xref: /openbmc/linux/arch/x86/kvm/vmx/pmu_intel.c (revision 2154aca2)
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 		reprogram_counter(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 /* function is called when global control register has been updated. */
72 static void global_ctrl_changed(struct kvm_pmu *pmu, u64 data)
73 {
74 	int bit;
75 	u64 diff = pmu->global_ctrl ^ data;
76 	struct kvm_pmc *pmc;
77 
78 	pmu->global_ctrl = data;
79 
80 	for_each_set_bit(bit, (unsigned long *)&diff, X86_PMC_IDX_MAX) {
81 		pmc = intel_pmc_idx_to_pmc(pmu, bit);
82 		if (pmc)
83 			reprogram_counter(pmc);
84 	}
85 }
86 
87 static bool intel_hw_event_available(struct kvm_pmc *pmc)
88 {
89 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
90 	u8 event_select = pmc->eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
91 	u8 unit_mask = (pmc->eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
92 	int i;
93 
94 	for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++) {
95 		if (intel_arch_events[i].eventsel != event_select ||
96 		    intel_arch_events[i].unit_mask != unit_mask)
97 			continue;
98 
99 		/* disable event that reported as not present by cpuid */
100 		if ((i < 7) && !(pmu->available_event_types & (1 << i)))
101 			return false;
102 
103 		break;
104 	}
105 
106 	return true;
107 }
108 
109 /* check if a PMC is enabled by comparing it with globl_ctrl bits. */
110 static bool intel_pmc_is_enabled(struct kvm_pmc *pmc)
111 {
112 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
113 
114 	if (!intel_pmu_has_perf_global_ctrl(pmu))
115 		return true;
116 
117 	return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
118 }
119 
120 static bool intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
121 {
122 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
123 	bool fixed = idx & (1u << 30);
124 
125 	idx &= ~(3u << 30);
126 
127 	return fixed ? idx < pmu->nr_arch_fixed_counters
128 		     : idx < pmu->nr_arch_gp_counters;
129 }
130 
131 static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
132 					    unsigned int idx, u64 *mask)
133 {
134 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
135 	bool fixed = idx & (1u << 30);
136 	struct kvm_pmc *counters;
137 	unsigned int num_counters;
138 
139 	idx &= ~(3u << 30);
140 	if (fixed) {
141 		counters = pmu->fixed_counters;
142 		num_counters = pmu->nr_arch_fixed_counters;
143 	} else {
144 		counters = pmu->gp_counters;
145 		num_counters = pmu->nr_arch_gp_counters;
146 	}
147 	if (idx >= num_counters)
148 		return NULL;
149 	*mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
150 	return &counters[array_index_nospec(idx, num_counters)];
151 }
152 
153 static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
154 {
155 	if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
156 		return 0;
157 
158 	return vcpu->arch.perf_capabilities;
159 }
160 
161 static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
162 {
163 	return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
164 }
165 
166 static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
167 {
168 	if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
169 		return NULL;
170 
171 	return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
172 }
173 
174 static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
175 {
176 	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
177 	bool ret = false;
178 
179 	if (!intel_pmu_lbr_is_enabled(vcpu))
180 		return ret;
181 
182 	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
183 		(index >= records->from && index < records->from + records->nr) ||
184 		(index >= records->to && index < records->to + records->nr);
185 
186 	if (!ret && records->info)
187 		ret = (index >= records->info && index < records->info + records->nr);
188 
189 	return ret;
190 }
191 
192 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
193 {
194 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
195 	u64 perf_capabilities;
196 	int ret;
197 
198 	switch (msr) {
199 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
200 	case MSR_CORE_PERF_GLOBAL_STATUS:
201 	case MSR_CORE_PERF_GLOBAL_CTRL:
202 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
