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_pkvm.h>
  40 #include <asm/kvm_emulate.h>
  41 #include <asm/sections.h>
  42
  43 #include <kvm/arm_hypercalls.h>
  44 #include <kvm/arm_pmu.h>
  45 #include <kvm/arm_psci.h>
  46
  47 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
  48
  49 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
  50
  51 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  52 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
  53
  54 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt);
  55
  56 static bool vgic_present, kvm_arm_initialised;
  57
  58 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled);
  59 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  60
  61 bool is_kvm_arm_initialised(void)
  62 {
  63         return kvm_arm_initialised;
  64 }
  65
  66 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  67 {
  68         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  69 }
  70
  71 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  72                             struct kvm_enable_cap *cap)
  73 {
  74         int r;
  75         u64 new_cap;
  76
  77         if (cap->flags)
  78                 return -EINVAL;
  79
  80         switch (cap->cap) {
  81         case KVM_CAP_ARM_NISV_TO_USER:
  82                 r = 0;
  83                 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER,
  84                         &kvm->arch.flags);
  85                 break;
  86         case KVM_CAP_ARM_MTE:
  87                 mutex_lock(&kvm->lock);
  88                 if (!system_supports_mte() || kvm->created_vcpus) {
  89                         r = -EINVAL;
  90                 } else {
  91                         r = 0;
  92                         set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags);
  93                 }
  94                 mutex_unlock(&kvm->lock);
  95                 break;
  96         case KVM_CAP_ARM_SYSTEM_SUSPEND:
  97                 r = 0;
  98                 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags);
  99                 break;
 100         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 101                 new_cap = cap->args[0];
 102
 103                 mutex_lock(&kvm->slots_lock);
 104                 /*
 105                  * To keep things simple, allow changing the chunk
 106                  * size only when no memory slots have been created.
 107                  */
 108                 if (!kvm_are_all_memslots_empty(kvm)) {
 109                         r = -EINVAL;
 110                 } else if (new_cap && !kvm_is_block_size_supported(new_cap)) {
 111                         r = -EINVAL;
 112                 } else {
 113                         r = 0;
 114                         kvm->arch.mmu.split_page_chunk_size = new_cap;
 115                 }
 116                 mutex_unlock(&kvm->slots_lock);
 117                 break;
 118         default:
 119                 r = -EINVAL;
 120                 break;
 121         }
 122
 123         return r;
 124 }
 125
 126 static int kvm_arm_default_max_vcpus(void)
 127 {
 128         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 129 }
 130
 131 /**
 132  * kvm_arch_init_vm - initializes a VM data structure
 133  * @kvm:        pointer to the KVM struct
 134  */
 135 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 136 {
 137         int ret;
 138
 139         mutex_init(&kvm->arch.config_lock);
 140
 141 #ifdef CONFIG_LOCKDEP
 142         /* Clue in lockdep that the config_lock must be taken inside kvm->lock */
 143         mutex_lock(&kvm->lock);
 144         mutex_lock(&kvm->arch.config_lock);
 145         mutex_unlock(&kvm->arch.config_lock);
 146         mutex_unlock(&kvm->lock);
 147 #endif
 148
 149         ret = kvm_share_hyp(kvm, kvm + 1);
 150         if (ret)
 151                 return ret;
 152
 153         ret = pkvm_init_host_vm(kvm);
 154         if (ret)
 155                 goto err_unshare_kvm;
 156
 157         if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) {
 158                 ret = -ENOMEM;
 159                 goto err_unshare_kvm;
 160         }
 161         cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask);
 162
 163         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type);
 164         if (ret)
 165                 goto err_free_cpumask;
 166
 167         kvm_vgic_early_init(kvm);
 168
 169         kvm_timer_init_vm(kvm);
 170
 171         /* The maximum number of VCPUs is limited by the host's GIC model */
 172         kvm->max_vcpus = kvm_arm_default_max_vcpus();
 173
 174         kvm_arm_init_hypercalls(kvm);
 175
 176         bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES);
 177
 178         return 0;
 179
 180 err_free_cpumask:
 181         free_cpumask_var(kvm->arch.supported_cpus);
 182 err_unshare_kvm:
 183         kvm_unshare_hyp(kvm, kvm + 1);
 184         return ret;
 185 }
 186
 187 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 188 {
 189         return VM_FAULT_SIGBUS;
 190 }
 191
 192
 193 /**
 194  * kvm_arch_destroy_vm - destroy the VM data structure
 195  * @kvm:        pointer to the KVM struct
 196  */
 197 void kvm_arch_destroy_vm(struct kvm *kvm)
 198 {
 199         bitmap_free(kvm->arch.pmu_filter);
 200         free_cpumask_var(kvm->arch.supported_cpus);
 201
 202         kvm_vgic_destroy(kvm);
 203
 204         if (is_protected_kvm_enabled())
 205                 pkvm_destroy_hyp_vm(kvm);
 206
 207         kvm_destroy_vcpus(kvm);
 208
 209         kvm_unshare_hyp(kvm, kvm + 1);
 210
 211         kvm_arm_teardown_hypercalls(kvm);
 212 }
 213
 214 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 215 {
 216         int r;
 217         switch (ext) {
 218         case KVM_CAP_IRQCHIP:
 219                 r = vgic_present;
 220                 break;
 221         case KVM_CAP_IOEVENTFD:
 222         case KVM_CAP_DEVICE_CTRL:
 223         case KVM_CAP_USER_MEMORY:
 224         case KVM_CAP_SYNC_MMU:
 225         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 226         case KVM_CAP_ONE_REG:
 227         case KVM_CAP_ARM_PSCI:
 228         case KVM_CAP_ARM_PSCI_0_2:
 229         case KVM_CAP_READONLY_MEM:
 230         case KVM_CAP_MP_STATE:
 231         case KVM_CAP_IMMEDIATE_EXIT:
 232         case KVM_CAP_VCPU_EVENTS:
 233         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 234         case KVM_CAP_ARM_NISV_TO_USER:
 235         case KVM_CAP_ARM_INJECT_EXT_DABT:
 236         case KVM_CAP_SET_GUEST_DEBUG:
 237         case KVM_CAP_VCPU_ATTRIBUTES:
 238         case KVM_CAP_PTP_KVM:
 239         case KVM_CAP_ARM_SYSTEM_SUSPEND:
 240         case KVM_CAP_IRQFD_RESAMPLE:
 241         case KVM_CAP_COUNTER_OFFSET:
 242                 r = 1;
 243                 break;
 244         case KVM_CAP_SET_GUEST_DEBUG2:
 245                 return KVM_GUESTDBG_VALID_MASK;
 246         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 247                 r = 1;
 248                 break;
 249         case KVM_CAP_NR_VCPUS:
 250                 /*
 251                  * ARM64 treats KVM_CAP_NR_CPUS differently from all other
 252                  * architectures, as it does not always bound it to
 253                  * KVM_CAP_MAX_VCPUS. It should not matter much because
 254                  * this is just an advisory value.
