xref: /openbmc/linux/arch/x86/kvm/vmx/pmu_intel.c (revision 67559900)
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 	/* Index must match CPUID 0x0A.EBX bit vector */
25 	[0] = { 0x3c, 0x00, PERF_COUNT_HW_CPU_CYCLES },
26 	[1] = { 0xc0, 0x00, PERF_COUNT_HW_INSTRUCTIONS },
27 	[2] = { 0x3c, 0x01, PERF_COUNT_HW_BUS_CYCLES  },
28 	[3] = { 0x2e, 0x4f, PERF_COUNT_HW_CACHE_REFERENCES },
29 	[4] = { 0x2e, 0x41, PERF_COUNT_HW_CACHE_MISSES },
30 	[5] = { 0xc4, 0x00, PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
31 	[6] = { 0xc5, 0x00, PERF_COUNT_HW_BRANCH_MISSES },
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 intel_find_arch_event(struct kvm_pmu *pmu,
72 				      u8 event_select,
73 				      u8 unit_mask)
74 {
75 	int i;
76 
77 	for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++)
78 		if (intel_arch_events[i].eventsel == event_select
79 		    && intel_arch_events[i].unit_mask == unit_mask
80 		    && (pmu->available_event_types & (1 << i)))
81 			break;
82 
83 	if (i == ARRAY_SIZE(intel_arch_events))
84 		return PERF_COUNT_HW_MAX;
85 
86 	return intel_arch_events[i].event_type;
87 }
88 
89 static unsigned intel_find_fixed_event(int idx)
90 {
91 	u32 event;
92 	size_t size = ARRAY_SIZE(fixed_pmc_events);
93 
94 	if (idx >= size)
95 		return PERF_COUNT_HW_MAX;
96 
97 	event = fixed_pmc_events[array_index_nospec(idx, size)];
98 	return intel_arch_events[event].event_type;
99 }
100 
101 /* check if a PMC is enabled by comparing it with globl_ctrl bits. */
102 static bool intel_pmc_is_enabled(struct kvm_pmc *pmc)
103 {
104 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
105 
106 	return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
107 }
108 
109 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
110 {
111 	if (pmc_idx < INTEL_PMC_IDX_FIXED)
112 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
113 				  MSR_P6_EVNTSEL0);
114 	else {
115 		u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
116 
117 		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
118 	}
119 }
120 
121 static bool intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
122 {
123 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
124 	bool fixed = idx & (1u << 30);
125 
126 	idx &= ~(3u << 30);
127 
128 	return fixed ? idx < pmu->nr_arch_fixed_counters
129 		     : idx < pmu->nr_arch_gp_counters;
130 }
131 
132 static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
133 					    unsigned int idx, u64 *mask)
134 {
135 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
136 	bool fixed = idx & (1u << 30);
137 	struct kvm_pmc *counters;
138 	unsigned int num_counters;
139 
140 	idx &= ~(3u << 30);
141 	if (fixed) {
142 		counters = pmu->fixed_counters;
143 		num_counters = pmu->nr_arch_fixed_counters;
144 	} else {
145 		counters = pmu->gp_counters;
146 		num_counters = pmu->nr_arch_gp_counters;
147 	}
148 	if (idx >= num_counters)
149 		return NULL;
150 	*mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
151 	return &counters[array_index_nospec(idx, num_counters)];
152 }
153 
154 static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
155 {
156 	if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
157 		return 0;
158 
159 	return vcpu->arch.perf_capabilities;
160 }
161 
162 static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
163 {
164 	return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
165 }
166 
167 static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
168 {
169 	if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
170 		return NULL;
171 
172 	return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
173 }
174 
175 bool intel_pmu_lbr_is_compatible(struct kvm_vcpu *vcpu)
176 {
177 	/*
178 	 * As a first step, a guest could only enable LBR feature if its
179 	 * cpu model is the same as the host because the LBR registers
180 	 * would be pass-through to the guest and they're model specific.
