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/entry-kvm.h>
  10 #include <linux/errno.h>
  11 #include <linux/err.h>
  12 #include <linux/kvm_host.h>
  13 #include <linux/list.h>
  14 #include <linux/module.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mman.h>
  18 #include <linux/sched.h>
  19 #include <linux/kvm.h>
  20 #include <linux/kvm_irqfd.h>
  21 #include <linux/irqbypass.h>
  22 #include <linux/sched/stat.h>
  23 #include <linux/psci.h>
  24 #include <trace/events/kvm.h>
  25
  26 #define CREATE_TRACE_POINTS
  27 #include "trace_arm.h"
  28
  29 #include <linux/uaccess.h>
  30 #include <asm/ptrace.h>
  31 #include <asm/mman.h>
  32 #include <asm/tlbflush.h>
  33 #include <asm/cacheflush.h>
  34 #include <asm/cpufeature.h>
  35 #include <asm/virt.h>
  36 #include <asm/kvm_arm.h>
  37 #include <asm/kvm_asm.h>
  38 #include <asm/kvm_mmu.h>
  39 #include <asm/kvm_pkvm.h>
  40 #include <asm/kvm_emulate.h>
  41 #include <asm/sections.h>
  42
  43 #include <kvm/arm_hypercalls.h>
  44 #include <kvm/arm_pmu.h>
  45 #include <kvm/arm_psci.h>
  46
  47 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
  48
  49 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
  50
  51 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  52 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
  53
  54 static bool vgic_present;
  55
  56 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled);
  57 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  58
  59 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  60 {
  61         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  62 }
  63
  64 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  65                             struct kvm_enable_cap *cap)
  66 {
  67         int r;
  68
  69         if (cap->flags)
  70                 return -EINVAL;
  71
  72         switch (cap->cap) {
  73         case KVM_CAP_ARM_NISV_TO_USER:
  74                 r = 0;
  75                 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER,
  76                         &kvm->arch.flags);
  77                 break;
  78         case KVM_CAP_ARM_MTE:
  79                 mutex_lock(&kvm->lock);
  80                 if (!system_supports_mte() || kvm->created_vcpus) {
  81                         r = -EINVAL;
  82                 } else {
  83                         r = 0;
  84                         set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags);
  85                 }
  86                 mutex_unlock(&kvm->lock);
  87                 break;
  88         case KVM_CAP_ARM_SYSTEM_SUSPEND:
  89                 r = 0;
  90                 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags);
  91                 break;
  92         default:
  93                 r = -EINVAL;
  94                 break;
  95         }
  96
  97         return r;
  98 }
  99
 100 static int kvm_arm_default_max_vcpus(void)
 101 {
 102         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 103 }
 104
 105 static void set_default_spectre(struct kvm *kvm)
 106 {
 107         /*
 108          * The default is to expose CSV2 == 1 if the HW isn't affected.
 109          * Although this is a per-CPU feature, we make it global because
 110          * asymmetric systems are just a nuisance.
 111          *
 112          * Userspace can override this as long as it doesn't promise
 113          * the impossible.
 114          */
 115         if (arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED)
 116                 kvm->arch.pfr0_csv2 = 1;
 117         if (arm64_get_meltdown_state() == SPECTRE_UNAFFECTED)
 118                 kvm->arch.pfr0_csv3 = 1;
 119 }
 120
 121 /**
 122  * kvm_arch_init_vm - initializes a VM data structure
 123  * @kvm:        pointer to the KVM struct
 124  */
 125 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 126 {
 127         int ret;
 128
 129         mutex_init(&kvm->arch.config_lock);
 130
 131 #ifdef CONFIG_LOCKDEP
 132         /* Clue in lockdep that the config_lock must be taken inside kvm->lock */
 133         mutex_lock(&kvm->lock);
 134         mutex_lock(&kvm->arch.config_lock);
 135         mutex_unlock(&kvm->arch.config_lock);
 136         mutex_unlock(&kvm->lock);
 137 #endif
 138
 139         ret = kvm_share_hyp(kvm, kvm + 1);
 140         if (ret)
 141                 return ret;
 142
 143         ret = pkvm_init_host_vm(kvm);
 144         if (ret)
 145                 goto err_unshare_kvm;
 146
 147         if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) {
 148                 ret = -ENOMEM;
 149                 goto err_unshare_kvm;
 150         }
 151         cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask);
 152
 153         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type);
 154         if (ret)
 155                 goto err_free_cpumask;
 156
 157         kvm_vgic_early_init(kvm);
 158
 159         kvm_timer_init_vm(kvm);
 160
 161         /* The maximum number of VCPUs is limited by the host's GIC model */
 162         kvm->max_vcpus = kvm_arm_default_max_vcpus();
 163
 164         set_default_spectre(kvm);
 165         kvm_arm_init_hypercalls(kvm);
 166
 167         /*
 168          * Initialise the default PMUver before there is a chance to
 169          * create an actual PMU.
 170          */
 171         kvm->arch.dfr0_pmuver.imp = kvm_arm_pmu_get_pmuver_limit();
 172
 173         return 0;
 174
 175 err_free_cpumask:
 176         free_cpumask_var(kvm->arch.supported_cpus);
 177 err_unshare_kvm:
 178         kvm_unshare_hyp(kvm, kvm + 1);
 179         return ret;
 180 }
 181
 182 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 183 {
 184         return VM_FAULT_SIGBUS;
 185 }
 186
 187
 188 /**
 189  * kvm_arch_destroy_vm - destroy the VM data structure
 190  * @kvm:        pointer to the KVM struct
 191  */
 192 void kvm_arch_destroy_vm(struct kvm *kvm)
 193 {
 194         bitmap_free(kvm->arch.pmu_filter);
 195         free_cpumask_var(kvm->arch.supported_cpus);
 196
 197         kvm_vgic_destroy(kvm);
 198
 199         if (is_protected_kvm_enabled())
 200                 pkvm_destroy_hyp_vm(kvm);
 201
 202         kvm_destroy_vcpus(kvm);
 203
 204         kvm_unshare_hyp(kvm, kvm + 1);
 205
 206         kvm_arm_teardown_hypercalls(kvm);
 207 }
 208
 209 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 210 {
 211         int r;
 212         switch (ext) {
 213         case KVM_CAP_IRQCHIP:
 214                 r = vgic_present;
 215                 break;
 216         case KVM_CAP_IOEVENTFD:
 217         case KVM_CAP_DEVICE_CTRL:
 218         case KVM_CAP_USER_MEMORY:
 219         case KVM_CAP_SYNC_MMU:
 220         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 221         case KVM_CAP_ONE_REG:
 222         case KVM_CAP_ARM_PSCI:
 223         case KVM_CAP_ARM_PSCI_0_2:
 224         case KVM_CAP_READONLY_MEM:
 225         case KVM_CAP_MP_STATE:
 226         case KVM_CAP_IMMEDIATE_EXIT:
 227         case KVM_CAP_VCPU_EVENTS:
 228         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 229         case KVM_CAP_ARM_NISV_TO_USER:
 230         case KVM_CAP_ARM_INJECT_EXT_DABT:
 231         case KVM_CAP_SET_GUEST_DEBUG:
 232         case KVM_CAP_VCPU_ATTRIBUTES:
 233         case KVM_CAP_PTP_KVM:
 234         case KVM_CAP_ARM_SYSTEM_SUSPEND:
 235         case KVM_CAP_IRQFD_RESAMPLE:
 236         case KVM_CAP_COUNTER_OFFSET:
 237                 r = 1;
 238                 break;
 239         case KVM_CAP_SET_GUEST_DEBUG2:
 240                 return KVM_GUESTDBG_VALID_MASK;
 241         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 242                 r = 1;
 243                 break;
 244         case KVM_CAP_NR_VCPUS:
 245                 /*
 246                  * ARM64 treats KVM_CAP_NR_CPUS differently from all other
 247                  * architectures, as it does not always bound it to
 248                  * KVM_CAP_MAX_VCPUS. It should not matter much because
 249                  * this is just an advisory value.
