2 * Routines to indentify caches on Intel CPU.
5 * Venkatesh Pallipadi : Adding cache identification through cpuid(4)
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <linux/device.h>
11 #include <linux/compiler.h>
12 #include <linux/cpu.h>
14 #include <asm/processor.h>
25 unsigned char descriptor;
30 /* all the cache descriptor types we care about (no TLB or trace cache entries) */
31 static struct _cache_table cache_table[] __devinitdata =
33 { 0x06, LVL_1_INST, 8 }, /* 4-way set assoc, 32 byte line size */
34 { 0x08, LVL_1_INST, 16 }, /* 4-way set assoc, 32 byte line size */
35 { 0x0a, LVL_1_DATA, 8 }, /* 2 way set assoc, 32 byte line size */
36 { 0x0c, LVL_1_DATA, 16 }, /* 4-way set assoc, 32 byte line size */
37 { 0x22, LVL_3, 512 }, /* 4-way set assoc, sectored cache, 64 byte line size */
38 { 0x23, LVL_3, 1024 }, /* 8-way set assoc, sectored cache, 64 byte line size */
39 { 0x25, LVL_3, 2048 }, /* 8-way set assoc, sectored cache, 64 byte line size */
40 { 0x29, LVL_3, 4096 }, /* 8-way set assoc, sectored cache, 64 byte line size */
41 { 0x2c, LVL_1_DATA, 32 }, /* 8-way set assoc, 64 byte line size */
42 { 0x30, LVL_1_INST, 32 }, /* 8-way set assoc, 64 byte line size */
43 { 0x39, LVL_2, 128 }, /* 4-way set assoc, sectored cache, 64 byte line size */
44 { 0x3b, LVL_2, 128 }, /* 2-way set assoc, sectored cache, 64 byte line size */
45 { 0x3c, LVL_2, 256 }, /* 4-way set assoc, sectored cache, 64 byte line size */
46 { 0x41, LVL_2, 128 }, /* 4-way set assoc, 32 byte line size */
47 { 0x42, LVL_2, 256 }, /* 4-way set assoc, 32 byte line size */
48 { 0x43, LVL_2, 512 }, /* 4-way set assoc, 32 byte line size */
49 { 0x44, LVL_2, 1024 }, /* 4-way set assoc, 32 byte line size */
50 { 0x45, LVL_2, 2048 }, /* 4-way set assoc, 32 byte line size */
51 { 0x60, LVL_1_DATA, 16 }, /* 8-way set assoc, sectored cache, 64 byte line size */
52 { 0x66, LVL_1_DATA, 8 }, /* 4-way set assoc, sectored cache, 64 byte line size */
53 { 0x67, LVL_1_DATA, 16 }, /* 4-way set assoc, sectored cache, 64 byte line size */
54 { 0x68, LVL_1_DATA, 32 }, /* 4-way set assoc, sectored cache, 64 byte line size */
55 { 0x70, LVL_TRACE, 12 }, /* 8-way set assoc */
56 { 0x71, LVL_TRACE, 16 }, /* 8-way set assoc */
57 { 0x72, LVL_TRACE, 32 }, /* 8-way set assoc */
58 { 0x78, LVL_2, 1024 }, /* 4-way set assoc, 64 byte line size */
59 { 0x79, LVL_2, 128 }, /* 8-way set assoc, sectored cache, 64 byte line size */
60 { 0x7a, LVL_2, 256 }, /* 8-way set assoc, sectored cache, 64 byte line size */
61 { 0x7b, LVL_2, 512 }, /* 8-way set assoc, sectored cache, 64 byte line size */
62 { 0x7c, LVL_2, 1024 }, /* 8-way set assoc, sectored cache, 64 byte line size */
63 { 0x7d, LVL_2, 2048 }, /* 8-way set assoc, 64 byte line size */
64 { 0x7f, LVL_2, 512 }, /* 2-way set assoc, 64 byte line size */
65 { 0x82, LVL_2, 256 }, /* 8-way set assoc, 32 byte line size */
66 { 0x83, LVL_2, 512 }, /* 8-way set assoc, 32 byte line size */
