Loading...
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};
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
23static 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 */
36static int fixed_pmc_events[] = {1, 0, 7};
37
38static 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. */
60static 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
71static unsigned intel_find_arch_event(struct kvm_pmu *pmu,
72 u8 event_select,
73 u8 unit_mask)
74{
75 int i;
76
77 for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++)
78 if (intel_arch_events[i].eventsel == event_select
79 && intel_arch_events[i].unit_mask == unit_mask
80 && (pmu->available_event_types & (1 << i)))
81 break;
82
83 if (i == ARRAY_SIZE(intel_arch_events))
84 return PERF_COUNT_HW_MAX;
85
86 return intel_arch_events[i].event_type;
87}
88
89static unsigned intel_find_fixed_event(int idx)
90{
91 u32 event;
92 size_t size = ARRAY_SIZE(fixed_pmc_events);
93
94 if (idx >= size)
95 return PERF_COUNT_HW_MAX;
96
97 event = fixed_pmc_events[array_index_nospec(idx, size)];
98 return intel_arch_events[event].event_type;
99}
100
101/* check if a PMC is enabled by comparing it with globl_ctrl bits. */
102static bool intel_pmc_is_enabled(struct kvm_pmc *pmc)
103{
104 struct kvm_pmu *pmu = pmc_to_pmu(pmc);
105
106 return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl);
107}
108
109static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx)
110{
111 if (pmc_idx < INTEL_PMC_IDX_FIXED)
112 return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx,
113 MSR_P6_EVNTSEL0);
114 else {
115 u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED;
116
117 return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0);
118 }
119}
120
121/* returns 0 if idx's corresponding MSR exists; otherwise returns 1. */
122static int intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
123{
124 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
125 bool fixed = idx & (1u << 30);
126
127 idx &= ~(3u << 30);
128
129 return (!fixed && idx >= pmu->nr_arch_gp_counters) ||
130 (fixed && idx >= pmu->nr_arch_fixed_counters);
131}
132
133static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu,
134 unsigned int idx, u64 *mask)
135{
136 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
137 bool fixed = idx & (1u << 30);
138 struct kvm_pmc *counters;
139 unsigned int num_counters;
140
141 idx &= ~(3u << 30);
142 if (fixed) {
143 counters = pmu->fixed_counters;
144 num_counters = pmu->nr_arch_fixed_counters;
145 } else {
146 counters = pmu->gp_counters;
147 num_counters = pmu->nr_arch_gp_counters;
148 }
149 if (idx >= num_counters)
150 return NULL;
151 *mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP];
152 return &counters[array_index_nospec(idx, num_counters)];
153}
154
155static inline u64 vcpu_get_perf_capabilities(struct kvm_vcpu *vcpu)
156{
157 if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM))
158 return 0;
159
160 return vcpu->arch.perf_capabilities;
161}
162
163static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu)
164{
165 return (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_FW_WRITES) != 0;
166}
167
168static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr)
169{
170 if (!fw_writes_is_enabled(pmu_to_vcpu(pmu)))
171 return NULL;
172
173 return get_gp_pmc(pmu, msr, MSR_IA32_PMC0);
174}
175
176bool intel_pmu_lbr_is_compatible(struct kvm_vcpu *vcpu)
177{
178 /*
179 * As a first step, a guest could only enable LBR feature if its
180 * cpu model is the same as the host because the LBR registers
181 * would be pass-through to the guest and they're model specific.
