virt/kvm/arm/pmu.c

   1 /*
   2  * Copyright (C) 2015 Linaro Ltd.
   3  * Author: Shannon Zhao <shannon.zhao@linaro.org>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License version 2 as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program.  If not, see <http://www.gnu.org/licenses/>.
  16  */
  17
  18 #include <linux/cpu.h>
  19 #include <linux/kvm.h>
  20 #include <linux/kvm_host.h>
  21 #include <linux/perf_event.h>
  22 #include <linux/uaccess.h>
  23 #include <asm/kvm_emulate.h>
  24 #include <kvm/arm_pmu.h>
  25 #include <kvm/arm_vgic.h>
  26
  27 /**
  28  * kvm_pmu_get_counter_value - get PMU counter value
  29  * @vcpu: The vcpu pointer
  30  * @select_idx: The counter index
  31  */
  32 u64 kvm_pmu_get_counter_value(struct kvm_vcpu *vcpu, u64 select_idx)
  33 {
  34         u64 counter, reg, enabled, running;
  35         struct kvm_pmu *pmu = &vcpu->arch.pmu;
  36         struct kvm_pmc *pmc = &pmu->pmc[select_idx];
  37
  38         reg = (select_idx == ARMV8_PMU_CYCLE_IDX)
  39               ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + select_idx;
  40         counter = vcpu_sys_reg(vcpu, reg);
  41
  42         /* The real counter value is equal to the value of counter register plus
  43          * the value perf event counts.
  44          */
  45         if (pmc->perf_event)
  46                 counter += perf_event_read_value(pmc->perf_event, &enabled,
  47                                                  &running);
  48
  49         return counter & pmc->bitmask;
  50 }
  51
  52 /**
  53  * kvm_pmu_set_counter_value - set PMU counter value
  54  * @vcpu: The vcpu pointer
  55  * @select_idx: The counter index
  56  * @val: The counter value
  57  */
  58 void kvm_pmu_set_counter_value(struct kvm_vcpu *vcpu, u64 select_idx, u64 val)
  59 {
  60         u64 reg;
  61
  62         reg = (select_idx == ARMV8_PMU_CYCLE_IDX)
  63               ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + select_idx;
  64         vcpu_sys_reg(vcpu, reg) += (s64)val - kvm_pmu_get_counter_value(vcpu, select_idx);
  65 }
  66
  67 /**
  68  * kvm_pmu_stop_counter - stop PMU counter
  69  * @pmc: The PMU counter pointer
  70  *
  71  * If this counter has been configured to monitor some event, release it here.
  72  */
  73 static void kvm_pmu_stop_counter(struct kvm_vcpu *vcpu, struct kvm_pmc *pmc)
  74 {
  75         u64 counter, reg;
  76
  77         if (pmc->perf_event) {
  78                 counter = kvm_pmu_get_counter_value(vcpu, pmc->idx);
  79                 reg = (pmc->idx == ARMV8_PMU_CYCLE_IDX)
  80                        ? PMCCNTR_EL0 : PMEVCNTR0_EL0 + pmc->idx;
  81                 vcpu_sys_reg(vcpu, reg) = counter;
  82                 perf_event_disable(pmc->perf_event);
  83                 perf_event_release_kernel(pmc->perf_event);
  84                 pmc->perf_event = NULL;
  85         }
  86 }
  87
  88 /**
  89  * kvm_pmu_vcpu_reset - reset pmu state for cpu
  90  * @vcpu: The vcpu pointer
  91  *
  92  */
  93 void kvm_pmu_vcpu_reset(struct kvm_vcpu *vcpu)
  94 {
  95         int i;
  96         struct kvm_pmu *pmu = &vcpu->arch.pmu;
  97
  98         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
  99                 kvm_pmu_stop_counter(vcpu, &pmu->pmc[i]);
 100                 pmu->pmc[i].idx = i;
 101                 pmu->pmc[i].bitmask = 0xffffffffUL;
 102         }
 103 }
 104
 105 /**
 106  * kvm_pmu_vcpu_destroy - free perf event of PMU for cpu
 107  * @vcpu: The vcpu pointer
 108  *
 109  */
 110 void kvm_pmu_vcpu_destroy(struct kvm_vcpu *vcpu)
 111 {
 112         int i;
 113         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 114
 115         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
 116                 struct kvm_pmc *pmc = &pmu->pmc[i];
 117
 118                 if (pmc->perf_event) {
 119                         perf_event_disable(pmc->perf_event);
 120                         perf_event_release_kernel(pmc->perf_event);
 121                         pmc->perf_event = NULL;
 122                 }
 123         }
 124 }
