xref: /linux/arch/x86/kvm/vmx/pmu_intel.c (revision 79997eda0d31bc68203c95ecb978773ee6ce7a1f)
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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
12 
13 #include <linux/types.h>
14 #include <linux/kvm_host.h>
15 #include <linux/perf_event.h>
16 #include <asm/perf_event.h>
17 #include "x86.h"
18 #include "cpuid.h"
19 #include "lapic.h"
20 #include "nested.h"
21 #include "pmu.h"
22 
23 #define MSR_PMC_FULL_WIDTH_BIT      (MSR_IA32_PMC0 - MSR_IA32_PERFCTR0)
24 
25 enum intel_pmu_architectural_events {
26 	/*
27 	 * The order of the architectural events matters as support for each
28 	 * event is enumerated via CPUID using the index of the event.
29 	 */
30 	INTEL_ARCH_CPU_CYCLES,
31 	INTEL_ARCH_INSTRUCTIONS_RETIRED,
32 	INTEL_ARCH_REFERENCE_CYCLES,
33 	INTEL_ARCH_LLC_REFERENCES,
34 	INTEL_ARCH_LLC_MISSES,
35 	INTEL_ARCH_BRANCHES_RETIRED,
36 	INTEL_ARCH_BRANCHES_MISPREDICTED,
37 
38 	NR_REAL_INTEL_ARCH_EVENTS,
39 
40 	/*
41 	 * Pseudo-architectural event used to implement IA32_FIXED_CTR2, a.k.a.
42 	 * TSC reference cycles.  The architectural reference cycles event may
43 	 * or may not actually use the TSC as the reference, e.g. might use the
44 	 * core crystal clock or the bus clock (yeah, "architectural").
45 	 */
46 	PSEUDO_ARCH_REFERENCE_CYCLES = NR_REAL_INTEL_ARCH_EVENTS,
47 	NR_INTEL_ARCH_EVENTS,
48 };
49 
50 static struct {
51 	u8 eventsel;
52 	u8 unit_mask;
53 } const intel_arch_events[] = {
54 	[INTEL_ARCH_CPU_CYCLES]			= { 0x3c, 0x00 },
55 	[INTEL_ARCH_INSTRUCTIONS_RETIRED]	= { 0xc0, 0x00 },
56 	[INTEL_ARCH_REFERENCE_CYCLES]		= { 0x3c, 0x01 },
57 	[INTEL_ARCH_LLC_REFERENCES]		= { 0x2e, 0x4f },
58 	[INTEL_ARCH_LLC_MISSES]			= { 0x2e, 0x41 },
59 	[INTEL_ARCH_BRANCHES_RETIRED]		= { 0xc4, 0x00 },
60 	[INTEL_ARCH_BRANCHES_MISPREDICTED]	= { 0xc5, 0x00 },
61 	[PSEUDO_ARCH_REFERENCE_CYCLES]		= { 0x00, 0x03 },
62 };
63 
64 /* mapping between fixed pmc index and intel_arch_events array */
65 static int fixed_pmc_events[] = {
66 	[0] = INTEL_ARCH_INSTRUCTIONS_RETIRED,
67 	[1] = INTEL_ARCH_CPU_CYCLES,
68 	[2] = PSEUDO_ARCH_REFERENCE_CYCLES,
69 };
70 
71 static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
72 {
73 	struct kvm_pmc *pmc;
74 	u8 old_fixed_ctr_ctrl = pmu->fixed_ctr_ctrl;
75 	int i;
76 
77 	pmu->fixed_ctr_ctrl = data;
78 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
79 		u8 new_ctrl = fixed_ctrl_field(data, i);
80 		u8 old_ctrl = fixed_ctrl_field(old_fixed_ctr_ctrl, i);
81 
82 		if (old_ctrl == new_ctrl)
83 			continue;
84 
85 		pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i);
86 
87 		__set_bit(INTEL_PMC_IDX_FIXED + i, pmu->pmc_in_use);
88 		kvm_pmu_request_counter_reprogram(pmc);
89 	}
90 }
91 
92 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
93 {
94 	if (pmc_idx < INTEL_PMC_IDX_FIXED) {
95 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
96 				  MSR_P6_EVNTSEL0);
97 	} else {
98 		u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
99 
100 		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
101 	}
102 }
103 
104 static bool intel_hw_event_available(struct kvm_pmc *pmc)
105 {
106 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
107 	u8 event_select = pmc->eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
108 	u8 unit_mask = (pmc->eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
109 	int i;
110 
111 	BUILD_BUG_ON(ARRAY_SIZE(intel_arch_events) != NR_INTEL_ARCH_EVENTS);
112 
113 	/*
114 	 * Disallow events reported as unavailable in guest CPUID.  Note, this
115 	 * doesn't apply to pseudo-architectural events.
