1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University 4 * Author: Christoffer Dall <c.dall@virtualopensystems.com> 5 */ 6 7 #include <linux/bug.h> 8 #include <linux/cpu_pm.h> 9 #include <linux/entry-kvm.h> 10 #include <linux/errno.h> 11 #include <linux/err.h> 12 #include <linux/kvm_host.h> 13 #include <linux/list.h> 14 #include <linux/module.h> 15 #include <linux/vmalloc.h> 16 #include <linux/fs.h> 17 #include <linux/mman.h> 18 #include <linux/sched.h> 19 #include <linux/kvm.h> 20 #include <linux/kvm_irqfd.h> 21 #include <linux/irqbypass.h> 22 #include <linux/sched/stat.h> 23 #include <linux/psci.h> 24 #include <trace/events/kvm.h> 25 26 #define CREATE_TRACE_POINTS 27 #include "trace_arm.h" 28 29 #include <linux/uaccess.h> 30 #include <asm/ptrace.h> 31 #include <asm/mman.h> 32 #include <asm/tlbflush.h> 33 #include <asm/cacheflush.h> 34 #include <asm/cpufeature.h> 35 #include <asm/virt.h> 36 #include <asm/kvm_arm.h> 37 #include <asm/kvm_asm.h> 38 #include <asm/kvm_mmu.h> 39 #include <asm/kvm_nested.h> 40 #include <asm/kvm_pkvm.h> 41 #include <asm/kvm_emulate.h> 42 #include <asm/sections.h> 43 44 #include <kvm/arm_hypercalls.h> 45 #include <kvm/arm_pmu.h> 46 #include <kvm/arm_psci.h> 47 48 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT; 49 50 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector); 51 52 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page); 53 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params); 54 55 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt); 56 57 static bool vgic_present, kvm_arm_initialised; 58 59 static DEFINE_PER_CPU(unsigned char, kvm_hyp_initialized); 60 61 bool is_kvm_arm_initialised(void) 62 { 63 return kvm_arm_initialised; 64 } 65 66 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu) 67 { 68 return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE; 69 } 70 71 int kvm_vm_ioctl_enable_cap(struct kvm *kvm, 72 struct kvm_enable_cap *cap) 73 { 74 int r; 75 u64 new_cap; 76 77 if (cap->flags) 78 return -EINVAL; 79 80 switch (cap->cap) { 81 case KVM_CAP_ARM_NISV_TO_USER: 82 r = 0; 83 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER, 84 &kvm->arch.flags); 85 break; 86 case KVM_CAP_ARM_MTE: 87 mutex_lock(&kvm->lock); 88 if (!system_supports_mte() || kvm->created_vcpus) { 89 r = -EINVAL; 90 } else { 91 r = 0; 92 set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags); 93 } 94 mutex_unlock(&kvm->lock); 95 break; 96 case KVM_CAP_ARM_SYSTEM_SUSPEND: 97 r = 0; 98 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags); 99 break; 100 case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE: 101 new_cap = cap->args[0]; 102 103 mutex_lock(&kvm->slots_lock); 104 /* 105 * To keep things simple, allow changing the chunk 106 * size only when no memory slots have been created. 107 */ 108 if (!kvm_are_all_memslots_empty(kvm)) { 109 r = -EINVAL; 110 } else if (new_cap && !kvm_is_block_size_supported(new_cap)) { 111 r = -EINVAL; 112 } else { 113 r = 0; 114 kvm->arch.mmu.split_page_chunk_size = new_cap; 115 } 116 mutex_unlock(&kvm->slots_lock); 117 break; 118 default: 119 r = -EINVAL; 120 break; 121 } 122 123 return r; 124 } 125 126 static int kvm_arm_default_max_vcpus(void) 127 { 128 return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS; 129 } 130 131 /** 132 * kvm_arch_init_vm - initializes a VM data structure 133 * @kvm: pointer to the KVM struct 134 */ 135 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type) 136 { 137 int ret; 138 139 mutex_init(&kvm->arch.config_lock); 140 141 #ifdef CONFIG_LOCKDEP 142 /* Clue in lockdep that the config_lock must be taken inside kvm->lock */ 143 mutex_lock(&kvm->lock); 144 mutex_lock(&kvm->arch.config_lock); 145 mutex_unlock(&kvm->arch.config_lock); 146 mutex_unlock(&kvm->lock); 147 #endif 148 149 ret = kvm_share_hyp(kvm, kvm + 1); 150 if (ret) 151 return ret; 152 153 ret = pkvm_init_host_vm(kvm); 154 if (ret) 155 goto err_unshare_kvm; 156 157 if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) { 158 ret = -ENOMEM; 159 goto err_unshare_kvm; 160 } 161 cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask); 162 163 ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type); 164 if (ret) 165 goto err_free_cpumask; 166 167 kvm_vgic_early_init(kvm); 168 169 kvm_timer_init_vm(kvm); 170 171 /* The maximum number of VCPUs is limited by the host's GIC model */ 172 kvm->max_vcpus = kvm_arm_default_max_vcpus(); 173 174 kvm_arm_init_hypercalls(kvm); 175 176 bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES); 177 178 return 0; 179 180 err_free_cpumask: 181 free_cpumask_var(kvm->arch.supported_cpus); 182 err_unshare_kvm: 183 kvm_unshare_hyp(kvm, kvm + 1); 184 return ret; 185 } 186 187 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf) 188 { 189 return VM_FAULT_SIGBUS; 190 } 191 192 193 /** 194 * kvm_arch_destroy_vm - destroy the VM data structure 195 * @kvm: pointer to the KVM struct 196 */ 197 void kvm_arch_destroy_vm(struct kvm *kvm) 198 { 199 bitmap_free(kvm->arch.pmu_filter); 200 free_cpumask_var(kvm->arch.supported_cpus); 201 202 kvm_vgic_destroy(kvm); 203 204 if (is_protected_kvm_enabled()) 205 pkvm_destroy_hyp_vm(kvm); 206 207 kvm_destroy_vcpus(kvm); 208 209 kvm_unshare_hyp(kvm, kvm + 1); 210 211 kvm_arm_teardown_hypercalls(kvm); 212 } 213 214 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) 215 { 216 int r; 217 switch (ext) { 218 case KVM_CAP_IRQCHIP: 219 r = vgic_present; 220 break; 221 case KVM_CAP_IOEVENTFD: 222 case KVM_CAP_DEVICE_CTRL: 223 case KVM_CAP_USER_MEMORY: 224 case KVM_CAP_SYNC_MMU: 225 case KVM_CAP_DESTROY_MEMORY_REGION_WORKS: 226 case KVM_CAP_ONE_REG: 227 case KVM_CAP_ARM_PSCI: 228 case KVM_CAP_ARM_PSCI_0_2: 229 case KVM_CAP_READONLY_MEM: 230 case KVM_CAP_MP_STATE: 231 case KVM_CAP_IMMEDIATE_EXIT: 232 case KVM_CAP_VCPU_EVENTS: 233 case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2: 234 case KVM_CAP_ARM_NISV_TO_USER: 235 case KVM_CAP_ARM_INJECT_EXT_DABT: 236 case KVM_CAP_SET_GUEST_DEBUG: 237 case KVM_CAP_VCPU_ATTRIBUTES: 238 case KVM_CAP_PTP_KVM: 239 case KVM_CAP_ARM_SYSTEM_SUSPEND: 240 case KVM_CAP_IRQFD_RESAMPLE: 241 case KVM_CAP_COUNTER_OFFSET: 242 r = 1; 243 break; 244 case KVM_CAP_SET_GUEST_DEBUG2: 245 return KVM_GUESTDBG_VALID_MASK; 246 case KVM_CAP_ARM_SET_DEVICE_ADDR: 247 r = 1; 248 break; 249 case KVM_CAP_NR_VCPUS: 250 /* 251 * ARM64 treats KVM_CAP_NR_CPUS differently from all other 252 * architectures, as it does not always bound it to 253 * KVM_CAP_MAX_VCPUS. It should not matter much because 254 * this is just an advisory value. 255 */ 256 r = min_t(unsigned int, num_online_cpus(), 257 kvm_arm_default_max_vcpus()); 258 break; 259 case KVM_CAP_MAX_VCPUS: 260 case KVM_CAP_MAX_VCPU_ID: 261 if (kvm) 262 r = kvm->max_vcpus; 263 else 264 r = kvm_arm_default_max_vcpus(); 265 break; 266 case KVM_CAP_MSI_DEVID: 267 if (!kvm) 268 r = -EINVAL; 269 else 270 r = kvm->arch.vgic.msis_require_devid; 271 break; 272 case KVM_CAP_ARM_USER_IRQ: 273 /* 274 * 1: EL1_VTIMER, EL1_PTIMER, and PMU. 275 * (bump this number if adding more devices) 276 */ 277 r = 1; 278 break; 279 case KVM_CAP_ARM_MTE: 280 r = system_supports_mte(); 281 break; 282 case KVM_CAP_STEAL_TIME: 283 r = kvm_arm_pvtime_supported(); 284 break; 285 case KVM_CAP_ARM_EL1_32BIT: 286 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1); 287 break; 288 case KVM_CAP_GUEST_DEBUG_HW_BPS: 289 r = get_num_brps(); 290 break; 291 case KVM_CAP_GUEST_DEBUG_HW_WPS: 292 r = get_num_wrps(); 293 break; 294 case KVM_CAP_ARM_PMU_V3: 295 r = kvm_arm_support_pmu_v3(); 296 break; 297 case KVM_CAP_ARM_INJECT_SERROR_ESR: 298 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN); 299 break; 300 case KVM_CAP_ARM_VM_IPA_SIZE: 301 r = get_kvm_ipa_limit(); 302 break; 303 case KVM_CAP_ARM_SVE: 304 r = system_supports_sve(); 305 break; 306 case KVM_CAP_ARM_PTRAUTH_ADDRESS: 307 case KVM_CAP_ARM_PTRAUTH_GENERIC: 308 r = system_has_full_ptr_auth(); 309 break; 310 case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE: 311 if (kvm) 312 r = kvm->arch.mmu.split_page_chunk_size; 313 else 314 r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT; 315 break; 316 case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES: 317 r = kvm_supported_block_sizes(); 318 break; 319 default: 320 r = 0; 321 } 322 323 return r; 324 } 325 326 long kvm_arch_dev_ioctl(struct file *filp, 327 unsigned int ioctl, unsigned long arg) 328 { 329 return -EINVAL; 330 } 331 332 struct kvm *kvm_arch_alloc_vm(void) 333 { 334 size_t sz = sizeof(struct kvm); 335 336 if (!has_vhe()) 337 return kzalloc(sz, GFP_KERNEL_ACCOUNT); 338 339 return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO); 