xref: /linux/arch/arm64/kvm/arm.c (revision 7c31c06e2d2d75859d773ba940e56d1db2bd1fcd)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
4  * Author: Christoffer Dall <c.dall@virtualopensystems.com>
5  */
6 
7 #include <linux/bug.h>
8 #include <linux/cpu_pm.h>
9 #include <linux/errno.h>
10 #include <linux/err.h>
11 #include <linux/kvm_host.h>
12 #include <linux/list.h>
13 #include <linux/module.h>
14 #include <linux/vmalloc.h>
15 #include <linux/fs.h>
16 #include <linux/mman.h>
17 #include <linux/sched.h>
18 #include <linux/kvm.h>
19 #include <linux/kvm_irqfd.h>
20 #include <linux/irqbypass.h>
21 #include <linux/sched/stat.h>
22 #include <linux/psci.h>
23 #include <trace/events/kvm.h>
24 
25 #define CREATE_TRACE_POINTS
26 #include "trace_arm.h"
27 
28 #include <linux/uaccess.h>
29 #include <asm/ptrace.h>
30 #include <asm/mman.h>
31 #include <asm/tlbflush.h>
32 #include <asm/cacheflush.h>
33 #include <asm/cpufeature.h>
34 #include <asm/virt.h>
35 #include <asm/kvm_arm.h>
36 #include <asm/kvm_asm.h>
37 #include <asm/kvm_emulate.h>
38 #include <asm/kvm_mmu.h>
39 #include <asm/kvm_nested.h>
40 #include <asm/kvm_pkvm.h>
41 #include <asm/kvm_ptrauth.h>
42 #include <asm/sections.h>
43 #include <asm/stacktrace/nvhe.h>
44 
45 #include <kvm/arm_hypercalls.h>
46 #include <kvm/arm_pmu.h>
47 #include <kvm/arm_psci.h>
48 #include <kvm/arm_vgic.h>
49 
50 #include <linux/irqchip/arm-gic-v5.h>
51 
52 #include "sys_regs.h"
53 
54 static enum kvm_mode kvm_mode = KVM_MODE_DEFAULT;
55 
56 enum kvm_wfx_trap_policy {
57 	KVM_WFX_NOTRAP_SINGLE_TASK, /* Default option */
58 	KVM_WFX_NOTRAP,
59 	KVM_WFX_TRAP,
60 };
61 
62 static enum kvm_wfx_trap_policy kvm_wfi_trap_policy __read_mostly = KVM_WFX_NOTRAP_SINGLE_TASK;
63 static enum kvm_wfx_trap_policy kvm_wfe_trap_policy __read_mostly = KVM_WFX_NOTRAP_SINGLE_TASK;
64 
65 /*
66  * Tracks KVM IOCTLs and their associated KVM capabilities.
67  */
68 struct kvm_ioctl_cap_map {
69 	unsigned int ioctl;
70 	long ext;
71 };
72 
73 /* Make KVM_CAP_NR_VCPUS the reference for features we always supported */
74 #define KVM_CAP_ARM_BASIC	KVM_CAP_NR_VCPUS
75 
76 /*
77  * Sorted by ioctl to allow for potential binary search,
78  * though linear scan is sufficient for this size.
79  */
80 static const struct kvm_ioctl_cap_map vm_ioctl_caps[] = {
81 	{ KVM_CREATE_IRQCHIP, KVM_CAP_IRQCHIP },
82 	{ KVM_ARM_SET_DEVICE_ADDR, KVM_CAP_ARM_SET_DEVICE_ADDR },
83 	{ KVM_ARM_MTE_COPY_TAGS, KVM_CAP_ARM_MTE },
84 	{ KVM_SET_DEVICE_ATTR, KVM_CAP_DEVICE_CTRL },
85 	{ KVM_GET_DEVICE_ATTR, KVM_CAP_DEVICE_CTRL },
86 	{ KVM_HAS_DEVICE_ATTR, KVM_CAP_DEVICE_CTRL },
87 	{ KVM_ARM_SET_COUNTER_OFFSET, KVM_CAP_COUNTER_OFFSET },
88 	{ KVM_ARM_GET_REG_WRITABLE_MASKS, KVM_CAP_ARM_SUPPORTED_REG_MASK_RANGES },
89 	{ KVM_ARM_PREFERRED_TARGET, KVM_CAP_ARM_BASIC },
90 };
91 
92 /*
93  * Set *ext to the capability.
94  * Return 0 if found, or -EINVAL if no IOCTL matches.
95  */
96 long kvm_get_cap_for_kvm_ioctl(unsigned int ioctl, long *ext)
97 {
98 	int i;
99 
100 	for (i = 0; i < ARRAY_SIZE(vm_ioctl_caps); i++) {
101 		if (vm_ioctl_caps[i].ioctl == ioctl) {
102 			*ext = vm_ioctl_caps[i].ext;
103 			return 0;
104 		}
105 	}
106 
107 	return -EINVAL;
108 }
109 
110 DECLARE_KVM_HYP_PER_CPU(unsigned long, kvm_hyp_vector);
111 
112 DEFINE_PER_CPU(unsigned long, kvm_arm_hyp_stack_base);
113 DECLARE_KVM_NVHE_PER_CPU(struct kvm_nvhe_init_params, kvm_init_params);
114 
115 DECLARE_KVM_NVHE_PER_CPU(struct kvm_cpu_context, kvm_hyp_ctxt);
116 
117 static bool vgic_present, kvm_arm_initialised;
118 
119 static DEFINE_PER_CPU(unsigned char, kvm_hyp_initialized);
120 
121 bool is_kvm_arm_initialised(void)
122 {
123 	return kvm_arm_initialised;
124 }
125 
126 int kvm_arch_vcpu_should_kick(struct kvm_vcpu *vcpu)
127 {
128 	return kvm_vcpu_exiting_guest_mode(vcpu) == IN_GUEST_MODE;
129 }
130 
131 int kvm_vm_ioctl_enable_cap(struct kvm *kvm,
132 			    struct kvm_enable_cap *cap)
133 {
134 	int r = -EINVAL;
135 
136 	if (cap->flags)
137 		return -EINVAL;
138 
139 	if (is_protected_kvm_enabled() && !kvm_pkvm_ext_allowed(kvm, cap->cap))
140 		return -EINVAL;
141 
142 	switch (cap->cap) {
143 	case KVM_CAP_ARM_NISV_TO_USER:
144 		r = 0;
145 		set_bit(KVM_ARCH_FLAG_RETURN_NISV_IO_ABORT_TO_USER,
146 			&kvm->arch.flags);
147 		break;
148 	case KVM_CAP_ARM_MTE:
149 		mutex_lock(&kvm->lock);
150 		if (system_supports_mte() && !kvm->created_vcpus) {
151 			r = 0;
152 			set_bit(KVM_ARCH_FLAG_MTE_ENABLED, &kvm->arch.flags);
153 		}
154 		mutex_unlock(&kvm->lock);
155 		break;
156 	case KVM_CAP_ARM_SYSTEM_SUSPEND:
157 		r = 0;
158 		set_bit(KVM_ARCH_FLAG_SYSTEM_SUSPEND_ENABLED, &kvm->arch.flags);
159 		break;
160 	case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
161 		mutex_lock(&kvm->slots_lock);
162 		/*
163 		 * To keep things simple, allow changing the chunk
164 		 * size only when no memory slots have been created.
165 		 */
166 		if (kvm_are_all_memslots_empty(kvm)) {
167 			u64 new_cap = cap->args[0];
168 
169 			if (!new_cap || kvm_is_block_size_supported(new_cap)) {
170 				r = 0;
171 				kvm->arch.mmu.split_page_chunk_size = new_cap;
172 			}
173 		}
174 		mutex_unlock(&kvm->slots_lock);
175 		break;
176 	case KVM_CAP_ARM_WRITABLE_IMP_ID_REGS:
177 		mutex_lock(&kvm->lock);
178 		if (!kvm->created_vcpus) {
179 			r = 0;
180 			set_bit(KVM_ARCH_FLAG_WRITABLE_IMP_ID_REGS, &kvm->arch.flags);
181 		}
182 		mutex_unlock(&kvm->lock);
183 		break;
184 	case KVM_CAP_ARM_SEA_TO_USER:
185 		r = 0;
186 		set_bit(KVM_ARCH_FLAG_EXIT_SEA, &kvm->arch.flags);
187 		break;
188 	default:
189 		break;
190 	}
191 
192 	return r;
193 }
194 
195 static int kvm_arm_default_max_vcpus(void)
196 {
197 	return vgic_present ? kvm_vgic_get_max_vcpus() : KVM_MAX_VCPUS;
198 }
199 
200 /**
201  * kvm_arch_init_vm - initializes a VM data structure
202  * @kvm:	pointer to the KVM struct
203  * @type:	kvm device type
204  */
205 int kvm_arch_init_vm(struct kvm *kvm, unsigned long type)
206 {
207 	int ret;
208 
209 	mutex_init(&kvm->arch.config_lock);
210 
211 #ifdef CONFIG_LOCKDEP
212 	/* Clue in lockdep that the config_lock must be taken inside kvm->lock */
213 	mutex_lock(&kvm->lock);
214 	mutex_lock(&kvm->arch.config_lock);
215 	mutex_unlock(&kvm->arch.config_lock);
216 	mutex_unlock(&kvm->lock);
217 #endif
218 
219 	kvm_init_nested(kvm);
220 
221 	ret = kvm_share_hyp(kvm, kvm + 1);
222 	if (ret)
223 		return ret;
224 
225 	if (!zalloc_cpumask_var(&kvm->arch.supported_cpus, GFP_KERNEL_ACCOUNT)) {
226 		ret = -ENOMEM;
227 		goto err_unshare_kvm;
228 	}
229 	cpumask_copy(kvm->arch.supported_cpus, cpu_possible_mask);
230 
231 	ret = kvm_init_stage2_mmu(kvm, &kvm->arch.mmu, type);
232 	if (ret)
233 		goto err_free_cpumask;
234 
235 	if (is_protected_kvm_enabled()) {
236 		/*
237 		 * If any failures occur after this is successful, make sure to
238 		 * call __pkvm_unreserve_vm to unreserve the VM in hyp.
239 		 */
240 		ret = pkvm_init_host_vm(kvm);
241 		if (ret)
242 			goto err_free_cpumask;
243 	}
244 
245 	kvm_vgic_early_init(kvm);
246 
247 	kvm_timer_init_vm(kvm);
248 
249 	/* The maximum number of VCPUs is limited by the host's GIC model */
250 	kvm->max_vcpus = kvm_arm_default_max_vcpus();
251 
252 	kvm_arm_init_hypercalls(kvm);
253 
254 	bitmap_zero(kvm->arch.vcpu_features, KVM_VCPU_MAX_FEATURES);
255 
256 	return 0;
257 
258 err_free_cpumask:
259 	free_cpumask_var(kvm->arch.supported_cpus);
260 err_unshare_kvm:
261 	kvm_unshare_hyp(kvm, kvm + 1);
262 	return ret;
263 }
264 
265 vm_fault_t kvm_arch_vcpu_fault(struct kvm_vcpu *vcpu, struct vm_fault *vmf)
266 {
267 	return VM_FAULT_SIGBUS;
268 }
269 
270 void kvm_arch_create_vm_debugfs(struct kvm *kvm)
271 {
272 	kvm_sys_regs_create_debugfs(kvm);
273 	kvm_s2_ptdump_create_debugfs(kvm);
274 }
275 
276 static void kvm_destroy_mpidr_data(struct kvm *kvm)
277 {
278 	struct kvm_mpidr_data *data;
279 
280 	mutex_lock(&kvm->arch.config_lock);
281 
282 	data = rcu_dereference_protected(kvm->arch.mpidr_data,
283 					 lockdep_is_held(&kvm->arch.config_lock));
284 	if (data) {
285 		rcu_assign_pointer(kvm->arch.mpidr_data, NULL);
286 		synchronize_rcu();
287 		kfree(data);
288 	}
289 
290 	mutex_unlock(&kvm->arch.config_lock);
291 }
292 
293 /**
294  * kvm_arch_destroy_vm - destroy the VM data structure
295  * @kvm:	pointer to the KVM struct
296  */
297 void kvm_arch_destroy_vm(struct kvm *kvm)
298 {
299 	bitmap_free(kvm->arch.pmu_filter);
300 	free_cpumask_var(kvm->arch.supported_cpus);
301 
302 	kvm_vgic_destroy(kvm);
303 
304 	if (is_protected_kvm_enabled())
305 		pkvm_destroy_hyp_vm(kvm);
306 
307 	kvm_destroy_mpidr_data(kvm);
308 
309 	kfree(kvm->arch.sysreg_masks);
310 	kvm_destroy_vcpus(kvm);
311 
312 	kvm_unshare_hyp(kvm, kvm + 1);
313 
314 	kvm_arm_teardown_hypercalls(kvm);
315 }
316 
317 static bool kvm_has_full_ptr_auth(void)
318 {
319 	bool apa, gpa, api, gpi, apa3, gpa3;
320 	u64 isar1, isar2, val;
321 
322 	/*
323 	 * Check that:
324 	 *
325 	 * - both Address and Generic auth are implemented for a given
326          *   algorithm (Q5, IMPDEF or Q3)
327 	 * - only a single algorithm is implemented.