182 */
183 return boot_cpu_data.x86_model == guest_cpuid_model(vcpu);
184}
185
186bool intel_pmu_lbr_is_enabled(struct kvm_vcpu *vcpu)
187{
188 struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
189
190 return lbr->nr && (vcpu_get_perf_capabilities(vcpu) & PMU_CAP_LBR_FMT);
191}
192
193static bool intel_pmu_is_valid_lbr_msr(struct kvm_vcpu *vcpu, u32 index)
194{
195 struct x86_pmu_lbr *records = vcpu_to_lbr_records(vcpu);
196 bool ret = false;
197
198 if (!intel_pmu_lbr_is_enabled(vcpu))
199 return ret;
200
201 ret = (index == MSR_LBR_SELECT) || (index == MSR_LBR_TOS) ||
202 (index >= records->from && index < records->from + records->nr) ||
203 (index >= records->to && index < records->to + records->nr);
204
205 if (!ret && records->info)
206 ret = (index >= records->info && index < records->info + records->nr);
207
208 return ret;
209}
210
211static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
212{
213 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
214 int ret;
215
216 switch (msr) {
217 case MSR_CORE_PERF_FIXED_CTR_CTRL:
218 case MSR_CORE_PERF_GLOBAL_STATUS:
219 case MSR_CORE_PERF_GLOBAL_CTRL:
220 case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
221 ret = pmu->version > 1;
222 break;
223 default:
224 ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) ||
225 get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) ||
226 get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr) ||
227 intel_pmu_is_valid_lbr_msr(vcpu, msr);
228 break;
229 }
230
231 return ret;
232}
233
234static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr)
235{
236 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
237 struct kvm_pmc *pmc;
238
239 pmc = get_fixed_pmc(pmu, msr);
240 pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0);
241 pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0);
242
243 return pmc;
244}
245
246static inline void intel_pmu_release_guest_lbr_event(struct kvm_vcpu *vcpu)
247{
248 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
249
250 if (lbr_desc->event) {
251 perf_event_release_kernel(lbr_desc->event);
252 lbr_desc->event = NULL;
253 vcpu_to_pmu(vcpu)->event_count--;
254 }
255}
256
257int intel_pmu_create_guest_lbr_event(struct kvm_vcpu *vcpu)
258{
259 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
260 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
261 struct perf_event *event;
262
263 /*
264 * The perf_event_attr is constructed in the minimum efficient way:
265 * - set 'pinned = true' to make it task pinned so that if another
266 * cpu pinned event reclaims LBR, the event->oncpu will be set to -1;
267 * - set '.exclude_host = true' to record guest branches behavior;
268 *
269 * - set '.config = INTEL_FIXED_VLBR_EVENT' to indicates host perf
270 * schedule the event without a real HW counter but a fake one;
271 * check is_guest_lbr_event() and __intel_get_event_constraints();
272 *
273 * - set 'sample_type = PERF_SAMPLE_BRANCH_STACK' and
274 * 'branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
275 * PERF_SAMPLE_BRANCH_USER' to configure it as a LBR callstack
276 * event, which helps KVM to save/restore guest LBR records
277 * during host context switches and reduces quite a lot overhead,
278 * check branch_user_callstack() and intel_pmu_lbr_sched_task();
279 */
280 struct perf_event_attr attr = {
281 .type = PERF_TYPE_RAW,
282 .size = sizeof(attr),
283 .config = INTEL_FIXED_VLBR_EVENT,
284 .sample_type = PERF_SAMPLE_BRANCH_STACK,
285 .pinned = true,
286 .exclude_host = true,
287 .branch_sample_type = PERF_SAMPLE_BRANCH_CALL_STACK |
288 PERF_SAMPLE_BRANCH_USER,
289 };
290
291 if (unlikely(lbr_desc->event)) {
292 __set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
293 return 0;
294 }
295
296 event = perf_event_create_kernel_counter(&attr, -1,
297 current, NULL, NULL);
298 if (IS_ERR(event)) {
299 pr_debug_ratelimited("%s: failed %ld\n",
300 __func__, PTR_ERR(event));
301 return PTR_ERR(event);
302 }
303 lbr_desc->event = event;
304 pmu->event_count++;
305 __set_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
306 return 0;
307}
308
309/*
310 * It's safe to access LBR msrs from guest when they have not
311 * been passthrough since the host would help restore or reset
312 * the LBR msrs records when the guest LBR event is scheduled in.
313 */
314static bool intel_pmu_handle_lbr_msrs_access(struct kvm_vcpu *vcpu,
315 struct msr_data *msr_info, bool read)
316{
317 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
318 u32 index = msr_info->index;
319
320 if (!intel_pmu_is_valid_lbr_msr(vcpu, index))
321 return false;
322
323 if (!lbr_desc->event && intel_pmu_create_guest_lbr_event(vcpu) < 0)
324 goto dummy;
325
326 /*
327 * Disable irq to ensure the LBR feature doesn't get reclaimed by the
328 * host at the time the value is read from the msr, and this avoids the
329 * host LBR value to be leaked to the guest. If LBR has been reclaimed,
330 * return 0 on guest reads.