 250                  */
 251                 r = min_t(unsigned int, num_online_cpus(),
 252                           kvm_arm_default_max_vcpus());
 253                 break;
 254         case KVM_CAP_MAX_VCPUS:
 255         case KVM_CAP_MAX_VCPU_ID:
 256                 if (kvm)
 257                         r = kvm->max_vcpus;
 258                 else
 259                         r = kvm_arm_default_max_vcpus();
 260                 break;
 261         case KVM_CAP_MSI_DEVID:
 262                 if (!kvm)
 263                         r = -EINVAL;
 264                 else
 265                         r = kvm->arch.vgic.msis_require_devid;
 266                 break;
 267         case KVM_CAP_ARM_USER_IRQ:
 268                 /*
 269                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 270                  * (bump this number if adding more devices)
 271                  */
 272                 r = 1;
 273                 break;
 274         case KVM_CAP_ARM_MTE:
 275                 r = system_supports_mte();
 276                 break;
 277         case KVM_CAP_STEAL_TIME:
 278                 r = kvm_arm_pvtime_supported();
 279                 break;
 280         case KVM_CAP_ARM_EL1_32BIT:
 281                 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1);
 282                 break;
 283         case KVM_CAP_GUEST_DEBUG_HW_BPS:
 284                 r = get_num_brps();
 285                 break;
 286         case KVM_CAP_GUEST_DEBUG_HW_WPS:
 287                 r = get_num_wrps();
 288                 break;
 289         case KVM_CAP_ARM_PMU_V3:
 290                 r = kvm_arm_support_pmu_v3();
 291                 break;
 292         case KVM_CAP_ARM_INJECT_SERROR_ESR:
 293                 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN);
 294                 break;
 295         case KVM_CAP_ARM_VM_IPA_SIZE:
 296                 r = get_kvm_ipa_limit();
 297                 break;
 298         case KVM_CAP_ARM_SVE:
 299                 r = system_supports_sve();
 300                 break;
 301         case KVM_CAP_ARM_PTRAUTH_ADDRESS:
 302         case KVM_CAP_ARM_PTRAUTH_GENERIC:
 303                 r = system_has_full_ptr_auth();
 304                 break;
 305         default:
 306                 r = 0;
 307         }
 308
 309         return r;
 310 }
 311
 312 long kvm_arch_dev_ioctl(struct file *filp,
 313                         unsigned int ioctl, unsigned long arg)
 314 {
 315         return -EINVAL;
 316 }
 317
 318 struct kvm *kvm_arch_alloc_vm(void)
 319 {
 320         size_t sz = sizeof(struct kvm);
 321
 322         if (!has_vhe())
 323                 return kzalloc(sz, GFP_KERNEL_ACCOUNT);
 324
 325         return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO);
 326 }
 327
 328 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 329 {
 330         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 331                 return -EBUSY;
 332
 333         if (id >= kvm->max_vcpus)
 334                 return -EINVAL;
 335
 336         return 0;
 337 }
 338
 339 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 340 {
 341         int err;
 342
 343         spin_lock_init(&vcpu->arch.mp_state_lock);
 344
 345 #ifdef CONFIG_LOCKDEP
 346         /* Inform lockdep that the config_lock is acquired after vcpu->mutex */
 347         mutex_lock(&vcpu->mutex);
 348         mutex_lock(&vcpu->kvm->arch.config_lock);
 349         mutex_unlock(&vcpu->kvm->arch.config_lock);
 350         mutex_unlock(&vcpu->mutex);
 351 #endif
 352
 353         /* Force users to call KVM_ARM_VCPU_INIT */
 354         vcpu->arch.target = -1;
 355         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 356
 357         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 358
 359         /*
 360          * Default value for the FP state, will be overloaded at load
 361          * time if we support FP (pretty likely)
 362          */
 363         vcpu->arch.fp_state = FP_STATE_FREE;
 364
 365         /* Set up the timer */
 366         kvm_timer_vcpu_init(vcpu);
 367
 368         kvm_pmu_vcpu_init(vcpu);
 369
 370         kvm_arm_reset_debug_ptr(vcpu);
 371
 372         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 373
 374         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 375
 376         err = kvm_vgic_vcpu_init(vcpu);
 377         if (err)
 378                 return err;
 379
 380         return kvm_share_hyp(vcpu, vcpu + 1);
 381 }
 382
 383 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 384 {
 385 }
 386
 387 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 388 {
 389         if (vcpu_has_run_once(vcpu) && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 390                 static_branch_dec(&userspace_irqchip_in_use);
 391
 392         kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
 393         kvm_timer_vcpu_terminate(vcpu);
 394         kvm_pmu_vcpu_destroy(vcpu);
 395
 396         kvm_arm_vcpu_destroy(vcpu);
 397 }
 398
 399 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 400 {
 401
 402 }
 403
 404 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 405 {
 406
 407 }
 408
 409 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 410 {
 411         struct kvm_s2_mmu *mmu;
 412         int *last_ran;
 413
 414         mmu = vcpu->arch.hw_mmu;
 415         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 416
 417         /*
 418          * We guarantee that both TLBs and I-cache are private to each
 419          * vcpu. If detecting that a vcpu from the same VM has
 420          * previously run on the same physical CPU, call into the
 421          * hypervisor code to nuke the relevant contexts.
 422          *
 423          * We might get preempted before the vCPU actually runs, but
 424          * over-invalidation doesn't affect correctness.
 425          */
 426         if (*last_ran != vcpu->vcpu_id) {
 427                 kvm_call_hyp(__kvm_flush_cpu_context, mmu);
 428                 *last_ran = vcpu->vcpu_id;
 429         }
 430
 431         vcpu->cpu = cpu;
 432
 433         kvm_vgic_load(vcpu);
 434         kvm_timer_vcpu_load(vcpu);
 435         if (has_vhe())
 436                 kvm_vcpu_load_sysregs_vhe(vcpu);
 437         kvm_arch_vcpu_load_fp(vcpu);
 438         kvm_vcpu_pmu_restore_guest(vcpu);
 439         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 440                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 441
 442         if (single_task_running())
 443                 vcpu_clear_wfx_traps(vcpu);
 444         else
 445                 vcpu_set_wfx_traps(vcpu);
 446
 447         if (vcpu_has_ptrauth(vcpu))
 448                 vcpu_ptrauth_disable(vcpu);
 449         kvm_arch_vcpu_load_debug_state_flags(vcpu);
 450
 451         if (!cpumask_test_cpu(smp_processor_id(), vcpu->kvm->arch.supported_cpus))
 452                 vcpu_set_on_unsupported_cpu(vcpu);
 453 }
 454
 455 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 456 {
 457         kvm_arch_vcpu_put_debug_state_flags(vcpu);
 458         kvm_arch_vcpu_put_fp(vcpu);
 459         if (has_vhe())
 460                 kvm_vcpu_put_sysregs_vhe(vcpu);
 461         kvm_timer_vcpu_put(vcpu);
 462         kvm_vgic_put(vcpu);
 463         kvm_vcpu_pmu_restore_host(vcpu);
 464         kvm_arm_vmid_clear_active();
 465
 466         vcpu_clear_on_unsupported_cpu(vcpu);
 467         vcpu->cpu = -1;
 468 }
 469
 470 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 471 {
 472         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED);
 473         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 474         kvm_vcpu_kick(vcpu);
 475 }
 476
 477 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
 478 {
 479         spin_lock(&vcpu->arch.mp_state_lock);
 480         __kvm_arm_vcpu_power_off(vcpu);
 481         spin_unlock(&vcpu->arch.mp_state_lock);
 482 }
 483
 484 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu)
 485 {
 486         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED;
 487 }
 488
 489 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu)
 490 {
 491         WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED);
 492         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 493         kvm_vcpu_kick(vcpu);
 494 }
 495
 496 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu)
 497 {
 498         return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED;
 499 }
 500
 501 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 502                                     struct kvm_mp_state *mp_state)
 503 {
 504         *mp_state = READ_ONCE(vcpu->arch.mp_state);
 505
 506         return 0;
 507 }
 508
 509 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 510                                     struct kvm_mp_state *mp_state)
 511 {
 512         int ret = 0;
 513
 514         spin_lock(&vcpu->arch.mp_state_lock);
 515
 516         switch (mp_state->mp_state) {
 517         case KVM_MP_STATE_RUNNABLE:
 518                 WRITE_ONCE(vcpu->arch.mp_state, *mp_state);
 519                 break;
 520         case KVM_MP_STATE_STOPPED:
 521                 __kvm_arm_vcpu_power_off(vcpu);
 522                 break;
 523         case KVM_MP_STATE_SUSPENDED:
 524                 kvm_arm_vcpu_suspend(vcpu);
 525                 break;
 526         default:
 527                 ret = -EINVAL;
 528         }
 529
 530         spin_unlock(&vcpu->arch.mp_state_lock);
 531
 532         return ret;
 533 }
 534
 535 /**
 536  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 537  * @v:          The VCPU pointer
 538  *
 539  * If the guest CPU is not waiting for interrupts or an interrupt line is
 540  * asserted, the CPU is by definition runnable.
 541  */
 542 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 543 {
 544         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 545         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 546                 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause);
 547 }
 548
 549 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 550 {
 551         return vcpu_mode_priv(vcpu);
 552 }
 553
 554 #ifdef CONFIG_GUEST_PERF_EVENTS
 555 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu)
 556 {
 557         return *vcpu_pc(vcpu);
 558 }
 559 #endif
 560
 561 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 562 {
 563         return vcpu->arch.target >= 0;
 564 }
 565
 566 /*
 567  * Handle both the initialisation that is being done when the vcpu is
 568  * run for the first time, as well as the updates that must be
 569  * performed each time we get a new thread dealing with this vcpu.
 570  */
 571 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
 572 {
 573         struct kvm *kvm = vcpu->kvm;
 574         int ret;
 575
 576         if (!kvm_vcpu_initialized(vcpu))
 577                 return -ENOEXEC;
 578
 579         if (!kvm_arm_vcpu_is_finalized(vcpu))
 580                 return -EPERM;
 581
 582         ret = kvm_arch_vcpu_run_map_fp(vcpu);
 583         if (ret)
 584                 return ret;
 585
 586         if (likely(vcpu_has_run_once(vcpu)))
 587                 return 0;
 588
 589         kvm_arm_vcpu_init_debug(vcpu);
 590
 591         if (likely(irqchip_in_kernel(kvm))) {
 592                 /*
 593                  * Map the VGIC hardware resources before running a vcpu the
 594                  * first time on this VM.
