xref: /openbmc/linux/arch/x86/events/amd/uncore.c (revision 7ff94669)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Advanced Micro Devices, Inc.
4  *
5  * Author: Jacob Shin <jacob.shin@amd.com>
6  */
7 
8 #include <linux/perf_event.h>
9 #include <linux/percpu.h>
10 #include <linux/types.h>
11 #include <linux/slab.h>
12 #include <linux/init.h>
13 #include <linux/cpu.h>
14 #include <linux/cpumask.h>
15 
16 #include <asm/cpufeature.h>
17 #include <asm/perf_event.h>
18 #include <asm/msr.h>
19 #include <asm/smp.h>
20 
21 #define NUM_COUNTERS_NB		4
22 #define NUM_COUNTERS_L2		4
23 #define NUM_COUNTERS_L3		6
24 #define MAX_COUNTERS		6
25 
26 #define RDPMC_BASE_NB		6
27 #define RDPMC_BASE_LLC		10
28 
29 #define COUNTER_SHIFT		16
30 
31 #undef pr_fmt
32 #define pr_fmt(fmt)	"amd_uncore: " fmt
33 
34 static int num_counters_llc;
35 static int num_counters_nb;
36 static bool l3_mask;
37 
38 static HLIST_HEAD(uncore_unused_list);
39 
40 struct amd_uncore {
41 	int id;
42 	int refcnt;
43 	int cpu;
44 	int num_counters;
45 	int rdpmc_base;
46 	u32 msr_base;
47 	cpumask_t *active_mask;
48 	struct pmu *pmu;
49 	struct perf_event *events[MAX_COUNTERS];
50 	struct hlist_node node;
51 };
52 
53 static struct amd_uncore * __percpu *amd_uncore_nb;
54 static struct amd_uncore * __percpu *amd_uncore_llc;
55 
56 static struct pmu amd_nb_pmu;
57 static struct pmu amd_llc_pmu;
58 
59 static cpumask_t amd_nb_active_mask;
60 static cpumask_t amd_llc_active_mask;
61 
62 static bool is_nb_event(struct perf_event *event)
63 {
64 	return event->pmu->type == amd_nb_pmu.type;
65 }
66 
67 static bool is_llc_event(struct perf_event *event)
68 {
69 	return event->pmu->type == amd_llc_pmu.type;
70 }
71 
72 static struct amd_uncore *event_to_amd_uncore(struct perf_event *event)
73 {
74 	if (is_nb_event(event) && amd_uncore_nb)
75 		return *per_cpu_ptr(amd_uncore_nb, event->cpu);
76 	else if (is_llc_event(event) && amd_uncore_llc)
77 		return *per_cpu_ptr(amd_uncore_llc, event->cpu);
78 
79 	return NULL;
80 }
81 
82 static void amd_uncore_read(struct perf_event *event)
83 {
84 	struct hw_perf_event *hwc = &event->hw;
85 	u64 prev, new;
86 	s64 delta;
87 
88 	/*
89 	 * since we do not enable counter overflow interrupts,
90 	 * we do not have to worry about prev_count changing on us
91 	 */
92 
93 	prev = local64_read(&hwc->prev_count);
94 	rdpmcl(hwc->event_base_rdpmc, new);
95 	local64_set(&hwc->prev_count, new);
96 	delta = (new << COUNTER_SHIFT) - (prev << COUNTER_SHIFT);
97 	delta >>= COUNTER_SHIFT;
98 	local64_add(delta, &event->count);
99 }
100 
101 static void amd_uncore_start(struct perf_event *event, int flags)
102 {
103 	struct hw_perf_event *hwc = &event->hw;
104 
105 	if (flags & PERF_EF_RELOAD)
