arch/arm64/kvm/arm.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   4  * Author: Christoffer Dall <c.dall@virtualopensystems.com>
   5  */
   6
   7 #include <linux/bug.h>
   8 #include <linux/cpu_pm.h>
   9 #include <linux/errno.h>
  10 #include <linux/err.h>
  11 #include <linux/kvm_host.h>
  12 #include <linux/list.h>
  13 #include <linux/module.h>
  14 #include <linux/vmalloc.h>
  15 #include <linux/fs.h>
  16 #include <linux/mman.h>
  17 #include <linux/sched.h>
  18 #include <linux/kvm.h>
  19 #include <linux/kvm_irqfd.h>
  20 #include <linux/irqbypass.h>
  21 #include <linux/sched/stat.h>
  22 #include <linux/psci.h>
  23 #include <trace/events/kvm.h>
  24
  25 #define CREATE_TRACE_POINTS
  26 #include "trace_arm.h"
  27
  28 #include <linux/uaccess.h>
  29 #include <asm/ptrace.h>
  30 #include <asm/mman.h>
  31 #include <asm/tlbflush.h>
  32 #include <asm/cacheflush.h>
  33 #include <asm/cpufeature.h>
  34 #include <asm/virt.h>
  35 #include <asm/kvm_arm.h>
  36 #include <asm/kvm_asm.h>
  37 #include <asm/kvm_mmu.h>
  38 #include <asm/kvm_emulate.h>
  39 #include <asm/sections.h>
  40
  41 #include <kvm/arm_hypercalls.h>
  42 #include <kvm/arm_pmu.h>
  43 #include <kvm/arm_psci.h>
  44
  45 #ifdef REQUIRES_VIRT
  46 __asm__(".arch_extension        virt");
  47 #endif
  48
  49 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
  50 DEFINE_STATIC_KEY_FALSE(kvm_protected_mode_initialized);
  51
  52 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
  53
  54 static DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  55 unsigned long kvm_arm_hyp_percpu_base[NR_CPUS];
  56 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
  57
  58 /* The VMID used in the VTTBR */
  59 static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1);
  60 static u32 kvm_next_vmid;
  61 static DEFINE_SPINLOCK(kvm_vmid_lock);
  62
  63 static bool vgic_present;
  64
  65 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled);
  66 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  67
  68 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  69 {
  70         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  71 }
  72
  73 int kvm_arch_hardware_setup(void *opaque)
  74 {
  75         return 0;
  76 }
  77
  78 int kvm_arch_check_processor_compat(void *opaque)
  79 {
  80         return 0;
  81 }
  82
  83 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  84                             struct kvm_enable_cap *cap)
  85 {
  86         int r;
  87
  88         if (cap->flags)
  89                 return -EINVAL;
  90
  91         switch (cap->cap) {
  92         case KVM_CAP_ARM_NISV_TO_USER:
  93                 r = 0;
  94                 kvm->arch.return_nisv_io_abort_to_user = true;
  95                 break;
  96         default:
  97                 r = -EINVAL;
  98                 break;
  99         }
 100
 101         return r;
 102 }
 103
 104 static int kvm_arm_default_max_vcpus(void)
 105 {
 106         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 107 }
 108
 109 static void set_default_spectre(struct kvm *kvm)
 110 {
 111         /*
 112          * The default is to expose CSV2 == 1 if the HW isn't affected.
 113          * Although this is a per-CPU feature, we make it global because
 114          * asymmetric systems are just a nuisance.
 115          *
 116          * Userspace can override this as long as it doesn't promise
 117          * the impossible.
 118          */
 119         if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED)
 120                 kvm->arch.pfr0_csv2 = 1;
 121         if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED)
 122                 kvm->arch.pfr0_csv3 = 1;
 123 }
 124
 125 /**
 126  * kvm_arch_init_vm - initializes a VM data structure
 127  * @kvm:        pointer to the KVM struct
 128  */
 129 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 130 {
 131         int ret;
 132
 133         ret = kvm_arm_setup_stage2(kvm, type);
 134         if (ret)
 135                 return ret;
 136
 137         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu);
 138         if (ret)
 139                 return ret;
 140
 141         ret = create_hyp_mappings(kvm, kvm + 1, PAGE_HYP);
 142         if (ret)
 143                 goto out_free_stage2_pgd;
 144
 145         kvm_vgic_early_init(kvm);
 146
 147         /* The maximum number of VCPUs is limited by the host's GIC model */
 148         kvm->arch.max_vcpus = kvm_arm_default_max_vcpus();
 149
 150         set_default_spectre(kvm);
 151
 152         return ret;
 153 out_free_stage2_pgd:
 154         kvm_free_stage2_pgd(&kvm->arch.mmu);
 155         return ret;
 156 }
 157
 158 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 159 {
 160         return VM_FAULT_SIGBUS;
 161 }
 162
 163
 164 /**
 165  * kvm_arch_destroy_vm - destroy the VM data structure
 166  * @kvm:        pointer to the KVM struct
 167  */
 168 void kvm_arch_destroy_vm(struct kvm *kvm)
 169 {
 170         int i;
 171
 172         bitmap_free(kvm->arch.pmu_filter);
 173
 174         kvm_vgic_destroy(kvm);
 175
 176         for (i = 0; i < KVM_MAX_VCPUS; ++i) {
 177                 if (kvm->vcpus[i]) {
 178                         kvm_vcpu_destroy(kvm->vcpus[i]);
 179                         kvm->vcpus[i] = NULL;
 180                 }
 181         }
 182         atomic_set(&kvm->online_vcpus, 0);
 183 }
 184
 185 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 186 {
 187         int r;
 188         switch (ext) {
 189         case KVM_CAP_IRQCHIP:
 190                 r = vgic_present;
 191                 break;
 192         case KVM_CAP_IOEVENTFD:
 193         case KVM_CAP_DEVICE_CTRL:
 194         case KVM_CAP_USER_MEMORY:
 195         case KVM_CAP_SYNC_MMU:
 196         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 197         case KVM_CAP_ONE_REG:
 198         case KVM_CAP_ARM_PSCI:
 199         case KVM_CAP_ARM_PSCI_0_2:
 200         case KVM_CAP_READONLY_MEM:
 201         case KVM_CAP_MP_STATE:
 202         case KVM_CAP_IMMEDIATE_EXIT:
 203         case KVM_CAP_VCPU_EVENTS:
 204         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 205         case KVM_CAP_ARM_NISV_TO_USER:
 206         case KVM_CAP_ARM_INJECT_EXT_DABT:
 207         case KVM_CAP_SET_GUEST_DEBUG:
 208         case KVM_CAP_VCPU_ATTRIBUTES:
 209         case KVM_CAP_PTP_KVM:
 210                 r = 1;
 211                 break;
 212         case KVM_CAP_SET_GUEST_DEBUG2:
 213                 return KVM_GUESTDBG_VALID_MASK;
 214         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 215                 r = 1;
 216                 break;
 217         case KVM_CAP_NR_VCPUS:
 218                 r = num_online_cpus();
 219                 break;
 220         case KVM_CAP_MAX_VCPUS:
 221         case KVM_CAP_MAX_VCPU_ID:
 222                 if (kvm)
 223                         r = kvm->arch.max_vcpus;
 224                 else
 225                         r = kvm_arm_default_max_vcpus();
 226                 break;
 227         case KVM_CAP_MSI_DEVID:
 228                 if (!kvm)
 229                         r = -EINVAL;
 230                 else
 231                         r = kvm->arch.vgic.msis_require_devid;
 232                 break;
 233         case KVM_CAP_ARM_USER_IRQ:
 234                 /*
 235                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 236                  * (bump this number if adding more devices)
 237                  */
 238                 r = 1;
 239                 break;
 240         case KVM_CAP_STEAL_TIME:
 241                 r = kvm_arm_pvtime_supported();
 242                 break;
 243         case KVM_CAP_ARM_EL1_32BIT:
 244                 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1);
 245                 break;
 246         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 247                 r = get_num_brps();
 248                 break;
 249         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 250                 r = get_num_wrps();
 251                 break;
 252         case KVM_CAP_ARM_PMU_V3:
 253                 r = kvm_arm_support_pmu_v3();
 254                 break;
 255         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 256                 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN);
 257                 break;
 258         case KVM_CAP_ARM_VM_IPA_SIZE:
 259                 r = get_kvm_ipa_limit();
 260                 break;
 261         case KVM_CAP_ARM_SVE:
 262                 r = system_supports_sve();
 263                 break;
 264         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 265         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 266                 r = system_has_full_ptr_auth();
 267                 break;
 268         default:
 269                 r = 0;
 270         }
 271
 272         return r;
 273 }
 274
 275 long kvm_arch_dev_ioctl(struct file *filp,
 276                         unsigned int ioctl, unsigned long arg)
 277 {
 278         return -EINVAL;
 279 }
 280
 281 struct kvm *kvm_arch_alloc_vm(void)
 282 {
 283         if (!has_vhe())
 284                 return kzalloc(sizeof(struct kvm), GFP_KERNEL);
 285
 286         return vzalloc(sizeof(struct kvm));
 287 }
 288
 289 void kvm_arch_free_vm(struct kvm *kvm)
 290 {
 291         if (!has_vhe())
 292                 kfree(kvm);
 293         else
 294                 vfree(kvm);
 295 }
 296
 297 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 298 {
 299         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 300                 return -EBUSY;
 301
 302         if (id >= kvm->arch.max_vcpus)
 303                 return -EINVAL;
 304
 305         return 0;
 306 }
 307
 308 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 309 {
 310         int err;
 311
 312         /* Force users to call KVM_ARM_VCPU_INIT */
 313         vcpu->arch.target = -1;
 314         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 315
 316         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 317
 318         /* Set up the timer */
 319         kvm_timer_vcpu_init(vcpu);
 320
 321         kvm_pmu_vcpu_init(vcpu);
 322
 323         kvm_arm_reset_debug_ptr(vcpu);
 324
 325         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 326
 327         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 328
 329         err = kvm_vgic_vcpu_init(vcpu);
 330         if (err)
 331                 return err;
 332
 333         return create_hyp_mappings(vcpu, vcpu + 1, PAGE_HYP);
 334 }
 335
 336 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 337 {
 338 }
 339
 340 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 341 {
 342         if (vcpu->arch.has_run_once && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 343                 static_branch_dec(&userspace_irqchip_in_use);
 344
 345         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 346         kvm_timer_vcpu_terminate(vcpu);
 347         kvm_pmu_vcpu_destroy(vcpu);
 348
 349         kvm_arm_vcpu_destroy(vcpu);
 350 }
 351
 352 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
 353 {
 354         return kvm_timer_is_pending(vcpu);
 355 }
 356
 357 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 358 {
 359         /*
 360          * If we're about to block (most likely because we've just hit a
 361          * WFI), we need to sync back the state of the GIC CPU interface
 362          * so that we have the latest PMR and group enables. This ensures
 363          * that kvm_arch_vcpu_runnable has up-to-date data to decide
 364          * whether we have pending interrupts.
 365          *
 366          * For the same reason, we want to tell GICv4 that we need
 367          * doorbells to be signalled, should an interrupt become pending.
