arch/arm64/kvm/arm.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   4  * Author: Christoffer Dall <c.dall@virtualopensystems.com>
   5  */
   6
   7 #include <linux/bug.h>
   8 #include <linux/cpu_pm.h>
   9 #include <linux/entry-kvm.h>
  10 #include <linux/errno.h>
  11 #include <linux/err.h>
  12 #include <linux/kvm_host.h>
  13 #include <linux/list.h>
  14 #include <linux/module.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mman.h>
  18 #include <linux/sched.h>
  19 #include <linux/kvm.h>
  20 #include <linux/kvm_irqfd.h>
  21 #include <linux/irqbypass.h>
  22 #include <linux/sched/stat.h>
  23 #include <linux/psci.h>
  24 #include <trace/events/kvm.h>
  25
  26 #define CREATE_TRACE_POINTS
  27 #include "trace_arm.h"
  28
  29 #include <linux/uaccess.h>
  30 #include <asm/ptrace.h>
  31 #include <asm/mman.h>
  32 #include <asm/tlbflush.h>
  33 #include <asm/cacheflush.h>
  34 #include <asm/cpufeature.h>
  35 #include <asm/virt.h>
  36 #include <asm/kvm_arm.h>
  37 #include <asm/kvm_asm.h>
  38 #include <asm/kvm_mmu.h>
  39 #include <asm/kvm_nested.h>
  40 #include <asm/kvm_pkvm.h>
  41 #include <asm/kvm_emulate.h>
  42 #include <asm/sections.h>
  43
  44 #include <kvm/arm_hypercalls.h>
  45 #include <kvm/arm_pmu.h>
  46 #include <kvm/arm_psci.h>
  47
  48 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
  49
  50 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
  51
  52 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  53 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
  54
  55 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt);
  56
  57 static bool vgic_present, kvm_arm_initialised;
  58
  59 static DEFINE_PER_CPU(unsigned char, kvm_hyp_initialized);
  60 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  61
  62 bool is_kvm_arm_initialised(void)
  63 {
  64         return kvm_arm_initialised;
  65 }
  66
  67 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  68 {
  69         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  70 }
  71
  72 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  73                             struct kvm_enable_cap *cap)
  74 {
  75         int r;
  76         u64 new_cap;
  77
  78         if (cap->flags)
  79                 return -EINVAL;
  80
  81         switch (cap->cap) {
  82         case KVM_CAP_ARM_NISV_TO_USER:
  83                 r = 0;
  84                 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER,
  85                         &kvm->arch.flags);
  86                 break;
  87         case KVM_CAP_ARM_MTE:
  88                 mutex_lock(&kvm->lock);
  89                 if (!system_supports_mte() || kvm->created_vcpus) {
  90                         r = -EINVAL;
  91                 } else {
  92                         r = 0;
  93                         set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags);
  94                 }
  95                 mutex_unlock(&kvm->lock);
  96                 break;
  97         case KVM_CAP_ARM_SYSTEM_SUSPEND:
  98                 r = 0;
  99                 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags);
 100                 break;
 101         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 102                 new_cap = cap->args[0];
 103
 104                 mutex_lock(&kvm->slots_lock);
 105                 /*
 106                  * To keep things simple, allow changing the chunk
 107                  * size only when no memory slots have been created.
 108                  */
 109                 if (!kvm_are_all_memslots_empty(kvm)) {
 110                         r = -EINVAL;
 111                 } else if (new_cap && !kvm_is_block_size_supported(new_cap)) {
 112                         r = -EINVAL;
 113                 } else {
 114                         r = 0;
 115                         kvm->arch.mmu.split_page_chunk_size = new_cap;
 116                 }
 117                 mutex_unlock(&kvm->slots_lock);
 118                 break;
 119         default:
 120                 r = -EINVAL;
 121                 break;
 122         }
 123
 124         return r;
 125 }
 126
 127 static int kvm_arm_default_max_vcpus(void)
 128 {
 129         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 130 }
 131
 132 /**
 133  * kvm_arch_init_vm - initializes a VM data structure
 134  * @kvm:        pointer to the KVM struct
 135  */
 136 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 137 {
 138         int ret;
 139
 140         mutex_init(&kvm->arch.config_lock);
 141
 142 #ifdef CONFIG_LOCKDEP
 143         /* Clue in lockdep that the config_lock must be taken inside kvm->lock */
 144         mutex_lock(&kvm->lock);
 145         mutex_lock(&kvm->arch.config_lock);
 146         mutex_unlock(&kvm->arch.config_lock);
 147         mutex_unlock(&kvm->lock);
 148 #endif
 149
 150         ret = kvm_share_hyp(kvm, kvm + 1);
 151         if (ret)
 152                 return ret;
 153
 154         ret = pkvm_init_host_vm(kvm);
 155         if (ret)
 156                 goto err_unshare_kvm;
 157
 158         if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) {
 159                 ret = -ENOMEM;
 160                 goto err_unshare_kvm;
 161         }
 162         cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask);
 163
 164         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type);
 165         if (ret)
 166                 goto err_free_cpumask;
 167
 168         kvm_vgic_early_init(kvm);
 169
 170         kvm_timer_init_vm(kvm);
 171
 172         /* The maximum number of VCPUs is limited by the host's GIC model */
 173         kvm->max_vcpus = kvm_arm_default_max_vcpus();
 174
 175         kvm_arm_init_hypercalls(kvm);
 176
 177         bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES);
 178
 179         return 0;
 180
 181 err_free_cpumask:
 182         free_cpumask_var(kvm->arch.supported_cpus);
 183 err_unshare_kvm:
 184         kvm_unshare_hyp(kvm, kvm + 1);
 185         return ret;
 186 }
 187
 188 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 189 {
 190         return VM_FAULT_SIGBUS;
 191 }
 192
 193
 194 /**
 195  * kvm_arch_destroy_vm - destroy the VM data structure
 196  * @kvm:        pointer to the KVM struct
 197  */
 198 void kvm_arch_destroy_vm(struct kvm *kvm)
 199 {
 200         bitmap_free(kvm->arch.pmu_filter);
 201         free_cpumask_var(kvm->arch.supported_cpus);
 202
 203         kvm_vgic_destroy(kvm);
 204
 205         if (is_protected_kvm_enabled())
 206                 pkvm_destroy_hyp_vm(kvm);
 207
 208         kvm_destroy_vcpus(kvm);
 209
 210         kvm_unshare_hyp(kvm, kvm + 1);
 211
 212         kvm_arm_teardown_hypercalls(kvm);
 213 }
 214
 215 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 216 {
 217         int r;
 218         switch (ext) {
 219         case KVM_CAP_IRQCHIP:
 220                 r = vgic_present;
 221                 break;
 222         case KVM_CAP_IOEVENTFD:
 223         case KVM_CAP_DEVICE_CTRL:
 224         case KVM_CAP_USER_MEMORY:
 225         case KVM_CAP_SYNC_MMU:
 226         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 227         case KVM_CAP_ONE_REG:
 228         case KVM_CAP_ARM_PSCI:
 229         case KVM_CAP_ARM_PSCI_0_2:
 230         case KVM_CAP_READONLY_MEM:
 231         case KVM_CAP_MP_STATE:
 232         case KVM_CAP_IMMEDIATE_EXIT:
 233         case KVM_CAP_VCPU_EVENTS:
 234         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 235         case KVM_CAP_ARM_NISV_TO_USER:
 236         case KVM_CAP_ARM_INJECT_EXT_DABT:
 237         case KVM_CAP_SET_GUEST_DEBUG:
 238         case KVM_CAP_VCPU_ATTRIBUTES:
 239         case KVM_CAP_PTP_KVM:
 240         case KVM_CAP_ARM_SYSTEM_SUSPEND:
 241         case KVM_CAP_IRQFD_RESAMPLE:
 242         case KVM_CAP_COUNTER_OFFSET:
 243                 r = 1;
 244                 break;
 245         case KVM_CAP_SET_GUEST_DEBUG2:
 246                 return KVM_GUESTDBG_VALID_MASK;
 247         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 248                 r = 1;
 249                 break;
 250         case KVM_CAP_NR_VCPUS:
 251                 /*
 252                  * ARM64 treats KVM_CAP_NR_CPUS differently from all other
 253                  * architectures, as it does not always bound it to
 254                  * KVM_CAP_MAX_VCPUS. It should not matter much because
 255                  * this is just an advisory value.
