arch/x86/kvm/pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Kernel-based Virtual Machine -- Performance Monitoring Unit support
   4  *
   5  * Copyright 2015 Red Hat, Inc. and/or its affiliates.
   6  *
   7  * Authors:
   8  *   Avi Kivity   <avi@redhat.com>
   9  *   Gleb Natapov <gleb@redhat.com>
  10  *   Wei Huang    <wei@redhat.com>
  11  */
  12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  13
  14 #include <linux/types.h>
  15 #include <linux/kvm_host.h>
  16 #include <linux/perf_event.h>
  17 #include <linux/bsearch.h>
  18 #include <linux/sort.h>
  19 #include <asm/perf_event.h>
  20 #include <asm/cpu_device_id.h>
  21 #include "x86.h"
  22 #include "cpuid.h"
  23 #include "lapic.h"
  24 #include "pmu.h"
  25
  26 /* This is enough to filter the vast majority of currently defined events. */
  27 #define KVM_PMU_EVENT_FILTER_MAX_EVENTS 300
  28
  29 struct x86_pmu_capability __read_mostly kvm_pmu_cap;
  30 EXPORT_SYMBOL_GPL(kvm_pmu_cap);
  31
  32 /* Precise Distribution of Instructions Retired (PDIR) */
  33 static const struct x86_cpu_id vmx_pebs_pdir_cpu[] = {
  34         X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_D, NULL),
  35         X86_MATCH_INTEL_FAM6_MODEL(ICELAKE_X, NULL),
  36         /* Instruction-Accurate PDIR (PDIR++) */
  37         X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, NULL),
  38         {}
  39 };
  40
  41 /* Precise Distribution (PDist) */
  42 static const struct x86_cpu_id vmx_pebs_pdist_cpu[] = {
  43         X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, NULL),
  44         {}
  45 };
  46
  47 /* NOTE:
  48  * - Each perf counter is defined as "struct kvm_pmc";
  49  * - There are two types of perf counters: general purpose (gp) and fixed.
  50  *   gp counters are stored in gp_counters[] and fixed counters are stored
  51  *   in fixed_counters[] respectively. Both of them are part of "struct
  52  *   kvm_pmu";
  53  * - pmu.c understands the difference between gp counters and fixed counters.
  54  *   However AMD doesn't support fixed-counters;
  55  * - There are three types of index to access perf counters (PMC):
  56  *     1. MSR (named msr): For example Intel has MSR_IA32_PERFCTRn and AMD
  57  *        has MSR_K7_PERFCTRn and, for families 15H and later,
  58  *        MSR_F15H_PERF_CTRn, where MSR_F15H_PERF_CTR[0-3] are
  59  *        aliased to MSR_K7_PERFCTRn.
  60  *     2. MSR Index (named idx): This normally is used by RDPMC instruction.
  61  *        For instance AMD RDPMC instruction uses 0000_0003h in ECX to access
  62  *        C001_0007h (MSR_K7_PERCTR3). Intel has a similar mechanism, except
  63  *        that it also supports fixed counters. idx can be used to as index to
  64  *        gp and fixed counters.
  65  *     3. Global PMC Index (named pmc): pmc is an index specific to PMU
  66  *        code. Each pmc, stored in kvm_pmc.idx field, is unique across
  67  *        all perf counters (both gp and fixed). The mapping relationship
  68  *        between pmc and perf counters is as the following:
  69  *        * Intel: [0 .. KVM_INTEL_PMC_MAX_GENERIC-1] <=> gp counters
  70  *                 [INTEL_PMC_IDX_FIXED .. INTEL_PMC_IDX_FIXED + 2] <=> fixed
  71  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] and, for families 15H
  72  *          and later, [0 .. AMD64_NUM_COUNTERS_CORE-1] <=> gp counters
  73  */
  74
  75 static struct kvm_pmu_ops kvm_pmu_ops __read_mostly;
  76
  77 #define KVM_X86_PMU_OP(func)                                         \
  78         DEFINE_STATIC_CALL_NULL(kvm_x86_pmu_##func,                          \
  79                                 *(((struct kvm_pmu_ops *)0)->func));
  80 #define KVM_X86_PMU_OP_OPTIONAL KVM_X86_PMU_OP
  81 #include <asm/kvm-x86-pmu-ops.h>
  82
  83 void kvm_pmu_ops_update(const struct kvm_pmu_ops *pmu_ops)
  84 {
  85         memcpy(&kvm_pmu_ops, pmu_ops, sizeof(kvm_pmu_ops));
  86
  87 #define __KVM_X86_PMU_OP(func) \
  88         static_call_update(kvm_x86_pmu_##func, kvm_pmu_ops.func);
  89 #define KVM_X86_PMU_OP(func) \
  90         WARN_ON(!kvm_pmu_ops.func); __KVM_X86_PMU_OP(func)
  91 #define KVM_X86_PMU_OP_OPTIONAL __KVM_X86_PMU_OP
  92 #include <asm/kvm-x86-pmu-ops.h>
  93 #undef __KVM_X86_PMU_OP
  94 }
  95
  96 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
  97 {
  98         struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
  99         struct kvm_vcpu *vcpu = pmu_to_vcpu(pmu);
 100
 101         kvm_pmu_deliver_pmi(vcpu);
 102 }
 103
 104 static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
 105 {
 106         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
 107         bool skip_pmi = false;
 108
 109         if (pmc->perf_event && pmc->perf_event->attr.precise_ip) {
 110                 if (!in_pmi) {
 111                         /*
 112                          * TODO: KVM is currently _choosing_ to not generate records
 113                          * for emulated instructions, avoiding BUFFER_OVF PMI when
 114                          * there are no records. Strictly speaking, it should be done
 115                          * as well in the right context to improve sampling accuracy.