181 	 */
182 	return boot_cpu_data.x86_model == guest_cpuid_model(vcpu);
183 }
184 
185 bool intel_pmu_lbr_is_enabled(struct kvm_vcpu *vcpu)
186 {
187 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
188 
189 	return lbr->nr && (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_LBR_FMT);
190 }
191 
192 static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
193 {
194 	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
195 	bool ret = false;
196 
197 	if (!intel_pmu_lbr_is_enabled(vcpu))
198 		return ret;
199 
200 	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
201 		(index >= records->from && index < records->from + records->nr) ||
202 		(index >= records->to && index < records->to + records->nr);
203 
204 	if (!ret && records->info)
205 		ret = (index >= records->info && index < records->info + records->nr);
206 
207 	return ret;
208 }
209 
210 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
211 {
212 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
213 	int ret;
214 
215 	switch (msr) {
216 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
217 	case MSR_CORE_PERF_GLOBAL_STATUS:
218 	case MSR_CORE_PERF_GLOBAL_CTRL:
219 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
220 		ret = pmu->version > 1;
221 		break;
222 	default:
223 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
224 			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
225 			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
226 			intel_pmu_is_valid_lbr_msr(vcpu, msr);
227 		break;
228 	}
229 
230 	return ret;
231 }
232 
233 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
234 {
235 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
236 	struct kvm_pmc *pmc;
237 
238 	pmc = get_fixed_pmc(pmu, msr);
239 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
240 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
241 
242 	return pmc;
243 }
244 
245 static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
246 {
247 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
248 
249 	if (lbr_desc->event) {
250 		perf_event_release_kernel(lbr_desc->event);
251 		lbr_desc->event = NULL;
252 		vcpu_to_pmu(vcpu)->event_count--;
253 	}
254 }
255 
256 int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
257 {
258 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
259 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
260 	struct perf_event *event;
261 
262 	/*
263 	 * The perf_event_attr is constructed in the minimum efficient way:
264 	 * - set 'pinned = true' to make it task pinned so that if another
265 	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
266 	 * - set '.exclude_host = true' to record guest branches behavior;
267 	 *
268 	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
269 	 *   schedule the event without a real HW counter but a fake one;
270 	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
271 	 *
272 	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
273 	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
274 	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
275 	 *   event, which helps KVM to save/restore guest LBR records
276 	 *   during host context switches and reduces quite a lot overhead,
277 	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
278 	 */
279 	struct perf_event_attr attr = {
280 		.type = PERF_TYPE_RAW,
281 		.size = sizeof(attr),
282 		.config = INTEL_FIXED_VLBR_EVENT,
283 		.sample_type = PERF_SAMPLE_BRANCH_STACK,
284 		.pinned = true,
285 		.exclude_host = true,
286 		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
287 					PERF_SAMPLE_BRANCH_USER,
288 	};
289 
290 	if (unlikely(lbr_desc->event)) {
291 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
292 		return 0;
293 	}
294 
295 	event = perf_event_create_kernel_counter(&attr, -1,
296 						current, NULL, NULL);
297 	if (IS_ERR(event)) {
298 		pr_debug_ratelimited("%s: failed %ld\n",
299 					__func__, PTR_ERR(event));
300 		return PTR_ERR(event);
301 	}
302 	lbr_desc->event = event;
303 	pmu->event_count++;
304 	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
305 	return 0;
306 }
307 
308 /*
309  * It's safe to access LBR msrs from guest when they have not
310  * been passthrough since the host would help restore or reset
311  * the LBR msrs records when the guest LBR event is scheduled in.
312  */
313 static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
314 				     struct msr_data *msr_info, bool read)
315 {
316 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
317 	u32 index = msr_info->index;
318 
319 	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
320 		return false;
321 
322 	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
323 		goto dummy;
324 
325 	/*
326 	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
327 	 * host at the time the value is read from the msr, and this avoids the
328 	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
329 	 * return 0 on guest reads.