106 		wrmsrl(hwc->event_base, (u64)local64_read(&hwc->prev_count));
107 
108 	hwc->state = 0;
109 	wrmsrl(hwc->config_base, (hwc->config | ARCH_PERFMON_EVENTSEL_ENABLE));
110 	perf_event_update_userpage(event);
111 }
112 
113 static void amd_uncore_stop(struct perf_event *event, int flags)
114 {
115 	struct hw_perf_event *hwc = &event->hw;
116 
117 	wrmsrl(hwc->config_base, hwc->config);
118 	hwc->state |= PERF_HES_STOPPED;
119 
120 	if ((flags & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) {
121 		amd_uncore_read(event);
122 		hwc->state |= PERF_HES_UPTODATE;
123 	}
124 }
125 
126 static int amd_uncore_add(struct perf_event *event, int flags)
127 {
128 	int i;
129 	struct amd_uncore *uncore = event_to_amd_uncore(event);
130 	struct hw_perf_event *hwc = &event->hw;
131 
132 	/* are we already assigned? */
133 	if (hwc->idx != -1 && uncore->events[hwc->idx] == event)
134 		goto out;
135 
136 	for (i = 0; i < uncore->num_counters; i++) {
137 		if (uncore->events[i] == event) {
138 			hwc->idx = i;
139 			goto out;
140 		}
141 	}
142 
143 	/* if not, take the first available counter */
144 	hwc->idx = -1;
145 	for (i = 0; i < uncore->num_counters; i++) {
146 		if (cmpxchg(&uncore->events[i], NULL, event) == NULL) {
147 			hwc->idx = i;
148 			break;
149 		}
150 	}
151 
152 out:
153 	if (hwc->idx == -1)
154 		return -EBUSY;
155 
156 	hwc->config_base = uncore->msr_base + (2 * hwc->idx);
157 	hwc->event_base = uncore->msr_base + 1 + (2 * hwc->idx);
158 	hwc->event_base_rdpmc = uncore->rdpmc_base + hwc->idx;
159 	hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
160 
161 	if (flags & PERF_EF_START)
162 		amd_uncore_start(event, PERF_EF_RELOAD);
163 
164 	return 0;
165 }
166 
167 static void amd_uncore_del(struct perf_event *event, int flags)
168 {
169 	int i;
170 	struct amd_uncore *uncore = event_to_amd_uncore(event);
171 	struct hw_perf_event *hwc = &event->hw;
172 
173 	amd_uncore_stop(event, PERF_EF_UPDATE);
174 
175 	for (i = 0; i < uncore->num_counters; i++) {
176 		if (cmpxchg(&uncore->events[i], event, NULL) == event)
177 			break;
178 	}
179 
180 	hwc->idx = -1;
181 }
182 
183 /*
184  * Return a full thread and slice mask unless user
185  * has provided them
186  */
187 static u64 l3_thread_slice_mask(u64 config)
188 {
189 	if (boot_cpu_data.x86 <= 0x18)
190 		return ((config & AMD64_L3_SLICE_MASK) ? : AMD64_L3_SLICE_MASK) |
191 		       ((config & AMD64_L3_THREAD_MASK) ? : AMD64_L3_THREAD_MASK);
192 
193 	/*
194 	 * If the user doesn't specify a threadmask, they're not trying to
195 	 * count core 0, so we enable all cores & threads.
196 	 * We'll also assume that they want to count slice 0 if they specify
197 	 * a threadmask and leave sliceid and enallslices unpopulated.