67 { 0x84, LVL_2, 1024 }, /* 8-way set assoc, 32 byte line size */
68 { 0x85, LVL_2, 2048 }, /* 8-way set assoc, 32 byte line size */
69 { 0x86, LVL_2, 512 }, /* 4-way set assoc, 64 byte line size */
70 { 0x87, LVL_2, 1024 }, /* 8-way set assoc, 64 byte line size */
80 CACHE_TYPE_UNIFIED = 3
83 union _cpuid4_leaf_eax {
85 enum _cache_type type:5;
87 unsigned int is_self_initializing:1;
88 unsigned int is_fully_associative:1;
89 unsigned int reserved:4;
90 unsigned int num_threads_sharing:12;
91 unsigned int num_cores_on_die:6;
96 union _cpuid4_leaf_ebx {
98 unsigned int coherency_line_size:12;
99 unsigned int physical_line_partition:10;
100 unsigned int ways_of_associativity:10;
105 union _cpuid4_leaf_ecx {
107 unsigned int number_of_sets:32;
112 struct _cpuid4_info {
113 union _cpuid4_leaf_eax eax;
114 union _cpuid4_leaf_ebx ebx;
115 union _cpuid4_leaf_ecx ecx;
117 cpumask_t shared_cpu_map;
120 static unsigned short num_cache_leaves;
122 static int __devinit cpuid4_cache_lookup(int index, struct _cpuid4_info *this_leaf)
124 unsigned int eax, ebx, ecx, edx;
125 union _cpuid4_leaf_eax cache_eax;
127 cpuid_count(4, index, &eax, &ebx, &ecx, &edx);
128 cache_eax.full = eax;
129 if (cache_eax.split.type == CACHE_TYPE_NULL)
130 return -EIO; /* better error ? */
132 this_leaf->eax.full = eax;
133 this_leaf->ebx.full = ebx;
134 this_leaf->ecx.full = ecx;
135 this_leaf->size = (this_leaf->ecx.split.number_of_sets + 1) *
136 (this_leaf->ebx.split.coherency_line_size + 1) *
137 (this_leaf->ebx.split.physical_line_partition + 1) *
138 (this_leaf->ebx.split.ways_of_associativity + 1);
142 static int __init find_num_cache_leaves(void)
144 unsigned int eax, ebx, ecx, edx;
145 union _cpuid4_leaf_eax cache_eax;
150 /* Do cpuid(4) loop to find out num_cache_leaves */
151 cpuid_count(4, i, &eax, &ebx, &ecx, &edx);
152 cache_eax.full = eax;
153 } while (cache_eax.split.type != CACHE_TYPE_NULL);
157 unsigned int __devinit init_intel_cacheinfo(struct cpuinfo_x86 *c)
159 unsigned int trace = 0, l1i = 0, l1d = 0, l2 = 0, l3 = 0; /* Cache sizes */
160 unsigned int new_l1d = 0, new_l1i = 0; /* Cache sizes from cpuid(4) */
161 unsigned int new_l2 = 0, new_l3 = 0, i; /* Cache sizes from cpuid(4) */
163 if (c->cpuid_level > 4) {
164 static int is_initialized;
166 if (is_initialized == 0) {
167 /* Init num_cache_leaves from boot CPU */
168 num_cache_leaves = find_num_cache_leaves();
173 * Whenever possible use cpuid(4), deterministic cache
174 * parameters cpuid leaf to find the cache details
176 for (i = 0; i < num_cache_leaves; i++) {
177 struct _cpuid4_info this_leaf;
181 retval = cpuid4_cache_lookup(i, &this_leaf);
183 switch(this_leaf.eax.split.level) {
185 if (this_leaf.eax.split.type ==
187 new_l1d = this_leaf.size/1024;
188 else if (this_leaf.eax.split.type ==
190 new_l1i = this_leaf.size/1024;
193 new_l2 = this_leaf.size/1024;
196 new_l3 = this_leaf.size/1024;
204 if (c->cpuid_level > 1) {
205 /* supports eax=2 call */
