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
  13 #include <linux/types.h>
  14 #include <linux/kvm_host.h>
  15 #include <linux/perf_event.h>
  16 #include <linux/bsearch.h>
  17 #include <linux/sort.h>
  18 #include <asm/perf_event.h>
  19 #include "x86.h"
  20 #include "cpuid.h"
  21 #include "lapic.h"
  22 #include "pmu.h"
  23
  24 /* This is enough to filter the vast majority of currently defined events. */
  25 #define KVM_PMU_EVENT_FILTER_MAX_EVENTS 300
  26
  27 /* NOTE:
  28  * - Each perf counter is defined as "struct kvm_pmc";
  29  * - There are two types of perf counters: general purpose (gp) and fixed.
  30  *   gp counters are stored in gp_counters[] and fixed counters are stored
  31  *   in fixed_counters[] respectively. Both of them are part of "struct
  32  *   kvm_pmu";
  33  * - pmu.c understands the difference between gp counters and fixed counters.
  34  *   However AMD doesn't support fixed-counters;
  35  * - There are three types of index to access perf counters (PMC):
  36  *     1. MSR (named msr): For example Intel has MSR_IA32_PERFCTRn and AMD
  37  *        has MSR_K7_PERFCTRn.
  38  *     2. MSR Index (named idx): This normally is used by RDPMC instruction.
  39  *        For instance AMD RDPMC instruction uses 0000_0003h in ECX to access
  40  *        C001_0007h (MSR_K7_PERCTR3). Intel has a similar mechanism, except
  41  *        that it also supports fixed counters. idx can be used to as index to
  42  *        gp and fixed counters.
  43  *     3. Global PMC Index (named pmc): pmc is an index specific to PMU
  44  *        code. Each pmc, stored in kvm_pmc.idx field, is unique across
  45  *        all perf counters (both gp and fixed). The mapping relationship
  46  *        between pmc and perf counters is as the following:
  47  *        * Intel: [0 .. INTEL_PMC_MAX_GENERIC-1] <=> gp counters
  48  *                 [INTEL_PMC_IDX_FIXED .. INTEL_PMC_IDX_FIXED + 2] <=> fixed
  49  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] <=> gp counters
  50  */
  51
  52 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
  53 {
  54         struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
  55         struct kvm_vcpu *vcpu = pmu_to_vcpu(pmu);
  56
  57         kvm_pmu_deliver_pmi(vcpu);
  58 }
  59
  60 static inline void __kvm_perf_overflow(struct kvm_pmc *pmc, bool in_pmi)
  61 {
  62         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
  63
  64         /* Ignore counters that have been reprogrammed already. */
  65         if (test_and_set_bit(pmc->idx, pmu->reprogram_pmi))
  66                 return;
  67
  68         __set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
  69         kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
  70
  71         if (!pmc->intr)
  72                 return;
  73
  74         /*
  75          * Inject PMI. If vcpu was in a guest mode during NMI PMI
  76          * can be ejected on a guest mode re-entry. Otherwise we can't
  77          * be sure that vcpu wasn't executing hlt instruction at the
  78          * time of vmexit and is not going to re-enter guest mode until
  79          * woken up. So we should wake it, but this is impossible from
  80          * NMI context. Do it from irq work instead.
  81          */
  82         if (in_pmi && !kvm_handling_nmi_from_guest(pmc->vcpu))
  83                 irq_work_queue(&pmc_to_pmu(pmc)->irq_work);
  84         else
  85                 kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
  86 }
  87
  88 static void kvm_perf_overflow(struct perf_event *perf_event,
  89                               struct perf_sample_data *data,
  90                               struct pt_regs *regs)
  91 {
  92         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
  93
  94         __kvm_perf_overflow(pmc, true);
  95 }
  96
  97 static void pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type,
  98                                   u64 config, bool exclude_user,
  99                                   bool exclude_kernel, bool intr,
 100                                   bool in_tx, bool in_tx_cp)
 101 {
 102         struct perf_event *event;
 103         struct perf_event_attr attr = {
 104                 .type = type,
 105                 .size = sizeof(attr),
 106                 .pinned = true,
 107                 .exclude_idle = true,
 108                 .exclude_host = 1,
 109                 .exclude_user = exclude_user,
 110                 .exclude_kernel = exclude_kernel,
 111                 .config = config,
 112         };
 113
 114         if (type == PERF_TYPE_HARDWARE && config >= PERF_COUNT_HW_MAX)
 115                 return;
 116
 117         attr.sample_period = get_sample_period(pmc, pmc->counter);
 118
 119         if (in_tx)
 120                 attr.config |= HSW_IN_TX;
 121         if (in_tx_cp) {
 122                 /*
 123                  * HSW_IN_TX_CHECKPOINTED is not supported with nonzero
 124                  * period. Just clear the sample period so at least
 125                  * allocating the counter doesn't fail.