116 	 */
117 	for (i = 0; i < NR_REAL_INTEL_ARCH_EVENTS; i++) {
118 		if (intel_arch_events[i].eventsel != event_select ||
119 		    intel_arch_events[i].unit_mask != unit_mask)
120 			continue;
121 
122 		return pmu->available_event_types & BIT(i);
123 	}
124 
125 	return true;
126 }
127 
128 static bool intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
129 {
130 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
131 	bool fixed = idx & (1u << 30);
132 
133 	idx &= ~(3u << 30);
134 
135 	return fixed ? idx < pmu->nr_arch_fixed_counters
136 		     : idx < pmu->nr_arch_gp_counters;
137 }
138 
139 static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
140 					    unsigned int idx, u64 *mask)
141 {
142 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
143 	bool fixed = idx & (1u << 30);
144 	struct kvm_pmc *counters;
145 	unsigned int num_counters;
146 
147 	idx &= ~(3u << 30);
148 	if (fixed) {
149 		counters = pmu->fixed_counters;
150 		num_counters = pmu->nr_arch_fixed_counters;
151 	} else {
152 		counters = pmu->gp_counters;
153 		num_counters = pmu->nr_arch_gp_counters;
154 	}
155 	if (idx >= num_counters)
156 		return NULL;
157 	*mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
158 	return &counters[array_index_nospec(idx, num_counters)];
159 }
160 
161 static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
162 {
163 	if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
164 		return 0;
165 
166 	return vcpu->arch.perf_capabilities;
167 }
168 
169 static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
170 {
171 	return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
172 }
173 
174 static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
175 {
176 	if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
177 		return NULL;
178 
179 	return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
180 }
181 
182 static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
183 {
184 	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
185 	bool ret = false;
186 
187 	if (!intel_pmu_lbr_is_enabled(vcpu))
188 		return ret;
189 
190 	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
191 		(index >= records->from && index < records->from + records->nr) ||
192 		(index >= records->to && index < records->to + records->nr);
193 
194 	if (!ret && records->info)
195 		ret = (index >= records->info && index < records->info + records->nr);
196 
197 	return ret;
198 }
199 
200 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
201 {
202 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
203 	u64 perf_capabilities;
204 	int ret;
205 
206 	switch (msr) {
207 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
208 		return kvm_pmu_has_perf_global_ctrl(pmu);
209 	case MSR_IA32_PEBS_ENABLE:
210 		ret = vcpu_get_perf_capabilities(vcpu) & PERF_CAP_PEBS_FORMAT;
211 		break;
212 	case MSR_IA32_DS_AREA:
213 		ret = guest_cpuid_has(vcpu, X86_FEATURE_DS);
214 		break;
215 	case MSR_PEBS_DATA_CFG:
216 		perf_capabilities = vcpu_get_perf_capabilities(vcpu);
217 		ret = (perf_capabilities & PERF_CAP_PEBS_BASELINE) &&
218 			((perf_capabilities & PERF_CAP_PEBS_FORMAT) > 3);
219 		break;
220 	default:
221 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
222 			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
223 			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
224 			intel_pmu_is_valid_lbr_msr(vcpu, msr);
225 		break;
226 	}
227 
228 	return ret;
229 }
230 
231 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
232 {
233 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
234 	struct kvm_pmc *pmc;
235 
236 	pmc = get_fixed_pmc(pmu, msr);
237 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
238 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
239 
240 	return pmc;
241 }
242 
243 static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
244 {
245 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
246 
247 	if (lbr_desc->event) {
248 		perf_event_release_kernel(lbr_desc->event);
249 		lbr_desc->event = NULL;
250 		vcpu_to_pmu(vcpu)->event_count--;
251 	}
252 }
253 
254 int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
255 {
256 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
257 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
258 	struct perf_event *event;
259 
260 	/*
261 	 * The perf_event_attr is constructed in the minimum efficient way:
262 	 * - set 'pinned = true' to make it task pinned so that if another
263 	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
264 	 * - set '.exclude_host = true' to record guest branches behavior;
265 	 *
266 	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