330 	 */
331 	local_irq_disable();
332 	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
333 		if (read)
334 			rdmsrl(index, msr_info->data);
335 		else
336 			wrmsrl(index, msr_info->data);
337 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
338 		local_irq_enable();
339 		return true;
340 	}
341 	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
342 	local_irq_enable();
343 
344 dummy:
345 	if (read)
346 		msr_info->data = 0;
347 	return true;
348 }
349 
350 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
351 {
352 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
353 	struct kvm_pmc *pmc;
354 	u32 msr = msr_info->index;
355 
356 	switch (msr) {
357 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
358 		msr_info->data = pmu->fixed_ctr_ctrl;
359 		return 0;
360 	case MSR_CORE_PERF_GLOBAL_STATUS:
361 		msr_info->data = pmu->global_status;
362 		return 0;
363 	case MSR_CORE_PERF_GLOBAL_CTRL:
364 		msr_info->data = pmu->global_ctrl;
365 		return 0;
366 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
367 		msr_info->data = 0;
368 		return 0;
369 	default:
370 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
371 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
372 			u64 val = pmc_read_counter(pmc);
373 			msr_info->data =
374 				val & pmu->counter_bitmask[KVM_PMC_GP];
375 			return 0;
376 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
377 			u64 val = pmc_read_counter(pmc);
378 			msr_info->data =
379 				val & pmu->counter_bitmask[KVM_PMC_FIXED];
380 			return 0;
381 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
382 			msr_info->data = pmc->eventsel;
383 			return 0;
384 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true))
385 			return 0;
386 	}
387 
388 	return 1;
389 }
390 
391 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
392 {
393 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
394 	struct kvm_pmc *pmc;
395 	u32 msr = msr_info->index;
396 	u64 data = msr_info->data;
397 
398 	switch (msr) {
399 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
400 		if (pmu->fixed_ctr_ctrl == data)
401 			return 0;
402 		if (!(data & 0xfffffffffffff444ull)) {
403 			reprogram_fixed_counters(pmu, data);
404 			return 0;
405 		}
406 		break;
407 	case MSR_CORE_PERF_GLOBAL_STATUS:
408 		if (msr_info->host_initiated) {
409 			pmu->global_status = data;
410 			return 0;
411 		}
412 		break; /* RO MSR */
413 	case MSR_CORE_PERF_GLOBAL_CTRL:
414 		if (pmu->global_ctrl == data)
415 			return 0;
416 		if (kvm_valid_perf_global_ctrl(pmu, data)) {
417 			global_ctrl_changed(pmu, data);
418 			return 0;
419 		}
420 		break;
421 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
422 		if (!(data & pmu->global_ovf_ctrl_mask)) {
423 			if (!msr_info->host_initiated)
424 				pmu->global_status &= ~data;
425 			return 0;
426 		}
427 		break;
428 	default:
429 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
430 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
431 			if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
432 			    (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
433 				return 1;
434 			if (!msr_info->host_initiated &&
435 			    !(msr & MSR_PMC_FULL_WIDTH_BIT))
436 				data = (s64)(s32)data;
437 			pmc->counter += data - pmc_read_counter(pmc);
438 			if (pmc->perf_event && !pmc->is_paused)
439 				perf_event_period(pmc->perf_event,
440 						  get_sample_period(pmc, data));
441 			return 0;
442 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
443 			pmc->counter += data - pmc_read_counter(pmc);
444 			if (pmc->perf_event && !pmc->is_paused)
445 				perf_event_period(pmc->perf_event,
446 						  get_sample_period(pmc, data));
447 			return 0;
448 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
449 			if (data == pmc->eventsel)
450 				return 0;
451 			if (!(data & pmu->reserved_bits)) {
452 				reprogram_gp_counter(pmc, data);
453 				return 0;
454 			}
455 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false))
456 			return 0;
457 	}
458 
459 	return 1;
460 }
461 
462 static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
463 {
464 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
465 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
466 
467 	struct x86_pmu_capability x86_pmu;
468 	struct kvm_cpuid_entry2 *entry;
469 	union cpuid10_eax eax;
470 	union cpuid10_edx edx;
471 
472 	pmu->nr_arch_gp_counters = 0;