 126                  */
 127                 attr.sample_period = 0;
 128                 attr.config |= HSW_IN_TX_CHECKPOINTED;
 129         }
 130
 131         event = perf_event_create_kernel_counter(&attr, -1, current,
 132                                                  kvm_perf_overflow, pmc);
 133         if (IS_ERR(event)) {
 134                 pr_debug_ratelimited("kvm_pmu: event creation failed %ld for pmc->idx = %d\n",
 135                             PTR_ERR(event), pmc->idx);
 136                 return;
 137         }
 138
 139         pmc->perf_event = event;
 140         pmc_to_pmu(pmc)->event_count++;
 141         clear_bit(pmc->idx, pmc_to_pmu(pmc)->reprogram_pmi);
 142         pmc->is_paused = false;
 143         pmc->intr = intr;
 144 }
 145
 146 static void pmc_pause_counter(struct kvm_pmc *pmc)
 147 {
 148         u64 counter = pmc->counter;
 149
 150         if (!pmc->perf_event || pmc->is_paused)
 151                 return;
 152
 153         /* update counter, reset event value to avoid redundant accumulation */
 154         counter += perf_event_pause(pmc->perf_event, true);
 155         pmc->counter = counter & pmc_bitmask(pmc);
 156         pmc->is_paused = true;
 157 }
 158
 159 static bool pmc_resume_counter(struct kvm_pmc *pmc)
 160 {
 161         if (!pmc->perf_event)
 162                 return false;
 163
 164         /* recalibrate sample period and check if it's accepted by perf core */
 165         if (perf_event_period(pmc->perf_event,
 166                               get_sample_period(pmc, pmc->counter)))
 167                 return false;
 168
 169         /* reuse perf_event to serve as pmc_reprogram_counter() does*/
 170         perf_event_enable(pmc->perf_event);
 171         pmc->is_paused = false;
 172
 173         clear_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->reprogram_pmi);
 174         return true;
 175 }
 176
 177 static int cmp_u64(const void *a, const void *b)
 178 {
 179         return *(__u64 *)a - *(__u64 *)b;
 180 }
 181
 182 void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
 183 {
 184         u64 config;
 185         u32 type = PERF_TYPE_RAW;
 186         struct kvm *kvm = pmc->vcpu->kvm;
 187         struct kvm_pmu_event_filter *filter;
 188         bool allow_event = true;
 189
 190         if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
 191                 printk_once("kvm pmu: pin control bit is ignored\n");
 192
 193         pmc->eventsel = eventsel;
 194
 195         pmc_pause_counter(pmc);
 196
 197         if (!(eventsel & ARCH_PERFMON_EVENTSEL_ENABLE) || !pmc_is_enabled(pmc))
 198                 return;
 199
 200         filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
 201         if (filter) {
 202                 __u64 key = eventsel & AMD64_RAW_EVENT_MASK_NB;
 203
 204                 if (bsearch(&key, filter->events, filter->nevents,
 205                             sizeof(__u64), cmp_u64))
 206                         allow_event = filter->action == KVM_PMU_EVENT_ALLOW;
 207                 else
 208                         allow_event = filter->action == KVM_PMU_EVENT_DENY;
 209         }
 210         if (!allow_event)
 211                 return;
 212
 213         if (!(eventsel & (ARCH_PERFMON_EVENTSEL_EDGE |
 214                           ARCH_PERFMON_EVENTSEL_INV |
 215                           ARCH_PERFMON_EVENTSEL_CMASK |
 216                           HSW_IN_TX |
 217                           HSW_IN_TX_CHECKPOINTED))) {
 218                 config = kvm_x86_ops.pmu_ops->pmc_perf_hw_id(pmc);
 219                 if (config != PERF_COUNT_HW_MAX)
 220                         type = PERF_TYPE_HARDWARE;
 221         }
 222