208 unsigned char *dp = (unsigned char *)regs;
210 /* Number of times to iterate */
211 n = cpuid_eax(2) & 0xFF;
213 for ( i = 0 ; i < n ; i++ ) {
214 cpuid(2, ®s[0], ®s[1], ®s[2], ®s[3]);
216 /* If bit 31 is set, this is an unknown format */
217 for ( j = 0 ; j < 3 ; j++ ) {
218 if ( regs[j] < 0 ) regs[j] = 0;
221 /* Byte 0 is level count, not a descriptor */
222 for ( j = 1 ; j < 16 ; j++ ) {
223 unsigned char des = dp[j];
226 /* look up this descriptor in the table */
227 while (cache_table[k].descriptor != 0)
229 if (cache_table[k].descriptor == des) {
230 switch (cache_table[k].cache_type) {
232 l1i += cache_table[k].size;
235 l1d += cache_table[k].size;
238 l2 += cache_table[k].size;
241 l3 += cache_table[k].size;
244 trace += cache_table[k].size;
269 printk (KERN_INFO "CPU: Trace cache: %dK uops", trace);
271 printk (KERN_INFO "CPU: L1 I cache: %dK", l1i);
273 printk(", L1 D cache: %dK\n", l1d);
277 printk(KERN_INFO "CPU: L2 cache: %dK\n", l2);
279 printk(KERN_INFO "CPU: L3 cache: %dK\n", l3);
282 * This assumes the L3 cache is shared; it typically lives in
283 * the northbridge. The L1 caches are included by the L2
284 * cache, and so should not be included for the purpose of
285 * SMP switching weights.
287 c->x86_cache_size = l2 ? l2 : (l1i+l1d);
293 /* pointer to _cpuid4_info array (for each cache leaf) */
294 static struct _cpuid4_info *cpuid4_info[NR_CPUS];
295 #define CPUID4_INFO_IDX(x,y) (&((cpuid4_info[x])[y]))
298 static void __devinit cache_shared_cpu_map_setup(unsigned int cpu, int index)
300 struct _cpuid4_info *this_leaf;
301 unsigned long num_threads_sharing;
303 struct cpuinfo_x86 *c = cpu_data + cpu;
306 this_leaf = CPUID4_INFO_IDX(cpu, index);
307 num_threads_sharing = 1 + this_leaf->eax.split.num_threads_sharing;
309 if (num_threads_sharing == 1)
310 cpu_set(cpu, this_leaf->shared_cpu_map);
312 else if (num_threads_sharing == smp_num_siblings)
313 this_leaf->shared_cpu_map = cpu_sibling_map[cpu];
314 else if (num_threads_sharing == (c->x86_num_cores * smp_num_siblings))
315 this_leaf->shared_cpu_map = cpu_core_map[cpu];
317 printk(KERN_DEBUG "Number of CPUs sharing cache didn't match "
318 "any known set of CPUs\n");
322 static void __init cache_shared_cpu_map_setup(unsigned int cpu, int index) {}
325 static void free_cache_attributes(unsigned int cpu)
327 kfree(cpuid4_info[cpu]);
328 cpuid4_info[cpu] = NULL;
331 static int __devinit detect_cache_attributes(unsigned int cpu)
333 struct _cpuid4_info *this_leaf;
338 if (num_cache_leaves == 0)
341 cpuid4_info[cpu] = kmalloc(
342 sizeof(struct _cpuid4_info) * num_cache_leaves, GFP_KERNEL);
343 if (unlikely(cpuid4_info[cpu] == NULL))
345 memset(cpuid4_info[cpu], 0,
346 sizeof(struct _cpuid4_info) * num_cache_leaves);
348 oldmask = current->cpus_allowed;
349 retval = set_cpus_allowed(current, cpumask_of_cpu(cpu));
353 /* Do cpuid and store the results */
355 for (j = 0; j < num_cache_leaves; j++) {
356 this_leaf = CPUID4_INFO_IDX(cpu, j);