 125
 126 u64 kvm_pmu_valid_counter_mask(struct kvm_vcpu *vcpu)
 127 {
 128         u64 val = vcpu_sys_reg(vcpu, PMCR_EL0) >> ARMV8_PMU_PMCR_N_SHIFT;
 129
 130         val &= ARMV8_PMU_PMCR_N_MASK;
 131         if (val == 0)
 132                 return BIT(ARMV8_PMU_CYCLE_IDX);
 133         else
 134                 return GENMASK(val - 1, 0) | BIT(ARMV8_PMU_CYCLE_IDX);
 135 }
 136
 137 /**
 138  * kvm_pmu_enable_counter - enable selected PMU counter
 139  * @vcpu: The vcpu pointer
 140  * @val: the value guest writes to PMCNTENSET register
 141  *
 142  * Call perf_event_enable to start counting the perf event
 143  */
 144 void kvm_pmu_enable_counter(struct kvm_vcpu *vcpu, u64 val)
 145 {
 146         int i;
 147         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 148         struct kvm_pmc *pmc;
 149
 150         if (!(vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) || !val)
 151                 return;
 152
 153         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
 154                 if (!(val & BIT(i)))
 155                         continue;
 156
 157                 pmc = &pmu->pmc[i];
 158                 if (pmc->perf_event) {
 159                         perf_event_enable(pmc->perf_event);
 160                         if (pmc->perf_event->state != PERF_EVENT_STATE_ACTIVE)
 161                                 kvm_debug("fail to enable perf event\n");
 162                 }
 163         }
 164 }
 165
 166 /**
 167  * kvm_pmu_disable_counter - disable selected PMU counter
 168  * @vcpu: The vcpu pointer
 169  * @val: the value guest writes to PMCNTENCLR register
 170  *
 171  * Call perf_event_disable to stop counting the perf event
 172  */
 173 void kvm_pmu_disable_counter(struct kvm_vcpu *vcpu, u64 val)
 174 {
 175         int i;
 176         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 177         struct kvm_pmc *pmc;
 178
 179         if (!val)
 180                 return;
 181
 182         for (i = 0; i < ARMV8_PMU_MAX_COUNTERS; i++) {
 183                 if (!(val & BIT(i)))
 184                         continue;
 185
 186                 pmc = &pmu->pmc[i];
 187                 if (pmc->perf_event)
 188                         perf_event_disable(pmc->perf_event);
 189         }
 190 }
 191
 192 static u64 kvm_pmu_overflow_status(struct kvm_vcpu *vcpu)
 193 {
 194         u64 reg = 0;
 195
 196         if ((vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E)) {
 197                 reg = vcpu_sys_reg(vcpu, PMOVSSET_EL0);
 198                 reg &= vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
 199                 reg &= vcpu_sys_reg(vcpu, PMINTENSET_EL1);
 200                 reg &= kvm_pmu_valid_counter_mask(vcpu);
 201         }
 202
 203         return reg;
 204 }
 205
 206 /**
 207  * kvm_pmu_overflow_set - set PMU overflow interrupt
 208  * @vcpu: The vcpu pointer
 209  * @val: the value guest writes to PMOVSSET register
 210  */
 211 void kvm_pmu_overflow_set(struct kvm_vcpu *vcpu, u64 val)
 212 {
 213         u64 reg;
 214
 215         if (val == 0)
 216                 return;
 217
 218         vcpu_sys_reg(vcpu, PMOVSSET_EL0) |= val;
 219         reg = kvm_pmu_overflow_status(vcpu);
 220         if (reg != 0)
 221                 kvm_vcpu_kick(vcpu);
 222 }
 223
 224 static void kvm_pmu_update_state(struct kvm_vcpu *vcpu)
 225 {
 226         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 227         bool overflow;
 228
 229         if (!kvm_arm_pmu_v3_ready(vcpu))
 230                 return;
 231
 232         overflow = !!kvm_pmu_overflow_status(vcpu);
 233         if (pmu->irq_level != overflow) {
 234                 pmu->irq_level = overflow;
 235                 kvm_vgic_inject_irq(vcpu->kvm, vcpu->vcpu_id,
 236                                     pmu->irq_num, overflow);
 237         }
 238 }
 239
 240 /**
 241  * kvm_pmu_flush_hwstate - flush pmu state to cpu
 242  * @vcpu: The vcpu pointer
 243  *
 244  * Check if the PMU has overflowed while we were running in the host, and inject
 245  * an interrupt if that was the case.
