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