267 	 *   schedule the event without a real HW counter but a fake one;
268 	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
269 	 *
270 	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
271 	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
272 	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
273 	 *   event, which helps KVM to save/restore guest LBR records
274 	 *   during host context switches and reduces quite a lot overhead,
275 	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
276 	 */
277 	struct perf_event_attr attr = {
278 		.type = PERF_TYPE_RAW,
279 		.size = sizeof(attr),
280 		.config = INTEL_FIXED_VLBR_EVENT,
281 		.sample_type = PERF_SAMPLE_BRANCH_STACK,
282 		.pinned = true,
283 		.exclude_host = true,
284 		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
285 					PERF_SAMPLE_BRANCH_USER,
286 	};
287 
288 	if (unlikely(lbr_desc->event)) {
289 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
290 		return 0;
291 	}
292 
293 	event = perf_event_create_kernel_counter(&attr, -1,
294 						current, NULL, NULL);
295 	if (IS_ERR(event)) {
296 		pr_debug_ratelimited("%s: failed %ld\n",
297 					__func__, PTR_ERR(event));
298 		return PTR_ERR(event);
299 	}
300 	lbr_desc->event = event;
301 	pmu->event_count++;
302 	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
303 	return 0;
304 }
305 
306 /*
307  * It's safe to access LBR msrs from guest when they have not
308  * been passthrough since the host would help restore or reset
309  * the LBR msrs records when the guest LBR event is scheduled in.
310  */
311 static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
312 				     struct msr_data *msr_info, bool read)
313 {
314 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
315 	u32 index = msr_info->index;
316 
317 	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
318 		return false;
319 
320 	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
321 		goto dummy;
322 
323 	/*
324 	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
325 	 * host at the time the value is read from the msr, and this avoids the
326 	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
327 	 * return 0 on guest reads.
328 	 */
329 	local_irq_disable();
330 	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
331 		if (read)
332 			rdmsrl(index, msr_info->data);
333 		else
334 			wrmsrl(index, msr_info->data);
335 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
336 		local_irq_enable();
337 		return true;
338 	}
339 	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
340 	local_irq_enable();
341 
342 dummy:
343 	if (read)
344 		msr_info->data = 0;
345 	return true;
346 }
347 
348 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
349 {
350 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
351 	struct kvm_pmc *pmc;
352 	u32 msr = msr_info->index;
353 
354 	switch (msr) {
355 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
356 		msr_info->data = pmu->fixed_ctr_ctrl;
357 		break;
358 	case MSR_IA32_PEBS_ENABLE:
359 		msr_info->data = pmu->pebs_enable;
360 		break;
361 	case MSR_IA32_DS_AREA:
362 		msr_info->data = pmu->ds_area;
363 		break;
364 	case MSR_PEBS_DATA_CFG:
365 		msr_info->data = pmu->pebs_data_cfg;
366 		break;
367 	default:
368 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
369 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
370 			u64 val = pmc_read_counter(pmc);
371 			msr_info->data =
372 				val & pmu->counter_bitmask[KVM_PMC_GP];
373 			break;
374 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
375 			u64 val = pmc_read_counter(pmc);
376 			msr_info->data =
377 				val & pmu->counter_bitmask[KVM_PMC_FIXED];
378 			break;
379 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
380 			msr_info->data = pmc->eventsel;
381 			break;
382 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true)) {
383 			break;
384 		}
385 		return 1;
386 	}
387 
388 	return 0;
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 	u64 reserved_bits, diff;
398 
399 	switch (msr) {
400 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
401 		if (data & pmu->fixed_ctr_ctrl_mask)
402 			return 1;
403 
404 		if (pmu->fixed_ctr_ctrl != data)
405 			reprogram_fixed_counters(pmu, data);
406 		break;
407 	case MSR_IA32_PEBS_ENABLE:
408 		if (data & pmu->pebs_enable_mask)
409 			return 1;
410 
411 		if (pmu->pebs_enable != data) {
412 			diff = pmu->pebs_enable ^ data;
413 			pmu->pebs_enable = data;
414 			reprogram_counters(pmu, diff);
415 		}
416 		break;
417 	case MSR_IA32_DS_AREA:
418 		if (is_noncanonical_address(data, vcpu))
419 			return 1;
420 
421 		pmu->ds_area = data;
422 		break;
423 	case MSR_PEBS_DATA_CFG:
424 		if (data & pmu->pebs_data_cfg_mask)
425 			return 1;
