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_arm_hardware_enabled);
  60 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  61
  62 bool is_kvm_arm_initialised(void)
  63 {
  64         return kvm_arm_initialised;
  65 }
  66
  67 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  68 {
  69         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  70 }
  71
  72 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  73                             struct kvm_enable_cap *cap)
  74 {
  75         int r;
  76         u64 new_cap;
  77
  78         if (cap->flags)
  79                 return -EINVAL;
  80
  81         switch (cap->cap) {
  82         case KVM_CAP_ARM_NISV_TO_USER:
  83                 r = 0;
  84                 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER,
  85                         &kvm->arch.flags);
  86                 break;
  87         case KVM_CAP_ARM_MTE:
  88                 mutex_lock(&kvm->lock);
  89                 if (!system_supports_mte() || kvm->created_vcpus) {
  90                         r = -EINVAL;
  91                 } else {
  92                         r = 0;
  93                         set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags);
  94                 }
  95                 mutex_unlock(&kvm->lock);
  96                 break;
  97         case KVM_CAP_ARM_SYSTEM_SUSPEND:
  98                 r = 0;
  99                 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags);
 100                 break;
 101         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 102                 new_cap = cap->args[0];
 103
 104                 mutex_lock(&kvm->slots_lock);
 105                 /*
 106                  * To keep things simple, allow changing the chunk
 107                  * size only when no memory slots have been created.
 108                  */
 109                 if (!kvm_are_all_memslots_empty(kvm)) {
 110                         r = -EINVAL;
 111                 } else if (new_cap && !kvm_is_block_size_supported(new_cap)) {
 112                         r = -EINVAL;
 113                 } else {
 114                         r = 0;
 115                         kvm->arch.mmu.split_page_chunk_size = new_cap;
 116                 }
 117                 mutex_unlock(&kvm->slots_lock);
 118                 break;
 119         default:
 120                 r = -EINVAL;
 121                 break;
 122         }
 123
 124         return r;
 125 }
 126
 127 static int kvm_arm_default_max_vcpus(void)
 128 {
 129         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 130 }
 131
 132 /**
 133  * kvm_arch_init_vm - initializes a VM data structure
 134  * @kvm:        pointer to the KVM struct
 135  */
 136 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 137 {
 138         int ret;
 139
 140         mutex_init(&kvm->arch.config_lock);
 141
 142 #ifdef CONFIG_LOCKDEP
 143         /* Clue in lockdep that the config_lock must be taken inside kvm->lock */
 144         mutex_lock(&kvm->lock);
 145         mutex_lock(&kvm->arch.config_lock);
 146         mutex_unlock(&kvm->arch.config_lock);
 147         mutex_unlock(&kvm->lock);
 148 #endif
 149
 150         ret = kvm_share_hyp(kvm, kvm + 1);
 151         if (ret)
 152                 return ret;
 153
 154         ret = pkvm_init_host_vm(kvm);
 155         if (ret)
 156                 goto err_unshare_kvm;
 157
 158         if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) {
 159                 ret = -ENOMEM;
 160                 goto err_unshare_kvm;
 161         }
 162         cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask);
 163
 164         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type);
 165         if (ret)
 166                 goto err_free_cpumask;
 167
 168         kvm_vgic_early_init(kvm);
 169
 170         kvm_timer_init_vm(kvm);
 171
 172         /* The maximum number of VCPUs is limited by the host's GIC model */
 173         kvm->max_vcpus = kvm_arm_default_max_vcpus();
 174
 175         kvm_arm_init_hypercalls(kvm);
 176
 177         bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES);
 178
 179         return 0;
 180
 181 err_free_cpumask:
 182         free_cpumask_var(kvm->arch.supported_cpus);
 183 err_unshare_kvm:
 184         kvm_unshare_hyp(kvm, kvm + 1);
 185         return ret;
 186 }
 187
 188 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 189 {
 190         return VM_FAULT_SIGBUS;
 191 }
 192
 193
 194 /**
 195  * kvm_arch_destroy_vm - destroy the VM data structure
 196  * @kvm:        pointer to the KVM struct
 197  */
 198 void kvm_arch_destroy_vm(struct kvm *kvm)
 199 {
 200         bitmap_free(kvm->arch.pmu_filter);
 201         free_cpumask_var(kvm->arch.supported_cpus);
 202
 203         kvm_vgic_destroy(kvm);
 204
 205         if (is_protected_kvm_enabled())
 206                 pkvm_destroy_hyp_vm(kvm);
 207
 208         kvm_destroy_vcpus(kvm);
 209
 210         kvm_unshare_hyp(kvm, kvm + 1);
 211
 212         kvm_arm_teardown_hypercalls(kvm);
 213 }
 214
 215 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 216 {
 217         int r;
 218         switch (ext) {
 219         case KVM_CAP_IRQCHIP:
 220                 r = vgic_present;
 221                 break;
 222         case KVM_CAP_IOEVENTFD:
 223         case KVM_CAP_DEVICE_CTRL:
 224         case KVM_CAP_USER_MEMORY:
 225         case KVM_CAP_SYNC_MMU:
 226         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 227         case KVM_CAP_ONE_REG:
 228         case KVM_CAP_ARM_PSCI:
 229         case KVM_CAP_ARM_PSCI_0_2:
 230         case KVM_CAP_READONLY_MEM:
 231         case KVM_CAP_MP_STATE:
 232         case KVM_CAP_IMMEDIATE_EXIT:
 233         case KVM_CAP_VCPU_EVENTS:
 234         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 235         case KVM_CAP_ARM_NISV_TO_USER:
 236         case KVM_CAP_ARM_INJECT_EXT_DABT:
 237         case KVM_CAP_SET_GUEST_DEBUG:
 238         case KVM_CAP_VCPU_ATTRIBUTES:
 239         case KVM_CAP_PTP_KVM:
 240         case KVM_CAP_ARM_SYSTEM_SUSPEND:
 241         case KVM_CAP_IRQFD_RESAMPLE:
 242         case KVM_CAP_COUNTER_OFFSET:
 243                 r = 1;
 244                 break;
 245         case KVM_CAP_SET_GUEST_DEBUG2:
 246                 return KVM_GUESTDBG_VALID_MASK;
 247         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 248                 r = 1;
 249                 break;
 250         case KVM_CAP_NR_VCPUS:
 251                 /*
 252                  * ARM64 treats KVM_CAP_NR_CPUS differently from all other
 253                  * architectures, as it does not always bound it to
 254                  * KVM_CAP_MAX_VCPUS. It should not matter much because
 255                  * this is just an advisory value.
 256                  */
 257                 r = min_t(unsigned int, num_online_cpus(),
 258                           kvm_arm_default_max_vcpus());
 259                 break;
 260         case KVM_CAP_MAX_VCPUS:
 261         case KVM_CAP_MAX_VCPU_ID:
 262                 if (kvm)
 263                         r = kvm->max_vcpus;
 264                 else
 265                         r = kvm_arm_default_max_vcpus();
 266                 break;
 267         case KVM_CAP_MSI_DEVID:
 268                 if (!kvm)
 269                         r = -EINVAL;
 270                 else
 271                         r = kvm->arch.vgic.msis_require_devid;
 272                 break;
 273         case KVM_CAP_ARM_USER_IRQ:
 274                 /*
 275                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 276                  * (bump this number if adding more devices)
 277                  */
 278                 r = 1;
 279                 break;
 280         case KVM_CAP_ARM_MTE:
 281                 r = system_supports_mte();
 282                 break;
 283         case KVM_CAP_STEAL_TIME:
 284                 r = kvm_arm_pvtime_supported();
 285                 break;
 286         case KVM_CAP_ARM_EL1_32BIT:
 287                 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1);
 288                 break;
 289         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 290                 r = get_num_brps();
 291                 break;
 292         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 293                 r = get_num_wrps();
 294                 break;
 295         case KVM_CAP_ARM_PMU_V3:
 296                 r = kvm_arm_support_pmu_v3();
 297                 break;
 298         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 299                 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN);
 300                 break;
 301         case KVM_CAP_ARM_VM_IPA_SIZE:
 302                 r = get_kvm_ipa_limit();
 303                 break;
 304         case KVM_CAP_ARM_SVE:
 305                 r = system_supports_sve();
 306                 break;
 307         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 308         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 309                 r = system_has_full_ptr_auth();
 310                 break;
 311         case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
 312                 if (kvm)
 313                         r = kvm->arch.mmu.split_page_chunk_size;
 314                 else
 315                         r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
 316                 break;
 317         case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES:
 318                 r = kvm_supported_block_sizes();
 319                 break;
 320         default:
 321                 r = 0;
 322         }
 323
 324         return r;
 325 }
 326
 327 long kvm_arch_dev_ioctl(struct file *filp,
 328                         unsigned int ioctl, unsigned long arg)
 329 {
 330         return -EINVAL;
 331 }
 332
 333 struct kvm *kvm_arch_alloc_vm(void)
 334 {
 335         size_t sz = sizeof(struct kvm);
 336
 337         if (!has_vhe())
 338                 return kzalloc(sz, GFP_KERNEL_ACCOUNT);
 339
 340         return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO);
 341 }
 342
 343 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 344 {
 345         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 346                 return -EBUSY;
 347
 348         if (id >= kvm->max_vcpus)
 349                 return -EINVAL;
 350
 351         return 0;
 352 }
 353
 354 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 355 {
 356         int err;
 357
 358         spin_lock_init(&vcpu->arch.mp_state_lock);
 359
 360 #ifdef CONFIG_LOCKDEP
 361         /* Inform lockdep that the config_lock is acquired after vcpu->mutex */
 362         mutex_lock(&vcpu->mutex);
 363         mutex_lock(&vcpu->kvm->arch.config_lock);
 364         mutex_unlock(&vcpu->kvm->arch.config_lock);
 365         mutex_unlock(&vcpu->mutex);
 366 #endif
 367
 368         /* Force users to call KVM_ARM_VCPU_INIT */
 369         vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
 370         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 371
 372         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 373
 374         /*
 375          * Default value for the FP state, will be overloaded at load
 376          * time if we support FP (pretty likely)
 377          */
 378         vcpu->arch.fp_state = FP_STATE_FREE;
 379
 380         /* Set up the timer */
 381         kvm_timer_vcpu_init(vcpu);
 382
 383         kvm_pmu_vcpu_init(vcpu);
 384
 385         kvm_arm_reset_debug_ptr(vcpu);
 386
 387         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 388
 389         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 390
 391         err = kvm_vgic_vcpu_init(vcpu);
 392         if (err)
 393                 return err;
 394
 395         return kvm_share_hyp(vcpu, vcpu + 1);
 396 }
 397
 398 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 399 {
 400 }
 401
 402 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 403 {
 404         if (vcpu_has_run_once(vcpu) && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 405                 static_branch_dec(&userspace_irqchip_in_use);
 406
 407         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 408         kvm_timer_vcpu_terminate(vcpu);
 409         kvm_pmu_vcpu_destroy(vcpu);
 410
 411         kvm_arm_vcpu_destroy(vcpu);
 412 }
 413
 414 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 415 {
 416
 417 }
 418
 419 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 420 {
 421
 422 }
 423
 424 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 425 {
 426         struct kvm_s2_mmu *mmu;
 427         int *last_ran;
 428
 429         mmu = vcpu->arch.hw_mmu;
 430         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 431
 432         /*
 433          * We guarantee that both TLBs and I-cache are private to each
 434          * vcpu. If detecting that a vcpu from the same VM has
 435          * previously run on the same physical CPU, call into the
 436          * hypervisor code to nuke the relevant contexts.
