Linux Audio

Check our new training course

Loading...
v6.13.7
  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/*
 24 * Perf's "BASE" is wildly misleading, architectural PMUs use bits 31:16 of ECX
 25 * to encode the "type" of counter to read, i.e. this is not a "base".  And to
 26 * further confuse things, non-architectural PMUs use bit 31 as a flag for
 27 * "fast" reads, whereas the "type" is an explicit value.
 28 */
 29#define INTEL_RDPMC_GP		0
 30#define INTEL_RDPMC_FIXED	INTEL_PMC_FIXED_RDPMC_BASE
 31
 32#define INTEL_RDPMC_TYPE_MASK	GENMASK(31, 16)
 33#define INTEL_RDPMC_INDEX_MASK	GENMASK(15, 0)
 
 
 
 
 
 
 
 
 
 
 34
 35#define MSR_PMC_FULL_WIDTH_BIT      (MSR_IA32_PMC0 - MSR_IA32_PERFCTR0)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 36
 37static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
 38{
 39	struct kvm_pmc *pmc;
 40	u64 old_fixed_ctr_ctrl = pmu->fixed_ctr_ctrl;
 41	int i;
 42
 43	pmu->fixed_ctr_ctrl = data;
 44	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
 45		u8 new_ctrl = fixed_ctrl_field(data, i);
 46		u8 old_ctrl = fixed_ctrl_field(old_fixed_ctr_ctrl, i);
 47
 48		if (old_ctrl == new_ctrl)
 49			continue;
 50
 51		pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i);
 52
 53		__set_bit(KVM_FIXED_PMC_BASE_IDX + i, pmu->pmc_in_use);
 54		kvm_pmu_request_counter_reprogram(pmc);
 55	}
 56}
 57
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 58static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
 59					    unsigned int idx, u64 *mask)
 60{
 61	unsigned int type = idx & INTEL_RDPMC_TYPE_MASK;
 62	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 
 63	struct kvm_pmc *counters;
 64	unsigned int num_counters;
 65	u64 bitmask;
 66
 67	/*
 68	 * The encoding of ECX for RDPMC is different for architectural versus
 69	 * non-architecturals PMUs (PMUs with version '0').  For architectural
 70	 * PMUs, bits 31:16 specify the PMC type and bits 15:0 specify the PMC
 71	 * index.  For non-architectural PMUs, bit 31 is a "fast" flag, and
 72	 * bits 30:0 specify the PMC index.
 73	 *
 74	 * Yell and reject attempts to read PMCs for a non-architectural PMU,
 75	 * as KVM doesn't support such PMUs.
 76	 */
 77	if (WARN_ON_ONCE(!pmu->version))
 78		return NULL;
 79
 80	/*
 81	 * General Purpose (GP) PMCs are supported on all PMUs, and fixed PMCs
 82	 * are supported on all architectural PMUs, i.e. on all virtual PMUs
 83	 * supported by KVM.  Note, KVM only emulates fixed PMCs for PMU v2+,
 84	 * but the type itself is still valid, i.e. let RDPMC fail due to
 85	 * accessing a non-existent counter.  Reject attempts to read all other
 86	 * types, which are unknown/unsupported.
 87	 */
 88	switch (type) {
 89	case INTEL_RDPMC_FIXED:
 90		counters = pmu->fixed_counters;
 91		num_counters = pmu->nr_arch_fixed_counters;
 92		bitmask = pmu->counter_bitmask[KVM_PMC_FIXED];
 93		break;
 94	case INTEL_RDPMC_GP:
 95		counters = pmu->gp_counters;
 96		num_counters = pmu->nr_arch_gp_counters;
 97		bitmask = pmu->counter_bitmask[KVM_PMC_GP];
 98		break;
 99	default:
100		return NULL;
101	}
102
103	idx &= INTEL_RDPMC_INDEX_MASK;
104	if (idx >= num_counters)
105		return NULL;
106
107	*mask &= bitmask;
108	return &counters[array_index_nospec(idx, num_counters)];
109}
110
111static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
112{
113	if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
114		return 0;
115
116	return vcpu->arch.perf_capabilities;
117}
118
119static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
120{
121	return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
122}
123
124static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
125{
126	if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
127		return NULL;
128
129	return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
130}
131
132static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
133{
134	struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
135	bool ret = false;
136
137	if (!intel_pmu_lbr_is_enabled(vcpu))
138		return ret;
139
140	ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
141		(index >= records->from && index < records->from + records->nr) ||
142		(index >= records->to && index < records->to + records->nr);
143
144	if (!ret && records->info)
145		ret = (index >= records->info && index < records->info + records->nr);
146
147	return ret;
148}
149
150static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
151{
152	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
153	u64 perf_capabilities;
154	int ret;
155
156	switch (msr) {
157	case MSR_CORE_PERF_FIXED_CTR_CTRL:
158		return kvm_pmu_has_perf_global_ctrl(pmu);
159	case MSR_IA32_PEBS_ENABLE:
160		ret = vcpu_get_perf_capabilities(vcpu) & PERF_CAP_PEBS_FORMAT;
161		break;
162	case MSR_IA32_DS_AREA:
163		ret = guest_cpuid_has(vcpu, X86_FEATURE_DS);
164		break;
165	case MSR_PEBS_DATA_CFG:
166		perf_capabilities = vcpu_get_perf_capabilities(vcpu);
167		ret = (perf_capabilities & PERF_CAP_PEBS_BASELINE) &&
168			((perf_capabilities & PERF_CAP_PEBS_FORMAT) > 3);
169		break;
170	default:
171		ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
172			get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
173			get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
174			intel_pmu_is_valid_lbr_msr(vcpu, msr);
175		break;
176	}
177
178	return ret;
179}
180
181static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
182{
183	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
184	struct kvm_pmc *pmc;
185
186	pmc = get_fixed_pmc(pmu, msr);
187	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
188	pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
189
190	return pmc;
191}
192
193static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
194{
195	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
196
197	if (lbr_desc->event) {
198		perf_event_release_kernel(lbr_desc->event);
199		lbr_desc->event = NULL;
200		vcpu_to_pmu(vcpu)->event_count--;
201	}
202}
203
204int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
205{
206	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
207	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
208	struct perf_event *event;
209
210	/*
211	 * The perf_event_attr is constructed in the minimum efficient way:
212	 * - set 'pinned = true' to make it task pinned so that if another
213	 *   cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
214	 * - set '.exclude_host = true' to record guest branches behavior;
215	 *
216	 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
217	 *   schedule the event without a real HW counter but a fake one;
218	 *   check is_guest_lbr_event() and __intel_get_event_constraints();
219	 *
220	 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
221	 *   'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
222	 *   PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
223	 *   event, which helps KVM to save/restore guest LBR records
224	 *   during host context switches and reduces quite a lot overhead,
225	 *   check branch_user_callstack() and intel_pmu_lbr_sched_task();
226	 */
227	struct perf_event_attr attr = {