 246  */
 247 void kvm_pmu_flush_hwstate(struct kvm_vcpu *vcpu)
 248 {
 249         kvm_pmu_update_state(vcpu);
 250 }
 251
 252 /**
 253  * kvm_pmu_sync_hwstate - sync pmu state from cpu
 254  * @vcpu: The vcpu pointer
 255  *
 256  * Check if the PMU has overflowed while we were running in the guest, and
 257  * inject an interrupt if that was the case.
 258  */
 259 void kvm_pmu_sync_hwstate(struct kvm_vcpu *vcpu)
 260 {
 261         kvm_pmu_update_state(vcpu);
 262 }
 263
 264 static inline struct kvm_vcpu *kvm_pmc_to_vcpu(struct kvm_pmc *pmc)
 265 {
 266         struct kvm_pmu *pmu;
 267         struct kvm_vcpu_arch *vcpu_arch;
 268
 269         pmc -= pmc->idx;
 270         pmu = container_of(pmc, struct kvm_pmu, pmc[0]);
 271         vcpu_arch = container_of(pmu, struct kvm_vcpu_arch, pmu);
 272         return container_of(vcpu_arch, struct kvm_vcpu, arch);
 273 }
 274
 275 /**
 276  * When perf event overflows, call kvm_pmu_overflow_set to set overflow status.
 277  */
 278 static void kvm_pmu_perf_overflow(struct perf_event *perf_event,
 279                                   struct perf_sample_data *data,
 280                                   struct pt_regs *regs)
 281 {
 282         struct kvm_pmc *pmc = perf_event->overflow_handler_context;
 283         struct kvm_vcpu *vcpu = kvm_pmc_to_vcpu(pmc);
 284         int idx = pmc->idx;
 285
 286         kvm_pmu_overflow_set(vcpu, BIT(idx));
 287 }
 288
 289 /**
 290  * kvm_pmu_software_increment - do software increment
 291  * @vcpu: The vcpu pointer
 292  * @val: the value guest writes to PMSWINC register
 293  */
 294 void kvm_pmu_software_increment(struct kvm_vcpu *vcpu, u64 val)
 295 {
 296         int i;
 297         u64 type, enable, reg;
 298
 299         if (val == 0)
 300                 return;
 301
 302         enable = vcpu_sys_reg(vcpu, PMCNTENSET_EL0);
 303         for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++) {
 304                 if (!(val & BIT(i)))
 305                         continue;
 306                 type = vcpu_sys_reg(vcpu, PMEVTYPER0_EL0 + i)
 307                        & ARMV8_PMU_EVTYPE_EVENT;
 308                 if ((type == ARMV8_PMU_EVTYPE_EVENT_SW_INCR)
 309                     && (enable & BIT(i))) {
 310                         reg = vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) + 1;
 311                         reg = lower_32_bits(reg);
 312                         vcpu_sys_reg(vcpu, PMEVCNTR0_EL0 + i) = reg;
 313                         if (!reg)
 314                                 kvm_pmu_overflow_set(vcpu, BIT(i));
 315                 }
 316         }
 317 }
 318
 319 /**
 320  * kvm_pmu_handle_pmcr - handle PMCR register
 321  * @vcpu: The vcpu pointer
 322  * @val: the value guest writes to PMCR register
 323  */
 324 void kvm_pmu_handle_pmcr(struct kvm_vcpu *vcpu, u64 val)
 325 {
 326         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 327         struct kvm_pmc *pmc;
 328         u64 mask;
 329         int i;
 330
 331         mask = kvm_pmu_valid_counter_mask(vcpu);
 332         if (val & ARMV8_PMU_PMCR_E) {
 333                 kvm_pmu_enable_counter(vcpu,
 334                                 vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & mask);
 335         } else {
 336                 kvm_pmu_disable_counter(vcpu, mask);
 337         }
 338
 339         if (val & ARMV8_PMU_PMCR_C)
 340                 kvm_pmu_set_counter_value(vcpu, ARMV8_PMU_CYCLE_IDX, 0);
 341
 342         if (val & ARMV8_PMU_PMCR_P) {
 343                 for (i = 0; i < ARMV8_PMU_CYCLE_IDX; i++)
