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 }
 703
 704 /**
 705  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 706  * @vcpu:       The VCPU pointer
 707  *
 708  * This function is called through the VCPU_RUN ioctl called from user space. It
 709  * will execute VM code in a loop until the time slice for the process is used
 710  * or some emulation is needed from user space in which case the function will
 711  * return with return value 0 and with the kvm_run structure filled in with the
 712  * required data for the requested emulation.
 713  */
 714 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 715 {
 716         struct kvm_run *run = vcpu->run;
 717         int ret;
 718
 719         if (unlikely(!kvm_vcpu_initialized(vcpu)))
 720                 return -ENOEXEC;
 721
 722         ret = kvm_vcpu_first_run_init(vcpu);
 723         if (ret)
 724                 return ret;
 725
 726         if (run->exit_reason == KVM_EXIT_MMIO) {
 727                 ret = kvm_handle_mmio_return(vcpu);
 728                 if (ret)
 729                         return ret;
 730         }
 731
 732         vcpu_load(vcpu);
 733
 734         if (run->immediate_exit) {
 735                 ret = -EINTR;
 736                 goto out;
 737         }
 738
 739         kvm_sigset_activate(vcpu);
 740
 741         ret = 1;
 742         run->exit_reason = KVM_EXIT_UNKNOWN;
 743         while (ret > 0) {
 744                 /*
 745                  * Check conditions before entering the guest
 746                  */
 747                 cond_resched();
 748
 749                 update_vmid(&vcpu->arch.hw_mmu->vmid);
 750
 751                 check_vcpu_requests(vcpu);
 752
 753                 /*
 754                  * Preparing the interrupts to be injected also
 755                  * involves poking the GIC, which must be done in a
 756                  * non-preemptible context.
 757                  */
 758                 preempt_disable();
 759
 760                 kvm_pmu_flush_hwstate(vcpu);
 761
 762                 local_irq_disable();
 763
 764                 kvm_vgic_flush_hwstate(vcpu);
 765
 766                 /*
 767                  * Exit if we have a signal pending so that we can deliver the
 768                  * signal to user space.
 769                  */
 770                 if (signal_pending(current)) {
 771                         ret = -EINTR;
 772                         run->exit_reason = KVM_EXIT_INTR;
 773                 }
 774
 775                 /*
 776                  * If we're using a userspace irqchip, then check if we need
 777                  * to tell a userspace irqchip about timer or PMU level
 778                  * changes and if so, exit to userspace (the actual level
 779                  * state gets updated in kvm_timer_update_run and
 780                  * kvm_pmu_update_run below).
 781                  */
 782                 if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 783                         if (kvm_timer_should_notify_user(vcpu) ||
 784                             kvm_pmu_should_notify_user(vcpu)) {
 785                                 ret = -EINTR;
 786                                 run->exit_reason = KVM_EXIT_INTR;
 787                         }
 788                 }
 789
 790                 /*
 791                  * Ensure we set mode to IN_GUEST_MODE after we disable
 792                  * interrupts and before the final VCPU requests check.
 793                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 794                  * Documentation/virt/kvm/vcpu-requests.rst
 795                  */
 796                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 797
 798                 if (ret <= 0 || need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
 799                     kvm_request_pending(vcpu)) {
 800                         vcpu->mode = OUTSIDE_GUEST_MODE;
 801                         isb(); /* Ensure work in x_flush_hwstate is committed */
 802                         kvm_pmu_sync_hwstate(vcpu);
 803                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 804                                 kvm_timer_sync_user(vcpu);
 805                         kvm_vgic_sync_hwstate(vcpu);
 806                         local_irq_enable();
 807                         preempt_enable();
 808                         continue;
 809                 }
 810
 811                 kvm_arm_setup_debug(vcpu);
 812
 813                 /**************************************************************
 814                  * Enter the guest
 815                  */
 816                 trace_kvm_entry(*vcpu_pc(vcpu));
 817                 guest_enter_irqoff();
 818
 819                 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 820
 821                 vcpu->mode = OUTSIDE_GUEST_MODE;
 822                 vcpu->stat.exits++;
 823                 /*
 824                  * Back from guest
 825                  *************************************************************/
 826
 827                 kvm_arm_clear_debug(vcpu);
 828
 829                 /*
 830                  * We must sync the PMU state before the vgic state so
 831                  * that the vgic can properly sample the updated state of the
 832                  * interrupt line.
 833                  */
 834                 kvm_pmu_sync_hwstate(vcpu);
 835
 836                 /*
 837                  * Sync the vgic state before syncing the timer state because
 838                  * the timer code needs to know if the virtual timer
 839                  * interrupts are active.
 840                  */
 841                 kvm_vgic_sync_hwstate(vcpu);
 842
 843                 /*
 844                  * Sync the timer hardware state before enabling interrupts as
 845                  * we don't want vtimer interrupts to race with syncing the
 846                  * timer virtual interrupt state.
 847                  */
 848                 if (static_branch_unlikely(&userspace_irqchip_in_use))
 849                         kvm_timer_sync_user(vcpu);
 850
 851                 kvm_arch_vcpu_ctxsync_fp(vcpu);
 852
 853                 /*
 854                  * We may have taken a host interrupt in HYP mode (ie
 855                  * while executing the guest). This interrupt is still
 856                  * pending, as we haven't serviced it yet!
 857                  *
 858                  * We're now back in SVC mode, with interrupts
 859                  * disabled.  Enabling the interrupts now will have
 860                  * the effect of taking the interrupt again, in SVC
 861                  * mode this time.