 116                          */
 117                         skip_pmi = true;
 118                 } else {
 119                         /* Indicate PEBS overflow PMI to guest. */
 120                         skip_pmi = __test_and_set_bit(GLOBAL_STATUS_BUFFER_OVF_BIT,
 121                                                       (unsigned long *)&pmu->global_status);
 122                 }
 123         } else {
 124                 __set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
 125         }
 126
 127         if (!pmc->intr || skip_pmi)
 128                 return;
 129
 130         /*
 131          * Inject PMI. If vcpu was in a guest mode during NMI PMI
 132          * can be ejected on a guest mode re-entry. Otherwise we can't
 133          * be sure that vcpu wasn't executing hlt instruction at the
 134          * time of vmexit and is not going to re-enter guest mode until
 135          * woken up. So we should wake it, but this is impossible from
 136          * NMI context. Do it from irq work instead.
 137          */
 138         if (in_pmi && !kvm_handling_nmi_from_guest(pmc->vcpu))
 139                 irq_work_queue(&pmc_to_pmu(pmc)->irq_work);
 140         else
 141                 kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
 142 }
 143
 144 static void kvm_perf_overflow(struct perf_event *perf_event,
 145                               struct perf_sample_data *data,
 146                               struct pt_regs *regs)
 147 {
 148         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
 149
 150         /*
 151          * Ignore overflow events for counters that are scheduled to be
 152          * reprogrammed, e.g. if a PMI for the previous event races with KVM's
 153          * handling of a related guest WRMSR.
 154          */
 155         if (test_and_set_bit(pmc->idx, pmc_to_pmu(pmc)->reprogram_pmi))
 156                 return;
 157
 158         __kvm_perf_overflow(pmc, true);
 159
 160         kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
 161 }
 162
 163 static u64 pmc_get_pebs_precise_level(struct kvm_pmc *pmc)
 164 {
 165         /*
 166          * For some model specific pebs counters with special capabilities
 167          * (PDIR, PDIR++, PDIST), KVM needs to raise the event precise
 168          * level to the maximum value (currently 3, backwards compatible)
 169          * so that the perf subsystem would assign specific hardware counter
 170          * with that capability for vPMC.
 171          */
 172         if ((pmc->idx == 0 && x86_match_cpu(vmx_pebs_pdist_cpu)) ||
 173             (pmc->idx == 32 && x86_match_cpu(vmx_pebs_pdir_cpu)))
 174                 return 3;
 175
 176         /*
 177          * The non-zero precision level of guest event makes the ordinary
 178          * guest event becomes a guest PEBS event and triggers the host
 179          * PEBS PMI handler to determine whether the PEBS overflow PMI
 180          * comes from the host counters or the guest.
 181          */
 182         return 1;
 183 }
 184
 185 static int pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type, u64 config,
 186                                  bool exclude_user, bool exclude_kernel,
 187                                  bool intr)
 188 {
 189         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
 190         struct perf_event *event;
 191         struct perf_event_attr attr = {
 192                 .type = type,
 193                 .size = sizeof(attr),
 194                 .pinned = true,
 195                 .exclude_idle = true,
 196                 .exclude_host = 1,
 197                 .exclude_user = exclude_user,
 198                 .exclude_kernel = exclude_kernel,
 199                 .config = config,
 200         };
 201         bool pebs = test_bit(pmc->idx, (unsigned long *)&pmu->pebs_enable);
 202
 203         attr.sample_period = get_sample_period(pmc, pmc->counter);
 204
 205         if ((attr.config & HSW_IN_TX_CHECKPOINTED) &&
 206             guest_cpuid_is_intel(pmc->vcpu)) {
 207                 /*
 208                  * HSW_IN_TX_CHECKPOINTED is not supported with nonzero
 209                  * period. Just clear the sample period so at least
 210                  * allocating the counter doesn't fail.