357 retval = cpuid4_cache_lookup(j, this_leaf);
358 if (unlikely(retval < 0))
360 cache_shared_cpu_map_setup(cpu, j);
362 set_cpus_allowed(current, oldmask);
366 free_cache_attributes(cpu);
372 #include <linux/kobject.h>
373 #include <linux/sysfs.h>
375 extern struct sysdev_class cpu_sysdev_class; /* from drivers/base/cpu.c */
377 /* pointer to kobject for cpuX/cache */
378 static struct kobject * cache_kobject[NR_CPUS];
380 struct _index_kobject {
383 unsigned short index;
386 /* pointer to array of kobjects for cpuX/cache/indexY */
387 static struct _index_kobject *index_kobject[NR_CPUS];
388 #define INDEX_KOBJECT_PTR(x,y) (&((index_kobject[x])[y]))
390 #define show_one_plus(file_name, object, val) \
391 static ssize_t show_##file_name \
392 (struct _cpuid4_info *this_leaf, char *buf) \
394 return sprintf (buf, "%lu\n", (unsigned long)this_leaf->object + val); \
397 show_one_plus(level, eax.split.level, 0);
398 show_one_plus(coherency_line_size, ebx.split.coherency_line_size, 1);
399 show_one_plus(physical_line_partition, ebx.split.physical_line_partition, 1);
400 show_one_plus(ways_of_associativity, ebx.split.ways_of_associativity, 1);
401 show_one_plus(number_of_sets, ecx.split.number_of_sets, 1);
403 static ssize_t show_size(struct _cpuid4_info *this_leaf, char *buf)
405 return sprintf (buf, "%luK\n", this_leaf->size / 1024);
408 static ssize_t show_shared_cpu_map(struct _cpuid4_info *this_leaf, char *buf)
410 char mask_str[NR_CPUS];
411 cpumask_scnprintf(mask_str, NR_CPUS, this_leaf->shared_cpu_map);
412 return sprintf(buf, "%s\n", mask_str);
415 static ssize_t show_type(struct _cpuid4_info *this_leaf, char *buf) {
416 switch(this_leaf->eax.split.type) {
417 case CACHE_TYPE_DATA:
418 return sprintf(buf, "Data\n");
420 case CACHE_TYPE_INST:
421 return sprintf(buf, "Instruction\n");
423 case CACHE_TYPE_UNIFIED:
424 return sprintf(buf, "Unified\n");
427 return sprintf(buf, "Unknown\n");
433 struct attribute attr;
434 ssize_t (*show)(struct _cpuid4_info *, char *);
435 ssize_t (*store)(struct _cpuid4_info *, const char *, size_t count);
438 #define define_one_ro(_name) \
439 static struct _cache_attr _name = \
440 __ATTR(_name, 0444, show_##_name, NULL)
442 define_one_ro(level);
444 define_one_ro(coherency_line_size);
445 define_one_ro(physical_line_partition);
446 define_one_ro(ways_of_associativity);
447 define_one_ro(number_of_sets);
449 define_one_ro(shared_cpu_map);
451 static struct attribute * default_attrs[] = {
454 &coherency_line_size.attr,
455 &physical_line_partition.attr,
456 &ways_of_associativity.attr,
457 &number_of_sets.attr,
459 &shared_cpu_map.attr,
463 #define to_object(k) container_of(k, struct _index_kobject, kobj)
464 #define to_attr(a) container_of(a, struct _cache_attr, attr)
466 static ssize_t show(struct kobject * kobj, struct attribute * attr, char * buf)
468 struct _cache_attr *fattr = to_attr(attr);
469 struct _index_kobject *this_leaf = to_object(kobj);