 255                  */
 256                 r = min_t(unsigned int, num_online_cpus(),
 257                           kvm_arm_default_max_vcpus());
 258                 break;
 259         case KVM_CAP_MAX_VCPUS:
 260         case KVM_CAP_MAX_VCPU_ID:
 261                 if (kvm)
 262                         r = kvm->max_vcpus;
 263                 else
 264                         r = kvm_arm_default_max_vcpus();
 265                 break;
 266         case KVM_CAP_MSI_DEVID:
 267                 if (!kvm)
 268                         r = -EINVAL;
 269                 else
 270                         r = kvm->arch.vgic.msis_require_devid;
 271                 break;
 272         case KVM_CAP_ARM_USER_IRQ:
 273                 /*
 274                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 275                  * (bump this number if adding more devices)
 276                  */
 277                 r = 1;
 278                 break;
 279         case KVM_CAP_ARM_MTE:
 280                 r = system_supports_mte();
 281                 break;
 282         case KVM_CAP_STEAL_TIME:
 283                 r = kvm_arm_pvtime_supported();
 284                 break;
 285         case KVM_CAP_ARM_EL1_32BIT:
 286                 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1);
 287                 break;
 288         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 289                 r = get_num_brps();
 290                 break;
 291         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 292                 r = get_num_wrps();
 293                 break;
 294         case KVM_CAP_ARM_PMU_V3:
 295                 r = kvm_arm_support_pmu_v3();
 296                 break;
 297         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 298                 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN);
 299                 break;
 300         case KVM_CAP_ARM_VM_IPA_SIZE:
 301                 r = get_kvm_ipa_limit();
 302                 break;
 303         case KVM_CAP_ARM_SVE:
 304                 r = system_supports_sve();
 305                 break;
 306         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 307         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 308                 r = system_has_full_ptr_auth();
 309                 break;
 310         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 311                 if (kvm)
 312                         r = kvm->arch.mmu.split_page_chunk_size;
 313                 else
 314                         r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
 315                 break;
 316         case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES:
 317                 r = kvm_supported_block_sizes();
 318                 break;
 319         default:
 320                 r = 0;
 321         }
 322
 323         return r;
 324 }
 325
 326 long kvm_arch_dev_ioctl(struct file *filp,
 327                         unsigned int ioctl, unsigned long arg)
 328 {
 329         return -EINVAL;
 330 }
 331
 332 struct kvm *kvm_arch_alloc_vm(void)
 333 {
 334         size_t sz = sizeof(struct kvm);
 335
 336         if (!has_vhe())
 337                 return kzalloc(sz, GFP_KERNEL_ACCOUNT);
 338
 339         return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO);
 340 }
 341
 342 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 343 {
 344         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 345                 return -EBUSY;
 346
 347         if (id >= kvm->max_vcpus)
 348                 return -EINVAL;
 349
 350         return 0;
 351 }
 352
 353 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 354 {
 355         int err;
 356
 357         spin_lock_init(&vcpu->arch.mp_state_lock);
 358
 359 #ifdef CONFIG_LOCKDEP
 360         /* Inform lockdep that the config_lock is acquired after vcpu->mutex */
 361         mutex_lock(&vcpu->mutex);
 362         mutex_lock(&vcpu->kvm->arch.config_lock);
 363         mutex_unlock(&vcpu->kvm->arch.config_lock);
 364         mutex_unlock(&vcpu->mutex);
 365 #endif
 366
 367         /* Force users to call KVM_ARM_VCPU_INIT */
 368         vcpu->arch.target = -1;
 369         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 370
 371         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 372
 373         /*
 374          * Default value for the FP state, will be overloaded at load
 375          * time if we support FP (pretty likely)
 376          */
 377         vcpu->arch.fp_state = FP_STATE_FREE;
 378
 379         /* Set up the timer */
 380         kvm_timer_vcpu_init(vcpu);
 381
 382         kvm_pmu_vcpu_init(vcpu);
 383
 384         kvm_arm_reset_debug_ptr(vcpu);
 385
 386         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 387
 388         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 389
 390         err = kvm_vgic_vcpu_init(vcpu);
 391         if (err)
 392                 return err;
 393
 394         return kvm_share_hyp(vcpu, vcpu + 1);
 395 }
 396
 397 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 398 {
 399 }
 400
 401 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 402 {
 403         if (vcpu_has_run_once(vcpu) && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 404                 static_branch_dec(&userspace_irqchip_in_use);
 405
 406         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 407         kvm_timer_vcpu_terminate(vcpu);
 408         kvm_pmu_vcpu_destroy(vcpu);
 409
 410         kvm_arm_vcpu_destroy(vcpu);
 411 }
 412
 413 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 414 {
 415
 416 }
 417
 418 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 419 {
 420
 421 }
 422
 423 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 424 {
 425         struct kvm_s2_mmu *mmu;
 426         int *last_ran;
 427
 428         mmu = vcpu->arch.hw_mmu;
 429         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 430
 431         /*
 432          * We guarantee that both TLBs and I-cache are private to each
 433          * vcpu. If detecting that a vcpu from the same VM has
 434          * previously run on the same physical CPU, call into the
 435          * hypervisor code to nuke the relevant contexts.
 436          *
 437          * We might get preempted before the vCPU actually runs, but
 438          * over-invalidation doesn't affect correctness.
 439          */
 440         if (*last_ran != vcpu->vcpu_id) {
 441                 kvm_call_hyp(__kvm_flush_cpu_context, mmu);
 442                 *last_ran = vcpu->vcpu_id;
 443         }
 444
 445         vcpu->cpu = cpu;
 446
 447         kvm_vgic_load(vcpu);
 448         kvm_timer_vcpu_load(vcpu);
 449         if (has_vhe())
 450                 kvm_vcpu_load_sysregs_vhe(vcpu);
 451         kvm_arch_vcpu_load_fp(vcpu);
 452         kvm_vcpu_pmu_restore_guest(vcpu);
 453         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 454                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 455
 456         if (single_task_running())
 457                 vcpu_clear_wfx_traps(vcpu);
 458         else
 459                 vcpu_set_wfx_traps(vcpu);
 460
 461         if (vcpu_has_ptrauth(vcpu))
 462                 vcpu_ptrauth_disable(vcpu);
 463         kvm_arch_vcpu_load_debug_state_flags(vcpu);
 464
 465         if (!cpumask_test_cpu(smp_processor_id(), vcpu->kvm->arch.supported_cpus))
 466                 vcpu_set_on_unsupported_cpu(vcpu);
 467 }
 468
 469 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 470 {
 471         kvm_arch_vcpu_put_debug_state_flags(vcpu);
 472         kvm_arch_vcpu_put_fp(vcpu);
 473         if (has_vhe())
 474                 kvm_vcpu_put_sysregs_vhe(vcpu);
 475         kvm_timer_vcpu_put(vcpu);
 476         kvm_vgic_put(vcpu);
 477         kvm_vcpu_pmu_restore_host(vcpu);
 478         kvm_arm_vmid_clear_active();
 479
 480         vcpu_clear_on_unsupported_cpu(vcpu);
 481         vcpu->cpu = -1;
 482 }
 483
 484 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 485 {
 486         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED);
 487         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 488         kvm_vcpu_kick(vcpu);
 489 }
 490
 491 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 492 {
 493         spin_lock(&vcpu->arch.mp_state_lock);
 494         __kvm_arm_vcpu_power_off(vcpu);
 495         spin_unlock(&vcpu->arch.mp_state_lock);
 496 }
 497
 498 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu)
 499 {
 500         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED;
 501 }
 502
 503 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu)
 504 {
 505         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED);
 506         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 507         kvm_vcpu_kick(vcpu);
 508 }
 509
 510 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu)
 511 {
 512         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED;
 513 }
 514
 515 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 516                                     struct kvm_mp_state *mp_state)
 517 {
 518         *mp_state = READ_ONCE(vcpu->arch.mp_state);
 519
 520         return 0;
 521 }
 522
 523 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 524                                     struct kvm_mp_state *mp_state)
 525 {
 526         int ret = 0;
 527
 528         spin_lock(&vcpu->arch.mp_state_lock);
 529
 530         switch (mp_state->mp_state) {
 531         case KVM_MP_STATE_RUNNABLE:
 532                 WRITE_ONCE(vcpu->arch.mp_state, *mp_state);
 533                 break;
 534         case KVM_MP_STATE_STOPPED:
 535                 __kvm_arm_vcpu_power_off(vcpu);
 536                 break;
 537         case KVM_MP_STATE_SUSPENDED:
 538                 kvm_arm_vcpu_suspend(vcpu);
 539                 break;
 540         default:
 541                 ret = -EINVAL;
 542         }
 543
 544         spin_unlock(&vcpu->arch.mp_state_lock);
 545
 546         return ret;
 547 }
 548
 549 /**
 550  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 551  * @v:          The VCPU pointer
 552  *
 553  * If the guest CPU is not waiting for interrupts or an interrupt line is
 554  * asserted, the CPU is by definition runnable.
 555  */
 556 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 557 {
 558         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 559         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 560                 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause);
 561 }
 562
 563 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 564 {
 565         return vcpu_mode_priv(vcpu);
 566 }
 567
 568 #ifdef CONFIG_GUEST_PERF_EVENTS
 569 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu)
 570 {
 571         return *vcpu_pc(vcpu);
 572 }
 573 #endif
 574
 575 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 576 {
 577         return vcpu->arch.target >= 0;
 578 }
 579
 580 /*
 581  * Handle both the initialisation that is being done when the vcpu is
 582  * run for the first time, as well as the updates that must be
 583  * performed each time we get a new thread dealing with this vcpu.
 584  */
 585 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
 586 {
 587         struct kvm *kvm = vcpu->kvm;
 588         int ret;
 589
 590         if (!kvm_vcpu_initialized(vcpu))
 591                 return -ENOEXEC;
 592
 593         if (!kvm_arm_vcpu_is_finalized(vcpu))
 594                 return -EPERM;
 595
 596         ret = kvm_arch_vcpu_run_map_fp(vcpu);
 597         if (ret)
 598                 return ret;
 599
 600         if (likely(vcpu_has_run_once(vcpu)))
 601                 return 0;
 602
 603         kvm_arm_vcpu_init_debug(vcpu);
 604
 605         if (likely(irqchip_in_kernel(kvm))) {
 606                 /*
 607                  * Map the VGIC hardware resources before running a vcpu the
 608                  * first time on this VM.
 609                  */
 610                 ret = kvm_vgic_map_resources(kvm);
 611                 if (ret)
 612                         return ret;
 613         }
 614
 615         ret = kvm_timer_enable(vcpu);
 616         if (ret)
 617                 return ret;
 618
 619         ret = kvm_arm_pmu_v3_enable(vcpu);
 620         if (ret)
 621                 return ret;
 622
 623         if (is_protected_kvm_enabled()) {
 624                 ret = pkvm_create_hyp_vm(kvm);
 625                 if (ret)
 626                         return ret;
 627         }
 628
 629         if (!irqchip_in_kernel(kvm)) {
 630                 /*
 631                  * Tell the rest of the code that there are userspace irqchip
 632                  * VMs in the wild.
 633                  */
 634                 static_branch_inc(&userspace_irqchip_in_use);
 635         }
 636
 637         /*
 638          * Initialize traps for protected VMs.
 639          * NOTE: Move to run in EL2 directly, rather than via a hypercall, once
 640          * the code is in place for first run initialization at EL2.
