fedora core 6 1.2949 + vserver 2.2.0
[linux-2.6.git] / arch / i386 / kernel / cpu / intel_cacheinfo.c
index 2bba6df..80b4c5d 100644 (file)
@@ -1,5 +1,21 @@
+/*
+ *      Routines to indentify caches on Intel CPU.
+ *
+ *      Changes:
+ *      Venkatesh Pallipadi    : Adding cache identification through cpuid(4)
+ *             Ashok Raj <ashok.raj@intel.com>: Work with CPU hotplug infrastructure.
+ *     Andi Kleen              : CPUID4 emulation on AMD.
+ */
+
 #include <linux/init.h>
+#include <linux/slab.h>
+#include <linux/device.h>
+#include <linux/compiler.h>
+#include <linux/cpu.h>
+#include <linux/sched.h>
+
 #include <asm/processor.h>
+#include <asm/smp.h>
 
 #define LVL_1_INST     1
 #define LVL_1_DATA     2
@@ -15,7 +31,7 @@ struct _cache_table
 };
 
 /* all the cache descriptor types we care about (no TLB or trace cache entries) */
-static struct _cache_table cache_table[] __initdata =
+static struct _cache_table cache_table[] __cpuinitdata =
 {
        { 0x06, LVL_1_INST, 8 },        /* 4-way set assoc, 32 byte line size */
        { 0x08, LVL_1_INST, 16 },       /* 4-way set assoc, 32 byte line size */
@@ -28,13 +44,23 @@ static struct _cache_table cache_table[] __initdata =
        { 0x2c, LVL_1_DATA, 32 },       /* 8-way set assoc, 64 byte line size */
        { 0x30, LVL_1_INST, 32 },       /* 8-way set assoc, 64 byte line size */
        { 0x39, LVL_2,      128 },      /* 4-way set assoc, sectored cache, 64 byte line size */
+       { 0x3a, LVL_2,      192 },      /* 6-way set assoc, sectored cache, 64 byte line size */
        { 0x3b, LVL_2,      128 },      /* 2-way set assoc, sectored cache, 64 byte line size */
        { 0x3c, LVL_2,      256 },      /* 4-way set assoc, sectored cache, 64 byte line size */
+       { 0x3d, LVL_2,      384 },      /* 6-way set assoc, sectored cache, 64 byte line size */
+       { 0x3e, LVL_2,      512 },      /* 4-way set assoc, sectored cache, 64 byte line size */
        { 0x41, LVL_2,      128 },      /* 4-way set assoc, 32 byte line size */
        { 0x42, LVL_2,      256 },      /* 4-way set assoc, 32 byte line size */
        { 0x43, LVL_2,      512 },      /* 4-way set assoc, 32 byte line size */
        { 0x44, LVL_2,      1024 },     /* 4-way set assoc, 32 byte line size */
        { 0x45, LVL_2,      2048 },     /* 4-way set assoc, 32 byte line size */
+       { 0x46, LVL_3,      4096 },     /* 4-way set assoc, 64 byte line size */
+       { 0x47, LVL_3,      8192 },     /* 8-way set assoc, 64 byte line size */
+       { 0x49, LVL_3,      4096 },     /* 16-way set assoc, 64 byte line size */
+       { 0x4a, LVL_3,      6144 },     /* 12-way set assoc, 64 byte line size */
+       { 0x4b, LVL_3,      8192 },     /* 16-way set assoc, 64 byte line size */
+       { 0x4c, LVL_3,     12288 },     /* 12-way set assoc, 64 byte line size */
+       { 0x4d, LVL_3,     16384 },     /* 16-way set assoc, 64 byte line size */
        { 0x60, LVL_1_DATA, 16 },       /* 8-way set assoc, sectored cache, 64 byte line size */
        { 0x66, LVL_1_DATA, 8 },        /* 4-way set assoc, sectored cache, 64 byte line size */
