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/errno.h>
  10 #include <linux/err.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/list.h>
  13 #include <linux/module.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/fs.h>
  16 #include <linux/mman.h>
  17 #include <linux/sched.h>
  18 #include <linux/kvm.h>
  19 #include <linux/kvm_irqfd.h>
  20 #include <linux/irqbypass.h>
  21 #include <linux/sched/stat.h>
  22 #include <linux/psci.h>
  23 #include <trace/events/kvm.h>
  24
  25 #define CREATE_TRACE_POINTS
  26 #include "trace_arm.h"
  27
  28 #include <linux/uaccess.h>
  29 #include <asm/ptrace.h>
  30 #include <asm/mman.h>
  31 #include <asm/tlbflush.h>
  32 #include <asm/cacheflush.h>
  33 #include <asm/cpufeature.h>
  34 #include <asm/virt.h>
  35 #include <asm/kvm_arm.h>
  36 #include <asm/kvm_asm.h>
  37 #include <asm/kvm_mmu.h>
  38 #include <asm/kvm_emulate.h>
  39 #include <asm/sections.h>
  40
  41 #include <kvm/arm_hypercalls.h>
  42 #include <kvm/arm_pmu.h>
  43 #include <kvm/arm_psci.h>
  44
  45 #ifdef REQUIRES_VIRT
  46 __asm__(".arch_extension        virt");
  47 #endif
  48
  49 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
  50 DEFINE_STATIC_KEY_FALSE(kvm_protected_mode_initialized);
  51
  52 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
  53
  54 static DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  55 unsigned long kvm_arm_hyp_percpu_base[NR_CPUS];
  56 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
  57
  58 /* The VMID used in the VTTBR */
  59 static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1);
  60 static u32 kvm_next_vmid;
  61 static DEFINE_SPINLOCK(kvm_vmid_lock);
  62
  63 static bool vgic_present;
  64
  65 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled);
  66 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  67
  68 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  69 {
  70         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  71 }
  72
  73 int kvm_arch_hardware_setup(void *opaque)
  74 {
  75         return 0;
  76 }
  77
  78 int kvm_arch_check_processor_compat(void *opaque)
  79 {
  80         return 0;
  81 }
  82
  83 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  84                             struct kvm_enable_cap *cap)
  85 {
  86         int r;
  87
  88         if (cap->flags)
  89                 return -EINVAL;
  90
  91         switch (cap->cap) {
  92         case KVM_CAP_ARM_NISV_TO_USER:
  93                 r = 0;
  94                 kvm->arch.return_nisv_io_abort_to_user = true;
  95                 break;
  96         case KVM_CAP_ARM_MTE:
  97                 mutex_lock(&kvm->lock);
  98                 if (!system_supports_mte() || kvm->created_vcpus) {
  99                         r = -EINVAL;
 100                 } else {
 101                         r = 0;
 102                         kvm->arch.mte_enabled = true;
 103                 }
 104                 mutex_unlock(&kvm->lock);
 105                 break;
 106         default:
 107                 r = -EINVAL;
 108                 break;
 109         }
 110
 111         return r;
 112 }
 113
 114 static int kvm_arm_default_max_vcpus(void)
 115 {
 116         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 117 }
 118
 119 static void set_default_spectre(struct kvm *kvm)
 120 {
 121         /*
 122          * The default is to expose CSV2 == 1 if the HW isn't affected.
 123          * Although this is a per-CPU feature, we make it global because
 124          * asymmetric systems are just a nuisance.
 125          *
 126          * Userspace can override this as long as it doesn't promise
 127          * the impossible.
 128          */
 129         if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED)
 130                 kvm->arch.pfr0_csv2 = 1;
 131         if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED)
 132                 kvm->arch.pfr0_csv3 = 1;
 133 }
 134
 135 /**
 136  * kvm_arch_init_vm - initializes a VM data structure
 137  * @kvm:        pointer to the KVM struct
 138  */
 139 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 140 {
 141         int ret;
 142
 143         ret = kvm_arm_setup_stage2(kvm, type);
 144         if (ret)
 145                 return ret;
 146
 147         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu);
 148         if (ret)
 149                 return ret;
 150
 151         ret = create_hyp_mappings(kvm, kvm + 1, PAGE_HYP);
 152         if (ret)
 153                 goto out_free_stage2_pgd;
 154
 155         kvm_vgic_early_init(kvm);
 156
 157         /* The maximum number of VCPUs is limited by the host's GIC model */
 158         kvm->arch.max_vcpus = kvm_arm_default_max_vcpus();
 159
 160         set_default_spectre(kvm);
 161
 162         return ret;
 163 out_free_stage2_pgd:
 164         kvm_free_stage2_pgd(&kvm->arch.mmu);
 165         return ret;
 166 }
 167
 168 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 169 {
 170         return VM_FAULT_SIGBUS;
 171 }
 172
 173
 174 /**
 175  * kvm_arch_destroy_vm - destroy the VM data structure
 176  * @kvm:        pointer to the KVM struct
 177  */
 178 void kvm_arch_destroy_vm(struct kvm *kvm)
 179 {
 180         int i;
 181
 182         bitmap_free(kvm->arch.pmu_filter);
 183
 184         kvm_vgic_destroy(kvm);
 185
 186         for (i = 0; i < KVM_MAX_VCPUS; ++i) {
 187                 if (kvm->vcpus[i]) {
 188                         kvm_vcpu_destroy(kvm->vcpus[i]);
 189                         kvm->vcpus[i] = NULL;
 190                 }
 191         }
 192         atomic_set(&kvm->online_vcpus, 0);
 193 }
 194
 195 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 196 {
 197         int r;
 198         switch (ext) {
 199         case KVM_CAP_IRQCHIP:
 200                 r = vgic_present;
 201                 break;
 202         case KVM_CAP_IOEVENTFD:
 203         case KVM_CAP_DEVICE_CTRL:
 204         case KVM_CAP_USER_MEMORY:
 205         case KVM_CAP_SYNC_MMU:
 206         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 207         case KVM_CAP_ONE_REG:
 208         case KVM_CAP_ARM_PSCI:
 209         case KVM_CAP_ARM_PSCI_0_2:
 210         case KVM_CAP_READONLY_MEM:
 211         case KVM_CAP_MP_STATE:
 212         case KVM_CAP_IMMEDIATE_EXIT:
 213         case KVM_CAP_VCPU_EVENTS:
 214         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 215         case KVM_CAP_ARM_NISV_TO_USER:
 216         case KVM_CAP_ARM_INJECT_EXT_DABT:
 217         case KVM_CAP_SET_GUEST_DEBUG:
 218         case KVM_CAP_VCPU_ATTRIBUTES:
 219         case KVM_CAP_PTP_KVM:
 220                 r = 1;
 221                 break;
 222         case KVM_CAP_SET_GUEST_DEBUG2:
 223                 return KVM_GUESTDBG_VALID_MASK;
 224         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 225                 r = 1;
 226                 break;
 227         case KVM_CAP_NR_VCPUS:
 228                 r = num_online_cpus();
 229                 break;
 230         case KVM_CAP_MAX_VCPUS:
 231         case KVM_CAP_MAX_VCPU_ID:
 232                 if (kvm)
 233                         r = kvm->arch.max_vcpus;
 234                 else
 235                         r = kvm_arm_default_max_vcpus();
 236                 break;
 237         case KVM_CAP_MSI_DEVID:
 238                 if (!kvm)
 239                         r = -EINVAL;
 240                 else
 241                         r = kvm->arch.vgic.msis_require_devid;
 242                 break;
 243         case KVM_CAP_ARM_USER_IRQ:
 244                 /*
 245                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 246                  * (bump this number if adding more devices)
 247                  */
 248                 r = 1;
 249                 break;
 250         case KVM_CAP_ARM_MTE:
 251                 r = system_supports_mte();
 252                 break;
 253         case KVM_CAP_STEAL_TIME:
 254                 r = kvm_arm_pvtime_supported();
 255                 break;
 256         case KVM_CAP_ARM_EL1_32BIT:
 257                 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1);
 258                 break;
 259         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 260                 r = get_num_brps();
 261                 break;
 262         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 263                 r = get_num_wrps();
 264                 break;
 265         case KVM_CAP_ARM_PMU_V3:
 266                 r = kvm_arm_support_pmu_v3();
 267                 break;