 223         if (type == PERF_TYPE_RAW)
 224                 config = eventsel & AMD64_RAW_EVENT_MASK;
 225
 226         if (pmc->current_config == eventsel && pmc_resume_counter(pmc))
 227                 return;
 228
 229         pmc_release_perf_event(pmc);
 230
 231         pmc->current_config = eventsel;
 232         pmc_reprogram_counter(pmc, type, config,
 233                               !(eventsel & ARCH_PERFMON_EVENTSEL_USR),
 234                               !(eventsel & ARCH_PERFMON_EVENTSEL_OS),
 235                               eventsel & ARCH_PERFMON_EVENTSEL_INT,
 236                               (eventsel & HSW_IN_TX),
 237                               (eventsel & HSW_IN_TX_CHECKPOINTED));
 238 }
 239 EXPORT_SYMBOL_GPL(reprogram_gp_counter);
 240
 241 void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
 242 {
 243         unsigned en_field = ctrl & 0x3;
 244         bool pmi = ctrl & 0x8;
 245         struct kvm_pmu_event_filter *filter;
 246         struct kvm *kvm = pmc->vcpu->kvm;
 247
 248         pmc_pause_counter(pmc);
 249
 250         if (!en_field || !pmc_is_enabled(pmc))
 251                 return;
 252
 253         filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
 254         if (filter) {
 255                 if (filter->action == KVM_PMU_EVENT_DENY &&
 256                     test_bit(idx, (ulong *)&filter->fixed_counter_bitmap))
 257                         return;
 258                 if (filter->action == KVM_PMU_EVENT_ALLOW &&
 259                     !test_bit(idx, (ulong *)&filter->fixed_counter_bitmap))
 260                         return;
 261         }
 262
 263         if (pmc->current_config == (u64)ctrl && pmc_resume_counter(pmc))
 264                 return;
 265
 266         pmc_release_perf_event(pmc);
 267
 268         pmc->current_config = (u64)ctrl;
 269         pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
 270                               kvm_x86_ops.pmu_ops->pmc_perf_hw_id(pmc),
 271                               !(en_field & 0x2), /* exclude user */
 272                               !(en_field & 0x1), /* exclude kernel */
 273                               pmi, false, false);
 274 }
 275 EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
 276
 277 void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
 278 {
 279         struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, pmc_idx);
 280
 281         if (!pmc)
 282                 return;
 283
 284         if (pmc_is_gp(pmc))
 285                 reprogram_gp_counter(pmc, pmc->eventsel);
 286         else {
 287                 int idx = pmc_idx - INTEL_PMC_IDX_FIXED;
 288                 u8 ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl, idx);
 289
 290                 reprogram_fixed_counter(pmc, ctrl, idx);
 291         }
 292 }
 293 EXPORT_SYMBOL_GPL(reprogram_counter);
 294
 295 void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 296 {
 297         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 298         int bit;
 299
 300         for_each_set_bit(bit, pmu->reprogram_pmi, X86_PMC_IDX_MAX) {
 301                 struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, bit);
 302
 303                 if (unlikely(!pmc || !pmc->perf_event)) {
 304                         clear_bit(bit, pmu->reprogram_pmi);
 305                         continue;
 306                 }
 307
 308                 reprogram_counter(pmu, bit);
 309         }
 310
 311         /*
 312          * Unused perf_events are only released if the corresponding MSRs
 313          * weren't accessed during the last vCPU time slice. kvm_arch_sched_in
 314          * triggers KVM_REQ_PMU if cleanup is needed.
