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