cacheinfo *sib_cpu_ci =
428 get_cpu_cacheinfo(sibling);
429
430 if (sibling == cpu || !sib_cpu_ci->info_list)
431 continue;/* skip if itself or no cacheinfo */
432
433 for (sib_index = 0; sib_index < cache_leaves(sibling); sib_index++) {
434 sib_leaf = per_cpu_cacheinfo_idx(sibling, sib_index);
435
436 /*
437 * Comparing cache IDs only makes sense if the leaves
438 * belong to the same cache level of same type. Skip
439 * the check if level and type do not match.
440 */
441 if (sib_leaf->level != this_leaf->level ||
442 sib_leaf->type != this_leaf->type)
443 continue;
444
445 if (cache_leaves_are_shared(this_leaf, sib_leaf)) {
446 cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
447 cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
448 break;
449 }
450 }
451 }
452 }
453
454 /* cpu is no longer populated in the shared map */
455 this_cpu_ci->cpu_map_populated = false;
456}
457
458static void free_cache_attributes(unsigned int cpu)
459{
460 if (!per_cpu_cacheinfo(cpu))
461 return;
462
463 cache_shared_cpu_map_remove(cpu);
464}
465
466int __weak early_cache_level(unsigned int cpu)
467{
468 return -ENOENT;
469}
470
471int __weak init_cache_level(unsigned int cpu)
472{
473 return -ENOENT;
474}
475
476int __weak populate_cache_leaves(unsigned int cpu)
477{
478 return -ENOENT;
479}
480
481static inline
482int allocate_cache_info(int cpu)
483{
484 per_cpu_cacheinfo(cpu) = kcalloc(cache_leaves(cpu),
485 sizeof(struct cacheinfo), GFP_ATOMIC);
486 if (!per_cpu_cacheinfo(cpu)) {
487 cache_leaves(cpu) = 0;
488 return -ENOMEM;
489 }
490
491 return 0;
492}
493
494int fetch_cache_info(unsigned int cpu)
495{
496 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
497 unsigned int levels = 0, split_levels = 0;
498 int ret;
499
500 if (acpi_disabled) {
501 ret = init_of_cache_level(cpu);
502 } else {
503 ret = acpi_get_cache_info(cpu, &levels, &split_levels);
504 if (!ret) {
505 this_cpu_ci->num_levels = levels;
506 /*
507 * This assumes that:
508 * - there cannot be any split caches (data/instruction)
509 * above a unified cache
510 * - data/instruction caches come by pair
511 */
512 this_cpu_ci->num_leaves = levels + split_levels;
513 }
514 }
515
516 if (ret || !cache_leaves(cpu)) {
517 ret = early_cache_level(cpu);
518 if (ret)
519 return ret;
520
521 if (!cache_leaves(cpu))
522 return -ENOENT;
523
524 this_cpu_ci->early_ci_levels = true;
525 }
526
527 return allocate_cache_info(cpu);
528}
529
530static inline int init_level_allocate_ci(unsigned int cpu)
531{
532 unsigned int early_leaves = cache_leaves(cpu);
533
534 /* Since early initialization/allocation of the cacheinfo is allowed
535 * via fetch_cache_info() and this also gets called as CPU hotplug
536 * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped
537 * as it will happen only once (the cacheinfo memory is never freed).
538 * Just populate the cacheinfo. However, if the cacheinfo has been
539 * allocated early through the arch-specific early_cache_level() call,
540 * there is a chance the info is wrong (this can happen on arm64). In
541 * that case, call init_cache_level() anyway to give the arch-specific
542 * code a chance to make things right.
543 */
544 if (per_cpu_cacheinfo(cpu) && !ci_cacheinfo(cpu)->early_ci_levels)
545 return 0;
546
547 if (init_cache_level(cpu) || !cache_leaves(cpu))
548 return -ENOENT;
549
550 /*
551 * Now that we have properly initialized the cache level info, make
552 * sure we don't try to do that again the next time we are called
553 * (e.g. as CPU hotplug callbacks).
554 */
555 ci_cacheinfo(cpu)->early_ci_levels = false;
556
557 if (cache_leaves(cpu) <= early_leaves)
558 return 0;
559
560 kfree(per_cpu_cacheinfo(cpu));
561 return allocate_cache_info(cpu);
562}
563
564int detect_cache_attributes(unsigned int cpu)
565{
566 int ret;
567
568 ret = init_level_allocate_ci(cpu);
569 if (ret)
570 return ret;
571
572 /*
573 * If LLC is valid the cache leaves were already populated so just go to
574 * update the cpu map.
575 */
576 if (!last_level_cache_is_valid(cpu)) {
577 /*
578 * populate_cache_leaves() may completely setup the cache leaves and
579 * shared_cpu_map or it may leave it partially setup.