198 	 */
199 	if (!(config & AMD64_L3_F19H_THREAD_MASK))
200 		return AMD64_L3_F19H_THREAD_MASK | AMD64_L3_EN_ALL_SLICES |
201 		       AMD64_L3_EN_ALL_CORES;
202 
203 	return config & (AMD64_L3_F19H_THREAD_MASK | AMD64_L3_SLICEID_MASK |
204 			 AMD64_L3_EN_ALL_CORES | AMD64_L3_EN_ALL_SLICES |
205 			 AMD64_L3_COREID_MASK);
206 }
207 
208 static int amd_uncore_event_init(struct perf_event *event)
209 {
210 	struct amd_uncore *uncore;
211 	struct hw_perf_event *hwc = &event->hw;
212 
213 	if (event->attr.type != event->pmu->type)
214 		return -ENOENT;
215 
216 	/*
217 	 * NB and Last level cache counters (MSRs) are shared across all cores
218 	 * that share the same NB / Last level cache.  On family 16h and below,
219 	 * Interrupts can be directed to a single target core, however, event
220 	 * counts generated by processes running on other cores cannot be masked
221 	 * out. So we do not support sampling and per-thread events via
222 	 * CAP_NO_INTERRUPT, and we do not enable counter overflow interrupts:
223 	 */
224 	hwc->config = event->attr.config & AMD64_RAW_EVENT_MASK_NB;
225 	hwc->idx = -1;
226 
227 	if (event->cpu < 0)
228 		return -EINVAL;
229 
230 	/*
231 	 * SliceMask and ThreadMask need to be set for certain L3 events.
232 	 * For other events, the two fields do not affect the count.
233 	 */
234 	if (l3_mask && is_llc_event(event))
235 		hwc->config |= l3_thread_slice_mask(event->attr.config);
236 
237 	uncore = event_to_amd_uncore(event);
238 	if (!uncore)
239 		return -ENODEV;
240 
241 	/*
242 	 * since request can come in to any of the shared cores, we will remap
243 	 * to a single common cpu.
244 	 */
245 	event->cpu = uncore->cpu;
246 
247 	return 0;
248 }
249 
250 static ssize_t amd_uncore_attr_show_cpumask(struct device *dev,
251 					    struct device_attribute *attr,
252 					    char *buf)
253 {
254 	cpumask_t *active_mask;
255 	struct pmu *pmu = dev_get_drvdata(dev);
256 
257 	if (pmu->type == amd_nb_pmu.type)
258 		active_mask = &amd_nb_active_mask;
259 	else if (pmu->type == amd_llc_pmu.type)
260 		active_mask = &amd_llc_active_mask;
261 	else
262 		return 0;
263 
264 	return cpumap_print_to_pagebuf(true, buf, active_mask);
265 }
266 static DEVICE_ATTR(cpumask, S_IRUGO, amd_uncore_attr_show_cpumask, NULL);
267 
268 static struct attribute *amd_uncore_attrs[] = {
269 	&dev_attr_cpumask.attr,
270 	NULL,
271 };
272 
273 static struct attribute_group amd_uncore_attr_group = {
274 	.attrs = amd_uncore_attrs,
275 };
276 
277 #define DEFINE_UNCORE_FORMAT_ATTR(_var, _name, _format)			\
278 static ssize_t __uncore_##_var##_show(struct kobject *kobj,		\
279 				struct kobj_attribute *attr,		\
280 				char *page)				\
281 {									\
282 	BUILD_BUG_ON(sizeof(_format) >= PAGE_SIZE);			\
283 	return sprintf(page, _format "\n");				\
284 }									\
285 static struct kobj_attribute format_attr_##_var =			\
