xref: /linux/arch/x86/kvm/vmx/pmu_intel.c (revision f69e98a91a01fd7c5755dd710e94a17d6e9f583f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * KVM PMU support for Intel CPUs
4  *
5  * Copyright 2011 Red Hat, Inc. and/or its affiliates.
6  *
7  * Authors:
8  *   Avi Kivity   <avi@redhat.com>
9  *   Gleb Natapov <gleb@redhat.com>
10  */
11 #include <linux/types.h>
12 #include <linux/kvm_host.h>
13 #include <linux/perf_event.h>
14 #include <asm/perf_event.h>
15 #include "x86.h"
16 #include "cpuid.h"
17 #include "lapic.h"
18 #include "nested.h"
19 #include "pmu.h"
20 
21 #define MSR_PMC_FULL_WIDTH_BIT      (MSR_IA32_PMC0 - MSR_IA32_PERFCTR0)
22 
23 static struct kvm_event_hw_type_mapping intel_arch_events[] = {
24 	[0] = { 0x3c, 0x00, PERF_COUNT_HW_CPU_CYCLES },
25 	[1] = { 0xc0, 0x00, PERF_COUNT_HW_INSTRUCTIONS },
26 	[2] = { 0x3c, 0x01, PERF_COUNT_HW_BUS_CYCLES  },
27 	[3] = { 0x2e, 0x4f, PERF_COUNT_HW_CACHE_REFERENCES },
28 	[4] = { 0x2e, 0x41, PERF_COUNT_HW_CACHE_MISSES },
29 	[5] = { 0xc4, 0x00, PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
30 	[6] = { 0xc5, 0x00, PERF_COUNT_HW_BRANCH_MISSES },
31 	/* The above index must match CPUID 0x0A.EBX bit vector */
32 	[7] = { 0x00, 0x03, PERF_COUNT_HW_REF_CPU_CYCLES },
33 };
34 
35 /* mapping between fixed pmc index and intel_arch_events array */
36 static int fixed_pmc_events[] = {1, 0, 7};
37 
38 static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
39 {
40 	int i;
41 
42 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
43 		u8 new_ctrl = fixed_ctrl_field(data, i);
44 		u8 old_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl, i);
45 		struct kvm_pmc *pmc;
46 
47 		pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i);
48 
49 		if (old_ctrl == new_ctrl)
50 			continue;
51 
52 		__set_bit(INTEL_PMC_IDX_FIXED + i, pmu->pmc_in_use);
53 		reprogram_fixed_counter(pmc, new_ctrl, i);
54 	}
55 
56 	pmu->fixed_ctr_ctrl = data;
57 }
58 
59 /* function is called when global control register has been updated. */
60 static void global_ctrl_changed(struct kvm_pmu *pmu, u64 data)
61 {
62 	int bit;
63 	u64 diff = pmu->global_ctrl ^ data;
64 
65 	pmu->global_ctrl = data;
66 
67 	for_each_set_bit(bit, (unsigned long *)&diff, X86_PMC_IDX_MAX)
68 		reprogram_counter(pmu, bit);
69 }
70 
71 static unsigned int intel_pmc_perf_hw_id(struct kvm_pmc *pmc)
72 {
73 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
74 	u8 event_select = pmc->eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
75 	u8 unit_mask = (pmc->eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
76 	int i;
77 
78 	for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++) {
79 		if (intel_arch_events[i].eventsel != event_select ||
80 		    intel_arch_events[i].unit_mask != unit_mask)
81 			continue;
82 
83 		/* disable event that reported as not present by cpuid */
84 		if ((i < 7) && !(pmu->available_event_types & (1 << i)))
85 			return PERF_COUNT_HW_MAX + 1;
86 
87 		break;
88 	}
89 
90 	if (i == ARRAY_SIZE(intel_arch_events))
91 		return PERF_COUNT_HW_MAX;
92 
93 	return intel_arch_events[i].event_type;
94 }
95 
96 /* check if a PMC is enabled by comparing it with globl_ctrl bits. */
97 static bool intel_pmc_is_enabled(struct kvm_pmc *pmc)
98 {
99 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
100 
101 	return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
102 }
103 
104 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
105 {
106 	if (pmc_idx < INTEL_PMC_IDX_FIXED)
107 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
108 				  MSR_P6_EVNTSEL0);
109 	else {
110 		u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
111 
112 		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
113 	}
114 }
115 
116 static bool intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
117 {
118 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
119 	bool fixed = idx & (1u << 30);
120 
121 	idx &= ~(3u << 30);
122 
123 	return fixed ? idx < pmu->nr_arch_fixed_counters
124 		     : idx < pmu->nr_arch_gp_counters;
125 }
126 
127 static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
128 					    unsigned int idx, u64 *mask)
129 {
130 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
131 	bool fixed = idx & (1u << 30);
132 	struct kvm_pmc *counters;
133 	unsigned int num_counters;
134 
135 	idx &= ~(3u << 30);
136 	if (fixed) {
137 		counters = pmu->fixed_counters;
138 		num_counters = pmu->nr_arch_fixed_counters;
139 	} else {
140 		counters = pmu->gp_counters;
141 		num_counters = pmu->nr_arch_gp_counters;
142 	}
143 	if (idx >= num_counters)
144 		return NULL;
145 	*mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
146 	return &counters[array_index_nospec(idx, num_counters)];
147 }
148 
149 static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
150 {
151 	if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
152 		return 0;
153 
154 	return vcpu->arch.perf_capabilities;
155 }
156 
157 static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