 437          *
 438          * We might get preempted before the vCPU actually runs, but
 439          * over-invalidation doesn't affect correctness.
 440          */
 441         if (*last_ran != vcpu->vcpu_id) {
 442                 kvm_call_hyp(__kvm_flush_cpu_context, mmu);
 443                 *last_ran = vcpu->vcpu_id;
 444         }
 445
 446         vcpu->cpu = cpu;
 447
 448         kvm_vgic_load(vcpu);
 449         kvm_timer_vcpu_load(vcpu);
 450         if (has_vhe())
 451                 kvm_vcpu_load_sysregs_vhe(vcpu);
 452         kvm_arch_vcpu_load_fp(vcpu);
 453         kvm_vcpu_pmu_restore_guest(vcpu);
 454         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 455                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 456
 457         if (single_task_running())
 458                 vcpu_clear_wfx_traps(vcpu);
 459         else
 460                 vcpu_set_wfx_traps(vcpu);
 461
 462         if (vcpu_has_ptrauth(vcpu))
 463                 vcpu_ptrauth_disable(vcpu);
 464         kvm_arch_vcpu_load_debug_state_flags(vcpu);
 465
 466         if (!cpumask_test_cpu(smp_processor_id(), vcpu->kvm->arch.supported_cpus))
 467                 vcpu_set_on_unsupported_cpu(vcpu);
 468 }
 469
 470 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 471 {
 472         kvm_arch_vcpu_put_debug_state_flags(vcpu);
 473         kvm_arch_vcpu_put_fp(vcpu);
 474         if (has_vhe())
 475                 kvm_vcpu_put_sysregs_vhe(vcpu);
 476         kvm_timer_vcpu_put(vcpu);
 477         kvm_vgic_put(vcpu);
 478         kvm_vcpu_pmu_restore_host(vcpu);
 479         kvm_arm_vmid_clear_active();
 480
 481         vcpu_clear_on_unsupported_cpu(vcpu);
 482         vcpu->cpu = -1;
 483 }
 484
 485 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 486 {
 487         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED);
 488         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 489         kvm_vcpu_kick(vcpu);
 490 }
 491
 492 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 493 {
 494         spin_lock(&vcpu->arch.mp_state_lock);
 495         __kvm_arm_vcpu_power_off(vcpu);
 496         spin_unlock(&vcpu->arch.mp_state_lock);
 497 }
 498
 499 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu)
 500 {
 501         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED;
 502 }
 503
 504 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu)
 505 {
 506         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED);
 507         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 508         kvm_vcpu_kick(vcpu);
 509 }
 510
 511 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu)
 512 {
 513         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED;
 514 }
 515
 516 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 517                                     struct kvm_mp_state *mp_state)
 518 {
 519         *mp_state = READ_ONCE(vcpu->arch.mp_state);
 520
 521         return 0;
 522 }
 523
 524 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 525                                     struct kvm_mp_state *mp_state)
 526 {
 527         int ret = 0;
 528
 529         spin_lock(&vcpu->arch.mp_state_lock);
 530
 531         switch (mp_state->mp_state) {
 532         case KVM_MP_STATE_RUNNABLE:
 533                 WRITE_ONCE(vcpu->arch.mp_state, *mp_state);
 534                 break;
 535         case KVM_MP_STATE_STOPPED:
 536                 __kvm_arm_vcpu_power_off(vcpu);
 537                 break;
 538         case KVM_MP_STATE_SUSPENDED:
 539                 kvm_arm_vcpu_suspend(vcpu);
 540                 break;
 541         default:
 542                 ret = -EINVAL;
 543         }
 544
 545         spin_unlock(&vcpu->arch.mp_state_lock);
 546
 547         return ret;
 548 }
 549
 550 /**
 551  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 552  * @v:          The VCPU pointer
 553  *
 554  * If the guest CPU is not waiting for interrupts or an interrupt line is
 555  * asserted, the CPU is by definition runnable.
 556  */
 557 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 558 {
 559         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 560         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 561                 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause);
 562 }
 563
 564 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 565 {
 566         return vcpu_mode_priv(vcpu);
 567 }
 568
 569 #ifdef CONFIG_GUEST_PERF_EVENTS
 570 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu)
 571 {
 572         return *vcpu_pc(vcpu);
 573 }
 574 #endif
 575
 576 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 577 {
 578         return vcpu_get_flag(vcpu, VCPU_INITIALIZED);
 579 }
 580
 581 /*
 582  * Handle both the initialisation that is being done when the vcpu is
 583  * run for the first time, as well as the updates that must be
 584  * performed each time we get a new thread dealing with this vcpu.
 585  */
 586 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
 587 {
 588         struct kvm *kvm = vcpu->kvm;
 589         int ret;
 590
 591         if (!kvm_vcpu_initialized(vcpu))
 592                 return -ENOEXEC;
 593
 594         if (!kvm_arm_vcpu_is_finalized(vcpu))
 595                 return -EPERM;
 596
 597         ret = kvm_arch_vcpu_run_map_fp(vcpu);
 598         if (ret)
 599                 return ret;
 600
 601         if (likely(vcpu_has_run_once(vcpu)))
 602                 return 0;
 603
 604         kvm_arm_vcpu_init_debug(vcpu);
 605
 606         if (likely(irqchip_in_kernel(kvm))) {
 607                 /*
 608                  * Map the VGIC hardware resources before running a vcpu the
 609                  * first time on this VM.
 610                  */
 611                 ret = kvm_vgic_map_resources(kvm);
 612                 if (ret)
 613                         return ret;
 614         }
 615
 616         ret = kvm_timer_enable(vcpu);
 617         if (ret)
 618                 return ret;
 619
 620         ret = kvm_arm_pmu_v3_enable(vcpu);
 621         if (ret)
 622                 return ret;
 623
 624         if (is_protected_kvm_enabled()) {
 625                 ret = pkvm_create_hyp_vm(kvm);
 626                 if (ret)
 627                         return ret;
 628         }
 629
 630         if (!irqchip_in_kernel(kvm)) {
 631                 /*
 632                  * Tell the rest of the code that there are userspace irqchip
 633                  * VMs in the wild.
 634                  */
 635                 static_branch_inc(&userspace_irqchip_in_use);
 636         }
 637
 638         /*
 639          * Initialize traps for protected VMs.
 640          * NOTE: Move to run in EL2 directly, rather than via a hypercall, once
 641          * the code is in place for first run initialization at EL2.
