xref: /openbmc/linux/arch/x86/kvm/vmx/pmu_intel.c (revision 27e45f2e)
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 	struct kvm_pmc *pmc;
41 	u8 old_fixed_ctr_ctrl = pmu->fixed_ctr_ctrl;
42 	int i;
43 
44 	pmu->fixed_ctr_ctrl = data;
45 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
46 		u8 new_ctrl = fixed_ctrl_field(data, i);
47 		u8 old_ctrl = fixed_ctrl_field(old_fixed_ctr_ctrl, i);
48 
49 		if (old_ctrl == new_ctrl)
50 			continue;
51 
52 		pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i);
53 
54 		__set_bit(INTEL_PMC_IDX_FIXED + i, pmu->pmc_in_use);
55 		kvm_pmu_request_counter_reprogam(pmc);
56 	}
57 }
58 
59 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
60 {
61 	if (pmc_idx < INTEL_PMC_IDX_FIXED) {
62 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
63 				  MSR_P6_EVNTSEL0);
64 	} else {
65 		u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
66 
67 		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
68 	}
69 }
70 
71 static void reprogram_counters(struct kvm_pmu *pmu, u64 diff)
72 {
73 	int bit;
74 	struct kvm_pmc *pmc;
75 
76 	for_each_set_bit(bit, (unsigned long *)&diff, X86_PMC_IDX_MAX) {
77 		pmc = intel_pmc_idx_to_pmc(pmu, bit);
78 		if (pmc)
79 			kvm_pmu_request_counter_reprogam(pmc);
80 	}
81 }
82 
83 static bool intel_hw_event_available(struct kvm_pmc *pmc)
84 {
85 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
86 	u8 event_select = pmc->eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
87 	u8 unit_mask = (pmc->eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
88 	int i;
89 
90 	for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++) {
91 		if (intel_arch_events[i].eventsel != event_select ||
92 		    intel_arch_events[i].unit_mask != unit_mask)
93 			continue;
94 
95 		/* disable event that reported as not present by cpuid */
96 		if ((i < 7) && !(pmu->available_event_types & (1 << i)))
97 			return false;
98 
99 		break;
100 	}
101 
102 	return true;
103 }
104 
105 /* check if a PMC is enabled by comparing it with globl_ctrl bits. */
106 static bool intel_pmc_is_enabled(struct kvm_pmc *pmc)
107 {
108 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
109 
110 	if (!intel_pmu_has_perf_global_ctrl(pmu))
111 		return true;
112 
113 	return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
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 static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
171 {
172 	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
173 	bool ret = false;
174 
175 	if (!intel_pmu_lbr_is_enabled(vcpu))
176 		return ret;
177 
178 	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
179 		(index >= records->from && index < records->from + records->nr) ||
180 		(index >= records->to && index < records->to + records->nr);
181 
182 	if (!ret && records->info)
183 		ret = (index >= records->info && index < records->info + records->nr);
184 
185 	return ret;
186 }
187 
188 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
189 {
190 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
191 	u64 perf_capabilities;
192 	int ret;
193 
194 	switch (msr) {
195 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
196 	case MSR_CORE_PERF_GLOBAL_STATUS:
197 	case MSR_CORE_PERF_GLOBAL_CTRL:
198 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
199 		return intel_pmu_has_perf_global_ctrl(pmu);
200 		break;
201 	case MSR_IA32_PEBS_ENABLE:
202 		ret = vcpu_get_perf_capabilities(vcpu) & PERF_CAP_PEBS_FORMAT;
203 		break;
204 	case MSR_IA32_DS_AREA:
205 		ret = guest_cpuid_has(vcpu, X86_FEATURE_DS);
206 		break;
207 	case MSR_PEBS_DATA_CFG:
208 		perf_capabilities = vcpu_get_perf_capabilities(vcpu);
209 		ret = (perf_capabilities & PERF_CAP_PEBS_BASELINE) &&
210 			((perf_capabilities & PERF_CAP_PEBS_FORMAT) > 3);
211 		break;
212 	default:
213 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
214 			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
215 			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
216 			intel_pmu_is_valid_lbr_msr(vcpu, msr);
217 		break;
218 	}
219 
220 	return ret;
221 }
222 
223 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
224 {
225 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
226 	struct kvm_pmc *pmc;
227 
228 	pmc = get_fixed_pmc(pmu, msr);
229 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
230 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
231 
232 	return pmc;
233 }
234 
235 static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
236 {
237 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
238 
239 	if (lbr_desc->event) {
240 		perf_event_release_kernel(lbr_desc->event);