328 	 */
329 	if (!system_has_full_ptr_auth())
330 		return false;
331 
332 	isar1 = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
333 	isar2 = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
334 
335 	apa = !!FIELD_GET(ID_AA64ISAR1_EL1_APA_MASK, isar1);
336 	val = FIELD_GET(ID_AA64ISAR1_EL1_GPA_MASK, isar1);
337 	gpa = (val == ID_AA64ISAR1_EL1_GPA_IMP);
338 
339 	api = !!FIELD_GET(ID_AA64ISAR1_EL1_API_MASK, isar1);
340 	val = FIELD_GET(ID_AA64ISAR1_EL1_GPI_MASK, isar1);
341 	gpi = (val == ID_AA64ISAR1_EL1_GPI_IMP);
342 
343 	apa3 = !!FIELD_GET(ID_AA64ISAR2_EL1_APA3_MASK, isar2);
344 	val  = FIELD_GET(ID_AA64ISAR2_EL1_GPA3_MASK, isar2);
345 	gpa3 = (val == ID_AA64ISAR2_EL1_GPA3_IMP);
346 
347 	return (apa == gpa && api == gpi && apa3 == gpa3 &&
348 		(apa + api + apa3) == 1);
349 }
350 
351 int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
352 {
353 	int r;
354 
355 	if (is_protected_kvm_enabled() && !kvm_pkvm_ext_allowed(kvm, ext))
356 		return 0;
357 
358 	switch (ext) {
359 	case KVM_CAP_IRQCHIP:
360 		r = vgic_present;
361 		break;
362 	case KVM_CAP_IOEVENTFD:
363 	case KVM_CAP_USER_MEMORY:
364 	case KVM_CAP_DESTROY_MEMORY_REGION_WORKS:
365 	case KVM_CAP_ONE_REG:
366 	case KVM_CAP_ARM_PSCI:
367 	case KVM_CAP_ARM_PSCI_0_2:
368 	case KVM_CAP_READONLY_MEM:
369 	case KVM_CAP_MP_STATE:
370 	case KVM_CAP_IMMEDIATE_EXIT:
371 	case KVM_CAP_VCPU_EVENTS:
372 	case KVM_CAP_ARM_IRQ_LINE_LAYOUT_2:
373 	case KVM_CAP_ARM_NISV_TO_USER:
374 	case KVM_CAP_ARM_INJECT_EXT_DABT:
375 	case KVM_CAP_SET_GUEST_DEBUG:
376 	case KVM_CAP_VCPU_ATTRIBUTES:
377 	case KVM_CAP_PTP_KVM:
378 	case KVM_CAP_ARM_SYSTEM_SUSPEND:
379 	case KVM_CAP_IRQFD_RESAMPLE:
380 	case KVM_CAP_COUNTER_OFFSET:
381 	case KVM_CAP_ARM_WRITABLE_IMP_ID_REGS:
382 	case KVM_CAP_ARM_SEA_TO_USER:
383 		r = 1;
384 		break;
385 	case KVM_CAP_SET_GUEST_DEBUG2:
386 		return KVM_GUESTDBG_VALID_MASK;
387 	case KVM_CAP_ARM_SET_DEVICE_ADDR:
388 		r = 1;
389 		break;
390 	case KVM_CAP_NR_VCPUS:
391 		/*
392 		 * ARM64 treats KVM_CAP_NR_CPUS differently from all other
393 		 * architectures, as it does not always bound it to
394 		 * KVM_CAP_MAX_VCPUS. It should not matter much because
395 		 * this is just an advisory value.
396 		 */
397 		r = min_t(unsigned int, num_online_cpus(),
398 			  kvm_arm_default_max_vcpus());
399 		break;
400 	case KVM_CAP_MAX_VCPUS:
401 	case KVM_CAP_MAX_VCPU_ID:
402 		if (kvm)
403 			r = kvm->max_vcpus;
404 		else
405 			r = kvm_arm_default_max_vcpus();
406 		break;
407 	case KVM_CAP_MSI_DEVID:
408 		if (!kvm)
409 			r = -EINVAL;
410 		else
411 			r = kvm->arch.vgic.msis_require_devid;
412 		break;
413 	case KVM_CAP_ARM_USER_IRQ:
414 		/*
415 		 * 1: EL1_VTIMER, EL1_PTIMER, and PMU.
416 		 * (bump this number if adding more devices)
417 		 */
418 		r = 1;
419 		break;
420 	case KVM_CAP_ARM_MTE:
421 		r = system_supports_mte();
422 		break;
423 	case KVM_CAP_STEAL_TIME:
424 		r = kvm_arm_pvtime_supported();
425 		break;
426 	case KVM_CAP_ARM_EL1_32BIT:
427 		r = cpus_have_final_cap(ARM64_HAS_32BIT_EL1);
428 		break;
429 	case KVM_CAP_ARM_EL2:
430 		r = cpus_have_final_cap(ARM64_HAS_NESTED_VIRT);
431 		break;
432 	case KVM_CAP_ARM_EL2_E2H0:
433 		r = cpus_have_final_cap(ARM64_HAS_HCR_NV1);
434 		break;
435 	case KVM_CAP_GUEST_DEBUG_HW_BPS:
436 		r = get_num_brps();
437 		break;
438 	case KVM_CAP_GUEST_DEBUG_HW_WPS:
439 		r = get_num_wrps();
440 		break;
441 	case KVM_CAP_ARM_PMU_V3:
442 		r = kvm_supports_guest_pmuv3();
443 		break;
444 	case KVM_CAP_ARM_INJECT_SERROR_ESR:
445 		r = cpus_have_final_cap(ARM64_HAS_RAS_EXTN);
446 		break;
447 	case KVM_CAP_ARM_VM_IPA_SIZE:
448 		r = get_kvm_ipa_limit();
449 		break;
450 	case KVM_CAP_ARM_SVE:
451 		r = system_supports_sve();
452 		break;
453 	case KVM_CAP_ARM_PTRAUTH_ADDRESS:
454 	case KVM_CAP_ARM_PTRAUTH_GENERIC:
455 		r = kvm_has_full_ptr_auth();
456 		break;
457 	case KVM_CAP_ARM_EAGER_SPLIT_CHUNK_SIZE:
458 		if (kvm)
459 			r = kvm->arch.mmu.split_page_chunk_size;
460 		else
461 			r = KVM_ARM_EAGER_SPLIT_CHUNK_SIZE_DEFAULT;
462 		break;
463 	case KVM_CAP_ARM_SUPPORTED_BLOCK_SIZES:
464 		r = kvm_supported_block_sizes();
465 		break;
466 	case KVM_CAP_ARM_SUPPORTED_REG_MASK_RANGES:
467 		r = BIT(0);
468 		break;
469 	case KVM_CAP_ARM_CACHEABLE_PFNMAP_SUPPORTED:
470 		if (!kvm)
471 			r = -EINVAL;
472 		else
473 			r = kvm_supports_cacheable_pfnmap();
474 		break;
475 
476 	default:
477 		r = 0;
478 	}
479 
480 	return r;
481 }
482 
483 long kvm_arch_dev_ioctl(struct file *filp,
484 			unsigned int ioctl, unsigned long arg)
485 {
486 	return -EINVAL;
487 }
488 
489 struct kvm *kvm_arch_alloc_vm(void)
490 {
491 	size_t sz = sizeof(struct kvm);
492 
493 	if (!has_vhe())
494 		return kzalloc(sz, GFP_KERNEL_ACCOUNT);
495 
496 	return kvzalloc(sz, GFP_KERNEL_ACCOUNT);
497 }
498 
499 int kvm_arch_vcpu_precreate(struct kvm *kvm, unsigned int id)
500 {
501 	if (irqchip_in_kernel(kvm) && vgic_initialized(kvm))
502 		return -EBUSY;
503 
504 	if (id >= kvm->max_vcpus)
505 		return -EINVAL;
506 
507 	return 0;
508 }
509 
510 int kvm_arch_vcpu_create(struct kvm_vcpu *vcpu)
511 {
512 	int err;
513 
514 	spin_lock_init(&vcpu->arch.mp_state_lock);
515 
516 #ifdef CONFIG_LOCKDEP
517 	/* Inform lockdep that the config_lock is acquired after vcpu->mutex */
518 	mutex_lock(&vcpu->mutex);
519 	mutex_lock(&vcpu->kvm->arch.config_lock);
520 	mutex_unlock(&vcpu->kvm->arch.config_lock);
521 	mutex_unlock(&vcpu->mutex);
522 #endif
523 
524 	/* Force users to call KVM_ARM_VCPU_INIT */
525 	vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
526 
527 	vcpu->arch.mmu_page_cache.gfp_zero = __GFP_ZERO;
528 
529 	/* Set up the timer */
530 	kvm_timer_vcpu_init(vcpu);
531 
532 	kvm_pmu_vcpu_init(vcpu);
533 
534 	kvm_arm_pvtime_vcpu_init(&vcpu->arch);
535 
536 	vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
537 
538 	/*
539 	 * This vCPU may have been created after mpidr_data was initialized.
540 	 * Throw out the pre-computed mappings if that is the case which forces
541 	 * KVM to fall back to iteratively searching the vCPUs.
542 	 */
543 	kvm_destroy_mpidr_data(vcpu->kvm);
544 
545 	err = kvm_vgic_vcpu_init(vcpu);
546 	if (err)
547 		return err;
548 
549 	err = kvm_share_hyp(vcpu, vcpu + 1);
550 	if (err)
551 		kvm_vgic_vcpu_destroy(vcpu);
552 
553 	return err;
554 }
555 
556 void kvm_arch_vcpu_postcreate(struct kvm_vcpu *vcpu)
557 {
558 }
559 
560 void kvm_arch_vcpu_destroy(struct kvm_vcpu *vcpu)
561 {
562 	if (!is_protected_kvm_enabled())
563 		kvm_mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
564 	else
565 		free_hyp_memcache(&vcpu->arch.pkvm_memcache);
566 	kvm_timer_vcpu_terminate(vcpu);
567 	kvm_pmu_vcpu_destroy(vcpu);
568 	kvm_vgic_vcpu_destroy(vcpu);
569 	kvm_arm_vcpu_destroy(vcpu);
570 }
571 
572 void kvm_arch_vcpu_blocking(struct kvm_vcpu *vcpu)
573 {
574 
575 }
576 
577 void kvm_arch_vcpu_unblocking(struct kvm_vcpu *vcpu)
578 {
579 
580 }
581 
582 static void vcpu_set_pauth_traps(struct kvm_vcpu *vcpu)
583 {
584 	if (vcpu_has_ptrauth(vcpu) && !is_protected_kvm_enabled()) {
585 		/*
586 		 * Either we're running an L2 guest, and the API/APK bits come
587 		 * from L1's HCR_EL2, or API/APK are both set.
588 		 */
589 		if (unlikely(is_nested_ctxt(vcpu))) {
590 			u64 val;
591 
592 			val = __vcpu_sys_reg(vcpu, HCR_EL2);
593 			val &= (HCR_API | HCR_APK);
594 			vcpu->arch.hcr_el2 &= ~(HCR_API | HCR_APK);
595 			vcpu->arch.hcr_el2 |= val;
596 		} else {
597 			vcpu->arch.hcr_el2 |= (HCR_API | HCR_APK);
598 		}
599 
600 		/*
601 		 * Save the host keys if there is any chance for the guest
602 		 * to use pauth, as the entry code will reload the guest
603 		 * keys in that case.
604 		 */
605 		if (vcpu->arch.hcr_el2 & (HCR_API | HCR_APK)) {
606 			struct kvm_cpu_context *ctxt;
607 
608 			ctxt = this_cpu_ptr_hyp_sym(kvm_hyp_ctxt);
609 			ptrauth_save_keys(ctxt);
610 		}
611 	}
612 }
613 
614 static bool kvm_vcpu_should_clear_twi(struct kvm_vcpu *vcpu)
615 {
616 	if (unlikely(kvm_wfi_trap_policy != KVM_WFX_NOTRAP_SINGLE_TASK))
617 		return kvm_wfi_trap_policy == KVM_WFX_NOTRAP;
618 
619 	if (vgic_is_v5(vcpu->kvm))
620 		return single_task_running();
621 
622 	return single_task_running() &&
623 	       vcpu->kvm->arch.vgic.vgic_model == KVM_DEV_TYPE_ARM_VGIC_V3 &&
624 	       (atomic_read(&vcpu->arch.vgic_cpu.vgic_v3.its_vpe.vlpi_count) ||
625 		vcpu->kvm->arch.vgic.nassgireq);
626 }
627 
628 static bool kvm_vcpu_should_clear_twe(struct kvm_vcpu *vcpu)
629 {
630 	if (unlikely(kvm_wfe_trap_policy != KVM_WFX_NOTRAP_SINGLE_TASK))
631 		return kvm_wfe_trap_policy == KVM_WFX_NOTRAP;
632 
633 	return single_task_running();
634 }
635 
636 void kvm_arch_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
637 {
638 	struct kvm_s2_mmu *mmu;
639 	int *last_ran;
640 
641 	if (is_protected_kvm_enabled())
642 		goto nommu;
643 
644 	if (vcpu_has_nv(vcpu))
645 		kvm_vcpu_load_hw_mmu(vcpu);
646 
647 	mmu = vcpu->arch.hw_mmu;
648 	last_ran = this_cpu_ptr(mmu->last_vcpu_ran);
649 
650 	/*
651 	 * Ensure a VMID is allocated for the MMU before programming VTTBR_EL2,
652 	 * which happens eagerly in VHE.
653 	 *
654 	 * Also, the VMID allocator only preserves VMIDs that are active at the
655 	 * time of rollover, so KVM might need to grab a new VMID for the MMU if
656 	 * this is called from kvm_sched_in().
657 	 */
658 	kvm_arm_vmid_update(&mmu->vmid);
659 
660 	/*
661 	 * We guarantee that both TLBs and I-cache are private to each
662 	 * vcpu. If detecting that a vcpu from the same VM has
663 	 * previously run on the same physical CPU, call into the
664 	 * hypervisor code to nuke the relevant contexts.
665 	 *
666 	 * We might get preempted before the vCPU actually runs, but
667 	 * over-invalidation doesn't affect correctness.
668 	 */
669 	if (*last_ran != vcpu->vcpu_idx) {
670 		kvm_call_hyp(__kvm_flush_cpu_context, mmu);
671 		*last_ran = vcpu->vcpu_idx;
672 	}
673 
674 nommu:
675 	vcpu->cpu = cpu;
676 
677 	/*
678 	 * The timer must be loaded before the vgic to correctly set up physical
679 	 * interrupt deactivation in nested state (e.g. timer interrupt).