 256                  */
 257                 r = min_t(unsigned int, num_online_cpus(),
 258                           kvm_arm_default_max_vcpus());
 259                 break;
 260         case KVM_CAP_MAX_VCPUS:
 261         case KVM_CAP_MAX_VCPU_ID:
 262                 if (kvm)
 263                         r = kvm->max_vcpus;
 264                 else
 265                         r = kvm_arm_default_max_vcpus();
 266                 break;
 267         case KVM_CAP_MSI_DEVID:
 268                 if (!kvm)
 269                         r = -EINVAL;
 270                 else
 271                         r = kvm->arch.vgic.msis_require_devid;
 272                 break;
 273         case KVM_CAP_ARM_USER_IRQ:
 274                 /*
 275                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 276                  * (bump this number if adding more devices)
 277                  */
 278                 r = 1;
 279                 break;
 280         case KVM_CAP_ARM_MTE:
 281                 r = system_supports_mte();
 282                 break;
 283         case KVM_CAP_STEAL_TIME:
 284                 r = kvm_arm_pvtime_supported();
 285                 break;
 286         case KVM_CAP_ARM_EL1_32BIT:
 287                 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1);
 288                 break;
 289         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 290                 r = get_num_brps();
 291                 break;
 292         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 293                 r = get_num_wrps();
 294                 break;
 295         case KVM_CAP_ARM_PMU_V3:
 296                 r = kvm_arm_support_pmu_v3();
 297                 break;
 298         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 299                 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN);
 300                 break;
 301         case KVM_CAP_ARM_VM_IPA_SIZE:
 302                 r = get_kvm_ipa_limit();
 303                 break;
 304         case KVM_CAP_ARM_SVE:
 305                 r = system_supports_sve();
 306                 break;
 307         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 308         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 309                 r = system_has_full_ptr_auth();
 310                 break;
 311         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 312                 if (kvm)
 313                         r = kvm->arch.mmu.split_page_chunk_size;
 314                 else
 315                         r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
 316                 break;
 317         case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES:
 318                 r = kvm_supported_block_sizes();
 319                 break;
 320         default:
 321                 r = 0;
 322         }
 323
 324         return r;
 325 }
 326
 327 long kvm_arch_dev_ioctl(struct file *filp,
 328                         unsigned int ioctl, unsigned long arg)
 329 {
 330         return -EINVAL;
 331 }
 332
 333 struct kvm *kvm_arch_alloc_vm(void)
 334 {
 335         size_t sz = sizeof(struct kvm);
 336
 337         if (!has_vhe())
 338                 return kzalloc(sz, GFP_KERNEL_ACCOUNT);
 339
 340         return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO);
 341 }
 342
 343 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 344 {
 345         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 346                 return -EBUSY;
 347
 348         if (id >= kvm->max_vcpus)
 349                 return -EINVAL;
 350
 351         return 0;
 352 }
 353
 354 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 355 {
 356         int err;
 357
 358         spin_lock_init(&vcpu->arch.mp_state_lock);
 359
 360 #ifdef CONFIG_LOCKDEP
 361         /* Inform lockdep that the config_lock is acquired after vcpu->mutex */
 362         mutex_lock(&vcpu->mutex);
 363         mutex_lock(&vcpu->kvm->arch.config_lock);
 364         mutex_unlock(&vcpu->kvm->arch.config_lock);
 365         mutex_unlock(&vcpu->mutex);
 366 #endif
 367
 368         /* Force users to call KVM_ARM_VCPU_INIT */
 369         vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
 370
 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(cpu, 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_get_flag(vcpu, VCPU_INITIALIZED);
 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         vcpu_set_flag(vcpu, IN_WFI);
 722         vgic_v4_put(vcpu);
 723         preempt_enable();
 724
 725         kvm_vcpu_halt(vcpu);
 726         vcpu_clear_flag(vcpu, IN_WFIT);
 727
 728         preempt_disable();
 729         vcpu_clear_flag(vcpu, IN_WFI);
 730         vgic_v4_load(vcpu);
 731         preempt_enable();
 732 }
 733
 734 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu)
 735 {
 736         if (!kvm_arm_vcpu_suspended(vcpu))
 737                 return 1;
 738
 739         kvm_vcpu_wfi(vcpu);
 740
 741         /*
 742          * The suspend state is sticky; we do not leave it until userspace
 743          * explicitly marks the vCPU as runnable. Request that we suspend again
 744          * later.
 745          */
 746         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 747
 748         /*
 749          * Check to make sure the vCPU is actually runnable. If so, exit to
 750          * userspace informing it of the wakeup condition.
 751          */
 752         if (kvm_arch_vcpu_runnable(vcpu)) {
 753                 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event));
 754                 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP;
 755                 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
 756                 return 0;
 757         }
 758
 759         /*
 760          * Otherwise, we were unblocked to process a different event, such as a
 761          * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to
 762          * process the event.
 763          */
 764         return 1;
 765 }
 766
 767 /**
 768  * check_vcpu_requests - check and handle pending vCPU requests
 769  * @vcpu:       the VCPU pointer
 770  *
 771  * Return: 1 if we should enter the guest
 772  *         0 if we should exit to userspace
 773  *         < 0 if we should exit to userspace, where the return value indicates
 774  *         an error
 775  */
 776 static int check_vcpu_requests(struct kvm_vcpu *vcpu)
 777 {
 778         if (kvm_request_pending(vcpu)) {
 779                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 780                         kvm_vcpu_sleep(vcpu);
 781
 782                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 783                         kvm_reset_vcpu(vcpu);
 784
 785                 /*
 786                  * Clear IRQ_PENDING requests that were made to guarantee
 787                  * that a VCPU sees new virtual interrupts.
 788                  */
 789                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 790
 791                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 792                         kvm_update_stolen_time(vcpu);
 793
 794                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 795                         /* The distributor enable bits were changed */
 796                         preempt_disable();
 797                         vgic_v4_put(vcpu);
 798                         vgic_v4_load(vcpu);
 799                         preempt_enable();
 800                 }
 801
 802                 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
 803                         kvm_pmu_handle_pmcr(vcpu,
 804                                             __vcpu_sys_reg(vcpu, PMCR_EL0));
 805
 806                 if (kvm_check_request(KVM_REQ_RESYNC_PMU_EL0, vcpu))
 807                         kvm_vcpu_pmu_restore_guest(vcpu);
 808
 809                 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu))
 810                         return kvm_vcpu_suspend(vcpu);
 811
 812                 if (kvm_dirty_ring_check_request(vcpu))
 813                         return 0;
 814         }
 815
 816         return 1;
 817 }
 818
 819 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 820 {
 821         if (likely(!vcpu_mode_is_32bit(vcpu)))
 822                 return false;
 823
 824         if (vcpu_has_nv(vcpu))
 825                 return true;
 826
 827         return !kvm_supports_32bit_el0();
 828 }
 829
 830 /**
 831  * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest
 832  * @vcpu:       The VCPU pointer
 833  * @ret:        Pointer to write optional return code
 834  *
 835  * Returns: true if the VCPU needs to return to a preemptible + interruptible
 836  *          and skip guest entry.
 837  *
 838  * This function disambiguates between two different types of exits: exits to a
 839  * preemptible + interruptible kernel context and exits to userspace. For an
 840  * exit to userspace, this function will write the return code to ret and return
 841  * true. For an exit to preemptible + interruptible kernel context (i.e. check
 842  * for pending work and re-enter), return true without writing to ret.
 843  */
 844 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret)
 845 {
 846         struct kvm_run *run = vcpu->run;
 847
 848         /*
 849          * If we're using a userspace irqchip, then check if we need
 850          * to tell a userspace irqchip about timer or PMU level
 851          * changes and if so, exit to userspace (the actual level
 852          * state gets updated in kvm_timer_update_run and
 853          * kvm_pmu_update_run below).