 641          */
 642         if (kvm_vm_is_protected(kvm))
 643                 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu);
 644
 645         mutex_lock(&kvm->arch.config_lock);
 646         set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags);
 647         mutex_unlock(&kvm->arch.config_lock);
 648
 649         return ret;
 650 }
 651
 652 bool kvm_arch_intc_initialized(struct kvm *kvm)
 653 {
 654         return vgic_initialized(kvm);
 655 }
 656
 657 void kvm_arm_halt_guest(struct kvm *kvm)
 658 {
 659         unsigned long i;
 660         struct kvm_vcpu *vcpu;
 661
 662         kvm_for_each_vcpu(i, vcpu, kvm)
 663                 vcpu->arch.pause = true;
 664         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 665 }
 666
 667 void kvm_arm_resume_guest(struct kvm *kvm)
 668 {
 669         unsigned long i;
 670         struct kvm_vcpu *vcpu;
 671
 672         kvm_for_each_vcpu(i, vcpu, kvm) {
 673                 vcpu->arch.pause = false;
 674                 __kvm_vcpu_wake_up(vcpu);
 675         }
 676 }
 677
 678 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu)
 679 {
 680         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 681
 682         rcuwait_wait_event(wait,
 683                            (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause),
 684                            TASK_INTERRUPTIBLE);
 685
 686         if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) {
 687                 /* Awaken to handle a signal, request we sleep again later. */
 688                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 689         }
 690
 691         /*
 692          * Make sure we will observe a potential reset request if we've
 693          * observed a change to the power state. Pairs with the smp_wmb() in
 694          * kvm_psci_vcpu_on().
 695          */
 696         smp_rmb();
 697 }
 698
 699 /**
 700  * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior
 701  * @vcpu:       The VCPU pointer
 702  *
 703  * Suspend execution of a vCPU until a valid wake event is detected, i.e. until
 704  * the vCPU is runnable.  The vCPU may or may not be scheduled out, depending
 705  * on when a wake event arrives, e.g. there may already be a pending wake event.
 706  */
 707 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu)
 708 {
 709         /*
 710          * Sync back the state of the GIC CPU interface so that we have
 711          * the latest PMR and group enables. This ensures that
 712          * kvm_arch_vcpu_runnable has up-to-date data to decide whether
 713          * we have pending interrupts, e.g. when determining if the
 714          * vCPU should block.
 715          *
 716          * For the same reason, we want to tell GICv4 that we need
 717          * doorbells to be signalled, should an interrupt become pending.
 718          */
 719         preempt_disable();
 720         kvm_vgic_vmcr_sync(vcpu);
 721         vgic_v4_put(vcpu, true);
 722         preempt_enable();
 723
 724         kvm_vcpu_halt(vcpu);
 725         vcpu_clear_flag(vcpu, IN_WFIT);
 726
 727         preempt_disable();
 728         vgic_v4_load(vcpu);
 729         preempt_enable();
 730 }
 731
 732 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu)
 733 {
 734         if (!kvm_arm_vcpu_suspended(vcpu))
 735                 return 1;
 736
 737         kvm_vcpu_wfi(vcpu);
 738
 739         /*
 740          * The suspend state is sticky; we do not leave it until userspace
 741          * explicitly marks the vCPU as runnable. Request that we suspend again
 742          * later.
 743          */
 744         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 745
 746         /*
 747          * Check to make sure the vCPU is actually runnable. If so, exit to
 748          * userspace informing it of the wakeup condition.
 749          */
 750         if (kvm_arch_vcpu_runnable(vcpu)) {
 751                 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event));
 752                 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP;
 753                 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
 754                 return 0;
 755         }
 756
 757         /*
 758          * Otherwise, we were unblocked to process a different event, such as a
 759          * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to
 760          * process the event.
 761          */
 762         return 1;
 763 }
 764
 765 /**
 766  * check_vcpu_requests - check and handle pending vCPU requests
 767  * @vcpu:       the VCPU pointer
 768  *
 769  * Return: 1 if we should enter the guest
 770  *         0 if we should exit to userspace
 771  *         < 0 if we should exit to userspace, where the return value indicates
 772  *         an error
 773  */
 774 static int check_vcpu_requests(struct kvm_vcpu *vcpu)
 775 {
 776         if (kvm_request_pending(vcpu)) {
 777                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 778                         kvm_vcpu_sleep(vcpu);
 779
 780                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 781                         kvm_reset_vcpu(vcpu);
 782
 783                 /*
 784                  * Clear IRQ_PENDING requests that were made to guarantee
 785                  * that a VCPU sees new virtual interrupts.
 786                  */
 787                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 788
 789                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 790                         kvm_update_stolen_time(vcpu);
 791
 792                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 793                         /* The distributor enable bits were changed */
 794                         preempt_disable();
 795                         vgic_v4_put(vcpu, false);
 796                         vgic_v4_load(vcpu);
 797                         preempt_enable();
 798                 }
 799
 800                 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
 801                         kvm_pmu_handle_pmcr(vcpu,
 802                                             __vcpu_sys_reg(vcpu, PMCR_EL0));
 803
 804                 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu))
 805                         return kvm_vcpu_suspend(vcpu);
 806
 807                 if (kvm_dirty_ring_check_request(vcpu))
 808                         return 0;
 809         }
 810
 811         return 1;
 812 }
 813
 814 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 815 {
 816         if (likely(!vcpu_mode_is_32bit(vcpu)))
 817                 return false;
 818
 819         return !kvm_supports_32bit_el0();
 820 }
 821
 822 /**
 823  * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest
 824  * @vcpu:       The VCPU pointer
 825  * @ret:        Pointer to write optional return code
 826  *
 827  * Returns: true if the VCPU needs to return to a preemptible + interruptible
 828  *          and skip guest entry.
 829  *
 830  * This function disambiguates between two different types of exits: exits to a
 831  * preemptible + interruptible kernel context and exits to userspace. For an
 832  * exit to userspace, this function will write the return code to ret and return
 833  * true. For an exit to preemptible + interruptible kernel context (i.e. check
 834  * for pending work and re-enter), return true without writing to ret.
 835  */
 836 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret)
 837 {
 838         struct kvm_run *run = vcpu->run;
 839
 840         /*
 841          * If we're using a userspace irqchip, then check if we need
 842          * to tell a userspace irqchip about timer or PMU level
 843          * changes and if so, exit to userspace (the actual level
 844          * state gets updated in kvm_timer_update_run and
 845          * kvm_pmu_update_run below).
 846          */
 847         if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 848                 if (kvm_timer_should_notify_user(vcpu) ||
 849                     kvm_pmu_should_notify_user(vcpu)) {
 850                         *ret = -EINTR;
 851                         run->exit_reason = KVM_EXIT_INTR;
 852                         return true;
 853                 }
 854         }
 855
 856         if (unlikely(vcpu_on_unsupported_cpu(vcpu))) {
 857                 run->exit_reason = KVM_EXIT_FAIL_ENTRY;
 858                 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED;
 859                 run->fail_entry.cpu = smp_processor_id();
 860                 *ret = 0;
 861                 return true;
 862         }
 863
 864         return kvm_request_pending(vcpu) ||
 865                         xfer_to_guest_mode_work_pending();
 866 }
 867
 868 /*
 869  * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while
 870  * the vCPU is running.
 871  *
 872  * This must be noinstr as instrumentation may make use of RCU, and this is not
 873  * safe during the EQS.
 874  */
 875 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
 876 {
 877         int ret;
 878
 879         guest_state_enter_irqoff();
 880         ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 881         guest_state_exit_irqoff();
 882
 883         return ret;
 884 }
 885
 886 /**
 887  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 888  * @vcpu:       The VCPU pointer
 889  *
 890  * This function is called through the VCPU_RUN ioctl called from user space. It
 891  * will execute VM code in a loop until the time slice for the process is used
 892  * or some emulation is needed from user space in which case the function will
 893  * return with return value 0 and with the kvm_run structure filled in with the
 894  * required data for the requested emulation.
 895  */
 896 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 897 {
 898         struct kvm_run *run = vcpu->run;
 899         int ret;
 900
 901         if (run->exit_reason == KVM_EXIT_MMIO) {
 902                 ret = kvm_handle_mmio_return(vcpu);
 903                 if (ret)
 904                         return ret;
 905         }
 906
 907         vcpu_load(vcpu);
 908
 909         if (run->immediate_exit) {
 910                 ret = -EINTR;
 911                 goto out;
 912         }
 913
 914         kvm_sigset_activate(vcpu);
 915
 916         ret = 1;
 917         run->exit_reason = KVM_EXIT_UNKNOWN;
 918         run->flags = 0;
 919         while (ret > 0) {
 920                 /*
 921                  * Check conditions before entering the guest
 922                  */
 923                 ret = xfer_to_guest_mode_handle_work(vcpu);
 924                 if (!ret)
 925                         ret = 1;
 926
 927                 if (ret > 0)
 928                         ret = check_vcpu_requests(vcpu);
 929
 930                 /*
 931                  * Preparing the interrupts to be injected also
 932                  * involves poking the GIC, which must be done in a
 933                  * non-preemptible context.
