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 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use); 61 62 bool is_kvm_arm_initialised(void) 63 { 64 return kvm_arm_initialised; 65 } 66 67 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu) 68 { 69 return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE; 70 } 71 72 int kvm_vm_ioctl_enable_cap(struct kvm *kvm, 73 struct kvm_enable_cap *cap) 74 { 75 int r; 76 u64 new_cap; 77 78 if (cap->flags) 79 return -EINVAL; 80 81 switch (cap->cap) { 82 case KVM_CAP_ARM_NISV_TO_USER: 83 r = 0; 84 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER, 85 &kvm->arch.flags); 86 break; 87 case KVM_CAP_ARM_MTE: 88 mutex_lock(&kvm->lock); 89 if (!system_supports_mte() || kvm->created_vcpus) { 90 r = -EINVAL; 91 } else { 92 r = 0; 93 set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags); 94 } 95 mutex_unlock(&kvm->lock); 96 break; 97 case KVM_CAP_ARM_SYSTEM_SUSPEND: 98 r = 0; 99 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags); 100 break; 101 case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE: 102 new_cap = cap->args[0]; 103 104 mutex_lock(&kvm->slots_lock); 105 /* 106 * To keep things simple, allow changing the chunk 107 * size only when no memory slots have been created. 108 */ 109 if (!kvm_are_all_memslots_empty(kvm)) { 110 r = -EINVAL; 111 } else if (new_cap && !kvm_is_block_size_supported(new_cap)) { 112 r = -EINVAL; 113 } else { 114 r = 0; 115 kvm->arch.mmu.split_page_chunk_size = new_cap; 116 } 117 mutex_unlock(&kvm->slots_lock); 118 break; 119 default: 120 r = -EINVAL; 121 break; 122 } 123 124 return r; 125 } 126 127 static int kvm_arm_default_max_vcpus(void) 128 { 129 return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS; 130 } 131 132 /** 133 * kvm_arch_init_vm - initializes a VM data structure 134 * @kvm: pointer to the KVM struct 135 */ 136 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type) 137 { 138 int ret; 139 140 mutex_init(&kvm->arch.config_lock); 141 142 #ifdef CONFIG_LOCKDEP 143 /* Clue in lockdep that the config_lock must be taken inside kvm->lock */ 144 mutex_lock(&kvm->lock); 145 mutex_lock(&kvm->arch.config_lock); 146 mutex_unlock(&kvm->arch.config_lock); 147 mutex_unlock(&kvm->lock); 148 #endif 149 150 ret = kvm_share_hyp(kvm, kvm + 1); 151 if (ret) 152 return ret; 153 154 ret = pkvm_init_host_vm(kvm); 155 if (ret) 156 goto err_unshare_kvm; 157 158 if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) { 159 ret = -ENOMEM; 160 goto err_unshare_kvm; 161 } 162 cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask); 163 164 ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type); 165 if (ret) 166 goto err_free_cpumask; 167 168 kvm_vgic_early_init(kvm); 169 170 kvm_timer_init_vm(kvm); 171 172 /* The maximum number of VCPUs is limited by the host's GIC model */ 173 kvm->max_vcpus = kvm_arm_default_max_vcpus(); 174 175 kvm_arm_init_hypercalls(kvm); 176 177 bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES); 178 179 return 0; 180 181 err_free_cpumask: 182 free_cpumask_var(kvm->arch.supported_cpus); 183 err_unshare_kvm: 184 kvm_unshare_hyp(kvm, kvm + 1); 185 return ret; 186 } 187 188 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf) 189 { 190 return VM_FAULT_SIGBUS; 191 } 192 193 194 /** 195 * kvm_arch_destroy_vm - destroy the VM data structure 196 * @kvm: pointer to the KVM struct 197 */ 198 void kvm_arch_destroy_vm(struct kvm *kvm) 199 { 200 bitmap_free(kvm->arch.pmu_filter); 201 free_cpumask_var(kvm->arch.supported_cpus); 202 203 kvm_vgic_destroy(kvm); 204 205 if (is_protected_kvm_enabled()) 206 pkvm_destroy_hyp_vm(kvm); 207 208 kvm_destroy_vcpus(kvm); 209 210 kvm_unshare_hyp(kvm, kvm + 1); 211 212 kvm_arm_teardown_hypercalls(kvm); 213 } 214 215 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) 216 { 217 int r; 218 switch (ext) { 219 case KVM_CAP_IRQCHIP: 220 r = vgic_present; 221 break; 222 case KVM_CAP_IOEVENTFD: 223 case KVM_CAP_DEVICE_CTRL: 224 case KVM_CAP_USER_MEMORY: 225 case KVM_CAP_SYNC_MMU: 226 case KVM_CAP_DESTROY_MEMORY_REGION_WORKS: 227 case KVM_CAP_ONE_REG: 228 case KVM_CAP_ARM_PSCI: 229 case KVM_CAP_ARM_PSCI_0_2: 230 case KVM_CAP_READONLY_MEM: 231 case KVM_CAP_MP_STATE: 232 case KVM_CAP_IMMEDIATE_EXIT: 233 case KVM_CAP_VCPU_EVENTS: 234 case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2: 235 case KVM_CAP_ARM_NISV_TO_USER: 236 case KVM_CAP_ARM_INJECT_EXT_DABT: 237 case KVM_CAP_SET_GUEST_DEBUG: 238 case KVM_CAP_VCPU_ATTRIBUTES: 239 case KVM_CAP_PTP_KVM: 240 case KVM_CAP_ARM_SYSTEM_SUSPEND: 241 case KVM_CAP_IRQFD_RESAMPLE: 242 case KVM_CAP_COUNTER_OFFSET: 243 r = 1; 244 break; 245 case KVM_CAP_SET_GUEST_DEBUG2: 246 return KVM_GUESTDBG_VALID_MASK; 247 case KVM_CAP_ARM_SET_DEVICE_ADDR: 248 r = 1; 249 break; 250 case KVM_CAP_NR_VCPUS: 251 /* 252 * ARM64 treats KVM_CAP_NR_CPUS differently from all other 253 * architectures, as it does not always bound it to 254 * KVM_CAP_MAX_VCPUS. It should not matter much because 255 * this is just an advisory value. 256 */ 257 r = min_t(unsigned int, num_online_cpus(), 258 kvm_arm_default_max_vcpus()); 259 break; 260 case KVM_CAP_MAX_VCPUS: 261 case KVM_CAP_MAX_VCPU_ID: 262 if (kvm) 263 r = kvm->max_vcpus; 264 else 265 r = kvm_arm_default_max_vcpus(); 266 break; 267 case KVM_CAP_MSI_DEVID: 268 if (!kvm) 269 r = -EINVAL; 270 else 271 r = kvm->arch.vgic.msis_require_devid; 272 break; 273 case KVM_CAP_ARM_USER_IRQ: 274 /* 275 * 1: EL1_VTIMER, EL1_PTIMER, and PMU. 276 * (bump this number if adding more devices) 277 */ 278 r = 1; 279 break; 280 case KVM_CAP_ARM_MTE: 281 r = system_supports_mte(); 282 break; 283 case KVM_CAP_STEAL_TIME: 284 r = kvm_arm_pvtime_supported(); 285 break; 286 case KVM_CAP_ARM_EL1_32BIT: 287 r = cpus_have_const_cap(ARM64_HAS_32BIT_EL1); 288 break; 289 case KVM_CAP_GUEST_DEBUG_HW_BPS: 290 r = get_num_brps(); 291 break; 292 case KVM_CAP_GUEST_DEBUG_HW_WPS: 293 r = get_num_wrps(); 294 break; 295 case KVM_CAP_ARM_PMU_V3: 296 r = kvm_arm_support_pmu_v3(); 297 break; 298 case KVM_CAP_ARM_INJECT_SERROR_ESR: 299 r = cpus_have_const_cap(ARM64_HAS_RAS_EXTN); 300 break; 301 case KVM_CAP_ARM_VM_IPA_SIZE: 302 r = get_kvm_ipa_limit(); 303 break; 304 case KVM_CAP_ARM_SVE: 305 r = system_supports_sve(); 306 break; 307 case KVM_CAP_ARM_PTRAUTH_ADDRESS: 308 case KVM_CAP_ARM_PTRAUTH_GENERIC: 309 r = system_has_full_ptr_auth(); 310 break; 311 case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE: 312 if (kvm) 313 r = kvm->arch.mmu.split_page_chunk_size; 314 else 315 r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT; 316 break; 317 case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES: 318 r = kvm_supported_block_sizes(); 319 break; 320 default: 321 r = 0; 322 } 323 324 return r; 325 } 326 327 long kvm_arch_dev_ioctl(struct file *filp, 328 unsigned int ioctl, unsigned long arg) 329 { 330 return -EINVAL; 331 } 332 333 struct kvm *kvm_arch_alloc_vm(void) 334 { 335 size_t sz = sizeof(struct kvm); 336 337 if (!has_vhe()) 338 return kzalloc(sz, GFP_KERNEL_ACCOUNT); 339 340 return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO); 341 } 342 343 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id) 344 { 345 if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) 346 return -EBUSY; 347 348 if (id >= kvm->max_vcpus) 349 return -EINVAL; 350 351 return 0; 352 } 353 354 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu) 355 { 356 int err; 357 358 spin_lock_init(&vcpu->arch.mp_state_lock); 359 360 #ifdef CONFIG_LOCKDEP 361 /* Inform lockdep that the config_lock is acquired after vcpu->mutex */ 362 mutex_lock(&vcpu->mutex); 363 mutex_lock(&vcpu->kvm->arch.config_lock); 364 mutex_unlock(&vcpu->kvm->arch.config_lock); 365 mutex_unlock(&vcpu->mutex); 366 #endif 367 368 /* Force users to call KVM_ARM_VCPU_INIT */ 369 vcpu_clear_flag(vcpu, VCPU_INITIALIZED); 370 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES); 371 372 vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO; 373 374 /* 375 * Default value for the FP state, will be overloaded at load 376 * time if we support FP (pretty likely) 377 */ 378 vcpu->arch.fp_state = FP_STATE_FREE; 379 380 /* Set up the timer */ 381 kvm_timer_vcpu_init(vcpu); 382 383 kvm_pmu_vcpu_init(vcpu); 384 385 kvm_arm_reset_debug_ptr(vcpu); 386 387 kvm_arm_pvtime_vcpu_init(&vcpu->arch); 388 389 vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu; 390 391 err = kvm_vgic_vcpu_init(vcpu); 392 if (err) 393 return err; 394 395 return kvm_share_hyp(vcpu, vcpu + 1); 396 } 397 398 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu) 399 { 400 } 401 402 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu) 403 { 404 if (vcpu_has_run_once(vcpu) && unlikely(!irqchip_in_kernel(vcpu->kvm))) 405 static_branch_dec(&userspace_irqchip_in_use); 406 407 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache); 408 kvm_timer_vcpu_terminate(vcpu); 409 kvm_pmu_vcpu_destroy(vcpu); 410 411 kvm_arm_vcpu_destroy(vcpu); 412 } 413 414 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu) 415 { 416 417 } 418 419 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu) 420 { 421 422 } 423 424 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) 425 { 426 struct kvm_s2_mmu *mmu; 427 int *last_ran; 428 429 mmu = vcpu->arch.hw_mmu; 430 last_ran = this_cpu_ptr(mmu->last_vcpu_ran); 431 432 /* 433 * We guarantee that both TLBs and I-cache are private to each 434 * vcpu. If detecting that a vcpu from the same VM has 435 * previously run on the same physical CPU, call into the 436 * hypervisor code to nuke the relevant contexts. 