228		.type = PERF_TYPE_RAW,
229		.size = sizeof(attr),
230		.config = INTEL_FIXED_VLBR_EVENT,
231		.sample_type = PERF_SAMPLE_BRANCH_STACK,
232		.pinned = true,
233		.exclude_host = true,
234		.branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
235					PERF_SAMPLE_BRANCH_USER,
236	};
237
238	if (unlikely(lbr_desc->event)) {
239		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
240		return 0;
241	}
242
243	event = perf_event_create_kernel_counter(&attr, -1,
244						current, NULL, NULL);
245	if (IS_ERR(event)) {
246		pr_debug_ratelimited("%s: failed %ld\n",
247					__func__, PTR_ERR(event));
248		return PTR_ERR(event);
249	}
250	lbr_desc->event = event;
251	pmu->event_count++;
252	__set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
253	return 0;
254}
255
256/*
257 * It's safe to access LBR msrs from guest when they have not
258 * been passthrough since the host would help restore or reset
259 * the LBR msrs records when the guest LBR event is scheduled in.
260 */
261static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
262				     struct msr_data *msr_info, bool read)
263{
264	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
265	u32 index = msr_info->index;
266
267	if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
268		return false;
269
270	if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
271		goto dummy;
272
273	/*
274	 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
275	 * host at the time the value is read from the msr, and this avoids the
276	 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
277	 * return 0 on guest reads.
278	 */
279	local_irq_disable();
280	if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
281		if (read)
282			rdmsrl(index, msr_info->data);
283		else
284			wrmsrl(index, msr_info->data);
285		__set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
286		local_irq_enable();
287		return true;
288	}
289	clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
290	local_irq_enable();
291
292dummy:
293	if (read)
294		msr_info->data = 0;
295	return true;
296}
297
298static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
299{
300	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
301	struct kvm_pmc *pmc;
302	u32 msr = msr_info->index;
303
304	switch (msr) {
305	case MSR_CORE_PERF_FIXED_CTR_CTRL:
306		msr_info->data = pmu->fixed_ctr_ctrl;
307		break;
308	case MSR_IA32_PEBS_ENABLE:
309		msr_info->data = pmu->pebs_enable;
310		break;
311	case MSR_IA32_DS_AREA:
312		msr_info->data = pmu->ds_area;
313		break;
314	case MSR_PEBS_DATA_CFG:
315		msr_info->data = pmu->pebs_data_cfg;
316		break;
317	default:
318		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
319		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
320			u64 val = pmc_read_counter(pmc);
321			msr_info->data =
322				val & pmu->counter_bitmask[KVM_PMC_GP];
323			break;
324		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
325			u64 val = pmc_read_counter(pmc);
326			msr_info->data =
327				val & pmu->counter_bitmask[KVM_PMC_FIXED];
328			break;
329		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
330			msr_info->data = pmc->eventsel;
331			break;
332		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true)) {
333			break;
334		}
335		return 1;
336	}
337
338	return 0;
339}
340
341static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
342{
343	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
344	struct kvm_pmc *pmc;
345	u32 msr = msr_info->index;
346	u64 data = msr_info->data;
347	u64 reserved_bits, diff;
348
349	switch (msr) {
350	case MSR_CORE_PERF_FIXED_CTR_CTRL:
351		if (data & pmu->fixed_ctr_ctrl_rsvd)
352			return 1;
353
354		if (pmu->fixed_ctr_ctrl != data)
355			reprogram_fixed_counters(pmu, data);
356		break;
357	case MSR_IA32_PEBS_ENABLE:
358		if (data & pmu->pebs_enable_rsvd)
359			return 1;
360
361		if (pmu->pebs_enable != data) {
362			diff = pmu->pebs_enable ^ data;
363			pmu->pebs_enable = data;
364			reprogram_counters(pmu, diff);
365		}
366		break;
367	case MSR_IA32_DS_AREA:
368		if (is_noncanonical_msr_address(data, vcpu))
369			return 1;
370
371		pmu->ds_area = data;
372		break;
373	case MSR_PEBS_DATA_CFG:
374		if (data & pmu->pebs_data_cfg_rsvd)
375			return 1;
376
377		pmu->pebs_data_cfg = data;
378		break;
379	default:
380		if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
381		    (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
382			if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
383			    (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
384				return 1;
385
386			if (!msr_info->host_initiated &&
387			    !(msr & MSR_PMC_FULL_WIDTH_BIT))
388				data = (s64)(s32)data;
389			pmc_write_counter(pmc, data);
390			break;
391		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
392			pmc_write_counter(pmc, data);
393			break;
394		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
395			reserved_bits = pmu->reserved_bits;
396			if ((pmc->idx == 2) &&
397			    (pmu->raw_event_mask & HSW_IN_TX_CHECKPOINTED))
398				reserved_bits ^= HSW_IN_TX_CHECKPOINTED;
399			if (data & reserved_bits)
400				return 1;
401
402			if (data != pmc->eventsel) {
403				pmc->eventsel = data;
404				kvm_pmu_request_counter_reprogram(pmc);
405			}
406			break;
407		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false)) {
408			break;
409		}
410		/* Not a known PMU MSR. */
411		return 1;
412	}
413
414	return 0;
415}
416
417/*
418 * Map fixed counter events to architectural general purpose event encodings.
419 * Perf doesn't provide APIs to allow KVM to directly program a fixed counter,
420 * and so KVM instead programs the architectural event to effectively request
421 * the fixed counter.  Perf isn't guaranteed to use a fixed counter and may
422 * instead program the encoding into a general purpose counter, e.g. if a
423 * different perf_event is already utilizing the requested counter, but the end
424 * result is the same (ignoring the fact that using a general purpose counter
425 * will likely exacerbate counter contention).
426 *
427 * Forcibly inlined to allow asserting on @index at build time, and there should
428 * never be more than one user.
429 */
430static __always_inline u64 intel_get_fixed_pmc_eventsel(unsigned int index)
431{
432	const enum perf_hw_id fixed_pmc_perf_ids[] = {
433		[0] = PERF_COUNT_HW_INSTRUCTIONS,
434		[1] = PERF_COUNT_HW_CPU_CYCLES,
435		[2] = PERF_COUNT_HW_REF_CPU_CYCLES,
436	};
437	u64 eventsel;
438
439	BUILD_BUG_ON(ARRAY_SIZE(fixed_pmc_perf_ids) != KVM_MAX_NR_INTEL_FIXED_COUTNERS);
440	BUILD_BUG_ON(index >= KVM_MAX_NR_INTEL_FIXED_COUTNERS);
441
442	/*
443	 * Yell if perf reports support for a fixed counter but perf doesn't
444	 * have a known encoding for the associated general purpose event.
445	 */
446	eventsel = perf_get_hw_event_config(fixed_pmc_perf_ids[index]);
447	WARN_ON_ONCE(!eventsel && index < kvm_pmu_cap.num_counters_fixed);
448	return eventsel;
449}
450
451static void intel_pmu_enable_fixed_counter_bits(struct kvm_pmu *pmu, u64 bits)
452{
453	int i;
454
455	for (i = 0; i < pmu->nr_arch_fixed_counters; i++)
456		pmu->fixed_ctr_ctrl_rsvd &= ~intel_fixed_bits_by_idx(i, bits);
 
