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_emulate.h> 39 #include <asm/kvm_mmu.h> 40 #include <asm/kvm_nested.h> 41 #include <asm/kvm_pkvm.h> 42 #include <asm/kvm_ptrauth.h> 43 #include <asm/sections.h> 44 45 #include <kvm/arm_hypercalls.h> 46 #include <kvm/arm_pmu.h> 47 #include <kvm/arm_psci.h> 48 49 #include "sys_regs.h" 50 51 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT; 52 53 enum kvm_wfx_trap_policy { 54 KVM_WFX_NOTRAP_SINGLE_TASK, /* Default option */ 55 KVM_WFX_NOTRAP, 56 KVM_WFX_TRAP, 57 }; 58 59 static enum kvm_wfx_trap_policy kvm_wfi_trap_policy __read_mostly = KVM_WFX_NOTRAP_SINGLE_TASK; 60 static enum kvm_wfx_trap_policy kvm_wfe_trap_policy __read_mostly = KVM_WFX_NOTRAP_SINGLE_TASK; 61 62 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector); 63 64 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_base); 65 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params); 66 67 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt); 68 69 static bool vgic_present, kvm_arm_initialised; 70 71 static DEFINE_PER_CPU(unsigned char, kvm_hyp_initialized); 72 73 bool is_kvm_arm_initialised(void) 74 { 75 return kvm_arm_initialised; 76 } 77 78 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu) 79 { 80 return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE; 81 } 82 83 int kvm_vm_ioctl_enable_cap(struct kvm *kvm, 84 struct kvm_enable_cap *cap) 85 { 86 int r = -EINVAL; 87 88 if (cap->flags) 89 return -EINVAL; 90 91 if (kvm_vm_is_protected(kvm) && !kvm_pvm_ext_allowed(cap->cap)) 92 return -EINVAL; 93 94 switch (cap->cap) { 95 case KVM_CAP_ARM_NISV_TO_USER: 96 r = 0; 97 set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER, 98 &kvm->arch.flags); 99 break; 100 case KVM_CAP_ARM_MTE: 101 mutex_lock(&kvm->lock); 102 if (system_supports_mte() && !kvm->created_vcpus) { 103 r = 0; 104 set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags); 105 } 106 mutex_unlock(&kvm->lock); 107 break; 108 case KVM_CAP_ARM_SYSTEM_SUSPEND: 109 r = 0; 110 set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags); 111 break; 112 case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE: 113 mutex_lock(&kvm->slots_lock); 114 /* 115 * To keep things simple, allow changing the chunk 116 * size only when no memory slots have been created. 117 */ 118 if (kvm_are_all_memslots_empty(kvm)) { 119 u64 new_cap = cap->args[0]; 120 121 if (!new_cap || kvm_is_block_size_supported(new_cap)) { 122 r = 0; 123 kvm->arch.mmu.split_page_chunk_size = new_cap; 124 } 125 } 126 mutex_unlock(&kvm->slots_lock); 127 break; 128 default: 129 break; 130 } 131 132 return r; 133 } 134 135 static int kvm_arm_default_max_vcpus(void) 136 { 137 return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS; 138 } 139 140 /** 141 * kvm_arch_init_vm - initializes a VM data structure 142 * @kvm: pointer to the KVM struct 143 * @type: kvm device type 144 */ 145 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type) 146 { 147 int ret; 148 149 mutex_init(&kvm->arch.config_lock); 150 151 #ifdef CONFIG_LOCKDEP 152 /* Clue in lockdep that the config_lock must be taken inside kvm->lock */ 153 mutex_lock(&kvm->lock); 154 mutex_lock(&kvm->arch.config_lock); 155 mutex_unlock(&kvm->arch.config_lock); 156 mutex_unlock(&kvm->lock); 157 #endif 158 159 kvm_init_nested(kvm); 160 161 ret = kvm_share_hyp(kvm, kvm + 1); 162 if (ret) 163 return ret; 164 165 ret = pkvm_init_host_vm(kvm); 166 if (ret) 167 goto err_unshare_kvm; 168 169 if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) { 170 ret = -ENOMEM; 171 goto err_unshare_kvm; 172 } 173 cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask); 174 175 ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type); 176 if (ret) 177 goto err_free_cpumask; 178 179 kvm_vgic_early_init(kvm); 180 181 kvm_timer_init_vm(kvm); 182 183 /* The maximum number of VCPUs is limited by the host's GIC model */ 184 kvm->max_vcpus = kvm_arm_default_max_vcpus(); 185 186 kvm_arm_init_hypercalls(kvm); 187 188 bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES); 189 190 return 0; 191 192 err_free_cpumask: 193 free_cpumask_var(kvm->arch.supported_cpus); 194 err_unshare_kvm: 195 kvm_unshare_hyp(kvm, kvm + 1); 196 return ret; 197 } 198 199 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf) 200 { 201 return VM_FAULT_SIGBUS; 202 } 203 204 void kvm_arch_create_vm_debugfs(struct kvm *kvm) 205 { 206 kvm_sys_regs_create_debugfs(kvm); 207 kvm_s2_ptdump_create_debugfs(kvm); 208 } 209 210 static void kvm_destroy_mpidr_data(struct kvm *kvm) 211 { 212 struct kvm_mpidr_data *data; 213 214 mutex_lock(&kvm->arch.config_lock); 215 216 data = rcu_dereference_protected(kvm->arch.mpidr_data, 217 lockdep_is_held(&kvm->arch.config_lock)); 218 if (data) { 219 rcu_assign_pointer(kvm->arch.mpidr_data, NULL); 220 synchronize_rcu(); 221 kfree(data); 222 } 223 224 mutex_unlock(&kvm->arch.config_lock); 225 } 226 227 /** 228 * kvm_arch_destroy_vm - destroy the VM data structure 229 * @kvm: pointer to the KVM struct 230 */ 231 void kvm_arch_destroy_vm(struct kvm *kvm) 232 { 233 bitmap_free(kvm->arch.pmu_filter); 234 free_cpumask_var(kvm->arch.supported_cpus); 235 236 kvm_vgic_destroy(kvm); 237 238 if (is_protected_kvm_enabled()) 239 pkvm_destroy_hyp_vm(kvm); 240 241 kvm_destroy_mpidr_data(kvm); 242 243 kfree(kvm->arch.sysreg_masks); 244 kvm_destroy_vcpus(kvm); 245 246 kvm_unshare_hyp(kvm, kvm + 1); 247 248 kvm_arm_teardown_hypercalls(kvm); 249 } 250 251 static bool kvm_has_full_ptr_auth(void) 252 { 253 bool apa, gpa, api, gpi, apa3, gpa3; 254 u64 isar1, isar2, val; 255 256 /* 257 * Check that: 258 * 259 * - both Address and Generic auth are implemented for a given 260 * algorithm (Q5, IMPDEF or Q3) 261 * - only a single algorithm is implemented. 262 */ 263 if (!system_has_full_ptr_auth()) 264 return false; 265 266 isar1 = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1); 267 isar2 = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1); 268 269 apa = !!FIELD_GET(ID_AA64ISAR1_EL1_APA_MASK, isar1); 270 val = FIELD_GET(ID_AA64ISAR1_EL1_GPA_MASK, isar1); 271 gpa = (val == ID_AA64ISAR1_EL1_GPA_IMP); 272 273 api = !!FIELD_GET(ID_AA64ISAR1_EL1_API_MASK, isar1); 274 val = FIELD_GET(ID_AA64ISAR1_EL1_GPI_MASK, isar1); 275 gpi = (val == ID_AA64ISAR1_EL1_GPI_IMP); 276 277 apa3 = !!FIELD_GET(ID_AA64ISAR2_EL1_APA3_MASK, isar2); 278 val = FIELD_GET(ID_AA64ISAR2_EL1_GPA3_MASK, isar2); 279 gpa3 = (val == ID_AA64ISAR2_EL1_GPA3_IMP); 280 281 return (apa == gpa && api == gpi && apa3 == gpa3 && 282 (apa + api + apa3) == 1); 283 } 284 285 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) 286 { 287 int r; 288 289 if (kvm && kvm_vm_is_protected(kvm) && !kvm_pvm_ext_allowed(ext)) 290 return 0; 291 292 switch (ext) { 293 case KVM_CAP_IRQCHIP: 294 r = vgic_present; 295 break; 296 case KVM_CAP_IOEVENTFD: 297 case KVM_CAP_USER_MEMORY: 298 case KVM_CAP_SYNC_MMU: 299 case KVM_CAP_DESTROY_MEMORY_REGION_WORKS: 300 case KVM_CAP_ONE_REG: 301 case KVM_CAP_ARM_PSCI: 302 case KVM_CAP_ARM_PSCI_0_2: 303 case KVM_CAP_READONLY_MEM: 304 case KVM_CAP_MP_STATE: 305 case KVM_CAP_IMMEDIATE_EXIT: 306 case KVM_CAP_VCPU_EVENTS: 307 case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2: 308 case KVM_CAP_ARM_NISV_TO_USER: 309 case KVM_CAP_ARM_INJECT_EXT_DABT: 310 case KVM_CAP_SET_GUEST_DEBUG: 311 case KVM_CAP_VCPU_ATTRIBUTES: 312 case KVM_CAP_PTP_KVM: 313 case KVM_CAP_ARM_SYSTEM_SUSPEND: 314 case KVM_CAP_IRQFD_RESAMPLE: 315 case KVM_CAP_COUNTER_OFFSET: 316 r = 1; 317 break; 318 case KVM_CAP_SET_GUEST_DEBUG2: 319 return KVM_GUESTDBG_VALID_MASK; 320 case KVM_CAP_ARM_SET_DEVICE_ADDR: 321 r = 1; 322 break; 323 case KVM_CAP_NR_VCPUS: 324 /* 325 * ARM64 treats KVM_CAP_NR_CPUS differently from all other 326 * architectures, as it does not always bound it to 327 * KVM_CAP_MAX_VCPUS. It should not matter much because 328 * this is just an advisory value. 329 */ 330 r = min_t(unsigned int, num_online_cpus(), 331 kvm_arm_default_max_vcpus()); 332 break; 333 case KVM_CAP_MAX_VCPUS: 334 case KVM_CAP_MAX_VCPU_ID: 335 if (kvm) 336 r = kvm->max_vcpus; 337 else 338 r = kvm_arm_default_max_vcpus(); 339 break; 340 case KVM_CAP_MSI_DEVID: 341 if (!kvm) 342 r = -EINVAL; 343 else 344 r = kvm->arch.vgic.msis_require_devid; 345 break; 346 case KVM_CAP_ARM_USER_IRQ: 347 /* 348 * 1: EL1_VTIMER, EL1_PTIMER, and PMU. 349 * (bump this number if adding more devices) 350 */ 351 r = 1; 352 break; 353 case KVM_CAP_ARM_MTE: 354 r = system_supports_mte(); 355 break; 356 case KVM_CAP_STEAL_TIME: 357 r = kvm_arm_pvtime_supported(); 358 break; 359 case KVM_CAP_ARM_EL1_32BIT: 360 r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1); 361 break; 362 case KVM_CAP_GUEST_DEBUG_HW_BPS: 363 r = get_num_brps(); 364 break; 365 case KVM_CAP_GUEST_DEBUG_HW_WPS: 366 r = get_num_wrps(); 367 break; 368 case KVM_CAP_ARM_PMU_V3: 369 r = kvm_arm_support_pmu_v3(); 370 break; 371 case KVM_CAP_ARM_INJECT_SERROR_ESR: 372 r = cpus_have_final_cap(ARM64_HAS_RAS_EXTN); 373 break; 374 case KVM_CAP_ARM_VM_IPA_SIZE: 375 r = get_kvm_ipa_limit(); 376 break; 377 case KVM_CAP_ARM_SVE: 378 r = system_supports_sve(); 379 break; 380 case KVM_CAP_ARM_PTRAUTH_ADDRESS: 381 case KVM_CAP_ARM_PTRAUTH_GENERIC: 382 r = kvm_has_full_ptr_auth(); 383 break; 384 case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE: 385 if (kvm) 386 r = kvm->arch.mmu.split_page_chunk_size; 387 else 388 r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT; 389 break; 390 case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES: 391 r = kvm_supported_block_sizes(); 392 break; 393 case KVM_CAP_ARM_SUPPORTED_REG_MASK_RANGES: 394 r = BIT(0); 395 break; 396 default: 397 r = 0; 398 } 399 400 return r; 401 } 402 403 long kvm_arch_dev_ioctl(struct file *filp, 404 unsigned int ioctl, unsigned long arg) 405 { 406 return -EINVAL; 407 } 408 409 struct kvm *kvm_arch_alloc_vm(void) 410 { 411 size_t sz = sizeof(struct kvm); 412 413 if (!has_vhe()) 414 return kzalloc(sz, GFP_KERNEL_ACCOUNT); 415 416 return __vmalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_HIGHMEM | __GFP_ZERO); 417 } 418 419 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id) 420 { 421 if (irqchip_in_kernel(kvm) && vgic_initialized(kvm)) 422 return -EBUSY; 423 424 if (id >= kvm->max_vcpus) 425 return -EINVAL; 426 427 return 0; 428 } 429 430 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu) 431 { 432 int err; 433 434 spin_lock_init(&vcpu->arch.mp_state_lock); 435 436 #ifdef CONFIG_LOCKDEP 437 /* Inform lockdep that the config_lock is acquired after vcpu->mutex */ 438 mutex_lock(&vcpu->mutex); 439 mutex_lock(&vcpu->kvm->arch.config_lock); 440 mutex_unlock(&vcpu->kvm->arch.config_lock); 441 mutex_unlock(&vcpu->mutex); 442 #endif 443 444 /* Force users to call KVM_ARM_VCPU_INIT */ 445 vcpu_clear_flag(vcpu, VCPU_INITIALIZED); 446 447 vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO; 448 449 /* Set up the timer */ 450 kvm_timer_vcpu_init(vcpu); 451 452 kvm_pmu_vcpu_init(vcpu); 453 454 kvm_arm_pvtime_vcpu_init(&vcpu->arch); 455 456 vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu; 457 458 /* 459 * This vCPU may have been created after mpidr_data was initialized. 