437 * 438 * We might get preempted before the vCPU actually runs, but 439 * over-invalidation doesn't affect correctness. 440 */ 441 if (*last_ran != vcpu->vcpu_id) { 442 kvm_call_hyp(__kvm_flush_cpu_context, mmu); 443 *last_ran = vcpu->vcpu_id; 444 } 445 446 vcpu->cpu = cpu; 447 448 kvm_vgic_load(vcpu); 449 kvm_timer_vcpu_load(vcpu); 450 if (has_vhe()) 451 kvm_vcpu_load_sysregs_vhe(vcpu); 452 kvm_arch_vcpu_load_fp(vcpu); 453 kvm_vcpu_pmu_restore_guest(vcpu); 454 if (kvm_arm_is_pvtime_enabled(&vcpu->arch)) 455 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu); 456 457 if (single_task_running()) 458 vcpu_clear_wfx_traps(vcpu); 459 else 460 vcpu_set_wfx_traps(vcpu); 461 462 if (vcpu_has_ptrauth(vcpu)) 463 vcpu_ptrauth_disable(vcpu); 464 kvm_arch_vcpu_load_debug_state_flags(vcpu); 465 466 if (!cpumask_test_cpu(cpu, vcpu->kvm->arch.supported_cpus)) 467 vcpu_set_on_unsupported_cpu(vcpu); 468 } 469 470 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu) 471 { 472 kvm_arch_vcpu_put_debug_state_flags(vcpu); 473 kvm_arch_vcpu_put_fp(vcpu); 474 if (has_vhe()) 475 kvm_vcpu_put_sysregs_vhe(vcpu); 476 kvm_timer_vcpu_put(vcpu); 477 kvm_vgic_put(vcpu); 478 kvm_vcpu_pmu_restore_host(vcpu); 479 kvm_arm_vmid_clear_active(); 480 481 vcpu_clear_on_unsupported_cpu(vcpu); 482 vcpu->cpu = -1; 483 } 484 485 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 486 { 487 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED); 488 kvm_make_request(KVM_REQ_SLEEP, vcpu); 489 kvm_vcpu_kick(vcpu); 490 } 491 492 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 493 { 494 spin_lock(&vcpu->arch.mp_state_lock); 495 __kvm_arm_vcpu_power_off(vcpu); 496 spin_unlock(&vcpu->arch.mp_state_lock); 497 } 498 499 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu) 500 { 501 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED; 502 } 503 504 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu) 505 { 506 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED); 507 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 508 kvm_vcpu_kick(vcpu); 509 } 510 511 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu) 512 { 513 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED; 514 } 515 516 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu, 517 struct kvm_mp_state *mp_state) 518 { 519 *mp_state = READ_ONCE(vcpu->arch.mp_state); 520 521 return 0; 522 } 523 524 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu, 525 struct kvm_mp_state *mp_state) 526 { 527 int ret = 0; 528 529 spin_lock(&vcpu->arch.mp_state_lock); 530 531 switch (mp_state->mp_state) { 532 case KVM_MP_STATE_RUNNABLE: 533 WRITE_ONCE(vcpu->arch.mp_state, *mp_state); 534 break; 535 case KVM_MP_STATE_STOPPED: 536 __kvm_arm_vcpu_power_off(vcpu); 537 break; 538 case KVM_MP_STATE_SUSPENDED: 539 kvm_arm_vcpu_suspend(vcpu); 540 break; 541 default: 542 ret = -EINVAL; 543 } 544 545 spin_unlock(&vcpu->arch.mp_state_lock); 546 547 return ret; 548 } 549 550 /** 551 * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled 552 * @v: The VCPU pointer 553 * 554 * If the guest CPU is not waiting for interrupts or an interrupt line is 555 * asserted, the CPU is by definition runnable. 556 */ 557 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v) 558 { 559 bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF); 560 return ((irq_lines || kvm_vgic_vcpu_pending_irq(v)) 561 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause); 562 } 563 564 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu) 565 { 566 return vcpu_mode_priv(vcpu); 567 } 568 569 #ifdef CONFIG_GUEST_PERF_EVENTS 570 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu) 571 { 572 return *vcpu_pc(vcpu); 573 } 574 #endif 575 576 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu) 577 { 578 return vcpu_get_flag(vcpu, VCPU_INITIALIZED); 579 } 580 581 /* 582 * Handle both the initialisation that is being done when the vcpu is 583 * run for the first time, as well as the updates that must be 584 * performed each time we get a new thread dealing with this vcpu. 585 */ 586 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu) 587 { 588 struct kvm *kvm = vcpu->kvm; 589 int ret; 590 591 if (!kvm_vcpu_initialized(vcpu)) 592 return -ENOEXEC; 593 594 if (!kvm_arm_vcpu_is_finalized(vcpu)) 595 return -EPERM; 596 597 ret = kvm_arch_vcpu_run_map_fp(vcpu); 598 if (ret) 599 return ret; 600 601 if (likely(vcpu_has_run_once(vcpu))) 602 return 0; 603 604 kvm_arm_vcpu_init_debug(vcpu); 605 606 if (likely(irqchip_in_kernel(kvm))) { 607 /* 608 * Map the VGIC hardware resources before running a vcpu the 609 * first time on this VM. 610 */ 611 ret = kvm_vgic_map_resources(kvm); 612 if (ret) 613 return ret; 614 } 615 616 ret = kvm_timer_enable(vcpu); 617 if (ret) 618 return ret; 619 620 ret = kvm_arm_pmu_v3_enable(vcpu); 621 if (ret) 622 return ret; 623 624 if (is_protected_kvm_enabled()) { 625 ret = pkvm_create_hyp_vm(kvm); 626 if (ret) 627 return ret; 628 } 629 630 if (!irqchip_in_kernel(kvm)) { 631 /* 632 * Tell the rest of the code that there are userspace irqchip 633 * VMs in the wild. 634 */ 635 static_branch_inc(&userspace_irqchip_in_use); 636 } 637 638 /* 639 * Initialize traps for protected VMs. 640 * NOTE: Move to run in EL2 directly, rather than via a hypercall, once 641 * the code is in place for first run initialization at EL2. 642 */ 643 if (kvm_vm_is_protected(kvm)) 644 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu); 645 646 mutex_lock(&kvm->arch.config_lock); 647 set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags); 648 mutex_unlock(&kvm->arch.config_lock); 649 650 return ret; 651 } 652 653 bool kvm_arch_intc_initialized(struct kvm *kvm) 654 { 655 return vgic_initialized(kvm); 656 } 657 658 void kvm_arm_halt_guest(struct kvm *kvm) 659 { 660 unsigned long i; 661 struct kvm_vcpu *vcpu; 662 663 kvm_for_each_vcpu(i, vcpu, kvm) 664 vcpu->arch.pause = true; 665 kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP); 666 } 667 668 void kvm_arm_resume_guest(struct kvm *kvm) 669 { 670 unsigned long i; 671 struct kvm_vcpu *vcpu; 672 673 kvm_for_each_vcpu(i, vcpu, kvm) { 674 vcpu->arch.pause = false; 675 __kvm_vcpu_wake_up(vcpu); 676 } 677 } 678 679 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu) 680 { 681 struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); 682 683 rcuwait_wait_event(wait, 684 (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause), 685 TASK_INTERRUPTIBLE); 686 687 if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) { 688 /* Awaken to handle a signal, request we sleep again later. */ 689 kvm_make_request(KVM_REQ_SLEEP, vcpu); 690 } 691 692 /* 693 * Make sure we will observe a potential reset request if we've 694 * observed a change to the power state. Pairs with the smp_wmb() in 695 * kvm_psci_vcpu_on(). 696 */ 697 smp_rmb(); 698 } 699 700 /** 701 * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior 702 * @vcpu: The VCPU pointer 703 * 704 * Suspend execution of a vCPU until a valid wake event is detected, i.e. until 705 * the vCPU is runnable. The vCPU may or may not be scheduled out, depending 706 * on when a wake event arrives, e.g. there may already be a pending wake event. 