158 {
159 	return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
160 }
161 
162 static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
163 {
164 	if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
165 		return NULL;
166 
167 	return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
168 }
169 
170 bool intel_pmu_lbr_is_compatible(struct kvm_vcpu *vcpu)
171 {
172 	/*
173 	 * As a first step, a guest could only enable LBR feature if its
174 	 * cpu model is the same as the host because the LBR registers
175 	 * would be pass-through to the guest and they're model specific.
176 	 */
177 	return boot_cpu_data.x86_model == guest_cpuid_model(vcpu);
178 }
179 
180 bool intel_pmu_lbr_is_enabled(struct kvm_vcpu *vcpu)
181 {
182 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
183 
184 	return lbr->nr && (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_LBR_FMT);
185 }
186 
187 static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
188 {
189 	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
190 	bool ret = false;
191 
192 	if (!intel_pmu_lbr_is_enabled(vcpu))
193 		return ret;
194 
195 	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
196 		(index >= records->from && index < records->from + records->nr) ||
197 		(index >= records->to && index < records->to + records->nr);
198 
199 	if (!ret && records->info)
200 		ret = (index >= records->info && index < records->info + records->nr);
201 
202 	return ret;
203 }
204 
205 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
206 {
207 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
208 	int ret;
209 
210 	switch (msr) {
211 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
212 	case MSR_CORE_PERF_GLOBAL_STATUS:
213 	case MSR_CORE_PERF_GLOBAL_CTRL:
214 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
215 		ret = pmu->version > 1;
216 		break;
217 	default:
218 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
219 			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
220 			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
221 			intel_pmu_is_valid_lbr_msr(vcpu, msr);
222 		break;
223 	}
224 
225 	return ret;
226 }
227 
228 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
229 {
230 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
231 	struct kvm_pmc *pmc;
232 
233 	pmc = get_fixed_pmc(pmu, msr);
234 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
235 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
236 
237 	return pmc;
238 }
239 
240 static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
241 {
242 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
243 
244 	if (lbr_desc->event) {
245 		perf_event_release_kernel(lbr_desc->event);
246 		lbr_desc->event = NULL;
247 		vcpu_to_pmu(vcpu)->event_count--;
248 	}
249 }
250 
251 int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
252 {
253 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
254 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
255 	struct perf_event *event;
256 
257 	/*
258 	 * The perf_event_attr is constructed in the minimum efficient way:
259 	 * - set 'pinned = true' to make it task pinned so that if another
260 	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
261 	 * - set '.exclude_host = true' to record guest branches behavior;
262 	 *
263 	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
264 	 *   schedule the event without a real HW counter but a fake one;
265 	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
266 	 *
267 	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
268 	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
269 	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
270 	 *   event, which helps KVM to save/restore guest LBR records
271 	 *   during host context switches and reduces quite a lot overhead,
272 	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
273 	 */
274 	struct perf_event_attr attr = {
275 		.type = PERF_TYPE_RAW,
276 		.size = sizeof(attr),
277 		.config = INTEL_FIXED_VLBR_EVENT,
278 		.sample_type = PERF_SAMPLE_BRANCH_STACK,
279 		.pinned = true,
280 		.exclude_host = true,
281 		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
282 					PERF_SAMPLE_BRANCH_USER,
283 	};
284 
285 	if (unlikely(lbr_desc->event)) {
286 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
287 		return 0;
288 	}
289 
290 	event = perf_event_create_kernel_counter(&attr, -1,
291 						current, NULL, NULL);
292 	if (IS_ERR(event)) {
293 		pr_debug_ratelimited("%s: failed %ld\n",
294 					__func__, PTR_ERR(event));
295 		return PTR_ERR(event);
296 	}
297 	lbr_desc->event = event;
298 	pmu->event_count++;
299 	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
300 	return 0;
301 }
302 
303 /*
304  * It's safe to access LBR msrs from guest when they have not
305  * been passthrough since the host would help restore or reset
306  * the LBR msrs records when the guest LBR event is scheduled in.