 642          */
 643         if (kvm_vm_is_protected(kvm))
 644                 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu);
 645
 646         mutex_lock(&kvm->arch.config_lock);
 647         set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags);
 648         mutex_unlock(&kvm->arch.config_lock);
 649
 650         return ret;
 651 }
 652
 653 bool kvm_arch_intc_initialized(struct kvm *kvm)
 654 {
 655         return vgic_initialized(kvm);
 656 }
 657
 658 void kvm_arm_halt_guest(struct kvm *kvm)
 659 {
 660         unsigned long i;
 661         struct kvm_vcpu *vcpu;
 662
 663         kvm_for_each_vcpu(i, vcpu, kvm)
 664                 vcpu->arch.pause = true;
 665         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 666 }
 667
 668 void kvm_arm_resume_guest(struct kvm *kvm)
 669 {
 670         unsigned long i;
 671         struct kvm_vcpu *vcpu;
 672
 673         kvm_for_each_vcpu(i, vcpu, kvm) {
 674                 vcpu->arch.pause = false;
 675                 __kvm_vcpu_wake_up(vcpu);
 676         }
 677 }
 678
 679 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu)
 680 {
 681         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 682
 683         rcuwait_wait_event(wait,
 684                            (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause),
 685                            TASK_INTERRUPTIBLE);
 686
 687         if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) {
 688                 /* Awaken to handle a signal, request we sleep again later. */
 689                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 690         }
 691
 692         /*
 693          * Make sure we will observe a potential reset request if we've
 694          * observed a change to the power state. Pairs with the smp_wmb() in
 695          * kvm_psci_vcpu_on().
 696          */
 697         smp_rmb();
 698 }
 699
 700 /**
 701  * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior
 702  * @vcpu:       The VCPU pointer
 703  *
 704  * Suspend execution of a vCPU until a valid wake event is detected, i.e. until
 705  * the vCPU is runnable.  The vCPU may or may not be scheduled out, depending
 706  * on when a wake event arrives, e.g. there may already be a pending wake event.
 707  */
 708 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu)
 709 {
 710         /*
 711          * Sync back the state of the GIC CPU interface so that we have
 712          * the latest PMR and group enables. This ensures that
 713          * kvm_arch_vcpu_runnable has up-to-date data to decide whether
 714          * we have pending interrupts, e.g. when determining if the
 715          * vCPU should block.
 716          *
 717          * For the same reason, we want to tell GICv4 that we need
 718          * doorbells to be signalled, should an interrupt become pending.
 719          */
 720         preempt_disable();
 721         kvm_vgic_vmcr_sync(vcpu);
 722         vcpu_set_flag(vcpu, IN_WFI);
 723         vgic_v4_put(vcpu);
 724         preempt_enable();
 725
 726         kvm_vcpu_halt(vcpu);
 727         vcpu_clear_flag(vcpu, IN_WFIT);
 728
 729         preempt_disable();
 730         vcpu_clear_flag(vcpu, IN_WFI);
 731         vgic_v4_load(vcpu);
 732         preempt_enable();
 733 }
 734
 735 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu)
 736 {
 737         if (!kvm_arm_vcpu_suspended(vcpu))
 738                 return 1;
 739
 740         kvm_vcpu_wfi(vcpu);
 741
 742         /*
 743          * The suspend state is sticky; we do not leave it until userspace
 744          * explicitly marks the vCPU as runnable. Request that we suspend again
 745          * later.
 746          */
 747         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 748
 749         /*
 750          * Check to make sure the vCPU is actually runnable. If so, exit to
 751          * userspace informing it of the wakeup condition.
 752          */
 753         if (kvm_arch_vcpu_runnable(vcpu)) {
 754                 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event));
 755                 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP;
 756                 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
 757                 return 0;
 758         }
 759
 760         /*
 761          * Otherwise, we were unblocked to process a different event, such as a
 762          * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to
 763          * process the event.
 764          */
 765         return 1;
 766 }
 767
 768 /**
 769  * check_vcpu_requests - check and handle pending vCPU requests
 770  * @vcpu:       the VCPU pointer
 771  *
 772  * Return: 1 if we should enter the guest
 773  *         0 if we should exit to userspace
 774  *         < 0 if we should exit to userspace, where the return value indicates
 775  *         an error
 776  */
 777 static int check_vcpu_requests(struct kvm_vcpu *vcpu)
 778 {
 779         if (kvm_request_pending(vcpu)) {
 780                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 781                         kvm_vcpu_sleep(vcpu);
 782
 783                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 784                         kvm_reset_vcpu(vcpu);
 785
 786                 /*
 787                  * Clear IRQ_PENDING requests that were made to guarantee
 788                  * that a VCPU sees new virtual interrupts.
 789                  */
 790                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 791
 792                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 793                         kvm_update_stolen_time(vcpu);
 794
 795                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 796                         /* The distributor enable bits were changed */
 797                         preempt_disable();
 798                         vgic_v4_put(vcpu);
 799                         vgic_v4_load(vcpu);
 800                         preempt_enable();
 801                 }
 802
 803                 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
 804                         kvm_pmu_handle_pmcr(vcpu,
 805                                             __vcpu_sys_reg(vcpu, PMCR_EL0));
 806
 807                 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu))
 808                         return kvm_vcpu_suspend(vcpu);
 809
 810                 if (kvm_dirty_ring_check_request(vcpu))
 811                         return 0;
 812         }
 813
 814         return 1;
 815 }
 816
 817 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 818 {
 819         if (likely(!vcpu_mode_is_32bit(vcpu)))
 820                 return false;
 821
 822         if (vcpu_has_nv(vcpu))
 823                 return true;
 824
 825         return !kvm_supports_32bit_el0();
 826 }
 827
 828 /**
 829  * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest
 830  * @vcpu:       The VCPU pointer
 831  * @ret:        Pointer to write optional return code
 832  *
 833  * Returns: true if the VCPU needs to return to a preemptible + interruptible
 834  *          and skip guest entry.
 835  *
 836  * This function disambiguates between two different types of exits: exits to a
 837  * preemptible + interruptible kernel context and exits to userspace. For an
 838  * exit to userspace, this function will write the return code to ret and return
 839  * true. For an exit to preemptible + interruptible kernel context (i.e. check
 840  * for pending work and re-enter), return true without writing to ret.
 841  */
 842 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret)
 843 {
 844         struct kvm_run *run = vcpu->run;
 845
 846         /*
 847          * If we're using a userspace irqchip, then check if we need
 848          * to tell a userspace irqchip about timer or PMU level
 849          * changes and if so, exit to userspace (the actual level
 850          * state gets updated in kvm_timer_update_run and
 851          * kvm_pmu_update_run below).
 852          */
 853         if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 854                 if (kvm_timer_should_notify_user(vcpu) ||
 855                     kvm_pmu_should_notify_user(vcpu)) {
 856                         *ret = -EINTR;
 857                         run->exit_reason = KVM_EXIT_INTR;
 858                         return true;
 859                 }
 860         }
 861
 862         if (unlikely(vcpu_on_unsupported_cpu(vcpu))) {
 863                 run->exit_reason = KVM_EXIT_FAIL_ENTRY;
 864                 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED;
 865                 run->fail_entry.cpu = smp_processor_id();
 866                 *ret = 0;
 867                 return true;
 868         }
 869
 870         return kvm_request_pending(vcpu) ||
 871                         xfer_to_guest_mode_work_pending();
 872 }
 873
 874 /*
 875  * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while
 876  * the vCPU is running.
 877  *
 878  * This must be noinstr as instrumentation may make use of RCU, and this is not
 879  * safe during the EQS.
 880  */
 881 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
 882 {
 883         int ret;
 884
 885         guest_state_enter_irqoff();
 886         ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 887         guest_state_exit_irqoff();
 888
 889         return ret;
 890 }
 891
 892 /**
 893  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 894  * @vcpu:       The VCPU pointer
 895  *
 896  * This function is called through the VCPU_RUN ioctl called from user space. It
 897  * will execute VM code in a loop until the time slice for the process is used
 898  * or some emulation is needed from user space in which case the function will
 899  * return with return value 0 and with the kvm_run structure filled in with the
 900  * required data for the requested emulation.
 901  */
 902 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 903 {
 904         struct kvm_run *run = vcpu->run;
 905         int ret;
 906
 907         if (run->exit_reason == KVM_EXIT_MMIO) {
 908                 ret = kvm_handle_mmio_return(vcpu);
 909                 if (ret)
 910                         return ret;
 911         }
 912
 913         vcpu_load(vcpu);
 914
 915         if (run->immediate_exit) {
 916                 ret = -EINTR;
 917                 goto out;
 918         }
 919
 920         kvm_sigset_activate(vcpu);
 921
 922         ret = 1;
 923         run->exit_reason = KVM_EXIT_UNKNOWN;
 924         run->flags = 0;
 925         while (ret > 0) {
 926                 /*
 927                  * Check conditions before entering the guest
 928                  */
 929                 ret = xfer_to_guest_mode_handle_work(vcpu);
 930                 if (!ret)
 931                         ret = 1;
 932
 933                 if (ret > 0)
 934                         ret = check_vcpu_requests(vcpu);
 935
 936                 /*
 937                  * Preparing the interrupts to be injected also
 938                  * involves poking the GIC, which must be done in a
 939                  * non-preemptible context.
 940                  */
 941                 preempt_disable();
 942
 943                 /*
 944                  * The VMID allocator only tracks active VMIDs per
 945                  * physical CPU, and therefore the VMID allocated may not be
 946                  * preserved on VMID roll-over if the task was preempted,
 947                  * making a thread's VMID inactive. So we need to call
 948                  * kvm_arm_vmid_update() in non-premptible context.
