xref: /openbmc/linux/arch/x86/kernel/cpu/cacheinfo.c (revision 828ff2ad)
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/amd_nb.h>
21 #include <asm/smp.h>
22 
23 #include "cpu.h"
24 
25 #define LVL_1_INST	1
26 #define LVL_1_DATA	2
27 #define LVL_2		3
28 #define LVL_3		4
29 #define LVL_TRACE	5
30 
31 struct _cache_table {
32 	unsigned char descriptor;
33 	char cache_type;
34 	short size;
35 };
36 
37 #define MB(x)	((x) * 1024)
38 
39 /* All the cache descriptor types we care about (no TLB or
40    trace cache entries) */
41 
42 static const struct _cache_table cache_table[] =
43 {
44 	{ 0x06, LVL_1_INST, 8 },	/* 4-way set assoc, 32 byte line size */
45 	{ 0x08, LVL_1_INST, 16 },	/* 4-way set assoc, 32 byte line size */
46 	{ 0x09, LVL_1_INST, 32 },	/* 4-way set assoc, 64 byte line size */
47 	{ 0x0a, LVL_1_DATA, 8 },	/* 2 way set assoc, 32 byte line size */
48 	{ 0x0c, LVL_1_DATA, 16 },	/* 4-way set assoc, 32 byte line size */
49 	{ 0x0d, LVL_1_DATA, 16 },	/* 4-way set assoc, 64 byte line size */
50 	{ 0x0e, LVL_1_DATA, 24 },	/* 6-way set assoc, 64 byte line size */
51 	{ 0x21, LVL_2,      256 },	/* 8-way set assoc, 64 byte line size */
52 	{ 0x22, LVL_3,      512 },	/* 4-way set assoc, sectored cache, 64 byte line size */
53 	{ 0x23, LVL_3,      MB(1) },	/* 8-way set assoc, sectored cache, 64 byte line size */
54 	{ 0x25, LVL_3,      MB(2) },	/* 8-way set assoc, sectored cache, 64 byte line size */
55 	{ 0x29, LVL_3,      MB(4) },	/* 8-way set assoc, sectored cache, 64 byte line size */
56 	{ 0x2c, LVL_1_DATA, 32 },	/* 8-way set assoc, 64 byte line size */
57 	{ 0x30, LVL_1_INST, 32 },	/* 8-way set assoc, 64 byte line size */
58 	{ 0x39, LVL_2,      128 },	/* 4-way set assoc, sectored cache, 64 byte line size */
59 	{ 0x3a, LVL_2,      192 },	/* 6-way set assoc, sectored cache, 64 byte line size */
60 	{ 0x3b, LVL_2,      128 },	/* 2-way set assoc, sectored cache, 64 byte line size */
61 	{ 0x3c, LVL_2,      256 },	/* 4-way set assoc, sectored cache, 64 byte line size */
62 	{ 0x3d, LVL_2,      384 },	/* 6-way set assoc, sectored cache, 64 byte line size */
63 	{ 0x3e, LVL_2,      512 },	/* 4-way set assoc, sectored cache, 64 byte line size */
64 	{ 0x3f, LVL_2,      256 },	/* 2-way set assoc, 64 byte line size */
65 	{ 0x41, LVL_2,      128 },	/* 4-way set assoc, 32 byte line size */
66 	{ 0x42, LVL_2,      256 },	/* 4-way set assoc, 32 byte line size */
67 	{ 0x43, LVL_2,      512 },	/* 4-way set assoc, 32 byte line size */
68 	{ 0x44, LVL_2,      MB(1) },	/* 4-way set assoc, 32 byte line size */
69 	{ 0x45, LVL_2,      MB(2) },	/* 4-way set assoc, 32 byte line size */
70 	{ 0x46, LVL_3,      MB(4) },	/* 4-way set assoc, 64 byte line size */
71 	{ 0x47, LVL_3,      MB(8) },	/* 8-way set assoc, 64 byte line size */
72 	{ 0x48, LVL_2,      MB(3) },	/* 12-way set assoc, 64 byte line size */
73 	{ 0x49, LVL_3,      MB(4) },	/* 16-way set assoc, 64 byte line size */
74 	{ 0x4a, LVL_3,      MB(6) },	/* 12-way set assoc, 64 byte line size */
75 	{ 0x4b, LVL_3,      MB(8) },	/* 16-way set assoc, 64 byte line size */
76 	{ 0x4c, LVL_3,      MB(12) },	/* 12-way set assoc, 64 byte line size */
77 	{ 0x4d, LVL_3,      MB(16) },	/* 16-way set assoc, 64 byte line size */
78 	{ 0x4e, LVL_2,      MB(6) },	/* 24-way set assoc, 64 byte line size */
79 	{ 0x60, LVL_1_DATA, 16 },	/* 8-way set assoc, sectored cache, 64 byte line size */
80 	{ 0x66, LVL_1_DATA, 8 },	/* 4-way set assoc, sectored cache, 64 byte line size */
81 	{ 0x67, LVL_1_DATA, 16 },	/* 4-way set assoc, sectored cache, 64 byte line size */
82 	{ 0x68, LVL_1_DATA, 32 },	/* 4-way set assoc, sectored cache, 64 byte line size */