340 } 341 342 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id) 343 { 344 if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) 345 return -EBUSY; 346 347 if (id >= kvm->max_vcpus) 348 return -EINVAL; 349 350 return 0; 351 } 352 353 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu) 354 { 355 int err; 356 357 spin_lock_init(&vcpu->arch.mp_state_lock); 358 359 #ifdef CONFIG_LOCKDEP 360 /* Inform lockdep that the config_lock is acquired after vcpu->mutex */ 361 mutex_lock(&vcpu->mutex); 362 mutex_lock(&vcpu->kvm->arch.config_lock); 363 mutex_unlock(&vcpu->kvm->arch.config_lock); 364 mutex_unlock(&vcpu->mutex); 365 #endif 366 367 /* Force users to call KVM_ARM_VCPU_INIT */ 368 vcpu_clear_flag(vcpu, VCPU_INITIALIZED); 369 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES); 370 371 vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO; 372 373 /* 374 * Default value for the FP state, will be overloaded at load 375 * time if we support FP (pretty likely) 376 */ 377 vcpu->arch.fp_state = FP_STATE_FREE; 378 379 /* Set up the timer */ 380 kvm_timer_vcpu_init(vcpu); 381 382 kvm_pmu_vcpu_init(vcpu); 383 384 kvm_arm_reset_debug_ptr(vcpu); 385 386 kvm_arm_pvtime_vcpu_init(&vcpu->arch); 387 388 vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu; 389 390 err = kvm_vgic_vcpu_init(vcpu); 391 if (err) 392 return err; 393 394 return kvm_share_hyp(vcpu, vcpu + 1); 395 } 396 397 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu) 398 { 399 } 400 401 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu) 402 { 403 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache); 404 kvm_timer_vcpu_terminate(vcpu); 405 kvm_pmu_vcpu_destroy(vcpu); 406 kvm_vgic_vcpu_destroy(vcpu); 407 kvm_arm_vcpu_destroy(vcpu); 408 } 409 410 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu) 411 { 412 413 } 414 415 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu) 416 { 417 418 } 419 420 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) 421 { 422 struct kvm_s2_mmu *mmu; 423 int *last_ran; 424 425 mmu = vcpu->arch.hw_mmu; 426 last_ran = this_cpu_ptr(mmu->last_vcpu_ran); 427 428 /* 429 * We guarantee that both TLBs and I-cache are private to each 430 * vcpu. If detecting that a vcpu from the same VM has 431 * previously run on the same physical CPU, call into the 432 * hypervisor code to nuke the relevant contexts. 433 * 434 * We might get preempted before the vCPU actually runs, but 435 * over-invalidation doesn't affect correctness. 436 */ 437 if (*last_ran != vcpu->vcpu_id) { 438 kvm_call_hyp(__kvm_flush_cpu_context, mmu); 439 *last_ran = vcpu->vcpu_id; 440 } 441 442 vcpu->cpu = cpu; 443 444 kvm_vgic_load(vcpu); 445 kvm_timer_vcpu_load(vcpu); 446 if (has_vhe()) 447 kvm_vcpu_load_sysregs_vhe(vcpu); 448 kvm_arch_vcpu_load_fp(vcpu); 449 kvm_vcpu_pmu_restore_guest(vcpu); 450 if (kvm_arm_is_pvtime_enabled(&vcpu->arch)) 451 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu); 452 453 if (single_task_running()) 454 vcpu_clear_wfx_traps(vcpu); 455 else 456 vcpu_set_wfx_traps(vcpu); 457 458 if (vcpu_has_ptrauth(vcpu)) 459 vcpu_ptrauth_disable(vcpu); 460 kvm_arch_vcpu_load_debug_state_flags(vcpu); 461 462 if (!cpumask_test_cpu(cpu, vcpu->kvm->arch.supported_cpus)) 463 vcpu_set_on_unsupported_cpu(vcpu); 464 } 465 466 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu) 467 { 468 kvm_arch_vcpu_put_debug_state_flags(vcpu); 469 kvm_arch_vcpu_put_fp(vcpu); 470 if (has_vhe()) 471 kvm_vcpu_put_sysregs_vhe(vcpu); 472 kvm_timer_vcpu_put(vcpu); 473 kvm_vgic_put(vcpu); 474 kvm_vcpu_pmu_restore_host(vcpu); 475 kvm_arm_vmid_clear_active(); 476 477 vcpu_clear_on_unsupported_cpu(vcpu); 478 vcpu->cpu = -1; 479 } 480 481 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 482 { 483 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED); 484 kvm_make_request(KVM_REQ_SLEEP, vcpu); 485 kvm_vcpu_kick(vcpu); 486 } 487 488 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 489 { 490 spin_lock(&vcpu->arch.mp_state_lock); 491 __kvm_arm_vcpu_power_off(vcpu); 492 spin_unlock(&vcpu->arch.mp_state_lock); 493 } 494 495 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu) 496 { 497 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED; 498 } 499 500 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu) 501 { 502 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED); 503 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 504 kvm_vcpu_kick(vcpu); 505 } 506 507 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu) 508 { 509 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED; 510 } 511 512 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu, 513 struct kvm_mp_state *mp_state) 514 { 515 *mp_state = READ_ONCE(vcpu->arch.mp_state); 516 517 return 0; 518 } 519 520 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu, 521 struct kvm_mp_state *mp_state) 522 { 523 int ret = 0; 524 525 spin_lock(&vcpu->arch.mp_state_lock); 526 527 switch (mp_state->mp_state) { 528 case KVM_MP_STATE_RUNNABLE: 529 WRITE_ONCE(vcpu->arch.mp_state, *mp_state); 530 break; 531 case KVM_MP_STATE_STOPPED: 532 __kvm_arm_vcpu_power_off(vcpu); 533 break; 534 case KVM_MP_STATE_SUSPENDED: 535 kvm_arm_vcpu_suspend(vcpu); 536 break; 537 default: 538 ret = -EINVAL; 539 } 540 541 spin_unlock(&vcpu->arch.mp_state_lock); 542 543 return ret; 544 } 545 546 /** 547 * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled 548 * @v: The VCPU pointer 549 * 550 * If the guest CPU is not waiting for interrupts or an interrupt line is 551 * asserted, the CPU is by definition runnable. 552 */ 553 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v) 554 { 555 bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF); 556 return ((irq_lines || kvm_vgic_vcpu_pending_irq(v)) 557 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause); 558 } 559 560 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu) 561 { 562 return vcpu_mode_priv(vcpu); 563 } 564 565 #ifdef CONFIG_GUEST_PERF_EVENTS 566 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu) 567 { 568 return *vcpu_pc(vcpu); 569 } 570 #endif 571 572 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu) 573 { 574 return vcpu_get_flag(vcpu, VCPU_INITIALIZED); 575 } 576 577 /* 578 * Handle both the initialisation that is being done when the vcpu is 579 * run for the first time, as well as the updates that must be 580 * performed each time we get a new thread dealing with this vcpu. 581 */ 582 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu) 583 { 584 struct kvm *kvm = vcpu->kvm; 585 int ret; 586 587 if (!kvm_vcpu_initialized(vcpu)) 588 return -ENOEXEC; 589 590 if (!kvm_arm_vcpu_is_finalized(vcpu)) 591 return -EPERM; 592 593 ret = kvm_arch_vcpu_run_map_fp(vcpu); 594 if (ret) 595 return ret; 596 597 if (likely(vcpu_has_run_once(vcpu))) 598 return 0; 599 600 kvm_arm_vcpu_init_debug(vcpu); 601 602 if (likely(irqchip_in_kernel(kvm))) { 603 /* 604 * Map the VGIC hardware resources before running a vcpu the 605 * first time on this VM. 606 */ 607 ret = kvm_vgic_map_resources(kvm); 608 if (ret) 609 return ret; 610 } 611 612 ret = kvm_timer_enable(vcpu); 613 if (ret) 614 return ret; 615 616 ret = kvm_arm_pmu_v3_enable(vcpu); 617 if (ret) 618 return ret; 619 620 if (is_protected_kvm_enabled()) { 621 ret = pkvm_create_hyp_vm(kvm); 622 if (ret) 623 return ret; 624 } 625 626 /* 627 * Initialize traps for protected VMs. 628 * NOTE: Move to run in EL2 directly, rather than via a hypercall, once 629 * the code is in place for first run initialization at EL2. 630 */ 631 if (kvm_vm_is_protected(kvm)) 632 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu); 633 634 mutex_lock(&kvm->arch.config_lock); 635 set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags); 636 mutex_unlock(&kvm->arch.config_lock); 637 638 return ret; 639 } 640 641 bool kvm_arch_intc_initialized(struct kvm *kvm) 642 { 643 return vgic_initialized(kvm); 644 } 645 646 void kvm_arm_halt_guest(struct kvm *kvm) 647 { 648 unsigned long i; 649 struct kvm_vcpu *vcpu; 650 651 kvm_for_each_vcpu(i, vcpu, kvm) 652 vcpu->arch.pause = true; 653 kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP); 654 } 655 656 void kvm_arm_resume_guest(struct kvm *kvm) 657 { 658 unsigned long i; 659 struct kvm_vcpu *vcpu; 660 661 kvm_for_each_vcpu(i, vcpu, kvm) { 662 vcpu->arch.pause = false; 663 __kvm_vcpu_wake_up(vcpu); 664 } 665 } 666 667 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu) 668 { 669 struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); 670 671 rcuwait_wait_event(wait, 672 (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause), 673 TASK_INTERRUPTIBLE); 674 675 if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) { 676 /* Awaken to handle a signal, request we sleep again later. */ 677 kvm_make_request(KVM_REQ_SLEEP, vcpu); 678 } 679 680 /* 681 * Make sure we will observe a potential reset request if we've 682 * observed a change to the power state. Pairs with the smp_wmb() in 683 * kvm_psci_vcpu_on(). 