1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2012 - Virtual Open Systems and Columbia University 4 * Author: Christoffer Dall <c.dall@virtualopensystems.com> 5 */ 6 7 #include <linux/bug.h> 8 #include <linux/cpu_pm.h> 9 #include <linux/entry-kvm.h> 10 #include <linux/errno.h> 11 #include <linux/err.h> 12 #include <linux/kvm_host.h> 13 #include <linux/list.h> 14 #include <linux/module.h> 15 #include <linux/vmalloc.h> 16 #include <linux/fs.h> 17 #include <linux/mman.h> 18 #include <linux/sched.h> 19 #include <linux/kvm.h> 20 #include <linux/kvm_irqfd.h> 21 #include <linux/irqbypass.h> 22 #include <linux/sched/stat.h> 23 #include <linux/psci.h> 24 #include <trace/events/kvm.h> 25 26 #define CREATE_TRACE_POINTS 27 #include "trace_arm.h" 28 29 #include <linux/uaccess.h> 30 #include <asm/ptrace.h> 31 #include <asm/mman.h> 32 #include <asm/tlbflush.h> 33 #include <asm/cacheflush.h> 34 #include <asm/cpufeature.h> 35 #include <asm/virt.h> 36 #include <asm/kvm_arm.h> 37 #include <asm/kvm_asm.h> 38 #include <asm/kvm_mmu.h> 39 #include <asm/kvm_pkvm.h> 40 #include <asm/kvm_emulate.h> 41 #include <asm/sections.h> 42 43 #include <kvm/arm_hypercalls.h> 44 #include <kvm/arm_pmu.h> 45 #include <kvm/arm_psci.h> 46 47 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT; 48 49 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector); 50 51 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_page); 52 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params); 53 54 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt); 55 56 static bool vgic_present, kvm_arm_initialised; 57 58 static DEFINE_PER_CPU(unsigned char, kvm_arm_hardware_enabled); 59 DEFINE_STATIC_KEY_FALSE(userspace_irqchip_in_use); 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->arch.target = -1; 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 if (vcpu_has_run_once(vcpu) && unlikely(!irqchip_in_kernel(vcpu->kvm))) 404 static_branch_dec(&userspace_irqchip_in_use); 405 406 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache); 407 kvm_timer_vcpu_terminate(vcpu); 408 kvm_pmu_vcpu_destroy(vcpu); 409 410 kvm_arm_vcpu_destroy(vcpu); 411 } 412 413 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu) 414 { 415 416 } 417 418 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu) 419 { 420 421 } 422 423 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) 424 { 425 struct kvm_s2_mmu *mmu; 426 int *last_ran; 427 428 mmu = vcpu->arch.hw_mmu; 429 last_ran = this_cpu_ptr(mmu->last_vcpu_ran); 430 431 /* 432 * We guarantee that both TLBs and I-cache are private to each 433 * vcpu. If detecting that a vcpu from the same VM has 434 * previously run on the same physical CPU, call into the 435 * hypervisor code to nuke the relevant contexts. 436 * 437 * We might get preempted before the vCPU actually runs, but 438 * over-invalidation doesn't affect correctness. 439 */ 440 if (*last_ran != vcpu->vcpu_id) { 441 kvm_call_hyp(__kvm_flush_cpu_context, mmu); 442 *last_ran = vcpu->vcpu_id; 443 } 444 445 vcpu->cpu = cpu; 446 447 kvm_vgic_load(vcpu); 448 kvm_timer_vcpu_load(vcpu); 449 if (has_vhe()) 450 kvm_vcpu_load_sysregs_vhe(vcpu); 451 kvm_arch_vcpu_load_fp(vcpu); 452 kvm_vcpu_pmu_restore_guest(vcpu); 453 if (kvm_arm_is_pvtime_enabled(&vcpu->arch)) 454 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu); 455 456 if (single_task_running()) 457 vcpu_clear_wfx_traps(vcpu); 458 else 459 vcpu_set_wfx_traps(vcpu); 460 461 if (vcpu_has_ptrauth(vcpu)) 462 vcpu_ptrauth_disable(vcpu); 463 kvm_arch_vcpu_load_debug_state_flags(vcpu); 464 465 if (!cpumask_test_cpu(smp_processor_id(), vcpu->kvm->arch.supported_cpus)) 466 vcpu_set_on_unsupported_cpu(vcpu); 467 } 468 469 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu) 470 { 471 kvm_arch_vcpu_put_debug_state_flags(vcpu); 472 kvm_arch_vcpu_put_fp(vcpu); 473 if (has_vhe()) 474 kvm_vcpu_put_sysregs_vhe(vcpu); 475 kvm_timer_vcpu_put(vcpu); 476 kvm_vgic_put(vcpu); 477 kvm_vcpu_pmu_restore_host(vcpu); 478 kvm_arm_vmid_clear_active(); 479 480 vcpu_clear_on_unsupported_cpu(vcpu); 481 vcpu->cpu = -1; 482 } 483 484 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 485 { 486 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED); 487 kvm_make_request(KVM_REQ_SLEEP, vcpu); 488 kvm_vcpu_kick(vcpu); 489 } 490 491 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 492 { 493 spin_lock(&vcpu->arch.mp_state_lock); 494 __kvm_arm_vcpu_power_off(vcpu); 495 spin_unlock(&vcpu->arch.mp_state_lock); 496 } 497 498 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu) 499 { 500 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED; 501 } 502 503 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu) 504 { 505 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED); 506 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 507 kvm_vcpu_kick(vcpu); 508 } 509 510 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu) 511 { 512 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED; 513 } 514 515 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu, 516 struct kvm_mp_state *mp_state) 517 { 518 *mp_state = READ_ONCE(vcpu->arch.mp_state); 519 520 return 0; 521 } 522 523 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu, 524 struct kvm_mp_state *mp_state) 525 { 526 int ret = 0; 527 528 spin_lock(&vcpu->arch.mp_state_lock); 529 530 switch (mp_state->mp_state) { 531 case KVM_MP_STATE_RUNNABLE: 532 WRITE_ONCE(vcpu->arch.mp_state, *mp_state); 533 break; 534 case KVM_MP_STATE_STOPPED: 535 __kvm_arm_vcpu_power_off(vcpu); 536 break; 537 case KVM_MP_STATE_SUSPENDED: 538 kvm_arm_vcpu_suspend(vcpu); 539 break; 540 default: 541 ret = -EINVAL; 542 } 543 544 spin_unlock(&vcpu->arch.mp_state_lock); 545 546 return ret; 547 } 548 549 /** 550 * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled 551 * @v: The VCPU pointer 552 * 553 * If the guest CPU is not waiting for interrupts or an interrupt line is 554 * asserted, the CPU is by definition runnable. 555 */ 556 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v) 557 { 558 bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF); 559 return ((irq_lines || kvm_vgic_vcpu_pending_irq(v)) 560 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause); 561 } 562 563 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu) 564 { 565 return vcpu_mode_priv(vcpu); 566 } 567 568 #ifdef CONFIG_GUEST_PERF_EVENTS 569 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu) 570 { 571 return *vcpu_pc(vcpu); 572 } 573 #endif 574 575 static int kvm_vcpu_initialized(struct kvm_vcpu *vcpu) 576 { 577 return vcpu->arch.target >= 0; 578 } 579 580 /* 581 * Handle both the initialisation that is being done when the vcpu is 582 * run for the first time, as well as the updates that must be 583 * performed each time we get a new thread dealing with this vcpu. 584 */ 585 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu) 586 { 587 struct kvm *kvm = vcpu->kvm; 588 int ret; 589 590 if (!kvm_vcpu_initialized(vcpu)) 591 return -ENOEXEC; 592 593 if (!kvm_arm_vcpu_is_finalized(vcpu)) 594 return -EPERM; 595 596 ret = kvm_arch_vcpu_run_map_fp(vcpu); 597 if (ret) 598 return ret; 599 600 if (likely(vcpu_has_run_once(vcpu))) 601 return 0; 602 603 kvm_arm_vcpu_init_debug(vcpu); 604 605 if (likely(irqchip_in_kernel(kvm))) { 606 /* 607 * Map the VGIC hardware resources before running a vcpu the 608 * first time on this VM. 609 */ 610 ret = kvm_vgic_map_resources(kvm); 611 if (ret) 612 return ret; 613 } 614 615 ret = kvm_timer_enable(vcpu); 616 if (ret) 617 return ret; 618 619 ret = kvm_arm_pmu_v3_enable(vcpu); 620 if (ret) 621 return ret; 622 623 if (is_protected_kvm_enabled()) { 624 ret = pkvm_create_hyp_vm(kvm); 625 if (ret) 626 return ret; 627 } 628 629 if (!irqchip_in_kernel(kvm)) { 630 /* 631 * Tell the rest of the code that there are userspace irqchip 632 * VMs in the wild. 633 */ 634 static_branch_inc(&userspace_irqchip_in_use); 635 } 636 637 /* 638 * Initialize traps for protected VMs. 639 * NOTE: Move to run in EL2 directly, rather than via a hypercall, once 640 * the code is in place for first run initialization at EL2. 