707 */ 708 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu) 709 { 710 /* 711 * Sync back the state of the GIC CPU interface so that we have 712 * the latest PMR and group enables. This ensures that 713 * kvm_arch_vcpu_runnable has up-to-date data to decide whether 714 * we have pending interrupts, e.g. when determining if the 715 * vCPU should block. 716 * 717 * For the same reason, we want to tell GICv4 that we need 718 * doorbells to be signalled, should an interrupt become pending. 719 */ 720 preempt_disable(); 721 kvm_vgic_vmcr_sync(vcpu); 722 vcpu_set_flag(vcpu, IN_WFI); 723 vgic_v4_put(vcpu); 724 preempt_enable(); 725 726 kvm_vcpu_halt(vcpu); 727 vcpu_clear_flag(vcpu, IN_WFIT); 728 729 preempt_disable(); 730 vcpu_clear_flag(vcpu, IN_WFI); 731 vgic_v4_load(vcpu); 732 preempt_enable(); 733 } 734 735 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu) 736 { 737 if (!kvm_arm_vcpu_suspended(vcpu)) 738 return 1; 739 740 kvm_vcpu_wfi(vcpu); 741 742 /* 743 * The suspend state is sticky; we do not leave it until userspace 744 * explicitly marks the vCPU as runnable. Request that we suspend again 745 * later. 746 */ 747 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 748 749 /* 750 * Check to make sure the vCPU is actually runnable. If so, exit to 751 * userspace informing it of the wakeup condition. 752 */ 753 if (kvm_arch_vcpu_runnable(vcpu)) { 754 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event)); 755 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP; 756 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT; 757 return 0; 758 } 759 760 /* 761 * Otherwise, we were unblocked to process a different event, such as a 762 * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to 763 * process the event. 764 */ 765 return 1; 766 } 767 768 /** 769 * check_vcpu_requests - check and handle pending vCPU requests 770 * @vcpu: the VCPU pointer 771 * 772 * Return: 1 if we should enter the guest 773 * 0 if we should exit to userspace 774 * < 0 if we should exit to userspace, where the return value indicates 775 * an error 776 */ 777 static int check_vcpu_requests(struct kvm_vcpu *vcpu) 778 { 779 if (kvm_request_pending(vcpu)) { 780 if (kvm_check_request(KVM_REQ_SLEEP, vcpu)) 781 kvm_vcpu_sleep(vcpu); 782 783 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 784 kvm_reset_vcpu(vcpu); 785 786 /* 787 * Clear IRQ_PENDING requests that were made to guarantee 788 * that a VCPU sees new virtual interrupts. 789 */ 790 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu); 791 792 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu)) 793 kvm_update_stolen_time(vcpu); 794 795 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) { 796 /* The distributor enable bits were changed */ 797 preempt_disable(); 798 vgic_v4_put(vcpu); 799 vgic_v4_load(vcpu); 800 preempt_enable(); 801 } 802 803 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu)) 804 kvm_pmu_handle_pmcr(vcpu, 805 __vcpu_sys_reg(vcpu, PMCR_EL0)); 806 807 if (kvm_check_request(KVM_REQ_RESYNC_PMU_EL0, vcpu)) 808 kvm_vcpu_pmu_restore_guest(vcpu); 809 810 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu)) 811 return kvm_vcpu_suspend(vcpu); 812 813 if (kvm_dirty_ring_check_request(vcpu)) 814 return 0; 815 } 816 817 return 1; 818 } 819 820 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) 821 { 822 if (likely(!vcpu_mode_is_32bit(vcpu))) 823 return false; 824 825 if (vcpu_has_nv(vcpu)) 826 return true; 827 828 return !kvm_supports_32bit_el0(); 829 } 830 831 /** 832 * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest 833 * @vcpu: The VCPU pointer 834 * @ret: Pointer to write optional return code 835 * 836 * Returns: true if the VCPU needs to return to a preemptible + interruptible 837 * and skip guest entry. 838 * 839 * This function disambiguates between two different types of exits: exits to a 840 * preemptible + interruptible kernel context and exits to userspace. For an 841 * exit to userspace, this function will write the return code to ret and return 842 * true. For an exit to preemptible + interruptible kernel context (i.e. check 843 * for pending work and re-enter), return true without writing to ret. 844 */ 845 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret) 846 { 847 struct kvm_run *run = vcpu->run; 848 849 /* 850 * If we're using a userspace irqchip, then check if we need 851 * to tell a userspace irqchip about timer or PMU level 852 * changes and if so, exit to userspace (the actual level 853 * state gets updated in kvm_timer_update_run and 854 * kvm_pmu_update_run below). 855 */ 856 if (static_branch_unlikely(&userspace_irqchip_in_use)) { 857 if (kvm_timer_should_notify_user(vcpu) || 858 kvm_pmu_should_notify_user(vcpu)) { 859 *ret = -EINTR; 860 run->exit_reason = KVM_EXIT_INTR; 861 return true; 862 } 863 } 864 865 if (unlikely(vcpu_on_unsupported_cpu(vcpu))) { 866 run->exit_reason = KVM_EXIT_FAIL_ENTRY; 867 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED; 868 run->fail_entry.cpu = smp_processor_id(); 869 *ret = 0; 870 return true; 871 } 872 873 return kvm_request_pending(vcpu) || 874 xfer_to_guest_mode_work_pending(); 875 } 876 877 /* 878 * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while 879 * the vCPU is running. 880 * 881 * This must be noinstr as instrumentation may make use of RCU, and this is not 882 * safe during the EQS. 883 */ 884 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu) 885 { 886 int ret; 887 888 guest_state_enter_irqoff(); 889 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu); 890 guest_state_exit_irqoff(); 891 892 return ret; 893 } 894 895 /** 896 * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code 897 * @vcpu: The VCPU pointer 898 * 899 * This function is called through the VCPU_RUN ioctl called from user space. It 900 * will execute VM code in a loop until the time slice for the process is used 901 * or some emulation is needed from user space in which case the function will 902 * return with return value 0 and with the kvm_run structure filled in with the 903 * required data for the requested emulation. 904 */ 905 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) 906 { 907 struct kvm_run *run = vcpu->run; 908 int ret; 909 910 if (run->exit_reason == KVM_EXIT_MMIO) { 911 ret = kvm_handle_mmio_return(vcpu); 912 if (ret) 913 return ret; 914 } 915 916 vcpu_load(vcpu); 917 918 if (run->immediate_exit) { 919 ret = -EINTR; 920 goto out; 921 } 922 923 kvm_sigset_activate(vcpu); 924 925 ret = 1; 926 run->exit_reason = KVM_EXIT_UNKNOWN; 927 run->flags = 0; 928 while (ret > 0) { 929 /* 930 * Check conditions before entering the guest 931 */ 932 ret = xfer_to_guest_mode_handle_work(vcpu); 933 if (!ret) 934 ret = 1; 935 936 if (ret > 0) 937 ret = check_vcpu_requests(vcpu); 938 939 /* 940 * Preparing the interrupts to be injected also 941 * involves poking the GIC, which must be done in a 942 * non-preemptible context. 943 */ 944 preempt_disable(); 945 946 /* 947 * The VMID allocator only tracks active VMIDs per 948 * physical CPU, and therefore the VMID allocated may not be 949 * preserved on VMID roll-over if the task was preempted, 950 * making a thread's VMID inactive. So we need to call 951 * kvm_arm_vmid_update() in non-premptible context. 952 */ 953 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid); 954 955 kvm_pmu_flush_hwstate(vcpu); 956 957 local_irq_disable(); 958 959 kvm_vgic_flush_hwstate(vcpu); 960 961 kvm_pmu_update_vcpu_events(vcpu); 962 963 /* 964 * Ensure we set mode to IN_GUEST_MODE after we disable 965 * interrupts and before the final VCPU requests check. 966 * See the comment in kvm_vcpu_exiting_guest_mode() and 967 * Documentation/virt/kvm/vcpu-requests.rst 968 */ 969 smp_store_mb(vcpu->mode, IN_GUEST_MODE); 970 971 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) { 972 vcpu->mode = OUTSIDE_GUEST_MODE; 973 isb(); /* Ensure work in x_flush_hwstate is committed */ 974 kvm_pmu_sync_hwstate(vcpu); 975 if (static_branch_unlikely(&userspace_irqchip_in_use)) 976 kvm_timer_sync_user(vcpu); 977 kvm_vgic_sync_hwstate(vcpu); 978 local_irq_enable(); 979 preempt_enable(); 980 continue; 981 } 982 983 kvm_arm_setup_debug(vcpu); 984 kvm_arch_vcpu_ctxflush_fp(vcpu); 985 986 /************************************************************** 987 * Enter the guest 988 */ 989 trace_kvm_entry(*vcpu_pc(vcpu)); 990 guest_timing_enter_irqoff(); 991 992 ret = kvm_arm_vcpu_enter_exit(vcpu); 993 994 vcpu->mode = OUTSIDE_GUEST_MODE; 995 vcpu->stat.exits++; 996 /* 997 * Back from guest 998 *************************************************************/ 999 1000 kvm_arm_clear_debug(vcpu); 1001 1002 /* 1003 * We must sync the PMU state before the vgic state so 1004 * that the vgic can properly sample the updated state of the 1005 * interrupt line. 1006 */ 1007 kvm_pmu_sync_hwstate(vcpu); 1008 1009 /* 1010 * Sync the vgic state before syncing the timer state because 1011 * the timer code needs to know if the virtual timer 1012 * interrupts are active. 