684 */ 685 smp_rmb(); 686 } 687 688 /** 689 * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior 690 * @vcpu: The VCPU pointer 691 * 692 * Suspend execution of a vCPU until a valid wake event is detected, i.e. until 693 * the vCPU is runnable. The vCPU may or may not be scheduled out, depending 694 * on when a wake event arrives, e.g. there may already be a pending wake event. 695 */ 696 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu) 697 { 698 /* 699 * Sync back the state of the GIC CPU interface so that we have 700 * the latest PMR and group enables. This ensures that 701 * kvm_arch_vcpu_runnable has up-to-date data to decide whether 702 * we have pending interrupts, e.g. when determining if the 703 * vCPU should block. 704 * 705 * For the same reason, we want to tell GICv4 that we need 706 * doorbells to be signalled, should an interrupt become pending. 707 */ 708 preempt_disable(); 709 kvm_vgic_vmcr_sync(vcpu); 710 vcpu_set_flag(vcpu, IN_WFI); 711 vgic_v4_put(vcpu); 712 preempt_enable(); 713 714 kvm_vcpu_halt(vcpu); 715 vcpu_clear_flag(vcpu, IN_WFIT); 716 717 preempt_disable(); 718 vcpu_clear_flag(vcpu, IN_WFI); 719 vgic_v4_load(vcpu); 720 preempt_enable(); 721 } 722 723 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu) 724 { 725 if (!kvm_arm_vcpu_suspended(vcpu)) 726 return 1; 727 728 kvm_vcpu_wfi(vcpu); 729 730 /* 731 * The suspend state is sticky; we do not leave it until userspace 732 * explicitly marks the vCPU as runnable. Request that we suspend again 733 * later. 734 */ 735 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 736 737 /* 738 * Check to make sure the vCPU is actually runnable. If so, exit to 739 * userspace informing it of the wakeup condition. 740 */ 741 if (kvm_arch_vcpu_runnable(vcpu)) { 742 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event)); 743 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP; 744 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT; 745 return 0; 746 } 747 748 /* 749 * Otherwise, we were unblocked to process a different event, such as a 750 * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to 751 * process the event. 752 */ 753 return 1; 754 } 755 756 /** 757 * check_vcpu_requests - check and handle pending vCPU requests 758 * @vcpu: the VCPU pointer 759 * 760 * Return: 1 if we should enter the guest 761 * 0 if we should exit to userspace 762 * < 0 if we should exit to userspace, where the return value indicates 763 * an error 764 */ 765 static int check_vcpu_requests(struct kvm_vcpu *vcpu) 766 { 767 if (kvm_request_pending(vcpu)) { 768 if (kvm_check_request(KVM_REQ_VM_DEAD, vcpu)) 769 return -EIO; 770 771 if (kvm_check_request(KVM_REQ_SLEEP, vcpu)) 772 kvm_vcpu_sleep(vcpu); 773 774 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 775 kvm_reset_vcpu(vcpu); 776 777 /* 778 * Clear IRQ_PENDING requests that were made to guarantee 779 * that a VCPU sees new virtual interrupts. 780 */ 781 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu); 782 783 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu)) 784 kvm_update_stolen_time(vcpu); 785 786 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) { 787 /* The distributor enable bits were changed */ 788 preempt_disable(); 789 vgic_v4_put(vcpu); 790 vgic_v4_load(vcpu); 791 preempt_enable(); 792 } 793 794 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu)) 795 kvm_pmu_handle_pmcr(vcpu, 796 __vcpu_sys_reg(vcpu, PMCR_EL0)); 797 798 if (kvm_check_request(KVM_REQ_RESYNC_PMU_EL0, vcpu)) 799 kvm_vcpu_pmu_restore_guest(vcpu); 800 801 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu)) 802 return kvm_vcpu_suspend(vcpu); 803 804 if (kvm_dirty_ring_check_request(vcpu)) 805 return 0; 806 } 807 808 return 1; 809 } 810 811 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) 812 { 813 if (likely(!vcpu_mode_is_32bit(vcpu))) 814 return false; 815 816 if (vcpu_has_nv(vcpu)) 817 return true; 818 819 return !kvm_supports_32bit_el0(); 820 } 821 822 /** 823 * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest 824 * @vcpu: The VCPU pointer 825 * @ret: Pointer to write optional return code 826 * 827 * Returns: true if the VCPU needs to return to a preemptible + interruptible 828 * and skip guest entry. 829 * 830 * This function disambiguates between two different types of exits: exits to a 831 * preemptible + interruptible kernel context and exits to userspace. For an 832 * exit to userspace, this function will write the return code to ret and return 833 * true. For an exit to preemptible + interruptible kernel context (i.e. check 834 * for pending work and re-enter), return true without writing to ret. 835 */ 836 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret) 837 { 838 struct kvm_run *run = vcpu->run; 839 840 /* 841 * If we're using a userspace irqchip, then check if we need 842 * to tell a userspace irqchip about timer or PMU level 843 * changes and if so, exit to userspace (the actual level 844 * state gets updated in kvm_timer_update_run and 845 * kvm_pmu_update_run below). 846 */ 847 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) { 848 if (kvm_timer_should_notify_user(vcpu) || 849 kvm_pmu_should_notify_user(vcpu)) { 850 *ret = -EINTR; 851 run->exit_reason = KVM_EXIT_INTR; 852 return true; 853 } 854 } 855 856 if (unlikely(vcpu_on_unsupported_cpu(vcpu))) { 857 run->exit_reason = KVM_EXIT_FAIL_ENTRY; 858 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED; 859 run->fail_entry.cpu = smp_processor_id(); 860 *ret = 0; 861 return true; 862 } 863 864 return kvm_request_pending(vcpu) || 865 xfer_to_guest_mode_work_pending(); 866 } 867 868 /* 869 * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while 870 * the vCPU is running. 871 * 872 * This must be noinstr as instrumentation may make use of RCU, and this is not 873 * safe during the EQS. 874 */ 875 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu) 876 { 877 int ret; 878 879 guest_state_enter_irqoff(); 880 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu); 881 guest_state_exit_irqoff(); 882 883 return ret; 884 } 885 886 /** 887 * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code 888 * @vcpu: The VCPU pointer 889 * 890 * This function is called through the VCPU_RUN ioctl called from user space. It 891 * will execute VM code in a loop until the time slice for the process is used 892 * or some emulation is needed from user space in which case the function will 893 * return with return value 0 and with the kvm_run structure filled in with the 894 * required data for the requested emulation. 895 */ 896 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) 897 { 898 struct kvm_run *run = vcpu->run; 899 int ret; 900 901 if (run->exit_reason == KVM_EXIT_MMIO) { 902 ret = kvm_handle_mmio_return(vcpu); 903 if (ret) 904 return ret; 905 } 906 907 vcpu_load(vcpu); 908 909 if (run->immediate_exit) { 910 ret = -EINTR; 911 goto out; 912 } 913 914 kvm_sigset_activate(vcpu); 915 916 ret = 1; 917 run->exit_reason = KVM_EXIT_UNKNOWN; 918 run->flags = 0; 919 while (ret > 0) { 920 /* 921 * Check conditions before entering the guest 922 */ 923 ret = xfer_to_guest_mode_handle_work(vcpu); 924 if (!ret) 925 ret = 1; 926 927 if (ret > 0) 928 ret = check_vcpu_requests(vcpu); 929 930 /* 931 * Preparing the interrupts to be injected also 932 * involves poking the GIC, which must be done in a 933 * non-preemptible context. 934 */ 935 preempt_disable(); 936 937 /* 938 * The VMID allocator only tracks active VMIDs per 939 * physical CPU, and therefore the VMID allocated may not be 940 * preserved on VMID roll-over if the task was preempted, 941 * making a thread's VMID inactive. So we need to call 942 * kvm_arm_vmid_update() in non-premptible context. 943 */ 944 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid); 945 946 kvm_pmu_flush_hwstate(vcpu); 947 948 local_irq_disable(); 949 950 kvm_vgic_flush_hwstate(vcpu); 951 952 kvm_pmu_update_vcpu_events(vcpu); 953 954 /* 955 * Ensure we set mode to IN_GUEST_MODE after we disable 956 * interrupts and before the final VCPU requests check. 