1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * KVM PMU support for Intel CPUs 4 * 5 * Copyright 2011 Red Hat, Inc. and/or its affiliates. 6 * 7 * Authors: 8 * Avi Kivity <avi@redhat.com> 9 * Gleb Natapov <gleb@redhat.com> 10 */ 11 #include <linux/types.h> 12 #include <linux/kvm_host.h> 13 #include <linux/perf_event.h> 14 #include <asm/perf_event.h> 15 #include "x86.h" 16 #include "cpuid.h" 17 #include "lapic.h" 18 #include "nested.h" 19 #include "pmu.h" 20 21 #define MSR_PMC_FULL_WIDTH_BIT (MSR_IA32_PMC0 - MSR_IA32_PERFCTR0) 22 23 static struct kvm_event_hw_type_mapping intel_arch_events[] = { 24 /* Index must match CPUID 0x0A.EBX bit vector */ 25 [0] = { 0x3c, 0x00, PERF_COUNT_HW_CPU_CYCLES }, 26 [1] = { 0xc0, 0x00, PERF_COUNT_HW_INSTRUCTIONS }, 27 [2] = { 0x3c, 0x01, PERF_COUNT_HW_BUS_CYCLES }, 28 [3] = { 0x2e, 0x4f, PERF_COUNT_HW_CACHE_REFERENCES }, 29 [4] = { 0x2e, 0x41, PERF_COUNT_HW_CACHE_MISSES }, 30 [5] = { 0xc4, 0x00, PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 31 [6] = { 0xc5, 0x00, PERF_COUNT_HW_BRANCH_MISSES }, 32 [7] = { 0x00, 0x03, PERF_COUNT_HW_REF_CPU_CYCLES }, 33 }; 34 35 /* mapping between fixed pmc index and intel_arch_events array */ 36 static int fixed_pmc_events[] = {1, 0, 7}; 37 38 static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data) 39 { 40 int i; 41 42 for (i = 0; i < pmu->nr_arch_fixed_counters; i++) { 43 u8 new_ctrl = fixed_ctrl_field(data, i); 44 u8 old_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl, i); 45 struct kvm_pmc *pmc; 46 47 pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i); 48 49 if (old_ctrl == new_ctrl) 50 continue; 51 52 __set_bit(INTEL_PMC_IDX_FIXED + i, pmu->pmc_in_use); 53 reprogram_fixed_counter(pmc, new_ctrl, i); 54 } 55 56 pmu->fixed_ctr_ctrl = data; 57 } 58 59 /* function is called when global control register has been updated. */ 60 static void global_ctrl_changed(struct kvm_pmu *pmu, u64 data) 61 { 62 int bit; 63 u64 diff = pmu->global_ctrl ^ data; 64 65 pmu->global_ctrl = data; 66 67 for_each_set_bit(bit, (unsigned long *)&diff, X86_PMC_IDX_MAX) 68 reprogram_counter(pmu, bit); 69 } 70 71 static 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 89 static 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. */ 102 static 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 } 108 109 static 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. */ 122 static 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 } 132 133 static 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 } 154 155 static inline bool fw_writes_is_enabled(struct kvm_vcpu *vcpu) 156 { 157 if (!guest_cpuid_has(vcpu, X86_FEATURE_PDCM)) 158 return false; 159 160 return vcpu->arch.perf_capabilities & PMU_CAP_FW_WRITES; 161 } 162 163 static inline struct kvm_pmc *get_fw_gp_pmc(struct kvm_pmu *pmu, u32 msr) 164 { 165 if (!fw_writes_is_enabled(pmu_to_vcpu(pmu))) 166 return NULL; 167 168 return get_gp_pmc(pmu, msr, MSR_IA32_PMC0); 169 } 170 171 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr) 172 { 173 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 174 int ret; 175 176 switch (msr) { 177 case MSR_CORE_PERF_FIXED_CTR_CTRL: 178 case MSR_CORE_PERF_GLOBAL_STATUS: 179 case MSR_CORE_PERF_GLOBAL_CTRL: 180 case MSR_CORE_PERF_GLOBAL_OVF_CTRL: 181 ret = pmu->version > 1; 182 break; 183 default: 184 ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) || 185 get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) || 186 get_fixed_pmc(pmu, msr) || get_fw_gp_pmc(pmu, msr); 187 break; 188 } 189 190 return ret; 191 } 192 193 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr) 194 { 195 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 196 struct kvm_pmc *pmc; 197 198 pmc = get_fixed_pmc(pmu, msr); 199 pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0); 200 pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0); 