1013 */ 1014 kvm_vgic_sync_hwstate(vcpu); 1015 1016 /* 1017 * Sync the timer hardware state before enabling interrupts as 1018 * we don't want vtimer interrupts to race with syncing the 1019 * timer virtual interrupt state. 1020 */ 1021 if (static_branch_unlikely(&userspace_irqchip_in_use)) 1022 kvm_timer_sync_user(vcpu); 1023 1024 kvm_arch_vcpu_ctxsync_fp(vcpu); 1025 1026 /* 1027 * We must ensure that any pending interrupts are taken before 1028 * we exit guest timing so that timer ticks are accounted as 1029 * guest time. Transiently unmask interrupts so that any 1030 * pending interrupts are taken. 1031 * 1032 * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other 1033 * context synchronization event) is necessary to ensure that 1034 * pending interrupts are taken. 1035 */ 1036 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) { 1037 local_irq_enable(); 1038 isb(); 1039 local_irq_disable(); 1040 } 1041 1042 guest_timing_exit_irqoff(); 1043 1044 local_irq_enable(); 1045 1046 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu)); 1047 1048 /* Exit types that need handling before we can be preempted */ 1049 handle_exit_early(vcpu, ret); 1050 1051 preempt_enable(); 1052 1053 /* 1054 * The ARMv8 architecture doesn't give the hypervisor 1055 * a mechanism to prevent a guest from dropping to AArch32 EL0 1056 * if implemented by the CPU. If we spot the guest in such 1057 * state and that we decided it wasn't supposed to do so (like 1058 * with the asymmetric AArch32 case), return to userspace with 1059 * a fatal error. 1060 */ 1061 if (vcpu_mode_is_bad_32bit(vcpu)) { 1062 /* 1063 * As we have caught the guest red-handed, decide that 1064 * it isn't fit for purpose anymore by making the vcpu 1065 * invalid. The VMM can try and fix it by issuing a 1066 * KVM_ARM_VCPU_INIT if it really wants to. 1067 */ 1068 vcpu_clear_flag(vcpu, VCPU_INITIALIZED); 1069 ret = ARM_EXCEPTION_IL; 1070 } 1071 1072 ret = handle_exit(vcpu, ret); 1073 } 1074 1075 /* Tell userspace about in-kernel device output levels */ 1076 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) { 1077 kvm_timer_update_run(vcpu); 1078 kvm_pmu_update_run(vcpu); 1079 } 1080 1081 kvm_sigset_deactivate(vcpu); 1082 1083 out: 1084 /* 1085 * In the unlikely event that we are returning to userspace 1086 * with pending exceptions or PC adjustment, commit these 1087 * adjustments in order to give userspace a consistent view of 1088 * the vcpu state. Note that this relies on __kvm_adjust_pc() 1089 * being preempt-safe on VHE. 1090 */ 1091 if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) || 1092 vcpu_get_flag(vcpu, INCREMENT_PC))) 1093 kvm_call_hyp(__kvm_adjust_pc, vcpu); 1094 1095 vcpu_put(vcpu); 1096 return ret; 1097 } 1098 1099 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level) 1100 { 1101 int bit_index; 1102 bool set; 1103 unsigned long *hcr; 1104 1105 if (number == KVM_ARM_IRQ_CPU_IRQ) 1106 bit_index = __ffs(HCR_VI); 1107 else /* KVM_ARM_IRQ_CPU_FIQ */ 1108 bit_index = __ffs(HCR_VF); 1109 1110 hcr = vcpu_hcr(vcpu); 1111 if (level) 1112 set = test_and_set_bit(bit_index, hcr); 1113 else 1114 set = test_and_clear_bit(bit_index, hcr); 1115 1116 /* 1117 * If we didn't change anything, no need to wake up or kick other CPUs 1118 */ 1119 if (set == level) 1120 return 0; 1121 1122 /* 1123 * The vcpu irq_lines field was updated, wake up sleeping VCPUs and 1124 * trigger a world-switch round on the running physical CPU to set the 1125 * virtual IRQ/FIQ fields in the HCR appropriately. 1126 */ 1127 kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); 1128 kvm_vcpu_kick(vcpu); 1129 1130 return 0; 1131 } 1132 1133 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level, 1134 bool line_status) 1135 { 1136 u32 irq = irq_level->irq; 1137 unsigned int irq_type, vcpu_idx, irq_num; 1138 int nrcpus = atomic_read(&kvm->online_vcpus); 1139 struct kvm_vcpu *vcpu = NULL; 1140 bool level = irq_level->level; 1141 1142 irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK; 1143 vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK; 1144 vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1); 1145 irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK; 1146 1147 trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level); 1148 1149 switch (irq_type) { 1150 case KVM_ARM_IRQ_TYPE_CPU: 1151 if (irqchip_in_kernel(kvm)) 1152 return -ENXIO; 1153 1154 if (vcpu_idx >= nrcpus) 1155 return -EINVAL; 1156 1157 vcpu = kvm_get_vcpu(kvm, vcpu_idx); 1158 if (!vcpu) 1159 return -EINVAL; 1160 1161 if (irq_num > KVM_ARM_IRQ_CPU_FIQ) 1162 return -EINVAL; 1163 1164 return vcpu_interrupt_line(vcpu, irq_num, level); 1165 case KVM_ARM_IRQ_TYPE_PPI: 1166 if (!irqchip_in_kernel(kvm)) 1167 return -ENXIO; 1168 1169 if (vcpu_idx >= nrcpus) 1170 return -EINVAL; 1171 1172 vcpu = kvm_get_vcpu(kvm, vcpu_idx); 1173 if (!vcpu) 1174 return -EINVAL; 1175 1176 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS) 1177 return -EINVAL; 1178 1179 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL); 1180 case KVM_ARM_IRQ_TYPE_SPI: 1181 if (!irqchip_in_kernel(kvm)) 1182 return -ENXIO; 1183 1184 if (irq_num < VGIC_NR_PRIVATE_IRQS) 1185 return -EINVAL; 1186 1187 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL); 1188 } 1189 1190 return -EINVAL; 1191 } 1192 1193 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu, 1194 const struct kvm_vcpu_init *init) 1195 { 1196 unsigned long features = init->features[0]; 1197 int i; 1198 1199 if (features & ~KVM_VCPU_VALID_FEATURES) 1200 return -ENOENT; 1201 1202 for (i = 1; i < ARRAY_SIZE(init->features); i++) { 1203 if (init->features[i]) 1204 return -ENOENT; 1205 } 1206 1207 if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features)) 1208 return 0; 1209 1210 if (!cpus_have_const_cap(ARM64_HAS_32BIT_EL1)) 1211 return -EINVAL; 1212 1213 /* MTE is incompatible with AArch32 */ 1214 if (kvm_has_mte(vcpu->kvm)) 1215 return -EINVAL; 1216 1217 /* NV is incompatible with AArch32 */ 1218 if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features)) 1219 return -EINVAL; 1220 1221 return 0; 1222 } 1223 1224 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu, 1225 const struct kvm_vcpu_init *init) 1226 { 1227 unsigned long features = init->features[0]; 1228 1229 return !bitmap_equal(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES); 1230 } 1231 1232 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1233 const struct kvm_vcpu_init *init) 1234 { 1235 unsigned long features = init->features[0]; 1236 struct kvm *kvm = vcpu->kvm; 1237 int ret = -EINVAL; 1238 1239 mutex_lock(&kvm->arch.config_lock); 1240 1241 if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) && 1242 !bitmap_equal(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES)) 1243 goto out_unlock; 1244 1245 bitmap_copy(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES); 1246 1247 /* Now we know what it is, we can reset it. */ 1248 ret = kvm_reset_vcpu(vcpu); 1249 if (ret) { 1250 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES); 1251 goto out_unlock; 1252 } 1253 1254 bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES); 1255 set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags); 1256 vcpu_set_flag(vcpu, VCPU_INITIALIZED); 1257 out_unlock: 1258 mutex_unlock(&kvm->arch.config_lock); 1259 return ret; 1260 } 1261 1262 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1263 const struct kvm_vcpu_init *init) 1264 { 1265 int ret; 1266 1267 if (init->target != KVM_ARM_TARGET_GENERIC_V8 && 1268 init->target != kvm_target_cpu()) 1269 return -EINVAL; 1270 1271 ret = kvm_vcpu_init_check_features(vcpu, init); 1272 if (ret) 1273 return ret; 1274 1275 if (!kvm_vcpu_initialized(vcpu)) 1276 return __kvm_vcpu_set_target(vcpu, init); 1277 1278 if (kvm_vcpu_init_changed(vcpu, init)) 1279 return -EINVAL; 1280 1281 return kvm_reset_vcpu(vcpu); 1282 } 1283 1284 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu, 1285 struct kvm_vcpu_init *init) 1286 { 1287 bool power_off = false; 1288 int ret; 1289 1290 /* 1291 * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid 1292 * reflecting it in the finalized feature set, thus limiting its scope 1293 * to a single KVM_ARM_VCPU_INIT call. 1294 */ 1295 if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) { 1296 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF); 1297 power_off = true; 1298 } 1299 1300 ret = kvm_vcpu_set_target(vcpu, init); 1301 if (ret) 1302 return ret; 1303 1304 /* 1305 * Ensure a rebooted VM will fault in RAM pages and detect if the 1306 * guest MMU is turned off and flush the caches as needed. 