641 */ 642 if (kvm_vm_is_protected(kvm)) 643 kvm_call_hyp_nvhe(__pkvm_vcpu_init_traps, vcpu); 644 645 mutex_lock(&kvm->arch.config_lock); 646 set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags); 647 mutex_unlock(&kvm->arch.config_lock); 648 649 return ret; 650 } 651 652 bool kvm_arch_intc_initialized(struct kvm *kvm) 653 { 654 return vgic_initialized(kvm); 655 } 656 657 void kvm_arm_halt_guest(struct kvm *kvm) 658 { 659 unsigned long i; 660 struct kvm_vcpu *vcpu; 661 662 kvm_for_each_vcpu(i, vcpu, kvm) 663 vcpu->arch.pause = true; 664 kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP); 665 } 666 667 void kvm_arm_resume_guest(struct kvm *kvm) 668 { 669 unsigned long i; 670 struct kvm_vcpu *vcpu; 671 672 kvm_for_each_vcpu(i, vcpu, kvm) { 673 vcpu->arch.pause = false; 674 __kvm_vcpu_wake_up(vcpu); 675 } 676 } 677 678 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu) 679 { 680 struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); 681 682 rcuwait_wait_event(wait, 683 (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause), 684 TASK_INTERRUPTIBLE); 685 686 if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) { 687 /* Awaken to handle a signal, request we sleep again later. */ 688 kvm_make_request(KVM_REQ_SLEEP, vcpu); 689 } 690 691 /* 692 * Make sure we will observe a potential reset request if we've 693 * observed a change to the power state. Pairs with the smp_wmb() in 694 * kvm_psci_vcpu_on(). 695 */ 696 smp_rmb(); 697 } 698 699 /** 700 * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior 701 * @vcpu: The VCPU pointer 702 * 703 * Suspend execution of a vCPU until a valid wake event is detected, i.e. until 704 * the vCPU is runnable. The vCPU may or may not be scheduled out, depending 705 * on when a wake event arrives, e.g. there may already be a pending wake event. 706 */ 707 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu) 708 { 709 /* 710 * Sync back the state of the GIC CPU interface so that we have 711 * the latest PMR and group enables. This ensures that 712 * kvm_arch_vcpu_runnable has up-to-date data to decide whether 713 * we have pending interrupts, e.g. when determining if the 714 * vCPU should block. 715 * 716 * For the same reason, we want to tell GICv4 that we need 717 * doorbells to be signalled, should an interrupt become pending. 718 */ 719 preempt_disable(); 720 kvm_vgic_vmcr_sync(vcpu); 721 vcpu_set_flag(vcpu, IN_WFI); 722 vgic_v4_put(vcpu); 723 preempt_enable(); 724 725 kvm_vcpu_halt(vcpu); 726 vcpu_clear_flag(vcpu, IN_WFIT); 727 728 preempt_disable(); 729 vcpu_clear_flag(vcpu, IN_WFI); 730 vgic_v4_load(vcpu); 731 preempt_enable(); 732 } 733 734 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu) 735 { 736 if (!kvm_arm_vcpu_suspended(vcpu)) 737 return 1; 738 739 kvm_vcpu_wfi(vcpu); 740 741 /* 742 * The suspend state is sticky; we do not leave it until userspace 743 * explicitly marks the vCPU as runnable. Request that we suspend again 744 * later. 745 */ 746 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 747 748 /* 749 * Check to make sure the vCPU is actually runnable. If so, exit to 750 * userspace informing it of the wakeup condition. 751 */ 752 if (kvm_arch_vcpu_runnable(vcpu)) { 753 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event)); 754 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP; 755 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT; 756 return 0; 757 } 758 759 /* 760 * Otherwise, we were unblocked to process a different event, such as a 761 * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to 762 * process the event. 763 */ 764 return 1; 765 } 766 767 /** 768 * check_vcpu_requests - check and handle pending vCPU requests 769 * @vcpu: the VCPU pointer 770 * 771 * Return: 1 if we should enter the guest 772 * 0 if we should exit to userspace 773 * < 0 if we should exit to userspace, where the return value indicates 774 * an error 775 */ 776 static int check_vcpu_requests(struct kvm_vcpu *vcpu) 777 { 778 if (kvm_request_pending(vcpu)) { 779 if (kvm_check_request(KVM_REQ_SLEEP, vcpu)) 780 kvm_vcpu_sleep(vcpu); 781 782 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 783 kvm_reset_vcpu(vcpu); 784 785 /* 786 * Clear IRQ_PENDING requests that were made to guarantee 787 * that a VCPU sees new virtual interrupts. 788 */ 789 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu); 790 791 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu)) 792 kvm_update_stolen_time(vcpu); 793 794 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) { 795 /* The distributor enable bits were changed */ 796 preempt_disable(); 797 vgic_v4_put(vcpu); 798 vgic_v4_load(vcpu); 799 preempt_enable(); 800 } 801 802 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu)) 803 kvm_pmu_handle_pmcr(vcpu, 804 __vcpu_sys_reg(vcpu, PMCR_EL0)); 805 806 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu)) 807 return kvm_vcpu_suspend(vcpu); 808 809 if (kvm_dirty_ring_check_request(vcpu)) 810 return 0; 811 } 812 813 return 1; 814 } 815 816 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) 817 { 818 if (likely(!vcpu_mode_is_32bit(vcpu))) 819 return false; 820 821 return !kvm_supports_32bit_el0(); 822 } 823 824 /** 825 * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest 826 * @vcpu: The VCPU pointer 827 * @ret: Pointer to write optional return code 828 * 829 * Returns: true if the VCPU needs to return to a preemptible + interruptible 830 * and skip guest entry. 831 * 832 * This function disambiguates between two different types of exits: exits to a 833 * preemptible + interruptible kernel context and exits to userspace. For an 834 * exit to userspace, this function will write the return code to ret and return 835 * true. For an exit to preemptible + interruptible kernel context (i.e. check 836 * for pending work and re-enter), return true without writing to ret. 837 */ 838 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret) 839 { 840 struct kvm_run *run = vcpu->run; 841 842 /* 843 * If we're using a userspace irqchip, then check if we need 844 * to tell a userspace irqchip about timer or PMU level 845 * changes and if so, exit to userspace (the actual level 846 * state gets updated in kvm_timer_update_run and 847 * kvm_pmu_update_run below). 848 */ 849 if (static_branch_unlikely(&userspace_irqchip_in_use)) { 850 if (kvm_timer_should_notify_user(vcpu) || 851 kvm_pmu_should_notify_user(vcpu)) { 852 *ret = -EINTR; 853 run->exit_reason = KVM_EXIT_INTR; 854 return true; 855 } 856 } 857 858 if (unlikely(vcpu_on_unsupported_cpu(vcpu))) { 859 run->exit_reason = KVM_EXIT_FAIL_ENTRY; 860 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED; 861 run->fail_entry.cpu = smp_processor_id(); 862 *ret = 0; 863 return true; 864 } 865 866 return kvm_request_pending(vcpu) || 867 xfer_to_guest_mode_work_pending(); 868 } 869 870 /* 871 * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while 872 * the vCPU is running. 873 * 874 * This must be noinstr as instrumentation may make use of RCU, and this is not 875 * safe during the EQS. 876 */ 877 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu) 878 { 879 int ret; 880 881 guest_state_enter_irqoff(); 882 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu); 883 guest_state_exit_irqoff(); 884 885 return ret; 886 } 887 888 /** 889 * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code 890 * @vcpu: The VCPU pointer 891 * 892 * This function is called through the VCPU_RUN ioctl called from user space. It 893 * will execute VM code in a loop until the time slice for the process is used 894 * or some emulation is needed from user space in which case the function will 895 * return with return value 0 and with the kvm_run structure filled in with the 896 * required data for the requested emulation. 