 949                  */
 950                 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid);
 951
 952                 kvm_pmu_flush_hwstate(vcpu);
 953
 954                 local_irq_disable();
 955
 956                 kvm_vgic_flush_hwstate(vcpu);
 957
 958                 kvm_pmu_update_vcpu_events(vcpu);
 959
 960                 /*
 961                  * Ensure we set mode to IN_GUEST_MODE after we disable
 962                  * interrupts and before the final VCPU requests check.
 963                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 964                  * Documentation/virt/kvm/vcpu-requests.rst
 965                  */
 966                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 967
 968                 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) {
 969                         vcpu->mode = OUTSIDE_GUEST_MODE;
 970                         isb(); /* Ensure work in x_flush_hwstate is committed */
 971                         kvm_pmu_sync_hwstate(vcpu);
 972                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 973                                 kvm_timer_sync_user(vcpu);
 974                         kvm_vgic_sync_hwstate(vcpu);
 975                         local_irq_enable();
 976                         preempt_enable();
 977                         continue;
 978                 }
 979
 980                 kvm_arm_setup_debug(vcpu);
 981                 kvm_arch_vcpu_ctxflush_fp(vcpu);
 982
 983                 /**************************************************************
 984                  * Enter the guest
 985                  */
 986                 trace_kvm_entry(*vcpu_pc(vcpu));
 987                 guest_timing_enter_irqoff();
 988
 989                 ret = kvm_arm_vcpu_enter_exit(vcpu);
 990
 991                 vcpu->mode = OUTSIDE_GUEST_MODE;
 992                 vcpu->stat.exits++;
 993                 /*
 994                  * Back from guest
 995                  *************************************************************/
 996
 997                 kvm_arm_clear_debug(vcpu);
 998
 999                 /*
1000                  * We must sync the PMU state before the vgic state so
1001                  * that the vgic can properly sample the updated state of the
1002                  * interrupt line.
1003                  */
1004                 kvm_pmu_sync_hwstate(vcpu);
1005
1006                 /*
1007                  * Sync the vgic state before syncing the timer state because
1008                  * the timer code needs to know if the virtual timer
1009                  * interrupts are active.
1010                  */
1011                 kvm_vgic_sync_hwstate(vcpu);
1012
1013                 /*
1014                  * Sync the timer hardware state before enabling interrupts as
1015                  * we don't want vtimer interrupts to race with syncing the
1016                  * timer virtual interrupt state.
1017                  */
1018                 if (static_branch_unlikely(&userspace_irqchip_in_use))
1019                         kvm_timer_sync_user(vcpu);
1020
1021                 kvm_arch_vcpu_ctxsync_fp(vcpu);
1022
1023                 /*
1024                  * We must ensure that any pending interrupts are taken before
1025                  * we exit guest timing so that timer ticks are accounted as
1026                  * guest time. Transiently unmask interrupts so that any
1027                  * pending interrupts are taken.
1028                  *
1029                  * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other
1030                  * context synchronization event) is necessary to ensure that
1031                  * pending interrupts are taken.
1032                  */
1033                 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) {
1034                         local_irq_enable();
1035                         isb();
1036                         local_irq_disable();
1037                 }
1038
1039                 guest_timing_exit_irqoff();
1040
1041                 local_irq_enable();
1042
1043                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
1044
1045                 /* Exit types that need handling before we can be preempted */
1046                 handle_exit_early(vcpu, ret);
1047
1048                 preempt_enable();
1049
1050                 /*
1051                  * The ARMv8 architecture doesn't give the hypervisor
1052                  * a mechanism to prevent a guest from dropping to AArch32 EL0
1053                  * if implemented by the CPU. If we spot the guest in such
1054                  * state and that we decided it wasn't supposed to do so (like
1055                  * with the asymmetric AArch32 case), return to userspace with
1056                  * a fatal error.
1057                  */
1058                 if (vcpu_mode_is_bad_32bit(vcpu)) {
1059                         /*
1060                          * As we have caught the guest red-handed, decide that
1061                          * it isn't fit for purpose anymore by making the vcpu
1062                          * invalid. The VMM can try and fix it by issuing  a
1063                          * KVM_ARM_VCPU_INIT if it really wants to.
1064                          */
1065                         vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
1066                         ret = ARM_EXCEPTION_IL;
1067                 }
1068
1069                 ret = handle_exit(vcpu, ret);
1070         }
1071
1072         /* Tell userspace about in-kernel device output levels */
1073         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1074                 kvm_timer_update_run(vcpu);
1075                 kvm_pmu_update_run(vcpu);
1076         }
1077
1078         kvm_sigset_deactivate(vcpu);
1079
1080 out:
1081         /*
1082          * In the unlikely event that we are returning to userspace
1083          * with pending exceptions or PC adjustment, commit these
1084          * adjustments in order to give userspace a consistent view of
1085          * the vcpu state. Note that this relies on __kvm_adjust_pc()
1086          * being preempt-safe on VHE.
1087          */
1088         if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) ||
1089                      vcpu_get_flag(vcpu, INCREMENT_PC)))
1090                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
1091
1092         vcpu_put(vcpu);
1093         return ret;
1094 }
1095
1096 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
1097 {
1098         int bit_index;
1099         bool set;
1100         unsigned long *hcr;
1101
1102         if (number == KVM_ARM_IRQ_CPU_IRQ)
1103                 bit_index = __ffs(HCR_VI);
1104         else /* KVM_ARM_IRQ_CPU_FIQ */
1105                 bit_index = __ffs(HCR_VF);
1106
1107         hcr = vcpu_hcr(vcpu);
1108         if (level)
1109                 set = test_and_set_bit(bit_index, hcr);
1110         else
1111                 set = test_and_clear_bit(bit_index, hcr);
1112
1113         /*
1114          * If we didn't change anything, no need to wake up or kick other CPUs
1115          */
1116         if (set == level)
1117                 return 0;
1118
1119         /*
1120          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
1121          * trigger a world-switch round on the running physical CPU to set the
1122          * virtual IRQ/FIQ fields in the HCR appropriately.
1123          */
1124         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
1125         kvm_vcpu_kick(vcpu);
1126
1127         return 0;
1128 }
1129
1130 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
1131                           bool line_status)
1132 {
1133         u32 irq = irq_level->irq;
1134         unsigned int irq_type, vcpu_idx, irq_num;
1135         int nrcpus = atomic_read(&kvm->online_vcpus);
1136         struct kvm_vcpu *vcpu = NULL;
1137         bool level = irq_level->level;
1138
1139         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
1140         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
1141         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
1142         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
1143
1144         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
1145
1146         switch (irq_type) {
1147         case KVM_ARM_IRQ_TYPE_CPU:
1148                 if (irqchip_in_kernel(kvm))
1149                         return -ENXIO;
1150
1151                 if (vcpu_idx >= nrcpus)
1152                         return -EINVAL;
1153
1154                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1155                 if (!vcpu)
1156                         return -EINVAL;
1157
1158                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1159                         return -EINVAL;
1160
1161                 return vcpu_interrupt_line(vcpu, irq_num, level);
1162         case KVM_ARM_IRQ_TYPE_PPI:
1163                 if (!irqchip_in_kernel(kvm))
1164                         return -ENXIO;
1165
1166                 if (vcpu_idx >= nrcpus)
1167                         return -EINVAL;
1168
1169                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1170                 if (!vcpu)
1171                         return -EINVAL;
1172
1173                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1174                         return -EINVAL;
1175
1176                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
1177         case KVM_ARM_IRQ_TYPE_SPI:
1178                 if (!irqchip_in_kernel(kvm))
1179                         return -ENXIO;
1180
1181                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1182                         return -EINVAL;
1183
1184                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
1185         }
1186
1187         return -EINVAL;
1188 }
1189
1190 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
1191                                         const struct kvm_vcpu_init *init)
1192 {
1193         unsigned long features = init->features[0];
1194         int i;
1195
1196         if (features & ~KVM_VCPU_VALID_FEATURES)
1197                 return -ENOENT;
1198
1199         for (i = 1; i < ARRAY_SIZE(init->features); i++) {
1200                 if (init->features[i])
1201                         return -ENOENT;
1202         }
1203
1204         if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
1205                 return 0;
1206
1207         if (!cpus_have_const_cap(ARM64_HAS_32BIT_EL1))
1208                 return -EINVAL;
1209
1210         /* MTE is incompatible with AArch32 */
1211         if (kvm_has_mte(vcpu->kvm))
1212                 return -EINVAL;
1213
1214         /* NV is incompatible with AArch32 */
1215         if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
1216                 return -EINVAL;
1217
1218         return 0;
1219 }
1220
1221 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu,
1222                                   const struct kvm_vcpu_init *init)
1223 {
1224         unsigned long features = init->features[0];
1225
1226         return !bitmap_equal(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES);
1227 }
1228
1229 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1230                                  const struct kvm_vcpu_init *init)
1231 {
1232         unsigned long features = init->features[0];
1233         struct kvm *kvm = vcpu->kvm;
1234         int ret = -EINVAL;
1235
1236         mutex_lock(&kvm->arch.config_lock);
1237
1238         if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) &&
1239             !bitmap_equal(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES))
1240                 goto out_unlock;
1241
1242         bitmap_copy(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES);
1243
1244         /* Now we know what it is, we can reset it. */
1245         ret = kvm_reset_vcpu(vcpu);
1246         if (ret) {
1247                 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
1248                 goto out_unlock;
1249         }
1250
1251         bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES);
1252         set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags);
1253         vcpu_set_flag(vcpu, VCPU_INITIALIZED);
1254 out_unlock:
1255         mutex_unlock(&kvm->arch.config_lock);
1256         return ret;
1257 }
1258
1259 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1260                                const struct kvm_vcpu_init *init)
1261 {
1262         int ret;
1263
1264         if (init->target != KVM_ARM_TARGET_GENERIC_V8 &&
1265             init->target != kvm_target_cpu())
1266                 return -EINVAL;
1267
1268         ret = kvm_vcpu_init_check_features(vcpu, init);
1269         if (ret)
1270                 return ret;
1271
1272         if (!kvm_vcpu_initialized(vcpu))
1273                 return __kvm_vcpu_set_target(vcpu, init);
1274
1275         if (kvm_vcpu_init_changed(vcpu, init))
1276                 return -EINVAL;
1277
1278         return kvm_reset_vcpu(vcpu);
1279 }
1280
1281 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1282                                          struct kvm_vcpu_init *init)
1283 {
1284         bool power_off = false;
1285         int ret;
1286
1287         /*
1288          * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid
1289          * reflecting it in the finalized feature set, thus limiting its scope
1290          * to a single KVM_ARM_VCPU_INIT call.