1307 * 1308 * S2FWB enforces all memory accesses to RAM being cacheable, 1309 * ensuring that the data side is always coherent. We still 1310 * need to invalidate the I-cache though, as FWB does *not* 1311 * imply CTR_EL0.DIC. 1312 */ 1313 if (vcpu_has_run_once(vcpu)) { 1314 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB)) 1315 stage2_unmap_vm(vcpu->kvm); 1316 else 1317 icache_inval_all_pou(); 1318 } 1319 1320 vcpu_reset_hcr(vcpu); 1321 vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu); 1322 1323 /* 1324 * Handle the "start in power-off" case. 1325 */ 1326 spin_lock(&vcpu->arch.mp_state_lock); 1327 1328 if (power_off) 1329 __kvm_arm_vcpu_power_off(vcpu); 1330 else 1331 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE); 1332 1333 spin_unlock(&vcpu->arch.mp_state_lock); 1334 1335 return 0; 1336 } 1337 1338 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu, 1339 struct kvm_device_attr *attr) 1340 { 1341 int ret = -ENXIO; 1342 1343 switch (attr->group) { 1344 default: 1345 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr); 1346 break; 1347 } 1348 1349 return ret; 1350 } 1351 1352 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu, 1353 struct kvm_device_attr *attr) 1354 { 1355 int ret = -ENXIO; 1356 1357 switch (attr->group) { 1358 default: 1359 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr); 1360 break; 1361 } 1362 1363 return ret; 1364 } 1365 1366 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu, 1367 struct kvm_device_attr *attr) 1368 { 1369 int ret = -ENXIO; 1370 1371 switch (attr->group) { 1372 default: 1373 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr); 1374 break; 1375 } 1376 1377 return ret; 1378 } 1379 1380 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu, 1381 struct kvm_vcpu_events *events) 1382 { 1383 memset(events, 0, sizeof(*events)); 1384 1385 return __kvm_arm_vcpu_get_events(vcpu, events); 1386 } 1387 1388 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu, 1389 struct kvm_vcpu_events *events) 1390 { 1391 int i; 1392 1393 /* check whether the reserved field is zero */ 1394 for (i = 0; i < ARRAY_SIZE(events->reserved); i++) 1395 if (events->reserved[i]) 1396 return -EINVAL; 1397 1398 /* check whether the pad field is zero */ 1399 for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++) 1400 if (events->exception.pad[i]) 1401 return -EINVAL; 1402 1403 return __kvm_arm_vcpu_set_events(vcpu, events); 1404 } 1405 1406 long kvm_arch_vcpu_ioctl(struct file *filp, 1407 unsigned int ioctl, unsigned long arg) 1408 { 1409 struct kvm_vcpu *vcpu = filp->private_data; 1410 void __user *argp = (void __user *)arg; 1411 struct kvm_device_attr attr; 1412 long r; 1413 1414 switch (ioctl) { 1415 case KVM_ARM_VCPU_INIT: { 1416 struct kvm_vcpu_init init; 1417 1418 r = -EFAULT; 1419 if (copy_from_user(&init, argp, sizeof(init))) 1420 break; 1421 1422 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init); 1423 break; 1424 } 1425 case KVM_SET_ONE_REG: 1426 case KVM_GET_ONE_REG: { 1427 struct kvm_one_reg reg; 1428 1429 r = -ENOEXEC; 1430 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1431 break; 1432 1433 r = -EFAULT; 1434 if (copy_from_user(®, argp, sizeof(reg))) 1435 break; 1436 1437 /* 1438 * We could owe a reset due to PSCI. Handle the pending reset 1439 * here to ensure userspace register accesses are ordered after 1440 * the reset. 1441 */ 1442 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 1443 kvm_reset_vcpu(vcpu); 1444 1445 if (ioctl == KVM_SET_ONE_REG) 1446 r = kvm_arm_set_reg(vcpu, ®); 1447 else 1448 r = kvm_arm_get_reg(vcpu, ®); 1449 break; 1450 } 1451 case KVM_GET_REG_LIST: { 1452 struct kvm_reg_list __user *user_list = argp; 1453 struct kvm_reg_list reg_list; 1454 unsigned n; 1455 1456 r = -ENOEXEC; 1457 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1458 break; 1459 1460 r = -EPERM; 1461 if (!kvm_arm_vcpu_is_finalized(vcpu)) 1462 break; 1463 1464 r = -EFAULT; 1465 if (copy_from_user(®_list, user_list, sizeof(reg_list))) 1466 break; 1467 n = reg_list.n; 1468 reg_list.n = kvm_arm_num_regs(vcpu); 1469 if (copy_to_user(user_list, ®_list, sizeof(reg_list))) 1470 break; 1471 r = -E2BIG; 1472 if (n < reg_list.n) 1473 break; 1474 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg); 1475 break; 1476 } 1477 case KVM_SET_DEVICE_ATTR: { 1478 r = -EFAULT; 1479 if (copy_from_user(&attr, argp, sizeof(attr))) 1480 break; 1481 r = kvm_arm_vcpu_set_attr(vcpu, &attr); 1482 break; 1483 } 1484 case KVM_GET_DEVICE_ATTR: { 1485 r = -EFAULT; 1486 if (copy_from_user(&attr, argp, sizeof(attr))) 1487 break; 1488 r = kvm_arm_vcpu_get_attr(vcpu, &attr); 1489 break; 1490 } 1491 case KVM_HAS_DEVICE_ATTR: { 1492 r = -EFAULT; 1493 if (copy_from_user(&attr, argp, sizeof(attr))) 1494 break; 1495 r = kvm_arm_vcpu_has_attr(vcpu, &attr); 1496 break; 1497 } 1498 case KVM_GET_VCPU_EVENTS: { 1499 struct kvm_vcpu_events events; 1500 1501 if (kvm_arm_vcpu_get_events(vcpu, &events)) 1502 return -EINVAL; 1503 1504 if (copy_to_user(argp, &events, sizeof(events))) 1505 return -EFAULT; 1506 1507 return 0; 1508 } 1509 case KVM_SET_VCPU_EVENTS: { 1510 struct kvm_vcpu_events events; 1511 1512 if (copy_from_user(&events, argp, sizeof(events))) 1513 return -EFAULT; 1514 1515 return kvm_arm_vcpu_set_events(vcpu, &events); 1516 } 1517 case KVM_ARM_VCPU_FINALIZE: { 1518 int what; 1519 1520 if (!kvm_vcpu_initialized(vcpu)) 1521 return -ENOEXEC; 1522 1523 if (get_user(what, (const int __user *)argp)) 1524 return -EFAULT; 1525 1526 return kvm_arm_vcpu_finalize(vcpu, what); 1527 } 1528 default: 1529 r = -EINVAL; 1530 } 1531 1532 return r; 1533 } 1534 1535 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot) 1536 { 1537 1538 } 1539 1540 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm, 1541 struct kvm_arm_device_addr *dev_addr) 1542 { 1543 switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) { 1544 case KVM_ARM_DEVICE_VGIC_V2: 1545 if (!vgic_present) 1546 return -ENXIO; 1547 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr); 1548 default: 1549 return -ENODEV; 1550 } 1551 } 1552 1553 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1554 { 1555 switch (attr->group) { 1556 case KVM_ARM_VM_SMCCC_CTRL: 1557 return kvm_vm_smccc_has_attr(kvm, attr); 1558 default: 1559 return -ENXIO; 1560 } 1561 } 1562 1563 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1564 { 1565 switch (attr->group) { 1566 case KVM_ARM_VM_SMCCC_CTRL: 1567 return kvm_vm_smccc_set_attr(kvm, attr); 1568 default: 1569 return -ENXIO; 1570 } 1571 } 1572 1573 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) 1574 { 1575 struct kvm *kvm = filp->private_data; 1576 void __user *argp = (void __user *)arg; 1577 struct kvm_device_attr attr; 1578 1579 switch (ioctl) { 1580 case KVM_CREATE_IRQCHIP: { 1581 int ret; 1582 if (!vgic_present) 1583 return -ENXIO; 1584 mutex_lock(&kvm->lock); 1585 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2); 1586 mutex_unlock(&kvm->lock); 1587 return ret; 1588 } 1589 case KVM_ARM_SET_DEVICE_ADDR: { 1590 struct kvm_arm_device_addr dev_addr; 1591 1592 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr))) 1593 return -EFAULT; 1594 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr); 1595 } 1596 case KVM_ARM_PREFERRED_TARGET: { 1597 struct kvm_vcpu_init init = { 1598 .target = KVM_ARM_TARGET_GENERIC_V8, 1599 }; 1600 1601 if (copy_to_user(argp, &init, sizeof(init))) 1602 return -EFAULT; 1603 1604 return 0; 1605 } 1606 case KVM_ARM_MTE_COPY_TAGS: { 1607 struct kvm_arm_copy_mte_tags copy_tags; 1608 1609 if (copy_from_user(©_tags, argp, sizeof(copy_tags))) 1610 return -EFAULT; 1611 return kvm_vm_ioctl_mte_copy_tags(kvm, ©_tags); 1612 } 1613 case KVM_ARM_SET_COUNTER_OFFSET: { 1614 struct kvm_arm_counter_offset offset; 1615 1616 if (copy_from_user(&offset, argp, sizeof(offset))) 1617 return -EFAULT; 1618 return kvm_vm_ioctl_set_counter_offset(kvm, &offset); 1619 } 1620 case KVM_HAS_DEVICE_ATTR: { 1621 if (copy_from_user(&attr, argp, sizeof(attr))) 1622 return -EFAULT; 1623 1624 return kvm_vm_has_attr(kvm, &attr); 1625 } 1626 case KVM_SET_DEVICE_ATTR: { 1627 if (copy_from_user(&attr, argp, sizeof(attr))) 1628 return -EFAULT; 1629 1630 return kvm_vm_set_attr(kvm, &attr); 1631 } 1632 default: 1633 return -EINVAL; 1634 } 1635 } 1636 1637 /* unlocks vcpus from @vcpu_lock_idx and smaller */ 1638 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx) 1639 { 1640 struct kvm_vcpu *tmp_vcpu; 1641 1642 for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) { 1643 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx); 1644 mutex_unlock(&tmp_vcpu->mutex); 1645 } 1646 } 1647 1648 void unlock_all_vcpus(struct kvm *kvm) 1649 { 1650 lockdep_assert_held(&kvm->lock); 1651 1652 unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1); 1653 } 1654 1655 /* Returns true if all vcpus were locked, false otherwise */ 1656 bool lock_all_vcpus(struct kvm *kvm) 1657 { 1658 struct kvm_vcpu *tmp_vcpu; 1659 unsigned long c; 1660 1661 lockdep_assert_held(&kvm->lock); 1662 1663 /* 1664 * Any time a vcpu is in an ioctl (including running), the 1665 * core KVM code tries to grab the vcpu->mutex. 