arch/arm64/kvm/pmu-emul.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2015 Linaro Ltd.
   4  * Author: Shannon Zhao <shannon.zhao@linaro.org>
   5  */
   6
   7 #include <linux/cpu.h>
   8 #include <linux/kvm.h>
   9 #include <linux/kvm_host.h>
  10 #include <linux/list.h>
  11 #include <linux/perf_event.h>
  12 #include <linux/perf/arm_pmu.h>
  13 #include <linux/uaccess.h>
  14 #include <asm/kvm_emulate.h>
  15 #include <kvm/arm_pmu.h>
  16 #include <kvm/arm_vgic.h>
  17
  18 #define PERF_ATTR_CFG1_COUNTER_64BIT    BIT(0)
  19
  20 DEFINE_STATIC_KEY_FALSE(kvm_arm_pmu_available);
  21
  22 static LIST_HEAD(arm_pmus);
  23 static DEFINE_MUTEX(arm_pmus_lock);
  24
  25 static void kvm_pmu_create_perf_event(struct kvm_pmc *pmc);
  26 static void kvm_pmu_release_perf_event(struct kvm_pmc *pmc);
  27
  28 static struct kvm_vcpu *kvm_pmc_to_vcpu(const struct kvm_pmc *pmc)
  29 {
  30         return container_of(pmc, struct kvm_vcpu, arch.pmu.pmc[pmc->idx]);
  31 }
  32
  33 static struct kvm_pmc *kvm_vcpu_idx_to_pmc(struct kvm_vcpu *vcpu, int cnt_idx)
  34 {
  35         return &vcpu->arch.pmu.pmc[cnt_idx];
  36 }
  37
  38 static u32 kvm_pmu_event_mask(struct kvm *kvm)
  39 {
  40         unsigned int pmuver;
  41
  42         pmuver = kvm->arch.arm_pmu->pmuver;
  43
  44         switch (pmuver) {
  45         case ID_AA64DFR0_EL1_PMUVer_IMP:
  46                 return GENMASK(9, 0);
  47         case ID_AA64DFR0_EL1_PMUVer_V3P1:
  48         case ID_AA64DFR0_EL1_PMUVer_V3P4:
  49         case ID_AA64DFR0_EL1_PMUVer_V3P5:
  50         case ID_AA64DFR0_EL1_PMUVer_V3P7:
  51                 return GENMASK(15, 0);
  52         default:                /* Shouldn't be here, just for sanity */
  53                 WARN_ONCE(1, "Unknown PMU version %d\n", pmuver);
  54                 return 0;
  55         }
  56 }
  57
  58 /**
  59  * kvm_pmc_is_64bit - determine if counter is 64bit
  60  * @pmc: counter context
  61  */
  62 static bool kvm_pmc_is_64bit(struct kvm_pmc *pmc)
  63 {
  64         return (pmc->idx == ARMV8_PMU_CYCLE_IDX ||
  65                 kvm_pmu_is_3p5(kvm_pmc_to_vcpu(pmc)));
  66 }
  67
  68 static bool kvm_pmc_has_64bit_overflow(struct kvm_pmc *pmc)
  69 {
  70         u64 val = __vcpu_sys_reg(kvm_pmc_to_vcpu(pmc), PMCR_EL0);
  71
  72         return (pmc->idx < ARMV8_PMU_CYCLE_IDX && (val & ARMV8_PMU_PMCR_LP)) ||
  73                (pmc->idx == ARMV8_PMU_CYCLE_IDX && (val & ARMV8_PMU_PMCR_LC));
  74 }
  75
  76 static bool kvm_pmu_counter_can_chain(struct kvm_pmc *pmc)
  77 {
  78         return (!(pmc->idx & 1) && (pmc->idx + 1) < ARMV8_PMU_CYCLE_IDX &&
  79                 !kvm_pmc_has_64bit_overflow(pmc));
  80 }
  81
  82 static u32 counter_index_to_reg(u64 idx)
  83 {
  84         return (idx == ARMV8_PMU_CYCLE_IDX) ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + idx;
  85 }
  86
  87 static u32 counter_index_to_evtreg(u64 idx)
  88 {
  89         return (idx == ARMV8_PMU_CYCLE_IDX) ? PMCCFILTR_EL0 : PMEVTYPER0_EL0 + idx;
  90 }
  91
  92 static u64 kvm_pmu_get_pmc_value(struct kvm_pmc *pmc)
  93 {
  94         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
  95         u64 counter, reg, enabled, running;
  96
  97         reg = counter_index_to_reg(pmc->idx);
  98         counter = __vcpu_sys_reg(vcpu, reg);
  99
 100         /*
 101          * The real counter value is equal to the value of counter register plus
 102          * the value perf event counts.