 368          */
 369         preempt_disable();
 370         kvm_vgic_vmcr_sync(vcpu);
 371         vgic_v4_put(vcpu, true);
 372         preempt_enable();
 373 }
 374
 375 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 376 {
 377         preempt_disable();
 378         vgic_v4_load(vcpu);
 379         preempt_enable();
 380 }
 381
 382 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 383 {
 384         struct kvm_s2_mmu *mmu;
 385         int *last_ran;
 386
 387         mmu = vcpu->arch.hw_mmu;
 388         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 389
 390         /*
 391          * We guarantee that both TLBs and I-cache are private to each
 392          * vcpu. If detecting that a vcpu from the same VM has
 393          * previously run on the same physical CPU, call into the
 394          * hypervisor code to nuke the relevant contexts.
 395          *
 396          * We might get preempted before the vCPU actually runs, but
 397          * over-invalidation doesn't affect correctness.
 398          */
 399         if (*last_ran != vcpu->vcpu_id) {
 400                 kvm_call_hyp(__kvm_flush_cpu_context, mmu);
 401                 *last_ran = vcpu->vcpu_id;
 402         }
 403
 404         vcpu->cpu = cpu;
 405
 406         kvm_vgic_load(vcpu);
 407         kvm_timer_vcpu_load(vcpu);
 408         if (has_vhe())
 409                 kvm_vcpu_load_sysregs_vhe(vcpu);
 410         kvm_arch_vcpu_load_fp(vcpu);
 411         kvm_vcpu_pmu_restore_guest(vcpu);
 412         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 413                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 414
 415         if (single_task_running())
 416                 vcpu_clear_wfx_traps(vcpu);
 417         else
 418                 vcpu_set_wfx_traps(vcpu);
 419
 420         if (vcpu_has_ptrauth(vcpu))
 421                 vcpu_ptrauth_disable(vcpu);
 422         kvm_arch_vcpu_load_debug_state_flags(vcpu);
 423 }
 424
 425 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 426 {
 427         kvm_arch_vcpu_put_debug_state_flags(vcpu);
 428         kvm_arch_vcpu_put_fp(vcpu);
 429         if (has_vhe())
 430                 kvm_vcpu_put_sysregs_vhe(vcpu);
 431         kvm_timer_vcpu_put(vcpu);
 432         kvm_vgic_put(vcpu);
 433         kvm_vcpu_pmu_restore_host(vcpu);
 434
 435         vcpu->cpu = -1;
 436 }
 437
 438 static void vcpu_power_off(struct kvm_vcpu *vcpu)
 439 {
 440         vcpu->arch.power_off = true;
 441         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 442         kvm_vcpu_kick(vcpu);
 443 }
 444
 445 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 446                                     struct kvm_mp_state *mp_state)
 447 {
 448         if (vcpu->arch.power_off)
 449                 mp_state->mp_state = KVM_MP_STATE_STOPPED;
 450         else
 451                 mp_state->mp_state = KVM_MP_STATE_RUNNABLE;
 452
 453         return 0;
 454 }
 455
 456 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 457                                     struct kvm_mp_state *mp_state)
 458 {
 459         int ret = 0;
 460
 461         switch (mp_state->mp_state) {
 462         case KVM_MP_STATE_RUNNABLE:
 463                 vcpu->arch.power_off = false;
 464                 break;
 465         case KVM_MP_STATE_STOPPED:
 466                 vcpu_power_off(vcpu);
 467                 break;
 468         default:
 469                 ret = -EINVAL;
 470         }
 471
 472         return ret;
 473 }
 474
 475 /**
 476  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 477  * @v:          The VCPU pointer
 478  *
 479  * If the guest CPU is not waiting for interrupts or an interrupt line is
 480  * asserted, the CPU is by definition runnable.
 481  */
 482 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 483 {
 484         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 485         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 486                 && !v->arch.power_off && !v->arch.pause);
 487 }
 488
 489 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 490 {
 491         return vcpu_mode_priv(vcpu);
 492 }
 493
 494 /* Just ensure a guest exit from a particular CPU */
 495 static void exit_vm_noop(void *info)
 496 {
 497 }
 498
 499 void force_vm_exit(const cpumask_t *mask)
 500 {
 501         preempt_disable();
 502         smp_call_function_many(mask, exit_vm_noop, NULL, true);
 503         preempt_enable();
 504 }
 505
 506 /**
 507  * need_new_vmid_gen - check that the VMID is still valid
 508  * @vmid: The VMID to check
 509  *
 510  * return true if there is a new generation of VMIDs being used
 511  *
 512  * The hardware supports a limited set of values with the value zero reserved
 513  * for the host, so we check if an assigned value belongs to a previous
 514  * generation, which requires us to assign a new value. If we're the first to
 515  * use a VMID for the new generation, we must flush necessary caches and TLBs
 516  * on all CPUs.
 517  */
 518 static bool need_new_vmid_gen(struct kvm_vmid *vmid)
 519 {
 520         u64 current_vmid_gen = atomic64_read(&kvm_vmid_gen);
 521         smp_rmb(); /* Orders read of kvm_vmid_gen and kvm->arch.vmid */
 522         return unlikely(READ_ONCE(vmid->vmid_gen) != current_vmid_gen);
 523 }
 524
 525 /**
 526  * update_vmid - Update the vmid with a valid VMID for the current generation
 527  * @vmid: The stage-2 VMID information struct
 528  */
 529 static void update_vmid(struct kvm_vmid *vmid)
 530 {
 531         if (!need_new_vmid_gen(vmid))
 532                 return;
 533
 534         spin_lock(&kvm_vmid_lock);
 535
 536         /*
 537          * We need to re-check the vmid_gen here to ensure that if another vcpu
 538          * already allocated a valid vmid for this vm, then this vcpu should
 539          * use the same vmid.