 268         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 269                 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN);
 270                 break;
 271         case KVM_CAP_ARM_VM_IPA_SIZE:
 272                 r = get_kvm_ipa_limit();
 273                 break;
 274         case KVM_CAP_ARM_SVE:
 275                 r = system_supports_sve();
 276                 break;
 277         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 278         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 279                 r = system_has_full_ptr_auth();
 280                 break;
 281         default:
 282                 r = 0;
 283         }
 284
 285         return r;
 286 }
 287
 288 long kvm_arch_dev_ioctl(struct file *filp,
 289                         unsigned int ioctl, unsigned long arg)
 290 {
 291         return -EINVAL;
 292 }
 293
 294 struct kvm *kvm_arch_alloc_vm(void)
 295 {
 296         if (!has_vhe())
 297                 return kzalloc(sizeof(struct kvm), GFP_KERNEL);
 298
 299         return vzalloc(sizeof(struct kvm));
 300 }
 301
 302 void kvm_arch_free_vm(struct kvm *kvm)
 303 {
 304         if (!has_vhe())
 305                 kfree(kvm);
 306         else
 307                 vfree(kvm);
 308 }
 309
 310 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 311 {
 312         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 313                 return -EBUSY;
 314
 315         if (id >= kvm->arch.max_vcpus)
 316                 return -EINVAL;
 317
 318         return 0;
 319 }
 320
 321 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 322 {
 323         int err;
 324
 325         /* Force users to call KVM_ARM_VCPU_INIT */
 326         vcpu->arch.target = -1;
 327         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 328
 329         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 330
 331         /* Set up the timer */
 332         kvm_timer_vcpu_init(vcpu);
 333
 334         kvm_pmu_vcpu_init(vcpu);
 335
 336         kvm_arm_reset_debug_ptr(vcpu);
 337
 338         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 339
 340         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 341
 342         err = kvm_vgic_vcpu_init(vcpu);
 343         if (err)
 344                 return err;
 345
 346         return create_hyp_mappings(vcpu, vcpu + 1, PAGE_HYP);
 347 }
 348
 349 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 350 {
 351 }
 352
 353 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 354 {
 355         if (vcpu->arch.has_run_once && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 356                 static_branch_dec(&userspace_irqchip_in_use);
 357
 358         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 359         kvm_timer_vcpu_terminate(vcpu);
 360         kvm_pmu_vcpu_destroy(vcpu);
 361
 362         kvm_arm_vcpu_destroy(vcpu);
 363 }
 364
 365 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
 366 {
 367         return kvm_timer_is_pending(vcpu);
 368 }
 369
 370 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 371 {
 372         /*
 373          * If we're about to block (most likely because we've just hit a
 374          * WFI), we need to sync back the state of the GIC CPU interface
 375          * so that we have the latest PMR and group enables. This ensures
 376          * that kvm_arch_vcpu_runnable has up-to-date data to decide
 377          * whether we have pending interrupts.
 378          *
 379          * For the same reason, we want to tell GICv4 that we need
 380          * doorbells to be signalled, should an interrupt become pending.
 381          */
 382         preempt_disable();
 383         kvm_vgic_vmcr_sync(vcpu);
 384         vgic_v4_put(vcpu, true);
 385         preempt_enable();
 386 }
 387
 388 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 389 {
 390         preempt_disable();
 391         vgic_v4_load(vcpu);
 392         preempt_enable();
 393 }
 394
 395 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 396 {
 397         struct kvm_s2_mmu *mmu;
 398         int *last_ran;
 399
 400         mmu = vcpu->arch.hw_mmu;
 401         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 402
 403         /*
 404          * We guarantee that both TLBs and I-cache are private to each
 405          * vcpu. If detecting that a vcpu from the same VM has
 406          * previously run on the same physical CPU, call into the
 407          * hypervisor code to nuke the relevant contexts.
 408          *
 409          * We might get preempted before the vCPU actually runs, but
 410          * over-invalidation doesn't affect correctness.
 411          */
 412         if (*last_ran != vcpu->vcpu_id) {
 413                 kvm_call_hyp(__kvm_flush_cpu_context, mmu);
 414                 *last_ran = vcpu->vcpu_id;
 415         }
 416
 417         vcpu->cpu = cpu;
 418
 419         kvm_vgic_load(vcpu);
 420         kvm_timer_vcpu_load(vcpu);
 421         if (has_vhe())
 422                 kvm_vcpu_load_sysregs_vhe(vcpu);
 423         kvm_arch_vcpu_load_fp(vcpu);
 424         kvm_vcpu_pmu_restore_guest(vcpu);
 425         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 426                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 427
 428         if (single_task_running())
 429                 vcpu_clear_wfx_traps(vcpu);
 430         else
 431                 vcpu_set_wfx_traps(vcpu);
 432
 433         if (vcpu_has_ptrauth(vcpu))
 434                 vcpu_ptrauth_disable(vcpu);
 435         kvm_arch_vcpu_load_debug_state_flags(vcpu);
 436 }
 437
 438 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 439 {
 440         kvm_arch_vcpu_put_debug_state_flags(vcpu);
 441         kvm_arch_vcpu_put_fp(vcpu);
 442         if (has_vhe())
 443                 kvm_vcpu_put_sysregs_vhe(vcpu);
 444         kvm_timer_vcpu_put(vcpu);
 445         kvm_vgic_put(vcpu);
 446         kvm_vcpu_pmu_restore_host(vcpu);
 447
 448         vcpu->cpu = -1;
 449 }
 450
 451 static void vcpu_power_off(struct kvm_vcpu *vcpu)
 452 {
 453         vcpu->arch.power_off = true;
 454         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 455         kvm_vcpu_kick(vcpu);
 456 }
 457
 458 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 459                                     struct kvm_mp_state *mp_state)
 460 {
 461         if (vcpu->arch.power_off)
 462                 mp_state->mp_state = KVM_MP_STATE_STOPPED;
 463         else
 464                 mp_state->mp_state = KVM_MP_STATE_RUNNABLE;
 465
 466         return 0;
 467 }
 468
 469 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 470                                     struct kvm_mp_state *mp_state)
 471 {
 472         int ret = 0;
 473
 474         switch (mp_state->mp_state) {
 475         case KVM_MP_STATE_RUNNABLE:
 476                 vcpu->arch.power_off = false;
 477                 break;
 478         case KVM_MP_STATE_STOPPED:
 479                 vcpu_power_off(vcpu);
 480                 break;
 481         default:
 482                 ret = -EINVAL;
 483         }
 484
 485         return ret;
 486 }
 487
 488 /**
 489  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 490  * @v:          The VCPU pointer
 491  *
 492  * If the guest CPU is not waiting for interrupts or an interrupt line is
 493  * asserted, the CPU is by definition runnable.
 494  */
 495 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 496 {
 497         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 498         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 499                 && !v->arch.power_off && !v->arch.pause);
 500 }
 501
 502 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 503 {
 504         return vcpu_mode_priv(vcpu);
 505 }
 506
 507 /* Just ensure a guest exit from a particular CPU */
 508 static void exit_vm_noop(void *info)
 509 {
 510 }
 511
 512 void force_vm_exit(const cpumask_t *mask)
 513 {
 514         preempt_disable();
 515         smp_call_function_many(mask, exit_vm_noop, NULL, true);
 516         preempt_enable();
 517 }
 518
 519 /**
 520  * need_new_vmid_gen - check that the VMID is still valid
 521  * @vmid: The VMID to check
 522  *
 523  * return true if there is a new generation of VMIDs being used
 524  *
 525  * The hardware supports a limited set of values with the value zero reserved
 526  * for the host, so we check if an assigned value belongs to a previous
 527  * generation, which requires us to assign a new value. If we're the first to
 528  * use a VMID for the new generation, we must flush necessary caches and TLBs
 529  * on all CPUs.