 862                  */
 863                 local_irq_enable();
 864
 865                 /*
 866                  * We do local_irq_enable() before calling guest_exit() so
 867                  * that if a timer interrupt hits while running the guest we
 868                  * account that tick as being spent in the guest.  We enable
 869                  * preemption after calling guest_exit() so that if we get
 870                  * preempted we make sure ticks after that is not counted as
 871                  * guest time.
 872                  */
 873                 guest_exit();
 874                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
 875
 876                 /* Exit types that need handling before we can be preempted */
 877                 handle_exit_early(vcpu, ret);
 878
 879                 preempt_enable();
 880
 881                 /*
 882                  * The ARMv8 architecture doesn't give the hypervisor
 883                  * a mechanism to prevent a guest from dropping to AArch32 EL0
 884                  * if implemented by the CPU. If we spot the guest in such
 885                  * state and that we decided it wasn't supposed to do so (like
 886                  * with the asymmetric AArch32 case), return to userspace with
 887                  * a fatal error.
 888                  */
 889                 if (!system_supports_32bit_el0() && vcpu_mode_is_32bit(vcpu)) {
 890                         /*
 891                          * As we have caught the guest red-handed, decide that
 892                          * it isn't fit for purpose anymore by making the vcpu
 893                          * invalid. The VMM can try and fix it by issuing  a
 894                          * KVM_ARM_VCPU_INIT if it really wants to.
 895                          */
 896                         vcpu->arch.target = -1;
 897                         ret = ARM_EXCEPTION_IL;
 898                 }
 899
 900                 ret = handle_exit(vcpu, ret);
 901         }
 902
 903         /* Tell userspace about in-kernel device output levels */
 904         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
 905                 kvm_timer_update_run(vcpu);
 906                 kvm_pmu_update_run(vcpu);
 907         }
 908
 909         kvm_sigset_deactivate(vcpu);
 910
 911 out:
 912         /*
 913          * In the unlikely event that we are returning to userspace
 914          * with pending exceptions or PC adjustment, commit these
 915          * adjustments in order to give userspace a consistent view of
 916          * the vcpu state. Note that this relies on __kvm_adjust_pc()
 917          * being preempt-safe on VHE.
 918          */
 919         if (unlikely(vcpu->arch.flags & (KVM_ARM64_PENDING_EXCEPTION |
 920                                          KVM_ARM64_INCREMENT_PC)))
 921                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
 922
 923         vcpu_put(vcpu);
 924         return ret;
 925 }
 926
 927 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
 928 {
 929         int bit_index;
 930         bool set;
 931         unsigned long *hcr;
 932
 933         if (number == KVM_ARM_IRQ_CPU_IRQ)
 934                 bit_index = __ffs(HCR_VI);
 935         else /* KVM_ARM_IRQ_CPU_FIQ */
 936                 bit_index = __ffs(HCR_VF);
 937
 938         hcr = vcpu_hcr(vcpu);
 939         if (level)
 940                 set = test_and_set_bit(bit_index, hcr);
 941         else
 942                 set = test_and_clear_bit(bit_index, hcr);
 943
 944         /*
 945          * If we didn't change anything, no need to wake up or kick other CPUs
 946          */
 947         if (set == level)
 948                 return 0;
 949
 950         /*
 951          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
 952          * trigger a world-switch round on the running physical CPU to set the
 953          * virtual IRQ/FIQ fields in the HCR appropriately.
 954          */
 955         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 956         kvm_vcpu_kick(vcpu);
 957
 958         return 0;
 959 }
 960
 961 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
 962                           bool line_status)
 963 {
 964         u32 irq = irq_level->irq;
 965         unsigned int irq_type, vcpu_idx, irq_num;
 966         int nrcpus = atomic_read(&kvm->online_vcpus);
 967         struct kvm_vcpu *vcpu = NULL;
 968         bool level = irq_level->level;
 969
 970         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
 971         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
 972         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
 973         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
 974
 975         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
 976
 977         switch (irq_type) {
 978         case KVM_ARM_IRQ_TYPE_CPU:
 979                 if (irqchip_in_kernel(kvm))
 980                         return -ENXIO;
 981
 982                 if (vcpu_idx >= nrcpus)
 983                         return -EINVAL;
 984
 985                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 986                 if (!vcpu)
 987                         return -EINVAL;
 988
 989                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
 990                         return -EINVAL;
 991
 992                 return vcpu_interrupt_line(vcpu, irq_num, level);
 993         case KVM_ARM_IRQ_TYPE_PPI:
 994                 if (!irqchip_in_kernel(kvm))
 995                         return -ENXIO;
 996
 997                 if (vcpu_idx >= nrcpus)
 998                         return -EINVAL;
 999
1000                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1001                 if (!vcpu)
1002                         return -EINVAL;
1003
1004                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1005                         return -EINVAL;
1006
1007                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
1008         case KVM_ARM_IRQ_TYPE_SPI:
1009                 if (!irqchip_in_kernel(kvm))
1010                         return -ENXIO;
1011
1012                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1013                         return -EINVAL;
1014
1015                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
1016         }
1017
1018         return -EINVAL;
1019 }
1020
1021 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1022                                const struct kvm_vcpu_init *init)
1023 {
1024         unsigned int i, ret;
1025         int phys_target = kvm_target_cpu();
1026
1027         if (init->target != phys_target)
1028                 return -EINVAL;
1029
1030         /*
1031          * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
1032          * use the same target.