580 */
581 ret = populate_cache_leaves(cpu);
582 if (ret)
583 goto free_ci;
584 }
585
586 /*
587 * For systems using DT for cache hierarchy, fw_token
588 * and shared_cpu_map will be set up here only if they are
589 * not populated already
590 */
591 ret = cache_shared_cpu_map_setup(cpu);
592 if (ret) {
593 pr_warn("Unable to detect cache hierarchy for CPU %d\n", cpu);
594 goto free_ci;
595 }
596
597 return 0;
598
599free_ci:
600 free_cache_attributes(cpu);
601 return ret;
602}
603
604/* pointer to cpuX/cache device */
605static DEFINE_PER_CPU(struct device *, ci_cache_dev);
606#define per_cpu_cache_dev(cpu) (per_cpu(ci_cache_dev, cpu))
607
608static cpumask_t cache_dev_map;
609
610/* pointer to array of devices for cpuX/cache/indexY */
611static DEFINE_PER_CPU(struct device **, ci_index_dev);
612#define per_cpu_index_dev(cpu) (per_cpu(ci_index_dev, cpu))
613#define per_cache_index_dev(cpu, idx) ((per_cpu_index_dev(cpu))[idx])
614
615#define show_one(file_name, object) \
616static ssize_t file_name##_show(struct device *dev, \
617 struct device_attribute *attr, char *buf) \
618{ \
619 struct cacheinfo *this_leaf = dev_get_drvdata(dev); \
620 return sysfs_emit(buf, "%u\n", this_leaf->object); \
621}
622
623show_one(id, id);
624show_one(level, level);
625show_one(coherency_line_size, coherency_line_size);
626show_one(number_of_sets, number_of_sets);
627show_one(physical_line_partition, physical_line_partition);
628show_one(ways_of_associativity, ways_of_associativity);
629
630static ssize_t size_show(struct device *dev,
631 struct device_attribute *attr, char *buf)
632{
633 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
634
635 return sysfs_emit(buf, "%uK\n", this_leaf->size >> 10);
636}
637
638static ssize_t shared_cpu_map_show(struct device *dev,
639 struct device_attribute *attr, char *buf)
640{
641 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
642 const struct cpumask *mask = &this_leaf->shared_cpu_map;
643
644 return sysfs_emit(buf, "%*pb\n", nr_cpu_ids, mask);
645}
646
647static ssize_t shared_cpu_list_show(struct device *dev,
648 struct device_attribute *attr, char *buf)
649{
650 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
651 const struct cpumask *mask = &this_leaf->shared_cpu_map;
652
653 return sysfs_emit(buf, "%*pbl\n", nr_cpu_ids, mask);
654}
655
656static ssize_t type_show(struct device *dev,
657 struct device_attribute *attr, char *buf)
658{
659 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
660 const char *output;
661
662 switch (this_leaf->type) {
663 case CACHE_TYPE_DATA:
664 output = "Data";
665 break;
666 case CACHE_TYPE_INST:
667 output = "Instruction";
668 break;
669 case CACHE_TYPE_UNIFIED:
670 output = "Unified";
671 break;
672 default:
673 return -EINVAL;
674 }
675
676 return sysfs_emit(buf, "%s\n", output);
677}
678
679static ssize_t allocation_policy_show(struct device *dev,
680 struct device_attribute *attr, char *buf)
681{
682 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
683 unsigned int ci_attr = this_leaf->attributes;
684 const char *output;
685
686 if ((ci_attr & CACHE_READ_ALLOCATE) && (ci_attr & CACHE_WRITE_ALLOCATE))
687 output = "ReadWriteAllocate";
688 else if (ci_attr & CACHE_READ_ALLOCATE)
689 output = "ReadAllocate";
690 else if (ci_attr & CACHE_WRITE_ALLOCATE)
691 output = "WriteAllocate";
692 else
693 return 0;
694
695 return sysfs_emit(buf, "%s\n", output);
696}
697
698static ssize_t write_policy_show(struct device *dev,
699 struct device_attribute *attr, char *buf)
700{
701 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
702 unsigned int ci_attr = this_leaf->attributes;
703 int n = 0;
704
705 if (ci_attr & CACHE_WRITE_THROUGH)
706 n = sysfs_emit(buf, "WriteThrough\n");
707 else if (ci_attr & CACHE_WRITE_BACK)
708 n = sysfs_emit(buf, "WriteBack\n");
709 return n;
710}
711
712static DEVICE_ATTR_RO(id);
713static DEVICE_ATTR_RO(level);
714static DEVICE_ATTR_RO(type);
715static DEVICE_ATTR_RO(coherency_line_size);
716static DEVICE_ATTR_RO(ways_of_associativity);
717static DEVICE_ATTR_RO(number_of_sets);
718static DEVICE_ATTR_RO(size);
719static DEVICE_ATTR_RO(allocation_policy);
720static DEVICE_ATTR_RO(write_policy);
721static DEVICE_ATTR_RO(shared_cpu_map);
722static DEVICE_ATTR_RO(shared_cpu_list);
723static DEVICE_ATTR_RO(physical_line_partition);
724
725static struct attribute *cache_default_attrs[] = {
726 &dev_attr_id.attr,
727 &dev_attr_type.attr,
728 &dev_attr_level.attr,
729 &dev_attr_shared_cpu_map.attr,
730 &dev_attr_shared_cpu_list.attr,
731 &dev_attr_coherency_line_size.attr,
732 &dev_attr_ways_of_associativity.attr,
733 &dev_attr_number_of_sets.attr,
734 &dev_attr_size.attr,
735 &dev_attr_allocation_policy.attr,
736 &dev_attr_write_policy.attr,
737 &dev_attr_physical_line_partition.attr,
738 NULL
739};
740
741static umode_t
742cache_default_attrs_is_visible(struct kobject *kobj,
743 struct attribute *attr, int unused)
744{
745 struct device *dev = kobj_to_dev(kobj);
746 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
747 const struct cpumask *mask = &this_leaf->shared_cpu_map;
748 umode_t mode = attr->mode;
749
750 if ((attr == &dev_attr_id.attr) && (this_leaf->attributes & CACHE_ID))
751 return mode;
752 if ((attr == &dev_attr_type.attr) && this_leaf->type)
753 return mode;
754 if ((attr == &dev_attr_level.attr) && this_leaf->level)
755 return mode;
756 if ((attr == &dev_attr_shared_cpu_map.attr) && !cpumask_empty(mask))
757 return mode;
758 if ((attr == &dev_attr_shared_cpu_list.attr) && !cpumask_empty(mask))
759 return mode;
760 if ((attr == &dev_attr_coherency_line_size.attr) &&
761 this_leaf->coherency_line_size)
762 return mode;
763 if ((attr == &dev_attr_ways_of_associativity.attr) &&
764 this_leaf->size) /* allow 0 = full associativity */
765 return mode;
766 if ((attr == &dev_attr_number_of_sets.attr) &&
767 this_leaf->number_of_sets)
768 return mode;
769 if ((attr == &dev_attr_size.attr) && this_leaf->size)
770 return mode;
771 if ((attr == &dev_attr_write_policy.attr) &&
772 (this_leaf->attributes & CACHE_WRITE_POLICY_MASK))
773 return mode;
774 if ((attr == &dev_attr_allocation_policy.attr) &&
775 (this_leaf->attributes & CACHE_ALLOCATE_POLICY_MASK))
776 return mode;
777 if ((attr == &dev_attr_physical_line_partition.attr) &&
778 this_leaf->physical_line_partition)
779 return mode;
780
781 return 0;
782}
783
784static const struct attribute_group cache_default_group = {
785 .attrs = cache_default_attrs,
786 .is_visible = cache_default_attrs_is_visible,
787};
788
789static const struct attribute_group *cache_default_groups[] = {
790 &cache_default_group,
791 NULL,
792};
793
794static const struct attribute_group *cache_private_groups[] = {
795 &cache_default_group,
796 NULL, /* Place holder for private group */
797 NULL,
798};
799
800const struct attribute_group *
801__weak cache_get_priv_group(struct cacheinfo *this_leaf)
802{
803 return NULL;
804}
805
806static const struct attribute_group **
807cache_get_attribute_groups(struct cacheinfo *this_leaf)