331 */
332 local_irq_disable();
333 if (lbr_desc->event->state == PERF_EVENT_STATE_ACTIVE) {
334 if (read)
335 rdmsrl(index, msr_info->data);
336 else
337 wrmsrl(index, msr_info->data);
338 __set_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
339 local_irq_enable();
340 return true;
341 }
342 clear_bit(INTEL_PMC_IDX_FIXED_VLBR, vcpu_to_pmu(vcpu)->pmc_in_use);
343 local_irq_enable();
344
345dummy:
346 if (read)
347 msr_info->data = 0;
348 return true;
349}
350
351static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
352{
353 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
354 struct kvm_pmc *pmc;
355 u32 msr = msr_info->index;
356
357 switch (msr) {
358 case MSR_CORE_PERF_FIXED_CTR_CTRL:
359 msr_info->data = pmu->fixed_ctr_ctrl;
360 return 0;
361 case MSR_CORE_PERF_GLOBAL_STATUS:
362 msr_info->data = pmu->global_status;
363 return 0;
364 case MSR_CORE_PERF_GLOBAL_CTRL:
365 msr_info->data = pmu->global_ctrl;
366 return 0;
367 case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
368 msr_info->data = pmu->global_ovf_ctrl;
369 return 0;
370 default:
371 if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
372 (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
373 u64 val = pmc_read_counter(pmc);
374 msr_info->data =
375 val & pmu->counter_bitmask[KVM_PMC_GP];
376 return 0;
377 } else if ((pmc = get_fixed_pmc(pmu, msr))) {
378 u64 val = pmc_read_counter(pmc);
379 msr_info->data =
380 val & pmu->counter_bitmask[KVM_PMC_FIXED];
381 return 0;
382 } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
383 msr_info->data = pmc->eventsel;
384 return 0;
385 } else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, true))
386 return 0;
387 }
388
389 return 1;
390}
391
392static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
393{
394 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
395 struct kvm_pmc *pmc;
396 u32 msr = msr_info->index;
397 u64 data = msr_info->data;
398
399 switch (msr) {
400 case MSR_CORE_PERF_FIXED_CTR_CTRL:
401 if (pmu->fixed_ctr_ctrl == data)
402 return 0;
403 if (!(data & 0xfffffffffffff444ull)) {
404 reprogram_fixed_counters(pmu, data);
405 return 0;
406 }
407 break;
408 case MSR_CORE_PERF_GLOBAL_STATUS:
409 if (msr_info->host_initiated) {
410 pmu->global_status = data;
411 return 0;
412 }
413 break; /* RO MSR */
414 case MSR_CORE_PERF_GLOBAL_CTRL:
415 if (pmu->global_ctrl == data)
416 return 0;
417 if (kvm_valid_perf_global_ctrl(pmu, data)) {
418 global_ctrl_changed(pmu, data);
419 return 0;
420 }
421 break;
422 case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
423 if (!(data & pmu->global_ovf_ctrl_mask)) {
424 if (!msr_info->host_initiated)
425 pmu->global_status &= ~data;
426 pmu->global_ovf_ctrl = data;
427 return 0;
428 }
429 break;
430 default:
431 if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) ||
432 (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) {
433 if ((msr & MSR_PMC_FULL_WIDTH_BIT) &&
434 (data & ~pmu->counter_bitmask[KVM_PMC_GP]))
435 return 1;
436 if (!msr_info->host_initiated &&
437 !(msr & MSR_PMC_FULL_WIDTH_BIT))
438 data = (s64)(s32)data;
439 pmc->counter += data - pmc_read_counter(pmc);
440 if (pmc->perf_event)
441 perf_event_period(pmc->perf_event,
442 get_sample_period(pmc, data));
443 return 0;
444 } else if ((pmc = get_fixed_pmc(pmu, msr))) {
445 pmc->counter += data - pmc_read_counter(pmc);
446 if (pmc->perf_event)
447 perf_event_period(pmc->perf_event,
448 get_sample_period(pmc, data));
449 return 0;
450 } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) {
451 if (data == pmc->eventsel)
452 return 0;
453 if (!(data & pmu->reserved_bits)) {
454 reprogram_gp_counter(pmc, data);
455 return 0;
456 }
457 } else if (intel_pmu_handle_lbr_msrs_access(vcpu, msr_info, false))
458 return 0;
459 }
460
461 return 1;
462}
463
464static void intel_pmu_refresh(struct kvm_vcpu *vcpu)
465{
466 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
467 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
468
469 struct x86_pmu_capability x86_pmu;
470 struct kvm_cpuid_entry2 *entry;
471 union cpuid10_eax eax;
472 union cpuid10_edx edx;
473
474 pmu->nr_arch_gp_counters = 0;
475 pmu->nr_arch_fixed_counters = 0;
476 pmu->counter_bitmask[KVM_PMC_GP] = 0;