1033          */
1034         if (vcpu->arch.target != -1 && vcpu->arch.target != init->target)
1035                 return -EINVAL;
1036
1037         /* -ENOENT for unknown features, -EINVAL for invalid combinations. */
1038         for (i = 0; i < sizeof(init->features) * 8; i++) {
1039                 bool set = (init->features[i / 32] & (1 << (i % 32)));
1040
1041                 if (set && i >= KVM_VCPU_MAX_FEATURES)
1042                         return -ENOENT;
1043
1044                 /*
1045                  * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
1046                  * use the same feature set.
1047                  */
1048                 if (vcpu->arch.target != -1 && i < KVM_VCPU_MAX_FEATURES &&
1049                     test_bit(i, vcpu->arch.features) != set)
1050                         return -EINVAL;
1051
1052                 if (set)
1053                         set_bit(i, vcpu->arch.features);
1054         }
1055
1056         vcpu->arch.target = phys_target;
1057
1058         /* Now we know what it is, we can reset it. */
1059         ret = kvm_reset_vcpu(vcpu);
1060         if (ret) {
1061                 vcpu->arch.target = -1;
1062                 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
1063         }
1064
1065         return ret;
1066 }
1067
1068 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1069                                          struct kvm_vcpu_init *init)
1070 {
1071         int ret;
1072
1073         ret = kvm_vcpu_set_target(vcpu, init);
1074         if (ret)
1075                 return ret;
1076
1077         /*
1078          * Ensure a rebooted VM will fault in RAM pages and detect if the
1079          * guest MMU is turned off and flush the caches as needed.
1080          *
1081          * S2FWB enforces all memory accesses to RAM being cacheable,
1082          * ensuring that the data side is always coherent. We still
1083          * need to invalidate the I-cache though, as FWB does *not*
1084          * imply CTR_EL0.DIC.
1085          */
1086         if (vcpu->arch.has_run_once) {
1087                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1088                         stage2_unmap_vm(vcpu->kvm);
1089                 else
1090                         __flush_icache_all();
1091         }
1092
1093         vcpu_reset_hcr(vcpu);
1094
1095         /*
1096          * Handle the "start in power-off" case.
1097          */
1098         if (test_bit(KVM_ARM_VCPU_POWER_OFF, vcpu->arch.features))
1099                 vcpu_power_off(vcpu);
1100         else
1101                 vcpu->arch.power_off = false;
1102
1103         return 0;
1104 }
1105
1106 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1107                                  struct kvm_device_attr *attr)
1108 {
1109         int ret = -ENXIO;
1110
1111         switch (attr->group) {
1112         default:
1113                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1114                 break;
1115         }
1116
1117         return ret;
1118 }
1119
1120 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1121                                  struct kvm_device_attr *attr)
1122 {
1123         int ret = -ENXIO;
1124
1125         switch (attr->group) {
1126         default:
1127                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1128                 break;
1129         }
1130
1131         return ret;
1132 }
1133
1134 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1135                                  struct kvm_device_attr *attr)
1136 {
1137         int ret = -ENXIO;
1138
1139         switch (attr->group) {
1140         default:
1141                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1142                 break;
1143         }
1144
1145         return ret;
1146 }
1147
1148 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1149                                    struct kvm_vcpu_events *events)
1150 {
1151         memset(events, 0, sizeof(*events));
1152
1153         return __kvm_arm_vcpu_get_events(vcpu, events);
1154 }
1155
1156 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1157                                    struct kvm_vcpu_events *events)
1158 {
1159         int i;
1160
1161         /* check whether the reserved field is zero */
1162         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1163                 if (events->reserved[i])
1164                         return -EINVAL;
1165
1166         /* check whether the pad field is zero */
1167         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1168                 if (events->exception.pad[i])
1169                         return -EINVAL;
1170
1171         return __kvm_arm_vcpu_set_events(vcpu, events);
1172 }
1173
1174 long kvm_arch_vcpu_ioctl(struct file *filp,
1175                          unsigned int ioctl, unsigned long arg)
1176 {
1177         struct kvm_vcpu *vcpu = filp->private_data;
1178         void __user *argp = (void __user *)arg;
1179         struct kvm_device_attr attr;
1180         long r;
1181
1182         switch (ioctl) {
1183         case KVM_ARM_VCPU_INIT: {
1184                 struct kvm_vcpu_init init;
1185
1186                 r = -EFAULT;
1187                 if (copy_from_user(&init, argp, sizeof(init)))
1188                         break;
1189
1190                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1191                 break;
1192         }
1193         case KVM_SET_ONE_REG:
1194         case KVM_GET_ONE_REG: {
1195                 struct kvm_one_reg reg;
1196
1197                 r = -ENOEXEC;
1198                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1199                         break;
1200
1201                 r = -EFAULT;
1202                 if (copy_from_user(&reg, argp, sizeof(reg)))
1203                         break;
1204
1205                 if (ioctl == KVM_SET_ONE_REG)
1206                         r = kvm_arm_set_reg(vcpu, &reg);
1207                 else
1208                         r = kvm_arm_get_reg(vcpu, &reg);
1209                 break;
1210         }
1211         case KVM_GET_REG_LIST: {
1212                 struct kvm_reg_list __user *user_list = argp;
1213                 struct kvm_reg_list reg_list;
1214                 unsigned n;
1215
1216                 r = -ENOEXEC;
1217                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1218                         break;
1219
1220                 r = -EPERM;
1221                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1222                         break;
1223
1224                 r = -EFAULT;
1225                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1226                         break;
1227                 n = reg_list.n;
1228                 reg_list.n = kvm_arm_num_regs(vcpu);
1229                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1230                         break;
1231                 r = -E2BIG;
1232                 if (n < reg_list.n)