460 * Throw out the pre-computed mappings if that is the case which forces 461 * KVM to fall back to iteratively searching the vCPUs. 462 */ 463 kvm_destroy_mpidr_data(vcpu->kvm); 464 465 err = kvm_vgic_vcpu_init(vcpu); 466 if (err) 467 return err; 468 469 return kvm_share_hyp(vcpu, vcpu + 1); 470 } 471 472 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu) 473 { 474 } 475 476 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu) 477 { 478 if (!is_protected_kvm_enabled()) 479 kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache); 480 else 481 free_hyp_memcache(&vcpu->arch.pkvm_memcache); 482 kvm_timer_vcpu_terminate(vcpu); 483 kvm_pmu_vcpu_destroy(vcpu); 484 kvm_vgic_vcpu_destroy(vcpu); 485 kvm_arm_vcpu_destroy(vcpu); 486 } 487 488 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu) 489 { 490 491 } 492 493 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu) 494 { 495 496 } 497 498 static void vcpu_set_pauth_traps(struct kvm_vcpu *vcpu) 499 { 500 if (vcpu_has_ptrauth(vcpu) && !is_protected_kvm_enabled()) { 501 /* 502 * Either we're running an L2 guest, and the API/APK bits come 503 * from L1's HCR_EL2, or API/APK are both set. 504 */ 505 if (unlikely(vcpu_has_nv(vcpu) && !is_hyp_ctxt(vcpu))) { 506 u64 val; 507 508 val = __vcpu_sys_reg(vcpu, HCR_EL2); 509 val &= (HCR_API | HCR_APK); 510 vcpu->arch.hcr_el2 &= ~(HCR_API | HCR_APK); 511 vcpu->arch.hcr_el2 |= val; 512 } else { 513 vcpu->arch.hcr_el2 |= (HCR_API | HCR_APK); 514 } 515 516 /* 517 * Save the host keys if there is any chance for the guest 518 * to use pauth, as the entry code will reload the guest 519 * keys in that case. 520 */ 521 if (vcpu->arch.hcr_el2 & (HCR_API | HCR_APK)) { 522 struct kvm_cpu_context *ctxt; 523 524 ctxt = this_cpu_ptr_hyp_sym(kvm_hyp_ctxt); 525 ptrauth_save_keys(ctxt); 526 } 527 } 528 } 529 530 static bool kvm_vcpu_should_clear_twi(struct kvm_vcpu *vcpu) 531 { 532 if (unlikely(kvm_wfi_trap_policy != KVM_WFX_NOTRAP_SINGLE_TASK)) 533 return kvm_wfi_trap_policy == KVM_WFX_NOTRAP; 534 535 return single_task_running() && 536 (atomic_read(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count) || 537 vcpu->kvm->arch.vgic.nassgireq); 538 } 539 540 static bool kvm_vcpu_should_clear_twe(struct kvm_vcpu *vcpu) 541 { 542 if (unlikely(kvm_wfe_trap_policy != KVM_WFX_NOTRAP_SINGLE_TASK)) 543 return kvm_wfe_trap_policy == KVM_WFX_NOTRAP; 544 545 return single_task_running(); 546 } 547 548 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu) 549 { 550 struct kvm_s2_mmu *mmu; 551 int *last_ran; 552 553 if (is_protected_kvm_enabled()) 554 goto nommu; 555 556 if (vcpu_has_nv(vcpu)) 557 kvm_vcpu_load_hw_mmu(vcpu); 558 559 mmu = vcpu->arch.hw_mmu; 560 last_ran = this_cpu_ptr(mmu->last_vcpu_ran); 561 562 /* 563 * Ensure a VMID is allocated for the MMU before programming VTTBR_EL2, 564 * which happens eagerly in VHE. 565 * 566 * Also, the VMID allocator only preserves VMIDs that are active at the 567 * time of rollover, so KVM might need to grab a new VMID for the MMU if 568 * this is called from kvm_sched_in(). 569 */ 570 kvm_arm_vmid_update(&mmu->vmid); 571 572 /* 573 * We guarantee that both TLBs and I-cache are private to each 574 * vcpu. If detecting that a vcpu from the same VM has 575 * previously run on the same physical CPU, call into the 576 * hypervisor code to nuke the relevant contexts. 577 * 578 * We might get preempted before the vCPU actually runs, but 579 * over-invalidation doesn't affect correctness. 580 */ 581 if (*last_ran != vcpu->vcpu_idx) { 582 kvm_call_hyp(__kvm_flush_cpu_context, mmu); 583 *last_ran = vcpu->vcpu_idx; 584 } 585 586 nommu: 587 vcpu->cpu = cpu; 588 589 kvm_vgic_load(vcpu); 590 kvm_timer_vcpu_load(vcpu); 591 kvm_vcpu_load_debug(vcpu); 592 if (has_vhe()) 593 kvm_vcpu_load_vhe(vcpu); 594 kvm_arch_vcpu_load_fp(vcpu); 595 kvm_vcpu_pmu_restore_guest(vcpu); 596 if (kvm_arm_is_pvtime_enabled(&vcpu->arch)) 597 kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu); 598 599 if (kvm_vcpu_should_clear_twe(vcpu)) 600 vcpu->arch.hcr_el2 &= ~HCR_TWE; 601 else 602 vcpu->arch.hcr_el2 |= HCR_TWE; 603 604 if (kvm_vcpu_should_clear_twi(vcpu)) 605 vcpu->arch.hcr_el2 &= ~HCR_TWI; 606 else 607 vcpu->arch.hcr_el2 |= HCR_TWI; 608 609 vcpu_set_pauth_traps(vcpu); 610 611 if (is_protected_kvm_enabled()) { 612 kvm_call_hyp_nvhe(__pkvm_vcpu_load, 613 vcpu->kvm->arch.pkvm.handle, 614 vcpu->vcpu_idx, vcpu->arch.hcr_el2); 615 kvm_call_hyp(__vgic_v3_restore_vmcr_aprs, 616 &vcpu->arch.vgic_cpu.vgic_v3); 617 } 618 619 if (!cpumask_test_cpu(cpu, vcpu->kvm->arch.supported_cpus)) 620 vcpu_set_on_unsupported_cpu(vcpu); 621 } 622 623 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu) 624 { 625 if (is_protected_kvm_enabled()) { 626 kvm_call_hyp(__vgic_v3_save_vmcr_aprs, 627 &vcpu->arch.vgic_cpu.vgic_v3); 628 kvm_call_hyp_nvhe(__pkvm_vcpu_put); 629 } 630 631 kvm_vcpu_put_debug(vcpu); 632 kvm_arch_vcpu_put_fp(vcpu); 633 if (has_vhe()) 634 kvm_vcpu_put_vhe(vcpu); 635 kvm_timer_vcpu_put(vcpu); 636 kvm_vgic_put(vcpu); 637 kvm_vcpu_pmu_restore_host(vcpu); 638 if (vcpu_has_nv(vcpu)) 639 kvm_vcpu_put_hw_mmu(vcpu); 640 kvm_arm_vmid_clear_active(); 641 642 vcpu_clear_on_unsupported_cpu(vcpu); 643 vcpu->cpu = -1; 644 } 645 646 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 647 { 648 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED); 649 kvm_make_request(KVM_REQ_SLEEP, vcpu); 650 kvm_vcpu_kick(vcpu); 651 } 652 653 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu) 654 { 655 spin_lock(&vcpu->arch.mp_state_lock); 656 __kvm_arm_vcpu_power_off(vcpu); 657 spin_unlock(&vcpu->arch.mp_state_lock); 658 } 659 660 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu) 661 { 662 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED; 663 } 664 665 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu) 666 { 667 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED); 668 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 669 kvm_vcpu_kick(vcpu); 670 } 671 672 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu) 673 { 674 return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED; 675 } 676 677 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu, 678 struct kvm_mp_state *mp_state) 679 { 680 *mp_state = READ_ONCE(vcpu->arch.mp_state); 681 682 return 0; 683 } 684 685 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu, 686 struct kvm_mp_state *mp_state) 687 { 688 int ret = 0; 689 690 spin_lock(&vcpu->arch.mp_state_lock); 691 692 switch (mp_state->mp_state) { 693 case KVM_MP_STATE_RUNNABLE: 694 WRITE_ONCE(vcpu->arch.mp_state, *mp_state); 695 break; 696 case KVM_MP_STATE_STOPPED: 697 __kvm_arm_vcpu_power_off(vcpu); 698 break; 699 case KVM_MP_STATE_SUSPENDED: 700 kvm_arm_vcpu_suspend(vcpu); 701 break; 702 default: 703 ret = -EINVAL; 704 } 705 706 spin_unlock(&vcpu->arch.mp_state_lock); 707 708 return ret; 709 } 710 711 /** 712 * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled 713 * @v: The VCPU pointer 714 * 715 * If the guest CPU is not waiting for interrupts or an interrupt line is 716 * asserted, the CPU is by definition runnable. 717 */ 718 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v) 719 { 720 bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF); 721 return ((irq_lines || kvm_vgic_vcpu_pending_irq(v)) 722 && !kvm_arm_vcpu_stopped(v) && !v->arch.pause); 723 } 724 725 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu) 726 { 727 return vcpu_mode_priv(vcpu); 728 } 729 730 #ifdef CONFIG_GUEST_PERF_EVENTS 731 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu) 732 { 733 return *vcpu_pc(vcpu); 734 } 735 #endif 736 737 static void kvm_init_mpidr_data(struct kvm *kvm) 738 { 739 struct kvm_mpidr_data *data = NULL; 740 unsigned long c, mask, nr_entries; 741 u64 aff_set = 0, aff_clr = ~0UL; 742 struct kvm_vcpu *vcpu; 743 744 mutex_lock(&kvm->arch.config_lock); 745 746 if (rcu_access_pointer(kvm->arch.mpidr_data) || 747 atomic_read(&kvm->online_vcpus) == 1) 748 goto out; 749 750 kvm_for_each_vcpu(c, vcpu, kvm) { 751 u64 aff = kvm_vcpu_get_mpidr_aff(vcpu); 752 aff_set |= aff; 753 aff_clr &= aff; 754 } 755 756 /* 757 * A significant bit can be either 0 or 1, and will only appear in 758 * aff_set. Use aff_clr to weed out the useless stuff. 759 */ 760 mask = aff_set ^ aff_clr; 761 nr_entries = BIT_ULL(hweight_long(mask)); 762 763 /* 764 * Don't let userspace fool us. If we need more than a single page 765 * to describe the compressed MPIDR array, just fall back to the 766 * iterative method. Single vcpu VMs do not need this either. 