 934                  */
 935                 preempt_disable();
 936
 937                 /*
 938                  * The VMID allocator only tracks active VMIDs per
 939                  * physical CPU, and therefore the VMID allocated may not be
 940                  * preserved on VMID roll-over if the task was preempted,
 941                  * making a thread's VMID inactive. So we need to call
 942                  * kvm_arm_vmid_update() in non-premptible context.
 943                  */
 944                 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid);
 945
 946                 kvm_pmu_flush_hwstate(vcpu);
 947
 948                 local_irq_disable();
 949
 950                 kvm_vgic_flush_hwstate(vcpu);
 951
 952                 kvm_pmu_update_vcpu_events(vcpu);
 953
 954                 /*
 955                  * Ensure we set mode to IN_GUEST_MODE after we disable
 956                  * interrupts and before the final VCPU requests check.
 957                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 958                  * Documentation/virt/kvm/vcpu-requests.rst
 959                  */
 960                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 961
 962                 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) {
 963                         vcpu->mode = OUTSIDE_GUEST_MODE;
 964                         isb(); /* Ensure work in x_flush_hwstate is committed */
 965                         kvm_pmu_sync_hwstate(vcpu);
 966                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 967                                 kvm_timer_sync_user(vcpu);
 968                         kvm_vgic_sync_hwstate(vcpu);
 969                         local_irq_enable();
 970                         preempt_enable();
 971                         continue;
 972                 }
 973
 974                 kvm_arm_setup_debug(vcpu);
 975                 kvm_arch_vcpu_ctxflush_fp(vcpu);
 976
 977                 /**************************************************************
 978                  * Enter the guest
 979                  */
 980                 trace_kvm_entry(*vcpu_pc(vcpu));
 981                 guest_timing_enter_irqoff();
 982
 983                 ret = kvm_arm_vcpu_enter_exit(vcpu);
 984
 985                 vcpu->mode = OUTSIDE_GUEST_MODE;
 986                 vcpu->stat.exits++;
 987                 /*
 988                  * Back from guest
 989                  *************************************************************/
 990
 991                 kvm_arm_clear_debug(vcpu);
 992
 993                 /*
 994                  * We must sync the PMU state before the vgic state so
 995                  * that the vgic can properly sample the updated state of the
 996                  * interrupt line.
 997                  */
 998                 kvm_pmu_sync_hwstate(vcpu);
 999
1000                 /*
1001                  * Sync the vgic state before syncing the timer state because
1002                  * the timer code needs to know if the virtual timer
1003                  * interrupts are active.
1004                  */
1005                 kvm_vgic_sync_hwstate(vcpu);
1006
1007                 /*
1008                  * Sync the timer hardware state before enabling interrupts as
1009                  * we don't want vtimer interrupts to race with syncing the
1010                  * timer virtual interrupt state.
1011                  */
1012                 if (static_branch_unlikely(&userspace_irqchip_in_use))
1013                         kvm_timer_sync_user(vcpu);
1014
1015                 kvm_arch_vcpu_ctxsync_fp(vcpu);
1016
1017                 /*
1018                  * We must ensure that any pending interrupts are taken before
1019                  * we exit guest timing so that timer ticks are accounted as
1020                  * guest time. Transiently unmask interrupts so that any
1021                  * pending interrupts are taken.
1022                  *
1023                  * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other
1024                  * context synchronization event) is necessary to ensure that
1025                  * pending interrupts are taken.
1026                  */
1027                 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) {
1028                         local_irq_enable();
1029                         isb();
1030                         local_irq_disable();
1031                 }
1032
1033                 guest_timing_exit_irqoff();
1034
1035                 local_irq_enable();
1036
1037                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
1038
1039                 /* Exit types that need handling before we can be preempted */
1040                 handle_exit_early(vcpu, ret);
1041
1042                 preempt_enable();
1043
1044                 /*
1045                  * The ARMv8 architecture doesn't give the hypervisor
1046                  * a mechanism to prevent a guest from dropping to AArch32 EL0
1047                  * if implemented by the CPU. If we spot the guest in such
1048                  * state and that we decided it wasn't supposed to do so (like
1049                  * with the asymmetric AArch32 case), return to userspace with
1050                  * a fatal error.
1051                  */
1052                 if (vcpu_mode_is_bad_32bit(vcpu)) {
1053                         /*
1054                          * As we have caught the guest red-handed, decide that
1055                          * it isn't fit for purpose anymore by making the vcpu
1056                          * invalid. The VMM can try and fix it by issuing  a
1057                          * KVM_ARM_VCPU_INIT if it really wants to.
1058                          */
1059                         vcpu->arch.target = -1;
1060                         ret = ARM_EXCEPTION_IL;
1061                 }
1062
1063                 ret = handle_exit(vcpu, ret);
1064         }
1065
1066         /* Tell userspace about in-kernel device output levels */
1067         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1068                 kvm_timer_update_run(vcpu);
1069                 kvm_pmu_update_run(vcpu);
1070         }
1071
1072         kvm_sigset_deactivate(vcpu);
1073
1074 out:
1075         /*
1076          * In the unlikely event that we are returning to userspace
1077          * with pending exceptions or PC adjustment, commit these
1078          * adjustments in order to give userspace a consistent view of
1079          * the vcpu state. Note that this relies on __kvm_adjust_pc()
1080          * being preempt-safe on VHE.
1081          */
1082         if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) ||
1083                      vcpu_get_flag(vcpu, INCREMENT_PC)))
1084                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
1085
1086         vcpu_put(vcpu);
1087         return ret;
1088 }
1089
1090 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
1091 {
1092         int bit_index;
1093         bool set;
1094         unsigned long *hcr;
1095
1096         if (number == KVM_ARM_IRQ_CPU_IRQ)
1097                 bit_index = __ffs(HCR_VI);
1098         else /* KVM_ARM_IRQ_CPU_FIQ */
1099                 bit_index = __ffs(HCR_VF);
1100
1101         hcr = vcpu_hcr(vcpu);
1102         if (level)
1103                 set = test_and_set_bit(bit_index, hcr);
1104         else
1105                 set = test_and_clear_bit(bit_index, hcr);
1106
1107         /*
1108          * If we didn't change anything, no need to wake up or kick other CPUs
1109          */
1110         if (set == level)
1111                 return 0;
1112
1113         /*
1114          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
1115          * trigger a world-switch round on the running physical CPU to set the
1116          * virtual IRQ/FIQ fields in the HCR appropriately.
1117          */
1118         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
1119         kvm_vcpu_kick(vcpu);
1120
1121         return 0;
1122 }
1123
1124 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
1125                           bool line_status)
1126 {
1127         u32 irq = irq_level->irq;
1128         unsigned int irq_type, vcpu_idx, irq_num;
1129         int nrcpus = atomic_read(&kvm->online_vcpus);
1130         struct kvm_vcpu *vcpu = NULL;
1131         bool level = irq_level->level;
1132
1133         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
1134         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
1135         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
1136         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
1137
1138         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
1139
1140         switch (irq_type) {
1141         case KVM_ARM_IRQ_TYPE_CPU:
1142                 if (irqchip_in_kernel(kvm))
1143                         return -ENXIO;
1144
1145                 if (vcpu_idx >= nrcpus)
1146                         return -EINVAL;
1147
1148                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1149                 if (!vcpu)
1150                         return -EINVAL;
1151
1152                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1153                         return -EINVAL;
1154
1155                 return vcpu_interrupt_line(vcpu, irq_num, level);
1156         case KVM_ARM_IRQ_TYPE_PPI:
1157                 if (!irqchip_in_kernel(kvm))
1158                         return -ENXIO;
1159
1160                 if (vcpu_idx >= nrcpus)
1161                         return -EINVAL;
1162
1163                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1164                 if (!vcpu)
1165                         return -EINVAL;
1166
1167                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1168                         return -EINVAL;
1169
1170                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
1171         case KVM_ARM_IRQ_TYPE_SPI:
1172                 if (!irqchip_in_kernel(kvm))
1173                         return -ENXIO;
1174
1175                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1176                         return -EINVAL;
1177
1178                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
1179         }
1180
1181         return -EINVAL;
1182 }
1183
1184 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
1185                                         const struct kvm_vcpu_init *init)
1186 {
1187         unsigned long features = init->features[0];
1188         int i;
1189
1190         if (features & ~KVM_VCPU_VALID_FEATURES)
1191                 return -ENOENT;
1192
1193         for (i = 1; i < ARRAY_SIZE(init->features); i++) {
1194                 if (init->features[i])
1195                         return -ENOENT;
1196         }
1197
1198         if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
1199                 return 0;
1200
1201         if (!cpus_have_const_cap(ARM64_HAS_32BIT_EL1))
1202                 return -EINVAL;
1203
1204         /* MTE is incompatible with AArch32 */
1205         if (kvm_has_mte(vcpu->kvm))
1206                 return -EINVAL;
1207
1208         /* NV is incompatible with AArch32 */
1209         if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
1210                 return -EINVAL;
1211
1212         return 0;
1213 }
1214
1215 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu,
1216                                   const struct kvm_vcpu_init *init)
1217 {
1218         unsigned long features = init->features[0];
1219
1220         return !bitmap_equal(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES) ||
1221                         vcpu->arch.target != init->target;
1222 }
1223
1224 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1225                                  const struct kvm_vcpu_init *init)
1226 {
1227         unsigned long features = init->features[0];
1228         struct kvm *kvm = vcpu->kvm;
1229         int ret = -EINVAL;
1230
1231         mutex_lock(&kvm->arch.config_lock);
1232
1233         if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) &&
1234             !bitmap_equal(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES))
1235                 goto out_unlock;
1236
1237         vcpu->arch.target = init->target;
1238         bitmap_copy(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES);
1239
1240         /* Now we know what it is, we can reset it. */
1241         ret = kvm_reset_vcpu(vcpu);
1242         if (ret) {
1243                 vcpu->arch.target = -1;
1244                 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
1245                 goto out_unlock;
1246         }
1247
1248         bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES);
1249         set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags);
1250
1251 out_unlock:
1252         mutex_unlock(&kvm->arch.config_lock);
1253         return ret;
1254 }
1255
1256 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1257                                const struct kvm_vcpu_init *init)
1258 {
1259         int ret;
1260
1261         if (init->target != kvm_target_cpu())
1262                 return -EINVAL;
1263
1264         ret = kvm_vcpu_init_check_features(vcpu, init);
1265         if (ret)
1266                 return ret;
1267
1268         if (vcpu->arch.target == -1)
1269                 return __kvm_vcpu_set_target(vcpu, init);
1270
1271         if (kvm_vcpu_init_changed(vcpu, init))
1272                 return -EINVAL;
1273
1274         return kvm_reset_vcpu(vcpu);
1275 }
1276
1277 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1278                                          struct kvm_vcpu_init *init)
1279 {
1280         bool power_off = false;
1281         int ret;
1282
1283         /*
1284          * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid
1285          * reflecting it in the finalized feature set, thus limiting its scope
1286          * to a single KVM_ARM_VCPU_INIT call.