203 		return intel_pmu_has_perf_global_ctrl(pmu);
204 		break;
205 	case MSR_IA32_PEBS_ENABLE:
206 		ret = vcpu_get_perf_capabilities(vcpu) & PERF_CAP_PEBS_FORMAT;
207 		break;
208 	case MSR_IA32_DS_AREA:
209 		ret = guest_cpuid_has(vcpu, X86_FEATURE_DS);
210 		break;
211 	case MSR_PEBS_DATA_CFG:
212 		perf_capabilities = vcpu_get_perf_capabilities(vcpu);
213 		ret = (perf_capabilities & PERF_CAP_PEBS_BASELINE) &&
214 			((perf_capabilities & PERF_CAP_PEBS_FORMAT) > 3);
215 		break;
216 	default:
217 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
218 			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
219 			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
220 			intel_pmu_is_valid_lbr_msr(vcpu, msr);
221 		break;
222 	}
223 
224 	return ret;
225 }
226 
227 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
228 {
229 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
230 	struct kvm_pmc *pmc;
231 
232 	pmc = get_fixed_pmc(pmu, msr);
233 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
234 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
235 
236 	return pmc;
237 }
238 
239 static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
240 {
241 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
242 
243 	if (lbr_desc->event) {
244 		perf_event_release_kernel(lbr_desc->event);
245 		lbr_desc->event = NULL;
246 		vcpu_to_pmu(vcpu)->event_count--;
247 	}
248 }
249 
250 int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
251 {
252 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
253 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
254 	struct perf_event *event;
255 
256 	/*
257 	 * The perf_event_attr is constructed in the minimum efficient way:
258 	 * - set 'pinned = true' to make it task pinned so that if another
259 	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
260 	 * - set '.exclude_host = true' to record guest branches behavior;
261 	 *
262 	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
263 	 *   schedule the event without a real HW counter but a fake one;
264 	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
265 	 *
266 	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
267 	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
268 	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
269 	 *   event, which helps KVM to save/restore guest LBR records
270 	 *   during host context switches and reduces quite a lot overhead,
271 	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
272 	 */
273 	struct perf_event_attr attr = {
274 		.type = PERF_TYPE_RAW,
275 		.size = sizeof(attr),
276 		.config = INTEL_FIXED_VLBR_EVENT,
277 		.sample_type = PERF_SAMPLE_BRANCH_STACK,
278 		.pinned = true,
279 		.exclude_host = true,
280 		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
281 					PERF_SAMPLE_BRANCH_USER,
282 	};
283 
284 	if (unlikely(lbr_desc->event)) {
285 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
286 		return 0;
287 	}
288 
289 	event = perf_event_create_kernel_counter(&attr, -1,
290 						current, NULL, NULL);
291 	if (IS_ERR(event)) {
292 		pr_debug_ratelimited("%s: failed %ld\n",
293 					__func__, PTR_ERR(event));
294 		return PTR_ERR(event);
295 	}
296 	lbr_desc->event = event;
297 	pmu->event_count++;
298 	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
299 	return 0;
300 }
301 
302 /*
303  * It's safe to access LBR msrs from guest when they have not
304  * been passthrough since the host would help restore or reset
305  * the LBR msrs records when the guest LBR event is scheduled in.
306  */
307 static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
308 				     struct msr_data *msr_info, bool read)
309 {
310 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
311 	u32 index = msr_info->index;
312 
313 	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
314 		return false;
315 
316 	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
317 		goto dummy;
318 
319 	/*
320 	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
321 	 * host at the time the value is read from the msr, and this avoids the
322 	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
323 	 * return 0 on guest reads.