767 */ 768 if (struct_size(data, cmpidr_to_idx, nr_entries) <= PAGE_SIZE) 769 data = kzalloc(struct_size(data, cmpidr_to_idx, nr_entries), 770 GFP_KERNEL_ACCOUNT); 771 772 if (!data) 773 goto out; 774 775 data->mpidr_mask = mask; 776 777 kvm_for_each_vcpu(c, vcpu, kvm) { 778 u64 aff = kvm_vcpu_get_mpidr_aff(vcpu); 779 u16 index = kvm_mpidr_index(data, aff); 780 781 data->cmpidr_to_idx[index] = c; 782 } 783 784 rcu_assign_pointer(kvm->arch.mpidr_data, data); 785 out: 786 mutex_unlock(&kvm->arch.config_lock); 787 } 788 789 /* 790 * Handle both the initialisation that is being done when the vcpu is 791 * run for the first time, as well as the updates that must be 792 * performed each time we get a new thread dealing with this vcpu. 793 */ 794 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu) 795 { 796 struct kvm *kvm = vcpu->kvm; 797 int ret; 798 799 if (!kvm_vcpu_initialized(vcpu)) 800 return -ENOEXEC; 801 802 if (!kvm_arm_vcpu_is_finalized(vcpu)) 803 return -EPERM; 804 805 ret = kvm_arch_vcpu_run_map_fp(vcpu); 806 if (ret) 807 return ret; 808 809 if (likely(vcpu_has_run_once(vcpu))) 810 return 0; 811 812 kvm_init_mpidr_data(kvm); 813 814 if (likely(irqchip_in_kernel(kvm))) { 815 /* 816 * Map the VGIC hardware resources before running a vcpu the 817 * first time on this VM. 818 */ 819 ret = kvm_vgic_map_resources(kvm); 820 if (ret) 821 return ret; 822 } 823 824 ret = kvm_finalize_sys_regs(vcpu); 825 if (ret) 826 return ret; 827 828 /* 829 * This needs to happen after any restriction has been applied 830 * to the feature set. 831 */ 832 kvm_calculate_traps(vcpu); 833 834 ret = kvm_timer_enable(vcpu); 835 if (ret) 836 return ret; 837 838 ret = kvm_arm_pmu_v3_enable(vcpu); 839 if (ret) 840 return ret; 841 842 if (is_protected_kvm_enabled()) { 843 ret = pkvm_create_hyp_vm(kvm); 844 if (ret) 845 return ret; 846 } 847 848 mutex_lock(&kvm->arch.config_lock); 849 set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags); 850 mutex_unlock(&kvm->arch.config_lock); 851 852 return ret; 853 } 854 855 bool kvm_arch_intc_initialized(struct kvm *kvm) 856 { 857 return vgic_initialized(kvm); 858 } 859 860 void kvm_arm_halt_guest(struct kvm *kvm) 861 { 862 unsigned long i; 863 struct kvm_vcpu *vcpu; 864 865 kvm_for_each_vcpu(i, vcpu, kvm) 866 vcpu->arch.pause = true; 867 kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP); 868 } 869 870 void kvm_arm_resume_guest(struct kvm *kvm) 871 { 872 unsigned long i; 873 struct kvm_vcpu *vcpu; 874 875 kvm_for_each_vcpu(i, vcpu, kvm) { 876 vcpu->arch.pause = false; 877 __kvm_vcpu_wake_up(vcpu); 878 } 879 } 880 881 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu) 882 { 883 struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu); 884 885 rcuwait_wait_event(wait, 886 (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause), 887 TASK_INTERRUPTIBLE); 888 889 if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) { 890 /* Awaken to handle a signal, request we sleep again later. */ 891 kvm_make_request(KVM_REQ_SLEEP, vcpu); 892 } 893 894 /* 895 * Make sure we will observe a potential reset request if we've 896 * observed a change to the power state. Pairs with the smp_wmb() in 897 * kvm_psci_vcpu_on(). 898 */ 899 smp_rmb(); 900 } 901 902 /** 903 * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior 904 * @vcpu: The VCPU pointer 905 * 906 * Suspend execution of a vCPU until a valid wake event is detected, i.e. until 907 * the vCPU is runnable. The vCPU may or may not be scheduled out, depending 908 * on when a wake event arrives, e.g. there may already be a pending wake event. 909 */ 910 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu) 911 { 912 /* 913 * Sync back the state of the GIC CPU interface so that we have 914 * the latest PMR and group enables. This ensures that 915 * kvm_arch_vcpu_runnable has up-to-date data to decide whether 916 * we have pending interrupts, e.g. when determining if the 917 * vCPU should block. 918 * 919 * For the same reason, we want to tell GICv4 that we need 920 * doorbells to be signalled, should an interrupt become pending. 921 */ 922 preempt_disable(); 923 vcpu_set_flag(vcpu, IN_WFI); 924 kvm_vgic_put(vcpu); 925 preempt_enable(); 926 927 kvm_vcpu_halt(vcpu); 928 vcpu_clear_flag(vcpu, IN_WFIT); 929 930 preempt_disable(); 931 vcpu_clear_flag(vcpu, IN_WFI); 932 kvm_vgic_load(vcpu); 933 preempt_enable(); 934 } 935 936 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu) 937 { 938 if (!kvm_arm_vcpu_suspended(vcpu)) 939 return 1; 940 941 kvm_vcpu_wfi(vcpu); 942 943 /* 944 * The suspend state is sticky; we do not leave it until userspace 945 * explicitly marks the vCPU as runnable. Request that we suspend again 946 * later. 947 */ 948 kvm_make_request(KVM_REQ_SUSPEND, vcpu); 949 950 /* 951 * Check to make sure the vCPU is actually runnable. If so, exit to 952 * userspace informing it of the wakeup condition. 953 */ 954 if (kvm_arch_vcpu_runnable(vcpu)) { 955 memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event)); 956 vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP; 957 vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT; 958 return 0; 959 } 960 961 /* 962 * Otherwise, we were unblocked to process a different event, such as a 963 * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to 964 * process the event. 965 */ 966 return 1; 967 } 968 969 /** 970 * check_vcpu_requests - check and handle pending vCPU requests 971 * @vcpu: the VCPU pointer 972 * 973 * Return: 1 if we should enter the guest 974 * 0 if we should exit to userspace 975 * < 0 if we should exit to userspace, where the return value indicates 976 * an error 977 */ 978 static int check_vcpu_requests(struct kvm_vcpu *vcpu) 979 { 980 if (kvm_request_pending(vcpu)) { 981 if (kvm_check_request(KVM_REQ_VM_DEAD, vcpu)) 982 return -EIO; 983 984 if (kvm_check_request(KVM_REQ_SLEEP, vcpu)) 985 kvm_vcpu_sleep(vcpu); 986 987 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 988 kvm_reset_vcpu(vcpu); 989 990 /* 991 * Clear IRQ_PENDING requests that were made to guarantee 992 * that a VCPU sees new virtual interrupts. 993 */ 994 kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu); 995 996 if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu)) 997 kvm_update_stolen_time(vcpu); 998 999 if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) { 1000 /* The distributor enable bits were changed */ 1001 preempt_disable(); 1002 vgic_v4_put(vcpu); 1003 vgic_v4_load(vcpu); 1004 preempt_enable(); 1005 } 1006 1007 if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu)) 1008 kvm_vcpu_reload_pmu(vcpu); 1009 1010 if (kvm_check_request(KVM_REQ_RESYNC_PMU_EL0, vcpu)) 1011 kvm_vcpu_pmu_restore_guest(vcpu); 1012 1013 if (kvm_check_request(KVM_REQ_SUSPEND, vcpu)) 1014 return kvm_vcpu_suspend(vcpu); 1015 1016 if (kvm_dirty_ring_check_request(vcpu)) 1017 return 0; 1018 1019 check_nested_vcpu_requests(vcpu); 1020 } 1021 1022 return 1; 1023 } 1024 1025 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu) 1026 { 1027 if (likely(!vcpu_mode_is_32bit(vcpu))) 1028 return false; 1029 1030 if (vcpu_has_nv(vcpu)) 1031 return true; 1032 1033 return !kvm_supports_32bit_el0(); 1034 } 1035 1036 /** 1037 * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest 1038 * @vcpu: The VCPU pointer 1039 * @ret: Pointer to write optional return code 1040 * 1041 * Returns: true if the VCPU needs to return to a preemptible + interruptible 1042 * and skip guest entry. 1043 * 1044 * This function disambiguates between two different types of exits: exits to a 1045 * preemptible + interruptible kernel context and exits to userspace. For an 1046 * exit to userspace, this function will write the return code to ret and return 1047 * true. For an exit to preemptible + interruptible kernel context (i.e. check 1048 * for pending work and re-enter), return true without writing to ret. 1049 */ 1050 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret) 1051 { 1052 struct kvm_run *run = vcpu->run; 1053 1054 /* 1055 * If we're using a userspace irqchip, then check if we need 1056 * to tell a userspace irqchip about timer or PMU level 1057 * changes and if so, exit to userspace (the actual level 1058 * state gets updated in kvm_timer_update_run and 1059 * kvm_pmu_update_run below). 1060 */ 1061 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) { 1062 if (kvm_timer_should_notify_user(vcpu) || 1063 kvm_pmu_should_notify_user(vcpu)) { 1064 *ret = -EINTR; 1065 run->exit_reason = KVM_EXIT_INTR; 1066 return true; 1067 } 1068 } 1069 1070 if (unlikely(vcpu_on_unsupported_cpu(vcpu))) { 1071 run->exit_reason = KVM_EXIT_FAIL_ENTRY; 1072 run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED; 1073 run->fail_entry.cpu = smp_processor_id(); 1074 *ret = 0; 1075 return true; 1076 } 1077 1078 return kvm_request_pending(vcpu) || 1079 xfer_to_guest_mode_work_pending(); 1080 } 1081 1082 /* 1083 * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while 1084 * the vCPU is running. 1085 * 1086 * This must be noinstr as instrumentation may make use of RCU, and this is not 1087 * safe during the EQS. 1088 */ 1089 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu) 1090 { 1091 int ret; 1092 1093 guest_state_enter_irqoff(); 1094 ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu); 1095 guest_state_exit_irqoff(); 1096 1097 return ret; 1098 } 1099 1100 /** 1101 * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code 1102 * @vcpu: The VCPU pointer 1103 * 1104 * This function is called through the VCPU_RUN ioctl called from user space. It 1105 * will execute VM code in a loop until the time slice for the process is used 1106 * or some emulation is needed from user space in which case the function will 1107 * return with return value 0 and with the kvm_run structure filled in with the 1108 * required data for the requested emulation. 