957 * See the comment in kvm_vcpu_exiting_guest_mode() and 958 * Documentation/virt/kvm/vcpu-requests.rst 959 */ 960 smp_store_mb(vcpu->mode, IN_GUEST_MODE); 961 962 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) { 963 vcpu->mode = OUTSIDE_GUEST_MODE; 964 isb(); /* Ensure work in x_flush_hwstate is committed */ 965 kvm_pmu_sync_hwstate(vcpu); 966 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) 967 kvm_timer_sync_user(vcpu); 968 kvm_vgic_sync_hwstate(vcpu); 969 local_irq_enable(); 970 preempt_enable(); 971 continue; 972 } 973 974 kvm_arm_setup_debug(vcpu); 975 kvm_arch_vcpu_ctxflush_fp(vcpu); 976 977 /************************************************************** 978 * Enter the guest 979 */ 980 trace_kvm_entry(*vcpu_pc(vcpu)); 981 guest_timing_enter_irqoff(); 982 983 ret = kvm_arm_vcpu_enter_exit(vcpu); 984 985 vcpu->mode = OUTSIDE_GUEST_MODE; 986 vcpu->stat.exits++; 987 /* 988 * Back from guest 989 *************************************************************/ 990 991 kvm_arm_clear_debug(vcpu); 992 993 /* 994 * We must sync the PMU state before the vgic state so 995 * that the vgic can properly sample the updated state of the 996 * interrupt line. 997 */ 998 kvm_pmu_sync_hwstate(vcpu); 999 1000 /* 1001 * Sync the vgic state before syncing the timer state because 1002 * the timer code needs to know if the virtual timer 1003 * interrupts are active. 1004 */ 1005 kvm_vgic_sync_hwstate(vcpu); 1006 1007 /* 1008 * Sync the timer hardware state before enabling interrupts as 1009 * we don't want vtimer interrupts to race with syncing the 1010 * timer virtual interrupt state. 1011 */ 1012 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) 1013 kvm_timer_sync_user(vcpu); 1014 1015 kvm_arch_vcpu_ctxsync_fp(vcpu); 1016 1017 /* 1018 * We must ensure that any pending interrupts are taken before 1019 * we exit guest timing so that timer ticks are accounted as 1020 * guest time. Transiently unmask interrupts so that any 1021 * pending interrupts are taken. 1022 * 1023 * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other 1024 * context synchronization event) is necessary to ensure that 1025 * pending interrupts are taken. 1026 */ 1027 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) { 1028 local_irq_enable(); 1029 isb(); 1030 local_irq_disable(); 1031 } 1032 1033 guest_timing_exit_irqoff(); 1034 1035 local_irq_enable(); 1036 1037 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu)); 1038 1039 /* Exit types that need handling before we can be preempted */ 1040 handle_exit_early(vcpu, ret); 1041 1042 preempt_enable(); 1043 1044 /* 1045 * The ARMv8 architecture doesn't give the hypervisor 1046 * a mechanism to prevent a guest from dropping to AArch32 EL0 1047 * if implemented by the CPU. If we spot the guest in such 1048 * state and that we decided it wasn't supposed to do so (like 1049 * with the asymmetric AArch32 case), return to userspace with 1050 * a fatal error. 1051 */ 1052 if (vcpu_mode_is_bad_32bit(vcpu)) { 1053 /* 1054 * As we have caught the guest red-handed, decide that 1055 * it isn't fit for purpose anymore by making the vcpu 1056 * invalid. The VMM can try and fix it by issuing a 1057 * KVM_ARM_VCPU_INIT if it really wants to. 1058 */ 1059 vcpu_clear_flag(vcpu, VCPU_INITIALIZED); 1060 ret = ARM_EXCEPTION_IL; 1061 } 1062 1063 ret = handle_exit(vcpu, ret); 1064 } 1065 1066 /* Tell userspace about in-kernel device output levels */ 1067 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) { 1068 kvm_timer_update_run(vcpu); 1069 kvm_pmu_update_run(vcpu); 1070 } 1071 1072 kvm_sigset_deactivate(vcpu); 1073 1074 out: 1075 /* 1076 * In the unlikely event that we are returning to userspace 1077 * with pending exceptions or PC adjustment, commit these 1078 * adjustments in order to give userspace a consistent view of 1079 * the vcpu state. Note that this relies on __kvm_adjust_pc() 1080 * being preempt-safe on VHE. 1081 */ 1082 if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) || 1083 vcpu_get_flag(vcpu, INCREMENT_PC))) 1084 kvm_call_hyp(__kvm_adjust_pc, vcpu); 1085 1086 vcpu_put(vcpu); 1087 return ret; 1088 } 1089 1090 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level) 1091 { 1092 int bit_index; 1093 bool set; 1094 unsigned long *hcr; 1095 1096 if (number == KVM_ARM_IRQ_CPU_IRQ) 1097 bit_index = __ffs(HCR_VI); 1098 else /* KVM_ARM_IRQ_CPU_FIQ */ 1099 bit_index = __ffs(HCR_VF); 1100 1101 hcr = vcpu_hcr(vcpu); 1102 if (level) 1103 set = test_and_set_bit(bit_index, hcr); 1104 else 1105 set = test_and_clear_bit(bit_index, hcr); 1106 1107 /* 1108 * If we didn't change anything, no need to wake up or kick other CPUs 1109 */ 1110 if (set == level) 1111 return 0; 1112 1113 /* 1114 * The vcpu irq_lines field was updated, wake up sleeping VCPUs and 1115 * trigger a world-switch round on the running physical CPU to set the 1116 * virtual IRQ/FIQ fields in the HCR appropriately. 1117 */ 1118 kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); 1119 kvm_vcpu_kick(vcpu); 1120 1121 return 0; 1122 } 1123 1124 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level, 1125 bool line_status) 1126 { 1127 u32 irq = irq_level->irq; 1128 unsigned int irq_type, vcpu_idx, irq_num; 1129 int nrcpus = atomic_read(&kvm->online_vcpus); 1130 struct kvm_vcpu *vcpu = NULL; 1131 bool level = irq_level->level; 1132 1133 irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK; 1134 vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK; 1135 vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1); 1136 irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK; 1137 1138 trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level); 1139 1140 switch (irq_type) { 1141 case KVM_ARM_IRQ_TYPE_CPU: 1142 if (irqchip_in_kernel(kvm)) 1143 return -ENXIO; 1144 1145 if (vcpu_idx >= nrcpus) 1146 return -EINVAL; 1147 1148 vcpu = kvm_get_vcpu(kvm, vcpu_idx); 1149 if (!vcpu) 1150 return -EINVAL; 1151 1152 if (irq_num > KVM_ARM_IRQ_CPU_FIQ) 1153 return -EINVAL; 1154 1155 return vcpu_interrupt_line(vcpu, irq_num, level); 1156 case KVM_ARM_IRQ_TYPE_PPI: 1157 if (!irqchip_in_kernel(kvm)) 1158 return -ENXIO; 1159 1160 if (vcpu_idx >= nrcpus) 1161 return -EINVAL; 1162 1163 vcpu = kvm_get_vcpu(kvm, vcpu_idx); 1164 if (!vcpu) 1165 return -EINVAL; 1166 1167 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS) 1168 return -EINVAL; 1169 1170 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL); 1171 case KVM_ARM_IRQ_TYPE_SPI: 1172 if (!irqchip_in_kernel(kvm)) 1173 return -ENXIO; 1174 1175 if (irq_num < VGIC_NR_PRIVATE_IRQS) 1176 return -EINVAL; 1177 1178 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL); 1179 } 1180 1181 return -EINVAL; 1182 } 1183 1184 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu, 1185 const struct kvm_vcpu_init *init) 1186 { 1187 unsigned long features = init->features[0]; 1188 int i; 1189 1190 if (features & ~KVM_VCPU_VALID_FEATURES) 1191 return -ENOENT; 1192 1193 for (i = 1; i < ARRAY_SIZE(init->features); i++) { 1194 if (init->features[i]) 1195 return -ENOENT; 1196 } 1197 1198 if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features)) 1199 return 0; 1200 1201 if (!cpus_have_const_cap(ARM64_HAS_32BIT_EL1)) 1202 return -EINVAL; 1203 1204 /* MTE is incompatible with AArch32 */ 1205 if (kvm_has_mte(vcpu->kvm)) 1206 return -EINVAL; 1207 1208 /* NV is incompatible with AArch32 */ 1209 if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features)) 1210 return -EINVAL; 1211 1212 return 0; 1213 } 1214 1215 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu, 1216 const struct kvm_vcpu_init *init) 1217 { 1218 unsigned long features = init->features[0]; 1219 1220 return !bitmap_equal(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES); 1221 } 1222 1223 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1224 const struct kvm_vcpu_init *init) 1225 { 1226 unsigned long features = init->features[0]; 1227 struct kvm *kvm = vcpu->kvm; 1228 int ret = -EINVAL; 1229 1230 mutex_lock(&kvm->arch.config_lock); 1231 1232 if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) && 1233 !bitmap_equal(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES)) 1234 goto out_unlock; 1235 1236 bitmap_copy(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES); 1237 1238 /* Now we know what it is, we can reset it. */ 1239 ret = kvm_reset_vcpu(vcpu); 1240 if (ret) { 1241 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES); 1242 goto out_unlock; 1243 } 1244 1245 bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES); 1246 set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags); 1247 vcpu_set_flag(vcpu, VCPU_INITIALIZED); 1248 out_unlock: 1249 mutex_unlock(&kvm->arch.config_lock); 1250 return ret; 1251 } 1252 1253 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1254 const struct kvm_vcpu_init *init) 1255 { 1256 int ret; 1257 1258 if (init->target != KVM_ARM_TARGET_GENERIC_V8 && 1259 init->target != kvm_target_cpu()) 1260 return -EINVAL; 1261 1262 ret = kvm_vcpu_init_check_features(vcpu, init); 1263 if (ret) 1264 return ret; 1265 1266 if (!kvm_vcpu_initialized(vcpu)) 1267 return __kvm_vcpu_set_target(vcpu, init); 1268 1269 if (kvm_vcpu_init_changed(vcpu, init)) 1270 return -EINVAL; 1271 1272 return kvm_reset_vcpu(vcpu); 1273 } 1274 1275 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu, 1276 struct kvm_vcpu_init *init) 1277 { 1278 bool power_off = false; 1279 int ret; 1280 1281 /* 1282 * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid 1283 * reflecting it in the finalized feature set, thus limiting its scope 1284 * to a single KVM_ARM_VCPU_INIT call. 1285 */ 1286 if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) { 1287 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF); 1288 power_off = true; 1289 } 1290 1291 ret = kvm_vcpu_set_target(vcpu, init); 1292 if (ret) 1293 return ret; 1294 1295 /* 1296 * Ensure a rebooted VM will fault in RAM pages and detect if the 1297 * guest MMU is turned off and flush the caches as needed. 