 854          */
 855         if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 856                 if (kvm_timer_should_notify_user(vcpu) ||
 857                     kvm_pmu_should_notify_user(vcpu)) {
 858                         *ret = -EINTR;
 859                         run->exit_reason = KVM_EXIT_INTR;
 860                         return true;
 861                 }
 862         }
 863
 864         if (unlikely(vcpu_on_unsupported_cpu(vcpu))) {
 865                 run->exit_reason = KVM_EXIT_FAIL_ENTRY;
 866                 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED;
 867                 run->fail_entry.cpu = smp_processor_id();
 868                 *ret = 0;
 869                 return true;
 870         }
 871
 872         return kvm_request_pending(vcpu) ||
 873                         xfer_to_guest_mode_work_pending();
 874 }
 875
 876 /*
 877  * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while
 878  * the vCPU is running.
 879  *
 880  * This must be noinstr as instrumentation may make use of RCU, and this is not
 881  * safe during the EQS.
 882  */
 883 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
 884 {
 885         int ret;
 886
 887         guest_state_enter_irqoff();
 888         ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 889         guest_state_exit_irqoff();
 890
 891         return ret;
 892 }
 893
 894 /**
 895  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 896  * @vcpu:       The VCPU pointer
 897  *
 898  * This function is called through the VCPU_RUN ioctl called from user space. It
 899  * will execute VM code in a loop until the time slice for the process is used
 900  * or some emulation is needed from user space in which case the function will
 901  * return with return value 0 and with the kvm_run structure filled in with the
 902  * required data for the requested emulation.
 903  */
 904 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 905 {
 906         struct kvm_run *run = vcpu->run;
 907         int ret;
 908
 909         if (run->exit_reason == KVM_EXIT_MMIO) {
 910                 ret = kvm_handle_mmio_return(vcpu);
 911                 if (ret)
 912                         return ret;
 913         }
 914
 915         vcpu_load(vcpu);
 916
 917         if (run->immediate_exit) {
 918                 ret = -EINTR;
 919                 goto out;
 920         }
 921
 922         kvm_sigset_activate(vcpu);
 923
 924         ret = 1;
 925         run->exit_reason = KVM_EXIT_UNKNOWN;
 926         run->flags = 0;
 927         while (ret > 0) {
 928                 /*
 929                  * Check conditions before entering the guest
 930                  */
 931                 ret = xfer_to_guest_mode_handle_work(vcpu);
 932                 if (!ret)
 933                         ret = 1;
 934
 935                 if (ret > 0)
 936                         ret = check_vcpu_requests(vcpu);
 937
 938                 /*
 939                  * Preparing the interrupts to be injected also
 940                  * involves poking the GIC, which must be done in a
 941                  * non-preemptible context.
 942                  */
 943                 preempt_disable();
 944
 945                 /*
 946                  * The VMID allocator only tracks active VMIDs per
 947                  * physical CPU, and therefore the VMID allocated may not be
 948                  * preserved on VMID roll-over if the task was preempted,
 949                  * making a thread's VMID inactive. So we need to call
 950                  * kvm_arm_vmid_update() in non-premptible context.
 951                  */
 952                 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid);
 953
 954                 kvm_pmu_flush_hwstate(vcpu);
 955
 956                 local_irq_disable();
 957
 958                 kvm_vgic_flush_hwstate(vcpu);
 959
 960                 kvm_pmu_update_vcpu_events(vcpu);
 961
 962                 /*
 963                  * Ensure we set mode to IN_GUEST_MODE after we disable
 964                  * interrupts and before the final VCPU requests check.
 965                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 966                  * Documentation/virt/kvm/vcpu-requests.rst
 967                  */
 968                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 969
 970                 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) {
 971                         vcpu->mode = OUTSIDE_GUEST_MODE;
 972                         isb(); /* Ensure work in x_flush_hwstate is committed */
 973                         kvm_pmu_sync_hwstate(vcpu);
 974                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 975                                 kvm_timer_sync_user(vcpu);
 976                         kvm_vgic_sync_hwstate(vcpu);
 977                         local_irq_enable();
 978                         preempt_enable();
 979                         continue;
 980                 }
 981
 982                 kvm_arm_setup_debug(vcpu);
 983                 kvm_arch_vcpu_ctxflush_fp(vcpu);
 984
 985                 /**************************************************************
 986                  * Enter the guest
 987                  */
 988                 trace_kvm_entry(*vcpu_pc(vcpu));
 989                 guest_timing_enter_irqoff();
 990
 991                 ret = kvm_arm_vcpu_enter_exit(vcpu);
 992
 993                 vcpu->mode = OUTSIDE_GUEST_MODE;
 994                 vcpu->stat.exits++;
 995                 /*
 996                  * Back from guest
 997                  *************************************************************/
 998
 999                 kvm_arm_clear_debug(vcpu);
1000
1001                 /*
1002                  * We must sync the PMU state before the vgic state so
1003                  * that the vgic can properly sample the updated state of the
1004                  * interrupt line.
1005                  */
1006                 kvm_pmu_sync_hwstate(vcpu);
1007
1008                 /*
1009                  * Sync the vgic state before syncing the timer state because
1010                  * the timer code needs to know if the virtual timer
1011                  * interrupts are active.
1012                  */
1013                 kvm_vgic_sync_hwstate(vcpu);
1014
1015                 /*
1016                  * Sync the timer hardware state before enabling interrupts as
1017                  * we don't want vtimer interrupts to race with syncing the
1018                  * timer virtual interrupt state.
1019                  */
1020                 if (static_branch_unlikely(&userspace_irqchip_in_use))
1021                         kvm_timer_sync_user(vcpu);
1022
1023                 kvm_arch_vcpu_ctxsync_fp(vcpu);
1024
1025                 /*
1026                  * We must ensure that any pending interrupts are taken before
1027                  * we exit guest timing so that timer ticks are accounted as
1028                  * guest time. Transiently unmask interrupts so that any
1029                  * pending interrupts are taken.
1030                  *
1031                  * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other
1032                  * context synchronization event) is necessary to ensure that
1033                  * pending interrupts are taken.
1034                  */
1035                 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) {
1036                         local_irq_enable();
1037                         isb();
1038                         local_irq_disable();
1039                 }
1040
1041                 guest_timing_exit_irqoff();
1042
1043                 local_irq_enable();
1044
1045                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
1046
1047                 /* Exit types that need handling before we can be preempted */
1048                 handle_exit_early(vcpu, ret);
1049
1050                 preempt_enable();
1051
1052                 /*
1053                  * The ARMv8 architecture doesn't give the hypervisor
1054                  * a mechanism to prevent a guest from dropping to AArch32 EL0
1055                  * if implemented by the CPU. If we spot the guest in such
1056                  * state and that we decided it wasn't supposed to do so (like
1057                  * with the asymmetric AArch32 case), return to userspace with
1058                  * a fatal error.
1059                  */
1060                 if (vcpu_mode_is_bad_32bit(vcpu)) {
1061                         /*
1062                          * As we have caught the guest red-handed, decide that
1063                          * it isn't fit for purpose anymore by making the vcpu
1064                          * invalid. The VMM can try and fix it by issuing  a
1065                          * KVM_ARM_VCPU_INIT if it really wants to.
1066                          */
1067                         vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
1068                         ret = ARM_EXCEPTION_IL;
1069                 }
1070
1071                 ret = handle_exit(vcpu, ret);
1072         }
1073
1074         /* Tell userspace about in-kernel device output levels */
1075         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1076                 kvm_timer_update_run(vcpu);
1077                 kvm_pmu_update_run(vcpu);
1078         }
1079
1080         kvm_sigset_deactivate(vcpu);
1081
1082 out:
1083         /*
1084          * In the unlikely event that we are returning to userspace
1085          * with pending exceptions or PC adjustment, commit these
1086          * adjustments in order to give userspace a consistent view of
1087          * the vcpu state. Note that this relies on __kvm_adjust_pc()
1088          * being preempt-safe on VHE.
1089          */
1090         if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) ||
1091                      vcpu_get_flag(vcpu, INCREMENT_PC)))
1092                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
1093
1094         vcpu_put(vcpu);
1095         return ret;
1096 }
1097
1098 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
1099 {
1100         int bit_index;
1101         bool set;
1102         unsigned long *hcr;
1103
1104         if (number == KVM_ARM_IRQ_CPU_IRQ)
1105                 bit_index = __ffs(HCR_VI);
1106         else /* KVM_ARM_IRQ_CPU_FIQ */
1107                 bit_index = __ffs(HCR_VF);
1108
1109         hcr = vcpu_hcr(vcpu);
1110         if (level)
1111                 set = test_and_set_bit(bit_index, hcr);
1112         else
1113                 set = test_and_clear_bit(bit_index, hcr);
1114
1115         /*
1116          * If we didn't change anything, no need to wake up or kick other CPUs
1117          */
1118         if (set == level)
1119                 return 0;
1120
1121         /*
1122          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
1123          * trigger a world-switch round on the running physical CPU to set the
1124          * virtual IRQ/FIQ fields in the HCR appropriately.
