arch/arm64/kvm/arm.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   4  * Author: Christoffer Dall <c.dall@virtualopensystems.com>
   5  */
   6
   7 #include <linux/bug.h>
   8 #include <linux/cpu_pm.h>
   9 #include <linux/entry-kvm.h>
  10 #include <linux/errno.h>
  11 #include <linux/err.h>
  12 #include <linux/kvm_host.h>
  13 #include <linux/list.h>
  14 #include <linux/module.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mman.h>
  18 #include <linux/sched.h>
  19 #include <linux/kvm.h>
  20 #include <linux/kvm_irqfd.h>
  21 #include <linux/irqbypass.h>
  22 #include <linux/sched/stat.h>
  23 #include <linux/psci.h>
  24 #include <trace/events/kvm.h>
  25
  26 #define CREATE_TRACE_POINTS
  27 #include "trace_arm.h"
  28
  29 #include <linux/uaccess.h>
  30 #include <asm/ptrace.h>
  31 #include <asm/mman.h>
  32 #include <asm/tlbflush.h>
  33 #include <asm/cacheflush.h>
  34 #include <asm/cpufeature.h>
  35 #include <asm/virt.h>
  36 #include <asm/kvm_arm.h>
  37 #include <asm/kvm_asm.h>
  38 #include <asm/kvm_mmu.h>
  39 #include <asm/kvm_nested.h>
  40 #include <asm/kvm_pkvm.h>
  41 #include <asm/kvm_emulate.h>
  42 #include <asm/sections.h>
  43
  44 #include <kvm/arm_hypercalls.h>
  45 #include <kvm/arm_pmu.h>
  46 #include <kvm/arm_psci.h>
  47
  48 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
  49
  50 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
  51
  52 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  53 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
  54
  55 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt);
  56
  57 static bool vgic_present, kvm_arm_initialised;
  58
  59 static DEFINE_PER_CPU(unsigned char, kvm_hyp_initialized);
  60 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  61
  62 bool is_kvm_arm_initialised(void)
  63 {
  64         return kvm_arm_initialised;
  65 }
  66
  67 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  68 {
  69         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  70 }
  71
  72 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  73                             struct kvm_enable_cap *cap)
  74 {
  75         int r;
  76         u64 new_cap;
  77
  78         if (cap->flags)
  79                 return -EINVAL;
  80
  81         switch (cap->cap) {
  82         case KVM_CAP_ARM_NISV_TO_USER:
  83                 r = 0;
  84                 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER,
  85                         &kvm->arch.flags);
  86                 break;
  87         case KVM_CAP_ARM_MTE:
  88                 mutex_lock(&kvm->lock);
  89                 if (!system_supports_mte() || kvm->created_vcpus) {
  90                         r = -EINVAL;
  91                 } else {
  92                         r = 0;
  93                         set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags);
  94                 }
  95                 mutex_unlock(&kvm->lock);
  96                 break;
  97         case KVM_CAP_ARM_SYSTEM_SUSPEND:
  98                 r = 0;
  99                 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags);
 100                 break;
 101         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 102                 new_cap = cap->args[0];
 103
 104                 mutex_lock(&kvm->slots_lock);
 105                 /*
 106                  * To keep things simple, allow changing the chunk
 107                  * size only when no memory slots have been created.
 108                  */
 109                 if (!kvm_are_all_memslots_empty(kvm)) {
 110                         r = -EINVAL;
 111                 } else if (new_cap && !kvm_is_block_size_supported(new_cap)) {
 112                         r = -EINVAL;
 113                 } else {
 114                         r = 0;
 115                         kvm->arch.mmu.split_page_chunk_size = new_cap;
 116                 }
 117                 mutex_unlock(&kvm->slots_lock);
 118                 break;
 119         default:
 120                 r = -EINVAL;
 121                 break;
 122         }
 123
 124         return r;
 125 }
 126
 127 static int kvm_arm_default_max_vcpus(void)
 128 {
 129         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 130 }
 131
 132 /**
 133  * kvm_arch_init_vm - initializes a VM data structure
 134  * @kvm:        pointer to the KVM struct
 135  */
 136 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 137 {
 138         int ret;
 139
 140         mutex_init(&kvm->arch.config_lock);
 141
 142 #ifdef CONFIG_LOCKDEP
 143         /* Clue in lockdep that the config_lock must be taken inside kvm->lock */
 144         mutex_lock(&kvm->lock);
 145         mutex_lock(&kvm->arch.config_lock);
 146         mutex_unlock(&kvm->arch.config_lock);
 147         mutex_unlock(&kvm->lock);
 148 #endif
 149
 150         ret = kvm_share_hyp(kvm, kvm + 1);
 151         if (ret)
 152                 return ret;
 153
 154         ret = pkvm_init_host_vm(kvm);
 155         if (ret)
 156                 goto err_unshare_kvm;
 157
 158         if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) {
 159                 ret = -ENOMEM;
 160                 goto err_unshare_kvm;
 161         }
 162         cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask);
 163
 164         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type);
 165         if (ret)
 166                 goto err_free_cpumask;
 167
 168         kvm_vgic_early_init(kvm);
 169
 170         kvm_timer_init_vm(kvm);
 171
 172         /* The maximum number of VCPUs is limited by the host's GIC model */
 173         kvm->max_vcpus = kvm_arm_default_max_vcpus();
 174
 175         kvm_arm_init_hypercalls(kvm);
 176
 177         bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES);
 178
 179         return 0;
 180
 181 err_free_cpumask:
 182         free_cpumask_var(kvm->arch.supported_cpus);
 183 err_unshare_kvm:
 184         kvm_unshare_hyp(kvm, kvm + 1);
 185         return ret;
 186 }
 187
 188 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 189 {
 190         return VM_FAULT_SIGBUS;
 191 }
 192
 193
 194 /**
 195  * kvm_arch_destroy_vm - destroy the VM data structure
 196  * @kvm:        pointer to the KVM struct
 197  */
 198 void kvm_arch_destroy_vm(struct kvm *kvm)
 199 {
 200         bitmap_free(kvm->arch.pmu_filter);
 201         free_cpumask_var(kvm->arch.supported_cpus);
 202
 203         kvm_vgic_destroy(kvm);
 204
 205         if (is_protected_kvm_enabled())
 206                 pkvm_destroy_hyp_vm(kvm);
 207
 208         kfree(kvm->arch.mpidr_data);
 209         kvm_destroy_vcpus(kvm);
 210
 211         kvm_unshare_hyp(kvm, kvm + 1);
 212
 213         kvm_arm_teardown_hypercalls(kvm);
 214 }
 215
 216 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 217 {
 218         int r;
 219         switch (ext) {
 220         case KVM_CAP_IRQCHIP:
 221                 r = vgic_present;
 222                 break;
 223         case KVM_CAP_IOEVENTFD:
 224         case KVM_CAP_DEVICE_CTRL:
 225         case KVM_CAP_USER_MEMORY:
 226         case KVM_CAP_SYNC_MMU:
 227         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 228         case KVM_CAP_ONE_REG:
 229         case KVM_CAP_ARM_PSCI:
 230         case KVM_CAP_ARM_PSCI_0_2:
 231         case KVM_CAP_READONLY_MEM:
 232         case KVM_CAP_MP_STATE:
 233         case KVM_CAP_IMMEDIATE_EXIT:
 234         case KVM_CAP_VCPU_EVENTS:
 235         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 236         case KVM_CAP_ARM_NISV_TO_USER:
 237         case KVM_CAP_ARM_INJECT_EXT_DABT:
 238         case KVM_CAP_SET_GUEST_DEBUG:
 239         case KVM_CAP_VCPU_ATTRIBUTES:
 240         case KVM_CAP_PTP_KVM:
 241         case KVM_CAP_ARM_SYSTEM_SUSPEND:
 242         case KVM_CAP_IRQFD_RESAMPLE:
 243         case KVM_CAP_COUNTER_OFFSET:
 244                 r = 1;
 245                 break;
 246         case KVM_CAP_SET_GUEST_DEBUG2:
 247                 return KVM_GUESTDBG_VALID_MASK;
 248         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 249                 r = 1;
 250                 break;
 251         case KVM_CAP_NR_VCPUS:
 252                 /*
 253                  * ARM64 treats KVM_CAP_NR_CPUS differently from all other
 254                  * architectures, as it does not always bound it to
 255                  * KVM_CAP_MAX_VCPUS. It should not matter much because
 256                  * this is just an advisory value.