 315          */
 316         if (unlikely(pmu->need_cleanup))
 317                 kvm_pmu_cleanup(vcpu);
 318 }
 319
 320 /* check if idx is a valid index to access PMU */
 321 bool kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
 322 {
 323         return kvm_x86_ops.pmu_ops->is_valid_rdpmc_ecx(vcpu, idx);
 324 }
 325
 326 bool is_vmware_backdoor_pmc(u32 pmc_idx)
 327 {
 328         switch (pmc_idx) {
 329         case VMWARE_BACKDOOR_PMC_HOST_TSC:
 330         case VMWARE_BACKDOOR_PMC_REAL_TIME:
 331         case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
 332                 return true;
 333         }
 334         return false;
 335 }
 336
 337 static int kvm_pmu_rdpmc_vmware(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 338 {
 339         u64 ctr_val;
 340
 341         switch (idx) {
 342         case VMWARE_BACKDOOR_PMC_HOST_TSC:
 343                 ctr_val = rdtsc();
 344                 break;
 345         case VMWARE_BACKDOOR_PMC_REAL_TIME:
 346                 ctr_val = ktime_get_boottime_ns();
 347                 break;
 348         case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
 349                 ctr_val = ktime_get_boottime_ns() +
 350                         vcpu->kvm->arch.kvmclock_offset;
 351                 break;
 352         default:
 353                 return 1;
 354         }
 355
 356         *data = ctr_val;
 357         return 0;
 358 }
 359
 360 int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 361 {
 362         bool fast_mode = idx & (1u << 31);
 363         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 364         struct kvm_pmc *pmc;
 365         u64 mask = fast_mode ? ~0u : ~0ull;
 366
 367         if (!pmu->version)
 368                 return 1;
 369
 370         if (is_vmware_backdoor_pmc(idx))
 371                 return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
 372
 373         pmc = kvm_x86_ops.pmu_ops->rdpmc_ecx_to_pmc(vcpu, idx, &mask);
 374         if (!pmc)
 375                 return 1;
 376
 377         if (!(kvm_read_cr4(vcpu) & X86_CR4_PCE) &&
 378             (static_call(kvm_x86_get_cpl)(vcpu) != 0) &&
 379             (kvm_read_cr0(vcpu) & X86_CR0_PE))
 380                 return 1;
 381
 382         *data = pmc_read_counter(pmc) & mask;
 383         return 0;
 384 }
 385
 386 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
 387 {
 388         if (lapic_in_kernel(vcpu)) {
 389                 if (kvm_x86_ops.pmu_ops->deliver_pmi)
 390                         kvm_x86_ops.pmu_ops->deliver_pmi(vcpu);
 391                 kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
 392         }
 393 }
 394
 395 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
 396 {
 397         return kvm_x86_ops.pmu_ops->msr_idx_to_pmc(vcpu, msr) ||
 398                 kvm_x86_ops.pmu_ops->is_valid_msr(vcpu, msr);
 399 }
 400
 401 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 402 {
 403         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 404         struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->msr_idx_to_pmc(vcpu, msr);
 405
 406         if (pmc)
 407                 __set_bit(pmc->idx, pmu->pmc_in_use);
 408 }
 409
 410 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 411 {
 412         return kvm_x86_ops.pmu_ops->get_msr(vcpu, msr_info);
 413 }
 414
 415 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 416 {
 417         kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
 418         return kvm_x86_ops.pmu_ops->set_msr(vcpu, msr_info);
 419 }
 420
 421 /* refresh PMU settings. This function generally is called when underlying
 422  * settings are changed (such as changes of PMU CPUID by guest VMs), which
 423  * should rarely happen.
 424  */
 425 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 426 {
 427         kvm_x86_ops.pmu_ops->refresh(vcpu);
 428 }
 429
 430 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 431 {
 432         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 433
 434         irq_work_sync(&pmu->irq_work);
 435         kvm_x86_ops.pmu_ops->reset(vcpu);
 436 }
 437
 438 void kvm_pmu_init(struct kvm_vcpu *vcpu)
 439 {
 440         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 441
 442         memset(pmu, 0, sizeof(*pmu));
 443         kvm_x86_ops.pmu_ops->init(vcpu);
 444         init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
 445         pmu->event_count = 0;
 446         pmu->need_cleanup = false;
 447         kvm_pmu_refresh(vcpu);
 448 }
 449
 450 static inline bool pmc_speculative_in_use(struct kvm_pmc *pmc)
 451 {
 452         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
 453
 454         if (pmc_is_fixed(pmc))
 455                 return fixed_ctrl_field(pmu->fixed_ctr_ctrl,
 456                         pmc->idx - INTEL_PMC_IDX_FIXED) & 0x3;
 457
 458         return pmc->eventsel & ARCH_PERFMON_EVENTSEL_ENABLE;
 459 }
 460