457}
458
459static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
460{
461	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
462	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
463	struct kvm_cpuid_entry2 *entry;
464	union cpuid10_eax eax;
465	union cpuid10_edx edx;
466	u64 perf_capabilities;
467	u64 counter_rsvd;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
468
469	memset(&lbr_desc->records, 0, sizeof(lbr_desc->records));
470
471	/*
472	 * Setting passthrough of LBR MSRs is done only in the VM-Entry loop,
473	 * and PMU refresh is disallowed after the vCPU has run, i.e. this code
474	 * should never be reached while KVM is passing through MSRs.
475	 */
476	if (KVM_BUG_ON(lbr_desc->msr_passthrough, vcpu->kvm))
477		return;
478
479	entry = kvm_find_cpuid_entry(vcpu, 0xa);
480	if (!entry)
481		return;
482
483	eax.full = entry->eax;
484	edx.full = entry->edx;
485
486	pmu->version = eax.split.version_id;
487	if (!pmu->version)
488		return;
489
490	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
491					 kvm_pmu_cap.num_counters_gp);
492	eax.split.bit_width = min_t(int, eax.split.bit_width,
493				    kvm_pmu_cap.bit_width_gp);
494	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
495	eax.split.mask_length = min_t(int, eax.split.mask_length,
496				      kvm_pmu_cap.events_mask_len);
497	pmu->available_event_types = ~entry->ebx &
498					((1ull << eax.split.mask_length) - 1);
499
500	if (pmu->version == 1) {
501		pmu->nr_arch_fixed_counters = 0;
502	} else {
503		pmu->nr_arch_fixed_counters = min_t(int, edx.split.num_counters_fixed,
504						    kvm_pmu_cap.num_counters_fixed);
505		edx.split.bit_width_fixed = min_t(int, edx.split.bit_width_fixed,
506						  kvm_pmu_cap.bit_width_fixed);
507		pmu->counter_bitmask[KVM_PMC_FIXED] =
508			((u64)1 << edx.split.bit_width_fixed) - 1;
 