 530  */
 531 static bool need_new_vmid_gen(struct kvm_vmid *vmid)
 532 {
 533         u64 current_vmid_gen = atomic64_read(&kvm_vmid_gen);
 534         smp_rmb(); /* Orders read of kvm_vmid_gen and kvm->arch.vmid */
 535         return unlikely(READ_ONCE(vmid->vmid_gen) != current_vmid_gen);
 536 }
 537
 538 /**
 539  * update_vmid - Update the vmid with a valid VMID for the current generation
 540  * @vmid: The stage-2 VMID information struct
 541  */
 542 static void update_vmid(struct kvm_vmid *vmid)
 543 {
 544         if (!need_new_vmid_gen(vmid))
 545                 return;
 546
 547         spin_lock(&kvm_vmid_lock);
 548
 549         /*
 550          * We need to re-check the vmid_gen here to ensure that if another vcpu
 551          * already allocated a valid vmid for this vm, then this vcpu should
 552          * use the same vmid.
 553          */
 554         if (!need_new_vmid_gen(vmid)) {
 555                 spin_unlock(&kvm_vmid_lock);
 556                 return;
 557         }
 558
 559         /* First user of a new VMID generation? */
 560         if (unlikely(kvm_next_vmid == 0)) {
 561                 atomic64_inc(&kvm_vmid_gen);
 562                 kvm_next_vmid = 1;
 563
 564                 /*
 565                  * On SMP we know no other CPUs can use this CPU's or each
 566                  * other's VMID after force_vm_exit returns since the
 567                  * kvm_vmid_lock blocks them from reentry to the guest.
 568                  */
 569                 force_vm_exit(cpu_all_mask);
 570                 /*
 571                  * Now broadcast TLB + ICACHE invalidation over the inner
 572                  * shareable domain to make sure all data structures are
 573                  * clean.
 574                  */
 575                 kvm_call_hyp(__kvm_flush_vm_context);
 576         }
 577
 578         vmid->vmid = kvm_next_vmid;
 579         kvm_next_vmid++;
 580         kvm_next_vmid &= (1 << kvm_get_vmid_bits()) - 1;
 581
 582         smp_wmb();
 583         WRITE_ONCE(vmid->vmid_gen, atomic64_read(&kvm_vmid_gen));
 584
 585         spin_unlock(&kvm_vmid_lock);
 586 }
 587
 588 static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 589 {
 590         struct kvm *kvm = vcpu->kvm;
 591         int ret = 0;
 592
 593         if (likely(vcpu->arch.has_run_once))
 594                 return 0;
 595
 596         if (!kvm_arm_vcpu_is_finalized(vcpu))
 597                 return -EPERM;
 598
 599         vcpu->arch.has_run_once = true;
 600
 601         kvm_arm_vcpu_init_debug(vcpu);
 602
 603         if (likely(irqchip_in_kernel(kvm))) {
 604                 /*
 605                  * Map the VGIC hardware resources before running a vcpu the
 606                  * first time on this VM.
 607                  */
 608                 ret = kvm_vgic_map_resources(kvm);
 609                 if (ret)
 610                         return ret;
 611         } else {
 612                 /*
 613                  * Tell the rest of the code that there are userspace irqchip
 614                  * VMs in the wild.
 615                  */
 616                 static_branch_inc(&userspace_irqchip_in_use);
 617         }
 618
 619         ret = kvm_timer_enable(vcpu);
 620         if (ret)
 621                 return ret;
 622
 623         ret = kvm_arm_pmu_v3_enable(vcpu);
 624
 625         return ret;
 626 }
 627
 628 bool kvm_arch_intc_initialized(struct kvm *kvm)
 629 {
 630         return vgic_initialized(kvm);
 631 }
 632
 633 void kvm_arm_halt_guest(struct kvm *kvm)
 634 {
 635         int i;
 636         struct kvm_vcpu *vcpu;
 637
 638         kvm_for_each_vcpu(i, vcpu, kvm)
 639                 vcpu->arch.pause = true;
 640         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 641 }
 642
 643 void kvm_arm_resume_guest(struct kvm *kvm)
 644 {
 645         int i;
 646         struct kvm_vcpu *vcpu;
 647
 648         kvm_for_each_vcpu(i, vcpu, kvm) {
 649                 vcpu->arch.pause = false;
 650                 rcuwait_wake_up(kvm_arch_vcpu_get_wait(vcpu));
 651         }
 652 }
 653
 654 static void vcpu_req_sleep(struct kvm_vcpu *vcpu)
 655 {
 656         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 657
 658         rcuwait_wait_event(wait,
 659                            (!vcpu->arch.power_off) &&(!vcpu->arch.pause),
 660                            TASK_INTERRUPTIBLE);
 661
 662         if (vcpu->arch.power_off || vcpu->arch.pause) {
 663                 /* Awaken to handle a signal, request we sleep again later. */
 664                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 665         }
 666
 667         /*
 668          * Make sure we will observe a potential reset request if we've
 669          * observed a change to the power state. Pairs with the smp_wmb() in
 670          * kvm_psci_vcpu_on().
 671          */
 672         smp_rmb();
 673 }
 674
 675 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 676 {
 677         return vcpu->arch.target >= 0;
 678 }
 679
 680 static void check_vcpu_requests(struct kvm_vcpu *vcpu)
 681 {
 682         if (kvm_request_pending(vcpu)) {
 683                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 684                         vcpu_req_sleep(vcpu);
 685
 686                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 687                         kvm_reset_vcpu(vcpu);
 688
 689                 /*
 690                  * Clear IRQ_PENDING requests that were made to guarantee
 691                  * that a VCPU sees new virtual interrupts.
 692                  */
 693                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 694
 695                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 696                         kvm_update_stolen_time(vcpu);
 697
 698                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 699                         /* The distributor enable bits were changed */
 700                         preempt_disable();
 701                         vgic_v4_put(vcpu, false);
 702                         vgic_v4_load(vcpu);
 703                         preempt_enable();
 704                 }
 705
 706                 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
 707                         kvm_pmu_handle_pmcr(vcpu,
 708                                             __vcpu_sys_reg(vcpu, PMCR_EL0));
 709         }
 710 }
 711
 712 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 713 {
 714         if (likely(!vcpu_mode_is_32bit(vcpu)))
 715                 return false;
 716
 717         return !system_supports_32bit_el0() ||
 718                 static_branch_unlikely(&arm64_mismatched_32bit_el0);
 719 }
 720
 721 /**
 722  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 723  * @vcpu:       The VCPU pointer
 724  *
 725  * This function is called through the VCPU_RUN ioctl called from user space. It
 726  * will execute VM code in a loop until the time slice for the process is used
 727  * or some emulation is needed from user space in which case the function will
 728  * return with return value 0 and with the kvm_run structure filled in with the
 729  * required data for the requested emulation.
 730  */
 731 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 732 {
 733         struct kvm_run *run = vcpu->run;
 734         int ret;
 735
 736         if (unlikely(!kvm_vcpu_initialized(vcpu)))
 737                 return -ENOEXEC;
 738
 739         ret = kvm_vcpu_first_run_init(vcpu);
 740         if (ret)
 741                 return ret;
 742
 743         if (run->exit_reason == KVM_EXIT_MMIO) {
 744                 ret = kvm_handle_mmio_return(vcpu);
 745                 if (ret)
 746                         return ret;
 747         }
 748
 749         vcpu_load(vcpu);
 750
 751         if (run->immediate_exit) {
 752                 ret = -EINTR;
 753                 goto out;
 754         }
 755
 756         kvm_sigset_activate(vcpu);
 757
 758         ret = 1;
 759         run->exit_reason = KVM_EXIT_UNKNOWN;
 760         while (ret > 0) {
 761                 /*
 762                  * Check conditions before entering the guest
 763                  */
 764                 cond_resched();
 765
 766                 update_vmid(&vcpu->arch.hw_mmu->vmid);
 767
 768                 check_vcpu_requests(vcpu);
 769
 770                 /*
 771                  * Preparing the interrupts to be injected also
 772                  * involves poking the GIC, which must be done in a
 773                  * non-preemptible context.
 774                  */
 775                 preempt_disable();
 776
 777                 kvm_pmu_flush_hwstate(vcpu);
 778
 779                 local_irq_disable();
 780
 781                 kvm_vgic_flush_hwstate(vcpu);
 782
 783                 /*
 784                  * Exit if we have a signal pending so that we can deliver the
 785                  * signal to user space.
