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