1109 */ 1110 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) 1111 { 1112 struct kvm_run *run = vcpu->run; 1113 int ret; 1114 1115 if (run->exit_reason == KVM_EXIT_MMIO) { 1116 ret = kvm_handle_mmio_return(vcpu); 1117 if (ret <= 0) 1118 return ret; 1119 } 1120 1121 vcpu_load(vcpu); 1122 1123 if (!vcpu->wants_to_run) { 1124 ret = -EINTR; 1125 goto out; 1126 } 1127 1128 kvm_sigset_activate(vcpu); 1129 1130 ret = 1; 1131 run->exit_reason = KVM_EXIT_UNKNOWN; 1132 run->flags = 0; 1133 while (ret > 0) { 1134 /* 1135 * Check conditions before entering the guest 1136 */ 1137 ret = xfer_to_guest_mode_handle_work(vcpu); 1138 if (!ret) 1139 ret = 1; 1140 1141 if (ret > 0) 1142 ret = check_vcpu_requests(vcpu); 1143 1144 /* 1145 * Preparing the interrupts to be injected also 1146 * involves poking the GIC, which must be done in a 1147 * non-preemptible context. 1148 */ 1149 preempt_disable(); 1150 1151 kvm_pmu_flush_hwstate(vcpu); 1152 1153 local_irq_disable(); 1154 1155 kvm_vgic_flush_hwstate(vcpu); 1156 1157 kvm_pmu_update_vcpu_events(vcpu); 1158 1159 /* 1160 * Ensure we set mode to IN_GUEST_MODE after we disable 1161 * interrupts and before the final VCPU requests check. 1162 * See the comment in kvm_vcpu_exiting_guest_mode() and 1163 * Documentation/virt/kvm/vcpu-requests.rst 1164 */ 1165 smp_store_mb(vcpu->mode, IN_GUEST_MODE); 1166 1167 if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) { 1168 vcpu->mode = OUTSIDE_GUEST_MODE; 1169 isb(); /* Ensure work in x_flush_hwstate is committed */ 1170 kvm_pmu_sync_hwstate(vcpu); 1171 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) 1172 kvm_timer_sync_user(vcpu); 1173 kvm_vgic_sync_hwstate(vcpu); 1174 local_irq_enable(); 1175 preempt_enable(); 1176 continue; 1177 } 1178 1179 kvm_arch_vcpu_ctxflush_fp(vcpu); 1180 1181 /************************************************************** 1182 * Enter the guest 1183 */ 1184 trace_kvm_entry(*vcpu_pc(vcpu)); 1185 guest_timing_enter_irqoff(); 1186 1187 ret = kvm_arm_vcpu_enter_exit(vcpu); 1188 1189 vcpu->mode = OUTSIDE_GUEST_MODE; 1190 vcpu->stat.exits++; 1191 /* 1192 * Back from guest 1193 *************************************************************/ 1194 1195 /* 1196 * We must sync the PMU state before the vgic state so 1197 * that the vgic can properly sample the updated state of the 1198 * interrupt line. 1199 */ 1200 kvm_pmu_sync_hwstate(vcpu); 1201 1202 /* 1203 * Sync the vgic state before syncing the timer state because 1204 * the timer code needs to know if the virtual timer 1205 * interrupts are active. 1206 */ 1207 kvm_vgic_sync_hwstate(vcpu); 1208 1209 /* 1210 * Sync the timer hardware state before enabling interrupts as 1211 * we don't want vtimer interrupts to race with syncing the 1212 * timer virtual interrupt state. 1213 */ 1214 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) 1215 kvm_timer_sync_user(vcpu); 1216 1217 if (is_hyp_ctxt(vcpu)) 1218 kvm_timer_sync_nested(vcpu); 1219 1220 kvm_arch_vcpu_ctxsync_fp(vcpu); 1221 1222 /* 1223 * We must ensure that any pending interrupts are taken before 1224 * we exit guest timing so that timer ticks are accounted as 1225 * guest time. Transiently unmask interrupts so that any 1226 * pending interrupts are taken. 1227 * 1228 * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other 1229 * context synchronization event) is necessary to ensure that 1230 * pending interrupts are taken. 1231 */ 1232 if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) { 1233 local_irq_enable(); 1234 isb(); 1235 local_irq_disable(); 1236 } 1237 1238 guest_timing_exit_irqoff(); 1239 1240 local_irq_enable(); 1241 1242 trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu)); 1243 1244 /* Exit types that need handling before we can be preempted */ 1245 handle_exit_early(vcpu, ret); 1246 1247 preempt_enable(); 1248 1249 /* 1250 * The ARMv8 architecture doesn't give the hypervisor 1251 * a mechanism to prevent a guest from dropping to AArch32 EL0 1252 * if implemented by the CPU. If we spot the guest in such 1253 * state and that we decided it wasn't supposed to do so (like 1254 * with the asymmetric AArch32 case), return to userspace with 1255 * a fatal error. 1256 */ 1257 if (vcpu_mode_is_bad_32bit(vcpu)) { 1258 /* 1259 * As we have caught the guest red-handed, decide that 1260 * it isn't fit for purpose anymore by making the vcpu 1261 * invalid. The VMM can try and fix it by issuing a 1262 * KVM_ARM_VCPU_INIT if it really wants to. 1263 */ 1264 vcpu_clear_flag(vcpu, VCPU_INITIALIZED); 1265 ret = ARM_EXCEPTION_IL; 1266 } 1267 1268 ret = handle_exit(vcpu, ret); 1269 } 1270 1271 /* Tell userspace about in-kernel device output levels */ 1272 if (unlikely(!irqchip_in_kernel(vcpu->kvm))) { 1273 kvm_timer_update_run(vcpu); 1274 kvm_pmu_update_run(vcpu); 1275 } 1276 1277 kvm_sigset_deactivate(vcpu); 1278 1279 out: 1280 /* 1281 * In the unlikely event that we are returning to userspace 1282 * with pending exceptions or PC adjustment, commit these 1283 * adjustments in order to give userspace a consistent view of 1284 * the vcpu state. Note that this relies on __kvm_adjust_pc() 1285 * being preempt-safe on VHE. 1286 */ 1287 if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) || 1288 vcpu_get_flag(vcpu, INCREMENT_PC))) 1289 kvm_call_hyp(__kvm_adjust_pc, vcpu); 1290 1291 vcpu_put(vcpu); 1292 return ret; 1293 } 1294 1295 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level) 1296 { 1297 int bit_index; 1298 bool set; 1299 unsigned long *hcr; 1300 1301 if (number == KVM_ARM_IRQ_CPU_IRQ) 1302 bit_index = __ffs(HCR_VI); 1303 else /* KVM_ARM_IRQ_CPU_FIQ */ 1304 bit_index = __ffs(HCR_VF); 1305 1306 hcr = vcpu_hcr(vcpu); 1307 if (level) 1308 set = test_and_set_bit(bit_index, hcr); 1309 else 1310 set = test_and_clear_bit(bit_index, hcr); 1311 1312 /* 1313 * If we didn't change anything, no need to wake up or kick other CPUs 1314 */ 1315 if (set == level) 1316 return 0; 1317 1318 /* 1319 * The vcpu irq_lines field was updated, wake up sleeping VCPUs and 1320 * trigger a world-switch round on the running physical CPU to set the 1321 * virtual IRQ/FIQ fields in the HCR appropriately. 1322 */ 1323 kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu); 1324 kvm_vcpu_kick(vcpu); 1325 1326 return 0; 1327 } 1328 1329 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level, 1330 bool line_status) 1331 { 1332 u32 irq = irq_level->irq; 1333 unsigned int irq_type, vcpu_id, irq_num; 1334 struct kvm_vcpu *vcpu = NULL; 1335 bool level = irq_level->level; 1336 1337 irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK; 1338 vcpu_id = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK; 1339 vcpu_id += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1); 1340 irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK; 1341 1342 trace_kvm_irq_line(irq_type, vcpu_id, irq_num, irq_level->level); 1343 1344 switch (irq_type) { 1345 case KVM_ARM_IRQ_TYPE_CPU: 1346 if (irqchip_in_kernel(kvm)) 1347 return -ENXIO; 1348 1349 vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id); 1350 if (!vcpu) 1351 return -EINVAL; 1352 1353 if (irq_num > KVM_ARM_IRQ_CPU_FIQ) 1354 return -EINVAL; 1355 1356 return vcpu_interrupt_line(vcpu, irq_num, level); 1357 case KVM_ARM_IRQ_TYPE_PPI: 1358 if (!irqchip_in_kernel(kvm)) 1359 return -ENXIO; 1360 1361 vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id); 1362 if (!vcpu) 1363 return -EINVAL; 1364 1365 if (irq_num < VGIC_NR_SGIS || irq_num >= VGIC_NR_PRIVATE_IRQS) 1366 return -EINVAL; 1367 1368 return kvm_vgic_inject_irq(kvm, vcpu, irq_num, level, NULL); 1369 case KVM_ARM_IRQ_TYPE_SPI: 1370 if (!irqchip_in_kernel(kvm)) 1371 return -ENXIO; 1372 1373 if (irq_num < VGIC_NR_PRIVATE_IRQS) 1374 return -EINVAL; 1375 1376 return kvm_vgic_inject_irq(kvm, NULL, irq_num, level, NULL); 1377 } 1378 1379 return -EINVAL; 1380 } 1381 1382 static unsigned long system_supported_vcpu_features(void) 1383 { 1384 unsigned long features = KVM_VCPU_VALID_FEATURES; 1385 1386 if (!cpus_have_final_cap(ARM64_HAS_32BIT_EL1)) 1387 clear_bit(KVM_ARM_VCPU_EL1_32BIT, &features); 1388 1389 if (!kvm_arm_support_pmu_v3()) 1390 clear_bit(KVM_ARM_VCPU_PMU_V3, &features); 1391 1392 if (!system_supports_sve()) 1393 clear_bit(KVM_ARM_VCPU_SVE, &features); 1394 1395 if (!kvm_has_full_ptr_auth()) { 1396 clear_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features); 1397 clear_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features); 1398 } 1399 1400 if (!cpus_have_final_cap(ARM64_HAS_NESTED_VIRT)) 1401 clear_bit(KVM_ARM_VCPU_HAS_EL2, &features); 1402 1403 return features; 1404 } 1405 1406 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu, 1407 const struct kvm_vcpu_init *init) 1408 { 1409 unsigned long features = init->features[0]; 1410 int i; 1411 1412 if (features & ~KVM_VCPU_VALID_FEATURES) 1413 return -ENOENT; 1414 1415 for (i = 1; i < ARRAY_SIZE(init->features); i++) { 1416 if (init->features[i]) 1417 return -ENOENT; 1418 } 1419 1420 if (features & ~system_supported_vcpu_features()) 1421 return -EINVAL; 1422 1423 /* 1424 * For now make sure that both address/generic pointer authentication 1425 * features are requested by the userspace together. 