 211                  */
 212                 attr.sample_period = 0;
 213         }
 214         if (pebs) {
 215                 /*
 216                  * For most PEBS hardware events, the difference in the software
 217                  * precision levels of guest and host PEBS events will not affect
 218                  * the accuracy of the PEBS profiling result, because the "event IP"
 219                  * in the PEBS record is calibrated on the guest side.
 220                  */
 221                 attr.precise_ip = pmc_get_pebs_precise_level(pmc);
 222         }
 223
 224         event = perf_event_create_kernel_counter(&attr, -1, current,
 225                                                  kvm_perf_overflow, pmc);
 226         if (IS_ERR(event)) {
 227                 pr_debug_ratelimited("kvm_pmu: event creation failed %ld for pmc->idx = %d\n",
 228                             PTR_ERR(event), pmc->idx);
 229                 return PTR_ERR(event);
 230         }
 231
 232         pmc->perf_event = event;
 233         pmc_to_pmu(pmc)->event_count++;
 234         pmc->is_paused = false;
 235         pmc->intr = intr || pebs;
 236         return 0;
 237 }
 238
 239 static void pmc_pause_counter(struct kvm_pmc *pmc)
 240 {
 241         u64 counter = pmc->counter;
 242
 243         if (!pmc->perf_event || pmc->is_paused)
 244                 return;
 245
 246         /* update counter, reset event value to avoid redundant accumulation */
 247         counter += perf_event_pause(pmc->perf_event, true);
 248         pmc->counter = counter & pmc_bitmask(pmc);
 249         pmc->is_paused = true;
 250 }
 251
 252 static bool pmc_resume_counter(struct kvm_pmc *pmc)
 253 {
 254         if (!pmc->perf_event)
 255                 return false;
 256
 257         /* recalibrate sample period and check if it's accepted by perf core */
 258         if (is_sampling_event(pmc->perf_event) &&
 259             perf_event_period(pmc->perf_event,
 260                               get_sample_period(pmc, pmc->counter)))
 261                 return false;
 262
 263         if (test_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->pebs_enable) !=
 264             (!!pmc->perf_event->attr.precise_ip))
 265                 return false;
 266
 267         /* reuse perf_event to serve as pmc_reprogram_counter() does*/
 268         perf_event_enable(pmc->perf_event);
 269         pmc->is_paused = false;
 270
 271         return true;
 272 }
 273
 274 static int filter_cmp(const void *pa, const void *pb, u64 mask)
 275 {
 276         u64 a = *(u64 *)pa & mask;
 277         u64 b = *(u64 *)pb & mask;
 278
 279         return (a > b) - (a < b);
 280 }
 281
 282
 283 static int filter_sort_cmp(const void *pa, const void *pb)
 284 {
 285         return filter_cmp(pa, pb, (KVM_PMU_MASKED_ENTRY_EVENT_SELECT |
 286                                    KVM_PMU_MASKED_ENTRY_EXCLUDE));
 287 }
 288
 289 /*
 290  * For the event filter, searching is done on the 'includes' list and
 291  * 'excludes' list separately rather than on the 'events' list (which
 292  * has both).  As a result the exclude bit can be ignored.
 293  */
 294 static int filter_event_cmp(const void *pa, const void *pb)
 295 {
 296         return filter_cmp(pa, pb, (KVM_PMU_MASKED_ENTRY_EVENT_SELECT));
 297 }
 298
 299 static int find_filter_index(u64 *events, u64 nevents, u64 key)
 300 {
 301         u64 *fe = bsearch(&key, events, nevents, sizeof(events[0]),
 302                           filter_event_cmp);
 303
 304         if (!fe)
 305                 return -1;
 306
 307         return fe - events;
 308 }
 309
 310 static bool is_filter_entry_match(u64 filter_event, u64 umask)
 311 {
 312         u64 mask = filter_event >> (KVM_PMU_MASKED_ENTRY_UMASK_MASK_SHIFT - 8);
 313         u64 match = filter_event & KVM_PMU_MASKED_ENTRY_UMASK_MATCH;
 314
 315         BUILD_BUG_ON((KVM_PMU_ENCODE_MASKED_ENTRY(0, 0xff, 0, false) >>
 316                      (KVM_PMU_MASKED_ENTRY_UMASK_MASK_SHIFT - 8)) !=
 317                      ARCH_PERFMON_EVENTSEL_UMASK);
 318
 319         return (umask & mask) == match;
 320 }
 321
 322 static bool filter_contains_match(u64 *events, u64 nevents, u64 eventsel)
 323 {
 324         u64 event_select = eventsel & kvm_pmu_ops.EVENTSEL_EVENT;
 325         u64 umask = eventsel & ARCH_PERFMON_EVENTSEL_UMASK;
 326         int i, index;
 327
 328         index = find_filter_index(events, nevents, event_select);
 329         if (index < 0)
 330                 return false;
 331
 332         /*
 333          * Entries are sorted by the event select.  Walk the list in both
 334          * directions to process all entries with the targeted event select.