 257                  */
 258                 r = min_t(unsigned int, num_online_cpus(),
 259                           kvm_arm_default_max_vcpus());
 260                 break;
 261         case KVM_CAP_MAX_VCPUS:
 262         case KVM_CAP_MAX_VCPU_ID:
 263                 if (kvm)
 264                         r = kvm->max_vcpus;
 265                 else
 266                         r = kvm_arm_default_max_vcpus();
 267                 break;
 268         case KVM_CAP_MSI_DEVID:
 269                 if (!kvm)
 270                         r = -EINVAL;
 271                 else
 272                         r = kvm->arch.vgic.msis_require_devid;
 273                 break;
 274         case KVM_CAP_ARM_USER_IRQ:
 275                 /*
 276                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 277                  * (bump this number if adding more devices)
 278                  */
 279                 r = 1;
 280                 break;
 281         case KVM_CAP_ARM_MTE:
 282                 r = system_supports_mte();
 283                 break;
 284         case KVM_CAP_STEAL_TIME:
 285                 r = kvm_arm_pvtime_supported();
 286                 break;
 287         case KVM_CAP_ARM_EL1_32BIT:
 288                 r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);
 289                 break;
 290         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 291                 r = get_num_brps();
 292                 break;
 293         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 294                 r = get_num_wrps();
 295                 break;
 296         case KVM_CAP_ARM_PMU_V3:
 297                 r = kvm_arm_support_pmu_v3();
 298                 break;
 299         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 300                 r = cpus_have_final_cap(ARM64_HAS_RAS_EXTN);
 301                 break;
 302         case KVM_CAP_ARM_VM_IPA_SIZE:
 303                 r = get_kvm_ipa_limit();
 304                 break;
 305         case KVM_CAP_ARM_SVE:
 306                 r = system_supports_sve();
 307                 break;
 308         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 309         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 310                 r = system_has_full_ptr_auth();
 311                 break;
 312         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 313                 if (kvm)
 314                         r = kvm->arch.mmu.split_page_chunk_size;
 315                 else
 316                         r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
 317                 break;
 318         case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES:
 319                 r = kvm_supported_block_sizes();
 320                 break;
 321         case KVM_CAP_ARM_SUPPORTED_REG_MASK_RANGES:
 322                 r = BIT(0);
 323                 break;
 324         default:
 325                 r = 0;
 326         }
 327
 328         return r;
 329 }
 330
 331 long kvm_arch_dev_ioctl(struct file *filp,
 332                         unsigned int ioctl, unsigned long arg)
 333 {
 334         return -EINVAL;
 335 }
 336
 337 struct kvm *kvm_arch_alloc_vm(void)
 338 {
 339         size_t sz = sizeof(struct kvm);
 340
 341         if (!has_vhe())
 342                 return kzalloc(sz, GFP_KERNEL_ACCOUNT);
 343
 344         return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO);
 345 }
 346
 347 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 348 {
 349         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 350                 return -EBUSY;
 351
 352         if (id >= kvm->max_vcpus)
 353                 return -EINVAL;
 354
 355         return 0;
 356 }
 357
 358 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 359 {
 360         int err;
 361
 362         spin_lock_init(&vcpu->arch.mp_state_lock);
 363
 364 #ifdef CONFIG_LOCKDEP
 365         /* Inform lockdep that the config_lock is acquired after vcpu->mutex */
 366         mutex_lock(&vcpu->mutex);
 367         mutex_lock(&vcpu->kvm->arch.config_lock);
 368         mutex_unlock(&vcpu->kvm->arch.config_lock);
 369         mutex_unlock(&vcpu->mutex);
 370 #endif
 371
 372         /* Force users to call KVM_ARM_VCPU_INIT */
 373         vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
 374
 375         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 376
 377         /*
 378          * Default value for the FP state, will be overloaded at load
 379          * time if we support FP (pretty likely)
 380          */
 381         vcpu->arch.fp_state = FP_STATE_FREE;
 382
 383         /* Set up the timer */
 384         kvm_timer_vcpu_init(vcpu);
 385
 386         kvm_pmu_vcpu_init(vcpu);
 387
 388         kvm_arm_reset_debug_ptr(vcpu);
 389
 390         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 391
 392         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 393
 394         err = kvm_vgic_vcpu_init(vcpu);
 395         if (err)
 396                 return err;
 397
 398         return kvm_share_hyp(vcpu, vcpu + 1);
 399 }
 400
 401 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 402 {
 403 }
 404
 405 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 406 {
 407         if (vcpu_has_run_once(vcpu) && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 408                 static_branch_dec(&userspace_irqchip_in_use);
 409
 410         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 411         kvm_timer_vcpu_terminate(vcpu);
 412         kvm_pmu_vcpu_destroy(vcpu);
 413
 414         kvm_arm_vcpu_destroy(vcpu);
 415 }
 416
 417 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 418 {
 419
 420 }
 421
 422 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 423 {
 424
 425 }
 426
 427 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 428 {
 429         struct kvm_s2_mmu *mmu;
 430         int *last_ran;
 431
 432         mmu = vcpu->arch.hw_mmu;
 433         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 434
 435         /*
 436          * We guarantee that both TLBs and I-cache are private to each
 437          * vcpu. If detecting that a vcpu from the same VM has
 438          * previously run on the same physical CPU, call into the
 439          * hypervisor code to nuke the relevant contexts.
 440          *
 441          * We might get preempted before the vCPU actually runs, but
 442          * over-invalidation doesn't affect correctness.
 443          */
 444         if (*last_ran != vcpu->vcpu_idx) {
 445                 kvm_call_hyp(__kvm_flush_cpu_context, mmu);
 446                 *last_ran = vcpu->vcpu_idx;
 447         }
 448
 449         vcpu->cpu = cpu;
 450
 451         kvm_vgic_load(vcpu);
 452         kvm_timer_vcpu_load(vcpu);
 453         if (has_vhe())
 454                 kvm_vcpu_load_vhe(vcpu);
 455         kvm_arch_vcpu_load_fp(vcpu);
 456         kvm_vcpu_pmu_restore_guest(vcpu);
 457         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 458                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 459
 460         if (single_task_running())
 461                 vcpu_clear_wfx_traps(vcpu);
 462         else
 463                 vcpu_set_wfx_traps(vcpu);
 464
 465         if (vcpu_has_ptrauth(vcpu))
 466                 vcpu_ptrauth_disable(vcpu);
 467         kvm_arch_vcpu_load_debug_state_flags(vcpu);
 468
 469         if (!cpumask_test_cpu(cpu, vcpu->kvm->arch.supported_cpus))
 470                 vcpu_set_on_unsupported_cpu(vcpu);
 471 }
 472
 473 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 474 {
 475         kvm_arch_vcpu_put_debug_state_flags(vcpu);
 476         kvm_arch_vcpu_put_fp(vcpu);
 477         if (has_vhe())
 478                 kvm_vcpu_put_vhe(vcpu);
 479         kvm_timer_vcpu_put(vcpu);
 480         kvm_vgic_put(vcpu);
 481         kvm_vcpu_pmu_restore_host(vcpu);
 482         kvm_arm_vmid_clear_active();
 483
 484         vcpu_clear_on_unsupported_cpu(vcpu);
 485         vcpu->cpu = -1;
 486 }
 487
 488 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 489 {
 490         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED);
 491         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 492         kvm_vcpu_kick(vcpu);
 493 }
 494
 495 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 496 {
 497         spin_lock(&vcpu->arch.mp_state_lock);
 498         __kvm_arm_vcpu_power_off(vcpu);
 499         spin_unlock(&vcpu->arch.mp_state_lock);
 500 }
 501
 502 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu)
 503 {
 504         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED;
 505 }
 506
 507 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu)
 508 {
 509         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED);
 510         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 511         kvm_vcpu_kick(vcpu);
 512 }
 513
 514 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu)
 515 {
 516         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED;
 517 }
 518
 519 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 520                                     struct kvm_mp_state *mp_state)
 521 {
 522         *mp_state = READ_ONCE(vcpu->arch.mp_state);
 523
 524         return 0;
 525 }
 526
 527 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 528                                     struct kvm_mp_state *mp_state)
 529 {
 530         int ret = 0;
 531
 532         spin_lock(&vcpu->arch.mp_state_lock);
 533
 534         switch (mp_state->mp_state) {
 535         case KVM_MP_STATE_RUNNABLE:
 536                 WRITE_ONCE(vcpu->arch.mp_state, *mp_state);
 537                 break;
 538         case KVM_MP_STATE_STOPPED:
 539                 __kvm_arm_vcpu_power_off(vcpu);
 540                 break;
 541         case KVM_MP_STATE_SUSPENDED:
 542                 kvm_arm_vcpu_suspend(vcpu);
 543                 break;
 544         default:
 545                 ret = -EINVAL;
 546         }
 547
 548         spin_unlock(&vcpu->arch.mp_state_lock);
 549
 550         return ret;
 551 }
 552
 553 /**
 554  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 555  * @v:          The VCPU pointer
 556  *
 557  * If the guest CPU is not waiting for interrupts or an interrupt line is
 558  * asserted, the CPU is by definition runnable.
 559  */
 560 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 561 {
 562         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 563         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 564                 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause);
 565 }
 566
 567 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 568 {
 569         return vcpu_mode_priv(vcpu);
 570 }
 571
 572 #ifdef CONFIG_GUEST_PERF_EVENTS
 573 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu)
 574 {
 575         return *vcpu_pc(vcpu);
 576 }
 577 #endif
 578
 579 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 580 {
 581         return vcpu_get_flag(vcpu, VCPU_INITIALIZED);
 582 }
 583
 584 static void kvm_init_mpidr_data(struct kvm *kvm)
 585 {
 586         struct kvm_mpidr_data *data = NULL;
 587         unsigned long c, mask, nr_entries;
 588         u64 aff_set = 0, aff_clr = ~0UL;
 589         struct kvm_vcpu *vcpu;
 590
 591         mutex_lock(&kvm->arch.config_lock);
 592
 593         if (kvm->arch.mpidr_data || atomic_read(&kvm->online_vcpus) == 1)
 594                 goto out;
 595
 596         kvm_for_each_vcpu(c, vcpu, kvm) {
 597                 u64 aff = kvm_vcpu_get_mpidr_aff(vcpu);
 598                 aff_set |= aff;
 599                 aff_clr &= aff;
 600         }
 601
 602         /*
 603          * A significant bit can be either 0 or 1, and will only appear in
 604          * aff_set. Use aff_clr to weed out the useless stuff.
 605          */
 606         mask = aff_set ^ aff_clr;
 607         nr_entries = BIT_ULL(hweight_long(mask));
 608
 609         /*
 610          * Don't let userspace fool us. If we need more than a single page
 611          * to describe the compressed MPIDR array, just fall back to the
 612          * iterative method. Single vcpu VMs do not need this either.
 613          */
 614         if (struct_size(data, cmpidr_to_idx, nr_entries) <= PAGE_SIZE)
 615                 data = kzalloc(struct_size(data, cmpidr_to_idx, nr_entries),
 616                                GFP_KERNEL_ACCOUNT);
 617
 618         if (!data)
 619                 goto out;
 620
 621         data->mpidr_mask = mask;
 622
 623         kvm_for_each_vcpu(c, vcpu, kvm) {
 624                 u64 aff = kvm_vcpu_get_mpidr_aff(vcpu);
 625                 u16 index = kvm_mpidr_index(data, aff);
 626
 627                 data->cmpidr_to_idx[index] = c;
 628         }
 629
 630         kvm->arch.mpidr_data = data;
 631 out:
 632         mutex_unlock(&kvm->arch.config_lock);
 633 }
 634
 635 /*
 636  * Handle both the initialisation that is being done when the vcpu is
 637  * run for the first time, as well as the updates that must be
 638  * performed each time we get a new thread dealing with this vcpu.