307  */
308 static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
309 				     struct msr_data *msr_info, bool read)
310 {
311 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
312 	u32 index = msr_info->index;
313 
314 	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
315 		return false;
316 
317 	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
318 		goto dummy;
319 
320 	/*
321 	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
322 	 * host at the time the value is read from the msr, and this avoids the
323 	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
324 	 * return 0 on guest reads.
325 	 */
326 	local_irq_disable();
327 	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
328 		if (read)
329 			rdmsrl(index, msr_info->data);
330 		else
331 			wrmsrl(index, msr_info->data);
332 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
333 		local_irq_enable();
334 		return true;
335 	}
336 	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
337 	local_irq_enable();
338 
339 dummy:
340 	if (read)
341 		msr_info->data = 0;
342 	return true;
343 }
344 
345 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
346 {
347 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
348 	struct kvm_pmc *pmc;
349 	u32 msr = msr_info->index;
350 
351 	switch (msr) {
352 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
353 		msr_info->data = pmu->fixed_ctr_ctrl;
354 		return 0;
355 	case MSR_CORE_PERF_GLOBAL_STATUS:
356 		msr_info->data = pmu->global_status;
357 		return 0;
358 	case MSR_CORE_PERF_GLOBAL_CTRL:
359 		msr_info->data = pmu->global_ctrl;
360 		return 0;
361 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
362 		msr_info->data = 0;
363 		return 0;
364 	default:
365 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
366 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
367 			u64 val = pmc_read_counter(pmc);
368 			msr_info->data =
369 				val & pmu->counter_bitmask[KVM_PMC_GP];
370 			return 0;
371 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
372 			u64 val = pmc_read_counter(pmc);
373 			msr_info->data =
374 				val & pmu->counter_bitmask[KVM_PMC_FIXED];
375 			return 0;
376 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
377 			msr_info->data = pmc->eventsel;
378 			return 0;
379 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true))
380 			return 0;
381 	}
382 
383 	return 1;
384 }
385 
386 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
387 {
388 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
389 	struct kvm_pmc *pmc;
390 	u32 msr = msr_info->index;
391 	u64 data = msr_info->data;
392 	u64 reserved_bits;
393 
394 	switch (msr) {
395 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
396 		if (pmu->fixed_ctr_ctrl == data)
397 			return 0;
398 		if (!(data & 0xfffffffffffff444ull)) {
399 			reprogram_fixed_counters(pmu, data);
400 			return 0;
401 		}
402 		break;
403 	case MSR_CORE_PERF_GLOBAL_STATUS:
404 		if (msr_info->host_initiated) {
405 			pmu->global_status = data;
406 			return 0;
407 		}
408 		break; /* RO MSR */
409 	case MSR_CORE_PERF_GLOBAL_CTRL:
410 		if (pmu->global_ctrl == data)
411 			return 0;
412 		if (kvm_valid_perf_global_ctrl(pmu, data)) {
413 			global_ctrl_changed(pmu, data);
414 			return 0;
415 		}
416 		break;
417 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
418 		if (!(data & pmu->global_ovf_ctrl_mask)) {
419 			if (!msr_info->host_initiated)
420 				pmu->global_status &= ~data;
421 			return 0;
422 		}
423 		break;
424 	default:
425 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
426 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
427 			if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
428 			    (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
429 				return 1;
430 			if (!msr_info->host_initiated &&
431 			    !(msr & MSR_PMC_FULL_WIDTH_BIT))
432 				data = (s64)(s32)data;
433 			pmc->counter += data - pmc_read_counter(pmc);
434 			pmc_update_sample_period(pmc);
435 			return 0;
436 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
437 			pmc->counter += data - pmc_read_counter(pmc);
438 			pmc_update_sample_period(pmc);
439 			return 0;
440 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
441 			if (data == pmc->eventsel)
442 				return 0;
443 			reserved_bits = pmu->reserved_bits;
444 			if ((pmc->idx == 2) &&
445 			    (pmu->raw_event_mask & HSW_IN_TX_CHECKPOINTED))
446 				reserved_bits ^= HSW_IN_TX_CHECKPOINTED;
447 			if (!(data & reserved_bits)) {
448 				reprogram_gp_counter(pmc, data);
449 				return 0;
450 			}
451 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false))
452 			return 0;
453 	}
454 
455 	return 1;
456 }
457 
458 static void setup_fixed_pmc_eventsel(struct kvm_pmu *pmu)
459 {
460 	size_t size = ARRAY_SIZE(fixed_pmc_events);
461 	struct kvm_pmc *pmc;
462 	u32 event;
463 	int i;
464 
465 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
466 		pmc = &pmu->fixed_counters[i];
467 		event = fixed_pmc_events[array_index_nospec(i, size)];
468 		pmc->eventsel = (intel_arch_events[event].unit_mask << 8) |