509	}
510
511	intel_pmu_enable_fixed_counter_bits(pmu, INTEL_FIXED_0_KERNEL |
512						 INTEL_FIXED_0_USER |
513						 INTEL_FIXED_0_ENABLE_PMI);
514
515	counter_rsvd = ~(((1ull << pmu->nr_arch_gp_counters) - 1) |
516		(((1ull << pmu->nr_arch_fixed_counters) - 1) << KVM_FIXED_PMC_BASE_IDX));
517	pmu->global_ctrl_rsvd = counter_rsvd;
518
519	/*
520	 * GLOBAL_STATUS and GLOBAL_OVF_CONTROL (a.k.a. GLOBAL_STATUS_RESET)
521	 * share reserved bit definitions.  The kernel just happens to use
522	 * OVF_CTRL for the names.
523	 */
524	pmu->global_status_rsvd = pmu->global_ctrl_rsvd
525			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
526			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
527	if (vmx_pt_mode_is_host_guest())
528		pmu->global_status_rsvd &=
529				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
530
531	entry = kvm_find_cpuid_entry_index(vcpu, 7, 0);
532	if (entry &&
533	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
534	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) {
535		pmu->reserved_bits ^= HSW_IN_TX;
536		pmu->raw_event_mask |= (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
537	}
538
539	bitmap_set(pmu->all_valid_pmc_idx,
540		0, pmu->nr_arch_gp_counters);
541	bitmap_set(pmu->all_valid_pmc_idx,
542		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
543
544	perf_capabilities = vcpu_get_perf_capabilities(vcpu);
545	if (cpuid_model_is_consistent(vcpu) &&
546	    (perf_capabilities & PMU_CAP_LBR_FMT))
547		memcpy(&lbr_desc->records, &vmx_lbr_caps, sizeof(vmx_lbr_caps));
548	else
549		lbr_desc->records.nr = 0;
550
551	if (lbr_desc->records.nr)
552		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
553
554	if (perf_capabilities & PERF_CAP_PEBS_FORMAT) {
555		if (perf_capabilities & PERF_CAP_PEBS_BASELINE) {
556			pmu->pebs_enable_rsvd = counter_rsvd;
557			pmu->reserved_bits &= ~ICL_EVENTSEL_ADAPTIVE;
558			pmu->pebs_data_cfg_rsvd = ~0xff00000full;
559			intel_pmu_enable_fixed_counter_bits(pmu, ICL_FIXED_0_ADAPTIVE);
 