 344                         kvm_pmu_set_counter_value(vcpu, i, 0);
 345         }
 346
 347         if (val & ARMV8_PMU_PMCR_LC) {
 348                 pmc = &pmu->pmc[ARMV8_PMU_CYCLE_IDX];
 349                 pmc->bitmask = 0xffffffffffffffffUL;
 350         }
 351 }
 352
 353 static bool kvm_pmu_counter_is_enabled(struct kvm_vcpu *vcpu, u64 select_idx)
 354 {
 355         return (vcpu_sys_reg(vcpu, PMCR_EL0) & ARMV8_PMU_PMCR_E) &&
 356                (vcpu_sys_reg(vcpu, PMCNTENSET_EL0) & BIT(select_idx));
 357 }
 358
 359 /**
 360  * kvm_pmu_set_counter_event_type - set selected counter to monitor some event
 361  * @vcpu: The vcpu pointer
 362  * @data: The data guest writes to PMXEVTYPER_EL0
 363  * @select_idx: The number of selected counter
 364  *
 365  * When OS accesses PMXEVTYPER_EL0, that means it wants to set a PMC to count an
 366  * event with given hardware event number. Here we call perf_event API to
 367  * emulate this action and create a kernel perf event for it.
 368  */
 369 void kvm_pmu_set_counter_event_type(struct kvm_vcpu *vcpu, u64 data,
 370                                     u64 select_idx)
 371 {
 372         struct kvm_pmu *pmu = &vcpu->arch.pmu;
 373         struct kvm_pmc *pmc = &pmu->pmc[select_idx];
 374         struct perf_event *event;
 375         struct perf_event_attr attr;
 376         u64 eventsel, counter;
 377
 378         kvm_pmu_stop_counter(vcpu, pmc);
 379         eventsel = data & ARMV8_PMU_EVTYPE_EVENT;
 380
 381         /* Software increment event does't need to be backed by a perf event */
 382         if (eventsel == ARMV8_PMU_EVTYPE_EVENT_SW_INCR)
 383                 return;
 384
 385         memset(&attr, 0, sizeof(struct perf_event_attr));
 386         attr.type = PERF_TYPE_RAW;
 387         attr.size = sizeof(attr);
 388         attr.pinned = 1;
 389         attr.disabled = !kvm_pmu_counter_is_enabled(vcpu, select_idx);
 390         attr.exclude_user = data & ARMV8_PMU_EXCLUDE_EL0 ? 1 : 0;
 391         attr.exclude_kernel = data & ARMV8_PMU_EXCLUDE_EL1 ? 1 : 0;
 392         attr.exclude_hv = 1; /* Don't count EL2 events */
 393         attr.exclude_host = 1; /* Don't count host events */
 394         attr.config = eventsel;
 395
 396         counter = kvm_pmu_get_counter_value(vcpu, select_idx);
 397         /* The initial sample period (overflow count) of an event. */
 398         attr.sample_period = (-counter) & pmc->bitmask;
 399
 400         event = perf_event_create_kernel_counter(&attr, -1, current,
 401                                                  kvm_pmu_perf_overflow, pmc);
 402         if (IS_ERR(event)) {
 403                 pr_err_once("kvm: pmu event creation failed %ld\n",
 404                             PTR_ERR(event));
 405                 return;
 406         }
 407
 408         pmc->perf_event = event;
 409 }
 410
 411 bool kvm_arm_support_pmu_v3(void)
 412 {
 413         /*
 414          * Check if HW_PERF_EVENTS are supported by checking the number of
 415          * hardware performance counters. This could ensure the presence of
 416          * a physical PMU and CONFIG_PERF_EVENT is selected.
 417          */
 418         return (perf_num_counters() > 0);
 419 }
 420
 421 static int kvm_arm_pmu_v3_init(struct kvm_vcpu *vcpu)
 422 {
 423         if (!kvm_arm_support_pmu_v3())
 424                 return -ENODEV;
 425
 426         /*
 427          * We currently require an in-kernel VGIC to use the PMU emulation,
 428          * because we do not support forwarding PMU overflow interrupts to
 429          * userspace yet.
