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 <trace/events/kvm.h>
  23
  24 #define CREATE_TRACE_POINTS
  25 #include "trace_arm.h"
  26
  27 #include <linux/uaccess.h>
  28 #include <asm/ptrace.h>
  29 #include <asm/mman.h>
  30 #include <asm/tlbflush.h>
  31 #include <asm/cacheflush.h>
  32 #include <asm/cpufeature.h>
  33 #include <asm/virt.h>
  34 #include <asm/kvm_arm.h>
  35 #include <asm/kvm_asm.h>
  36 #include <asm/kvm_mmu.h>
  37 #include <asm/kvm_emulate.h>
  38 #include <asm/kvm_coproc.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 DEFINE_PER_CPU(kvm_host_data_t, kvm_host_data);
  50 static DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page);
  51
  52 /* The VMID used in the VTTBR */
  53 static atomic64_t kvm_vmid_gen = ATOMIC64_INIT(1);
  54 static u32 kvm_next_vmid;
  55 static DEFINE_SPINLOCK(kvm_vmid_lock);
  56
  57 static bool vgic_present;
  58
  59 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled);
  60 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use);
  61
  62 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
  63 {
  64         return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
  65 }
  66
  67 int kvm_arch_hardware_setup(void *opaque)
  68 {
  69         return 0;
  70 }
  71
  72 int kvm_arch_check_processor_compat(void *opaque)
  73 {
  74         return 0;
  75 }
  76
  77 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
  78                             struct kvm_enable_cap *cap)
  79 {
  80         int r;
  81
  82         if (cap->flags)
  83                 return -EINVAL;
  84
  85         switch (cap->cap) {
  86         case KVM_CAP_ARM_NISV_TO_USER:
  87                 r = 0;
  88                 kvm->arch.return_nisv_io_abort_to_user = true;
  89                 break;
  90         default:
  91                 r = -EINVAL;
  92                 break;
  93         }
  94
  95         return r;
  96 }
  97
  98 static int kvm_arm_default_max_vcpus(void)
  99 {
 100         return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
 101 }
 102
 103 /**
 104  * kvm_arch_init_vm - initializes a VM data structure
 105  * @kvm:        pointer to the KVM struct
 106  */
 107 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
 108 {
 109         int ret;
 110
 111         ret = kvm_arm_setup_stage2(kvm, type);
 112         if (ret)
 113                 return ret;
 114
 115         ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu);
 116         if (ret)
 117                 return ret;
 118
 119         ret = create_hyp_mappings(kvm, kvm + 1, PAGE_HYP);
 120         if (ret)
 121                 goto out_free_stage2_pgd;
 122
 123         kvm_vgic_early_init(kvm);
 124
 125         /* The maximum number of VCPUs is limited by the host's GIC model */
 126         kvm->arch.max_vcpus = kvm_arm_default_max_vcpus();
 127
 128         return ret;
 129 out_free_stage2_pgd:
 130         kvm_free_stage2_pgd(&kvm->arch.mmu);
 131         return ret;
 132 }
 133
 134 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
 135 {
 136         return VM_FAULT_SIGBUS;
 137 }
 138
 139
 140 /**
 141  * kvm_arch_destroy_vm - destroy the VM data structure
 142  * @kvm:        pointer to the KVM struct
 143  */
 144 void kvm_arch_destroy_vm(struct kvm *kvm)
 145 {
 146         int i;
 147
 148         kvm_vgic_destroy(kvm);
 149
 150         for (i = 0; i < KVM_MAX_VCPUS; ++i) {
 151                 if (kvm->vcpus[i]) {
 152                         kvm_vcpu_destroy(kvm->vcpus[i]);
 153                         kvm->vcpus[i] = NULL;
 154                 }
 155         }
 156         atomic_set(&kvm->online_vcpus, 0);
 157 }
 158
 159 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 160 {
 161         int r;
 162         switch (ext) {
 163         case KVM_CAP_IRQCHIP:
 164                 r = vgic_present;
 165                 break;
 166         case KVM_CAP_IOEVENTFD:
 167         case KVM_CAP_DEVICE_CTRL:
 168         case KVM_CAP_USER_MEMORY:
 169         case KVM_CAP_SYNC_MMU:
 170         case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
 171         case KVM_CAP_ONE_REG:
 172         case KVM_CAP_ARM_PSCI:
 173         case KVM_CAP_ARM_PSCI_0_2:
 174         case KVM_CAP_READONLY_MEM:
 175         case KVM_CAP_MP_STATE:
 176         case KVM_CAP_IMMEDIATE_EXIT:
 177         case KVM_CAP_VCPU_EVENTS:
 178         case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
 179         case KVM_CAP_ARM_NISV_TO_USER:
 180         case KVM_CAP_ARM_INJECT_EXT_DABT:
 181                 r = 1;
 182                 break;
 183         case KVM_CAP_ARM_SET_DEVICE_ADDR:
 184                 r = 1;
 185                 break;
 186         case KVM_CAP_NR_VCPUS:
 187                 r = num_online_cpus();
 188                 break;
 189         case KVM_CAP_MAX_VCPUS:
 190         case KVM_CAP_MAX_VCPU_ID:
 191                 if (kvm)
 192                         r = kvm->arch.max_vcpus;
 193                 else
 194                         r = kvm_arm_default_max_vcpus();
 195                 break;
 196         case KVM_CAP_MSI_DEVID:
 197                 if (!kvm)
 198                         r = -EINVAL;
 199                 else
 200                         r = kvm->arch.vgic.msis_require_devid;
 201                 break;
 202         case KVM_CAP_ARM_USER_IRQ:
 203                 /*
 204                  * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
 205                  * (bump this number if adding more devices)
 206                  */
 207                 r = 1;
 208                 break;
 209         case KVM_CAP_STEAL_TIME:
 210                 r = kvm_arm_pvtime_supported();
 211                 break;
 212         default:
 213                 r = kvm_arch_vm_ioctl_check_extension(kvm, ext);
 214                 break;
 215         }
 216         return r;
 217 }
 218
 219 long kvm_arch_dev_ioctl(struct file *filp,
 220                         unsigned int ioctl, unsigned long arg)
 221 {
 222         return -EINVAL;
 223 }
 224
 225 struct kvm *kvm_arch_alloc_vm(void)
 226 {
 227         if (!has_vhe())
 228                 return kzalloc(sizeof(struct kvm), GFP_KERNEL);
 229
 230         return vzalloc(sizeof(struct kvm));
 231 }
 232
 233 void kvm_arch_free_vm(struct kvm *kvm)
 234 {
 235         if (!has_vhe())
 236                 kfree(kvm);
 237         else
 238                 vfree(kvm);
 239 }
 240
 241 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
 242 {
 243         if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
 244                 return -EBUSY;
 245
 246         if (id >= kvm->arch.max_vcpus)
 247                 return -EINVAL;
 248
 249         return 0;
 250 }
 251
 252 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
 253 {
 254         int err;
 255
 256         /* Force users to call KVM_ARM_VCPU_INIT */
 257         vcpu->arch.target = -1;
 258         bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 259
 260         vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
 261
 262         /* Set up the timer */
 263         kvm_timer_vcpu_init(vcpu);
 264
 265         kvm_pmu_vcpu_init(vcpu);
 266
 267         kvm_arm_reset_debug_ptr(vcpu);
 268
 269         kvm_arm_pvtime_vcpu_init(&vcpu->arch);
 270
 271         vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
 272
 273         err = kvm_vgic_vcpu_init(vcpu);
 274         if (err)
 275                 return err;
 276
 277         return create_hyp_mappings(vcpu, vcpu + 1, PAGE_HYP);
 278 }
 279
 280 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
 281 {
 282 }
 283
 284 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
 285 {
 286         if (vcpu->arch.has_run_once && unlikely(!irqchip_in_kernel(vcpu->kvm)))
 287                 static_branch_dec(&userspace_irqchip_in_use);
 288
 289         kvm_mmu_free_memory_caches(vcpu);
 290         kvm_timer_vcpu_terminate(vcpu);
 291         kvm_pmu_vcpu_destroy(vcpu);
 292
 293         kvm_arm_vcpu_destroy(vcpu);
 294 }
 295
 296 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu)
 297 {
 298         return kvm_timer_is_pending(vcpu);
 299 }
 300
 301 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
 302 {
 303         /*
 304          * If we're about to block (most likely because we've just hit a
 305          * WFI), we need to sync back the state of the GIC CPU interface
 306          * so that we have the latest PMR and group enables. This ensures
 307          * that kvm_arch_vcpu_runnable has up-to-date data to decide
 308          * whether we have pending interrupts.