        { 0x67, LVL_1_DATA, 16 },       /* 4-way set assoc, sectored cache, 64 byte line size */
@@ -42,6 +68,7 @@ static struct _cache_table cache_table[] __initdata =
        { 0x70, LVL_TRACE,  12 },       /* 8-way set assoc */
        { 0x71, LVL_TRACE,  16 },       /* 8-way set assoc */
        { 0x72, LVL_TRACE,  32 },       /* 8-way set assoc */
+       { 0x73, LVL_TRACE,  64 },       /* 8-way set assoc */
        { 0x78, LVL_2,    1024 },       /* 4-way set assoc, 64 byte line size */
        { 0x79, LVL_2,     128 },       /* 8-way set assoc, sectored cache, 64 byte line size */
        { 0x7a, LVL_2,     256 },       /* 8-way set assoc, sectored cache, 64 byte line size */
@@ -58,15 +85,246 @@ static struct _cache_table cache_table[] __initdata =
        { 0x00, 0, 0}
 };
 
-unsigned int init_intel_cacheinfo(struct cpuinfo_x86 *c)
+
+enum _cache_type
+{
+       CACHE_TYPE_NULL = 0,
+       CACHE_TYPE_DATA = 1,
+       CACHE_TYPE_INST = 2,
+       CACHE_TYPE_UNIFIED = 3
+};
+
+union _cpuid4_leaf_eax {
+       struct {
+               enum _cache_type        type:5;
+               unsigned int            level:3;
+               unsigned int            is_self_initializing:1;
+               unsigned int            is_fully_associative:1;
+               unsigned int            reserved:4;
+               unsigned int            num_threads_sharing:12;
+               unsigned int            num_cores_on_die:6;
+       } split;
+       u32 full;
+};
+
+union _cpuid4_leaf_ebx {
+       struct {
+               unsigned int            coherency_line_size:12;
+               unsigned int            physical_line_partition:10;
+               unsigned int            ways_of_associativity:10;
+       } split;
+       u32 full;
+};
+
+union _cpuid4_leaf_ecx {
+       struct {
+               unsigned int            number_of_sets:32;
+       } split;
+       u32 full;
+};
+
+struct _cpuid4_info {
+       union _cpuid4_leaf_eax eax;
+       union _cpuid4_leaf_ebx ebx;
+       union _cpuid4_leaf_ecx ecx;
+       unsigned long size;
+       cpumask_t shared_cpu_map;
+};
+
+unsigned short                 num_cache_leaves;
+
+/* AMD doesn't have CPUID4. Emulate it here to report the same
+   information to the user.  This makes some assumptions about the machine:
+   No L3, L2 not shared, no SMT etc. that is currently true on AMD CPUs.
+
+   In theory the TLBs could be reported as fake type (they are in "dummy").
+   Maybe later */
+union l1_cache {
+       struct {
+               unsigned line_size : 8;
+               unsigned lines_per_tag : 8;
+               unsigned assoc : 8;
+               unsigned size_in_kb : 8;
+       };
+       unsigned val;
+};
+
+union l2_cache {
+       struct {
+               unsigned line_size : 8;
+               unsigned lines_per_tag : 4;
+               unsigned assoc : 4;
+               unsigned size_in_kb : 16;
+       };
+       unsigned val;
+};
+
+static const unsigned short assocs[] = {
+       [1] = 1, [2] = 2, [4] = 4, [6] = 8,
+       [8] = 16,
+       [0xf] = 0xffff // ??
