• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
reprogram_fixed_counters(struct kvm_pmu * pmu,u64 data)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. */
global_ctrl_changed(struct kvm_pmu * pmu,u64 data)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 
intel_pmc_perf_hw_id(struct kvm_pmc * pmc)71 static unsigned int intel_pmc_perf_hw_id(struct kvm_pmc *pmc)
72 {
73 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
74 	u8 event_select = pmc->eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
75 	u8 unit_mask = (pmc->eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
76 	int i;
77 
78 	for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++)
79 		if (intel_arch_events[i].eventsel == event_select
80 		    && intel_arch_events[i].unit_mask == unit_mask
81 		    && (pmu->available_event_types & (1 << i)))
82 			break;
83 
84 	if (i == ARRAY_SIZE(intel_arch_events))
85 		return PERF_COUNT_HW_MAX;
86 
87 	return intel_arch_events[i].event_type;
88 }
89 
intel_find_fixed_event(int idx)90 static unsigned intel_find_fixed_event(int idx)
91 {
92 	u32 event;
93 	size_t size = ARRAY_SIZE(fixed_pmc_events);
94 
95 	if (idx >= size)
96 		return PERF_COUNT_HW_MAX;
97 
98 	event = fixed_pmc_events[array_index_nospec(idx, size)];
99 	return intel_arch_events[event].event_type;
100 }
101 
102 /* check if a PMC is enabled by comparing it with globl_ctrl bits. */
intel_pmc_is_enabled(struct kvm_pmc * pmc)103 static bool intel_pmc_is_enabled(struct kvm_pmc *pmc)
104 {
105 	struct kvm_pmu *pmu = pmc_to_pmu(pmc);
106 
107 	if (!intel_pmu_has_perf_global_ctrl(pmu))
108 		return true;
109 
110 	return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
111 }
112 
intel_pmc_idx_to_pmc(struct kvm_pmu * pmu,int pmc_idx)113 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
114 {
115 	if (pmc_idx < INTEL_PMC_IDX_FIXED)
116 		return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
117 				  MSR_P6_EVNTSEL0);
118 	else {
119 		u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
120 
121 		return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
122 	}
123 }
124 
125 /* returns 0 if idx's corresponding MSR exists; otherwise returns 1. */
intel_is_valid_rdpmc_ecx(struct kvm_vcpu * vcpu,unsigned int idx)126 static int intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
127 {
128 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
129 	bool fixed = idx & (1u << 30);
130 
131 	idx &= ~(3u << 30);
132 
133 	return (!fixed && idx >= pmu->nr_arch_gp_counters) ||
134 		(fixed && idx >= pmu->nr_arch_fixed_counters);
135 }
136 
intel_rdpmc_ecx_to_pmc(struct kvm_vcpu * vcpu,unsigned int idx,u64 * mask)137 static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
138 					    unsigned int idx, u64 *mask)
139 {
140 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
141 	bool fixed = idx & (1u << 30);
142 	struct kvm_pmc *counters;
143 	unsigned int num_counters;
144 
145 	idx &= ~(3u << 30);
146 	if (fixed) {
147 		counters = pmu->fixed_counters;
148 		num_counters = pmu->nr_arch_fixed_counters;
149 	} else {
150 		counters = pmu->gp_counters;
151 		num_counters = pmu->nr_arch_gp_counters;
152 	}
153 	if (idx >= num_counters)
154 		return NULL;
155 	*mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
156 	return &counters[array_index_nospec(idx, num_counters)];
157 }
158 
vcpu_get_perf_capabilities(struct kvm_vcpu * vcpu)159 static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
160 {
161 	if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
162 		return 0;
163 
164 	return vcpu->arch.perf_capabilities;
165 }
166 
fw_writes_is_enabled(struct kvm_vcpu * vcpu)167 static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
168 {
169 	return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
170 }
171 
get_fw_gp_pmc(struct kvm_pmu * pmu,u32 msr)172 static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
173 {
174 	if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
175 		return NULL;
176 
177 	return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
178 }
179 
intel_pmu_lbr_is_compatible(struct kvm_vcpu * vcpu)180 bool intel_pmu_lbr_is_compatible(struct kvm_vcpu *vcpu)
181 {
182 	/*
183 	 * As a first step, a guest could only enable LBR feature if its
184 	 * cpu model is the same as the host because the LBR registers
185 	 * would be pass-through to the guest and they're model specific.