241 		lbr_desc->event = NULL;
242 		vcpu_to_pmu(vcpu)->event_count--;
243 	}
244 }
245 
246 int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
247 {
248 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
249 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
250 	struct perf_event *event;
251 
252 	/*
253 	 * The perf_event_attr is constructed in the minimum efficient way:
254 	 * - set 'pinned = true' to make it task pinned so that if another
255 	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
256 	 * - set '.exclude_host = true' to record guest branches behavior;
257 	 *
258 	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
259 	 *   schedule the event without a real HW counter but a fake one;
260 	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
261 	 *
262 	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
263 	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
264 	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
265 	 *   event, which helps KVM to save/restore guest LBR records
266 	 *   during host context switches and reduces quite a lot overhead,
267 	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
268 	 */
269 	struct perf_event_attr attr = {
270 		.type = PERF_TYPE_RAW,
271 		.size = sizeof(attr),
272 		.config = INTEL_FIXED_VLBR_EVENT,
273 		.sample_type = PERF_SAMPLE_BRANCH_STACK,
274 		.pinned = true,
275 		.exclude_host = true,
276 		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
277 					PERF_SAMPLE_BRANCH_USER,
278 	};
279 
280 	if (unlikely(lbr_desc->event)) {
281 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
282 		return 0;
283 	}
284 
285 	event = perf_event_create_kernel_counter(&attr, -1,
286 						current, NULL, NULL);
287 	if (IS_ERR(event)) {
288 		pr_debug_ratelimited("%s: failed %ld\n",
289 					__func__, PTR_ERR(event));
290 		return PTR_ERR(event);
291 	}
292 	lbr_desc->event = event;
293 	pmu->event_count++;
294 	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
295 	return 0;
296 }
297 
298 /*
299  * It's safe to access LBR msrs from guest when they have not
300  * been passthrough since the host would help restore or reset
301  * the LBR msrs records when the guest LBR event is scheduled in.
302  */
303 static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
304 				     struct msr_data *msr_info, bool read)
305 {
306 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
307 	u32 index = msr_info->index;
308 
309 	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
310 		return false;
311 
312 	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
313 		goto dummy;
314 
315 	/*
316 	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
317 	 * host at the time the value is read from the msr, and this avoids the
318 	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
319 	 * return 0 on guest reads.
320 	 */
321 	local_irq_disable();
322 	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
323 		if (read)
324 			rdmsrl(index, msr_info->data);
325 		else
326 			wrmsrl(index, msr_info->data);
327 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
328 		local_irq_enable();
329 		return true;
330 	}
331 	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
332 	local_irq_enable();
333 
334 dummy:
335 	if (read)
336 		msr_info->data = 0;
337 	return true;
338 }
339 
340 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
341 {
342 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
343 	struct kvm_pmc *pmc;
344 	u32 msr = msr_info->index;
345 
346 	switch (msr) {
347 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
348 		msr_info->data = pmu->fixed_ctr_ctrl;
349 		return 0;
350 	case MSR_CORE_PERF_GLOBAL_STATUS:
351 		msr_info->data = pmu->global_status;
352 		return 0;
353 	case MSR_CORE_PERF_GLOBAL_CTRL:
354 		msr_info->data = pmu->global_ctrl;
355 		return 0;
356 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
357 		msr_info->data = 0;
358 		return 0;
359 	case MSR_IA32_PEBS_ENABLE:
360 		msr_info->data = pmu->pebs_enable;
361 		return 0;
362 	case MSR_IA32_DS_AREA:
363 		msr_info->data = pmu->ds_area;
364 		return 0;
365 	case MSR_PEBS_DATA_CFG:
366 		msr_info->data = pmu->pebs_data_cfg;
367 		return 0;
368 	default:
369 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
370 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
371 			u64 val = pmc_read_counter(pmc);
372 			msr_info->data =
373 				val & pmu->counter_bitmask[KVM_PMC_GP];
374 			return 0;
375 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
376 			u64 val = pmc_read_counter(pmc);
377 			msr_info->data =
378 				val & pmu->counter_bitmask[KVM_PMC_FIXED];