897 */ 898 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) 899 { 900 struct kvm_run *run = vcpu->run; 901 int ret; 902 903 if (run->exit_reason == KVM_EXIT_MMIO) { 904 ret = kvm_handle_mmio_return(vcpu); 905 if (ret) 906 return ret; 907 } 908 909 vcpu_load(vcpu); 910 911 if (run->immediate_exit) { 912 ret = -EINTR; 913 goto out; 914 } 915 916 kvm_sigset_activate(vcpu); 917 918 ret = 1; 919 run->exit_reason = KVM_EXIT_UNKNOWN; 920 run->flags = 0; 921 while (ret > 0) { 922 /* 923 * Check conditions before entering the guest 924 */ 925 ret = xfer_to_guest_mode_handle_work(vcpu); 926 if (!ret) 927 ret = 1; 928 929 if (ret > 0) 930 ret = check_vcpu_requests(vcpu); 931 932 /* 933 * Preparing the interrupts to be injected also 934 * involves poking the GIC, which must be done in a 935 * non-preemptible context. 936 */ 937 preempt_disable(); 938 939 /* 940 * The VMID allocator only tracks active VMIDs per 941 * physical CPU, and therefore the VMID allocated may not be 942 * preserved on VMID roll-over if the task was preempted, 943 * making a thread's VMID inactive. So we need to call 944 * kvm_arm_vmid_update() in non-premptible context. 945 */ 946 kvm_arm_vmid_update(&vcpu->arch.hw_mmu->vmid); 947 948 kvm_pmu_flush_hwstate(vcpu); 949 950 local_irq_disable(); 951 952 kvm_vgic_flush_hwstate(vcpu); 953 954 kvm_pmu_update_vcpu_events(vcpu); 955 956 /* 957 * Ensure we set mode to IN_GUEST_MODE after we disable 958 * interrupts and before the final VCPU requests check. 959 * See the comment in kvm_vcpu_exiting_guest_mode() and 960 * Documentation/virt/kvm/vcpu-requests.rst 961 */ 962 smp_store_mb(vcpu->mode, IN_GUEST_MODE); 963 964 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) { 965 vcpu->mode = OUTSIDE_GUEST_MODE; 966 isb(); /* Ensure work in x_flush_hwstate is committed */ 967 kvm_pmu_sync_hwstate(vcpu); 968 if (static_branch_unlikely(&userspace_irqchip_in_use)) 969 kvm_timer_sync_user(vcpu); 970 kvm_vgic_sync_hwstate(vcpu); 971 local_irq_enable(); 972 preempt_enable(); 973 continue; 974 } 975 976 kvm_arm_setup_debug(vcpu); 977 kvm_arch_vcpu_ctxflush_fp(vcpu); 978 979 /************************************************************** 980 * Enter the guest 981 */ 982 trace_kvm_entry(*vcpu_pc(vcpu)); 983 guest_timing_enter_irqoff(); 984 985 ret = kvm_arm_vcpu_enter_exit(vcpu); 986 987 vcpu->mode = OUTSIDE_GUEST_MODE; 988 vcpu->stat.exits++; 989 /* 990 * Back from guest 991 *************************************************************/ 992 993 kvm_arm_clear_debug(vcpu); 994 995 /* 996 * We must sync the PMU state before the vgic state so 997 * that the vgic can properly sample the updated state of the 998 * interrupt line. 999 */ 1000 kvm_pmu_sync_hwstate(vcpu); 1001 1002 /* 1003 * Sync the vgic state before syncing the timer state because 1004 * the timer code needs to know if the virtual timer 1005 * interrupts are active. 1006 */ 1007 kvm_vgic_sync_hwstate(vcpu); 1008 1009 /* 1010 * Sync the timer hardware state before enabling interrupts as 1011 * we don't want vtimer interrupts to race with syncing the 1012 * timer virtual interrupt state. 1013 */ 1014 if (static_branch_unlikely(&userspace_irqchip_in_use)) 1015 kvm_timer_sync_user(vcpu); 1016 1017 kvm_arch_vcpu_ctxsync_fp(vcpu); 1018 1019 /* 1020 * We must ensure that any pending interrupts are taken before 1021 * we exit guest timing so that timer ticks are accounted as 1022 * guest time. Transiently unmask interrupts so that any 1023 * pending interrupts are taken. 1024 * 1025 * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other 1026 * context synchronization event) is necessary to ensure that 1027 * pending interrupts are taken. 1028 */ 1029 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) { 1030 local_irq_enable(); 1031 isb(); 1032 local_irq_disable(); 1033 } 1034 1035 guest_timing_exit_irqoff(); 1036 1037 local_irq_enable(); 1038 1039 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu)); 1040 1041 /* Exit types that need handling before we can be preempted */ 1042 handle_exit_early(vcpu, ret); 1043 1044 preempt_enable(); 1045 1046 /* 1047 * The ARMv8 architecture doesn't give the hypervisor 1048 * a mechanism to prevent a guest from dropping to AArch32 EL0 1049 * if implemented by the CPU. If we spot the guest in such 1050 * state and that we decided it wasn't supposed to do so (like 1051 * with the asymmetric AArch32 case), return to userspace with 1052 * a fatal error. 1053 */ 1054 if (vcpu_mode_is_bad_32bit(vcpu)) { 1055 /* 1056 * As we have caught the guest red-handed, decide that 1057 * it isn't fit for purpose anymore by making the vcpu 1058 * invalid. The VMM can try and fix it by issuing a 1059 * KVM_ARM_VCPU_INIT if it really wants to. 1060 */ 1061 vcpu->arch.target = -1; 1062 ret = ARM_EXCEPTION_IL; 1063 } 1064 1065 ret = handle_exit(vcpu, ret); 1066 } 1067 1068 /* Tell userspace about in-kernel device output levels */ 1069 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) { 1070 kvm_timer_update_run(vcpu); 1071 kvm_pmu_update_run(vcpu); 1072 } 1073 1074 kvm_sigset_deactivate(vcpu); 1075 1076 out: 1077 /* 1078 * In the unlikely event that we are returning to userspace 1079 * with pending exceptions or PC adjustment, commit these 1080 * adjustments in order to give userspace a consistent view of 1081 * the vcpu state. Note that this relies on __kvm_adjust_pc() 1082 * being preempt-safe on VHE. 1083 */ 1084 if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) || 1085 vcpu_get_flag(vcpu, INCREMENT_PC))) 1086 kvm_call_hyp(__kvm_adjust_pc, vcpu); 1087 1088 vcpu_put(vcpu); 1089 return ret; 1090 } 1091 1092 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level) 1093 { 1094 int bit_index; 1095 bool set; 1096 unsigned long *hcr; 1097 1098 if (number == KVM_ARM_IRQ_CPU_IRQ) 1099 bit_index = __ffs(HCR_VI); 1100 else /* KVM_ARM_IRQ_CPU_FIQ */ 1101 bit_index = __ffs(HCR_VF); 1102 1103 hcr = vcpu_hcr(vcpu); 1104 if (level) 1105 set = test_and_set_bit(bit_index, hcr); 1106 else 1107 set = test_and_clear_bit(bit_index, hcr); 1108 1109 /* 1110 * If we didn't change anything, no need to wake up or kick other CPUs 1111 */ 1112 if (set == level) 1113 return 0; 1114 1115 /* 1116 * The vcpu irq_lines field was updated, wake up sleeping VCPUs and 1117 * trigger a world-switch round on the running physical CPU to set the 1118 * virtual IRQ/FIQ fields in the HCR appropriately. 