1287          */
1288         if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) {
1289                 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF);
1290                 power_off = true;
1291         }
1292
1293         ret = kvm_vcpu_set_target(vcpu, init);
1294         if (ret)
1295                 return ret;
1296
1297         /*
1298          * Ensure a rebooted VM will fault in RAM pages and detect if the
1299          * guest MMU is turned off and flush the caches as needed.
1300          *
1301          * S2FWB enforces all memory accesses to RAM being cacheable,
1302          * ensuring that the data side is always coherent. We still
1303          * need to invalidate the I-cache though, as FWB does *not*
1304          * imply CTR_EL0.DIC.
1305          */
1306         if (vcpu_has_run_once(vcpu)) {
1307                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1308                         stage2_unmap_vm(vcpu->kvm);
1309                 else
1310                         icache_inval_all_pou();
1311         }
1312
1313         vcpu_reset_hcr(vcpu);
1314         vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu);
1315
1316         /*
1317          * Handle the "start in power-off" case.
1318          */
1319         spin_lock(&vcpu->arch.mp_state_lock);
1320
1321         if (power_off)
1322                 __kvm_arm_vcpu_power_off(vcpu);
1323         else
1324                 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE);
1325
1326         spin_unlock(&vcpu->arch.mp_state_lock);
1327
1328         return 0;
1329 }
1330
1331 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1332                                  struct kvm_device_attr *attr)
1333 {
1334         int ret = -ENXIO;
1335
1336         switch (attr->group) {
1337         default:
1338                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1339                 break;
1340         }
1341
1342         return ret;
1343 }
1344
1345 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1346                                  struct kvm_device_attr *attr)
1347 {
1348         int ret = -ENXIO;
1349
1350         switch (attr->group) {
1351         default:
1352                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1353                 break;
1354         }
1355
1356         return ret;
1357 }
1358
1359 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1360                                  struct kvm_device_attr *attr)
1361 {
1362         int ret = -ENXIO;
1363
1364         switch (attr->group) {
1365         default:
1366                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1367                 break;
1368         }
1369
1370         return ret;
1371 }
1372
1373 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1374                                    struct kvm_vcpu_events *events)
1375 {
1376         memset(events, 0, sizeof(*events));
1377
1378         return __kvm_arm_vcpu_get_events(vcpu, events);
1379 }
1380
1381 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1382                                    struct kvm_vcpu_events *events)
1383 {
1384         int i;
1385
1386         /* check whether the reserved field is zero */
1387         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1388                 if (events->reserved[i])
1389                         return -EINVAL;
1390
1391         /* check whether the pad field is zero */
1392         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1393                 if (events->exception.pad[i])
1394                         return -EINVAL;
1395
1396         return __kvm_arm_vcpu_set_events(vcpu, events);
1397 }
1398
1399 long kvm_arch_vcpu_ioctl(struct file *filp,
1400                          unsigned int ioctl, unsigned long arg)
1401 {
1402         struct kvm_vcpu *vcpu = filp->private_data;
1403         void __user *argp = (void __user *)arg;
1404         struct kvm_device_attr attr;
1405         long r;
1406
1407         switch (ioctl) {
1408         case KVM_ARM_VCPU_INIT: {
1409                 struct kvm_vcpu_init init;
1410
1411                 r = -EFAULT;
1412                 if (copy_from_user(&init, argp, sizeof(init)))
1413                         break;
1414
1415                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1416                 break;
1417         }
1418         case KVM_SET_ONE_REG:
1419         case KVM_GET_ONE_REG: {
1420                 struct kvm_one_reg reg;
1421
1422                 r = -ENOEXEC;
1423                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1424                         break;
1425
1426                 r = -EFAULT;
1427                 if (copy_from_user(&reg, argp, sizeof(reg)))
1428                         break;
1429
1430                 /*
1431                  * We could owe a reset due to PSCI. Handle the pending reset
1432                  * here to ensure userspace register accesses are ordered after
1433                  * the reset.
1434                  */
1435                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1436                         kvm_reset_vcpu(vcpu);
1437
1438                 if (ioctl == KVM_SET_ONE_REG)
1439                         r = kvm_arm_set_reg(vcpu, &reg);
1440                 else
1441                         r = kvm_arm_get_reg(vcpu, &reg);
1442                 break;
1443         }
1444         case KVM_GET_REG_LIST: {
1445                 struct kvm_reg_list __user *user_list = argp;
1446                 struct kvm_reg_list reg_list;
1447                 unsigned n;
1448
1449                 r = -ENOEXEC;
1450                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1451                         break;
1452
1453                 r = -EPERM;
1454                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1455                         break;
1456
1457                 r = -EFAULT;
1458                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1459                         break;
1460                 n = reg_list.n;
1461                 reg_list.n = kvm_arm_num_regs(vcpu);
1462                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1463                         break;
1464                 r = -E2BIG;
1465                 if (n < reg_list.n)
1466                         break;
1467                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1468                 break;
1469         }
1470         case KVM_SET_DEVICE_ATTR: {
1471                 r = -EFAULT;
1472                 if (copy_from_user(&attr, argp, sizeof(attr)))
1473                         break;
1474                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1475                 break;
1476         }
1477         case KVM_GET_DEVICE_ATTR: {
1478                 r = -EFAULT;
1479                 if (copy_from_user(&attr, argp, sizeof(attr)))
1480                         break;
1481                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1482                 break;
1483         }
1484         case KVM_HAS_DEVICE_ATTR: {
1485                 r = -EFAULT;
1486                 if (copy_from_user(&attr, argp, sizeof(attr)))
1487                         break;
1488                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1489                 break;
1490         }
1491         case KVM_GET_VCPU_EVENTS: {
1492                 struct kvm_vcpu_events events;
1493
1494                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1495                         return -EINVAL;
1496
1497                 if (copy_to_user(argp, &events, sizeof(events)))
1498                         return -EFAULT;
1499
1500                 return 0;
1501         }
1502         case KVM_SET_VCPU_EVENTS: {
1503                 struct kvm_vcpu_events events;
1504
1505                 if (copy_from_user(&events, argp, sizeof(events)))
1506                         return -EFAULT;
1507
1508                 return kvm_arm_vcpu_set_events(vcpu, &events);
1509         }
1510         case KVM_ARM_VCPU_FINALIZE: {
1511                 int what;
1512
1513                 if (!kvm_vcpu_initialized(vcpu))
1514                         return -ENOEXEC;
1515
1516                 if (get_user(what, (const int __user *)argp))
1517                         return -EFAULT;
1518
1519                 return kvm_arm_vcpu_finalize(vcpu, what);
1520         }
1521         default:
1522                 r = -EINVAL;
1523         }
1524
1525         return r;
1526 }
1527
1528 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1529 {
1530
1531 }
1532
1533 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
1534                                         const struct kvm_memory_slot *memslot)
1535 {
1536         kvm_flush_remote_tlbs(kvm);
1537 }
1538
1539 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1540                                         struct kvm_arm_device_addr *dev_addr)
1541 {
1542         switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) {
1543         case KVM_ARM_DEVICE_VGIC_V2:
1544                 if (!vgic_present)
1545                         return -ENXIO;
1546                 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr);
1547         default:
1548                 return -ENODEV;
1549         }
1550 }
1551
1552 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1553 {
1554         switch (attr->group) {
1555         case KVM_ARM_VM_SMCCC_CTRL:
1556                 return kvm_vm_smccc_has_attr(kvm, attr);
1557         default:
1558                 return -ENXIO;
1559         }
1560 }
1561
1562 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1563 {
1564         switch (attr->group) {
1565         case KVM_ARM_VM_SMCCC_CTRL:
1566                 return kvm_vm_smccc_set_attr(kvm, attr);
1567         default:
1568                 return -ENXIO;
1569         }
1570 }
1571
1572 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
1573 {
1574         struct kvm *kvm = filp->private_data;
1575         void __user *argp = (void __user *)arg;
1576         struct kvm_device_attr attr;
1577
1578         switch (ioctl) {
1579         case KVM_CREATE_IRQCHIP: {
1580                 int ret;
1581                 if (!vgic_present)
1582                         return -ENXIO;
1583                 mutex_lock(&kvm->lock);
1584                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1585                 mutex_unlock(&kvm->lock);
1586                 return ret;
1587         }
1588         case KVM_ARM_SET_DEVICE_ADDR: {
1589                 struct kvm_arm_device_addr dev_addr;
1590
1591                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1592                         return -EFAULT;
1593                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1594         }
1595         case KVM_ARM_PREFERRED_TARGET: {
1596                 struct kvm_vcpu_init init;
1597
1598                 kvm_vcpu_preferred_target(&init);
1599
1600                 if (copy_to_user(argp, &init, sizeof(init)))