1233                         break;
1234                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1235                 break;
1236         }
1237         case KVM_SET_DEVICE_ATTR: {
1238                 r = -EFAULT;
1239                 if (copy_from_user(&attr, argp, sizeof(attr)))
1240                         break;
1241                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1242                 break;
1243         }
1244         case KVM_GET_DEVICE_ATTR: {
1245                 r = -EFAULT;
1246                 if (copy_from_user(&attr, argp, sizeof(attr)))
1247                         break;
1248                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1249                 break;
1250         }
1251         case KVM_HAS_DEVICE_ATTR: {
1252                 r = -EFAULT;
1253                 if (copy_from_user(&attr, argp, sizeof(attr)))
1254                         break;
1255                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1256                 break;
1257         }
1258         case KVM_GET_VCPU_EVENTS: {
1259                 struct kvm_vcpu_events events;
1260
1261                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1262                         return -EINVAL;
1263
1264                 if (copy_to_user(argp, &events, sizeof(events)))
1265                         return -EFAULT;
1266
1267                 return 0;
1268         }
1269         case KVM_SET_VCPU_EVENTS: {
1270                 struct kvm_vcpu_events events;
1271
1272                 if (copy_from_user(&events, argp, sizeof(events)))
1273                         return -EFAULT;
1274
1275                 return kvm_arm_vcpu_set_events(vcpu, &events);
1276         }
1277         case KVM_ARM_VCPU_FINALIZE: {
1278                 int what;
1279
1280                 if (!kvm_vcpu_initialized(vcpu))
1281                         return -ENOEXEC;
1282
1283                 if (get_user(what, (const int __user *)argp))
1284                         return -EFAULT;
1285
1286                 return kvm_arm_vcpu_finalize(vcpu, what);
1287         }
1288         default:
1289                 r = -EINVAL;
1290         }
1291
1292         return r;
1293 }
1294
1295 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1296 {
1297
1298 }
1299
1300 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
1301                                         const struct kvm_memory_slot *memslot)
1302 {
1303         kvm_flush_remote_tlbs(kvm);
1304 }
1305
1306 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1307                                         struct kvm_arm_device_addr *dev_addr)
1308 {
1309         unsigned long dev_id, type;
1310
1311         dev_id = (dev_addr->id & KVM_ARM_DEVICE_ID_MASK) >>
1312                 KVM_ARM_DEVICE_ID_SHIFT;
1313         type = (dev_addr->id & KVM_ARM_DEVICE_TYPE_MASK) >>
1314                 KVM_ARM_DEVICE_TYPE_SHIFT;
1315
1316         switch (dev_id) {
1317         case KVM_ARM_DEVICE_VGIC_V2:
1318                 if (!vgic_present)
1319                         return -ENXIO;
1320                 return kvm_vgic_addr(kvm, type, &dev_addr->addr, true);
1321         default:
1322                 return -ENODEV;
1323         }
1324 }
1325
1326 long kvm_arch_vm_ioctl(struct file *filp,
1327                        unsigned int ioctl, unsigned long arg)
1328 {
1329         struct kvm *kvm = filp->private_data;
1330         void __user *argp = (void __user *)arg;
1331
1332         switch (ioctl) {
1333         case KVM_CREATE_IRQCHIP: {
1334                 int ret;
1335                 if (!vgic_present)
1336                         return -ENXIO;
1337                 mutex_lock(&kvm->lock);
1338                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1339                 mutex_unlock(&kvm->lock);
1340                 return ret;
1341         }
1342         case KVM_ARM_SET_DEVICE_ADDR: {
1343                 struct kvm_arm_device_addr dev_addr;
1344
1345                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1346                         return -EFAULT;
1347                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1348         }
1349         case KVM_ARM_PREFERRED_TARGET: {
1350                 int err;
1351                 struct kvm_vcpu_init init;
1352
1353                 err = kvm_vcpu_preferred_target(&init);
1354                 if (err)
1355                         return err;
1356
1357                 if (copy_to_user(argp, &init, sizeof(init)))
1358                         return -EFAULT;
1359
1360                 return 0;
1361         }
1362         case KVM_ARM_MTE_COPY_TAGS: {
1363                 struct kvm_arm_copy_mte_tags copy_tags;
1364
1365                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1366                         return -EFAULT;
1367                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1368         }
1369         default:
1370                 return -EINVAL;
1371         }
1372 }
1373
1374 static unsigned long nvhe_percpu_size(void)
1375 {
1376         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1377                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1378 }
1379
1380 static unsigned long nvhe_percpu_order(void)
1381 {
1382         unsigned long size = nvhe_percpu_size();
1383
1384         return size ? get_order(size) : 0;
1385 }
1386
1387 /* A lookup table holding the hypervisor VA for each vector slot */
1388 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1389
1390 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1391 {
1392         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1393 }
1394
1395 static int kvm_init_vector_slots(void)
1396 {
1397         int err;
1398         void *base;
1399
1400         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1401         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1402
1403         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1404         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1405
1406         if (!cpus_have_const_cap(ARM64_SPECTRE_V3A))
1407                 return 0;
1408
1409         if (!has_vhe()) {
1410                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1411                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1412                 if (err)
1413                         return err;
1414         }
1415
1416         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1417         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1418         return 0;
1419 }
1420
1421 static void cpu_prepare_hyp_mode(int cpu)
1422 {
1423         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1424         unsigned long tcr;
1425
1426         /*
1427          * Calculate the raw per-cpu offset without a translation from the
1428          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1429          * so that we can use adr_l to access per-cpu variables in EL2.