379 			return 0;
380 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
381 			msr_info->data = pmc->eventsel;
382 			return 0;
383 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true))
384 			return 0;
385 	}
386 
387 	return 1;
388 }
389 
390 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
391 {
392 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
393 	struct kvm_pmc *pmc;
394 	u32 msr = msr_info->index;
395 	u64 data = msr_info->data;
396 	u64 reserved_bits, diff;
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 & pmu->fixed_ctr_ctrl_mask)) {
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 			diff = pmu->global_ctrl ^ data;
418 			pmu->global_ctrl = data;
419 			reprogram_counters(pmu, diff);
420 			return 0;
421 		}
422 		break;
423 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
424 		if (!(data & pmu->global_ovf_ctrl_mask)) {
425 			if (!msr_info->host_initiated)
426 				pmu->global_status &= ~data;
427 			return 0;
428 		}
429 		break;
430 	case MSR_IA32_PEBS_ENABLE:
431 		if (pmu->pebs_enable == data)
432 			return 0;
433 		if (!(data & pmu->pebs_enable_mask)) {
434 			diff = pmu->pebs_enable ^ data;
435 			pmu->pebs_enable = data;
436 			reprogram_counters(pmu, diff);
437 			return 0;
438 		}
439 		break;
440 	case MSR_IA32_DS_AREA:
441 		if (msr_info->host_initiated && data && !guest_cpuid_has(vcpu, X86_FEATURE_DS))
442 			return 1;
443 		if (is_noncanonical_address(data, vcpu))
444 			return 1;
445 		pmu->ds_area = data;
446 		return 0;
447 	case MSR_PEBS_DATA_CFG:
448 		if (pmu->pebs_data_cfg == data)
449 			return 0;
450 		if (!(data & pmu->pebs_data_cfg_mask)) {
451 			pmu->pebs_data_cfg = data;
452 			return 0;
453 		}
454 		break;
455 	default:
456 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
457 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
458 			if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
459 			    (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
460 				return 1;
461 			if (!msr_info->host_initiated &&
462 			    !(msr & MSR_PMC_FULL_WIDTH_BIT))
463 				data = (s64)(s32)data;
464 			pmc->counter += data - pmc_read_counter(pmc);
465 			pmc_update_sample_period(pmc);
466 			return 0;
467 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
468 			pmc->counter += data - pmc_read_counter(pmc);
469 			pmc_update_sample_period(pmc);
470 			return 0;
471 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
472 			if (data == pmc->eventsel)
473 				return 0;
474 			reserved_bits = pmu->reserved_bits;
475 			if ((pmc->idx == 2) &&
476 			    (pmu->raw_event_mask & HSW_IN_TX_CHECKPOINTED))
477 				reserved_bits ^= HSW_IN_TX_CHECKPOINTED;
478 			if (!(data & reserved_bits)) {
479 				pmc->eventsel = data;
480 				kvm_pmu_request_counter_reprogam(pmc);
481 				return 0;
482 			}
483 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false))
484 			return 0;
485 	}
486 
487 	return 1;
488 }
489 
490 static void setup_fixed_pmc_eventsel(struct kvm_pmu *pmu)
491 {
492 	size_t size = ARRAY_SIZE(fixed_pmc_events);
493 	struct kvm_pmc *pmc;
494 	u32 event;
495 	int i;
496 
497 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
498 		pmc = &pmu->fixed_counters[i];
499 		event = fixed_pmc_events[array_index_nospec(i, size)];
500 		pmc->eventsel = (intel_arch_events[event].unit_mask << 8) |
501 			intel_arch_events[event].eventsel;
502 	}
503 }
504 
505 static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
506 {
507 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
508 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
509 	struct kvm_cpuid_entry2 *entry;
510 	union cpuid10_eax eax;
511 	union cpuid10_edx edx;
512 	u64 perf_capabilities;
513 	u64 counter_mask;
514 	int i;
515 
516 	pmu->nr_arch_gp_counters = 0;
517 	pmu->nr_arch_fixed_counters = 0;
518 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
519 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
520 	pmu->version = 0;
521 	pmu->reserved_bits = 0xffffffff00200000ull;
522 	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
523 	pmu->global_ctrl_mask = ~0ull;
524 	pmu->global_ovf_ctrl_mask = ~0ull;
525 	pmu->fixed_ctr_ctrl_mask = ~0ull;
526 	pmu->pebs_enable_mask = ~0ull;
527 	pmu->pebs_data_cfg_mask = ~0ull;
528 
529 	entry = kvm_find_cpuid_entry(vcpu, 0xa);
530 	if (!entry || !vcpu->kvm->arch.enable_pmu)
531 		return;
532 	eax.full = entry->eax;
533 	edx.full = entry->edx;
534 
535 	pmu->version = eax.split.version_id;
536 	if (!pmu->version)
537 		return;
538 
539 	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
540 					 kvm_pmu_cap.num_counters_gp);
