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 static struct kvm_event_hw_type_mapping intel_arch_events[] = { 22 /* Index must match CPUID 0x0A.EBX bit vector */ 23 [0] = { 0x3c, 0x00, PERF_COUNT_HW_CPU_CYCLES }, 24 [1] = { 0xc0, 0x00, PERF_COUNT_HW_INSTRUCTIONS }, 25 [2] = { 0x3c, 0x01, PERF_COUNT_HW_BUS_CYCLES }, 26 [3] = { 0x2e, 0x4f, PERF_COUNT_HW_CACHE_REFERENCES }, 27 [4] = { 0x2e, 0x41, PERF_COUNT_HW_CACHE_MISSES }, 28 [5] = { 0xc4, 0x00, PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 29 [6] = { 0xc5, 0x00, PERF_COUNT_HW_BRANCH_MISSES }, 30 [7] = { 0x00, 0x30, PERF_COUNT_HW_REF_CPU_CYCLES }, 31 }; 32 33 /* mapping between fixed pmc index and intel_arch_events array */ 34 static int fixed_pmc_events[] = {1, 0, 7}; 35 36 static void reprogram_fixed_counters(struct kvm_pmu *pmu, u64 data) 37 { 38 int i; 39 40 for (i = 0; i < pmu->nr_arch_fixed_counters; i++) { 41 u8 new_ctrl = fixed_ctrl_field(data, i); 42 u8 old_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl, i); 43 struct kvm_pmc *pmc; 44 45 pmc = get_fixed_pmc(pmu, MSR_CORE_PERF_FIXED_CTR0 + i); 46 47 if (old_ctrl == new_ctrl) 48 continue; 49 50 __set_bit(INTEL_PMC_IDX_FIXED + i, pmu->pmc_in_use); 51 reprogram_fixed_counter(pmc, new_ctrl, i); 52 } 53 54 pmu->fixed_ctr_ctrl = data; 55 } 56 57 /* function is called when global control register has been updated. */ 58 static void global_ctrl_changed(struct kvm_pmu *pmu, u64 data) 59 { 60 int bit; 61 u64 diff = pmu->global_ctrl ^ data; 62 63 pmu->global_ctrl = data; 64 65 for_each_set_bit(bit, (unsigned long *)&diff, X86_PMC_IDX_MAX) 66 reprogram_counter(pmu, bit); 67 } 68 69 static unsigned intel_find_arch_event(struct kvm_pmu *pmu, 70 u8 event_select, 71 u8 unit_mask) 72 { 73 int i; 74 75 for (i = 0; i < ARRAY_SIZE(intel_arch_events); i++) 76 if (intel_arch_events[i].eventsel == event_select 77 && intel_arch_events[i].unit_mask == unit_mask 78 && (pmu->available_event_types & (1 << i))) 79 break; 80 81 if (i == ARRAY_SIZE(intel_arch_events)) 82 return PERF_COUNT_HW_MAX; 83 84 return intel_arch_events[i].event_type; 85 } 86 87 static unsigned intel_find_fixed_event(int idx) 88 { 89 u32 event; 90 size_t size = ARRAY_SIZE(fixed_pmc_events); 91 92 if (idx >= size) 93 return PERF_COUNT_HW_MAX; 94 95 event = fixed_pmc_events[array_index_nospec(idx, size)]; 96 return intel_arch_events[event].event_type; 97 } 98 99 /* check if a PMC is enabled by comparing it with globl_ctrl bits. */ 100 static bool intel_pmc_is_enabled(struct kvm_pmc *pmc) 101 { 102 struct kvm_pmu *pmu = pmc_to_pmu(pmc); 103 104 return test_bit(pmc->idx, (unsigned long *)&pmu->global_ctrl); 105 } 106 107 static struct kvm_pmc *intel_pmc_idx_to_pmc(struct kvm_pmu *pmu, int pmc_idx) 108 { 109 if (pmc_idx < INTEL_PMC_IDX_FIXED) 110 return get_gp_pmc(pmu, MSR_P6_EVNTSEL0 + pmc_idx, 111 MSR_P6_EVNTSEL0); 112 else { 113 u32 idx = pmc_idx - INTEL_PMC_IDX_FIXED; 114 115 return get_fixed_pmc(pmu, idx + MSR_CORE_PERF_FIXED_CTR0); 