 309          *
 310          * For the same reason, we want to tell GICv4 that we need
 311          * doorbells to be signalled, should an interrupt become pending.
 312          */
 313         preempt_disable();
 314         kvm_vgic_vmcr_sync(vcpu);
 315         vgic_v4_put(vcpu, true);
 316         preempt_enable();
 317 }
 318
 319 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
 320 {
 321         preempt_disable();
 322         vgic_v4_load(vcpu);
 323         preempt_enable();
 324 }
 325
 326 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
 327 {
 328         struct kvm_s2_mmu *mmu;
 329         int *last_ran;
 330
 331         mmu = vcpu->arch.hw_mmu;
 332         last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
 333
 334         /*
 335          * We might get preempted before the vCPU actually runs, but
 336          * over-invalidation doesn't affect correctness.
 337          */
 338         if (*last_ran != vcpu->vcpu_id) {
 339                 kvm_call_hyp(__kvm_tlb_flush_local_vmid, mmu);
 340                 *last_ran = vcpu->vcpu_id;
 341         }
 342
 343         vcpu->cpu = cpu;
 344
 345         kvm_vgic_load(vcpu);
 346         kvm_timer_vcpu_load(vcpu);
 347         if (has_vhe())
 348                 kvm_vcpu_load_sysregs_vhe(vcpu);
 349         kvm_arch_vcpu_load_fp(vcpu);
 350         kvm_vcpu_pmu_restore_guest(vcpu);
 351         if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
 352                 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
 353
 354         if (single_task_running())
 355                 vcpu_clear_wfx_traps(vcpu);
 356         else
 357                 vcpu_set_wfx_traps(vcpu);
 358
 359         if (vcpu_has_ptrauth(vcpu))
 360                 vcpu_ptrauth_disable(vcpu);
 361 }
 362
 363 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
 364 {
 365         kvm_arch_vcpu_put_fp(vcpu);
 366         if (has_vhe())
 367                 kvm_vcpu_put_sysregs_vhe(vcpu);
 368         kvm_timer_vcpu_put(vcpu);
 369         kvm_vgic_put(vcpu);
 370         kvm_vcpu_pmu_restore_host(vcpu);
 371
 372         vcpu->cpu = -1;
 373 }
 374
 375 static void vcpu_power_off(struct kvm_vcpu *vcpu)
 376 {
 377         vcpu->arch.power_off = true;
 378         kvm_make_request(KVM_REQ_SLEEP, vcpu);
 379         kvm_vcpu_kick(vcpu);
 380 }
 381
 382 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
 383                                     struct kvm_mp_state *mp_state)
 384 {
 385         if (vcpu->arch.power_off)
 386                 mp_state->mp_state = KVM_MP_STATE_STOPPED;
 387         else
 388                 mp_state->mp_state = KVM_MP_STATE_RUNNABLE;
 389
 390         return 0;
 391 }
 392
 393 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
 394                                     struct kvm_mp_state *mp_state)
 395 {
 396         int ret = 0;
 397
 398         switch (mp_state->mp_state) {
 399         case KVM_MP_STATE_RUNNABLE:
 400                 vcpu->arch.power_off = false;
 401                 break;
 402         case KVM_MP_STATE_STOPPED:
 403                 vcpu_power_off(vcpu);
 404                 break;
 405         default:
 406                 ret = -EINVAL;
 407         }
 408
 409         return ret;
 410 }
 411
 412 /**
 413  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
 414  * @v:          The VCPU pointer
 415  *
 416  * If the guest CPU is not waiting for interrupts or an interrupt line is
 417  * asserted, the CPU is by definition runnable.
 418  */
 419 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
 420 {
 421         bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF);
 422         return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
 423                 && !v->arch.power_off && !v->arch.pause);
 424 }
 425
 426 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
 427 {
 428         return vcpu_mode_priv(vcpu);
 429 }
 430
 431 /* Just ensure a guest exit from a particular CPU */
 432 static void exit_vm_noop(void *info)
 433 {
 434 }
 435
 436 void force_vm_exit(const cpumask_t *mask)
 437 {
 438         preempt_disable();
 439         smp_call_function_many(mask, exit_vm_noop, NULL, true);
 440         preempt_enable();
 441 }
 442
 443 /**
 444  * need_new_vmid_gen - check that the VMID is still valid
 445  * @vmid: The VMID to check
 446  *
 447  * return true if there is a new generation of VMIDs being used
 448  *
 449  * The hardware supports a limited set of values with the value zero reserved
 450  * for the host, so we check if an assigned value belongs to a previous
 451  * generation, which requires us to assign a new value. If we're the first to
 452  * use a VMID for the new generation, we must flush necessary caches and TLBs
 453  * on all CPUs.