286 	__ATTR(_name, 0444, __uncore_##_var##_show, NULL)
287 
288 DEFINE_UNCORE_FORMAT_ATTR(event12,	event,		"config:0-7,32-35");
289 DEFINE_UNCORE_FORMAT_ATTR(event14,	event,		"config:0-7,32-35,59-60"); /* F17h+ DF */
290 DEFINE_UNCORE_FORMAT_ATTR(event8,	event,		"config:0-7");		   /* F17h+ L3 */
291 DEFINE_UNCORE_FORMAT_ATTR(umask,	umask,		"config:8-15");
292 DEFINE_UNCORE_FORMAT_ATTR(coreid,	coreid,		"config:42-44");	   /* F19h L3 */
293 DEFINE_UNCORE_FORMAT_ATTR(slicemask,	slicemask,	"config:48-51");	   /* F17h L3 */
294 DEFINE_UNCORE_FORMAT_ATTR(threadmask8,	threadmask,	"config:56-63");	   /* F17h L3 */
295 DEFINE_UNCORE_FORMAT_ATTR(threadmask2,	threadmask,	"config:56-57");	   /* F19h L3 */
296 DEFINE_UNCORE_FORMAT_ATTR(enallslices,	enallslices,	"config:46");		   /* F19h L3 */
297 DEFINE_UNCORE_FORMAT_ATTR(enallcores,	enallcores,	"config:47");		   /* F19h L3 */
298 DEFINE_UNCORE_FORMAT_ATTR(sliceid,	sliceid,	"config:48-50");	   /* F19h L3 */
299 
300 static struct attribute *amd_uncore_df_format_attr[] = {
301 	&format_attr_event12.attr, /* event14 if F17h+ */
302 	&format_attr_umask.attr,
303 	NULL,
304 };
305 
306 static struct attribute *amd_uncore_l3_format_attr[] = {
307 	&format_attr_event12.attr, /* event8 if F17h+ */
308 	&format_attr_umask.attr,
309 	NULL, /* slicemask if F17h,	coreid if F19h */
310 	NULL, /* threadmask8 if F17h,	enallslices if F19h */
311 	NULL, /*			enallcores if F19h */
312 	NULL, /*			sliceid if F19h */
313 	NULL, /*			threadmask2 if F19h */
314 	NULL,
315 };
316 
317 static struct attribute_group amd_uncore_df_format_group = {
318 	.name = "format",
319 	.attrs = amd_uncore_df_format_attr,
320 };
321 
322 static struct attribute_group amd_uncore_l3_format_group = {
323 	.name = "format",
324 	.attrs = amd_uncore_l3_format_attr,
325 };
326 
327 static const struct attribute_group *amd_uncore_df_attr_groups[] = {
328 	&amd_uncore_attr_group,
329 	&amd_uncore_df_format_group,
330 	NULL,
331 };
332 
333 static const struct attribute_group *amd_uncore_l3_attr_groups[] = {
334 	&amd_uncore_attr_group,
335 	&amd_uncore_l3_format_group,
336 	NULL,
337 };
338 
339 static struct pmu amd_nb_pmu = {
340 	.task_ctx_nr	= perf_invalid_context,
341 	.attr_groups	= amd_uncore_df_attr_groups,
342 	.name		= "amd_nb",
343 	.event_init	= amd_uncore_event_init,
344 	.add		= amd_uncore_add,
345 	.del		= amd_uncore_del,
346 	.start		= amd_uncore_start,
347 	.stop		= amd_uncore_stop,
348 	.read		= amd_uncore_read,
349 	.capabilities	= PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
350 };
351 
352 static struct pmu amd_llc_pmu = {
353 	.task_ctx_nr	= perf_invalid_context,
354 	.attr_groups	= amd_uncore_l3_attr_groups,
355 	.name		= "amd_l2",
356 	.event_init	= amd_uncore_event_init,
357 	.add		= amd_uncore_add,
358 	.del		= amd_uncore_del,
359 	.start		= amd_uncore_start,
360 	.stop		= amd_uncore_stop,
361 	.read		= amd_uncore_read,