 461 /* Release perf_events for vPMCs that have been unused for a full time slice.  */
 462 void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
 463 {
 464         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 465         struct kvm_pmc *pmc = NULL;
 466         DECLARE_BITMAP(bitmask, X86_PMC_IDX_MAX);
 467         int i;
 468
 469         pmu->need_cleanup = false;
 470
 471         bitmap_andnot(bitmask, pmu->all_valid_pmc_idx,
 472                       pmu->pmc_in_use, X86_PMC_IDX_MAX);
 473
 474         for_each_set_bit(i, bitmask, X86_PMC_IDX_MAX) {
 475                 pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, i);
 476
 477                 if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
 478                         pmc_stop_counter(pmc);
 479         }
 480
 481         if (kvm_x86_ops.pmu_ops->cleanup)
 482                 kvm_x86_ops.pmu_ops->cleanup(vcpu);
 483
 484         bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
 485 }
 486
 487 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
 488 {
 489         kvm_pmu_reset(vcpu);
 490 }
 491
 492 static void kvm_pmu_incr_counter(struct kvm_pmc *pmc)
 493 {
 494         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
 495         u64 prev_count;
 496
 497         prev_count = pmc->counter;
 498         pmc->counter = (pmc->counter + 1) & pmc_bitmask(pmc);
 499
 500         reprogram_counter(pmu, pmc->idx);
 501         if (pmc->counter < prev_count)
 502                 __kvm_perf_overflow(pmc, false);
 503 }
 504
 505 static inline bool eventsel_match_perf_hw_id(struct kvm_pmc *pmc,
 506         unsigned int perf_hw_id)
 507 {
 508         u64 old_eventsel = pmc->eventsel;
 509         unsigned int config;
 510
 511         pmc->eventsel &= (ARCH_PERFMON_EVENTSEL_EVENT | ARCH_PERFMON_EVENTSEL_UMASK);
 512         config = kvm_x86_ops.pmu_ops->pmc_perf_hw_id(pmc);
 513         pmc->eventsel = old_eventsel;
 514         return config == perf_hw_id;
 515 }
 516
 517 static inline bool cpl_is_matched(struct kvm_pmc *pmc)
 518 {
 519         bool select_os, select_user;
 520         u64 config = pmc->current_config;
 521
 522         if (pmc_is_gp(pmc)) {
 523                 select_os = config & ARCH_PERFMON_EVENTSEL_OS;
 524                 select_user = config & ARCH_PERFMON_EVENTSEL_USR;
 525         } else {
 526                 select_os = config & 0x1;
 527                 select_user = config & 0x2;
 528         }
 529
 530         return (static_call(kvm_x86_get_cpl)(pmc->vcpu) == 0) ? select_os : select_user;
 531 }
 532
 533 void kvm_pmu_trigger_event(struct kvm_vcpu *vcpu, u64 perf_hw_id)
 534 {
 535         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 536         struct kvm_pmc *pmc;
 537         int i;
 538
 539         for_each_set_bit(i, pmu->all_valid_pmc_idx, X86_PMC_IDX_MAX) {
 540                 pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, i);
 541
 542                 if (!pmc || !pmc_is_enabled(pmc) || !pmc_speculative_in_use(pmc))
 543                         continue;
 544
 545                 /* Ignore checks for edge detect, pin control, invert and CMASK bits */
 546                 if (eventsel_match_perf_hw_id(pmc, perf_hw_id) && cpl_is_matched(pmc))
 547                         kvm_pmu_incr_counter(pmc);
 548         }
 549 }
 550 EXPORT_SYMBOL_GPL(kvm_pmu_trigger_event);
 551
 552 int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp)
 553 {
 554         struct kvm_pmu_event_filter tmp, *filter;
 555         size_t size;
 556         int r;
 557
 558         if (copy_from_user(&tmp, argp, sizeof(tmp)))
 559                 return -EFAULT;
 560
 561         if (tmp.action != KVM_PMU_EVENT_ALLOW &&
 562             tmp.action != KVM_PMU_EVENT_DENY)
 563                 return -EINVAL;
 564
 565         if (tmp.flags != 0)
 566                 return -EINVAL;
 567
 568         if (tmp.nevents > KVM_PMU_EVENT_FILTER_MAX_EVENTS)
 569                 return -E2BIG;
 570
 571         size = struct_size(filter, events, tmp.nevents);
 572         filter = kmalloc(size, GFP_KERNEL_ACCOUNT);
 573         if (!filter)
 574                 return -ENOMEM;
 575
 576         r = -EFAULT;
 577         if (copy_from_user(filter, argp, size))
 578                 goto cleanup;
 579
 580         /* Ensure nevents can't be changed between the user copies. */
 581         *filter = tmp;
 582
 583         /*
 584          * Sort the in-kernel list so that we can search it with bsearch.
 585          */
 586         sort(&filter->events, filter->nevents, sizeof(__u64), cmp_u64, NULL);
 587
 588         mutex_lock(&kvm->lock);
 589         filter = rcu_replace_pointer(kvm->arch.pmu_event_filter, filter,
 590                                      mutex_is_locked(&kvm->lock));
 591         mutex_unlock(&kvm->lock);
 592
 593         synchronize_srcu_expedited(&kvm->srcu);
 594         r = 0;
 595 cleanup:
 596         kfree(filter);
 597         return r;
 598 }