 639  */
 640 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
 641 {
 642         struct kvm *kvm = vcpu->kvm;
 643         int ret;
 644
 645         if (!kvm_vcpu_initialized(vcpu))
 646                 return -ENOEXEC;
 647
 648         if (!kvm_arm_vcpu_is_finalized(vcpu))
 649                 return -EPERM;
 650
 651         ret = kvm_arch_vcpu_run_map_fp(vcpu);
 652         if (ret)
 653                 return ret;
 654
 655         if (likely(vcpu_has_run_once(vcpu)))
 656                 return 0;
 657
 658         kvm_init_mpidr_data(kvm);
 659
 660         kvm_arm_vcpu_init_debug(vcpu);
 661
 662         if (likely(irqchip_in_kernel(kvm))) {
 663                 /*
 664                  * Map the VGIC hardware resources before running a vcpu the
 665                  * first time on this VM.
 666                  */
 667                 ret = kvm_vgic_map_resources(kvm);
 668                 if (ret)
 669                         return ret;
 670         }
 671
 672         ret = kvm_timer_enable(vcpu);
 673         if (ret)
 674                 return ret;
 675
 676         ret = kvm_arm_pmu_v3_enable(vcpu);
 677         if (ret)
 678                 return ret;
 679
 680         if (is_protected_kvm_enabled()) {
 681                 ret = pkvm_create_hyp_vm(kvm);
 682                 if (ret)
 683                         return ret;
 684         }
 685
 686         if (!irqchip_in_kernel(kvm)) {
 687                 /*
 688                  * Tell the rest of the code that there are userspace irqchip
 689                  * VMs in the wild.
 690                  */
 691                 static_branch_inc(&userspace_irqchip_in_use);
 692         }
 693
 694         /*
 695          * Initialize traps for protected VMs.
 696          * NOTE: Move to run in EL2 directly, rather than via a hypercall, once
 697          * the code is in place for first run initialization at EL2.
 698          */
 699         if (kvm_vm_is_protected(kvm))
 700                 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu);
 701
 702         mutex_lock(&kvm->arch.config_lock);
 703         set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags);
 704         mutex_unlock(&kvm->arch.config_lock);
 705
 706         return ret;
 707 }
 708
 709 bool kvm_arch_intc_initialized(struct kvm *kvm)
 710 {
 711         return vgic_initialized(kvm);
 712 }
 713
 714 void kvm_arm_halt_guest(struct kvm *kvm)
 715 {
 716         unsigned long i;
 717         struct kvm_vcpu *vcpu;
 718
 719         kvm_for_each_vcpu(i, vcpu, kvm)
 720                 vcpu->arch.pause = true;
 721         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 722 }
 723
 724 void kvm_arm_resume_guest(struct kvm *kvm)
 725 {
 726         unsigned long i;
 727         struct kvm_vcpu *vcpu;
 728
 729         kvm_for_each_vcpu(i, vcpu, kvm) {
 730                 vcpu->arch.pause = false;
 731                 __kvm_vcpu_wake_up(vcpu);
 732         }
 733 }
 734
 735 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu)
 736 {
 737         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 738
 739         rcuwait_wait_event(wait,
 740                            (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause),
 741                            TASK_INTERRUPTIBLE);
 742
 743         if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) {
 744                 /* Awaken to handle a signal, request we sleep again later. */
 745                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 746         }
 747
 748         /*
 749          * Make sure we will observe a potential reset request if we've
 750          * observed a change to the power state. Pairs with the smp_wmb() in
 751          * kvm_psci_vcpu_on().
 752          */
 753         smp_rmb();
 754 }
 755
 756 /**
 757  * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior
 758  * @vcpu:       The VCPU pointer
 759  *
 760  * Suspend execution of a vCPU until a valid wake event is detected, i.e. until
 761  * the vCPU is runnable.  The vCPU may or may not be scheduled out, depending
 762  * on when a wake event arrives, e.g. there may already be a pending wake event.
 763  */
 764 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu)
 765 {
 766         /*
 767          * Sync back the state of the GIC CPU interface so that we have
 768          * the latest PMR and group enables. This ensures that
 769          * kvm_arch_vcpu_runnable has up-to-date data to decide whether
 770          * we have pending interrupts, e.g. when determining if the
 771          * vCPU should block.
 772          *
 773          * For the same reason, we want to tell GICv4 that we need
 774          * doorbells to be signalled, should an interrupt become pending.
 775          */
 776         preempt_disable();
 777         kvm_vgic_vmcr_sync(vcpu);
 778         vcpu_set_flag(vcpu, IN_WFI);
 779         vgic_v4_put(vcpu);
 780         preempt_enable();
 781
 782         kvm_vcpu_halt(vcpu);
 783         vcpu_clear_flag(vcpu, IN_WFIT);
 784
 785         preempt_disable();
 786         vcpu_clear_flag(vcpu, IN_WFI);
 787         vgic_v4_load(vcpu);
 788         preempt_enable();
 789 }
 790
 791 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu)
 792 {
 793         if (!kvm_arm_vcpu_suspended(vcpu))
 794                 return 1;
 795
 796         kvm_vcpu_wfi(vcpu);
 797
 798         /*
 799          * The suspend state is sticky; we do not leave it until userspace
 800          * explicitly marks the vCPU as runnable. Request that we suspend again
 801          * later.
 802          */
 803         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 804
 805         /*
 806          * Check to make sure the vCPU is actually runnable. If so, exit to
 807          * userspace informing it of the wakeup condition.
 808          */
 809         if (kvm_arch_vcpu_runnable(vcpu)) {
 810                 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event));
 811                 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP;
 812                 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
 813                 return 0;
 814         }
 815
 816         /*
 817          * Otherwise, we were unblocked to process a different event, such as a
 818          * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to
 819          * process the event.
 820          */
 821         return 1;
 822 }
 823
 824 /**
 825  * check_vcpu_requests - check and handle pending vCPU requests
 826  * @vcpu:       the VCPU pointer
 827  *
 828  * Return: 1 if we should enter the guest
 829  *         0 if we should exit to userspace
 830  *         < 0 if we should exit to userspace, where the return value indicates
 831  *         an error
 832  */
 833 static int check_vcpu_requests(struct kvm_vcpu *vcpu)
 834 {
 835         if (kvm_request_pending(vcpu)) {
 836                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 837                         kvm_vcpu_sleep(vcpu);
 838
 839                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 840                         kvm_reset_vcpu(vcpu);
 841
 842                 /*
 843                  * Clear IRQ_PENDING requests that were made to guarantee
 844                  * that a VCPU sees new virtual interrupts.
 845                  */
 846                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 847
 848                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 849                         kvm_update_stolen_time(vcpu);
 850
 851                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 852                         /* The distributor enable bits were changed */
 853                         preempt_disable();
 854                         vgic_v4_put(vcpu);
 855                         vgic_v4_load(vcpu);
 856                         preempt_enable();
 857                 }
 858
 859                 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
 860                         kvm_vcpu_reload_pmu(vcpu);
 861
 862                 if (kvm_check_request(KVM_REQ_RESYNC_PMU_EL0, vcpu))
 863                         kvm_vcpu_pmu_restore_guest(vcpu);
 864
 865                 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu))
 866                         return kvm_vcpu_suspend(vcpu);
 867
 868                 if (kvm_dirty_ring_check_request(vcpu))
 869                         return 0;
 870         }
 871
 872         return 1;
 873 }
 874
 875 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 876 {
 877         if (likely(!vcpu_mode_is_32bit(vcpu)))
 878                 return false;
 879
 880         if (vcpu_has_nv(vcpu))
 881                 return true;
 882
 883         return !kvm_supports_32bit_el0();
 884 }
 885
 886 /**
 887  * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest
 888  * @vcpu:       The VCPU pointer
 889  * @ret:        Pointer to write optional return code
 890  *
 891  * Returns: true if the VCPU needs to return to a preemptible + interruptible
 892  *          and skip guest entry.
 893  *
 894  * This function disambiguates between two different types of exits: exits to a
 895  * preemptible + interruptible kernel context and exits to userspace. For an
 896  * exit to userspace, this function will write the return code to ret and return
 897  * true. For an exit to preemptible + interruptible kernel context (i.e. check
 898  * for pending work and re-enter), return true without writing to ret.
 899  */
 900 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret)
 901 {
 902         struct kvm_run *run = vcpu->run;
 903
 904         /*
 905          * If we're using a userspace irqchip, then check if we need
 906          * to tell a userspace irqchip about timer or PMU level
 907          * changes and if so, exit to userspace (the actual level
 908          * state gets updated in kvm_timer_update_run and
 909          * kvm_pmu_update_run below).
 910          */
 911         if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 912                 if (kvm_timer_should_notify_user(vcpu) ||
 913                     kvm_pmu_should_notify_user(vcpu)) {
 914                         *ret = -EINTR;
 915                         run->exit_reason = KVM_EXIT_INTR;
 916                         return true;
 917                 }
 918         }
 919
 920         if (unlikely(vcpu_on_unsupported_cpu(vcpu))) {
 921                 run->exit_reason = KVM_EXIT_FAIL_ENTRY;
 922                 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED;
 923                 run->fail_entry.cpu = smp_processor_id();
 924                 *ret = 0;
 925                 return true;
 926         }
 927
 928         return kvm_request_pending(vcpu) ||
 929                         xfer_to_guest_mode_work_pending();
 930 }
 931
 932 /*
 933  * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while
 934  * the vCPU is running.
 935  *
 936  * This must be noinstr as instrumentation may make use of RCU, and this is not
 937  * safe during the EQS.