1426 */ 1427 if (test_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features) != 1428 test_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features)) 1429 return -EINVAL; 1430 1431 if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features)) 1432 return 0; 1433 1434 /* MTE is incompatible with AArch32 */ 1435 if (kvm_has_mte(vcpu->kvm)) 1436 return -EINVAL; 1437 1438 /* NV is incompatible with AArch32 */ 1439 if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features)) 1440 return -EINVAL; 1441 1442 return 0; 1443 } 1444 1445 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu, 1446 const struct kvm_vcpu_init *init) 1447 { 1448 unsigned long features = init->features[0]; 1449 1450 return !bitmap_equal(vcpu->kvm->arch.vcpu_features, &features, 1451 KVM_VCPU_MAX_FEATURES); 1452 } 1453 1454 static int kvm_setup_vcpu(struct kvm_vcpu *vcpu) 1455 { 1456 struct kvm *kvm = vcpu->kvm; 1457 int ret = 0; 1458 1459 /* 1460 * When the vCPU has a PMU, but no PMU is set for the guest 1461 * yet, set the default one. 1462 */ 1463 if (kvm_vcpu_has_pmu(vcpu) && !kvm->arch.arm_pmu) 1464 ret = kvm_arm_set_default_pmu(kvm); 1465 1466 /* Prepare for nested if required */ 1467 if (!ret && vcpu_has_nv(vcpu)) 1468 ret = kvm_vcpu_init_nested(vcpu); 1469 1470 return ret; 1471 } 1472 1473 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1474 const struct kvm_vcpu_init *init) 1475 { 1476 unsigned long features = init->features[0]; 1477 struct kvm *kvm = vcpu->kvm; 1478 int ret = -EINVAL; 1479 1480 mutex_lock(&kvm->arch.config_lock); 1481 1482 if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) && 1483 kvm_vcpu_init_changed(vcpu, init)) 1484 goto out_unlock; 1485 1486 bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES); 1487 1488 ret = kvm_setup_vcpu(vcpu); 1489 if (ret) 1490 goto out_unlock; 1491 1492 /* Now we know what it is, we can reset it. */ 1493 kvm_reset_vcpu(vcpu); 1494 1495 set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags); 1496 vcpu_set_flag(vcpu, VCPU_INITIALIZED); 1497 ret = 0; 1498 out_unlock: 1499 mutex_unlock(&kvm->arch.config_lock); 1500 return ret; 1501 } 1502 1503 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu, 1504 const struct kvm_vcpu_init *init) 1505 { 1506 int ret; 1507 1508 if (init->target != KVM_ARM_TARGET_GENERIC_V8 && 1509 init->target != kvm_target_cpu()) 1510 return -EINVAL; 1511 1512 ret = kvm_vcpu_init_check_features(vcpu, init); 1513 if (ret) 1514 return ret; 1515 1516 if (!kvm_vcpu_initialized(vcpu)) 1517 return __kvm_vcpu_set_target(vcpu, init); 1518 1519 if (kvm_vcpu_init_changed(vcpu, init)) 1520 return -EINVAL; 1521 1522 kvm_reset_vcpu(vcpu); 1523 return 0; 1524 } 1525 1526 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu, 1527 struct kvm_vcpu_init *init) 1528 { 1529 bool power_off = false; 1530 int ret; 1531 1532 /* 1533 * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid 1534 * reflecting it in the finalized feature set, thus limiting its scope 1535 * to a single KVM_ARM_VCPU_INIT call. 1536 */ 1537 if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) { 1538 init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF); 1539 power_off = true; 1540 } 1541 1542 ret = kvm_vcpu_set_target(vcpu, init); 1543 if (ret) 1544 return ret; 1545 1546 /* 1547 * Ensure a rebooted VM will fault in RAM pages and detect if the 1548 * guest MMU is turned off and flush the caches as needed. 1549 * 1550 * S2FWB enforces all memory accesses to RAM being cacheable, 1551 * ensuring that the data side is always coherent. We still 1552 * need to invalidate the I-cache though, as FWB does *not* 1553 * imply CTR_EL0.DIC. 1554 */ 1555 if (vcpu_has_run_once(vcpu)) { 1556 if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB)) 1557 stage2_unmap_vm(vcpu->kvm); 1558 else 1559 icache_inval_all_pou(); 1560 } 1561 1562 vcpu_reset_hcr(vcpu); 1563 1564 /* 1565 * Handle the "start in power-off" case. 1566 */ 1567 spin_lock(&vcpu->arch.mp_state_lock); 1568 1569 if (power_off) 1570 __kvm_arm_vcpu_power_off(vcpu); 1571 else 1572 WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE); 1573 1574 spin_unlock(&vcpu->arch.mp_state_lock); 1575 1576 return 0; 1577 } 1578 1579 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu, 1580 struct kvm_device_attr *attr) 1581 { 1582 int ret = -ENXIO; 1583 1584 switch (attr->group) { 1585 default: 1586 ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr); 1587 break; 1588 } 1589 1590 return ret; 1591 } 1592 1593 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu, 1594 struct kvm_device_attr *attr) 1595 { 1596 int ret = -ENXIO; 1597 1598 switch (attr->group) { 1599 default: 1600 ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr); 1601 break; 1602 } 1603 1604 return ret; 1605 } 1606 1607 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu, 1608 struct kvm_device_attr *attr) 1609 { 1610 int ret = -ENXIO; 1611 1612 switch (attr->group) { 1613 default: 1614 ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr); 1615 break; 1616 } 1617 1618 return ret; 1619 } 1620 1621 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu, 1622 struct kvm_vcpu_events *events) 1623 { 1624 memset(events, 0, sizeof(*events)); 1625 1626 return __kvm_arm_vcpu_get_events(vcpu, events); 1627 } 1628 1629 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu, 1630 struct kvm_vcpu_events *events) 1631 { 1632 int i; 1633 1634 /* check whether the reserved field is zero */ 1635 for (i = 0; i < ARRAY_SIZE(events->reserved); i++) 1636 if (events->reserved[i]) 1637 return -EINVAL; 1638 1639 /* check whether the pad field is zero */ 1640 for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++) 1641 if (events->exception.pad[i]) 1642 return -EINVAL; 1643 1644 return __kvm_arm_vcpu_set_events(vcpu, events); 1645 } 1646 1647 long kvm_arch_vcpu_ioctl(struct file *filp, 1648 unsigned int ioctl, unsigned long arg) 1649 { 1650 struct kvm_vcpu *vcpu = filp->private_data; 1651 void __user *argp = (void __user *)arg; 1652 struct kvm_device_attr attr; 1653 long r; 1654 1655 switch (ioctl) { 1656 case KVM_ARM_VCPU_INIT: { 1657 struct kvm_vcpu_init init; 1658 1659 r = -EFAULT; 1660 if (copy_from_user(&init, argp, sizeof(init))) 1661 break; 1662 1663 r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init); 1664 break; 1665 } 1666 case KVM_SET_ONE_REG: 1667 case KVM_GET_ONE_REG: { 1668 struct kvm_one_reg reg; 1669 1670 r = -ENOEXEC; 1671 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1672 break; 1673 1674 r = -EFAULT; 1675 if (copy_from_user(®, argp, sizeof(reg))) 1676 break; 1677 1678 /* 1679 * We could owe a reset due to PSCI. Handle the pending reset 1680 * here to ensure userspace register accesses are ordered after 1681 * the reset. 1682 */ 1683 if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu)) 1684 kvm_reset_vcpu(vcpu); 1685 1686 if (ioctl == KVM_SET_ONE_REG) 1687 r = kvm_arm_set_reg(vcpu, ®); 1688 else 1689 r = kvm_arm_get_reg(vcpu, ®); 1690 break; 1691 } 1692 case KVM_GET_REG_LIST: { 1693 struct kvm_reg_list __user *user_list = argp; 1694 struct kvm_reg_list reg_list; 1695 unsigned n; 1696 1697 r = -ENOEXEC; 1698 if (unlikely(!kvm_vcpu_initialized(vcpu))) 1699 break; 1700 1701 r = -EPERM; 1702 if (!kvm_arm_vcpu_is_finalized(vcpu)) 1703 break; 1704 1705 r = -EFAULT; 1706 if (copy_from_user(®_list, user_list, sizeof(reg_list))) 1707 break; 1708 n = reg_list.n; 1709 reg_list.n = kvm_arm_num_regs(vcpu); 1710 if (copy_to_user(user_list, ®_list, sizeof(reg_list))) 1711 break; 1712 r = -E2BIG; 1713 if (n < reg_list.n) 1714 break; 1715 r = kvm_arm_copy_reg_indices(vcpu, user_list->reg); 1716 break; 1717 } 1718 case KVM_SET_DEVICE_ATTR: { 1719 r = -EFAULT; 1720 if (copy_from_user(&attr, argp, sizeof(attr))) 1721 break; 1722 r = kvm_arm_vcpu_set_attr(vcpu, &attr); 1723 break; 1724 } 1725 case KVM_GET_DEVICE_ATTR: { 1726 r = -EFAULT; 1727 if (copy_from_user(&attr, argp, sizeof(attr))) 1728 break; 1729 r = kvm_arm_vcpu_get_attr(vcpu, &attr); 1730 break; 1731 } 1732 case KVM_HAS_DEVICE_ATTR: { 1733 r = -EFAULT; 1734 if (copy_from_user(&attr, argp, sizeof(attr))) 1735 break; 1736 r = kvm_arm_vcpu_has_attr(vcpu, &attr); 1737 break; 1738 } 1739 case KVM_GET_VCPU_EVENTS: { 1740 struct kvm_vcpu_events events; 1741 1742 if (kvm_arm_vcpu_get_events(vcpu, &events)) 1743 return -EINVAL; 1744 1745 if (copy_to_user(argp, &events, sizeof(events))) 1746 return -EFAULT; 1747 1748 return 0; 1749 } 1750 case KVM_SET_VCPU_EVENTS: { 1751 struct kvm_vcpu_events events; 1752 1753 if (copy_from_user(&events, argp, sizeof(events))) 1754 return -EFAULT; 1755 1756 return kvm_arm_vcpu_set_events(vcpu, &events); 1757 } 1758 case KVM_ARM_VCPU_FINALIZE: { 1759 int what; 1760 1761 if (!kvm_vcpu_initialized(vcpu)) 1762 return -ENOEXEC; 1763 1764 if (get_user(what, (const int __user *)argp)) 1765 return -EFAULT; 1766 1767 return kvm_arm_vcpu_finalize(vcpu, what); 1768 } 1769 default: 1770 r = -EINVAL; 1771 } 1772 1773 return r; 1774 } 1775 1776 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot) 1777 { 1778 1779 } 1780 1781 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm, 1782 struct kvm_arm_device_addr *dev_addr) 1783 { 1784 switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) { 1785 case KVM_ARM_DEVICE_VGIC_V2: 1786 if (!vgic_present) 1787 return -ENXIO; 1788 return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr); 1789 default: 1790 return -ENODEV; 1791 } 1792 } 1793 1794 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1795 { 1796 switch (attr->group) { 1797 case KVM_ARM_VM_SMCCC_CTRL: 1798 return kvm_vm_smccc_has_attr(kvm, attr); 1799 default: 1800 return -ENXIO; 1801 } 1802 } 1803 1804 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr) 1805 { 1806 switch (attr->group) { 1807 case KVM_ARM_VM_SMCCC_CTRL: 1808 return kvm_vm_smccc_set_attr(kvm, attr); 1809 default: 1810 return -ENXIO; 1811 } 1812 } 1813 1814 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg) 1815 { 1816 struct kvm *kvm = filp->private_data; 1817 void __user *argp = (void __user *)arg; 1818 struct kvm_device_attr attr; 1819 1820 switch (ioctl) { 1821 case KVM_CREATE_IRQCHIP: { 1822 int ret; 1823 if (!vgic_present) 1824 return -ENXIO; 1825 mutex_lock(&kvm->lock); 1826 ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2); 1827 mutex_unlock(&kvm->lock); 1828 return ret; 1829 } 1830 case KVM_ARM_SET_DEVICE_ADDR: { 1831 struct kvm_arm_device_addr dev_addr; 1832 1833 if (copy_from_user(&dev_addr, argp, sizeof(dev_addr))) 1834 return -EFAULT; 1835 return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr); 1836 } 1837 case KVM_ARM_PREFERRED_TARGET: { 1838 struct kvm_vcpu_init init = { 1839 .target = KVM_ARM_TARGET_GENERIC_V8, 1840 }; 1841 1842 if (copy_to_user(argp, &init, sizeof(init))) 1843 return -EFAULT; 1844 1845 return 0; 1846 } 1847 case KVM_ARM_MTE_COPY_TAGS: { 1848 struct kvm_arm_copy_mte_tags copy_tags; 1849 1850 if (copy_from_user(©_tags, argp, sizeof(copy_tags))) 1851 return -EFAULT; 1852 return kvm_vm_ioctl_mte_copy_tags(kvm, ©_tags); 1853 } 1854 case KVM_ARM_SET_COUNTER_OFFSET: { 1855 struct kvm_arm_counter_offset offset; 1856 1857 if (copy_from_user(&offset, argp, sizeof(offset))) 1858 return -EFAULT; 1859 return kvm_vm_ioctl_set_counter_offset(kvm, &offset); 1860 } 1861 case KVM_HAS_DEVICE_ATTR: { 1862 if (copy_from_user(&attr, argp, sizeof(attr))) 1863 return -EFAULT; 1864 1865 return kvm_vm_has_attr(kvm, &attr); 1866 } 1867 case KVM_SET_DEVICE_ATTR: { 1868 if (copy_from_user(&attr, argp, sizeof(attr))) 1869 return -EFAULT; 1870 1871 return kvm_vm_set_attr(kvm, &attr); 1872 } 1873 case KVM_ARM_GET_REG_WRITABLE_MASKS: { 1874 struct reg_mask_range range; 1875 1876 if (copy_from_user(&range, argp, sizeof(range))) 1877 return -EFAULT; 1878 return kvm_vm_ioctl_get_reg_writable_masks(kvm, &range); 1879 } 1880 default: 1881 return -EINVAL; 1882 } 1883 } 1884 1885 /* unlocks vcpus from @vcpu_lock_idx and smaller */ 1886 static void unlock_vcpus(struct kvm *kvm, int vcpu_lock_idx) 1887 { 1888 struct kvm_vcpu *tmp_vcpu; 1889 1890 for (; vcpu_lock_idx >= 0; vcpu_lock_idx--) { 1891 tmp_vcpu = kvm_get_vcpu(kvm, vcpu_lock_idx); 1892 mutex_unlock(&tmp_vcpu->mutex); 1893 } 1894 } 1895 1896 void unlock_all_vcpus(struct kvm *kvm) 1897 { 1898 lockdep_assert_held(&kvm->lock); 1899 1900 unlock_vcpus(kvm, atomic_read(&kvm->online_vcpus) - 1); 1901 } 1902 1903 /* Returns true if all vcpus were locked, false otherwise */ 1904 bool lock_all_vcpus(struct kvm *kvm) 1905 { 1906 struct kvm_vcpu *tmp_vcpu; 1907 unsigned long c; 1908 1909 lockdep_assert_held(&kvm->lock); 1910 1911 /* 1912 * Any time a vcpu is in an ioctl (including running), the 1913 * core KVM code tries to grab the vcpu->mutex. 1914 * 1915 * By grabbing the vcpu->mutex of all VCPUs we ensure that no 1916 * other VCPUs can fiddle with the state while we access it. 1917 */ 1918 kvm_for_each_vcpu(c, tmp_vcpu, kvm) { 1919 if (!mutex_trylock(&tmp_vcpu->mutex)) { 1920 unlock_vcpus(kvm, c - 1); 1921 return false; 1922 } 1923 } 1924 1925 return true; 1926 } 1927 1928 static unsigned long nvhe_percpu_size(void) 1929 { 1930 return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) - 1931 (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start); 1932 } 1933 1934 static unsigned long nvhe_percpu_order(void) 1935 { 1936 unsigned long size = nvhe_percpu_size(); 1937 1938 return size ? get_order(size) : 0; 1939 } 1940 1941 static size_t pkvm_host_sve_state_order(void) 1942 { 1943 return get_order(pkvm_host_sve_state_size()); 1944 } 1945 1946 /* A lookup table holding the hypervisor VA for each vector slot */ 1947 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS]; 1948 1949 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot) 1950 { 1951 hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot); 1952 } 1953 1954 static int kvm_init_vector_slots(void) 1955 { 1956 int err; 1957 void *base; 1958 1959 base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector)); 1960 kvm_init_vector_slot(base, HYP_VECTOR_DIRECT); 1961 1962 base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs)); 1963 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT); 1964 1965 if (kvm_system_needs_idmapped_vectors() && 1966 !is_protected_kvm_enabled()) { 1967 err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs), 1968 __BP_HARDEN_HYP_VECS_SZ, &base); 1969 if (err) 1970 return err; 1971 } 1972 1973 kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT); 1974 kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT); 1975 return 0; 1976 } 1977 1978 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits) 1979 { 1980 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 1981 unsigned long tcr; 1982 1983 /* 1984 * Calculate the raw per-cpu offset without a translation from the 1985 * kernel's mapping to the linear mapping, and store it in tpidr_el2 1986 * so that we can use adr_l to access per-cpu variables in EL2. 1987 * Also drop the KASAN tag which gets in the way... 1988 */ 1989 params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) - 1990 (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start)); 1991 1992 params->mair_el2 = read_sysreg(mair_el1); 1993 1994 tcr = read_sysreg(tcr_el1); 1995 if (cpus_have_final_cap(ARM64_KVM_HVHE)) { 1996 tcr &= ~(TCR_HD | TCR_HA | TCR_A1 | TCR_T0SZ_MASK); 1997 tcr |= TCR_EPD1_MASK; 1998 } else { 1999 unsigned long ips = FIELD_GET(TCR_IPS_MASK, tcr); 2000 2001 tcr &= TCR_EL2_MASK; 2002 tcr |= TCR_EL2_RES1 | FIELD_PREP(TCR_EL2_PS_MASK, ips); 2003 if (lpa2_is_enabled()) 2004 tcr |= TCR_EL2_DS; 2005 } 2006 tcr |= TCR_T0SZ(hyp_va_bits); 2007 params->tcr_el2 = tcr; 2008 2009 params->pgd_pa = kvm_mmu_get_httbr(); 2010 if (is_protected_kvm_enabled()) 2011 params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS; 2012 else 2013 params->hcr_el2 = HCR_HOST_NVHE_FLAGS; 2014 if (cpus_have_final_cap(ARM64_KVM_HVHE)) 2015 params->hcr_el2 |= HCR_E2H; 2016 params->vttbr = params->vtcr = 0; 2017 2018 /* 2019 * Flush the init params from the data cache because the struct will 2020 * be read while the MMU is off. 2021 */ 2022 kvm_flush_dcache_to_poc(params, sizeof(*params)); 2023 } 2024 2025 static void hyp_install_host_vector(void) 2026 { 2027 struct kvm_nvhe_init_params *params; 2028 struct arm_smccc_res res; 2029 2030 /* Switch from the HYP stub to our own HYP init vector */ 2031 __hyp_set_vectors(kvm_get_idmap_vector()); 2032 2033 /* 2034 * Call initialization code, and switch to the full blown HYP code. 2035 * If the cpucaps haven't been finalized yet, something has gone very 2036 * wrong, and hyp will crash and burn when it uses any 2037 * cpus_have_*_cap() wrapper. 2038 */ 2039 BUG_ON(!system_capabilities_finalized()); 2040 params = this_cpu_ptr_nvhe_sym(kvm_init_params); 2041 arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res); 2042 WARN_ON(res.a0 != SMCCC_RET_SUCCESS); 2043 } 2044 2045 static void cpu_init_hyp_mode(void) 2046 { 2047 hyp_install_host_vector(); 2048 2049 /* 2050 * Disabling SSBD on a non-VHE system requires us to enable SSBS 2051 * at EL2. 2052 */ 2053 if (this_cpu_has_cap(ARM64_SSBS) && 2054 arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) { 2055 kvm_call_hyp_nvhe(__kvm_enable_ssbs); 2056 } 2057 } 2058 2059 static void cpu_hyp_reset(void) 2060 { 2061 if (!is_kernel_in_hyp_mode()) 2062 __hyp_reset_vectors(); 2063 } 2064 2065 /* 2066 * EL2 vectors can be mapped and rerouted in a number of ways, 2067 * depending on the kernel configuration and CPU present: 2068 * 2069 * - If the CPU is affected by Spectre-v2, the hardening sequence is 2070 * placed in one of the vector slots, which is executed before jumping 2071 * to the real vectors. 2072 * 2073 * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot 2074 * containing the hardening sequence is mapped next to the idmap page, 2075 * and executed before jumping to the real vectors. 2076 * 2077 * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an 2078 * empty slot is selected, mapped next to the idmap page, and 2079 * executed before jumping to the real vectors. 2080 * 2081 * Note that ARM64_SPECTRE_V3A is somewhat incompatible with 2082 * VHE, as we don't have hypervisor-specific mappings. If the system 2083 * is VHE and yet selects this capability, it will be ignored. 2084 */ 2085 static void cpu_set_hyp_vector(void) 2086 { 2087 struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data); 2088 void *vector = hyp_spectre_vector_selector[data->slot]; 2089 2090 if (!is_protected_kvm_enabled()) 2091 *this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector; 2092 else 2093 kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot); 2094 } 2095 2096 static void cpu_hyp_init_context(void) 2097 { 2098 kvm_init_host_cpu_context(host_data_ptr(host_ctxt)); 2099 kvm_init_host_debug_data(); 2100 2101 if (!is_kernel_in_hyp_mode()) 2102 cpu_init_hyp_mode(); 2103 } 2104 2105 static void cpu_hyp_init_features(void) 2106 { 2107 cpu_set_hyp_vector(); 2108 2109 if (is_kernel_in_hyp_mode()) 2110 kvm_timer_init_vhe(); 2111 2112 if (vgic_present) 2113 kvm_vgic_init_cpu_hardware(); 2114 } 2115 2116 static void cpu_hyp_reinit(void) 2117 { 2118 cpu_hyp_reset(); 2119 cpu_hyp_init_context(); 2120 cpu_hyp_init_features(); 2121 } 2122 2123 static void cpu_hyp_init(void *discard) 2124 { 2125 if (!__this_cpu_read(kvm_hyp_initialized)) { 2126 cpu_hyp_reinit(); 2127 __this_cpu_write(kvm_hyp_initialized, 1); 2128 } 2129 } 2130 2131 static void cpu_hyp_uninit(void *discard) 2132 { 2133 if (__this_cpu_read(kvm_hyp_initialized)) { 2134 cpu_hyp_reset(); 2135 __this_cpu_write(kvm_hyp_initialized, 0); 2136 } 2137 } 2138 2139 int kvm_arch_enable_virtualization_cpu(void) 2140 { 2141 /* 2142 * Most calls to this function are made with migration 2143 * disabled, but not with preemption disabled. The former is 2144 * enough to ensure correctness, but most of the helpers 2145 * expect the later and will throw a tantrum otherwise. 2146 */ 2147 preempt_disable(); 2148 2149 cpu_hyp_init(NULL); 2150 2151 kvm_vgic_cpu_up(); 2152 kvm_timer_cpu_up(); 2153 2154 preempt_enable(); 2155 2156 return 0; 2157 } 2158 2159 void kvm_arch_disable_virtualization_cpu(void) 2160 { 2161 kvm_timer_cpu_down(); 2162 kvm_vgic_cpu_down(); 2163 2164 if (!is_protected_kvm_enabled()) 2165 cpu_hyp_uninit(NULL); 2166 } 2167 2168 #ifdef CONFIG_CPU_PM 2169 static int hyp_init_cpu_pm_notifier(struct notifier_block *self, 2170 unsigned long cmd, 2171 void *v) 2172 { 2173 /* 2174 * kvm_hyp_initialized is left with its old value over 2175 * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should 2176 * re-enable hyp. 2177 */ 2178 switch (cmd) { 2179 case CPU_PM_ENTER: 2180 if (__this_cpu_read(kvm_hyp_initialized)) 2181 /* 2182 * don't update kvm_hyp_initialized here 2183 * so that the hyp will be re-enabled 2184 * when we resume. See below. 