83 	{ 0x70, LVL_TRACE,  12 },	/* 8-way set assoc */
84 	{ 0x71, LVL_TRACE,  16 },	/* 8-way set assoc */
85 	{ 0x72, LVL_TRACE,  32 },	/* 8-way set assoc */
86 	{ 0x73, LVL_TRACE,  64 },	/* 8-way set assoc */
87 	{ 0x78, LVL_2,      MB(1) },	/* 4-way set assoc, 64 byte line size */
88 	{ 0x79, LVL_2,      128 },	/* 8-way set assoc, sectored cache, 64 byte line size */
89 	{ 0x7a, LVL_2,      256 },	/* 8-way set assoc, sectored cache, 64 byte line size */
90 	{ 0x7b, LVL_2,      512 },	/* 8-way set assoc, sectored cache, 64 byte line size */
91 	{ 0x7c, LVL_2,      MB(1) },	/* 8-way set assoc, sectored cache, 64 byte line size */
92 	{ 0x7d, LVL_2,      MB(2) },	/* 8-way set assoc, 64 byte line size */
93 	{ 0x7f, LVL_2,      512 },	/* 2-way set assoc, 64 byte line size */
94 	{ 0x80, LVL_2,      512 },	/* 8-way set assoc, 64 byte line size */
95 	{ 0x82, LVL_2,      256 },	/* 8-way set assoc, 32 byte line size */
96 	{ 0x83, LVL_2,      512 },	/* 8-way set assoc, 32 byte line size */
97 	{ 0x84, LVL_2,      MB(1) },	/* 8-way set assoc, 32 byte line size */
98 	{ 0x85, LVL_2,      MB(2) },	/* 8-way set assoc, 32 byte line size */
99 	{ 0x86, LVL_2,      512 },	/* 4-way set assoc, 64 byte line size */
100 	{ 0x87, LVL_2,      MB(1) },	/* 8-way set assoc, 64 byte line size */
101 	{ 0xd0, LVL_3,      512 },	/* 4-way set assoc, 64 byte line size */
102 	{ 0xd1, LVL_3,      MB(1) },	/* 4-way set assoc, 64 byte line size */
103 	{ 0xd2, LVL_3,      MB(2) },	/* 4-way set assoc, 64 byte line size */
104 	{ 0xd6, LVL_3,      MB(1) },	/* 8-way set assoc, 64 byte line size */
105 	{ 0xd7, LVL_3,      MB(2) },	/* 8-way set assoc, 64 byte line size */
106 	{ 0xd8, LVL_3,      MB(4) },	/* 12-way set assoc, 64 byte line size */
107 	{ 0xdc, LVL_3,      MB(2) },	/* 12-way set assoc, 64 byte line size */
108 	{ 0xdd, LVL_3,      MB(4) },	/* 12-way set assoc, 64 byte line size */
109 	{ 0xde, LVL_3,      MB(8) },	/* 12-way set assoc, 64 byte line size */
110 	{ 0xe2, LVL_3,      MB(2) },	/* 16-way set assoc, 64 byte line size */
111 	{ 0xe3, LVL_3,      MB(4) },	/* 16-way set assoc, 64 byte line size */
112 	{ 0xe4, LVL_3,      MB(8) },	/* 16-way set assoc, 64 byte line size */
113 	{ 0xea, LVL_3,      MB(12) },	/* 24-way set assoc, 64 byte line size */
114 	{ 0xeb, LVL_3,      MB(18) },	/* 24-way set assoc, 64 byte line size */
115 	{ 0xec, LVL_3,      MB(24) },	/* 24-way set assoc, 64 byte line size */
116 	{ 0x00, 0, 0}
117 };
118 
119 
120 enum _cache_type {
121 	CTYPE_NULL = 0,
122 	CTYPE_DATA = 1,
123 	CTYPE_INST = 2,
124 	CTYPE_UNIFIED = 3
125 };
126 
127 union _cpuid4_leaf_eax {
128 	struct {
129 		enum _cache_type	type:5;
130 		unsigned int		level:3;
131 		unsigned int		is_self_initializing:1;
132 		unsigned int		is_fully_associative:1;
133 		unsigned int		reserved:4;
134 		unsigned int		num_threads_sharing:12;
135 		unsigned int		num_cores_on_die:6;
136 	} split;
137 	u32 full;
138 };
139 
140 union _cpuid4_leaf_ebx {
141 	struct {
142 		unsigned int		coherency_line_size:12;
143 		unsigned int		physical_line_partition:10;
144 		unsigned int		ways_of_associativity:10;
145 	} split;
146 	u32 full;
147 };
148 
149 union _cpuid4_leaf_ecx {
150 	struct {
151 		unsigned int		number_of_sets:32;
152 	} split;
153 	u32 full;
154 };
155 
156 struct _cpuid4_info_regs {
157 	union _cpuid4_leaf_eax eax;
158 	union _cpuid4_leaf_ebx ebx;
159 	union _cpuid4_leaf_ecx ecx;
160 	unsigned int id;
161 	unsigned long size;
162 	struct amd_northbridge *nb;
163 };
164 
165 static unsigned short num_cache_leaves;
166 
167 /* AMD doesn't have CPUID4. Emulate it here to report the same
168    information to the user.  This makes some assumptions about the machine:
169    L2 not shared, no SMT etc. that is currently true on AMD CPUs.
170 
171    In theory the TLBs could be reported as fake type (they are in "dummy").
172    Maybe later */
173 union l1_cache {
174 	struct {
175 		unsigned line_size:8;
176 		unsigned lines_per_tag:8;
177 		unsigned assoc:8;
178 		unsigned size_in_kb:8;
179 	};
180 	unsigned val;
181 };
182 
183 union l2_cache {
184 	struct {
185 		unsigned line_size:8;
186 		unsigned lines_per_tag:4;
187 		unsigned assoc:4;
188 		unsigned size_in_kb:16;
189 	};
190 	unsigned val;
191 };
192 
193 union l3_cache {
194 	struct {
195 		unsigned line_size:8;
196 		unsigned lines_per_tag:4;
197 		unsigned assoc:4;
198 		unsigned res:2;
199 		unsigned size_encoded:14;
200 	};
201 	unsigned val;
202 };
203 
204 static const unsigned short assocs[] = {
205 	[1] = 1,
206 	[2] = 2,
207 	[4] = 4,
208 	[6] = 8,
209 	[8] = 16,
210 	[0xa] = 32,
211 	[0xb] = 48,
212 	[0xc] = 64,
213 	[0xd] = 96,
214 	[0xe] = 128,
215 	[0xf] = 0xffff /* fully associative - no way to show this currently */
216 };
217 
218 static const unsigned char levels[] = { 1, 1, 2, 3 };
219 static const unsigned char types[] = { 1, 2, 3, 3 };
220 
221 static const enum cache_type cache_type_map[] = {
222 	[CTYPE_NULL] = CACHE_TYPE_NOCACHE,
223 	[CTYPE_DATA] = CACHE_TYPE_DATA,
224 	[CTYPE_INST] = CACHE_TYPE_INST,
225 	[CTYPE_UNIFIED] = CACHE_TYPE_UNIFIED,
226 };
227 
228 static void
229 amd_cpuid4(int leaf, union _cpuid4_leaf_eax *eax,
230 		     union _cpuid4_leaf_ebx *ebx,
231 		     union _cpuid4_leaf_ecx *ecx)
232 {
233 	unsigned dummy;
234 	unsigned line_size, lines_per_tag, assoc, size_in_kb;
235 	union l1_cache l1i, l1d;
236 	union l2_cache l2;
237 	union l3_cache l3;
238 	union l1_cache *l1 = &l1d;
239 
240 	eax->full = 0;
241 	ebx->full = 0;
242 	ecx->full = 0;
243 
244 	cpuid(0x80000005, &dummy, &dummy, &l1d.val, &l1i.val);
245 	cpuid(0x80000006, &dummy, &dummy, &l2.val, &l3.val);
246 
247 	switch (leaf) {
248 	case 1:
249 		l1 = &l1i;
250 	case 0:
251 		if (!l1->val)
252 			return;
253 		assoc = assocs[l1->assoc];
254 		line_size = l1->line_size;
255 		lines_per_tag = l1->lines_per_tag;
256 		size_in_kb = l1->size_in_kb;
257 		break;
258 	case 2:
259 		if (!l2.val)
260 			return;
261 		assoc = assocs[l2.assoc];
262 		line_size = l2.line_size;
263 		lines_per_tag = l2.lines_per_tag;
264 		/* cpu_data has errata corrections for K7 applied */
265 		size_in_kb = __this_cpu_read(cpu_info.x86_cache_size);
266 		break;
267 	case 3:
268 		if (!l3.val)
269 			return;
270 		assoc = assocs[l3.assoc];
271 		line_size = l3.line_size;
272 		lines_per_tag = l3.lines_per_tag;
273 		size_in_kb = l3.size_encoded * 512;
274 		if (boot_cpu_has(X86_FEATURE_AMD_DCM)) {
275 			size_in_kb = size_in_kb >> 1;
276 			assoc = assoc >> 1;
277 		}
278 		break;
279 	default:
280 		return;
281 	}
282 
283 	eax->split.is_self_initializing = 1;
284 	eax->split.type = types[leaf];
285 	eax->split.level = levels[leaf];
286 	eax->split.num_threads_sharing = 0;
287 	eax->split.num_cores_on_die = __this_cpu_read(cpu_info.x86_max_cores) - 1;
288 
289 
290 	if (assoc == 0xffff)
291 		eax->split.is_fully_associative = 1;
292 	ebx->split.coherency_line_size = line_size - 1;
293 	ebx->split.ways_of_associativity = assoc - 1;
294 	ebx->split.physical_line_partition = lines_per_tag - 1;
295 	ecx->split.number_of_sets = (size_in_kb * 1024) / line_size /
296 		(ebx->split.ways_of_associativity + 1) - 1;
297 }
298 
299 #if defined(CONFIG_AMD_NB) && defined(CONFIG_SYSFS)
300 
301 /*
302  * L3 cache descriptors
303  */
304 static void amd_calc_l3_indices(struct amd_northbridge *nb)
305 {
306 	struct amd_l3_cache *l3 = &nb->l3_cache;
307 	unsigned int sc0, sc1, sc2, sc3;
308 	u32 val = 0;
309 
310 	pci_read_config_dword(nb->misc, 0x1C4, &val);
311 
312 	/* calculate subcache sizes */
313 	l3->subcaches[0] = sc0 = !(val & BIT(0));
314 	l3->subcaches[1] = sc1 = !(val & BIT(4));
315 
316 	if (boot_cpu_data.x86 == 0x15) {
317 		l3->subcaches[0] = sc0 += !(val & BIT(1));
318 		l3->subcaches[1] = sc1 += !(val & BIT(5));
319 	}
320 
321 	l3->subcaches[2] = sc2 = !(val & BIT(8))  + !(val & BIT(9));
322 	l3->subcaches[3] = sc3 = !(val & BIT(12)) + !(val & BIT(13));
323 
324 	l3->indices = (max(max3(sc0, sc1, sc2), sc3) << 10) - 1;
325 }
326 
327 /*
328  * check whether a slot used for disabling an L3 index is occupied.