 786                  */
 787                 if (signal_pending(current)) {
 788                         ret = -EINTR;
 789                         run->exit_reason = KVM_EXIT_INTR;
 790                 }
 791
 792                 /*
 793                  * If we're using a userspace irqchip, then check if we need
 794                  * to tell a userspace irqchip about timer or PMU level
 795                  * changes and if so, exit to userspace (the actual level
 796                  * state gets updated in kvm_timer_update_run and
 797                  * kvm_pmu_update_run below).
 798                  */
 799                 if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 800                         if (kvm_timer_should_notify_user(vcpu) ||
 801                             kvm_pmu_should_notify_user(vcpu)) {
 802                                 ret = -EINTR;
 803                                 run->exit_reason = KVM_EXIT_INTR;
 804                         }
 805                 }
 806
 807                 /*
 808                  * Ensure we set mode to IN_GUEST_MODE after we disable
 809                  * interrupts and before the final VCPU requests check.
 810                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 811                  * Documentation/virt/kvm/vcpu-requests.rst
 812                  */
 813                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 814
 815                 if (ret <= 0 || need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
 816                     kvm_request_pending(vcpu)) {
 817                         vcpu->mode = OUTSIDE_GUEST_MODE;
 818                         isb(); /* Ensure work in x_flush_hwstate is committed */
 819                         kvm_pmu_sync_hwstate(vcpu);
 820                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 821                                 kvm_timer_sync_user(vcpu);
 822                         kvm_vgic_sync_hwstate(vcpu);
 823                         local_irq_enable();
 824                         preempt_enable();
 825                         continue;
 826                 }
 827
 828                 kvm_arm_setup_debug(vcpu);
 829
 830                 /**************************************************************
 831                  * Enter the guest
 832                  */
 833                 trace_kvm_entry(*vcpu_pc(vcpu));
 834                 guest_enter_irqoff();
 835
 836                 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 837
 838                 vcpu->mode = OUTSIDE_GUEST_MODE;
 839                 vcpu->stat.exits++;
 840                 /*
 841                  * Back from guest
 842                  *************************************************************/
 843
 844                 kvm_arm_clear_debug(vcpu);
 845
 846                 /*
 847                  * We must sync the PMU state before the vgic state so
 848                  * that the vgic can properly sample the updated state of the
 849                  * interrupt line.
 850                  */
 851                 kvm_pmu_sync_hwstate(vcpu);
 852
 853                 /*
 854                  * Sync the vgic state before syncing the timer state because
 855                  * the timer code needs to know if the virtual timer
 856                  * interrupts are active.
 857                  */
 858                 kvm_vgic_sync_hwstate(vcpu);
 859
 860                 /*
 861                  * Sync the timer hardware state before enabling interrupts as
 862                  * we don't want vtimer interrupts to race with syncing the
 863                  * timer virtual interrupt state.
 864                  */
 865                 if (static_branch_unlikely(&userspace_irqchip_in_use))
 866                         kvm_timer_sync_user(vcpu);
 867
 868                 kvm_arch_vcpu_ctxsync_fp(vcpu);
 869
 870                 /*
 871                  * We may have taken a host interrupt in HYP mode (ie
 872                  * while executing the guest). This interrupt is still
 873                  * pending, as we haven't serviced it yet!
 874                  *
 875                  * We're now back in SVC mode, with interrupts
 876                  * disabled.  Enabling the interrupts now will have
 877                  * the effect of taking the interrupt again, in SVC
 878                  * mode this time.
 879                  */
 880                 local_irq_enable();
 881
 882                 /*
 883                  * We do local_irq_enable() before calling guest_exit() so
 884                  * that if a timer interrupt hits while running the guest we
 885                  * account that tick as being spent in the guest.  We enable
 886                  * preemption after calling guest_exit() so that if we get
 887                  * preempted we make sure ticks after that is not counted as
 888                  * guest time.
 889                  */
 890                 guest_exit();
 891                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
 892
 893                 /* Exit types that need handling before we can be preempted */
 894                 handle_exit_early(vcpu, ret);
 895
 896                 preempt_enable();
 897
 898                 /*
 899                  * The ARMv8 architecture doesn't give the hypervisor
 900                  * a mechanism to prevent a guest from dropping to AArch32 EL0
 901                  * if implemented by the CPU. If we spot the guest in such
 902                  * state and that we decided it wasn't supposed to do so (like
 903                  * with the asymmetric AArch32 case), return to userspace with
 904                  * a fatal error.
 905                  */
 906                 if (vcpu_mode_is_bad_32bit(vcpu)) {
 907                         /*
 908                          * As we have caught the guest red-handed, decide that
 909                          * it isn't fit for purpose anymore by making the vcpu
 910                          * invalid. The VMM can try and fix it by issuing  a
 911                          * KVM_ARM_VCPU_INIT if it really wants to.
 912                          */
 913                         vcpu->arch.target = -1;
 914                         ret = ARM_EXCEPTION_IL;
 915                 }
 916
 917                 ret = handle_exit(vcpu, ret);
 918         }
 919
 920         /* Tell userspace about in-kernel device output levels */
 921         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
 922                 kvm_timer_update_run(vcpu);
 923                 kvm_pmu_update_run(vcpu);
 924         }
 925
 926         kvm_sigset_deactivate(vcpu);
 927
 928 out:
 929         /*
 930          * In the unlikely event that we are returning to userspace
 931          * with pending exceptions or PC adjustment, commit these
 932          * adjustments in order to give userspace a consistent view of
 933          * the vcpu state. Note that this relies on __kvm_adjust_pc()
 934          * being preempt-safe on VHE.
 935          */
 936         if (unlikely(vcpu->arch.flags & (KVM_ARM64_PENDING_EXCEPTION |
 937                                          KVM_ARM64_INCREMENT_PC)))
 938                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
 939
 940         vcpu_put(vcpu);
 941         return ret;
 942 }
 943
 944 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
 945 {
 946         int bit_index;
 947         bool set;
 948         unsigned long *hcr;
 949
 950         if (number == KVM_ARM_IRQ_CPU_IRQ)
 951                 bit_index = __ffs(HCR_VI);
 952         else /* KVM_ARM_IRQ_CPU_FIQ */
 953                 bit_index = __ffs(HCR_VF);
 954
 955         hcr = vcpu_hcr(vcpu);
 956         if (level)
 957                 set = test_and_set_bit(bit_index, hcr);
 958         else
 959                 set = test_and_clear_bit(bit_index, hcr);
 960
 961         /*
 962          * If we didn't change anything, no need to wake up or kick other CPUs
 963          */
 964         if (set == level)
 965                 return 0;
 966
 967         /*
 968          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
 969          * trigger a world-switch round on the running physical CPU to set the
 970          * virtual IRQ/FIQ fields in the HCR appropriately.
 971          */
 972         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 973         kvm_vcpu_kick(vcpu);
 974
 975         return 0;
 976 }
 977
 978 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
 979                           bool line_status)
 980 {
 981         u32 irq = irq_level->irq;
 982         unsigned int irq_type, vcpu_idx, irq_num;
 983         int nrcpus = atomic_read(&kvm->online_vcpus);
 984         struct kvm_vcpu *vcpu = NULL;
 985         bool level = irq_level->level;
 986
 987         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
 988         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
 989         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
 990         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
 991
 992         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
 993
 994         switch (irq_type) {
 995         case KVM_ARM_IRQ_TYPE_CPU:
 996                 if (irqchip_in_kernel(kvm))
 997                         return -ENXIO;
 998
 999                 if (vcpu_idx >= nrcpus)
1000                         return -EINVAL;
1001
1002                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1003                 if (!vcpu)
1004                         return -EINVAL;
1005
1006                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1007                         return -EINVAL;
1008
1009                 return vcpu_interrupt_line(vcpu, irq_num, level);
1010         case KVM_ARM_IRQ_TYPE_PPI:
1011                 if (!irqchip_in_kernel(kvm))
1012                         return -ENXIO;
1013
1014                 if (vcpu_idx >= nrcpus)
1015                         return -EINVAL;
1016
1017                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1018                 if (!vcpu)
1019                         return -EINVAL;
1020
1021                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1022                         return -EINVAL;
1023
1024                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
1025         case KVM_ARM_IRQ_TYPE_SPI:
1026                 if (!irqchip_in_kernel(kvm))
1027                         return -ENXIO;
1028
1029                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1030                         return -EINVAL;
1031
1032                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
1033         }
1034
1035         return -EINVAL;
1036 }
1037
1038 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1039                                const struct kvm_vcpu_init *init)
1040 {
1041         unsigned int i, ret;
1042         int phys_target = kvm_target_cpu();
1043
1044         if (init->target != phys_target)
1045                 return -EINVAL;
1046
1047         /*
1048          * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
1049          * use the same target.