362 	.capabilities	= PERF_PMU_CAP_NO_EXCLUDE | PERF_PMU_CAP_NO_INTERRUPT,
363 };
364 
365 static struct amd_uncore *amd_uncore_alloc(unsigned int cpu)
366 {
367 	return kzalloc_node(sizeof(struct amd_uncore), GFP_KERNEL,
368 			cpu_to_node(cpu));
369 }
370 
371 static int amd_uncore_cpu_up_prepare(unsigned int cpu)
372 {
373 	struct amd_uncore *uncore_nb = NULL, *uncore_llc;
374 
375 	if (amd_uncore_nb) {
376 		uncore_nb = amd_uncore_alloc(cpu);
377 		if (!uncore_nb)
378 			goto fail;
379 		uncore_nb->cpu = cpu;
380 		uncore_nb->num_counters = num_counters_nb;
381 		uncore_nb->rdpmc_base = RDPMC_BASE_NB;
382 		uncore_nb->msr_base = MSR_F15H_NB_PERF_CTL;
383 		uncore_nb->active_mask = &amd_nb_active_mask;
384 		uncore_nb->pmu = &amd_nb_pmu;
385 		uncore_nb->id = -1;
386 		*per_cpu_ptr(amd_uncore_nb, cpu) = uncore_nb;
387 	}
388 
389 	if (amd_uncore_llc) {
390 		uncore_llc = amd_uncore_alloc(cpu);
391 		if (!uncore_llc)
392 			goto fail;
393 		uncore_llc->cpu = cpu;
394 		uncore_llc->num_counters = num_counters_llc;
395 		uncore_llc->rdpmc_base = RDPMC_BASE_LLC;
396 		uncore_llc->msr_base = MSR_F16H_L2I_PERF_CTL;
397 		uncore_llc->active_mask = &amd_llc_active_mask;
398 		uncore_llc->pmu = &amd_llc_pmu;
399 		uncore_llc->id = -1;
400 		*per_cpu_ptr(amd_uncore_llc, cpu) = uncore_llc;
401 	}
402 
403 	return 0;
404 
405 fail:
406 	if (amd_uncore_nb)
407 		*per_cpu_ptr(amd_uncore_nb, cpu) = NULL;
408 	kfree(uncore_nb);
409 	return -ENOMEM;
410 }
411 
412 static struct amd_uncore *
413 amd_uncore_find_online_sibling(struct amd_uncore *this,
414 			       struct amd_uncore * __percpu *uncores)
415 {
416 	unsigned int cpu;
417 	struct amd_uncore *that;
418 
419 	for_each_online_cpu(cpu) {
420 		that = *per_cpu_ptr(uncores, cpu);
421 
422 		if (!that)
423 			continue;
424 
425 		if (this == that)
426 			continue;
427 
428 		if (this->id == that->id) {
429 			hlist_add_head(&this->node, &uncore_unused_list);
430 			this = that;
431 			break;
432 		}
433 	}
434 
435 	this->refcnt++;
436 	return this;
437 }
438 
439 static int amd_uncore_cpu_starting(unsigned int cpu)
440 {
441 	unsigned int eax, ebx, ecx, edx;
442 	struct amd_uncore *uncore;
443 
444 	if (amd_uncore_nb) {
445 		uncore = *per_cpu_ptr(amd_uncore_nb, cpu);
446 		cpuid(0x8000001e, &eax, &ebx, &ecx, &edx);
447 		uncore->id = ecx & 0xff;
448 
449 		uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_nb);
450 		*per_cpu_ptr(amd_uncore_nb, cpu) = uncore;
451 	}
452 
453 	if (amd_uncore_llc) {
454 		uncore = *per_cpu_ptr(amd_uncore_llc, cpu);
455 		uncore->id = per_cpu(cpu_llc_id, cpu);
456 
457 		uncore = amd_uncore_find_online_sibling(uncore, amd_uncore_llc);
458 		*per_cpu_ptr(amd_uncore_llc, cpu) = uncore;
459 	}
460 
461 	return 0;
462 }
463 
464 static void uncore_clean_online(void)
465 {
466 	struct amd_uncore *uncore;
467 	struct hlist_node *n;
468 
469 	hlist_for_each_entry_safe(uncore, n, &uncore_unused_list, node) {