1125          */
1126         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
1127         kvm_vcpu_kick(vcpu);
1128
1129         return 0;
1130 }
1131
1132 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
1133                           bool line_status)
1134 {
1135         u32 irq = irq_level->irq;
1136         unsigned int irq_type, vcpu_idx, irq_num;
1137         int nrcpus = atomic_read(&kvm->online_vcpus);
1138         struct kvm_vcpu *vcpu = NULL;
1139         bool level = irq_level->level;
1140
1141         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
1142         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
1143         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
1144         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
1145
1146         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
1147
1148         switch (irq_type) {
1149         case KVM_ARM_IRQ_TYPE_CPU:
1150                 if (irqchip_in_kernel(kvm))
1151                         return -ENXIO;
1152
1153                 if (vcpu_idx >= nrcpus)
1154                         return -EINVAL;
1155
1156                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1157                 if (!vcpu)
1158                         return -EINVAL;
1159
1160                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1161                         return -EINVAL;
1162
1163                 return vcpu_interrupt_line(vcpu, irq_num, level);
1164         case KVM_ARM_IRQ_TYPE_PPI:
1165                 if (!irqchip_in_kernel(kvm))
1166                         return -ENXIO;
1167
1168                 if (vcpu_idx >= nrcpus)
1169                         return -EINVAL;
1170
1171                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1172                 if (!vcpu)
1173                         return -EINVAL;
1174
1175                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1176                         return -EINVAL;
1177
1178                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
1179         case KVM_ARM_IRQ_TYPE_SPI:
1180                 if (!irqchip_in_kernel(kvm))
1181                         return -ENXIO;
1182
1183                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1184                         return -EINVAL;
1185
1186                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
1187         }
1188
1189         return -EINVAL;
1190 }
1191
1192 static unsigned long system_supported_vcpu_features(void)
1193 {
1194         unsigned long features = KVM_VCPU_VALID_FEATURES;
1195
1196         if (!cpus_have_final_cap(ARM64_HAS_32BIT_EL1))
1197                 clear_bit(KVM_ARM_VCPU_EL1_32BIT, &features);
1198
1199         if (!kvm_arm_support_pmu_v3())
1200                 clear_bit(KVM_ARM_VCPU_PMU_V3, &features);
1201
1202         if (!system_supports_sve())
1203                 clear_bit(KVM_ARM_VCPU_SVE, &features);
1204
1205         if (!system_has_full_ptr_auth()) {
1206                 clear_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features);
1207                 clear_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features);
1208         }
1209
1210         if (!cpus_have_final_cap(ARM64_HAS_NESTED_VIRT))
1211                 clear_bit(KVM_ARM_VCPU_HAS_EL2, &features);
1212
1213         return features;
1214 }
1215
1216 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
1217                                         const struct kvm_vcpu_init *init)
1218 {
1219         unsigned long features = init->features[0];
1220         int i;
1221
1222         if (features & ~KVM_VCPU_VALID_FEATURES)
1223                 return -ENOENT;
1224
1225         for (i = 1; i < ARRAY_SIZE(init->features); i++) {
1226                 if (init->features[i])
1227                         return -ENOENT;
1228         }
1229
1230         if (features & ~system_supported_vcpu_features())
1231                 return -EINVAL;
1232
1233         /*
1234          * For now make sure that both address/generic pointer authentication
1235          * features are requested by the userspace together.
1236          */
1237         if (test_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features) !=
1238             test_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features))
1239                 return -EINVAL;
1240
1241         /* Disallow NV+SVE for the time being */
1242         if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features) &&
1243             test_bit(KVM_ARM_VCPU_SVE, &features))
1244                 return -EINVAL;
1245
1246         if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
1247                 return 0;
1248
1249         /* MTE is incompatible with AArch32 */
1250         if (kvm_has_mte(vcpu->kvm))
1251                 return -EINVAL;
1252
1253         /* NV is incompatible with AArch32 */
1254         if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
1255                 return -EINVAL;
1256
1257         return 0;
1258 }
1259
1260 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu,
1261                                   const struct kvm_vcpu_init *init)
1262 {
1263         unsigned long features = init->features[0];
1264
1265         return !bitmap_equal(vcpu->kvm->arch.vcpu_features, &features,
1266                              KVM_VCPU_MAX_FEATURES);
1267 }
1268
1269 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1270                                  const struct kvm_vcpu_init *init)
1271 {
1272         unsigned long features = init->features[0];
1273         struct kvm *kvm = vcpu->kvm;
1274         int ret = -EINVAL;
1275
1276         mutex_lock(&kvm->arch.config_lock);
1277
1278         if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) &&
1279             kvm_vcpu_init_changed(vcpu, init))
1280                 goto out_unlock;
1281
1282         bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES);
1283
1284         /* Now we know what it is, we can reset it. */
1285         kvm_reset_vcpu(vcpu);
1286
1287         set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags);
1288         vcpu_set_flag(vcpu, VCPU_INITIALIZED);
1289         ret = 0;
1290 out_unlock:
1291         mutex_unlock(&kvm->arch.config_lock);
1292         return ret;
1293 }
1294
1295 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1296                                const struct kvm_vcpu_init *init)
1297 {
1298         int ret;
1299
1300         if (init->target != KVM_ARM_TARGET_GENERIC_V8 &&
1301             init->target != kvm_target_cpu())
1302                 return -EINVAL;
1303
1304         ret = kvm_vcpu_init_check_features(vcpu, init);
1305         if (ret)
1306                 return ret;
1307
1308         if (!kvm_vcpu_initialized(vcpu))
1309                 return __kvm_vcpu_set_target(vcpu, init);
1310
1311         if (kvm_vcpu_init_changed(vcpu, init))
1312                 return -EINVAL;
1313
1314         kvm_reset_vcpu(vcpu);
1315         return 0;
1316 }
1317
1318 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1319                                          struct kvm_vcpu_init *init)
1320 {
1321         bool power_off = false;
1322         int ret;
1323
1324         /*
1325          * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid
1326          * reflecting it in the finalized feature set, thus limiting its scope
1327          * to a single KVM_ARM_VCPU_INIT call.
1328          */
1329         if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) {
1330                 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF);
1331                 power_off = true;
1332         }
1333
1334         ret = kvm_vcpu_set_target(vcpu, init);
1335         if (ret)
1336                 return ret;
1337
1338         /*
1339          * Ensure a rebooted VM will fault in RAM pages and detect if the
1340          * guest MMU is turned off and flush the caches as needed.
1341          *
1342          * S2FWB enforces all memory accesses to RAM being cacheable,
1343          * ensuring that the data side is always coherent. We still
1344          * need to invalidate the I-cache though, as FWB does *not*
1345          * imply CTR_EL0.DIC.
1346          */
1347         if (vcpu_has_run_once(vcpu)) {
1348                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1349                         stage2_unmap_vm(vcpu->kvm);
1350                 else
1351                         icache_inval_all_pou();
1352         }
1353
1354         vcpu_reset_hcr(vcpu);
1355         vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu);
1356
1357         /*
1358          * Handle the "start in power-off" case.