 454  */
 455 static bool need_new_vmid_gen(struct kvm_vmid *vmid)
 456 {
 457         u64 current_vmid_gen = atomic64_read(&kvm_vmid_gen);
 458         smp_rmb(); /* Orders read of kvm_vmid_gen and kvm->arch.vmid */
 459         return unlikely(READ_ONCE(vmid->vmid_gen) != current_vmid_gen);
 460 }
 461
 462 /**
 463  * update_vmid - Update the vmid with a valid VMID for the current generation
 464  * @vmid: The stage-2 VMID information struct
 465  */
 466 static void update_vmid(struct kvm_vmid *vmid)
 467 {
 468         if (!need_new_vmid_gen(vmid))
 469                 return;
 470
 471         spin_lock(&kvm_vmid_lock);
 472
 473         /*
 474          * We need to re-check the vmid_gen here to ensure that if another vcpu
 475          * already allocated a valid vmid for this vm, then this vcpu should
 476          * use the same vmid.
 477          */
 478         if (!need_new_vmid_gen(vmid)) {
 479                 spin_unlock(&kvm_vmid_lock);
 480                 return;
 481         }
 482
 483         /* First user of a new VMID generation? */
 484         if (unlikely(kvm_next_vmid == 0)) {
 485                 atomic64_inc(&kvm_vmid_gen);
 486                 kvm_next_vmid = 1;
 487
 488                 /*
 489                  * On SMP we know no other CPUs can use this CPU's or each
 490                  * other's VMID after force_vm_exit returns since the
 491                  * kvm_vmid_lock blocks them from reentry to the guest.
 492                  */
 493                 force_vm_exit(cpu_all_mask);
 494                 /*
 495                  * Now broadcast TLB + ICACHE invalidation over the inner
 496                  * shareable domain to make sure all data structures are
 497                  * clean.
 498                  */
 499                 kvm_call_hyp(__kvm_flush_vm_context);
 500         }
 501
 502         vmid->vmid = kvm_next_vmid;
 503         kvm_next_vmid++;
 504         kvm_next_vmid &= (1 << kvm_get_vmid_bits()) - 1;
 505
 506         smp_wmb();
 507         WRITE_ONCE(vmid->vmid_gen, atomic64_read(&kvm_vmid_gen));
 508
 509         spin_unlock(&kvm_vmid_lock);
 510 }
 511
 512 static int kvm_vcpu_first_run_init(struct kvm_vcpu *vcpu)
 513 {
 514         struct kvm *kvm = vcpu->kvm;
 515         int ret = 0;
 516
 517         if (likely(vcpu->arch.has_run_once))
 518                 return 0;
 519
 520         if (!kvm_arm_vcpu_is_finalized(vcpu))
 521                 return -EPERM;
 522
 523         vcpu->arch.has_run_once = true;
 524
 525         if (likely(irqchip_in_kernel(kvm))) {
 526                 /*
 527                  * Map the VGIC hardware resources before running a vcpu the
 528                  * first time on this VM.
 529                  */
 530                 if (unlikely(!vgic_ready(kvm))) {
 531                         ret = kvm_vgic_map_resources(kvm);
 532                         if (ret)
 533                                 return ret;
 534                 }
 535         } else {
 536                 /*
 537                  * Tell the rest of the code that there are userspace irqchip
 538                  * VMs in the wild.
 539                  */
 540                 static_branch_inc(&userspace_irqchip_in_use);
 541         }
 542
 543         ret = kvm_timer_enable(vcpu);
 544         if (ret)
 545                 return ret;
 546
 547         ret = kvm_arm_pmu_v3_enable(vcpu);
 548
 549         return ret;
 550 }
 551
 552 bool kvm_arch_intc_initialized(struct kvm *kvm)
 553 {
 554         return vgic_initialized(kvm);
 555 }
 556
 557 void kvm_arm_halt_guest(struct kvm *kvm)
 558 {
 559         int i;
 560         struct kvm_vcpu *vcpu;
 561
 562         kvm_for_each_vcpu(i, vcpu, kvm)
 563                 vcpu->arch.pause = true;
 564         kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
 565 }
 566
 567 void kvm_arm_resume_guest(struct kvm *kvm)
 568 {
 569         int i;
 570         struct kvm_vcpu *vcpu;
 571
 572         kvm_for_each_vcpu(i, vcpu, kvm) {
 573                 vcpu->arch.pause = false;
 574                 rcuwait_wake_up(kvm_arch_vcpu_get_wait(vcpu));
 575         }
 576 }
 577
 578 static void vcpu_req_sleep(struct kvm_vcpu *vcpu)
 579 {
 580         struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
 581
 582         rcuwait_wait_event(wait,
 583                            (!vcpu->arch.power_off) &&(!vcpu->arch.pause),
 584                            TASK_INTERRUPTIBLE);
 585
 586         if (vcpu->arch.power_off || vcpu->arch.pause) {
 587                 /* Awaken to handle a signal, request we sleep again later. */
 588                 kvm_make_request(KVM_REQ_SLEEP, vcpu);
 589         }
 590
 591         /*
 592          * Make sure we will observe a potential reset request if we've
 593          * observed a change to the power state. Pairs with the smp_wmb() in
 594          * kvm_psci_vcpu_on().
 595          */
 596         smp_rmb();
 597 }
 598
 599 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu)
 600 {
 601         return vcpu->arch.target >= 0;
 602 }
 603
 604 static void check_vcpu_requests(struct kvm_vcpu *vcpu)
 605 {
 606         if (kvm_request_pending(vcpu)) {
 607                 if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
 608                         vcpu_req_sleep(vcpu);
 609
 610                 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
 611                         kvm_reset_vcpu(vcpu);
 612
 613                 /*
 614                  * Clear IRQ_PENDING requests that were made to guarantee
 615                  * that a VCPU sees new virtual interrupts.
 616                  */
 617                 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
 618
 619                 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
 620                         kvm_update_stolen_time(vcpu);
 621
 622                 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
 623                         /* The distributor enable bits were changed */
 624                         preempt_disable();
 625                         vgic_v4_put(vcpu, false);
 626                         vgic_v4_load(vcpu);
 627                         preempt_enable();
 628                 }
 629         }
 630 }
 631
 632 /**
 633  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
 634  * @vcpu:       The VCPU pointer
 635  *
 636  * This function is called through the VCPU_RUN ioctl called from user space. It
 637  * will execute VM code in a loop until the time slice for the process is used
 638  * or some emulation is needed from user space in which case the function will
 639  * return with return value 0 and with the kvm_run structure filled in with the
 640  * required data for the requested emulation.