1430          * Also drop the KASAN tag which gets in the way...
1431          */
1432         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1433                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1434
1435         params->mair_el2 = read_sysreg(mair_el1);
1436
1437         /*
1438          * The ID map may be configured to use an extended virtual address
1439          * range. This is only the case if system RAM is out of range for the
1440          * currently configured page size and VA_BITS, in which case we will
1441          * also need the extended virtual range for the HYP ID map, or we won't
1442          * be able to enable the EL2 MMU.
1443          *
1444          * However, at EL2, there is only one TTBR register, and we can't switch
1445          * between translation tables *and* update TCR_EL2.T0SZ at the same
1446          * time. Bottom line: we need to use the extended range with *both* our
1447          * translation tables.
1448          *
1449          * So use the same T0SZ value we use for the ID map.
1450          */
1451         tcr = (read_sysreg(tcr_el1) & TCR_EL2_MASK) | TCR_EL2_RES1;
1452         tcr &= ~TCR_T0SZ_MASK;
1453         tcr |= (idmap_t0sz & GENMASK(TCR_TxSZ_WIDTH - 1, 0)) << TCR_T0SZ_OFFSET;
1454         params->tcr_el2 = tcr;
1455
1456         params->stack_hyp_va = kern_hyp_va(per_cpu(kvm_arm_hyp_stack_page, cpu) + PAGE_SIZE);
1457         params->pgd_pa = kvm_mmu_get_httbr();
1458         if (is_protected_kvm_enabled())
1459                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1460         else
1461                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1462         params->vttbr = params->vtcr = 0;
1463
1464         /*
1465          * Flush the init params from the data cache because the struct will
1466          * be read while the MMU is off.
1467          */
1468         kvm_flush_dcache_to_poc(params, sizeof(*params));
1469 }
1470
1471 static void hyp_install_host_vector(void)
1472 {
1473         struct kvm_nvhe_init_params *params;
1474         struct arm_smccc_res res;
1475
1476         /* Switch from the HYP stub to our own HYP init vector */
1477         __hyp_set_vectors(kvm_get_idmap_vector());
1478
1479         /*
1480          * Call initialization code, and switch to the full blown HYP code.
1481          * If the cpucaps haven't been finalized yet, something has gone very
1482          * wrong, and hyp will crash and burn when it uses any
1483          * cpus_have_const_cap() wrapper.
1484          */
1485         BUG_ON(!system_capabilities_finalized());
1486         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1487         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1488         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1489 }
1490
1491 static void cpu_init_hyp_mode(void)
1492 {
1493         hyp_install_host_vector();
1494
1495         /*
1496          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1497          * at EL2.
1498          */
1499         if (this_cpu_has_cap(ARM64_SSBS) &&
1500             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1501                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1502         }
1503 }
1504
1505 static void cpu_hyp_reset(void)
1506 {
1507         if (!is_kernel_in_hyp_mode())
1508                 __hyp_reset_vectors();
1509 }
1510
1511 /*
1512  * EL2 vectors can be mapped and rerouted in a number of ways,
1513  * depending on the kernel configuration and CPU present:
1514  *
1515  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1516  *   placed in one of the vector slots, which is executed before jumping
1517  *   to the real vectors.
1518  *
1519  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1520  *   containing the hardening sequence is mapped next to the idmap page,
1521  *   and executed before jumping to the real vectors.
1522  *
1523  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1524  *   empty slot is selected, mapped next to the idmap page, and
1525  *   executed before jumping to the real vectors.
1526  *
1527  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1528  * VHE, as we don't have hypervisor-specific mappings. If the system
1529  * is VHE and yet selects this capability, it will be ignored.
1530  */
1531 static void cpu_set_hyp_vector(void)
1532 {
1533         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1534         void *vector = hyp_spectre_vector_selector[data->slot];
1535
1536         if (!is_protected_kvm_enabled())
1537                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1538         else
1539                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1540 }
1541
1542 static void cpu_hyp_reinit(void)
1543 {
1544         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1545
1546         cpu_hyp_reset();
1547
1548         if (is_kernel_in_hyp_mode())
1549                 kvm_timer_init_vhe();
1550         else
1551                 cpu_init_hyp_mode();
1552
1553         cpu_set_hyp_vector();
1554
1555         kvm_arm_init_debug();
1556
1557         if (vgic_present)
1558                 kvm_vgic_init_cpu_hardware();
1559 }
1560
1561 static void _kvm_arch_hardware_enable(void *discard)
1562 {
1563         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1564                 cpu_hyp_reinit();
1565                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1566         }
1567 }
1568
1569 int kvm_arch_hardware_enable(void)
1570 {
1571         _kvm_arch_hardware_enable(NULL);
1572         return 0;
1573 }
1574
1575 static void _kvm_arch_hardware_disable(void *discard)
1576 {
1577         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1578                 cpu_hyp_reset();
1579                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1580         }
1581 }
1582
1583 void kvm_arch_hardware_disable(void)
1584 {
1585         if (!is_protected_kvm_enabled())
1586                 _kvm_arch_hardware_disable(NULL);
1587 }
1588
1589 #ifdef CONFIG_CPU_PM
1590 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1591                                     unsigned long cmd,
1592                                     void *v)
1593 {
1594         /*
1595          * kvm_arm_hardware_enabled is left with its old value over
1596          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1597          * re-enable hyp.
