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 <asm/perf_event.h>
  17 #include "x86.h"
  18 #include "cpuid.h"
  19 #include "lapic.h"
  20 #include "pmu.h"
  21
  22 /* This is enough to filter the vast majority of currently defined events. */
  23 #define KVM_PMU_EVENT_FILTER_MAX_EVENTS 300
  24
  25 /* NOTE:
  26  * - Each perf counter is defined as "struct kvm_pmc";
  27  * - There are two types of perf counters: general purpose (gp) and fixed.
  28  *   gp counters are stored in gp_counters[] and fixed counters are stored
  29  *   in fixed_counters[] respectively. Both of them are part of "struct
  30  *   kvm_pmu";
  31  * - pmu.c understands the difference between gp counters and fixed counters.
  32  *   However AMD doesn't support fixed-counters;
  33  * - There are three types of index to access perf counters (PMC):
  34  *     1. MSR (named msr): For example Intel has MSR_IA32_PERFCTRn and AMD
  35  *        has MSR_K7_PERFCTRn.
  36  *     2. MSR Index (named idx): This normally is used by RDPMC instruction.
  37  *        For instance AMD RDPMC instruction uses 0000_0003h in ECX to access
  38  *        C001_0007h (MSR_K7_PERCTR3). Intel has a similar mechanism, except
  39  *        that it also supports fixed counters. idx can be used to as index to
  40  *        gp and fixed counters.
  41  *     3. Global PMC Index (named pmc): pmc is an index specific to PMU
  42  *        code. Each pmc, stored in kvm_pmc.idx field, is unique across
  43  *        all perf counters (both gp and fixed). The mapping relationship
  44  *        between pmc and perf counters is as the following:
  45  *        * Intel: [0 .. INTEL_PMC_MAX_GENERIC-1] <=> gp counters
  46  *                 [INTEL_PMC_IDX_FIXED .. INTEL_PMC_IDX_FIXED + 2] <=> fixed
  47  *        * AMD:   [0 .. AMD64_NUM_COUNTERS-1] <=> gp counters
  48  */
  49
  50 static void kvm_pmi_trigger_fn(struct irq_work *irq_work)
  51 {
  52         struct kvm_pmu *pmu = container_of(irq_work, struct kvm_pmu, irq_work);
  53         struct kvm_vcpu *vcpu = pmu_to_vcpu(pmu);
  54
  55         kvm_pmu_deliver_pmi(vcpu);
  56 }
  57
  58 static void kvm_perf_overflow(struct perf_event *perf_event,
  59                               struct perf_sample_data *data,
  60                               struct pt_regs *regs)
  61 {
  62         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
  63         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
  64
  65         if (!test_and_set_bit(pmc->idx, pmu->reprogram_pmi)) {
  66                 __set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
  67                 kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
  68         }
  69 }
  70
  71 static void kvm_perf_overflow_intr(struct perf_event *perf_event,
  72                                    struct perf_sample_data *data,
  73                                    struct pt_regs *regs)
  74 {
  75         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
  76         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
  77
  78         if (!test_and_set_bit(pmc->idx, pmu->reprogram_pmi)) {
  79                 __set_bit(pmc->idx, (unsigned long *)&pmu->global_status);
  80                 kvm_make_request(KVM_REQ_PMU, pmc->vcpu);
  81
  82                 /*
  83                  * Inject PMI. If vcpu was in a guest mode during NMI PMI
  84                  * can be ejected on a guest mode re-entry. Otherwise we can't
  85                  * be sure that vcpu wasn't executing hlt instruction at the
  86                  * time of vmexit and is not going to re-enter guest mode until
  87                  * woken up. So we should wake it, but this is impossible from
  88                  * NMI context. Do it from irq work instead.
  89                  */
  90                 if (!kvm_is_in_guest())
  91                         irq_work_queue(&pmc_to_pmu(pmc)->irq_work);
  92                 else
  93                         kvm_make_request(KVM_REQ_PMI, pmc->vcpu);
  94         }
  95 }
  96
  97 static void pmc_reprogram_counter(struct kvm_pmc *pmc, u32 type,
  98                                   unsigned 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         attr.sample_period = get_sample_period(pmc, pmc->counter);
 115
 116         if (in_tx)
 117                 attr.config |= HSW_IN_TX;
 118         if (in_tx_cp) {
 119                 /*
 120                  * HSW_IN_TX_CHECKPOINTED is not supported with nonzero
 121                  * period. Just clear the sample period so at least
 122                  * allocating the counter doesn't fail.