 641  */
 642 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 643 {
 644         struct kvm_run *run = vcpu->run;
 645         int ret;
 646
 647         if (unlikely(!kvm_vcpu_initialized(vcpu)))
 648                 return -ENOEXEC;
 649
 650         ret = kvm_vcpu_first_run_init(vcpu);
 651         if (ret)
 652                 return ret;
 653
 654         if (run->exit_reason == KVM_EXIT_MMIO) {
 655                 ret = kvm_handle_mmio_return(vcpu);
 656                 if (ret)
 657                         return ret;
 658         }
 659
 660         if (run->immediate_exit)
 661                 return -EINTR;
 662
 663         vcpu_load(vcpu);
 664
 665         kvm_sigset_activate(vcpu);
 666
 667         ret = 1;
 668         run->exit_reason = KVM_EXIT_UNKNOWN;
 669         while (ret > 0) {
 670                 /*
 671                  * Check conditions before entering the guest
 672                  */
 673                 cond_resched();
 674
 675                 update_vmid(&vcpu->arch.hw_mmu->vmid);
 676
 677                 check_vcpu_requests(vcpu);
 678
 679                 /*
 680                  * Preparing the interrupts to be injected also
 681                  * involves poking the GIC, which must be done in a
 682                  * non-preemptible context.
 683                  */
 684                 preempt_disable();
 685
 686                 kvm_pmu_flush_hwstate(vcpu);
 687
 688                 local_irq_disable();
 689
 690                 kvm_vgic_flush_hwstate(vcpu);
 691
 692                 /*
 693                  * Exit if we have a signal pending so that we can deliver the
 694                  * signal to user space.
 695                  */
 696                 if (signal_pending(current)) {
 697                         ret = -EINTR;
 698                         run->exit_reason = KVM_EXIT_INTR;
 699                 }
 700
 701                 /*
 702                  * If we're using a userspace irqchip, then check if we need
 703                  * to tell a userspace irqchip about timer or PMU level
 704                  * changes and if so, exit to userspace (the actual level
 705                  * state gets updated in kvm_timer_update_run and
 706                  * kvm_pmu_update_run below).
 707                  */
 708                 if (static_branch_unlikely(&userspace_irqchip_in_use)) {
 709                         if (kvm_timer_should_notify_user(vcpu) ||
 710                             kvm_pmu_should_notify_user(vcpu)) {
 711                                 ret = -EINTR;
 712                                 run->exit_reason = KVM_EXIT_INTR;
 713                         }
 714                 }
 715
 716                 /*
 717                  * Ensure we set mode to IN_GUEST_MODE after we disable
 718                  * interrupts and before the final VCPU requests check.
 719                  * See the comment in kvm_vcpu_exiting_guest_mode() and
 720                  * Documentation/virt/kvm/vcpu-requests.rst
 721                  */
 722                 smp_store_mb(vcpu->mode, IN_GUEST_MODE);
 723
 724                 if (ret <= 0 || need_new_vmid_gen(&vcpu->arch.hw_mmu->vmid) ||
 725                     kvm_request_pending(vcpu)) {
 726                         vcpu->mode = OUTSIDE_GUEST_MODE;
 727                         isb(); /* Ensure work in x_flush_hwstate is committed */
 728                         kvm_pmu_sync_hwstate(vcpu);
 729                         if (static_branch_unlikely(&userspace_irqchip_in_use))
 730                                 kvm_timer_sync_user(vcpu);
 731                         kvm_vgic_sync_hwstate(vcpu);
 732                         local_irq_enable();
 733                         preempt_enable();
 734                         continue;
 735                 }
 736
 737                 kvm_arm_setup_debug(vcpu);
 738
 739                 /**************************************************************
 740                  * Enter the guest
 741                  */
 742                 trace_kvm_entry(*vcpu_pc(vcpu));
 743                 guest_enter_irqoff();
 744
 745                 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
 746
 747                 vcpu->mode = OUTSIDE_GUEST_MODE;
 748                 vcpu->stat.exits++;
 749                 /*
 750                  * Back from guest
 751                  *************************************************************/
 752
 753                 kvm_arm_clear_debug(vcpu);
 754
 755                 /*
 756                  * We must sync the PMU state before the vgic state so
 757                  * that the vgic can properly sample the updated state of the
 758                  * interrupt line.
 759                  */
 760                 kvm_pmu_sync_hwstate(vcpu);
 761
 762                 /*
 763                  * Sync the vgic state before syncing the timer state because
 764                  * the timer code needs to know if the virtual timer
 765                  * interrupts are active.
 766                  */
 767                 kvm_vgic_sync_hwstate(vcpu);
 768
 769                 /*
 770                  * Sync the timer hardware state before enabling interrupts as
 771                  * we don't want vtimer interrupts to race with syncing the
 772                  * timer virtual interrupt state.
 773                  */
 774                 if (static_branch_unlikely(&userspace_irqchip_in_use))
 775                         kvm_timer_sync_user(vcpu);
 776
 777                 kvm_arch_vcpu_ctxsync_fp(vcpu);
 778
 779                 /*
 780                  * We may have taken a host interrupt in HYP mode (ie
 781                  * while executing the guest). This interrupt is still
 782                  * pending, as we haven't serviced it yet!
 783                  *
 784                  * We're now back in SVC mode, with interrupts
 785                  * disabled.  Enabling the interrupts now will have
 786                  * the effect of taking the interrupt again, in SVC
 787                  * mode this time.
 788                  */
 789                 local_irq_enable();
 790
 791                 /*
 792                  * We do local_irq_enable() before calling guest_exit() so
 793                  * that if a timer interrupt hits while running the guest we
 794                  * account that tick as being spent in the guest.  We enable
 795                  * preemption after calling guest_exit() so that if we get
 796                  * preempted we make sure ticks after that is not counted as
 797                  * guest time.
 798                  */
 799                 guest_exit();
 800                 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
 801
 802                 /* Exit types that need handling before we can be preempted */
 803                 handle_exit_early(vcpu, ret);
 804
 805                 preempt_enable();
 806
 807                 ret = handle_exit(vcpu, ret);
 808         }
 809
 810         /* Tell userspace about in-kernel device output levels */
 811         if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
 812                 kvm_timer_update_run(vcpu);
 813                 kvm_pmu_update_run(vcpu);
 814         }
 815
 816         kvm_sigset_deactivate(vcpu);
 817
 818         vcpu_put(vcpu);
 819         return ret;
 820 }
 821
 822 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
 823 {
 824         int bit_index;
 825         bool set;
 826         unsigned long *hcr;
 827
 828         if (number == KVM_ARM_IRQ_CPU_IRQ)
 829                 bit_index = __ffs(HCR_VI);
 830         else /* KVM_ARM_IRQ_CPU_FIQ */
 831                 bit_index = __ffs(HCR_VF);
 832
 833         hcr = vcpu_hcr(vcpu);
 834         if (level)
 835                 set = test_and_set_bit(bit_index, hcr);
 836         else
 837                 set = test_and_clear_bit(bit_index, hcr);
 838
 839         /*
 840          * If we didn't change anything, no need to wake up or kick other CPUs
 841          */
 842         if (set == level)
 843                 return 0;
 844
 845         /*
 846          * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
 847          * trigger a world-switch round on the running physical CPU to set the
 848          * virtual IRQ/FIQ fields in the HCR appropriately.