 335          */
 336         for (i = index; i < nevents; i++) {
 337                 if (filter_event_cmp(&events[i], &event_select))
 338                         break;
 339
 340                 if (is_filter_entry_match(events[i], umask))
 341                         return true;
 342         }
 343
 344         for (i = index - 1; i >= 0; i--) {
 345                 if (filter_event_cmp(&events[i], &event_select))
 346                         break;
 347
 348                 if (is_filter_entry_match(events[i], umask))
 349                         return true;
 350         }
 351
 352         return false;
 353 }
 354
 355 static bool is_gp_event_allowed(struct kvm_x86_pmu_event_filter *f,
 356                                 u64 eventsel)
 357 {
 358         if (filter_contains_match(f->includes, f->nr_includes, eventsel) &&
 359             !filter_contains_match(f->excludes, f->nr_excludes, eventsel))
 360                 return f->action == KVM_PMU_EVENT_ALLOW;
 361
 362         return f->action == KVM_PMU_EVENT_DENY;
 363 }
 364
 365 static bool is_fixed_event_allowed(struct kvm_x86_pmu_event_filter *filter,
 366                                    int idx)
 367 {
 368         int fixed_idx = idx - INTEL_PMC_IDX_FIXED;
 369
 370         if (filter->action == KVM_PMU_EVENT_DENY &&
 371             test_bit(fixed_idx, (ulong *)&filter->fixed_counter_bitmap))
 372                 return false;
 373         if (filter->action == KVM_PMU_EVENT_ALLOW &&
 374             !test_bit(fixed_idx, (ulong *)&filter->fixed_counter_bitmap))
 375                 return false;
 376
 377         return true;
 378 }
 379
 380 static bool check_pmu_event_filter(struct kvm_pmc *pmc)
 381 {
 382         struct kvm_x86_pmu_event_filter *filter;
 383         struct kvm *kvm = pmc->vcpu->kvm;
 384
 385         if (!static_call(kvm_x86_pmu_hw_event_available)(pmc))
 386                 return false;
 387
 388         filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
 389         if (!filter)
 390                 return true;
 391
 392         if (pmc_is_gp(pmc))
 393                 return is_gp_event_allowed(filter, pmc->eventsel);
 394
 395         return is_fixed_event_allowed(filter, pmc->idx);
 396 }
 397
 398 static bool pmc_event_is_allowed(struct kvm_pmc *pmc)
 399 {
 400         return pmc_is_globally_enabled(pmc) && pmc_speculative_in_use(pmc) &&
 401                check_pmu_event_filter(pmc);
 402 }
 403
 404 static void reprogram_counter(struct kvm_pmc *pmc)
 405 {
 406         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
 407         u64 eventsel = pmc->eventsel;
 408         u64 new_config = eventsel;
 409         u8 fixed_ctr_ctrl;
 410
 411         pmc_pause_counter(pmc);
 412
 413         if (!pmc_event_is_allowed(pmc))
 414                 goto reprogram_complete;
 415
 416         if (pmc->counter < pmc->prev_counter)
 417                 __kvm_perf_overflow(pmc, false);
 418
 419         if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
 420                 printk_once("kvm pmu: pin control bit is ignored\n");
 421
 422         if (pmc_is_fixed(pmc)) {
 423                 fixed_ctr_ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl,
 424                                                   pmc->idx - INTEL_PMC_IDX_FIXED);
 425                 if (fixed_ctr_ctrl & 0x1)
 426                         eventsel |= ARCH_PERFMON_EVENTSEL_OS;
 427                 if (fixed_ctr_ctrl & 0x2)
 428                         eventsel |= ARCH_PERFMON_EVENTSEL_USR;
 429                 if (fixed_ctr_ctrl & 0x8)
 430                         eventsel |= ARCH_PERFMON_EVENTSEL_INT;
 431                 new_config = (u64)fixed_ctr_ctrl;
 432         }
 433
 434         if (pmc->current_config == new_config && pmc_resume_counter(pmc))
 435                 goto reprogram_complete;
 436
 437         pmc_release_perf_event(pmc);
 438
 439         pmc->current_config = new_config;
 440
 441         /*
 442          * If reprogramming fails, e.g. due to contention, leave the counter's
 443          * regprogram bit set, i.e. opportunistically try again on the next PMU
 444          * refresh.  Don't make a new request as doing so can stall the guest
 445          * if reprogramming repeatedly fails.