 595                  */
 596                 ret = kvm_vgic_map_resources(kvm);
 597                 if (ret)
 598                         return ret;
 599         }
 600
 601         ret = kvm_timer_enable(vcpu);
 602         if (ret)
 603                 return ret;
 604
 605         ret = kvm_arm_pmu_v3_enable(vcpu);
 606         if (ret)
 607                 return ret;
 608
 609         if (is_protected_kvm_enabled()) {
 610                 ret = pkvm_create_hyp_vm(kvm);
 611                 if (ret)
 612                         return ret;
 613         }
 614
 615         if (!irqchip_in_kernel(kvm)) {
 616                 /*
 617                  * Tell the rest of the code that there are userspace irqchip
 618                  * VMs in the wild.
 619                  */
 620                 static_branch_inc(&userspace_irqchip_in_use);
 621         }
 622
 623         /*
 624          * Initialize traps for protected VMs.
 625          * NOTE: Move to run in EL2 directly, rather than via a hypercall, once
 626          * the code is in place for first run initialization at EL2.
 627          */
 628         if (kvm_vm_is_protected(kvm))
 629                 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu);
 630
 631         mutex_lock(&kvm->arch.config_lock);
 632         set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags);
 633         mutex_unlock(&kvm->arch.config_lock);
 634
 635         return ret;
 636 }
 637
 638 bool kvm_arch_intc_initialized(struct kvm *kvm)
 639 {
 640         return vgic_initialized(kvm);
 641 }
 642
 643 void kvm_arm_halt_guest(struct kvm *kvm)
 644 {
 645         unsigned long i;
 646         struct kvm_vcpu *vcpu;
 647
 648         kvm_for_each_vcpu(i, vcpu, kvm)
 649                 vcpu->arch.pause = true;
 650         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 651 }
 652
 653 void kvm_arm_resume_guest(struct kvm *kvm)
 654 {
 655         unsigned long i;
 656         struct kvm_vcpu *vcpu;
 657
 658         kvm_for_each_vcpu(i, vcpu, kvm) {
 659                 vcpu->arch.pause = false;
 660                 __kvm_vcpu_wake_up(vcpu);
 661         }
 662 }
 663
 664 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu)
 665 {
 666         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 667
 668         rcuwait_wait_event(wait,
 669                            (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause),
 670                            TASK_INTERRUPTIBLE);
 671
 672         if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) {
 673                 /* Awaken to handle a signal, request we sleep again later. */
 674                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 675         }
 676
 677         /*
 678          * Make sure we will observe a potential reset request if we've
 679          * observed a change to the power state. Pairs with the smp_wmb() in
 680          * kvm_psci_vcpu_on().
 681          */
 682         smp_rmb();
 683 }
 684
 685 /**
 686  * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior
 687  * @vcpu:       The VCPU pointer
 688  *
 689  * Suspend execution of a vCPU until a valid wake event is detected, i.e. until
 690  * the vCPU is runnable.  The vCPU may or may not be scheduled out, depending
 691  * on when a wake event arrives, e.g. there may already be a pending wake event.
 692  */
 693 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu)
 694 {
 695         /*
 696          * Sync back the state of the GIC CPU interface so that we have
 697          * the latest PMR and group enables. This ensures that
 698          * kvm_arch_vcpu_runnable has up-to-date data to decide whether
 699          * we have pending interrupts, e.g. when determining if the
 700          * vCPU should block.
 701          *
 702          * For the same reason, we want to tell GICv4 that we need
 703          * doorbells to be signalled, should an interrupt become pending.
 704          */
 705         preempt_disable();
 706         kvm_vgic_vmcr_sync(vcpu);
 707         vgic_v4_put(vcpu, true);
 708         preempt_enable();
 709
 710         kvm_vcpu_halt(vcpu);
 711         vcpu_clear_flag(vcpu, IN_WFIT);
 712
 713         preempt_disable();
 714         vgic_v4_load(vcpu);
 715         preempt_enable();
 716 }
 717
 718 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu)
 719 {
 720         if (!kvm_arm_vcpu_suspended(vcpu))
 721                 return 1;
 722
 723         kvm_vcpu_wfi(vcpu);
 724
 725         /*
 726          * The suspend state is sticky; we do not leave it until userspace
 727          * explicitly marks the vCPU as runnable. Request that we suspend again
 728          * later.
 729          */
 730         kvm_make_request(KVM_REQ_SUSPEND, vcpu);
 731
 732         /*
 733          * Check to make sure the vCPU is actually runnable. If so, exit to
 734          * userspace informing it of the wakeup condition.
 735          */
 736         if (kvm_arch_vcpu_runnable(vcpu)) {
 737                 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event));
 738                 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP;
 739                 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
 740                 return 0;
 741         }
 742
 743         /*
 744          * Otherwise, we were unblocked to process a different event, such as a
 745          * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to
 746          * process the event.
 747          */
 748         return 1;
 749 }
 750
 751 /**
 752  * check_vcpu_requests - check and handle pending vCPU requests
 753  * @vcpu:       the VCPU pointer
 754  *
 755  * Return: 1 if we should enter the guest
 756  *         0 if we should exit to userspace
 757  *         < 0 if we should exit to userspace, where the return value indicates
 758  *         an error
 759  */
 760 static int check_vcpu_requests(struct kvm_vcpu *vcpu)
 761 {
 762         if (kvm_request_pending(vcpu)) {
 763                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 764                         kvm_vcpu_sleep(vcpu);
 765
 766                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 767                         kvm_reset_vcpu(vcpu);
 768
 769                 /*
 770                  * Clear IRQ_PENDING requests that were made to guarantee
 771                  * that a VCPU sees new virtual interrupts.
 772                  */
 773                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 774
 775                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 776                         kvm_update_stolen_time(vcpu);
 777
 778                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 779                         /* The distributor enable bits were changed */
 780                         preempt_disable();
 781                         vgic_v4_put(vcpu, false);
 782                         vgic_v4_load(vcpu);
 783                         preempt_enable();
 784                 }
 785
 786                 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
 787                         kvm_pmu_handle_pmcr(vcpu,
 788                                             __vcpu_sys_reg(vcpu, PMCR_EL0));
 789
 790                 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu))
 791                         return kvm_vcpu_suspend(vcpu);
 792
 793                 if (kvm_dirty_ring_check_request(vcpu))
 794                         return 0;
 795         }
 796
 797         return 1;
 798 }
 799
 800 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
 801 {
 802         if (likely(!vcpu_mode_is_32bit(vcpu)))
 803                 return false;
 804
 805         return !kvm_supports_32bit_el0();
 806 }
 807
 808 /**
 809  * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest
 810  * @vcpu:       The VCPU pointer
 811  * @ret:        Pointer to write optional return code
 812  *
 813  * Returns: true if the VCPU needs to return to a preemptible + interruptible
 814  *          and skip guest entry.
 815  *
 816  * This function disambiguates between two different types of exits: exits to a
 817  * preemptible + interruptible kernel context and exits to userspace. For an
 818  * exit to userspace, this function will write the return code to ret and return
 819  * true. For an exit to preemptible + interruptible kernel context (i.e. check
 820  * for pending work and re-enter), return true without writing to ret.
 821  */
 822 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret)
 823 {
 824         struct kvm_run *run = vcpu->run;
 825
 826         /*
 827          * If we're using a userspace irqchip, then check if we need
 828          * to tell a userspace irqchip about timer or PMU level
 829          * changes and if so, exit to userspace (the actual level
 830          * state gets updated in kvm_timer_update_run and
 831          * kvm_pmu_update_run below).
 832          */
 833         if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 834                 if (kvm_timer_should_notify_user(vcpu) ||
 835                     kvm_pmu_should_notify_user(vcpu)) {
 836                         *ret = -EINTR;
 837                         run->exit_reason = KVM_EXIT_INTR;
 838                         return true;
 839                 }
 840         }
 841
 842         if (unlikely(vcpu_on_unsupported_cpu(vcpu))) {
 843                 run->exit_reason = KVM_EXIT_FAIL_ENTRY;
 844                 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED;
 845                 run->fail_entry.cpu = smp_processor_id();
 846                 *ret = 0;
 847                 return true;
 848         }
 849
 850         return kvm_request_pending(vcpu) ||
 851                         xfer_to_guest_mode_work_pending();
 852 }
 853
 854 /*
 855  * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while
 856  * the vCPU is running.
 857  *
 858  * This must be noinstr as instrumentation may make use of RCU, and this is not
 859  * safe during the EQS.
 860  */
 861 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
 862 {
 863         int ret;
 864
 865         guest_state_enter_irqoff();
 866         ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 867         guest_state_exit_irqoff();
 868
 869         return ret;
 870 }
 871
 872 /**
 873  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 874  * @vcpu:       The VCPU pointer
 875  *
 876  * This function is called through the VCPU_RUN ioctl called from user space. It
 877  * will execute VM code in a loop until the time slice for the process is used
 878  * or some emulation is needed from user space in which case the function will
 879  * return with return value 0 and with the kvm_run structure filled in with the
 880  * required data for the requested emulation.