808{
809 const struct attribute_group *priv_group =
810 cache_get_priv_group(this_leaf);
811
812 if (!priv_group)
813 return cache_default_groups;
814
815 if (!cache_private_groups[1])
816 cache_private_groups[1] = priv_group;
817
818 return cache_private_groups;
819}
820
821/* Add/Remove cache interface for CPU device */
822static void cpu_cache_sysfs_exit(unsigned int cpu)
823{
824 int i;
825 struct device *ci_dev;
826
827 if (per_cpu_index_dev(cpu)) {
828 for (i = 0; i < cache_leaves(cpu); i++) {
829 ci_dev = per_cache_index_dev(cpu, i);
830 if (!ci_dev)
831 continue;
832 device_unregister(ci_dev);
833 }
834 kfree(per_cpu_index_dev(cpu));
835 per_cpu_index_dev(cpu) = NULL;
836 }
837 device_unregister(per_cpu_cache_dev(cpu));
838 per_cpu_cache_dev(cpu) = NULL;
839}
840
841static int cpu_cache_sysfs_init(unsigned int cpu)
842{
843 struct device *dev = get_cpu_device(cpu);
844
845 if (per_cpu_cacheinfo(cpu) == NULL)
846 return -ENOENT;
847
848 per_cpu_cache_dev(cpu) = cpu_device_create(dev, NULL, NULL, "cache");
849 if (IS_ERR(per_cpu_cache_dev(cpu)))
850 return PTR_ERR(per_cpu_cache_dev(cpu));
851
852 /* Allocate all required memory */
853 per_cpu_index_dev(cpu) = kcalloc(cache_leaves(cpu),
854 sizeof(struct device *), GFP_KERNEL);
855 if (unlikely(per_cpu_index_dev(cpu) == NULL))
856 goto err_out;
857
858 return 0;
859
860err_out:
861 cpu_cache_sysfs_exit(cpu);
862 return -ENOMEM;
863}
864
865static int cache_add_dev(unsigned int cpu)
866{
867 unsigned int i;
868 int rc;
869 struct device *ci_dev, *parent;
870 struct cacheinfo *this_leaf;
871 const struct attribute_group **cache_groups;
872
873 rc = cpu_cache_sysfs_init(cpu);
874 if (unlikely(rc < 0))
875 return rc;
876
877 parent = per_cpu_cache_dev(cpu);
878 for (i = 0; i < cache_leaves(cpu); i++) {
879 this_leaf = per_cpu_cacheinfo_idx(cpu, i);
880 if (this_leaf->disable_sysfs)
881 continue;
882 if (this_leaf->type == CACHE_TYPE_NOCACHE)
883 break;
884 cache_groups = cache_get_attribute_groups(this_leaf);
885 ci_dev = cpu_device_create(parent, this_leaf, cache_groups,
886 "index%1u", i);
887 if (IS_ERR(ci_dev)) {
888 rc = PTR_ERR(ci_dev);
889 goto err;
890 }
891 per_cache_index_dev(cpu, i) = ci_dev;
892 }
893 cpumask_set_cpu(cpu, &cache_dev_map);
894
895 return 0;
896err:
897 cpu_cache_sysfs_exit(cpu);
898 return rc;
899}
900
901static unsigned int cpu_map_shared_cache(bool online, unsigned int cpu,
902 cpumask_t **map)
903{
904 struct cacheinfo *llc, *sib_llc;
905 unsigned int sibling;
906
907 if (!last_level_cache_is_valid(cpu))
908 return 0;
909
910 llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
911
912 if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
913 return 0;
914
915 if (online) {
916 *map = &llc->shared_cpu_map;
917 return cpumask_weight(*map);
918 }
919
920 /* shared_cpu_map of offlined CPU will be cleared, so use sibling map */
921 for_each_cpu(sibling, &llc->shared_cpu_map) {
922 if (sibling == cpu || !last_level_cache_is_valid(sibling))
923 continue;
924 sib_llc = per_cpu_cacheinfo_idx(sibling, cache_leaves(sibling) - 1);
925 *map = &sib_llc->shared_cpu_map;
926 return cpumask_weight(*map);
927 }
928
929 return 0;
930}
931
932/*
933 * Calculate the size of the per-CPU data cache slice. This can be
934 * used to estimate the size of the data cache slice that can be used
935 * by one CPU under ideal circumstances. UNIFIED caches are counted
936 * in addition to DATA caches. So, please consider code cache usage
937 * when use the result.
938 *
939 * Because the cache inclusive/non-inclusive information isn't
940 * available, we just use the size of the per-CPU slice of LLC to make
941 * the result more predictable across architectures.
942 */
943static void update_per_cpu_data_slice_size_cpu(unsigned int cpu)
944{
945 struct cpu_cacheinfo *ci;
946 struct cacheinfo *llc;
947 unsigned int nr_shared;
948
949 if (!last_level_cache_is_valid(cpu))
950 return;
951
952 ci = ci_cacheinfo(cpu);
953 llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
954
955 if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
956 return;
957
958 nr_shared = cpumask_weight(&llc->shared_cpu_map);
959 if (nr_shared)
960 ci->per_cpu_data_slice_size = llc->size / nr_shared;
961}
962
963static void update_per_cpu_data_slice_size(bool cpu_online, unsigned int cpu,
964 cpumask_t *cpu_map)
965{
966 unsigned int icpu;
967
968 for_each_cpu(icpu, cpu_map) {
969 if (!cpu_online && icpu == cpu)
970 continue;
971 update_per_cpu_data_slice_size_cpu(icpu);
972 setup_pcp_cacheinfo(icpu);
973 }
974}
975
976static int cacheinfo_cpu_online(unsigned int cpu)
977{
978 int rc = detect_cache_attributes(cpu);
979 cpumask_t *cpu_map;
980
981 if (rc)
982 return rc;
983 rc = cache_add_dev(cpu);
984 if (rc)
985 goto err;
986 if (cpu_map_shared_cache(true, cpu, &cpu_map))
987 update_per_cpu_data_slice_size(true, cpu, cpu_map);
988 return 0;
989err:
990 free_cache_attributes(cpu);
991 return rc;
992}
993
994static int cacheinfo_cpu_pre_down(unsigned int cpu)
995{
996 cpumask_t *cpu_map;
997 unsigned int nr_shared;
998
999 nr_shared = cpu_map_shared_cache(false, cpu, &cpu_map);
1000 if (cpumask_test_and_clear_cpu(cpu, &cache_dev_map))
1001 cpu_cache_sysfs_exit(cpu);
1002
1003 free_cache_attributes(cpu);
1004 if (nr_shared > 1)
1005 update_per_cpu_data_slice_size(false, cpu, cpu_map);
1006 return 0;
1007}
1008
1009static int __init cacheinfo_sysfs_init(void)
1010{
1011 return cpuhp_setup_state(CPUHP_AP_BASE_CACHEINFO_ONLINE,
1012 "base/cacheinfo:online",
1013 cacheinfo_cpu_online, cacheinfo_cpu_pre_down);
1014}
1015device_initcall(cacheinfo_sysfs_init);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * cacheinfo support - processor cache information via sysfs
4 *
5 * Based on arch/x86/kernel/cpu/intel_cacheinfo.c
6 * Author: Sudeep Holla <sudeep.holla@arm.com>
7 */
8#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
9
10#include <linux/acpi.h>
11#include <linux/bitops.h>
12#include <linux/cacheinfo.h>
13#include <linux/compiler.h>
14#include <linux/cpu.h>
15#include <linux/device.h>
16#include <linux/init.h>
17#include <linux/of.h>
18#include <linux/sched.h>
19#include <linux/slab.h>
20#include <linux/smp.h>
21#include <linux/sysfs.h>
22
23/* pointer to per cpu cacheinfo */
24static DEFINE_PER_CPU(struct cpu_cacheinfo, ci_cpu_cacheinfo);
25#define ci_cacheinfo(cpu) (&per_cpu(ci_cpu_cacheinfo, cpu))
26#define cache_leaves(cpu) (ci_cacheinfo(cpu)->num_leaves)
27#define per_cpu_cacheinfo(cpu) (ci_cacheinfo(cpu)->info_list)
28#define per_cpu_cacheinfo_idx(cpu, idx) \
29 (per_cpu_cacheinfo(cpu) + (idx))
30
31/* Set if no cache information is found in DT/ACPI. */
32static bool use_arch_info;
33
34struct cpu_cacheinfo *get_cpu_cacheinfo(unsigned int cpu)
35{
36 return ci_cacheinfo(cpu);
37}
38
39static inline bool cache_leaves_are_shared(struct cacheinfo *this_leaf,
40 struct cacheinfo *sib_leaf)
41{
42 /*
43 * For non DT/ACPI systems, assume unique level 1 caches,
44 * system-wide shared caches for all other levels.