1601                         return -EFAULT;
1602
1603                 return 0;
1604         }
1605         case KVM_ARM_MTE_COPY_TAGS: {
1606                 struct kvm_arm_copy_mte_tags copy_tags;
1607
1608                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1609                         return -EFAULT;
1610                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1611         }
1612         case KVM_ARM_SET_COUNTER_OFFSET: {
1613                 struct kvm_arm_counter_offset offset;
1614
1615                 if (copy_from_user(&offset, argp, sizeof(offset)))
1616                         return -EFAULT;
1617                 return kvm_vm_ioctl_set_counter_offset(kvm, &offset);
1618         }
1619         case KVM_HAS_DEVICE_ATTR: {
1620                 if (copy_from_user(&attr, argp, sizeof(attr)))
1621                         return -EFAULT;
1622
1623                 return kvm_vm_has_attr(kvm, &attr);
1624         }
1625         case KVM_SET_DEVICE_ATTR: {
1626                 if (copy_from_user(&attr, argp, sizeof(attr)))
1627                         return -EFAULT;
1628
1629                 return kvm_vm_set_attr(kvm, &attr);
1630         }
1631         default:
1632                 return -EINVAL;
1633         }
1634 }
1635
1636 /* unlocks vcpus from @vcpu_lock_idx and smaller */
1637 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx)
1638 {
1639         struct kvm_vcpu *tmp_vcpu;
1640
1641         for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
1642                 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
1643                 mutex_unlock(&tmp_vcpu->mutex);
1644         }
1645 }
1646
1647 void unlock_all_vcpus(struct kvm *kvm)
1648 {
1649         lockdep_assert_held(&kvm->lock);
1650
1651         unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1);
1652 }
1653
1654 /* Returns true if all vcpus were locked, false otherwise */
1655 bool lock_all_vcpus(struct kvm *kvm)
1656 {
1657         struct kvm_vcpu *tmp_vcpu;
1658         unsigned long c;
1659
1660         lockdep_assert_held(&kvm->lock);
1661
1662         /*
1663          * Any time a vcpu is in an ioctl (including running), the
1664          * core KVM code tries to grab the vcpu->mutex.
1665          *
1666          * By grabbing the vcpu->mutex of all VCPUs we ensure that no
1667          * other VCPUs can fiddle with the state while we access it.
1668          */
1669         kvm_for_each_vcpu(c, tmp_vcpu, kvm) {
1670                 if (!mutex_trylock(&tmp_vcpu->mutex)) {
1671                         unlock_vcpus(kvm, c - 1);
1672                         return false;
1673                 }
1674         }
1675
1676         return true;
1677 }
1678
1679 static unsigned long nvhe_percpu_size(void)
1680 {
1681         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1682                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1683 }
1684
1685 static unsigned long nvhe_percpu_order(void)
1686 {
1687         unsigned long size = nvhe_percpu_size();
1688
1689         return size ? get_order(size) : 0;
1690 }
1691
1692 /* A lookup table holding the hypervisor VA for each vector slot */
1693 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1694
1695 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1696 {
1697         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1698 }
1699
1700 static int kvm_init_vector_slots(void)
1701 {
1702         int err;
1703         void *base;
1704
1705         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1706         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1707
1708         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1709         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1710
1711         if (kvm_system_needs_idmapped_vectors() &&
1712             !is_protected_kvm_enabled()) {
1713                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1714                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1715                 if (err)
1716                         return err;
1717         }
1718
1719         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1720         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1721         return 0;
1722 }
1723
1724 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits)
1725 {
1726         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1727         unsigned long tcr;
1728
1729         /*
1730          * Calculate the raw per-cpu offset without a translation from the
1731          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1732          * so that we can use adr_l to access per-cpu variables in EL2.
1733          * Also drop the KASAN tag which gets in the way...
1734          */
1735         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1736                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1737
1738         params->mair_el2 = read_sysreg(mair_el1);
1739
1740         tcr = read_sysreg(tcr_el1);
1741         if (cpus_have_final_cap(ARM64_KVM_HVHE)) {
1742                 tcr |= TCR_EPD1_MASK;
1743         } else {
1744                 tcr &= TCR_EL2_MASK;
1745                 tcr |= TCR_EL2_RES1;
1746         }
1747         tcr &= ~TCR_T0SZ_MASK;
1748         tcr |= TCR_T0SZ(hyp_va_bits);
1749         params->tcr_el2 = tcr;
1750
1751         params->pgd_pa = kvm_mmu_get_httbr();
1752         if (is_protected_kvm_enabled())
1753                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1754         else
1755                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1756         if (cpus_have_final_cap(ARM64_KVM_HVHE))
1757                 params->hcr_el2 |= HCR_E2H;
1758         params->vttbr = params->vtcr = 0;
1759
1760         /*
1761          * Flush the init params from the data cache because the struct will
1762          * be read while the MMU is off.
1763          */
1764         kvm_flush_dcache_to_poc(params, sizeof(*params));
1765 }
1766
1767 static void hyp_install_host_vector(void)
1768 {
1769         struct kvm_nvhe_init_params *params;
1770         struct arm_smccc_res res;
1771
1772         /* Switch from the HYP stub to our own HYP init vector */
1773         __hyp_set_vectors(kvm_get_idmap_vector());
1774
1775         /*
1776          * Call initialization code, and switch to the full blown HYP code.
1777          * If the cpucaps haven't been finalized yet, something has gone very
1778          * wrong, and hyp will crash and burn when it uses any
1779          * cpus_have_const_cap() wrapper.
1780          */
1781         BUG_ON(!system_capabilities_finalized());
1782         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1783         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1784         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1785 }
1786
1787 static void cpu_init_hyp_mode(void)
1788 {
1789         hyp_install_host_vector();
1790
1791         /*
1792          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1793          * at EL2.
1794          */
1795         if (this_cpu_has_cap(ARM64_SSBS) &&
1796             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1797                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1798         }
1799 }
1800
1801 static void cpu_hyp_reset(void)
1802 {
1803         if (!is_kernel_in_hyp_mode())
1804                 __hyp_reset_vectors();
1805 }
1806
1807 /*
1808  * EL2 vectors can be mapped and rerouted in a number of ways,
1809  * depending on the kernel configuration and CPU present:
1810  *
1811  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1812  *   placed in one of the vector slots, which is executed before jumping
1813  *   to the real vectors.
1814  *
1815  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1816  *   containing the hardening sequence is mapped next to the idmap page,
1817  *   and executed before jumping to the real vectors.
1818  *
1819  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1820  *   empty slot is selected, mapped next to the idmap page, and
1821  *   executed before jumping to the real vectors.
1822  *
1823  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1824  * VHE, as we don't have hypervisor-specific mappings. If the system
1825  * is VHE and yet selects this capability, it will be ignored.
1826  */
1827 static void cpu_set_hyp_vector(void)
1828 {
1829         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1830         void *vector = hyp_spectre_vector_selector[data->slot];
1831
1832         if (!is_protected_kvm_enabled())
1833                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1834         else
1835                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1836 }
1837
1838 static void cpu_hyp_init_context(void)
1839 {
1840         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1841
1842         if (!is_kernel_in_hyp_mode())
1843                 cpu_init_hyp_mode();
1844 }
1845
1846 static void cpu_hyp_init_features(void)
1847 {
1848         cpu_set_hyp_vector();
1849         kvm_arm_init_debug();
1850
1851         if (is_kernel_in_hyp_mode())
1852                 kvm_timer_init_vhe();
1853
1854         if (vgic_present)
1855                 kvm_vgic_init_cpu_hardware();
1856 }
1857
1858 static void cpu_hyp_reinit(void)
1859 {
1860         cpu_hyp_reset();
1861         cpu_hyp_init_context();
1862         cpu_hyp_init_features();
1863 }
1864
1865 static void _kvm_arch_hardware_enable(void *discard)
1866 {
1867         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1868                 cpu_hyp_reinit();
1869                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1870         }
1871 }
1872
1873 int kvm_arch_hardware_enable(void)
1874 {
1875         int was_enabled = __this_cpu_read(kvm_arm_hardware_enabled);
1876
1877         _kvm_arch_hardware_enable(NULL);
1878
1879         if (!was_enabled) {
1880                 kvm_vgic_cpu_up();
1881                 kvm_timer_cpu_up();
1882         }
1883
1884         return 0;
1885 }
1886
1887 static void _kvm_arch_hardware_disable(void *discard)
1888 {
1889         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1890                 cpu_hyp_reset();
1891                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1892         }
1893 }
1894
1895 void kvm_arch_hardware_disable(void)
1896 {
1897         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1898                 kvm_timer_cpu_down();
1899                 kvm_vgic_cpu_down();
1900         }
1901
1902         if (!is_protected_kvm_enabled())
1903                 _kvm_arch_hardware_disable(NULL);
1904 }
1905
1906 #ifdef CONFIG_CPU_PM
1907 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1908                                     unsigned long cmd,
1909                                     void *v)
1910 {
1911         /*
1912          * kvm_arm_hardware_enabled is left with its old value over
1913          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1914          * re-enable hyp.