680 	 */
681 	kvm_timer_vcpu_load(vcpu);
682 	kvm_vgic_load(vcpu);
683 	kvm_vcpu_load_debug(vcpu);
684 	kvm_vcpu_load_fgt(vcpu);
685 	if (has_vhe())
686 		kvm_vcpu_load_vhe(vcpu);
687 	kvm_arch_vcpu_load_fp(vcpu);
688 	kvm_vcpu_pmu_restore_guest(vcpu);
689 	if (kvm_arm_is_pvtime_enabled(&vcpu->arch))
690 		kvm_make_request(KVM_REQ_RECORD_STEAL, vcpu);
691 
692 	if (kvm_vcpu_should_clear_twe(vcpu))
693 		vcpu->arch.hcr_el2 &= ~HCR_TWE;
694 	else
695 		vcpu->arch.hcr_el2 |= HCR_TWE;
696 
697 	if (kvm_vcpu_should_clear_twi(vcpu))
698 		vcpu->arch.hcr_el2 &= ~HCR_TWI;
699 	else
700 		vcpu->arch.hcr_el2 |= HCR_TWI;
701 
702 	vcpu_set_pauth_traps(vcpu);
703 
704 	if (is_protected_kvm_enabled()) {
705 		kvm_call_hyp_nvhe(__pkvm_vcpu_load,
706 				  vcpu->kvm->arch.pkvm.handle,
707 				  vcpu->vcpu_idx, vcpu->arch.hcr_el2);
708 		kvm_call_hyp(__vgic_v3_restore_vmcr_aprs,
709 			     &vcpu->arch.vgic_cpu.vgic_v3);
710 	}
711 
712 	if (!cpumask_test_cpu(cpu, vcpu->kvm->arch.supported_cpus))
713 		vcpu_set_on_unsupported_cpu(vcpu);
714 }
715 
716 void kvm_arch_vcpu_put(struct kvm_vcpu *vcpu)
717 {
718 	if (is_protected_kvm_enabled()) {
719 		kvm_call_hyp(__vgic_v3_save_aprs, &vcpu->arch.vgic_cpu.vgic_v3);
720 		kvm_call_hyp_nvhe(__pkvm_vcpu_put);
721 	}
722 
723 	kvm_vcpu_put_debug(vcpu);
724 	kvm_arch_vcpu_put_fp(vcpu);
725 	if (has_vhe())
726 		kvm_vcpu_put_vhe(vcpu);
727 	kvm_timer_vcpu_put(vcpu);
728 	kvm_vgic_put(vcpu);
729 	kvm_vcpu_pmu_restore_host(vcpu);
730 	if (vcpu_has_nv(vcpu))
731 		kvm_vcpu_put_hw_mmu(vcpu);
732 	kvm_arm_vmid_clear_active();
733 
734 	vcpu_clear_on_unsupported_cpu(vcpu);
735 	vcpu->cpu = -1;
736 }
737 
738 static void __kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
739 {
740 	WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_STOPPED);
741 	kvm_make_request(KVM_REQ_SLEEP, vcpu);
742 	kvm_vcpu_kick(vcpu);
743 }
744 
745 void kvm_arm_vcpu_power_off(struct kvm_vcpu *vcpu)
746 {
747 	spin_lock(&vcpu->arch.mp_state_lock);
748 	__kvm_arm_vcpu_power_off(vcpu);
749 	spin_unlock(&vcpu->arch.mp_state_lock);
750 }
751 
752 bool kvm_arm_vcpu_stopped(struct kvm_vcpu *vcpu)
753 {
754 	return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_STOPPED;
755 }
756 
757 static void kvm_arm_vcpu_suspend(struct kvm_vcpu *vcpu)
758 {
759 	WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_SUSPENDED);
760 	kvm_make_request(KVM_REQ_SUSPEND, vcpu);
761 	kvm_vcpu_kick(vcpu);
762 }
763 
764 static bool kvm_arm_vcpu_suspended(struct kvm_vcpu *vcpu)
765 {
766 	return READ_ONCE(vcpu->arch.mp_state.mp_state) == KVM_MP_STATE_SUSPENDED;
767 }
768 
769 int kvm_arch_vcpu_ioctl_get_mpstate(struct kvm_vcpu *vcpu,
770 				    struct kvm_mp_state *mp_state)
771 {
772 	*mp_state = READ_ONCE(vcpu->arch.mp_state);
773 
774 	return 0;
775 }
776 
777 int kvm_arch_vcpu_ioctl_set_mpstate(struct kvm_vcpu *vcpu,
778 				    struct kvm_mp_state *mp_state)
779 {
780 	int ret = 0;
781 
782 	spin_lock(&vcpu->arch.mp_state_lock);
783 
784 	switch (mp_state->mp_state) {
785 	case KVM_MP_STATE_RUNNABLE:
786 		WRITE_ONCE(vcpu->arch.mp_state, *mp_state);
787 		break;
788 	case KVM_MP_STATE_STOPPED:
789 		__kvm_arm_vcpu_power_off(vcpu);
790 		break;
791 	case KVM_MP_STATE_SUSPENDED:
792 		kvm_arm_vcpu_suspend(vcpu);
793 		break;
794 	default:
795 		ret = -EINVAL;
796 	}
797 
798 	spin_unlock(&vcpu->arch.mp_state_lock);
799 
800 	return ret;
801 }
802 
803 /**
804  * kvm_arch_vcpu_runnable - determine if the vcpu can be scheduled
805  * @v:		The VCPU pointer
806  *
807  * If the guest CPU is not waiting for interrupts or an interrupt line is
808  * asserted, the CPU is by definition runnable.
809  */
810 int kvm_arch_vcpu_runnable(struct kvm_vcpu *v)
811 {
812 	bool irq_lines = *vcpu_hcr(v) & (HCR_VI | HCR_VF | HCR_VSE);
813 
814 	return ((irq_lines || kvm_vgic_vcpu_pending_irq(v))
815 		&& !kvm_arm_vcpu_stopped(v) && !v->arch.pause);
816 }
817 
818 bool kvm_arch_vcpu_in_kernel(struct kvm_vcpu *vcpu)
819 {
820 	return vcpu_mode_priv(vcpu);
821 }
822 
823 #ifdef CONFIG_GUEST_PERF_EVENTS
824 unsigned long kvm_arch_vcpu_get_ip(struct kvm_vcpu *vcpu)
825 {
826 	return *vcpu_pc(vcpu);
827 }
828 #endif
829 
830 static void kvm_init_mpidr_data(struct kvm *kvm)
831 {
832 	struct kvm_mpidr_data *data = NULL;
833 	unsigned long c, mask, nr_entries;
834 	u64 aff_set = 0, aff_clr = ~0UL;
835 	struct kvm_vcpu *vcpu;
836 
837 	mutex_lock(&kvm->arch.config_lock);
838 
839 	if (rcu_access_pointer(kvm->arch.mpidr_data) ||
840 	    atomic_read(&kvm->online_vcpus) == 1)
841 		goto out;
842 
843 	kvm_for_each_vcpu(c, vcpu, kvm) {
844 		u64 aff = kvm_vcpu_get_mpidr_aff(vcpu);
845 		aff_set |= aff;
846 		aff_clr &= aff;
847 	}
848 
849 	/*
850 	 * A significant bit can be either 0 or 1, and will only appear in
851 	 * aff_set. Use aff_clr to weed out the useless stuff.
852 	 */
853 	mask = aff_set ^ aff_clr;
854 	nr_entries = BIT_ULL(hweight_long(mask));
855 
856 	/*
857 	 * Don't let userspace fool us. If we need more than a single page
858 	 * to describe the compressed MPIDR array, just fall back to the
859 	 * iterative method. Single vcpu VMs do not need this either.
860 	 */
861 	if (struct_size(data, cmpidr_to_idx, nr_entries) <= PAGE_SIZE)
862 		data = kzalloc_flex(*data, cmpidr_to_idx, nr_entries,
863 				    GFP_KERNEL_ACCOUNT);
864 
865 	if (!data)
866 		goto out;
867 
868 	data->mpidr_mask = mask;
869 
870 	kvm_for_each_vcpu(c, vcpu, kvm) {
871 		u64 aff = kvm_vcpu_get_mpidr_aff(vcpu);
872 		u16 index = kvm_mpidr_index(data, aff);
873 
874 		data->cmpidr_to_idx[index] = c;
875 	}
876 
877 	rcu_assign_pointer(kvm->arch.mpidr_data, data);
878 out:
879 	mutex_unlock(&kvm->arch.config_lock);
880 }
881 
882 /*
883  * Handle both the initialisation that is being done when the vcpu is
884  * run for the first time, as well as the updates that must be
885  * performed each time we get a new thread dealing with this vcpu.
886  */
887 int kvm_arch_vcpu_run_pid_change(struct kvm_vcpu *vcpu)
888 {
889 	struct kvm *kvm = vcpu->kvm;
890 	int ret;
891 
892 	if (!kvm_vcpu_initialized(vcpu))
893 		return -ENOEXEC;
894 
895 	if (!kvm_arm_vcpu_is_finalized(vcpu))
896 		return -EPERM;
897 
898 	if (likely(vcpu_has_run_once(vcpu)))
899 		return 0;
900 
901 	kvm_init_mpidr_data(kvm);
902 
903 	if (likely(irqchip_in_kernel(kvm))) {
904 		/*
905 		 * Map the VGIC hardware resources before running a vcpu the
906 		 * first time on this VM.
907 		 */
908 		ret = kvm_vgic_map_resources(kvm);
909 		if (ret)
910 			return ret;
911 	}
912 
913 	ret = kvm_finalize_sys_regs(vcpu);
914 	if (ret)
915 		return ret;
916 
917 	if (vcpu_has_nv(vcpu)) {
918 		ret = kvm_vcpu_allocate_vncr_tlb(vcpu);
919 		if (ret)
920 			return ret;
921 
922 		ret = kvm_vgic_vcpu_nv_init(vcpu);
923 		if (ret)
924 			return ret;
925 	}
926 
927 	/*
928 	 * This needs to happen after any restriction has been applied
929 	 * to the feature set.
930 	 */
931 	kvm_calculate_traps(vcpu);
932 
933 	ret = kvm_timer_enable(vcpu);
934 	if (ret)
935 		return ret;
936 
937 	if (kvm_vcpu_has_pmu(vcpu)) {
938 		ret = kvm_arm_pmu_v3_enable(vcpu);
939 		if (ret)
940 			return ret;
941 	}
942 
943 	ret = vgic_v5_finalize_ppi_state(kvm);
944 	if (ret)
945 		return ret;
946 
947 	if (is_protected_kvm_enabled()) {
948 		ret = pkvm_create_hyp_vm(kvm);
949 		if (ret)
950 			return ret;
951 
952 		ret = pkvm_create_hyp_vcpu(vcpu);
953 		if (ret)
954 			return ret;
955 	}
956 
957 	mutex_lock(&kvm->arch.config_lock);
958 	set_bit(KVM_ARCH_FLAG_HAS_RAN_ONCE, &kvm->arch.flags);
959 	mutex_unlock(&kvm->arch.config_lock);
960 
961 	return ret;
962 }
963 
964 bool kvm_arch_intc_initialized(struct kvm *kvm)
965 {
966 	return vgic_initialized(kvm);
967 }
968 
969 void kvm_arm_halt_guest(struct kvm *kvm)
970 {
971 	unsigned long i;
972 	struct kvm_vcpu *vcpu;
973 
974 	kvm_for_each_vcpu(i, vcpu, kvm)
975 		vcpu->arch.pause = true;
976 	kvm_make_all_cpus_request(kvm, KVM_REQ_SLEEP);
977 }
978 
979 void kvm_arm_resume_guest(struct kvm *kvm)
980 {
981 	unsigned long i;
982 	struct kvm_vcpu *vcpu;
983 
984 	kvm_for_each_vcpu(i, vcpu, kvm) {
985 		vcpu->arch.pause = false;
986 		__kvm_vcpu_wake_up(vcpu);
987 	}
988 }
989 
990 static void kvm_vcpu_sleep(struct kvm_vcpu *vcpu)
991 {
992 	struct rcuwait *wait = kvm_arch_vcpu_get_wait(vcpu);
993 
994 	rcuwait_wait_event(wait,
995 			   (!kvm_arm_vcpu_stopped(vcpu)) && (!vcpu->arch.pause),
996 			   TASK_INTERRUPTIBLE);
997 
998 	if (kvm_arm_vcpu_stopped(vcpu) || vcpu->arch.pause) {
999 		/* Awaken to handle a signal, request we sleep again later. */
1000 		kvm_make_request(KVM_REQ_SLEEP, vcpu);
1001 	}
1002 
1003 	/*
1004 	 * Make sure we will observe a potential reset request if we've
1005 	 * observed a change to the power state. Pairs with the smp_wmb() in
1006 	 * kvm_psci_vcpu_on().
1007 	 */
1008 	smp_rmb();
1009 }
1010 
1011 /**
1012  * kvm_vcpu_wfi - emulate Wait-For-Interrupt behavior
1013  * @vcpu:	The VCPU pointer
1014  *
1015  * Suspend execution of a vCPU until a valid wake event is detected, i.e. until
1016  * the vCPU is runnable.  The vCPU may or may not be scheduled out, depending
1017  * on when a wake event arrives, e.g. there may already be a pending wake event.
1018  */
1019 void kvm_vcpu_wfi(struct kvm_vcpu *vcpu)
1020 {
1021 	/*
1022 	 * Sync back the state of the GIC CPU interface so that we have
1023 	 * the latest PMR and group enables. This ensures that
1024 	 * kvm_arch_vcpu_runnable has up-to-date data to decide whether
1025 	 * we have pending interrupts, e.g. when determining if the
1026 	 * vCPU should block.