470 		hlist_del(&uncore->node);
471 		kfree(uncore);
472 	}
473 }
474 
475 static void uncore_online(unsigned int cpu,
476 			  struct amd_uncore * __percpu *uncores)
477 {
478 	struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
479 
480 	uncore_clean_online();
481 
482 	if (cpu == uncore->cpu)
483 		cpumask_set_cpu(cpu, uncore->active_mask);
484 }
485 
486 static int amd_uncore_cpu_online(unsigned int cpu)
487 {
488 	if (amd_uncore_nb)
489 		uncore_online(cpu, amd_uncore_nb);
490 
491 	if (amd_uncore_llc)
492 		uncore_online(cpu, amd_uncore_llc);
493 
494 	return 0;
495 }
496 
497 static void uncore_down_prepare(unsigned int cpu,
498 				struct amd_uncore * __percpu *uncores)
499 {
500 	unsigned int i;
501 	struct amd_uncore *this = *per_cpu_ptr(uncores, cpu);
502 
503 	if (this->cpu != cpu)
504 		return;
505 
506 	/* this cpu is going down, migrate to a shared sibling if possible */
507 	for_each_online_cpu(i) {
508 		struct amd_uncore *that = *per_cpu_ptr(uncores, i);
509 
510 		if (cpu == i)
511 			continue;
512 
513 		if (this == that) {
514 			perf_pmu_migrate_context(this->pmu, cpu, i);
515 			cpumask_clear_cpu(cpu, that->active_mask);
516 			cpumask_set_cpu(i, that->active_mask);
517 			that->cpu = i;
518 			break;
519 		}
520 	}
521 }
522 
523 static int amd_uncore_cpu_down_prepare(unsigned int cpu)
524 {
525 	if (amd_uncore_nb)
526 		uncore_down_prepare(cpu, amd_uncore_nb);
527 
528 	if (amd_uncore_llc)
529 		uncore_down_prepare(cpu, amd_uncore_llc);
530 
531 	return 0;
532 }
533 
534 static void uncore_dead(unsigned int cpu, struct amd_uncore * __percpu *uncores)
535 {
536 	struct amd_uncore *uncore = *per_cpu_ptr(uncores, cpu);
537 
538 	if (cpu == uncore->cpu)
539 		cpumask_clear_cpu(cpu, uncore->active_mask);
540 
541 	if (!--uncore->refcnt)
542 		kfree(uncore);
543 	*per_cpu_ptr(uncores, cpu) = NULL;
544 }
545 
546 static int amd_uncore_cpu_dead(unsigned int cpu)
547 {
548 	if (amd_uncore_nb)
549 		uncore_dead(cpu, amd_uncore_nb);
550 
551 	if (amd_uncore_llc)
552 		uncore_dead(cpu, amd_uncore_llc);
553 
554 	return 0;
555 }
556 
557 static int __init amd_uncore_init(void)
558 {
559 	struct attribute **df_attr = amd_uncore_df_format_attr;
560 	struct attribute **l3_attr = amd_uncore_l3_format_attr;
561 	int ret = -ENODEV;
562 
563 	if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD &&
564 	    boot_cpu_data.x86_vendor != X86_VENDOR_HYGON)
565 		return -ENODEV;
566 
567 	if (!boot_cpu_has(X86_FEATURE_TOPOEXT))
568 		return -ENODEV;
569 
570 	num_counters_nb	= NUM_COUNTERS_NB;
571 	num_counters_llc = NUM_COUNTERS_L2;
572 	if (boot_cpu_data.x86 >= 0x17) {
573 		/*
574 		 * For F17h and above, the Northbridge counters are
575 		 * repurposed as Data Fabric counters. Also, L3
576 		 * counters are supported too. The PMUs are exported
577 		 * based on family as either L2 or L3 and NB or DF.