 881  */
 882 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 883 {
 884         struct kvm_run *run = vcpu->run;
 885         int ret;
 886
 887         if (run->exit_reason == KVM_EXIT_MMIO) {
 888                 ret = kvm_handle_mmio_return(vcpu);
 889                 if (ret)
 890                         return ret;
 891         }
 892
 893         vcpu_load(vcpu);
 894
 895         if (run->immediate_exit) {
 896                 ret = -EINTR;
 897                 goto out;
 898         }
 899
 900         kvm_sigset_activate(vcpu);
 901
 902         ret = 1;
 903         run->exit_reason = KVM_EXIT_UNKNOWN;
 904         run->flags = 0;
 905         while (ret > 0) {
 906                 /*
 907                  * Check conditions before entering the guest
 908                  */
 909                 ret = xfer_to_guest_mode_handle_work(vcpu);
 910                 if (!ret)
 911                         ret = 1;
 912
 913                 if (ret > 0)
 914                         ret = check_vcpu_requests(vcpu);
 915
 916                 /*
 917                  * Preparing the interrupts to be injected also
 918                  * involves poking the GIC, which must be done in a
 919                  * non-preemptible context.
 920                  */
 921                 preempt_disable();
 922
 923                 /*
 924                  * The VMID allocator only tracks active VMIDs per
 925                  * physical CPU, and therefore the VMID allocated may not be
 926                  * preserved on VMID roll-over if the task was preempted,
 927                  * making a thread's VMID inactive. So we need to call
 928                  * kvm_arm_vmid_update() in non-premptible context.
 929                  */
 930                 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid);
 931
 932                 kvm_pmu_flush_hwstate(vcpu);
 933
 934                 local_irq_disable();
 935
 936                 kvm_vgic_flush_hwstate(vcpu);
 937
 938                 kvm_pmu_update_vcpu_events(vcpu);
 939
 940                 /*
 941                  * Ensure we set mode to IN_GUEST_MODE after we disable
 942                  * interrupts and before the final VCPU requests check.
 943                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 944                  * Documentation/virt/kvm/vcpu-requests.rst
 945                  */
 946                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 947
 948                 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) {
 949                         vcpu->mode = OUTSIDE_GUEST_MODE;
 950                         isb(); /* Ensure work in x_flush_hwstate is committed */
 951                         kvm_pmu_sync_hwstate(vcpu);
 952                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 953                                 kvm_timer_sync_user(vcpu);
 954                         kvm_vgic_sync_hwstate(vcpu);
 955                         local_irq_enable();
 956                         preempt_enable();
 957                         continue;
 958                 }
 959
 960                 kvm_arm_setup_debug(vcpu);
 961                 kvm_arch_vcpu_ctxflush_fp(vcpu);
 962
 963                 /**************************************************************
 964                  * Enter the guest
 965                  */
 966                 trace_kvm_entry(*vcpu_pc(vcpu));
 967                 guest_timing_enter_irqoff();
 968
 969                 ret = kvm_arm_vcpu_enter_exit(vcpu);
 970
 971                 vcpu->mode = OUTSIDE_GUEST_MODE;
 972                 vcpu->stat.exits++;
 973                 /*
 974                  * Back from guest
 975                  *************************************************************/
 976
 977                 kvm_arm_clear_debug(vcpu);
 978
 979                 /*
 980                  * We must sync the PMU state before the vgic state so
 981                  * that the vgic can properly sample the updated state of the
 982                  * interrupt line.
 983                  */
 984                 kvm_pmu_sync_hwstate(vcpu);
 985
 986                 /*
 987                  * Sync the vgic state before syncing the timer state because
 988                  * the timer code needs to know if the virtual timer
 989                  * interrupts are active.
 990                  */
 991                 kvm_vgic_sync_hwstate(vcpu);
 992
 993                 /*
 994                  * Sync the timer hardware state before enabling interrupts as
 995                  * we don't want vtimer interrupts to race with syncing the
 996                  * timer virtual interrupt state.
 997                  */
 998                 if (static_branch_unlikely(&userspace_irqchip_in_use))
 999                         kvm_timer_sync_user(vcpu);
1000
1001                 kvm_arch_vcpu_ctxsync_fp(vcpu);
1002
1003                 /*
1004                  * We must ensure that any pending interrupts are taken before
1005                  * we exit guest timing so that timer ticks are accounted as
1006                  * guest time. Transiently unmask interrupts so that any
1007                  * pending interrupts are taken.
1008                  *
1009                  * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other
1010                  * context synchronization event) is necessary to ensure that
1011                  * pending interrupts are taken.
1012                  */
1013                 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) {
1014                         local_irq_enable();
1015                         isb();
1016                         local_irq_disable();
1017                 }
1018
1019                 guest_timing_exit_irqoff();
1020
1021                 local_irq_enable();
1022
1023                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
1024
1025                 /* Exit types that need handling before we can be preempted */
1026                 handle_exit_early(vcpu, ret);
1027
1028                 preempt_enable();
1029
1030                 /*
1031                  * The ARMv8 architecture doesn't give the hypervisor
1032                  * a mechanism to prevent a guest from dropping to AArch32 EL0
1033                  * if implemented by the CPU. If we spot the guest in such
1034                  * state and that we decided it wasn't supposed to do so (like
1035                  * with the asymmetric AArch32 case), return to userspace with
1036                  * a fatal error.
1037                  */
1038                 if (vcpu_mode_is_bad_32bit(vcpu)) {
1039                         /*
1040                          * As we have caught the guest red-handed, decide that
1041                          * it isn't fit for purpose anymore by making the vcpu
1042                          * invalid. The VMM can try and fix it by issuing  a
1043                          * KVM_ARM_VCPU_INIT if it really wants to.
1044                          */
1045                         vcpu->arch.target = -1;
1046                         ret = ARM_EXCEPTION_IL;
1047                 }
1048
1049                 ret = handle_exit(vcpu, ret);
1050         }
1051
1052         /* Tell userspace about in-kernel device output levels */
1053         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1054                 kvm_timer_update_run(vcpu);
1055                 kvm_pmu_update_run(vcpu);
1056         }
1057
1058         kvm_sigset_deactivate(vcpu);
1059
1060 out:
1061         /*
1062          * In the unlikely event that we are returning to userspace
1063          * with pending exceptions or PC adjustment, commit these
1064          * adjustments in order to give userspace a consistent view of
1065          * the vcpu state. Note that this relies on __kvm_adjust_pc()
1066          * being preempt-safe on VHE.
1067          */
1068         if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) ||
1069                      vcpu_get_flag(vcpu, INCREMENT_PC)))
1070                 kvm_call_hyp(__kvm_adjust_pc, vcpu);
1071
1072         vcpu_put(vcpu);
1073         return ret;
1074 }
1075
1076 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
1077 {
1078         int bit_index;
1079         bool set;
1080         unsigned long *hcr;
1081
1082         if (number == KVM_ARM_IRQ_CPU_IRQ)
1083                 bit_index = __ffs(HCR_VI);
1084         else /* KVM_ARM_IRQ_CPU_FIQ */
1085                 bit_index = __ffs(HCR_VF);
1086
1087         hcr = vcpu_hcr(vcpu);
1088         if (level)
1089                 set = test_and_set_bit(bit_index, hcr);
1090         else
1091                 set = test_and_clear_bit(bit_index, hcr);
1092
1093         /*
1094          * If we didn't change anything, no need to wake up or kick other CPUs
1095          */
1096         if (set == level)
1097                 return 0;
1098
1099         /*
1100          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
1101          * trigger a world-switch round on the running physical CPU to set the
1102          * virtual IRQ/FIQ fields in the HCR appropriately.