186 	 */
187 	return boot_cpu_data.x86_model == guest_cpuid_model(vcpu);
188 }
189 
intel_pmu_lbr_is_enabled(struct kvm_vcpu * vcpu)190 bool intel_pmu_lbr_is_enabled(struct kvm_vcpu *vcpu)
191 {
192 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
193 
194 	return lbr->nr && (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_LBR_FMT);
195 }
196 
intel_pmu_is_valid_lbr_msr(struct kvm_vcpu * vcpu,u32 index)197 static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
198 {
199 	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
200 	bool ret = false;
201 
202 	if (!intel_pmu_lbr_is_enabled(vcpu))
203 		return ret;
204 
205 	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
206 		(index >= records->from && index < records->from + records->nr) ||
207 		(index >= records->to && index < records->to + records->nr);
208 
209 	if (!ret && records->info)
210 		ret = (index >= records->info && index < records->info + records->nr);
211 
212 	return ret;
213 }
214 
intel_is_valid_msr(struct kvm_vcpu * vcpu,u32 msr)215 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
216 {
217 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
218 	int ret;
219 
220 	switch (msr) {
221 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
222 	case MSR_CORE_PERF_GLOBAL_STATUS:
223 	case MSR_CORE_PERF_GLOBAL_CTRL:
224 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
225 		return intel_pmu_has_perf_global_ctrl(pmu);
226 		break;
227 	default:
228 		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
229 			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
230 			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
231 			intel_pmu_is_valid_lbr_msr(vcpu, msr);
232 		break;
233 	}
234 
235 	return ret;
236 }
237 
intel_msr_idx_to_pmc(struct kvm_vcpu * vcpu,u32 msr)238 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
239 {
240 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
241 	struct kvm_pmc *pmc;
242 
243 	pmc = get_fixed_pmc(pmu, msr);
244 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
245 	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
246 
247 	return pmc;
248 }
249 
intel_pmu_release_guest_lbr_event(struct kvm_vcpu * vcpu)250 static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
251 {
252 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
253 
254 	if (lbr_desc->event) {
255 		perf_event_release_kernel(lbr_desc->event);
256 		lbr_desc->event = NULL;
257 		vcpu_to_pmu(vcpu)->event_count--;
258 	}
259 }
260 
intel_pmu_create_guest_lbr_event(struct kvm_vcpu * vcpu)261 int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
262 {
263 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
264 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
265 	struct perf_event *event;
266 
267 	/*
268 	 * The perf_event_attr is constructed in the minimum efficient way:
269 	 * - set 'pinned = true' to make it task pinned so that if another
270 	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
271 	 * - set '.exclude_host = true' to record guest branches behavior;
272 	 *
273 	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
274 	 *   schedule the event without a real HW counter but a fake one;
275 	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
276 	 *
277 	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
278 	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
279 	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
280 	 *   event, which helps KVM to save/restore guest LBR records
281 	 *   during host context switches and reduces quite a lot overhead,
282 	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
283 	 */
284 	struct perf_event_attr attr = {
285 		.type = PERF_TYPE_RAW,
286 		.size = sizeof(attr),
287 		.config = INTEL_FIXED_VLBR_EVENT,
288 		.sample_type = PERF_SAMPLE_BRANCH_STACK,
289 		.pinned = true,
290 		.exclude_host = true,
291 		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
292 					PERF_SAMPLE_BRANCH_USER,
293 	};
294 
295 	if (unlikely(lbr_desc->event)) {
296 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
297 		return 0;
298 	}
299 
300 	event = perf_event_create_kernel_counter(&attr, -1,
301 						current, NULL, NULL);
302 	if (IS_ERR(event)) {
303 		pr_debug_ratelimited("%s: failed %ld\n",
304 					__func__, PTR_ERR(event));
305 		return PTR_ERR(event);
306 	}
307 	lbr_desc->event = event;
308 	pmu->event_count++;
309 	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
310 	return 0;
311 }
312 
313 /*
314  * It's safe to access LBR msrs from guest when they have not
315  * been passthrough since the host would help restore or reset
316  * the LBR msrs records when the guest LBR event is scheduled in.
317  */
intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu * vcpu,struct msr_data * msr_info,bool read)318 static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
319 				     struct msr_data *msr_info, bool read)
320 {
321 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
322 	u32 index = msr_info->index;
323 
324 	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
325 		return false;
326 
327 	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
328 		goto dummy;
329 
330 	/*
331 	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
332 	 * host at the time the value is read from the msr, and this avoids the
333 	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
334 	 * return 0 on guest reads.