1291          */
1292         if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) {
1293                 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF);
1294                 power_off = true;
1295         }
1296
1297         ret = kvm_vcpu_set_target(vcpu, init);
1298         if (ret)
1299                 return ret;
1300
1301         /*
1302          * Ensure a rebooted VM will fault in RAM pages and detect if the
1303          * guest MMU is turned off and flush the caches as needed.
1304          *
1305          * S2FWB enforces all memory accesses to RAM being cacheable,
1306          * ensuring that the data side is always coherent. We still
1307          * need to invalidate the I-cache though, as FWB does *not*
1308          * imply CTR_EL0.DIC.
1309          */
1310         if (vcpu_has_run_once(vcpu)) {
1311                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1312                         stage2_unmap_vm(vcpu->kvm);
1313                 else
1314                         icache_inval_all_pou();
1315         }
1316
1317         vcpu_reset_hcr(vcpu);
1318         vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu);
1319
1320         /*
1321          * Handle the "start in power-off" case.
1322          */
1323         spin_lock(&vcpu->arch.mp_state_lock);
1324
1325         if (power_off)
1326                 __kvm_arm_vcpu_power_off(vcpu);
1327         else
1328                 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE);
1329
1330         spin_unlock(&vcpu->arch.mp_state_lock);
1331
1332         return 0;
1333 }
1334
1335 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1336                                  struct kvm_device_attr *attr)
1337 {
1338         int ret = -ENXIO;
1339
1340         switch (attr->group) {
1341         default:
1342                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1343                 break;
1344         }
1345
1346         return ret;
1347 }
1348
1349 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1350                                  struct kvm_device_attr *attr)
1351 {
1352         int ret = -ENXIO;
1353
1354         switch (attr->group) {
1355         default:
1356                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1357                 break;
1358         }
1359
1360         return ret;
1361 }
1362
1363 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1364                                  struct kvm_device_attr *attr)
1365 {
1366         int ret = -ENXIO;
1367
1368         switch (attr->group) {
1369         default:
1370                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1371                 break;
1372         }
1373
1374         return ret;
1375 }
1376
1377 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1378                                    struct kvm_vcpu_events *events)
1379 {
1380         memset(events, 0, sizeof(*events));
1381
1382         return __kvm_arm_vcpu_get_events(vcpu, events);
1383 }
1384
1385 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1386                                    struct kvm_vcpu_events *events)
1387 {
1388         int i;
1389
1390         /* check whether the reserved field is zero */
1391         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1392                 if (events->reserved[i])
1393                         return -EINVAL;
1394
1395         /* check whether the pad field is zero */
1396         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1397                 if (events->exception.pad[i])
1398                         return -EINVAL;
1399
1400         return __kvm_arm_vcpu_set_events(vcpu, events);
1401 }
1402
1403 long kvm_arch_vcpu_ioctl(struct file *filp,
1404                          unsigned int ioctl, unsigned long arg)
1405 {
1406         struct kvm_vcpu *vcpu = filp->private_data;
1407         void __user *argp = (void __user *)arg;
1408         struct kvm_device_attr attr;
1409         long r;
1410
1411         switch (ioctl) {
1412         case KVM_ARM_VCPU_INIT: {
1413                 struct kvm_vcpu_init init;
1414
1415                 r = -EFAULT;
1416                 if (copy_from_user(&init, argp, sizeof(init)))
1417                         break;
1418
1419                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1420                 break;
1421         }
1422         case KVM_SET_ONE_REG:
1423         case KVM_GET_ONE_REG: {
1424                 struct kvm_one_reg reg;
1425
1426                 r = -ENOEXEC;
1427                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1428                         break;
1429
1430                 r = -EFAULT;
1431                 if (copy_from_user(&reg, argp, sizeof(reg)))
1432                         break;
1433
1434                 /*
1435                  * We could owe a reset due to PSCI. Handle the pending reset
1436                  * here to ensure userspace register accesses are ordered after
1437                  * the reset.
1438                  */
1439                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1440                         kvm_reset_vcpu(vcpu);
1441
1442                 if (ioctl == KVM_SET_ONE_REG)
1443                         r = kvm_arm_set_reg(vcpu, &reg);
1444                 else
1445                         r = kvm_arm_get_reg(vcpu, &reg);
1446                 break;
1447         }
1448         case KVM_GET_REG_LIST: {
1449                 struct kvm_reg_list __user *user_list = argp;
1450                 struct kvm_reg_list reg_list;
1451                 unsigned n;
1452
1453                 r = -ENOEXEC;
1454                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1455                         break;
1456
1457                 r = -EPERM;
1458                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1459                         break;
1460
1461                 r = -EFAULT;
1462                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1463                         break;
1464                 n = reg_list.n;
1465                 reg_list.n = kvm_arm_num_regs(vcpu);
1466                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1467                         break;
1468                 r = -E2BIG;
1469                 if (n < reg_list.n)
1470                         break;
1471                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1472                 break;
1473         }
1474         case KVM_SET_DEVICE_ATTR: {
1475                 r = -EFAULT;
1476                 if (copy_from_user(&attr, argp, sizeof(attr)))
1477                         break;
1478                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1479                 break;
1480         }
1481         case KVM_GET_DEVICE_ATTR: {
1482                 r = -EFAULT;
1483                 if (copy_from_user(&attr, argp, sizeof(attr)))
1484                         break;
1485                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1486                 break;
1487         }
1488         case KVM_HAS_DEVICE_ATTR: {
1489                 r = -EFAULT;
1490                 if (copy_from_user(&attr, argp, sizeof(attr)))
1491                         break;
1492                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1493                 break;
1494         }
1495         case KVM_GET_VCPU_EVENTS: {
1496                 struct kvm_vcpu_events events;
1497
1498                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1499                         return -EINVAL;
1500
1501                 if (copy_to_user(argp, &events, sizeof(events)))
1502                         return -EFAULT;
1503
1504                 return 0;
1505         }
1506         case KVM_SET_VCPU_EVENTS: {
1507                 struct kvm_vcpu_events events;
1508
1509                 if (copy_from_user(&events, argp, sizeof(events)))
1510                         return -EFAULT;
1511
1512                 return kvm_arm_vcpu_set_events(vcpu, &events);
1513         }
1514         case KVM_ARM_VCPU_FINALIZE: {
1515                 int what;
1516
1517                 if (!kvm_vcpu_initialized(vcpu))
1518                         return -ENOEXEC;
1519
1520                 if (get_user(what, (const int __user *)argp))
1521                         return -EFAULT;
1522
1523                 return kvm_arm_vcpu_finalize(vcpu, what);
1524         }
1525         default:
1526                 r = -EINVAL;
1527         }
1528
1529         return r;
1530 }
1531
1532 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1533 {
1534
1535 }
1536
1537 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1538                                         struct kvm_arm_device_addr *dev_addr)
1539 {
1540         switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) {
1541         case KVM_ARM_DEVICE_VGIC_V2:
1542                 if (!vgic_present)
1543                         return -ENXIO;
1544                 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr);
1545         default:
1546                 return -ENODEV;
1547         }
1548 }
1549
1550 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1551 {
1552         switch (attr->group) {
1553         case KVM_ARM_VM_SMCCC_CTRL:
1554                 return kvm_vm_smccc_has_attr(kvm, attr);
1555         default:
1556                 return -ENXIO;
1557         }
1558 }
1559
1560 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1561 {
1562         switch (attr->group) {
1563         case KVM_ARM_VM_SMCCC_CTRL:
1564                 return kvm_vm_smccc_set_attr(kvm, attr);
1565         default:
1566                 return -ENXIO;
1567         }
1568 }
1569
1570 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
1571 {
1572         struct kvm *kvm = filp->private_data;
1573         void __user *argp = (void __user *)arg;
1574         struct kvm_device_attr attr;
1575
1576         switch (ioctl) {
1577         case KVM_CREATE_IRQCHIP: {
1578                 int ret;
1579                 if (!vgic_present)
1580                         return -ENXIO;
1581                 mutex_lock(&kvm->lock);
1582                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1583                 mutex_unlock(&kvm->lock);
1584                 return ret;
1585         }
1586         case KVM_ARM_SET_DEVICE_ADDR: {
1587                 struct kvm_arm_device_addr dev_addr;
1588
1589                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1590                         return -EFAULT;