 
 
560		} else {
561			pmu->pebs_enable_rsvd =
562				~((1ull << pmu->nr_arch_gp_counters) - 1);
563		}
564	}
565}
566
567static void intel_pmu_init(struct kvm_vcpu *vcpu)
568{
569	int i;
570	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
571	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
572
573	for (i = 0; i < KVM_MAX_NR_INTEL_GP_COUNTERS; i++) {
574		pmu->gp_counters[i].type = KVM_PMC_GP;
575		pmu->gp_counters[i].vcpu = vcpu;
576		pmu->gp_counters[i].idx = i;
577		pmu->gp_counters[i].current_config = 0;
578	}
579
580	for (i = 0; i < KVM_MAX_NR_INTEL_FIXED_COUTNERS; i++) {
581		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
582		pmu->fixed_counters[i].vcpu = vcpu;
583		pmu->fixed_counters[i].idx = i + KVM_FIXED_PMC_BASE_IDX;
584		pmu->fixed_counters[i].current_config = 0;
585		pmu->fixed_counters[i].eventsel = intel_get_fixed_pmc_eventsel(i);
586	}
587
588	lbr_desc->records.nr = 0;
589	lbr_desc->event = NULL;
590	lbr_desc->msr_passthrough = false;
591}
592
593static void intel_pmu_reset(struct kvm_vcpu *vcpu)
594{
595	intel_pmu_release_guest_lbr_event(vcpu);
596}
597
598/*
599 * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
600 *
601 * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
602 * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
603 *
604 * Guest needs to re-enable LBR to resume branches recording.
605 */
606static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
607{
608	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
609
610	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
611		data &= ~DEBUGCTLMSR_LBR;
612		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
613	}
614}
615
616static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
617{
618	u8 version = vcpu_to_pmu(vcpu)->version;
619
620	if (!intel_pmu_lbr_is_enabled(vcpu))
621		return;
622
623	if (version > 1 && version < 4)
624		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
625}
626
627static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
628{
629	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
630	int i;
631
632	for (i = 0; i < lbr->nr; i++) {
633		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
634		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
635		if (lbr->info)
636			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
637	}
638
639	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
640	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
641}
642
643static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
644{
645	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
646
647	if (!lbr_desc->msr_passthrough)
648		return;
649
650	vmx_update_intercept_for_lbr_msrs(vcpu, true);
651	lbr_desc->msr_passthrough = false;
652}
653
654static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
655{
656	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
657
658	if (lbr_desc->msr_passthrough)
659		return;
660
661	vmx_update_intercept_for_lbr_msrs(vcpu, false);
662	lbr_desc->msr_passthrough = true;
663}
664
665/*
666 * Higher priority host perf events (e.g. cpu pinned) could reclaim the
667 * pmu resources (e.g. LBR) that were assigned to the guest. This is
668 * usually done via ipi calls (more details in perf_install_in_context).
669 *
670 * Before entering the non-root mode (with irq disabled here), double
671 * confirm that the pmu features enabled to the guest are not reclaimed
672 * by higher priority host events. Otherwise, disallow vcpu's access to
673 * the reclaimed features.
674 */
675void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
676{
677	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
678	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
679
680	if (!lbr_desc->event) {
681		vmx_disable_lbr_msrs_passthrough(vcpu);
682		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
683			goto warn;
684		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
685			goto warn;
686		return;
687	}
688
689	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
690		vmx_disable_lbr_msrs_passthrough(vcpu);
691		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
692		goto warn;
693	} else
694		vmx_enable_lbr_msrs_passthrough(vcpu);
695
696	return;
697
698warn:
699	pr_warn_ratelimited("vcpu-%d: fail to passthrough LBR.\n", vcpu->vcpu_id);
700}
701
702static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
703{
704	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
705		intel_pmu_release_guest_lbr_event(vcpu);
706}
707
708void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
709{
710	struct kvm_pmc *pmc = NULL;
711	int bit, hw_idx;
712
713	kvm_for_each_pmc(pmu, pmc, bit, (unsigned long *)&pmu->global_ctrl) {
714		if (!pmc_speculative_in_use(pmc) ||
 
 
 
715		    !pmc_is_globally_enabled(pmc) || !pmc->perf_event)
716			continue;
717
718		/*
719		 * A negative index indicates the event isn't mapped to a
720		 * physical counter in the host, e.g. due to contention.
721		 */
722		hw_idx = pmc->perf_event->hw.idx;
723		if (hw_idx != pmc->idx && hw_idx > -1)
724			pmu->host_cross_mapped_mask |= BIT_ULL(hw_idx);
725	}
726}
727
728struct kvm_pmu_ops intel_pmu_ops __initdata = {
 