 123                  */
 124                 attr.sample_period = 0;
 125                 attr.config |= HSW_IN_TX_CHECKPOINTED;
 126         }
 127
 128         event = perf_event_create_kernel_counter(&attr, -1, current,
 129                                                  intr ? kvm_perf_overflow_intr :
 130                                                  kvm_perf_overflow, pmc);
 131         if (IS_ERR(event)) {
 132                 pr_debug_ratelimited("kvm_pmu: event creation failed %ld for pmc->idx = %d\n",
 133                             PTR_ERR(event), pmc->idx);
 134                 return;
 135         }
 136
 137         pmc->perf_event = event;
 138         pmc_to_pmu(pmc)->event_count++;
 139         clear_bit(pmc->idx, pmc_to_pmu(pmc)->reprogram_pmi);
 140 }
 141
 142 static void pmc_pause_counter(struct kvm_pmc *pmc)
 143 {
 144         u64 counter = pmc->counter;
 145
 146         if (!pmc->perf_event)
 147                 return;
 148
 149         /* update counter, reset event value to avoid redundant accumulation */
 150         counter += perf_event_pause(pmc->perf_event, true);
 151         pmc->counter = counter & pmc_bitmask(pmc);
 152 }
 153
 154 static bool pmc_resume_counter(struct kvm_pmc *pmc)
 155 {
 156         if (!pmc->perf_event)
 157                 return false;
 158
 159         /* recalibrate sample period and check if it's accepted by perf core */
 160         if (perf_event_period(pmc->perf_event,
 161                               get_sample_period(pmc, pmc->counter)))
 162                 return false;
 163
 164         /* reuse perf_event to serve as pmc_reprogram_counter() does*/
 165         perf_event_enable(pmc->perf_event);
 166
 167         clear_bit(pmc->idx, (unsigned long *)&pmc_to_pmu(pmc)->reprogram_pmi);
 168         return true;
 169 }
 170
 171 void reprogram_gp_counter(struct kvm_pmc *pmc, u64 eventsel)
 172 {
 173         unsigned config, type = PERF_TYPE_RAW;
 174         u8 event_select, unit_mask;
 175         struct kvm *kvm = pmc->vcpu->kvm;
 176         struct kvm_pmu_event_filter *filter;
 177         int i;
 178         bool allow_event = true;
 179
 180         if (eventsel & ARCH_PERFMON_EVENTSEL_PIN_CONTROL)
 181                 printk_once("kvm pmu: pin control bit is ignored\n");
 182
 183         pmc->eventsel = eventsel;
 184
 185         pmc_pause_counter(pmc);
 186
 187         if (!(eventsel & ARCH_PERFMON_EVENTSEL_ENABLE) || !pmc_is_enabled(pmc))
 188                 return;
 189
 190         filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
 191         if (filter) {
 192                 for (i = 0; i < filter->nevents; i++)
 193                         if (filter->events[i] ==
 194                             (eventsel & AMD64_RAW_EVENT_MASK_NB))
 195                                 break;
 196                 if (filter->action == KVM_PMU_EVENT_ALLOW &&
 197                     i == filter->nevents)
 198                         allow_event = false;
 199                 if (filter->action == KVM_PMU_EVENT_DENY &&
 200                     i < filter->nevents)
 201                         allow_event = false;
 202         }
 203         if (!allow_event)
 204                 return;
 205
 206         event_select = eventsel & ARCH_PERFMON_EVENTSEL_EVENT;
 207         unit_mask = (eventsel & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
 208
 209         if (!(eventsel & (ARCH_PERFMON_EVENTSEL_EDGE |
 210                           ARCH_PERFMON_EVENTSEL_INV |
 211                           ARCH_PERFMON_EVENTSEL_CMASK |
 212                           HSW_IN_TX |
 213                           HSW_IN_TX_CHECKPOINTED))) {
 214                 config = kvm_x86_ops.pmu_ops->find_arch_event(pmc_to_pmu(pmc),
 215                                                       event_select,
