1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Routines to identify caches on Intel CPU.
4 *
5 * Changes:
6 * Venkatesh Pallipadi : Adding cache identification through cpuid(4)
7 * Ashok Raj <ashok.raj@intel.com>: Work with CPU hotplug infrastructure.
8 * Andi Kleen / Andreas Herrmann : CPUID4 emulation on AMD.
9 */
10
11 #include <linux/slab.h>
12 #include <linux/cacheinfo.h>
13 #include <linux/cpu.h>
14 #include <linux/sched.h>
15 #include <linux/capability.h>
16 #include <linux/sysfs.h>
17 #include <linux/pci.h>
18
19 #include <asm/cpufeature.h>
20 #include <asm/cacheinfo.h>
21 #include <asm/amd_nb.h>
22 #include <asm/smp.h>
23
24 #include "cpu.h"
25
26 #define LVL_1_INST 1
27 #define LVL_1_DATA 2
28 #define LVL_2 3
29 #define LVL_3 4
30 #define LVL_TRACE 5
31
32 struct _cache_table {
33 unsigned char descriptor;
34 char cache_type;
35 short size;
36 };
37
38 #define MB(x) ((x) * 1024)
39
40 /* All the cache descriptor types we care about (no TLB or
41 trace cache entries) */
42
43 static const struct _cache_table cache_table[] =
44 {
45 { 0x06, LVL_1_INST, 8 }, /* 4-way set assoc, 32 byte line size */
46 { 0x08, LVL_1_INST, 16 }, /* 4-way set assoc, 32 byte line size */
47 { 0x09, LVL_1_INST, 32 }, /* 4-way set assoc, 64 byte line size */
48 { 0x0a, LVL_1_DATA, 8 }, /* 2 way set assoc, 32 byte line size */
49 { 0x0c, LVL_1_DATA, 16 }, /* 4-way set assoc, 32 byte line size */
50 { 0x0d, LVL_1_DATA, 16 }, /* 4-way set assoc, 64 byte line size */
51 { 0x0e, LVL_1_DATA, 24 }, /* 6-way set assoc, 64 byte line size */
52 { 0x21, LVL_2, 256 }, /* 8-way set assoc, 64 byte line size */
53 { 0x22, LVL_3, 512 }, /* 4-way set assoc, sectored cache, 64 byte line size */
54 { 0x23, LVL_3, MB(1) }, /* 8-way set assoc, sectored cache, 64 byte line size */
55 { 0x25, LVL_3, MB(2) }, /* 8-way set assoc, sectored cache, 64 byte line size */
56 { 0x29, LVL_3, MB(4) }, /* 8-way set assoc, sectored cache, 64 byte line size */
57 { 0x2c, LVL_1_DATA, 32 }, /* 8-way set assoc, 64 byte line size */
58 { 0x30, LVL_1_INST, 32 }, /* 8-way set assoc, 64 byte line size */
59 { 0x39, LVL_2, 128 }, /* 4-way set assoc, sectored cache, 64 byte line size */
60 { 0x3a, LVL_2, 192 }, /* 6-way set assoc, sectored cache, 64 byte line size */
61 { 0x3b, LVL_2, 128 }, /* 2-way set assoc, sectored cache, 64 byte line size */
62 { 0x3c, LVL_2, 256 }, /* 4-way set assoc, sectored cache, 64 byte line size */
63 { 0x3d, LVL_2, 384 }, /* 6-way set assoc, sectored cache, 64 byte line size */
64 { 0x3e, LVL_2, 512 }, /* 4-way set assoc, sectored cache, 64 byte line size */
65 { 0x3f, LVL_2, 256 }, /* 2-way set assoc, 64 byte line size */
66 { 0x41, LVL_2, 128 }, /* 4-way set assoc, 32 byte line size */
67 { 0x42, LVL_2, 256 }, /* 4-way set assoc, 32 byte line size */
68 { 0x43, LVL_2, 512 }, /* 4-way set assoc, 32 byte line size */
69 { 0x44, LVL_2, MB(1) }, /* 4-way set assoc, 32 byte line size */
70 { 0x45, LVL_2, MB(2) }, /* 4-way set assoc, 32 byte line size */
71 { 0x46, LVL_3, MB(4) }, /* 4-way set assoc, 64 byte line size */
72 { 0x47, LVL_3, MB(8) }, /* 8-way set assoc, 64 byte line size */
73 { 0x48, LVL_2, MB(3) }, /* 12-way set assoc, 64 byte line size */
74 { 0x49, LVL_3, MB(4) }, /* 16-way set assoc, 64 byte line size */
75 { 0x4a, LVL_3, MB(6) }, /* 12-way set assoc, 64 byte line size */
76 { 0x4b, LVL_3, MB(8) }, /* 16-way set assoc, 64 byte line size */
77 { 0x4c, LVL_3, MB(12) }, /* 12-way set assoc, 64 byte line size */
78 { 0x4d, LVL_3, MB(16) }, /* 16-way set assoc, 64 byte line size */
79 { 0x4e, LVL_2, MB(6) }, /* 24-way set assoc, 64 byte line size */
80 { 0x60, LVL_1_DATA, 16 }, /* 8-way set assoc, sectored cache, 64 byte line size */