 849          */
 850         kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
 851         kvm_vcpu_kick(vcpu);
 852
 853         return 0;
 854 }
 855
 856 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
 857                           bool line_status)
 858 {
 859         u32 irq = irq_level->irq;
 860         unsigned int irq_type, vcpu_idx, irq_num;
 861         int nrcpus = atomic_read(&kvm->online_vcpus);
 862         struct kvm_vcpu *vcpu = NULL;
 863         bool level = irq_level->level;
 864
 865         irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
 866         vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
 867         vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
 868         irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
 869
 870         trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level);
 871
 872         switch (irq_type) {
 873         case KVM_ARM_IRQ_TYPE_CPU:
 874                 if (irqchip_in_kernel(kvm))
 875                         return -ENXIO;
 876
 877                 if (vcpu_idx >= nrcpus)
 878                         return -EINVAL;
 879
 880                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 881                 if (!vcpu)
 882                         return -EINVAL;
 883
 884                 if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
 885                         return -EINVAL;
 886
 887                 return vcpu_interrupt_line(vcpu, irq_num, level);
 888         case KVM_ARM_IRQ_TYPE_PPI:
 889                 if (!irqchip_in_kernel(kvm))
 890                         return -ENXIO;
 891
 892                 if (vcpu_idx >= nrcpus)
 893                         return -EINVAL;
 894
 895                 vcpu = kvm_get_vcpu(kvm, vcpu_idx);
 896                 if (!vcpu)
 897                         return -EINVAL;
 898
 899                 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS)
 900                         return -EINVAL;
 901
 902                 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL);
 903         case KVM_ARM_IRQ_TYPE_SPI:
 904                 if (!irqchip_in_kernel(kvm))
 905                         return -ENXIO;
 906
 907                 if (irq_num < VGIC_NR_PRIVATE_IRQS)
 908                         return -EINVAL;
 909
 910                 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL);
 911         }
 912
 913         return -EINVAL;
 914 }
 915
 916 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
 917                                const struct kvm_vcpu_init *init)
 918 {
 919         unsigned int i, ret;
 920         int phys_target = kvm_target_cpu();
 921
 922         if (init->target != phys_target)
 923                 return -EINVAL;
 924
 925         /*
 926          * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
 927          * use the same target.
 928          */
 929         if (vcpu->arch.target != -1 && vcpu->arch.target != init->target)
 930                 return -EINVAL;
 931
 932         /* -ENOENT for unknown features, -EINVAL for invalid combinations. */
 933         for (i = 0; i < sizeof(init->features) * 8; i++) {
 934                 bool set = (init->features[i / 32] & (1 << (i % 32)));
 935
 936                 if (set && i >= KVM_VCPU_MAX_FEATURES)
 937                         return -ENOENT;
 938
 939                 /*
 940                  * Secondary and subsequent calls to KVM_ARM_VCPU_INIT must
 941                  * use the same feature set.
 942                  */
 943                 if (vcpu->arch.target != -1 && i < KVM_VCPU_MAX_FEATURES &&
 944                     test_bit(i, vcpu->arch.features) != set)
 945                         return -EINVAL;
 946
 947                 if (set)
 948                         set_bit(i, vcpu->arch.features);
 949         }
 950
 951         vcpu->arch.target = phys_target;
 952
 953         /* Now we know what it is, we can reset it. */
 954         ret = kvm_reset_vcpu(vcpu);
 955         if (ret) {
 956                 vcpu->arch.target = -1;
 957                 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES);
 958         }
 959
 960         return ret;
 961 }
 962
 963 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
 964                                          struct kvm_vcpu_init *init)
 965 {
 966         int ret;
 967
 968         ret = kvm_vcpu_set_target(vcpu, init);
 969         if (ret)
 970                 return ret;
 971
 972         /*
 973          * Ensure a rebooted VM will fault in RAM pages and detect if the
 974          * guest MMU is turned off and flush the caches as needed.
 975          *
 976          * S2FWB enforces all memory accesses to RAM being cacheable,
 977          * ensuring that the data side is always coherent. We still
 978          * need to invalidate the I-cache though, as FWB does *not*
 979          * imply CTR_EL0.DIC.
 980          */
 981         if (vcpu->arch.has_run_once) {
 982                 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
 983                         stage2_unmap_vm(vcpu->kvm);
 984                 else
 985                         __flush_icache_all();
 986         }
 987
 988         vcpu_reset_hcr(vcpu);
 989
 990         /*
 991          * Handle the "start in power-off" case.