1591                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1592         }
1593         case KVM_ARM_PREFERRED_TARGET: {
1594                 struct kvm_vcpu_init init = {
1595                         .target = KVM_ARM_TARGET_GENERIC_V8,
1596                 };
1597
1598                 if (copy_to_user(argp, &init, sizeof(init)))
1599                         return -EFAULT;
1600
1601                 return 0;
1602         }
1603         case KVM_ARM_MTE_COPY_TAGS: {
1604                 struct kvm_arm_copy_mte_tags copy_tags;
1605
1606                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1607                         return -EFAULT;
1608                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1609         }
1610         case KVM_ARM_SET_COUNTER_OFFSET: {
1611                 struct kvm_arm_counter_offset offset;
1612
1613                 if (copy_from_user(&offset, argp, sizeof(offset)))
1614                         return -EFAULT;
1615                 return kvm_vm_ioctl_set_counter_offset(kvm, &offset);
1616         }
1617         case KVM_HAS_DEVICE_ATTR: {
1618                 if (copy_from_user(&attr, argp, sizeof(attr)))
1619                         return -EFAULT;
1620
1621                 return kvm_vm_has_attr(kvm, &attr);
1622         }
1623         case KVM_SET_DEVICE_ATTR: {
1624                 if (copy_from_user(&attr, argp, sizeof(attr)))
1625                         return -EFAULT;
1626
1627                 return kvm_vm_set_attr(kvm, &attr);
1628         }
1629         default:
1630                 return -EINVAL;
1631         }
1632 }
1633
1634 /* unlocks vcpus from @vcpu_lock_idx and smaller */
1635 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx)
1636 {
1637         struct kvm_vcpu *tmp_vcpu;
1638
1639         for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
1640                 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
1641                 mutex_unlock(&tmp_vcpu->mutex);
1642         }
1643 }
1644
1645 void unlock_all_vcpus(struct kvm *kvm)
1646 {
1647         lockdep_assert_held(&kvm->lock);
1648
1649         unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1);
1650 }
1651
1652 /* Returns true if all vcpus were locked, false otherwise */
1653 bool lock_all_vcpus(struct kvm *kvm)
1654 {
1655         struct kvm_vcpu *tmp_vcpu;
1656         unsigned long c;
1657
1658         lockdep_assert_held(&kvm->lock);
1659
1660         /*
1661          * Any time a vcpu is in an ioctl (including running), the
1662          * core KVM code tries to grab the vcpu->mutex.
1663          *
1664          * By grabbing the vcpu->mutex of all VCPUs we ensure that no
1665          * other VCPUs can fiddle with the state while we access it.
1666          */
1667         kvm_for_each_vcpu(c, tmp_vcpu, kvm) {
1668                 if (!mutex_trylock(&tmp_vcpu->mutex)) {
1669                         unlock_vcpus(kvm, c - 1);
1670                         return false;
1671                 }
1672         }
1673
1674         return true;
1675 }
1676
1677 static unsigned long nvhe_percpu_size(void)
1678 {
1679         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1680                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1681 }
1682
1683 static unsigned long nvhe_percpu_order(void)
1684 {
1685         unsigned long size = nvhe_percpu_size();
1686
1687         return size ? get_order(size) : 0;
1688 }
1689
1690 /* A lookup table holding the hypervisor VA for each vector slot */
1691 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1692
1693 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1694 {
1695         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1696 }
1697
1698 static int kvm_init_vector_slots(void)
1699 {
1700         int err;
1701         void *base;
1702
1703         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1704         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1705
1706         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1707         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1708
1709         if (kvm_system_needs_idmapped_vectors() &&
1710             !is_protected_kvm_enabled()) {
1711                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1712                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1713                 if (err)
1714                         return err;
1715         }
1716
1717         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1718         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1719         return 0;
1720 }
1721
1722 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits)
1723 {
1724         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1725         unsigned long tcr;
1726
1727         /*
1728          * Calculate the raw per-cpu offset without a translation from the
1729          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1730          * so that we can use adr_l to access per-cpu variables in EL2.
1731          * Also drop the KASAN tag which gets in the way...
1732          */
1733         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1734                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1735
1736         params->mair_el2 = read_sysreg(mair_el1);
1737
1738         tcr = read_sysreg(tcr_el1);
1739         if (cpus_have_final_cap(ARM64_KVM_HVHE)) {
1740                 tcr |= TCR_EPD1_MASK;
1741         } else {
1742                 tcr &= TCR_EL2_MASK;
1743                 tcr |= TCR_EL2_RES1;
1744         }
1745         tcr &= ~TCR_T0SZ_MASK;
1746         tcr |= TCR_T0SZ(hyp_va_bits);
1747         params->tcr_el2 = tcr;
1748
1749         params->pgd_pa = kvm_mmu_get_httbr();
1750         if (is_protected_kvm_enabled())
1751                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1752         else
1753                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1754         if (cpus_have_final_cap(ARM64_KVM_HVHE))
1755                 params->hcr_el2 |= HCR_E2H;
1756         params->vttbr = params->vtcr = 0;
1757
1758         /*
1759          * Flush the init params from the data cache because the struct will
1760          * be read while the MMU is off.
1761          */
1762         kvm_flush_dcache_to_poc(params, sizeof(*params));
1763 }
1764
1765 static void hyp_install_host_vector(void)
1766 {
1767         struct kvm_nvhe_init_params *params;
1768         struct arm_smccc_res res;
1769
1770         /* Switch from the HYP stub to our own HYP init vector */
1771         __hyp_set_vectors(kvm_get_idmap_vector());
1772
1773         /*
1774          * Call initialization code, and switch to the full blown HYP code.
1775          * If the cpucaps haven't been finalized yet, something has gone very
1776          * wrong, and hyp will crash and burn when it uses any
1777          * cpus_have_const_cap() wrapper.
1778          */
1779         BUG_ON(!system_capabilities_finalized());
1780         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1781         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1782         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1783 }
1784
1785 static void cpu_init_hyp_mode(void)
1786 {
1787         hyp_install_host_vector();
1788
1789         /*
1790          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1791          * at EL2.
1792          */
1793         if (this_cpu_has_cap(ARM64_SSBS) &&
1794             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1795                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1796         }
1797 }
1798
1799 static void cpu_hyp_reset(void)
1800 {
1801         if (!is_kernel_in_hyp_mode())
1802                 __hyp_reset_vectors();
1803 }
1804
1805 /*
1806  * EL2 vectors can be mapped and rerouted in a number of ways,
1807  * depending on the kernel configuration and CPU present:
1808  *
1809  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1810  *   placed in one of the vector slots, which is executed before jumping
1811  *   to the real vectors.
1812  *
1813  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1814  *   containing the hardening sequence is mapped next to the idmap page,
1815  *   and executed before jumping to the real vectors.
1816  *
1817  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1818  *   empty slot is selected, mapped next to the idmap page, and
1819  *   executed before jumping to the real vectors.
1820  *
1821  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1822  * VHE, as we don't have hypervisor-specific mappings. If the system
1823  * is VHE and yet selects this capability, it will be ignored.
1824  */
1825 static void cpu_set_hyp_vector(void)
1826 {
1827         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1828         void *vector = hyp_spectre_vector_selector[data->slot];
1829
1830         if (!is_protected_kvm_enabled())
1831                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1832         else
1833                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1834 }
1835
1836 static void cpu_hyp_init_context(void)
1837 {
1838         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1839
1840         if (!is_kernel_in_hyp_mode())
1841                 cpu_init_hyp_mode();
1842 }
1843
1844 static void cpu_hyp_init_features(void)
1845 {
1846         cpu_set_hyp_vector();
1847         kvm_arm_init_debug();
1848
1849         if (is_kernel_in_hyp_mode())
1850                 kvm_timer_init_vhe();
1851
1852         if (vgic_present)
1853                 kvm_vgic_init_cpu_hardware();
1854 }
1855
1856 static void cpu_hyp_reinit(void)
1857 {
1858         cpu_hyp_reset();
1859         cpu_hyp_init_context();
1860         cpu_hyp_init_features();
1861 }
1862
1863 static void _kvm_arch_hardware_enable(void *discard)
1864 {
1865         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1866                 cpu_hyp_reinit();
1867                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1868         }
1869 }
1870
1871 int kvm_arch_hardware_enable(void)
1872 {
1873         int was_enabled;
1874
1875         /*
1876          * Most calls to this function are made with migration
1877          * disabled, but not with preemption disabled. The former is
1878          * enough to ensure correctness, but most of the helpers
1879          * expect the later and will throw a tantrum otherwise.