45 */
46 if (!(IS_ENABLED(CONFIG_OF) || IS_ENABLED(CONFIG_ACPI)) ||
47 use_arch_info)
48 return (this_leaf->level != 1) && (sib_leaf->level != 1);
49
50 if ((sib_leaf->attributes & CACHE_ID) &&
51 (this_leaf->attributes & CACHE_ID))
52 return sib_leaf->id == this_leaf->id;
53
54 return sib_leaf->fw_token == this_leaf->fw_token;
55}
56
57bool last_level_cache_is_valid(unsigned int cpu)
58{
59 struct cacheinfo *llc;
60
61 if (!cache_leaves(cpu) || !per_cpu_cacheinfo(cpu))
62 return false;
63
64 llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
65
66 return (llc->attributes & CACHE_ID) || !!llc->fw_token;
67
68}
69
70bool last_level_cache_is_shared(unsigned int cpu_x, unsigned int cpu_y)
71{
72 struct cacheinfo *llc_x, *llc_y;
73
74 if (!last_level_cache_is_valid(cpu_x) ||
75 !last_level_cache_is_valid(cpu_y))
76 return false;
77
78 llc_x = per_cpu_cacheinfo_idx(cpu_x, cache_leaves(cpu_x) - 1);
79 llc_y = per_cpu_cacheinfo_idx(cpu_y, cache_leaves(cpu_y) - 1);
80
81 return cache_leaves_are_shared(llc_x, llc_y);
82}
83
84#ifdef CONFIG_OF
85
86static bool of_check_cache_nodes(struct device_node *np);
87
88/* OF properties to query for a given cache type */
89struct cache_type_info {
90 const char *size_prop;
91 const char *line_size_props[2];
92 const char *nr_sets_prop;
93};
94
95static const struct cache_type_info cache_type_info[] = {
96 {
97 .size_prop = "cache-size",
98 .line_size_props = { "cache-line-size",
99 "cache-block-size", },
100 .nr_sets_prop = "cache-sets",
101 }, {
102 .size_prop = "i-cache-size",
103 .line_size_props = { "i-cache-line-size",
104 "i-cache-block-size", },
105 .nr_sets_prop = "i-cache-sets",
106 }, {
107 .size_prop = "d-cache-size",
108 .line_size_props = { "d-cache-line-size",
109 "d-cache-block-size", },
110 .nr_sets_prop = "d-cache-sets",
111 },
112};
113
114static inline int get_cacheinfo_idx(enum cache_type type)
115{
116 if (type == CACHE_TYPE_UNIFIED)
117 return 0;
118 return type;
119}
120
121static void cache_size(struct cacheinfo *this_leaf, struct device_node *np)
122{
123 const char *propname;
124 int ct_idx;
125
126 ct_idx = get_cacheinfo_idx(this_leaf->type);
127 propname = cache_type_info[ct_idx].size_prop;
128
129 of_property_read_u32(np, propname, &this_leaf->size);
130}
131
132/* not cache_line_size() because that's a macro in include/linux/cache.h */
133static void cache_get_line_size(struct cacheinfo *this_leaf,
134 struct device_node *np)
135{
136 int i, lim, ct_idx;
137
138 ct_idx = get_cacheinfo_idx(this_leaf->type);
139 lim = ARRAY_SIZE(cache_type_info[ct_idx].line_size_props);
140
141 for (i = 0; i < lim; i++) {
142 int ret;
143 u32 line_size;
144 const char *propname;
145
146 propname = cache_type_info[ct_idx].line_size_props[i];
147 ret = of_property_read_u32(np, propname, &line_size);
148 if (!ret) {
149 this_leaf->coherency_line_size = line_size;
150 break;
151 }
152 }
153}
154
155static void cache_nr_sets(struct cacheinfo *this_leaf, struct device_node *np)
156{
157 const char *propname;
158 int ct_idx;
159
160 ct_idx = get_cacheinfo_idx(this_leaf->type);
161 propname = cache_type_info[ct_idx].nr_sets_prop;
162
163 of_property_read_u32(np, propname, &this_leaf->number_of_sets);
164}
165
166static void cache_associativity(struct cacheinfo *this_leaf)
167{
168 unsigned int line_size = this_leaf->coherency_line_size;
169 unsigned int nr_sets = this_leaf->number_of_sets;
170 unsigned int size = this_leaf->size;
171
172 /*
173 * If the cache is fully associative, there is no need to
174 * check the other properties.