1359          */
1360         spin_lock(&vcpu->arch.mp_state_lock);
1361
1362         if (power_off)
1363                 __kvm_arm_vcpu_power_off(vcpu);
1364         else
1365                 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE);
1366
1367         spin_unlock(&vcpu->arch.mp_state_lock);
1368
1369         return 0;
1370 }
1371
1372 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1373                                  struct kvm_device_attr *attr)
1374 {
1375         int ret = -ENXIO;
1376
1377         switch (attr->group) {
1378         default:
1379                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1380                 break;
1381         }
1382
1383         return ret;
1384 }
1385
1386 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1387                                  struct kvm_device_attr *attr)
1388 {
1389         int ret = -ENXIO;
1390
1391         switch (attr->group) {
1392         default:
1393                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1394                 break;
1395         }
1396
1397         return ret;
1398 }
1399
1400 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1401                                  struct kvm_device_attr *attr)
1402 {
1403         int ret = -ENXIO;
1404
1405         switch (attr->group) {
1406         default:
1407                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1408                 break;
1409         }
1410
1411         return ret;
1412 }
1413
1414 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1415                                    struct kvm_vcpu_events *events)
1416 {
1417         memset(events, 0, sizeof(*events));
1418
1419         return __kvm_arm_vcpu_get_events(vcpu, events);
1420 }
1421
1422 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1423                                    struct kvm_vcpu_events *events)
1424 {
1425         int i;
1426
1427         /* check whether the reserved field is zero */
1428         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1429                 if (events->reserved[i])
1430                         return -EINVAL;
1431
1432         /* check whether the pad field is zero */
1433         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1434                 if (events->exception.pad[i])
1435                         return -EINVAL;
1436
1437         return __kvm_arm_vcpu_set_events(vcpu, events);
1438 }
1439
1440 long kvm_arch_vcpu_ioctl(struct file *filp,
1441                          unsigned int ioctl, unsigned long arg)
1442 {
1443         struct kvm_vcpu *vcpu = filp->private_data;
1444         void __user *argp = (void __user *)arg;
1445         struct kvm_device_attr attr;
1446         long r;
1447
1448         switch (ioctl) {
1449         case KVM_ARM_VCPU_INIT: {
1450                 struct kvm_vcpu_init init;
1451
1452                 r = -EFAULT;
1453                 if (copy_from_user(&init, argp, sizeof(init)))
1454                         break;
1455
1456                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1457                 break;
1458         }
1459         case KVM_SET_ONE_REG:
1460         case KVM_GET_ONE_REG: {
1461                 struct kvm_one_reg reg;
1462
1463                 r = -ENOEXEC;
1464                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1465                         break;
1466
1467                 r = -EFAULT;
1468                 if (copy_from_user(&reg, argp, sizeof(reg)))
1469                         break;
1470
1471                 /*
1472                  * We could owe a reset due to PSCI. Handle the pending reset
1473                  * here to ensure userspace register accesses are ordered after
1474                  * the reset.
1475                  */
1476                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1477                         kvm_reset_vcpu(vcpu);
1478
1479                 if (ioctl == KVM_SET_ONE_REG)
1480                         r = kvm_arm_set_reg(vcpu, &reg);
1481                 else
1482                         r = kvm_arm_get_reg(vcpu, &reg);
1483                 break;
1484         }
1485         case KVM_GET_REG_LIST: {
1486                 struct kvm_reg_list __user *user_list = argp;
1487                 struct kvm_reg_list reg_list;
1488                 unsigned n;
1489
1490                 r = -ENOEXEC;
1491                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1492                         break;
1493
1494                 r = -EPERM;
1495                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1496                         break;
1497
1498                 r = -EFAULT;
1499                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1500                         break;
1501                 n = reg_list.n;
1502                 reg_list.n = kvm_arm_num_regs(vcpu);
1503                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1504                         break;
1505                 r = -E2BIG;
1506                 if (n < reg_list.n)
1507                         break;
1508                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1509                 break;
1510         }
1511         case KVM_SET_DEVICE_ATTR: {
1512                 r = -EFAULT;
1513                 if (copy_from_user(&attr, argp, sizeof(attr)))
1514                         break;
1515                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1516                 break;
1517         }
1518         case KVM_GET_DEVICE_ATTR: {
1519                 r = -EFAULT;
1520                 if (copy_from_user(&attr, argp, sizeof(attr)))
1521                         break;
1522                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1523                 break;
1524         }
1525         case KVM_HAS_DEVICE_ATTR: {
1526                 r = -EFAULT;
1527                 if (copy_from_user(&attr, argp, sizeof(attr)))
1528                         break;
1529                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1530                 break;
1531         }
1532         case KVM_GET_VCPU_EVENTS: {
1533                 struct kvm_vcpu_events events;
1534
1535                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1536                         return -EINVAL;
1537
1538                 if (copy_to_user(argp, &events, sizeof(events)))
1539                         return -EFAULT;
1540
1541                 return 0;
1542         }
1543         case KVM_SET_VCPU_EVENTS: {
1544                 struct kvm_vcpu_events events;
1545
1546                 if (copy_from_user(&events, argp, sizeof(events)))
1547                         return -EFAULT;
1548
1549                 return kvm_arm_vcpu_set_events(vcpu, &events);
1550         }
1551         case KVM_ARM_VCPU_FINALIZE: {
1552                 int what;
1553
1554                 if (!kvm_vcpu_initialized(vcpu))
1555                         return -ENOEXEC;
1556
1557                 if (get_user(what, (const int __user *)argp))
1558                         return -EFAULT;
1559
1560                 return kvm_arm_vcpu_finalize(vcpu, what);
1561         }
1562         default:
1563                 r = -EINVAL;
1564         }
1565
1566         return r;
1567 }
1568
1569 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1570 {
1571
1572 }
1573
1574 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1575                                         struct kvm_arm_device_addr *dev_addr)
1576 {
1577         switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) {
1578         case KVM_ARM_DEVICE_VGIC_V2:
1579                 if (!vgic_present)
1580                         return -ENXIO;
1581                 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr);
1582         default:
1583                 return -ENODEV;
1584         }
1585 }
1586
1587 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1588 {
1589         switch (attr->group) {
1590         case KVM_ARM_VM_SMCCC_CTRL:
1591                 return kvm_vm_smccc_has_attr(kvm, attr);
1592         default:
1593                 return -ENXIO;
1594         }
1595 }
1596
1597 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1598 {
1599         switch (attr->group) {
1600         case KVM_ARM_VM_SMCCC_CTRL:
1601                 return kvm_vm_smccc_set_attr(kvm, attr);
1602         default:
1603                 return -ENXIO;
1604         }
1605 }
1606
1607 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
1608 {
1609         struct kvm *kvm = filp->private_data;
1610         void __user *argp = (void __user *)arg;
1611         struct kvm_device_attr attr;
1612
1613         switch (ioctl) {
1614         case KVM_CREATE_IRQCHIP: {
1615                 int ret;
1616                 if (!vgic_present)
1617                         return -ENXIO;
1618                 mutex_lock(&kvm->lock);
1619                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1620                 mutex_unlock(&kvm->lock);
1621                 return ret;
1622         }
1623         case KVM_ARM_SET_DEVICE_ADDR: {
1624                 struct kvm_arm_device_addr dev_addr;
1625
1626                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1627                         return -EFAULT;
1628                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1629         }
1630         case KVM_ARM_PREFERRED_TARGET: {
1631                 struct kvm_vcpu_init init = {
1632                         .target = KVM_ARM_TARGET_GENERIC_V8,
1633                 };
1634
1635                 if (copy_to_user(argp, &init, sizeof(init)))
1636                         return -EFAULT;
1637
1638                 return 0;
1639         }
1640         case KVM_ARM_MTE_COPY_TAGS: {
1641                 struct kvm_arm_copy_mte_tags copy_tags;
1642
1643                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1644                         return -EFAULT;
1645                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1646         }
1647         case KVM_ARM_SET_COUNTER_OFFSET: {
1648                 struct kvm_arm_counter_offset offset;
1649
1650                 if (copy_from_user(&offset, argp, sizeof(offset)))
1651                         return -EFAULT;
1652                 return kvm_vm_ioctl_set_counter_offset(kvm, &offset);
1653         }
1654         case KVM_HAS_DEVICE_ATTR: {
1655                 if (copy_from_user(&attr, argp, sizeof(attr)))
1656                         return -EFAULT;
1657
1658                 return kvm_vm_has_attr(kvm, &attr);
1659         }
1660         case KVM_SET_DEVICE_ATTR: {
1661                 if (copy_from_user(&attr, argp, sizeof(attr)))
1662                         return -EFAULT;
1663
1664                 return kvm_vm_set_attr(kvm, &attr);
1665         }
1666         default:
1667                 return -EINVAL;
1668         }
1669 }
1670
1671 /* unlocks vcpus from @vcpu_lock_idx and smaller */
1672 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx)
1673 {
1674         struct kvm_vcpu *tmp_vcpu;
1675
1676         for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
1677                 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
1678                 mutex_unlock(&tmp_vcpu->mutex);
1679         }
1680 }
1681
1682 void unlock_all_vcpus(struct kvm *kvm)
1683 {
1684         lockdep_assert_held(&kvm->lock);
1685
1686         unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1);
1687 }
1688
1689 /* Returns true if all vcpus were locked, false otherwise */
1690 bool lock_all_vcpus(struct kvm *kvm)
1691 {
1692         struct kvm_vcpu *tmp_vcpu;
1693         unsigned long c;
1694
1695         lockdep_assert_held(&kvm->lock);
1696
1697         /*
1698          * Any time a vcpu is in an ioctl (including running), the
1699          * core KVM code tries to grab the vcpu->mutex.