1050          */
1051         if (vcpu->arch.target != -1 && vcpu->arch.target != init->target)
1052                 return -EINVAL;
1053
1054         /* -ENOENT for unknown features, -EINVAL for invalid combinations. */
1055         for (i = 0; i < sizeof(init->features) * 8; i++) {
1056                 bool set = (init->features[i / 32] & (1 << (i % 32)));
1057
1058                 if (set && i >= KVM_VCPU_MAX_FEATURES)
1059                         return -ENOENT;
1060
1061                 /*
1062                  * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
1063                  * use the same feature set.
1064                  */
1065                 if (vcpu->arch.target != -1 && i < KVM_VCPU_MAX_FEATURES &&
1066                     test_bit(i, vcpu->arch.features) != set)
1067                         return -EINVAL;
1068
1069                 if (set)
1070                         set_bit(i, vcpu->arch.features);
1071         }
1072
1073         vcpu->arch.target = phys_target;
1074
1075         /* Now we know what it is, we can reset it. */
1076         ret = kvm_reset_vcpu(vcpu);
1077         if (ret) {
1078                 vcpu->arch.target = -1;
1079                 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
1080         }
1081
1082         return ret;
1083 }
1084
1085 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1086                                          struct kvm_vcpu_init *init)
1087 {
1088         int ret;
1089
1090         ret = kvm_vcpu_set_target(vcpu, init);
1091         if (ret)
1092                 return ret;
1093
1094         /*
1095          * Ensure a rebooted VM will fault in RAM pages and detect if the
1096          * guest MMU is turned off and flush the caches as needed.
1097          *
1098          * S2FWB enforces all memory accesses to RAM being cacheable,
1099          * ensuring that the data side is always coherent. We still
1100          * need to invalidate the I-cache though, as FWB does *not*
1101          * imply CTR_EL0.DIC.
1102          */
1103         if (vcpu->arch.has_run_once) {
1104                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1105                         stage2_unmap_vm(vcpu->kvm);
1106                 else
1107                         icache_inval_all_pou();
1108         }
1109
1110         vcpu_reset_hcr(vcpu);
1111
1112         /*
1113          * Handle the "start in power-off" case.
1114          */
1115         if (test_bit(KVM_ARM_VCPU_POWER_OFF, vcpu->arch.features))
1116                 vcpu_power_off(vcpu);
1117         else
1118                 vcpu->arch.power_off = false;
1119
1120         return 0;
1121 }
1122
1123 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1124                                  struct kvm_device_attr *attr)
1125 {
1126         int ret = -ENXIO;
1127
1128         switch (attr->group) {
1129         default:
1130                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1131                 break;
1132         }
1133
1134         return ret;
1135 }
1136
1137 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1138                                  struct kvm_device_attr *attr)
1139 {
1140         int ret = -ENXIO;
1141
1142         switch (attr->group) {
1143         default:
1144                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1145                 break;
1146         }
1147
1148         return ret;
1149 }
1150
1151 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1152                                  struct kvm_device_attr *attr)
1153 {
1154         int ret = -ENXIO;
1155
1156         switch (attr->group) {
1157         default:
1158                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1159                 break;
1160         }
1161
1162         return ret;
1163 }
1164
1165 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1166                                    struct kvm_vcpu_events *events)
1167 {
1168         memset(events, 0, sizeof(*events));
1169
1170         return __kvm_arm_vcpu_get_events(vcpu, events);
1171 }
1172
1173 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1174                                    struct kvm_vcpu_events *events)
1175 {
1176         int i;
1177
1178         /* check whether the reserved field is zero */
1179         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1180                 if (events->reserved[i])
1181                         return -EINVAL;
1182
1183         /* check whether the pad field is zero */
1184         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1185                 if (events->exception.pad[i])
1186                         return -EINVAL;
1187
1188         return __kvm_arm_vcpu_set_events(vcpu, events);
1189 }
1190
1191 long kvm_arch_vcpu_ioctl(struct file *filp,
1192                          unsigned int ioctl, unsigned long arg)
1193 {
1194         struct kvm_vcpu *vcpu = filp->private_data;
1195         void __user *argp = (void __user *)arg;
1196         struct kvm_device_attr attr;
1197         long r;
1198
1199         switch (ioctl) {
1200         case KVM_ARM_VCPU_INIT: {
1201                 struct kvm_vcpu_init init;
1202
1203                 r = -EFAULT;
1204                 if (copy_from_user(&init, argp, sizeof(init)))
1205                         break;
1206
1207                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1208                 break;
1209         }
1210         case KVM_SET_ONE_REG:
1211         case KVM_GET_ONE_REG: {
1212                 struct kvm_one_reg reg;
1213
1214                 r = -ENOEXEC;
1215                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1216                         break;
1217
1218                 r = -EFAULT;
1219                 if (copy_from_user(&reg, argp, sizeof(reg)))
1220                         break;
1221
1222                 if (ioctl == KVM_SET_ONE_REG)
1223                         r = kvm_arm_set_reg(vcpu, &reg);
1224                 else
1225                         r = kvm_arm_get_reg(vcpu, &reg);
1226                 break;
1227         }
1228         case KVM_GET_REG_LIST: {
1229                 struct kvm_reg_list __user *user_list = argp;
1230                 struct kvm_reg_list reg_list;
1231                 unsigned n;
1232
1233                 r = -ENOEXEC;
1234                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1235                         break;
1236
1237                 r = -EPERM;
1238                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1239                         break;
1240
1241                 r = -EFAULT;
1242                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1243                         break;
1244                 n = reg_list.n;
1245                 reg_list.n = kvm_arm_num_regs(vcpu);
1246                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1247                         break;
1248                 r = -E2BIG;
1249                 if (n < reg_list.n)
1250                         break;
1251                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1252                 break;
1253         }
1254         case KVM_SET_DEVICE_ATTR: {
1255                 r = -EFAULT;
1256                 if (copy_from_user(&attr, argp, sizeof(attr)))
1257                         break;
1258                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1259                 break;
1260         }
1261         case KVM_GET_DEVICE_ATTR: {
1262                 r = -EFAULT;
1263                 if (copy_from_user(&attr, argp, sizeof(attr)))
1264                         break;
1265                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1266                 break;
1267         }
1268         case KVM_HAS_DEVICE_ATTR: {
1269                 r = -EFAULT;
1270                 if (copy_from_user(&attr, argp, sizeof(attr)))
1271                         break;
1272                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1273                 break;
1274         }
1275         case KVM_GET_VCPU_EVENTS: {
1276                 struct kvm_vcpu_events events;
1277
1278                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1279                         return -EINVAL;
1280
1281                 if (copy_to_user(argp, &events, sizeof(events)))
1282                         return -EFAULT;
1283
1284                 return 0;
1285         }
1286         case KVM_SET_VCPU_EVENTS: {
1287                 struct kvm_vcpu_events events;
1288
1289                 if (copy_from_user(&events, argp, sizeof(events)))
1290                         return -EFAULT;
1291
1292                 return kvm_arm_vcpu_set_events(vcpu, &events);
1293         }
1294         case KVM_ARM_VCPU_FINALIZE: {
1295                 int what;
1296
1297                 if (!kvm_vcpu_initialized(vcpu))
1298                         return -ENOEXEC;
1299
1300                 if (get_user(what, (const int __user *)argp))
1301                         return -EFAULT;
1302
1303                 return kvm_arm_vcpu_finalize(vcpu, what);
1304         }
1305         default:
1306                 r = -EINVAL;
1307         }
1308
1309         return r;
1310 }
1311
1312 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1313 {
1314
1315 }
1316
1317 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
1318                                         const struct kvm_memory_slot *memslot)
1319 {
1320         kvm_flush_remote_tlbs(kvm);
1321 }
1322
1323 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1324                                         struct kvm_arm_device_addr *dev_addr)
1325 {
1326         unsigned long dev_id, type;
1327
1328         dev_id = (dev_addr->id & KVM_ARM_DEVICE_ID_MASK) >>
1329                 KVM_ARM_DEVICE_ID_SHIFT;
1330         type = (dev_addr->id & KVM_ARM_DEVICE_TYPE_MASK) >>
1331                 KVM_ARM_DEVICE_TYPE_SHIFT;
1332
1333         switch (dev_id) {
1334         case KVM_ARM_DEVICE_VGIC_V2:
1335                 if (!vgic_present)
1336                         return -ENXIO;
1337                 return kvm_vgic_addr(kvm, type, &dev_addr->addr, true);
1338         default:
1339                 return -ENODEV;
1340         }
1341 }
1342
1343 long kvm_arch_vm_ioctl(struct file *filp,
1344                        unsigned int ioctl, unsigned long arg)
1345 {
1346         struct kvm *kvm = filp->private_data;
1347         void __user *argp = (void __user *)arg;
1348
1349         switch (ioctl) {
1350         case KVM_CREATE_IRQCHIP: {
1351                 int ret;
1352                 if (!vgic_present)
1353                         return -ENXIO;
1354                 mutex_lock(&kvm->lock);
1355                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1356                 mutex_unlock(&kvm->lock);
1357                 return ret;
1358         }
1359         case KVM_ARM_SET_DEVICE_ADDR: {
1360                 struct kvm_arm_device_addr dev_addr;
1361
1362                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1363                         return -EFAULT;
1364                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1365         }
1366         case KVM_ARM_PREFERRED_TARGET: {
1367                 int err;
1368                 struct kvm_vcpu_init init;
1369
1370                 err = kvm_vcpu_preferred_target(&init);
1371                 if (err)
1372                         return err;
1373
1374                 if (copy_to_user(argp, &init, sizeof(init)))
1375                         return -EFAULT;
1376
1377                 return 0;
1378         }
1379         case KVM_ARM_MTE_COPY_TAGS: {
1380                 struct kvm_arm_copy_mte_tags copy_tags;
1381
1382                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1383                         return -EFAULT;
1384                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1385         }
1386         default:
1387                 return -EINVAL;
1388         }
1389 }
1390
1391 static unsigned long nvhe_percpu_size(void)
1392 {
1393         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1394                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1395 }
1396
1397 static unsigned long nvhe_percpu_order(void)
1398 {
1399         unsigned long size = nvhe_percpu_size();
1400
1401         return size ? get_order(size) : 0;
1402 }
1403
1404 /* A lookup table holding the hypervisor VA for each vector slot */
1405 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1406
1407 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1408 {
1409         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1410 }
1411
1412 static int kvm_init_vector_slots(void)
1413 {
1414         int err;
1415         void *base;
1416
1417         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1418         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1419
1420         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1421         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1422
1423         if (!cpus_have_const_cap(ARM64_SPECTRE_V3A))
1424                 return 0;
1425
1426         if (!has_vhe()) {
1427                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1428                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1429                 if (err)
1430                         return err;
1431         }
1432
1433         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1434         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1435         return 0;
1436 }
1437
1438 static void cpu_prepare_hyp_mode(int cpu)
1439 {
1440         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1441         unsigned long tcr;
1442
1443         /*
1444          * Calculate the raw per-cpu offset without a translation from the
1445          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1446          * so that we can use adr_l to access per-cpu variables in EL2.
1447          * Also drop the KASAN tag which gets in the way...
1448          */
1449         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1450                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1451
1452         params->mair_el2 = read_sysreg(mair_el1);
1453
1454         /*
1455          * The ID map may be configured to use an extended virtual address
1456          * range. This is only the case if system RAM is out of range for the
1457          * currently configured page size and VA_BITS, in which case we will
1458          * also need the extended virtual range for the HYP ID map, or we won't
1459          * be able to enable the EL2 MMU.
1460          *
1461          * However, at EL2, there is only one TTBR register, and we can't switch
1462          * between translation tables *and* update TCR_EL2.T0SZ at the same
1463          * time. Bottom line: we need to use the extended range with *both* our
1464          * translation tables.
1465          *
1466          * So use the same T0SZ value we use for the ID map.
1467          */
1468         tcr = (read_sysreg(tcr_el1) & TCR_EL2_MASK) | TCR_EL2_RES1;
1469         tcr &= ~TCR_T0SZ_MASK;
1470         tcr |= (idmap_t0sz & GENMASK(TCR_TxSZ_WIDTH - 1, 0)) << TCR_T0SZ_OFFSET;
1471         params->tcr_el2 = tcr;
1472
1473         params->stack_hyp_va = kern_hyp_va(per_cpu(kvm_arm_hyp_stack_page, cpu) + PAGE_SIZE);
1474         params->pgd_pa = kvm_mmu_get_httbr();
1475         if (is_protected_kvm_enabled())
1476                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1477         else
1478                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1479         params->vttbr = params->vtcr = 0;
1480
1481         /*
1482          * Flush the init params from the data cache because the struct will
1483          * be read while the MMU is off.
1484          */
1485         kvm_flush_dcache_to_poc(params, sizeof(*params));
1486 }
1487
1488 static void hyp_install_host_vector(void)
1489 {
1490         struct kvm_nvhe_init_params *params;
1491         struct arm_smccc_res res;
1492
1493         /* Switch from the HYP stub to our own HYP init vector */
1494         __hyp_set_vectors(kvm_get_idmap_vector());
1495
1496         /*
1497          * Call initialization code, and switch to the full blown HYP code.
1498          * If the cpucaps haven't been finalized yet, something has gone very
1499          * wrong, and hyp will crash and burn when it uses any
1500          * cpus_have_const_cap() wrapper.
1501          */
1502         BUG_ON(!system_capabilities_finalized());
1503         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1504         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1505         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1506 }
1507
1508 static void cpu_init_hyp_mode(void)
1509 {
1510         hyp_install_host_vector();
1511
1512         /*
1513          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1514          * at EL2.
1515          */
1516         if (this_cpu_has_cap(ARM64_SSBS) &&
1517             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1518                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1519         }
1520 }
1521
1522 static void cpu_hyp_reset(void)
1523 {
1524         if (!is_kernel_in_hyp_mode())
1525                 __hyp_reset_vectors();
1526 }
1527
1528 /*
1529  * EL2 vectors can be mapped and rerouted in a number of ways,
1530  * depending on the kernel configuration and CPU present:
1531  *
1532  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1533  *   placed in one of the vector slots, which is executed before jumping
1534  *   to the real vectors.
1535  *
1536  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1537  *   containing the hardening sequence is mapped next to the idmap page,
1538  *   and executed before jumping to the real vectors.
1539  *
1540  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1541  *   empty slot is selected, mapped next to the idmap page, and
1542  *   executed before jumping to the real vectors.
1543  *
1544  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1545  * VHE, as we don't have hypervisor-specific mappings. If the system
1546  * is VHE and yet selects this capability, it will be ignored.
1547  */
1548 static void cpu_set_hyp_vector(void)
1549 {
1550         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1551         void *vector = hyp_spectre_vector_selector[data->slot];
1552
1553         if (!is_protected_kvm_enabled())
1554                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1555         else
1556                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1557 }
1558
1559 static void cpu_hyp_reinit(void)
1560 {
1561         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1562
1563         cpu_hyp_reset();
1564
1565         if (is_kernel_in_hyp_mode())
1566                 kvm_timer_init_vhe();
1567         else
1568                 cpu_init_hyp_mode();
1569
1570         cpu_set_hyp_vector();
1571
1572         kvm_arm_init_debug();
1573
1574         if (vgic_present)
1575                 kvm_vgic_init_cpu_hardware();
1576 }
1577
1578 static void _kvm_arch_hardware_enable(void *discard)
1579 {
1580         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1581                 cpu_hyp_reinit();
1582                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1583         }
1584 }
1585
1586 int kvm_arch_hardware_enable(void)
1587 {
1588         _kvm_arch_hardware_enable(NULL);
1589         return 0;
1590 }
1591
1592 static void _kvm_arch_hardware_disable(void *discard)
1593 {
1594         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1595                 cpu_hyp_reset();
1596                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1597         }
1598 }
1599
1600 void kvm_arch_hardware_disable(void)
1601 {
1602         if (!is_protected_kvm_enabled())
1603                 _kvm_arch_hardware_disable(NULL);
1604 }
1605
1606 #ifdef CONFIG_CPU_PM
1607 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1608                                     unsigned long cmd,
1609                                     void *v)
1610 {
1611         /*
1612          * kvm_arm_hardware_enabled is left with its old value over
1613          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1614          * re-enable hyp.