469 			intel_arch_events[event].eventsel;
470 	}
471 }
472 
473 static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
474 {
475 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
476 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
477 
478 	struct x86_pmu_capability x86_pmu;
479 	struct kvm_cpuid_entry2 *entry;
480 	union cpuid10_eax eax;
481 	union cpuid10_edx edx;
482 
483 	pmu->nr_arch_gp_counters = 0;
484 	pmu->nr_arch_fixed_counters = 0;
485 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
486 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
487 	pmu->version = 0;
488 	pmu->reserved_bits = 0xffffffff00200000ull;
489 	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
490 
491 	entry = kvm_find_cpuid_entry(vcpu, 0xa, 0);
492 	if (!entry || !vcpu->kvm->arch.enable_pmu)
493 		return;
494 	eax.full = entry->eax;
495 	edx.full = entry->edx;
496 
497 	pmu->version = eax.split.version_id;
498 	if (!pmu->version)
499 		return;
500 
501 	perf_get_x86_pmu_capability(&x86_pmu);
502 
503 	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
504 					 x86_pmu.num_counters_gp);
505 	eax.split.bit_width = min_t(int, eax.split.bit_width, x86_pmu.bit_width_gp);
506 	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
507 	eax.split.mask_length = min_t(int, eax.split.mask_length, x86_pmu.events_mask_len);
508 	pmu->available_event_types = ~entry->ebx &
509 					((1ull << eax.split.mask_length) - 1);
510 
511 	if (pmu->version == 1) {
512 		pmu->nr_arch_fixed_counters = 0;
513 	} else {
514 		pmu->nr_arch_fixed_counters =
515 			min3(ARRAY_SIZE(fixed_pmc_events),
516 			     (size_t) edx.split.num_counters_fixed,
517 			     (size_t) x86_pmu.num_counters_fixed);
518 		edx.split.bit_width_fixed = min_t(int,
519 			edx.split.bit_width_fixed, x86_pmu.bit_width_fixed);
520 		pmu->counter_bitmask[KVM_PMC_FIXED] =
521 			((u64)1 << edx.split.bit_width_fixed) - 1;
522 		setup_fixed_pmc_eventsel(pmu);
523 	}
524 
525 	pmu->global_ctrl = ((1ull << pmu->nr_arch_gp_counters) - 1) |
526 		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED);
527 	pmu->global_ctrl_mask = ~pmu->global_ctrl;
528 	pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask
529 			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
530 			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
531 	if (vmx_pt_mode_is_host_guest())
532 		pmu->global_ovf_ctrl_mask &=
533 				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
534 
535 	entry = kvm_find_cpuid_entry(vcpu, 7, 0);
536 	if (entry &&
537 	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
538 	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) {
539 		pmu->reserved_bits ^= HSW_IN_TX;
540 		pmu->raw_event_mask |= (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
541 	}
542 
543 	bitmap_set(pmu->all_valid_pmc_idx,
544 		0, pmu->nr_arch_gp_counters);
545 	bitmap_set(pmu->all_valid_pmc_idx,
546 		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
547 
548 	nested_vmx_pmu_refresh(vcpu,
549 			       intel_is_valid_msr(vcpu, MSR_CORE_PERF_GLOBAL_CTRL));
550 
551 	if (intel_pmu_lbr_is_compatible(vcpu))
552 		x86_perf_get_lbr(&lbr_desc->records);
553 	else
554 		lbr_desc->records.nr = 0;
555 
556 	if (lbr_desc->records.nr)
557 		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
558 }
559 
560 static void intel_pmu_init(struct kvm_vcpu *vcpu)
561 {
562 	int i;
563 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
564 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
565 
566 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
567 		pmu->gp_counters[i].type = KVM_PMC_GP;
568 		pmu->gp_counters[i].vcpu = vcpu;
569 		pmu->gp_counters[i].idx = i;
570 		pmu->gp_counters[i].current_config = 0;
571 	}
572 
573 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
574 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
575 		pmu->fixed_counters[i].vcpu = vcpu;
576 		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
577 		pmu->fixed_counters[i].current_config = 0;
578 	}
579 
580 	vcpu->arch.perf_capabilities = vmx_get_perf_capabilities();
581 	lbr_desc->records.nr = 0;
582 	lbr_desc->event = NULL;
583 	lbr_desc->msr_passthrough = false;
584 }
585 
586 static void intel_pmu_reset(struct kvm_vcpu *vcpu)
587 {
588 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
589 	struct kvm_pmc *pmc = NULL;
590 	int i;
591 
592 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
593 		pmc = &pmu->gp_counters[i];
594 
595 		pmc_stop_counter(pmc);
596 		pmc->counter = pmc->eventsel = 0;
597 	}
598 
599 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
600 		pmc = &pmu->fixed_counters[i];
601 
602 		pmc_stop_counter(pmc);
603 		pmc->counter = 0;
604 	}
605 
606 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 0;
607 
608 	intel_pmu_release_guest_lbr_event(vcpu);
609 }
610 
611 /*
612  * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
613  *
614  * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
615  * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
616  *
617  * Guest needs to re-enable LBR to resume branches recording.