 992          */
 993         if (test_bit(KVM_ARM_VCPU_POWER_OFF, vcpu->arch.features))
 994                 vcpu_power_off(vcpu);
 995         else
 996                 vcpu->arch.power_off = false;
 997
 998         return 0;
 999 }
1000
1001 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1002                                  struct kvm_device_attr *attr)
1003 {
1004         int ret = -ENXIO;
1005
1006         switch (attr->group) {
1007         default:
1008                 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1009                 break;
1010         }
1011
1012         return ret;
1013 }
1014
1015 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1016                                  struct kvm_device_attr *attr)
1017 {
1018         int ret = -ENXIO;
1019
1020         switch (attr->group) {
1021         default:
1022                 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1023                 break;
1024         }
1025
1026         return ret;
1027 }
1028
1029 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1030                                  struct kvm_device_attr *attr)
1031 {
1032         int ret = -ENXIO;
1033
1034         switch (attr->group) {
1035         default:
1036                 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1037                 break;
1038         }
1039
1040         return ret;
1041 }
1042
1043 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1044                                    struct kvm_vcpu_events *events)
1045 {
1046         memset(events, 0, sizeof(*events));
1047
1048         return __kvm_arm_vcpu_get_events(vcpu, events);
1049 }
1050
1051 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1052                                    struct kvm_vcpu_events *events)
1053 {
1054         int i;
1055
1056         /* check whether the reserved field is zero */
1057         for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1058                 if (events->reserved[i])
1059                         return -EINVAL;
1060
1061         /* check whether the pad field is zero */
1062         for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1063                 if (events->exception.pad[i])
1064                         return -EINVAL;
1065
1066         return __kvm_arm_vcpu_set_events(vcpu, events);
1067 }
1068
1069 long kvm_arch_vcpu_ioctl(struct file *filp,
1070                          unsigned int ioctl, unsigned long arg)
1071 {
1072         struct kvm_vcpu *vcpu = filp->private_data;
1073         void __user *argp = (void __user *)arg;
1074         struct kvm_device_attr attr;
1075         long r;
1076
1077         switch (ioctl) {
1078         case KVM_ARM_VCPU_INIT: {
1079                 struct kvm_vcpu_init init;
1080
1081                 r = -EFAULT;
1082                 if (copy_from_user(&init, argp, sizeof(init)))
1083                         break;
1084
1085                 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1086                 break;
1087         }
1088         case KVM_SET_ONE_REG:
1089         case KVM_GET_ONE_REG: {
1090                 struct kvm_one_reg reg;
1091
1092                 r = -ENOEXEC;
1093                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1094                         break;
1095
1096                 r = -EFAULT;
1097                 if (copy_from_user(&reg, argp, sizeof(reg)))
1098                         break;
1099
1100                 if (ioctl == KVM_SET_ONE_REG)
1101                         r = kvm_arm_set_reg(vcpu, &reg);
1102                 else
1103                         r = kvm_arm_get_reg(vcpu, &reg);
1104                 break;
1105         }
1106         case KVM_GET_REG_LIST: {
1107                 struct kvm_reg_list __user *user_list = argp;
1108                 struct kvm_reg_list reg_list;
1109                 unsigned n;
1110
1111                 r = -ENOEXEC;
1112                 if (unlikely(!kvm_vcpu_initialized(vcpu)))
1113                         break;
1114
1115                 r = -EPERM;
1116                 if (!kvm_arm_vcpu_is_finalized(vcpu))
1117                         break;
1118
1119                 r = -EFAULT;
1120                 if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1121                         break;
1122                 n = reg_list.n;
1123                 reg_list.n = kvm_arm_num_regs(vcpu);
1124                 if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1125                         break;
1126                 r = -E2BIG;
1127                 if (n < reg_list.n)
1128                         break;
1129                 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1130                 break;
1131         }
1132         case KVM_SET_DEVICE_ATTR: {
1133                 r = -EFAULT;
1134                 if (copy_from_user(&attr, argp, sizeof(attr)))
1135                         break;
1136                 r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1137                 break;
1138         }
1139         case KVM_GET_DEVICE_ATTR: {
1140                 r = -EFAULT;
1141                 if (copy_from_user(&attr, argp, sizeof(attr)))
1142                         break;
1143                 r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1144                 break;
1145         }
1146         case KVM_HAS_DEVICE_ATTR: {
1147                 r = -EFAULT;
1148                 if (copy_from_user(&attr, argp, sizeof(attr)))
1149                         break;
1150                 r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1151                 break;
1152         }
1153         case KVM_GET_VCPU_EVENTS: {
1154                 struct kvm_vcpu_events events;
1155
1156                 if (kvm_arm_vcpu_get_events(vcpu, &events))
1157                         return -EINVAL;
1158
1159                 if (copy_to_user(argp, &events, sizeof(events)))
1160                         return -EFAULT;
1161
1162                 return 0;
1163         }
1164         case KVM_SET_VCPU_EVENTS: {
1165                 struct kvm_vcpu_events events;
1166
1167                 if (copy_from_user(&events, argp, sizeof(events)))
1168                         return -EFAULT;
1169
1170                 return kvm_arm_vcpu_set_events(vcpu, &events);
1171         }
1172         case KVM_ARM_VCPU_FINALIZE: {
1173                 int what;
1174
1175                 if (!kvm_vcpu_initialized(vcpu))
1176                         return -ENOEXEC;
1177
1178                 if (get_user(what, (const int __user *)argp))
1179                         return -EFAULT;
1180
1181                 return kvm_arm_vcpu_finalize(vcpu, what);
1182         }
1183         default:
1184                 r = -EINVAL;
1185         }
1186
1187         return r;
1188 }
1189
1190 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1191 {
1192
1193 }
1194
1195 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm,
1196                                         struct kvm_memory_slot *memslot)
1197 {
1198         kvm_flush_remote_tlbs(kvm);
1199 }
1200
1201 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1202                                         struct kvm_arm_device_addr *dev_addr)
1203 {
1204         unsigned long dev_id, type;
1205
1206         dev_id = (dev_addr->id & KVM_ARM_DEVICE_ID_MASK) >>
1207                 KVM_ARM_DEVICE_ID_SHIFT;
1208         type = (dev_addr->id & KVM_ARM_DEVICE_TYPE_MASK) >>
1209                 KVM_ARM_DEVICE_TYPE_SHIFT;
1210
1211         switch (dev_id) {
1212         case KVM_ARM_DEVICE_VGIC_V2:
1213                 if (!vgic_present)
1214                         return -ENXIO;
1215                 return kvm_vgic_addr(kvm, type, &dev_addr->addr, true);
1216         default:
1217                 return -ENODEV;
1218         }
1219 }
1220
1221 long kvm_arch_vm_ioctl(struct file *filp,
1222                        unsigned int ioctl, unsigned long arg)
1223 {
1224         struct kvm *kvm = filp->private_data;
1225         void __user *argp = (void __user *)arg;
1226
1227         switch (ioctl) {
1228         case KVM_CREATE_IRQCHIP: {
1229                 int ret;
1230                 if (!vgic_present)
1231                         return -ENXIO;
1232                 mutex_lock(&kvm->lock);
1233                 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1234                 mutex_unlock(&kvm->lock);
1235                 return ret;
1236         }
1237         case KVM_ARM_SET_DEVICE_ADDR: {
1238                 struct kvm_arm_device_addr dev_addr;
1239
1240                 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1241                         return -EFAULT;
1242                 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1243         }
1244         case KVM_ARM_PREFERRED_TARGET: {
1245                 int err;
1246                 struct kvm_vcpu_init init;
1247
1248                 err = kvm_vcpu_preferred_target(&init);
1249                 if (err)
1250                         return err;
1251
1252                 if (copy_to_user(argp, &init, sizeof(init)))
1253                         return -EFAULT;
1254
1255                 return 0;
1256         }
1257         default:
1258                 return -EINVAL;
1259         }
1260 }
1261
1262 static void cpu_init_hyp_mode(void)
1263 {
1264         phys_addr_t pgd_ptr;
1265         unsigned long hyp_stack_ptr;
1266         unsigned long vector_ptr;
1267         unsigned long tpidr_el2;
1268
1269         /* Switch from the HYP stub to our own HYP init vector */
1270         __hyp_set_vectors(kvm_get_idmap_vector());
1271
1272         /*
1273          * Calculate the raw per-cpu offset without a translation from the
1274          * kernel's mapping to the linear mapping, and store it in tpidr_el2
1275          * so that we can use adr_l to access per-cpu variables in EL2.