1598          */
1599         switch (cmd) {
1600         case CPU_PM_ENTER:
1601                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1602                         /*
1603                          * don't update kvm_arm_hardware_enabled here
1604                          * so that the hardware will be re-enabled
1605                          * when we resume. See below.
1606                          */
1607                         cpu_hyp_reset();
1608
1609                 return NOTIFY_OK;
1610         case CPU_PM_ENTER_FAILED:
1611         case CPU_PM_EXIT:
1612                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1613                         /* The hardware was enabled before suspend. */
1614                         cpu_hyp_reinit();
1615
1616                 return NOTIFY_OK;
1617
1618         default:
1619                 return NOTIFY_DONE;
1620         }
1621 }
1622
1623 static struct notifier_block hyp_init_cpu_pm_nb = {
1624         .notifier_call = hyp_init_cpu_pm_notifier,
1625 };
1626
1627 static void hyp_cpu_pm_init(void)
1628 {
1629         if (!is_protected_kvm_enabled())
1630                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1631 }
1632 static void hyp_cpu_pm_exit(void)
1633 {
1634         if (!is_protected_kvm_enabled())
1635                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1636 }
1637 #else
1638 static inline void hyp_cpu_pm_init(void)
1639 {
1640 }
1641 static inline void hyp_cpu_pm_exit(void)
1642 {
1643 }
1644 #endif
1645
1646 static void init_cpu_logical_map(void)
1647 {
1648         unsigned int cpu;
1649
1650         /*
1651          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
1652          * Only copy the set of online CPUs whose features have been chacked
1653          * against the finalized system capabilities. The hypervisor will not
1654          * allow any other CPUs from the `possible` set to boot.
1655          */
1656         for_each_online_cpu(cpu)
1657                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
1658 }
1659
1660 #define init_psci_0_1_impl_state(config, what)  \
1661         config.psci_0_1_ ## what ## _implemented = psci_ops.what
1662
1663 static bool init_psci_relay(void)
1664 {
1665         /*
1666          * If PSCI has not been initialized, protected KVM cannot install
1667          * itself on newly booted CPUs.
1668          */
1669         if (!psci_ops.get_version) {
1670                 kvm_err("Cannot initialize protected mode without PSCI\n");
1671                 return false;
1672         }
1673
1674         kvm_host_psci_config.version = psci_ops.get_version();
1675
1676         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
1677                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
1678                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
1679                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
1680                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
1681                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
1682         }
1683         return true;
1684 }
1685
1686 static int init_common_resources(void)
1687 {
1688         return kvm_set_ipa_limit();
1689 }
1690
1691 static int init_subsystems(void)
1692 {
1693         int err = 0;
1694
1695         /*
1696          * Enable hardware so that subsystem initialisation can access EL2.
1697          */
1698         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
1699
1700         /*
1701          * Register CPU lower-power notifier
1702          */
1703         hyp_cpu_pm_init();
1704
1705         /*
1706          * Init HYP view of VGIC
1707          */
1708         err = kvm_vgic_hyp_init();
1709         switch (err) {
1710         case 0:
1711                 vgic_present = true;
1712                 break;
1713         case -ENODEV:
1714         case -ENXIO:
1715                 vgic_present = false;
1716                 err = 0;
1717                 break;
1718         default:
1719                 goto out;
1720         }
1721
1722         /*
1723          * Init HYP architected timer support
1724          */
1725         err = kvm_timer_hyp_init(vgic_present);
1726         if (err)
1727                 goto out;
1728
1729         kvm_perf_init();
1730         kvm_sys_reg_table_init();
1731
1732 out:
1733         if (err || !is_protected_kvm_enabled())
1734                 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
1735
1736         return err;
1737 }
1738
1739 static void teardown_hyp_mode(void)
1740 {
1741         int cpu;
1742
1743         free_hyp_pgds();
1744         for_each_possible_cpu(cpu) {
1745                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
1746                 free_pages(kvm_arm_hyp_percpu_base[cpu], nvhe_percpu_order());
1747         }
1748 }
1749
1750 static int do_pkvm_init(u32 hyp_va_bits)
1751 {
1752         void *per_cpu_base = kvm_ksym_ref(kvm_arm_hyp_percpu_base);
1753         int ret;
1754
1755         preempt_disable();
1756         hyp_install_host_vector();
1757         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
1758                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
1759                                 hyp_va_bits);
1760         preempt_enable();
1761
1762         return ret;
1763 }
1764
1765 static int kvm_hyp_init_protection(u32 hyp_va_bits)
1766 {
1767         void *addr = phys_to_virt(hyp_mem_base);
1768         int ret;
1769
1770         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
1771         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
1772
1773         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
1774         if (ret)
1775                 return ret;
1776
1777         ret = do_pkvm_init(hyp_va_bits);
1778         if (ret)
1779                 return ret;
1780
1781         free_hyp_pgds();
1782
1783         return 0;
1784 }
1785
1786 /**
1787  * Inits Hyp-mode on all online CPUs
1788  */
1789 static int init_hyp_mode(void)
1790 {
1791         u32 hyp_va_bits;
1792         int cpu;
1793         int err = -ENOMEM;
1794
1795         /*
1796          * The protected Hyp-mode cannot be initialized if the memory pool
1797          * allocation has failed.