1666 * 1667 * By grabbing the vcpu->mutex of all VCPUs we ensure that no 1668 * other VCPUs can fiddle with the state while we access it. 1669 */ 1670 kvm_for_each_vcpu(c, tmp_vcpu, kvm) { 1671 if (!mutex_trylock(&tmp_vcpu->mutex)) { 1672 unlock_vcpus(kvm, c - 1); 1673 return false; 1674 } 1675 } 1676 1677 return true; 1678 } 1679 1680 static unsigned long nvhe_percpu_size(void) 1681 { 1682 return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) - 1683 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start); 1684 } 1685 1686 static unsigned long nvhe_percpu_order(void) 1687 { 1688 unsigned long size = nvhe_percpu_size(); 1689 1690 return size ? get_order(size) : 0; 1691 } 1692 1693 /* A lookup table holding the hypervisor VA for each vector slot */ 1694 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS]; 1695 1696 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot) 1697 { 1698 hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot); 1699 } 1700 1701 static int kvm_init_vector_slots(void) 1702 { 1703 int err; 1704 void *base; 1705 1706 base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector)); 1707 kvm_init_vector_slot(base, HYP_VECTOR_DIRECT); 1708 1709 base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs)); 1710 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT); 1711 1712 if (kvm_system_needs_idmapped_vectors() && 1713 !is_protected_kvm_enabled()) { 1714 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs), 1715 __BP_HARDEN_HYP_VECS_SZ, &base); 1716 if (err) 1717 return err; 1718 } 1719 1720 kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT); 1721 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT); 1722 return 0; 1723 } 1724 1725 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits) 1726 { 1727 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 1728 unsigned long tcr; 1729 1730 /* 1731 * Calculate the raw per-cpu offset without a translation from the 1732 * kernel's mapping to the linear mapping, and store it in tpidr_el2 1733 * so that we can use adr_l to access per-cpu variables in EL2. 1734 * Also drop the KASAN tag which gets in the way... 1735 */ 1736 params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) - 1737 (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start)); 1738 1739 params->mair_el2 = read_sysreg(mair_el1); 1740 1741 tcr = read_sysreg(tcr_el1); 1742 if (cpus_have_final_cap(ARM64_KVM_HVHE)) { 1743 tcr |= TCR_EPD1_MASK; 1744 } else { 1745 tcr &= TCR_EL2_MASK; 1746 tcr |= TCR_EL2_RES1; 1747 } 1748 tcr &= ~TCR_T0SZ_MASK; 1749 tcr |= TCR_T0SZ(hyp_va_bits); 1750 params->tcr_el2 = tcr; 1751 1752 params->pgd_pa = kvm_mmu_get_httbr(); 1753 if (is_protected_kvm_enabled()) 1754 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS; 1755 else 1756 params->hcr_el2 = HCR_HOST_NVHE_FLAGS; 1757 if (cpus_have_final_cap(ARM64_KVM_HVHE)) 1758 params->hcr_el2 |= HCR_E2H; 1759 params->vttbr = params->vtcr = 0; 1760 1761 /* 1762 * Flush the init params from the data cache because the struct will 1763 * be read while the MMU is off. 1764 */ 1765 kvm_flush_dcache_to_poc(params, sizeof(*params)); 1766 } 1767 1768 static void hyp_install_host_vector(void) 1769 { 1770 struct kvm_nvhe_init_params *params; 1771 struct arm_smccc_res res; 1772 1773 /* Switch from the HYP stub to our own HYP init vector */ 1774 __hyp_set_vectors(kvm_get_idmap_vector()); 1775 1776 /* 1777 * Call initialization code, and switch to the full blown HYP code. 1778 * If the cpucaps haven't been finalized yet, something has gone very 1779 * wrong, and hyp will crash and burn when it uses any 1780 * cpus_have_const_cap() wrapper. 1781 */ 1782 BUG_ON(!system_capabilities_finalized()); 1783 params = this_cpu_ptr_nvhe_sym(kvm_init_params); 1784 arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res); 1785 WARN_ON(res.a0 != SMCCC_RET_SUCCESS); 1786 } 1787 1788 static void cpu_init_hyp_mode(void) 1789 { 1790 hyp_install_host_vector(); 1791 1792 /* 1793 * Disabling SSBD on a non-VHE system requires us to enable SSBS 1794 * at EL2. 1795 */ 1796 if (this_cpu_has_cap(ARM64_SSBS) && 1797 arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) { 1798 kvm_call_hyp_nvhe(__kvm_enable_ssbs); 1799 } 1800 } 1801 1802 static void cpu_hyp_reset(void) 1803 { 1804 if (!is_kernel_in_hyp_mode()) 1805 __hyp_reset_vectors(); 1806 } 1807 1808 /* 1809 * EL2 vectors can be mapped and rerouted in a number of ways, 1810 * depending on the kernel configuration and CPU present: 1811 * 1812 * - If the CPU is affected by Spectre-v2, the hardening sequence is 1813 * placed in one of the vector slots, which is executed before jumping 1814 * to the real vectors. 1815 * 1816 * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot 1817 * containing the hardening sequence is mapped next to the idmap page, 1818 * and executed before jumping to the real vectors. 1819 * 1820 * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an 1821 * empty slot is selected, mapped next to the idmap page, and 1822 * executed before jumping to the real vectors. 1823 * 1824 * Note that ARM64_SPECTRE_V3A is somewhat incompatible with 1825 * VHE, as we don't have hypervisor-specific mappings. If the system 1826 * is VHE and yet selects this capability, it will be ignored. 1827 */ 1828 static void cpu_set_hyp_vector(void) 1829 { 1830 struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data); 1831 void *vector = hyp_spectre_vector_selector[data->slot]; 1832 1833 if (!is_protected_kvm_enabled()) 1834 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector; 1835 else 1836 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot); 1837 } 1838 1839 static void cpu_hyp_init_context(void) 1840 { 1841 kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt); 1842 1843 if (!is_kernel_in_hyp_mode()) 1844 cpu_init_hyp_mode(); 1845 } 1846 1847 static void cpu_hyp_init_features(void) 1848 { 1849 cpu_set_hyp_vector(); 1850 kvm_arm_init_debug(); 1851 1852 if (is_kernel_in_hyp_mode()) 1853 kvm_timer_init_vhe(); 1854 1855 if (vgic_present) 1856 kvm_vgic_init_cpu_hardware(); 1857 } 1858 1859 static void cpu_hyp_reinit(void) 1860 { 1861 cpu_hyp_reset(); 1862 cpu_hyp_init_context(); 1863 cpu_hyp_init_features(); 1864 } 1865 1866 static void cpu_hyp_init(void *discard) 1867 { 1868 if (!__this_cpu_read(kvm_hyp_initialized)) { 1869 cpu_hyp_reinit(); 1870 __this_cpu_write(kvm_hyp_initialized, 1); 1871 } 1872 } 1873 1874 static void cpu_hyp_uninit(void *discard) 1875 { 1876 if (__this_cpu_read(kvm_hyp_initialized)) { 1877 cpu_hyp_reset(); 1878 __this_cpu_write(kvm_hyp_initialized, 0); 1879 } 1880 } 1881 1882 int kvm_arch_hardware_enable(void) 1883 { 1884 /* 1885 * Most calls to this function are made with migration 1886 * disabled, but not with preemption disabled. The former is 1887 * enough to ensure correctness, but most of the helpers 1888 * expect the later and will throw a tantrum otherwise. 