+       };
+static const unsigned char levels[] = { 1, 1, 2 };
+static const unsigned char types[] = { 1, 2, 3 };
+
+static void __cpuinit amd_cpuid4(int leaf, union _cpuid4_leaf_eax *eax,
+                      union _cpuid4_leaf_ebx *ebx,
+                      union _cpuid4_leaf_ecx *ecx)
+{
+       unsigned dummy;
+       unsigned line_size, lines_per_tag, assoc, size_in_kb;
+       union l1_cache l1i, l1d;
+       union l2_cache l2;
+
+       eax->full = 0;
+       ebx->full = 0;
+       ecx->full = 0;
+
+       cpuid(0x80000005, &dummy, &dummy, &l1d.val, &l1i.val);
+       cpuid(0x80000006, &dummy, &dummy, &l2.val, &dummy);
+
+       if (leaf > 2 || !l1d.val || !l1i.val || !l2.val)
+               return;
+
+       eax->split.is_self_initializing = 1;
+       eax->split.type = types[leaf];
+       eax->split.level = levels[leaf];
+       eax->split.num_threads_sharing = 0;
+       eax->split.num_cores_on_die = current_cpu_data.x86_max_cores - 1;
+
+       if (leaf <= 1) {
+               union l1_cache *l1 = leaf == 0 ? &l1d : &l1i;
+               assoc = l1->assoc;
+               line_size = l1->line_size;
+               lines_per_tag = l1->lines_per_tag;
+               size_in_kb = l1->size_in_kb;
+       } else {
+               assoc = l2.assoc;
+               line_size = l2.line_size;
+               lines_per_tag = l2.lines_per_tag;
+               /* cpu_data has errata corrections for K7 applied */
+               size_in_kb = current_cpu_data.x86_cache_size;
+       }
+
+       if (assoc == 0xf)
+               eax->split.is_fully_associative = 1;
+       ebx->split.coherency_line_size = line_size - 1;
+       ebx->split.ways_of_associativity = assocs[assoc] - 1;
+       ebx->split.physical_line_partition = lines_per_tag - 1;
+       ecx->split.number_of_sets = (size_in_kb * 1024) / line_size /
+               (ebx->split.ways_of_associativity + 1) - 1;
+}
+
+static int __cpuinit cpuid4_cache_lookup(int index, struct _cpuid4_info *this_leaf)
+{
+       union _cpuid4_leaf_eax  eax;
+       union _cpuid4_leaf_ebx  ebx;
+       union _cpuid4_leaf_ecx  ecx;
+       unsigned                edx;
+
+       if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
+               amd_cpuid4(index, &eax, &ebx, &ecx);
+       else
+               cpuid_count(4, index, &eax.full, &ebx.full, &ecx.full,  &edx);
+       if (eax.split.type == CACHE_TYPE_NULL)
+               return -EIO; /* better error ? */
+
+       this_leaf->eax = eax;
+       this_leaf->ebx = ebx;
+       this_leaf->ecx = ecx;
+       this_leaf->size = (ecx.split.number_of_sets + 1) *
+               (ebx.split.coherency_line_size + 1) *
+               (ebx.split.physical_line_partition + 1) *
+               (ebx.split.ways_of_associativity + 1);
+       return 0;
+}
+
+/* will only be called once; __init is safe here */
+static int __init find_num_cache_leaves(void)
+{
+       unsigned int            eax, ebx, ecx, edx;
+       union _cpuid4_leaf_eax  cache_eax;
+       int                     i = -1;
+
+       do {
+               ++i;
+               /* Do cpuid(4) loop to find out num_cache_leaves */
+               cpuid_count(4, i, &eax, &ebx, &ecx, &edx);
+               cache_eax.full = eax;
+       } while (cache_eax.split.type != CACHE_TYPE_NULL);
+       return i;
+}
+
+unsigned int __cpuinit init_intel_cacheinfo(struct cpuinfo_x86 *c)
 {
        unsigned int trace = 0, l1i = 0, l1d = 0, l2 = 0, l3 = 0; /* Cache sizes */
+       unsigned int new_l1d = 0, new_l1i = 0; /* Cache sizes from cpuid(4) */
+       unsigned int new_l2 = 0, new_l3 = 0, i; /* Cache sizes from cpuid(4) */
+       unsigned int l2_id = 0, l3_id = 0, num_threads_sharing, index_msb;
+#ifdef CONFIG_X86_HT
+       unsigned int cpu = (c == &boot_cpu_data) ? 0 : (c - cpu_data);
+#endif
+
+       if (c->cpuid_level > 3) {
+               static int is_initialized;
 
-       if (c->cpuid_level > 1) {
+               if (is_initialized == 0) {
+                       /* Init num_cache_leaves from boot CPU */