1298 * 1299 * S2FWB enforces all memory accesses to RAM being cacheable, 1300 * ensuring that the data side is always coherent. We still 1301 * need to invalidate the I-cache though, as FWB does *not* 1302 * imply CTR_EL0.DIC. 1303 */ 1304 if (vcpu_has_run_once(vcpu)) { 1305 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB)) 1306 stage2_unmap_vm(vcpu->kvm); 1307 else 1308 icache_inval_all_pou(); 1309 } 1310 1311 vcpu_reset_hcr(vcpu); 1312 vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu); 1313 1314 /* 1315 * Handle the "start in power-off" case. 1316 */ 1317 spin_lock(&vcpu->arch.mp_state_lock); 1318 1319 if (power_off) 1320 __kvm_arm_vcpu_power_off(vcpu); 1321 else 1322 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE); 1323 1324 spin_unlock(&vcpu->arch.mp_state_lock); 1325 1326 return 0; 1327 } 1328 1329 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu, 1330 struct kvm_device_attr *attr) 1331 { 1332 int ret = -ENXIO; 1333 1334 switch (attr->group) { 1335 default: 1336 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr); 1337 break; 1338 } 1339 1340 return ret; 1341 } 1342 1343 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu, 1344 struct kvm_device_attr *attr) 1345 { 1346 int ret = -ENXIO; 1347 1348 switch (attr->group) { 1349 default: 1350 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr); 1351 break; 1352 } 1353 1354 return ret; 1355 } 1356 1357 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu, 1358 struct kvm_device_attr *attr) 1359 { 1360 int ret = -ENXIO; 1361 1362 switch (attr->group) { 1363 default: 1364 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr); 1365 break; 1366 } 1367 1368 return ret; 1369 } 1370 1371 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu, 1372 struct kvm_vcpu_events *events) 1373 { 1374 memset(events, 0, sizeof(*events)); 1375 1376 return __kvm_arm_vcpu_get_events(vcpu, events); 1377 } 1378 1379 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu, 1380 struct kvm_vcpu_events *events) 1381 { 1382 int i; 1383 1384 /* check whether the reserved field is zero */ 1385 for (i = 0; i < ARRAY_SIZE(events->reserved); i++) 1386 if (events->reserved[i]) 1387 return -EINVAL; 1388 1389 /* check whether the pad field is zero */ 1390 for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++) 1391 if (events->exception.pad[i]) 1392 return -EINVAL; 1393 1394 return __kvm_arm_vcpu_set_events(vcpu, events); 1395 } 1396 1397 long kvm_arch_vcpu_ioctl(struct file *filp, 1398 unsigned int ioctl, unsigned long arg) 1399 { 1400 struct kvm_vcpu *vcpu = filp->private_data; 1401 void __user *argp = (void __user *)arg; 1402 struct kvm_device_attr attr; 1403 long r; 1404 1405 switch (ioctl) { 1406 case KVM_ARM_VCPU_INIT: { 1407 struct kvm_vcpu_init init; 1408 1409 r = -EFAULT; 1410 if (copy_from_user(&init, argp, sizeof(init))) 1411 break; 1412 1413 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init); 1414 break; 1415 } 1416 case KVM_SET_ONE_REG: 1417 case KVM_GET_ONE_REG: { 1418 struct kvm_one_reg reg; 1419 1420 r = -ENOEXEC; 1421 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1422 break; 1423 1424 r = -EFAULT; 1425 if (copy_from_user(®, argp, sizeof(reg))) 1426 break; 1427 1428 /* 1429 * We could owe a reset due to PSCI. Handle the pending reset 1430 * here to ensure userspace register accesses are ordered after 1431 * the reset. 1432 */ 1433 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 1434 kvm_reset_vcpu(vcpu); 1435 1436 if (ioctl == KVM_SET_ONE_REG) 1437 r = kvm_arm_set_reg(vcpu, ®); 1438 else 1439 r = kvm_arm_get_reg(vcpu, ®); 1440 break; 1441 } 1442 case KVM_GET_REG_LIST: { 1443 struct kvm_reg_list __user *user_list = argp; 1444 struct kvm_reg_list reg_list; 1445 unsigned n; 1446 1447 r = -ENOEXEC; 1448 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1449 break; 1450 1451 r = -EPERM; 1452 if (!kvm_arm_vcpu_is_finalized(vcpu)) 1453 break; 1454 1455 r = -EFAULT; 1456 if (copy_from_user(®_list, user_list, sizeof(reg_list))) 1457 break; 1458 n = reg_list.n; 1459 reg_list.n = kvm_arm_num_regs(vcpu); 1460 if (copy_to_user(user_list, ®_list, sizeof(reg_list))) 1461 break; 1462 r = -E2BIG; 1463 if (n < reg_list.n) 1464 break; 1465 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg); 1466 break; 1467 } 1468 case KVM_SET_DEVICE_ATTR: { 1469 r = -EFAULT; 1470 if (copy_from_user(&attr, argp, sizeof(attr))) 1471 break; 1472 r = kvm_arm_vcpu_set_attr(vcpu, &attr); 1473 break; 1474 } 1475 case KVM_GET_DEVICE_ATTR: { 1476 r = -EFAULT; 1477 if (copy_from_user(&attr, argp, sizeof(attr))) 1478 break; 1479 r = kvm_arm_vcpu_get_attr(vcpu, &attr); 1480 break; 1481 } 1482 case KVM_HAS_DEVICE_ATTR: { 1483 r = -EFAULT; 1484 if (copy_from_user(&attr, argp, sizeof(attr))) 1485 break; 1486 r = kvm_arm_vcpu_has_attr(vcpu, &attr); 1487 break; 1488 } 1489 case KVM_GET_VCPU_EVENTS: { 1490 struct kvm_vcpu_events events; 1491 1492 if (kvm_arm_vcpu_get_events(vcpu, &events)) 1493 return -EINVAL; 1494 1495 if (copy_to_user(argp, &events, sizeof(events))) 1496 return -EFAULT; 1497 1498 return 0; 1499 } 1500 case KVM_SET_VCPU_EVENTS: { 1501 struct kvm_vcpu_events events; 1502 1503 if (copy_from_user(&events, argp, sizeof(events))) 1504 return -EFAULT; 1505 1506 return kvm_arm_vcpu_set_events(vcpu, &events); 1507 } 1508 case KVM_ARM_VCPU_FINALIZE: { 1509 int what; 1510 1511 if (!kvm_vcpu_initialized(vcpu)) 1512 return -ENOEXEC; 1513 1514 if (get_user(what, (const int __user *)argp)) 1515 return -EFAULT; 1516 1517 return kvm_arm_vcpu_finalize(vcpu, what); 1518 } 1519 default: 1520 r = -EINVAL; 1521 } 1522 1523 return r; 1524 } 1525 1526 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot) 1527 { 1528 1529 } 1530 1531 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm, 1532 struct kvm_arm_device_addr *dev_addr) 1533 { 1534 switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) { 1535 case KVM_ARM_DEVICE_VGIC_V2: 1536 if (!vgic_present) 1537 return -ENXIO; 1538 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr); 1539 default: 1540 return -ENODEV; 1541 } 1542 } 1543 1544 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1545 { 1546 switch (attr->group) { 1547 case KVM_ARM_VM_SMCCC_CTRL: 1548 return kvm_vm_smccc_has_attr(kvm, attr); 1549 default: 1550 return -ENXIO; 1551 } 1552 } 1553 1554 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1555 { 1556 switch (attr->group) { 1557 case KVM_ARM_VM_SMCCC_CTRL: 1558 return kvm_vm_smccc_set_attr(kvm, attr); 1559 default: 1560 return -ENXIO; 1561 } 1562 } 1563 1564 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) 1565 { 1566 struct kvm *kvm = filp->private_data; 1567 void __user *argp = (void __user *)arg; 1568 struct kvm_device_attr attr; 1569 1570 switch (ioctl) { 1571 case KVM_CREATE_IRQCHIP: { 1572 int ret; 1573 if (!vgic_present) 1574 return -ENXIO; 1575 mutex_lock(&kvm->lock); 1576 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2); 1577 mutex_unlock(&kvm->lock); 1578 return ret; 1579 } 1580 case KVM_ARM_SET_DEVICE_ADDR: { 1581 struct kvm_arm_device_addr dev_addr; 1582 1583 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr))) 1584 return -EFAULT; 1585 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr); 1586 } 1587 case KVM_ARM_PREFERRED_TARGET: { 1588 struct kvm_vcpu_init init = { 1589 .target = KVM_ARM_TARGET_GENERIC_V8, 1590 }; 1591 1592 if (copy_to_user(argp, &init, sizeof(init))) 1593 return -EFAULT; 1594 1595 return 0; 1596 } 1597 case