473 	pmu->nr_arch_fixed_counters = 0;
474 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
475 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
476 	pmu->version = 0;
477 	pmu->reserved_bits = 0xffffffff00200000ull;
478 
479 	entry = kvm_find_cpuid_entry(vcpu, 0xa, 0);
480 	if (!entry)
481 		return;
482 	eax.full = entry->eax;
483 	edx.full = entry->edx;
484 
485 	pmu->version = eax.split.version_id;
486 	if (!pmu->version)
487 		return;
488 
489 	perf_get_x86_pmu_capability(&x86_pmu);
490 
491 	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
492 					 x86_pmu.num_counters_gp);
493 	eax.split.bit_width = min_t(int, eax.split.bit_width, x86_pmu.bit_width_gp);
494 	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
495 	eax.split.mask_length = min_t(int, eax.split.mask_length, x86_pmu.events_mask_len);
496 	pmu->available_event_types = ~entry->ebx &
497 					((1ull << eax.split.mask_length) - 1);
498 
499 	if (pmu->version == 1) {
500 		pmu->nr_arch_fixed_counters = 0;
501 	} else {
502 		pmu->nr_arch_fixed_counters =
503 			min_t(int, edx.split.num_counters_fixed,
504 			      x86_pmu.num_counters_fixed);
505 		edx.split.bit_width_fixed = min_t(int,
506 			edx.split.bit_width_fixed, x86_pmu.bit_width_fixed);
507 		pmu->counter_bitmask[KVM_PMC_FIXED] =
508 			((u64)1 << edx.split.bit_width_fixed) - 1;
509 	}
510 
511 	pmu->global_ctrl = ((1ull << pmu->nr_arch_gp_counters) - 1) |
512 		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED);
513 	pmu->global_ctrl_mask = ~pmu->global_ctrl;
514 	pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask
515 			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
516 			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
517 	if (vmx_pt_mode_is_host_guest())
518 		pmu->global_ovf_ctrl_mask &=
519 				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
520 
521 	entry = kvm_find_cpuid_entry(vcpu, 7, 0);
522 	if (entry &&
523 	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
524 	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM)))
525 		pmu->reserved_bits ^= HSW_IN_TX|HSW_IN_TX_CHECKPOINTED;
526 
527 	bitmap_set(pmu->all_valid_pmc_idx,
528 		0, pmu->nr_arch_gp_counters);
529 	bitmap_set(pmu->all_valid_pmc_idx,
530 		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
531 
532 	nested_vmx_pmu_entry_exit_ctls_update(vcpu);
533 
534 	if (intel_pmu_lbr_is_compatible(vcpu))
535 		x86_perf_get_lbr(&lbr_desc->records);
536 	else
537 		lbr_desc->records.nr = 0;
538 
539 	if (lbr_desc->records.nr)
540 		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
541 }
542 
543 static void intel_pmu_init(struct kvm_vcpu *vcpu)
544 {
545 	int i;
546 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
547 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
548 
549 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
550 		pmu->gp_counters[i].type = KVM_PMC_GP;
551 		pmu->gp_counters[i].vcpu = vcpu;
552 		pmu->gp_counters[i].idx = i;
553 		pmu->gp_counters[i].current_config = 0;
554 	}
555 
556 	for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) {
557 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
558 		pmu->fixed_counters[i].vcpu = vcpu;
559 		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
560 		pmu->fixed_counters[i].current_config = 0;
561 	}
562 
563 	vcpu->arch.perf_capabilities = vmx_get_perf_capabilities();
564 	lbr_desc->records.nr = 0;
565 	lbr_desc->event = NULL;
566 	lbr_desc->msr_passthrough = false;
567 }
568 
569 static void intel_pmu_reset(struct kvm_vcpu *vcpu)
570 {
571 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
572 	struct kvm_pmc *pmc = NULL;
573 	int i;
574 
575 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
576 		pmc = &pmu->gp_counters[i];
577 
578 		pmc_stop_counter(pmc);
579 		pmc->counter = pmc->eventsel = 0;
580 	}
581 
582 	for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) {
583 		pmc = &pmu->fixed_counters[i];
584 
585 		pmc_stop_counter(pmc);
586 		pmc->counter = 0;
587 	}
588 
589 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 0;
590 
591 	intel_pmu_release_guest_lbr_event(vcpu);
592 }
593 
594 /*
595  * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
596  *
597  * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
598  * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
599  *
600  * Guest needs to re-enable LBR to resume branches recording.