1027 	 *
1028 	 * For the same reason, we want to tell GICv4 that we need
1029 	 * doorbells to be signalled, should an interrupt become pending.
1030 	 */
1031 	preempt_disable();
1032 	vcpu_set_flag(vcpu, IN_WFI);
1033 	kvm_vgic_put(vcpu);
1034 	preempt_enable();
1035 
1036 	kvm_vcpu_halt(vcpu);
1037 	vcpu_clear_flag(vcpu, IN_WFIT);
1038 
1039 	preempt_disable();
1040 	vcpu_clear_flag(vcpu, IN_WFI);
1041 	kvm_vgic_load(vcpu);
1042 	preempt_enable();
1043 }
1044 
1045 static int kvm_vcpu_suspend(struct kvm_vcpu *vcpu)
1046 {
1047 	if (!kvm_arm_vcpu_suspended(vcpu))
1048 		return 1;
1049 
1050 	kvm_vcpu_wfi(vcpu);
1051 
1052 	/*
1053 	 * The suspend state is sticky; we do not leave it until userspace
1054 	 * explicitly marks the vCPU as runnable. Request that we suspend again
1055 	 * later.
1056 	 */
1057 	kvm_make_request(KVM_REQ_SUSPEND, vcpu);
1058 
1059 	/*
1060 	 * Check to make sure the vCPU is actually runnable. If so, exit to
1061 	 * userspace informing it of the wakeup condition.
1062 	 */
1063 	if (kvm_arch_vcpu_runnable(vcpu)) {
1064 		memset(&vcpu->run->system_event, 0, sizeof(vcpu->run->system_event));
1065 		vcpu->run->system_event.type = KVM_SYSTEM_EVENT_WAKEUP;
1066 		vcpu->run->exit_reason = KVM_EXIT_SYSTEM_EVENT;
1067 		return 0;
1068 	}
1069 
1070 	/*
1071 	 * Otherwise, we were unblocked to process a different event, such as a
1072 	 * pending signal. Return 1 and allow kvm_arch_vcpu_ioctl_run() to
1073 	 * process the event.
1074 	 */
1075 	return 1;
1076 }
1077 
1078 /**
1079  * check_vcpu_requests - check and handle pending vCPU requests
1080  * @vcpu:	the VCPU pointer
1081  *
1082  * Return: 1 if we should enter the guest
1083  *	   0 if we should exit to userspace
1084  *	   < 0 if we should exit to userspace, where the return value indicates
1085  *	   an error
1086  */
1087 static int check_vcpu_requests(struct kvm_vcpu *vcpu)
1088 {
1089 	if (kvm_request_pending(vcpu)) {
1090 		if (kvm_check_request(KVM_REQ_VM_DEAD, vcpu))
1091 			return -EIO;
1092 
1093 		if (kvm_check_request(KVM_REQ_SLEEP, vcpu))
1094 			kvm_vcpu_sleep(vcpu);
1095 
1096 		if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1097 			kvm_reset_vcpu(vcpu);
1098 
1099 		/*
1100 		 * Clear IRQ_PENDING requests that were made to guarantee
1101 		 * that a VCPU sees new virtual interrupts.
1102 		 */
1103 		kvm_check_request(KVM_REQ_IRQ_PENDING, vcpu);
1104 
1105 		/* Process interrupts deactivated through a trap */
1106 		if (kvm_check_request(KVM_REQ_VGIC_PROCESS_UPDATE, vcpu))
1107 			kvm_vgic_process_async_update(vcpu);
1108 
1109 		if (kvm_check_request(KVM_REQ_RECORD_STEAL, vcpu))
1110 			kvm_update_stolen_time(vcpu);
1111 
1112 		if (kvm_check_request(KVM_REQ_RELOAD_GICv4, vcpu)) {
1113 			/* The distributor enable bits were changed */
1114 			preempt_disable();
1115 			vgic_v4_put(vcpu);
1116 			vgic_v4_load(vcpu);
1117 			preempt_enable();
1118 		}
1119 
1120 		if (kvm_check_request(KVM_REQ_RELOAD_PMU, vcpu))
1121 			kvm_vcpu_reload_pmu(vcpu);
1122 
1123 		if (kvm_check_request(KVM_REQ_RESYNC_PMU_EL0, vcpu))
1124 			kvm_vcpu_pmu_restore_guest(vcpu);
1125 
1126 		if (kvm_check_request(KVM_REQ_SUSPEND, vcpu))
1127 			return kvm_vcpu_suspend(vcpu);
1128 
1129 		if (kvm_dirty_ring_check_request(vcpu))
1130 			return 0;
1131 
1132 		check_nested_vcpu_requests(vcpu);
1133 	}
1134 
1135 	return 1;
1136 }
1137 
1138 static bool vcpu_mode_is_bad_32bit(struct kvm_vcpu *vcpu)
1139 {
1140 	if (likely(!vcpu_mode_is_32bit(vcpu)))
1141 		return false;
1142 
1143 	if (vcpu_has_nv(vcpu))
1144 		return true;
1145 
1146 	return !kvm_supports_32bit_el0();
1147 }
1148 
1149 /**
1150  * kvm_vcpu_exit_request - returns true if the VCPU should *not* enter the guest
1151  * @vcpu:	The VCPU pointer
1152  * @ret:	Pointer to write optional return code
1153  *
1154  * Returns: true if the VCPU needs to return to a preemptible + interruptible
1155  *	    and skip guest entry.
1156  *
1157  * This function disambiguates between two different types of exits: exits to a
1158  * preemptible + interruptible kernel context and exits to userspace. For an
1159  * exit to userspace, this function will write the return code to ret and return
1160  * true. For an exit to preemptible + interruptible kernel context (i.e. check
1161  * for pending work and re-enter), return true without writing to ret.
1162  */
1163 static bool kvm_vcpu_exit_request(struct kvm_vcpu *vcpu, int *ret)
1164 {
1165 	struct kvm_run *run = vcpu->run;
1166 
1167 	/*
1168 	 * If we're using a userspace irqchip, then check if we need
1169 	 * to tell a userspace irqchip about timer or PMU level
1170 	 * changes and if so, exit to userspace (the actual level
1171 	 * state gets updated in kvm_timer_update_run and
1172 	 * kvm_pmu_update_run below).
1173 	 */
1174 	if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1175 		if (kvm_timer_should_notify_user(vcpu) ||
1176 		    kvm_pmu_should_notify_user(vcpu)) {
1177 			*ret = -EINTR;
1178 			run->exit_reason = KVM_EXIT_INTR;
1179 			return true;
1180 		}
1181 	}
1182 
1183 	if (unlikely(vcpu_on_unsupported_cpu(vcpu))) {
1184 		run->exit_reason = KVM_EXIT_FAIL_ENTRY;
1185 		run->fail_entry.hardware_entry_failure_reason = KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED;
1186 		run->fail_entry.cpu = smp_processor_id();
1187 		*ret = 0;
1188 		return true;
1189 	}
1190 
1191 	return kvm_request_pending(vcpu) ||
1192 			xfer_to_guest_mode_work_pending();
1193 }
1194 
1195 /*
1196  * Actually run the vCPU, entering an RCU extended quiescent state (EQS) while
1197  * the vCPU is running.
1198  *
1199  * This must be noinstr as instrumentation may make use of RCU, and this is not
1200  * safe during the EQS.
1201  */
1202 static int noinstr kvm_arm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
1203 {
1204 	int ret;
1205 
1206 	guest_state_enter_irqoff();
1207 	ret = kvm_call_hyp_ret(__kvm_vcpu_run, vcpu);
1208 	guest_state_exit_irqoff();
1209 
1210 	return ret;
1211 }
1212 
1213 /**
1214  * kvm_arch_vcpu_ioctl_run - the main VCPU run function to execute guest code
1215  * @vcpu:	The VCPU pointer
1216  *
1217  * This function is called through the VCPU_RUN ioctl called from user space. It
1218  * will execute VM code in a loop until the time slice for the process is used
1219  * or some emulation is needed from user space in which case the function will
1220  * return with return value 0 and with the kvm_run structure filled in with the
1221  * required data for the requested emulation.
1222  */
1223 int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
1224 {
1225 	struct kvm_run *run = vcpu->run;
1226 	int ret;
1227 
1228 	if (run->exit_reason == KVM_EXIT_MMIO) {
1229 		ret = kvm_handle_mmio_return(vcpu);
1230 		if (ret <= 0)
1231 			return ret;
1232 	}
1233 
1234 	vcpu_load(vcpu);
1235 
1236 	if (!vcpu->wants_to_run) {
1237 		ret = -EINTR;
1238 		goto out;
1239 	}
1240 
1241 	kvm_sigset_activate(vcpu);
1242 
1243 	ret = 1;
1244 	run->exit_reason = KVM_EXIT_UNKNOWN;
1245 	run->flags = 0;
1246 	while (ret > 0) {
1247 		/*
1248 		 * Check conditions before entering the guest
1249 		 */
1250 		ret = kvm_xfer_to_guest_mode_handle_work(vcpu);
1251 		if (!ret)
1252 			ret = 1;
1253 
1254 		if (ret > 0)
1255 			ret = check_vcpu_requests(vcpu);
1256 
1257 		/*
1258 		 * Preparing the interrupts to be injected also
1259 		 * involves poking the GIC, which must be done in a
1260 		 * non-preemptible context.
1261 		 */
1262 		preempt_disable();
1263 
1264 		kvm_nested_flush_hwstate(vcpu);
1265 
1266 		if (kvm_vcpu_has_pmu(vcpu))
1267 			kvm_pmu_flush_hwstate(vcpu);
1268 
1269 		local_irq_disable();
1270 
1271 		kvm_vgic_flush_hwstate(vcpu);
1272 
1273 		kvm_pmu_update_vcpu_events(vcpu);
1274 
1275 		/*
1276 		 * Ensure we set mode to IN_GUEST_MODE after we disable
1277 		 * interrupts and before the final VCPU requests check.
1278 		 * See the comment in kvm_vcpu_exiting_guest_mode() and
1279 		 * Documentation/virt/kvm/vcpu-requests.rst
1280 		 */
1281 		smp_store_mb(vcpu->mode, IN_GUEST_MODE);
1282 
1283 		if (ret <= 0 || kvm_vcpu_exit_request(vcpu, &ret)) {
1284 			vcpu->mode = OUTSIDE_GUEST_MODE;
1285 			isb(); /* Ensure work in x_flush_hwstate is committed */
1286 			if (kvm_vcpu_has_pmu(vcpu))
1287 				kvm_pmu_sync_hwstate(vcpu);
1288 			if (unlikely(!irqchip_in_kernel(vcpu->kvm)))
1289 				kvm_timer_sync_user(vcpu);
1290 			kvm_vgic_sync_hwstate(vcpu);
1291 			local_irq_enable();
1292 			preempt_enable();
1293 			continue;
1294 		}
1295 
1296 		kvm_arch_vcpu_ctxflush_fp(vcpu);
1297 
1298 		/**************************************************************
1299 		 * Enter the guest
1300 		 */
1301 		trace_kvm_entry(*vcpu_pc(vcpu));
1302 		guest_timing_enter_irqoff();
1303 
1304 		ret = kvm_arm_vcpu_enter_exit(vcpu);
1305 
1306 		vcpu->mode = OUTSIDE_GUEST_MODE;
1307 		vcpu->stat.exits++;
1308 		/*
1309 		 * Back from guest
1310 		 *************************************************************/
1311 
1312 		/*
1313 		 * We must sync the PMU state before the vgic state so
1314 		 * that the vgic can properly sample the updated state of the
1315 		 * interrupt line.
1316 		 */
1317 		if (kvm_vcpu_has_pmu(vcpu))
1318 			kvm_pmu_sync_hwstate(vcpu);
1319 
1320 		/*
1321 		 * Sync the vgic state before syncing the timer state because
1322 		 * the timer code needs to know if the virtual timer
1323 		 * interrupts are active.
1324 		 */
1325 		kvm_vgic_sync_hwstate(vcpu);
1326 
1327 		/*
1328 		 * Sync the timer hardware state before enabling interrupts as
1329 		 * we don't want vtimer interrupts to race with syncing the
1330 		 * timer virtual interrupt state.
1331 		 */
1332 		if (unlikely(!irqchip_in_kernel(vcpu->kvm)))
1333 			kvm_timer_sync_user(vcpu);
1334 
1335 		if (is_hyp_ctxt(vcpu))
1336 			kvm_timer_sync_nested(vcpu);
1337 
1338 		kvm_arch_vcpu_ctxsync_fp(vcpu);
1339 
1340 		/*
1341 		 * We must ensure that any pending interrupts are taken before
1342 		 * we exit guest timing so that timer ticks are accounted as
1343 		 * guest time. Transiently unmask interrupts so that any
1344 		 * pending interrupts are taken.
1345 		 *
1346 		 * Per ARM DDI 0487G.b section D1.13.4, an ISB (or other
1347 		 * context synchronization event) is necessary to ensure that
1348 		 * pending interrupts are taken.
1349 		 */
1350 		if (ARM_EXCEPTION_CODE(ret) == ARM_EXCEPTION_IRQ) {
1351 			local_irq_enable();
1352 			isb();
1353 			local_irq_disable();
1354 		}
1355 
1356 		guest_timing_exit_irqoff();
1357 
1358 		local_irq_enable();
1359 
1360 		trace_kvm_exit(ret, kvm_vcpu_trap_get_class(vcpu), *vcpu_pc(vcpu));
1361 
1362 		/* Exit types that need handling before we can be preempted */
1363 		handle_exit_early(vcpu, ret);
1364 
1365 		kvm_nested_sync_hwstate(vcpu);
1366 
1367 		preempt_enable();
1368 
1369 		/*
1370 		 * The ARMv8 architecture doesn't give the hypervisor
1371 		 * a mechanism to prevent a guest from dropping to AArch32 EL0
1372 		 * if implemented by the CPU. If we spot the guest in such
1373 		 * state and that we decided it wasn't supposed to do so (like
1374 		 * with the asymmetric AArch32 case), return to userspace with
1375 		 * a fatal error.