1880          */
1881         preempt_disable();
1882
1883         was_enabled = __this_cpu_read(kvm_arm_hardware_enabled);
1884         _kvm_arch_hardware_enable(NULL);
1885
1886         if (!was_enabled) {
1887                 kvm_vgic_cpu_up();
1888                 kvm_timer_cpu_up();
1889         }
1890
1891         preempt_enable();
1892
1893         return 0;
1894 }
1895
1896 static void _kvm_arch_hardware_disable(void *discard)
1897 {
1898         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1899                 cpu_hyp_reset();
1900                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1901         }
1902 }
1903
1904 void kvm_arch_hardware_disable(void)
1905 {
1906         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1907                 kvm_timer_cpu_down();
1908                 kvm_vgic_cpu_down();
1909         }
1910
1911         if (!is_protected_kvm_enabled())
1912                 _kvm_arch_hardware_disable(NULL);
1913 }
1914
1915 #ifdef CONFIG_CPU_PM
1916 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1917                                     unsigned long cmd,
1918                                     void *v)
1919 {
1920         /*
1921          * kvm_arm_hardware_enabled is left with its old value over
1922          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1923          * re-enable hyp.
1924          */
1925         switch (cmd) {
1926         case CPU_PM_ENTER:
1927                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1928                         /*
1929                          * don't update kvm_arm_hardware_enabled here
1930                          * so that the hardware will be re-enabled
1931                          * when we resume. See below.
1932                          */
1933                         cpu_hyp_reset();
1934
1935                 return NOTIFY_OK;
1936         case CPU_PM_ENTER_FAILED:
1937         case CPU_PM_EXIT:
1938                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1939                         /* The hardware was enabled before suspend. */
1940                         cpu_hyp_reinit();
1941
1942                 return NOTIFY_OK;
1943
1944         default:
1945                 return NOTIFY_DONE;
1946         }
1947 }
1948
1949 static struct notifier_block hyp_init_cpu_pm_nb = {
1950         .notifier_call = hyp_init_cpu_pm_notifier,
1951 };
1952
1953 static void __init hyp_cpu_pm_init(void)
1954 {
1955         if (!is_protected_kvm_enabled())
1956                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1957 }
1958 static void __init hyp_cpu_pm_exit(void)
1959 {
1960         if (!is_protected_kvm_enabled())
1961                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1962 }
1963 #else
1964 static inline void __init hyp_cpu_pm_init(void)
1965 {
1966 }
1967 static inline void __init hyp_cpu_pm_exit(void)
1968 {
1969 }
1970 #endif
1971
1972 static void __init init_cpu_logical_map(void)
1973 {
1974         unsigned int cpu;
1975
1976         /*
1977          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
1978          * Only copy the set of online CPUs whose features have been checked
1979          * against the finalized system capabilities. The hypervisor will not
1980          * allow any other CPUs from the `possible` set to boot.
1981          */
1982         for_each_online_cpu(cpu)
1983                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
1984 }
1985
1986 #define init_psci_0_1_impl_state(config, what)  \
1987         config.psci_0_1_ ## what ## _implemented = psci_ops.what
1988
1989 static bool __init init_psci_relay(void)
1990 {
1991         /*
1992          * If PSCI has not been initialized, protected KVM cannot install
1993          * itself on newly booted CPUs.
1994          */
1995         if (!psci_ops.get_version) {
1996                 kvm_err("Cannot initialize protected mode without PSCI\n");
1997                 return false;
1998         }
1999
2000         kvm_host_psci_config.version = psci_ops.get_version();
2001         kvm_host_psci_config.smccc_version = arm_smccc_get_version();
2002
2003         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
2004                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
2005                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
2006                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
2007                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
2008                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
2009         }
2010         return true;
2011 }
2012
2013 static int __init init_subsystems(void)
2014 {
2015         int err = 0;
2016
2017         /*
2018          * Enable hardware so that subsystem initialisation can access EL2.
2019          */
2020         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
2021
2022         /*
2023          * Register CPU lower-power notifier
2024          */
2025         hyp_cpu_pm_init();
2026
2027         /*
2028          * Init HYP view of VGIC
2029          */
2030         err = kvm_vgic_hyp_init();
2031         switch (err) {
2032         case 0:
2033                 vgic_present = true;
2034                 break;
2035         case -ENODEV:
2036         case -ENXIO:
2037                 vgic_present = false;
2038                 err = 0;
2039                 break;
2040         default:
2041                 goto out;
2042         }
2043
2044         /*
2045          * Init HYP architected timer support
2046          */
2047         err = kvm_timer_hyp_init(vgic_present);
2048         if (err)
2049                 goto out;
2050
2051         kvm_register_perf_callbacks(NULL);
2052
2053 out:
2054         if (err)
2055                 hyp_cpu_pm_exit();
2056
2057         if (err || !is_protected_kvm_enabled())
2058                 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
2059
2060         return err;
2061 }
2062
2063 static void __init teardown_subsystems(void)
2064 {
2065         kvm_unregister_perf_callbacks();
2066         hyp_cpu_pm_exit();
2067 }
2068
2069 static void __init teardown_hyp_mode(void)
2070 {
2071         int cpu;
2072
2073         free_hyp_pgds();
2074         for_each_possible_cpu(cpu) {
2075                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
2076                 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order());
2077         }
2078 }
2079
2080 static int __init do_pkvm_init(u32 hyp_va_bits)
2081 {
2082         void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base));
2083         int ret;
2084
2085         preempt_disable();
2086         cpu_hyp_init_context();
2087         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
2088                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
2089                                 hyp_va_bits);
2090         cpu_hyp_init_features();
2091
2092         /*
2093          * The stub hypercalls are now disabled, so set our local flag to
2094          * prevent a later re-init attempt in kvm_arch_hardware_enable().
2095          */
2096         __this_cpu_write(kvm_arm_hardware_enabled, 1);
2097         preempt_enable();
2098
2099         return ret;
2100 }
2101
2102 static u64 get_hyp_id_aa64pfr0_el1(void)
2103 {
2104         /*
2105          * Track whether the system isn't affected by spectre/meltdown in the
2106          * hypervisor's view of id_aa64pfr0_el1, used for protected VMs.
2107          * Although this is per-CPU, we make it global for simplicity, e.g., not
2108          * to have to worry about vcpu migration.
2109          *
2110          * Unlike for non-protected VMs, userspace cannot override this for
2111          * protected VMs.
2112          */
2113         u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
2114
2115         val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
2116                  ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));
2117
2118         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
2119                           arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED);
2120         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
2121                           arm64_get_meltdown_state() == SPECTRE_UNAFFECTED);
2122
2123         return val;
2124 }
2125
2126 static void kvm_hyp_init_symbols(void)
2127 {
2128         kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1();
2129         kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1);
2130         kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
2131         kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
2132         kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
2133         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
2134         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
2135         kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1);
2136         kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1);
2137         kvm_nvhe_sym(__icache_flags) = __icache_flags;
2138         kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits;
2139 }
2140
2141 static int __init kvm_hyp_init_protection(u32 hyp_va_bits)
2142 {
2143         void *addr = phys_to_virt(hyp_mem_base);
2144         int ret;
2145
2146         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
2147         if (ret)
2148                 return ret;
2149
2150         ret = do_pkvm_init(hyp_va_bits);
2151         if (ret)
2152                 return ret;
2153
2154         free_hyp_pgds();
2155
2156         return 0;
2157 }
2158
2159 static void pkvm_hyp_init_ptrauth(void)
2160 {
2161         struct kvm_cpu_context *hyp_ctxt;
2162         int cpu;
2163
2164         for_each_possible_cpu(cpu) {
2165                 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu);
2166                 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long();
2167                 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long();
2168                 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long();
2169                 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long();
2170                 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long();
2171                 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long();
2172                 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long();
2173                 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long();
2174                 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long();
2175                 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long();
2176         }
2177 }
2178
2179 /* Inits Hyp-mode on all online CPUs */
2180 static int __init init_hyp_mode(void)
2181 {
2182         u32 hyp_va_bits;
2183         int cpu;
2184         int err = -ENOMEM;
2185
2186         /*
2187          * The protected Hyp-mode cannot be initialized if the memory pool
2188          * allocation has failed.
2189          */
2190         if (is_protected_kvm_enabled() && !hyp_mem_base)
2191                 goto out_err;
2192
2193         /*
2194          * Allocate Hyp PGD and setup Hyp identity mapping
2195          */
2196         err = kvm_mmu_init(&hyp_va_bits);
2197         if (err)
2198                 goto out_err;
2199
2200         /*
2201          * Allocate stack pages for Hypervisor-mode
2202          */
2203         for_each_possible_cpu(cpu) {
2204                 unsigned long stack_page;
2205
2206                 stack_page = __get_free_page(GFP_KERNEL);
2207                 if (!stack_page) {
2208                         err = -ENOMEM;
2209                         goto out_err;
2210                 }
2211
2212                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
2213         }
2214
2215         /*
2216          * Allocate and initialize pages for Hypervisor-mode percpu regions.