1915          */
1916         switch (cmd) {
1917         case CPU_PM_ENTER:
1918                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1919                         /*
1920                          * don't update kvm_arm_hardware_enabled here
1921                          * so that the hardware will be re-enabled
1922                          * when we resume. See below.
1923                          */
1924                         cpu_hyp_reset();
1925
1926                 return NOTIFY_OK;
1927         case CPU_PM_ENTER_FAILED:
1928         case CPU_PM_EXIT:
1929                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1930                         /* The hardware was enabled before suspend. */
1931                         cpu_hyp_reinit();
1932
1933                 return NOTIFY_OK;
1934
1935         default:
1936                 return NOTIFY_DONE;
1937         }
1938 }
1939
1940 static struct notifier_block hyp_init_cpu_pm_nb = {
1941         .notifier_call = hyp_init_cpu_pm_notifier,
1942 };
1943
1944 static void __init hyp_cpu_pm_init(void)
1945 {
1946         if (!is_protected_kvm_enabled())
1947                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1948 }
1949 static void __init hyp_cpu_pm_exit(void)
1950 {
1951         if (!is_protected_kvm_enabled())
1952                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1953 }
1954 #else
1955 static inline void __init hyp_cpu_pm_init(void)
1956 {
1957 }
1958 static inline void __init hyp_cpu_pm_exit(void)
1959 {
1960 }
1961 #endif
1962
1963 static void __init init_cpu_logical_map(void)
1964 {
1965         unsigned int cpu;
1966
1967         /*
1968          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
1969          * Only copy the set of online CPUs whose features have been checked
1970          * against the finalized system capabilities. The hypervisor will not
1971          * allow any other CPUs from the `possible` set to boot.
1972          */
1973         for_each_online_cpu(cpu)
1974                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
1975 }
1976
1977 #define init_psci_0_1_impl_state(config, what)  \
1978         config.psci_0_1_ ## what ## _implemented = psci_ops.what
1979
1980 static bool __init init_psci_relay(void)
1981 {
1982         /*
1983          * If PSCI has not been initialized, protected KVM cannot install
1984          * itself on newly booted CPUs.
1985          */
1986         if (!psci_ops.get_version) {
1987                 kvm_err("Cannot initialize protected mode without PSCI\n");
1988                 return false;
1989         }
1990
1991         kvm_host_psci_config.version = psci_ops.get_version();
1992         kvm_host_psci_config.smccc_version = arm_smccc_get_version();
1993
1994         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
1995                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
1996                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
1997                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
1998                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
1999                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
2000         }
2001         return true;
2002 }
2003
2004 static int __init init_subsystems(void)
2005 {
2006         int err = 0;
2007
2008         /*
2009          * Enable hardware so that subsystem initialisation can access EL2.
2010          */
2011         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
2012
2013         /*
2014          * Register CPU lower-power notifier
2015          */
2016         hyp_cpu_pm_init();
2017
2018         /*
2019          * Init HYP view of VGIC
2020          */
2021         err = kvm_vgic_hyp_init();
2022         switch (err) {
2023         case 0:
2024                 vgic_present = true;
2025                 break;
2026         case -ENODEV:
2027         case -ENXIO:
2028                 vgic_present = false;
2029                 err = 0;
2030                 break;
2031         default:
2032                 goto out;
2033         }
2034
2035         /*
2036          * Init HYP architected timer support
2037          */
2038         err = kvm_timer_hyp_init(vgic_present);
2039         if (err)
2040                 goto out;
2041
2042         kvm_register_perf_callbacks(NULL);
2043
2044 out:
2045         if (err)
2046                 hyp_cpu_pm_exit();
2047
2048         if (err || !is_protected_kvm_enabled())
2049                 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
2050
2051         return err;
2052 }
2053
2054 static void __init teardown_subsystems(void)
2055 {
2056         kvm_unregister_perf_callbacks();
2057         hyp_cpu_pm_exit();
2058 }
2059
2060 static void __init teardown_hyp_mode(void)
2061 {
2062         int cpu;
2063
2064         free_hyp_pgds();
2065         for_each_possible_cpu(cpu) {
2066                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
2067                 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order());
2068         }
2069 }
2070
2071 static int __init do_pkvm_init(u32 hyp_va_bits)
2072 {
2073         void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base));
2074         int ret;
2075
2076         preempt_disable();
2077         cpu_hyp_init_context();
2078         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
2079                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
2080                                 hyp_va_bits);
2081         cpu_hyp_init_features();
2082
2083         /*
2084          * The stub hypercalls are now disabled, so set our local flag to
2085          * prevent a later re-init attempt in kvm_arch_hardware_enable().
2086          */
2087         __this_cpu_write(kvm_arm_hardware_enabled, 1);
2088         preempt_enable();
2089
2090         return ret;
2091 }
2092
2093 static u64 get_hyp_id_aa64pfr0_el1(void)
2094 {
2095         /*
2096          * Track whether the system isn't affected by spectre/meltdown in the
2097          * hypervisor's view of id_aa64pfr0_el1, used for protected VMs.
2098          * Although this is per-CPU, we make it global for simplicity, e.g., not
2099          * to have to worry about vcpu migration.
2100          *
2101          * Unlike for non-protected VMs, userspace cannot override this for
2102          * protected VMs.
2103          */
2104         u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
2105
2106         val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
2107                  ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));
2108
2109         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
2110                           arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED);
2111         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
2112                           arm64_get_meltdown_state() == SPECTRE_UNAFFECTED);
2113
2114         return val;
2115 }
2116
2117 static void kvm_hyp_init_symbols(void)
2118 {
2119         kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1();
2120         kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1);
2121         kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
2122         kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
2123         kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
2124         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
2125         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
2126         kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1);
2127         kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1);
2128         kvm_nvhe_sym(__icache_flags) = __icache_flags;
2129         kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits;
2130 }
2131
2132 static int __init kvm_hyp_init_protection(u32 hyp_va_bits)
2133 {
2134         void *addr = phys_to_virt(hyp_mem_base);
2135         int ret;
2136
2137         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
2138         if (ret)
2139                 return ret;
2140
2141         ret = do_pkvm_init(hyp_va_bits);
2142         if (ret)
2143                 return ret;
2144
2145         free_hyp_pgds();
2146
2147         return 0;
2148 }
2149
2150 static void pkvm_hyp_init_ptrauth(void)
2151 {
2152         struct kvm_cpu_context *hyp_ctxt;
2153         int cpu;
2154
2155         for_each_possible_cpu(cpu) {
2156                 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu);
2157                 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long();
2158                 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long();
2159                 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long();
2160                 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long();
2161                 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long();
2162                 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long();
2163                 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long();
2164                 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long();
2165                 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long();
2166                 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long();
2167         }
2168 }
2169
2170 /* Inits Hyp-mode on all online CPUs */
2171 static int __init init_hyp_mode(void)
2172 {
2173         u32 hyp_va_bits;
2174         int cpu;
2175         int err = -ENOMEM;
2176
2177         /*
2178          * The protected Hyp-mode cannot be initialized if the memory pool
2179          * allocation has failed.
2180          */
2181         if (is_protected_kvm_enabled() && !hyp_mem_base)
2182                 goto out_err;
2183
2184         /*
2185          * Allocate Hyp PGD and setup Hyp identity mapping
2186          */
2187         err = kvm_mmu_init(&hyp_va_bits);
2188         if (err)
2189                 goto out_err;
2190
2191         /*
2192          * Allocate stack pages for Hypervisor-mode
2193          */
2194         for_each_possible_cpu(cpu) {
2195                 unsigned long stack_page;
2196
2197                 stack_page = __get_free_page(GFP_KERNEL);
2198                 if (!stack_page) {
2199                         err = -ENOMEM;
2200                         goto out_err;
2201                 }
2202
2203                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
2204         }
2205
2206         /*
2207          * Allocate and initialize pages for Hypervisor-mode percpu regions.