81 { 0x66, LVL_1_DATA, 8 }, /* 4-way set assoc, sectored cache, 64 byte line size */
82 { 0x67, LVL_1_DATA, 16 }, /* 4-way set assoc, sectored cache, 64 byte line size */
83 { 0x68, LVL_1_DATA, 32 }, /* 4-way set assoc, sectored cache, 64 byte line size */
84 { 0x70, LVL_TRACE, 12 }, /* 8-way set assoc */
85 { 0x71, LVL_TRACE, 16 }, /* 8-way set assoc */
86 { 0x72, LVL_TRACE, 32 }, /* 8-way set assoc */
87 { 0x73, LVL_TRACE, 64 }, /* 8-way set assoc */
88 { 0x78, LVL_2, MB(1) }, /* 4-way set assoc, 64 byte line size */
89 { 0x79, LVL_2, 128 }, /* 8-way set assoc, sectored cache, 64 byte line size */
90 { 0x7a, LVL_2, 256 }, /* 8-way set assoc, sectored cache, 64 byte line size */
91 { 0x7b, LVL_2, 512 }, /* 8-way set assoc, sectored cache, 64 byte line size */
92 { 0x7c, LVL_2, MB(1) }, /* 8-way set assoc, sectored cache, 64 byte line size */
93 { 0x7d, LVL_2, MB(2) }, /* 8-way set assoc, 64 byte line size */
94 { 0x7f, LVL_2, 512 }, /* 2-way set assoc, 64 byte line size */
95 { 0x80, LVL_2, 512 }, /* 8-way set assoc, 64 byte line size */
96 { 0x82, LVL_2, 256 }, /* 8-way set assoc, 32 byte line size */
97 { 0x83, LVL_2, 512 }, /* 8-way set assoc, 32 byte line size */
98 { 0x84, LVL_2, MB(1) }, /* 8-way set assoc, 32 byte line size */
99 { 0x85, LVL_2, MB(2) }, /* 8-way set assoc, 32 byte line size */
100 { 0x86, LVL_2, 512 }, /* 4-way set assoc, 64 byte line size */
101 { 0x87, LVL_2, MB(1) }, /* 8-way set assoc, 64 byte line size */
102 { 0xd0, LVL_3, 512 }, /* 4-way set assoc, 64 byte line size */
103 { 0xd1, LVL_3, MB(1) }, /* 4-way set assoc, 64 byte line size */
104 { 0xd2, LVL_3, MB(2) }, /* 4-way set assoc, 64 byte line size */
105 { 0xd6, LVL_3, MB(1) }, /* 8-way set assoc, 64 byte line size */
106 { 0xd7, LVL_3, MB(2) }, /* 8-way set assoc, 64 byte line size */
107 { 0xd8, LVL_3, MB(4) }, /* 12-way set assoc, 64 byte line size */
108 { 0xdc, LVL_3, MB(2) }, /* 12-way set assoc, 64 byte line size */
109 { 0xdd, LVL_3, MB(4) }, /* 12-way set assoc, 64 byte line size */
110 { 0xde, LVL_3, MB(8) }, /* 12-way set assoc, 64 byte line size */
111 { 0xe2, LVL_3, MB(2) }, /* 16-way set assoc, 64 byte line size */
112 { 0xe3, LVL_3, MB(4) }, /* 16-way set assoc, 64 byte line size */
113 { 0xe4, LVL_3, MB(8) }, /* 16-way set assoc, 64 byte line size */
114 { 0xea, LVL_3, MB(12) }, /* 24-way set assoc, 64 byte line size */
115 { 0xeb, LVL_3, MB(18) }, /* 24-way set assoc, 64 byte line size */
116 { 0xec, LVL_3, MB(24) }, /* 24-way set assoc, 64 byte line size */
117 { 0x00, 0, 0}
118 };
119
120
121 enum _cache_type {
122 CTYPE_NULL = 0,
123 CTYPE_DATA = 1,
124 CTYPE_INST = 2,
125 CTYPE_UNIFIED = 3
126 };
127
128 union _cpuid4_leaf_eax {
129 struct {
130 enum _cache_type type:5;
131 unsigned int level:3;
132 unsigned int is_self_initializing:1;
133 unsigned int is_fully_associative:1;
134 unsigned int reserved:4;
135 unsigned int num_threads_sharing:12;
136 unsigned int num_cores_on_die:6;
137 } split;
138 u32 full;
139 };
140
141 union _cpuid4_leaf_ebx {
142 struct {
143 unsigned int coherency_line_size:12;
144 unsigned int physical_line_partition:10;
145 unsigned int ways_of_associativity:10;
146 } split;
147 u32 full;
148 };
149
150 union _cpuid4_leaf_ecx {
151 struct {
152 unsigned int number_of_sets:32;
153 } split;
154 u32 full;
155 };
156
157 struct _cpuid4_info_regs {
158 union _cpuid4_leaf_eax eax;
159 union _cpuid4_leaf_ebx ebx;
160 union _cpuid4_leaf_ecx ecx;
161 unsigned int id;
162 unsigned long size;
163 struct amd_northbridge *nb;
164 };
165
166 static unsigned short num_cache_leaves;
167
168 /* AMD doesn't have CPUID4. Emulate it here to report the same
169 information to the user. This makes some assumptions about the machine:
170 L2 not shared, no SMT etc. that is currently true on AMD CPUs.
171
172 In theory the TLBs could be reported as fake type (they are in "dummy").
173 Maybe later */
174 union l1_cache {
175 struct {
176 unsigned line_size:8;
177 unsigned lines_per_tag:8;
178 unsigned assoc:8;
179 unsigned size_in_kb:8;
180 };
181 unsigned val;
182 };
183
184 union l2_cache {
185 struct {
186 unsigned line_size:8;
187 unsigned lines_per_tag:4;
188 unsigned assoc:4;
189 unsigned size_in_kb:16;
190 };
191 unsigned val;
192 };
193
194 union l3_cache {
195 struct {
196 unsigned line_size:8;
197 unsigned lines_per_tag:4;
198 unsigned assoc:4;
199 unsigned res:2;
200 unsigned size_encoded:14;
201 };
202 unsigned val;
203 };
204
205 static const unsigned short assocs[] = {
206 [1] = 1,
207 [2] = 2,
208 [4] = 4,
209 [6] = 8,
210 [8] = 16,
211 [0xa] = 32,
212 [0xb] = 48,
213 [0xc] = 64,
214 [0xd] = 96,
215 [0xe] = 128,
216 [0xf] = 0xffff /* fully associative - no way to show this currently */
217 };
218
219 static const unsigned char levels[] = { 1, 1, 2, 3 };
220 static const unsigned char types[] = { 1, 2, 3, 3 };
221
222 static const enum cache_type cache_type_map[] = {
223 [CTYPE_NULL] = CACHE_TYPE_NOCACHE,
224 [CTYPE_DATA] = CACHE_TYPE_DATA,
225 [CTYPE_INST] = CACHE_TYPE_INST,
226 [CTYPE_UNIFIED] = CACHE_TYPE_UNIFIED,
227 };
228
229 static void
amd_cpuid4(int leaf,union _cpuid4_leaf_eax * eax,union _cpuid4_leaf_ebx * ebx,union _cpuid4_leaf_ecx * ecx)230 amd_cpuid4(int leaf, union _cpuid4_leaf_eax *eax,
231 union _cpuid4_leaf_ebx *ebx,
232 union _cpuid4_leaf_ecx *ecx)
233 {
234 unsigned dummy;
235 unsigned line_size, lines_per_tag, assoc, size_in_kb;
236 union l1_cache l1i, l1d;
237 union l2_cache l2;
238 union l3_cache l3;
239 union l1_cache *l1 = &l1d;
240
241 eax->full = 0;
242 ebx->full = 0;
243 ecx->full = 0;
244
245 cpuid(0x80000005, &dummy, &dummy, &l1d.val, &l1i.val);
246 cpuid(0x80000006, &dummy, &dummy, &l2.val, &l3.val);
247
248 switch (leaf) {
249 case 1:
250 l1 = &l1i;
251 /* fall through */
252 case 0:
253 if (!l1->val)
254 return;
255 assoc = assocs[l1->assoc];
256 line_size = l1->line_size;
257 lines_per_tag = l1->lines_per_tag;
258 size_in_kb = l1->size_in_kb;
259 break;
260 case 2:
261 if (!l2.val)
262 return;
263 assoc = assocs[l2.assoc];
264 line_size = l2.line_size;
265 lines_per_tag = l2.lines_per_tag;
266 /* cpu_data has errata corrections for K7 applied */
267 size_in_kb = __this_cpu_read(cpu_info.x86_cache_size);
268 break;
269 case 3:
270 if (!l3.val)
271 return;
272 assoc = assocs[l3.assoc];
273 line_size = l3.line_size;
274 lines_per_tag = l3.lines_per_tag;
275 size_in_kb = l3.size_encoded * 512;
276 if (boot_cpu_has(X86_FEATURE_AMD_DCM)) {
277 size_in_kb = size_in_kb >> 1;
278 assoc = assoc >> 1;
279 }
280 break;
281 default:
282 return;
283 }
284
285 eax->split.is_self_initializing = 1;
286 eax->split.type = types[leaf];
287 eax->split.level = levels[leaf];
288 eax->split.num_threads_sharing = 0;
289 eax->split.num_cores_on_die = __this_cpu_read(cpu_info.x86_max_cores) - 1;
290
291
292 if (assoc == 0xffff)
293 eax->split.is_fully_associative = 1;
294 ebx->split.coherency_line_size = line_size - 1;
295 ebx->split.ways_of_associativity = assoc - 1;
296 ebx->split.physical_line_partition = lines_per_tag - 1;
297 ecx->split.number_of_sets = (size_in_kb * 1024) / line_size /
298 (ebx->split.ways_of_associativity + 1) - 1;
299 }
300
301 #if defined(CONFIG_AMD_NB) && defined(CONFIG_SYSFS)
302
303 /*
304 * L3 cache descriptors
305 */
amd_calc_l3_indices(struct amd_northbridge * nb)306 static void amd_calc_l3_indices(struct amd_northbridge *nb)
307 {
308 struct amd_l3_cache *l3 = &nb->l3_cache;
309 unsigned int sc0, sc1, sc2, sc3;
310 u32 val = 0;
311
312 pci_read_config_dword(nb->misc, 0x1C4, &val);
313
314 /* calculate subcache sizes */
315 l3->subcaches[0] = sc0 = !(val & BIT(0));
316 l3->subcaches[1] = sc1 = !(val & BIT(4));
317
318 if (boot_cpu_data.x86 == 0x15) {
319 l3->subcaches[0] = sc0 += !(val & BIT(1));
320 l3->subcaches[1] = sc1 += !(val & BIT(5));
321 }
322
323 l3->subcaches[2] = sc2 = !(val & BIT(8)) + !(val & BIT(9));
324 l3->subcaches[3] = sc3 = !(val & BIT(12)) + !(val & BIT(13));
325
326 l3->indices = (max(max3(sc0, sc1, sc2), sc3) << 10) - 1;
327 }
328
329 /*
330 * check whether a slot used for disabling an L3 index is occupied.