1119 */ 1120 kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); 1121 kvm_vcpu_kick(vcpu); 1122 1123 return 0; 1124 } 1125 1126 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level, 1127 bool line_status) 1128 { 1129 u32 irq = irq_level->irq; 1130 unsigned int irq_type, vcpu_idx, irq_num; 1131 int nrcpus = atomic_read(&kvm->online_vcpus); 1132 struct kvm_vcpu *vcpu = NULL; 1133 bool level = irq_level->level; 1134 1135 irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK; 1136 vcpu_idx = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK; 1137 vcpu_idx += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1); 1138 irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK; 1139 1140 trace_kvm_irq_line(irq_type, vcpu_idx, irq_num, irq_level->level); 1141 1142 switch (irq_type) { 1143 case KVM_ARM_IRQ_TYPE_CPU: 1144 if (irqchip_in_kernel(kvm)) 1145 return -ENXIO; 1146 1147 if (vcpu_idx >= nrcpus) 1148 return -EINVAL; 1149 1150 vcpu = kvm_get_vcpu(kvm, vcpu_idx); 1151 if (!vcpu) 1152 return -EINVAL; 1153 1154 if (irq_num > KVM_ARM_IRQ_CPU_FIQ) 1155 return -EINVAL; 1156 1157 return vcpu_interrupt_line(vcpu, irq_num, level); 1158 case KVM_ARM_IRQ_TYPE_PPI: 1159 if (!irqchip_in_kernel(kvm)) 1160 return -ENXIO; 1161 1162 if (vcpu_idx >= nrcpus) 1163 return -EINVAL; 1164 1165 vcpu = kvm_get_vcpu(kvm, vcpu_idx); 1166 if (!vcpu) 1167 return -EINVAL; 1168 1169 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS) 1170 return -EINVAL; 1171 1172 return kvm_vgic_inject_irq(kvm, vcpu->vcpu_id, irq_num, level, NULL); 1173 case KVM_ARM_IRQ_TYPE_SPI: 1174 if (!irqchip_in_kernel(kvm)) 1175 return -ENXIO; 1176 1177 if (irq_num < VGIC_NR_PRIVATE_IRQS) 1178 return -EINVAL; 1179 1180 return kvm_vgic_inject_irq(kvm, 0, irq_num, level, NULL); 1181 } 1182 1183 return -EINVAL; 1184 } 1185 1186 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu, 1187 const struct kvm_vcpu_init *init) 1188 { 1189 unsigned long features = init->features[0]; 1190 int i; 1191 1192 if (features & ~KVM_VCPU_VALID_FEATURES) 1193 return -ENOENT; 1194 1195 for (i = 1; i < ARRAY_SIZE(init->features); i++) { 1196 if (init->features[i]) 1197 return -ENOENT; 1198 } 1199 1200 if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features)) 1201 return 0; 1202 1203 if (!cpus_have_const_cap(ARM64_HAS_32BIT_EL1)) 1204 return -EINVAL; 1205 1206 /* MTE is incompatible with AArch32 */ 1207 if (kvm_has_mte(vcpu->kvm)) 1208 return -EINVAL; 1209 1210 /* NV is incompatible with AArch32 */ 1211 if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features)) 1212 return -EINVAL; 1213 1214 return 0; 1215 } 1216 1217 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu, 1218 const struct kvm_vcpu_init *init) 1219 { 1220 unsigned long features = init->features[0]; 1221 1222 return !bitmap_equal(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES) || 1223 vcpu->arch.target != init->target; 1224 } 1225 1226 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1227 const struct kvm_vcpu_init *init) 1228 { 1229 unsigned long features = init->features[0]; 1230 struct kvm *kvm = vcpu->kvm; 1231 int ret = -EINVAL; 1232 1233 mutex_lock(&kvm->arch.config_lock); 1234 1235 if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) && 1236 !bitmap_equal(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES)) 1237 goto out_unlock; 1238 1239 vcpu->arch.target = init->target; 1240 bitmap_copy(vcpu->arch.features, &features, KVM_VCPU_MAX_FEATURES); 1241 1242 /* Now we know what it is, we can reset it. */ 1243 ret = kvm_reset_vcpu(vcpu); 1244 if (ret) { 1245 vcpu->arch.target = -1; 1246 bitmap_zero(vcpu->arch.features, KVM_VCPU_MAX_FEATURES); 1247 goto out_unlock; 1248 } 1249 1250 bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES); 1251 set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags); 1252 1253 out_unlock: 1254 mutex_unlock(&kvm->arch.config_lock); 1255 return ret; 1256 } 1257 1258 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1259 const struct kvm_vcpu_init *init) 1260 { 1261 int ret; 1262 1263 if (init->target != kvm_target_cpu()) 1264 return -EINVAL; 1265 1266 ret = kvm_vcpu_init_check_features(vcpu, init); 1267 if (ret) 1268 return ret; 1269 1270 if (vcpu->arch.target == -1) 1271 return __kvm_vcpu_set_target(vcpu, init); 1272 1273 if (kvm_vcpu_init_changed(vcpu, init)) 1274 return -EINVAL; 1275 1276 return kvm_reset_vcpu(vcpu); 1277 } 1278 1279 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu, 1280 struct kvm_vcpu_init *init) 1281 { 1282 bool power_off = false; 1283 int ret; 1284 1285 /* 1286 * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid 1287 * reflecting it in the finalized feature set, thus limiting its scope 1288 * to a single KVM_ARM_VCPU_INIT call. 1289 */ 1290 if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) { 1291 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF); 1292 power_off = true; 1293 } 1294 1295 ret = kvm_vcpu_set_target(vcpu, init); 1296 if (ret) 1297 return ret; 1298 1299 /* 1300 * Ensure a rebooted VM will fault in RAM pages and detect if the 1301 * guest MMU is turned off and flush the caches as needed. 1302 * 1303 * S2FWB enforces all memory accesses to RAM being cacheable, 1304 * ensuring that the data side is always coherent. We still 1305 * need to invalidate the I-cache though, as FWB does *not* 1306 * imply CTR_EL0.DIC. 1307 */ 1308 if (vcpu_has_run_once(vcpu)) { 1309 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB)) 1310 stage2_unmap_vm(vcpu->kvm); 1311 else 1312 icache_inval_all_pou(); 1313 } 1314 1315 vcpu_reset_hcr(vcpu); 1316 vcpu->arch.cptr_el2 = kvm_get_reset_cptr_el2(vcpu); 1317 1318 /* 1319 * Handle the "start in power-off" case. 1320 */ 1321 spin_lock(&vcpu->arch.mp_state_lock); 1322 1323 if (power_off) 1324 __kvm_arm_vcpu_power_off(vcpu); 1325 else 1326 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE); 1327 1328 spin_unlock(&vcpu->arch.mp_state_lock); 1329 1330 return 0; 1331 } 1332 1333 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu, 1334 struct kvm_device_attr *attr) 1335 { 1336 int ret = -ENXIO; 1337 1338 switch (attr->group) { 1339 default: 1340 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr); 1341 break; 1342 } 1343 1344 return ret; 1345 } 1346 1347 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu, 1348 struct kvm_device_attr *attr) 1349 { 1350 int ret = -ENXIO; 1351 1352 switch (attr->group) { 1353 default: 1354 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr); 1355 break; 1356 } 1357 1358 return ret; 1359 } 1360 1361 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu, 1362 struct kvm_device_attr *attr) 1363 { 1364 int ret = -ENXIO; 1365 1366 switch (attr->group) { 1367 default: 1368 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr); 1369 break; 1370 } 1371 1372 return ret; 1373 } 1374 1375 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu, 1376 struct kvm_vcpu_events *events) 1377 { 1378 memset(events, 0, sizeof(*events)); 1379 1380 return __kvm_arm_vcpu_get_events(vcpu, events); 1381 } 1382 1383 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu, 1384 struct kvm_vcpu_events *events) 1385 { 1386 int i; 1387 1388 /* check whether the reserved field is zero */ 1389 for (i = 0; i < ARRAY_SIZE(events->reserved); i++) 1390 if (events->reserved[i]) 1391 return -EINVAL; 1392 1393 /* check whether the pad field is zero */ 1394 for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++) 1395 if (events->exception.pad[i]) 1396 return -EINVAL; 1397 1398 return __kvm_arm_vcpu_set_events(vcpu, events); 1399 } 1400 1401 long kvm_arch_vcpu_ioctl(struct file *filp, 1402 unsigned int ioctl, unsigned long arg) 1403 { 1404 struct kvm_vcpu *vcpu = filp->private_data; 1405 void __user *argp = (void __user *)arg; 1406 struct kvm_device_attr attr; 1407 long r; 1408 1409 switch (ioctl) { 1410 case KVM_ARM_VCPU_INIT: { 1411 struct kvm_vcpu_init init; 1412 1413 r = -EFAULT; 1414 if (copy_from_user(&init, argp, sizeof(init))) 1415 break; 1416 1417 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init); 1418 break; 1419 } 1420 case KVM_SET_ONE_REG: 1421 case KVM_GET_ONE_REG: { 1422 struct kvm_one_reg reg; 1423 1424 r = -ENOEXEC; 1425 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1426 break; 1427 1428 r = -EFAULT; 1429 if (copy_from_user(®, argp, sizeof(reg))) 1430 break; 1431 1432 /* 1433 * We could owe a reset due to PSCI. Handle the pending reset 1434 * here to ensure userspace register accesses are ordered after 1435 * the reset. 