477 pmu->counter_bitmask[KVM_PMC_FIXED] = 0;
478 pmu->version = 0;
479 pmu->reserved_bits = 0xffffffff00200000ull;
480
481 entry = kvm_find_cpuid_entry(vcpu, 0xa, 0);
482 if (!entry)
483 return;
484 eax.full = entry->eax;
485 edx.full = entry->edx;
486
487 pmu->version = eax.split.version_id;
488 if (!pmu->version)
489 return;
490
491 perf_get_x86_pmu_capability(&x86_pmu);
492
493 pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters,
494 x86_pmu.num_counters_gp);
495 eax.split.bit_width = min_t(int, eax.split.bit_width, x86_pmu.bit_width_gp);
496 pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1;
497 eax.split.mask_length = min_t(int, eax.split.mask_length, x86_pmu.events_mask_len);
498 pmu->available_event_types = ~entry->ebx &
499 ((1ull << eax.split.mask_length) - 1);
500
501 if (pmu->version == 1) {
502 pmu->nr_arch_fixed_counters = 0;
503 } else {
504 pmu->nr_arch_fixed_counters =
505 min_t(int, edx.split.num_counters_fixed,
506 x86_pmu.num_counters_fixed);
507 edx.split.bit_width_fixed = min_t(int,
508 edx.split.bit_width_fixed, x86_pmu.bit_width_fixed);
509 pmu->counter_bitmask[KVM_PMC_FIXED] =
510 ((u64)1 << edx.split.bit_width_fixed) - 1;
511 }
512
513 pmu->global_ctrl = ((1ull << pmu->nr_arch_gp_counters) - 1) |
514 (((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED);
515 pmu->global_ctrl_mask = ~pmu->global_ctrl;
516 pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask
517 & ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF |
518 MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD);
519 if (vmx_pt_mode_is_host_guest())
520 pmu->global_ovf_ctrl_mask &=
521 ~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI;
522
523 entry = kvm_find_cpuid_entry(vcpu, 7, 0);
524 if (entry &&
525 (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) &&
526 (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM)))
527 pmu->reserved_bits ^= HSW_IN_TX|HSW_IN_TX_CHECKPOINTED;
528
529 bitmap_set(pmu->all_valid_pmc_idx,
530 0, pmu->nr_arch_gp_counters);
531 bitmap_set(pmu->all_valid_pmc_idx,
532 INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters);
533
534 nested_vmx_pmu_entry_exit_ctls_update(vcpu);
535
536 if (intel_pmu_lbr_is_compatible(vcpu))
537 x86_perf_get_lbr(&lbr_desc->records);
538 else
539 lbr_desc->records.nr = 0;
540
541 if (lbr_desc->records.nr)
542 bitmap_set(pmu->all_valid_pmc_idx, INTEL_PMC_IDX_FIXED_VLBR, 1);
543}
544
545static void intel_pmu_init(struct kvm_vcpu *vcpu)
546{
547 int i;
548 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
549 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
550
551 for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
552 pmu->gp_counters[i].type = KVM_PMC_GP;
553 pmu->gp_counters[i].vcpu = vcpu;
554 pmu->gp_counters[i].idx = i;
555 pmu->gp_counters[i].current_config = 0;
556 }
557
558 for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) {
559 pmu->fixed_counters[i].type = KVM_PMC_FIXED;
560 pmu->fixed_counters[i].vcpu = vcpu;
561 pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED;
562 pmu->fixed_counters[i].current_config = 0;
563 }
564
565 vcpu->arch.perf_capabilities = vmx_get_perf_capabilities();
566 lbr_desc->records.nr = 0;
567 lbr_desc->event = NULL;
568 lbr_desc->msr_passthrough = false;
569}
570
571static void intel_pmu_reset(struct kvm_vcpu *vcpu)
572{
573 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
574 struct kvm_pmc *pmc = NULL;
575 int i;
576
577 for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) {
578 pmc = &pmu->gp_counters[i];
579
580 pmc_stop_counter(pmc);
581 pmc->counter = pmc->eventsel = 0;
582 }
583
584 for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) {
585 pmc = &pmu->fixed_counters[i];
586
587 pmc_stop_counter(pmc);
588 pmc->counter = 0;
589 }
590
591 pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status =
592 pmu->global_ovf_ctrl = 0;
593
594 intel_pmu_release_guest_lbr_event(vcpu);
595}
596
597/*
598 * Emulate LBR_On_PMI behavior for 1 < pmu.version < 4.
599 *
600 * If Freeze_LBR_On_PMI = 1, the LBR is frozen on PMI and
601 * the KVM emulates to clear the LBR bit (bit 0) in IA32_DEBUGCTL.
602 *
603 * Guest needs to re-enable LBR to resume branches recording.