331 * @l3: L3 cache descriptor
332 * @slot: slot number (0..1)
333 *
334 * @returns: the disabled index if used or negative value if slot free.
335 */
amd_get_l3_disable_slot(struct amd_northbridge * nb,unsigned slot)336 static int amd_get_l3_disable_slot(struct amd_northbridge *nb, unsigned slot)
337 {
338 unsigned int reg = 0;
339
340 pci_read_config_dword(nb->misc, 0x1BC + slot * 4, ®);
341
342 /* check whether this slot is activated already */
343 if (reg & (3UL << 30))
344 return reg & 0xfff;
345
346 return -1;
347 }
348
show_cache_disable(struct cacheinfo * this_leaf,char * buf,unsigned int slot)349 static ssize_t show_cache_disable(struct cacheinfo *this_leaf, char *buf,
350 unsigned int slot)
351 {
352 int index;
353 struct amd_northbridge *nb = this_leaf->priv;
354
355 index = amd_get_l3_disable_slot(nb, slot);
356 if (index >= 0)
357 return sprintf(buf, "%d\n", index);
358
359 return sprintf(buf, "FREE\n");
360 }
361
362 #define SHOW_CACHE_DISABLE(slot) \
363 static ssize_t \
364 cache_disable_##slot##_show(struct device *dev, \
365 struct device_attribute *attr, char *buf) \
366 { \
367 struct cacheinfo *this_leaf = dev_get_drvdata(dev); \
368 return show_cache_disable(this_leaf, buf, slot); \
369 }
370 SHOW_CACHE_DISABLE(0)
371 SHOW_CACHE_DISABLE(1)
372
amd_l3_disable_index(struct amd_northbridge * nb,int cpu,unsigned slot,unsigned long idx)373 static void amd_l3_disable_index(struct amd_northbridge *nb, int cpu,
374 unsigned slot, unsigned long idx)
375 {
376 int i;
377
378 idx |= BIT(30);
379
380 /*
381 * disable index in all 4 subcaches
382 */
383 for (i = 0; i < 4; i++) {
384 u32 reg = idx | (i << 20);
385
386 if (!nb->l3_cache.subcaches[i])
387 continue;
388
389 pci_write_config_dword(nb->misc, 0x1BC + slot * 4, reg);
390
391 /*
392 * We need to WBINVD on a core on the node containing the L3
393 * cache which indices we disable therefore a simple wbinvd()
394 * is not sufficient.
395 */
396 wbinvd_on_cpu(cpu);
397
398 reg |= BIT(31);
399 pci_write_config_dword(nb->misc, 0x1BC + slot * 4, reg);
400 }
401 }
402
403 /*
404 * disable a L3 cache index by using a disable-slot
405 *
406 * @l3: L3 cache descriptor
407 * @cpu: A CPU on the node containing the L3 cache
408 * @slot: slot number (0..1)
409 * @index: index to disable
410 *
411 * @return: 0 on success, error status on failure
412 */
amd_set_l3_disable_slot(struct amd_northbridge * nb,int cpu,unsigned slot,unsigned long index)413 static int amd_set_l3_disable_slot(struct amd_northbridge *nb, int cpu,
414 unsigned slot, unsigned long index)
415 {
416 int ret = 0;
417
418 /* check if @slot is already used or the index is already disabled */
419 ret = amd_get_l3_disable_slot(nb, slot);
420 if (ret >= 0)
421 return -EEXIST;
422
423 if (index > nb->l3_cache.indices)
424 return -EINVAL;
425
426 /* check whether the other slot has disabled the same index already */
427 if (index == amd_get_l3_disable_slot(nb, !slot))
428 return -EEXIST;
429
430 amd_l3_disable_index(nb, cpu, slot, index);
431
432 return 0;
433 }
434
store_cache_disable(struct cacheinfo * this_leaf,const char * buf,size_t count,unsigned int slot)435 static ssize_t store_cache_disable(struct cacheinfo *this_leaf,
436 const char *buf, size_t count,
437 unsigned int slot)
438 {
439 unsigned long val = 0;
440 int cpu, err = 0;
441 struct amd_northbridge *nb = this_leaf->priv;
442
443 if (!capable(CAP_SYS_ADMIN))
444 return -EPERM;
445
446 cpu = cpumask_first(&this_leaf->shared_cpu_map);
447
448 if (kstrtoul(buf, 10, &val) < 0)
449 return -EINVAL;
450
451 err = amd_set_l3_disable_slot(nb, cpu, slot, val);
452 if (err) {
453 if (err == -EEXIST)
454 pr_warn("L3 slot %d in use/index already disabled!\n",
455 slot);
456 return err;
457 }
458 return count;
459 }
460
461 #define STORE_CACHE_DISABLE(slot) \
462 static ssize_t \
463 cache_disable_##slot##_store(struct device *dev, \
464 struct device_attribute *attr, \
465 const char *buf, size_t count) \
466 { \
467 struct cacheinfo *this_leaf = dev_get_drvdata(dev); \
468 return store_cache_disable(this_leaf, buf, count, slot); \
469 }
470 STORE_CACHE_DISABLE(0)
471 STORE_CACHE_DISABLE(1)
472
subcaches_show(struct device * dev,struct device_attribute * attr,char * buf)473 static ssize_t subcaches_show(struct device *dev,
474 struct device_attribute *attr, char *buf)
475 {
476 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
477 int cpu = cpumask_first(&this_leaf->shared_cpu_map);
478
479 return sprintf(buf, "%x\n", amd_get_subcaches(cpu));
480 }
481
subcaches_store(struct device * dev,struct device_attribute * attr,const char * buf,size_t count)482 static ssize_t subcaches_store(struct device *dev,
483 struct device_attribute *attr,
484 const char *buf, size_t count)
485 {
486 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
487 int cpu = cpumask_first(&this_leaf->shared_cpu_map);
488 unsigned long val;
489