1376 		 */
1377 		if (vcpu_mode_is_bad_32bit(vcpu)) {
1378 			/*
1379 			 * As we have caught the guest red-handed, decide that
1380 			 * it isn't fit for purpose anymore by making the vcpu
1381 			 * invalid. The VMM can try and fix it by issuing  a
1382 			 * KVM_ARM_VCPU_INIT if it really wants to.
1383 			 */
1384 			vcpu_clear_flag(vcpu, VCPU_INITIALIZED);
1385 			ret = ARM_EXCEPTION_IL;
1386 		}
1387 
1388 		ret = handle_exit(vcpu, ret);
1389 	}
1390 
1391 	/* Tell userspace about in-kernel device output levels */
1392 	if (unlikely(!irqchip_in_kernel(vcpu->kvm))) {
1393 		kvm_timer_update_run(vcpu);
1394 		kvm_pmu_update_run(vcpu);
1395 	}
1396 
1397 	kvm_sigset_deactivate(vcpu);
1398 
1399 out:
1400 	/*
1401 	 * In the unlikely event that we are returning to userspace
1402 	 * with pending exceptions or PC adjustment, commit these
1403 	 * adjustments in order to give userspace a consistent view of
1404 	 * the vcpu state. Note that this relies on __kvm_adjust_pc()
1405 	 * being preempt-safe on VHE.
1406 	 */
1407 	if (unlikely(vcpu_get_flag(vcpu, PENDING_EXCEPTION) ||
1408 		     vcpu_get_flag(vcpu, INCREMENT_PC)))
1409 		kvm_call_hyp(__kvm_adjust_pc, vcpu);
1410 
1411 	vcpu_put(vcpu);
1412 	return ret;
1413 }
1414 
1415 static int vcpu_interrupt_line(struct kvm_vcpu *vcpu, int number, bool level)
1416 {
1417 	int bit_index;
1418 	bool set;
1419 	unsigned long *hcr;
1420 
1421 	if (number == KVM_ARM_IRQ_CPU_IRQ)
1422 		bit_index = __ffs(HCR_VI);
1423 	else /* KVM_ARM_IRQ_CPU_FIQ */
1424 		bit_index = __ffs(HCR_VF);
1425 
1426 	hcr = vcpu_hcr(vcpu);
1427 	if (level)
1428 		set = test_and_set_bit(bit_index, hcr);
1429 	else
1430 		set = test_and_clear_bit(bit_index, hcr);
1431 
1432 	/*
1433 	 * If we didn't change anything, no need to wake up or kick other CPUs
1434 	 */
1435 	if (set == level)
1436 		return 0;
1437 
1438 	/*
1439 	 * The vcpu irq_lines field was updated, wake up sleeping VCPUs and
1440 	 * trigger a world-switch round on the running physical CPU to set the
1441 	 * virtual IRQ/FIQ fields in the HCR appropriately.
1442 	 */
1443 	kvm_make_request(KVM_REQ_IRQ_PENDING, vcpu);
1444 	kvm_vcpu_kick(vcpu);
1445 
1446 	return 0;
1447 }
1448 
1449 int kvm_vm_ioctl_irq_line(struct kvm *kvm, struct kvm_irq_level *irq_level,
1450 			  bool line_status)
1451 {
1452 	u32 irq = irq_level->irq;
1453 	unsigned int irq_type, vcpu_id, irq_num;
1454 	struct kvm_vcpu *vcpu = NULL;
1455 	bool level = irq_level->level;
1456 
1457 	irq_type = (irq >> KVM_ARM_IRQ_TYPE_SHIFT) & KVM_ARM_IRQ_TYPE_MASK;
1458 	vcpu_id = (irq >> KVM_ARM_IRQ_VCPU_SHIFT) & KVM_ARM_IRQ_VCPU_MASK;
1459 	vcpu_id += ((irq >> KVM_ARM_IRQ_VCPU2_SHIFT) & KVM_ARM_IRQ_VCPU2_MASK) * (KVM_ARM_IRQ_VCPU_MASK + 1);
1460 	irq_num = (irq >> KVM_ARM_IRQ_NUM_SHIFT) & KVM_ARM_IRQ_NUM_MASK;
1461 
1462 	trace_kvm_irq_line(irq_type, vcpu_id, irq_num, irq_level->level);
1463 
1464 	switch (irq_type) {
1465 	case KVM_ARM_IRQ_TYPE_CPU:
1466 		if (irqchip_in_kernel(kvm))
1467 			return -ENXIO;
1468 
1469 		vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
1470 		if (!vcpu)
1471 			return -EINVAL;
1472 
1473 		if (irq_num > KVM_ARM_IRQ_CPU_FIQ)
1474 			return -EINVAL;
1475 
1476 		return vcpu_interrupt_line(vcpu, irq_num, level);
1477 	case KVM_ARM_IRQ_TYPE_PPI:
1478 		if (!irqchip_in_kernel(kvm))
1479 			return -ENXIO;
1480 
1481 		vcpu = kvm_get_vcpu_by_id(kvm, vcpu_id);
1482 		if (!vcpu)
1483 			return -EINVAL;
1484 
1485 		if (vgic_is_v5(kvm)) {
1486 			if (irq_num >= VGIC_V5_NR_PRIVATE_IRQS)
1487 				return -EINVAL;
1488 
1489 			/* Build a GICv5-style IntID here */
1490 			irq_num = vgic_v5_make_ppi(irq_num);
1491 		} else if (irq_num < VGIC_NR_SGIS ||
1492 			   irq_num >= VGIC_NR_PRIVATE_IRQS) {
1493 			return -EINVAL;
1494 		}
1495 
1496 		return kvm_vgic_inject_irq(kvm, vcpu, irq_num, level, NULL);
1497 	case KVM_ARM_IRQ_TYPE_SPI:
1498 		if (!irqchip_in_kernel(kvm))
1499 			return -ENXIO;
1500 
1501 		if (vgic_is_v5(kvm)) {
1502 			/* Build a GICv5-style IntID here */
1503 			irq_num = vgic_v5_make_spi(irq_num);
1504 		} else {
1505 			if (irq_num < VGIC_NR_PRIVATE_IRQS)
1506 				return -EINVAL;
1507 		}
1508 
1509 		return kvm_vgic_inject_irq(kvm, NULL, irq_num, level, NULL);
1510 	}
1511 
1512 	return -EINVAL;
1513 }
1514 
1515 static unsigned long system_supported_vcpu_features(void)
1516 {
1517 	unsigned long features = KVM_VCPU_VALID_FEATURES;
1518 
1519 	if (!cpus_have_final_cap(ARM64_HAS_32BIT_EL1))
1520 		clear_bit(KVM_ARM_VCPU_EL1_32BIT, &features);
1521 
1522 	if (!kvm_supports_guest_pmuv3())
1523 		clear_bit(KVM_ARM_VCPU_PMU_V3, &features);
1524 
1525 	if (!system_supports_sve())
1526 		clear_bit(KVM_ARM_VCPU_SVE, &features);
1527 
1528 	if (!kvm_has_full_ptr_auth()) {
1529 		clear_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features);
1530 		clear_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features);
1531 	}
1532 
1533 	if (!cpus_have_final_cap(ARM64_HAS_NESTED_VIRT))
1534 		clear_bit(KVM_ARM_VCPU_HAS_EL2, &features);
1535 
1536 	return features;
1537 }
1538 
1539 static int kvm_vcpu_init_check_features(struct kvm_vcpu *vcpu,
1540 					const struct kvm_vcpu_init *init)
1541 {
1542 	unsigned long features = init->features[0];
1543 	int i;
1544 
1545 	if (features & ~KVM_VCPU_VALID_FEATURES)
1546 		return -ENOENT;
1547 
1548 	for (i = 1; i < ARRAY_SIZE(init->features); i++) {
1549 		if (init->features[i])
1550 			return -ENOENT;
1551 	}
1552 
1553 	if (features & ~system_supported_vcpu_features())
1554 		return -EINVAL;
1555 
1556 	/*
1557 	 * For now make sure that both address/generic pointer authentication
1558 	 * features are requested by the userspace together.
1559 	 */
1560 	if (test_bit(KVM_ARM_VCPU_PTRAUTH_ADDRESS, &features) !=
1561 	    test_bit(KVM_ARM_VCPU_PTRAUTH_GENERIC, &features))
1562 		return -EINVAL;
1563 
1564 	if (!test_bit(KVM_ARM_VCPU_EL1_32BIT, &features))
1565 		return 0;
1566 
1567 	/* MTE is incompatible with AArch32 */
1568 	if (kvm_has_mte(vcpu->kvm))
1569 		return -EINVAL;
1570 
1571 	/* NV is incompatible with AArch32 */
1572 	if (test_bit(KVM_ARM_VCPU_HAS_EL2, &features))
1573 		return -EINVAL;
1574 
1575 	return 0;
1576 }
1577 
1578 static bool kvm_vcpu_init_changed(struct kvm_vcpu *vcpu,
1579 				  const struct kvm_vcpu_init *init)
1580 {
1581 	unsigned long features = init->features[0];
1582 
1583 	return !bitmap_equal(vcpu->kvm->arch.vcpu_features, &features,
1584 			     KVM_VCPU_MAX_FEATURES);
1585 }
1586 
1587 static int kvm_setup_vcpu(struct kvm_vcpu *vcpu)
1588 {
1589 	struct kvm *kvm = vcpu->kvm;
1590 	int ret = 0;
1591 
1592 	/*
1593 	 * When the vCPU has a PMU, but no PMU is set for the guest
1594 	 * yet, set the default one.
1595 	 */
1596 	if (kvm_vcpu_has_pmu(vcpu) && !kvm->arch.arm_pmu)
1597 		ret = kvm_arm_set_default_pmu(kvm);
1598 
1599 	/* Prepare for nested if required */
1600 	if (!ret && vcpu_has_nv(vcpu))
1601 		ret = kvm_vcpu_init_nested(vcpu);
1602 
1603 	return ret;
1604 }
1605 
1606 static int __kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1607 				 const struct kvm_vcpu_init *init)
1608 {
1609 	unsigned long features = init->features[0];
1610 	struct kvm *kvm = vcpu->kvm;
1611 	int ret = -EINVAL;
1612 
1613 	mutex_lock(&kvm->arch.config_lock);
1614 
1615 	if (test_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags) &&
1616 	    kvm_vcpu_init_changed(vcpu, init))
1617 		goto out_unlock;
1618 
1619 	bitmap_copy(kvm->arch.vcpu_features, &features, KVM_VCPU_MAX_FEATURES);
1620 
1621 	ret = kvm_setup_vcpu(vcpu);
1622 	if (ret)
1623 		goto out_unlock;
1624 
1625 	/* Now we know what it is, we can reset it. */
1626 	kvm_reset_vcpu(vcpu);
1627 
1628 	set_bit(KVM_ARCH_FLAG_VCPU_FEATURES_CONFIGURED, &kvm->arch.flags);
1629 	vcpu_set_flag(vcpu, VCPU_INITIALIZED);
1630 	ret = 0;
1631 out_unlock:
1632 	mutex_unlock(&kvm->arch.config_lock);
1633 	return ret;
1634 }
1635 
1636 static int kvm_vcpu_set_target(struct kvm_vcpu *vcpu,
1637 			       const struct kvm_vcpu_init *init)
1638 {
1639 	int ret;
1640 
1641 	if (init->target != KVM_ARM_TARGET_GENERIC_V8 &&
1642 	    init->target != kvm_target_cpu())
1643 		return -EINVAL;
1644 
1645 	ret = kvm_vcpu_init_check_features(vcpu, init);
1646 	if (ret)
1647 		return ret;
1648 
1649 	if (!kvm_vcpu_initialized(vcpu))
1650 		return __kvm_vcpu_set_target(vcpu, init);
1651 
1652 	if (kvm_vcpu_init_changed(vcpu, init))
1653 		return -EINVAL;
1654 
1655 	kvm_reset_vcpu(vcpu);
1656 	return 0;
1657 }
1658 
1659 static int kvm_arch_vcpu_ioctl_vcpu_init(struct kvm_vcpu *vcpu,
1660 					 struct kvm_vcpu_init *init)
1661 {
1662 	bool power_off = false;
1663 	int ret;
1664 
1665 	/*
1666 	 * Treat the power-off vCPU feature as ephemeral. Clear the bit to avoid
1667 	 * reflecting it in the finalized feature set, thus limiting its scope
1668 	 * to a single KVM_ARM_VCPU_INIT call.
1669 	 */
1670 	if (init->features[0] & BIT(KVM_ARM_VCPU_POWER_OFF)) {
1671 		init->features[0] &= ~BIT(KVM_ARM_VCPU_POWER_OFF);
1672 		power_off = true;
1673 	}
1674 
1675 	ret = kvm_vcpu_set_target(vcpu, init);
1676 	if (ret)
1677 		return ret;
1678 
1679 	/*
1680 	 * Ensure a rebooted VM will fault in RAM pages and detect if the
1681 	 * guest MMU is turned off and flush the caches as needed.
1682 	 *
1683 	 * S2FWB enforces all memory accesses to RAM being cacheable,
1684 	 * ensuring that the data side is always coherent. We still
1685 	 * need to invalidate the I-cache though, as FWB does *not*
1686 	 * imply CTR_EL0.DIC.
1687 	 */
1688 	if (vcpu_has_run_once(vcpu)) {
1689 		if (!cpus_have_final_cap(ARM64_HAS_STAGE2_FWB))
1690 			stage2_unmap_vm(vcpu->kvm);
1691 		else
1692 			icache_inval_all_pou();
1693 	}
1694 
1695 	vcpu_reset_hcr(vcpu);
1696 
1697 	/*
1698 	 * Handle the "start in power-off" case.