426 
427 		pmu->pebs_data_cfg = data;
428 		break;
429 	default:
430 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
431 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
432 			if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
433 			    (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
434 				return 1;
435 
436 			if (!msr_info->host_initiated &&
437 			    !(msr & MSR_PMC_FULL_WIDTH_BIT))
438 				data = (s64)(s32)data;
439 			pmc_write_counter(pmc, data);
440 			pmc_update_sample_period(pmc);
441 			break;
442 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
443 			pmc_write_counter(pmc, data);
444 			pmc_update_sample_period(pmc);
445 			break;
446 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
447 			reserved_bits = pmu->reserved_bits;
448 			if ((pmc->idx == 2) &&
449 			    (pmu->raw_event_mask & HSW_IN_TX_CHECKPOINTED))
450 				reserved_bits ^= HSW_IN_TX_CHECKPOINTED;
451 			if (data & reserved_bits)
452 				return 1;
453 
454 			if (data != pmc->eventsel) {
455 				pmc->eventsel = data;
456 				kvm_pmu_request_counter_reprogram(pmc);
457 			}
458 			break;
459 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false)) {
460 			break;
461 		}
462 		/* Not a known PMU MSR. */
463 		return 1;
464 	}
465 
466 	return 0;
467 }
468 
469 static void setup_fixed_pmc_eventsel(struct kvm_pmu *pmu)
470 {
471 	int i;
472 
473 	BUILD_BUG_ON(ARRAY_SIZE(fixed_pmc_events) != KVM_PMC_MAX_FIXED);
474 
475 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
476 		int index = array_index_nospec(i, KVM_PMC_MAX_FIXED);
477 		struct kvm_pmc *pmc = &pmu->fixed_counters[index];
478 		u32 event = fixed_pmc_events[index];
479 
480 		pmc->eventsel = (intel_arch_events[event].unit_mask << 8) |
481 				 intel_arch_events[event].eventsel;
482 	}
483 }
484 
485 static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
486 {
487 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
488 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
489 	struct kvm_cpuid_entry2 *entry;
490 	union cpuid10_eax eax;
491 	union cpuid10_edx edx;
492 	u64 perf_capabilities;
493 	u64 counter_mask;
494 	int i;
495 
496 	pmu->nr_arch_gp_counters = 0;
497 	pmu->nr_arch_fixed_counters = 0;
498 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
499 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
500 	pmu->version = 0;
501 	pmu->reserved_bits = 0xffffffff00200000ull;
502 	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
503 	pmu->global_ctrl_mask = ~0ull;
504 	pmu->global_status_mask = ~0ull;
505 	pmu->fixed_ctr_ctrl_mask = ~0ull;
506 	pmu->pebs_enable_mask = ~0ull;
507 	pmu->pebs_data_cfg_mask = ~0ull;
508 
509 	memset(&lbr_desc->records, 0, sizeof(lbr_desc->records));
510 
511 	/*
512 	 * Setting passthrough of LBR MSRs is done only in the VM-Entry loop,
513 	 * and PMU refresh is disallowed after the vCPU has run, i.e. this code
514 	 * should never be reached while KVM is passing through MSRs.
515 	 */
516 	if (KVM_BUG_ON(lbr_desc->msr_passthrough, vcpu->kvm))
517 		return;
518 
519 	entry = kvm_find_cpuid_entry(vcpu, 0xa);
520 	if (!entry || !vcpu->kvm->arch.enable_pmu)
521 		return;
522 	eax.full = entry->eax;
523 	edx.full = entry->edx;
524 
525 	pmu->version = eax.split.version_id;
526 	if (!pmu->version)
527 		return;
528 
529 	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
530 					 kvm_pmu_cap.num_counters_gp);
531 	eax.split.bit_width = min_t(int, eax.split.bit_width,
532 				    kvm_pmu_cap.bit_width_gp);
533 	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
534 	eax.split.mask_length = min_t(int, eax.split.mask_length,
535 				      kvm_pmu_cap.events_mask_len);
536 	pmu->available_event_types = ~entry->ebx &
537 					((1ull << eax.split.mask_length) - 1);
538 
539 	if (pmu->version == 1) {
540 		pmu->nr_arch_fixed_counters = 0;
541 	} else {
542 		pmu->nr_arch_fixed_counters = min_t(int, edx.split.num_counters_fixed,
543 						    kvm_pmu_cap.num_counters_fixed);
544 		edx.split.bit_width_fixed = min_t(int, edx.split.bit_width_fixed,
545 						  kvm_pmu_cap.bit_width_fixed);
546 		pmu->counter_bitmask[KVM_PMC_FIXED] =
547 			((u64)1 << edx.split.bit_width_fixed) - 1;
548 		setup_fixed_pmc_eventsel(pmu);
549 	}
550 
551 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++)
552 		pmu->fixed_ctr_ctrl_mask &= ~(0xbull << (i * 4));
553 	counter_mask = ~(((1ull << pmu->nr_arch_gp_counters) - 1) |
554 		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED));
555 	pmu->global_ctrl_mask = counter_mask;
556 
557 	/*
558 	 * GLOBAL_STATUS and GLOBAL_OVF_CONTROL (a.k.a. GLOBAL_STATUS_RESET)
559 	 * share reserved bit definitions.  The kernel just happens to use
560 	 * OVF_CTRL for the names.