175 */
176 if (!(nr_sets == 1) && (nr_sets > 0 && size > 0 && line_size > 0))
177 this_leaf->ways_of_associativity = (size / nr_sets) / line_size;
178}
179
180static bool cache_node_is_unified(struct cacheinfo *this_leaf,
181 struct device_node *np)
182{
183 return of_property_read_bool(np, "cache-unified");
184}
185
186static void cache_of_set_props(struct cacheinfo *this_leaf,
187 struct device_node *np)
188{
189 /*
190 * init_cache_level must setup the cache level correctly
191 * overriding the architecturally specified levels, so
192 * if type is NONE at this stage, it should be unified
193 */
194 if (this_leaf->type == CACHE_TYPE_NOCACHE &&
195 cache_node_is_unified(this_leaf, np))
196 this_leaf->type = CACHE_TYPE_UNIFIED;
197 cache_size(this_leaf, np);
198 cache_get_line_size(this_leaf, np);
199 cache_nr_sets(this_leaf, np);
200 cache_associativity(this_leaf);
201}
202
203static int cache_setup_of_node(unsigned int cpu)
204{
205 struct cacheinfo *this_leaf;
206 unsigned int index = 0;
207
208 struct device_node *np __free(device_node) = of_cpu_device_node_get(cpu);
209 if (!np) {
210 pr_err("Failed to find cpu%d device node\n", cpu);
211 return -ENOENT;
212 }
213
214 if (!of_check_cache_nodes(np)) {
215 return -ENOENT;
216 }
217
218 while (index < cache_leaves(cpu)) {
219 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
220 if (this_leaf->level != 1) {
221 struct device_node *prev __free(device_node) = np;
222 np = of_find_next_cache_node(np);
223 if (!np)
224 break;
225 }
226 cache_of_set_props(this_leaf, np);
227 this_leaf->fw_token = np;
228 index++;
229 }
230
231 if (index != cache_leaves(cpu)) /* not all OF nodes populated */
232 return -ENOENT;
233
234 return 0;
235}
236
237static bool of_check_cache_nodes(struct device_node *np)
238{
239 if (of_property_present(np, "cache-size") ||
240 of_property_present(np, "i-cache-size") ||
241 of_property_present(np, "d-cache-size") ||
242 of_property_present(np, "cache-unified"))
243 return true;
244
245 struct device_node *next __free(device_node) = of_find_next_cache_node(np);
246 if (next) {
247 return true;
248 }
249
250 return false;
251}
252
253static int of_count_cache_leaves(struct device_node *np)
254{
255 unsigned int leaves = 0;
256
257 if (of_property_present(np, "cache-size"))
258 ++leaves;
259 if (of_property_present(np, "i-cache-size"))
260 ++leaves;
261 if (of_property_present(np, "d-cache-size"))
262 ++leaves;
263
264 if (!leaves) {
265 /* The '[i-|d-|]cache-size' property is required, but
266 * if absent, fallback on the 'cache-unified' property.
267 */
268 if (of_property_read_bool(np, "cache-unified"))
269 return 1;
270 else
271 return 2;
272 }
273
274 return leaves;
275}
276
277int init_of_cache_level(unsigned int cpu)
278{
279 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
280 struct device_node *np __free(device_node) = of_cpu_device_node_get(cpu);
281 unsigned int levels = 0, leaves, level;
282
283 if (!of_check_cache_nodes(np)) {
284 return -ENOENT;
285 }
286
287 leaves = of_count_cache_leaves(np);
288 if (leaves > 0)
289 levels = 1;
290
291 while (1) {
292 struct device_node *prev __free(device_node) = np;
293 np = of_find_next_cache_node(np);
294 if (!np)
295 break;
296
297 if (!of_device_is_compatible(np, "cache"))
298 return -EINVAL;
299 if (of_property_read_u32(np, "cache-level", &level))
300 return -EINVAL;
301 if (level <= levels)
302 return -EINVAL;
303
304 leaves += of_count_cache_leaves(np);
305 levels = level;
306 }
307
308 this_cpu_ci->num_levels = levels;
309 this_cpu_ci->num_leaves = leaves;
310
311 return 0;
312}
313
314#else
315static inline int cache_setup_of_node(unsigned int cpu) { return 0; }
316int init_of_cache_level(unsigned int cpu) { return 0; }
317#endif
318
319int __weak cache_setup_acpi(unsigned int cpu)
320{
321 return -ENOTSUPP;
322}
323
324unsigned int coherency_max_size;
325
326static int cache_setup_properties(unsigned int cpu)
327{
328 int ret = 0;
329
330 if (of_have_populated_dt())
331 ret = cache_setup_of_node(cpu);
332 else if (!acpi_disabled)
333 ret = cache_setup_acpi(cpu);
334
335 // Assume there is no cache information available in DT/ACPI from now.
336 if (ret && use_arch_cache_info())
337 use_arch_info = true;
338
339 return ret;
340}
341
342static int cache_shared_cpu_map_setup(unsigned int cpu)
343{
344 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
345 struct cacheinfo *this_leaf, *sib_leaf;
346 unsigned int index, sib_index;
347 int ret = 0;
348
349 if (this_cpu_ci->cpu_map_populated)
350 return 0;
351
352 /*
353 * skip setting up cache properties if LLC is valid, just need
354 * to update the shared cpu_map if the cache attributes were
355 * populated early before all the cpus are brought online
356 */
357 if (!last_level_cache_is_valid(cpu) && !use_arch_info) {
358 ret = cache_setup_properties(cpu);
359 if (ret)
360 return ret;
361 }
362
363 for (index = 0; index < cache_leaves(cpu); index++) {
364 unsigned int i;
365
366 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
367
368 cpumask_set_cpu(cpu, &this_leaf->shared_cpu_map);
369 for_each_online_cpu(i) {
370 if (i == cpu || !per_cpu_cacheinfo(i))
371 continue;/* skip if itself or no cacheinfo */
372 for (sib_index = 0; sib_index < cache_leaves(i); sib_index++) {
373 sib_leaf = per_cpu_cacheinfo_idx(i, sib_index);
374
375 /*
376 * Comparing cache IDs only makes sense if the leaves
377 * belong to the same cache level of same type. Skip
378 * the check if level and type do not match.
379 */
380 if (sib_leaf->level != this_leaf->level ||
381 sib_leaf->type != this_leaf->type)
382 continue;
383
384 if (cache_leaves_are_shared(this_leaf, sib_leaf)) {
385 cpumask_set_cpu(cpu, &sib_leaf->shared_cpu_map);
386 cpumask_set_cpu(i, &this_leaf->shared_cpu_map);
387 break;
388 }
389 }
390 }
391 /* record the maximum cache line size */
392 if (this_leaf->coherency_line_size > coherency_max_size)
393 coherency_max_size = this_leaf->coherency_line_size;
394 }
395
396 /* shared_cpu_map is now populated for the cpu */
397 this_cpu_ci->cpu_map_populated = true;
398 return 0;
399}
400
401static void cache_shared_cpu_map_remove(unsigned int cpu)
402{
403 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
404 struct cacheinfo *this_leaf, *sib_leaf;
405 unsigned int sibling, index, sib_index;
406
407 for (index = 0; index < cache_leaves(cpu); index++) {
408 this_leaf = per_cpu_cacheinfo_idx(cpu, index);
409 for_each_cpu(sibling, &this_leaf->shared_cpu_map) {
410 if (sibling == cpu || !per_cpu_cacheinfo(sibling))
411 continue;/* skip if itself or no cacheinfo */
412
413 for (sib_index = 0; sib_index < cache_leaves(sibling); sib_index++) {
414 sib_leaf = per_cpu_cacheinfo_idx(sibling, sib_index);
415
416 /*
417 * Comparing cache IDs only makes sense if the leaves
418 * belong to the same cache level of same type. Skip
419 * the check if level and type do not match.
