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