1798          */
1799         if (is_protected_kvm_enabled() && !hyp_mem_base)
1800                 goto out_err;
1801
1802         /*
1803          * Allocate Hyp PGD and setup Hyp identity mapping
1804          */
1805         err = kvm_mmu_init(&hyp_va_bits);
1806         if (err)
1807                 goto out_err;
1808
1809         /*
1810          * Allocate stack pages for Hypervisor-mode
1811          */
1812         for_each_possible_cpu(cpu) {
1813                 unsigned long stack_page;
1814
1815                 stack_page = __get_free_page(GFP_KERNEL);
1816                 if (!stack_page) {
1817                         err = -ENOMEM;
1818                         goto out_err;
1819                 }
1820
1821                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
1822         }
1823
1824         /*
1825          * Allocate and initialize pages for Hypervisor-mode percpu regions.
1826          */
1827         for_each_possible_cpu(cpu) {
1828                 struct page *page;
1829                 void *page_addr;
1830
1831                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
1832                 if (!page) {
1833                         err = -ENOMEM;
1834                         goto out_err;
1835                 }
1836
1837                 page_addr = page_address(page);
1838                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
1839                 kvm_arm_hyp_percpu_base[cpu] = (unsigned long)page_addr;
1840         }
1841
1842         /*
1843          * Map the Hyp-code called directly from the host
1844          */
1845         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
1846                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
1847         if (err) {
1848                 kvm_err("Cannot map world-switch code\n");
1849                 goto out_err;
1850         }
1851
1852         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
1853                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
1854         if (err) {
1855                 kvm_err("Cannot map .hyp.rodata section\n");
1856                 goto out_err;
1857         }
1858
1859         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
1860                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
1861         if (err) {
1862                 kvm_err("Cannot map rodata section\n");
1863                 goto out_err;
1864         }
1865
1866         /*
1867          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
1868          * section thanks to an assertion in the linker script. Map it RW and
1869          * the rest of .bss RO.
1870          */
1871         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
1872                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
1873         if (err) {
1874                 kvm_err("Cannot map hyp bss section: %d\n", err);
1875                 goto out_err;
1876         }
1877
1878         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
1879                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
1880         if (err) {
1881                 kvm_err("Cannot map bss section\n");
1882                 goto out_err;
1883         }
1884
1885         /*
1886          * Map the Hyp stack pages
1887          */
1888         for_each_possible_cpu(cpu) {
1889                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
1890                 err = create_hyp_mappings(stack_page, stack_page + PAGE_SIZE,
1891                                           PAGE_HYP);
1892
1893                 if (err) {
1894                         kvm_err("Cannot map hyp stack\n");
1895                         goto out_err;
1896                 }
1897         }
1898
1899         for_each_possible_cpu(cpu) {
1900                 char *percpu_begin = (char *)kvm_arm_hyp_percpu_base[cpu];
1901                 char *percpu_end = percpu_begin + nvhe_percpu_size();
1902
1903                 /* Map Hyp percpu pages */
1904                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
1905                 if (err) {
1906                         kvm_err("Cannot map hyp percpu region\n");
1907                         goto out_err;
1908                 }
1909
1910                 /* Prepare the CPU initialization parameters */
1911                 cpu_prepare_hyp_mode(cpu);
1912         }
1913
1914         if (is_protected_kvm_enabled()) {
1915                 init_cpu_logical_map();
1916
1917                 if (!init_psci_relay()) {
1918                         err = -ENODEV;
1919                         goto out_err;
1920                 }
1921         }
1922
1923         if (is_protected_kvm_enabled()) {
1924                 err = kvm_hyp_init_protection(hyp_va_bits);
1925                 if (err) {
1926                         kvm_err("Failed to init hyp memory protection\n");
1927                         goto out_err;
1928                 }
1929         }
1930
1931         return 0;
1932
1933 out_err:
1934         teardown_hyp_mode();
1935         kvm_err("error initializing Hyp mode: %d\n", err);
1936         return err;
1937 }
1938
1939 static void _kvm_host_prot_finalize(void *discard)
1940 {
1941         WARN_ON(kvm_call_hyp_nvhe(__pkvm_prot_finalize));
1942 }
1943
1944 static inline int pkvm_mark_hyp(phys_addr_t start, phys_addr_t end)
1945 {
1946         return kvm_call_hyp_nvhe(__pkvm_mark_hyp, start, end);
1947 }
1948
1949 #define pkvm_mark_hyp_section(__section)                \
1950         pkvm_mark_hyp(__pa_symbol(__section##_start),   \
1951                         __pa_symbol(__section##_end))
1952
1953 static int finalize_hyp_mode(void)
1954 {
1955         int cpu, ret;
1956
1957         if (!is_protected_kvm_enabled())
1958                 return 0;
1959
1960         ret = pkvm_mark_hyp_section(__hyp_idmap_text);
1961         if (ret)
1962                 return ret;
1963
1964         ret = pkvm_mark_hyp_section(__hyp_text);
1965         if (ret)
1966                 return ret;
1967
1968         ret = pkvm_mark_hyp_section(__hyp_rodata);
1969         if (ret)
1970                 return ret;
1971
1972         ret = pkvm_mark_hyp_section(__hyp_bss);
1973         if (ret)
1974                 return ret;
1975
1976         ret = pkvm_mark_hyp(hyp_mem_base, hyp_mem_base + hyp_mem_size);
1977         if (ret)
1978                 return ret;
1979
1980         for_each_possible_cpu(cpu) {
1981                 phys_addr_t start = virt_to_phys((void *)kvm_arm_hyp_percpu_base[cpu]);
1982                 phys_addr_t end = start + (PAGE_SIZE << nvhe_percpu_order());
1983
1984                 ret = pkvm_mark_hyp(start, end);
1985                 if (ret)
1986                         return ret;
1987
1988                 start = virt_to_phys((void *)per_cpu(kvm_arm_hyp_stack_page, cpu));
1989                 end = start + PAGE_SIZE;
1990                 ret = pkvm_mark_hyp(start, end);
1991                 if (ret)
1992                         return ret;
1993         }
1994
1995         /*
1996          * Flip the static key upfront as that may no longer be possible
1997          * once the host stage 2 is installed.