 540          */
 541         if (!need_new_vmid_gen(vmid)) {
 542                 spin_unlock(&kvm_vmid_lock);
 543                 return;
 544         }
 545
 546         /* First user of a new VMID generation? */
 547         if (unlikely(kvm_next_vmid == 0)) {
 548                 atomic64_inc(&kvm_vmid_gen);
 549                 kvm_next_vmid = 1;
 550
 551                 /*
 552                  * On SMP we know no other CPUs can use this CPU's or each
 553                  * other's VMID after force_vm_exit returns since the
 554                  * kvm_vmid_lock blocks them from reentry to the guest.
 555                  */
 556                 force_vm_exit(cpu_all_mask);
 557                 /*
 558                  * Now broadcast TLB + ICACHE invalidation over the inner
 559                  * shareable domain to make sure all data structures are
 560                  * clean.
 561                  */
 562                 kvm_call_hyp(__kvm_flush_vm_context);
 563         }
 564
 565         vmid->vmid = kvm_next_vmid;
 566         kvm_next_vmid++;
 567         kvm_next_vmid &= (1 << kvm_get_vmid_bits()) - 1;
 568
 569         smp_wmb();
 570         WRITE_ONCE(vmid->vmid_gen, atomic64_read(&kvm_vmid_gen));
 571
 572         spin_unlock(&kvm_vmid_lock);
 573 }
 574
 575 static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 576 {
 577         struct kvm *kvm = vcpu->kvm;
 578         int ret = 0;
 579
 580         if (likely(vcpu->arch.has_run_once))
 581                 return 0;
 582
 583         if (!kvm_arm_vcpu_is_finalized(vcpu))
 584                 return -EPERM;
 585
 586         vcpu->arch.has_run_once = true;
 587
 588         kvm_arm_vcpu_init_debug(vcpu);
 589
 590         if (likely(irqchip_in_kernel(kvm))) {
 591                 /*
 592                  * Map the VGIC hardware resources before running a vcpu the
 593                  * first time on this VM.
 594                  */
 595                 ret = kvm_vgic_map_resources(kvm);
 596                 if (ret)
 597                         return ret;
 598         } else {
 599                 /*
 600                  * Tell the rest of the code that there are userspace irqchip
 601                  * VMs in the wild.
 602                  */
 603                 static_branch_inc(&userspace_irqchip_in_use);
 604         }
 605
 606         ret = kvm_timer_enable(vcpu);
 607         if (ret)
 608                 return ret;
 609
 610         ret = kvm_arm_pmu_v3_enable(vcpu);
 611
 612         return ret;
 613 }
 614
 615 bool kvm_arch_intc_initialized(struct kvm *kvm)
 616 {
 617         return vgic_initialized(kvm);
 618 }
 619
 620 void kvm_arm_halt_guest(struct kvm *kvm)
 621 {
 622         int i;
 623         struct kvm_vcpu *vcpu;
 624
 625         kvm_for_each_vcpu(i, vcpu, kvm)
 626                 vcpu->arch.pause = true;
 627         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 628 }
 629
 630 void kvm_arm_resume_guest(struct kvm *kvm)
 631 {
 632         int i;
 633         struct kvm_vcpu *vcpu;
 634
 635         kvm_for_each_vcpu(i, vcpu, kvm) {
 636                 vcpu->arch.pause = false;
 637                 rcuwait_wake_up(kvm_arch_vcpu_get_wait(vcpu));
 638         }
 639 }
 640
 641 static void vcpu_req_sleep(struct kvm_vcpu *vcpu)
 642 {
 643         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 644
 645         rcuwait_wait_event(wait,
 646                            (!vcpu->arch.power_off) &&(!vcpu->arch.pause),
 647                            TASK_INTERRUPTIBLE);
 648
 649         if (vcpu->arch.power_off || vcpu->arch.pause) {
 650                 /* Awaken to handle a signal, request we sleep again later. */
 651                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 652         }
 653
 654         /*
 655          * Make sure we will observe a potential reset request if we've
 656          * observed a change to the power state. Pairs with the smp_wmb() in
 657          * kvm_psci_vcpu_on().
 658          */
 659         smp_rmb();
 660 }
 661
 662 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 663 {
 664         return vcpu->arch.target >= 0;
 665 }
 666
 667 static void check_vcpu_requests(struct kvm_vcpu *vcpu)
 668 {
 669         if (kvm_request_pending(vcpu)) {
 670                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 671                         vcpu_req_sleep(vcpu);
 672
 673                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 674                         kvm_reset_vcpu(vcpu);
 675
 676                 /*
 677                  * Clear IRQ_PENDING requests that were made to guarantee
 678                  * that a VCPU sees new virtual interrupts.