561 	 */
562 	pmu->global_status_mask = pmu->global_ctrl_mask
563 			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
564 			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
565 	if (vmx_pt_mode_is_host_guest())
566 		pmu->global_status_mask &=
567 				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
568 
569 	entry = kvm_find_cpuid_entry_index(vcpu, 7, 0);
570 	if (entry &&
571 	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
572 	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) {
573 		pmu->reserved_bits ^= HSW_IN_TX;
574 		pmu->raw_event_mask |= (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
575 	}
576 
577 	bitmap_set(pmu->all_valid_pmc_idx,
578 		0, pmu->nr_arch_gp_counters);
579 	bitmap_set(pmu->all_valid_pmc_idx,
580 		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
581 
582 	perf_capabilities = vcpu_get_perf_capabilities(vcpu);
583 	if (cpuid_model_is_consistent(vcpu) &&
584 	    (perf_capabilities & PMU_CAP_LBR_FMT))
585 		x86_perf_get_lbr(&lbr_desc->records);
586 	else
587 		lbr_desc->records.nr = 0;
588 
589 	if (lbr_desc->records.nr)
590 		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
591 
592 	if (perf_capabilities & PERF_CAP_PEBS_FORMAT) {
593 		if (perf_capabilities & PERF_CAP_PEBS_BASELINE) {
594 			pmu->pebs_enable_mask = counter_mask;
595 			pmu->reserved_bits &= ~ICL_EVENTSEL_ADAPTIVE;
596 			for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
597 				pmu->fixed_ctr_ctrl_mask &=
598 					~(1ULL << (INTEL_PMC_IDX_FIXED + i * 4));
599 			}
600 			pmu->pebs_data_cfg_mask = ~0xff00000full;
601 		} else {
602 			pmu->pebs_enable_mask =
603 				~((1ull << pmu->nr_arch_gp_counters) - 1);
604 		}
605 	}
606 }
607 
608 static void intel_pmu_init(struct kvm_vcpu *vcpu)
609 {
610 	int i;
611 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
612 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
613 
614 	for (i = 0; i < KVM_INTEL_PMC_MAX_GENERIC; i++) {
615 		pmu->gp_counters[i].type = KVM_PMC_GP;
616 		pmu->gp_counters[i].vcpu = vcpu;
617 		pmu->gp_counters[i].idx = i;
618 		pmu->gp_counters[i].current_config = 0;
619 	}
620 
621 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
622 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
623 		pmu->fixed_counters[i].vcpu = vcpu;
624 		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
625 		pmu->fixed_counters[i].current_config = 0;
626 	}
627 
628 	lbr_desc->records.nr = 0;
629 	lbr_desc->event = NULL;
630 	lbr_desc->msr_passthrough = false;
631 }
632 
633 static void intel_pmu_reset(struct kvm_vcpu *vcpu)
634 {
635 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
636 	struct kvm_pmc *pmc = NULL;
637 	int i;
638 
639 	for (i = 0; i < KVM_INTEL_PMC_MAX_GENERIC; i++) {
640 		pmc = &pmu->gp_counters[i];
641 
642 		pmc_stop_counter(pmc);
643 		pmc->counter = pmc->prev_counter = pmc->eventsel = 0;
644 	}
645 
646 	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
647 		pmc = &pmu->fixed_counters[i];
648 
649 		pmc_stop_counter(pmc);
650 		pmc->counter = pmc->prev_counter = 0;
651 	}
652 
653 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 0;
654 
655 	intel_pmu_release_guest_lbr_event(vcpu);
656 }
657 
658 /*
659  * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
660  *
661  * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
662  * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
663  *
664  * Guest needs to re-enable LBR to resume branches recording.