420 */
421 if (sib_leaf->level != this_leaf->level ||
422 sib_leaf->type != this_leaf->type)
423 continue;
424
425 if (cache_leaves_are_shared(this_leaf, sib_leaf)) {
426 cpumask_clear_cpu(cpu, &sib_leaf->shared_cpu_map);
427 cpumask_clear_cpu(sibling, &this_leaf->shared_cpu_map);
428 break;
429 }
430 }
431 }
432 }
433
434 /* cpu is no longer populated in the shared map */
435 this_cpu_ci->cpu_map_populated = false;
436}
437
438static void free_cache_attributes(unsigned int cpu)
439{
440 if (!per_cpu_cacheinfo(cpu))
441 return;
442
443 cache_shared_cpu_map_remove(cpu);
444}
445
446int __weak early_cache_level(unsigned int cpu)
447{
448 return -ENOENT;
449}
450
451int __weak init_cache_level(unsigned int cpu)
452{
453 return -ENOENT;
454}
455
456int __weak populate_cache_leaves(unsigned int cpu)
457{
458 return -ENOENT;
459}
460
461static inline int allocate_cache_info(int cpu)
462{
463 per_cpu_cacheinfo(cpu) = kcalloc(cache_leaves(cpu), sizeof(struct cacheinfo), GFP_ATOMIC);
464 if (!per_cpu_cacheinfo(cpu)) {
465 cache_leaves(cpu) = 0;
466 return -ENOMEM;
467 }
468
469 return 0;
470}
471
472int fetch_cache_info(unsigned int cpu)
473{
474 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
475 unsigned int levels = 0, split_levels = 0;
476 int ret;
477
478 if (acpi_disabled) {
479 ret = init_of_cache_level(cpu);
480 } else {
481 ret = acpi_get_cache_info(cpu, &levels, &split_levels);
482 if (!ret) {
483 this_cpu_ci->num_levels = levels;
484 /*
485 * This assumes that:
486 * - there cannot be any split caches (data/instruction)
487 * above a unified cache
488 * - data/instruction caches come by pair
489 */
490 this_cpu_ci->num_leaves = levels + split_levels;
491 }
492 }
493
494 if (ret || !cache_leaves(cpu)) {
495 ret = early_cache_level(cpu);
496 if (ret)
497 return ret;
498
499 if (!cache_leaves(cpu))
500 return -ENOENT;
501
502 this_cpu_ci->early_ci_levels = true;
503 }
504
505 return allocate_cache_info(cpu);
506}
507
508static inline int init_level_allocate_ci(unsigned int cpu)
509{
510 unsigned int early_leaves = cache_leaves(cpu);
511
512 /* Since early initialization/allocation of the cacheinfo is allowed
513 * via fetch_cache_info() and this also gets called as CPU hotplug
514 * callbacks via cacheinfo_cpu_online, the init/alloc can be skipped
515 * as it will happen only once (the cacheinfo memory is never freed).
516 * Just populate the cacheinfo. However, if the cacheinfo has been
517 * allocated early through the arch-specific early_cache_level() call,
518 * there is a chance the info is wrong (this can happen on arm64). In
519 * that case, call init_cache_level() anyway to give the arch-specific
520 * code a chance to make things right.
521 */
522 if (per_cpu_cacheinfo(cpu) && !ci_cacheinfo(cpu)->early_ci_levels)
523 return 0;
524
525 if (init_cache_level(cpu) || !cache_leaves(cpu))
526 return -ENOENT;
527
528 /*
529 * Now that we have properly initialized the cache level info, make
530 * sure we don't try to do that again the next time we are called
531 * (e.g. as CPU hotplug callbacks).
532 */
533 ci_cacheinfo(cpu)->early_ci_levels = false;
534
535 /*
536 * Some architectures (e.g., x86) do not use early initialization.
537 * Allocate memory now in such case.
538 */
539 if (cache_leaves(cpu) <= early_leaves && per_cpu_cacheinfo(cpu))
540 return 0;
541
542 kfree(per_cpu_cacheinfo(cpu));
543 return allocate_cache_info(cpu);
544}
545
546int detect_cache_attributes(unsigned int cpu)
547{
548 int ret;
549
550 ret = init_level_allocate_ci(cpu);
551 if (ret)
552 return ret;
553
554 /*
555 * If LLC is valid the cache leaves were already populated so just go to
556 * update the cpu map.
557 */
558 if (!last_level_cache_is_valid(cpu)) {
559 /*
560 * populate_cache_leaves() may completely setup the cache leaves and
561 * shared_cpu_map or it may leave it partially setup.