1276          */
1277         tpidr_el2 = ((unsigned long)this_cpu_ptr(&kvm_host_data) -
1278                      (unsigned long)kvm_ksym_ref(&kvm_host_data));
1279
1280         pgd_ptr = kvm_mmu_get_httbr();
1281         hyp_stack_ptr = __this_cpu_read(kvm_arm_hyp_stack_page) + PAGE_SIZE;
1282         vector_ptr = (unsigned long)kvm_get_hyp_vector();
1283
1284         /*
1285          * Call initialization code, and switch to the full blown HYP code.
1286          * If the cpucaps haven't been finalized yet, something has gone very
1287          * wrong, and hyp will crash and burn when it uses any
1288          * cpus_have_const_cap() wrapper.
1289          */
1290         BUG_ON(!system_capabilities_finalized());
1291         __kvm_call_hyp((void *)pgd_ptr, hyp_stack_ptr, vector_ptr, tpidr_el2);
1292
1293         /*
1294          * Disabling SSBD on a non-VHE system requires us to enable SSBS
1295          * at EL2.
1296          */
1297         if (this_cpu_has_cap(ARM64_SSBS) &&
1298             arm64_get_ssbd_state() == ARM64_SSBD_FORCE_DISABLE) {
1299                 kvm_call_hyp_nvhe(__kvm_enable_ssbs);
1300         }
1301 }
1302
1303 static void cpu_hyp_reset(void)
1304 {
1305         if (!is_kernel_in_hyp_mode())
1306                 __hyp_reset_vectors();
1307 }
1308
1309 static void cpu_hyp_reinit(void)
1310 {
1311         kvm_init_host_cpu_context(&this_cpu_ptr(&kvm_host_data)->host_ctxt);
1312
1313         cpu_hyp_reset();
1314
1315         if (is_kernel_in_hyp_mode())
1316                 kvm_timer_init_vhe();
1317         else
1318                 cpu_init_hyp_mode();
1319
1320         kvm_arm_init_debug();
1321
1322         if (vgic_present)
1323                 kvm_vgic_init_cpu_hardware();
1324 }
1325
1326 static void _kvm_arch_hardware_enable(void *discard)
1327 {
1328         if (!__this_cpu_read(kvm_arm_hardware_enabled)) {
1329                 cpu_hyp_reinit();
1330                 __this_cpu_write(kvm_arm_hardware_enabled, 1);
1331         }
1332 }
1333
1334 int kvm_arch_hardware_enable(void)
1335 {
1336         _kvm_arch_hardware_enable(NULL);
1337         return 0;
1338 }
1339
1340 static void _kvm_arch_hardware_disable(void *discard)
1341 {
1342         if (__this_cpu_read(kvm_arm_hardware_enabled)) {
1343                 cpu_hyp_reset();
1344                 __this_cpu_write(kvm_arm_hardware_enabled, 0);
1345         }
1346 }
1347
1348 void kvm_arch_hardware_disable(void)
1349 {
1350         _kvm_arch_hardware_disable(NULL);
1351 }
1352
1353 #ifdef CONFIG_CPU_PM
1354 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
1355                                     unsigned long cmd,
1356                                     void *v)
1357 {
1358         /*
1359          * kvm_arm_hardware_enabled is left with its old value over
1360          * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
1361          * re-enable hyp.
1362          */
1363         switch (cmd) {
1364         case CPU_PM_ENTER:
1365                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1366                         /*
1367                          * don't update kvm_arm_hardware_enabled here
1368                          * so that the hardware will be re-enabled
1369                          * when we resume. See below.
1370                          */
1371                         cpu_hyp_reset();
1372
1373                 return NOTIFY_OK;
1374         case CPU_PM_ENTER_FAILED:
1375         case CPU_PM_EXIT:
1376                 if (__this_cpu_read(kvm_arm_hardware_enabled))
1377                         /* The hardware was enabled before suspend. */
1378                         cpu_hyp_reinit();
1379
1380                 return NOTIFY_OK;
1381
1382         default:
1383                 return NOTIFY_DONE;
1384         }
1385 }
1386
1387 static struct notifier_block hyp_init_cpu_pm_nb = {
1388         .notifier_call = hyp_init_cpu_pm_notifier,
1389 };
1390
1391 static void __init hyp_cpu_pm_init(void)
1392 {
1393         cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
1394 }
1395 static void __init hyp_cpu_pm_exit(void)
1396 {
1397         cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
1398 }
1399 #else
1400 static inline void hyp_cpu_pm_init(void)
1401 {
1402 }
1403 static inline void hyp_cpu_pm_exit(void)
1404 {
1405 }
1406 #endif
1407
1408 static int init_common_resources(void)
1409 {
1410         return kvm_set_ipa_limit();
1411 }
1412
1413 static int init_subsystems(void)
1414 {
1415         int err = 0;
1416
1417         /*
1418          * Enable hardware so that subsystem initialisation can access EL2.