1436 */ 1437 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 1438 kvm_reset_vcpu(vcpu); 1439 1440 if (ioctl == KVM_SET_ONE_REG) 1441 r = kvm_arm_set_reg(vcpu, ®); 1442 else 1443 r = kvm_arm_get_reg(vcpu, ®); 1444 break; 1445 } 1446 case KVM_GET_REG_LIST: { 1447 struct kvm_reg_list __user *user_list = argp; 1448 struct kvm_reg_list reg_list; 1449 unsigned n; 1450 1451 r = -ENOEXEC; 1452 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1453 break; 1454 1455 r = -EPERM; 1456 if (!kvm_arm_vcpu_is_finalized(vcpu)) 1457 break; 1458 1459 r = -EFAULT; 1460 if (copy_from_user(®_list, user_list, sizeof(reg_list))) 1461 break; 1462 n = reg_list.n; 1463 reg_list.n = kvm_arm_num_regs(vcpu); 1464 if (copy_to_user(user_list, ®_list, sizeof(reg_list))) 1465 break; 1466 r = -E2BIG; 1467 if (n < reg_list.n) 1468 break; 1469 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg); 1470 break; 1471 } 1472 case KVM_SET_DEVICE_ATTR: { 1473 r = -EFAULT; 1474 if (copy_from_user(&attr, argp, sizeof(attr))) 1475 break; 1476 r = kvm_arm_vcpu_set_attr(vcpu, &attr); 1477 break; 1478 } 1479 case KVM_GET_DEVICE_ATTR: { 1480 r = -EFAULT; 1481 if (copy_from_user(&attr, argp, sizeof(attr))) 1482 break; 1483 r = kvm_arm_vcpu_get_attr(vcpu, &attr); 1484 break; 1485 } 1486 case KVM_HAS_DEVICE_ATTR: { 1487 r = -EFAULT; 1488 if (copy_from_user(&attr, argp, sizeof(attr))) 1489 break; 1490 r = kvm_arm_vcpu_has_attr(vcpu, &attr); 1491 break; 1492 } 1493 case KVM_GET_VCPU_EVENTS: { 1494 struct kvm_vcpu_events events; 1495 1496 if (kvm_arm_vcpu_get_events(vcpu, &events)) 1497 return -EINVAL; 1498 1499 if (copy_to_user(argp, &events, sizeof(events))) 1500 return -EFAULT; 1501 1502 return 0; 1503 } 1504 case KVM_SET_VCPU_EVENTS: { 1505 struct kvm_vcpu_events events; 1506 1507 if (copy_from_user(&events, argp, sizeof(events))) 1508 return -EFAULT; 1509 1510 return kvm_arm_vcpu_set_events(vcpu, &events); 1511 } 1512 case KVM_ARM_VCPU_FINALIZE: { 1513 int what; 1514 1515 if (!kvm_vcpu_initialized(vcpu)) 1516 return -ENOEXEC; 1517 1518 if (get_user(what, (const int __user *)argp)) 1519 return -EFAULT; 1520 1521 return kvm_arm_vcpu_finalize(vcpu, what); 1522 } 1523 default: 1524 r = -EINVAL; 1525 } 1526 1527 return r; 1528 } 1529 1530 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot) 1531 { 1532 1533 } 1534 1535 void kvm_arch_flush_remote_tlbs_memslot(struct kvm *kvm, 1536 const struct kvm_memory_slot *memslot) 1537 { 1538 kvm_flush_remote_tlbs(kvm); 1539 } 1540 1541 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm, 1542 struct kvm_arm_device_addr *dev_addr) 1543 { 1544 switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) { 1545 case KVM_ARM_DEVICE_VGIC_V2: 1546 if (!vgic_present) 1547 return -ENXIO; 1548 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr); 1549 default: 1550 return -ENODEV; 1551 } 1552 } 1553 1554 static int kvm_vm_has_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_has_attr(kvm, attr); 1559 default: 1560 return -ENXIO; 1561 } 1562 } 1563 1564 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1565 { 1566 switch (attr->group) { 1567 case KVM_ARM_VM_SMCCC_CTRL: 1568 return kvm_vm_smccc_set_attr(kvm, attr); 1569 default: 1570 return -ENXIO; 1571 } 1572 } 1573 1574 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) 1575 { 1576 struct kvm *kvm = filp->private_data; 1577 void __user *argp = (void __user *)arg; 1578 struct kvm_device_attr attr; 1579 1580 switch (ioctl) { 1581 case KVM_CREATE_IRQCHIP: { 1582 int ret; 1583 if (!vgic_present) 1584 return -ENXIO; 1585 mutex_lock(&kvm->lock); 1586 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2); 1587 mutex_unlock(&kvm->lock); 1588 return ret; 1589 } 1590 case KVM_ARM_SET_DEVICE_ADDR: { 1591 struct kvm_arm_device_addr dev_addr; 1592 1593 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr))) 1594 return -EFAULT; 1595 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr); 1596 } 1597 case KVM_ARM_PREFERRED_TARGET: { 1598 struct kvm_vcpu_init init; 1599 1600 kvm_vcpu_preferred_target(&init); 1601 1602 if (copy_to_user(argp, &init, sizeof(init))) 1603 return -EFAULT; 1604 1605 return 0; 1606 } 1607 case KVM_ARM_MTE_COPY_TAGS: { 1608 struct kvm_arm_copy_mte_tags copy_tags; 1609 1610 if (copy_from_user(©_tags, argp, sizeof(copy_tags))) 1611 return -EFAULT; 1612 return kvm_vm_ioctl_mte_copy_tags(kvm, ©_tags); 1613 } 1614 case KVM_ARM_SET_COUNTER_OFFSET: { 1615 struct kvm_arm_counter_offset offset; 1616 1617 if (copy_from_user(&offset, argp, sizeof(offset))) 1618 return -EFAULT; 1619 return kvm_vm_ioctl_set_counter_offset(kvm, &offset); 1620 } 1621 case KVM_HAS_DEVICE_ATTR: { 1622 if (copy_from_user(&attr, argp, sizeof(attr))) 1623 return -EFAULT; 1624 1625 return kvm_vm_has_attr(kvm, &attr); 1626 } 1627 case KVM_SET_DEVICE_ATTR: { 1628 if (copy_from_user(&attr, argp, sizeof(attr))) 1629 return -EFAULT; 1630 1631 return kvm_vm_set_attr(kvm, &attr); 1632 } 1633 default: 1634 return -EINVAL; 1635 } 1636 } 1637 1638 /* unlocks vcpus from @vcpu_lock_idx and smaller */ 1639 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx) 1640 { 1641 struct kvm_vcpu *tmp_vcpu; 1642 1643 for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) { 1644 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx); 1645 mutex_unlock(&tmp_vcpu->mutex); 1646 } 1647 } 1648 1649 void unlock_all_vcpus(struct kvm *kvm) 1650 { 1651 lockdep_assert_held(&kvm->lock); 1652 1653 unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1); 1654 } 1655 1656 /* Returns true if all vcpus were locked, false otherwise */ 1657 bool lock_all_vcpus(struct kvm *kvm) 1658 { 1659 struct kvm_vcpu *tmp_vcpu; 1660 unsigned long c; 1661 1662 lockdep_assert_held(&kvm->lock); 1663 1664 /* 1665 * Any time a vcpu is in an ioctl (including running), the 1666 * core KVM code tries to grab the vcpu->mutex. 1667 * 1668 * By grabbing the vcpu->mutex of all VCPUs we ensure that no 1669 * other VCPUs can fiddle with the state while we access it. 1670 */ 1671 kvm_for_each_vcpu(c, tmp_vcpu, kvm) { 1672 if (!mutex_trylock(&tmp_vcpu->mutex)) { 1673 unlock_vcpus(kvm, c - 1); 1674 return false; 1675 } 1676 } 1677 1678 return true; 1679 } 1680 1681 static unsigned long nvhe_percpu_size(void) 1682 { 1683 return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) - 1684 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start); 1685 } 1686 1687 static unsigned long nvhe_percpu_order(void) 1688 { 1689 unsigned long size = nvhe_percpu_size(); 1690 1691 return size ? get_order(size) : 0; 1692 } 1693 1694 /* A lookup table holding the hypervisor VA for each vector slot */ 1695 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS]; 1696 1697 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot) 1698 { 1699 hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot); 1700 } 1701 1702 static int kvm_init_vector_slots(void) 1703 { 1704 int err; 1705 void *base; 1706 1707 base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector)); 1708 kvm_init_vector_slot(base, HYP_VECTOR_DIRECT); 1709 1710 base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs)); 1711 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT); 1712 1713 if (kvm_system_needs_idmapped_vectors() && 1714 !is_protected_kvm_enabled()) { 1715 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs), 1716 __BP_HARDEN_HYP_VECS_SZ, &base); 1717 if (err) 1718 return err; 1719 } 1720 1721 kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT); 1722 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT); 1723 return 0; 1724 } 1725 1726 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits) 1727 { 1728 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 1729 unsigned long tcr; 1730 1731 /* 1732 * Calculate the raw per-cpu offset without a translation from the 1733 * kernel's mapping to the linear mapping, and store it in tpidr_el2 1734 * so that we can use adr_l to access per-cpu variables in EL2. 