1615          */
1616         switch (cmd) {
1617         case CPU_PM_ENTER:
1618                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1619                         /*
1620                          * don't update kvm_arm_hardware_enabled here
1621                          * so that the hardware will be re-enabled
1622                          * when we resume. See below.
1623                          */
1624                         cpu_hyp_reset();
1625
1626                 return NOTIFY_OK;
1627         case CPU_PM_ENTER_FAILED:
1628         case CPU_PM_EXIT:
1629                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1630                         /* The hardware was enabled before suspend. */
1631                         cpu_hyp_reinit();
1632
1633                 return NOTIFY_OK;
1634
1635         default:
1636                 return NOTIFY_DONE;
1637         }
1638 }
1639
1640 static struct notifier_block hyp_init_cpu_pm_nb = {
1641         .notifier_call = hyp_init_cpu_pm_notifier,
1642 };
1643
1644 static void hyp_cpu_pm_init(void)
1645 {
1646         if (!is_protected_kvm_enabled())
1647                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1648 }
1649 static void hyp_cpu_pm_exit(void)
1650 {
1651         if (!is_protected_kvm_enabled())
1652                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1653 }
1654 #else
1655 static inline void hyp_cpu_pm_init(void)
1656 {
1657 }
1658 static inline void hyp_cpu_pm_exit(void)
1659 {
1660 }
1661 #endif
1662
1663 static void init_cpu_logical_map(void)
1664 {
1665         unsigned int cpu;
1666
1667         /*
1668          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
1669          * Only copy the set of online CPUs whose features have been chacked
1670          * against the finalized system capabilities. The hypervisor will not
1671          * allow any other CPUs from the `possible` set to boot.
1672          */
1673         for_each_online_cpu(cpu)
1674                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
1675 }
1676
1677 #define init_psci_0_1_impl_state(config, what)  \
1678         config.psci_0_1_ ## what ## _implemented = psci_ops.what
1679
1680 static bool init_psci_relay(void)
1681 {
1682         /*
1683          * If PSCI has not been initialized, protected KVM cannot install
1684          * itself on newly booted CPUs.
1685          */
1686         if (!psci_ops.get_version) {
1687                 kvm_err("Cannot initialize protected mode without PSCI\n");
1688                 return false;
1689         }
1690
1691         kvm_host_psci_config.version = psci_ops.get_version();
1692
1693         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
1694                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
1695                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
1696                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
1697                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
1698                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
1699         }
1700         return true;
1701 }
1702
1703 static int init_common_resources(void)
1704 {
1705         return kvm_set_ipa_limit();
1706 }
1707
1708 static int init_subsystems(void)
1709 {
1710         int err = 0;
1711
1712         /*
1713          * Enable hardware so that subsystem initialisation can access EL2.
1714          */
1715         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
1716
1717         /*
1718          * Register CPU lower-power notifier
1719          */
1720         hyp_cpu_pm_init();
1721
1722         /*
1723          * Init HYP view of VGIC
1724          */
1725         err = kvm_vgic_hyp_init();
1726         switch (err) {
1727         case 0:
1728                 vgic_present = true;
1729                 break;
1730         case -ENODEV:
1731         case -ENXIO:
1732                 vgic_present = false;
1733                 err = 0;
1734                 break;
1735         default:
1736                 goto out;
1737         }
1738
1739         /*
1740          * Init HYP architected timer support
1741          */
1742         err = kvm_timer_hyp_init(vgic_present);
1743         if (err)
1744                 goto out;
1745
1746         kvm_perf_init();
1747         kvm_sys_reg_table_init();
1748
1749 out:
1750         if (err || !is_protected_kvm_enabled())
1751                 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
1752
1753         return err;
1754 }
1755
1756 static void teardown_hyp_mode(void)
1757 {
1758         int cpu;
1759
1760         free_hyp_pgds();
1761         for_each_possible_cpu(cpu) {
1762                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
1763                 free_pages(kvm_arm_hyp_percpu_base[cpu], nvhe_percpu_order());
1764         }
1765 }
1766
1767 static int do_pkvm_init(u32 hyp_va_bits)
1768 {
1769         void *per_cpu_base = kvm_ksym_ref(kvm_arm_hyp_percpu_base);
1770         int ret;
1771
1772         preempt_disable();
1773         hyp_install_host_vector();
1774         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
1775                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
1776                                 hyp_va_bits);
1777         preempt_enable();
1778
1779         return ret;
1780 }
1781
1782 static int kvm_hyp_init_protection(u32 hyp_va_bits)
1783 {
1784         void *addr = phys_to_virt(hyp_mem_base);
1785         int ret;
1786
1787         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
1788         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
1789
1790         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
1791         if (ret)
1792                 return ret;
1793
1794         ret = do_pkvm_init(hyp_va_bits);
1795         if (ret)
1796                 return ret;
1797
1798         free_hyp_pgds();
1799
1800         return 0;
1801 }
1802
1803 /**
1804  * Inits Hyp-mode on all online CPUs
1805  */
1806 static int init_hyp_mode(void)
1807 {
1808         u32 hyp_va_bits;
1809         int cpu;
1810         int err = -ENOMEM;
1811
1812         /*
1813          * The protected Hyp-mode cannot be initialized if the memory pool
1814          * allocation has failed.
1815          */
1816         if (is_protected_kvm_enabled() && !hyp_mem_base)
1817                 goto out_err;
1818
1819         /*
1820          * Allocate Hyp PGD and setup Hyp identity mapping
1821          */
1822         err = kvm_mmu_init(&hyp_va_bits);
1823         if (err)
1824                 goto out_err;
1825
1826         /*
1827          * Allocate stack pages for Hypervisor-mode
1828          */
1829         for_each_possible_cpu(cpu) {
1830                 unsigned long stack_page;
1831
1832                 stack_page = __get_free_page(GFP_KERNEL);
1833                 if (!stack_page) {
1834                         err = -ENOMEM;
1835                         goto out_err;
1836                 }
1837
1838                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
1839         }
1840
1841         /*
1842          * Allocate and initialize pages for Hypervisor-mode percpu regions.
1843          */
1844         for_each_possible_cpu(cpu) {
1845                 struct page *page;
1846                 void *page_addr;
1847
1848                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
1849                 if (!page) {
1850                         err = -ENOMEM;
1851                         goto out_err;
1852                 }
1853
1854                 page_addr = page_address(page);
1855                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
1856                 kvm_arm_hyp_percpu_base[cpu] = (unsigned long)page_addr;
1857         }
1858
1859         /*
1860          * Map the Hyp-code called directly from the host
1861          */
1862         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
1863                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
1864         if (err) {
1865                 kvm_err("Cannot map world-switch code\n");
1866                 goto out_err;
1867         }
1868
1869         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
1870                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
1871         if (err) {
1872                 kvm_err("Cannot map .hyp.rodata section\n");
1873                 goto out_err;
1874         }
1875
1876         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
1877                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
1878         if (err) {
1879                 kvm_err("Cannot map rodata section\n");
1880                 goto out_err;
1881         }
1882
1883         /*
1884          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
1885          * section thanks to an assertion in the linker script. Map it RW and
1886          * the rest of .bss RO.