329  * @l3: L3 cache descriptor
330  * @slot: slot number (0..1)
331  *
332  * @returns: the disabled index if used or negative value if slot free.
333  */
334 static int amd_get_l3_disable_slot(struct amd_northbridge *nb, unsigned slot)
335 {
336 	unsigned int reg = 0;
337 
338 	pci_read_config_dword(nb->misc, 0x1BC + slot * 4, &reg);
339 
340 	/* check whether this slot is activated already */
341 	if (reg & (3UL << 30))
342 		return reg & 0xfff;
343 
344 	return -1;
345 }
346 
347 static ssize_t show_cache_disable(struct cacheinfo *this_leaf, char *buf,
348 				  unsigned int slot)
349 {
350 	int index;
351 	struct amd_northbridge *nb = this_leaf->priv;
352 
353 	index = amd_get_l3_disable_slot(nb, slot);
354 	if (index >= 0)
355 		return sprintf(buf, "%d\n", index);
356 
357 	return sprintf(buf, "FREE\n");
358 }
359 
360 #define SHOW_CACHE_DISABLE(slot)					\
361 static ssize_t								\
362 cache_disable_##slot##_show(struct device *dev,				\
363 			    struct device_attribute *attr, char *buf)	\
364 {									\
365 	struct cacheinfo *this_leaf = dev_get_drvdata(dev);		\
366 	return show_cache_disable(this_leaf, buf, slot);		\
367 }
368 SHOW_CACHE_DISABLE(0)
369 SHOW_CACHE_DISABLE(1)
370 
371 static void amd_l3_disable_index(struct amd_northbridge *nb, int cpu,
372 				 unsigned slot, unsigned long idx)
373 {
374 	int i;
375 
376 	idx |= BIT(30);
377 
378 	/*
379 	 *  disable index in all 4 subcaches
380 	 */
381 	for (i = 0; i < 4; i++) {
382 		u32 reg = idx | (i << 20);
383 
384 		if (!nb->l3_cache.subcaches[i])
385 			continue;
386 
387 		pci_write_config_dword(nb->misc, 0x1BC + slot * 4, reg);
388 
389 		/*
390 		 * We need to WBINVD on a core on the node containing the L3
391 		 * cache which indices we disable therefore a simple wbinvd()
392 		 * is not sufficient.
393 		 */
394 		wbinvd_on_cpu(cpu);
395 
396 		reg |= BIT(31);
397 		pci_write_config_dword(nb->misc, 0x1BC + slot * 4, reg);
398 	}
399 }
400 
401 /*
402  * disable a L3 cache index by using a disable-slot
403  *
404  * @l3:    L3 cache descriptor
405  * @cpu:   A CPU on the node containing the L3 cache
406  * @slot:  slot number (0..1)
407  * @index: index to disable
408  *
409  * @return: 0 on success, error status on failure
410  */
411 static int amd_set_l3_disable_slot(struct amd_northbridge *nb, int cpu,
412 			    unsigned slot, unsigned long index)
413 {
414 	int ret = 0;
415 
416 	/*  check if @slot is already used or the index is already disabled */
417 	ret = amd_get_l3_disable_slot(nb, slot);
418 	if (ret >= 0)
419 		return -EEXIST;
420 
421 	if (index > nb->l3_cache.indices)
422 		return -EINVAL;
423 
424 	/* check whether the other slot has disabled the same index already */
425 	if (index == amd_get_l3_disable_slot(nb, !slot))
426 		return -EEXIST;
427 
428 	amd_l3_disable_index(nb, cpu, slot, index);
429 
430 	return 0;
431 }
432 
433 static ssize_t store_cache_disable(struct cacheinfo *this_leaf,
434 				   const char *buf, size_t count,
435 				   unsigned int slot)
436 {
437 	unsigned long val = 0;
438 	int cpu, err = 0;
439 	struct amd_northbridge *nb = this_leaf->priv;
440 
441 	if (!capable(CAP_SYS_ADMIN))
442 		return -EPERM;
443 
444 	cpu = cpumask_first(&this_leaf->shared_cpu_map);
445 
446 	if (kstrtoul(buf, 10, &val) < 0)
447 		return -EINVAL;
448 
449 	err = amd_set_l3_disable_slot(nb, cpu, slot, val);
450 	if (err) {
451 		if (err == -EEXIST)
452 			pr_warn("L3 slot %d in use/index already disabled!\n",
453 				   slot);
454 		return err;
455 	}
456 	return count;
457 }
458 
459 #define STORE_CACHE_DISABLE(slot)					\
460 static ssize_t								\
461 cache_disable_##slot##_store(struct device *dev,			\
462 			     struct device_attribute *attr,		\
463 			     const char *buf, size_t count)		\
464 {									\
465 	struct cacheinfo *this_leaf = dev_get_drvdata(dev);		\
466 	return store_cache_disable(this_leaf, buf, count, slot);	\
467 }
468 STORE_CACHE_DISABLE(0)
469 STORE_CACHE_DISABLE(1)
470 
471 static ssize_t subcaches_show(struct device *dev,
472 			      struct device_attribute *attr, char *buf)
473 {
474 	struct cacheinfo *this_leaf = dev_get_drvdata(dev);
475 	int cpu = cpumask_first(&this_leaf->shared_cpu_map);
476 
477 	return sprintf(buf, "%x\n", amd_get_subcaches(cpu));
478 }
479 
480 static ssize_t subcaches_store(struct device *dev,
481 			       struct device_attribute *attr,
482 			       const char *buf, size_t count)
483 {
484 	struct cacheinfo *this_leaf = dev_get_drvdata(dev);
485 	int cpu = cpumask_first(&this_leaf->shared_cpu_map);
486 	unsigned long val;
487 
488 	if (!capable(CAP_SYS_ADMIN))
489 		return -EPERM;
490 
491 	if (kstrtoul(buf, 16, &val) < 0)
