Linux Audio

Check our new training course

Loading...
v6.8
  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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
901/*
902 * Calculate the size of the per-CPU data cache slice.  This can be
903 * used to estimate the size of the data cache slice that can be used
904 * by one CPU under ideal circumstances.  UNIFIED caches are counted
905 * in addition to DATA caches.  So, please consider code cache usage
906 * when use the result.
907 *
908 * Because the cache inclusive/non-inclusive information isn't
909 * available, we just use the size of the per-CPU slice of LLC to make
910 * the result more predictable across architectures.
911 */
912static void update_per_cpu_data_slice_size_cpu(unsigned int cpu)
913{
914	struct cpu_cacheinfo *ci;
915	struct cacheinfo *llc;
916	unsigned int nr_shared;
917
918	if (!last_level_cache_is_valid(cpu))
919		return;
920
921	ci = ci_cacheinfo(cpu);
922	llc = per_cpu_cacheinfo_idx(cpu, cache_leaves(cpu) - 1);
923
924	if (llc->type != CACHE_TYPE_DATA && llc->type != CACHE_TYPE_UNIFIED)
925		return;
926
927	nr_shared = cpumask_weight(&llc->shared_cpu_map);
928	if (nr_shared)
929		ci->per_cpu_data_slice_size = llc->size / nr_shared;
930}
931
932static void update_per_cpu_data_slice_size(bool cpu_online, unsigned int cpu)
 
933{
934	unsigned int icpu;
935
936	for_each_online_cpu(icpu) {
937		if (!cpu_online && icpu == cpu)
938			continue;
939		update_per_cpu_data_slice_size_cpu(icpu);
 
940	}
941}
942
943static int cacheinfo_cpu_online(unsigned int cpu)
944{
945	int rc = detect_cache_attributes(cpu);
 
946
947	if (rc)
948		return rc;
949	rc = cache_add_dev(cpu);
950	if (rc)
951		goto err;
952	update_per_cpu_data_slice_size(true, cpu);
953	setup_pcp_cacheinfo();
954	return 0;
955err:
956	free_cache_attributes(cpu);
957	return rc;
958}
959
960static int cacheinfo_cpu_pre_down(unsigned int cpu)
961{
 
 
 
 
962	if (cpumask_test_and_clear_cpu(cpu, &cache_dev_map))
963		cpu_cache_sysfs_exit(cpu);
964
965	free_cache_attributes(cpu);
966	update_per_cpu_data_slice_size(false, cpu);
967	setup_pcp_cacheinfo();
968	return 0;
969}
970
971static int __init cacheinfo_sysfs_init(void)
972{
973	return cpuhp_setup_state(CPUHP_AP_BASE_CACHEINFO_ONLINE,
974				 "base/cacheinfo:online",
975				 cacheinfo_cpu_online, cacheinfo_cpu_pre_down);
976}
977device_initcall(cacheinfo_sysfs_init);
v6.9.4
   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);