1103          */
1104         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
1105         kvm_vcpu_kick(vcpu);
1106
1107         return 0;
1108 }
1109
1110 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
1111                           bool line_status)
1112 {
1113         u32 irq = irq_level->irq;
1114         unsigned int irq_type, vcpu_idx, irq_num;
1115         int nrcpus = atomic_read(&kvm->online_vcpus);
1116         struct kvm_vcpu *vcpu = NULL;
1117         bool level = irq_level->level;
1118
1119         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
1120         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
1121         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
1122         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
1123
1124         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
1125
1126         switch (irq_type) {
1127         case KVM_ARM_IRQ_TYPE_CPU:
1128                 if (irqchip_in_kernel(kvm))
1129                         return -ENXIO;
1130
1131                 if (vcpu_idx >= nrcpus)
1132                         return -EINVAL;
1133
1134                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1135                 if (!vcpu)
1136                         return -EINVAL;
1137
1138                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1139                         return -EINVAL;
1140
1141                 return vcpu_interrupt_line(vcpu, irq_num, level);
1142         case KVM_ARM_IRQ_TYPE_PPI:
1143                 if (!irqchip_in_kernel(kvm))
1144                         return -ENXIO;
1145
1146                 if (vcpu_idx >= nrcpus)
1147                         return -EINVAL;
1148
1149                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
1150                 if (!vcpu)
1151                         return -EINVAL;
1152
1153                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
1154                         return -EINVAL;
1155
1156                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
1157         case KVM_ARM_IRQ_TYPE_SPI:
1158                 if (!irqchip_in_kernel(kvm))
1159                         return -ENXIO;
1160
1161                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
1162                         return -EINVAL;
1163
1164                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
1165         }
1166
1167         return -EINVAL;
1168 }
1169
1170 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
1171                                         const struct kvm_vcpu_init *init)
1172 {
1173         unsigned long features = init->features[0];
1174         int i;
1175
1176         if (features & ~KVM_VCPU_VALID_FEATURES)
1177                 return -ENOENT;
1178
1179         for (i = 1; i < ARRAY_SIZE(init->features); i++) {
1180                 if (init->features[i])
1181                         return -ENOENT;
1182         }
1183
1184         return 0;
1185 }
1186
1187 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu,
1188                                   const struct kvm_vcpu_init *init)
1189 {
1190         unsigned long features = init->features[0];
1191
1192         return !bitmap_equal(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES) ||
1193                         vcpu->arch.target != init->target;
1194 }
1195
1196 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1197                                  const struct kvm_vcpu_init *init)
1198 {
1199         unsigned long features = init->features[0];
1200         int ret;
1201
1202         vcpu->arch.target = init->target;
1203         bitmap_copy(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES);
1204
1205         /* Now we know what it is, we can reset it. */
1206         ret = kvm_reset_vcpu(vcpu);
1207         if (ret) {
1208                 vcpu->arch.target = -1;
1209                 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
1210         }
1211
1212         return ret;
1213 }
1214
1215 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1216                                const struct kvm_vcpu_init *init)
1217 {
1218         int ret;
1219
1220         if (init->target != kvm_target_cpu())
1221                 return -EINVAL;
1222
1223         ret = kvm_vcpu_init_check_features(vcpu, init);
1224         if (ret)
1225                 return ret;
1226
1227         if (vcpu->arch.target == -1)
1228                 return __kvm_vcpu_set_target(vcpu, init);
1229
1230         if (kvm_vcpu_init_changed(vcpu, init))
1231                 return -EINVAL;
1232
1233         return kvm_reset_vcpu(vcpu);
1234 }
1235
1236 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1237                                          struct kvm_vcpu_init *init)
1238 {
1239         int ret;
1240
1241         ret = kvm_vcpu_set_target(vcpu, init);
1242         if (ret)
1243                 return ret;
1244
1245         /*
1246          * Ensure a rebooted VM will fault in RAM pages and detect if the
1247          * guest MMU is turned off and flush the caches as needed.
1248          *
1249          * S2FWB enforces all memory accesses to RAM being cacheable,
1250          * ensuring that the data side is always coherent. We still
1251          * need to invalidate the I-cache though, as FWB does *not*
1252          * imply CTR_EL0.DIC.
1253          */
1254         if (vcpu_has_run_once(vcpu)) {
1255                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1256                         stage2_unmap_vm(vcpu->kvm);
1257                 else
1258                         icache_inval_all_pou();
1259         }
1260
1261         vcpu_reset_hcr(vcpu);
1262         vcpu->arch.cptr_el2 = CPTR_EL2_DEFAULT;
1263
1264         /*
1265          * Handle the "start in power-off" case.
1266          */
1267         spin_lock(&vcpu->arch.mp_state_lock);
1268
1269         if (test_bit(KVM_ARM_VCPU_POWER_OFF, vcpu->arch.features))
1270                 __kvm_arm_vcpu_power_off(vcpu);
1271         else
1272                 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE);
1273
1274         spin_unlock(&vcpu->arch.mp_state_lock);
1275
1276         return 0;
1277 }
1278
1279 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1280                                  struct kvm_device_attr *attr)
1281 {
1282         int ret = -ENXIO;
1283
1284         switch (attr->group) {
1285         default:
1286                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1287                 break;
1288         }
1289
1290         return ret;
1291 }
1292
1293 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1294                                  struct kvm_device_attr *attr)
1295 {
1296         int ret = -ENXIO;
1297
1298         switch (attr->group) {
1299         default:
1300                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1301                 break;
1302         }
1303
1304         return ret;
1305 }
1306
1307 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1308                                  struct kvm_device_attr *attr)
1309 {
1310         int ret = -ENXIO;
1311
1312         switch (attr->group) {
1313         default:
1314                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1315                 break;
1316         }
1317
1318         return ret;
1319 }
1320
1321 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1322                                    struct kvm_vcpu_events *events)
1323 {
1324         memset(events, 0, sizeof(*events));
1325
1326         return __kvm_arm_vcpu_get_events(vcpu, events);
1327 }
1328
1329 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1330                                    struct kvm_vcpu_events *events)
1331 {
1332         int i;
1333
1334         /* check whether the reserved field is zero */
1335         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1336                 if (events->reserved[i])
1337                         return -EINVAL;
1338
1339         /* check whether the pad field is zero */
1340         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1341                 if (events->exception.pad[i])
1342                         return -EINVAL;
1343
1344         return __kvm_arm_vcpu_set_events(vcpu, events);
1345 }
1346
1347 long kvm_arch_vcpu_ioctl(struct file *filp,
1348                          unsigned int ioctl, unsigned long arg)
1349 {
1350         struct kvm_vcpu *vcpu = filp->private_data;
1351         void __user *argp = (void __user *)arg;
1352         struct kvm_device_attr attr;
1353         long r;
1354
1355         switch (ioctl) {
1356         case KVM_ARM_VCPU_INIT: {
1357                 struct kvm_vcpu_init init;
1358
1359                 r = -EFAULT;
1360                 if (copy_from_user(&init, argp, sizeof(init)))
1361                         break;
1362
1363                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1364                 break;
1365         }
1366         case KVM_SET_ONE_REG:
1367         case KVM_GET_ONE_REG: {
1368                 struct kvm_one_reg reg;
1369
1370                 r = -ENOEXEC;
1371                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1372                         break;
1373
1374                 r = -EFAULT;
1375                 if (copy_from_user(&reg, argp, sizeof(reg)))
1376                         break;
1377
1378                 /*
1379                  * We could owe a reset due to PSCI. Handle the pending reset
1380                  * here to ensure userspace register accesses are ordered after
1381                  * the reset.
1382                  */
1383                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1384                         kvm_reset_vcpu(vcpu);
1385
1386                 if (ioctl == KVM_SET_ONE_REG)
1387                         r = kvm_arm_set_reg(vcpu, &reg);
1388                 else
1389                         r = kvm_arm_get_reg(vcpu, &reg);
1390                 break;
1391         }
1392         case KVM_GET_REG_LIST: {
1393                 struct kvm_reg_list __user *user_list = argp;
1394                 struct kvm_reg_list reg_list;
1395                 unsigned n;
1396
1397                 r = -ENOEXEC;
1398                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1399                         break;
1400
1401                 r = -EPERM;
1402                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1403                         break;
1404
1405                 r = -EFAULT;
1406                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1407                         break;
1408                 n = reg_list.n;
1409                 reg_list.n = kvm_arm_num_regs(vcpu);
1410                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1411                         break;
1412                 r = -E2BIG;
1413                 if (n < reg_list.n)
1414                         break;
1415                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1416                 break;
1417         }
1418         case KVM_SET_DEVICE_ATTR: {
1419                 r = -EFAULT;
1420                 if (copy_from_user(&attr, argp, sizeof(attr)))
1421                         break;
1422                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1423                 break;
1424         }
1425         case KVM_GET_DEVICE_ATTR: {
1426                 r = -EFAULT;
1427                 if (copy_from_user(&attr, argp, sizeof(attr)))
1428                         break;
1429                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1430                 break;
1431         }
1432         case KVM_HAS_DEVICE_ATTR: {
1433                 r = -EFAULT;
1434                 if (copy_from_user(&attr, argp, sizeof(attr)))
1435                         break;
1436                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1437                 break;