 
729	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
730	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
 
731	.is_valid_msr = intel_is_valid_msr,
732	.get_msr = intel_pmu_get_msr,
733	.set_msr = intel_pmu_set_msr,
734	.refresh = intel_pmu_refresh,
735	.init = intel_pmu_init,
736	.reset = intel_pmu_reset,
737	.deliver_pmi = intel_pmu_deliver_pmi,
738	.cleanup = intel_pmu_cleanup,
739	.EVENTSEL_EVENT = ARCH_PERFMON_EVENTSEL_EVENT,
740	.MAX_NR_GP_COUNTERS = KVM_MAX_NR_INTEL_GP_COUNTERS,
741	.MIN_NR_GP_COUNTERS = 1,
742};
v6.8
  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
 25enum 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
 50static 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 */
 65static int fixed_pmc_events[] = {
 66	[0] = INTEL_ARCH_INSTRUCTIONS_RETIRED,
 67	[1] = INTEL_ARCH_CPU_CYCLES,
 68	[2] = PSEUDO_ARCH_REFERENCE_CYCLES,
 69};
 70
 71static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data)
 72{
 73	struct kvm_pmc *pmc;
 74	u64 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
 92static 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
104static 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
128static 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
139static 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
161static 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
169static 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
174static 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
182static 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
200static 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
231static 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
243static 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
254int 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 */
311static 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
342dummy:
343	if (read)
344		msr_info->data = 0;
345	return true;
346}
347
348static 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
391static 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			break;
441		} else if ((pmc = get_fixed_pmc(pmu, msr))) {
442			pmc_write_counter(pmc, data);
443			break;
444		} else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
445			reserved_bits = pmu->reserved_bits;
446			if ((pmc->idx == 2) &&
447			    (pmu->raw_event_mask & HSW_IN_TX_CHECKPOINTED))
448				reserved_bits ^= HSW_IN_TX_CHECKPOINTED;
449			if (data & reserved_bits)
450				return 1;
451
452			if (data != pmc->eventsel) {
453				pmc->eventsel = data;
454				kvm_pmu_request_counter_reprogram(pmc);
455			}
456			break;
457		} else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false)) {
458			break;
459		}
460		/* Not a known PMU MSR. */
461		return 1;
462	}
463
464	return 0;
465}
466
467static void setup_fixed_pmc_eventsel(struct kvm_pmu *pmu)
 
 
 
 
 
 
 
 
 
 
 
 
 
468{
469	int i;
 
 
 
 
 
470
471	BUILD_BUG_ON(ARRAY_SIZE(fixed_pmc_events) != KVM_PMC_MAX_FIXED);
 
472
473	for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
474		int index = array_index_nospec(i, KVM_PMC_MAX_FIXED);
475		struct kvm_pmc *pmc = &pmu->fixed_counters[index];
476		u32 event = fixed_pmc_events[index];
 
 
 
 
 
 
 
 
477
478		pmc->eventsel = (intel_arch_events[event].unit_mask << 8) |
479				 intel_arch_events[event].eventsel;
480	}
481}
482
483static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
484{
485	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
486	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
487	struct kvm_cpuid_entry2 *entry;
488	union cpuid10_eax eax;
489	union cpuid10_edx edx;
490	u64 perf_capabilities;
491	u64 counter_mask;
492	int i;
493
494	pmu->nr_arch_gp_counters = 0;
495	pmu->nr_arch_fixed_counters = 0;
496	pmu->counter_bitmask[KVM_PMC_GP] = 0;
497	pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
498	pmu->version = 0;
499	pmu->reserved_bits = 0xffffffff00200000ull;
500	pmu->raw_event_mask = X86_RAW_EVENT_MASK;
501	pmu->global_ctrl_mask = ~0ull;
502	pmu->global_status_mask = ~0ull;
503	pmu->fixed_ctr_ctrl_mask = ~0ull;
504	pmu->pebs_enable_mask = ~0ull;
505	pmu->pebs_data_cfg_mask = ~0ull;
506
507	memset(&lbr_desc->records, 0, sizeof(lbr_desc->records));
508
509	/*
510	 * Setting passthrough of LBR MSRs is done only in the VM-Entry loop,
511	 * and PMU refresh is disallowed after the vCPU has run, i.e. this code
512	 * should never be reached while KVM is passing through MSRs.
513	 */
514	if (KVM_BUG_ON(lbr_desc->msr_passthrough, vcpu->kvm))
515		return;
516
517	entry = kvm_find_cpuid_entry(vcpu, 0xa);
518	if (!entry || !vcpu->kvm->arch.enable_pmu)
519		return;
 