 216                                                       unit_mask);
 217                 if (config != PERF_COUNT_HW_MAX)
 218                         type = PERF_TYPE_HARDWARE;
 219         }
 220
 221         if (type == PERF_TYPE_RAW)
 222                 config = eventsel & X86_RAW_EVENT_MASK;
 223
 224         if (pmc->current_config == eventsel && pmc_resume_counter(pmc))
 225                 return;
 226
 227         pmc_release_perf_event(pmc);
 228
 229         pmc->current_config = eventsel;
 230         pmc_reprogram_counter(pmc, type, config,
 231                               !(eventsel & ARCH_PERFMON_EVENTSEL_USR),
 232                               !(eventsel & ARCH_PERFMON_EVENTSEL_OS),
 233                               eventsel & ARCH_PERFMON_EVENTSEL_INT,
 234                               (eventsel & HSW_IN_TX),
 235                               (eventsel & HSW_IN_TX_CHECKPOINTED));
 236 }
 237 EXPORT_SYMBOL_GPL(reprogram_gp_counter);
 238
 239 void reprogram_fixed_counter(struct kvm_pmc *pmc, u8 ctrl, int idx)
 240 {
 241         unsigned en_field = ctrl & 0x3;
 242         bool pmi = ctrl & 0x8;
 243         struct kvm_pmu_event_filter *filter;
 244         struct kvm *kvm = pmc->vcpu->kvm;
 245
 246         pmc_pause_counter(pmc);
 247
 248         if (!en_field || !pmc_is_enabled(pmc))
 249                 return;
 250
 251         filter = srcu_dereference(kvm->arch.pmu_event_filter, &kvm->srcu);
 252         if (filter) {
 253                 if (filter->action == KVM_PMU_EVENT_DENY &&
 254                     test_bit(idx, (ulong *)&filter->fixed_counter_bitmap))
 255                         return;
 256                 if (filter->action == KVM_PMU_EVENT_ALLOW &&
 257                     !test_bit(idx, (ulong *)&filter->fixed_counter_bitmap))
 258                         return;
 259         }
 260
 261         if (pmc->current_config == (u64)ctrl && pmc_resume_counter(pmc))
 262                 return;
 263
 264         pmc_release_perf_event(pmc);
 265
 266         pmc->current_config = (u64)ctrl;
 267         pmc_reprogram_counter(pmc, PERF_TYPE_HARDWARE,
 268                               kvm_x86_ops.pmu_ops->find_fixed_event(idx),
 269                               !(en_field & 0x2), /* exclude user */
 270                               !(en_field & 0x1), /* exclude kernel */
 271                               pmi, false, false);
 272 }
 273 EXPORT_SYMBOL_GPL(reprogram_fixed_counter);
 274
 275 void reprogram_counter(struct kvm_pmu *pmu, int pmc_idx)
 276 {
 277         struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, pmc_idx);
 278
 279         if (!pmc)
 280                 return;
 281
 282         if (pmc_is_gp(pmc))
 283                 reprogram_gp_counter(pmc, pmc->eventsel);
 284         else {
 285                 int idx = pmc_idx - INTEL_PMC_IDX_FIXED;
 286                 u8 ctrl = fixed_ctrl_field(pmu->fixed_ctr_ctrl, idx);
 287
 288                 reprogram_fixed_counter(pmc, ctrl, idx);
 289         }
 290 }
 291 EXPORT_SYMBOL_GPL(reprogram_counter);
 292
 293 void kvm_pmu_handle_event(struct kvm_vcpu *vcpu)
 294 {
 295         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 296         int bit;
 297
 298         for_each_set_bit(bit, pmu->reprogram_pmi, X86_PMC_IDX_MAX) {
 299                 struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, bit);
 300
 301                 if (unlikely(!pmc || !pmc->perf_event)) {
 302                         clear_bit(bit, pmu->reprogram_pmi);
 303                         continue;
 304                 }
 305
 306                 reprogram_counter(pmu, bit);
 307         }
 308
 309         /*
 310          * Unused perf_events are only released if the corresponding MSRs
 311          * weren't accessed during the last vCPU time slice. kvm_arch_sched_in
 312          * triggers KVM_REQ_PMU if cleanup is needed.