1438         }
1439         case KVM_GET_VCPU_EVENTS: {
1440                 struct kvm_vcpu_events events;
1441
1442                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1443                         return -EINVAL;
1444
1445                 if (copy_to_user(argp, &events, sizeof(events)))
1446                         return -EFAULT;
1447
1448                 return 0;
1449         }
1450         case KVM_SET_VCPU_EVENTS: {
1451                 struct kvm_vcpu_events events;
1452
1453                 if (copy_from_user(&events, argp, sizeof(events)))
1454                         return -EFAULT;
1455
1456                 return kvm_arm_vcpu_set_events(vcpu, &events);
1457         }
1458         case KVM_ARM_VCPU_FINALIZE: {
1459                 int what;
1460
1461                 if (!kvm_vcpu_initialized(vcpu))
1462                         return -ENOEXEC;
1463
1464                 if (get_user(what, (const int __user *)argp))
1465                         return -EFAULT;
1466
1467                 return kvm_arm_vcpu_finalize(vcpu, what);
1468         }
1469         default:
1470                 r = -EINVAL;
1471         }
1472
1473         return r;
1474 }
1475
1476 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1477 {
1478
1479 }
1480
1481 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
1482                                         const struct kvm_memory_slot *memslot)
1483 {
1484         kvm_flush_remote_tlbs(kvm);
1485 }
1486
1487 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1488                                         struct kvm_arm_device_addr *dev_addr)
1489 {
1490         switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) {
1491         case KVM_ARM_DEVICE_VGIC_V2:
1492                 if (!vgic_present)
1493                         return -ENXIO;
1494                 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr);
1495         default:
1496                 return -ENODEV;
1497         }
1498 }
1499
1500 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1501 {
1502         switch (attr->group) {
1503         case KVM_ARM_VM_SMCCC_CTRL:
1504                 return kvm_vm_smccc_has_attr(kvm, attr);
1505         default:
1506                 return -ENXIO;
1507         }
1508 }
1509
1510 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1511 {
1512         switch (attr->group) {
1513         case KVM_ARM_VM_SMCCC_CTRL:
1514                 return kvm_vm_smccc_set_attr(kvm, attr);
1515         default:
1516                 return -ENXIO;
1517         }
1518 }
1519
1520 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
1521 {
1522         struct kvm *kvm = filp->private_data;
1523         void __user *argp = (void __user *)arg;
1524         struct kvm_device_attr attr;
1525
1526         switch (ioctl) {
1527         case KVM_CREATE_IRQCHIP: {
1528                 int ret;
1529                 if (!vgic_present)
1530                         return -ENXIO;
1531                 mutex_lock(&kvm->lock);
1532                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1533                 mutex_unlock(&kvm->lock);
1534                 return ret;
1535         }
1536         case KVM_ARM_SET_DEVICE_ADDR: {
1537                 struct kvm_arm_device_addr dev_addr;
1538
1539                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1540                         return -EFAULT;
1541                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1542         }
1543         case KVM_ARM_PREFERRED_TARGET: {
1544                 struct kvm_vcpu_init init;
1545
1546                 kvm_vcpu_preferred_target(&init);
1547
1548                 if (copy_to_user(argp, &init, sizeof(init)))
1549                         return -EFAULT;
1550
1551                 return 0;
1552         }
1553         case KVM_ARM_MTE_COPY_TAGS: {
1554                 struct kvm_arm_copy_mte_tags copy_tags;
1555
1556                 if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1557                         return -EFAULT;
1558                 return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
1559         }
1560         case KVM_ARM_SET_COUNTER_OFFSET: {
1561                 struct kvm_arm_counter_offset offset;
1562
1563                 if (copy_from_user(&offset, argp, sizeof(offset)))
1564                         return -EFAULT;
1565                 return kvm_vm_ioctl_set_counter_offset(kvm, &offset);
1566         }
1567         case KVM_HAS_DEVICE_ATTR: {
1568                 if (copy_from_user(&attr, argp, sizeof(attr)))
1569                         return -EFAULT;
1570
1571                 return kvm_vm_has_attr(kvm, &attr);
1572         }
1573         case KVM_SET_DEVICE_ATTR: {
1574                 if (copy_from_user(&attr, argp, sizeof(attr)))
1575                         return -EFAULT;
1576
1577                 return kvm_vm_set_attr(kvm, &attr);
1578         }
1579         default:
1580                 return -EINVAL;
1581         }
1582 }
1583
1584 /* unlocks vcpus from @vcpu_lock_idx and smaller */
1585 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx)
1586 {
1587         struct kvm_vcpu *tmp_vcpu;
1588
1589         for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) {
1590                 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx);
1591                 mutex_unlock(&tmp_vcpu->mutex);
1592         }
1593 }
1594
1595 void unlock_all_vcpus(struct kvm *kvm)
1596 {
1597         lockdep_assert_held(&kvm->lock);
1598
1599         unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1);
1600 }
1601
1602 /* Returns true if all vcpus were locked, false otherwise */
1603 bool lock_all_vcpus(struct kvm *kvm)
1604 {
1605         struct kvm_vcpu *tmp_vcpu;
1606         unsigned long c;
1607
1608         lockdep_assert_held(&kvm->lock);
1609
1610         /*
1611          * Any time a vcpu is in an ioctl (including running), the
1612          * core KVM code tries to grab the vcpu->mutex.
1613          *
1614          * By grabbing the vcpu->mutex of all VCPUs we ensure that no
1615          * other VCPUs can fiddle with the state while we access it.
1616          */
1617         kvm_for_each_vcpu(c, tmp_vcpu, kvm) {
1618                 if (!mutex_trylock(&tmp_vcpu->mutex)) {
1619                         unlock_vcpus(kvm, c - 1);
1620                         return false;
1621                 }
1622         }
1623
1624         return true;
1625 }
1626
1627 static unsigned long nvhe_percpu_size(void)
1628 {
1629         return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
1630                 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
1631 }
1632
1633 static unsigned long nvhe_percpu_order(void)
1634 {
1635         unsigned long size = nvhe_percpu_size();
1636
1637         return size ? get_order(size) : 0;
1638 }
1639
1640 /* A lookup table holding the hypervisor VA for each vector slot */
1641 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
1642
1643 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
1644 {
1645         hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
1646 }
1647
1648 static int kvm_init_vector_slots(void)
1649 {
1650         int err;
1651         void *base;
1652
1653         base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
1654         kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
1655
1656         base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
1657         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
1658
1659         if (kvm_system_needs_idmapped_vectors() &&
1660             !is_protected_kvm_enabled()) {
1661                 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
1662                                                __BP_HARDEN_HYP_VECS_SZ, &base);
1663                 if (err)
1664                         return err;
1665         }
1666
1667         kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
1668         kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
1669         return 0;
1670 }
1671
1672 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits)
1673 {
1674         struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
1675         unsigned long tcr;
1676
1677         /*
1678          * Calculate the raw per-cpu offset without a translation from the
1679          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1680          * so that we can use adr_l to access per-cpu variables in EL2.
1681          * Also drop the KASAN tag which gets in the way...
1682          */
1683         params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
1684                             (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
1685
1686         params->mair_el2 = read_sysreg(mair_el1);
1687
1688         tcr = (read_sysreg(tcr_el1) & TCR_EL2_MASK) | TCR_EL2_RES1;
1689         tcr &= ~TCR_T0SZ_MASK;
1690         tcr |= TCR_T0SZ(hyp_va_bits);
1691         params->tcr_el2 = tcr;
1692
1693         params->pgd_pa = kvm_mmu_get_httbr();
1694         if (is_protected_kvm_enabled())
1695                 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
1696         else
1697                 params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
1698         params->vttbr = params->vtcr = 0;
1699
1700         /*
1701          * Flush the init params from the data cache because the struct will
1702          * be read while the MMU is off.
1703          */
1704         kvm_flush_dcache_to_poc(params, sizeof(*params));
1705 }
1706
1707 static void hyp_install_host_vector(void)
1708 {
1709         struct kvm_nvhe_init_params *params;
1710         struct arm_smccc_res res;
1711
1712         /* Switch from the HYP stub to our own HYP init vector */
1713         __hyp_set_vectors(kvm_get_idmap_vector());
1714
1715         /*
1716          * Call initialization code, and switch to the full blown HYP code.
1717          * If the cpucaps haven't been finalized yet, something has gone very
1718          * wrong, and hyp will crash and burn when it uses any
1719          * cpus_have_const_cap() wrapper.
1720          */
1721         BUG_ON(!system_capabilities_finalized());
1722         params = this_cpu_ptr_nvhe_sym(kvm_init_params);
1723         arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
1724         WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
1725 }
1726
1727 static void cpu_init_hyp_mode(void)
1728 {
1729         hyp_install_host_vector();
1730
1731         /*
1732          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1733          * at EL2.
1734          */
1735         if (this_cpu_has_cap(ARM64_SSBS) &&
1736             arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
1737                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1738         }
1739 }
1740
1741 static void cpu_hyp_reset(void)
1742 {
1743         if (!is_kernel_in_hyp_mode())
1744                 __hyp_reset_vectors();
1745 }
1746
1747 /*
1748  * EL2 vectors can be mapped and rerouted in a number of ways,
1749  * depending on the kernel configuration and CPU present:
1750  *
1751  * - If the CPU is affected by Spectre-v2, the hardening sequence is
1752  *   placed in one of the vector slots, which is executed before jumping
1753  *   to the real vectors.
1754  *
1755  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
1756  *   containing the hardening sequence is mapped next to the idmap page,
1757  *   and executed before jumping to the real vectors.
1758  *
1759  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
1760  *   empty slot is selected, mapped next to the idmap page, and
1761  *   executed before jumping to the real vectors.
1762  *
1763  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
1764  * VHE, as we don't have hypervisor-specific mappings. If the system
1765  * is VHE and yet selects this capability, it will be ignored.