 679                  */
 680                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 681
 682                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 683                         kvm_update_stolen_time(vcpu);
 684
 685                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 686                         /* The distributor enable bits were changed */
 687                         preempt_disable();
 688                         vgic_v4_put(vcpu, false);
 689                         vgic_v4_load(vcpu);
 690                         preempt_enable();
 691                 }
 692         }
 693 }
 694
 695 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 696 {
 697         if (likely(!vcpu_mode_is_32bit(vcpu)))
 698                 return false;
 699
 700         return !system_supports_32bit_el0() ||
 701                 static_branch_unlikely(&arm64_mismatched_32bit_el0);
 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 (vcpu_mode_is_bad_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                         icache_inval_all_pou();
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         default:
1363                 return -EINVAL;
1364         }
1365 }
1366
1367 static unsigned long nvhe_percpu_size(void)
1368 {
1369         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1370                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1371 }
1372
1373 static unsigned long nvhe_percpu_order(void)
1374 {
1375         unsigned long size = nvhe_percpu_size();
1376
1377         return size ? get_order(size) : 0;
1378 }
1379
1380 /* A lookup table holding the hypervisor VA for each vector slot */
1381 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1382
1383 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1384 {
1385         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1386 }
1387
1388 static int kvm_init_vector_slots(void)
1389 {
1390         int err;
1391         void *base;
1392
1393         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1394         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1395
1396         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1397         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1398
1399         if (!cpus_have_const_cap(ARM64_SPECTRE_V3A))
1400                 return 0;
1401
1402         if (!has_vhe()) {
1403                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1404                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1405                 if (err)
1406                         return err;
1407         }
1408
1409         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1410         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1411         return 0;
1412 }
1413
1414 static void cpu_prepare_hyp_mode(int cpu)
1415 {
1416         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1417         unsigned long tcr;
1418
1419         /*
1420          * Calculate the raw per-cpu offset without a translation from the
1421          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1422          * so that we can use adr_l to access per-cpu variables in EL2.
1423          * Also drop the KASAN tag which gets in the way...
1424          */
1425         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1426                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1427
1428         params->mair_el2 = read_sysreg(mair_el1);
1429
1430         /*
1431          * The ID map may be configured to use an extended virtual address
1432          * range. This is only the case if system RAM is out of range for the
1433          * currently configured page size and VA_BITS, in which case we will
1434          * also need the extended virtual range for the HYP ID map, or we won't
1435          * be able to enable the EL2 MMU.
1436          *
1437          * However, at EL2, there is only one TTBR register, and we can't switch
1438          * between translation tables *and* update TCR_EL2.T0SZ at the same
1439          * time. Bottom line: we need to use the extended range with *both* our
1440          * translation tables.
1441          *
1442          * So use the same T0SZ value we use for the ID map.
1443          */
1444         tcr = (read_sysreg(tcr_el1) & TCR_EL2_MASK) | TCR_EL2_RES1;
1445         tcr &= ~TCR_T0SZ_MASK;
1446         tcr |= (idmap_t0sz & GENMASK(TCR_TxSZ_WIDTH - 1, 0)) << TCR_T0SZ_OFFSET;
1447         params->tcr_el2 = tcr;
1448
1449         params->stack_hyp_va = kern_hyp_va(per_cpu(kvm_arm_hyp_stack_page, cpu) + PAGE_SIZE);
1450         params->pgd_pa = kvm_mmu_get_httbr();
1451         if (is_protected_kvm_enabled())
1452                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1453         else
1454                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1455         params->vttbr = params->vtcr = 0;
1456
1457         /*
1458          * Flush the init params from the data cache because the struct will
1459          * be read while the MMU is off.
1460          */
1461         kvm_flush_dcache_to_poc(params, sizeof(*params));
1462 }
1463
1464 static void hyp_install_host_vector(void)
1465 {
1466         struct kvm_nvhe_init_params *params;
1467         struct arm_smccc_res res;
1468
1469         /* Switch from the HYP stub to our own HYP init vector */
1470         __hyp_set_vectors(kvm_get_idmap_vector());
1471
1472         /*
1473          * Call initialization code, and switch to the full blown HYP code.
1474          * If the cpucaps haven't been finalized yet, something has gone very
1475          * wrong, and hyp will crash and burn when it uses any
1476          * cpus_have_const_cap() wrapper.
1477          */
1478         BUG_ON(!system_capabilities_finalized());
1479         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1480         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1481         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1482 }
1483
1484 static void cpu_init_hyp_mode(void)
1485 {
1486         hyp_install_host_vector();
1487
1488         /*
1489          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1490          * at EL2.
1491          */
1492         if (this_cpu_has_cap(ARM64_SSBS) &&
1493             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1494                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1495         }
1496 }
1497
1498 static void cpu_hyp_reset(void)
1499 {
1500         if (!is_kernel_in_hyp_mode())
1501                 __hyp_reset_vectors();
1502 }
1503
1504 /*
1505  * EL2 vectors can be mapped and rerouted in a number of ways,
1506  * depending on the kernel configuration and CPU present:
1507  *
1508  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1509  *   placed in one of the vector slots, which is executed before jumping
1510  *   to the real vectors.
1511  *
1512  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1513  *   containing the hardening sequence is mapped next to the idmap page,
1514  *   and executed before jumping to the real vectors.
1515  *
1516  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1517  *   empty slot is selected, mapped next to the idmap page, and
1518  *   executed before jumping to the real vectors.
1519  *
1520  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1521  * VHE, as we don't have hypervisor-specific mappings. If the system
1522  * is VHE and yet selects this capability, it will be ignored.
1523  */
1524 static void cpu_set_hyp_vector(void)
1525 {
1526         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1527         void *vector = hyp_spectre_vector_selector[data->slot];
1528
1529         if (!is_protected_kvm_enabled())
1530                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1531         else
1532                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1533 }
1534
1535 static void cpu_hyp_reinit(void)
1536 {
1537         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1538
1539         cpu_hyp_reset();
1540
1541         if (is_kernel_in_hyp_mode())
1542                 kvm_timer_init_vhe();
1543         else
1544                 cpu_init_hyp_mode();
1545
1546         cpu_set_hyp_vector();
1547
1548         kvm_arm_init_debug();
1549
1550         if (vgic_present)
1551                 kvm_vgic_init_cpu_hardware();
1552 }
1553
1554 static void _kvm_arch_hardware_enable(void *discard)
1555 {
1556         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1557                 cpu_hyp_reinit();
1558                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1559         }
1560 }
1561
1562 int kvm_arch_hardware_enable(void)
1563 {
1564         _kvm_arch_hardware_enable(NULL);
1565         return 0;
1566 }
1567
1568 static void _kvm_arch_hardware_disable(void *discard)
1569 {
1570         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1571                 cpu_hyp_reset();
1572                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1573         }
1574 }
1575
1576 void kvm_arch_hardware_disable(void)
1577 {
1578         if (!is_protected_kvm_enabled())
1579                 _kvm_arch_hardware_disable(NULL);
1580 }
1581
1582 #ifdef CONFIG_CPU_PM
1583 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1584                                     unsigned long cmd,
1585                                     void *v)
1586 {
1587         /*
1588          * kvm_arm_hardware_enabled is left with its old value over
1589          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1590          * re-enable hyp.