1735 * Also drop the KASAN tag which gets in the way... 1736 */ 1737 params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) - 1738 (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start)); 1739 1740 params->mair_el2 = read_sysreg(mair_el1); 1741 1742 tcr = read_sysreg(tcr_el1); 1743 if (cpus_have_final_cap(ARM64_KVM_HVHE)) { 1744 tcr |= TCR_EPD1_MASK; 1745 } else { 1746 tcr &= TCR_EL2_MASK; 1747 tcr |= TCR_EL2_RES1; 1748 } 1749 tcr &= ~TCR_T0SZ_MASK; 1750 tcr |= TCR_T0SZ(hyp_va_bits); 1751 params->tcr_el2 = tcr; 1752 1753 params->pgd_pa = kvm_mmu_get_httbr(); 1754 if (is_protected_kvm_enabled()) 1755 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS; 1756 else 1757 params->hcr_el2 = HCR_HOST_NVHE_FLAGS; 1758 if (cpus_have_final_cap(ARM64_KVM_HVHE)) 1759 params->hcr_el2 |= HCR_E2H; 1760 params->vttbr = params->vtcr = 0; 1761 1762 /* 1763 * Flush the init params from the data cache because the struct will 1764 * be read while the MMU is off. 1765 */ 1766 kvm_flush_dcache_to_poc(params, sizeof(*params)); 1767 } 1768 1769 static void hyp_install_host_vector(void) 1770 { 1771 struct kvm_nvhe_init_params *params; 1772 struct arm_smccc_res res; 1773 1774 /* Switch from the HYP stub to our own HYP init vector */ 1775 __hyp_set_vectors(kvm_get_idmap_vector()); 1776 1777 /* 1778 * Call initialization code, and switch to the full blown HYP code. 1779 * If the cpucaps haven't been finalized yet, something has gone very 1780 * wrong, and hyp will crash and burn when it uses any 1781 * cpus_have_const_cap() wrapper. 1782 */ 1783 BUG_ON(!system_capabilities_finalized()); 1784 params = this_cpu_ptr_nvhe_sym(kvm_init_params); 1785 arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res); 1786 WARN_ON(res.a0 != SMCCC_RET_SUCCESS); 1787 } 1788 1789 static void cpu_init_hyp_mode(void) 1790 { 1791 hyp_install_host_vector(); 1792 1793 /* 1794 * Disabling SSBD on a non-VHE system requires us to enable SSBS 1795 * at EL2. 1796 */ 1797 if (this_cpu_has_cap(ARM64_SSBS) && 1798 arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) { 1799 kvm_call_hyp_nvhe(__kvm_enable_ssbs); 1800 } 1801 } 1802 1803 static void cpu_hyp_reset(void) 1804 { 1805 if (!is_kernel_in_hyp_mode()) 1806 __hyp_reset_vectors(); 1807 } 1808 1809 /* 1810 * EL2 vectors can be mapped and rerouted in a number of ways, 1811 * depending on the kernel configuration and CPU present: 1812 * 1813 * - If the CPU is affected by Spectre-v2, the hardening sequence is 1814 * placed in one of the vector slots, which is executed before jumping 1815 * to the real vectors. 1816 * 1817 * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot 1818 * containing the hardening sequence is mapped next to the idmap page, 1819 * and executed before jumping to the real vectors. 1820 * 1821 * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an 1822 * empty slot is selected, mapped next to the idmap page, and 1823 * executed before jumping to the real vectors. 1824 * 1825 * Note that ARM64_SPECTRE_V3A is somewhat incompatible with 1826 * VHE, as we don't have hypervisor-specific mappings. If the system 1827 * is VHE and yet selects this capability, it will be ignored. 1828 */ 1829 static void cpu_set_hyp_vector(void) 1830 { 1831 struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data); 1832 void *vector = hyp_spectre_vector_selector[data->slot]; 1833 1834 if (!is_protected_kvm_enabled()) 1835 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector; 1836 else 1837 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot); 1838 } 1839 1840 static void cpu_hyp_init_context(void) 1841 { 1842 kvm_init_host_cpu_context(&this_cpu_ptr_hyp_sym(kvm_host_data)->host_ctxt); 1843 1844 if (!is_kernel_in_hyp_mode()) 1845 cpu_init_hyp_mode(); 1846 } 1847 1848 static void cpu_hyp_init_features(void) 1849 { 1850 cpu_set_hyp_vector(); 1851 kvm_arm_init_debug(); 1852 1853 if (is_kernel_in_hyp_mode()) 1854 kvm_timer_init_vhe(); 1855 1856 if (vgic_present) 1857 kvm_vgic_init_cpu_hardware(); 1858 } 1859 1860 static void cpu_hyp_reinit(void) 1861 { 1862 cpu_hyp_reset(); 1863 cpu_hyp_init_context(); 1864 cpu_hyp_init_features(); 1865 } 1866 1867 static void _kvm_arch_hardware_enable(void *discard) 1868 { 1869 if (!__this_cpu_read(kvm_arm_hardware_enabled)) { 1870 cpu_hyp_reinit(); 1871 __this_cpu_write(kvm_arm_hardware_enabled, 1); 1872 } 1873 } 1874 1875 int kvm_arch_hardware_enable(void) 1876 { 1877 int was_enabled; 1878 1879 /* 1880 * Most calls to this function are made with migration 1881 * disabled, but not with preemption disabled. The former is 1882 * enough to ensure correctness, but most of the helpers 1883 * expect the later and will throw a tantrum otherwise. 1884 */ 1885 preempt_disable(); 1886 1887 was_enabled = __this_cpu_read(kvm_arm_hardware_enabled); 1888 _kvm_arch_hardware_enable(NULL); 1889 1890 if (!was_enabled) { 1891 kvm_vgic_cpu_up(); 1892 kvm_timer_cpu_up(); 1893 } 1894 1895 preempt_enable(); 1896 1897 return 0; 1898 } 1899 1900 static void _kvm_arch_hardware_disable(void *discard) 1901 { 1902 if (__this_cpu_read(kvm_arm_hardware_enabled)) { 1903 cpu_hyp_reset(); 1904 __this_cpu_write(kvm_arm_hardware_enabled, 0); 1905 } 1906 } 1907 1908 void kvm_arch_hardware_disable(void) 1909 { 1910 if (__this_cpu_read(kvm_arm_hardware_enabled)) { 1911 kvm_timer_cpu_down(); 1912 kvm_vgic_cpu_down(); 1913 } 1914 1915 if (!is_protected_kvm_enabled()) 1916 _kvm_arch_hardware_disable(NULL); 1917 } 1918 1919 #ifdef CONFIG_CPU_PM 1920 static int hyp_init_cpu_pm_notifier(struct notifier_block *self, 1921 unsigned long cmd, 1922 void *v) 1923 { 1924 /* 1925 * kvm_arm_hardware_enabled is left with its old value over 1926 * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should 1927 * re-enable hyp. 1928 */ 1929 switch (cmd) { 1930 case CPU_PM_ENTER: 1931 if (__this_cpu_read(kvm_arm_hardware_enabled)) 1932 /* 1933 * don't update kvm_arm_hardware_enabled here 1934 * so that the hardware will be re-enabled 1935 * when we resume. See below. 1936 */ 1937 cpu_hyp_reset(); 1938 1939 return NOTIFY_OK; 1940 case CPU_PM_ENTER_FAILED: 1941 case CPU_PM_EXIT: 1942 if (__this_cpu_read(kvm_arm_hardware_enabled)) 1943 /* The hardware was enabled before suspend. */ 1944 cpu_hyp_reinit(); 1945 1946 return NOTIFY_OK; 1947 1948 default: 1949 return NOTIFY_DONE; 1950 } 1951 } 1952 1953 static struct notifier_block hyp_init_cpu_pm_nb = { 1954 .notifier_call = hyp_init_cpu_pm_notifier, 1955 }; 1956 1957 static void __init hyp_cpu_pm_init(void) 1958 { 1959 if (!is_protected_kvm_enabled()) 1960 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb); 1961 } 1962 static void __init hyp_cpu_pm_exit(void) 1963 { 1964 if (!is_protected_kvm_enabled()) 1965 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb); 1966 } 1967 #else 1968 static inline void __init hyp_cpu_pm_init(void) 1969 { 1970 } 1971 static inline void __init hyp_cpu_pm_exit(void) 1972 { 1973 } 1974 #endif 1975 1976 static void __init init_cpu_logical_map(void) 1977 { 1978 unsigned int cpu; 1979 1980 /* 1981 * Copy the MPIDR <-> logical CPU ID mapping to hyp. 1982 * Only copy the set of online CPUs whose features have been checked 1983 * against the finalized system capabilities. The hypervisor will not 1984 * allow any other CPUs from the `possible` set to boot. 1985 */ 1986 for_each_online_cpu(cpu) 1987 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu); 1988 } 1989 1990 #define init_psci_0_1_impl_state(config, what) \ 1991 config.psci_0_1_ ## what ## _implemented = psci_ops.what 1992 1993 static bool __init init_psci_relay(void) 1994 { 1995 /* 1996 * If PSCI has not been initialized, protected KVM cannot install 1997 * itself on newly booted CPUs. 1998 */ 1999 if (!psci_ops.get_version) { 2000 kvm_err("Cannot initialize protected mode without PSCI\n"); 2001 return false; 2002 } 2003 2004 kvm_host_psci_config.version = psci_ops.get_version(); 2005 kvm_host_psci_config.smccc_version = arm_smccc_get_version(); 2006 2007 if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) { 2008 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids(); 2009 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend); 2010 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on); 2011 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off); 2012 init_psci_0_1_impl_state(kvm_host_psci_config, migrate); 2013 } 2014 return true; 2015 } 2016 2017 static int __init init_subsystems(void) 2018 { 2019 int err = 0; 2020 2021 /* 2022 * Enable hardware so that subsystem initialisation can access EL2. 