335 	 */
336 	local_irq_disable();
337 	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
338 		if (read)
339 			rdmsrl(index, msr_info->data);
340 		else
341 			wrmsrl(index, msr_info->data);
342 		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
343 		local_irq_enable();
344 		return true;
345 	}
346 	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
347 	local_irq_enable();
348 
349 dummy:
350 	if (read)
351 		msr_info->data = 0;
352 	return true;
353 }
354 
intel_pmu_get_msr(struct kvm_vcpu * vcpu,struct msr_data * msr_info)355 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
356 {
357 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
358 	struct kvm_pmc *pmc;
359 	u32 msr = msr_info->index;
360 
361 	switch (msr) {
362 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
363 		msr_info->data = pmu->fixed_ctr_ctrl;
364 		return 0;
365 	case MSR_CORE_PERF_GLOBAL_STATUS:
366 		msr_info->data = pmu->global_status;
367 		return 0;
368 	case MSR_CORE_PERF_GLOBAL_CTRL:
369 		msr_info->data = pmu->global_ctrl;
370 		return 0;
371 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
372 		msr_info->data = pmu->global_ovf_ctrl;
373 		return 0;
374 	default:
375 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
376 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
377 			u64 val = pmc_read_counter(pmc);
378 			msr_info->data =
379 				val & pmu->counter_bitmask[KVM_PMC_GP];
380 			return 0;
381 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
382 			u64 val = pmc_read_counter(pmc);
383 			msr_info->data =
384 				val & pmu->counter_bitmask[KVM_PMC_FIXED];
385 			return 0;
386 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
387 			msr_info->data = pmc->eventsel;
388 			return 0;
389 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true))
390 			return 0;
391 	}
392 
393 	return 1;
394 }
395 
intel_pmu_set_msr(struct kvm_vcpu * vcpu,struct msr_data * msr_info)396 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
397 {
398 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
399 	struct kvm_pmc *pmc;
400 	u32 msr = msr_info->index;
401 	u64 data = msr_info->data;
402 	u64 reserved_bits;
403 
404 	switch (msr) {
405 	case MSR_CORE_PERF_FIXED_CTR_CTRL:
406 		if (pmu->fixed_ctr_ctrl == data)
407 			return 0;
408 		if (!(data & pmu->fixed_ctr_ctrl_mask)) {
409 			reprogram_fixed_counters(pmu, data);
410 			return 0;
411 		}
412 		break;
413 	case MSR_CORE_PERF_GLOBAL_STATUS:
414 		if (msr_info->host_initiated) {
415 			pmu->global_status = data;
416 			return 0;
417 		}
418 		break; /* RO MSR */
419 	case MSR_CORE_PERF_GLOBAL_CTRL:
420 		if (pmu->global_ctrl == data)
421 			return 0;
422 		if (kvm_valid_perf_global_ctrl(pmu, data)) {
423 			global_ctrl_changed(pmu, data);
424 			return 0;
425 		}
426 		break;
427 	case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
428 		if (!(data & pmu->global_ovf_ctrl_mask)) {
429 			if (!msr_info->host_initiated)
430 				pmu->global_status &= ~data;
431 			pmu->global_ovf_ctrl = data;
432 			return 0;
433 		}
434 		break;
435 	default:
436 		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
437 		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
438 			if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
439 			    (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
440 				return 1;
441 			if (!msr_info->host_initiated &&
442 			    !(msr & MSR_PMC_FULL_WIDTH_BIT))
443 				data = (s64)(s32)data;
444 			pmc->counter += data - pmc_read_counter(pmc);
445 			pmc_update_sample_period(pmc);
446 			return 0;
447 		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
448 			pmc->counter += data - pmc_read_counter(pmc);
449 			pmc_update_sample_period(pmc);
450 			return 0;
451 		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
452 			if (data == pmc->eventsel)
453 				return 0;
454 			reserved_bits = pmu->reserved_bits;
455 			if ((pmc->idx == 2) &&
456 			    (pmu->raw_event_mask & HSW_IN_TX_CHECKPOINTED))
457 				reserved_bits ^= HSW_IN_TX_CHECKPOINTED;
458 			if (!(data & reserved_bits)) {
459 				reprogram_gp_counter(pmc, data);
460 				return 0;
461 			}
462 		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false))
463 			return 0;
464 	}
465 
466 	return 1;
467 }
468 
intel_pmu_refresh(struct kvm_vcpu * vcpu)469 static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
470 {
471 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
472 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
473 
474 	struct x86_pmu_capability x86_pmu;
475 	struct kvm_cpuid_entry2 *entry;
476 	union cpuid10_eax eax;
477 	union cpuid10_edx edx;
478 	int i;
479 
480 	pmu->nr_arch_gp_counters = 0;
481 	pmu->nr_arch_fixed_counters = 0;
482 	pmu->counter_bitmask[KVM_PMC_GP] = 0;