 103          */
 104         if (pmc->perf_event)
 105                 counter += perf_event_read_value(pmc->perf_event, &enabled,
 106                                                  &running);
 107
 108         if (!kvm_pmc_is_64bit(pmc))
 109                 counter = lower_32_bits(counter);
 110
 111         return counter;
 112 }
 113
 114 /**
 115  * kvm_pmu_get_counter_value - get PMU counter value
 116  * @vcpu: The vcpu pointer
 117  * @select_idx: The counter index
 118  */
 119 u64 kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u64 select_idx)
 120 {
 121         if (!kvm_vcpu_has_pmu(vcpu))
 122                 return 0;
 123
 124         return kvm_pmu_get_pmc_value(kvm_vcpu_idx_to_pmc(vcpu, select_idx));
 125 }
 126
 127 static void kvm_pmu_set_pmc_value(struct kvm_pmc *pmc, u64 val, bool force)
 128 {
 129         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
 130         u64 reg;
 131
 132         kvm_pmu_release_perf_event(pmc);
 133
 134         reg = counter_index_to_reg(pmc->idx);
 135
 136         if (vcpu_mode_is_32bit(vcpu) && pmc->idx != ARMV8_PMU_CYCLE_IDX &&
 137             !force) {
 138                 /*
 139                  * Even with PMUv3p5, AArch32 cannot write to the top
 140                  * 32bit of the counters. The only possible course of
 141                  * action is to use PMCR.P, which will reset them to
 142                  * 0 (the only use of the 'force' parameter).
 143                  */
 144                 val  = __vcpu_sys_reg(vcpu, reg) & GENMASK(63, 32);
 145                 val |= lower_32_bits(val);
 146         }
 147
 148         __vcpu_sys_reg(vcpu, reg) = val;
 149
 150         /* Recreate the perf event to reflect the updated sample_period */
 151         kvm_pmu_create_perf_event(pmc);
 152 }
 153
 154 /**
 155  * kvm_pmu_set_counter_value - set PMU counter value
 156  * @vcpu: The vcpu pointer
 157  * @select_idx: The counter index
 158  * @val: The counter value
 159  */
 160 void kvm_pmu_set_counter_value(struct kvm_vcpu *vcpu, u64 select_idx, u64 val)
 161 {
 162         if (!kvm_vcpu_has_pmu(vcpu))
 163                 return;
 164
 165         kvm_pmu_set_pmc_value(kvm_vcpu_idx_to_pmc(vcpu, select_idx), val, false);
 166 }
 167
 168 /**
 169  * kvm_pmu_release_perf_event - remove the perf event
 170  * @pmc: The PMU counter pointer
 171  */
 172 static void kvm_pmu_release_perf_event(struct kvm_pmc *pmc)
 173 {
 174         if (pmc->perf_event) {
 175                 perf_event_disable(pmc->perf_event);
 176                 perf_event_release_kernel(pmc->perf_event);
 177                 pmc->perf_event = NULL;
 178         }
 179 }
 180
 181 /**
 182  * kvm_pmu_stop_counter - stop PMU counter
 183  * @pmc: The PMU counter pointer
 184  *
 185  * If this counter has been configured to monitor some event, release it here.
 186  */
 187 static void kvm_pmu_stop_counter(struct kvm_pmc *pmc)
 188 {
 189         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
 190         u64 reg, val;
 191
 192         if (!pmc->perf_event)
 193                 return;
 194
 195         val = kvm_pmu_get_pmc_value(pmc);
 196
 197         reg = counter_index_to_reg(pmc->idx);
 198
 199         __vcpu_sys_reg(vcpu, reg) = val;
 200
 201         kvm_pmu_release_perf_event(pmc);
 202 }
 203
 204 /**
 205  * kvm_pmu_vcpu_init - assign pmu counter idx for cpu
 206  * @vcpu: The vcpu pointer
 207  *
 208  */
 209 void kvm_pmu_vcpu_init(struct kvm_vcpu *vcpu)
 210 {
 211         int i;
 212         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 213
 214         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++)
 215                 pmu->pmc[i].idx = i;
 216 }
 217
 218 /**
 219  * kvm_pmu_vcpu_reset - reset pmu state for cpu
 220  * @vcpu: The vcpu pointer
 221  *
 222  */
 223 void kvm_pmu_vcpu_reset(struct kvm_vcpu *vcpu)
 224 {
 225         unsigned long mask = kvm_pmu_valid_counter_mask(vcpu);
 226         int i;
 227
 228         for_each_set_bit(i, &mask, 32)
 229                 kvm_pmu_stop_counter(kvm_vcpu_idx_to_pmc(vcpu, i));
 230 }
 231
 232 /**
 233  * kvm_pmu_vcpu_destroy - free perf event of PMU for cpu
 234  * @vcpu: The vcpu pointer
 235  *
 236  */
 237 void kvm_pmu_vcpu_destroy(struct kvm_vcpu *vcpu)
 238 {
 239         int i;
 240
 241         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++)
 242                 kvm_pmu_release_perf_event(kvm_vcpu_idx_to_pmc(vcpu, i));
 243         irq_work_sync(&vcpu->arch.pmu.overflow_work);
 244 }
 245
 246 u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu)
 247 {
 248         u64 val = __vcpu_sys_reg(vcpu, PMCR_EL0) >> ARMV8_PMU_PMCR_N_SHIFT;
 249
 250         val &= ARMV8_PMU_PMCR_N_MASK;
 251         if (val == 0)
 252                 return BIT(ARMV8_PMU_CYCLE_IDX);
 253         else
 254                 return GENMASK(val - 1, 0) | BIT(ARMV8_PMU_CYCLE_IDX);
 255 }
 256
 257 /**
 258  * kvm_pmu_enable_counter_mask - enable selected PMU counters