1766  */
1767 static void cpu_set_hyp_vector(void)
1768 {
1769         struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
1770         void *vector = hyp_spectre_vector_selector[data->slot];
1771
1772         if (!is_protected_kvm_enabled())
1773                 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
1774         else
1775                 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
1776 }
1777
1778 static void cpu_hyp_init_context(void)
1779 {
1780         kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt);
1781
1782         if (!is_kernel_in_hyp_mode())
1783                 cpu_init_hyp_mode();
1784 }
1785
1786 static void cpu_hyp_init_features(void)
1787 {
1788         cpu_set_hyp_vector();
1789         kvm_arm_init_debug();
1790
1791         if (is_kernel_in_hyp_mode())
1792                 kvm_timer_init_vhe();
1793
1794         if (vgic_present)
1795                 kvm_vgic_init_cpu_hardware();
1796 }
1797
1798 static void cpu_hyp_reinit(void)
1799 {
1800         cpu_hyp_reset();
1801         cpu_hyp_init_context();
1802         cpu_hyp_init_features();
1803 }
1804
1805 static void _kvm_arch_hardware_enable(void *discard)
1806 {
1807         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1808                 cpu_hyp_reinit();
1809                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1810         }
1811 }
1812
1813 int kvm_arch_hardware_enable(void)
1814 {
1815         int was_enabled = __this_cpu_read(kvm_arm_hardware_enabled);
1816
1817         _kvm_arch_hardware_enable(NULL);
1818
1819         if (!was_enabled) {
1820                 kvm_vgic_cpu_up();
1821                 kvm_timer_cpu_up();
1822         }
1823
1824         return 0;
1825 }
1826
1827 static void _kvm_arch_hardware_disable(void *discard)
1828 {
1829         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1830                 cpu_hyp_reset();
1831                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1832         }
1833 }
1834
1835 void kvm_arch_hardware_disable(void)
1836 {
1837         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1838                 kvm_timer_cpu_down();
1839                 kvm_vgic_cpu_down();
1840         }
1841
1842         if (!is_protected_kvm_enabled())
1843                 _kvm_arch_hardware_disable(NULL);
1844 }
1845
1846 #ifdef CONFIG_CPU_PM
1847 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1848                                     unsigned long cmd,
1849                                     void *v)
1850 {
1851         /*
1852          * kvm_arm_hardware_enabled is left with its old value over
1853          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1854          * re-enable hyp.
1855          */
1856         switch (cmd) {
1857         case CPU_PM_ENTER:
1858                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1859                         /*
1860                          * don't update kvm_arm_hardware_enabled here
1861                          * so that the hardware will be re-enabled
1862                          * when we resume. See below.
1863                          */
1864                         cpu_hyp_reset();
1865
1866                 return NOTIFY_OK;
1867         case CPU_PM_ENTER_FAILED:
1868         case CPU_PM_EXIT:
1869                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1870                         /* The hardware was enabled before suspend. */
1871                         cpu_hyp_reinit();
1872
1873                 return NOTIFY_OK;
1874
1875         default:
1876                 return NOTIFY_DONE;
1877         }
1878 }
1879
1880 static struct notifier_block hyp_init_cpu_pm_nb = {
1881         .notifier_call = hyp_init_cpu_pm_notifier,
1882 };
1883
1884 static void __init hyp_cpu_pm_init(void)
1885 {
1886         if (!is_protected_kvm_enabled())
1887                 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1888 }
1889 static void __init hyp_cpu_pm_exit(void)
1890 {
1891         if (!is_protected_kvm_enabled())
1892                 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1893 }
1894 #else
1895 static inline void __init hyp_cpu_pm_init(void)
1896 {
1897 }
1898 static inline void __init hyp_cpu_pm_exit(void)
1899 {
1900 }
1901 #endif
1902
1903 static void __init init_cpu_logical_map(void)
1904 {
1905         unsigned int cpu;
1906
1907         /*
1908          * Copy the MPIDR <-> logical CPU ID mapping to hyp.
1909          * Only copy the set of online CPUs whose features have been checked
1910          * against the finalized system capabilities. The hypervisor will not
1911          * allow any other CPUs from the `possible` set to boot.
1912          */
1913         for_each_online_cpu(cpu)
1914                 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
1915 }
1916
1917 #define init_psci_0_1_impl_state(config, what)  \
1918         config.psci_0_1_ ## what ## _implemented = psci_ops.what
1919
1920 static bool __init init_psci_relay(void)
1921 {
1922         /*
1923          * If PSCI has not been initialized, protected KVM cannot install
1924          * itself on newly booted CPUs.
1925          */
1926         if (!psci_ops.get_version) {
1927                 kvm_err("Cannot initialize protected mode without PSCI\n");
1928                 return false;
1929         }
1930
1931         kvm_host_psci_config.version = psci_ops.get_version();
1932
1933         if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
1934                 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
1935                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
1936                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
1937                 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
1938                 init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
1939         }
1940         return true;
1941 }
1942
1943 static int __init init_subsystems(void)
1944 {
1945         int err = 0;
1946
1947         /*
1948          * Enable hardware so that subsystem initialisation can access EL2.
1949          */
1950         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
1951
1952         /*
1953          * Register CPU lower-power notifier
1954          */
1955         hyp_cpu_pm_init();
1956
1957         /*
1958          * Init HYP view of VGIC
1959          */
1960         err = kvm_vgic_hyp_init();
1961         switch (err) {
1962         case 0:
1963                 vgic_present = true;
1964                 break;
1965         case -ENODEV:
1966         case -ENXIO:
1967                 vgic_present = false;
1968                 err = 0;
1969                 break;
1970         default:
1971                 goto out;
1972         }
1973
1974         /*
1975          * Init HYP architected timer support
1976          */
1977         err = kvm_timer_hyp_init(vgic_present);
1978         if (err)
1979                 goto out;
1980
1981         kvm_register_perf_callbacks(NULL);
1982
1983 out:
1984         if (err)
1985                 hyp_cpu_pm_exit();
1986
1987         if (err || !is_protected_kvm_enabled())
1988                 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
1989
1990         return err;
1991 }
1992
1993 static void __init teardown_subsystems(void)
1994 {
1995         kvm_unregister_perf_callbacks();
1996         hyp_cpu_pm_exit();
1997 }
1998
1999 static void __init teardown_hyp_mode(void)
2000 {
2001         int cpu;
2002
2003         free_hyp_pgds();
2004         for_each_possible_cpu(cpu) {
2005                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
2006                 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order());
2007         }
2008 }
2009
2010 static int __init do_pkvm_init(u32 hyp_va_bits)
2011 {
2012         void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base));
2013         int ret;
2014
2015         preempt_disable();
2016         cpu_hyp_init_context();
2017         ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
2018                                 num_possible_cpus(), kern_hyp_va(per_cpu_base),
2019                                 hyp_va_bits);
2020         cpu_hyp_init_features();
2021
2022         /*
2023          * The stub hypercalls are now disabled, so set our local flag to
2024          * prevent a later re-init attempt in kvm_arch_hardware_enable().
2025          */
2026         __this_cpu_write(kvm_arm_hardware_enabled, 1);
2027         preempt_enable();
2028
2029         return ret;
2030 }
2031
2032 static u64 get_hyp_id_aa64pfr0_el1(void)
2033 {
2034         /*
2035          * Track whether the system isn't affected by spectre/meltdown in the
2036          * hypervisor's view of id_aa64pfr0_el1, used for protected VMs.
2037          * Although this is per-CPU, we make it global for simplicity, e.g., not
2038          * to have to worry about vcpu migration.
2039          *
2040          * Unlike for non-protected VMs, userspace cannot override this for
2041          * protected VMs.
2042          */
2043         u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
2044
2045         val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) |
2046                  ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3));
2047
2048         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2),
2049                           arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED);
2050         val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3),
2051                           arm64_get_meltdown_state() == SPECTRE_UNAFFECTED);
2052
2053         return val;
2054 }
2055
2056 static void kvm_hyp_init_symbols(void)
2057 {
2058         kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1();
2059         kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1);
2060         kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
2061         kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
2062         kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
2063         kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
2064         kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
2065         kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1);
2066         kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1);
2067         kvm_nvhe_sym(__icache_flags) = __icache_flags;
2068         kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits;
2069 }
2070
2071 static int __init kvm_hyp_init_protection(u32 hyp_va_bits)
2072 {
2073         void *addr = phys_to_virt(hyp_mem_base);
2074         int ret;
2075
2076         ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
2077         if (ret)
2078                 return ret;
2079
2080         ret = do_pkvm_init(hyp_va_bits);
2081         if (ret)
2082                 return ret;
2083
2084         free_hyp_pgds();
2085
2086         return 0;
2087 }
2088
2089 /* Inits Hyp-mode on all online CPUs */
2090 static int __init init_hyp_mode(void)
2091 {
2092         u32 hyp_va_bits;
2093         int cpu;
2094         int err = -ENOMEM;
2095
2096         /*
2097          * The protected Hyp-mode cannot be initialized if the memory pool
2098          * allocation has failed.
2099          */
2100         if (is_protected_kvm_enabled() && !hyp_mem_base)
2101                 goto out_err;
2102
2103         /*
2104          * Allocate Hyp PGD and setup Hyp identity mapping
2105          */
2106         err = kvm_mmu_init(&hyp_va_bits);
2107         if (err)
2108                 goto out_err;
2109
2110         /*
2111          * Allocate stack pages for Hypervisor-mode
2112          */
2113         for_each_possible_cpu(cpu) {
2114                 unsigned long stack_page;
2115
2116                 stack_page = __get_free_page(GFP_KERNEL);
2117                 if (!stack_page) {
2118                         err = -ENOMEM;
2119                         goto out_err;
2120                 }
2121
2122                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
2123         }
2124
2125         /*
2126          * Allocate and initialize pages for Hypervisor-mode percpu regions.