1591          */
1592         switch (cmd) {
1593         case CPU_PM_ENTER:
1594                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1595                         /*
1596                          * don't update kvm_arm_hardware_enabled here
1597                          * so that the hardware will be re-enabled
1598                          * when we resume. See below.
1599                          */
1600                         cpu_hyp_reset();
1601
1602                 return NOTIFY_OK;
1603         case CPU_PM_ENTER_FAILED:
1604         case CPU_PM_EXIT:
1605                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1606                         /* The hardware was enabled before suspend. */
1607                         cpu_hyp_reinit();
1608
1609                 return NOTIFY_OK;
1610
1611         default:
1612                 return NOTIFY_DONE;
1613         }
1614 }
1615
1616 static struct notifier_block hyp_init_cpu_pm_nb = {
1617         .notifier_call = hyp_init_cpu_pm_notifier,
1618 };
1619
1620 static void hyp_cpu_pm_init(void)
1621 {
1622         if (!is_protected_kvm_enabled())
1623                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1624 }
1625 static void hyp_cpu_pm_exit(void)
1626 {
1627         if (!is_protected_kvm_enabled())
1628                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1629 }
1630 #else
1631 static inline void hyp_cpu_pm_init(void)
1632 {
1633 }
1634 static inline void hyp_cpu_pm_exit(void)
1635 {
1636 }
1637 #endif
1638
1639 static void init_cpu_logical_map(void)
1640 {
1641         unsigned int cpu;
1642
1643         /*
1644          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
1645          * Only copy the set of online CPUs whose features have been chacked
1646          * against the finalized system capabilities. The hypervisor will not
1647          * allow any other CPUs from the `possible` set to boot.
1648          */
1649         for_each_online_cpu(cpu)
1650                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
1651 }
1652
1653 #define init_psci_0_1_impl_state(config, what)  \
1654         config.psci_0_1_ ## what ## _implemented = psci_ops.what
1655
1656 static bool init_psci_relay(void)
1657 {
1658         /*
1659          * If PSCI has not been initialized, protected KVM cannot install
1660          * itself on newly booted CPUs.
1661          */
1662         if (!psci_ops.get_version) {
1663                 kvm_err("Cannot initialize protected mode without PSCI\n");
1664                 return false;
1665         }
1666
1667         kvm_host_psci_config.version = psci_ops.get_version();
1668
1669         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
1670                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
1671                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
1672                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
1673                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
1674                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
1675         }
1676         return true;
1677 }
1678
1679 static int init_common_resources(void)
1680 {
1681         return kvm_set_ipa_limit();
1682 }
1683
1684 static int init_subsystems(void)
1685 {
1686         int err = 0;
1687
1688         /*
1689          * Enable hardware so that subsystem initialisation can access EL2.
1690          */
1691         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
1692
1693         /*
1694          * Register CPU lower-power notifier
1695          */
1696         hyp_cpu_pm_init();
1697
1698         /*
1699          * Init HYP view of VGIC
1700          */
1701         err = kvm_vgic_hyp_init();
1702         switch (err) {
1703         case 0:
1704                 vgic_present = true;
1705                 break;
1706         case -ENODEV:
1707         case -ENXIO:
1708                 vgic_present = false;
1709                 err = 0;
1710                 break;
1711         default:
1712                 goto out;
1713         }
1714
1715         /*
1716          * Init HYP architected timer support
1717          */
1718         err = kvm_timer_hyp_init(vgic_present);
1719         if (err)
1720                 goto out;
1721
1722         kvm_perf_init();
1723         kvm_sys_reg_table_init();
1724
1725 out:
1726         if (err || !is_protected_kvm_enabled())
1727                 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
1728
1729         return err;
1730 }
1731
1732 static void teardown_hyp_mode(void)
1733 {
1734         int cpu;
1735
1736         free_hyp_pgds();
1737         for_each_possible_cpu(cpu) {
1738                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
1739                 free_pages(kvm_arm_hyp_percpu_base[cpu], nvhe_percpu_order());
1740         }
1741 }
1742
1743 static int do_pkvm_init(u32 hyp_va_bits)
1744 {
1745         void *per_cpu_base = kvm_ksym_ref(kvm_arm_hyp_percpu_base);
1746         int ret;
1747
1748         preempt_disable();
1749         hyp_install_host_vector();
1750         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
1751                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
1752                                 hyp_va_bits);
1753         preempt_enable();
1754
1755         return ret;
1756 }
1757
1758 static int kvm_hyp_init_protection(u32 hyp_va_bits)
1759 {
1760         void *addr = phys_to_virt(hyp_mem_base);
1761         int ret;
1762
1763         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
1764         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
1765
1766         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
1767         if (ret)
1768                 return ret;
1769
1770         ret = do_pkvm_init(hyp_va_bits);
1771         if (ret)
1772                 return ret;
1773
1774         free_hyp_pgds();
1775
1776         return 0;
1777 }
1778
1779 /**
1780  * Inits Hyp-mode on all online CPUs
1781  */
1782 static int init_hyp_mode(void)
1783 {
1784         u32 hyp_va_bits;
1785         int cpu;
1786         int err = -ENOMEM;
1787
1788         /*
1789          * The protected Hyp-mode cannot be initialized if the memory pool
1790          * allocation has failed.