1998          */
1999         static_branch_enable(&kvm_protected_mode_initialized);
2000         on_each_cpu(_kvm_host_prot_finalize, NULL, 1);
2001
2002         return 0;
2003 }
2004
2005 static void check_kvm_target_cpu(void *ret)
2006 {
2007         *(int *)ret = kvm_target_cpu();
2008 }
2009
2010 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2011 {
2012         struct kvm_vcpu *vcpu;
2013         int i;
2014
2015         mpidr &= MPIDR_HWID_BITMASK;
2016         kvm_for_each_vcpu(i, vcpu, kvm) {
2017                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2018                         return vcpu;
2019         }
2020         return NULL;
2021 }
2022
2023 bool kvm_arch_has_irq_bypass(void)
2024 {
2025         return true;
2026 }
2027
2028 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2029                                       struct irq_bypass_producer *prod)
2030 {
2031         struct kvm_kernel_irqfd *irqfd =
2032                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2033
2034         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2035                                           &irqfd->irq_entry);
2036 }
2037 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2038                                       struct irq_bypass_producer *prod)
2039 {
2040         struct kvm_kernel_irqfd *irqfd =
2041                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2042
2043         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2044                                      &irqfd->irq_entry);
2045 }
2046
2047 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2048 {
2049         struct kvm_kernel_irqfd *irqfd =
2050                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2051
2052         kvm_arm_halt_guest(irqfd->kvm);
2053 }
2054
2055 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2056 {
2057         struct kvm_kernel_irqfd *irqfd =
2058                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2059
2060         kvm_arm_resume_guest(irqfd->kvm);
2061 }
2062
2063 /**
2064  * Initialize Hyp-mode and memory mappings on all CPUs.
2065  */
2066 int kvm_arch_init(void *opaque)
2067 {
2068         int err;
2069         int ret, cpu;
2070         bool in_hyp_mode;
2071
2072         if (!is_hyp_mode_available()) {
2073                 kvm_info("HYP mode not available\n");
2074                 return -ENODEV;
2075         }
2076
2077         in_hyp_mode = is_kernel_in_hyp_mode();
2078
2079         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2080             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2081                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2082                          "Only trusted guests should be used on this system.\n");
2083
2084         for_each_online_cpu(cpu) {
2085                 smp_call_function_single(cpu, check_kvm_target_cpu, &ret, 1);
2086                 if (ret < 0) {
2087                         kvm_err("Error, CPU %d not supported!\n", cpu);
2088                         return -ENODEV;
2089                 }
2090         }
2091
2092         err = init_common_resources();
2093         if (err)
2094                 return err;
2095
2096         err = kvm_arm_init_sve();
2097         if (err)
2098                 return err;
2099
2100         if (!in_hyp_mode) {
2101                 err = init_hyp_mode();
2102                 if (err)
2103                         goto out_err;
2104         }
2105
2106         err = kvm_init_vector_slots();
2107         if (err) {
2108                 kvm_err("Cannot initialise vector slots\n");
2109                 goto out_err;
2110         }
2111
2112         err = init_subsystems();
2113         if (err)
2114                 goto out_hyp;
2115
2116         if (!in_hyp_mode) {
2117                 err = finalize_hyp_mode();
2118                 if (err) {
2119                         kvm_err("Failed to finalize Hyp protection\n");
2120                         goto out_hyp;
2121                 }
2122         }
2123
2124         if (is_protected_kvm_enabled()) {
2125                 kvm_info("Protected nVHE mode initialized successfully\n");
2126         } else if (in_hyp_mode) {
2127                 kvm_info("VHE mode initialized successfully\n");
2128         } else {
2129                 kvm_info("Hyp mode initialized successfully\n");
2130         }
2131
2132         return 0;
2133
2134 out_hyp:
2135         hyp_cpu_pm_exit();
2136         if (!in_hyp_mode)
2137                 teardown_hyp_mode();
2138 out_err:
2139         return err;
2140 }
2141
2142 /* NOP: Compiling as a module not supported */
2143 void kvm_arch_exit(void)
2144 {
2145         kvm_perf_teardown();
2146 }
2147
2148 static int __init early_kvm_mode_cfg(char *arg)
2149 {
2150         if (!arg)
2151                 return -EINVAL;
2152
2153         if (strcmp(arg, "protected") == 0) {
2154                 kvm_mode = KVM_MODE_PROTECTED;
2155                 return 0;
2156         }
2157
2158         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode()))
2159                 return 0;
2160
2161         return -EINVAL;
2162 }
2163 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2164
2165 enum kvm_mode kvm_get_mode(void)
2166 {
2167         return kvm_mode;
2168 }
2169
2170 static int arm_init(void)
2171 {
2172         int rc = kvm_init(NULL, sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2173         return rc;
2174 }
2175
2176 module_init(arm_init);