 938  */
 939 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
 940 {
 941         int ret;
 942
 943         guest_state_enter_irqoff();
 944         ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 945         guest_state_exit_irqoff();
 946
 947         return ret;
 948 }
 949
 950 /**
 951  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 952  * @vcpu:       The VCPU pointer
 953  *
 954  * This function is called through the VCPU_RUN ioctl called from user space. It
 955  * will execute VM code in a loop until the time slice for the process is used
 956  * or some emulation is needed from user space in which case the function will
 957  * return with return value 0 and with the kvm_run structure filled in with the
 958  * required data for the requested emulation.
 959  */
 960 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 961 {
 962         struct kvm_run *run = vcpu->run;
 963         int ret;
 964
 965         if (run->exit_reason == KVM_EXIT_MMIO) {
 966                 ret = kvm_handle_mmio_return(vcpu);
 967                 if (ret)
 968                         return ret;
 969         }
 970
 971         vcpu_load(vcpu);
 972
 973         if (run->immediate_exit) {
 974                 ret = -EINTR;
 975                 goto out;
 976         }
 977
 978         kvm_sigset_activate(vcpu);
 979
 980         ret = 1;
 981         run->exit_reason = KVM_EXIT_UNKNOWN;
 982         run->flags = 0;
 983         while (ret > 0) {
 984                 /*
 985                  * Check conditions before entering the guest
 986                  */
 987                 ret = xfer_to_guest_mode_handle_work(vcpu);
 988                 if (!ret)
 989                         ret = 1;
 990
 991                 if (ret > 0)
 992                         ret = check_vcpu_requests(vcpu);
 993
 994                 /*
 995                  * Preparing the interrupts to be injected also
 996                  * involves poking the GIC, which must be done in a
 997                  * non-preemptible context.
 998                  */
 999                 preempt_disable();
1000
1001                 /*
1002                  * The VMID allocator only tracks active VMIDs per
1003                  * physical CPU, and therefore the VMID allocated may not be
1004                  * preserved on VMID roll-over if the task was preempted,
1005                  * making a thread's VMID inactive. So we need to call
1006                  * kvm_arm_vmid_update() in non-premptible context.
1007                  */
1008                 if (kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid) &&
1009                     has_vhe())
1010                         __load_stage2(vcpu->arch.hw_mmu,
1011                                       vcpu->arch.hw_mmu->arch);
1012
1013                 kvm_pmu_flush_hwstate(vcpu);
1014
1015                 local_irq_disable();
1016
1017                 kvm_vgic_flush_hwstate(vcpu);
1018
1019                 kvm_pmu_update_vcpu_events(vcpu);
1020
1021                 /*
1022                  * Ensure we set mode to IN_GUEST_MODE after we disable
1023                  * interrupts and before the final VCPU requests check.
1024                  * See the comment in kvm_vcpu_exiting_guest_mode() and
1025                  * Documentation/virt/kvm/vcpu-requests.rst
1026                  */
1027                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
1028
1029                 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) {
1030                         vcpu->mode = OUTSIDE_GUEST_MODE;
1031                         isb(); /* Ensure work in x_flush_hwstate is committed */
1032                         kvm_pmu_sync_hwstate(vcpu);
1033                         if (static_branch_unlikely(&userspace_irqchip_in_use))
1034                                 kvm_timer_sync_user(vcpu);
1035                         kvm_vgic_sync_hwstate(vcpu);
1036                         local_irq_enable();
1037                         preempt_enable();
1038                         continue;
1039                 }
1040
1041                 kvm_arm_setup_debug(vcpu);
1042                 kvm_arch_vcpu_ctxflush_fp(vcpu);
1043
1044                 /**************************************************************
1045                  * Enter the guest
1046                  */
1047                 trace_kvm_entry(*vcpu_pc(vcpu));
1048                 guest_timing_enter_irqoff();
1049
1050                 ret = kvm_arm_vcpu_enter_exit(vcpu);
1051
1052                 vcpu->mode = OUTSIDE_GUEST_MODE;
1053                 vcpu->stat.exits++;
1054                 /*
1055                  * Back from guest
1056                  *************************************************************/
1057
1058                 kvm_arm_clear_debug(vcpu);
1059
1060                 /*
1061                  * We must sync the PMU state before the vgic state so
1062                  * that the vgic can properly sample the updated state of the
1063                  * interrupt line.
1064                  */
1065                 kvm_pmu_sync_hwstate(vcpu);
1066
1067                 /*
1068                  * Sync the vgic state before syncing the timer state because
1069                  * the timer code needs to know if the virtual timer
1070                  * interrupts are active.
1071                  */
1072                 kvm_vgic_sync_hwstate(vcpu);
1073
1074                 /*
1075                  * Sync the timer hardware state before enabling interrupts as
1076                  * we don't want vtimer interrupts to race with syncing the
1077                  * timer virtual interrupt state.
1078                  */
1079                 if (static_branch_unlikely(&userspace_irqchip_in_use))
1080                         kvm_timer_sync_user(vcpu);
1081
1082                 kvm_arch_vcpu_ctxsync_fp(vcpu);
1083
1084                 /*
1085                  * We must ensure that any pending interrupts are taken before
1086                  * we exit guest timing so that timer ticks are accounted as
1087                  * guest time. Transiently unmask interrupts so that any
1088                  * pending interrupts are taken.
1089                  *
1090                  * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other
1091                  * context synchronization event) is necessary to ensure that
1092                  * pending interrupts are taken.
1093                  */
1094                 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) {
1095                         local_irq_enable();
1096                         isb();
1097                         local_irq_disable();
1098                 }
1099
1100                 guest_timing_exit_irqoff();
1101
1102                 local_irq_enable();
1103
1104                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
1105
1106                 /* Exit types that need handling before we can be preempted */
1107                 handle_exit_early(vcpu, ret);
1108
1109                 preempt_enable();
1110
1111                 /*
1112                  * The ARMv8 architecture doesn't give the hypervisor
1113                  * a mechanism to prevent a guest from dropping to AArch32 EL0
1114                  * if implemented by the CPU. If we spot the guest in such
1115                  * state and that we decided it wasn't supposed to do so (like
1116                  * with the asymmetric AArch32 case), return to userspace with
1117                  * a fatal error.
1118                  */
1119                 if (vcpu_mode_is_bad_32bit(vcpu)) {
1120                         /*
1121                          * As we have caught the guest red-handed, decide that
1122                          * it isn't fit for purpose anymore by making the vcpu
1123                          * invalid. The VMM can try and fix it by issuing  a
1124                          * KVM_ARM_VCPU_INIT if it really wants to.
1125                          */
1126                         vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
1127                         ret = ARM_EXCEPTION_IL;
1128                 }
1129
1130                 ret = handle_exit(vcpu, ret);
1131         }
1132
1133         /* Tell userspace about in-kernel device output levels */
1134         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1135                 kvm_timer_update_run(vcpu);
1136                 kvm_pmu_update_run(vcpu);
1137         }
1138
1139         kvm_sigset_deactivate(vcpu);
1140
1141 out:
1142         /*
1143          * In the unlikely event that we are returning to userspace
1144          * with pending exceptions or PC adjustment, commit these
1145          * adjustments in order to give userspace a consistent view of
1146          * the vcpu state. Note that this relies on __kvm_adjust_pc()
1147          * being preempt-safe on VHE.
1148          */
1149         if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) ||
1150                      vcpu_get_flag(vcpu, INCREMENT_PC)))
1151                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
1152
1153         vcpu_put(vcpu);
1154         return ret;
1155 }
1156
1157 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
1158 {
1159         int bit_index;
1160         bool set;
1161         unsigned long *hcr;
1162
1163         if (number == KVM_ARM_IRQ_CPU_IRQ)
1164                 bit_index = __ffs(HCR_VI);
1165         else /* KVM_ARM_IRQ_CPU_FIQ */
1166                 bit_index = __ffs(HCR_VF);
1167
1168         hcr = vcpu_hcr(vcpu);
1169         if (level)
1170                 set = test_and_set_bit(bit_index, hcr);
1171         else
1172                 set = test_and_clear_bit(bit_index, hcr);
1173
1174         /*
1175          * If we didn't change anything, no need to wake up or kick other CPUs
1176          */
1177         if (set == level)
1178                 return 0;
1179
1180         /*
1181          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
1182          * trigger a world-switch round on the running physical CPU to set the
1183          * virtual IRQ/FIQ fields in the HCR appropriately.
1184          */
1185         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
1186         kvm_vcpu_kick(vcpu);
1187
1188         return 0;
1189 }
1190
1191 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
1192                           bool line_status)
1193 {
1194         u32 irq = irq_level->irq;
1195         unsigned int irq_type, vcpu_id, irq_num;
1196         struct kvm_vcpu *vcpu = NULL;
1197         bool level = irq_level->level;
1198
1199         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
1200         vcpu_id = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
1201         vcpu_id += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
1202         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
1203
1204         trace_kvm_irq_line(irq_type, vcpu_id, irq_num, irq_level->level);
1205
1206         switch (irq_type) {
1207         case KVM_ARM_IRQ_TYPE_CPU:
1208                 if (irqchip_in_kernel(kvm))
1209                         return -ENXIO;
1210
1211                 vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
1212                 if (!vcpu)
1213                         return -EINVAL;
1214
1215                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1216                         return -EINVAL;
1217
1218                 return vcpu_interrupt_line(vcpu, irq_num, level);
1219         case KVM_ARM_IRQ_TYPE_PPI:
1220                 if (!irqchip_in_kernel(kvm))
1221                         return -ENXIO;
1222
1223                 vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
1224                 if (!vcpu)
1225                         return -EINVAL;
1226
1227                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1228                         return -EINVAL;
1229
1230                 return kvm_vgic_inject_irq(kvm, vcpu, irq_num, level, NULL);
1231         case KVM_ARM_IRQ_TYPE_SPI:
1232                 if (!irqchip_in_kernel(kvm))
1233                         return -ENXIO;
1234
1235                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1236                         return -EINVAL;
1237
1238                 return kvm_vgic_inject_irq(kvm, NULL, irq_num, level, NULL);
1239         }
1240
1241         return -EINVAL;
1242 }
1243
1244 static unsigned long system_supported_vcpu_features(void)
1245 {
1246         unsigned long features = KVM_VCPU_VALID_FEATURES;
1247
1248         if (!cpus_have_final_cap(ARM64_HAS_32BIT_EL1))
1249                 clear_bit(KVM_ARM_VCPU_EL1_32BIT, &features);
1250
1251         if (!kvm_arm_support_pmu_v3())
1252                 clear_bit(KVM_ARM_VCPU_PMU_V3, &features);
1253
1254         if (!system_supports_sve())
1255                 clear_bit(KVM_ARM_VCPU_SVE, &features);
1256
1257         if (!system_has_full_ptr_auth()) {
1258                 clear_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features);
1259                 clear_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features);
1260         }
1261
1262         if (!cpus_have_final_cap(ARM64_HAS_NESTED_VIRT))
1263                 clear_bit(KVM_ARM_VCPU_HAS_EL2, &features);
1264
1265         return features;
1266 }
1267
1268 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
1269                                         const struct kvm_vcpu_init *init)
1270 {
1271         unsigned long features = init->features[0];
1272         int i;
1273
1274         if (features & ~KVM_VCPU_VALID_FEATURES)
1275                 return -ENOENT;
1276
1277         for (i = 1; i < ARRAY_SIZE(init->features); i++) {
1278                 if (init->features[i])
1279                         return -ENOENT;
1280         }
1281
1282         if (features & ~system_supported_vcpu_features())
1283                 return -EINVAL;
1284
1285         /*
1286          * For now make sure that both address/generic pointer authentication
1287          * features are requested by the userspace together.