1791          */
1792         if (is_protected_kvm_enabled() && !hyp_mem_base)
1793                 goto out_err;
1794
1795         /*
1796          * Allocate Hyp PGD and setup Hyp identity mapping
1797          */
1798         err = kvm_mmu_init(&hyp_va_bits);
1799         if (err)
1800                 goto out_err;
1801
1802         /*
1803          * Allocate stack pages for Hypervisor-mode
1804          */
1805         for_each_possible_cpu(cpu) {
1806                 unsigned long stack_page;
1807
1808                 stack_page = __get_free_page(GFP_KERNEL);
1809                 if (!stack_page) {
1810                         err = -ENOMEM;
1811                         goto out_err;
1812                 }
1813
1814                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
1815         }
1816
1817         /*
1818          * Allocate and initialize pages for Hypervisor-mode percpu regions.
1819          */
1820         for_each_possible_cpu(cpu) {
1821                 struct page *page;
1822                 void *page_addr;
1823
1824                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
1825                 if (!page) {
1826                         err = -ENOMEM;
1827                         goto out_err;
1828                 }
1829
1830                 page_addr = page_address(page);
1831                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
1832                 kvm_arm_hyp_percpu_base[cpu] = (unsigned long)page_addr;
1833         }
1834
1835         /*
1836          * Map the Hyp-code called directly from the host
1837          */
1838         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
1839                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
1840         if (err) {
1841                 kvm_err("Cannot map world-switch code\n");
1842                 goto out_err;
1843         }
1844
1845         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
1846                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
1847         if (err) {
1848                 kvm_err("Cannot map .hyp.rodata section\n");
1849                 goto out_err;
1850         }
1851
1852         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
1853                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
1854         if (err) {
1855                 kvm_err("Cannot map rodata section\n");
1856                 goto out_err;
1857         }
1858
1859         /*
1860          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
1861          * section thanks to an assertion in the linker script. Map it RW and
1862          * the rest of .bss RO.
1863          */
1864         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
1865                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
1866         if (err) {
1867                 kvm_err("Cannot map hyp bss section: %d\n", err);
1868                 goto out_err;
1869         }
1870
1871         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
1872                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
1873         if (err) {
1874                 kvm_err("Cannot map bss section\n");
1875                 goto out_err;
1876         }
1877
1878         /*
1879          * Map the Hyp stack pages
1880          */
1881         for_each_possible_cpu(cpu) {
1882                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
1883                 err = create_hyp_mappings(stack_page, stack_page + PAGE_SIZE,
1884                                           PAGE_HYP);
1885
1886                 if (err) {
1887                         kvm_err("Cannot map hyp stack\n");
1888                         goto out_err;
1889                 }
1890         }
1891
1892         for_each_possible_cpu(cpu) {
1893                 char *percpu_begin = (char *)kvm_arm_hyp_percpu_base[cpu];
1894                 char *percpu_end = percpu_begin + nvhe_percpu_size();
1895
1896                 /* Map Hyp percpu pages */
1897                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
1898                 if (err) {
1899                         kvm_err("Cannot map hyp percpu region\n");
1900                         goto out_err;
1901                 }
1902
1903                 /* Prepare the CPU initialization parameters */
1904                 cpu_prepare_hyp_mode(cpu);
1905         }
1906
1907         if (is_protected_kvm_enabled()) {
1908                 init_cpu_logical_map();
1909
1910                 if (!init_psci_relay()) {
1911                         err = -ENODEV;
1912                         goto out_err;
1913                 }
1914         }
1915
1916         if (is_protected_kvm_enabled()) {
1917                 err = kvm_hyp_init_protection(hyp_va_bits);
1918                 if (err) {
1919                         kvm_err("Failed to init hyp memory protection\n");
1920                         goto out_err;
1921                 }
1922         }
1923
1924         return 0;
1925
1926 out_err:
1927         teardown_hyp_mode();
1928         kvm_err("error initializing Hyp mode: %d\n", err);
1929         return err;
1930 }
1931
1932 static void _kvm_host_prot_finalize(void *discard)
1933 {
1934         WARN_ON(kvm_call_hyp_nvhe(__pkvm_prot_finalize));
1935 }
1936
1937 static inline int pkvm_mark_hyp(phys_addr_t start, phys_addr_t end)
1938 {
1939         return kvm_call_hyp_nvhe(__pkvm_mark_hyp, start, end);
1940 }
1941
1942 #define pkvm_mark_hyp_section(__section)                \
1943         pkvm_mark_hyp(__pa_symbol(__section##_start),   \
1944                         __pa_symbol(__section##_end))
1945
1946 static int finalize_hyp_mode(void)
1947 {
1948         int cpu, ret;
1949
1950         if (!is_protected_kvm_enabled())
1951                 return 0;
1952
1953         ret = pkvm_mark_hyp_section(__hyp_idmap_text);
1954         if (ret)
1955                 return ret;
1956
1957         ret = pkvm_mark_hyp_section(__hyp_text);
1958         if (ret)
1959                 return ret;
1960
1961         ret = pkvm_mark_hyp_section(__hyp_rodata);
1962         if (ret)
1963                 return ret;
1964
1965         ret = pkvm_mark_hyp_section(__hyp_bss);
1966         if (ret)
1967                 return ret;
1968
1969         ret = pkvm_mark_hyp(hyp_mem_base, hyp_mem_base + hyp_mem_size);
1970         if (ret)
1971                 return ret;
1972
1973         for_each_possible_cpu(cpu) {
1974                 phys_addr_t start = virt_to_phys((void *)kvm_arm_hyp_percpu_base[cpu]);
1975                 phys_addr_t end = start + (PAGE_SIZE << nvhe_percpu_order());
1976
1977                 ret = pkvm_mark_hyp(start, end);
1978                 if (ret)
1979                         return ret;
1980
1981                 start = virt_to_phys((void *)per_cpu(kvm_arm_hyp_stack_page, cpu));
1982                 end = start + PAGE_SIZE;
1983                 ret = pkvm_mark_hyp(start, end);
1984                 if (ret)
1985                         return ret;
1986         }
1987
1988         /*
1989          * Flip the static key upfront as that may no longer be possible
1990          * once the host stage 2 is installed.