 446          */
 447         if (pmc_reprogram_counter(pmc, PERF_TYPE_RAW,
 448                                   (eventsel & pmu->raw_event_mask),
 449                                   !(eventsel & ARCH_PERFMON_EVENTSEL_USR),
 450                                   !(eventsel & ARCH_PERFMON_EVENTSEL_OS),
 451                                   eventsel & ARCH_PERFMON_EVENTSEL_INT))
 452                 return;
 453
 454 reprogram_complete:
 455         clear_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->reprogram_pmi);
 456         pmc->prev_counter = 0;
 457 }
 458
 459 void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 460 {
 461         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 462         int bit;
 463
 464         for_each_set_bit(bit, pmu->reprogram_pmi, X86_PMC_IDX_MAX) {
 465                 struct kvm_pmc *pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, bit);
 466
 467                 if (unlikely(!pmc)) {
 468                         clear_bit(bit, pmu->reprogram_pmi);
 469                         continue;
 470                 }
 471
 472                 reprogram_counter(pmc);
 473         }
 474
 475         /*
 476          * Unused perf_events are only released if the corresponding MSRs
 477          * weren't accessed during the last vCPU time slice. kvm_arch_sched_in
 478          * triggers KVM_REQ_PMU if cleanup is needed.
 479          */
 480         if (unlikely(pmu->need_cleanup))
 481                 kvm_pmu_cleanup(vcpu);
 482 }
 483
 484 /* check if idx is a valid index to access PMU */
 485 bool kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
 486 {
 487         return static_call(kvm_x86_pmu_is_valid_rdpmc_ecx)(vcpu, idx);
 488 }
 489
 490 bool is_vmware_backdoor_pmc(u32 pmc_idx)
 491 {
 492         switch (pmc_idx) {
 493         case VMWARE_BACKDOOR_PMC_HOST_TSC:
 494         case VMWARE_BACKDOOR_PMC_REAL_TIME:
 495         case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
 496                 return true;
 497         }
 498         return false;
 499 }
 500
 501 static int kvm_pmu_rdpmc_vmware(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 502 {
 503         u64 ctr_val;
 504
 505         switch (idx) {
 506         case VMWARE_BACKDOOR_PMC_HOST_TSC:
 507                 ctr_val = rdtsc();
 508                 break;
 509         case VMWARE_BACKDOOR_PMC_REAL_TIME:
 510                 ctr_val = ktime_get_boottime_ns();
 511                 break;
 512         case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
 513                 ctr_val = ktime_get_boottime_ns() +
 514                         vcpu->kvm->arch.kvmclock_offset;
 515                 break;
 516         default:
 517                 return 1;
 518         }
 519
 520         *data = ctr_val;
 521         return 0;
 522 }
 523
 524 int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 525 {
 526         bool fast_mode = idx & (1u << 31);
 527         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 528         struct kvm_pmc *pmc;
 529         u64 mask = fast_mode ? ~0u : ~0ull;
 530
 531         if (!pmu->version)
 532                 return 1;
 533
 534         if (is_vmware_backdoor_pmc(idx))
 535                 return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
 536
 537         pmc = static_call(kvm_x86_pmu_rdpmc_ecx_to_pmc)(vcpu, idx, &mask);
 538         if (!pmc)
 539                 return 1;
 540
 541         if (!kvm_is_cr4_bit_set(vcpu, X86_CR4_PCE) &&
 542             (static_call(kvm_x86_get_cpl)(vcpu) != 0) &&
 543             kvm_is_cr0_bit_set(vcpu, X86_CR0_PE))
 544                 return 1;
 545
 546         *data = pmc_read_counter(pmc) & mask;
 547         return 0;
 548 }
 549
 550 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
 551 {
 552         if (lapic_in_kernel(vcpu)) {
 553                 static_call_cond(kvm_x86_pmu_deliver_pmi)(vcpu);
 554                 kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
 555         }
 556 }
 557
 558 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
 559 {
 560         switch (msr) {
 561         case MSR_CORE_PERF_GLOBAL_STATUS:
 562         case MSR_CORE_PERF_GLOBAL_CTRL:
 563         case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