201 202 return pmc; 203 } 204 205 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) 206 { 207 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 208 struct kvm_pmc *pmc; 209 u32 msr = msr_info->index; 210 211 switch (msr) { 212 case MSR_CORE_PERF_FIXED_CTR_CTRL: 213 msr_info->data = pmu->fixed_ctr_ctrl; 214 return 0; 215 case MSR_CORE_PERF_GLOBAL_STATUS: 216 msr_info->data = pmu->global_status; 217 return 0; 218 case MSR_CORE_PERF_GLOBAL_CTRL: 219 msr_info->data = pmu->global_ctrl; 220 return 0; 221 case MSR_CORE_PERF_GLOBAL_OVF_CTRL: 222 msr_info->data = pmu->global_ovf_ctrl; 223 return 0; 224 default: 225 if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) || 226 (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) { 227 u64 val = pmc_read_counter(pmc); 228 msr_info->data = 229 val & pmu->counter_bitmask[KVM_PMC_GP]; 230 return 0; 231 } else if ((pmc = get_fixed_pmc(pmu, msr))) { 232 u64 val = pmc_read_counter(pmc); 233 msr_info->data = 234 val & pmu->counter_bitmask[KVM_PMC_FIXED]; 235 return 0; 236 } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) { 237 msr_info->data = pmc->eventsel; 238 return 0; 239 } 240 } 241 242 return 1; 243 } 244 245 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) 246 { 247 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 248 struct kvm_pmc *pmc; 249 u32 msr = msr_info->index; 250 u64 data = msr_info->data; 251 252 switch (msr) { 253 case MSR_CORE_PERF_FIXED_CTR_CTRL: 254 if (pmu->fixed_ctr_ctrl == data) 255 return 0; 256 if (!(data & 0xfffffffffffff444ull)) { 257 reprogram_fixed_counters(pmu, data); 258 return 0; 259 } 260 break; 261 case MSR_CORE_PERF_GLOBAL_STATUS: 262 if (msr_info->host_initiated) { 263 pmu->global_status = data; 264 return 0; 265 } 266 break; /* RO MSR */ 267 case MSR_CORE_PERF_GLOBAL_CTRL: 268 if (pmu->global_ctrl == data) 269 return 0; 270 if (kvm_valid_perf_global_ctrl(pmu, data)) { 271 global_ctrl_changed(pmu, data); 272 return 0; 273 } 274 break; 275 case MSR_CORE_PERF_GLOBAL_OVF_CTRL: 276 if (!(data & pmu->global_ovf_ctrl_mask)) { 277 if (!msr_info->host_initiated) 278 pmu->global_status &= ~data; 279 pmu->global_ovf_ctrl = data; 280 return 0; 281 } 282 break; 283 default: 284 if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0)) || 285 (pmc = get_gp_pmc(pmu, msr, MSR_IA32_PMC0))) { 286 if ((msr & MSR_PMC_FULL_WIDTH_BIT) && 287 (data & ~pmu->counter_bitmask[KVM_PMC_GP])) 288 return 1; 289 if (!msr_info->host_initiated && 290 !(msr & MSR_PMC_FULL_WIDTH_BIT)) 291 data = (s64)(s32)data; 292 pmc->counter += data - pmc_read_counter(pmc); 293 if (pmc->perf_event) 294 perf_event_period(pmc->perf_event, 295 get_sample_period(pmc, data)); 296 return 0; 297 } else if ((pmc = get_fixed_pmc(pmu, msr))) { 298 pmc->counter += data - pmc_read_counter(pmc); 299 if (pmc->perf_event) 300 perf_event_period(pmc->perf_event, 301 get_sample_period(pmc, data)); 302 return 0; 303 } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) { 304 if (data == pmc->eventsel) 305 return 0; 306 if (!(data & pmu->reserved_bits)) { 307 reprogram_gp_counter(pmc, data); 308 return 0; 309 } 310 } 311 } 312 313 return 1; 314 } 315 316 static void intel_pmu_refresh(struct kvm_vcpu *vcpu) 317 { 318 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 319 struct x86_pmu_capability x86_pmu; 320 struct kvm_cpuid_entry2 *entry; 321 union cpuid10_eax eax; 322 union cpuid10_edx edx; 323 324 pmu->nr_arch_gp_counters = 0; 325 pmu->nr_arch_fixed_counters = 0; 326 pmu->counter_bitmask[KVM_PMC_GP] = 0; 327 pmu->counter_bitmask[KVM_PMC_FIXED] = 0; 328 pmu->version = 0; 329 pmu->reserved_bits = 0xffffffff00200000ull; 330 vcpu->arch.perf_capabilities = 0; 331 332 entry = kvm_find_cpuid_entry(vcpu, 0xa, 0); 333 if (!entry) 334 return; 335 eax.full = entry->eax; 336 edx.full = entry->edx; 337 338 pmu->version = eax.split.version_id; 