578 		 */
579 		num_counters_llc	  = NUM_COUNTERS_L3;
580 		amd_nb_pmu.name		  = "amd_df";
581 		amd_llc_pmu.name	  = "amd_l3";
582 		l3_mask			  = true;
583 	}
584 
585 	if (boot_cpu_has(X86_FEATURE_PERFCTR_NB)) {
586 		if (boot_cpu_data.x86 >= 0x17)
587 			*df_attr = &format_attr_event14.attr;
588 
589 		amd_uncore_nb = alloc_percpu(struct amd_uncore *);
590 		if (!amd_uncore_nb) {
591 			ret = -ENOMEM;
592 			goto fail_nb;
593 		}
594 		ret = perf_pmu_register(&amd_nb_pmu, amd_nb_pmu.name, -1);
595 		if (ret)
596 			goto fail_nb;
597 
598 		pr_info("%d %s %s counters detected\n", num_counters_nb,
599 			boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ?  "HYGON" : "",
600 			amd_nb_pmu.name);
601 
602 		ret = 0;
603 	}
604 
605 	if (boot_cpu_has(X86_FEATURE_PERFCTR_LLC)) {
606 		if (boot_cpu_data.x86 >= 0x19) {
607 			*l3_attr++ = &format_attr_event8.attr;
608 			*l3_attr++ = &format_attr_umask.attr;
609 			*l3_attr++ = &format_attr_coreid.attr;
610 			*l3_attr++ = &format_attr_enallslices.attr;
611 			*l3_attr++ = &format_attr_enallcores.attr;
612 			*l3_attr++ = &format_attr_sliceid.attr;
613 			*l3_attr++ = &format_attr_threadmask2.attr;
614 		} else if (boot_cpu_data.x86 >= 0x17) {
615 			*l3_attr++ = &format_attr_event8.attr;
616 			*l3_attr++ = &format_attr_umask.attr;
617 			*l3_attr++ = &format_attr_slicemask.attr;
618 			*l3_attr++ = &format_attr_threadmask8.attr;
619 		}
620 
621 		amd_uncore_llc = alloc_percpu(struct amd_uncore *);
622 		if (!amd_uncore_llc) {
623 			ret = -ENOMEM;
624 			goto fail_llc;
625 		}
626 		ret = perf_pmu_register(&amd_llc_pmu, amd_llc_pmu.name, -1);
627 		if (ret)
628 			goto fail_llc;
629 
630 		pr_info("%d %s %s counters detected\n", num_counters_llc,
631 			boot_cpu_data.x86_vendor == X86_VENDOR_HYGON ?  "HYGON" : "",
632 			amd_llc_pmu.name);
633 		ret = 0;
634 	}
635 
636 	/*
637 	 * Install callbacks. Core will call them for each online cpu.
638 	 */
639 	if (cpuhp_setup_state(CPUHP_PERF_X86_AMD_UNCORE_PREP,
640 			      "perf/x86/amd/uncore:prepare",
641 			      amd_uncore_cpu_up_prepare, amd_uncore_cpu_dead))
642 		goto fail_llc;
643 
644 	if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING,
645 			      "perf/x86/amd/uncore:starting",
646 			      amd_uncore_cpu_starting, NULL))
647 		goto fail_prep;
648 	if (cpuhp_setup_state(CPUHP_AP_PERF_X86_AMD_UNCORE_ONLINE,
649 			      "perf/x86/amd/uncore:online",
650 			      amd_uncore_cpu_online,
651 			      amd_uncore_cpu_down_prepare))
652 		goto fail_start;
653 	return 0;
654 
655 fail_start:
656 	cpuhp_remove_state(CPUHP_AP_PERF_X86_AMD_UNCORE_STARTING);
657 fail_prep:
658 	cpuhp_remove_state(CPUHP_PERF_X86_AMD_UNCORE_PREP);
659 fail_llc:
660 	if (boot_cpu_has(X86_FEATURE_PERFCTR_NB))
661 		perf_pmu_unregister(&amd_nb_pmu);
662 	if (amd_uncore_llc)
663 		free_percpu(amd_uncore_llc);
664 fail_nb:
665 	if (amd_uncore_nb)
666 		free_percpu(amd_uncore_nb);
667 
668 	return ret;
669 }
670 device_initcall(amd_uncore_init);
671