562 */
563 ret = populate_cache_leaves(cpu);
564 if (ret)
565 goto free_ci;
566 }
567
568 /*
569 * For systems using DT for cache hierarchy, fw_token
570 * and shared_cpu_map will be set up here only if they are
571 * not populated already
572 */
573 ret = cache_shared_cpu_map_setup(cpu);
574 if (ret) {
575 pr_warn("Unable to detect cache hierarchy for CPU %d\n", cpu);
576 goto free_ci;
577 }
578
579 return 0;
580
581free_ci:
582 free_cache_attributes(cpu);
583 return ret;
584}
585
586/* pointer to cpuX/cache device */
587static DEFINE_PER_CPU(struct device *, ci_cache_dev);
588#define per_cpu_cache_dev(cpu) (per_cpu(ci_cache_dev, cpu))
589
590static cpumask_t cache_dev_map;
591
592/* pointer to array of devices for cpuX/cache/indexY */
593static DEFINE_PER_CPU(struct device **, ci_index_dev);
594#define per_cpu_index_dev(cpu) (per_cpu(ci_index_dev, cpu))
595#define per_cache_index_dev(cpu, idx) ((per_cpu_index_dev(cpu))[idx])
596
597#define show_one(file_name, object) \
598static ssize_t file_name##_show(struct device *dev, \
599 struct device_attribute *attr, char *buf) \
600{ \
601 struct cacheinfo *this_leaf = dev_get_drvdata(dev); \
602 return sysfs_emit(buf, "%u\n", this_leaf->object); \
603}
604
605show_one(id, id);
606show_one(level, level);
607show_one(coherency_line_size, coherency_line_size);
608show_one(number_of_sets, number_of_sets);
609show_one(physical_line_partition, physical_line_partition);
610show_one(ways_of_associativity, ways_of_associativity);
611
612static ssize_t size_show(struct device *dev,
613 struct device_attribute *attr, char *buf)
614{
615 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
616
617 return sysfs_emit(buf, "%uK\n", this_leaf->size >> 10);
618}
619
620static ssize_t shared_cpu_map_show(struct device *dev,
621 struct device_attribute *attr, char *buf)
622{
623 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
624 const struct cpumask *mask = &this_leaf->shared_cpu_map;
625
626 return sysfs_emit(buf, "%*pb\n", nr_cpu_ids, mask);
627}
628
629static ssize_t shared_cpu_list_show(struct device *dev,
630 struct device_attribute *attr, char *buf)
631{
632 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
633 const struct cpumask *mask = &this_leaf->shared_cpu_map;
634
635 return sysfs_emit(buf, "%*pbl\n", nr_cpu_ids, mask);
636}
637
638static ssize_t type_show(struct device *dev,
639 struct device_attribute *attr, char *buf)
640{
641 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
642 const char *output;
643
644 switch (this_leaf->type) {
645 case CACHE_TYPE_DATA:
646 output = "Data";
647 break;
648 case CACHE_TYPE_INST:
649 output = "Instruction";
650 break;
651 case CACHE_TYPE_UNIFIED:
652 output = "Unified";
653 break;
654 default:
655 return -EINVAL;
656 }
657
658 return sysfs_emit(buf, "%s\n", output);
659}
660
661static ssize_t allocation_policy_show(struct device *dev,
662 struct device_attribute *attr, char *buf)
663{
664 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
665 unsigned int ci_attr = this_leaf->attributes;
666 const char *output;
667
668 if ((ci_attr & CACHE_READ_ALLOCATE) && (ci_attr & CACHE_WRITE_ALLOCATE))
669 output = "ReadWriteAllocate";
670 else if (ci_attr & CACHE_READ_ALLOCATE)
671 output = "ReadAllocate";
672 else if (ci_attr & CACHE_WRITE_ALLOCATE)
673 output = "WriteAllocate";
674 else
675 return 0;
676
677 return sysfs_emit(buf, "%s\n", output);
678}
679
680static ssize_t write_policy_show(struct device *dev,
681 struct device_attribute *attr, char *buf)
682{
683 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
684 unsigned int ci_attr = this_leaf->attributes;
685 int n = 0;
686
687 if (ci_attr & CACHE_WRITE_THROUGH)
688 n = sysfs_emit(buf, "WriteThrough\n");
689 else if (ci_attr & CACHE_WRITE_BACK)
690 n = sysfs_emit(buf, "WriteBack\n");
691 return n;
692}
693
694static DEVICE_ATTR_RO(id);
695static DEVICE_ATTR_RO(level);
696static DEVICE_ATTR_RO(type);
697static DEVICE_ATTR_RO(coherency_line_size);
698static DEVICE_ATTR_RO(ways_of_associativity);
699static DEVICE_ATTR_RO(number_of_sets);
700static DEVICE_ATTR_RO(size);
701static DEVICE_ATTR_RO(allocation_policy);
702static DEVICE_ATTR_RO(write_policy);
703static DEVICE_ATTR_RO(shared_cpu_map);
704static DEVICE_ATTR_RO(shared_cpu_list);
705static DEVICE_ATTR_RO(physical_line_partition);
706
707static struct attribute *cache_default_attrs[] = {
708 &dev_attr_id.attr,
709 &dev_attr_type.attr,
710 &dev_attr_level.attr,
711 &dev_attr_shared_cpu_map.attr,
712 &dev_attr_shared_cpu_list.attr,
713 &dev_attr_coherency_line_size.attr,
714 &dev_attr_ways_of_associativity.attr,
715 &dev_attr_number_of_sets.attr,
716 &dev_attr_size.attr,
717 &dev_attr_allocation_policy.attr,
718 &dev_attr_write_policy.attr,
719 &dev_attr_physical_line_partition.attr,
720 NULL
721};
722
723static umode_t
724cache_default_attrs_is_visible(struct kobject *kobj,
725 struct attribute *attr, int unused)
726{
727 struct device *dev = kobj_to_dev(kobj);
728 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
729 const struct cpumask *mask = &this_leaf->shared_cpu_map;
730 umode_t mode = attr->mode;
731
732 if ((attr == &dev_attr_id.attr) && (this_leaf->attributes & CACHE_ID))
733 return mode;
734 if ((attr == &dev_attr_type.attr) && this_leaf->type)
735 return mode;
736 if ((attr == &dev_attr_level.attr) && this_leaf->level)
737 return mode;
738 if ((attr == &dev_attr_shared_cpu_map.attr) && !cpumask_empty(mask))
739 return mode;
740 if ((attr == &dev_attr_shared_cpu_list.attr) && !cpumask_empty(mask))
741 return mode;
742 if ((attr == &dev_attr_coherency_line_size.attr) &&
743 this_leaf->coherency_line_size)
744 return mode;
745 if ((attr == &dev_attr_ways_of_associativity.attr) &&
746 this_leaf->size) /* allow 0 = full associativity */
747 return mode;
748 if ((attr == &dev_attr_number_of_sets.attr) &&
749 this_leaf->number_of_sets)
750 return mode;
751 if ((attr == &dev_attr_size.attr) && this_leaf->size)
752 return mode;
753 if ((attr == &dev_attr_write_policy.attr) &&
754 (this_leaf->attributes & CACHE_WRITE_POLICY_MASK))
755 return mode;
756 if ((attr == &dev_attr_allocation_policy.attr) &&
757 (this_leaf->attributes & CACHE_ALLOCATE_POLICY_MASK))
758 return mode;
759 if ((attr == &dev_attr_physical_line_partition.attr) &&
760 this_leaf->physical_line_partition)