1288          */
1289         if (test_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features) !=
1290             test_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features))
1291                 return -EINVAL;
1292
1293         /* Disallow NV+SVE for the time being */
1294         if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features) &&
1295             test_bit(KVM_ARM_VCPU_SVE, &features))
1296                 return -EINVAL;
1297
1298         if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
1299                 return 0;
1300
1301         /* MTE is incompatible with AArch32 */
1302         if (kvm_has_mte(vcpu->kvm))
1303                 return -EINVAL;
1304
1305         /* NV is incompatible with AArch32 */
1306         if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
1307                 return -EINVAL;
1308
1309         return 0;
1310 }
1311
1312 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu,
1313                                   const struct kvm_vcpu_init *init)
1314 {
1315         unsigned long features = init->features[0];
1316
1317         return !bitmap_equal(vcpu->kvm->arch.vcpu_features, &features,
1318                              KVM_VCPU_MAX_FEATURES);
1319 }
1320
1321 static int kvm_setup_vcpu(struct kvm_vcpu *vcpu)
1322 {
1323         struct kvm *kvm = vcpu->kvm;
1324         int ret = 0;
1325
1326         /*
1327          * When the vCPU has a PMU, but no PMU is set for the guest
1328          * yet, set the default one.
1329          */
1330         if (kvm_vcpu_has_pmu(vcpu) && !kvm->arch.arm_pmu)
1331                 ret = kvm_arm_set_default_pmu(kvm);
1332
1333         return ret;
1334 }
1335
1336 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1337                                  const struct kvm_vcpu_init *init)
1338 {
1339         unsigned long features = init->features[0];
1340         struct kvm *kvm = vcpu->kvm;
1341         int ret = -EINVAL;
1342
1343         mutex_lock(&kvm->arch.config_lock);
1344
1345         if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) &&
1346             kvm_vcpu_init_changed(vcpu, init))
1347                 goto out_unlock;
1348
1349         bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES);
1350
1351         ret = kvm_setup_vcpu(vcpu);
1352         if (ret)
1353                 goto out_unlock;
1354
1355         /* Now we know what it is, we can reset it. */
1356         kvm_reset_vcpu(vcpu);
1357
1358         set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags);
1359         vcpu_set_flag(vcpu, VCPU_INITIALIZED);
1360         ret = 0;
1361 out_unlock:
1362         mutex_unlock(&kvm->arch.config_lock);
1363         return ret;
1364 }
1365
1366 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1367                                const struct kvm_vcpu_init *init)
1368 {
1369         int ret;
1370
1371         if (init->target != KVM_ARM_TARGET_GENERIC_V8 &&
1372             init->target != kvm_target_cpu())
1373                 return -EINVAL;
1374
1375         ret = kvm_vcpu_init_check_features(vcpu, init);
1376         if (ret)
1377                 return ret;
1378
1379         if (!kvm_vcpu_initialized(vcpu))
1380                 return __kvm_vcpu_set_target(vcpu, init);
1381
1382         if (kvm_vcpu_init_changed(vcpu, init))
1383                 return -EINVAL;
1384
1385         kvm_reset_vcpu(vcpu);
1386         return 0;
1387 }
1388
1389 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1390                                          struct kvm_vcpu_init *init)
1391 {
1392         bool power_off = false;
1393         int ret;
1394
1395         /*
1396          * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid
1397          * reflecting it in the finalized feature set, thus limiting its scope
1398          * to a single KVM_ARM_VCPU_INIT call.
1399          */
1400         if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) {
1401                 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF);
1402                 power_off = true;
1403         }
1404
1405         ret = kvm_vcpu_set_target(vcpu, init);
1406         if (ret)
1407                 return ret;
1408
1409         /*
1410          * Ensure a rebooted VM will fault in RAM pages and detect if the
1411          * guest MMU is turned off and flush the caches as needed.
1412          *
1413          * S2FWB enforces all memory accesses to RAM being cacheable,
1414          * ensuring that the data side is always coherent. We still
1415          * need to invalidate the I-cache though, as FWB does *not*
1416          * imply CTR_EL0.DIC.
1417          */
1418         if (vcpu_has_run_once(vcpu)) {
1419                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1420                         stage2_unmap_vm(vcpu->kvm);
1421                 else
1422                         icache_inval_all_pou();
1423         }
1424
1425         vcpu_reset_hcr(vcpu);
1426         vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu);
1427
1428         /*
1429          * Handle the "start in power-off" case.
1430          */
1431         spin_lock(&vcpu->arch.mp_state_lock);
1432
1433         if (power_off)
1434                 __kvm_arm_vcpu_power_off(vcpu);
1435         else
1436                 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE);
1437
1438         spin_unlock(&vcpu->arch.mp_state_lock);
1439
1440         return 0;
1441 }
1442
1443 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1444                                  struct kvm_device_attr *attr)
1445 {
1446         int ret = -ENXIO;
1447
1448         switch (attr->group) {
1449         default:
1450                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1451                 break;
1452         }
1453
1454         return ret;
1455 }
1456
1457 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1458                                  struct kvm_device_attr *attr)
1459 {
1460         int ret = -ENXIO;
1461
1462         switch (attr->group) {
1463         default:
1464                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1465                 break;
1466         }
1467
1468         return ret;
1469 }
1470
1471 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1472                                  struct kvm_device_attr *attr)
1473 {
1474         int ret = -ENXIO;
1475
1476         switch (attr->group) {
1477         default:
1478                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1479                 break;
1480         }
1481
1482         return ret;
1483 }
1484
1485 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1486                                    struct kvm_vcpu_events *events)
1487 {
1488         memset(events, 0, sizeof(*events));
1489
1490         return __kvm_arm_vcpu_get_events(vcpu, events);
1491 }
1492
1493 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1494                                    struct kvm_vcpu_events *events)
1495 {
1496         int i;
1497
1498         /* check whether the reserved field is zero */
1499         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1500                 if (events->reserved[i])
1501                         return -EINVAL;
1502
1503         /* check whether the pad field is zero */
1504         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1505                 if (events->exception.pad[i])
1506                         return -EINVAL;
1507
1508         return __kvm_arm_vcpu_set_events(vcpu, events);
1509 }
1510
1511 long kvm_arch_vcpu_ioctl(struct file *filp,
1512                          unsigned int ioctl, unsigned long arg)
1513 {
1514         struct kvm_vcpu *vcpu = filp->private_data;
1515         void __user *argp = (void __user *)arg;
1516         struct kvm_device_attr attr;
1517         long r;
1518
1519         switch (ioctl) {
1520         case KVM_ARM_VCPU_INIT: {
1521                 struct kvm_vcpu_init init;
1522
1523                 r = -EFAULT;
1524                 if (copy_from_user(&init, argp, sizeof(init)))
1525                         break;
1526
1527                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1528                 break;
1529         }
1530         case KVM_SET_ONE_REG:
1531         case KVM_GET_ONE_REG: {
1532                 struct kvm_one_reg reg;
1533
1534                 r = -ENOEXEC;
1535                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1536                         break;
1537
1538                 r = -EFAULT;
1539                 if (copy_from_user(&reg, argp, sizeof(reg)))
1540                         break;
1541
1542                 /*
1543                  * We could owe a reset due to PSCI. Handle the pending reset
1544                  * here to ensure userspace register accesses are ordered after
1545                  * the reset.