483 	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
484 	pmu->version = 0;
485 	pmu->reserved_bits = 0xffffffff00200000ull;
486 	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
487 	pmu->global_ctrl_mask = ~0ull;
488 	pmu->global_ovf_ctrl_mask = ~0ull;
489 	pmu->fixed_ctr_ctrl_mask = ~0ull;
490 
491 	entry = kvm_find_cpuid_entry(vcpu, 0xa, 0);
492 	if (!entry)
493 		return;
494 	eax.full = entry->eax;
495 	edx.full = entry->edx;
496 
497 	pmu->version = eax.split.version_id;
498 	if (!pmu->version)
499 		return;
500 
501 	perf_get_x86_pmu_capability(&x86_pmu);
502 
503 	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
504 					 x86_pmu.num_counters_gp);
505 	eax.split.bit_width = min_t(int, eax.split.bit_width, x86_pmu.bit_width_gp);
506 	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
507 	eax.split.mask_length = min_t(int, eax.split.mask_length, x86_pmu.events_mask_len);
508 	pmu->available_event_types = ~entry->ebx &
509 					((1ull << eax.split.mask_length) - 1);
510 
511 	if (pmu->version == 1) {
512 		pmu->nr_arch_fixed_counters = 0;
513 	} else {
514 		pmu->nr_arch_fixed_counters =
515 			min_t(int, edx.split.num_counters_fixed,
516 			      x86_pmu.num_counters_fixed);
517 		edx.split.bit_width_fixed = min_t(int,
518 			edx.split.bit_width_fixed, x86_pmu.bit_width_fixed);
519 		pmu->counter_bitmask[KVM_PMC_FIXED] =
520 			((u64)1 << edx.split.bit_width_fixed) - 1;
521 	}
522 
523 	for (i = 0; i < pmu->nr_arch_fixed_counters; i++)
524 		pmu->fixed_ctr_ctrl_mask &= ~(0xbull << (i * 4));
525 	pmu->global_ctrl = ((1ull << pmu->nr_arch_gp_counters) - 1) |
526 		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED);
527 	pmu->global_ctrl_mask = ~pmu->global_ctrl;
528 	pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask
529 			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
530 			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
531 	if (vmx_pt_mode_is_host_guest())
532 		pmu->global_ovf_ctrl_mask &=
533 				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
534 
535 	entry = kvm_find_cpuid_entry(vcpu, 7, 0);
536 	if (entry &&
537 	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
538 	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) {
539 		pmu->reserved_bits ^= HSW_IN_TX;
540 		pmu->raw_event_mask |= (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
541 	}
542 
543 	bitmap_set(pmu->all_valid_pmc_idx,
544 		0, pmu->nr_arch_gp_counters);
545 	bitmap_set(pmu->all_valid_pmc_idx,
546 		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
547 
548 	nested_vmx_pmu_entry_exit_ctls_update(vcpu);
549 
550 	if (intel_pmu_lbr_is_compatible(vcpu))
551 		x86_perf_get_lbr(&lbr_desc->records);
552 	else
553 		lbr_desc->records.nr = 0;
554 
555 	if (lbr_desc->records.nr)
556 		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
557 }
558 
intel_pmu_init(struct kvm_vcpu * vcpu)559 static void intel_pmu_init(struct kvm_vcpu *vcpu)
560 {
561 	int i;
562 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
563 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
564 
565 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
566 		pmu->gp_counters[i].type = KVM_PMC_GP;
567 		pmu->gp_counters[i].vcpu = vcpu;
568 		pmu->gp_counters[i].idx = i;
569 		pmu->gp_counters[i].current_config = 0;
570 	}
571 
572 	for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) {
573 		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
574 		pmu->fixed_counters[i].vcpu = vcpu;
575 		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
576 		pmu->fixed_counters[i].current_config = 0;
577 	}
578 
579 	vcpu->arch.perf_capabilities = vmx_get_perf_capabilities();
580 	lbr_desc->records.nr = 0;
581 	lbr_desc->event = NULL;
582 	lbr_desc->msr_passthrough = false;
583 }
584 
intel_pmu_reset(struct kvm_vcpu * vcpu)585 static void intel_pmu_reset(struct kvm_vcpu *vcpu)
586 {
587 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
588 	struct kvm_pmc *pmc = NULL;
589 	int i;
590 
591 	for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
592 		pmc = &pmu->gp_counters[i];
593 
594 		pmc_stop_counter(pmc);
595 		pmc->counter = pmc->eventsel = 0;
596 	}
597 
598 	for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) {
599 		pmc = &pmu->fixed_counters[i];
600 
601 		pmc_stop_counter(pmc);
602 		pmc->counter = 0;
603 	}
604 
605 	pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status =
606 		pmu->global_ovf_ctrl = 0;
607 
608 	intel_pmu_release_guest_lbr_event(vcpu);
609 }
610 
611 /*
612  * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
613  *
614  * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
615  * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
616  *
617  * Guest needs to re-enable LBR to resume branches recording.