 259  * @vcpu: The vcpu pointer
 260  * @val: the value guest writes to PMCNTENSET register
 261  *
 262  * Call perf_event_enable to start counting the perf event
 263  */
 264 void kvm_pmu_enable_counter_mask(struct kvm_vcpu *vcpu, u64 val)
 265 {
 266         int i;
 267         if (!kvm_vcpu_has_pmu(vcpu))
 268                 return;
 269
 270         if (!(__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) || !val)
 271                 return;
 272
 273         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
 274                 struct kvm_pmc *pmc;
 275
 276                 if (!(val & BIT(i)))
 277                         continue;
 278
 279                 pmc = kvm_vcpu_idx_to_pmc(vcpu, i);
 280
 281                 if (!pmc->perf_event) {
 282                         kvm_pmu_create_perf_event(pmc);
 283                 } else {
 284                         perf_event_enable(pmc->perf_event);
 285                         if (pmc->perf_event->state != PERF_EVENT_STATE_ACTIVE)
 286                                 kvm_debug("fail to enable perf event\n");
 287                 }
 288         }
 289 }
 290
 291 /**
 292  * kvm_pmu_disable_counter_mask - disable selected PMU counters
 293  * @vcpu: The vcpu pointer
 294  * @val: the value guest writes to PMCNTENCLR register
 295  *
 296  * Call perf_event_disable to stop counting the perf event
 297  */
 298 void kvm_pmu_disable_counter_mask(struct kvm_vcpu *vcpu, u64 val)
 299 {
 300         int i;
 301
 302         if (!kvm_vcpu_has_pmu(vcpu) || !val)
 303                 return;
 304
 305         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
 306                 struct kvm_pmc *pmc;
 307
 308                 if (!(val & BIT(i)))
 309                         continue;
 310
 311                 pmc = kvm_vcpu_idx_to_pmc(vcpu, i);
 312
 313                 if (pmc->perf_event)
 314                         perf_event_disable(pmc->perf_event);
 315         }
 316 }
 317
 318 static u64 kvm_pmu_overflow_status(struct kvm_vcpu *vcpu)
 319 {
 320         u64 reg = 0;
 321
 322         if ((__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E)) {
 323                 reg = __vcpu_sys_reg(vcpu, PMOVSSET_EL0);
 324                 reg &= __vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
 325                 reg &= __vcpu_sys_reg(vcpu, PMINTENSET_EL1);
 326         }
 327
 328         return reg;
 329 }
 330
 331 static void kvm_pmu_update_state(struct kvm_vcpu *vcpu)
 332 {
 333         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 334         bool overflow;
 335
 336         if (!kvm_vcpu_has_pmu(vcpu))
 337                 return;
 338
 339         overflow = !!kvm_pmu_overflow_status(vcpu);
 340         if (pmu->irq_level == overflow)
 341                 return;
 342
 343         pmu->irq_level = overflow;
 344
 345         if (likely(irqchip_in_kernel(vcpu->kvm))) {
 346                 int ret = kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
 347                                               pmu->irq_num, overflow, pmu);
 348                 WARN_ON(ret);
 349         }
 350 }
 351
 352 bool kvm_pmu_should_notify_user(struct kvm_vcpu *vcpu)
 353 {
 354         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 355         struct kvm_sync_regs *sregs = &vcpu->run->s.regs;
 356         bool run_level = sregs->device_irq_level & KVM_ARM_DEV_PMU;
 357
 358         if (likely(irqchip_in_kernel(vcpu->kvm)))
 359                 return false;
 360
 361         return pmu->irq_level != run_level;
 362 }
 363
 364 /*
 365  * Reflect the PMU overflow interrupt output level into the kvm_run structure
 366  */
 367 void kvm_pmu_update_run(struct kvm_vcpu *vcpu)
 368 {
 369         struct kvm_sync_regs *regs = &vcpu->run->s.regs;
 370
 371         /* Populate the timer bitmap for user space */
 372         regs->device_irq_level &= ~KVM_ARM_DEV_PMU;
 373         if (vcpu->arch.pmu.irq_level)
 374                 regs->device_irq_level |= KVM_ARM_DEV_PMU;
 375 }
 376
 377 /**
 378  * kvm_pmu_flush_hwstate - flush pmu state to cpu
 379  * @vcpu: The vcpu pointer
 380  *
 381  * Check if the PMU has overflowed while we were running in the host, and inject
 382  * an interrupt if that was the case.
 383  */
 384 void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu)
 385 {
 386         kvm_pmu_update_state(vcpu);
 387 }
 388
 389 /**
 390  * kvm_pmu_sync_hwstate - sync pmu state from cpu
 391  * @vcpu: The vcpu pointer
 392  *
 393  * Check if the PMU has overflowed while we were running in the guest, and
 394  * inject an interrupt if that was the case.
 395  */
 396 void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
 397 {
 398         kvm_pmu_update_state(vcpu);
 399 }
 400
 401 /**
 402  * When perf interrupt is an NMI, we cannot safely notify the vcpu corresponding
 403  * to the event.
 404  * This is why we need a callback to do it once outside of the NMI context.