 564                 return kvm_pmu_has_perf_global_ctrl(vcpu_to_pmu(vcpu));
 565         default:
 566                 break;
 567         }
 568         return static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr) ||
 569                 static_call(kvm_x86_pmu_is_valid_msr)(vcpu, msr);
 570 }
 571
 572 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 573 {
 574         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 575         struct kvm_pmc *pmc = static_call(kvm_x86_pmu_msr_idx_to_pmc)(vcpu, msr);
 576
 577         if (pmc)
 578                 __set_bit(pmc->idx, pmu->pmc_in_use);
 579 }
 580
 581 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 582 {
 583         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 584         u32 msr = msr_info->index;
 585
 586         switch (msr) {
 587         case MSR_CORE_PERF_GLOBAL_STATUS:
 588         case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS:
 589                 msr_info->data = pmu->global_status;
 590                 break;
 591         case MSR_AMD64_PERF_CNTR_GLOBAL_CTL:
 592         case MSR_CORE_PERF_GLOBAL_CTRL:
 593                 msr_info->data = pmu->global_ctrl;
 594                 break;
 595         case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR:
 596         case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
 597                 msr_info->data = 0;
 598                 break;
 599         default:
 600                 return static_call(kvm_x86_pmu_get_msr)(vcpu, msr_info);
 601         }
 602
 603         return 0;
 604 }
 605
 606 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 607 {
 608         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 609         u32 msr = msr_info->index;
 610         u64 data = msr_info->data;
 611         u64 diff;
 612
 613         /*
 614          * Note, AMD ignores writes to reserved bits and read-only PMU MSRs,
 615          * whereas Intel generates #GP on attempts to write reserved/RO MSRs.
 616          */
 617         switch (msr) {
 618         case MSR_CORE_PERF_GLOBAL_STATUS:
 619                 if (!msr_info->host_initiated)
 620                         return 1; /* RO MSR */
 621                 fallthrough;
 622         case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS:
 623                 /* Per PPR, Read-only MSR. Writes are ignored. */
 624                 if (!msr_info->host_initiated)
 625                         break;
 626
 627                 if (data & pmu->global_status_mask)
 628                         return 1;
 629
 630                 pmu->global_status = data;
 631                 break;
 632         case MSR_AMD64_PERF_CNTR_GLOBAL_CTL:
 633                 data &= ~pmu->global_ctrl_mask;
 634                 fallthrough;
 635         case MSR_CORE_PERF_GLOBAL_CTRL:
 636                 if (!kvm_valid_perf_global_ctrl(pmu, data))
 637                         return 1;
 638
 639                 if (pmu->global_ctrl != data) {
 640                         diff = pmu->global_ctrl ^ data;
 641                         pmu->global_ctrl = data;
 642                         reprogram_counters(pmu, diff);
 643                 }
 644                 break;
 645         case MSR_CORE_PERF_GLOBAL_OVF_CTRL:
 646                 /*
 647                  * GLOBAL_OVF_CTRL, a.k.a. GLOBAL STATUS_RESET, clears bits in
 648                  * GLOBAL_STATUS, and so the set of reserved bits is the same.
 649                  */
 650                 if (data & pmu->global_status_mask)
 651                         return 1;
 652                 fallthrough;
 653         case MSR_AMD64_PERF_CNTR_GLOBAL_STATUS_CLR:
 654                 if (!msr_info->host_initiated)
 655                         pmu->global_status &= ~data;
 656                 break;
 657         default:
 658                 kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
 659                 return static_call(kvm_x86_pmu_set_msr)(vcpu, msr_info);
 660         }
 661
 662         return 0;
 663 }
 664
 665 /* refresh PMU settings. This function generally is called when underlying
 666  * settings are changed (such as changes of PMU CPUID by guest VMs), which
 667  * should rarely happen.