1546                  */
1547                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1548                         kvm_reset_vcpu(vcpu);
1549
1550                 if (ioctl == KVM_SET_ONE_REG)
1551                         r = kvm_arm_set_reg(vcpu, &reg);
1552                 else
1553                         r = kvm_arm_get_reg(vcpu, &reg);
1554                 break;
1555         }
1556         case KVM_GET_REG_LIST: {
1557                 struct kvm_reg_list __user *user_list = argp;
1558                 struct kvm_reg_list reg_list;
1559                 unsigned n;
1560
1561                 r = -ENOEXEC;
1562                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1563                         break;
1564
1565                 r = -EPERM;
1566                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1567                         break;
1568
1569                 r = -EFAULT;
1570                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1571                         break;
1572                 n = reg_list.n;
1573                 reg_list.n = kvm_arm_num_regs(vcpu);
1574                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1575                         break;
1576                 r = -E2BIG;
1577                 if (n < reg_list.n)
1578                         break;
1579                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1580                 break;
1581         }
1582         case KVM_SET_DEVICE_ATTR: {
1583                 r = -EFAULT;
1584                 if (copy_from_user(&attr, argp, sizeof(attr)))
1585                         break;
1586                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1587                 break;
1588         }
1589         case KVM_GET_DEVICE_ATTR: {
1590                 r = -EFAULT;
1591                 if (copy_from_user(&attr, argp, sizeof(attr)))
1592                         break;
1593                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1594                 break;
1595         }
1596         case KVM_HAS_DEVICE_ATTR: {
1597                 r = -EFAULT;
1598                 if (copy_from_user(&attr, argp, sizeof(attr)))
1599                         break;
1600                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1601                 break;
1602         }
1603         case KVM_GET_VCPU_EVENTS: {
1604                 struct kvm_vcpu_events events;
1605
1606                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1607                         return -EINVAL;
1608
1609                 if (copy_to_user(argp, &events, sizeof(events)))
1610                         return -EFAULT;
1611
1612                 return 0;
1613         }
1614         case KVM_SET_VCPU_EVENTS: {
1615                 struct kvm_vcpu_events events;
1616
1617                 if (copy_from_user(&events, argp, sizeof(events)))
1618                         return -EFAULT;
1619
1620                 return kvm_arm_vcpu_set_events(vcpu, &events);
1621         }
1622         case KVM_ARM_VCPU_FINALIZE: {
1623                 int what;
1624
1625                 if (!kvm_vcpu_initialized(vcpu))
1626                         return -ENOEXEC;
1627
1628                 if (get_user(what, (const int __user *)argp))
1629                         return -EFAULT;
1630
1631                 return kvm_arm_vcpu_finalize(vcpu, what);
1632         }
1633         default:
1634                 r = -EINVAL;
1635         }
1636
1637         return r;
1638 }
1639
1640 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1641 {
1642
1643 }
1644
1645 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1646                                         struct kvm_arm_device_addr *dev_addr)
1647 {
1648         switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) {
1649         case KVM_ARM_DEVICE_VGIC_V2:
1650                 if (!vgic_present)
1651                         return -ENXIO;
1652                 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr);
1653         default:
1654                 return -ENODEV;
1655         }
1656 }
1657
1658 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1659 {
1660         switch (attr->group) {
1661         case KVM_ARM_VM_SMCCC_CTRL:
1662                 return kvm_vm_smccc_has_attr(kvm, attr);
1663         default:
1664                 return -ENXIO;
1665         }
1666 }
1667
1668 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1669 {
1670         switch (attr->group) {
1671         case KVM_ARM_VM_SMCCC_CTRL:
1672                 return kvm_vm_smccc_set_attr(kvm, attr);
1673         default:
1674                 return -ENXIO;
1675         }
1676 }
1677
1678 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
1679 {
1680         struct kvm *kvm = filp->private_data;
1681         void __user *argp = (void __user *)arg;
1682         struct kvm_device_attr attr;
1683
1684         switch (ioctl) {
1685         case KVM_CREATE_IRQCHIP: {
1686                 int ret;
1687                 if (!vgic_present)
1688                         return -ENXIO;
1689                 mutex_lock(&kvm->lock);
1690                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1691                 mutex_unlock(&kvm->lock);
1692                 return ret;
1693         }
1694         case KVM_ARM_SET_DEVICE_ADDR: {
1695                 struct kvm_arm_device_addr dev_addr;
1696
1697                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1698                         return -EFAULT;
1699                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1700         }
1701         case KVM_ARM_PREFERRED_TARGET: {
1702                 struct kvm_vcpu_init init = {
1703                         .target = KVM_ARM_TARGET_GENERIC_V8,
1704                 };
1705
1706                 if (copy_to_user(argp, &init, sizeof(init)))
1707                         return -EFAULT;
1708
1709                 return 0;
1710         }
1711         case KVM_ARM_MTE_COPY_TAGS: {
1712                 struct kvm_arm_copy_mte_tags copy_tags;
1713
1714                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1715                         return -EFAULT;
1716                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1717         }
1718         case KVM_ARM_SET_COUNTER_OFFSET: {
1719                 struct kvm_arm_counter_offset offset;
1720
1721                 if (copy_from_user(&offset, argp, sizeof(offset)))
1722                         return -EFAULT;
1723                 return kvm_vm_ioctl_set_counter_offset(kvm, &offset);
1724         }
1725         case KVM_HAS_DEVICE_ATTR: {
1726                 if (copy_from_user(&attr, argp, sizeof(attr)))
1727                         return -EFAULT;
1728
1729                 return kvm_vm_has_attr(kvm, &attr);
1730         }
1731         case KVM_SET_DEVICE_ATTR: {
1732                 if (copy_from_user(&attr, argp, sizeof(attr)))
1733                         return -EFAULT;
1734
1735                 return kvm_vm_set_attr(kvm, &attr);
1736         }
1737         case KVM_ARM_GET_REG_WRITABLE_MASKS: {
1738                 struct reg_mask_range range;
1739
1740                 if (copy_from_user(&range, argp, sizeof(range)))
1741                         return -EFAULT;
1742                 return kvm_vm_ioctl_get_reg_writable_masks(kvm, &range);
1743         }
1744         default:
1745                 return -EINVAL;
1746         }
1747 }
1748
1749 /* unlocks vcpus from @vcpu_lock_idx and smaller */
1750 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx)
1751 {
1752         struct kvm_vcpu *tmp_vcpu;
1753
1754         for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
1755                 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
1756                 mutex_unlock(&tmp_vcpu->mutex);
1757         }
1758 }
1759
1760 void unlock_all_vcpus(struct kvm *kvm)
1761 {
1762         lockdep_assert_held(&kvm->lock);
1763
1764         unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1);
1765 }
1766
1767 /* Returns true if all vcpus were locked, false otherwise */
1768 bool lock_all_vcpus(struct kvm *kvm)
1769 {
1770         struct kvm_vcpu *tmp_vcpu;
1771         unsigned long c;
1772
1773         lockdep_assert_held(&kvm->lock);
1774
1775         /*
1776          * Any time a vcpu is in an ioctl (including running), the
1777          * core KVM code tries to grab the vcpu->mutex.
1778          *
1779          * By grabbing the vcpu->mutex of all VCPUs we ensure that no
1780          * other VCPUs can fiddle with the state while we access it.
1781          */
1782         kvm_for_each_vcpu(c, tmp_vcpu, kvm) {
1783                 if (!mutex_trylock(&tmp_vcpu->mutex)) {
1784                         unlock_vcpus(kvm, c - 1);
1785                         return false;
1786                 }
1787         }
1788
1789         return true;
1790 }
1791
1792 static unsigned long nvhe_percpu_size(void)
1793 {
1794         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1795                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1796 }
1797
1798 static unsigned long nvhe_percpu_order(void)
1799 {
1800         unsigned long size = nvhe_percpu_size();
1801
1802         return size ? get_order(size) : 0;
1803 }
1804
1805 /* A lookup table holding the hypervisor VA for each vector slot */
1806 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1807
1808 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1809 {
1810         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1811 }
1812
1813 static int kvm_init_vector_slots(void)
1814 {
1815         int err;
1816         void *base;
1817
1818         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1819         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1820
1821         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1822         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1823
1824         if (kvm_system_needs_idmapped_vectors() &&
1825             !is_protected_kvm_enabled()) {
1826                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1827                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1828                 if (err)
1829                         return err;
1830         }
1831
1832         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1833         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1834         return 0;
1835 }
1836
1837 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits)
1838 {
1839         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1840         unsigned long tcr;
1841
1842         /*
1843          * Calculate the raw per-cpu offset without a translation from the
1844          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1845          * so that we can use adr_l to access per-cpu variables in EL2.
1846          * Also drop the KASAN tag which gets in the way...
1847          */
1848         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1849                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1850
1851         params->mair_el2 = read_sysreg(mair_el1);
1852
1853         tcr = read_sysreg(tcr_el1);
1854         if (cpus_have_final_cap(ARM64_KVM_HVHE)) {
1855                 tcr |= TCR_EPD1_MASK;
1856         } else {
1857                 tcr &= TCR_EL2_MASK;
1858                 tcr |= TCR_EL2_RES1;
1859         }
1860         tcr &= ~TCR_T0SZ_MASK;
1861         tcr |= TCR_T0SZ(hyp_va_bits);
1862         params->tcr_el2 = tcr;
1863
1864         params->pgd_pa = kvm_mmu_get_httbr();
1865         if (is_protected_kvm_enabled())
1866                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1867         else
1868                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1869         if (cpus_have_final_cap(ARM64_KVM_HVHE))
1870                 params->hcr_el2 |= HCR_E2H;
1871         params->vttbr = params->vtcr = 0;
1872
1873         /*
1874          * Flush the init params from the data cache because the struct will
1875          * be read while the MMU is off.
1876          */
1877         kvm_flush_dcache_to_poc(params, sizeof(*params));
1878 }
1879
1880 static void hyp_install_host_vector(void)
1881 {
1882         struct kvm_nvhe_init_params *params;
1883         struct arm_smccc_res res;
1884
1885         /* Switch from the HYP stub to our own HYP init vector */
1886         __hyp_set_vectors(kvm_get_idmap_vector());
1887
1888         /*
1889          * Call initialization code, and switch to the full blown HYP code.
1890          * If the cpucaps haven't been finalized yet, something has gone very
1891          * wrong, and hyp will crash and burn when it uses any
1892          * cpus_have_*_cap() wrapper.
1893          */
1894         BUG_ON(!system_capabilities_finalized());
1895         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1896         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1897         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1898 }
1899
1900 static void cpu_init_hyp_mode(void)
1901 {
1902         hyp_install_host_vector();
1903
1904         /*
1905          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1906          * at EL2.
1907          */
1908         if (this_cpu_has_cap(ARM64_SSBS) &&
1909             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1910                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1911         }
1912 }
1913
1914 static void cpu_hyp_reset(void)
1915 {
1916         if (!is_kernel_in_hyp_mode())
1917                 __hyp_reset_vectors();
1918 }
1919
1920 /*
1921  * EL2 vectors can be mapped and rerouted in a number of ways,
1922  * depending on the kernel configuration and CPU present:
1923  *
1924  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1925  *   placed in one of the vector slots, which is executed before jumping
1926  *   to the real vectors.
1927  *
1928  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1929  *   containing the hardening sequence is mapped next to the idmap page,
1930  *   and executed before jumping to the real vectors.