490 if (!capable(CAP_SYS_ADMIN))
491 return -EPERM;
492
493 if (kstrtoul(buf, 16, &val) < 0)
494 return -EINVAL;
495
496 if (amd_set_subcaches(cpu, val))
497 return -EINVAL;
498
499 return count;
500 }
501
502 static DEVICE_ATTR_RW(cache_disable_0);
503 static DEVICE_ATTR_RW(cache_disable_1);
504 static DEVICE_ATTR_RW(subcaches);
505
506 static umode_t
cache_private_attrs_is_visible(struct kobject * kobj,struct attribute * attr,int unused)507 cache_private_attrs_is_visible(struct kobject *kobj,
508 struct attribute *attr, int unused)
509 {
510 struct device *dev = kobj_to_dev(kobj);
511 struct cacheinfo *this_leaf = dev_get_drvdata(dev);
512 umode_t mode = attr->mode;
513
514 if (!this_leaf->priv)
515 return 0;
516
517 if ((attr == &dev_attr_subcaches.attr) &&
518 amd_nb_has_feature(AMD_NB_L3_PARTITIONING))
519 return mode;
520
521 if ((attr == &dev_attr_cache_disable_0.attr ||
522 attr == &dev_attr_cache_disable_1.attr) &&
523 amd_nb_has_feature(AMD_NB_L3_INDEX_DISABLE))
524 return mode;
525
526 return 0;
527 }
528
529 static struct attribute_group cache_private_group = {
530 .is_visible = cache_private_attrs_is_visible,
531 };
532
init_amd_l3_attrs(void)533 static void init_amd_l3_attrs(void)
534 {
535 int n = 1;
536 static struct attribute **amd_l3_attrs;
537
538 if (amd_l3_attrs) /* already initialized */
539 return;
540
541 if (amd_nb_has_feature(AMD_NB_L3_INDEX_DISABLE))
542 n += 2;
543 if (amd_nb_has_feature(AMD_NB_L3_PARTITIONING))
544 n += 1;
545
546 amd_l3_attrs = kcalloc(n, sizeof(*amd_l3_attrs), GFP_KERNEL);
547 if (!amd_l3_attrs)
548 return;
549
550 n = 0;
551 if (amd_nb_has_feature(AMD_NB_L3_INDEX_DISABLE)) {
552 amd_l3_attrs[n++] = &dev_attr_cache_disable_0.attr;
553 amd_l3_attrs[n++] = &dev_attr_cache_disable_1.attr;
554 }
555 if (amd_nb_has_feature(AMD_NB_L3_PARTITIONING))
556 amd_l3_attrs[n++] = &dev_attr_subcaches.attr;
557
558 cache_private_group.attrs = amd_l3_attrs;
559 }
560
561 const struct attribute_group *
cache_get_priv_group(struct cacheinfo * this_leaf)562 cache_get_priv_group(struct cacheinfo *this_leaf)
563 {
564 struct amd_northbridge *nb = this_leaf->priv;
565
566 if (this_leaf->level < 3 || !nb)
567 return NULL;
568
569 if (nb && nb->l3_cache.indices)
570 init_amd_l3_attrs();
571
572 return &cache_private_group;
573 }
574
amd_init_l3_cache(struct _cpuid4_info_regs * this_leaf,int index)575 static void amd_init_l3_cache(struct _cpuid4_info_regs *this_leaf, int index)
576 {
577 int node;
578
579 /* only for L3, and not in virtualized environments */
580 if (index < 3)
581 return;
582
583 node = amd_get_nb_id(smp_processor_id());
584 this_leaf->nb = node_to_amd_nb(node);
585 if (this_leaf->nb && !this_leaf->nb->l3_cache.indices)
586 amd_calc_l3_indices(this_leaf->nb);
587 }
588 #else
589 #define amd_init_l3_cache(x, y)
590 #endif /* CONFIG_AMD_NB && CONFIG_SYSFS */
591
592 static int
cpuid4_cache_lookup_regs(int index,struct _cpuid4_info_regs * this_leaf)593 cpuid4_cache_lookup_regs(int index, struct _cpuid4_info_regs *this_leaf)
594 {
595 union _cpuid4_leaf_eax eax;
596 union _cpuid4_leaf_ebx ebx;
597 union _cpuid4_leaf_ecx ecx;
598 unsigned edx;
599
600 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
601 if (boot_cpu_has(X86_FEATURE_TOPOEXT))
602 cpuid_count(0x8000001d, index, &eax.full,
603 &ebx.full, &ecx.full, &edx);
604 else
605 amd_cpuid4(index, &eax, &ebx, &ecx);
606 amd_init_l3_cache(this_leaf, index);
607 } else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) {
608 cpuid_count(0x8000001d, index, &eax.full,
609 &ebx.full, &ecx.full, &edx);
610 amd_init_l3_cache(this_leaf, index);
611 } else {
612 cpuid_count(4, index, &eax.full, &ebx.full, &ecx.full, &edx);
613 }
614
615 if (eax.split.type == CTYPE_NULL)
616 return -EIO; /* better error ? */
617
618 this_leaf->eax = eax;
619 this_leaf->ebx = ebx;
620 this_leaf->ecx = ecx;
621 this_leaf->size = (ecx.split.number_of_sets + 1) *
622 (ebx.split.coherency_line_size + 1) *
623 (ebx.split.physical_line_partition + 1) *
624 (ebx.split.ways_of_associativity + 1);
625 return 0;
626 }
627
find_num_cache_leaves(struct cpuinfo_x86 * c)628 static int find_num_cache_leaves(struct cpuinfo_x86 *c)
629 {
630 unsigned int eax, ebx, ecx, edx, op;
631 union _cpuid4_leaf_eax cache_eax;
632 int i = -1;
633
634 if (c->x86_vendor == X86_VENDOR_AMD ||
635 c->x86_vendor == X86_VENDOR_HYGON)
636 op = 0x8000001d;
637 else
638 op = 4;
639
640 do {
641 ++i;
642 /* Do cpuid(op) loop to find out num_cache_leaves */
643 cpuid_count(op, i, &eax, &ebx, &ecx, &edx);
644 cache_eax.full = eax;
645 } while (cache_eax.split.type != CTYPE_NULL);
646 return i;
647 }
648
cacheinfo_amd_init_llc_id(struct cpuinfo_x86 * c,int cpu,u8 node_id)649 void cacheinfo_amd_init_llc_id(struct cpuinfo_x86 *c, int cpu, u8 node_id)
650 {
651 /*
652 * We may have multiple LLCs if L3 caches exist, so check if we
653 * have an L3 cache by looking at the L3 cache CPUID leaf.