 668  */
 669 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 670 {
 671         if (KVM_BUG_ON(kvm_vcpu_has_run(vcpu), vcpu->kvm))
 672                 return;
 673
 674         bitmap_zero(vcpu_to_pmu(vcpu)->all_valid_pmc_idx, X86_PMC_IDX_MAX);
 675         static_call(kvm_x86_pmu_refresh)(vcpu);
 676 }
 677
 678 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 679 {
 680         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 681
 682         irq_work_sync(&pmu->irq_work);
 683         static_call(kvm_x86_pmu_reset)(vcpu);
 684 }
 685
 686 void kvm_pmu_init(struct kvm_vcpu *vcpu)
 687 {
 688         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 689
 690         memset(pmu, 0, sizeof(*pmu));
 691         static_call(kvm_x86_pmu_init)(vcpu);
 692         init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
 693         pmu->event_count = 0;
 694         pmu->need_cleanup = false;
 695         kvm_pmu_refresh(vcpu);
 696 }
 697
 698 /* Release perf_events for vPMCs that have been unused for a full time slice.  */
 699 void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
 700 {
 701         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 702         struct kvm_pmc *pmc = NULL;
 703         DECLARE_BITMAP(bitmask, X86_PMC_IDX_MAX);
 704         int i;
 705
 706         pmu->need_cleanup = false;
 707
 708         bitmap_andnot(bitmask, pmu->all_valid_pmc_idx,
 709                       pmu->pmc_in_use, X86_PMC_IDX_MAX);
 710
 711         for_each_set_bit(i, bitmask, X86_PMC_IDX_MAX) {
 712                 pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
 713
 714                 if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
 715                         pmc_stop_counter(pmc);
 716         }
 717
 718         static_call_cond(kvm_x86_pmu_cleanup)(vcpu);
 719
 720         bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
 721 }
 722
 723 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
 724 {
 725         kvm_pmu_reset(vcpu);
 726 }
 727
 728 static void kvm_pmu_incr_counter(struct kvm_pmc *pmc)
 729 {
 730         pmc->prev_counter = pmc->counter;
 731         pmc->counter = (pmc->counter + 1) & pmc_bitmask(pmc);
 732         kvm_pmu_request_counter_reprogram(pmc);
 733 }
 734
 735 static inline bool eventsel_match_perf_hw_id(struct kvm_pmc *pmc,
 736         unsigned int perf_hw_id)
 737 {
 738         return !((pmc->eventsel ^ perf_get_hw_event_config(perf_hw_id)) &
 739                 AMD64_RAW_EVENT_MASK_NB);
 740 }
 741
 742 static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 743 {
 744         bool select_os, select_user;
 745         u64 config;
 746
 747         if (pmc_is_gp(pmc)) {
 748                 config = pmc->eventsel;
 749                 select_os = config & ARCH_PERFMON_EVENTSEL_OS;
 750                 select_user = config & ARCH_PERFMON_EVENTSEL_USR;
 751         } else {
 752                 config = fixed_ctrl_field(pmc_to_pmu(pmc)->fixed_ctr_ctrl,
 753                                           pmc->idx - INTEL_PMC_IDX_FIXED);
 754                 select_os = config & 0x1;
 755                 select_user = config & 0x2;
 756         }
 757
 758         return (static_call(kvm_x86_get_cpl)(pmc->vcpu) == 0) ? select_os : select_user;
 759 }
 760
 761 void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
 762 {
 763         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 764         struct kvm_pmc *pmc;
 765         int i;
 766
 767         for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
 768                 pmc = static_call(kvm_x86_pmu_pmc_idx_to_pmc)(pmu, i);
 769
 770                 if (!pmc || !pmc_event_is_allowed(pmc))
 771                         continue;
 772
 773                 /* Ignore checks for edge detect, pin control, invert and CMASK bits */
 774                 if (eventsel_match_perf_hw_id(pmc, perf_hw_id) && cpl_is_matched(pmc))
 775                         kvm_pmu_incr_counter(pmc);
 776         }
 777 }
 778 EXPORT_SYMBOL_GPL(kvm_pmu_trigger_event);
 779
 780 static bool is_masked_filter_valid(const struct kvm_x86_pmu_event_filter *filter)
 781 {
 782         u64 mask = kvm_pmu_ops.EVENTSEL_EVENT |
 783                    KVM_PMU_MASKED_ENTRY_UMASK_MASK |
 784                    KVM_PMU_MASKED_ENTRY_UMASK_MATCH |
 785                    KVM_PMU_MASKED_ENTRY_EXCLUDE;
 786         int i;
 787
 788         for (i = 0; i < filter->nevents; i++) {
 789                 if (filter->events[i] & ~mask)
 790                         return false;
 791         }
 792
 793         return true;
 794 }
 795
 796 static void convert_to_masked_filter(struct kvm_x86_pmu_event_filter *filter)
 797 {
 798         int i, j;
 799
 800         for (i = 0, j = 0; i < filter->nevents; i++) {
 801                 /*
 802                  * Skip events that are impossible to match against a guest
 803                  * event.  When filtering, only the event select + unit mask
 804                  * of the guest event is used.  To maintain backwards
 805                  * compatibility, impossible filters can't be rejected :-(
 806                  */
 807                 if (filter->events[i] & ~(kvm_pmu_ops.EVENTSEL_EVENT |
 808                                           ARCH_PERFMON_EVENTSEL_UMASK))
 809                         continue;
 810                 /*
 811                  * Convert userspace events to a common in-kernel event so
 812                  * only one code path is needed to support both events.  For
 813                  * the in-kernel events use masked events because they are
 814                  * flexible enough to handle both cases.  To convert to masked
 815                  * events all that's needed is to add an "all ones" umask_mask,
 816                  * (unmasked filter events don't support EXCLUDE).