2217          */
2218         for_each_possible_cpu(cpu) {
2219                 struct page *page;
2220                 void *page_addr;
2221
2222                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
2223                 if (!page) {
2224                         err = -ENOMEM;
2225                         goto out_err;
2226                 }
2227
2228                 page_addr = page_address(page);
2229                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
2230                 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr;
2231         }
2232
2233         /*
2234          * Map the Hyp-code called directly from the host
2235          */
2236         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
2237                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
2238         if (err) {
2239                 kvm_err("Cannot map world-switch code\n");
2240                 goto out_err;
2241         }
2242
2243         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
2244                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
2245         if (err) {
2246                 kvm_err("Cannot map .hyp.rodata section\n");
2247                 goto out_err;
2248         }
2249
2250         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
2251                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
2252         if (err) {
2253                 kvm_err("Cannot map rodata section\n");
2254                 goto out_err;
2255         }
2256
2257         /*
2258          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
2259          * section thanks to an assertion in the linker script. Map it RW and
2260          * the rest of .bss RO.
2261          */
2262         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
2263                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
2264         if (err) {
2265                 kvm_err("Cannot map hyp bss section: %d\n", err);
2266                 goto out_err;
2267         }
2268
2269         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
2270                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
2271         if (err) {
2272                 kvm_err("Cannot map bss section\n");
2273                 goto out_err;
2274         }
2275
2276         /*
2277          * Map the Hyp stack pages
2278          */
2279         for_each_possible_cpu(cpu) {
2280                 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2281                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
2282                 unsigned long hyp_addr;
2283
2284                 /*
2285                  * Allocate a contiguous HYP private VA range for the stack
2286                  * and guard page. The allocation is also aligned based on
2287                  * the order of its size.
2288                  */
2289                 err = hyp_alloc_private_va_range(PAGE_SIZE * 2, &hyp_addr);
2290                 if (err) {
2291                         kvm_err("Cannot allocate hyp stack guard page\n");
2292                         goto out_err;
2293                 }
2294
2295                 /*
2296                  * Since the stack grows downwards, map the stack to the page
2297                  * at the higher address and leave the lower guard page
2298                  * unbacked.
2299                  *
2300                  * Any valid stack address now has the PAGE_SHIFT bit as 1
2301                  * and addresses corresponding to the guard page have the
2302                  * PAGE_SHIFT bit as 0 - this is used for overflow detection.
2303                  */
2304                 err = __create_hyp_mappings(hyp_addr + PAGE_SIZE, PAGE_SIZE,
2305                                             __pa(stack_page), PAGE_HYP);
2306                 if (err) {
2307                         kvm_err("Cannot map hyp stack\n");
2308                         goto out_err;
2309                 }
2310
2311                 /*
2312                  * Save the stack PA in nvhe_init_params. This will be needed
2313                  * to recreate the stack mapping in protected nVHE mode.
2314                  * __hyp_pa() won't do the right thing there, since the stack
2315                  * has been mapped in the flexible private VA space.
2316                  */
2317                 params->stack_pa = __pa(stack_page);
2318
2319                 params->stack_hyp_va = hyp_addr + (2 * PAGE_SIZE);
2320         }
2321
2322         for_each_possible_cpu(cpu) {
2323                 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu];
2324                 char *percpu_end = percpu_begin + nvhe_percpu_size();
2325
2326                 /* Map Hyp percpu pages */
2327                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
2328                 if (err) {
2329                         kvm_err("Cannot map hyp percpu region\n");
2330                         goto out_err;
2331                 }
2332
2333                 /* Prepare the CPU initialization parameters */
2334                 cpu_prepare_hyp_mode(cpu, hyp_va_bits);
2335         }
2336
2337         kvm_hyp_init_symbols();
2338
2339         if (is_protected_kvm_enabled()) {
2340                 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) &&
2341                     cpus_have_const_cap(ARM64_HAS_ADDRESS_AUTH))
2342                         pkvm_hyp_init_ptrauth();
2343
2344                 init_cpu_logical_map();
2345
2346                 if (!init_psci_relay()) {
2347                         err = -ENODEV;
2348                         goto out_err;
2349                 }
2350
2351                 err = kvm_hyp_init_protection(hyp_va_bits);
2352                 if (err) {
2353                         kvm_err("Failed to init hyp memory protection\n");
2354                         goto out_err;
2355                 }
2356         }
2357
2358         return 0;
2359
2360 out_err:
2361         teardown_hyp_mode();
2362         kvm_err("error initializing Hyp mode: %d\n", err);
2363         return err;
2364 }
2365
2366 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2367 {
2368         struct kvm_vcpu *vcpu;
2369         unsigned long i;
2370
2371         mpidr &= MPIDR_HWID_BITMASK;
2372         kvm_for_each_vcpu(i, vcpu, kvm) {
2373                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2374                         return vcpu;
2375         }
2376         return NULL;
2377 }
2378
2379 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm)
2380 {
2381         return irqchip_in_kernel(kvm);
2382 }
2383
2384 bool kvm_arch_has_irq_bypass(void)
2385 {
2386         return true;
2387 }
2388
2389 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2390                                       struct irq_bypass_producer *prod)
2391 {
2392         struct kvm_kernel_irqfd *irqfd =
2393                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2394
2395         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2396                                           &irqfd->irq_entry);
2397 }
2398 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2399                                       struct irq_bypass_producer *prod)
2400 {
2401         struct kvm_kernel_irqfd *irqfd =
2402                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2403
2404         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2405                                      &irqfd->irq_entry);
2406 }
2407
2408 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2409 {
2410         struct kvm_kernel_irqfd *irqfd =
2411                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2412
2413         kvm_arm_halt_guest(irqfd->kvm);
2414 }
2415
2416 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2417 {
2418         struct kvm_kernel_irqfd *irqfd =
2419                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2420
2421         kvm_arm_resume_guest(irqfd->kvm);
2422 }
2423
2424 /* Initialize Hyp-mode and memory mappings on all CPUs */
2425 static __init int kvm_arm_init(void)
2426 {
2427         int err;
2428         bool in_hyp_mode;
2429
2430         if (!is_hyp_mode_available()) {
2431                 kvm_info("HYP mode not available\n");
2432                 return -ENODEV;
2433         }
2434
2435         if (kvm_get_mode() == KVM_MODE_NONE) {
2436                 kvm_info("KVM disabled from command line\n");
2437                 return -ENODEV;
2438         }
2439
2440         err = kvm_sys_reg_table_init();
2441         if (err) {
2442                 kvm_info("Error initializing system register tables");
2443                 return err;
2444         }
2445
2446         in_hyp_mode = is_kernel_in_hyp_mode();
2447
2448         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2449             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2450                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2451                          "Only trusted guests should be used on this system.\n");
2452
2453         err = kvm_set_ipa_limit();
2454         if (err)
2455                 return err;
2456
2457         err = kvm_arm_init_sve();
2458         if (err)
2459                 return err;
2460
2461         err = kvm_arm_vmid_alloc_init();
2462         if (err) {
2463                 kvm_err("Failed to initialize VMID allocator.\n");
2464                 return err;
2465         }
2466
2467         if (!in_hyp_mode) {
2468                 err = init_hyp_mode();
2469                 if (err)
2470                         goto out_err;
2471         }
2472
2473         err = kvm_init_vector_slots();
2474         if (err) {
2475                 kvm_err("Cannot initialise vector slots\n");
2476                 goto out_hyp;
2477         }
2478
2479         err = init_subsystems();
2480         if (err)
2481                 goto out_hyp;
2482
2483         if (is_protected_kvm_enabled()) {
2484                 kvm_info("Protected nVHE mode initialized successfully\n");
2485         } else if (in_hyp_mode) {
2486                 kvm_info("VHE mode initialized successfully\n");
2487         } else {
2488                 kvm_info("Hyp mode initialized successfully\n");
2489         }
2490
2491         /*
2492          * FIXME: Do something reasonable if kvm_init() fails after pKVM
2493          * hypervisor protection is finalized.
2494          */
2495         err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2496         if (err)
2497                 goto out_subs;
2498
2499         kvm_arm_initialised = true;
2500
2501         return 0;
2502
2503 out_subs:
2504         teardown_subsystems();
2505 out_hyp:
2506         if (!in_hyp_mode)
2507                 teardown_hyp_mode();
2508 out_err:
2509         kvm_arm_vmid_alloc_free();
2510         return err;
2511 }
2512
2513 static int __init early_kvm_mode_cfg(char *arg)
2514 {
2515         if (!arg)
2516                 return -EINVAL;
2517
2518         if (strcmp(arg, "none") == 0) {
2519                 kvm_mode = KVM_MODE_NONE;
2520                 return 0;
2521         }
2522
2523         if (!is_hyp_mode_available()) {
2524                 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n");
2525                 return 0;
2526         }
2527
2528         if (strcmp(arg, "protected") == 0) {
2529                 if (!is_kernel_in_hyp_mode())
2530                         kvm_mode = KVM_MODE_PROTECTED;
2531                 else
2532                         pr_warn_once("Protected KVM not available with VHE\n");
2533
2534                 return 0;
2535         }
2536
2537         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) {
2538                 kvm_mode = KVM_MODE_DEFAULT;
2539                 return 0;
2540         }
2541
2542         if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) {
2543                 kvm_mode = KVM_MODE_NV;
2544                 return 0;
2545         }
2546
2547         return -EINVAL;
2548 }
2549 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2550
2551 enum kvm_mode kvm_get_mode(void)
2552 {
2553         return kvm_mode;
2554 }
2555
2556 module_init(kvm_arm_init);