520	eax.full = entry->eax;
521	edx.full = entry->edx;
522
523	pmu->version = eax.split.version_id;
524	if (!pmu->version)
525		return;
526
527	pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
528					 kvm_pmu_cap.num_counters_gp);
529	eax.split.bit_width = min_t(int, eax.split.bit_width,
530				    kvm_pmu_cap.bit_width_gp);
531	pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
532	eax.split.mask_length = min_t(int, eax.split.mask_length,
533				      kvm_pmu_cap.events_mask_len);
534	pmu->available_event_types = ~entry->ebx &
535					((1ull << eax.split.mask_length) - 1);
536
537	if (pmu->version == 1) {
538		pmu->nr_arch_fixed_counters = 0;
539	} else {
540		pmu->nr_arch_fixed_counters = min_t(int, edx.split.num_counters_fixed,
541						    kvm_pmu_cap.num_counters_fixed);
542		edx.split.bit_width_fixed = min_t(int, edx.split.bit_width_fixed,
543						  kvm_pmu_cap.bit_width_fixed);
544		pmu->counter_bitmask[KVM_PMC_FIXED] =
545			((u64)1 << edx.split.bit_width_fixed) - 1;
546		setup_fixed_pmc_eventsel(pmu);
547	}
548
549	for (i = 0; i < pmu->nr_arch_fixed_counters; i++)
550		pmu->fixed_ctr_ctrl_mask &= ~(0xbull << (i * 4));
551	counter_mask = ~(((1ull << pmu->nr_arch_gp_counters) - 1) |
552		(((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED));
553	pmu->global_ctrl_mask = counter_mask;
 
 
554
555	/*
556	 * GLOBAL_STATUS and GLOBAL_OVF_CONTROL (a.k.a. GLOBAL_STATUS_RESET)
557	 * share reserved bit definitions.  The kernel just happens to use
558	 * OVF_CTRL for the names.
559	 */
560	pmu->global_status_mask = pmu->global_ctrl_mask
561			& ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
562			    MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
563	if (vmx_pt_mode_is_host_guest())
564		pmu->global_status_mask &=
565				~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
566
567	entry = kvm_find_cpuid_entry_index(vcpu, 7, 0);
568	if (entry &&
569	    (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
570	    (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) {
571		pmu->reserved_bits ^= HSW_IN_TX;
572		pmu->raw_event_mask |= (HSW_IN_TX|HSW_IN_TX_CHECKPOINTED);
573	}
574
575	bitmap_set(pmu->all_valid_pmc_idx,
576		0, pmu->nr_arch_gp_counters);
577	bitmap_set(pmu->all_valid_pmc_idx,
578		INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
579
580	perf_capabilities = vcpu_get_perf_capabilities(vcpu);
581	if (cpuid_model_is_consistent(vcpu) &&
582	    (perf_capabilities & PMU_CAP_LBR_FMT))
583		x86_perf_get_lbr(&lbr_desc->records);
584	else
585		lbr_desc->records.nr = 0;
586
587	if (lbr_desc->records.nr)
588		bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
589
590	if (perf_capabilities & PERF_CAP_PEBS_FORMAT) {
591		if (perf_capabilities & PERF_CAP_PEBS_BASELINE) {
592			pmu->pebs_enable_mask = counter_mask;
593			pmu->reserved_bits &= ~ICL_EVENTSEL_ADAPTIVE;
594			for (i = 0; i < pmu->nr_arch_fixed_counters; i++) {
595				pmu->fixed_ctr_ctrl_mask &=
596					~(1ULL << (INTEL_PMC_IDX_FIXED + i * 4));
597			}
598			pmu->pebs_data_cfg_mask = ~0xff00000full;
599		} else {
600			pmu->pebs_enable_mask =
601				~((1ull << pmu->nr_arch_gp_counters) - 1);
602		}
603	}
604}
605
606static void intel_pmu_init(struct kvm_vcpu *vcpu)
607{
608	int i;
609	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
610	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
611
612	for (i = 0; i < KVM_INTEL_PMC_MAX_GENERIC; i++) {
613		pmu->gp_counters[i].type = KVM_PMC_GP;
614		pmu->gp_counters[i].vcpu = vcpu;
615		pmu->gp_counters[i].idx = i;
616		pmu->gp_counters[i].current_config = 0;
617	}
618
619	for (i = 0; i < KVM_PMC_MAX_FIXED; i++) {
620		pmu->fixed_counters[i].type = KVM_PMC_FIXED;
621		pmu->fixed_counters[i].vcpu = vcpu;
622		pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
623		pmu->fixed_counters[i].current_config = 0;
 