492 		return -EINVAL;
493 
494 	if (amd_set_subcaches(cpu, val))
495 		return -EINVAL;
496 
497 	return count;
498 }
499 
500 static DEVICE_ATTR_RW(cache_disable_0);
501 static DEVICE_ATTR_RW(cache_disable_1);
502 static DEVICE_ATTR_RW(subcaches);
503 
504 static umode_t
505 cache_private_attrs_is_visible(struct kobject *kobj,
506 			       struct attribute *attr, int unused)
507 {
508 	struct device *dev = kobj_to_dev(kobj);
509 	struct cacheinfo *this_leaf = dev_get_drvdata(dev);
510 	umode_t mode = attr->mode;
511 
512 	if (!this_leaf->priv)
513 		return 0;
514 
515 	if ((attr == &dev_attr_subcaches.attr) &&
516 	    amd_nb_has_feature(AMD_NB_L3_PARTITIONING))
517 		return mode;
518 
519 	if ((attr == &dev_attr_cache_disable_0.attr ||
520 	     attr == &dev_attr_cache_disable_1.attr) &&
521 	    amd_nb_has_feature(AMD_NB_L3_INDEX_DISABLE))
522 		return mode;
523 
524 	return 0;
525 }
526 
527 static struct attribute_group cache_private_group = {
528 	.is_visible = cache_private_attrs_is_visible,
529 };
530 
531 static void init_amd_l3_attrs(void)
532 {
533 	int n = 1;
534 	static struct attribute **amd_l3_attrs;
535 
536 	if (amd_l3_attrs) /* already initialized */
537 		return;
538 
539 	if (amd_nb_has_feature(AMD_NB_L3_INDEX_DISABLE))
540 		n += 2;
541 	if (amd_nb_has_feature(AMD_NB_L3_PARTITIONING))
542 		n += 1;
543 
544 	amd_l3_attrs = kcalloc(n, sizeof(*amd_l3_attrs), GFP_KERNEL);
545 	if (!amd_l3_attrs)
546 		return;
547 
548 	n = 0;
549 	if (amd_nb_has_feature(AMD_NB_L3_INDEX_DISABLE)) {
550 		amd_l3_attrs[n++] = &dev_attr_cache_disable_0.attr;
551 		amd_l3_attrs[n++] = &dev_attr_cache_disable_1.attr;
552 	}
553 	if (amd_nb_has_feature(AMD_NB_L3_PARTITIONING))
554 		amd_l3_attrs[n++] = &dev_attr_subcaches.attr;
555 
556 	cache_private_group.attrs = amd_l3_attrs;
557 }
558 
559 const struct attribute_group *
560 cache_get_priv_group(struct cacheinfo *this_leaf)
561 {
562 	struct amd_northbridge *nb = this_leaf->priv;
563 
564 	if (this_leaf->level < 3 || !nb)
565 		return NULL;
566 
567 	if (nb && nb->l3_cache.indices)
568 		init_amd_l3_attrs();
569 
570 	return &cache_private_group;
571 }
572 
573 static void amd_init_l3_cache(struct _cpuid4_info_regs *this_leaf, int index)
574 {
575 	int node;
576 
577 	/* only for L3, and not in virtualized environments */
578 	if (index < 3)
579 		return;
580 
581 	node = amd_get_nb_id(smp_processor_id());
582 	this_leaf->nb = node_to_amd_nb(node);
583 	if (this_leaf->nb && !this_leaf->nb->l3_cache.indices)
584 		amd_calc_l3_indices(this_leaf->nb);
585 }
586 #else
587 #define amd_init_l3_cache(x, y)
588 #endif  /* CONFIG_AMD_NB && CONFIG_SYSFS */
589 
590 static int
591 cpuid4_cache_lookup_regs(int index, struct _cpuid4_info_regs *this_leaf)
592 {
593 	union _cpuid4_leaf_eax	eax;
594 	union _cpuid4_leaf_ebx	ebx;
595 	union _cpuid4_leaf_ecx	ecx;
596 	unsigned		edx;
597 
598 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
599 		if (boot_cpu_has(X86_FEATURE_TOPOEXT))
600 			cpuid_count(0x8000001d, index, &eax.full,
601 				    &ebx.full, &ecx.full, &edx);
602 		else
603 			amd_cpuid4(index, &eax, &ebx, &ecx);
604 		amd_init_l3_cache(this_leaf, index);
605 	} else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) {
606 		cpuid_count(0x8000001d, index, &eax.full,
607 			    &ebx.full, &ecx.full, &edx);
608 		amd_init_l3_cache(this_leaf, index);
609 	} else {
610 		cpuid_count(4, index, &eax.full, &ebx.full, &ecx.full, &edx);
611 	}
612 
613 	if (eax.split.type == CTYPE_NULL)
614 		return -EIO; /* better error ? */
615 
616 	this_leaf->eax = eax;
617 	this_leaf->ebx = ebx;
618 	this_leaf->ecx = ecx;
619 	this_leaf->size = (ecx.split.number_of_sets          + 1) *
620 			  (ebx.split.coherency_line_size     + 1) *
621 			  (ebx.split.physical_line_partition + 1) *
622 			  (ebx.split.ways_of_associativity   + 1);
623 	return 0;
624 }
625 
626 static int find_num_cache_leaves(struct cpuinfo_x86 *c)
627 {
628 	unsigned int		eax, ebx, ecx, edx, op;
629 	union _cpuid4_leaf_eax	cache_eax;
630 	int 			i = -1;
631 
632 	if (c->x86_vendor == X86_VENDOR_AMD ||
633 	    c->x86_vendor == X86_VENDOR_HYGON)
634 		op = 0x8000001d;
635 	else
636 		op = 4;
637 
638 	do {
639 		++i;
640 		/* Do cpuid(op) loop to find out num_cache_leaves */
641 		cpuid_count(op, i, &eax, &ebx, &ecx, &edx);
642 		cache_eax.full = eax;
643 	} while (cache_eax.split.type != CTYPE_NULL);
644 	return i;
645 }
646 
647 void cacheinfo_amd_init_llc_id(struct cpuinfo_x86 *c, int cpu, u8 node_id)
648 {
649 	/*
650 	 * We may have multiple LLCs if L3 caches exist, so check if we
651 	 * have an L3 cache by looking at the L3 cache CPUID leaf.