+                       num_cache_leaves = find_num_cache_leaves();
+                       is_initialized++;
+               }
+
+               /*
+                * Whenever possible use cpuid(4), deterministic cache
+                * parameters cpuid leaf to find the cache details
+                */
+               for (i = 0; i < num_cache_leaves; i++) {
+                       struct _cpuid4_info this_leaf;
+
+                       int retval;
+
+                       retval = cpuid4_cache_lookup(i, &this_leaf);
+                       if (retval >= 0) {
+                               switch(this_leaf.eax.split.level) {
+                                   case 1:
+                                       if (this_leaf.eax.split.type ==
+                                                       CACHE_TYPE_DATA)
+                                               new_l1d = this_leaf.size/1024;
+                                       else if (this_leaf.eax.split.type ==
+                                                       CACHE_TYPE_INST)
+                                               new_l1i = this_leaf.size/1024;
+                                       break;
+                                   case 2:
+                                       new_l2 = this_leaf.size/1024;
+                                       num_threads_sharing = 1 + this_leaf.eax.split.num_threads_sharing;
+                                       index_msb = get_count_order(num_threads_sharing);
+                                       l2_id = c->apicid >> index_msb;
+                                       break;
+                                   case 3:
+                                       new_l3 = this_leaf.size/1024;
+                                       num_threads_sharing = 1 + this_leaf.eax.split.num_threads_sharing;
+                                       index_msb = get_count_order(num_threads_sharing);
+                                       l3_id = c->apicid >> index_msb;
+                                       break;
+                                   default:
+                                       break;
+                               }
+                       }
+               }
+       }
+       /*
+        * Don't use cpuid2 if cpuid4 is supported. For P4, we use cpuid2 for
+        * trace cache
+        */
+       if ((num_cache_leaves == 0 || c->x86 == 15) && c->cpuid_level > 1) {
                /* supports eax=2  call */
                int i, j, n;
                int regs[4];
                unsigned char *dp = (unsigned char *)regs;
+               int only_trace = 0;
+
+               if (num_cache_leaves != 0 && c->x86 == 15)
+                       only_trace = 1;
 
                /* Number of times to iterate */
                n = cpuid_eax(2) & 0xFF;
@@ -88,6 +346,8 @@ unsigned int init_intel_cacheinfo(struct cpuinfo_x86 *c)
                                while (cache_table[k].descriptor != 0)
                                {
                                        if (cache_table[k].descriptor == des) {
+                                               if (only_trace && cache_table[k].cache_type != LVL_TRACE)
+                                                       break;
                                                switch (cache_table[k].cache_type) {
                                                case LVL_1_INST:
                                                        l1i += cache_table[k].size;
@@ -113,28 +373,397 @@ unsigned int init_intel_cacheinfo(struct cpuinfo_x86 *c)
                                }
                        }
                }
+       }
 
-               if ( trace )
-                       printk (KERN_INFO "CPU: Trace cache: %dK uops", trace);
-               else if ( l1i )
-                       printk (KERN_INFO "CPU: L1 I cache: %dK", l1i);
-               if ( l1d )
-                       printk(", L1 D cache: %dK\n", l1d);
-               else
-                       printk("\n");
-               if ( l2 )