116 } 117 } 118 119 /* returns 0 if idx's corresponding MSR exists; otherwise returns 1. */ 120 static int intel_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx) 121 { 122 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 123 bool fixed = idx & (1u << 30); 124 125 idx &= ~(3u << 30); 126 127 return (!fixed && idx >= pmu->nr_arch_gp_counters) || 128 (fixed && idx >= pmu->nr_arch_fixed_counters); 129 } 130 131 static struct kvm_pmc *intel_rdpmc_ecx_to_pmc(struct kvm_vcpu *vcpu, 132 unsigned int idx, u64 *mask) 133 { 134 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 135 bool fixed = idx & (1u << 30); 136 struct kvm_pmc *counters; 137 unsigned int num_counters; 138 139 idx &= ~(3u << 30); 140 if (fixed) { 141 counters = pmu->fixed_counters; 142 num_counters = pmu->nr_arch_fixed_counters; 143 } else { 144 counters = pmu->gp_counters; 145 num_counters = pmu->nr_arch_gp_counters; 146 } 147 if (idx >= num_counters) 148 return NULL; 149 *mask &= pmu->counter_bitmask[fixed ? KVM_PMC_FIXED : KVM_PMC_GP]; 150 return &counters[array_index_nospec(idx, num_counters)]; 151 } 152 153 static bool intel_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr) 154 { 155 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 156 int ret; 157 158 switch (msr) { 159 case MSR_CORE_PERF_FIXED_CTR_CTRL: 160 case MSR_CORE_PERF_GLOBAL_STATUS: 161 case MSR_CORE_PERF_GLOBAL_CTRL: 162 case MSR_CORE_PERF_GLOBAL_OVF_CTRL: 163 ret = pmu->version > 1; 164 break; 165 default: 166 ret = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0) || 167 get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0) || 168 get_fixed_pmc(pmu, msr); 169 break; 170 } 171 172 return ret; 173 } 174 175 static struct kvm_pmc *intel_msr_idx_to_pmc(struct kvm_vcpu *vcpu, u32 msr) 176 { 177 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 178 struct kvm_pmc *pmc; 179 180 pmc = get_fixed_pmc(pmu, msr); 181 pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0); 182 pmc = pmc ? pmc : get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0); 183 184 return pmc; 185 } 186 187 static int intel_pmu_get_msr(struct kvm_vcpu *vcpu, u32 msr, u64 *data) 188 { 189 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 190 struct kvm_pmc *pmc; 191 192 switch (msr) { 193 case MSR_CORE_PERF_FIXED_CTR_CTRL: 194 *data = pmu->fixed_ctr_ctrl; 195 return 0; 196 case MSR_CORE_PERF_GLOBAL_STATUS: 197 *data = pmu->global_status; 198 return 0; 199 case MSR_CORE_PERF_GLOBAL_CTRL: 200 *data = pmu->global_ctrl; 201 return 0; 202 case MSR_CORE_PERF_GLOBAL_OVF_CTRL: 203 *data = pmu->global_ovf_ctrl; 204 return 0; 205 default: 206 if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0))) { 207 u64 val = pmc_read_counter(pmc); 208 *data = val & pmu->counter_bitmask[KVM_PMC_GP]; 209 return 0; 210 } else if ((pmc = get_fixed_pmc(pmu, msr))) { 211 u64 val = pmc_read_counter(pmc); 212 *data = val & pmu->counter_bitmask[KVM_PMC_FIXED]; 213 return 0; 214 } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) { 215 *data = pmc->eventsel; 216 return 0; 217 } 218 } 219 220 return 1; 221 } 222 223 static int intel_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info) 224 { 225 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 226 struct kvm_pmc *pmc; 227 u32 msr = msr_info->index; 228 u64 data = msr_info->data; 229 230 switch (msr) { 231 case MSR_CORE_PERF_FIXED_CTR_CTRL: 232 if (pmu->fixed_ctr_ctrl == data) 233 return 0; 234 if (!(data & 0xfffffffffffff444ull)) { 235 reprogram_fixed_counters(pmu, data); 236 return 0; 237 } 238 break; 239 case MSR_CORE_PERF_GLOBAL_STATUS: 240 if (msr_info->host_initiated) { 241 pmu->global_status = data; 242 return 0; 243 } 244 break; /* RO MSR */ 245 case MSR_CORE_PERF_GLOBAL_CTRL: 246 if (pmu->global_ctrl == data) 247 return 0; 248 if (kvm_valid_perf_global_ctrl(pmu, data)) { 249 global_ctrl_changed(pmu, data); 250 return 0; 251 } 252 break; 253 case MSR_CORE_PERF_GLOBAL_OVF_CTRL: 254 if (!(data & pmu->global_ovf_ctrl_mask)) { 255 if (!msr_info->host_initiated) 256 pmu->global_status &= ~data; 257 pmu->global_ovf_ctrl = data; 258 return 0; 259 } 260 break; 261 default: 262 if ((pmc = get_gp_pmc(pmu, msr, MSR_IA32_PERFCTR0))) { 263 if (!msr_info->host_initiated) 264 data = (s64)(s32)data; 265 pmc->counter += data - pmc_read_counter(pmc); 266 if (pmc->perf_event) 267 perf_event_period(pmc->perf_event, 268 get_sample_period(pmc, data)); 269 return 0; 270 } else if ((pmc = get_fixed_pmc(pmu, msr))) { 271 pmc->counter += data - pmc_read_counter(pmc); 272 if (pmc->perf_event) 273 perf_event_period(pmc->perf_event, 274 get_sample_period(pmc, data)); 275 return 0; 276 } else if ((pmc = get_gp_pmc(pmu, msr, MSR_P6_EVNTSEL0))) { 277 if (data == pmc->eventsel) 278 return 0; 279 if (!(data & pmu->reserved_bits)) { 280 reprogram_gp_counter(pmc, data); 281 return 0; 282 } 283 } 284 } 285 286 return 1; 287 } 288 289 static void intel_pmu_refresh(struct kvm_vcpu *vcpu) 290 { 291 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 292 struct x86_pmu_capability x86_pmu; 293 struct kvm_cpuid_entry2 *entry; 294 union cpuid10_eax eax; 295 union cpuid10_edx edx; 296 297 pmu->nr_arch_gp_counters = 0; 298 pmu->nr_arch_fixed_counters = 0; 299 pmu->counter_bitmask[KVM_PMC_GP] = 0; 300 pmu->counter_bitmask[KVM_PMC_FIXED] = 0; 301 pmu->version = 0; 302 pmu->reserved_bits = 0xffffffff00200000ull; 303 304 entry = kvm_find_cpuid_entry(vcpu, 0xa, 0); 305 if (!entry) 306 return; 307 eax.full = entry->eax; 308 edx.full = entry->edx; 309 310 pmu->version = eax.split.version_id; 311 if (!pmu->version) 312 return; 313 314 perf_get_x86_pmu_capability(&x86_pmu); 315 316 pmu->nr_arch_gp_counters = min_t(int, eax.split.num_counters, 317 x86_pmu.num_counters_gp); 318 pmu->counter_bitmask[KVM_PMC_GP] = ((u64)1 << eax.split.bit_width) - 1; 319 pmu->available_event_types = ~entry->ebx & 320 ((1ull << eax.split.mask_length) - 1); 321 322 if (pmu->version == 1) { 323 pmu->nr_arch_fixed_counters = 0; 324 } else { 325 pmu->nr_arch_fixed_counters = 326 min_t(int, edx.split.num_counters_fixed, 327 x86_pmu.num_counters_fixed); 328 pmu->counter_bitmask[KVM_PMC_FIXED] = 329 ((u64)1 << edx.split.bit_width_fixed) - 1; 330 } 331 332 pmu->global_ctrl = ((1ull << pmu->nr_arch_gp_counters) - 1) | 333 (((1ull << pmu->nr_arch_fixed_counters) - 1) << INTEL_PMC_IDX_FIXED); 334 pmu->global_ctrl_mask = ~pmu->global_ctrl; 335 pmu->global_ovf_ctrl_mask = pmu->global_ctrl_mask 336 & ~(MSR_CORE_PERF_GLOBAL_OVF_CTRL_OVF_BUF | 337 MSR_CORE_PERF_GLOBAL_OVF_CTRL_COND_CHGD); 338 if (vmx_pt_mode_is_host_guest()) 339 pmu->global_ovf_ctrl_mask &= 340 ~MSR_CORE_PERF_GLOBAL_OVF_CTRL_TRACE_TOPA_PMI; 341 342 entry = kvm_find_cpuid_entry(vcpu, 7, 0); 343 if (entry && 344 (boot_cpu_has(X86_FEATURE_HLE) || boot_cpu_has(X86_FEATURE_RTM)) && 345 (entry->ebx & (X86_FEATURE_HLE|X86_FEATURE_RTM))) 346 pmu->reserved_bits ^= HSW_IN_TX|HSW_IN_TX_CHECKPOINTED; 347 348 bitmap_set(pmu->all_valid_pmc_idx, 349 0, pmu->nr_arch_gp_counters); 350 bitmap_set(pmu->all_valid_pmc_idx, 351 INTEL_PMC_MAX_GENERIC, pmu->nr_arch_fixed_counters); 352 353 nested_vmx_pmu_entry_exit_ctls_update(vcpu); 354 } 355 356 static void intel_pmu_init(struct kvm_vcpu *vcpu) 357 { 358 int i; 359 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 360 361 for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) { 362 pmu->gp_counters[i].type = KVM_PMC_GP; 363 pmu->gp_counters[i].vcpu = vcpu; 364 pmu->gp_counters[i].idx = i; 365 pmu->gp_counters[i].current_config = 0; 366 } 367 368 for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) { 369 pmu->fixed_counters[i].type = KVM_PMC_FIXED; 370 pmu->fixed_counters[i].vcpu = vcpu; 371 pmu->fixed_counters[i].idx = i + INTEL_PMC_IDX_FIXED; 372 pmu->fixed_counters[i].current_config = 0; 373 } 374 } 375 376 static void intel_pmu_reset(struct kvm_vcpu *vcpu) 377 { 378 struct kvm_pmu *pmu = vcpu_to_pmu(vcpu); 379 struct kvm_pmc *pmc = NULL; 380 int i; 381 382 for (i = 0; i < INTEL_PMC_MAX_GENERIC; i++) { 383 pmc = &pmu->gp_counters[i]; 384 385 pmc_stop_counter(pmc); 386 pmc->counter = pmc->eventsel = 0; 387 } 388 389 for (i = 0; i < INTEL_PMC_MAX_FIXED; i++) { 390 pmc = &pmu->fixed_counters[i]; 391 392 pmc_stop_counter(pmc); 393 pmc->counter = 0; 394 } 395 396 pmu->fixed_ctr_ctrl = pmu->global_ctrl = pmu->global_status = 397 pmu->global_ovf_ctrl = 0; 398 } 399 400 struct kvm_pmu_ops intel_pmu_ops = { 401 .find_arch_event = intel_find_arch_event, 402 .find_fixed_event = intel_find_fixed_event, 403 .pmc_is_enabled = intel_pmc_is_enabled, 404 .pmc_idx_to_pmc = intel_pmc_idx_to_pmc, 405 .rdpmc_ecx_to_pmc = intel_rdpmc_ecx_to_pmc, 406 .msr_idx_to_pmc = intel_msr_idx_to_pmc, 407 .is_valid_rdpmc_ecx = intel_is_valid_rdpmc_ecx, 408 .is_valid_msr = intel_is_valid_msr, 409 .get_msr = intel_pmu_get_msr, 410 .set_msr = intel_pmu_set_msr, 411 .refresh = intel_pmu_refresh, 412 .init = intel_pmu_init, 413 .reset = intel_pmu_reset, 414 }; 415