652 	 */
653 	if (!cpuid_edx(0x80000006))
654 		return;
655 
656 	if (c->x86 < 0x17) {
657 		/* LLC is at the node level. */
658 		per_cpu(cpu_llc_id, cpu) = node_id;
659 	} else if (c->x86 == 0x17 &&
660 		   c->x86_model >= 0 && c->x86_model <= 0x1F) {
661 		/*
662 		 * LLC is at the core complex level.
663 		 * Core complex ID is ApicId[3] for these processors.
664 		 */
665 		per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
666 	} else {
667 		/*
668 		 * LLC ID is calculated from the number of threads sharing the
669 		 * cache.
670 		 * */
671 		u32 eax, ebx, ecx, edx, num_sharing_cache = 0;
672 		u32 llc_index = find_num_cache_leaves(c) - 1;
673 
674 		cpuid_count(0x8000001d, llc_index, &eax, &ebx, &ecx, &edx);
675 		if (eax)
676 			num_sharing_cache = ((eax >> 14) & 0xfff) + 1;
677 
678 		if (num_sharing_cache) {
679 			int bits = get_count_order(num_sharing_cache);
680 
681 			per_cpu(cpu_llc_id, cpu) = c->apicid >> bits;
682 		}
683 	}
684 }
685 
686 void cacheinfo_hygon_init_llc_id(struct cpuinfo_x86 *c, int cpu, u8 node_id)
687 {
688 	/*
689 	 * We may have multiple LLCs if L3 caches exist, so check if we
690 	 * have an L3 cache by looking at the L3 cache CPUID leaf.
691 	 */
692 	if (!cpuid_edx(0x80000006))
693 		return;
694 
695 	/*
696 	 * LLC is at the core complex level.
697 	 * Core complex ID is ApicId[3] for these processors.
698 	 */
699 	per_cpu(cpu_llc_id, cpu) = c->apicid >> 3;
700 }
701 
702 void init_amd_cacheinfo(struct cpuinfo_x86 *c)
703 {
704 
705 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
706 		num_cache_leaves = find_num_cache_leaves(c);
707 	} else if (c->extended_cpuid_level >= 0x80000006) {
708 		if (cpuid_edx(0x80000006) & 0xf000)
709 			num_cache_leaves = 4;
710 		else
711 			num_cache_leaves = 3;
712 	}
713 }
714 
715 void init_hygon_cacheinfo(struct cpuinfo_x86 *c)
716 {
717 	num_cache_leaves = find_num_cache_leaves(c);
718 }
719 
720 void init_intel_cacheinfo(struct cpuinfo_x86 *c)
721 {
722 	/* Cache sizes */
723 	unsigned int trace = 0, l1i = 0, l1d = 0, l2 = 0, l3 = 0;
724 	unsigned int new_l1d = 0, new_l1i = 0; /* Cache sizes from cpuid(4) */
725 	unsigned int new_l2 = 0, new_l3 = 0, i; /* Cache sizes from cpuid(4) */
726 	unsigned int l2_id = 0, l3_id = 0, num_threads_sharing, index_msb;
727 #ifdef CONFIG_SMP
728 	unsigned int cpu = c->cpu_index;
729 #endif
730 
731 	if (c->cpuid_level > 3) {
732 		static int is_initialized;
733 
734 		if (is_initialized == 0) {
735 			/* Init num_cache_leaves from boot CPU */
736 			num_cache_leaves = find_num_cache_leaves(c);
737 			is_initialized++;
738 		}
739 
740 		/*
741 		 * Whenever possible use cpuid(4), deterministic cache
742 		 * parameters cpuid leaf to find the cache details
743 		 */
744 		for (i = 0; i < num_cache_leaves; i++) {
745 			struct _cpuid4_info_regs this_leaf = {};
746 			int retval;
747 
748 			retval = cpuid4_cache_lookup_regs(i, &this_leaf);
749 			if (retval < 0)
750 				continue;
751 
752 			switch (this_leaf.eax.split.level) {
753 			case 1:
754 				if (this_leaf.eax.split.type == CTYPE_DATA)
755 					new_l1d = this_leaf.size/1024;
756 				else if (this_leaf.eax.split.type == CTYPE_INST)
757 					new_l1i = this_leaf.size/1024;
758 				break;
759 			case 2:
760 				new_l2 = this_leaf.size/1024;
761 				num_threads_sharing = 1 + this_leaf.eax.split.num_threads_sharing;