2185 */ 2186 cpu_hyp_reset(); 2187 2188 return NOTIFY_OK; 2189 case CPU_PM_ENTER_FAILED: 2190 case CPU_PM_EXIT: 2191 if (__this_cpu_read(kvm_hyp_initialized)) 2192 /* The hyp was enabled before suspend. */ 2193 cpu_hyp_reinit(); 2194 2195 return NOTIFY_OK; 2196 2197 default: 2198 return NOTIFY_DONE; 2199 } 2200 } 2201 2202 static struct notifier_block hyp_init_cpu_pm_nb = { 2203 .notifier_call = hyp_init_cpu_pm_notifier, 2204 }; 2205 2206 static void __init hyp_cpu_pm_init(void) 2207 { 2208 if (!is_protected_kvm_enabled()) 2209 cpu_pm_register_notifier(&hyp_init_cpu_pm_nb); 2210 } 2211 static void __init hyp_cpu_pm_exit(void) 2212 { 2213 if (!is_protected_kvm_enabled()) 2214 cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb); 2215 } 2216 #else 2217 static inline void __init hyp_cpu_pm_init(void) 2218 { 2219 } 2220 static inline void __init hyp_cpu_pm_exit(void) 2221 { 2222 } 2223 #endif 2224 2225 static void __init init_cpu_logical_map(void) 2226 { 2227 unsigned int cpu; 2228 2229 /* 2230 * Copy the MPIDR <-> logical CPU ID mapping to hyp. 2231 * Only copy the set of online CPUs whose features have been checked 2232 * against the finalized system capabilities. The hypervisor will not 2233 * allow any other CPUs from the `possible` set to boot. 2234 */ 2235 for_each_online_cpu(cpu) 2236 hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu); 2237 } 2238 2239 #define init_psci_0_1_impl_state(config, what) \ 2240 config.psci_0_1_ ## what ## _implemented = psci_ops.what 2241 2242 static bool __init init_psci_relay(void) 2243 { 2244 /* 2245 * If PSCI has not been initialized, protected KVM cannot install 2246 * itself on newly booted CPUs. 2247 */ 2248 if (!psci_ops.get_version) { 2249 kvm_err("Cannot initialize protected mode without PSCI\n"); 2250 return false; 2251 } 2252 2253 kvm_host_psci_config.version = psci_ops.get_version(); 2254 kvm_host_psci_config.smccc_version = arm_smccc_get_version(); 2255 2256 if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) { 2257 kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids(); 2258 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend); 2259 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on); 2260 init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off); 2261 init_psci_0_1_impl_state(kvm_host_psci_config, migrate); 2262 } 2263 return true; 2264 } 2265 2266 static int __init init_subsystems(void) 2267 { 2268 int err = 0; 2269 2270 /* 2271 * Enable hardware so that subsystem initialisation can access EL2. 2272 */ 2273 on_each_cpu(cpu_hyp_init, NULL, 1); 2274 2275 /* 2276 * Register CPU lower-power notifier 2277 */ 2278 hyp_cpu_pm_init(); 2279 2280 /* 2281 * Init HYP view of VGIC 2282 */ 2283 err = kvm_vgic_hyp_init(); 2284 switch (err) { 2285 case 0: 2286 vgic_present = true; 2287 break; 2288 case -ENODEV: 2289 case -ENXIO: 2290 /* 2291 * No VGIC? No pKVM for you. 2292 * 2293 * Protected mode assumes that VGICv3 is present, so no point 2294 * in trying to hobble along if vgic initialization fails. 2295 */ 2296 if (is_protected_kvm_enabled()) 2297 goto out; 2298 2299 /* 2300 * Otherwise, userspace could choose to implement a GIC for its 2301 * guest on non-cooperative hardware. 2302 */ 2303 vgic_present = false; 2304 err = 0; 2305 break; 2306 default: 2307 goto out; 2308 } 2309 2310 /* 2311 * Init HYP architected timer support 2312 */ 2313 err = kvm_timer_hyp_init(vgic_present); 2314 if (err) 2315 goto out; 2316 2317 kvm_register_perf_callbacks(NULL); 2318 2319 out: 2320 if (err) 2321 hyp_cpu_pm_exit(); 2322 2323 if (err || !is_protected_kvm_enabled()) 2324 on_each_cpu(cpu_hyp_uninit, NULL, 1); 2325 2326 return err; 2327 } 2328 2329 static void __init teardown_subsystems(void) 2330 { 2331 kvm_unregister_perf_callbacks(); 2332 hyp_cpu_pm_exit(); 2333 } 2334 2335 static void __init teardown_hyp_mode(void) 2336 { 2337 bool free_sve = system_supports_sve() && is_protected_kvm_enabled(); 2338 int cpu; 2339 2340 free_hyp_pgds(); 2341 for_each_possible_cpu(cpu) { 2342 free_pages(per_cpu(kvm_arm_hyp_stack_base, cpu), NVHE_STACK_SHIFT - PAGE_SHIFT); 2343 free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order()); 2344 2345 if (free_sve) { 2346 struct cpu_sve_state *sve_state; 2347 2348 sve_state = per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state; 2349 free_pages((unsigned long) sve_state, pkvm_host_sve_state_order()); 2350 } 2351 } 2352 } 2353 2354 static int __init do_pkvm_init(u32 hyp_va_bits) 2355 { 2356 void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)); 2357 int ret; 2358 2359 preempt_disable(); 2360 cpu_hyp_init_context(); 2361 ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size, 2362 num_possible_cpus(), kern_hyp_va(per_cpu_base), 2363 hyp_va_bits); 2364 cpu_hyp_init_features(); 2365 2366 /* 2367 * The stub hypercalls are now disabled, so set our local flag to 2368 * prevent a later re-init attempt in kvm_arch_enable_virtualization_cpu(). 2369 */ 2370 __this_cpu_write(kvm_hyp_initialized, 1); 2371 preempt_enable(); 2372 2373 return ret; 2374 } 2375 2376 static u64 get_hyp_id_aa64pfr0_el1(void) 2377 { 2378 /* 2379 * Track whether the system isn't affected by spectre/meltdown in the 2380 * hypervisor's view of id_aa64pfr0_el1, used for protected VMs. 2381 * Although this is per-CPU, we make it global for simplicity, e.g., not 2382 * to have to worry about vcpu migration. 2383 * 2384 * Unlike for non-protected VMs, userspace cannot override this for 2385 * protected VMs. 2386 */ 2387 u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1); 2388 2389 val &= ~(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2) | 2390 ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3)); 2391 2392 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV2), 2393 arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED); 2394 val |= FIELD_PREP(ARM64_FEATURE_MASK(ID_AA64PFR0_EL1_CSV3), 2395 arm64_get_meltdown_state() == SPECTRE_UNAFFECTED); 2396 2397 return val; 2398 } 2399 2400 static void kvm_hyp_init_symbols(void) 2401 { 2402 kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1(); 2403 kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1); 2404 kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1); 2405 kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1); 2406 kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1); 2407 kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1); 2408 kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1); 2409 kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1); 2410 kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1); 2411 kvm_nvhe_sym(__icache_flags) = __icache_flags; 2412 kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits; 2413 2414 /* 2415 * Flush entire BSS since part of its data containing init symbols is read 2416 * while the MMU is off. 2417 */ 2418 kvm_flush_dcache_to_poc(kvm_ksym_ref(__hyp_bss_start), 2419 kvm_ksym_ref(__hyp_bss_end) - kvm_ksym_ref(__hyp_bss_start)); 2420 } 2421 2422 static int __init kvm_hyp_init_protection(u32 hyp_va_bits) 2423 { 2424 void *addr = phys_to_virt(hyp_mem_base); 2425 int ret; 2426 2427 ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP); 2428 if (ret) 2429 return ret; 2430 2431 ret = do_pkvm_init(hyp_va_bits); 2432 if (ret) 2433 return ret; 2434 2435 free_hyp_pgds(); 2436 2437 return 0; 2438 } 2439 2440 static int init_pkvm_host_sve_state(void) 2441 { 2442 int cpu; 2443 2444 if (!system_supports_sve()) 2445 return 0; 2446 2447 /* Allocate pages for host sve state in protected mode. */ 2448 for_each_possible_cpu(cpu) { 2449 struct page *page = alloc_pages(GFP_KERNEL, pkvm_host_sve_state_order()); 2450 2451 if (!page) 2452 return -ENOMEM; 2453 2454 per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state = page_address(page); 2455 } 2456 2457 /* 2458 * Don't map the pages in hyp since these are only used in protected 2459 * mode, which will (re)create its own mapping when initialized. 2460 */ 2461 2462 return 0; 2463 } 2464 2465 /* 2466 * Finalizes the initialization of hyp mode, once everything else is initialized 2467 * and the initialziation process cannot fail. 2468 */ 2469 static void finalize_init_hyp_mode(void) 2470 { 2471 int cpu; 2472 2473 if (system_supports_sve() && is_protected_kvm_enabled()) { 2474 for_each_possible_cpu(cpu) { 2475 struct cpu_sve_state *sve_state; 2476 2477 sve_state = per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state; 2478 per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state = 2479 kern_hyp_va(sve_state); 2480 } 2481 } 2482 } 2483 2484 static void pkvm_hyp_init_ptrauth(void) 2485 { 2486 struct kvm_cpu_context *hyp_ctxt; 2487 int cpu; 2488 2489 for_each_possible_cpu(cpu) { 2490 hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu); 2491 hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long(); 2492 hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long(); 2493 hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long(); 2494 hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long(); 2495 hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long(); 2496 hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long(); 2497 hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long(); 2498 hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long(); 2499 hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long(); 2500 hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long(); 2501 } 2502 } 2503 2504 /* Inits Hyp-mode on all online CPUs */ 2505 static int __init init_hyp_mode(void) 2506 { 2507 u32 hyp_va_bits; 2508 int cpu; 2509 int err = -ENOMEM; 2510 2511 /* 2512 * The protected Hyp-mode cannot be initialized if the memory pool 2513 * allocation has failed. 2514 */ 2515 if (is_protected_kvm_enabled() && !hyp_mem_base) 2516 goto out_err; 2517 2518 /* 2519 * Allocate Hyp PGD and setup Hyp identity mapping 2520 */ 2521 err = kvm_mmu_init(&hyp_va_bits); 2522 if (err) 2523 goto out_err; 2524 2525 /* 2526 * Allocate stack pages for Hypervisor-mode 2527 */ 2528 for_each_possible_cpu(cpu) { 2529 unsigned long stack_base; 2530 2531 stack_base = __get_free_pages(GFP_KERNEL, NVHE_STACK_SHIFT - PAGE_SHIFT); 2532 if (!stack_base) { 2533 err = -ENOMEM; 2534 goto out_err; 2535 } 2536 2537 per_cpu(kvm_arm_hyp_stack_base, cpu) = stack_base; 2538 } 2539 2540 /* 2541 * Allocate and initialize pages for Hypervisor-mode percpu regions. 