 405  */
 406 static void kvm_pmu_perf_overflow_notify_vcpu(struct irq_work *work)
 407 {
 408         struct kvm_vcpu *vcpu;
 409
 410         vcpu = container_of(work, struct kvm_vcpu, arch.pmu.overflow_work);
 411         kvm_vcpu_kick(vcpu);
 412 }
 413
 414 /*
 415  * Perform an increment on any of the counters described in @mask,
 416  * generating the overflow if required, and propagate it as a chained
 417  * event if possible.
 418  */
 419 static void kvm_pmu_counter_increment(struct kvm_vcpu *vcpu,
 420                                       unsigned long mask, u32 event)
 421 {
 422         int i;
 423
 424         if (!(__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E))
 425                 return;
 426
 427         /* Weed out disabled counters */
 428         mask &= __vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
 429
 430         for_each_set_bit(i, &mask, ARMV8_PMU_CYCLE_IDX) {
 431                 struct kvm_pmc *pmc = kvm_vcpu_idx_to_pmc(vcpu, i);
 432                 u64 type, reg;
 433
 434                 /* Filter on event type */
 435                 type = __vcpu_sys_reg(vcpu, counter_index_to_evtreg(i));
 436                 type &= kvm_pmu_event_mask(vcpu->kvm);
 437                 if (type != event)
 438                         continue;
 439
 440                 /* Increment this counter */
 441                 reg = __vcpu_sys_reg(vcpu, counter_index_to_reg(i)) + 1;
 442                 if (!kvm_pmc_is_64bit(pmc))
 443                         reg = lower_32_bits(reg);
 444                 __vcpu_sys_reg(vcpu, counter_index_to_reg(i)) = reg;
 445
 446                 /* No overflow? move on */
 447                 if (kvm_pmc_has_64bit_overflow(pmc) ? reg : lower_32_bits(reg))
 448                         continue;
 449
 450                 /* Mark overflow */
 451                 __vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(i);
 452
 453                 if (kvm_pmu_counter_can_chain(pmc))
 454                         kvm_pmu_counter_increment(vcpu, BIT(i + 1),
 455                                                   ARMV8_PMUV3_PERFCTR_CHAIN);
 456         }
 457 }
 458
 459 /* Compute the sample period for a given counter value */
 460 static u64 compute_period(struct kvm_pmc *pmc, u64 counter)
 461 {
 462         u64 val;
 463
 464         if (kvm_pmc_is_64bit(pmc) && kvm_pmc_has_64bit_overflow(pmc))
 465                 val = (-counter) & GENMASK(63, 0);
 466         else
 467                 val = (-counter) & GENMASK(31, 0);
 468
 469         return val;
 470 }
 471
 472 /**
 473  * When the perf event overflows, set the overflow status and inform the vcpu.
 474  */
 475 static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
 476                                   struct perf_sample_data *data,
 477                                   struct pt_regs *regs)
 478 {
 479         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
 480         struct arm_pmu *cpu_pmu = to_arm_pmu(perf_event->pmu);
 481         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
 482         int idx = pmc->idx;
 483         u64 period;
 484
 485         cpu_pmu->pmu.stop(perf_event, PERF_EF_UPDATE);
 486
 487         /*
 488          * Reset the sample period to the architectural limit,
 489          * i.e. the point where the counter overflows.
 490          */
 491         period = compute_period(pmc, local64_read(&perf_event->count));
 492
 493         local64_set(&perf_event->hw.period_left, 0);
 494         perf_event->attr.sample_period = period;
 495         perf_event->hw.sample_period = period;
 496
 497         __vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= BIT(idx);
 498
 499         if (kvm_pmu_counter_can_chain(pmc))
 500                 kvm_pmu_counter_increment(vcpu, BIT(idx + 1),
 501                                           ARMV8_PMUV3_PERFCTR_CHAIN);
 502
 503         if (kvm_pmu_overflow_status(vcpu)) {
 504                 kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 505
 506                 if (!in_nmi())
 507                         kvm_vcpu_kick(vcpu);
 508                 else
 509                         irq_work_queue(&vcpu->arch.pmu.overflow_work);
 510         }
 511
 512         cpu_pmu->pmu.start(perf_event, PERF_EF_RELOAD);
 513 }
 514
 515 /**
 516  * kvm_pmu_software_increment - do software increment
 517  * @vcpu: The vcpu pointer
 518  * @val: the value guest writes to PMSWINC register
 519  */
 520 void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val)
 521 {
 522         kvm_pmu_counter_increment(vcpu, val, ARMV8_PMUV3_PERFCTR_SW_INCR);
 523 }
 524
 525 /**
 526  * kvm_pmu_handle_pmcr - handle PMCR register
 527  * @vcpu: The vcpu pointer
 528  * @val: the value guest writes to PMCR register
 529  */
 530 void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
 531 {
 532         int i;