 817                  */
 818                 filter->events[j++] = filter->events[i] |
 819                                       (0xFFULL << KVM_PMU_MASKED_ENTRY_UMASK_MASK_SHIFT);
 820         }
 821
 822         filter->nevents = j;
 823 }
 824
 825 static int prepare_filter_lists(struct kvm_x86_pmu_event_filter *filter)
 826 {
 827         int i;
 828
 829         if (!(filter->flags & KVM_PMU_EVENT_FLAG_MASKED_EVENTS))
 830                 convert_to_masked_filter(filter);
 831         else if (!is_masked_filter_valid(filter))
 832                 return -EINVAL;
 833
 834         /*
 835          * Sort entries by event select and includes vs. excludes so that all
 836          * entries for a given event select can be processed efficiently during
 837          * filtering.  The EXCLUDE flag uses a more significant bit than the
 838          * event select, and so the sorted list is also effectively split into
 839          * includes and excludes sub-lists.
 840          */
 841         sort(&filter->events, filter->nevents, sizeof(filter->events[0]),
 842              filter_sort_cmp, NULL);
 843
 844         i = filter->nevents;
 845         /* Find the first EXCLUDE event (only supported for masked events). */
 846         if (filter->flags & KVM_PMU_EVENT_FLAG_MASKED_EVENTS) {
 847                 for (i = 0; i < filter->nevents; i++) {
 848                         if (filter->events[i] & KVM_PMU_MASKED_ENTRY_EXCLUDE)
 849                                 break;
 850                 }
 851         }
 852
 853         filter->nr_includes = i;
 854         filter->nr_excludes = filter->nevents - filter->nr_includes;
 855         filter->includes = filter->events;
 856         filter->excludes = filter->events + filter->nr_includes;
 857
 858         return 0;
 859 }
 860
 861 int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp)
 862 {
 863         struct kvm_pmu_event_filter __user *user_filter = argp;
 864         struct kvm_x86_pmu_event_filter *filter;
 865         struct kvm_pmu_event_filter tmp;
 866         struct kvm_vcpu *vcpu;
 867         unsigned long i;
 868         size_t size;
 869         int r;
 870
 871         if (copy_from_user(&tmp, user_filter, sizeof(tmp)))
 872                 return -EFAULT;
 873
 874         if (tmp.action != KVM_PMU_EVENT_ALLOW &&
 875             tmp.action != KVM_PMU_EVENT_DENY)
 876                 return -EINVAL;
 877
 878         if (tmp.flags & ~KVM_PMU_EVENT_FLAGS_VALID_MASK)
 879                 return -EINVAL;
 880
 881         if (tmp.nevents > KVM_PMU_EVENT_FILTER_MAX_EVENTS)
 882                 return -E2BIG;
 883
 884         size = struct_size(filter, events, tmp.nevents);
 885         filter = kzalloc(size, GFP_KERNEL_ACCOUNT);
 886         if (!filter)
 887                 return -ENOMEM;
 888
 889         filter->action = tmp.action;
 890         filter->nevents = tmp.nevents;
 891         filter->fixed_counter_bitmap = tmp.fixed_counter_bitmap;
 892         filter->flags = tmp.flags;
 893
 894         r = -EFAULT;
 895         if (copy_from_user(filter->events, user_filter->events,
 896                            sizeof(filter->events[0]) * filter->nevents))
 897                 goto cleanup;
 898
 899         r = prepare_filter_lists(filter);
 900         if (r)
 901                 goto cleanup;
 902
 903         mutex_lock(&kvm->lock);
 904         filter = rcu_replace_pointer(kvm->arch.pmu_event_filter, filter,
 905                                      mutex_is_locked(&kvm->lock));
 906         mutex_unlock(&kvm->lock);
 907         synchronize_srcu_expedited(&kvm->srcu);
 908
 909         BUILD_BUG_ON(sizeof(((struct kvm_pmu *)0)->reprogram_pmi) >
 910                      sizeof(((struct kvm_pmu *)0)->__reprogram_pmi));
 911
 912         kvm_for_each_vcpu(i, vcpu, kvm)
 913                 atomic64_set(&vcpu_to_pmu(vcpu)->__reprogram_pmi, -1ull);
 914
 915         kvm_make_all_cpus_request(kvm, KVM_REQ_PMU);
 916
 917         r = 0;
 918 cleanup:
 919         kfree(filter);
 920         return r;
 921 }