1700          *
1701          * By grabbing the vcpu->mutex of all VCPUs we ensure that no
1702          * other VCPUs can fiddle with the state while we access it.
1703          */
1704         kvm_for_each_vcpu(c, tmp_vcpu, kvm) {
1705                 if (!mutex_trylock(&tmp_vcpu->mutex)) {
1706                         unlock_vcpus(kvm, c - 1);
1707                         return false;
1708                 }
1709         }
1710
1711         return true;
1712 }
1713
1714 static unsigned long nvhe_percpu_size(void)
1715 {
1716         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1717                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1718 }
1719
1720 static unsigned long nvhe_percpu_order(void)
1721 {
1722         unsigned long size = nvhe_percpu_size();
1723
1724         return size ? get_order(size) : 0;
1725 }
1726
1727 /* A lookup table holding the hypervisor VA for each vector slot */
1728 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1729
1730 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1731 {
1732         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1733 }
1734
1735 static int kvm_init_vector_slots(void)
1736 {
1737         int err;
1738         void *base;
1739
1740         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1741         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1742
1743         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1744         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1745
1746         if (kvm_system_needs_idmapped_vectors() &&
1747             !is_protected_kvm_enabled()) {
1748                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1749                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1750                 if (err)
1751                         return err;
1752         }
1753
1754         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1755         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1756         return 0;
1757 }
1758
1759 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits)
1760 {
1761         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1762         unsigned long tcr;
1763
1764         /*
1765          * Calculate the raw per-cpu offset without a translation from the
1766          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1767          * so that we can use adr_l to access per-cpu variables in EL2.
1768          * Also drop the KASAN tag which gets in the way...
1769          */
1770         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1771                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1772
1773         params->mair_el2 = read_sysreg(mair_el1);
1774
1775         tcr = read_sysreg(tcr_el1);
1776         if (cpus_have_final_cap(ARM64_KVM_HVHE)) {
1777                 tcr |= TCR_EPD1_MASK;
1778         } else {
1779                 tcr &= TCR_EL2_MASK;
1780                 tcr |= TCR_EL2_RES1;
1781         }
1782         tcr &= ~TCR_T0SZ_MASK;
1783         tcr |= TCR_T0SZ(hyp_va_bits);
1784         params->tcr_el2 = tcr;
1785
1786         params->pgd_pa = kvm_mmu_get_httbr();
1787         if (is_protected_kvm_enabled())
1788                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1789         else
1790                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1791         if (cpus_have_final_cap(ARM64_KVM_HVHE))
1792                 params->hcr_el2 |= HCR_E2H;
1793         params->vttbr = params->vtcr = 0;
1794
1795         /*
1796          * Flush the init params from the data cache because the struct will
1797          * be read while the MMU is off.
1798          */
1799         kvm_flush_dcache_to_poc(params, sizeof(*params));
1800 }
1801
1802 static void hyp_install_host_vector(void)
1803 {
1804         struct kvm_nvhe_init_params *params;
1805         struct arm_smccc_res res;
1806
1807         /* Switch from the HYP stub to our own HYP init vector */
1808         __hyp_set_vectors(kvm_get_idmap_vector());
1809
1810         /*
1811          * Call initialization code, and switch to the full blown HYP code.
1812          * If the cpucaps haven't been finalized yet, something has gone very
1813          * wrong, and hyp will crash and burn when it uses any
1814          * cpus_have_const_cap() wrapper.
1815          */
1816         BUG_ON(!system_capabilities_finalized());
1817         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1818         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1819         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1820 }
1821
1822 static void cpu_init_hyp_mode(void)
1823 {
1824         hyp_install_host_vector();
1825
1826         /*
1827          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1828          * at EL2.
1829          */
1830         if (this_cpu_has_cap(ARM64_SSBS) &&
1831             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1832                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1833         }
1834 }
1835
1836 static void cpu_hyp_reset(void)
1837 {
1838         if (!is_kernel_in_hyp_mode())
1839                 __hyp_reset_vectors();
1840 }
1841
1842 /*
1843  * EL2 vectors can be mapped and rerouted in a number of ways,
1844  * depending on the kernel configuration and CPU present:
1845  *
1846  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1847  *   placed in one of the vector slots, which is executed before jumping
1848  *   to the real vectors.
1849  *
1850  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1851  *   containing the hardening sequence is mapped next to the idmap page,
1852  *   and executed before jumping to the real vectors.
1853  *
1854  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1855  *   empty slot is selected, mapped next to the idmap page, and
1856  *   executed before jumping to the real vectors.
1857  *
1858  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1859  * VHE, as we don't have hypervisor-specific mappings. If the system
1860  * is VHE and yet selects this capability, it will be ignored.
1861  */
1862 static void cpu_set_hyp_vector(void)
1863 {
1864         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1865         void *vector = hyp_spectre_vector_selector[data->slot];
1866
1867         if (!is_protected_kvm_enabled())
1868                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1869         else
1870                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1871 }
1872
1873 static void cpu_hyp_init_context(void)
1874 {
1875         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1876
1877         if (!is_kernel_in_hyp_mode())
1878                 cpu_init_hyp_mode();
1879 }
1880
1881 static void cpu_hyp_init_features(void)
1882 {
1883         cpu_set_hyp_vector();
1884         kvm_arm_init_debug();
1885
1886         if (is_kernel_in_hyp_mode())
1887                 kvm_timer_init_vhe();
1888
1889         if (vgic_present)
1890                 kvm_vgic_init_cpu_hardware();
1891 }
1892
1893 static void cpu_hyp_reinit(void)
1894 {
1895         cpu_hyp_reset();
1896         cpu_hyp_init_context();
1897         cpu_hyp_init_features();
1898 }
1899
1900 static void cpu_hyp_init(void *discard)
1901 {
1902         if (!__this_cpu_read(kvm_hyp_initialized)) {
1903                 cpu_hyp_reinit();
1904                 __this_cpu_write(kvm_hyp_initialized, 1);
1905         }
1906 }
1907
1908 static void cpu_hyp_uninit(void *discard)
1909 {
1910         if (__this_cpu_read(kvm_hyp_initialized)) {
1911                 cpu_hyp_reset();
1912                 __this_cpu_write(kvm_hyp_initialized, 0);
1913         }
1914 }
1915
1916 int kvm_arch_hardware_enable(void)
1917 {
1918         /*
1919          * Most calls to this function are made with migration
1920          * disabled, but not with preemption disabled. The former is
1921          * enough to ensure correctness, but most of the helpers
1922          * expect the later and will throw a tantrum otherwise.
1923          */
1924         preempt_disable();
1925
1926         cpu_hyp_init(NULL);
1927
1928         kvm_vgic_cpu_up();
1929         kvm_timer_cpu_up();
1930
1931         preempt_enable();
1932
1933         return 0;
1934 }
1935
1936 void kvm_arch_hardware_disable(void)
1937 {
1938         kvm_timer_cpu_down();
1939         kvm_vgic_cpu_down();
1940
1941         if (!is_protected_kvm_enabled())
1942                 cpu_hyp_uninit(NULL);
1943 }
1944
1945 #ifdef CONFIG_CPU_PM
1946 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1947                                     unsigned long cmd,
1948                                     void *v)
1949 {
1950         /*
1951          * kvm_hyp_initialized is left with its old value over
1952          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1953          * re-enable hyp.