-                       printk(KERN_INFO "CPU: L2 cache: %dK\n", l2);
-               if ( l3 )
-                       printk(KERN_INFO "CPU: L3 cache: %dK\n", l3);
+       if (new_l1d)
+               l1d = new_l1d;
 
-               /*
-                * This assumes the L3 cache is shared; it typically lives in
-                * the northbridge.  The L1 caches are included by the L2
-                * cache, and so should not be included for the purpose of
-                * SMP switching weights.
-                */
-               c->x86_cache_size = l2 ? l2 : (l1i+l1d);
+       if (new_l1i)
+               l1i = new_l1i;
+
+       if (new_l2) {
+               l2 = new_l2;
+#ifdef CONFIG_X86_HT
+               cpu_llc_id[cpu] = l2_id;
+#endif
+       }
+
+       if (new_l3) {
+               l3 = new_l3;
+#ifdef CONFIG_X86_HT
+               cpu_llc_id[cpu] = l3_id;
+#endif
        }
 
+       if (trace)
+               printk (KERN_INFO "CPU: Trace cache: %dK uops", trace);
+       else if ( l1i )
+               printk (KERN_INFO "CPU: L1 I cache: %dK", l1i);
+
+       if (l1d)
+               printk(", L1 D cache: %dK\n", l1d);
+       else
+               printk("\n");
+
+       if (l2)
+               printk(KERN_INFO "CPU: L2 cache: %dK\n", l2);
+
+       if (l3)
+               printk(KERN_INFO "CPU: L3 cache: %dK\n", l3);
+
+       c->x86_cache_size = l3 ? l3 : (l2 ? l2 : (l1i+l1d));
+
        return l2;
 }
+
+/* pointer to _cpuid4_info array (for each cache leaf) */
+static struct _cpuid4_info *cpuid4_info[NR_CPUS];
+#define CPUID4_INFO_IDX(x,y)    (&((cpuid4_info[x])[y]))
+
+#ifdef CONFIG_SMP
+static void __cpuinit cache_shared_cpu_map_setup(unsigned int cpu, int index)
+{
+       struct _cpuid4_info     *this_leaf, *sibling_leaf;
+       unsigned long num_threads_sharing;
+       int index_msb, i;
+       struct cpuinfo_x86 *c = cpu_data;
+
+       this_leaf = CPUID4_INFO_IDX(cpu, index);
+       num_threads_sharing = 1 + this_leaf->eax.split.num_threads_sharing;
+
+       if (num_threads_sharing == 1)
+               cpu_set(cpu, this_leaf->shared_cpu_map);
+       else {
+               index_msb = get_count_order(num_threads_sharing);
+
+               for_each_online_cpu(i) {
+                       if (c[i].apicid >> index_msb ==
+                           c[cpu].apicid >> index_msb) {
+                               cpu_set(i, this_leaf->shared_cpu_map);
+                               if (i != cpu && cpuid4_info[i])  {
+                                       sibling_leaf = CPUID4_INFO_IDX(i, index);
+                                       cpu_set(cpu, sibling_leaf->shared_cpu_map);
+                               }
+                       }
+               }
+       }
+}
+static void __cpuinit cache_remove_shared_cpu_map(unsigned int cpu, int index)
+{
+       struct _cpuid4_info     *this_leaf, *sibling_leaf;
+       int sibling;
+
+       this_leaf = CPUID4_INFO_IDX(cpu, index);
+       for_each_cpu_mask(sibling, this_leaf->shared_cpu_map) {
+               sibling_leaf = CPUID4_INFO_IDX(sibling, index); 
+               cpu_clear(cpu, sibling_leaf->shared_cpu_map);
+       }
+}
+#else
+static void __init cache_shared_cpu_map_setup(unsigned int cpu, int index) {}
+static void __init cache_remove_shared_cpu_map(unsigned int cpu, int index) {}
+#endif
+
+static void free_cache_attributes(unsigned int cpu)
+{
+       kfree(cpuid4_info[cpu]);
+       cpuid4_info[cpu] = NULL;
+}
+
+static int __cpuinit detect_cache_attributes(unsigned int cpu)
+{
+       struct _cpuid4_info     *this_leaf;
+       unsigned long           j;
+       int                     retval;
+       cpumask_t               oldmask;
+
+       if (num_cache_leaves == 0)
+               return -ENOENT;
+
+       cpuid4_info[cpu] = kzalloc(
+           sizeof(struct _cpuid4_info) * num_cache_leaves, GFP_KERNEL);
+       if (unlikely(cpuid4_info[cpu] == NULL))
+               return -ENOMEM;
+
+       oldmask = current->cpus_allowed;