1889 */ 1890 preempt_disable(); 1891 1892 cpu_hyp_init(NULL); 1893 1894 kvm_vgic_cpu_up(); 1895 kvm_timer_cpu_up(); 1896 1897 preempt_enable(); 1898 1899 return 0; 1900 } 1901 1902 void kvm_arch_hardware_disable(void) 1903 { 1904 kvm_timer_cpu_down(); 1905 kvm_vgic_cpu_down(); 1906 1907 if (!is_protected_kvm_enabled()) 1908 cpu_hyp_uninit(NULL); 1909 } 1910 1911 #ifdef CONFIG_CPU_PM 1912 static int hyp_init_cpu_pm_notifier(struct notifier_block *self, 1913 unsigned long cmd, 1914 void *v) 1915 { 1916 /* 1917 * kvm_hyp_initialized is left with its old value over 1918 * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should 1919 * re-enable hyp. 1920 */ 1921 switch (cmd) { 1922 case CPU_PM_ENTER: 1923 if (__this_cpu_read(kvm_hyp_initialized)) 1924 /* 1925 * don't update kvm_hyp_initialized here 1926 * so that the hyp will be re-enabled 1927 * when we resume. See below. 1928 */ 1929 cpu_hyp_reset(); 1930 1931 return NOTIFY_OK; 1932 case CPU_PM_ENTER_FAILED: 1933 case CPU_PM_EXIT: 1934 if (__this_cpu_read(kvm_hyp_initialized)) 1935 /* The hyp was enabled before suspend. */ 1936 cpu_hyp_reinit(); 1937 1938 return NOTIFY_OK; 1939 1940 default: 1941 return NOTIFY_DONE; 1942 } 1943 } 1944 1945 static struct notifier_block hyp_init_cpu_pm_nb = { 1946 .notifier_call = hyp_init_cpu_pm_notifier, 1947 }; 1948 1949 static void __init hyp_cpu_pm_init(void) 1950 { 1951 if (!is_protected_kvm_enabled()) 1952 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb); 1953 } 1954 static void __init hyp_cpu_pm_exit(void) 1955 { 1956 if (!is_protected_kvm_enabled()) 1957 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb); 1958 } 1959 #else 1960 static inline void __init hyp_cpu_pm_init(void) 1961 { 1962 } 1963 static inline void __init hyp_cpu_pm_exit(void) 1964 { 1965 } 1966 #endif 1967 1968 static void __init init_cpu_logical_map(void) 1969 { 1970 unsigned int cpu; 1971 1972 /* 1973 * Copy the MPIDR <-> logical CPU ID mapping to hyp. 1974 * Only copy the set of online CPUs whose features have been checked 1975 * against the finalized system capabilities. The hypervisor will not 1976 * allow any other CPUs from the `possible` set to boot. 1977 */ 1978 for_each_online_cpu(cpu) 1979 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu); 1980 } 1981 1982 #define init_psci_0_1_impl_state(config, what) \ 1983 config.psci_0_1_ ## what ## _implemented = psci_ops.what 1984 1985 static bool __init init_psci_relay(void) 1986 { 1987 /* 1988 * If PSCI has not been initialized, protected KVM cannot install 1989 * itself on newly booted CPUs. 1990 */ 1991 if (!psci_ops.get_version) { 1992 kvm_err("Cannot initialize protected mode without PSCI\n"); 1993 return false; 1994 } 1995 1996 kvm_host_psci_config.version = psci_ops.get_version(); 1997 kvm_host_psci_config.smccc_version = arm_smccc_get_version(); 1998 1999 if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) { 2000 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids(); 2001 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend); 2002 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on); 2003 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off); 2004 init_psci_0_1_impl_state(kvm_host_psci_config, migrate); 2005 } 2006 return true; 2007 } 2008 2009 static int __init init_subsystems(void) 2010 { 2011 int err = 0; 2012 2013 /* 2014 * Enable hardware so that subsystem initialisation can access EL2. 2015 */ 2016 on_each_cpu(cpu_hyp_init, NULL, 1); 2017 2018 /* 2019 * Register CPU lower-power notifier 2020 */ 2021 hyp_cpu_pm_init(); 2022 2023 /* 2024 * Init HYP view of VGIC 2025 */ 2026 err = kvm_vgic_hyp_init(); 2027 switch (err) { 2028 case 0: 2029 vgic_present = true; 2030 break; 2031 case -ENODEV: 2032 case -ENXIO: 2033 vgic_present = false; 2034 err = 0; 2035 break; 2036 default: 2037 goto out; 2038 } 2039 2040 /* 2041 * Init HYP architected timer support 2042 */ 2043 err = kvm_timer_hyp_init(vgic_present); 2044 if (err) 2045 goto out; 2046 2047 kvm_register_perf_callbacks(NULL); 2048 2049 out: 2050 if (err) 2051 hyp_cpu_pm_exit(); 2052 2053 if (err || !is_protected_kvm_enabled()) 2054 on_each_cpu(cpu_hyp_uninit, NULL, 1); 2055 2056 return err; 2057 } 2058 2059 static void __init teardown_subsystems(void) 2060 { 2061 kvm_unregister_perf_callbacks(); 2062 hyp_cpu_pm_exit(); 2063 } 2064 2065 static void __init teardown_hyp_mode(void) 2066 { 2067 int cpu; 2068 2069 free_hyp_pgds(); 2070 for_each_possible_cpu(cpu) { 2071 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu)); 2072 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order()); 2073 } 2074 } 2075 2076 static int __init do_pkvm_init(u32 hyp_va_bits) 2077 { 2078 void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)); 2079 int ret; 2080 2081 preempt_disable(); 2082 cpu_hyp_init_context(); 2083 ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size, 2084 num_possible_cpus(), kern_hyp_va(per_cpu_base), 2085 hyp_va_bits); 2086 cpu_hyp_init_features(); 2087 2088 /* 2089 * The stub hypercalls are now disabled, so set our local flag to 2090 * prevent a later re-init attempt in kvm_arch_hardware_enable(). 2091 */ 2092 __this_cpu_write(kvm_hyp_initialized, 1); 2093 preempt_enable(); 2094 2095 return ret; 2096 } 2097 2098 static u64 get_hyp_id_aa64pfr0_el1(void) 2099 { 2100 /* 2101 * Track whether the system isn't affected by spectre/meltdown in the 2102 * hypervisor's view of id_aa64pfr0_el1, used for protected VMs. 2103 * Although this is per-CPU, we make it global for simplicity, e.g., not 2104 * to have to worry about vcpu migration. 2105 * 2106 * Unlike for non-protected VMs, userspace cannot override this for 2107 * protected VMs. 2108 */ 2109 u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1); 2110 2111 val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) | 2112 ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3)); 2113 2114 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), 2115 arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED); 2116 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), 2117 arm64_get_meltdown_state() == SPECTRE_UNAFFECTED); 2118 2119 return val; 2120 } 2121 2122 static void kvm_hyp_init_symbols(void) 2123 { 2124 kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1(); 2125 kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1); 2126 kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1); 2127 kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1); 2128 kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1); 2129 kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1); 2130 kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1); 2131 kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1); 2132 kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1); 2133 kvm_nvhe_sym(__icache_flags) = __icache_flags; 2134 kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits; 2135 } 2136 2137 static int __init kvm_hyp_init_protection(u32 hyp_va_bits) 2138 { 2139 void *addr = phys_to_virt(hyp_mem_base); 2140 int ret; 2141 2142 ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP); 2143 if (ret) 2144 return ret; 2145 2146 ret = do_pkvm_init(hyp_va_bits); 2147 if (ret) 2148 return ret; 2149 2150 free_hyp_pgds(); 2151 2152 return 0; 2153 } 2154 2155 static void pkvm_hyp_init_ptrauth(void) 2156 { 2157 struct kvm_cpu_context *hyp_ctxt; 2158 int cpu; 2159 2160 for_each_possible_cpu(cpu) { 2161 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu); 2162 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long(); 2163 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long(); 2164 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long(); 2165 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long(); 2166 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long(); 2167 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long(); 2168 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long(); 2169 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long(); 2170 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long(); 2171 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long(); 2172 } 2173 } 2174 2175 /* Inits Hyp-mode on all online CPUs */ 2176 static int __init init_hyp_mode(void) 2177 { 2178 u32 hyp_va_bits; 2179 int cpu; 2180 int err = -ENOMEM; 2181 2182 /* 2183 * The protected Hyp-mode cannot be initialized if the memory pool 2184 * allocation has failed. 