 533
 534         if (!kvm_vcpu_has_pmu(vcpu))
 535                 return;
 536
 537         /* Fixup PMCR_EL0 to reconcile the PMU version and the LP bit */
 538         if (!kvm_pmu_is_3p5(vcpu))
 539                 val &= ~ARMV8_PMU_PMCR_LP;
 540
 541         /* The reset bits don't indicate any state, and shouldn't be saved. */
 542         __vcpu_sys_reg(vcpu, PMCR_EL0) = val & ~(ARMV8_PMU_PMCR_C | ARMV8_PMU_PMCR_P);
 543
 544         if (val & ARMV8_PMU_PMCR_E) {
 545                 kvm_pmu_enable_counter_mask(vcpu,
 546                        __vcpu_sys_reg(vcpu, PMCNTENSET_EL0));
 547         } else {
 548                 kvm_pmu_disable_counter_mask(vcpu,
 549                        __vcpu_sys_reg(vcpu, PMCNTENSET_EL0));
 550         }
 551
 552         if (val & ARMV8_PMU_PMCR_C)
 553                 kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
 554
 555         if (val & ARMV8_PMU_PMCR_P) {
 556                 unsigned long mask = kvm_pmu_valid_counter_mask(vcpu);
 557                 mask &= ~BIT(ARMV8_PMU_CYCLE_IDX);
 558                 for_each_set_bit(i, &mask, 32)
 559                         kvm_pmu_set_pmc_value(kvm_vcpu_idx_to_pmc(vcpu, i), 0, true);
 560         }
 561         kvm_vcpu_pmu_restore_guest(vcpu);
 562 }
 563
 564 static bool kvm_pmu_counter_is_enabled(struct kvm_pmc *pmc)
 565 {
 566         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
 567         return (__vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) &&
 568                (__vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & BIT(pmc->idx));
 569 }
 570
 571 /**
 572  * kvm_pmu_create_perf_event - create a perf event for a counter
 573  * @pmc: Counter context
 574  */
 575 static void kvm_pmu_create_perf_event(struct kvm_pmc *pmc)
 576 {
 577         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
 578         struct arm_pmu *arm_pmu = vcpu->kvm->arch.arm_pmu;
 579         struct perf_event *event;
 580         struct perf_event_attr attr;
 581         u64 eventsel, reg, data;
 582
 583         reg = counter_index_to_evtreg(pmc->idx);
 584         data = __vcpu_sys_reg(vcpu, reg);
 585
 586         kvm_pmu_stop_counter(pmc);
 587         if (pmc->idx == ARMV8_PMU_CYCLE_IDX)
 588                 eventsel = ARMV8_PMUV3_PERFCTR_CPU_CYCLES;
 589         else
 590                 eventsel = data & kvm_pmu_event_mask(vcpu->kvm);
 591
 592         /*
 593          * Neither SW increment nor chained events need to be backed
 594          * by a perf event.
 595          */
 596         if (eventsel == ARMV8_PMUV3_PERFCTR_SW_INCR ||
 597             eventsel == ARMV8_PMUV3_PERFCTR_CHAIN)
 598                 return;
 599
 600         /*
 601          * If we have a filter in place and that the event isn't allowed, do
 602          * not install a perf event either.
 603          */
 604         if (vcpu->kvm->arch.pmu_filter &&
 605             !test_bit(eventsel, vcpu->kvm->arch.pmu_filter))
 606                 return;
 607
 608         memset(&attr, 0, sizeof(struct perf_event_attr));
 609         attr.type = arm_pmu->pmu.type;
 610         attr.size = sizeof(attr);
 611         attr.pinned = 1;
 612         attr.disabled = !kvm_pmu_counter_is_enabled(pmc);
 613         attr.exclude_user = data & ARMV8_PMU_EXCLUDE_EL0 ? 1 : 0;
 614         attr.exclude_kernel = data & ARMV8_PMU_EXCLUDE_EL1 ? 1 : 0;
 615         attr.exclude_hv = 1; /* Don't count EL2 events */
 616         attr.exclude_host = 1; /* Don't count host events */
 617         attr.config = eventsel;
 618
 619         /*
 620          * If counting with a 64bit counter, advertise it to the perf
 621          * code, carefully dealing with the initial sample period
 622          * which also depends on the overflow.
 623          */
 624         if (kvm_pmc_is_64bit(pmc))
 625                 attr.config1 |= PERF_ATTR_CFG1_COUNTER_64BIT;
 626
 627         attr.sample_period = compute_period(pmc, kvm_pmu_get_pmc_value(pmc));
 628
 629         event = perf_event_create_kernel_counter(&attr, -1, current,
 630                                                  kvm_pmu_perf_overflow, pmc);
 631
 632         if (IS_ERR(event)) {
 633                 pr_err_once("kvm: pmu event creation failed %ld\n",
 634                             PTR_ERR(event));
 635                 return;
 636         }
 637
 638         pmc->perf_event = event;
 639 }
 640
 641 /**
 642  * kvm_pmu_set_counter_event_type - set selected counter to monitor some event
 643  * @vcpu: The vcpu pointer
 644  * @data: The data guest writes to PMXEVTYPER_EL0
 645  * @select_idx: The number of selected counter
 646  *
 647  * When OS accesses PMXEVTYPER_EL0, that means it wants to set a PMC to count an
 648  * event with given hardware event number. Here we call perf_event API to
 649  * emulate this action and create a kernel perf event for it.