541 	eax.split.bit_width = min_t(int, eax.split.bit_width,
542 				    kvm_pmu_cap.bit_width_gp);
543 	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
544 	eax.split.mask_length = min_t(int, eax.split.mask_length,
545 				      kvm_pmu_cap.events_mask_len);
546 	pmu->available_event_types = ~entry->ebx &
547 					((1ull << eax.split.mask_length) - 1);
548 
549 	if (pmu->version == 1) {
550 		pmu->nr_arch_fixed_counters = 0;
551 	} else {
552 		pmu->nr_arch_fixed_counters =
553 			min3(ARRAY_SIZE(fixed_pmc_events),
554 			     (size_t) edx.split.num_counters_fixed,
555 			     (size_t)kvm_pmu_cap.num_counters_fixed);
556 		edx.split.bit_width_fixed = min_t(int, edx.split.bit_width_fixed,
557 						  kvm_pmu_cap.bit_width_fixed);
558 		pmu->counter_bitmask[KVM_PMC_FIXED] =
559 			((u64)1 << edx.split.bit_width_fixed) - 1;
560 		setup_fixed_pmc_eventsel(pmu);
561 	}
562 
563 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++)
564 		pmu->fixed_ctr_ctrl_mask &= ~(0xbull << (i * 4));
565 	counter_mask = ~(((1ull << pmu->nr_arch_gp_counters) - 1) |
566 		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED));
567 	pmu->global_ctrl_mask = counter_mask;
568 	pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask
569 			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
570 			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
571 	if (vmx_pt_mode_is_host_guest())
572 		pmu->global_ovf_ctrl_mask &=
573 				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
574 
575 	entry = kvm_find_cpuid_entry_index(vcpu, 7, 0);
576 	if (entry &&
577 	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
578 	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) {
579 		pmu->reserved_bits ^= HSW_IN_TX;
580 		pmu->raw_event_mask |= (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
581 	}
582 
583 	bitmap_set(pmu->all_valid_pmc_idx,
584 		0, pmu->nr_arch_gp_counters);
585 	bitmap_set(pmu->all_valid_pmc_idx,
586 		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
587 
588 	perf_capabilities = vcpu_get_perf_capabilities(vcpu);
589 	if (cpuid_model_is_consistent(vcpu) &&
590 	    (perf_capabilities & PMU_CAP_LBR_FMT))
591 		x86_perf_get_lbr(&lbr_desc->records);
592 	else
593 		lbr_desc->records.nr = 0;
594 
595 	if (lbr_desc->records.nr)
596 		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
597 
598 	if (perf_capabilities & PERF_CAP_PEBS_FORMAT) {
599 		if (perf_capabilities & PERF_CAP_PEBS_BASELINE) {
600 			pmu->pebs_enable_mask = counter_mask;
601 			pmu->reserved_bits &= ~ICL_EVENTSEL_ADAPTIVE;
602 			for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
603 				pmu->fixed_ctr_ctrl_mask &=
604 					~(1ULL << (INTEL_PMC_IDX_FIXED + i * 4));
605 			}
606 			pmu->pebs_data_cfg_mask = ~0xff00000full;
607 		} else {
608 			pmu->pebs_enable_mask =
609 				~((1ull << pmu->nr_arch_gp_counters) - 1);
610 		}
611 	}
612 }
613 
614 static void intel_pmu_init(struct kvm_vcpu *vcpu)
615 {
616 	int i;
617 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
618 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
619 
620 	for (i = 0; i < KVM_INTEL_PMC_MAX_GENERIC; i++) {
621 		pmu->gp_counters[i].type = KVM_PMC_GP;
622 		pmu->gp_counters[i].vcpu = vcpu;
623 		pmu->gp_counters[i].idx = i;
624 		pmu->gp_counters[i].current_config = 0;
625 	}
626 
627 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
628 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
629 		pmu->fixed_counters[i].vcpu = vcpu;
630 		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
631 		pmu->fixed_counters[i].current_config = 0;
632 	}
633 
634 	lbr_desc->records.nr = 0;
635 	lbr_desc->event = NULL;
636 	lbr_desc->msr_passthrough = false;
637 }
638 
639 static void intel_pmu_reset(struct kvm_vcpu *vcpu)
640 {
641 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
642 	struct kvm_pmc *pmc = NULL;
643 	int i;
644 
645 	for (i = 0; i < KVM_INTEL_PMC_MAX_GENERIC; i++) {
646 		pmc = &pmu->gp_counters[i];
647 
648 		pmc_stop_counter(pmc);
649 		pmc->counter = pmc->prev_counter = pmc->eventsel = 0;
650 	}
651 
652 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
653 		pmc = &pmu->fixed_counters[i];
654 
655 		pmc_stop_counter(pmc);
656 		pmc->counter = pmc->prev_counter = 0;
657 	}
658 
659 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 0;
660 
661 	intel_pmu_release_guest_lbr_event(vcpu);
662 }
663 
664 /*
665  * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
666  *
667  * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
668  * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
669  *
670  * Guest needs to re-enable LBR to resume branches recording.