324 	 */
325 	local_irq_disable();
326 	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
327 		if (read)
328 			rdmsrl(index, msr_info->data);
329 		else
330 			wrmsrl(index, msr_info->data);
331 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
332 		local_irq_enable();
333 		return true;
334 	}
335 	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
336 	local_irq_enable();
337 
338 dummy:
339 	if (read)
340 		msr_info->data = 0;
341 	return true;
342 }
343 
344 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
345 {
346 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
347 	struct kvm_pmc *pmc;
348 	u32 msr = msr_info->index;
349 
350 	switch (msr) {
351 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
352 		msr_info->data = pmu->fixed_ctr_ctrl;
353 		return 0;
354 	case MSR_CORE_PERF_GLOBAL_STATUS:
355 		msr_info->data = pmu->global_status;
356 		return 0;
357 	case MSR_CORE_PERF_GLOBAL_CTRL:
358 		msr_info->data = pmu->global_ctrl;
359 		return 0;
360 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
361 		msr_info->data = 0;
362 		return 0;
363 	case MSR_IA32_PEBS_ENABLE:
364 		msr_info->data = pmu->pebs_enable;
365 		return 0;
366 	case MSR_IA32_DS_AREA:
367 		msr_info->data = pmu->ds_area;
368 		return 0;
369 	case MSR_PEBS_DATA_CFG:
370 		msr_info->data = pmu->pebs_data_cfg;
371 		return 0;
372 	default:
373 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
374 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
375 			u64 val = pmc_read_counter(pmc);
376 			msr_info->data =
377 				val & pmu->counter_bitmask[KVM_PMC_GP];
378 			return 0;
379 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
380 			u64 val = pmc_read_counter(pmc);
381 			msr_info->data =
382 				val & pmu->counter_bitmask[KVM_PMC_FIXED];
383 			return 0;
384 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
385 			msr_info->data = pmc->eventsel;
386 			return 0;
387 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true))
388 			return 0;
389 	}
390 
391 	return 1;
392 }
393 
394 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
395 {
396 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
397 	struct kvm_pmc *pmc;
398 	u32 msr = msr_info->index;
399 	u64 data = msr_info->data;
400 	u64 reserved_bits;
401 
402 	switch (msr) {
403 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
404 		if (pmu->fixed_ctr_ctrl == data)
405 			return 0;
406 		if (!(data & pmu->fixed_ctr_ctrl_mask)) {
407 			reprogram_fixed_counters(pmu, data);
408 			return 0;
409 		}
410 		break;
411 	case MSR_CORE_PERF_GLOBAL_STATUS:
412 		if (msr_info->host_initiated) {
413 			pmu->global_status = data;
414 			return 0;
415 		}
416 		break; /* RO MSR */
417 	case MSR_CORE_PERF_GLOBAL_CTRL:
418 		if (pmu->global_ctrl == data)
419 			return 0;
420 		if (kvm_valid_perf_global_ctrl(pmu, data)) {
421 			global_ctrl_changed(pmu, data);
422 			return 0;
423 		}
424 		break;
425 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
426 		if (!(data & pmu->global_ovf_ctrl_mask)) {
427 			if (!msr_info->host_initiated)
428 				pmu->global_status &= ~data;
429 			return 0;
430 		}
431 		break;
432 	case MSR_IA32_PEBS_ENABLE:
433 		if (pmu->pebs_enable == data)
434 			return 0;
435 		if (!(data & pmu->pebs_enable_mask)) {
436 			pmu->pebs_enable = data;
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 				reprogram_counter(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 < 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 	vcpu->arch.perf_capabilities = vmx_get_perf_capabilities();
635 	lbr_desc->records.nr = 0;
636 	lbr_desc->event = NULL;
637 	lbr_desc->msr_passthrough = false;
638 }
639 
640 static void intel_pmu_reset(struct kvm_vcpu *vcpu)
641 {
642 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
643 	struct kvm_pmc *pmc = NULL;
644 	int i;
645 
646 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
647 		pmc = &pmu->gp_counters[i];
648 
649 		pmc_stop_counter(pmc);
650 		pmc->counter = pmc->eventsel = 0;
651 	}
652 
653 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
654 		pmc = &pmu->fixed_counters[i];
655 
656 		pmc_stop_counter(pmc);
657 		pmc->counter = 0;
658 	}
659 
660 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 0;
661 
662 	intel_pmu_release_guest_lbr_event(vcpu);
663 }
664 
665 /*
666  * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
667  *
668  * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
669  * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
670  *
671  * Guest needs to re-enable LBR to resume branches recording.