473 fattr->show(CPUID4_INFO_IDX(this_leaf->cpu, this_leaf->index),
479 static ssize_t store(struct kobject * kobj, struct attribute * attr,
480 const char * buf, size_t count)
485 static struct sysfs_ops sysfs_ops = {
490 static struct kobj_type ktype_cache = {
491 .sysfs_ops = &sysfs_ops,
492 .default_attrs = default_attrs,
495 static struct kobj_type ktype_percpu_entry = {
496 .sysfs_ops = &sysfs_ops,
499 static void cpuid4_cache_sysfs_exit(unsigned int cpu)
501 kfree(cache_kobject[cpu]);
502 kfree(index_kobject[cpu]);
503 cache_kobject[cpu] = NULL;
504 index_kobject[cpu] = NULL;
505 free_cache_attributes(cpu);
508 static int __devinit cpuid4_cache_sysfs_init(unsigned int cpu)
511 if (num_cache_leaves == 0)
514 detect_cache_attributes(cpu);
515 if (cpuid4_info[cpu] == NULL)
518 /* Allocate all required memory */
519 cache_kobject[cpu] = kmalloc(sizeof(struct kobject), GFP_KERNEL);
520 if (unlikely(cache_kobject[cpu] == NULL))
522 memset(cache_kobject[cpu], 0, sizeof(struct kobject));
524 index_kobject[cpu] = kmalloc(
525 sizeof(struct _index_kobject ) * num_cache_leaves, GFP_KERNEL);
526 if (unlikely(index_kobject[cpu] == NULL))
528 memset(index_kobject[cpu], 0,
529 sizeof(struct _index_kobject) * num_cache_leaves);
534 cpuid4_cache_sysfs_exit(cpu);
538 /* Add/Remove cache interface for CPU device */
539 static int __devinit cache_add_dev(struct sys_device * sys_dev)
541 unsigned int cpu = sys_dev->id;
543 struct _index_kobject *this_object;
546 retval = cpuid4_cache_sysfs_init(cpu);
547 if (unlikely(retval < 0))
550 cache_kobject[cpu]->parent = &sys_dev->kobj;
551 kobject_set_name(cache_kobject[cpu], "%s", "cache");
552 cache_kobject[cpu]->ktype = &ktype_percpu_entry;
553 retval = kobject_register(cache_kobject[cpu]);
555 for (i = 0; i < num_cache_leaves; i++) {
556 this_object = INDEX_KOBJECT_PTR(cpu,i);
557 this_object->cpu = cpu;
558 this_object->index = i;
559 this_object->kobj.parent = cache_kobject[cpu];
560 kobject_set_name(&(this_object->kobj), "index%1lu", i);
561 this_object->kobj.ktype = &ktype_cache;
562 retval = kobject_register(&(this_object->kobj));
563 if (unlikely(retval)) {
564 for (j = 0; j < i; j++) {
566 &(INDEX_KOBJECT_PTR(cpu,j)->kobj));
568 kobject_unregister(cache_kobject[cpu]);
569 cpuid4_cache_sysfs_exit(cpu);
576 static int __devexit cache_remove_dev(struct sys_device * sys_dev)
578 unsigned int cpu = sys_dev->id;
581 for (i = 0; i < num_cache_leaves; i++)
582 kobject_unregister(&(INDEX_KOBJECT_PTR(cpu,i)->kobj));
583 kobject_unregister(cache_kobject[cpu]);
584 cpuid4_cache_sysfs_exit(cpu);
588 static struct sysdev_driver cache_sysdev_driver = {
589 .add = cache_add_dev,
590 .remove = __devexit_p(cache_remove_dev),
593 /* Register/Unregister the cpu_cache driver */
594 static int __devinit cache_register_driver(void)
596 if (num_cache_leaves == 0)
599 return sysdev_driver_register(&cpu_sysdev_class,&cache_sysdev_driver);
602 device_initcall(cache_register_driver);