671  */
672 static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
673 {
674 	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
675 
676 	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
677 		data &= ~DEBUGCTLMSR_LBR;
678 		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
679 	}
680 }
681 
682 static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
683 {
684 	u8 version = vcpu_to_pmu(vcpu)->version;
685 
686 	if (!intel_pmu_lbr_is_enabled(vcpu))
687 		return;
688 
689 	if (version > 1 && version < 4)
690 		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
691 }
692 
693 static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
694 {
695 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
696 	int i;
697 
698 	for (i = 0; i < lbr->nr; i++) {
699 		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
700 		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
701 		if (lbr->info)
702 			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
703 	}
704 
705 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
706 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
707 }
708 
709 static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
710 {
711 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
712 
713 	if (!lbr_desc->msr_passthrough)
714 		return;
715 
716 	vmx_update_intercept_for_lbr_msrs(vcpu, true);
717 	lbr_desc->msr_passthrough = false;
718 }
719 
720 static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
721 {
722 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
723 
724 	if (lbr_desc->msr_passthrough)
725 		return;
726 
727 	vmx_update_intercept_for_lbr_msrs(vcpu, false);
728 	lbr_desc->msr_passthrough = true;
729 }
730 
731 /*
732  * Higher priority host perf events (e.g. cpu pinned) could reclaim the
733  * pmu resources (e.g. LBR) that were assigned to the guest. This is
734  * usually done via ipi calls (more details in perf_install_in_context).
735  *
736  * Before entering the non-root mode (with irq disabled here), double
737  * confirm that the pmu features enabled to the guest are not reclaimed
738  * by higher priority host events. Otherwise, disallow vcpu's access to
739  * the reclaimed features.
740  */
741 void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
742 {
743 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
744 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
745 
746 	if (!lbr_desc->event) {
747 		vmx_disable_lbr_msrs_passthrough(vcpu);
748 		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
749 			goto warn;
750 		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
751 			goto warn;
752 		return;
753 	}
754 
755 	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
756 		vmx_disable_lbr_msrs_passthrough(vcpu);
757 		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
758 		goto warn;
759 	} else
760 		vmx_enable_lbr_msrs_passthrough(vcpu);
761 
762 	return;
763 
764 warn:
765 	pr_warn_ratelimited("kvm: vcpu-%d: fail to passthrough LBR.\n",
766 		vcpu->vcpu_id);
767 }
768 
769 static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
770 {
771 	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
772 		intel_pmu_release_guest_lbr_event(vcpu);
773 }
774 
775 void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
776 {
777 	struct kvm_pmc *pmc = NULL;
778 	int bit, hw_idx;
779 
780 	for_each_set_bit(bit, (unsigned long *)&pmu->global_ctrl,
781 			 X86_PMC_IDX_MAX) {
782 		pmc = intel_pmc_idx_to_pmc(pmu, bit);
783 
784 		if (!pmc || !pmc_speculative_in_use(pmc) ||
785 		    !intel_pmc_is_enabled(pmc) || !pmc->perf_event)
786 			continue;
787 
788 		/*
789 		 * A negative index indicates the event isn't mapped to a
790 		 * physical counter in the host, e.g. due to contention.
791 		 */
792 		hw_idx = pmc->perf_event->hw.idx;
793 		if (hw_idx != pmc->idx && hw_idx > -1)
794 			pmu->host_cross_mapped_mask |= BIT_ULL(hw_idx);
795 	}
796 }
797 
798 struct kvm_pmu_ops intel_pmu_ops __initdata = {
799 	.hw_event_available = intel_hw_event_available,
800 	.pmc_is_enabled = intel_pmc_is_enabled,
801 	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
802 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
803 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
804 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
805 	.is_valid_msr = intel_is_valid_msr,
806 	.get_msr = intel_pmu_get_msr,
807 	.set_msr = intel_pmu_set_msr,
808 	.refresh = intel_pmu_refresh,
809 	.init = intel_pmu_init,
810 	.reset = intel_pmu_reset,
811 	.deliver_pmi = intel_pmu_deliver_pmi,
812 	.cleanup = intel_pmu_cleanup,
813 };
814