2023 */ 2024 on_each_cpu(_kvm_arch_hardware_enable, NULL, 1); 2025 2026 /* 2027 * Register CPU lower-power notifier 2028 */ 2029 hyp_cpu_pm_init(); 2030 2031 /* 2032 * Init HYP view of VGIC 2033 */ 2034 err = kvm_vgic_hyp_init(); 2035 switch (err) { 2036 case 0: 2037 vgic_present = true; 2038 break; 2039 case -ENODEV: 2040 case -ENXIO: 2041 vgic_present = false; 2042 err = 0; 2043 break; 2044 default: 2045 goto out; 2046 } 2047 2048 /* 2049 * Init HYP architected timer support 2050 */ 2051 err = kvm_timer_hyp_init(vgic_present); 2052 if (err) 2053 goto out; 2054 2055 kvm_register_perf_callbacks(NULL); 2056 2057 out: 2058 if (err) 2059 hyp_cpu_pm_exit(); 2060 2061 if (err || !is_protected_kvm_enabled()) 2062 on_each_cpu(_kvm_arch_hardware_disable, NULL, 1); 2063 2064 return err; 2065 } 2066 2067 static void __init teardown_subsystems(void) 2068 { 2069 kvm_unregister_perf_callbacks(); 2070 hyp_cpu_pm_exit(); 2071 } 2072 2073 static void __init teardown_hyp_mode(void) 2074 { 2075 int cpu; 2076 2077 free_hyp_pgds(); 2078 for_each_possible_cpu(cpu) { 2079 free_page(per_cpu(kvm_arm_hyp_stack_page, cpu)); 2080 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order()); 2081 } 2082 } 2083 2084 static int __init do_pkvm_init(u32 hyp_va_bits) 2085 { 2086 void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)); 2087 int ret; 2088 2089 preempt_disable(); 2090 cpu_hyp_init_context(); 2091 ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size, 2092 num_possible_cpus(), kern_hyp_va(per_cpu_base), 2093 hyp_va_bits); 2094 cpu_hyp_init_features(); 2095 2096 /* 2097 * The stub hypercalls are now disabled, so set our local flag to 2098 * prevent a later re-init attempt in kvm_arch_hardware_enable(). 2099 */ 2100 __this_cpu_write(kvm_arm_hardware_enabled, 1); 2101 preempt_enable(); 2102 2103 return ret; 2104 } 2105 2106 static u64 get_hyp_id_aa64pfr0_el1(void) 2107 { 2108 /* 2109 * Track whether the system isn't affected by spectre/meltdown in the 2110 * hypervisor's view of id_aa64pfr0_el1, used for protected VMs. 2111 * Although this is per-CPU, we make it global for simplicity, e.g., not 2112 * to have to worry about vcpu migration. 2113 * 2114 * Unlike for non-protected VMs, userspace cannot override this for 2115 * protected VMs. 2116 */ 2117 u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1); 2118 2119 val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) | 2120 ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3)); 2121 2122 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), 2123 arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED); 2124 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), 2125 arm64_get_meltdown_state() == SPECTRE_UNAFFECTED); 2126 2127 return val; 2128 } 2129 2130 static void kvm_hyp_init_symbols(void) 2131 { 2132 kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1(); 2133 kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1); 2134 kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1); 2135 kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1); 2136 kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1); 2137 kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1); 2138 kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1); 2139 kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1); 2140 kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1); 2141 kvm_nvhe_sym(__icache_flags) = __icache_flags; 2142 kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits; 2143 } 2144 2145 static int __init kvm_hyp_init_protection(u32 hyp_va_bits) 2146 { 2147 void *addr = phys_to_virt(hyp_mem_base); 2148 int ret; 2149 2150 ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP); 2151 if (ret) 2152 return ret; 2153 2154 ret = do_pkvm_init(hyp_va_bits); 2155 if (ret) 2156 return ret; 2157 2158 free_hyp_pgds(); 2159 2160 return 0; 2161 } 2162 2163 static void pkvm_hyp_init_ptrauth(void) 2164 { 2165 struct kvm_cpu_context *hyp_ctxt; 2166 int cpu; 2167 2168 for_each_possible_cpu(cpu) { 2169 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu); 2170 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long(); 2171 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long(); 2172 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long(); 2173 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long(); 2174 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long(); 2175 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long(); 2176 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long(); 2177 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long(); 2178 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long(); 2179 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long(); 2180 } 2181 } 2182 2183 /* Inits Hyp-mode on all online CPUs */ 2184 static int __init init_hyp_mode(void) 2185 { 2186 u32 hyp_va_bits; 2187 int cpu; 2188 int err = -ENOMEM; 2189 2190 /* 2191 * The protected Hyp-mode cannot be initialized if the memory pool 2192 * allocation has failed. 2193 */ 2194 if (is_protected_kvm_enabled() && !hyp_mem_base) 2195 goto out_err; 2196 2197 /* 2198 * Allocate Hyp PGD and setup Hyp identity mapping 2199 */ 2200 err = kvm_mmu_init(&hyp_va_bits); 2201 if (err) 2202 goto out_err; 2203 2204 /* 2205 * Allocate stack pages for Hypervisor-mode 2206 */ 2207 for_each_possible_cpu(cpu) { 2208 unsigned long stack_page; 2209 2210 stack_page = __get_free_page(GFP_KERNEL); 2211 if (!stack_page) { 2212 err = -ENOMEM; 2213 goto out_err; 2214 } 2215 2216 per_cpu(kvm_arm_hyp_stack_page, cpu) = stack_page; 2217 } 2218 2219 /* 2220 * Allocate and initialize pages for Hypervisor-mode percpu regions. 2221 */ 2222 for_each_possible_cpu(cpu) { 2223 struct page *page; 2224 void *page_addr; 2225 2226 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order()); 2227 if (!page) { 2228 err = -ENOMEM; 2229 goto out_err; 2230 } 2231 2232 page_addr = page_address(page); 2233 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size()); 2234 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr; 2235 } 2236 2237 /* 2238 * Map the Hyp-code called directly from the host 2239 */ 2240 err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start), 2241 kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC); 2242 if (err) { 2243 kvm_err("Cannot map world-switch code\n"); 2244 goto out_err; 2245 } 2246 2247 err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start), 2248 kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO); 2249 if (err) { 2250 kvm_err("Cannot map .hyp.rodata section\n"); 2251 goto out_err; 2252 } 2253 2254 err = create_hyp_mappings(kvm_ksym_ref(__start_rodata), 2255 kvm_ksym_ref(__end_rodata), PAGE_HYP_RO); 2256 if (err) { 2257 kvm_err("Cannot map rodata section\n"); 2258 goto out_err; 2259 } 2260 2261 /* 2262 * .hyp.bss is guaranteed to be placed at the beginning of the .bss 2263 * section thanks to an assertion in the linker script. Map it RW and 2264 * the rest of .bss RO. 2265 */ 2266 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start), 2267 kvm_ksym_ref(__hyp_bss_end), PAGE_HYP); 2268 if (err) { 2269 kvm_err("Cannot map hyp bss section: %d\n", err); 2270 goto out_err; 2271 } 2272 2273 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end), 2274 kvm_ksym_ref(__bss_stop), PAGE_HYP_RO); 2275 if (err) { 2276 kvm_err("Cannot map bss section\n"); 2277 goto out_err; 2278 } 2279 2280 /* 2281 * Map the Hyp stack pages 2282 */ 2283 for_each_possible_cpu(cpu) { 2284 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 2285 char *stack_page = (char *)per_cpu(kvm_arm_hyp_stack_page, cpu); 2286 unsigned long hyp_addr; 2287 2288 /* 2289 * Allocate a contiguous HYP private VA range for the stack 2290 * and guard page. The allocation is also aligned based on 2291 * the order of its size. 2292 */ 2293 err = hyp_alloc_private_va_range(PAGE_SIZE * 2, &hyp_addr); 2294 if (err) { 2295 kvm_err("Cannot allocate hyp stack guard page\n"); 2296 goto out_err; 2297 } 2298 2299 /* 2300 * Since the stack grows downwards, map the stack to the page 2301 * at the higher address and leave the lower guard page 2302 * unbacked. 