KVM_ARM_MTE_COPY_TAGS: { 1598 struct kvm_arm_copy_mte_tags copy_tags; 1599 1600 if (copy_from_user(©_tags, argp, sizeof(copy_tags))) 1601 return -EFAULT; 1602 return kvm_vm_ioctl_mte_copy_tags(kvm, ©_tags); 1603 } 1604 case KVM_ARM_SET_COUNTER_OFFSET: { 1605 struct kvm_arm_counter_offset offset; 1606 1607 if (copy_from_user(&offset, argp, sizeof(offset))) 1608 return -EFAULT; 1609 return kvm_vm_ioctl_set_counter_offset(kvm, &offset); 1610 } 1611 case KVM_HAS_DEVICE_ATTR: { 1612 if (copy_from_user(&attr, argp, sizeof(attr))) 1613 return -EFAULT; 1614 1615 return kvm_vm_has_attr(kvm, &attr); 1616 } 1617 case KVM_SET_DEVICE_ATTR: { 1618 if (copy_from_user(&attr, argp, sizeof(attr))) 1619 return -EFAULT; 1620 1621 return kvm_vm_set_attr(kvm, &attr); 1622 } 1623 default: 1624 return -EINVAL; 1625 } 1626 } 1627 1628 /* unlocks vcpus from @vcpu_lock_idx and smaller */ 1629 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx) 1630 { 1631 struct kvm_vcpu *tmp_vcpu; 1632 1633 for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) { 1634 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx); 1635 mutex_unlock(&tmp_vcpu->mutex); 1636 } 1637 } 1638 1639 void unlock_all_vcpus(struct kvm *kvm) 1640 { 1641 lockdep_assert_held(&kvm->lock); 1642 1643 unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1); 1644 } 1645 1646 /* Returns true if all vcpus were locked, false otherwise */ 1647 bool lock_all_vcpus(struct kvm *kvm) 1648 { 1649 struct kvm_vcpu *tmp_vcpu; 1650 unsigned long c; 1651 1652 lockdep_assert_held(&kvm->lock); 1653 1654 /* 1655 * Any time a vcpu is in an ioctl (including running), the 1656 * core KVM code tries to grab the vcpu->mutex. 1657 * 1658 * By grabbing the vcpu->mutex of all VCPUs we ensure that no 1659 * other VCPUs can fiddle with the state while we access it. 1660 */ 1661 kvm_for_each_vcpu(c, tmp_vcpu, kvm) { 1662 if (!mutex_trylock(&tmp_vcpu->mutex)) { 1663 unlock_vcpus(kvm, c - 1); 1664 return false; 1665 } 1666 } 1667 1668 return true; 1669 } 1670 1671 static unsigned long nvhe_percpu_size(void) 1672 { 1673 return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) - 1674 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start); 1675 } 1676 1677 static unsigned long nvhe_percpu_order(void) 1678 { 1679 unsigned long size = nvhe_percpu_size(); 1680 1681 return size ? get_order(size) : 0; 1682 } 1683 1684 /* A lookup table holding the hypervisor VA for each vector slot */ 1685 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS]; 1686 1687 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot) 1688 { 1689 hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot); 1690 } 1691 1692 static int kvm_init_vector_slots(void) 1693 { 1694 int err; 1695 void *base; 1696 1697 base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector)); 1698 kvm_init_vector_slot(base, HYP_VECTOR_DIRECT); 1699 1700 base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs)); 1701 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT); 1702 1703 if (kvm_system_needs_idmapped_vectors() && 1704 !is_protected_kvm_enabled()) { 1705 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs), 1706 __BP_HARDEN_HYP_VECS_SZ, &base); 1707 if (err) 1708 return err; 1709 } 1710 1711 kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT); 1712 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT); 1713 return 0; 1714 } 1715 1716 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits) 1717 { 1718 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 1719 unsigned long tcr; 1720 1721 /* 1722 * Calculate the raw per-cpu offset without a translation from the 1723 * kernel's mapping to the linear mapping, and store it in tpidr_el2 1724 * so that we can use adr_l to access per-cpu variables in EL2. 1725 * Also drop the KASAN tag which gets in the way... 1726 */ 1727 params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) - 1728 (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start)); 1729 1730 params->mair_el2 = read_sysreg(mair_el1); 1731 1732 tcr = read_sysreg(tcr_el1); 1733 if (cpus_have_final_cap(ARM64_KVM_HVHE)) { 1734 tcr |= TCR_EPD1_MASK; 1735 } else { 1736 tcr &= TCR_EL2_MASK; 1737 tcr |= TCR_EL2_RES1; 1738 } 1739 tcr &= ~TCR_T0SZ_MASK; 1740 tcr |= TCR_T0SZ(hyp_va_bits); 1741 params->tcr_el2 = tcr; 1742 1743 params->pgd_pa = kvm_mmu_get_httbr(); 1744 if (is_protected_kvm_enabled()) 1745 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS; 1746 else 1747 params->hcr_el2 = HCR_HOST_NVHE_FLAGS; 1748 if (cpus_have_final_cap(ARM64_KVM_HVHE)) 1749 params->hcr_el2 |= HCR_E2H; 1750 params->vttbr = params->vtcr = 0; 1751 1752 /* 1753 * Flush the init params from the data cache because the struct will 1754 * be read while the MMU is off. 1755 */ 1756 kvm_flush_dcache_to_poc(params, sizeof(*params)); 1757 } 1758 1759 static void hyp_install_host_vector(void) 1760 { 1761 struct kvm_nvhe_init_params *params; 1762 struct arm_smccc_res res; 1763 1764 /* Switch from the HYP stub to our own HYP init vector */ 1765 __hyp_set_vectors(kvm_get_idmap_vector()); 1766 1767 /* 1768 * Call initialization code, and switch to the full blown HYP code. 1769 * If the cpucaps haven't been finalized yet, something has gone very 1770 * wrong, and hyp will crash and burn when it uses any 1771 * cpus_have_const_cap() wrapper. 1772 */ 1773 BUG_ON(!system_capabilities_finalized()); 1774 params = this_cpu_ptr_nvhe_sym(kvm_init_params); 1775 arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res); 1776 WARN_ON(res.a0 != SMCCC_RET_SUCCESS); 1777 } 1778 1779 static void cpu_init_hyp_mode(void) 1780 { 1781 hyp_install_host_vector(); 1782 1783 /* 1784 * Disabling SSBD on a non-VHE system requires us to enable SSBS 1785 * at EL2. 1786 */ 1787 if (this_cpu_has_cap(ARM64_SSBS) && 1788 arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) { 1789 kvm_call_hyp_nvhe(__kvm_enable_ssbs); 1790 } 1791 } 1792 1793 static void cpu_hyp_reset(void) 1794 { 1795 if (!is_kernel_in_hyp_mode()) 1796 __hyp_reset_vectors(); 1797 } 1798 1799 /* 1800 * EL2 vectors can be mapped and rerouted in a number of ways, 1801 * depending on the kernel configuration and CPU present: 1802 * 1803 * - If the CPU is affected by Spectre-v2, the hardening sequence is 1804 * placed in one of the vector slots, which is executed before jumping 1805 * to the real vectors. 1806 * 1807 * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot 1808 * containing the hardening sequence is mapped next to the idmap page, 1809 * and executed before jumping to the real vectors. 1810 * 1811 * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an 1812 * empty slot is selected, mapped next to the idmap page, and 1813 * executed before jumping to the real vectors. 1814 * 1815 * Note that ARM64_SPECTRE_V3A is somewhat incompatible with 1816 * VHE, as we don't have hypervisor-specific mappings. If the system 1817 * is VHE and yet selects this capability, it will be ignored. 1818 */ 1819 static void cpu_set_hyp_vector(void) 1820 { 1821 struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data); 1822 void *vector = hyp_spectre_vector_selector[data->slot]; 1823 1824 if (!is_protected_kvm_enabled()) 1825 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector; 1826 else 1827 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot); 1828 } 1829 1830 static void cpu_hyp_init_context(void) 1831 { 1832 kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt); 1833 1834 if (!is_kernel_in_hyp_mode()) 1835 cpu_init_hyp_mode(); 1836 } 1837 1838 static void cpu_hyp_init_features(void) 1839 { 1840 cpu_set_hyp_vector(); 1841 kvm_arm_init_debug(); 1842 1843 if (is_kernel_in_hyp_mode()) 1844 kvm_timer_init_vhe(); 1845 1846 if (vgic_present) 1847 kvm_vgic_init_cpu_hardware(); 1848 } 1849 1850 static void cpu_hyp_reinit(void) 1851 { 1852 cpu_hyp_reset(); 1853 cpu_hyp_init_context(); 1854 cpu_hyp_init_features(); 1855 } 1856 1857 static void cpu_hyp_init(void *discard) 1858 { 1859 if (!__this_cpu_read(kvm_hyp_initialized)) { 1860 cpu_hyp_reinit(); 1861 __this_cpu_write(kvm_hyp_initialized, 1); 1862 } 1863 } 1864 1865 static void cpu_hyp_uninit(void *discard) 1866 { 1867 if (__this_cpu_read(kvm_hyp_initialized)) { 1868 cpu_hyp_reset(); 1869 __this_cpu_write(kvm_hyp_initialized, 0); 1870 } 1871 } 1872 1873 int kvm_arch_hardware_enable(void) 1874 { 1875 /* 1876 * Most calls to this function are made with migration 1877 * disabled, but not with preemption disabled. The former is 1878 * enough to ensure correctness, but most of the helpers 1879 * expect the later and will throw a tantrum otherwise. 