2185 */ 2186 if (is_protected_kvm_enabled() && !hyp_mem_base) 2187 goto out_err; 2188 2189 /* 2190 * Allocate Hyp PGD and setup Hyp identity mapping 2191 */ 2192 err = kvm_mmu_init(&hyp_va_bits); 2193 if (err) 2194 goto out_err; 2195 2196 /* 2197 * Allocate stack pages for Hypervisor-mode 2198 */ 2199 for_each_possible_cpu(cpu) { 2200 unsigned long stack_page; 2201 2202 stack_page = __get_free_page(GFP_KERNEL); 2203 if (!stack_page) { 2204 err = -ENOMEM; 2205 goto out_err; 2206 } 2207 2208 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page; 2209 } 2210 2211 /* 2212 * Allocate and initialize pages for Hypervisor-mode percpu regions. 2213 */ 2214 for_each_possible_cpu(cpu) { 2215 struct page *page; 2216 void *page_addr; 2217 2218 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order()); 2219 if (!page) { 2220 err = -ENOMEM; 2221 goto out_err; 2222 } 2223 2224 page_addr = page_address(page); 2225 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size()); 2226 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr; 2227 } 2228 2229 /* 2230 * Map the Hyp-code called directly from the host 2231 */ 2232 err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start), 2233 kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC); 2234 if (err) { 2235 kvm_err("Cannot map world-switch code\n"); 2236 goto out_err; 2237 } 2238 2239 err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start), 2240 kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO); 2241 if (err) { 2242 kvm_err("Cannot map .hyp.rodata section\n"); 2243 goto out_err; 2244 } 2245 2246 err = create_hyp_mappings(kvm_ksym_ref(__start_rodata), 2247 kvm_ksym_ref(__end_rodata), PAGE_HYP_RO); 2248 if (err) { 2249 kvm_err("Cannot map rodata section\n"); 2250 goto out_err; 2251 } 2252 2253 /* 2254 * .hyp.bss is guaranteed to be placed at the beginning of the .bss 2255 * section thanks to an assertion in the linker script. Map it RW and 2256 * the rest of .bss RO. 2257 */ 2258 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start), 2259 kvm_ksym_ref(__hyp_bss_end), PAGE_HYP); 2260 if (err) { 2261 kvm_err("Cannot map hyp bss section: %d\n", err); 2262 goto out_err; 2263 } 2264 2265 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end), 2266 kvm_ksym_ref(__bss_stop), PAGE_HYP_RO); 2267 if (err) { 2268 kvm_err("Cannot map bss section\n"); 2269 goto out_err; 2270 } 2271 2272 /* 2273 * Map the Hyp stack pages 2274 */ 2275 for_each_possible_cpu(cpu) { 2276 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 2277 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu); 2278 2279 err = create_hyp_stack(__pa(stack_page), ¶ms->stack_hyp_va); 2280 if (err) { 2281 kvm_err("Cannot map hyp stack\n"); 2282 goto out_err; 2283 } 2284 2285 /* 2286 * Save the stack PA in nvhe_init_params. This will be needed 2287 * to recreate the stack mapping in protected nVHE mode. 2288 * __hyp_pa() won't do the right thing there, since the stack 2289 * has been mapped in the flexible private VA space. 2290 */ 2291 params->stack_pa = __pa(stack_page); 2292 } 2293 2294 for_each_possible_cpu(cpu) { 2295 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu]; 2296 char *percpu_end = percpu_begin + nvhe_percpu_size(); 2297 2298 /* Map Hyp percpu pages */ 2299 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP); 2300 if (err) { 2301 kvm_err("Cannot map hyp percpu region\n"); 2302 goto out_err; 2303 } 2304 2305 /* Prepare the CPU initialization parameters */ 2306 cpu_prepare_hyp_mode(cpu, hyp_va_bits); 2307 } 2308 2309 kvm_hyp_init_symbols(); 2310 2311 if (is_protected_kvm_enabled()) { 2312 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) && 2313 cpus_have_const_cap(ARM64_HAS_ADDRESS_AUTH)) 2314 pkvm_hyp_init_ptrauth(); 2315 2316 init_cpu_logical_map(); 2317 2318 if (!init_psci_relay()) { 2319 err = -ENODEV; 2320 goto out_err; 2321 } 2322 2323 err = kvm_hyp_init_protection(hyp_va_bits); 2324 if (err) { 2325 kvm_err("Failed to init hyp memory protection\n"); 2326 goto out_err; 2327 } 2328 } 2329 2330 return 0; 2331 2332 out_err: 2333 teardown_hyp_mode(); 2334 kvm_err("error initializing Hyp mode: %d\n", err); 2335 return err; 2336 } 2337 2338 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr) 2339 { 2340 struct kvm_vcpu *vcpu; 2341 unsigned long i; 2342 2343 mpidr &= MPIDR_HWID_BITMASK; 2344 kvm_for_each_vcpu(i, vcpu, kvm) { 2345 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu)) 2346 return vcpu; 2347 } 2348 return NULL; 2349 } 2350 2351 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm) 2352 { 2353 return irqchip_in_kernel(kvm); 2354 } 2355 2356 bool kvm_arch_has_irq_bypass(void) 2357 { 2358 return true; 2359 } 2360 2361 int kvm_arch_irq_bypass_add_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 return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq, 2368 &irqfd->irq_entry); 2369 } 2370 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons, 2371 struct irq_bypass_producer *prod) 2372 { 2373 struct kvm_kernel_irqfd *irqfd = 2374 container_of(cons, struct kvm_kernel_irqfd, consumer); 2375 2376 kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq, 2377 &irqfd->irq_entry); 2378 } 2379 2380 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons) 2381 { 2382 struct kvm_kernel_irqfd *irqfd = 2383 container_of(cons, struct kvm_kernel_irqfd, consumer); 2384 2385 kvm_arm_halt_guest(irqfd->kvm); 2386 } 2387 2388 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons) 2389 { 2390 struct kvm_kernel_irqfd *irqfd = 2391 container_of(cons, struct kvm_kernel_irqfd, consumer); 2392 2393 kvm_arm_resume_guest(irqfd->kvm); 2394 } 2395 2396 /* Initialize Hyp-mode and memory mappings on all CPUs */ 2397 static __init int kvm_arm_init(void) 2398 { 2399 int err; 2400 bool in_hyp_mode; 2401 2402 if (!is_hyp_mode_available()) { 2403 kvm_info("HYP mode not available\n"); 2404 return -ENODEV; 2405 } 2406 2407 if (kvm_get_mode() == KVM_MODE_NONE) { 2408 kvm_info("KVM disabled from command line\n"); 2409 return -ENODEV; 2410 } 2411 2412 err = kvm_sys_reg_table_init(); 2413 if (err) { 2414 kvm_info("Error initializing system register tables"); 2415 return err; 2416 } 2417 2418 in_hyp_mode = is_kernel_in_hyp_mode(); 2419 2420 if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) || 2421 cpus_have_final_cap(ARM64_WORKAROUND_1508412)) 2422 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \ 2423 "Only trusted guests should be used on this system.\n"); 2424 2425 err = kvm_set_ipa_limit(); 2426 if (err) 2427 return err; 2428 2429 err = kvm_arm_init_sve(); 2430 if (err) 2431 return err; 2432 2433 err = kvm_arm_vmid_alloc_init(); 2434 if (err) { 2435 kvm_err("Failed to initialize VMID allocator.\n"); 2436 return err; 2437 } 2438 2439 if (!in_hyp_mode) { 2440 err = init_hyp_mode(); 2441 if (err) 2442 goto out_err; 2443 } 2444 2445 err = kvm_init_vector_slots(); 2446 if (err) { 2447 kvm_err("Cannot initialise vector slots\n"); 2448 goto out_hyp; 2449 } 2450 2451 err = init_subsystems(); 2452 if (err) 2453 goto out_hyp; 2454 2455 if (is_protected_kvm_enabled()) { 2456 kvm_info("Protected nVHE mode initialized successfully\n"); 2457 } else if (in_hyp_mode) { 2458 kvm_info("VHE mode initialized successfully\n"); 2459 } else { 2460 kvm_info("Hyp mode initialized successfully\n"); 2461 } 2462 2463 /* 2464 * FIXME: Do something reasonable if kvm_init() fails after pKVM 2465 * hypervisor protection is finalized. 2466 */ 2467 err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE); 2468 if (err) 2469 goto out_subs; 2470 2471 kvm_arm_initialised = true; 2472 2473 return 0; 2474 2475 out_subs: 2476 teardown_subsystems(); 2477 out_hyp: 2478 if (!in_hyp_mode) 2479 teardown_hyp_mode(); 2480 out_err: 2481 kvm_arm_vmid_alloc_free(); 2482 return err; 2483 } 2484 2485 static int __init early_kvm_mode_cfg(char *arg) 2486 { 2487 if (!arg) 2488 return -EINVAL; 2489 2490 if (strcmp(arg, "none") == 0) { 2491 kvm_mode = KVM_MODE_NONE; 2492 return 0; 2493 } 2494 2495 if (!is_hyp_mode_available()) { 2496 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n"); 2497 return 0; 2498 } 2499 2500 if (strcmp(arg, "protected") == 0) { 2501 if (!is_kernel_in_hyp_mode()) 2502 kvm_mode = KVM_MODE_PROTECTED; 2503 else 2504 pr_warn_once("Protected KVM not available with VHE\n"); 2505 2506 return 0; 2507 } 2508 2509 if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) { 2510 kvm_mode = KVM_MODE_DEFAULT; 2511 return 0; 2512 } 2513 2514 if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) { 2515 kvm_mode = KVM_MODE_NV; 2516 return 0; 2517 } 2518 2519 return -EINVAL; 2520 } 2521 early_param("kvm-arm.mode", early_kvm_mode_cfg); 2522 2523 enum kvm_mode kvm_get_mode(void) 2524 { 2525 return kvm_mode; 2526 } 2527 2528 module_init(kvm_arm_init); 2529