2303 * 2304 * Any valid stack address now has the PAGE_SHIFT bit as 1 2305 * and addresses corresponding to the guard page have the 2306 * PAGE_SHIFT bit as 0 - this is used for overflow detection. 2307 */ 2308 err = __create_hyp_mappings(hyp_addr + PAGE_SIZE, PAGE_SIZE, 2309 __pa(stack_page), PAGE_HYP); 2310 if (err) { 2311 kvm_err("Cannot map hyp stack\n"); 2312 goto out_err; 2313 } 2314 2315 /* 2316 * Save the stack PA in nvhe_init_params. This will be needed 2317 * to recreate the stack mapping in protected nVHE mode. 2318 * __hyp_pa() won't do the right thing there, since the stack 2319 * has been mapped in the flexible private VA space. 2320 */ 2321 params->stack_pa = __pa(stack_page); 2322 2323 params->stack_hyp_va = hyp_addr + (2 * PAGE_SIZE); 2324 } 2325 2326 for_each_possible_cpu(cpu) { 2327 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu]; 2328 char *percpu_end = percpu_begin + nvhe_percpu_size(); 2329 2330 /* Map Hyp percpu pages */ 2331 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP); 2332 if (err) { 2333 kvm_err("Cannot map hyp percpu region\n"); 2334 goto out_err; 2335 } 2336 2337 /* Prepare the CPU initialization parameters */ 2338 cpu_prepare_hyp_mode(cpu, hyp_va_bits); 2339 } 2340 2341 kvm_hyp_init_symbols(); 2342 2343 if (is_protected_kvm_enabled()) { 2344 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) && 2345 cpus_have_const_cap(ARM64_HAS_ADDRESS_AUTH)) 2346 pkvm_hyp_init_ptrauth(); 2347 2348 init_cpu_logical_map(); 2349 2350 if (!init_psci_relay()) { 2351 err = -ENODEV; 2352 goto out_err; 2353 } 2354 2355 err = kvm_hyp_init_protection(hyp_va_bits); 2356 if (err) { 2357 kvm_err("Failed to init hyp memory protection\n"); 2358 goto out_err; 2359 } 2360 } 2361 2362 return 0; 2363 2364 out_err: 2365 teardown_hyp_mode(); 2366 kvm_err("error initializing Hyp mode: %d\n", err); 2367 return err; 2368 } 2369 2370 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr) 2371 { 2372 struct kvm_vcpu *vcpu; 2373 unsigned long i; 2374 2375 mpidr &= MPIDR_HWID_BITMASK; 2376 kvm_for_each_vcpu(i, vcpu, kvm) { 2377 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu)) 2378 return vcpu; 2379 } 2380 return NULL; 2381 } 2382 2383 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm) 2384 { 2385 return irqchip_in_kernel(kvm); 2386 } 2387 2388 bool kvm_arch_has_irq_bypass(void) 2389 { 2390 return true; 2391 } 2392 2393 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons, 2394 struct irq_bypass_producer *prod) 2395 { 2396 struct kvm_kernel_irqfd *irqfd = 2397 container_of(cons, struct kvm_kernel_irqfd, consumer); 2398 2399 return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq, 2400 &irqfd->irq_entry); 2401 } 2402 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons, 2403 struct irq_bypass_producer *prod) 2404 { 2405 struct kvm_kernel_irqfd *irqfd = 2406 container_of(cons, struct kvm_kernel_irqfd, consumer); 2407 2408 kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq, 2409 &irqfd->irq_entry); 2410 } 2411 2412 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons) 2413 { 2414 struct kvm_kernel_irqfd *irqfd = 2415 container_of(cons, struct kvm_kernel_irqfd, consumer); 2416 2417 kvm_arm_halt_guest(irqfd->kvm); 2418 } 2419 2420 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons) 2421 { 2422 struct kvm_kernel_irqfd *irqfd = 2423 container_of(cons, struct kvm_kernel_irqfd, consumer); 2424 2425 kvm_arm_resume_guest(irqfd->kvm); 2426 } 2427 2428 /* Initialize Hyp-mode and memory mappings on all CPUs */ 2429 static __init int kvm_arm_init(void) 2430 { 2431 int err; 2432 bool in_hyp_mode; 2433 2434 if (!is_hyp_mode_available()) { 2435 kvm_info("HYP mode not available\n"); 2436 return -ENODEV; 2437 } 2438 2439 if (kvm_get_mode() == KVM_MODE_NONE) { 2440 kvm_info("KVM disabled from command line\n"); 2441 return -ENODEV; 2442 } 2443 2444 err = kvm_sys_reg_table_init(); 2445 if (err) { 2446 kvm_info("Error initializing system register tables"); 2447 return err; 2448 } 2449 2450 in_hyp_mode = is_kernel_in_hyp_mode(); 2451 2452 if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) || 2453 cpus_have_final_cap(ARM64_WORKAROUND_1508412)) 2454 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \ 2455 "Only trusted guests should be used on this system.\n"); 2456 2457 err = kvm_set_ipa_limit(); 2458 if (err) 2459 return err; 2460 2461 err = kvm_arm_init_sve(); 2462 if (err) 2463 return err; 2464 2465 err = kvm_arm_vmid_alloc_init(); 2466 if (err) { 2467 kvm_err("Failed to initialize VMID allocator.\n"); 2468 return err; 2469 } 2470 2471 if (!in_hyp_mode) { 2472 err = init_hyp_mode(); 2473 if (err) 2474 goto out_err; 2475 } 2476 2477 err = kvm_init_vector_slots(); 2478 if (err) { 2479 kvm_err("Cannot initialise vector slots\n"); 2480 goto out_hyp; 2481 } 2482 2483 err = init_subsystems(); 2484 if (err) 2485 goto out_hyp; 2486 2487 if (is_protected_kvm_enabled()) { 2488 kvm_info("Protected nVHE mode initialized successfully\n"); 2489 } else if (in_hyp_mode) { 2490 kvm_info("VHE mode initialized successfully\n"); 2491 } else { 2492 kvm_info("Hyp mode initialized successfully\n"); 2493 } 2494 2495 /* 2496 * FIXME: Do something reasonable if kvm_init() fails after pKVM 2497 * hypervisor protection is finalized. 2498 */ 2499 err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE); 2500 if (err) 2501 goto out_subs; 2502 2503 kvm_arm_initialised = true; 2504 2505 return 0; 2506 2507 out_subs: 2508 teardown_subsystems(); 2509 out_hyp: 2510 if (!in_hyp_mode) 2511 teardown_hyp_mode(); 2512 out_err: 2513 kvm_arm_vmid_alloc_free(); 2514 return err; 2515 } 2516 2517 static int __init early_kvm_mode_cfg(char *arg) 2518 { 2519 if (!arg) 2520 return -EINVAL; 2521 2522 if (strcmp(arg, "none") == 0) { 2523 kvm_mode = KVM_MODE_NONE; 2524 return 0; 2525 } 2526 2527 if (!is_hyp_mode_available()) { 2528 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n"); 2529 return 0; 2530 } 2531 2532 if (strcmp(arg, "protected") == 0) { 2533 if (!is_kernel_in_hyp_mode()) 2534 kvm_mode = KVM_MODE_PROTECTED; 2535 else 2536 pr_warn_once("Protected KVM not available with VHE\n"); 2537 2538 return 0; 2539 } 2540 2541 if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) { 2542 kvm_mode = KVM_MODE_DEFAULT; 2543 return 0; 2544 } 2545 2546 if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) { 2547 kvm_mode = KVM_MODE_NV; 2548 return 0; 2549 } 2550 2551 return -EINVAL; 2552 } 2553 early_param("kvm-arm.mode", early_kvm_mode_cfg); 2554 2555 enum kvm_mode kvm_get_mode(void) 2556 { 2557 return kvm_mode; 2558 } 2559 2560 module_init(kvm_arm_init); 2561