1699 	 */
1700 	spin_lock(&vcpu->arch.mp_state_lock);
1701 
1702 	if (power_off)
1703 		__kvm_arm_vcpu_power_off(vcpu);
1704 	else
1705 		WRITE_ONCE(vcpu->arch.mp_state.mp_state, KVM_MP_STATE_RUNNABLE);
1706 
1707 	spin_unlock(&vcpu->arch.mp_state_lock);
1708 
1709 	return 0;
1710 }
1711 
1712 static int kvm_arm_vcpu_set_attr(struct kvm_vcpu *vcpu,
1713 				 struct kvm_device_attr *attr)
1714 {
1715 	int ret = -ENXIO;
1716 
1717 	switch (attr->group) {
1718 	default:
1719 		ret = kvm_arm_vcpu_arch_set_attr(vcpu, attr);
1720 		break;
1721 	}
1722 
1723 	return ret;
1724 }
1725 
1726 static int kvm_arm_vcpu_get_attr(struct kvm_vcpu *vcpu,
1727 				 struct kvm_device_attr *attr)
1728 {
1729 	int ret = -ENXIO;
1730 
1731 	switch (attr->group) {
1732 	default:
1733 		ret = kvm_arm_vcpu_arch_get_attr(vcpu, attr);
1734 		break;
1735 	}
1736 
1737 	return ret;
1738 }
1739 
1740 static int kvm_arm_vcpu_has_attr(struct kvm_vcpu *vcpu,
1741 				 struct kvm_device_attr *attr)
1742 {
1743 	int ret = -ENXIO;
1744 
1745 	switch (attr->group) {
1746 	default:
1747 		ret = kvm_arm_vcpu_arch_has_attr(vcpu, attr);
1748 		break;
1749 	}
1750 
1751 	return ret;
1752 }
1753 
1754 static int kvm_arm_vcpu_get_events(struct kvm_vcpu *vcpu,
1755 				   struct kvm_vcpu_events *events)
1756 {
1757 	memset(events, 0, sizeof(*events));
1758 
1759 	return __kvm_arm_vcpu_get_events(vcpu, events);
1760 }
1761 
1762 static int kvm_arm_vcpu_set_events(struct kvm_vcpu *vcpu,
1763 				   struct kvm_vcpu_events *events)
1764 {
1765 	int i;
1766 
1767 	/* check whether the reserved field is zero */
1768 	for (i = 0; i < ARRAY_SIZE(events->reserved); i++)
1769 		if (events->reserved[i])
1770 			return -EINVAL;
1771 
1772 	/* check whether the pad field is zero */
1773 	for (i = 0; i < ARRAY_SIZE(events->exception.pad); i++)
1774 		if (events->exception.pad[i])
1775 			return -EINVAL;
1776 
1777 	return __kvm_arm_vcpu_set_events(vcpu, events);
1778 }
1779 
1780 long kvm_arch_vcpu_ioctl(struct file *filp,
1781 			 unsigned int ioctl, unsigned long arg)
1782 {
1783 	struct kvm_vcpu *vcpu = filp->private_data;
1784 	void __user *argp = (void __user *)arg;
1785 	struct kvm_device_attr attr;
1786 	long r;
1787 
1788 	switch (ioctl) {
1789 	case KVM_ARM_VCPU_INIT: {
1790 		struct kvm_vcpu_init init;
1791 
1792 		r = -EFAULT;
1793 		if (copy_from_user(&init, argp, sizeof(init)))
1794 			break;
1795 
1796 		r = kvm_arch_vcpu_ioctl_vcpu_init(vcpu, &init);
1797 		break;
1798 	}
1799 	case KVM_SET_ONE_REG:
1800 	case KVM_GET_ONE_REG: {
1801 		struct kvm_one_reg reg;
1802 
1803 		r = -ENOEXEC;
1804 		if (unlikely(!kvm_vcpu_initialized(vcpu)))
1805 			break;
1806 
1807 		r = -EFAULT;
1808 		if (copy_from_user(&reg, argp, sizeof(reg)))
1809 			break;
1810 
1811 		/*
1812 		 * We could owe a reset due to PSCI. Handle the pending reset
1813 		 * here to ensure userspace register accesses are ordered after
1814 		 * the reset.
1815 		 */
1816 		if (kvm_check_request(KVM_REQ_VCPU_RESET, vcpu))
1817 			kvm_reset_vcpu(vcpu);
1818 
1819 		if (ioctl == KVM_SET_ONE_REG)
1820 			r = kvm_arm_set_reg(vcpu, &reg);
1821 		else
1822 			r = kvm_arm_get_reg(vcpu, &reg);
1823 		break;
1824 	}
1825 	case KVM_GET_REG_LIST: {
1826 		struct kvm_reg_list __user *user_list = argp;
1827 		struct kvm_reg_list reg_list;
1828 		unsigned n;
1829 
1830 		r = -ENOEXEC;
1831 		if (unlikely(!kvm_vcpu_initialized(vcpu)))
1832 			break;
1833 
1834 		r = -EPERM;
1835 		if (!kvm_arm_vcpu_is_finalized(vcpu))
1836 			break;
1837 
1838 		r = -EFAULT;
1839 		if (copy_from_user(&reg_list, user_list, sizeof(reg_list)))
1840 			break;
1841 		n = reg_list.n;
1842 		reg_list.n = kvm_arm_num_regs(vcpu);
1843 		if (copy_to_user(user_list, &reg_list, sizeof(reg_list)))
1844 			break;
1845 		r = -E2BIG;
1846 		if (n < reg_list.n)
1847 			break;
1848 		r = kvm_arm_copy_reg_indices(vcpu, user_list->reg);
1849 		break;
1850 	}
1851 	case KVM_SET_DEVICE_ATTR: {
1852 		r = -EFAULT;
1853 		if (copy_from_user(&attr, argp, sizeof(attr)))
1854 			break;
1855 		r = kvm_arm_vcpu_set_attr(vcpu, &attr);
1856 		break;
1857 	}
1858 	case KVM_GET_DEVICE_ATTR: {
1859 		r = -EFAULT;
1860 		if (copy_from_user(&attr, argp, sizeof(attr)))
1861 			break;
1862 		r = kvm_arm_vcpu_get_attr(vcpu, &attr);
1863 		break;
1864 	}
1865 	case KVM_HAS_DEVICE_ATTR: {
1866 		r = -EFAULT;
1867 		if (copy_from_user(&attr, argp, sizeof(attr)))
1868 			break;
1869 		r = kvm_arm_vcpu_has_attr(vcpu, &attr);
1870 		break;
1871 	}
1872 	case KVM_GET_VCPU_EVENTS: {
1873 		struct kvm_vcpu_events events;
1874 
1875 		if (!kvm_vcpu_initialized(vcpu))
1876 			return -ENOEXEC;
1877 
1878 		if (kvm_arm_vcpu_get_events(vcpu, &events))
1879 			return -EINVAL;
1880 
1881 		if (copy_to_user(argp, &events, sizeof(events)))
1882 			return -EFAULT;
1883 
1884 		return 0;
1885 	}
1886 	case KVM_SET_VCPU_EVENTS: {
1887 		struct kvm_vcpu_events events;
1888 
1889 		if (!kvm_vcpu_initialized(vcpu))
1890 			return -ENOEXEC;
1891 
1892 		if (copy_from_user(&events, argp, sizeof(events)))
1893 			return -EFAULT;
1894 
1895 		return kvm_arm_vcpu_set_events(vcpu, &events);
1896 	}
1897 	case KVM_ARM_VCPU_FINALIZE: {
1898 		int what;
1899 
1900 		if (!kvm_vcpu_initialized(vcpu))
1901 			return -ENOEXEC;
1902 
1903 		if (get_user(what, (const int __user *)argp))
1904 			return -EFAULT;
1905 
1906 		return kvm_arm_vcpu_finalize(vcpu, what);
1907 	}
1908 	default:
1909 		r = -EINVAL;
1910 	}
1911 
1912 	return r;
1913 }
1914 
1915 long kvm_arch_vcpu_unlocked_ioctl(struct file *filp, unsigned int ioctl,
1916 				  unsigned long arg)
1917 {
1918 	return -ENOIOCTLCMD;
1919 }
1920 
1921 void kvm_arch_sync_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot)
1922 {
1923 
1924 }
1925 
1926 static int kvm_vm_ioctl_set_device_addr(struct kvm *kvm,
1927 					struct kvm_arm_device_addr *dev_addr)
1928 {
1929 	switch (FIELD_GET(KVM_ARM_DEVICE_ID_MASK, dev_addr->id)) {
1930 	case KVM_ARM_DEVICE_VGIC_V2:
1931 		if (!vgic_present)
1932 			return -ENXIO;
1933 		return kvm_set_legacy_vgic_v2_addr(kvm, dev_addr);
1934 	default:
1935 		return -ENODEV;
1936 	}
1937 }
1938 
1939 static int kvm_vm_has_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1940 {
1941 	switch (attr->group) {
1942 	case KVM_ARM_VM_SMCCC_CTRL:
1943 		return kvm_vm_smccc_has_attr(kvm, attr);
1944 	default:
1945 		return -ENXIO;
1946 	}
1947 }
1948 
1949 static int kvm_vm_set_attr(struct kvm *kvm, struct kvm_device_attr *attr)
1950 {
1951 	switch (attr->group) {
1952 	case KVM_ARM_VM_SMCCC_CTRL:
1953 		return kvm_vm_smccc_set_attr(kvm, attr);
1954 	default:
1955 		return -ENXIO;
1956 	}
1957 }
1958 
1959 int kvm_arch_vm_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
1960 {
1961 	struct kvm *kvm = filp->private_data;
1962 	void __user *argp = (void __user *)arg;
1963 	struct kvm_device_attr attr;
1964 
1965 	if (is_protected_kvm_enabled() && !kvm_pkvm_ioctl_allowed(kvm, ioctl))
1966 		return -EINVAL;
1967 
1968 	switch (ioctl) {
1969 	case KVM_CREATE_IRQCHIP: {
1970 		int ret;
1971 		if (!vgic_present)
1972 			return -ENXIO;
1973 		mutex_lock(&kvm->lock);
1974 		ret = kvm_vgic_create(kvm, KVM_DEV_TYPE_ARM_VGIC_V2);
1975 		mutex_unlock(&kvm->lock);
1976 		return ret;
1977 	}
1978 	case KVM_ARM_SET_DEVICE_ADDR: {
1979 		struct kvm_arm_device_addr dev_addr;
1980 
1981 		if (copy_from_user(&dev_addr, argp, sizeof(dev_addr)))
1982 			return -EFAULT;
1983 		return kvm_vm_ioctl_set_device_addr(kvm, &dev_addr);
1984 	}
1985 	case KVM_ARM_PREFERRED_TARGET: {
1986 		struct kvm_vcpu_init init = {
1987 			.target = KVM_ARM_TARGET_GENERIC_V8,
1988 		};
1989 
1990 		if (copy_to_user(argp, &init, sizeof(init)))
1991 			return -EFAULT;
1992 
1993 		return 0;
1994 	}
1995 	case KVM_ARM_MTE_COPY_TAGS: {
1996 		struct kvm_arm_copy_mte_tags copy_tags;
1997 
1998 		if (copy_from_user(&copy_tags, argp, sizeof(copy_tags)))
1999 			return -EFAULT;
2000 		return kvm_vm_ioctl_mte_copy_tags(kvm, &copy_tags);
2001 	}
2002 	case KVM_ARM_SET_COUNTER_OFFSET: {
2003 		struct kvm_arm_counter_offset offset;
2004 
2005 		if (copy_from_user(&offset, argp, sizeof(offset)))
2006 			return -EFAULT;
2007 		return kvm_vm_ioctl_set_counter_offset(kvm, &offset);
2008 	}
2009 	case KVM_HAS_DEVICE_ATTR: {
2010 		if (copy_from_user(&attr, argp, sizeof(attr)))
2011 			return -EFAULT;
2012 
2013 		return kvm_vm_has_attr(kvm, &attr);
2014 	}
2015 	case KVM_SET_DEVICE_ATTR: {
2016 		if (copy_from_user(&attr, argp, sizeof(attr)))
2017 			return -EFAULT;
2018 
2019 		return kvm_vm_set_attr(kvm, &attr);
2020 	}
2021 	case KVM_ARM_GET_REG_WRITABLE_MASKS: {
2022 		struct reg_mask_range range;
2023 
2024 		if (copy_from_user(&range, argp, sizeof(range)))
2025 			return -EFAULT;
2026 		return kvm_vm_ioctl_get_reg_writable_masks(kvm, &range);
2027 	}
2028 	default:
2029 		return -EINVAL;
2030 	}
2031 }
2032 
2033 static unsigned long nvhe_percpu_size(void)
2034 {
2035 	return (unsigned long)CHOOSE_NVHE_SYM(__per_cpu_end) -
2036 		(unsigned long)CHOOSE_NVHE_SYM(__per_cpu_start);
2037 }
2038 
2039 static unsigned long nvhe_percpu_order(void)
2040 {
2041 	unsigned long size = nvhe_percpu_size();
2042 
2043 	return size ? get_order(size) : 0;
2044 }
2045 
2046 static size_t pkvm_host_sve_state_order(void)
2047 {
2048 	return get_order(pkvm_host_sve_state_size());
2049 }
2050 
2051 /* A lookup table holding the hypervisor VA for each vector slot */
2052 static void *hyp_spectre_vector_selector[BP_HARDEN_EL2_SLOTS];
2053 
2054 static void kvm_init_vector_slot(void *base, enum arm64_hyp_spectre_vector slot)
2055 {
2056 	hyp_spectre_vector_selector[slot] = __kvm_vector_slot2addr(base, slot);
2057 }
2058 
2059 static int kvm_init_vector_slots(void)
2060 {
2061 	int err;
2062 	void *base;
2063 
2064 	base = kern_hyp_va(kvm_ksym_ref(__kvm_hyp_vector));
2065 	kvm_init_vector_slot(base, HYP_VECTOR_DIRECT);
2066 
2067 	base = kern_hyp_va(kvm_ksym_ref(__bp_harden_hyp_vecs));
2068 	kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_DIRECT);
2069 
2070 	if (kvm_system_needs_idmapped_vectors() &&
2071 	    !is_protected_kvm_enabled()) {
2072 		err = create_hyp_exec_mappings(__pa_symbol(__bp_harden_hyp_vecs),
2073 					       __BP_HARDEN_HYP_VECS_SZ, &base);
2074 		if (err)
2075 			return err;
2076 	}
2077 
2078 	kvm_init_vector_slot(base, HYP_VECTOR_INDIRECT);
2079 	kvm_init_vector_slot(base, HYP_VECTOR_SPECTRE_INDIRECT);
2080 	return 0;
2081 }
2082 
2083 static void __init cpu_prepare_hyp_mode(int cpu, u32 hyp_va_bits)
2084 {
2085 	struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2086 	unsigned long tcr;
2087 
2088 	/*
2089 	 * Calculate the raw per-cpu offset without a translation from the
2090 	 * kernel's mapping to the linear mapping, and store it in tpidr_el2
2091 	 * so that we can use adr_l to access per-cpu variables in EL2.