+       retval = set_cpus_allowed(current, cpumask_of_cpu(cpu));
+       if (retval)
+               goto out;
+
+       /* Do cpuid and store the results */
+       retval = 0;
+       for (j = 0; j < num_cache_leaves; j++) {
+               this_leaf = CPUID4_INFO_IDX(cpu, j);
+               retval = cpuid4_cache_lookup(j, this_leaf);
+               if (unlikely(retval < 0))
+                       break;
+               cache_shared_cpu_map_setup(cpu, j);
+       }
+       set_cpus_allowed(current, oldmask);
+
+out:
+       if (retval)
+               free_cache_attributes(cpu);
+       return retval;
+}
+
+#ifdef CONFIG_SYSFS
+
+#include <linux/kobject.h>
+#include <linux/sysfs.h>
+
+extern struct sysdev_class cpu_sysdev_class; /* from drivers/base/cpu.c */
+
+/* pointer to kobject for cpuX/cache */
+static struct kobject * cache_kobject[NR_CPUS];
+
+struct _index_kobject {
+       struct kobject kobj;
+       unsigned int cpu;
+       unsigned short index;
+};
+
+/* pointer to array of kobjects for cpuX/cache/indexY */
+static struct _index_kobject *index_kobject[NR_CPUS];
+#define INDEX_KOBJECT_PTR(x,y)    (&((index_kobject[x])[y]))
+
+#define show_one_plus(file_name, object, val)                          \
+static ssize_t show_##file_name                                                \
+                       (struct _cpuid4_info *this_leaf, char *buf)     \
+{                                                                      \
+       return sprintf (buf, "%lu\n", (unsigned long)this_leaf->object + val); \
+}
+
+show_one_plus(level, eax.split.level, 0);
+show_one_plus(coherency_line_size, ebx.split.coherency_line_size, 1);
+show_one_plus(physical_line_partition, ebx.split.physical_line_partition, 1);
+show_one_plus(ways_of_associativity, ebx.split.ways_of_associativity, 1);
+show_one_plus(number_of_sets, ecx.split.number_of_sets, 1);
+
+static ssize_t show_size(struct _cpuid4_info *this_leaf, char *buf)
+{
+       return sprintf (buf, "%luK\n", this_leaf->size / 1024);
+}
+
+static ssize_t show_shared_cpu_map(struct _cpuid4_info *this_leaf, char *buf)
+{
+       char mask_str[NR_CPUS];
+       cpumask_scnprintf(mask_str, NR_CPUS, this_leaf->shared_cpu_map);
+       return sprintf(buf, "%s\n", mask_str);
+}
+
+static ssize_t show_type(struct _cpuid4_info *this_leaf, char *buf) {
+       switch(this_leaf->eax.split.type) {
+           case CACHE_TYPE_DATA:
+               return sprintf(buf, "Data\n");
+               break;
+           case CACHE_TYPE_INST:
+               return sprintf(buf, "Instruction\n");
+               break;
+           case CACHE_TYPE_UNIFIED:
+               return sprintf(buf, "Unified\n");
+               break;
+           default:
+               return sprintf(buf, "Unknown\n");
+               break;
+       }
+}
+
+struct _cache_attr {
+       struct attribute attr;
+       ssize_t (*show)(struct _cpuid4_info *, char *);
+       ssize_t (*store)(struct _cpuid4_info *, const char *, size_t count);
+};
+
+#define define_one_ro(_name) \
+static struct _cache_attr _name = \
+       __ATTR(_name, 0444, show_##_name, NULL)
+
+define_one_ro(level);
+define_one_ro(type);
+define_one_ro(coherency_line_size);
+define_one_ro(physical_line_partition);
+define_one_ro(ways_of_associativity);
+define_one_ro(number_of_sets);
+define_one_ro(size);
+define_one_ro(shared_cpu_map);
+
+static struct attribute * default_attrs[] = {
+       &type.attr,
+       &level.attr,
+       &coherency_line_size.attr,
+       &physical_line_partition.attr,
+       &ways_of_associativity.attr,
+       &number_of_sets.attr,
+       &size.attr,
+       &shared_cpu_map.attr,
+       NULL
+};
+
+#define to_object(k) container_of(k, struct _index_kobject, kobj)
+#define to_attr(a) container_of(a, struct _cache_attr, attr)
+
+static ssize_t show(struct kobject * kobj, struct attribute * attr, char * buf)
+{
+       struct _cache_attr *fattr = to_attr(attr);
+       struct _index_kobject *this_leaf = to_object(kobj);
+       ssize_t ret;
+
+       ret = fattr->show ?