1880 */ 1881 preempt_disable(); 1882 1883 cpu_hyp_init(NULL); 1884 1885 kvm_vgic_cpu_up(); 1886 kvm_timer_cpu_up(); 1887 1888 preempt_enable(); 1889 1890 return 0; 1891 } 1892 1893 void kvm_arch_hardware_disable(void) 1894 { 1895 kvm_timer_cpu_down(); 1896 kvm_vgic_cpu_down(); 1897 1898 if (!is_protected_kvm_enabled()) 1899 cpu_hyp_uninit(NULL); 1900 } 1901 1902 #ifdef CONFIG_CPU_PM 1903 static int hyp_init_cpu_pm_notifier(struct notifier_block *self, 1904 unsigned long cmd, 1905 void *v) 1906 { 1907 /* 1908 * kvm_hyp_initialized is left with its old value over 1909 * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should 1910 * re-enable hyp. 1911 */ 1912 switch (cmd) { 1913 case CPU_PM_ENTER: 1914 if (__this_cpu_read(kvm_hyp_initialized)) 1915 /* 1916 * don't update kvm_hyp_initialized here 1917 * so that the hyp will be re-enabled 1918 * when we resume. See below. 1919 */ 1920 cpu_hyp_reset(); 1921 1922 return NOTIFY_OK; 1923 case CPU_PM_ENTER_FAILED: 1924 case CPU_PM_EXIT: 1925 if (__this_cpu_read(kvm_hyp_initialized)) 1926 /* The hyp was enabled before suspend. */ 1927 cpu_hyp_reinit(); 1928 1929 return NOTIFY_OK; 1930 1931 default: 1932 return NOTIFY_DONE; 1933 } 1934 } 1935 1936 static struct notifier_block hyp_init_cpu_pm_nb = { 1937 .notifier_call = hyp_init_cpu_pm_notifier, 1938 }; 1939 1940 static void __init hyp_cpu_pm_init(void) 1941 { 1942 if (!is_protected_kvm_enabled()) 1943 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb); 1944 } 1945 static void __init hyp_cpu_pm_exit(void) 1946 { 1947 if (!is_protected_kvm_enabled()) 1948 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb); 1949 } 1950 #else 1951 static inline void __init hyp_cpu_pm_init(void) 1952 { 1953 } 1954 static inline void __init hyp_cpu_pm_exit(void) 1955 { 1956 } 1957 #endif 1958 1959 static void __init init_cpu_logical_map(void) 1960 { 1961 unsigned int cpu; 1962 1963 /* 1964 * Copy the MPIDR <-> logical CPU ID mapping to hyp. 1965 * Only copy the set of online CPUs whose features have been checked 1966 * against the finalized system capabilities. The hypervisor will not 1967 * allow any other CPUs from the `possible` set to boot. 1968 */ 1969 for_each_online_cpu(cpu) 1970 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu); 1971 } 1972 1973 #define init_psci_0_1_impl_state(config, what) \ 1974 config.psci_0_1_ ## what ## _implemented = psci_ops.what 1975 1976 static bool __init init_psci_relay(void) 1977 { 1978 /* 1979 * If PSCI has not been initialized, protected KVM cannot install 1980 * itself on newly booted CPUs. 1981 */ 1982 if (!psci_ops.get_version) { 1983 kvm_err("Cannot initialize protected mode without PSCI\n"); 1984 return false; 1985 } 1986 1987 kvm_host_psci_config.version = psci_ops.get_version(); 1988 kvm_host_psci_config.smccc_version = arm_smccc_get_version(); 1989 1990 if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) { 1991 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids(); 1992 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend); 1993 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on); 1994 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off); 1995 init_psci_0_1_impl_state(kvm_host_psci_config, migrate); 1996 } 1997 return true; 1998 } 1999 2000 static int __init init_subsystems(void) 2001 { 2002 int err = 0; 2003 2004 /* 2005 * Enable hardware so that subsystem initialisation can access EL2. 2006 */ 2007 on_each_cpu(cpu_hyp_init, NULL, 1); 2008 2009 /* 2010 * Register CPU lower-power notifier 2011 */ 2012 hyp_cpu_pm_init(); 2013 2014 /* 2015 * Init HYP view of VGIC 2016 */ 2017 err = kvm_vgic_hyp_init(); 2018 switch (err) { 2019 case 0: 2020 vgic_present = true; 2021 break; 2022 case -ENODEV: 2023 case -ENXIO: 2024 vgic_present = false; 2025 err = 0; 2026 break; 2027 default: 2028 goto out; 2029 } 2030 2031 /* 2032 * Init HYP architected timer support 2033 */ 2034 err = kvm_timer_hyp_init(vgic_present); 2035 if (err) 2036 goto out; 2037 2038 kvm_register_perf_callbacks(NULL); 2039 2040 out: 2041 if (err) 2042 hyp_cpu_pm_exit(); 2043 2044 if (err || !is_protected_kvm_enabled()) 2045 on_each_cpu(cpu_hyp_uninit, NULL, 1); 2046 2047 return err; 2048 } 2049 2050 static void __init teardown_subsystems(void) 2051 { 2052 kvm_unregister_perf_callbacks(); 2053 hyp_cpu_pm_exit(); 2054 } 2055 2056 static void __init teardown_hyp_mode(void) 2057 { 2058 int cpu; 2059 2060 free_hyp_pgds(); 2061 for_each_possible_cpu(cpu) { 2062 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu)); 2063 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order()); 2064 } 2065 } 2066 2067 static int __init do_pkvm_init(u32 hyp_va_bits) 2068 { 2069 void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)); 2070 int ret; 2071 2072 preempt_disable(); 2073 cpu_hyp_init_context(); 2074 ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size, 2075 num_possible_cpus(), kern_hyp_va(per_cpu_base), 2076 hyp_va_bits); 2077 cpu_hyp_init_features(); 2078 2079 /* 2080 * The stub hypercalls are now disabled, so set our local flag to 2081 * prevent a later re-init attempt in kvm_arch_hardware_enable(). 2082 */ 2083 __this_cpu_write(kvm_hyp_initialized, 1); 2084 preempt_enable(); 2085 2086 return ret; 2087 } 2088 2089 static u64 get_hyp_id_aa64pfr0_el1(void) 2090 { 2091 /* 2092 * Track whether the system isn't affected by spectre/meltdown in the 2093 * hypervisor's view of id_aa64pfr0_el1, used for protected VMs. 2094 * Although this is per-CPU, we make it global for simplicity, e.g., not 2095 * to have to worry about vcpu migration. 2096 * 2097 * Unlike for non-protected VMs, userspace cannot override this for 2098 * protected VMs. 2099 */ 2100 u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1); 2101 2102 val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) | 2103 ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3)); 2104 2105 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), 2106 arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED); 2107 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), 2108 arm64_get_meltdown_state() == SPECTRE_UNAFFECTED); 2109 2110 return val; 2111 } 2112 2113 static void kvm_hyp_init_symbols(void) 2114 { 2115 kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1(); 2116 kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1); 2117 kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1); 2118 kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1); 2119 kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1); 2120 kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1); 2121 kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1); 2122 kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1); 2123 kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1); 2124 kvm_nvhe_sym(__icache_flags) = __icache_flags; 2125 kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits; 2126 } 2127 2128 static int __init kvm_hyp_init_protection(u32 hyp_va_bits) 2129 { 2130 void *addr = phys_to_virt(hyp_mem_base); 2131 int ret; 2132 2133 ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP); 2134 if (ret) 2135 return ret; 2136 2137 ret = do_pkvm_init(hyp_va_bits); 2138 if (ret) 2139 return ret; 2140 2141 free_hyp_pgds(); 2142 2143 return 0; 2144 } 2145 2146 static void pkvm_hyp_init_ptrauth(void) 2147 { 2148 struct kvm_cpu_context *hyp_ctxt; 2149 int cpu; 2150 2151 for_each_possible_cpu(cpu) { 2152 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu); 2153 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long(); 2154 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long(); 2155 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long(); 2156 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long(); 2157 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long(); 2158 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long(); 2159 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long(); 2160 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long(); 2161 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long(); 2162 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long(); 2163 } 2164 } 2165 2166 /* Inits Hyp-mode on all online CPUs */ 2167 static int __init init_hyp_mode(void) 2168 { 2169 u32 hyp_va_bits; 2170 int cpu; 2171 int err = -ENOMEM; 2172 2173 /* 2174 * The protected Hyp-mode cannot be initialized if the memory pool 2175 * allocation has failed. 