654 */
655 if (!cpuid_edx(0x80000006))
656 return;
657
658 if (c->x86 < 0x17) {
659 /* LLC is at the node level. */
660 per_cpu(cpu_llc_id, cpu) = node_id;
661 } else if (c->x86 == 0x17 && c->x86_model <= 0x1F) {
662 /*
663 * LLC is at the core complex level.
664 * Core complex ID is ApicId[3] for these processors.
665 */
666 per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
667 } else {
668 /*
669 * LLC ID is calculated from the number of threads sharing the
670 * cache.
671 * */
672 u32 eax, ebx, ecx, edx, num_sharing_cache = 0;
673 u32 llc_index = find_num_cache_leaves(c) - 1;
674
675 cpuid_count(0x8000001d, llc_index, &eax, &ebx, &ecx, &edx);
676 if (eax)
677 num_sharing_cache = ((eax >> 14) & 0xfff) + 1;
678
679 if (num_sharing_cache) {
680 int bits = get_count_order(num_sharing_cache);
681
682 per_cpu(cpu_llc_id, cpu) = c->apicid >> bits;
683 }
684 }
685 }
686
cacheinfo_hygon_init_llc_id(struct cpuinfo_x86 * c,int cpu,u8 node_id)687 void cacheinfo_hygon_init_llc_id(struct cpuinfo_x86 *c, int cpu, u8 node_id)
688 {
689 /*
690 * We may have multiple LLCs if L3 caches exist, so check if we
691 * have an L3 cache by looking at the L3 cache CPUID leaf.
692 */
693 if (!cpuid_edx(0x80000006))
694 return;
695
696 /*
697 * LLC is at the core complex level.
698 * Core complex ID is ApicId[3] for these processors.
699 */
700 per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
701 }
702
init_amd_cacheinfo(struct cpuinfo_x86 * c)703 void init_amd_cacheinfo(struct cpuinfo_x86 *c)
704 {
705
706 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
707 num_cache_leaves = find_num_cache_leaves(c);
708 } else if (c->extended_cpuid_level >= 0x80000006) {
709 if (cpuid_edx(0x80000006) & 0xf000)
710 num_cache_leaves = 4;
711 else
712 num_cache_leaves = 3;
713 }
714 }
715
init_hygon_cacheinfo(struct cpuinfo_x86 * c)716 void init_hygon_cacheinfo(struct cpuinfo_x86 *c)
717 {
718 num_cache_leaves = find_num_cache_leaves(c);
719 }
720
init_intel_cacheinfo(struct cpuinfo_x86 * c)721 void init_intel_cacheinfo(struct cpuinfo_x86 *c)
722 {
723 /* Cache sizes */
724 unsigned int trace = 0, l1i = 0, l1d = 0, l2 = 0, l3 = 0;
725 unsigned int new_l1d = 0, new_l1i = 0; /* Cache sizes from cpuid(4) */
726 unsigned int new_l2 = 0, new_l3 = 0, i; /* Cache sizes from cpuid(4) */
727 unsigned int l2_id = 0, l3_id = 0, num_threads_sharing, index_msb;
728 #ifdef CONFIG_SMP
729 unsigned int cpu = c->cpu_index;
730 #endif
731
732 if (c->cpuid_level > 3) {
733 static int is_initialized;
734
735 if (is_initialized == 0) {
736 /* Init num_cache_leaves from boot CPU */
737 num_cache_leaves = find_num_cache_leaves(c);
738 is_initialized++;
739 }
740
741 /*
742 * Whenever possible use cpuid(4), deterministic cache
743 * parameters cpuid leaf to find the cache details
744 */
745 for (i = 0; i < num_cache_leaves; i++) {
746 struct _cpuid4_info_regs this_leaf = {};
747 int retval;
748
749 retval = cpuid4_cache_lookup_regs(i, &this_leaf);
750 if (retval < 0)
751 continue;
752
753 switch (this_leaf.eax.split.level) {
754 case 1:
755 if (this_leaf.eax.split.type == CTYPE_DATA)
756 new_l1d = this_leaf.size/1024;
757 else if (this_leaf.eax.split.type == CTYPE_INST)
758 new_l1i = this_leaf.size/1024;
759 break;
760 case 2:
761 new_l2 = this_leaf.size/1024;
762 num_threads_sharing = 1 + this_leaf.eax.split.num_threads_sharing;
763 index_msb = get_count_order(num_threads_sharing);
764 l2_id = c->apicid & ~((1 << index_msb) - 1);
765 break;
766 case 3:
767 new_l3 = this_leaf.size/1024;
768 num_threads_sharing = 1 + this_leaf.eax.split.num_threads_sharing;
769 index_msb = get_count_order(num_threads_sharing);
770 l3_id = c->apicid & ~((1 << index_msb) - 1);
771 break;
772 default:
773 break;
774 }
775 }
776 }
777 /*
778 * Don't use cpuid2 if cpuid4 is supported. For P4, we use cpuid2 for
779 * trace cache
780 */
781 if ((num_cache_leaves == 0 || c->x86 == 15) && c->cpuid_level > 1) {
782 /* supports eax=2 call */
783 int j, n;
784 unsigned int regs[4];
785 unsigned char *dp = (unsigned char *)regs;
786 int only_trace = 0;
787
788 if (num_cache_leaves != 0 && c->x86 == 15)
789 only_trace = 1;
790
791 /* Number of times to iterate */
792 n = cpuid_eax(2) & 0xFF;
793
794 for (i = 0 ; i < n ; i++) {
795 cpuid(2, ®s[0], ®s[1], ®s[2], ®s[3]);
796
797 /* If bit 31 is set, this is an unknown format */
798 for (j = 0 ; j < 3 ; j++)
799 if (regs[j] & (1 << 31))
800 regs[j] = 0;
801
802 /* Byte 0 is level count, not a descriptor */
803 for (j = 1 ; j < 16 ; j++) {
804 unsigned char des = dp[j];
805 unsigned char k = 0;
806
807 /* look up this descriptor in the table */
808 while (cache_table[k].descriptor != 0) {
809 if (cache_table[k].descriptor == des) {
810 if (only_trace && cache_table[k].cache_type != LVL_TRACE)
811 break;
812 switch (cache_table[k].cache_type) {
813 case LVL_1_INST:
814 l1i += cache_table[k].size;
815 break;
816 case LVL_1_DATA:
817 l1d += cache_table[k].size;
818 break;
819 case LVL_2:
820 l2 += cache_table[k].size;
821 break;
822 case LVL_3:
823 l3 += cache_table[k].size;
824 break;
825 case LVL_TRACE:
826 trace += cache_table[k].size;
827 break;
828 }
829
830 break;
831 }
832
833 k++;
834 }
835 }
836 }
837 }
838
839 if (new_l1d)
840 l1d = new_l1d;
841
842 if (new_l1i)
843 l1i = new_l1i;
844
845 if (new_l2) {
846 l2 = new_l2;
847 #ifdef CONFIG_SMP
848 per_cpu(cpu_llc_id, cpu) = l2_id;
849 #endif
850 }
851
852 if (new_l3) {
853 l3 = new_l3;
854 #ifdef CONFIG_SMP
855 per_cpu(cpu_llc_id, cpu) = l3_id;
856 #endif
857 }
858
859 #ifdef CONFIG_SMP
860 /*
861 * If cpu_llc_id is not yet set, this means cpuid_level < 4 which in
862 * turns means that the only possibility is SMT (as indicated in
863 * cpuid1). Since cpuid2 doesn't specify shared caches, and we know
864 * that SMT shares all caches, we can unconditionally set cpu_llc_id to
865 * c->phys_proc_id.
866 */
867 if (per_cpu(cpu_llc_id, cpu) == BAD_APICID)
868 per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
869 #endif
870
871 c->x86_cache_size = l3 ? l3 : (l2 ? l2 : (l1i+l1d));
872
873 if (!l2)
874 cpu_detect_cache_sizes(c);
875 }
876
__cache_amd_cpumap_setup(unsigned int cpu,int index,struct _cpuid4_info_regs * base)877 static int __cache_amd_cpumap_setup(unsigned int cpu, int index,
878 struct _cpuid4_info_regs *base)
879 {
880 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
881 struct cacheinfo *this_leaf;
882 int i, sibling;
883
884 /*
885 * For L3, always use the pre-calculated cpu_llc_shared_mask
886 * to derive shared_cpu_map.