2542 */ 2543 for_each_possible_cpu(cpu) { 2544 struct page *page; 2545 void *page_addr; 2546 2547 page = alloc_pages(GFP_KERNEL, nvhe_percpu_order()); 2548 if (!page) { 2549 err = -ENOMEM; 2550 goto out_err; 2551 } 2552 2553 page_addr = page_address(page); 2554 memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size()); 2555 kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr; 2556 } 2557 2558 /* 2559 * Map the Hyp-code called directly from the host 2560 */ 2561 err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start), 2562 kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC); 2563 if (err) { 2564 kvm_err("Cannot map world-switch code\n"); 2565 goto out_err; 2566 } 2567 2568 err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start), 2569 kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO); 2570 if (err) { 2571 kvm_err("Cannot map .hyp.rodata section\n"); 2572 goto out_err; 2573 } 2574 2575 err = create_hyp_mappings(kvm_ksym_ref(__start_rodata), 2576 kvm_ksym_ref(__end_rodata), PAGE_HYP_RO); 2577 if (err) { 2578 kvm_err("Cannot map rodata section\n"); 2579 goto out_err; 2580 } 2581 2582 /* 2583 * .hyp.bss is guaranteed to be placed at the beginning of the .bss 2584 * section thanks to an assertion in the linker script. Map it RW and 2585 * the rest of .bss RO. 2586 */ 2587 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start), 2588 kvm_ksym_ref(__hyp_bss_end), PAGE_HYP); 2589 if (err) { 2590 kvm_err("Cannot map hyp bss section: %d\n", err); 2591 goto out_err; 2592 } 2593 2594 err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end), 2595 kvm_ksym_ref(__bss_stop), PAGE_HYP_RO); 2596 if (err) { 2597 kvm_err("Cannot map bss section\n"); 2598 goto out_err; 2599 } 2600 2601 /* 2602 * Map the Hyp stack pages 2603 */ 2604 for_each_possible_cpu(cpu) { 2605 struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu); 2606 char *stack_base = (char *)per_cpu(kvm_arm_hyp_stack_base, cpu); 2607 2608 err = create_hyp_stack(__pa(stack_base), ¶ms->stack_hyp_va); 2609 if (err) { 2610 kvm_err("Cannot map hyp stack\n"); 2611 goto out_err; 2612 } 2613 2614 /* 2615 * Save the stack PA in nvhe_init_params. This will be needed 2616 * to recreate the stack mapping in protected nVHE mode. 2617 * __hyp_pa() won't do the right thing there, since the stack 2618 * has been mapped in the flexible private VA space. 2619 */ 2620 params->stack_pa = __pa(stack_base); 2621 } 2622 2623 for_each_possible_cpu(cpu) { 2624 char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu]; 2625 char *percpu_end = percpu_begin + nvhe_percpu_size(); 2626 2627 /* Map Hyp percpu pages */ 2628 err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP); 2629 if (err) { 2630 kvm_err("Cannot map hyp percpu region\n"); 2631 goto out_err; 2632 } 2633 2634 /* Prepare the CPU initialization parameters */ 2635 cpu_prepare_hyp_mode(cpu, hyp_va_bits); 2636 } 2637 2638 kvm_hyp_init_symbols(); 2639 2640 if (is_protected_kvm_enabled()) { 2641 if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) && 2642 cpus_have_final_cap(ARM64_HAS_ADDRESS_AUTH)) 2643 pkvm_hyp_init_ptrauth(); 2644 2645 init_cpu_logical_map(); 2646 2647 if (!init_psci_relay()) { 2648 err = -ENODEV; 2649 goto out_err; 2650 } 2651 2652 err = init_pkvm_host_sve_state(); 2653 if (err) 2654 goto out_err; 2655 2656 err = kvm_hyp_init_protection(hyp_va_bits); 2657 if (err) { 2658 kvm_err("Failed to init hyp memory protection\n"); 2659 goto out_err; 2660 } 2661 } 2662 2663 return 0; 2664 2665 out_err: 2666 teardown_hyp_mode(); 2667 kvm_err("error initializing Hyp mode: %d\n", err); 2668 return err; 2669 } 2670 2671 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr) 2672 { 2673 struct kvm_vcpu *vcpu = NULL; 2674 struct kvm_mpidr_data *data; 2675 unsigned long i; 2676 2677 mpidr &= MPIDR_HWID_BITMASK; 2678 2679 rcu_read_lock(); 2680 data = rcu_dereference(kvm->arch.mpidr_data); 2681 2682 if (data) { 2683 u16 idx = kvm_mpidr_index(data, mpidr); 2684 2685 vcpu = kvm_get_vcpu(kvm, data->cmpidr_to_idx[idx]); 2686 if (mpidr != kvm_vcpu_get_mpidr_aff(vcpu)) 2687 vcpu = NULL; 2688 } 2689 2690 rcu_read_unlock(); 2691 2692 if (vcpu) 2693 return vcpu; 2694 2695 kvm_for_each_vcpu(i, vcpu, kvm) { 2696 if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu)) 2697 return vcpu; 2698 } 2699 return NULL; 2700 } 2701 2702 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm) 2703 { 2704 return irqchip_in_kernel(kvm); 2705 } 2706 2707 bool kvm_arch_has_irq_bypass(void) 2708 { 2709 return true; 2710 } 2711 2712 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons, 2713 struct irq_bypass_producer *prod) 2714 { 2715 struct kvm_kernel_irqfd *irqfd = 2716 container_of(cons, struct kvm_kernel_irqfd, consumer); 2717 2718 return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq, 2719 &irqfd->irq_entry); 2720 } 2721 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons, 2722 struct irq_bypass_producer *prod) 2723 { 2724 struct kvm_kernel_irqfd *irqfd = 2725 container_of(cons, struct kvm_kernel_irqfd, consumer); 2726 2727 kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq, 2728 &irqfd->irq_entry); 2729 } 2730 2731 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons) 2732 { 2733 struct kvm_kernel_irqfd *irqfd = 2734 container_of(cons, struct kvm_kernel_irqfd, consumer); 2735 2736 kvm_arm_halt_guest(irqfd->kvm); 2737 } 2738 2739 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons) 2740 { 2741 struct kvm_kernel_irqfd *irqfd = 2742 container_of(cons, struct kvm_kernel_irqfd, consumer); 2743 2744 kvm_arm_resume_guest(irqfd->kvm); 2745 } 2746 2747 /* Initialize Hyp-mode and memory mappings on all CPUs */ 2748 static __init int kvm_arm_init(void) 2749 { 2750 int err; 2751 bool in_hyp_mode; 2752 2753 if (!is_hyp_mode_available()) { 2754 kvm_info("HYP mode not available\n"); 2755 return -ENODEV; 2756 } 2757 2758 if (kvm_get_mode() == KVM_MODE_NONE) { 2759 kvm_info("KVM disabled from command line\n"); 2760 return -ENODEV; 2761 } 2762 2763 err = kvm_sys_reg_table_init(); 2764 if (err) { 2765 kvm_info("Error initializing system register tables"); 2766 return err; 2767 } 2768 2769 in_hyp_mode = is_kernel_in_hyp_mode(); 2770 2771 if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) || 2772 cpus_have_final_cap(ARM64_WORKAROUND_1508412)) 2773 kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \ 2774 "Only trusted guests should be used on this system.\n"); 2775 2776 err = kvm_set_ipa_limit(); 2777 if (err) 2778 return err; 2779 2780 err = kvm_arm_init_sve(); 2781 if (err) 2782 return err; 2783 2784 err = kvm_arm_vmid_alloc_init(); 2785 if (err) { 2786 kvm_err("Failed to initialize VMID allocator.\n"); 2787 return err; 2788 } 2789 2790 if (!in_hyp_mode) { 2791 err = init_hyp_mode(); 2792 if (err) 2793 goto out_err; 2794 } 2795 2796 err = kvm_init_vector_slots(); 2797 if (err) { 2798 kvm_err("Cannot initialise vector slots\n"); 2799 goto out_hyp; 2800 } 2801 2802 err = init_subsystems(); 2803 if (err) 2804 goto out_hyp; 2805 2806 kvm_info("%s%sVHE mode initialized successfully\n", 2807 in_hyp_mode ? "" : (is_protected_kvm_enabled() ? 2808 "Protected " : "Hyp "), 2809 in_hyp_mode ? "" : (cpus_have_final_cap(ARM64_KVM_HVHE) ? 2810 "h" : "n")); 2811 2812 /* 2813 * FIXME: Do something reasonable if kvm_init() fails after pKVM 2814 * hypervisor protection is finalized. 2815 */ 2816 err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE); 2817 if (err) 2818 goto out_subs; 2819 2820 /* 2821 * This should be called after initialization is done and failure isn't 2822 * possible anymore. 2823 */ 2824 if (!in_hyp_mode) 2825 finalize_init_hyp_mode(); 2826 2827 kvm_arm_initialised = true; 2828 2829 return 0; 2830 2831 out_subs: 2832 teardown_subsystems(); 2833 out_hyp: 2834 if (!in_hyp_mode) 2835 teardown_hyp_mode(); 2836 out_err: 2837 kvm_arm_vmid_alloc_free(); 2838 return err; 2839 } 2840 2841 static int __init early_kvm_mode_cfg(char *arg) 2842 { 2843 if (!arg) 2844 return -EINVAL; 2845 2846 if (strcmp(arg, "none") == 0) { 2847 kvm_mode = KVM_MODE_NONE; 2848 return 0; 2849 } 2850 2851 if (!is_hyp_mode_available()) { 2852 pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n"); 2853 return 0; 2854 } 2855 2856 if (strcmp(arg, "protected") == 0) { 2857 if (!is_kernel_in_hyp_mode()) 2858 kvm_mode = KVM_MODE_PROTECTED; 2859 else 2860 pr_warn_once("Protected KVM not available with VHE\n"); 2861 2862 return 0; 2863 } 2864 2865 if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) { 2866 kvm_mode = KVM_MODE_DEFAULT; 2867 return 0; 2868 } 2869 2870 if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) { 2871 kvm_mode = KVM_MODE_NV; 2872 return 0; 2873 } 2874 2875 return -EINVAL; 2876 } 2877 early_param("kvm-arm.mode", early_kvm_mode_cfg); 2878 2879 static int __init early_kvm_wfx_trap_policy_cfg(char *arg, enum kvm_wfx_trap_policy *p) 2880 { 2881 if (!arg) 2882 return -EINVAL; 2883 2884 if (strcmp(arg, "trap") == 0) { 2885 *p = KVM_WFX_TRAP; 2886 return 0; 2887 } 2888 2889 if (strcmp(arg, "notrap") == 0) { 2890 *p = KVM_WFX_NOTRAP; 2891 return 0; 2892 } 2893 2894 return -EINVAL; 2895 } 2896 2897 static int __init early_kvm_wfi_trap_policy_cfg(char *arg) 2898 { 2899 return early_kvm_wfx_trap_policy_cfg(arg, &kvm_wfi_trap_policy); 2900 } 2901 early_param("kvm-arm.wfi_trap_policy", early_kvm_wfi_trap_policy_cfg); 2902 2903 static int __init early_kvm_wfe_trap_policy_cfg(char *arg) 2904 { 2905 return early_kvm_wfx_trap_policy_cfg(arg, &kvm_wfe_trap_policy); 2906 } 2907 early_param("kvm-arm.wfe_trap_policy", early_kvm_wfe_trap_policy_cfg); 2908 2909 enum kvm_mode kvm_get_mode(void) 2910 { 2911 return kvm_mode; 2912 } 2913 2914 module_init(kvm_arm_init); 2915