 650  */
 651 void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data,
 652                                     u64 select_idx)
 653 {
 654         struct kvm_pmc *pmc = kvm_vcpu_idx_to_pmc(vcpu, select_idx);
 655         u64 reg, mask;
 656
 657         if (!kvm_vcpu_has_pmu(vcpu))
 658                 return;
 659
 660         mask  =  ARMV8_PMU_EVTYPE_MASK;
 661         mask &= ~ARMV8_PMU_EVTYPE_EVENT;
 662         mask |= kvm_pmu_event_mask(vcpu->kvm);
 663
 664         reg = counter_index_to_evtreg(pmc->idx);
 665
 666         __vcpu_sys_reg(vcpu, reg) = data & mask;
 667
 668         kvm_pmu_create_perf_event(pmc);
 669 }
 670
 671 void kvm_host_pmu_init(struct arm_pmu *pmu)
 672 {
 673         struct arm_pmu_entry *entry;
 674
 675         if (pmu->pmuver == ID_AA64DFR0_EL1_PMUVer_NI ||
 676             pmu->pmuver == ID_AA64DFR0_EL1_PMUVer_IMP_DEF)
 677                 return;
 678
 679         mutex_lock(&arm_pmus_lock);
 680
 681         entry = kmalloc(sizeof(*entry), GFP_KERNEL);
 682         if (!entry)
 683                 goto out_unlock;
 684
 685         entry->arm_pmu = pmu;
 686         list_add_tail(&entry->entry, &arm_pmus);
 687
 688         if (list_is_singular(&arm_pmus))
 689                 static_branch_enable(&kvm_arm_pmu_available);
 690
 691 out_unlock:
 692         mutex_unlock(&arm_pmus_lock);
 693 }
 694
 695 static struct arm_pmu *kvm_pmu_probe_armpmu(void)
 696 {
 697         struct perf_event_attr attr = { };
 698         struct perf_event *event;
 699         struct arm_pmu *pmu = NULL;
 700
 701         /*
 702          * Create a dummy event that only counts user cycles. As we'll never
 703          * leave this function with the event being live, it will never
 704          * count anything. But it allows us to probe some of the PMU
 705          * details. Yes, this is terrible.
 706          */
 707         attr.type = PERF_TYPE_RAW;
 708         attr.size = sizeof(attr);
 709         attr.pinned = 1;
 710         attr.disabled = 0;
 711         attr.exclude_user = 0;
 712         attr.exclude_kernel = 1;
 713         attr.exclude_hv = 1;
 714         attr.exclude_host = 1;
 715         attr.config = ARMV8_PMUV3_PERFCTR_CPU_CYCLES;
 716         attr.sample_period = GENMASK(63, 0);
 717
 718         event = perf_event_create_kernel_counter(&attr, -1, current,
 719                                                  kvm_pmu_perf_overflow, &attr);
 720
 721         if (IS_ERR(event)) {
 722                 pr_err_once("kvm: pmu event creation failed %ld\n",
 723                             PTR_ERR(event));
 724                 return NULL;
 725         }
 726
 727         if (event->pmu) {
 728                 pmu = to_arm_pmu(event->pmu);
 729                 if (pmu->pmuver == ID_AA64DFR0_EL1_PMUVer_NI ||
 730                     pmu->pmuver == ID_AA64DFR0_EL1_PMUVer_IMP_DEF)
 731                         pmu = NULL;
 732         }
 733
 734         perf_event_disable(event);
 735         perf_event_release_kernel(event);
 736
 737         return pmu;
 738 }
 739
 740 u64 kvm_pmu_get_pmceid(struct kvm_vcpu *vcpu, bool pmceid1)
 741 {
 742         unsigned long *bmap = vcpu->kvm->arch.pmu_filter;
 743         u64 val, mask = 0;
 744         int base, i, nr_events;
 745
 746         if (!kvm_vcpu_has_pmu(vcpu))
 747                 return 0;
 748
 749         if (!pmceid1) {
 750                 val = read_sysreg(pmceid0_el0);
 751                 /* always support CHAIN */
 752                 val |= BIT(ARMV8_PMUV3_PERFCTR_CHAIN);
 753                 base = 0;
 754         } else {
 755                 val = read_sysreg(pmceid1_el0);
 756                 /*
 757                  * Don't advertise STALL_SLOT, as PMMIR_EL0 is handled
 758                  * as RAZ
 759                  */
 760                 if (vcpu->kvm->arch.arm_pmu->pmuver >= ID_AA64DFR0_EL1_PMUVer_V3P4)
 761                         val &= ~BIT_ULL(ARMV8_PMUV3_PERFCTR_STALL_SLOT - 32);
 762                 base = 32;
 763         }
 764
 765         if (!bmap)
 766                 return val;
 767
 768         nr_events = kvm_pmu_event_mask(vcpu->kvm) + 1;
 769
 770         for (i = 0; i < 32; i += 8) {
 771                 u64 byte;
 772
 773                 byte = bitmap_get_value8(bmap, base + i);
 774                 mask |= byte << i;
 775                 if (nr_events >= (0x4000 + base + 32)) {
 776                         byte = bitmap_get_value8(bmap, 0x4000 + base + i);
 777                         mask |= byte << (32 + i);
 778                 }
 779         }
 780
 781         return val & mask;
 782 }
 783
 784 int kvm_arm_pmu_v3_enable(struct kvm_vcpu *vcpu)
 785 {
 786         if (!kvm_vcpu_has_pmu(vcpu))
 787                 return 0;
 788
 789         if (!vcpu->arch.pmu.created)
 790                 return -EINVAL;
 791
 792         /*
 793          * A valid interrupt configuration for the PMU is either to have a
 794          * properly configured interrupt number and using an in-kernel
 795          * irqchip, or to not have an in-kernel GIC and not set an IRQ.