2208          */
2209         for_each_possible_cpu(cpu) {
2210                 struct page *page;
2211                 void *page_addr;
2212
2213                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
2214                 if (!page) {
2215                         err = -ENOMEM;
2216                         goto out_err;
2217                 }
2218
2219                 page_addr = page_address(page);
2220                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
2221                 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr;
2222         }
2223
2224         /*
2225          * Map the Hyp-code called directly from the host
2226          */
2227         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
2228                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
2229         if (err) {
2230                 kvm_err("Cannot map world-switch code\n");
2231                 goto out_err;
2232         }
2233
2234         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
2235                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
2236         if (err) {
2237                 kvm_err("Cannot map .hyp.rodata section\n");
2238                 goto out_err;
2239         }
2240
2241         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
2242                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
2243         if (err) {
2244                 kvm_err("Cannot map rodata section\n");
2245                 goto out_err;
2246         }
2247
2248         /*
2249          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
2250          * section thanks to an assertion in the linker script. Map it RW and
2251          * the rest of .bss RO.
2252          */
2253         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
2254                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
2255         if (err) {
2256                 kvm_err("Cannot map hyp bss section: %d\n", err);
2257                 goto out_err;
2258         }
2259
2260         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
2261                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
2262         if (err) {
2263                 kvm_err("Cannot map bss section\n");
2264                 goto out_err;
2265         }
2266
2267         /*
2268          * Map the Hyp stack pages
2269          */
2270         for_each_possible_cpu(cpu) {
2271                 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2272                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
2273                 unsigned long hyp_addr;
2274
2275                 /*
2276                  * Allocate a contiguous HYP private VA range for the stack
2277                  * and guard page. The allocation is also aligned based on
2278                  * the order of its size.
2279                  */
2280                 err = hyp_alloc_private_va_range(PAGE_SIZE * 2, &hyp_addr);
2281                 if (err) {
2282                         kvm_err("Cannot allocate hyp stack guard page\n");
2283                         goto out_err;
2284                 }
2285
2286                 /*
2287                  * Since the stack grows downwards, map the stack to the page
2288                  * at the higher address and leave the lower guard page
2289                  * unbacked.
2290                  *
2291                  * Any valid stack address now has the PAGE_SHIFT bit as 1
2292                  * and addresses corresponding to the guard page have the
2293                  * PAGE_SHIFT bit as 0 - this is used for overflow detection.
2294                  */
2295                 err = __create_hyp_mappings(hyp_addr + PAGE_SIZE, PAGE_SIZE,
2296                                             __pa(stack_page), PAGE_HYP);
2297                 if (err) {
2298                         kvm_err("Cannot map hyp stack\n");
2299                         goto out_err;
2300                 }
2301
2302                 /*
2303                  * Save the stack PA in nvhe_init_params. This will be needed
2304                  * to recreate the stack mapping in protected nVHE mode.
2305                  * __hyp_pa() won't do the right thing there, since the stack
2306                  * has been mapped in the flexible private VA space.
2307                  */
2308                 params->stack_pa = __pa(stack_page);
2309
2310                 params->stack_hyp_va = hyp_addr + (2 * PAGE_SIZE);
2311         }
2312
2313         for_each_possible_cpu(cpu) {
2314                 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu];
2315                 char *percpu_end = percpu_begin + nvhe_percpu_size();
2316
2317                 /* Map Hyp percpu pages */
2318                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
2319                 if (err) {
2320                         kvm_err("Cannot map hyp percpu region\n");
2321                         goto out_err;
2322                 }
2323
2324                 /* Prepare the CPU initialization parameters */
2325                 cpu_prepare_hyp_mode(cpu, hyp_va_bits);
2326         }
2327
2328         kvm_hyp_init_symbols();
2329
2330         if (is_protected_kvm_enabled()) {
2331                 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) &&
2332                     cpus_have_const_cap(ARM64_HAS_ADDRESS_AUTH))
2333                         pkvm_hyp_init_ptrauth();
2334
2335                 init_cpu_logical_map();
2336
2337                 if (!init_psci_relay()) {
2338                         err = -ENODEV;
2339                         goto out_err;
2340                 }
2341
2342                 err = kvm_hyp_init_protection(hyp_va_bits);
2343                 if (err) {
2344                         kvm_err("Failed to init hyp memory protection\n");
2345                         goto out_err;
2346                 }
2347         }
2348
2349         return 0;
2350
2351 out_err:
2352         teardown_hyp_mode();
2353         kvm_err("error initializing Hyp mode: %d\n", err);
2354         return err;
2355 }
2356
2357 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2358 {
2359         struct kvm_vcpu *vcpu;
2360         unsigned long i;
2361
2362         mpidr &= MPIDR_HWID_BITMASK;
2363         kvm_for_each_vcpu(i, vcpu, kvm) {
2364                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2365                         return vcpu;
2366         }
2367         return NULL;
2368 }
2369
2370 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm)
2371 {
2372         return irqchip_in_kernel(kvm);
2373 }
2374
2375 bool kvm_arch_has_irq_bypass(void)
2376 {
2377         return true;
2378 }
2379
2380 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2381                                       struct irq_bypass_producer *prod)
2382 {
2383         struct kvm_kernel_irqfd *irqfd =
2384                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2385
2386         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2387                                           &irqfd->irq_entry);
2388 }
2389 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2390                                       struct irq_bypass_producer *prod)
2391 {
2392         struct kvm_kernel_irqfd *irqfd =
2393                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2394
2395         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2396                                      &irqfd->irq_entry);
2397 }
2398
2399 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2400 {
2401         struct kvm_kernel_irqfd *irqfd =
2402                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2403
2404         kvm_arm_halt_guest(irqfd->kvm);
2405 }
2406
2407 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2408 {
2409         struct kvm_kernel_irqfd *irqfd =
2410                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2411
2412         kvm_arm_resume_guest(irqfd->kvm);
2413 }
2414
2415 /* Initialize Hyp-mode and memory mappings on all CPUs */
2416 static __init int kvm_arm_init(void)
2417 {
2418         int err;
2419         bool in_hyp_mode;
2420
2421         if (!is_hyp_mode_available()) {
2422                 kvm_info("HYP mode not available\n");
2423                 return -ENODEV;
2424         }
2425
2426         if (kvm_get_mode() == KVM_MODE_NONE) {
2427                 kvm_info("KVM disabled from command line\n");
2428                 return -ENODEV;
2429         }
2430
2431         err = kvm_sys_reg_table_init();
2432         if (err) {
2433                 kvm_info("Error initializing system register tables");
2434                 return err;
2435         }
2436
2437         in_hyp_mode = is_kernel_in_hyp_mode();
2438
2439         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2440             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2441                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2442                          "Only trusted guests should be used on this system.\n");
2443
2444         err = kvm_set_ipa_limit();
2445         if (err)
2446                 return err;
2447
2448         err = kvm_arm_init_sve();
2449         if (err)
2450                 return err;
2451
2452         err = kvm_arm_vmid_alloc_init();
2453         if (err) {
2454                 kvm_err("Failed to initialize VMID allocator.\n");
2455                 return err;
2456         }
2457
2458         if (!in_hyp_mode) {
2459                 err = init_hyp_mode();
2460                 if (err)
2461                         goto out_err;
2462         }
2463
2464         err = kvm_init_vector_slots();
2465         if (err) {
2466                 kvm_err("Cannot initialise vector slots\n");
2467                 goto out_hyp;
2468         }
2469
2470         err = init_subsystems();
2471         if (err)
2472                 goto out_hyp;
2473
2474         if (is_protected_kvm_enabled()) {
2475                 kvm_info("Protected nVHE mode initialized successfully\n");
2476         } else if (in_hyp_mode) {
2477                 kvm_info("VHE mode initialized successfully\n");
2478         } else {
2479                 kvm_info("Hyp mode initialized successfully\n");
2480         }
2481
2482         /*
2483          * FIXME: Do something reasonable if kvm_init() fails after pKVM
2484          * hypervisor protection is finalized.
2485          */
2486         err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2487         if (err)
2488                 goto out_subs;
2489
2490         kvm_arm_initialised = true;
2491
2492         return 0;
2493
2494 out_subs:
2495         teardown_subsystems();
2496 out_hyp:
2497         if (!in_hyp_mode)
2498                 teardown_hyp_mode();
2499 out_err:
2500         kvm_arm_vmid_alloc_free();
2501         return err;
2502 }
2503
2504 static int __init early_kvm_mode_cfg(char *arg)
2505 {
2506         if (!arg)
2507                 return -EINVAL;
2508
2509         if (strcmp(arg, "none") == 0) {
2510                 kvm_mode = KVM_MODE_NONE;
2511                 return 0;
2512         }
2513
2514         if (!is_hyp_mode_available()) {
2515                 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n");
2516                 return 0;
2517         }
2518
2519         if (strcmp(arg, "protected") == 0) {
2520                 if (!is_kernel_in_hyp_mode())
2521                         kvm_mode = KVM_MODE_PROTECTED;
2522                 else
2523                         pr_warn_once("Protected KVM not available with VHE\n");
2524
2525                 return 0;
2526         }
2527
2528         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) {
2529                 kvm_mode = KVM_MODE_DEFAULT;
2530                 return 0;
2531         }
2532
2533         if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) {
2534                 kvm_mode = KVM_MODE_NV;
2535                 return 0;
2536         }
2537
2538         return -EINVAL;
2539 }
2540 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2541
2542 enum kvm_mode kvm_get_mode(void)
2543 {
2544         return kvm_mode;
2545 }
2546
2547 module_init(kvm_arm_init);