1931  *
1932  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1933  *   empty slot is selected, mapped next to the idmap page, and
1934  *   executed before jumping to the real vectors.
1935  *
1936  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1937  * VHE, as we don't have hypervisor-specific mappings. If the system
1938  * is VHE and yet selects this capability, it will be ignored.
1939  */
1940 static void cpu_set_hyp_vector(void)
1941 {
1942         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1943         void *vector = hyp_spectre_vector_selector[data->slot];
1944
1945         if (!is_protected_kvm_enabled())
1946                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1947         else
1948                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1949 }
1950
1951 static void cpu_hyp_init_context(void)
1952 {
1953         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1954
1955         if (!is_kernel_in_hyp_mode())
1956                 cpu_init_hyp_mode();
1957 }
1958
1959 static void cpu_hyp_init_features(void)
1960 {
1961         cpu_set_hyp_vector();
1962         kvm_arm_init_debug();
1963
1964         if (is_kernel_in_hyp_mode())
1965                 kvm_timer_init_vhe();
1966
1967         if (vgic_present)
1968                 kvm_vgic_init_cpu_hardware();
1969 }
1970
1971 static void cpu_hyp_reinit(void)
1972 {
1973         cpu_hyp_reset();
1974         cpu_hyp_init_context();
1975         cpu_hyp_init_features();
1976 }
1977
1978 static void cpu_hyp_init(void *discard)
1979 {
1980         if (!__this_cpu_read(kvm_hyp_initialized)) {
1981                 cpu_hyp_reinit();
1982                 __this_cpu_write(kvm_hyp_initialized, 1);
1983         }
1984 }
1985
1986 static void cpu_hyp_uninit(void *discard)
1987 {
1988         if (__this_cpu_read(kvm_hyp_initialized)) {
1989                 cpu_hyp_reset();
1990                 __this_cpu_write(kvm_hyp_initialized, 0);
1991         }
1992 }
1993
1994 int kvm_arch_hardware_enable(void)
1995 {
1996         /*
1997          * Most calls to this function are made with migration
1998          * disabled, but not with preemption disabled. The former is
1999          * enough to ensure correctness, but most of the helpers
2000          * expect the later and will throw a tantrum otherwise.
2001          */
2002         preempt_disable();
2003
2004         cpu_hyp_init(NULL);
2005
2006         kvm_vgic_cpu_up();
2007         kvm_timer_cpu_up();
2008
2009         preempt_enable();
2010
2011         return 0;
2012 }
2013
2014 void kvm_arch_hardware_disable(void)
2015 {
2016         kvm_timer_cpu_down();
2017         kvm_vgic_cpu_down();
2018
2019         if (!is_protected_kvm_enabled())
2020                 cpu_hyp_uninit(NULL);
2021 }
2022
2023 #ifdef CONFIG_CPU_PM
2024 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
2025                                     unsigned long cmd,
2026                                     void *v)
2027 {
2028         /*
2029          * kvm_hyp_initialized is left with its old value over
2030          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
2031          * re-enable hyp.
2032          */
2033         switch (cmd) {
2034         case CPU_PM_ENTER:
2035                 if (__this_cpu_read(kvm_hyp_initialized))
2036                         /*
2037                          * don't update kvm_hyp_initialized here
2038                          * so that the hyp will be re-enabled
2039                          * when we resume. See below.
2040                          */
2041                         cpu_hyp_reset();
2042
2043                 return NOTIFY_OK;
2044         case CPU_PM_ENTER_FAILED:
2045         case CPU_PM_EXIT:
2046                 if (__this_cpu_read(kvm_hyp_initialized))
2047                         /* The hyp was enabled before suspend. */
2048                         cpu_hyp_reinit();
2049
2050                 return NOTIFY_OK;
2051
2052         default:
2053                 return NOTIFY_DONE;
2054         }
2055 }
2056
2057 static struct notifier_block hyp_init_cpu_pm_nb = {
2058         .notifier_call = hyp_init_cpu_pm_notifier,
2059 };
2060
2061 static void __init hyp_cpu_pm_init(void)
2062 {
2063         if (!is_protected_kvm_enabled())
2064                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
2065 }
2066 static void __init hyp_cpu_pm_exit(void)
2067 {
2068         if (!is_protected_kvm_enabled())
2069                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
2070 }
2071 #else
2072 static inline void __init hyp_cpu_pm_init(void)
2073 {
2074 }
2075 static inline void __init hyp_cpu_pm_exit(void)
2076 {
2077 }
2078 #endif
2079
2080 static void __init init_cpu_logical_map(void)
2081 {
2082         unsigned int cpu;
2083
2084         /*
2085          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
2086          * Only copy the set of online CPUs whose features have been checked
2087          * against the finalized system capabilities. The hypervisor will not
2088          * allow any other CPUs from the `possible` set to boot.
2089          */
2090         for_each_online_cpu(cpu)
2091                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
2092 }
2093
2094 #define init_psci_0_1_impl_state(config, what)  \
2095         config.psci_0_1_ ## what ## _implemented = psci_ops.what
2096
2097 static bool __init init_psci_relay(void)
2098 {
2099         /*
2100          * If PSCI has not been initialized, protected KVM cannot install
2101          * itself on newly booted CPUs.
2102          */
2103         if (!psci_ops.get_version) {
2104                 kvm_err("Cannot initialize protected mode without PSCI\n");
2105                 return false;
2106         }
2107
2108         kvm_host_psci_config.version = psci_ops.get_version();
2109         kvm_host_psci_config.smccc_version = arm_smccc_get_version();
2110
2111         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
2112                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
2113                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
2114                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
2115                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
2116                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
2117         }
2118         return true;
2119 }
2120
2121 static int __init init_subsystems(void)
2122 {
2123         int err = 0;
2124
2125         /*
2126          * Enable hardware so that subsystem initialisation can access EL2.
2127          */
2128         on_each_cpu(cpu_hyp_init, NULL, 1);
2129
2130         /*
2131          * Register CPU lower-power notifier
2132          */
2133         hyp_cpu_pm_init();
2134
2135         /*
2136          * Init HYP view of VGIC
2137          */
2138         err = kvm_vgic_hyp_init();
2139         switch (err) {
2140         case 0:
2141                 vgic_present = true;
2142                 break;
2143         case -ENODEV:
2144         case -ENXIO:
2145                 vgic_present = false;
2146                 err = 0;
2147                 break;
2148         default:
2149                 goto out;
2150         }
2151
2152         /*
2153          * Init HYP architected timer support
2154          */
2155         err = kvm_timer_hyp_init(vgic_present);
2156         if (err)
2157                 goto out;
2158
2159         kvm_register_perf_callbacks(NULL);
2160
2161 out:
2162         if (err)
2163                 hyp_cpu_pm_exit();
2164
2165         if (err || !is_protected_kvm_enabled())
2166                 on_each_cpu(cpu_hyp_uninit, NULL, 1);
2167
2168         return err;
2169 }
2170
2171 static void __init teardown_subsystems(void)
2172 {
2173         kvm_unregister_perf_callbacks();
2174         hyp_cpu_pm_exit();
2175 }
2176
2177 static void __init teardown_hyp_mode(void)
2178 {
2179         int cpu;
2180
2181         free_hyp_pgds();
2182         for_each_possible_cpu(cpu) {
2183                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
2184                 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order());
2185         }
2186 }
2187
2188 static int __init do_pkvm_init(u32 hyp_va_bits)
2189 {
2190         void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base));
2191         int ret;
2192
2193         preempt_disable();
2194         cpu_hyp_init_context();
2195         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
2196                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
2197                                 hyp_va_bits);
2198         cpu_hyp_init_features();
2199
2200         /*
2201          * The stub hypercalls are now disabled, so set our local flag to
2202          * prevent a later re-init attempt in kvm_arch_hardware_enable().
2203          */
2204         __this_cpu_write(kvm_hyp_initialized, 1);
2205         preempt_enable();
2206
2207         return ret;
2208 }
2209
2210 static u64 get_hyp_id_aa64pfr0_el1(void)
2211 {
2212         /*
2213          * Track whether the system isn't affected by spectre/meltdown in the
2214          * hypervisor's view of id_aa64pfr0_el1, used for protected VMs.
2215          * Although this is per-CPU, we make it global for simplicity, e.g., not
2216          * to have to worry about vcpu migration.
2217          *
2218          * Unlike for non-protected VMs, userspace cannot override this for
2219          * protected VMs.
2220          */
2221         u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
2222
2223         val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
2224                  ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));
2225
2226         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
2227                           arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED);
2228         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
2229                           arm64_get_meltdown_state() == SPECTRE_UNAFFECTED);
2230
2231         return val;
2232 }
2233
2234 static void kvm_hyp_init_symbols(void)
2235 {
2236         kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1();
2237         kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1);
2238         kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
2239         kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
2240         kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
2241         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
2242         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
2243         kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1);
2244         kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1);
2245         kvm_nvhe_sym(__icache_flags) = __icache_flags;
2246         kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits;
2247 }
2248
2249 static int __init kvm_hyp_init_protection(u32 hyp_va_bits)
2250 {
2251         void *addr = phys_to_virt(hyp_mem_base);
2252         int ret;
2253
2254         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
2255         if (ret)
2256                 return ret;
2257
2258         ret = do_pkvm_init(hyp_va_bits);
2259         if (ret)
2260                 return ret;
2261
2262         free_hyp_pgds();
2263
2264         return 0;
2265 }
2266
2267 static void pkvm_hyp_init_ptrauth(void)
2268 {
2269         struct kvm_cpu_context *hyp_ctxt;
2270         int cpu;
2271
2272         for_each_possible_cpu(cpu) {
2273                 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu);
2274                 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long();
2275                 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long();
2276                 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long();
2277                 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long();
2278                 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long();
2279                 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long();
2280                 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long();
2281                 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long();
2282                 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long();
2283                 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long();
2284         }
2285 }
2286
2287 /* Inits Hyp-mode on all online CPUs */
2288 static int __init init_hyp_mode(void)
2289 {
2290         u32 hyp_va_bits;
2291         int cpu;
2292         int err = -ENOMEM;
2293
2294         /*
2295          * The protected Hyp-mode cannot be initialized if the memory pool
2296          * allocation has failed.