761 return mode;
762
763 return 0;
764}
765
766static const struct attribute_group cache_default_group = {
767 .attrs = cache_default_attrs,
768 .is_visible = cache_default_attrs_is_visible,
769};
770
771static const struct attribute_group *cache_default_groups[] = {
772 &cache_default_group,
773 NULL,
774};
775
776static const struct attribute_group *cache_private_groups[] = {
777 &cache_default_group,
778 NULL, /* Place holder for private group */
779 NULL,
780};
781
782const struct attribute_group *
783__weak cache_get_priv_group(struct cacheinfo *this_leaf)
784{
785 return NULL;
786}
787
788static const struct attribute_group **
789cache_get_attribute_groups(struct cacheinfo *this_leaf)
790{
791 const struct attribute_group *priv_group =
792 cache_get_priv_group(this_leaf);
793
794 if (!priv_group)
795 return cache_default_groups;
796
797 if (!cache_private_groups[1])
798 cache_private_groups[1] = priv_group;
799
800 return cache_private_groups;
801}
802
803/* Add/Remove cache interface for CPU device */
804static void cpu_cache_sysfs_exit(unsigned int cpu)
805{
806 int i;
807 struct device *ci_dev;
808
809 if (per_cpu_index_dev(cpu)) {
810 for (i = 0; i < cache_leaves(cpu); i++) {
811 ci_dev = per_cache_index_dev(cpu, i);
812 if (!ci_dev)
813 continue;
814 device_unregister(ci_dev);
815 }
816 kfree(per_cpu_index_dev(cpu));
817 per_cpu_index_dev(cpu) = NULL;
818 }
819 device_unregister(per_cpu_cache_dev(cpu));
820 per_cpu_cache_dev(cpu) = NULL;
821}
822
823static int cpu_cache_sysfs_init(unsigned int cpu)
824{
825 struct device *dev = get_cpu_device(cpu);
826
827 if (per_cpu_cacheinfo(cpu) == NULL)
828 return -ENOENT;
829
830 per_cpu_cache_dev(cpu) = cpu_device_create(dev, NULL, NULL, "cache");
831 if (IS_ERR(per_cpu_cache_dev(cpu)))
832 return PTR_ERR(per_cpu_cache_dev(cpu));
833
834 /* Allocate all required memory */
835 per_cpu_index_dev(cpu) = kcalloc(cache_leaves(cpu),
836 sizeof(struct device *), GFP_KERNEL);
837 if (unlikely(per_cpu_index_dev(cpu) == NULL))
838 goto err_out;
839
840 return 0;
841
842err_out:
843 cpu_cache_sysfs_exit(cpu);
844 return -ENOMEM;
845}
846
847static int cache_add_dev(unsigned int cpu)
848{
849 unsigned int i;
850 int rc;
851 struct device *ci_dev, *parent;
852 struct cacheinfo *this_leaf;
853 const struct attribute_group **cache_groups;
854
855 rc = cpu_cache_sysfs_init(cpu);
856 if (unlikely(rc < 0))
857 return rc;
858
859 parent = per_cpu_cache_dev(cpu);
860 for (i = 0; i < cache_leaves(cpu); i++) {
861 this_leaf = per_cpu_cacheinfo_idx(cpu, i);
862 if (this_leaf->disable_sysfs)
863 continue;
864 if (this_leaf->type == CACHE_TYPE_NOCACHE)
865 break;
866 cache_groups = cache_get_attribute_groups(this_leaf);
867 ci_dev = cpu_device_create(parent, this_leaf, cache_groups,
868 "index%1u", i);
869 if (IS_ERR(ci_dev)) {
870 rc = PTR_ERR(ci_dev);
871 goto err;
872 }
873 per_cache_index_dev(cpu, i) = ci_dev;
874 }
875 cpumask_set_cpu(cpu, &cache_dev_map);
876
877 return 0;
878err:
879 cpu_cache_sysfs_exit(cpu);
880 return rc;
881}
882
883static unsigned int cpu_map_shared_cache(bool online, unsigned int cpu,
884 cpumask_t **map)
885{
886 struct cacheinfo *llc, *sib_llc;
887 unsigned int sibling;
888
889 if (!last_level_cache_is_valid(cpu))
890 return 0;
891
892 llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
893
894 if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
895 return 0;
896
897 if (online) {
898 *map = &llc->shared_cpu_map;
899 return cpumask_weight(*map);
900 }
901
902 /* shared_cpu_map of offlined CPU will be cleared, so use sibling map */
903 for_each_cpu(sibling, &llc->shared_cpu_map) {
904 if (sibling == cpu || !last_level_cache_is_valid(sibling))
905 continue;
906 sib_llc = per_cpu_cacheinfo_idx(sibling, cache_leaves(sibling) - 1);
907 *map = &sib_llc->shared_cpu_map;
908 return cpumask_weight(*map);
909 }
910
911 return 0;
912}
913
914/*
915 * Calculate the size of the per-CPU data cache slice. This can be
916 * used to estimate the size of the data cache slice that can be used
917 * by one CPU under ideal circumstances. UNIFIED caches are counted
918 * in addition to DATA caches. So, please consider code cache usage
919 * when use the result.
920 *
921 * Because the cache inclusive/non-inclusive information isn't
922 * available, we just use the size of the per-CPU slice of LLC to make
923 * the result more predictable across architectures.
924 */
925static void update_per_cpu_data_slice_size_cpu(unsigned int cpu)
926{
927 struct cpu_cacheinfo *ci;
928 struct cacheinfo *llc;
929 unsigned int nr_shared;
930
931 if (!last_level_cache_is_valid(cpu))
932 return;
933
934 ci = ci_cacheinfo(cpu);
935 llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
936
937 if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
938 return;
939
940 nr_shared = cpumask_weight(&llc->shared_cpu_map);
941 if (nr_shared)
942 ci->per_cpu_data_slice_size = llc->size / nr_shared;
943}
944
945static void update_per_cpu_data_slice_size(bool cpu_online, unsigned int cpu,
946 cpumask_t *cpu_map)
947{
948 unsigned int icpu;
949
950 for_each_cpu(icpu, cpu_map) {
951 if (!cpu_online && icpu == cpu)
952 continue;
953 update_per_cpu_data_slice_size_cpu(icpu);
954 setup_pcp_cacheinfo(icpu);
955 }
956}
957
958static int cacheinfo_cpu_online(unsigned int cpu)
959{
960 int rc = detect_cache_attributes(cpu);
961 cpumask_t *cpu_map;
962
963 if (rc)
964 return rc;
965 rc = cache_add_dev(cpu);
966 if (rc)
967 goto err;
968 if (cpu_map_shared_cache(true, cpu, &cpu_map))
969 update_per_cpu_data_slice_size(true, cpu, cpu_map);
970 return 0;
971err:
972 free_cache_attributes(cpu);
973 return rc;
974}
975
976static int cacheinfo_cpu_pre_down(unsigned int cpu)
977{
978 cpumask_t *cpu_map;
979 unsigned int nr_shared;
980
981 nr_shared = cpu_map_shared_cache(false, cpu, &cpu_map);
982 if (cpumask_test_and_clear_cpu(cpu, &cache_dev_map))
983 cpu_cache_sysfs_exit(cpu);
984
985 free_cache_attributes(cpu);
986 if (nr_shared > 1)
987 update_per_cpu_data_slice_size(false, cpu, cpu_map);
988 return 0;
989}
990
991static int __init cacheinfo_sysfs_init(void)
992{
993 return cpuhp_setup_state(CPUHP_AP_BASE_CACHEINFO_ONLINE,
994 "base/cacheinfo:online",
995 cacheinfo_cpu_online, cacheinfo_cpu_pre_down);
996}
997device_initcall(cacheinfo_sysfs_init);