665  */
666 static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
667 {
668 	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
669 
670 	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
671 		data &= ~DEBUGCTLMSR_LBR;
672 		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
673 	}
674 }
675 
676 static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
677 {
678 	u8 version = vcpu_to_pmu(vcpu)->version;
679 
680 	if (!intel_pmu_lbr_is_enabled(vcpu))
681 		return;
682 
683 	if (version > 1 && version < 4)
684 		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
685 }
686 
687 static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
688 {
689 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
690 	int i;
691 
692 	for (i = 0; i < lbr->nr; i++) {
693 		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
694 		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
695 		if (lbr->info)
696 			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
697 	}
698 
699 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
700 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
701 }
702 
703 static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
704 {
705 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
706 
707 	if (!lbr_desc->msr_passthrough)
708 		return;
709 
710 	vmx_update_intercept_for_lbr_msrs(vcpu, true);
711 	lbr_desc->msr_passthrough = false;
712 }
713 
714 static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
715 {
716 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
717 
718 	if (lbr_desc->msr_passthrough)
719 		return;
720 
721 	vmx_update_intercept_for_lbr_msrs(vcpu, false);
722 	lbr_desc->msr_passthrough = true;
723 }
724 
725 /*
726  * Higher priority host perf events (e.g. cpu pinned) could reclaim the
727  * pmu resources (e.g. LBR) that were assigned to the guest. This is
728  * usually done via ipi calls (more details in perf_install_in_context).
729  *
730  * Before entering the non-root mode (with irq disabled here), double
731  * confirm that the pmu features enabled to the guest are not reclaimed
732  * by higher priority host events. Otherwise, disallow vcpu's access to
733  * the reclaimed features.
734  */
735 void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
736 {
737 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
738 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
739 
740 	if (!lbr_desc->event) {
741 		vmx_disable_lbr_msrs_passthrough(vcpu);
742 		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
743 			goto warn;
744 		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
745 			goto warn;
746 		return;
747 	}
748 
749 	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
750 		vmx_disable_lbr_msrs_passthrough(vcpu);
751 		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
752 		goto warn;
753 	} else
754 		vmx_enable_lbr_msrs_passthrough(vcpu);
755 
756 	return;
757 
758 warn:
759 	pr_warn_ratelimited("vcpu-%d: fail to passthrough LBR.\n", vcpu->vcpu_id);
760 }
761 
762 static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
763 {
764 	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
765 		intel_pmu_release_guest_lbr_event(vcpu);
766 }
767 
768 void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
769 {
770 	struct kvm_pmc *pmc = NULL;
771 	int bit, hw_idx;
772 
773 	for_each_set_bit(bit, (unsigned long *)&pmu->global_ctrl,
774 			 X86_PMC_IDX_MAX) {
775 		pmc = intel_pmc_idx_to_pmc(pmu, bit);
776 
777 		if (!pmc || !pmc_speculative_in_use(pmc) ||
778 		    !pmc_is_globally_enabled(pmc) || !pmc->perf_event)
779 			continue;
780 
781 		/*
782 		 * A negative index indicates the event isn't mapped to a
783 		 * physical counter in the host, e.g. due to contention.
784 		 */
785 		hw_idx = pmc->perf_event->hw.idx;
786 		if (hw_idx != pmc->idx && hw_idx > -1)
787 			pmu->host_cross_mapped_mask |= BIT_ULL(hw_idx);
788 	}
789 }
790 
791 struct kvm_pmu_ops intel_pmu_ops __initdata = {
792 	.hw_event_available = intel_hw_event_available,
793 	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
794 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
795 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
796 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
797 	.is_valid_msr = intel_is_valid_msr,
798 	.get_msr = intel_pmu_get_msr,
799 	.set_msr = intel_pmu_set_msr,
800 	.refresh = intel_pmu_refresh,
801 	.init = intel_pmu_init,
802 	.reset = intel_pmu_reset,
803 	.deliver_pmi = intel_pmu_deliver_pmi,
804 	.cleanup = intel_pmu_cleanup,
805 	.EVENTSEL_EVENT = ARCH_PERFMON_EVENTSEL_EVENT,
806 	.MAX_NR_GP_COUNTERS = KVM_INTEL_PMC_MAX_GENERIC,
807 	.MIN_NR_GP_COUNTERS = 1,
808 };
809