1887          */
1888         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
1889                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
1890         if (err) {
1891                 kvm_err("Cannot map hyp bss section: %d\n", err);
1892                 goto out_err;
1893         }
1894
1895         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
1896                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
1897         if (err) {
1898                 kvm_err("Cannot map bss section\n");
1899                 goto out_err;
1900         }
1901
1902         /*
1903          * Map the Hyp stack pages
1904          */
1905         for_each_possible_cpu(cpu) {
1906                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
1907                 err = create_hyp_mappings(stack_page, stack_page + PAGE_SIZE,
1908                                           PAGE_HYP);
1909
1910                 if (err) {
1911                         kvm_err("Cannot map hyp stack\n");
1912                         goto out_err;
1913                 }
1914         }
1915
1916         for_each_possible_cpu(cpu) {
1917                 char *percpu_begin = (char *)kvm_arm_hyp_percpu_base[cpu];
1918                 char *percpu_end = percpu_begin + nvhe_percpu_size();
1919
1920                 /* Map Hyp percpu pages */
1921                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
1922                 if (err) {
1923                         kvm_err("Cannot map hyp percpu region\n");
1924                         goto out_err;
1925                 }
1926
1927                 /* Prepare the CPU initialization parameters */
1928                 cpu_prepare_hyp_mode(cpu);
1929         }
1930
1931         if (is_protected_kvm_enabled()) {
1932                 init_cpu_logical_map();
1933
1934                 if (!init_psci_relay()) {
1935                         err = -ENODEV;
1936                         goto out_err;
1937                 }
1938         }
1939
1940         if (is_protected_kvm_enabled()) {
1941                 err = kvm_hyp_init_protection(hyp_va_bits);
1942                 if (err) {
1943                         kvm_err("Failed to init hyp memory protection\n");
1944                         goto out_err;
1945                 }
1946         }
1947
1948         return 0;
1949
1950 out_err:
1951         teardown_hyp_mode();
1952         kvm_err("error initializing Hyp mode: %d\n", err);
1953         return err;
1954 }
1955
1956 static void _kvm_host_prot_finalize(void *discard)
1957 {
1958         WARN_ON(kvm_call_hyp_nvhe(__pkvm_prot_finalize));
1959 }
1960
1961 static inline int pkvm_mark_hyp(phys_addr_t start, phys_addr_t end)
1962 {
1963         return kvm_call_hyp_nvhe(__pkvm_mark_hyp, start, end);
1964 }
1965
1966 #define pkvm_mark_hyp_section(__section)                \
1967         pkvm_mark_hyp(__pa_symbol(__section##_start),   \
1968                         __pa_symbol(__section##_end))
1969
1970 static int finalize_hyp_mode(void)
1971 {
1972         int cpu, ret;
1973
1974         if (!is_protected_kvm_enabled())
1975                 return 0;
1976
1977         ret = pkvm_mark_hyp_section(__hyp_idmap_text);
1978         if (ret)
1979                 return ret;
1980
1981         ret = pkvm_mark_hyp_section(__hyp_text);
1982         if (ret)
1983                 return ret;
1984
1985         ret = pkvm_mark_hyp_section(__hyp_rodata);
1986         if (ret)
1987                 return ret;
1988
1989         ret = pkvm_mark_hyp_section(__hyp_bss);
1990         if (ret)
1991                 return ret;
1992
1993         ret = pkvm_mark_hyp(hyp_mem_base, hyp_mem_base + hyp_mem_size);
1994         if (ret)
1995                 return ret;
1996
1997         for_each_possible_cpu(cpu) {
1998                 phys_addr_t start = virt_to_phys((void *)kvm_arm_hyp_percpu_base[cpu]);
1999                 phys_addr_t end = start + (PAGE_SIZE << nvhe_percpu_order());
2000
2001                 ret = pkvm_mark_hyp(start, end);
2002                 if (ret)
2003                         return ret;
2004
2005                 start = virt_to_phys((void *)per_cpu(kvm_arm_hyp_stack_page, cpu));
2006                 end = start + PAGE_SIZE;
2007                 ret = pkvm_mark_hyp(start, end);
2008                 if (ret)
2009                         return ret;
2010         }
2011
2012         /*
2013          * Flip the static key upfront as that may no longer be possible
2014          * once the host stage 2 is installed.
2015          */
2016         static_branch_enable(&kvm_protected_mode_initialized);
2017         on_each_cpu(_kvm_host_prot_finalize, NULL, 1);
2018
2019         return 0;
2020 }
2021
2022 static void check_kvm_target_cpu(void *ret)
2023 {
2024         *(int *)ret = kvm_target_cpu();
2025 }
2026
2027 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2028 {
2029         struct kvm_vcpu *vcpu;
2030         int i;
2031
2032         mpidr &= MPIDR_HWID_BITMASK;
2033         kvm_for_each_vcpu(i, vcpu, kvm) {
2034                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2035                         return vcpu;
2036         }
2037         return NULL;
2038 }
2039
2040 bool kvm_arch_has_irq_bypass(void)
2041 {
2042         return true;
2043 }
2044
2045 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2046                                       struct irq_bypass_producer *prod)
2047 {
2048         struct kvm_kernel_irqfd *irqfd =
2049                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2050
2051         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2052                                           &irqfd->irq_entry);
2053 }
2054 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2055                                       struct irq_bypass_producer *prod)
2056 {
2057         struct kvm_kernel_irqfd *irqfd =
2058                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2059
2060         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2061                                      &irqfd->irq_entry);
2062 }
2063
2064 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2065 {
2066         struct kvm_kernel_irqfd *irqfd =
2067                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2068
2069         kvm_arm_halt_guest(irqfd->kvm);
2070 }
2071
2072 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2073 {
2074         struct kvm_kernel_irqfd *irqfd =
2075                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2076
2077         kvm_arm_resume_guest(irqfd->kvm);
2078 }
2079
2080 /**
2081  * Initialize Hyp-mode and memory mappings on all CPUs.
2082  */
2083 int kvm_arch_init(void *opaque)
2084 {
2085         int err;
2086         int ret, cpu;
2087         bool in_hyp_mode;
2088
2089         if (!is_hyp_mode_available()) {
2090                 kvm_info("HYP mode not available\n");
2091                 return -ENODEV;
2092         }
2093
2094         in_hyp_mode = is_kernel_in_hyp_mode();
2095
2096         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2097             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2098                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2099                          "Only trusted guests should be used on this system.\n");
2100
2101         for_each_online_cpu(cpu) {
2102                 smp_call_function_single(cpu, check_kvm_target_cpu, &ret, 1);
2103                 if (ret < 0) {
2104                         kvm_err("Error, CPU %d not supported!\n", cpu);
2105                         return -ENODEV;
2106                 }
2107         }
2108
2109         err = init_common_resources();
2110         if (err)
2111                 return err;
2112
2113         err = kvm_arm_init_sve();
2114         if (err)
2115                 return err;
2116
2117         if (!in_hyp_mode) {
2118                 err = init_hyp_mode();
2119                 if (err)
2120                         goto out_err;
2121         }
2122
2123         err = kvm_init_vector_slots();
2124         if (err) {
2125                 kvm_err("Cannot initialise vector slots\n");
2126                 goto out_err;
2127         }
2128
2129         err = init_subsystems();
2130         if (err)
2131                 goto out_hyp;
2132
2133         if (!in_hyp_mode) {
2134                 err = finalize_hyp_mode();
2135                 if (err) {
2136                         kvm_err("Failed to finalize Hyp protection\n");
2137                         goto out_hyp;
2138                 }
2139         }
2140
2141         if (is_protected_kvm_enabled()) {
2142                 kvm_info("Protected nVHE mode initialized successfully\n");
2143         } else if (in_hyp_mode) {
2144                 kvm_info("VHE mode initialized successfully\n");
2145         } else {
2146                 kvm_info("Hyp mode initialized successfully\n");
2147         }
2148
2149         return 0;
2150
2151 out_hyp:
2152         hyp_cpu_pm_exit();
2153         if (!in_hyp_mode)
2154                 teardown_hyp_mode();
2155 out_err:
2156         return err;
2157 }
2158
2159 /* NOP: Compiling as a module not supported */
2160 void kvm_arch_exit(void)
2161 {
2162         kvm_perf_teardown();
2163 }
2164
2165 static int __init early_kvm_mode_cfg(char *arg)
2166 {
2167         if (!arg)
2168                 return -EINVAL;
2169
2170         if (strcmp(arg, "protected") == 0) {
2171                 kvm_mode = KVM_MODE_PROTECTED;
2172                 return 0;
2173         }
2174
2175         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode()))
2176                 return 0;
2177
2178         return -EINVAL;
2179 }
2180 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2181
2182 enum kvm_mode kvm_get_mode(void)
2183 {
2184         return kvm_mode;
2185 }
2186
2187 static int arm_init(void)
2188 {
2189         int rc = kvm_init(NULL, sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2190         return rc;
2191 }
2192
2193 module_init(arm_init);