 430          */
 431         if (!irqchip_in_kernel(vcpu->kvm) || !vgic_initialized(vcpu->kvm))
 432                 return -ENODEV;
 433
 434         if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features) ||
 435             !kvm_arm_pmu_irq_initialized(vcpu))
 436                 return -ENXIO;
 437
 438         if (kvm_arm_pmu_v3_ready(vcpu))
 439                 return -EBUSY;
 440
 441         kvm_pmu_vcpu_reset(vcpu);
 442         vcpu->arch.pmu.ready = true;
 443
 444         return 0;
 445 }
 446
 447 #define irq_is_ppi(irq) ((irq) >= VGIC_NR_SGIS && (irq) < VGIC_NR_PRIVATE_IRQS)
 448
 449 /*
 450  * For one VM the interrupt type must be same for each vcpu.
 451  * As a PPI, the interrupt number is the same for all vcpus,
 452  * while as an SPI it must be a separate number per vcpu.
 453  */
 454 static bool pmu_irq_is_valid(struct kvm *kvm, int irq)
 455 {
 456         int i;
 457         struct kvm_vcpu *vcpu;
 458
 459         kvm_for_each_vcpu(i, vcpu, kvm) {
 460                 if (!kvm_arm_pmu_irq_initialized(vcpu))
 461                         continue;
 462
 463                 if (irq_is_ppi(irq)) {
 464                         if (vcpu->arch.pmu.irq_num != irq)
 465                                 return false;
 466                 } else {
 467                         if (vcpu->arch.pmu.irq_num == irq)
 468                                 return false;
 469                 }
 470         }
 471
 472         return true;
 473 }
 474
 475 int kvm_arm_pmu_v3_set_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 476 {
 477         switch (attr->attr) {
 478         case KVM_ARM_VCPU_PMU_V3_IRQ: {
 479                 int __user *uaddr = (int __user *)(long)attr->addr;
 480                 int irq;
 481
 482                 if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
 483                         return -ENODEV;
 484
 485                 if (get_user(irq, uaddr))
 486                         return -EFAULT;
 487
 488                 /* The PMU overflow interrupt can be a PPI or a valid SPI. */
 489                 if (!(irq_is_ppi(irq) || vgic_valid_spi(vcpu->kvm, irq)))
 490                         return -EINVAL;
 491
 492                 if (!pmu_irq_is_valid(vcpu->kvm, irq))
 493                         return -EINVAL;
 494
 495                 if (kvm_arm_pmu_irq_initialized(vcpu))
 496                         return -EBUSY;
 497
 498                 kvm_debug("Set kvm ARM PMU irq: %d\n", irq);
 499                 vcpu->arch.pmu.irq_num = irq;
 500                 return 0;
 501         }
 502         case KVM_ARM_VCPU_PMU_V3_INIT:
 503                 return kvm_arm_pmu_v3_init(vcpu);
 504         }
 505
 506         return -ENXIO;
 507 }
 508
 509 int kvm_arm_pmu_v3_get_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 510 {
 511         switch (attr->attr) {
 512         case KVM_ARM_VCPU_PMU_V3_IRQ: {
 513                 int __user *uaddr = (int __user *)(long)attr->addr;
 514                 int irq;
 515
 516                 if (!test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
 517                         return -ENODEV;
 518
 519                 if (!kvm_arm_pmu_irq_initialized(vcpu))
 520                         return -ENXIO;
 521
 522                 irq = vcpu->arch.pmu.irq_num;
 523                 return put_user(irq, uaddr);
 524         }
 525         }
 526
 527         return -ENXIO;
 528 }
 529
 530 int kvm_arm_pmu_v3_has_attr(struct kvm_vcpu *vcpu, struct kvm_device_attr *attr)
 531 {
 532         switch (attr->attr) {
 533         case KVM_ARM_VCPU_PMU_V3_IRQ:
 534         case KVM_ARM_VCPU_PMU_V3_INIT:
 535                 if (kvm_arm_support_pmu_v3() &&
 536                     test_bit(KVM_ARM_VCPU_PMU_V3, vcpu->arch.features))
 537                         return 0;
 538         }
 539
 540         return -ENXIO;
 541 }