1954          */
1955         switch (cmd) {
1956         case CPU_PM_ENTER:
1957                 if (__this_cpu_read(kvm_hyp_initialized))
1958                         /*
1959                          * don't update kvm_hyp_initialized here
1960                          * so that the hyp will be re-enabled
1961                          * when we resume. See below.
1962                          */
1963                         cpu_hyp_reset();
1964
1965                 return NOTIFY_OK;
1966         case CPU_PM_ENTER_FAILED:
1967         case CPU_PM_EXIT:
1968                 if (__this_cpu_read(kvm_hyp_initialized))
1969                         /* The hyp was enabled before suspend. */
1970                         cpu_hyp_reinit();
1971
1972                 return NOTIFY_OK;
1973
1974         default:
1975                 return NOTIFY_DONE;
1976         }
1977 }
1978
1979 static struct notifier_block hyp_init_cpu_pm_nb = {
1980         .notifier_call = hyp_init_cpu_pm_notifier,
1981 };
1982
1983 static void __init hyp_cpu_pm_init(void)
1984 {
1985         if (!is_protected_kvm_enabled())
1986                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1987 }
1988 static void __init hyp_cpu_pm_exit(void)
1989 {
1990         if (!is_protected_kvm_enabled())
1991                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1992 }
1993 #else
1994 static inline void __init hyp_cpu_pm_init(void)
1995 {
1996 }
1997 static inline void __init hyp_cpu_pm_exit(void)
1998 {
1999 }
2000 #endif
2001
2002 static void __init init_cpu_logical_map(void)
2003 {
2004         unsigned int cpu;
2005
2006         /*
2007          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
2008          * Only copy the set of online CPUs whose features have been checked
2009          * against the finalized system capabilities. The hypervisor will not
2010          * allow any other CPUs from the `possible` set to boot.
2011          */
2012         for_each_online_cpu(cpu)
2013                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
2014 }
2015
2016 #define init_psci_0_1_impl_state(config, what)  \
2017         config.psci_0_1_ ## what ## _implemented = psci_ops.what
2018
2019 static bool __init init_psci_relay(void)
2020 {
2021         /*
2022          * If PSCI has not been initialized, protected KVM cannot install
2023          * itself on newly booted CPUs.
2024          */
2025         if (!psci_ops.get_version) {
2026                 kvm_err("Cannot initialize protected mode without PSCI\n");
2027                 return false;
2028         }
2029
2030         kvm_host_psci_config.version = psci_ops.get_version();
2031         kvm_host_psci_config.smccc_version = arm_smccc_get_version();
2032
2033         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
2034                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
2035                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
2036                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
2037                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
2038                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
2039         }
2040         return true;
2041 }
2042
2043 static int __init init_subsystems(void)
2044 {
2045         int err = 0;
2046
2047         /*
2048          * Enable hardware so that subsystem initialisation can access EL2.
2049          */
2050         on_each_cpu(cpu_hyp_init, NULL, 1);
2051
2052         /*
2053          * Register CPU lower-power notifier
2054          */
2055         hyp_cpu_pm_init();
2056
2057         /*
2058          * Init HYP view of VGIC
2059          */
2060         err = kvm_vgic_hyp_init();
2061         switch (err) {
2062         case 0:
2063                 vgic_present = true;
2064                 break;
2065         case -ENODEV:
2066         case -ENXIO:
2067                 vgic_present = false;
2068                 err = 0;
2069                 break;
2070         default:
2071                 goto out;
2072         }
2073
2074         /*
2075          * Init HYP architected timer support
2076          */
2077         err = kvm_timer_hyp_init(vgic_present);
2078         if (err)
2079                 goto out;
2080
2081         kvm_register_perf_callbacks(NULL);
2082
2083 out:
2084         if (err)
2085                 hyp_cpu_pm_exit();
2086
2087         if (err || !is_protected_kvm_enabled())
2088                 on_each_cpu(cpu_hyp_uninit, NULL, 1);
2089
2090         return err;
2091 }
2092
2093 static void __init teardown_subsystems(void)
2094 {
2095         kvm_unregister_perf_callbacks();
2096         hyp_cpu_pm_exit();
2097 }
2098
2099 static void __init teardown_hyp_mode(void)
2100 {
2101         int cpu;
2102
2103         free_hyp_pgds();
2104         for_each_possible_cpu(cpu) {
2105                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
2106                 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order());
2107         }
2108 }
2109
2110 static int __init do_pkvm_init(u32 hyp_va_bits)
2111 {
2112         void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base));
2113         int ret;
2114
2115         preempt_disable();
2116         cpu_hyp_init_context();
2117         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
2118                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
2119                                 hyp_va_bits);
2120         cpu_hyp_init_features();
2121
2122         /*
2123          * The stub hypercalls are now disabled, so set our local flag to
2124          * prevent a later re-init attempt in kvm_arch_hardware_enable().
2125          */
2126         __this_cpu_write(kvm_hyp_initialized, 1);
2127         preempt_enable();
2128
2129         return ret;
2130 }
2131
2132 static u64 get_hyp_id_aa64pfr0_el1(void)
2133 {
2134         /*
2135          * Track whether the system isn't affected by spectre/meltdown in the
2136          * hypervisor's view of id_aa64pfr0_el1, used for protected VMs.
2137          * Although this is per-CPU, we make it global for simplicity, e.g., not
2138          * to have to worry about vcpu migration.
2139          *
2140          * Unlike for non-protected VMs, userspace cannot override this for
2141          * protected VMs.
2142          */
2143         u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
2144
2145         val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
2146                  ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));
2147
2148         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
2149                           arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED);
2150         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
2151                           arm64_get_meltdown_state() == SPECTRE_UNAFFECTED);
2152
2153         return val;
2154 }
2155
2156 static void kvm_hyp_init_symbols(void)
2157 {
2158         kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1();
2159         kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1);
2160         kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
2161         kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
2162         kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
2163         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
2164         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
2165         kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1);
2166         kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1);
2167         kvm_nvhe_sym(__icache_flags) = __icache_flags;
2168         kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits;
2169 }
2170
2171 static int __init kvm_hyp_init_protection(u32 hyp_va_bits)
2172 {
2173         void *addr = phys_to_virt(hyp_mem_base);
2174         int ret;
2175
2176         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
2177         if (ret)
2178                 return ret;
2179
2180         ret = do_pkvm_init(hyp_va_bits);
2181         if (ret)
2182                 return ret;
2183
2184         free_hyp_pgds();
2185
2186         return 0;
2187 }
2188
2189 static void pkvm_hyp_init_ptrauth(void)
2190 {
2191         struct kvm_cpu_context *hyp_ctxt;
2192         int cpu;
2193
2194         for_each_possible_cpu(cpu) {
2195                 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu);
2196                 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long();
2197                 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long();
2198                 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long();
2199                 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long();
2200                 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long();
2201                 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long();
2202                 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long();
2203                 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long();
2204                 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long();
2205                 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long();
2206         }
2207 }
2208
2209 /* Inits Hyp-mode on all online CPUs */
2210 static int __init init_hyp_mode(void)
2211 {
2212         u32 hyp_va_bits;
2213         int cpu;
2214         int err = -ENOMEM;
2215
2216         /*
2217          * The protected Hyp-mode cannot be initialized if the memory pool
2218          * allocation has failed.
2219          */
2220         if (is_protected_kvm_enabled() && !hyp_mem_base)
2221                 goto out_err;
2222
2223         /*
2224          * Allocate Hyp PGD and setup Hyp identity mapping
2225          */
2226         err = kvm_mmu_init(&hyp_va_bits);
2227         if (err)
2228                 goto out_err;
2229
2230         /*
2231          * Allocate stack pages for Hypervisor-mode
2232          */
2233         for_each_possible_cpu(cpu) {
2234                 unsigned long stack_page;
2235
2236                 stack_page = __get_free_page(GFP_KERNEL);
2237                 if (!stack_page) {
2238                         err = -ENOMEM;
2239                         goto out_err;
2240                 }
2241
2242                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
2243         }
2244
2245         /*
2246          * Allocate and initialize pages for Hypervisor-mode percpu regions.