 313          */
 314         if (unlikely(pmu->need_cleanup))
 315                 kvm_pmu_cleanup(vcpu);
 316 }
 317
 318 /* check if idx is a valid index to access PMU */
 319 int kvm_pmu_is_valid_rdpmc_ecx(struct kvm_vcpu *vcpu, unsigned int idx)
 320 {
 321         return kvm_x86_ops.pmu_ops->is_valid_rdpmc_ecx(vcpu, idx);
 322 }
 323
 324 bool is_vmware_backdoor_pmc(u32 pmc_idx)
 325 {
 326         switch (pmc_idx) {
 327         case VMWARE_BACKDOOR_PMC_HOST_TSC:
 328         case VMWARE_BACKDOOR_PMC_REAL_TIME:
 329         case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
 330                 return true;
 331         }
 332         return false;
 333 }
 334
 335 static int kvm_pmu_rdpmc_vmware(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 336 {
 337         u64 ctr_val;
 338
 339         switch (idx) {
 340         case VMWARE_BACKDOOR_PMC_HOST_TSC:
 341                 ctr_val = rdtsc();
 342                 break;
 343         case VMWARE_BACKDOOR_PMC_REAL_TIME:
 344                 ctr_val = ktime_get_boottime_ns();
 345                 break;
 346         case VMWARE_BACKDOOR_PMC_APPARENT_TIME:
 347                 ctr_val = ktime_get_boottime_ns() +
 348                         vcpu->kvm->arch.kvmclock_offset;
 349                 break;
 350         default:
 351                 return 1;
 352         }
 353
 354         *data = ctr_val;
 355         return 0;
 356 }
 357
 358 int kvm_pmu_rdpmc(struct kvm_vcpu *vcpu, unsigned idx, u64 *data)
 359 {
 360         bool fast_mode = idx & (1u << 31);
 361         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 362         struct kvm_pmc *pmc;
 363         u64 mask = fast_mode ? ~0u : ~0ull;
 364
 365         if (!pmu->version)
 366                 return 1;
 367
 368         if (is_vmware_backdoor_pmc(idx))
 369                 return kvm_pmu_rdpmc_vmware(vcpu, idx, data);
 370
 371         pmc = kvm_x86_ops.pmu_ops->rdpmc_ecx_to_pmc(vcpu, idx, &mask);
 372         if (!pmc)
 373                 return 1;
 374
 375         if (!(kvm_read_cr4(vcpu) & X86_CR4_PCE) &&
 376             (static_call(kvm_x86_get_cpl)(vcpu) != 0) &&
 377             (kvm_read_cr0(vcpu) & X86_CR0_PE))
 378                 return 1;
 379
 380         *data = pmc_read_counter(pmc) & mask;
 381         return 0;
 382 }
 383
 384 void kvm_pmu_deliver_pmi(struct kvm_vcpu *vcpu)
 385 {
 386         if (lapic_in_kernel(vcpu)) {
 387                 if (kvm_x86_ops.pmu_ops->deliver_pmi)
 388                         kvm_x86_ops.pmu_ops->deliver_pmi(vcpu);
 389                 kvm_apic_local_deliver(vcpu->arch.apic, APIC_LVTPC);
 390         }
 391 }
 392
 393 bool kvm_pmu_is_valid_msr(struct kvm_vcpu *vcpu, u32 msr)
 394 {
 395         return kvm_x86_ops.pmu_ops->msr_idx_to_pmc(vcpu, msr) ||
 396                 kvm_x86_ops.pmu_ops->is_valid_msr(vcpu, msr);
 397 }
 398
 399 static void kvm_pmu_mark_pmc_in_use(struct kvm_vcpu *vcpu, u32 msr)
 400 {
 401         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 402         struct kvm_pmc *pmc = kvm_x86_ops.pmu_ops->msr_idx_to_pmc(vcpu, msr);
 403
 404         if (pmc)
 405                 __set_bit(pmc->idx, pmu->pmc_in_use);
 406 }
 407
 408 int kvm_pmu_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 409 {
 410         return kvm_x86_ops.pmu_ops->get_msr(vcpu, msr_info);
 411 }
 412
 413 int kvm_pmu_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)
 414 {
 415         kvm_pmu_mark_pmc_in_use(vcpu, msr_info->index);
 416         return kvm_x86_ops.pmu_ops->set_msr(vcpu, msr_info);
 417 }
 418
 419 /* refresh PMU settings. This function generally is called when underlying
 420  * settings are changed (such as changes of PMU CPUID by guest VMs), which
 421  * should rarely happen.