1419          */
1420         on_each_cpu(_kvm_arch_hardware_enable, NULL, 1);
1421
1422         /*
1423          * Register CPU lower-power notifier
1424          */
1425         hyp_cpu_pm_init();
1426
1427         /*
1428          * Init HYP view of VGIC
1429          */
1430         err = kvm_vgic_hyp_init();
1431         switch (err) {
1432         case 0:
1433                 vgic_present = true;
1434                 break;
1435         case -ENODEV:
1436         case -ENXIO:
1437                 vgic_present = false;
1438                 err = 0;
1439                 break;
1440         default:
1441                 goto out;
1442         }
1443
1444         /*
1445          * Init HYP architected timer support
1446          */
1447         err = kvm_timer_hyp_init(vgic_present);
1448         if (err)
1449                 goto out;
1450
1451         kvm_perf_init();
1452         kvm_coproc_table_init();
1453
1454 out:
1455         on_each_cpu(_kvm_arch_hardware_disable, NULL, 1);
1456
1457         return err;
1458 }
1459
1460 static void teardown_hyp_mode(void)
1461 {
1462         int cpu;
1463
1464         free_hyp_pgds();
1465         for_each_possible_cpu(cpu)
1466                 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu));
1467 }
1468
1469 /**
1470  * Inits Hyp-mode on all online CPUs
1471  */
1472 static int init_hyp_mode(void)
1473 {
1474         int cpu;
1475         int err = 0;
1476
1477         /*
1478          * Allocate Hyp PGD and setup Hyp identity mapping
1479          */
1480         err = kvm_mmu_init();
1481         if (err)
1482                 goto out_err;
1483
1484         /*
1485          * Allocate stack pages for Hypervisor-mode
1486          */
1487         for_each_possible_cpu(cpu) {
1488                 unsigned long stack_page;
1489
1490                 stack_page = __get_free_page(GFP_KERNEL);
1491                 if (!stack_page) {
1492                         err = -ENOMEM;
1493                         goto out_err;
1494                 }
1495
1496                 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page;
1497         }
1498
1499         /*
1500          * Map the Hyp-code called directly from the host
1501          */
1502         err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
1503                                   kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
1504         if (err) {
1505                 kvm_err("Cannot map world-switch code\n");
1506                 goto out_err;
1507         }
1508
1509         err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
1510                                   kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
1511         if (err) {
1512                 kvm_err("Cannot map rodata section\n");
1513                 goto out_err;
1514         }
1515
1516         err = create_hyp_mappings(kvm_ksym_ref(__bss_start),
1517                                   kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
1518         if (err) {
1519                 kvm_err("Cannot map bss section\n");
1520                 goto out_err;
1521         }
1522
1523         err = kvm_map_vectors();
1524         if (err) {
1525                 kvm_err("Cannot map vectors\n");
1526                 goto out_err;
1527         }
1528
1529         /*
1530          * Map the Hyp stack pages
1531          */
1532         for_each_possible_cpu(cpu) {
1533                 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu);
1534                 err = create_hyp_mappings(stack_page, stack_page + PAGE_SIZE,
1535                                           PAGE_HYP);
1536
1537                 if (err) {
1538                         kvm_err("Cannot map hyp stack\n");
1539                         goto out_err;
1540                 }
1541         }
1542
1543         for_each_possible_cpu(cpu) {
1544                 kvm_host_data_t *cpu_data;
1545
1546                 cpu_data = per_cpu_ptr(&kvm_host_data, cpu);
1547                 err = create_hyp_mappings(cpu_data, cpu_data + 1, PAGE_HYP);
1548
1549                 if (err) {
1550                         kvm_err("Cannot map host CPU state: %d\n", err);
1551                         goto out_err;
1552                 }
1553         }
1554
1555         err = hyp_map_aux_data();
1556         if (err)
1557                 kvm_err("Cannot map host auxiliary data: %d\n", err);
1558
1559         return 0;
1560
1561 out_err:
1562         teardown_hyp_mode();
1563         kvm_err("error initializing Hyp mode: %d\n", err);
1564         return err;
1565 }
1566
1567 static void check_kvm_target_cpu(void *ret)
1568 {
1569         *(int *)ret = kvm_target_cpu();
1570 }
1571
1572 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
1573 {
1574         struct kvm_vcpu *vcpu;
1575         int i;
1576
1577         mpidr &= MPIDR_HWID_BITMASK;
1578         kvm_for_each_vcpu(i, vcpu, kvm) {
1579                 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
1580                         return vcpu;
1581         }
1582         return NULL;
1583 }
1584
1585 bool kvm_arch_has_irq_bypass(void)
1586 {
1587         return true;
1588 }
1589
1590 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
1591                                       struct irq_bypass_producer *prod)
1592 {
1593         struct kvm_kernel_irqfd *irqfd =
1594                 container_of(cons, struct kvm_kernel_irqfd, consumer);
1595
1596         return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
1597                                           &irqfd->irq_entry);
1598 }
1599 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
1600                                       struct irq_bypass_producer *prod)
1601 {
1602         struct kvm_kernel_irqfd *irqfd =
1603                 container_of(cons, struct kvm_kernel_irqfd, consumer);
1604
1605         kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq,
1606                                      &irqfd->irq_entry);
1607 }
1608
1609 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
1610 {
1611         struct kvm_kernel_irqfd *irqfd =
1612                 container_of(cons, struct kvm_kernel_irqfd, consumer);
1613
1614         kvm_arm_halt_guest(irqfd->kvm);
1615 }
1616
1617 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
1618 {
1619         struct kvm_kernel_irqfd *irqfd =
1620                 container_of(cons, struct kvm_kernel_irqfd, consumer);
1621
1622         kvm_arm_resume_guest(irqfd->kvm);
1623 }
1624
1625 /**
1626  * Initialize Hyp-mode and memory mappings on all CPUs.
1627  */
1628 int kvm_arch_init(void *opaque)
1629 {
1630         int err;
1631         int ret, cpu;
1632         bool in_hyp_mode;
1633
1634         if (!is_hyp_mode_available()) {
1635                 kvm_info("HYP mode not available\n");
1636                 return -ENODEV;
1637         }
1638
1639         in_hyp_mode = is_kernel_in_hyp_mode();
1640
1641         if (!in_hyp_mode && kvm_arch_requires_vhe()) {
1642                 kvm_pr_unimpl("CPU unsupported in non-VHE mode, not initializing\n");
1643                 return -ENODEV;
1644         }
1645
1646         if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE))
1647                 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
1648                          "Only trusted guests should be used on this system.\n");
1649
1650         for_each_online_cpu(cpu) {
1651                 smp_call_function_single(cpu, check_kvm_target_cpu, &ret, 1);
1652                 if (ret < 0) {
1653                         kvm_err("Error, CPU %d not supported!\n", cpu);
1654                         return -ENODEV;
1655                 }
1656         }
1657
1658         err = init_common_resources();
1659         if (err)
1660                 return err;
1661
1662         err = kvm_arm_init_sve();
1663         if (err)
1664                 return err;
1665
1666         if (!in_hyp_mode) {
1667                 err = init_hyp_mode();
1668                 if (err)
1669                         goto out_err;
1670         }
1671
1672         err = init_subsystems();
1673         if (err)
1674                 goto out_hyp;
1675
1676         if (in_hyp_mode)
1677                 kvm_info("VHE mode initialized successfully\n");
1678         else
1679                 kvm_info("Hyp mode initialized successfully\n");
1680
1681         return 0;
1682
1683 out_hyp:
1684         hyp_cpu_pm_exit();
1685         if (!in_hyp_mode)
1686                 teardown_hyp_mode();
1687 out_err:
1688         return err;
1689 }
1690
1691 /* NOP: Compiling as a module not supported */
1692 void kvm_arch_exit(void)
1693 {
1694         kvm_perf_teardown();
1695 }
1696
1697 static int arm_init(void)
1698 {
1699         int rc = kvm_init(NULL, sizeof(struct kvm_vcpu), 0, THIS_MODULE);
1700         return rc;
1701 }
1702
1703 module_init(arm_init);