2247          */
2248         for_each_possible_cpu(cpu) {
2249                 struct page *page;
2250                 void *page_addr;
2251
2252                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
2253                 if (!page) {
2254                         err = -ENOMEM;
2255                         goto out_err;
2256                 }
2257
2258                 page_addr = page_address(page);
2259                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
2260                 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr;
2261         }
2262
2263         /*
2264          * Map the Hyp-code called directly from the host
2265          */
2266         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
2267                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
2268         if (err) {
2269                 kvm_err("Cannot map world-switch code\n");
2270                 goto out_err;
2271         }
2272
2273         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
2274                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
2275         if (err) {
2276                 kvm_err("Cannot map .hyp.rodata section\n");
2277                 goto out_err;
2278         }
2279
2280         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
2281                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
2282         if (err) {
2283                 kvm_err("Cannot map rodata section\n");
2284                 goto out_err;
2285         }
2286
2287         /*
2288          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
2289          * section thanks to an assertion in the linker script. Map it RW and
2290          * the rest of .bss RO.
2291          */
2292         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
2293                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
2294         if (err) {
2295                 kvm_err("Cannot map hyp bss section: %d\n", err);
2296                 goto out_err;
2297         }
2298
2299         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
2300                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
2301         if (err) {
2302                 kvm_err("Cannot map bss section\n");
2303                 goto out_err;
2304         }
2305
2306         /*
2307          * Map the Hyp stack pages
2308          */
2309         for_each_possible_cpu(cpu) {
2310                 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2311                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
2312
2313                 err = create_hyp_stack(__pa(stack_page), &params->stack_hyp_va);
2314                 if (err) {
2315                         kvm_err("Cannot map hyp stack\n");
2316                         goto out_err;
2317                 }
2318
2319                 /*
2320                  * Save the stack PA in nvhe_init_params. This will be needed
2321                  * to recreate the stack mapping in protected nVHE mode.
2322                  * __hyp_pa() won't do the right thing there, since the stack
2323                  * has been mapped in the flexible private VA space.
2324                  */
2325                 params->stack_pa = __pa(stack_page);
2326         }
2327
2328         for_each_possible_cpu(cpu) {
2329                 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu];
2330                 char *percpu_end = percpu_begin + nvhe_percpu_size();
2331
2332                 /* Map Hyp percpu pages */
2333                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
2334                 if (err) {
2335                         kvm_err("Cannot map hyp percpu region\n");
2336                         goto out_err;
2337                 }
2338
2339                 /* Prepare the CPU initialization parameters */
2340                 cpu_prepare_hyp_mode(cpu, hyp_va_bits);
2341         }
2342
2343         kvm_hyp_init_symbols();
2344
2345         if (is_protected_kvm_enabled()) {
2346                 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) &&
2347                     cpus_have_const_cap(ARM64_HAS_ADDRESS_AUTH))
2348                         pkvm_hyp_init_ptrauth();
2349
2350                 init_cpu_logical_map();
2351
2352                 if (!init_psci_relay()) {
2353                         err = -ENODEV;
2354                         goto out_err;
2355                 }
2356
2357                 err = kvm_hyp_init_protection(hyp_va_bits);
2358                 if (err) {
2359                         kvm_err("Failed to init hyp memory protection\n");
2360                         goto out_err;
2361                 }
2362         }
2363
2364         return 0;
2365
2366 out_err:
2367         teardown_hyp_mode();
2368         kvm_err("error initializing Hyp mode: %d\n", err);
2369         return err;
2370 }
2371
2372 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2373 {
2374         struct kvm_vcpu *vcpu;
2375         unsigned long i;
2376
2377         mpidr &= MPIDR_HWID_BITMASK;
2378         kvm_for_each_vcpu(i, vcpu, kvm) {
2379                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2380                         return vcpu;
2381         }
2382         return NULL;
2383 }
2384
2385 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm)
2386 {
2387         return irqchip_in_kernel(kvm);
2388 }
2389
2390 bool kvm_arch_has_irq_bypass(void)
2391 {
2392         return true;
2393 }
2394
2395 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2396                                       struct irq_bypass_producer *prod)
2397 {
2398         struct kvm_kernel_irqfd *irqfd =
2399                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2400
2401         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2402                                           &irqfd->irq_entry);
2403 }
2404 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2405                                       struct irq_bypass_producer *prod)
2406 {
2407         struct kvm_kernel_irqfd *irqfd =
2408                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2409
2410         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2411                                      &irqfd->irq_entry);
2412 }
2413
2414 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2415 {
2416         struct kvm_kernel_irqfd *irqfd =
2417                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2418
2419         kvm_arm_halt_guest(irqfd->kvm);
2420 }
2421
2422 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2423 {
2424         struct kvm_kernel_irqfd *irqfd =
2425                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2426
2427         kvm_arm_resume_guest(irqfd->kvm);
2428 }
2429
2430 /* Initialize Hyp-mode and memory mappings on all CPUs */
2431 static __init int kvm_arm_init(void)
2432 {
2433         int err;
2434         bool in_hyp_mode;
2435
2436         if (!is_hyp_mode_available()) {
2437                 kvm_info("HYP mode not available\n");
2438                 return -ENODEV;
2439         }
2440
2441         if (kvm_get_mode() == KVM_MODE_NONE) {
2442                 kvm_info("KVM disabled from command line\n");
2443                 return -ENODEV;
2444         }
2445
2446         err = kvm_sys_reg_table_init();
2447         if (err) {
2448                 kvm_info("Error initializing system register tables");
2449                 return err;
2450         }
2451
2452         in_hyp_mode = is_kernel_in_hyp_mode();
2453
2454         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2455             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2456                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2457                          "Only trusted guests should be used on this system.\n");
2458
2459         err = kvm_set_ipa_limit();
2460         if (err)
2461                 return err;
2462
2463         err = kvm_arm_init_sve();
2464         if (err)
2465                 return err;
2466
2467         err = kvm_arm_vmid_alloc_init();
2468         if (err) {
2469                 kvm_err("Failed to initialize VMID allocator.\n");
2470                 return err;
2471         }
2472
2473         if (!in_hyp_mode) {
2474                 err = init_hyp_mode();
2475                 if (err)
2476                         goto out_err;
2477         }
2478
2479         err = kvm_init_vector_slots();
2480         if (err) {
2481                 kvm_err("Cannot initialise vector slots\n");
2482                 goto out_hyp;
2483         }
2484
2485         err = init_subsystems();
2486         if (err)
2487                 goto out_hyp;
2488
2489         if (is_protected_kvm_enabled()) {
2490                 kvm_info("Protected nVHE mode initialized successfully\n");
2491         } else if (in_hyp_mode) {
2492                 kvm_info("VHE mode initialized successfully\n");
2493         } else {
2494                 kvm_info("Hyp mode initialized successfully\n");
2495         }
2496
2497         /*
2498          * FIXME: Do something reasonable if kvm_init() fails after pKVM
2499          * hypervisor protection is finalized.
2500          */
2501         err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2502         if (err)
2503                 goto out_subs;
2504
2505         kvm_arm_initialised = true;
2506
2507         return 0;
2508
2509 out_subs:
2510         teardown_subsystems();
2511 out_hyp:
2512         if (!in_hyp_mode)
2513                 teardown_hyp_mode();
2514 out_err:
2515         kvm_arm_vmid_alloc_free();
2516         return err;
2517 }
2518
2519 static int __init early_kvm_mode_cfg(char *arg)
2520 {
2521         if (!arg)
2522                 return -EINVAL;
2523
2524         if (strcmp(arg, "none") == 0) {
2525                 kvm_mode = KVM_MODE_NONE;
2526                 return 0;
2527         }
2528
2529         if (!is_hyp_mode_available()) {
2530                 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n");
2531                 return 0;
2532         }
2533
2534         if (strcmp(arg, "protected") == 0) {
2535                 if (!is_kernel_in_hyp_mode())
2536                         kvm_mode = KVM_MODE_PROTECTED;
2537                 else
2538                         pr_warn_once("Protected KVM not available with VHE\n");
2539
2540                 return 0;
2541         }
2542
2543         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) {
2544                 kvm_mode = KVM_MODE_DEFAULT;
2545                 return 0;
2546         }
2547
2548         if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) {
2549                 kvm_mode = KVM_MODE_NV;
2550                 return 0;
2551         }
2552
2553         return -EINVAL;
2554 }
2555 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2556
2557 enum kvm_mode kvm_get_mode(void)
2558 {
2559         return kvm_mode;
2560 }
2561
2562 module_init(kvm_arm_init);