762 				index_msb = get_count_order(num_threads_sharing);
763 				l2_id = c->apicid & ~((1 << index_msb) - 1);
764 				break;
765 			case 3:
766 				new_l3 = this_leaf.size/1024;
767 				num_threads_sharing = 1 + this_leaf.eax.split.num_threads_sharing;
768 				index_msb = get_count_order(num_threads_sharing);
769 				l3_id = c->apicid & ~((1 << index_msb) - 1);
770 				break;
771 			default:
772 				break;
773 			}
774 		}
775 	}
776 	/*
777 	 * Don't use cpuid2 if cpuid4 is supported. For P4, we use cpuid2 for
778 	 * trace cache
779 	 */
780 	if ((num_cache_leaves == 0 || c->x86 == 15) && c->cpuid_level > 1) {
781 		/* supports eax=2  call */
782 		int j, n;
783 		unsigned int regs[4];
784 		unsigned char *dp = (unsigned char *)regs;
785 		int only_trace = 0;
786 
787 		if (num_cache_leaves != 0 && c->x86 == 15)
788 			only_trace = 1;
789 
790 		/* Number of times to iterate */
791 		n = cpuid_eax(2) & 0xFF;
792 
793 		for (i = 0 ; i < n ; i++) {
794 			cpuid(2, &regs[0], &regs[1], &regs[2], &regs[3]);
795 
796 			/* If bit 31 is set, this is an unknown format */
797 			for (j = 0 ; j < 3 ; j++)
798 				if (regs[j] & (1 << 31))
799 					regs[j] = 0;
800 
801 			/* Byte 0 is level count, not a descriptor */
802 			for (j = 1 ; j < 16 ; j++) {
803 				unsigned char des = dp[j];
804 				unsigned char k = 0;
805 
806 				/* look up this descriptor in the table */
807 				while (cache_table[k].descriptor != 0) {
808 					if (cache_table[k].descriptor == des) {
809 						if (only_trace && cache_table[k].cache_type != LVL_TRACE)
810 							break;
811 						switch (cache_table[k].cache_type) {
812 						case LVL_1_INST:
813 							l1i += cache_table[k].size;
814 							break;
815 						case LVL_1_DATA:
816 							l1d += cache_table[k].size;
817 							break;
818 						case LVL_2:
819 							l2 += cache_table[k].size;
820 							break;
821 						case LVL_3:
822 							l3 += cache_table[k].size;
823 							break;
824 						case LVL_TRACE:
825 							trace += cache_table[k].size;
826 							break;
827 						}
828 
829 						break;
830 					}
831 
832 					k++;
833 				}
834 			}
835 		}
836 	}
837 
838 	if (new_l1d)
839 		l1d = new_l1d;
840 
841 	if (new_l1i)
842 		l1i = new_l1i;
843 
844 	if (new_l2) {
845 		l2 = new_l2;
846 #ifdef CONFIG_SMP
847 		per_cpu(cpu_llc_id, cpu) = l2_id;
848 #endif
849 	}
850 
851 	if (new_l3) {
852 		l3 = new_l3;
853 #ifdef CONFIG_SMP
854 		per_cpu(cpu_llc_id, cpu) = l3_id;
855 #endif
856 	}
857 
858 #ifdef CONFIG_SMP
859 	/*
860 	 * If cpu_llc_id is not yet set, this means cpuid_level < 4 which in
861 	 * turns means that the only possibility is SMT (as indicated in
862 	 * cpuid1). Since cpuid2 doesn't specify shared caches, and we know
863 	 * that SMT shares all caches, we can unconditionally set cpu_llc_id to
864 	 * c->phys_proc_id.
865 	 */
866 	if (per_cpu(cpu_llc_id, cpu) == BAD_APICID)
867 		per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
868 #endif
869 
870 	c->x86_cache_size = l3 ? l3 : (l2 ? l2 : (l1i+l1d));
871 
872 	if (!l2)
873 		cpu_detect_cache_sizes(c);
874 }
875 
876 static int __cache_amd_cpumap_setup(unsigned int cpu, int index,
877 				    struct _cpuid4_info_regs *base)
878 {
879 	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
880 	struct cacheinfo *this_leaf;
881 	int i, sibling;
882 
883 	/*
884 	 * For L3, always use the pre-calculated cpu_llc_shared_mask
885 	 * to derive shared_cpu_map.