1991          */
1992         static_branch_enable(&kvm_protected_mode_initialized);
1993         on_each_cpu(_kvm_host_prot_finalize, NULL, 1);
1994
1995         return 0;
1996 }
1997
1998 static void check_kvm_target_cpu(void *ret)
1999 {
2000         *(int *)ret = kvm_target_cpu();
2001 }
2002
2003 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2004 {
2005         struct kvm_vcpu *vcpu;
2006         int i;
2007
2008         mpidr &= MPIDR_HWID_BITMASK;
2009         kvm_for_each_vcpu(i, vcpu, kvm) {
2010                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2011                         return vcpu;
2012         }
2013         return NULL;
2014 }
2015
2016 bool kvm_arch_has_irq_bypass(void)
2017 {
2018         return true;
2019 }
2020
2021 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2022                                       struct irq_bypass_producer *prod)
2023 {
2024         struct kvm_kernel_irqfd *irqfd =
2025                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2026
2027         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2028                                           &irqfd->irq_entry);
2029 }
2030 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2031                                       struct irq_bypass_producer *prod)
2032 {
2033         struct kvm_kernel_irqfd *irqfd =
2034                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2035
2036         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2037                                      &irqfd->irq_entry);
2038 }
2039
2040 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2041 {
2042         struct kvm_kernel_irqfd *irqfd =
2043                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2044
2045         kvm_arm_halt_guest(irqfd->kvm);
2046 }
2047
2048 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2049 {
2050         struct kvm_kernel_irqfd *irqfd =
2051                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2052
2053         kvm_arm_resume_guest(irqfd->kvm);
2054 }
2055
2056 /**
2057  * Initialize Hyp-mode and memory mappings on all CPUs.
2058  */
2059 int kvm_arch_init(void *opaque)
2060 {
2061         int err;
2062         int ret, cpu;
2063         bool in_hyp_mode;
2064
2065         if (!is_hyp_mode_available()) {
2066                 kvm_info("HYP mode not available\n");
2067                 return -ENODEV;
2068         }
2069
2070         in_hyp_mode = is_kernel_in_hyp_mode();
2071
2072         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2073             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2074                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2075                          "Only trusted guests should be used on this system.\n");
2076
2077         for_each_online_cpu(cpu) {
2078                 smp_call_function_single(cpu, check_kvm_target_cpu, &ret, 1);
2079                 if (ret < 0) {
2080                         kvm_err("Error, CPU %d not supported!\n", cpu);
2081                         return -ENODEV;
2082                 }
2083         }
2084
2085         err = init_common_resources();
2086         if (err)
2087                 return err;
2088
2089         err = kvm_arm_init_sve();
2090         if (err)
2091                 return err;
2092
2093         if (!in_hyp_mode) {
2094                 err = init_hyp_mode();
2095                 if (err)
2096                         goto out_err;
2097         }
2098
2099         err = kvm_init_vector_slots();
2100         if (err) {
2101                 kvm_err("Cannot initialise vector slots\n");
2102                 goto out_err;
2103         }
2104
2105         err = init_subsystems();
2106         if (err)
2107                 goto out_hyp;
2108
2109         if (!in_hyp_mode) {
2110                 err = finalize_hyp_mode();
2111                 if (err) {
2112                         kvm_err("Failed to finalize Hyp protection\n");
2113                         goto out_hyp;
2114                 }
2115         }
2116
2117         if (is_protected_kvm_enabled()) {
2118                 kvm_info("Protected nVHE mode initialized successfully\n");
2119         } else if (in_hyp_mode) {
2120                 kvm_info("VHE mode initialized successfully\n");
2121         } else {
2122                 kvm_info("Hyp mode initialized successfully\n");
2123         }
2124
2125         return 0;
2126
2127 out_hyp:
2128         hyp_cpu_pm_exit();
2129         if (!in_hyp_mode)
2130                 teardown_hyp_mode();
2131 out_err:
2132         return err;
2133 }
2134
2135 /* NOP: Compiling as a module not supported */
2136 void kvm_arch_exit(void)
2137 {
2138         kvm_perf_teardown();
2139 }
2140
2141 static int __init early_kvm_mode_cfg(char *arg)
2142 {
2143         if (!arg)
2144                 return -EINVAL;
2145
2146         if (strcmp(arg, "protected") == 0) {
2147                 kvm_mode = KVM_MODE_PROTECTED;
2148                 return 0;
2149         }
2150
2151         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode()))
2152                 return 0;
2153
2154         return -EINVAL;
2155 }
2156 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2157
2158 enum kvm_mode kvm_get_mode(void)
2159 {
2160         return kvm_mode;
2161 }
2162
2163 static int arm_init(void)
2164 {
2165         int rc = kvm_init(NULL, sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2166         return rc;
2167 }
2168
2169 module_init(arm_init);