2297          */
2298         if (is_protected_kvm_enabled() && !hyp_mem_base)
2299                 goto out_err;
2300
2301         /*
2302          * Allocate Hyp PGD and setup Hyp identity mapping
2303          */
2304         err = kvm_mmu_init(&hyp_va_bits);
2305         if (err)
2306                 goto out_err;
2307
2308         /*
2309          * Allocate stack pages for Hypervisor-mode
2310          */
2311         for_each_possible_cpu(cpu) {
2312                 unsigned long stack_page;
2313
2314                 stack_page = __get_free_page(GFP_KERNEL);
2315                 if (!stack_page) {
2316                         err = -ENOMEM;
2317                         goto out_err;
2318                 }
2319
2320                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
2321         }
2322
2323         /*
2324          * Allocate and initialize pages for Hypervisor-mode percpu regions.
2325          */
2326         for_each_possible_cpu(cpu) {
2327                 struct page *page;
2328                 void *page_addr;
2329
2330                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
2331                 if (!page) {
2332                         err = -ENOMEM;
2333                         goto out_err;
2334                 }
2335
2336                 page_addr = page_address(page);
2337                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
2338                 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr;
2339         }
2340
2341         /*
2342          * Map the Hyp-code called directly from the host
2343          */
2344         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
2345                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
2346         if (err) {
2347                 kvm_err("Cannot map world-switch code\n");
2348                 goto out_err;
2349         }
2350
2351         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
2352                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
2353         if (err) {
2354                 kvm_err("Cannot map .hyp.rodata section\n");
2355                 goto out_err;
2356         }
2357
2358         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
2359                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
2360         if (err) {
2361                 kvm_err("Cannot map rodata section\n");
2362                 goto out_err;
2363         }
2364
2365         /*
2366          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
2367          * section thanks to an assertion in the linker script. Map it RW and
2368          * the rest of .bss RO.
2369          */
2370         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
2371                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
2372         if (err) {
2373                 kvm_err("Cannot map hyp bss section: %d\n", err);
2374                 goto out_err;
2375         }
2376
2377         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
2378                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
2379         if (err) {
2380                 kvm_err("Cannot map bss section\n");
2381                 goto out_err;
2382         }
2383
2384         /*
2385          * Map the Hyp stack pages
2386          */
2387         for_each_possible_cpu(cpu) {
2388                 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2389                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
2390
2391                 err = create_hyp_stack(__pa(stack_page), &params->stack_hyp_va);
2392                 if (err) {
2393                         kvm_err("Cannot map hyp stack\n");
2394                         goto out_err;
2395                 }
2396
2397                 /*
2398                  * Save the stack PA in nvhe_init_params. This will be needed
2399                  * to recreate the stack mapping in protected nVHE mode.
2400                  * __hyp_pa() won't do the right thing there, since the stack
2401                  * has been mapped in the flexible private VA space.
2402                  */
2403                 params->stack_pa = __pa(stack_page);
2404         }
2405
2406         for_each_possible_cpu(cpu) {
2407                 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu];
2408                 char *percpu_end = percpu_begin + nvhe_percpu_size();
2409
2410                 /* Map Hyp percpu pages */
2411                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
2412                 if (err) {
2413                         kvm_err("Cannot map hyp percpu region\n");
2414                         goto out_err;
2415                 }
2416
2417                 /* Prepare the CPU initialization parameters */
2418                 cpu_prepare_hyp_mode(cpu, hyp_va_bits);
2419         }
2420
2421         kvm_hyp_init_symbols();
2422
2423         if (is_protected_kvm_enabled()) {
2424                 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) &&
2425                     cpus_have_final_cap(ARM64_HAS_ADDRESS_AUTH))
2426                         pkvm_hyp_init_ptrauth();
2427
2428                 init_cpu_logical_map();
2429
2430                 if (!init_psci_relay()) {
2431                         err = -ENODEV;
2432                         goto out_err;
2433                 }
2434
2435                 err = kvm_hyp_init_protection(hyp_va_bits);
2436                 if (err) {
2437                         kvm_err("Failed to init hyp memory protection\n");
2438                         goto out_err;
2439                 }
2440         }
2441
2442         return 0;
2443
2444 out_err:
2445         teardown_hyp_mode();
2446         kvm_err("error initializing Hyp mode: %d\n", err);
2447         return err;
2448 }
2449
2450 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2451 {
2452         struct kvm_vcpu *vcpu;
2453         unsigned long i;
2454
2455         mpidr &= MPIDR_HWID_BITMASK;
2456
2457         if (kvm->arch.mpidr_data) {
2458                 u16 idx = kvm_mpidr_index(kvm->arch.mpidr_data, mpidr);
2459
2460                 vcpu = kvm_get_vcpu(kvm,
2461                                     kvm->arch.mpidr_data->cmpidr_to_idx[idx]);
2462                 if (mpidr != kvm_vcpu_get_mpidr_aff(vcpu))
2463                         vcpu = NULL;
2464
2465                 return vcpu;
2466         }
2467
2468         kvm_for_each_vcpu(i, vcpu, kvm) {
2469                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2470                         return vcpu;
2471         }
2472         return NULL;
2473 }
2474
2475 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm)
2476 {
2477         return irqchip_in_kernel(kvm);
2478 }
2479
2480 bool kvm_arch_has_irq_bypass(void)
2481 {
2482         return true;
2483 }
2484
2485 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2486                                       struct irq_bypass_producer *prod)
2487 {
2488         struct kvm_kernel_irqfd *irqfd =
2489                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2490
2491         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2492                                           &irqfd->irq_entry);
2493 }
2494 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2495                                       struct irq_bypass_producer *prod)
2496 {
2497         struct kvm_kernel_irqfd *irqfd =
2498                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2499
2500         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2501                                      &irqfd->irq_entry);
2502 }
2503
2504 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2505 {
2506         struct kvm_kernel_irqfd *irqfd =
2507                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2508
2509         kvm_arm_halt_guest(irqfd->kvm);
2510 }
2511
2512 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2513 {
2514         struct kvm_kernel_irqfd *irqfd =
2515                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2516
2517         kvm_arm_resume_guest(irqfd->kvm);
2518 }
2519
2520 /* Initialize Hyp-mode and memory mappings on all CPUs */
2521 static __init int kvm_arm_init(void)
2522 {
2523         int err;
2524         bool in_hyp_mode;
2525
2526         if (!is_hyp_mode_available()) {
2527                 kvm_info("HYP mode not available\n");
2528                 return -ENODEV;
2529         }
2530
2531         if (kvm_get_mode() == KVM_MODE_NONE) {
2532                 kvm_info("KVM disabled from command line\n");
2533                 return -ENODEV;
2534         }
2535
2536         err = kvm_sys_reg_table_init();
2537         if (err) {
2538                 kvm_info("Error initializing system register tables");
2539                 return err;
2540         }
2541
2542         in_hyp_mode = is_kernel_in_hyp_mode();
2543
2544         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2545             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2546                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2547                          "Only trusted guests should be used on this system.\n");
2548
2549         err = kvm_set_ipa_limit();
2550         if (err)
2551                 return err;
2552
2553         err = kvm_arm_init_sve();
2554         if (err)
2555                 return err;
2556
2557         err = kvm_arm_vmid_alloc_init();
2558         if (err) {
2559                 kvm_err("Failed to initialize VMID allocator.\n");
2560                 return err;
2561         }
2562
2563         if (!in_hyp_mode) {
2564                 err = init_hyp_mode();
2565                 if (err)
2566                         goto out_err;
2567         }
2568
2569         err = kvm_init_vector_slots();
2570         if (err) {
2571                 kvm_err("Cannot initialise vector slots\n");
2572                 goto out_hyp;
2573         }
2574
2575         err = init_subsystems();
2576         if (err)
2577                 goto out_hyp;
2578
2579         if (is_protected_kvm_enabled()) {
2580                 kvm_info("Protected nVHE mode initialized successfully\n");
2581         } else if (in_hyp_mode) {
2582                 kvm_info("VHE mode initialized successfully\n");
2583         } else {
2584                 kvm_info("Hyp mode initialized successfully\n");
2585         }
2586
2587         /*
2588          * FIXME: Do something reasonable if kvm_init() fails after pKVM
2589          * hypervisor protection is finalized.
2590          */
2591         err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2592         if (err)
2593                 goto out_subs;
2594
2595         kvm_arm_initialised = true;
2596
2597         return 0;
2598
2599 out_subs:
2600         teardown_subsystems();
2601 out_hyp:
2602         if (!in_hyp_mode)
2603                 teardown_hyp_mode();
2604 out_err:
2605         kvm_arm_vmid_alloc_free();
2606         return err;
2607 }
2608
2609 static int __init early_kvm_mode_cfg(char *arg)
2610 {
2611         if (!arg)
2612                 return -EINVAL;
2613
2614         if (strcmp(arg, "none") == 0) {
2615                 kvm_mode = KVM_MODE_NONE;
2616                 return 0;
2617         }
2618
2619         if (!is_hyp_mode_available()) {
2620                 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n");
2621                 return 0;
2622         }
2623
2624         if (strcmp(arg, "protected") == 0) {
2625                 if (!is_kernel_in_hyp_mode())
2626                         kvm_mode = KVM_MODE_PROTECTED;
2627                 else
2628                         pr_warn_once("Protected KVM not available with VHE\n");
2629
2630                 return 0;
2631         }
2632
2633         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) {
2634                 kvm_mode = KVM_MODE_DEFAULT;
2635                 return 0;
2636         }
2637
2638         if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) {
2639                 kvm_mode = KVM_MODE_NV;
2640                 return 0;
2641         }
2642
2643         return -EINVAL;
2644 }
2645 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2646
2647 enum kvm_mode kvm_get_mode(void)
2648 {
2649         return kvm_mode;
2650 }
2651
2652 module_init(kvm_arm_init);