618  */
619 static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
620 {
621 	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
622 
623 	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
624 		data &= ~DEBUGCTLMSR_LBR;
625 		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
626 	}
627 }
628 
629 static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
630 {
631 	u8 version = vcpu_to_pmu(vcpu)->version;
632 
633 	if (!intel_pmu_lbr_is_enabled(vcpu))
634 		return;
635 
636 	if (version > 1 && version < 4)
637 		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
638 }
639 
640 static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
641 {
642 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
643 	int i;
644 
645 	for (i = 0; i < lbr->nr; i++) {
646 		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
647 		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
648 		if (lbr->info)
649 			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
650 	}
651 
652 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
653 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
654 }
655 
656 static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
657 {
658 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
659 
660 	if (!lbr_desc->msr_passthrough)
661 		return;
662 
663 	vmx_update_intercept_for_lbr_msrs(vcpu, true);
664 	lbr_desc->msr_passthrough = false;
665 }
666 
667 static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
668 {
669 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
670 
671 	if (lbr_desc->msr_passthrough)
672 		return;
673 
674 	vmx_update_intercept_for_lbr_msrs(vcpu, false);
675 	lbr_desc->msr_passthrough = true;
676 }
677 
678 /*
679  * Higher priority host perf events (e.g. cpu pinned) could reclaim the
680  * pmu resources (e.g. LBR) that were assigned to the guest. This is
681  * usually done via ipi calls (more details in perf_install_in_context).
682  *
683  * Before entering the non-root mode (with irq disabled here), double
684  * confirm that the pmu features enabled to the guest are not reclaimed
685  * by higher priority host events. Otherwise, disallow vcpu's access to
686  * the reclaimed features.
687  */
688 void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
689 {
690 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
691 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
692 
693 	if (!lbr_desc->event) {
694 		vmx_disable_lbr_msrs_passthrough(vcpu);
695 		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
696 			goto warn;
697 		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
698 			goto warn;
699 		return;
700 	}
701 
702 	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
703 		vmx_disable_lbr_msrs_passthrough(vcpu);
704 		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
705 		goto warn;
706 	} else
707 		vmx_enable_lbr_msrs_passthrough(vcpu);
708 
709 	return;
710 
711 warn:
712 	pr_warn_ratelimited("kvm: vcpu-%d: fail to passthrough LBR.\n",
713 		vcpu->vcpu_id);
714 }
715 
716 static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
717 {
718 	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
719 		intel_pmu_release_guest_lbr_event(vcpu);
720 }
721 
722 struct kvm_pmu_ops intel_pmu_ops = {
723 	.pmc_perf_hw_id = intel_pmc_perf_hw_id,
724 	.pmc_is_enabled = intel_pmc_is_enabled,
725 	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
726 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
727 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
728 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
729 	.is_valid_msr = intel_is_valid_msr,
730 	.get_msr = intel_pmu_get_msr,
731 	.set_msr = intel_pmu_set_msr,
732 	.refresh = intel_pmu_refresh,
733 	.init = intel_pmu_init,
734 	.reset = intel_pmu_reset,
735 	.deliver_pmi = intel_pmu_deliver_pmi,
736 	.cleanup = intel_pmu_cleanup,
737 };
738