601  */
602 static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
603 {
604 	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
605 
606 	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
607 		data &= ~DEBUGCTLMSR_LBR;
608 		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
609 	}
610 }
611 
612 static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
613 {
614 	u8 version = vcpu_to_pmu(vcpu)->version;
615 
616 	if (!intel_pmu_lbr_is_enabled(vcpu))
617 		return;
618 
619 	if (version > 1 && version < 4)
620 		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
621 }
622 
623 static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
624 {
625 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
626 	int i;
627 
628 	for (i = 0; i < lbr->nr; i++) {
629 		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
630 		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
631 		if (lbr->info)
632 			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
633 	}
634 
635 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
636 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
637 }
638 
639 static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
640 {
641 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
642 
643 	if (!lbr_desc->msr_passthrough)
644 		return;
645 
646 	vmx_update_intercept_for_lbr_msrs(vcpu, true);
647 	lbr_desc->msr_passthrough = false;
648 }
649 
650 static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
651 {
652 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
653 
654 	if (lbr_desc->msr_passthrough)
655 		return;
656 
657 	vmx_update_intercept_for_lbr_msrs(vcpu, false);
658 	lbr_desc->msr_passthrough = true;
659 }
660 
661 /*
662  * Higher priority host perf events (e.g. cpu pinned) could reclaim the
663  * pmu resources (e.g. LBR) that were assigned to the guest. This is
664  * usually done via ipi calls (more details in perf_install_in_context).
665  *
666  * Before entering the non-root mode (with irq disabled here), double
667  * confirm that the pmu features enabled to the guest are not reclaimed
668  * by higher priority host events. Otherwise, disallow vcpu's access to
669  * the reclaimed features.
670  */
671 void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
672 {
673 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
674 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
675 
676 	if (!lbr_desc->event) {
677 		vmx_disable_lbr_msrs_passthrough(vcpu);
678 		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
679 			goto warn;
680 		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
681 			goto warn;
682 		return;
683 	}
684 
685 	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
686 		vmx_disable_lbr_msrs_passthrough(vcpu);
687 		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
688 		goto warn;
689 	} else
690 		vmx_enable_lbr_msrs_passthrough(vcpu);
691 
692 	return;
693 
694 warn:
695 	pr_warn_ratelimited("kvm: vcpu-%d: fail to passthrough LBR.\n",
696 		vcpu->vcpu_id);
697 }
698 
699 static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
700 {
701 	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
702 		intel_pmu_release_guest_lbr_event(vcpu);
703 }
704 
705 struct kvm_pmu_ops intel_pmu_ops = {
706 	.find_arch_event = intel_find_arch_event,
707 	.find_fixed_event = intel_find_fixed_event,
708 	.pmc_is_enabled = intel_pmc_is_enabled,
709 	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
710 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
711 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
712 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
713 	.is_valid_msr = intel_is_valid_msr,
714 	.get_msr = intel_pmu_get_msr,
715 	.set_msr = intel_pmu_set_msr,
716 	.refresh = intel_pmu_refresh,
717 	.init = intel_pmu_init,
718 	.reset = intel_pmu_reset,
719 	.deliver_pmi = intel_pmu_deliver_pmi,
720 	.cleanup = intel_pmu_cleanup,
721 };
722