2092 	 * Also drop the KASAN tag which gets in the way...
2093 	 */
2094 	params->tpidr_el2 = (unsigned long)kasan_reset_tag(per_cpu_ptr_nvhe_sym(__per_cpu_start, cpu)) -
2095 			    (unsigned long)kvm_ksym_ref(CHOOSE_NVHE_SYM(__per_cpu_start));
2096 
2097 	params->mair_el2 = read_sysreg(mair_el1);
2098 
2099 	tcr = read_sysreg(tcr_el1);
2100 	if (cpus_have_final_cap(ARM64_KVM_HVHE)) {
2101 		tcr &= ~(TCR_HD | TCR_HA | TCR_A1 | TCR_T0SZ_MASK);
2102 		tcr |= TCR_EPD1_MASK;
2103 	} else {
2104 		unsigned long ips = FIELD_GET(TCR_IPS_MASK, tcr);
2105 
2106 		tcr &= TCR_EL2_MASK;
2107 		tcr |= TCR_EL2_RES1 | FIELD_PREP(TCR_EL2_PS_MASK, ips);
2108 		if (lpa2_is_enabled())
2109 			tcr |= TCR_EL2_DS;
2110 	}
2111 	tcr |= TCR_T0SZ(hyp_va_bits);
2112 	params->tcr_el2 = tcr;
2113 
2114 	params->pgd_pa = kvm_mmu_get_httbr();
2115 	if (is_protected_kvm_enabled())
2116 		params->hcr_el2 = HCR_HOST_NVHE_PROTECTED_FLAGS;
2117 	else
2118 		params->hcr_el2 = HCR_HOST_NVHE_FLAGS;
2119 
2120 	if (system_supports_mte())
2121 		params->hcr_el2 |= HCR_ATA;
2122 	else
2123 		params->hcr_el2 |= HCR_TID5;
2124 
2125 	if (cpus_have_final_cap(ARM64_KVM_HVHE))
2126 		params->hcr_el2 |= HCR_E2H;
2127 	params->vttbr = params->vtcr = 0;
2128 
2129 	/*
2130 	 * Flush the init params from the data cache because the struct will
2131 	 * be read while the MMU is off.
2132 	 */
2133 	kvm_flush_dcache_to_poc(params, sizeof(*params));
2134 }
2135 
2136 static void hyp_install_host_vector(void)
2137 {
2138 	struct kvm_nvhe_init_params *params;
2139 	struct arm_smccc_res res;
2140 
2141 	/* Switch from the HYP stub to our own HYP init vector */
2142 	__hyp_set_vectors(kvm_get_idmap_vector());
2143 
2144 	/*
2145 	 * Call initialization code, and switch to the full blown HYP code.
2146 	 * If the cpucaps haven't been finalized yet, something has gone very
2147 	 * wrong, and hyp will crash and burn when it uses any
2148 	 * cpus_have_*_cap() wrapper.
2149 	 */
2150 	BUG_ON(!system_capabilities_finalized());
2151 	params = this_cpu_ptr_nvhe_sym(kvm_init_params);
2152 	arm_smccc_1_1_hvc(KVM_HOST_SMCCC_FUNC(__kvm_hyp_init), virt_to_phys(params), &res);
2153 	WARN_ON(res.a0 != SMCCC_RET_SUCCESS);
2154 }
2155 
2156 static void cpu_init_hyp_mode(void)
2157 {
2158 	hyp_install_host_vector();
2159 
2160 	/*
2161 	 * Disabling SSBD on a non-VHE system requires us to enable SSBS
2162 	 * at EL2.
2163 	 */
2164 	if (this_cpu_has_cap(ARM64_SSBS) &&
2165 	    arm64_get_spectre_v4_state() == SPECTRE_VULNERABLE) {
2166 		kvm_call_hyp_nvhe(__kvm_enable_ssbs);
2167 	}
2168 }
2169 
2170 static void cpu_hyp_reset(void)
2171 {
2172 	if (!is_kernel_in_hyp_mode())
2173 		__hyp_reset_vectors();
2174 }
2175 
2176 /*
2177  * EL2 vectors can be mapped and rerouted in a number of ways,
2178  * depending on the kernel configuration and CPU present:
2179  *
2180  * - If the CPU is affected by Spectre-v2, the hardening sequence is
2181  *   placed in one of the vector slots, which is executed before jumping
2182  *   to the real vectors.
2183  *
2184  * - If the CPU also has the ARM64_SPECTRE_V3A cap, the slot
2185  *   containing the hardening sequence is mapped next to the idmap page,
2186  *   and executed before jumping to the real vectors.
2187  *
2188  * - If the CPU only has the ARM64_SPECTRE_V3A cap, then an
2189  *   empty slot is selected, mapped next to the idmap page, and
2190  *   executed before jumping to the real vectors.
2191  *
2192  * Note that ARM64_SPECTRE_V3A is somewhat incompatible with
2193  * VHE, as we don't have hypervisor-specific mappings. If the system
2194  * is VHE and yet selects this capability, it will be ignored.
2195  */
2196 static void cpu_set_hyp_vector(void)
2197 {
2198 	struct bp_hardening_data *data = this_cpu_ptr(&bp_hardening_data);
2199 	void *vector = hyp_spectre_vector_selector[data->slot];
2200 
2201 	if (!is_protected_kvm_enabled())
2202 		*this_cpu_ptr_hyp_sym(kvm_hyp_vector) = (unsigned long)vector;
2203 	else
2204 		kvm_call_hyp_nvhe(__pkvm_cpu_set_vector, data->slot);
2205 }
2206 
2207 static void cpu_hyp_init_context(void)
2208 {
2209 	kvm_init_host_cpu_context(host_data_ptr(host_ctxt));
2210 	kvm_init_host_debug_data();
2211 
2212 	if (!is_kernel_in_hyp_mode())
2213 		cpu_init_hyp_mode();
2214 }
2215 
2216 static void cpu_hyp_init_features(void)
2217 {
2218 	cpu_set_hyp_vector();
2219 
2220 	if (is_kernel_in_hyp_mode()) {
2221 		kvm_timer_init_vhe();
2222 		kvm_debug_init_vhe();
2223 	}
2224 
2225 	if (vgic_present)
2226 		kvm_vgic_init_cpu_hardware();
2227 }
2228 
2229 static void cpu_hyp_reinit(void)
2230 {
2231 	cpu_hyp_reset();
2232 	cpu_hyp_init_context();
2233 	cpu_hyp_init_features();
2234 }
2235 
2236 static void cpu_hyp_init(void *discard)
2237 {
2238 	if (!__this_cpu_read(kvm_hyp_initialized)) {
2239 		cpu_hyp_reinit();
2240 		__this_cpu_write(kvm_hyp_initialized, 1);
2241 	}
2242 }
2243 
2244 static void cpu_hyp_uninit(void *discard)
2245 {
2246 	if (!is_protected_kvm_enabled() && __this_cpu_read(kvm_hyp_initialized)) {
2247 		cpu_hyp_reset();
2248 		__this_cpu_write(kvm_hyp_initialized, 0);
2249 	}
2250 }
2251 
2252 int kvm_arch_enable_virtualization_cpu(void)
2253 {
2254 	/*
2255 	 * Most calls to this function are made with migration
2256 	 * disabled, but not with preemption disabled. The former is
2257 	 * enough to ensure correctness, but most of the helpers
2258 	 * expect the later and will throw a tantrum otherwise.
2259 	 */
2260 	preempt_disable();
2261 
2262 	cpu_hyp_init(NULL);
2263 
2264 	kvm_vgic_cpu_up();
2265 	kvm_timer_cpu_up();
2266 
2267 	preempt_enable();
2268 
2269 	return 0;
2270 }
2271 
2272 void kvm_arch_disable_virtualization_cpu(void)
2273 {
2274 	kvm_timer_cpu_down();
2275 	kvm_vgic_cpu_down();
2276 
2277 	if (!is_protected_kvm_enabled())
2278 		cpu_hyp_uninit(NULL);
2279 }
2280 
2281 #ifdef CONFIG_CPU_PM
2282 static int hyp_init_cpu_pm_notifier(struct notifier_block *self,
2283 				    unsigned long cmd,
2284 				    void *v)
2285 {
2286 	/*
2287 	 * kvm_hyp_initialized is left with its old value over
2288 	 * PM_ENTER->PM_EXIT. It is used to indicate PM_EXIT should
2289 	 * re-enable hyp.
2290 	 */
2291 	switch (cmd) {
2292 	case CPU_PM_ENTER:
2293 		if (__this_cpu_read(kvm_hyp_initialized))
2294 			/*
2295 			 * don't update kvm_hyp_initialized here
2296 			 * so that the hyp will be re-enabled
2297 			 * when we resume. See below.
2298 			 */
2299 			cpu_hyp_reset();
2300 
2301 		return NOTIFY_OK;
2302 	case CPU_PM_ENTER_FAILED:
2303 	case CPU_PM_EXIT:
2304 		if (__this_cpu_read(kvm_hyp_initialized))
2305 			/* The hyp was enabled before suspend. */
2306 			cpu_hyp_reinit();
2307 
2308 		return NOTIFY_OK;
2309 
2310 	default:
2311 		return NOTIFY_DONE;
2312 	}
2313 }
2314 
2315 static struct notifier_block hyp_init_cpu_pm_nb = {
2316 	.notifier_call = hyp_init_cpu_pm_notifier,
2317 };
2318 
2319 static void __init hyp_cpu_pm_init(void)
2320 {
2321 	if (!is_protected_kvm_enabled())
2322 		cpu_pm_register_notifier(&hyp_init_cpu_pm_nb);
2323 }
2324 static void __init hyp_cpu_pm_exit(void)
2325 {
2326 	if (!is_protected_kvm_enabled())
2327 		cpu_pm_unregister_notifier(&hyp_init_cpu_pm_nb);
2328 }
2329 #else
2330 static inline void __init hyp_cpu_pm_init(void)
2331 {
2332 }
2333 static inline void __init hyp_cpu_pm_exit(void)
2334 {
2335 }
2336 #endif
2337 
2338 static void __init init_cpu_logical_map(void)
2339 {
2340 	unsigned int cpu;
2341 
2342 	/*
2343 	 * Copy the MPIDR <-> logical CPU ID mapping to hyp.
2344 	 * Only copy the set of online CPUs whose features have been checked
2345 	 * against the finalized system capabilities. The hypervisor will not
2346 	 * allow any other CPUs from the `possible` set to boot.
2347 	 */
2348 	for_each_online_cpu(cpu)
2349 		hyp_cpu_logical_map[cpu] = cpu_logical_map(cpu);
2350 }
2351 
2352 #define init_psci_0_1_impl_state(config, what)	\
2353 	config.psci_0_1_ ## what ## _implemented = psci_ops.what
2354 
2355 static bool __init init_psci_relay(void)
2356 {
2357 	/*
2358 	 * If PSCI has not been initialized, protected KVM cannot install
2359 	 * itself on newly booted CPUs.
2360 	 */
2361 	if (!psci_ops.get_version) {
2362 		kvm_err("Cannot initialize protected mode without PSCI\n");
2363 		return false;
2364 	}
2365 
2366 	kvm_host_psci_config.version = psci_ops.get_version();
2367 	kvm_host_psci_config.smccc_version = arm_smccc_get_version();
2368 
2369 	if (kvm_host_psci_config.version == PSCI_VERSION(0, 1)) {
2370 		kvm_host_psci_config.function_ids_0_1 = get_psci_0_1_function_ids();
2371 		init_psci_0_1_impl_state(kvm_host_psci_config, cpu_suspend);
2372 		init_psci_0_1_impl_state(kvm_host_psci_config, cpu_on);
2373 		init_psci_0_1_impl_state(kvm_host_psci_config, cpu_off);
2374 		init_psci_0_1_impl_state(kvm_host_psci_config, migrate);
2375 	}
2376 	return true;
2377 }
2378 
2379 static int __init init_subsystems(void)
2380 {
2381 	int err = 0;
2382 
2383 	/*
2384 	 * Enable hardware so that subsystem initialisation can access EL2.
2385 	 */
2386 	on_each_cpu(cpu_hyp_init, NULL, 1);
2387 
2388 	/*
2389 	 * Register CPU lower-power notifier
2390 	 */
2391 	hyp_cpu_pm_init();
2392 
2393 	/*
2394 	 * Init HYP view of VGIC
2395 	 */
2396 	err = kvm_vgic_hyp_init();
2397 	switch (err) {
2398 	case 0:
2399 		vgic_present = true;
2400 		break;
2401 	case -ENODEV:
2402 	case -ENXIO:
2403 		/*
2404 		 * No VGIC? No pKVM for you.