886 	 */
887 	if (index == 3) {
888 		for_each_cpu(i, cpu_llc_shared_mask(cpu)) {
889 			this_cpu_ci = get_cpu_cacheinfo(i);
890 			if (!this_cpu_ci->info_list)
891 				continue;
892 			this_leaf = this_cpu_ci->info_list + index;
893 			for_each_cpu(sibling, cpu_llc_shared_mask(cpu)) {
894 				if (!cpu_online(sibling))
895 					continue;
896 				cpumask_set_cpu(sibling,
897 						&this_leaf->shared_cpu_map);
898 			}
899 		}
900 	} else if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
901 		unsigned int apicid, nshared, first, last;
902 
903 		nshared = base->eax.split.num_threads_sharing + 1;
904 		apicid = cpu_data(cpu).apicid;
905 		first = apicid - (apicid % nshared);
906 		last = first + nshared - 1;
907 
908 		for_each_online_cpu(i) {
909 			this_cpu_ci = get_cpu_cacheinfo(i);
910 			if (!this_cpu_ci->info_list)
911 				continue;
912 
913 			apicid = cpu_data(i).apicid;
914 			if ((apicid < first) || (apicid > last))
915 				continue;
916 
917 			this_leaf = this_cpu_ci->info_list + index;
918 
919 			for_each_online_cpu(sibling) {
920 				apicid = cpu_data(sibling).apicid;
921 				if ((apicid < first) || (apicid > last))
922 					continue;
923 				cpumask_set_cpu(sibling,
924 						&this_leaf->shared_cpu_map);
925 			}
926 		}
927 	} else
928 		return 0;
929 
930 	return 1;
931 }
932 
933 static void __cache_cpumap_setup(unsigned int cpu, int index,
934 				 struct _cpuid4_info_regs *base)
935 {
936 	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
937 	struct cacheinfo *this_leaf, *sibling_leaf;
938 	unsigned long num_threads_sharing;
939 	int index_msb, i;
940 	struct cpuinfo_x86 *c = &cpu_data(cpu);
941 
942 	if (c->x86_vendor == X86_VENDOR_AMD ||
943 	    c->x86_vendor == X86_VENDOR_HYGON) {
944 		if (__cache_amd_cpumap_setup(cpu, index, base))
945 			return;
946 	}
947 
948 	this_leaf = this_cpu_ci->info_list + index;
949 	num_threads_sharing = 1 + base->eax.split.num_threads_sharing;
950 
951 	cpumask_set_cpu(cpu, &this_leaf->shared_cpu_map);
952 	if (num_threads_sharing == 1)
953 		return;
954 
955 	index_msb = get_count_order(num_threads_sharing);
956 
957 	for_each_online_cpu(i)
958 		if (cpu_data(i).apicid >> index_msb == c->apicid >> index_msb) {
959 			struct cpu_cacheinfo *sib_cpu_ci = get_cpu_cacheinfo(i);
960 
961 			if (i == cpu || !sib_cpu_ci->info_list)
962 				continue;/* skip if itself or no cacheinfo */
963 			sibling_leaf = sib_cpu_ci->info_list + index;
964 			cpumask_set_cpu(i, &this_leaf->shared_cpu_map);
965 			cpumask_set_cpu(cpu, &sibling_leaf->shared_cpu_map);
966 		}
967 }
968 
969 static void ci_leaf_init(struct cacheinfo *this_leaf,
970 			 struct _cpuid4_info_regs *base)
971 {
972 	this_leaf->id = base->id;
973 	this_leaf->attributes = CACHE_ID;
974 	this_leaf->level = base->eax.split.level;
975 	this_leaf->type = cache_type_map[base->eax.split.type];
976 	this_leaf->coherency_line_size =
977 				base->ebx.split.coherency_line_size + 1;
978 	this_leaf->ways_of_associativity =
979 				base->ebx.split.ways_of_associativity + 1;
980 	this_leaf->size = base->size;
981 	this_leaf->number_of_sets = base->ecx.split.number_of_sets + 1;
982 	this_leaf->physical_line_partition =
983 				base->ebx.split.physical_line_partition + 1;
984 	this_leaf->priv = base->nb;
985 }
986 
987 static int __init_cache_level(unsigned int cpu)
988 {
989 	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
990 
991 	if (!num_cache_leaves)
992 		return -ENOENT;
993 	if (!this_cpu_ci)
994 		return -EINVAL;
995 	this_cpu_ci->num_levels = 3;
996 	this_cpu_ci->num_leaves = num_cache_leaves;
997 	return 0;
998 }
999 
1000 /*
1001  * The max shared threads number comes from CPUID.4:EAX[25-14] with input
1002  * ECX as cache index. Then right shift apicid by the number's order to get
1003  * cache id for this cache node.
1004  */
1005 static void get_cache_id(int cpu, struct _cpuid4_info_regs *id4_regs)
1006 {
1007 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1008 	unsigned long num_threads_sharing;
1009 	int index_msb;
1010 
1011 	num_threads_sharing = 1 + id4_regs->eax.split.num_threads_sharing;
1012 	index_msb = get_count_order(num_threads_sharing);
1013 	id4_regs->id = c->apicid >> index_msb;
1014 }
1015 
1016 static int __populate_cache_leaves(unsigned int cpu)
1017 {
1018 	unsigned int idx, ret;
1019 	struct cpu_cacheinfo *this_cpu_ci = get_cpu_cacheinfo(cpu);
1020 	struct cacheinfo *this_leaf = this_cpu_ci->info_list;
1021 	struct _cpuid4_info_regs id4_regs = {};
1022 
1023 	for (idx = 0; idx < this_cpu_ci->num_leaves; idx++) {
1024 		ret = cpuid4_cache_lookup_regs(idx, &id4_regs);
1025 		if (ret)
1026 			return ret;
1027 		get_cache_id(cpu, &id4_regs);
1028 		ci_leaf_init(this_leaf++, &id4_regs);
1029 		__cache_cpumap_setup(cpu, idx, &id4_regs);
1030 	}
1031 	this_cpu_ci->cpu_map_populated = true;
1032 
1033 	return 0;
1034 }
1035 
1036 DEFINE_SMP_CALL_CACHE_FUNCTION(init_cache_level)
1037 DEFINE_SMP_CALL_CACHE_FUNCTION(populate_cache_leaves)
1038