887 */
888 if (index == 3) {
889 for_each_cpu(i, cpu_llc_shared_mask(cpu)) {
890 this_cpu_ci = get_cpu_cacheinfo(i);
891 if (!this_cpu_ci->info_list)
892 continue;
893 this_leaf = this_cpu_ci->info_list + index;
894 for_each_cpu(sibling, cpu_llc_shared_mask(cpu)) {
895 if (!cpu_online(sibling))
896 continue;
897 cpumask_set_cpu(sibling,
898 &this_leaf->shared_cpu_map);
899 }
900 }
901 } else if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
902 unsigned int apicid, nshared, first, last;
903
904 nshared = base->eax.split.num_threads_sharing + 1;
905 apicid = cpu_data(cpu).apicid;
906 first = apicid - (apicid % nshared);
907 last = first + nshared - 1;
908
909 for_each_online_cpu(i) {
910 this_cpu_ci = get_cpu_cacheinfo(i);
911 if (!this_cpu_ci->info_list)
912 continue;
913
914 apicid = cpu_data(i).apicid;
915 if ((apicid < first) || (apicid > last))
916 continue;
917
918 this_leaf = this_cpu_ci->info_list + index;
919
920 for_each_online_cpu(sibling) {
921 apicid = cpu_data(sibling).apicid;
922 if ((apicid < first) || (apicid > last))
923 continue;
924 cpumask_set_cpu(sibling,
925 &this_leaf->shared_cpu_map);
926 }
927 }
928 } else
929 return 0;
930
931 return 1;
932 }
933
__cache_cpumap_setup(unsigned int cpu,int index,struct _cpuid4_info_regs * base)934 static void __cache_cpumap_setup(unsigned int cpu, int index,
935 struct _cpuid4_info_regs *base)
936 {
937 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
938 struct cacheinfo *this_leaf, *sibling_leaf;
939 unsigned long num_threads_sharing;
940 int index_msb, i;
941 struct cpuinfo_x86 *c = &cpu_data(cpu);
942
943 if (c->x86_vendor == X86_VENDOR_AMD ||
944 c->x86_vendor == X86_VENDOR_HYGON) {
945 if (__cache_amd_cpumap_setup(cpu, index, base))
946 return;
947 }
948
949 this_leaf = this_cpu_ci->info_list + index;
950 num_threads_sharing = 1 + base->eax.split.num_threads_sharing;
951
952 cpumask_set_cpu(cpu, &this_leaf->shared_cpu_map);
953 if (num_threads_sharing == 1)
954 return;
955
956 index_msb = get_count_order(num_threads_sharing);
957
958 for_each_online_cpu(i)
959 if (cpu_data(i).apicid >> index_msb == c->apicid >> index_msb) {
960 struct cpu_cacheinfo *sib_cpu_ci = get_cpu_cacheinfo(i);
961
962 if (i == cpu || !sib_cpu_ci->info_list)
963 continue;/* skip if itself or no cacheinfo */
964 sibling_leaf = sib_cpu_ci->info_list + index;
965 cpumask_set_cpu(i, &this_leaf->shared_cpu_map);
966 cpumask_set_cpu(cpu, &sibling_leaf->shared_cpu_map);
967 }
968 }
969
ci_leaf_init(struct cacheinfo * this_leaf,struct _cpuid4_info_regs * base)970 static void ci_leaf_init(struct cacheinfo *this_leaf,
971 struct _cpuid4_info_regs *base)
972 {
973 this_leaf->id = base->id;
974 this_leaf->attributes = CACHE_ID;
975 this_leaf->level = base->eax.split.level;
976 this_leaf->type = cache_type_map[base->eax.split.type];
977 this_leaf->coherency_line_size =
978 base->ebx.split.coherency_line_size + 1;
979 this_leaf->ways_of_associativity =
980 base->ebx.split.ways_of_associativity + 1;
981 this_leaf->size = base->size;
982 this_leaf->number_of_sets = base->ecx.split.number_of_sets + 1;
983 this_leaf->physical_line_partition =
984 base->ebx.split.physical_line_partition + 1;
985 this_leaf->priv = base->nb;
986 }
987
__init_cache_level(unsigned int cpu)988 static int __init_cache_level(unsigned int cpu)
989 {
990 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
991
992 if (!num_cache_leaves)
993 return -ENOENT;
994 if (!this_cpu_ci)
995 return -EINVAL;
996 this_cpu_ci->num_levels = 3;
997 this_cpu_ci->num_leaves = num_cache_leaves;
998 return 0;
999 }
1000
1001 /*
1002 * The max shared threads number comes from CPUID.4:EAX[25-14] with input
1003 * ECX as cache index. Then right shift apicid by the number's order to get
1004 * cache id for this cache node.
1005 */
get_cache_id(int cpu,struct _cpuid4_info_regs * id4_regs)1006 static void get_cache_id(int cpu, struct _cpuid4_info_regs *id4_regs)
1007 {
1008 struct cpuinfo_x86 *c = &cpu_data(cpu);
1009 unsigned long num_threads_sharing;
1010 int index_msb;
1011
1012 num_threads_sharing = 1 + id4_regs->eax.split.num_threads_sharing;
1013 index_msb = get_count_order(num_threads_sharing);
1014 id4_regs->id = c->apicid >> index_msb;
1015 }
1016
__populate_cache_leaves(unsigned int cpu)1017 static int __populate_cache_leaves(unsigned int cpu)
1018 {
1019 unsigned int idx, ret;
1020 struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
1021 struct cacheinfo *this_leaf = this_cpu_ci->info_list;
1022 struct _cpuid4_info_regs id4_regs = {};
1023
1024 for (idx = 0; idx < this_cpu_ci->num_leaves; idx++) {
1025 ret = cpuid4_cache_lookup_regs(idx, &id4_regs);
1026 if (ret)
1027 return ret;
1028 get_cache_id(cpu, &id4_regs);
1029 ci_leaf_init(this_leaf++, &id4_regs);
1030 __cache_cpumap_setup(cpu, idx, &id4_regs);
1031 }
1032 this_cpu_ci->cpu_map_populated = true;
1033
1034 return 0;
1035 }
1036
1037 DEFINE_SMP_CALL_CACHE_FUNCTION(init_cache_level)
1038 DEFINE_SMP_CALL_CACHE_FUNCTION(populate_cache_leaves)
1039