339 if (!pmu->version) 340 return; 341 342 perf_get_x86_pmu_capability(&x86_pmu); 343 if (guest_cpuid_has(vcpu, X86_FEATURE_PDCM)) 344 vcpu->arch.perf_capabilities = vmx_get_perf_capabilities(); 345 346 pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters, 347 x86_pmu.num_counters_gp); 348 eax.split.bit_width = min_t(int, eax.split.bit_width, x86_pmu.bit_width_gp); 349 pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1; 350 eax.split.mask_length = min_t(int, eax.split.mask_length, x86_pmu.events_mask_len); 351 pmu->available_event_types = ~entry->ebx & 352 ((1ull << eax.split.mask_length) - 1); 353 354 if (pmu->version == 1) { 355 pmu->nr_arch_fixed_counters = 0; 356 } else { 357 pmu->nr_arch_fixed_counters = 358 min_t(int, edx.split.num_counters_fixed, 359 x86_pmu.num_counters_fixed); 360 edx.split.bit_width_fixed = min_t(int, 361 edx.split.bit_width_fixed, x86_pmu.bit_width_fixed); 362 pmu->counter_bitmask[KVM_PMC_FIXED] = 363 ((u64)1 << edx.split.bit_width_fixed) - 1; 364 } 365 366 pmu->global_ctrl = ((1ull << pmu->nr_arch_gp_counters) - 1) | 367 (((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED); 368 pmu->global_ctrl_mask = ~pmu->global_ctrl; 369 pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask 370 & ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF | 371 MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD); 372 if (vmx_pt_mode_is_host_guest()) 373 pmu->global_ovf_ctrl_mask &= 374 ~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI; 375 376 entry = kvm_find_cpuid_entry(vcpu, 7, 0); 377 if (entry && 378 (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) && 379 (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) 380 pmu->reserved_bits ^= HSW_IN_TX|HSW_IN_TX_CHECKPOINTED; 381 382 bitmap_set(pmu->all_valid_pmc_idx, 383 0, pmu->nr_arch_gp_counters); 384 bitmap_set(pmu->all_valid_pmc_idx, 385 INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters); 386 387 nested_vmx_pmu_entry_exit_ctls_update(vcpu); 388 } 389 390 static void intel_pmu_init(struct kvm_vcpu *vcpu) 391 { 392 int i; 393 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 394 395 for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) { 396 pmu->gp_counters[i].type = KVM_PMC_GP; 397 pmu->gp_counters[i].vcpu = vcpu; 398 pmu->gp_counters[i].idx = i; 399 pmu->gp_counters[i].current_config = 0; 400 } 401 402 for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) { 403 pmu->fixed_counters[i].type = KVM_PMC_FIXED; 404 pmu->fixed_counters[i].vcpu = vcpu; 405 pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED; 406 pmu->fixed_counters[i].current_config = 0; 407 } 408 } 409 410 static void intel_pmu_reset(struct kvm_vcpu *vcpu) 411 { 412 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 413 struct kvm_pmc *pmc = NULL; 414 int i; 415 416 for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) { 417 pmc = &pmu->gp_counters[i]; 418 419 pmc_stop_counter(pmc); 420 pmc->counter = pmc->eventsel = 0; 421 } 422 423 for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) { 424 pmc = &pmu->fixed_counters[i]; 425 426 pmc_stop_counter(pmc); 427 pmc->counter = 0; 428 } 429 430 pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 431 pmu->global_ovf_ctrl = 0; 432 } 433 434 struct kvm_pmu_ops intel_pmu_ops = { 435 .find_arch_event = intel_find_arch_event, 436 .find_fixed_event = intel_find_fixed_event, 437 .pmc_is_enabled = intel_pmc_is_enabled, 438 .pmc_idx_to_pmc = intel_pmc_idx_to_pmc, 439 .rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc, 440 .msr_idx_to_pmc = intel_msr_idx_to_pmc, 441 .is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx, 442 .is_valid_msr = intel_is_valid_msr, 443 .get_msr = intel_pmu_get_msr, 444 .set_msr = intel_pmu_set_msr, 445 .refresh = intel_pmu_refresh, 446 .init = intel_pmu_init, 447 .reset = intel_pmu_reset, 448 }; 449