 796          */
 797         if (irqchip_in_kernel(vcpu->kvm)) {
 798                 int irq = vcpu->arch.pmu.irq_num;
 799                 /*
 800                  * If we are using an in-kernel vgic, at this point we know
 801                  * the vgic will be initialized, so we can check the PMU irq
 802                  * number against the dimensions of the vgic and make sure
 803                  * it's valid.
 804                  */
 805                 if (!irq_is_ppi(irq) && !vgic_valid_spi(vcpu->kvm, irq))
 806                         return -EINVAL;
 807         } else if (kvm_arm_pmu_irq_initialized(vcpu)) {
 808                    return -EINVAL;
 809         }
 810
 811         /* One-off reload of the PMU on first run */
 812         kvm_make_request(KVM_REQ_RELOAD_PMU, vcpu);
 813
 814         return 0;
 815 }
 816
 817 static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
 818 {
 819         if (irqchip_in_kernel(vcpu->kvm)) {
 820                 int ret;
 821
 822                 /*
 823                  * If using the PMU with an in-kernel virtual GIC
 824                  * implementation, we require the GIC to be already
 825                  * initialized when initializing the PMU.
 826                  */
 827                 if (!vgic_initialized(vcpu->kvm))
 828                         return -ENODEV;
 829
 830                 if (!kvm_arm_pmu_irq_initialized(vcpu))
 831                         return -ENXIO;
 832
 833                 ret = kvm_vgic_set_owner(vcpu, vcpu->arch.pmu.irq_num,
 834                                          &vcpu->arch.pmu);
 835                 if (ret)
 836                         return ret;
 837         }
 838
 839         init_irq_work(&vcpu->arch.pmu.overflow_work,
 840                       kvm_pmu_perf_overflow_notify_vcpu);
 841
 842         vcpu->arch.pmu.created = true;
 843         return 0;
 844 }
 845
 846 /*
 847  * For one VM the interrupt type must be same for each vcpu.
 848  * As a PPI, the interrupt number is the same for all vcpus,
 849  * while as an SPI it must be a separate number per vcpu.
 850  */
 851 static bool pmu_irq_is_valid(struct kvm *kvm, int irq)
 852 {
 853         unsigned long i;
 854         struct kvm_vcpu *vcpu;
 855
 856         kvm_for_each_vcpu(i, vcpu, kvm) {
 857                 if (!kvm_arm_pmu_irq_initialized(vcpu))
 858                         continue;
 859
 860                 if (irq_is_ppi(irq)) {
 861                         if (vcpu->arch.pmu.irq_num != irq)
 862                                 return false;
 863                 } else {
 864                         if (vcpu->arch.pmu.irq_num == irq)
 865                                 return false;
 866                 }
 867         }
 868
 869         return true;
 870 }
 871
 872 static int kvm_arm_pmu_v3_set_pmu(struct kvm_vcpu *vcpu, int pmu_id)
 873 {
 874         struct kvm *kvm = vcpu->kvm;
 875         struct arm_pmu_entry *entry;
 876         struct arm_pmu *arm_pmu;
 877         int ret = -ENXIO;
 878
 879         lockdep_assert_held(&kvm->arch.config_lock);
 880         mutex_lock(&arm_pmus_lock);
 881
 882         list_for_each_entry(entry, &arm_pmus, entry) {
 883                 arm_pmu = entry->arm_pmu;
 884                 if (arm_pmu->pmu.type == pmu_id) {
 885                         if (kvm_vm_has_ran_once(kvm) ||
 886                             (kvm->arch.pmu_filter && kvm->arch.arm_pmu != arm_pmu)) {
 887                                 ret = -EBUSY;
 888                                 break;
 889                         }
 890
 891                         kvm->arch.arm_pmu = arm_pmu;
 892                         cpumask_copy(kvm->arch.supported_cpus, &arm_pmu->supported_cpus);
 893                         ret = 0;
 894                         break;
 895                 }
 896         }
 897
 898         mutex_unlock(&arm_pmus_lock);
 899         return ret;
 900 }
 901
 902 int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 903 {
 904         struct kvm *kvm = vcpu->kvm;
 905
 906         lockdep_assert_held(&kvm->arch.config_lock);
 907
 908         if (!kvm_vcpu_has_pmu(vcpu))
 909                 return -ENODEV;
 910
 911         if (vcpu->arch.pmu.created)
 912                 return -EBUSY;
 913
 914         if (!kvm->arch.arm_pmu) {
 915                 /* No PMU set, get the default one */
 916                 kvm->arch.arm_pmu = kvm_pmu_probe_armpmu();
 917                 if (!kvm->arch.arm_pmu)
 918                         return -ENODEV;
 919         }
 920
 921         switch (attr->attr) {
 922         case KVM_ARM_VCPU_PMU_V3_IRQ: {
 923                 int __user *uaddr = (int __user *)(long)attr->addr;
 924                 int irq;
 925
 926                 if (!irqchip_in_kernel(kvm))
 927                         return -EINVAL;
 928
 929                 if (get_user(irq, uaddr))