672  */
673 static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
674 {
675 	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
676 
677 	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
678 		data &= ~DEBUGCTLMSR_LBR;
679 		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
680 	}
681 }
682 
683 static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
684 {
685 	u8 version = vcpu_to_pmu(vcpu)->version;
686 
687 	if (!intel_pmu_lbr_is_enabled(vcpu))
688 		return;
689 
690 	if (version > 1 && version < 4)
691 		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
692 }
693 
694 static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
695 {
696 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
697 	int i;
698 
699 	for (i = 0; i < lbr->nr; i++) {
700 		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
701 		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
702 		if (lbr->info)
703 			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
704 	}
705 
706 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
707 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
708 }
709 
710 static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
711 {
712 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
713 
714 	if (!lbr_desc->msr_passthrough)
715 		return;
716 
717 	vmx_update_intercept_for_lbr_msrs(vcpu, true);
718 	lbr_desc->msr_passthrough = false;
719 }
720 
721 static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
722 {
723 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
724 
725 	if (lbr_desc->msr_passthrough)
726 		return;
727 
728 	vmx_update_intercept_for_lbr_msrs(vcpu, false);
729 	lbr_desc->msr_passthrough = true;
730 }
731 
732 /*
733  * Higher priority host perf events (e.g. cpu pinned) could reclaim the
734  * pmu resources (e.g. LBR) that were assigned to the guest. This is
735  * usually done via ipi calls (more details in perf_install_in_context).
736  *
737  * Before entering the non-root mode (with irq disabled here), double
738  * confirm that the pmu features enabled to the guest are not reclaimed
739  * by higher priority host events. Otherwise, disallow vcpu's access to
740  * the reclaimed features.
741  */
742 void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
743 {
744 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
745 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
746 
747 	if (!lbr_desc->event) {
748 		vmx_disable_lbr_msrs_passthrough(vcpu);
749 		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
750 			goto warn;
751 		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
752 			goto warn;
753 		return;
754 	}
755 
756 	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
757 		vmx_disable_lbr_msrs_passthrough(vcpu);
758 		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
759 		goto warn;
760 	} else
761 		vmx_enable_lbr_msrs_passthrough(vcpu);
762 
763 	return;
764 
765 warn:
766 	pr_warn_ratelimited("kvm: vcpu-%d: fail to passthrough LBR.\n",
767 		vcpu->vcpu_id);
768 }
769 
770 static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
771 {
772 	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
773 		intel_pmu_release_guest_lbr_event(vcpu);
774 }
775 
776 void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
777 {
778 	struct kvm_pmc *pmc = NULL;
779 	int bit;
780 
781 	for_each_set_bit(bit, (unsigned long *)&pmu->global_ctrl,
782 			 X86_PMC_IDX_MAX) {
783 		pmc = intel_pmc_idx_to_pmc(pmu, bit);
784 
785 		if (!pmc || !pmc_speculative_in_use(pmc) ||
786 		    !intel_pmc_is_enabled(pmc))
787 			continue;
788 
789 		if (pmc->perf_event && pmc->idx != pmc->perf_event->hw.idx) {
790 			pmu->host_cross_mapped_mask |=
791 				BIT_ULL(pmc->perf_event->hw.idx);
792 		}
793 	}
794 }
795 
796 struct kvm_pmu_ops intel_pmu_ops __initdata = {
797 	.hw_event_available = intel_hw_event_available,
798 	.pmc_is_enabled = intel_pmc_is_enabled,
799 	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
800 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
801 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
802 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
803 	.is_valid_msr = intel_is_valid_msr,
804 	.get_msr = intel_pmu_get_msr,
805 	.set_msr = intel_pmu_set_msr,
806 	.refresh = intel_pmu_refresh,
807 	.init = intel_pmu_init,
808 	.reset = intel_pmu_reset,
809 	.deliver_pmi = intel_pmu_deliver_pmi,
810 	.cleanup = intel_pmu_cleanup,
811 };
812