+               fattr->show(CPUID4_INFO_IDX(this_leaf->cpu, this_leaf->index),
+                       buf) :
+               0;
+       return ret;
+}
+
+static ssize_t store(struct kobject * kobj, struct attribute * attr,
+                    const char * buf, size_t count)
+{
+       return 0;
+}
+
+static struct sysfs_ops sysfs_ops = {
+       .show   = show,
+       .store  = store,
+};
+
+static struct kobj_type ktype_cache = {
+       .sysfs_ops      = &sysfs_ops,
+       .default_attrs  = default_attrs,
+};
+
+static struct kobj_type ktype_percpu_entry = {
+       .sysfs_ops      = &sysfs_ops,
+};
+
+static void cpuid4_cache_sysfs_exit(unsigned int cpu)
+{
+       kfree(cache_kobject[cpu]);
+       kfree(index_kobject[cpu]);
+       cache_kobject[cpu] = NULL;
+       index_kobject[cpu] = NULL;
+       free_cache_attributes(cpu);
+}
+
+static int __cpuinit cpuid4_cache_sysfs_init(unsigned int cpu)
+{
+
+       if (num_cache_leaves == 0)
+               return -ENOENT;
+
+       detect_cache_attributes(cpu);
+       if (cpuid4_info[cpu] == NULL)
+               return -ENOENT;
+
+       /* Allocate all required memory */
+       cache_kobject[cpu] = kzalloc(sizeof(struct kobject), GFP_KERNEL);
+       if (unlikely(cache_kobject[cpu] == NULL))
+               goto err_out;
+
+       index_kobject[cpu] = kzalloc(
+           sizeof(struct _index_kobject ) * num_cache_leaves, GFP_KERNEL);
+       if (unlikely(index_kobject[cpu] == NULL))
+               goto err_out;
+
+       return 0;
+
+err_out:
+       cpuid4_cache_sysfs_exit(cpu);
+       return -ENOMEM;
+}
+
+/* Add/Remove cache interface for CPU device */
+static int __cpuinit cache_add_dev(struct sys_device * sys_dev)
+{
+       unsigned int cpu = sys_dev->id;
+       unsigned long i, j;
+       struct _index_kobject *this_object;
+       int retval = 0;
+
+       retval = cpuid4_cache_sysfs_init(cpu);
+       if (unlikely(retval < 0))
+               return retval;
+
+       cache_kobject[cpu]->parent = &sys_dev->kobj;
+       kobject_set_name(cache_kobject[cpu], "%s", "cache");
+       cache_kobject[cpu]->ktype = &ktype_percpu_entry;
+       retval = kobject_register(cache_kobject[cpu]);
+
+       for (i = 0; i < num_cache_leaves; i++) {
+               this_object = INDEX_KOBJECT_PTR(cpu,i);
+               this_object->cpu = cpu;
+               this_object->index = i;
+               this_object->kobj.parent = cache_kobject[cpu];
+               kobject_set_name(&(this_object->kobj), "index%1lu", i);
+               this_object->kobj.ktype = &ktype_cache;
+               retval = kobject_register(&(this_object->kobj));
+               if (unlikely(retval)) {
+                       for (j = 0; j < i; j++) {
+                               kobject_unregister(
+                                       &(INDEX_KOBJECT_PTR(cpu,j)->kobj));
+                       }
+                       kobject_unregister(cache_kobject[cpu]);
+                       cpuid4_cache_sysfs_exit(cpu);
+                       break;
+               }
+       }
+       return retval;
+}
+
+static void __cpuexit cache_remove_dev(struct sys_device * sys_dev)
+{
+       unsigned int cpu = sys_dev->id;
+       unsigned long i;
+
+       for (i = 0; i < num_cache_leaves; i++) {
+               cache_remove_shared_cpu_map(cpu, i);
+               kobject_unregister(&(INDEX_KOBJECT_PTR(cpu,i)->kobj));
+       }
+       kobject_unregister(cache_kobject[cpu]);
+       cpuid4_cache_sysfs_exit(cpu);
+       return;
+}
+
+static int __cpuinit cacheinfo_cpu_callback(struct notifier_block *nfb,
+                                       unsigned long action, void *hcpu)
+{
+       unsigned int cpu = (unsigned long)hcpu;
+       struct sys_device *sys_dev;
+
+       sys_dev = get_cpu_sysdev(cpu);
+       switch (action) {
+       case CPU_ONLINE:
+               cache_add_dev(sys_dev);
+               break;
+       case CPU_DEAD:
+               cache_remove_dev(sys_dev);
+               break;
+       }
+       return NOTIFY_OK;
+}
+
+static struct notifier_block __cpuinitdata cacheinfo_cpu_notifier =
+{
+    .notifier_call = cacheinfo_cpu_callback,
+};
+
+static int __cpuinit cache_sysfs_init(void)
+{
+       int i;
+
+       if (num_cache_leaves == 0)
+               return 0;
+
+       register_hotcpu_notifier(&cacheinfo_cpu_notifier);
+
+       for_each_online_cpu(i) {
+               cacheinfo_cpu_callback(&cacheinfo_cpu_notifier, CPU_ONLINE,
+                       (void *)(long)i);
+       }
+
+       return 0;
+}
+
+device_initcall(cache_sysfs_init);
+
+#endif