624	}
625
626	lbr_desc->records.nr = 0;
627	lbr_desc->event = NULL;
628	lbr_desc->msr_passthrough = false;
629}
630
631static void intel_pmu_reset(struct kvm_vcpu *vcpu)
632{
633	intel_pmu_release_guest_lbr_event(vcpu);
634}
635
636/*
637 * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
638 *
639 * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
640 * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
641 *
642 * Guest needs to re-enable LBR to resume branches recording.
643 */
644static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
645{
646	u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
647
648	if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
649		data &= ~DEBUGCTLMSR_LBR;
650		vmcs_write64(GUEST_IA32_DEBUGCTL, data);
651	}
652}
653
654static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
655{
656	u8 version = vcpu_to_pmu(vcpu)->version;
657
658	if (!intel_pmu_lbr_is_enabled(vcpu))
659		return;
660
661	if (version > 1 && version < 4)
662		intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
663}
664
665static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
666{
667	struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
668	int i;
669
670	for (i = 0; i < lbr->nr; i++) {
671		vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
672		vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
673		if (lbr->info)
674			vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
675	}
676
677	vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
678	vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
679}
680
681static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
682{
683	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
684
685	if (!lbr_desc->msr_passthrough)
686		return;
687
688	vmx_update_intercept_for_lbr_msrs(vcpu, true);
689	lbr_desc->msr_passthrough = false;
690}
691
692static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
693{
694	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
695
696	if (lbr_desc->msr_passthrough)
697		return;
698
699	vmx_update_intercept_for_lbr_msrs(vcpu, false);
700	lbr_desc->msr_passthrough = true;
701}
702
703/*
704 * Higher priority host perf events (e.g. cpu pinned) could reclaim the
705 * pmu resources (e.g. LBR) that were assigned to the guest. This is
706 * usually done via ipi calls (more details in perf_install_in_context).
707 *
708 * Before entering the non-root mode (with irq disabled here), double
709 * confirm that the pmu features enabled to the guest are not reclaimed
710 * by higher priority host events. Otherwise, disallow vcpu's access to
711 * the reclaimed features.
712 */
713void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
714{
715	struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
716	struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
717
718	if (!lbr_desc->event) {
719		vmx_disable_lbr_msrs_passthrough(vcpu);
720		if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
721			goto warn;
722		if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
723			goto warn;
724		return;
725	}
726
727	if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
728		vmx_disable_lbr_msrs_passthrough(vcpu);
729		__clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
730		goto warn;
731	} else
732		vmx_enable_lbr_msrs_passthrough(vcpu);
733
734	return;
735
736warn:
737	pr_warn_ratelimited("vcpu-%d: fail to passthrough LBR.\n", vcpu->vcpu_id);
738}
739
740static void intel_pmu_cleanup(struct kvm_vcpu *vcpu)
741{
742	if (!(vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR))
743		intel_pmu_release_guest_lbr_event(vcpu);
744}
745
746void intel_pmu_cross_mapped_check(struct kvm_pmu *pmu)
747{
748	struct kvm_pmc *pmc = NULL;
749	int bit, hw_idx;
750
751	for_each_set_bit(bit, (unsigned long *)&pmu->global_ctrl,
752			 X86_PMC_IDX_MAX) {
753		pmc = intel_pmc_idx_to_pmc(pmu, bit);
754
755		if (!pmc || !pmc_speculative_in_use(pmc) ||
756		    !pmc_is_globally_enabled(pmc) || !pmc->perf_event)
757			continue;
758
759		/*
760		 * A negative index indicates the event isn't mapped to a
761		 * physical counter in the host, e.g. due to contention.
762		 */
763		hw_idx = pmc->perf_event->hw.idx;
764		if (hw_idx != pmc->idx && hw_idx > -1)
765			pmu->host_cross_mapped_mask |= BIT_ULL(hw_idx);
766	}
767}
768
769struct kvm_pmu_ops intel_pmu_ops __initdata = {
770	.hw_event_available = intel_hw_event_available,
771	.pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
772	.rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
773	.msr_idx_to_pmc = intel_msr_idx_to_pmc,
774	.is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
775	.is_valid_msr = intel_is_valid_msr,
776	.get_msr = intel_pmu_get_msr,
777	.set_msr = intel_pmu_set_msr,
778	.refresh = intel_pmu_refresh,
779	.init = intel_pmu_init,
780	.reset = intel_pmu_reset,
781	.deliver_pmi = intel_pmu_deliver_pmi,
782	.cleanup = intel_pmu_cleanup,
783	.EVENTSEL_EVENT = ARCH_PERFMON_EVENTSEL_EVENT,
784	.MAX_NR_GP_COUNTERS = KVM_INTEL_PMC_MAX_GENERIC,
785	.MIN_NR_GP_COUNTERS = 1,
786};