 930                         return -EFAULT;
 931
 932                 /* The PMU overflow interrupt can be a PPI or a valid SPI. */
 933                 if (!(irq_is_ppi(irq) || irq_is_spi(irq)))
 934                         return -EINVAL;
 935
 936                 if (!pmu_irq_is_valid(kvm, irq))
 937                         return -EINVAL;
 938
 939                 if (kvm_arm_pmu_irq_initialized(vcpu))
 940                         return -EBUSY;
 941
 942                 kvm_debug("Set kvm ARM PMU irq: %d\n", irq);
 943                 vcpu->arch.pmu.irq_num = irq;
 944                 return 0;
 945         }
 946         case KVM_ARM_VCPU_PMU_V3_FILTER: {
 947                 struct kvm_pmu_event_filter __user *uaddr;
 948                 struct kvm_pmu_event_filter filter;
 949                 int nr_events;
 950
 951                 nr_events = kvm_pmu_event_mask(kvm) + 1;
 952
 953                 uaddr = (struct kvm_pmu_event_filter __user *)(long)attr->addr;
 954
 955                 if (copy_from_user(&filter, uaddr, sizeof(filter)))
 956                         return -EFAULT;
 957
 958                 if (((u32)filter.base_event + filter.nevents) > nr_events ||
 959                     (filter.action != KVM_PMU_EVENT_ALLOW &&
 960                      filter.action != KVM_PMU_EVENT_DENY))
 961                         return -EINVAL;
 962
 963                 if (kvm_vm_has_ran_once(kvm))
 964                         return -EBUSY;
 965
 966                 if (!kvm->arch.pmu_filter) {
 967                         kvm->arch.pmu_filter = bitmap_alloc(nr_events, GFP_KERNEL_ACCOUNT);
 968                         if (!kvm->arch.pmu_filter)
 969                                 return -ENOMEM;
 970
 971                         /*
 972                          * The default depends on the first applied filter.
 973                          * If it allows events, the default is to deny.
 974                          * Conversely, if the first filter denies a set of
 975                          * events, the default is to allow.
 976                          */
 977                         if (filter.action == KVM_PMU_EVENT_ALLOW)
 978                                 bitmap_zero(kvm->arch.pmu_filter, nr_events);
 979                         else
 980                                 bitmap_fill(kvm->arch.pmu_filter, nr_events);
 981                 }
 982
 983                 if (filter.action == KVM_PMU_EVENT_ALLOW)
 984                         bitmap_set(kvm->arch.pmu_filter, filter.base_event, filter.nevents);
 985                 else
 986                         bitmap_clear(kvm->arch.pmu_filter, filter.base_event, filter.nevents);
 987
 988                 return 0;
 989         }
 990         case KVM_ARM_VCPU_PMU_V3_SET_PMU: {
 991                 int __user *uaddr = (int __user *)(long)attr->addr;
 992                 int pmu_id;
 993
 994                 if (get_user(pmu_id, uaddr))
 995                         return -EFAULT;
 996
 997                 return kvm_arm_pmu_v3_set_pmu(vcpu, pmu_id);
 998         }
 999         case KVM_ARM_VCPU_PMU_V3_INIT:
1000                 return kvm_arm_pmu_v3_init(vcpu);
1001         }
1002
1003         return -ENXIO;
1004 }
1005
1006 int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1007 {
1008         switch (attr->attr) {
1009         case KVM_ARM_VCPU_PMU_V3_IRQ: {
1010                 int __user *uaddr = (int __user *)(long)attr->addr;
1011                 int irq;
1012
1013                 if (!irqchip_in_kernel(vcpu->kvm))
1014                         return -EINVAL;
1015
1016                 if (!kvm_vcpu_has_pmu(vcpu))
1017                         return -ENODEV;
1018
1019                 if (!kvm_arm_pmu_irq_initialized(vcpu))
1020                         return -ENXIO;
1021
1022                 irq = vcpu->arch.pmu.irq_num;
1023                 return put_user(irq, uaddr);
1024         }
1025         }
1026
1027         return -ENXIO;
1028 }
1029
1030 int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
1031 {
1032         switch (attr->attr) {
1033         case KVM_ARM_VCPU_PMU_V3_IRQ:
1034         case KVM_ARM_VCPU_PMU_V3_INIT:
1035         case KVM_ARM_VCPU_PMU_V3_FILTER:
1036         case KVM_ARM_VCPU_PMU_V3_SET_PMU:
1037                 if (kvm_vcpu_has_pmu(vcpu))
1038                         return 0;
1039         }
1040
1041         return -ENXIO;
1042 }
1043
1044 u8 kvm_arm_pmu_get_pmuver_limit(void)
1045 {
1046         u64 tmp;
1047
1048         tmp = read_sanitised_ftr_reg(SYS_ID_AA64DFR0_EL1);
1049         tmp = cpuid_feature_cap_perfmon_field(tmp,
1050                                               ID_AA64DFR0_EL1_PMUVer_SHIFT,
1051                                               ID_AA64DFR0_EL1_PMUVer_V3P5);
1052         return FIELD_GET(ARM64_FEATURE_MASK(ID_AA64DFR0_EL1_PMUVer), tmp);
1053 }