604 */
605static void intel_pmu_legacy_freezing_lbrs_on_pmi(struct kvm_vcpu *vcpu)
606{
607 u64 data = vmcs_read64(GUEST_IA32_DEBUGCTL);
608
609 if (data & DEBUGCTLMSR_FREEZE_LBRS_ON_PMI) {
610 data &= ~DEBUGCTLMSR_LBR;
611 vmcs_write64(GUEST_IA32_DEBUGCTL, data);
612 }
613}
614
615static void intel_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
616{
617 u8 version = vcpu_to_pmu(vcpu)->version;
618
619 if (!intel_pmu_lbr_is_enabled(vcpu))
620 return;
621
622 if (version > 1 && version < 4)
623 intel_pmu_legacy_freezing_lbrs_on_pmi(vcpu);
624}
625
626static void vmx_update_intercept_for_lbr_msrs(struct kvm_vcpu *vcpu, bool set)
627{
628 struct x86_pmu_lbr *lbr = vcpu_to_lbr_records(vcpu);
629 int i;
630
631 for (i = 0; i < lbr->nr; i++) {
632 vmx_set_intercept_for_msr(vcpu, lbr->from + i, MSR_TYPE_RW, set);
633 vmx_set_intercept_for_msr(vcpu, lbr->to + i, MSR_TYPE_RW, set);
634 if (lbr->info)
635 vmx_set_intercept_for_msr(vcpu, lbr->info + i, MSR_TYPE_RW, set);
636 }
637
638 vmx_set_intercept_for_msr(vcpu, MSR_LBR_SELECT, MSR_TYPE_RW, set);
639 vmx_set_intercept_for_msr(vcpu, MSR_LBR_TOS, MSR_TYPE_RW, set);
640}
641
642static inline void vmx_disable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
643{
644 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
645
646 if (!lbr_desc->msr_passthrough)
647 return;
648
649 vmx_update_intercept_for_lbr_msrs(vcpu, true);
650 lbr_desc->msr_passthrough = false;
651}
652
653static inline void vmx_enable_lbr_msrs_passthrough(struct kvm_vcpu *vcpu)
654{
655 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
656
657 if (lbr_desc->msr_passthrough)
658 return;
659
660 vmx_update_intercept_for_lbr_msrs(vcpu, false);
661 lbr_desc->msr_passthrough = true;
662}
663
664/*
665 * Higher priority host perf events (e.g. cpu pinned) could reclaim the
666 * pmu resources (e.g. LBR) that were assigned to the guest. This is
667 * usually done via ipi calls (more details in perf_install_in_context).
668 *
669 * Before entering the non-root mode (with irq disabled here), double
670 * confirm that the pmu features enabled to the guest are not reclaimed
671 * by higher priority host events. Otherwise, disallow vcpu's access to
672 * the reclaimed features.
673 */
674void vmx_passthrough_lbr_msrs(struct kvm_vcpu *vcpu)
675{
676 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
677 struct lbr_desc *lbr_desc = vcpu_to_lbr_desc(vcpu);
678
679 if (!lbr_desc->event) {
680 vmx_disable_lbr_msrs_passthrough(vcpu);
681 if (vmcs_read64(GUEST_IA32_DEBUGCTL) & DEBUGCTLMSR_LBR)
682 goto warn;
683 if (test_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use))
684 goto warn;
685 return;
686 }
687
688 if (lbr_desc->event->state < PERF_EVENT_STATE_ACTIVE) {
689 vmx_disable_lbr_msrs_passthrough(vcpu);
690 __clear_bit(INTEL_PMC_IDX_FIXED_VLBR, pmu->pmc_in_use);
691 goto warn;
692 } else
693 vmx_enable_lbr_msrs_passthrough(vcpu);
694
695 return;
696
697warn:
698 pr_warn_ratelimited("kvm: vcpu-%d: fail to passthrough LBR.\n",
699 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
708struct kvm_pmu_ops intel_pmu_ops = {
709 .find_arch_event = intel_find_arch_event,
710 .find_fixed_event = intel_find_fixed_event,
711 .pmc_is_enabled = intel_pmc_is_enabled,
712 .pmc_idx_to_pmc = intel_pmc_idx_to_pmc,
713 .rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc,
714 .msr_idx_to_pmc = intel_msr_idx_to_pmc,
715 .is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx,
716 .is_valid_msr = intel_is_valid_msr,
717 .get_msr = intel_pmu_get_msr,
718 .set_msr = intel_pmu_set_msr,
719 .refresh = intel_pmu_refresh,
720 .init = intel_pmu_init,
721 .reset = intel_pmu_reset,
722 .deliver_pmi = intel_pmu_deliver_pmi,
723 .cleanup = intel_pmu_cleanup,
724};