2127          */
2128         for_each_possible_cpu(cpu) {
2129                 struct page *page;
2130                 void *page_addr;
2131
2132                 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
2133                 if (!page) {
2134                         err = -ENOMEM;
2135                         goto out_err;
2136                 }
2137
2138                 page_addr = page_address(page);
2139                 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
2140                 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr;
2141         }
2142
2143         /*
2144          * Map the Hyp-code called directly from the host
2145          */
2146         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
2147                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
2148         if (err) {
2149                 kvm_err("Cannot map world-switch code\n");
2150                 goto out_err;
2151         }
2152
2153         err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
2154                                   kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
2155         if (err) {
2156                 kvm_err("Cannot map .hyp.rodata section\n");
2157                 goto out_err;
2158         }
2159
2160         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
2161                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
2162         if (err) {
2163                 kvm_err("Cannot map rodata section\n");
2164                 goto out_err;
2165         }
2166
2167         /*
2168          * .hyp.bss is guaranteed to be placed at the beginning of the .bss
2169          * section thanks to an assertion in the linker script. Map it RW and
2170          * the rest of .bss RO.
2171          */
2172         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
2173                                   kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
2174         if (err) {
2175                 kvm_err("Cannot map hyp bss section: %d\n", err);
2176                 goto out_err;
2177         }
2178
2179         err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
2180                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
2181         if (err) {
2182                 kvm_err("Cannot map bss section\n");
2183                 goto out_err;
2184         }
2185
2186         /*
2187          * Map the Hyp stack pages
2188          */
2189         for_each_possible_cpu(cpu) {
2190                 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2191                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
2192                 unsigned long hyp_addr;
2193
2194                 /*
2195                  * Allocate a contiguous HYP private VA range for the stack
2196                  * and guard page. The allocation is also aligned based on
2197                  * the order of its size.
2198                  */
2199                 err = hyp_alloc_private_va_range(PAGE_SIZE * 2, &hyp_addr);
2200                 if (err) {
2201                         kvm_err("Cannot allocate hyp stack guard page\n");
2202                         goto out_err;
2203                 }
2204
2205                 /*
2206                  * Since the stack grows downwards, map the stack to the page
2207                  * at the higher address and leave the lower guard page
2208                  * unbacked.
2209                  *
2210                  * Any valid stack address now has the PAGE_SHIFT bit as 1
2211                  * and addresses corresponding to the guard page have the
2212                  * PAGE_SHIFT bit as 0 - this is used for overflow detection.
2213                  */
2214                 err = __create_hyp_mappings(hyp_addr + PAGE_SIZE, PAGE_SIZE,
2215                                             __pa(stack_page), PAGE_HYP);
2216                 if (err) {
2217                         kvm_err("Cannot map hyp stack\n");
2218                         goto out_err;
2219                 }
2220
2221                 /*
2222                  * Save the stack PA in nvhe_init_params. This will be needed
2223                  * to recreate the stack mapping in protected nVHE mode.
2224                  * __hyp_pa() won't do the right thing there, since the stack
2225                  * has been mapped in the flexible private VA space.
2226                  */
2227                 params->stack_pa = __pa(stack_page);
2228
2229                 params->stack_hyp_va = hyp_addr + (2 * PAGE_SIZE);
2230         }
2231
2232         for_each_possible_cpu(cpu) {
2233                 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu];
2234                 char *percpu_end = percpu_begin + nvhe_percpu_size();
2235
2236                 /* Map Hyp percpu pages */
2237                 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
2238                 if (err) {
2239                         kvm_err("Cannot map hyp percpu region\n");
2240                         goto out_err;
2241                 }
2242
2243                 /* Prepare the CPU initialization parameters */
2244                 cpu_prepare_hyp_mode(cpu, hyp_va_bits);
2245         }
2246
2247         kvm_hyp_init_symbols();
2248
2249         if (is_protected_kvm_enabled()) {
2250                 init_cpu_logical_map();
2251
2252                 if (!init_psci_relay()) {
2253                         err = -ENODEV;
2254                         goto out_err;
2255                 }
2256
2257                 err = kvm_hyp_init_protection(hyp_va_bits);
2258                 if (err) {
2259                         kvm_err("Failed to init hyp memory protection\n");
2260                         goto out_err;
2261                 }
2262         }
2263
2264         return 0;
2265
2266 out_err:
2267         teardown_hyp_mode();
2268         kvm_err("error initializing Hyp mode: %d\n", err);
2269         return err;
2270 }
2271
2272 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2273 {
2274         struct kvm_vcpu *vcpu;
2275         unsigned long i;
2276
2277         mpidr &= MPIDR_HWID_BITMASK;
2278         kvm_for_each_vcpu(i, vcpu, kvm) {
2279                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2280                         return vcpu;
2281         }
2282         return NULL;
2283 }
2284
2285 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm)
2286 {
2287         return irqchip_in_kernel(kvm);
2288 }
2289
2290 bool kvm_arch_has_irq_bypass(void)
2291 {
2292         return true;
2293 }
2294
2295 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2296                                       struct irq_bypass_producer *prod)
2297 {
2298         struct kvm_kernel_irqfd *irqfd =
2299                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2300
2301         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2302                                           &irqfd->irq_entry);
2303 }
2304 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2305                                       struct irq_bypass_producer *prod)
2306 {
2307         struct kvm_kernel_irqfd *irqfd =
2308                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2309
2310         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
2311                                      &irqfd->irq_entry);
2312 }
2313
2314 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2315 {
2316         struct kvm_kernel_irqfd *irqfd =
2317                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2318
2319         kvm_arm_halt_guest(irqfd->kvm);
2320 }
2321
2322 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2323 {
2324         struct kvm_kernel_irqfd *irqfd =
2325                 container_of(cons, struct kvm_kernel_irqfd, consumer);
2326
2327         kvm_arm_resume_guest(irqfd->kvm);
2328 }
2329
2330 /* Initialize Hyp-mode and memory mappings on all CPUs */
2331 static __init int kvm_arm_init(void)
2332 {
2333         int err;
2334         bool in_hyp_mode;
2335
2336         if (!is_hyp_mode_available()) {
2337                 kvm_info("HYP mode not available\n");
2338                 return -ENODEV;
2339         }
2340
2341         if (kvm_get_mode() == KVM_MODE_NONE) {
2342                 kvm_info("KVM disabled from command line\n");
2343                 return -ENODEV;
2344         }
2345
2346         err = kvm_sys_reg_table_init();
2347         if (err) {
2348                 kvm_info("Error initializing system register tables");
2349                 return err;
2350         }
2351
2352         in_hyp_mode = is_kernel_in_hyp_mode();
2353
2354         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2355             cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2356                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2357                          "Only trusted guests should be used on this system.\n");
2358
2359         err = kvm_set_ipa_limit();
2360         if (err)
2361                 return err;
2362
2363         err = kvm_arm_init_sve();
2364         if (err)
2365                 return err;
2366
2367         err = kvm_arm_vmid_alloc_init();
2368         if (err) {
2369                 kvm_err("Failed to initialize VMID allocator.\n");
2370                 return err;
2371         }
2372
2373         if (!in_hyp_mode) {
2374                 err = init_hyp_mode();
2375                 if (err)
2376                         goto out_err;
2377         }
2378
2379         err = kvm_init_vector_slots();
2380         if (err) {
2381                 kvm_err("Cannot initialise vector slots\n");
2382                 goto out_hyp;
2383         }
2384
2385         err = init_subsystems();
2386         if (err)
2387                 goto out_hyp;
2388
2389         if (is_protected_kvm_enabled()) {
2390                 kvm_info("Protected nVHE mode initialized successfully\n");
2391         } else if (in_hyp_mode) {
2392                 kvm_info("VHE mode initialized successfully\n");
2393         } else {
2394                 kvm_info("Hyp mode initialized successfully\n");
2395         }
2396
2397         /*
2398          * FIXME: Do something reasonable if kvm_init() fails after pKVM
2399          * hypervisor protection is finalized.
2400          */
2401         err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2402         if (err)
2403                 goto out_subs;
2404
2405         return 0;
2406
2407 out_subs:
2408         teardown_subsystems();
2409 out_hyp:
2410         if (!in_hyp_mode)
2411                 teardown_hyp_mode();
2412 out_err:
2413         kvm_arm_vmid_alloc_free();
2414         return err;
2415 }
2416
2417 static int __init early_kvm_mode_cfg(char *arg)
2418 {
2419         if (!arg)
2420                 return -EINVAL;
2421
2422         if (strcmp(arg, "none") == 0) {
2423                 kvm_mode = KVM_MODE_NONE;
2424                 return 0;
2425         }
2426
2427         if (!is_hyp_mode_available()) {
2428                 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n");
2429                 return 0;
2430         }
2431
2432         if (strcmp(arg, "protected") == 0) {
2433                 if (!is_kernel_in_hyp_mode())
2434                         kvm_mode = KVM_MODE_PROTECTED;
2435                 else
2436                         pr_warn_once("Protected KVM not available with VHE\n");
2437
2438                 return 0;
2439         }
2440
2441         if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) {
2442                 kvm_mode = KVM_MODE_DEFAULT;
2443                 return 0;
2444         }
2445
2446         if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) {
2447                 kvm_mode = KVM_MODE_NV;
2448                 return 0;
2449         }
2450
2451         return -EINVAL;
2452 }
2453 early_param("kvm-arm.mode", early_kvm_mode_cfg);
2454
2455 enum kvm_mode kvm_get_mode(void)
2456 {
2457         return kvm_mode;
2458 }
2459
2460 module_init(kvm_arm_init);