618  */
intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu * vcpu)619 static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
620 {
621 	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
622 
623 	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
624 		data &= ~DEBUGCTLMSR_LBR;
625 		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
626 	}
627 }
628 
intel_pmu_deliver_pmi(struct kvm_vcpu * vcpu)629 static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
630 {
631 	u8 version = vcpu_to_pmu(vcpu)->version;
632 
633 	if (!intel_pmu_lbr_is_enabled(vcpu))
634 		return;
635 
636 	if (version > 1 && version < 4)
637 		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
638 }
639 
vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu * vcpu,bool set)640 static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
641 {
642 	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
643 	int i;
644 
645 	for (i = 0; i < lbr->nr; i++) {
646 		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
647 		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
648 		if (lbr->info)
649 			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
650 	}
651 
652 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
653 	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
654 }
655 
vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu * vcpu)656 static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
657 {
658 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
659 
660 	if (!lbr_desc->msr_passthrough)
661 		return;
662 
663 	vmx_update_intercept_for_lbr_msrs(vcpu, true);
664 	lbr_desc->msr_passthrough = false;
665 }
666 
vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu * vcpu)667 static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
668 {
669 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
670 
671 	if (lbr_desc->msr_passthrough)
672 		return;
673 
674 	vmx_update_intercept_for_lbr_msrs(vcpu, false);
675 	lbr_desc->msr_passthrough = true;
676 }
677 
678 /*
679  * Higher priority host perf events (e.g. cpu pinned) could reclaim the
680  * pmu resources (e.g. LBR) that were assigned to the guest. This is
681  * usually done via ipi calls (more details in perf_install_in_context).
682  *
683  * Before entering the non-root mode (with irq disabled here), double
684  * confirm that the pmu features enabled to the guest are not reclaimed
685  * by higher priority host events. Otherwise, disallow vcpu's access to
686  * the reclaimed features.
687  */
vmx_passthrough_lbr_msrs(struct kvm_vcpu * vcpu)688 void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
689 {
690 	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
691 	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
692 
693 	if (!lbr_desc->event) {
694 		vmx_disable_lbr_msrs_passthrough(vcpu);
695 		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
696 			goto warn;
697 		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
698 			goto warn;
699 		return;
700 	}
701 
702 	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
703 		vmx_disable_lbr_msrs_passthrough(vcpu);
704 		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
705 		goto warn;
706 	} else
707 		vmx_enable_lbr_msrs_passthrough(vcpu);
708 
709 	return;
710 
711 warn:
712 	pr_warn_ratelimited("kvm: vcpu-%d: fail to passthrough LBR.\n",
713 		vcpu->vcpu_id);
714 }
715 
intel_pmu_cleanup(struct kvm_vcpu * vcpu)716 static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
717 {
718 	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
719 		intel_pmu_release_guest_lbr_event(vcpu);
720 }
721 
722 struct kvm_pmu_ops intel_pmu_ops = {
723 	.pmc_perf_hw_id = intel_pmc_perf_hw_id,
724 	.find_fixed_event = intel_find_fixed_event,
725 	.pmc_is_enabled = intel_pmc_is_enabled,
726 	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
727 	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
728 	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
729 	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
730 	.is_valid_msr = intel_is_valid_msr,
731 	.get_msr = intel_pmu_get_msr,
732 	.set_msr = intel_pmu_set_msr,
733 	.refresh = intel_pmu_refresh,
734 	.init = intel_pmu_init,
735 	.reset = intel_pmu_reset,
736 	.deliver_pmi = intel_pmu_deliver_pmi,
737 	.cleanup = intel_pmu_cleanup,
738 };
739