 422  */
 423 void kvm_pmu_refresh(struct kvm_vcpu *vcpu)
 424 {
 425         kvm_x86_ops.pmu_ops->refresh(vcpu);
 426 }
 427
 428 void kvm_pmu_reset(struct kvm_vcpu *vcpu)
 429 {
 430         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 431
 432         irq_work_sync(&pmu->irq_work);
 433         kvm_x86_ops.pmu_ops->reset(vcpu);
 434 }
 435
 436 void kvm_pmu_init(struct kvm_vcpu *vcpu)
 437 {
 438         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 439
 440         memset(pmu, 0, sizeof(*pmu));
 441         kvm_x86_ops.pmu_ops->init(vcpu);
 442         init_irq_work(&pmu->irq_work, kvm_pmi_trigger_fn);
 443         pmu->event_count = 0;
 444         pmu->need_cleanup = false;
 445         kvm_pmu_refresh(vcpu);
 446 }
 447
 448 static inline bool pmc_speculative_in_use(struct kvm_pmc *pmc)
 449 {
 450         struct kvm_pmu *pmu = pmc_to_pmu(pmc);
 451
 452         if (pmc_is_fixed(pmc))
 453                 return fixed_ctrl_field(pmu->fixed_ctr_ctrl,
 454                         pmc->idx - INTEL_PMC_IDX_FIXED) & 0x3;
 455
 456         return pmc->eventsel & ARCH_PERFMON_EVENTSEL_ENABLE;
 457 }
 458
 459 /* Release perf_events for vPMCs that have been unused for a full time slice.  */
 460 void kvm_pmu_cleanup(struct kvm_vcpu *vcpu)
 461 {
 462         struct kvm_pmu *pmu = vcpu_to_pmu(vcpu);
 463         struct kvm_pmc *pmc = NULL;
 464         DECLARE_BITMAP(bitmask, X86_PMC_IDX_MAX);
 465         int i;
 466
 467         pmu->need_cleanup = false;
 468
 469         bitmap_andnot(bitmask, pmu->all_valid_pmc_idx,
 470                       pmu->pmc_in_use, X86_PMC_IDX_MAX);
 471
 472         for_each_set_bit(i, bitmask, X86_PMC_IDX_MAX) {
 473                 pmc = kvm_x86_ops.pmu_ops->pmc_idx_to_pmc(pmu, i);
 474
 475                 if (pmc && pmc->perf_event && !pmc_speculative_in_use(pmc))
 476                         pmc_stop_counter(pmc);
 477         }
 478
 479         if (kvm_x86_ops.pmu_ops->cleanup)
 480                 kvm_x86_ops.pmu_ops->cleanup(vcpu);
 481
 482         bitmap_zero(pmu->pmc_in_use, X86_PMC_IDX_MAX);
 483 }
 484
 485 void kvm_pmu_destroy(struct kvm_vcpu *vcpu)
 486 {
 487         kvm_pmu_reset(vcpu);
 488 }
 489
 490 int kvm_vm_ioctl_set_pmu_event_filter(struct kvm *kvm, void __user *argp)
 491 {
 492         struct kvm_pmu_event_filter tmp, *filter;
 493         size_t size;
 494         int r;
 495
 496         if (copy_from_user(&tmp, argp, sizeof(tmp)))
 497                 return -EFAULT;
 498
 499         if (tmp.action != KVM_PMU_EVENT_ALLOW &&
 500             tmp.action != KVM_PMU_EVENT_DENY)
 501                 return -EINVAL;
 502
 503         if (tmp.flags != 0)
 504                 return -EINVAL;
 505
 506         if (tmp.nevents > KVM_PMU_EVENT_FILTER_MAX_EVENTS)
 507                 return -E2BIG;
 508
 509         size = struct_size(filter, events, tmp.nevents);
 510         filter = kmalloc(size, GFP_KERNEL_ACCOUNT);
 511         if (!filter)
 512                 return -ENOMEM;
 513
 514         r = -EFAULT;
 515         if (copy_from_user(filter, argp, size))
 516                 goto cleanup;
 517
 518         /* Ensure nevents can't be changed between the user copies. */
 519         *filter = tmp;
 520
 521         mutex_lock(&kvm->lock);
 522         filter = rcu_replace_pointer(kvm->arch.pmu_event_filter, filter,
 523                                      mutex_is_locked(&kvm->lock));
 524         mutex_unlock(&kvm->lock);
 525
 526         synchronize_srcu_expedited(&kvm->srcu);
 527         r = 0;
 528 cleanup:
 529         kfree(filter);
 530         return r;
 531 }