2176 */ 2177 if (is_protected_kvm_enabled() && !hyp_mem_base) 2178 goto out_err; 2179 2180 /* 2181 * Allocate Hyp PGD and setup Hyp identity mapping 2182 */ 2183 err = kvm_mmu_init(&hyp_va_bits); 2184 if (err) 2185 goto out_err; 2186 2187 /* 2188 * Allocate stack pages for Hypervisor-mode 2189 */ 2190 for_each_possible_cpu(cpu) { 2191 unsigned long stack_page; 2192 2193 stack_page = __get_free_page(GFP_KERNEL); 2194 if (!stack_page) { 2195 err = -ENOMEM; 2196 goto out_err; 2197 } 2198 2199 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page; 2200 } 2201 2202 /* 2203 * Allocate and initialize pages for Hypervisor-mode percpu regions. 2204 */ 2205 for_each_possible_cpu(cpu) { 2206 struct page *page; 2207 void *page_addr; 2208 2209 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order()); 2210 if (!page) { 2211 err = -ENOMEM; 2212 goto out_err; 2213 } 2214 2215 page_addr = page_address(page); 2216 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size()); 2217 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr; 2218 } 2219 2220 /* 2221 * Map the Hyp-code called directly from the host 2222 */ 2223 err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start), 2224 kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC); 2225 if (err) { 2226 kvm_err("Cannot map world-switch code\n"); 2227 goto out_err; 2228 } 2229 2230 err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start), 2231 kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO); 2232 if (err) { 2233 kvm_err("Cannot map .hyp.rodata section\n"); 2234 goto out_err; 2235 } 2236 2237 err = create_hyp_mappings(kvm_ksym_ref(__start_rodata), 2238 kvm_ksym_ref(__end_rodata), PAGE_HYP_RO); 2239 if (err) { 2240 kvm_err("Cannot map rodata section\n"); 2241 goto out_err; 2242 } 2243 2244 /* 2245 * .hyp.bss is guaranteed to be placed at the beginning of the .bss 2246 * section thanks to an assertion in the linker script. Map it RW and 2247 * the rest of .bss RO. 2248 */ 2249 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start), 2250 kvm_ksym_ref(__hyp_bss_end), PAGE_HYP); 2251 if (err) { 2252 kvm_err("Cannot map hyp bss section: %d\n", err); 2253 goto out_err; 2254 } 2255 2256 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end), 2257 kvm_ksym_ref(__bss_stop), PAGE_HYP_RO); 2258 if (err) { 2259 kvm_err("Cannot map bss section\n"); 2260 goto out_err; 2261 } 2262 2263 /* 2264 * Map the Hyp stack pages 2265 */ 2266 for_each_possible_cpu(cpu) { 2267 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 2268 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu); 2269 2270 err = create_hyp_stack(__pa(stack_page), ¶ms->stack_hyp_va); 2271 if (err) { 2272 kvm_err("Cannot map hyp stack\n"); 2273 goto out_err; 2274 } 2275 2276 /* 2277 * Save the stack PA in nvhe_init_params. This will be needed 2278 * to recreate the stack mapping in protected nVHE mode. 2279 * __hyp_pa() won't do the right thing there, since the stack 2280 * has been mapped in the flexible private VA space. 2281 */ 2282 params->stack_pa = __pa(stack_page); 2283 } 2284 2285 for_each_possible_cpu(cpu) { 2286 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu]; 2287 char *percpu_end = percpu_begin + nvhe_percpu_size(); 2288 2289 /* Map Hyp percpu pages */ 2290 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP); 2291 if (err) { 2292 kvm_err("Cannot map hyp percpu region\n"); 2293 goto out_err; 2294 } 2295 2296 /* Prepare the CPU initialization parameters */ 2297 cpu_prepare_hyp_mode(cpu, hyp_va_bits); 2298 } 2299 2300 kvm_hyp_init_symbols(); 2301 2302 if (is_protected_kvm_enabled()) { 2303 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) && 2304 cpus_have_const_cap(ARM64_HAS_ADDRESS_AUTH)) 2305 pkvm_hyp_init_ptrauth(); 2306 2307 init_cpu_logical_map(); 2308 2309 if (!init_psci_relay()) { 2310 err = -ENODEV; 2311 goto out_err; 2312 } 2313 2314 err = kvm_hyp_init_protection(hyp_va_bits); 2315 if (err) { 2316 kvm_err("Failed to init hyp memory protection\n"); 2317 goto out_err; 2318 } 2319 } 2320 2321 return 0; 2322 2323 out_err: 2324 teardown_hyp_mode(); 2325 kvm_err("error initializing Hyp mode: %d\n", err); 2326 return err; 2327 } 2328 2329 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr) 2330 { 2331 struct kvm_vcpu *vcpu; 2332 unsigned long i; 2333 2334 mpidr &= MPIDR_HWID_BITMASK; 2335 kvm_for_each_vcpu(i, vcpu, kvm) { 2336 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu)) 2337 return vcpu; 2338 } 2339 return NULL; 2340 } 2341 2342 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm) 2343 { 2344 return irqchip_in_kernel(kvm); 2345 } 2346 2347 bool kvm_arch_has_irq_bypass(void) 2348 { 2349 return true; 2350 } 2351 2352 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons, 2353 struct irq_bypass_producer *prod) 2354 { 2355 struct kvm_kernel_irqfd *irqfd = 2356 container_of(cons, struct kvm_kernel_irqfd, consumer); 2357 2358 return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq, 2359 &irqfd->irq_entry); 2360 } 2361 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons, 2362 struct irq_bypass_producer *prod) 2363 { 2364 struct kvm_kernel_irqfd *irqfd = 2365 container_of(cons, struct kvm_kernel_irqfd, consumer); 2366 2367 kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq, 2368 &irqfd->irq_entry); 2369 } 2370 2371 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons) 2372 { 2373 struct kvm_kernel_irqfd *irqfd = 2374 container_of(cons, struct kvm_kernel_irqfd, consumer); 2375 2376 kvm_arm_halt_guest(irqfd->kvm); 2377 } 2378 2379 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons) 2380 { 2381 struct kvm_kernel_irqfd *irqfd = 2382 container_of(cons, struct kvm_kernel_irqfd, consumer); 2383 2384 kvm_arm_resume_guest(irqfd->kvm); 2385 } 2386 2387 /* Initialize Hyp-mode and memory mappings on all CPUs */ 2388 static __init int kvm_arm_init(void) 2389 { 2390 int err; 2391 bool in_hyp_mode; 2392 2393 if (!is_hyp_mode_available()) { 2394 kvm_info("HYP mode not available\n"); 2395 return -ENODEV; 2396 } 2397 2398 if (kvm_get_mode() == KVM_MODE_NONE) { 2399 kvm_info("KVM disabled from command line\n"); 2400 return -ENODEV; 2401 } 2402 2403 err = kvm_sys_reg_table_init(); 2404 if (err) { 2405 kvm_info("Error initializing system register tables"); 2406 return err; 2407 } 2408 2409 in_hyp_mode = is_kernel_in_hyp_mode(); 2410 2411 if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) || 2412 cpus_have_final_cap(ARM64_WORKAROUND_1508412)) 2413 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \ 2414 "Only trusted guests should be used on this system.\n"); 2415 2416 err = kvm_set_ipa_limit(); 2417 if (err) 2418 return err; 2419 2420 err = kvm_arm_init_sve(); 2421 if (err) 2422 return err; 2423 2424 err = kvm_arm_vmid_alloc_init(); 2425 if (err) { 2426 kvm_err("Failed to initialize VMID allocator.\n"); 2427 return err; 2428 } 2429 2430 if (!in_hyp_mode) { 2431 err = init_hyp_mode(); 2432 if (err) 2433 goto out_err; 2434 } 2435 2436 err = kvm_init_vector_slots(); 2437 if (err) { 2438 kvm_err("Cannot initialise vector slots\n"); 2439 goto out_hyp; 2440 } 2441 2442 err = init_subsystems(); 2443 if (err) 2444 goto out_hyp; 2445 2446 if (is_protected_kvm_enabled()) { 2447 kvm_info("Protected nVHE mode initialized successfully\n"); 2448 } else if (in_hyp_mode) { 2449 kvm_info("VHE mode initialized successfully\n"); 2450 } else { 2451 kvm_info("Hyp mode initialized successfully\n"); 2452 } 2453 2454 /* 2455 * FIXME: Do something reasonable if kvm_init() fails after pKVM 2456 * hypervisor protection is finalized. 2457 */ 2458 err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE); 2459 if (err) 2460 goto out_subs; 2461 2462 kvm_arm_initialised = true; 2463 2464 return 0; 2465 2466 out_subs: 2467 teardown_subsystems(); 2468 out_hyp: 2469 if (!in_hyp_mode) 2470 teardown_hyp_mode(); 2471 out_err: 2472 kvm_arm_vmid_alloc_free(); 2473 return err; 2474 } 2475 2476 static int __init early_kvm_mode_cfg(char *arg) 2477 { 2478 if (!arg) 2479 return -EINVAL; 2480 2481 if (strcmp(arg, "none") == 0) { 2482 kvm_mode = KVM_MODE_NONE; 2483 return 0; 2484 } 2485 2486 if (!is_hyp_mode_available()) { 2487 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n"); 2488 return 0; 2489 } 2490 2491 if (strcmp(arg, "protected") == 0) { 2492 if (!is_kernel_in_hyp_mode()) 2493 kvm_mode = KVM_MODE_PROTECTED; 2494 else 2495 pr_warn_once("Protected KVM not available with VHE\n"); 2496 2497 return 0; 2498 } 2499 2500 if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) { 2501 kvm_mode = KVM_MODE_DEFAULT; 2502 return 0; 2503 } 2504 2505 if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) { 2506 kvm_mode = KVM_MODE_NV; 2507 return 0; 2508 } 2509 2510 return -EINVAL; 2511 } 2512 early_param("kvm-arm.mode", early_kvm_mode_cfg); 2513 2514 enum kvm_mode kvm_get_mode(void) 2515 { 2516 return kvm_mode; 2517 } 2518 2519 module_init(kvm_arm_init); 2520