2405 		 *
2406 		 * Protected mode assumes that VGICv3 is present, so no point
2407 		 * in trying to hobble along if vgic initialization fails.
2408 		 */
2409 		if (is_protected_kvm_enabled())
2410 			goto out;
2411 
2412 		/*
2413 		 * Otherwise, userspace could choose to implement a GIC for its
2414 		 * guest on non-cooperative hardware.
2415 		 */
2416 		vgic_present = false;
2417 		err = 0;
2418 		break;
2419 	default:
2420 		goto out;
2421 	}
2422 
2423 	if (kvm_mode == KVM_MODE_NV &&
2424 		!(vgic_present && (kvm_vgic_global_state.type == VGIC_V3 ||
2425 				   kvm_vgic_global_state.has_gcie_v3_compat))) {
2426 		kvm_err("NV support requires GICv3 or GICv5 with legacy support, giving up\n");
2427 		err = -EINVAL;
2428 		goto out;
2429 	}
2430 
2431 	/*
2432 	 * Init HYP architected timer support
2433 	 */
2434 	err = kvm_timer_hyp_init(vgic_present);
2435 	if (err)
2436 		goto out;
2437 
2438 	kvm_register_perf_callbacks();
2439 
2440 out:
2441 	if (err)
2442 		hyp_cpu_pm_exit();
2443 
2444 	if (err || !is_protected_kvm_enabled())
2445 		on_each_cpu(cpu_hyp_uninit, NULL, 1);
2446 
2447 	return err;
2448 }
2449 
2450 static void __init teardown_subsystems(void)
2451 {
2452 	kvm_unregister_perf_callbacks();
2453 	hyp_cpu_pm_exit();
2454 }
2455 
2456 static void __init teardown_hyp_mode(void)
2457 {
2458 	bool free_sve = system_supports_sve() && is_protected_kvm_enabled();
2459 	int cpu;
2460 
2461 	free_hyp_pgds();
2462 	for_each_possible_cpu(cpu) {
2463 		if (per_cpu(kvm_hyp_initialized, cpu))
2464 			continue;
2465 
2466 		free_pages(per_cpu(kvm_arm_hyp_stack_base, cpu), NVHE_STACK_SHIFT - PAGE_SHIFT);
2467 
2468 		if (!kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu])
2469 			continue;
2470 
2471 		if (free_sve) {
2472 			struct cpu_sve_state *sve_state;
2473 
2474 			sve_state = per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state;
2475 			free_pages((unsigned long) sve_state, pkvm_host_sve_state_order());
2476 		}
2477 
2478 		free_pages(kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu], nvhe_percpu_order());
2479 
2480 	}
2481 }
2482 
2483 static int __init do_pkvm_init(u32 hyp_va_bits)
2484 {
2485 	void *per_cpu_base = kvm_ksym_ref(kvm_nvhe_sym(kvm_arm_hyp_percpu_base));
2486 	int ret;
2487 
2488 	preempt_disable();
2489 	cpu_hyp_init_context();
2490 	ret = kvm_call_hyp_nvhe(__pkvm_init, hyp_mem_base, hyp_mem_size,
2491 				num_possible_cpus(), kern_hyp_va(per_cpu_base),
2492 				hyp_va_bits);
2493 	cpu_hyp_init_features();
2494 
2495 	/*
2496 	 * The stub hypercalls are now disabled, so set our local flag to
2497 	 * prevent a later re-init attempt in kvm_arch_enable_virtualization_cpu().
2498 	 */
2499 	__this_cpu_write(kvm_hyp_initialized, 1);
2500 	preempt_enable();
2501 
2502 	return ret;
2503 }
2504 
2505 static u64 get_hyp_id_aa64pfr0_el1(void)
2506 {
2507 	/*
2508 	 * Track whether the system isn't affected by spectre/meltdown in the
2509 	 * hypervisor's view of id_aa64pfr0_el1, used for protected VMs.
2510 	 * Although this is per-CPU, we make it global for simplicity, e.g., not
2511 	 * to have to worry about vcpu migration.
2512 	 *
2513 	 * Unlike for non-protected VMs, userspace cannot override this for
2514 	 * protected VMs.
2515 	 */
2516 	u64 val = read_sanitised_ftr_reg(SYS_ID_AA64PFR0_EL1);
2517 
2518 	val &= ~(ID_AA64PFR0_EL1_CSV2 |
2519 		 ID_AA64PFR0_EL1_CSV3);
2520 
2521 	val |= FIELD_PREP(ID_AA64PFR0_EL1_CSV2,
2522 			  arm64_get_spectre_v2_state() == SPECTRE_UNAFFECTED);
2523 	val |= FIELD_PREP(ID_AA64PFR0_EL1_CSV3,
2524 			  arm64_get_meltdown_state() == SPECTRE_UNAFFECTED);
2525 
2526 	return val;
2527 }
2528 
2529 static void kvm_hyp_init_symbols(void)
2530 {
2531 	kvm_nvhe_sym(id_aa64pfr0_el1_sys_val) = get_hyp_id_aa64pfr0_el1();
2532 	kvm_nvhe_sym(id_aa64pfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64PFR1_EL1);
2533 	kvm_nvhe_sym(id_aa64isar0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR0_EL1);
2534 	kvm_nvhe_sym(id_aa64isar1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR1_EL1);
2535 	kvm_nvhe_sym(id_aa64isar2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64ISAR2_EL1);
2536 	kvm_nvhe_sym(id_aa64mmfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR0_EL1);
2537 	kvm_nvhe_sym(id_aa64mmfr1_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR1_EL1);
2538 	kvm_nvhe_sym(id_aa64mmfr2_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64MMFR2_EL1);
2539 	kvm_nvhe_sym(id_aa64smfr0_el1_sys_val) = read_sanitised_ftr_reg(SYS_ID_AA64SMFR0_EL1);
2540 	kvm_nvhe_sym(__icache_flags) = __icache_flags;
2541 	kvm_nvhe_sym(kvm_arm_vmid_bits) = kvm_arm_vmid_bits;
2542 
2543 	/* Propagate the FGT state to the nVHE side */
2544 	kvm_nvhe_sym(hfgrtr_masks)  = hfgrtr_masks;
2545 	kvm_nvhe_sym(hfgwtr_masks)  = hfgwtr_masks;
2546 	kvm_nvhe_sym(hfgitr_masks)  = hfgitr_masks;
2547 	kvm_nvhe_sym(hdfgrtr_masks) = hdfgrtr_masks;
2548 	kvm_nvhe_sym(hdfgwtr_masks) = hdfgwtr_masks;
2549 	kvm_nvhe_sym(hafgrtr_masks) = hafgrtr_masks;
2550 	kvm_nvhe_sym(hfgrtr2_masks) = hfgrtr2_masks;
2551 	kvm_nvhe_sym(hfgwtr2_masks) = hfgwtr2_masks;
2552 	kvm_nvhe_sym(hfgitr2_masks) = hfgitr2_masks;
2553 	kvm_nvhe_sym(hdfgrtr2_masks)= hdfgrtr2_masks;
2554 	kvm_nvhe_sym(hdfgwtr2_masks)= hdfgwtr2_masks;
2555 	kvm_nvhe_sym(ich_hfgrtr_masks) = ich_hfgrtr_masks;
2556 	kvm_nvhe_sym(ich_hfgwtr_masks) = ich_hfgwtr_masks;
2557 	kvm_nvhe_sym(ich_hfgitr_masks) = ich_hfgitr_masks;
2558 
2559 	/*
2560 	 * Flush entire BSS since part of its data containing init symbols is read
2561 	 * while the MMU is off.
2562 	 */
2563 	kvm_flush_dcache_to_poc(kvm_ksym_ref(__hyp_bss_start),
2564 				kvm_ksym_ref(__hyp_bss_end) - kvm_ksym_ref(__hyp_bss_start));
2565 }
2566 
2567 static int __init kvm_hyp_init_protection(u32 hyp_va_bits)
2568 {
2569 	void *addr = phys_to_virt(hyp_mem_base);
2570 	int ret;
2571 
2572 	ret = create_hyp_mappings(addr, addr + hyp_mem_size, PAGE_HYP);
2573 	if (ret)
2574 		return ret;
2575 
2576 	ret = do_pkvm_init(hyp_va_bits);
2577 	if (ret)
2578 		return ret;
2579 
2580 	free_hyp_pgds();
2581 
2582 	return 0;
2583 }
2584 
2585 static int init_pkvm_host_sve_state(void)
2586 {
2587 	int cpu;
2588 
2589 	if (!system_supports_sve())
2590 		return 0;
2591 
2592 	/* Allocate pages for host sve state in protected mode. */
2593 	for_each_possible_cpu(cpu) {
2594 		struct page *page = alloc_pages(GFP_KERNEL, pkvm_host_sve_state_order());
2595 
2596 		if (!page)
2597 			return -ENOMEM;
2598 
2599 		per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state = page_address(page);
2600 	}
2601 
2602 	/*
2603 	 * Don't map the pages in hyp since these are only used in protected
2604 	 * mode, which will (re)create its own mapping when initialized.
2605 	 */
2606 
2607 	return 0;
2608 }
2609 
2610 /*
2611  * Finalizes the initialization of hyp mode, once everything else is initialized
2612  * and the initialziation process cannot fail.
2613  */
2614 static void finalize_init_hyp_mode(void)
2615 {
2616 	int cpu;
2617 
2618 	if (system_supports_sve() && is_protected_kvm_enabled()) {
2619 		for_each_possible_cpu(cpu) {
2620 			struct cpu_sve_state *sve_state;
2621 
2622 			sve_state = per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state;
2623 			per_cpu_ptr_nvhe_sym(kvm_host_data, cpu)->sve_state =
2624 				kern_hyp_va(sve_state);
2625 		}
2626 	}
2627 }
2628 
2629 static void pkvm_hyp_init_ptrauth(void)
2630 {
2631 	struct kvm_cpu_context *hyp_ctxt;
2632 	int cpu;
2633 
2634 	for_each_possible_cpu(cpu) {
2635 		hyp_ctxt = per_cpu_ptr_nvhe_sym(kvm_hyp_ctxt, cpu);
2636 		hyp_ctxt->sys_regs[APIAKEYLO_EL1] = get_random_long();
2637 		hyp_ctxt->sys_regs[APIAKEYHI_EL1] = get_random_long();
2638 		hyp_ctxt->sys_regs[APIBKEYLO_EL1] = get_random_long();
2639 		hyp_ctxt->sys_regs[APIBKEYHI_EL1] = get_random_long();
2640 		hyp_ctxt->sys_regs[APDAKEYLO_EL1] = get_random_long();
2641 		hyp_ctxt->sys_regs[APDAKEYHI_EL1] = get_random_long();
2642 		hyp_ctxt->sys_regs[APDBKEYLO_EL1] = get_random_long();
2643 		hyp_ctxt->sys_regs[APDBKEYHI_EL1] = get_random_long();
2644 		hyp_ctxt->sys_regs[APGAKEYLO_EL1] = get_random_long();
2645 		hyp_ctxt->sys_regs[APGAKEYHI_EL1] = get_random_long();
2646 	}
2647 }
2648 
2649 /* Inits Hyp-mode on all online CPUs */
2650 static int __init init_hyp_mode(void)
2651 {
2652 	u32 hyp_va_bits = kvm_hyp_va_bits();
2653 	int cpu;
2654 	int err = -ENOMEM;
2655 
2656 	/*
2657 	 * The protected Hyp-mode cannot be initialized if the memory pool
2658 	 * allocation has failed.
2659 	 */
2660 	if (is_protected_kvm_enabled() && !hyp_mem_base)
2661 		goto out_err;
2662 
2663 	/*
2664 	 * Allocate Hyp PGD and setup Hyp identity mapping
2665 	 */
2666 	err = kvm_mmu_init(hyp_va_bits);
2667 	if (err)
2668 		goto out_err;
2669 
2670 	/*
2671 	 * Allocate stack pages for Hypervisor-mode
2672 	 */
2673 	for_each_possible_cpu(cpu) {
2674 		unsigned long stack_base;
2675 
2676 		stack_base = __get_free_pages(GFP_KERNEL, NVHE_STACK_SHIFT - PAGE_SHIFT);
2677 		if (!stack_base) {
2678 			err = -ENOMEM;
2679 			goto out_err;
2680 		}
2681 
2682 		per_cpu(kvm_arm_hyp_stack_base, cpu) = stack_base;
2683 	}
2684 
2685 	/*
2686 	 * Allocate and initialize pages for Hypervisor-mode percpu regions.
2687 	 */
2688 	for_each_possible_cpu(cpu) {
2689 		struct page *page;
2690 		void *page_addr;
2691 
2692 		page = alloc_pages(GFP_KERNEL, nvhe_percpu_order());
2693 		if (!page) {
2694 			err = -ENOMEM;
2695 			goto out_err;
2696 		}
2697 
2698 		page_addr = page_address(page);
2699 		memcpy(page_addr, CHOOSE_NVHE_SYM(__per_cpu_start), nvhe_percpu_size());
2700 		kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu] = (unsigned long)page_addr;
2701 	}
2702 
2703 	/*
2704 	 * Map the Hyp-code called directly from the host
2705 	 */
2706 	err = create_hyp_mappings(kvm_ksym_ref(__hyp_text_start),
2707 				  kvm_ksym_ref(__hyp_text_end), PAGE_HYP_EXEC);
2708 	if (err) {
2709 		kvm_err("Cannot map world-switch code\n");
2710 		goto out_err;
2711 	}
2712 
2713 	err = create_hyp_mappings(kvm_ksym_ref(__hyp_data_start),
2714 				  kvm_ksym_ref(__hyp_data_end), PAGE_HYP);
2715 	if (err) {
2716 		kvm_err("Cannot map .hyp.data section\n");
2717 		goto out_err;
2718 	}
2719 
2720 	err = create_hyp_mappings(kvm_ksym_ref(__hyp_rodata_start),
2721 				  kvm_ksym_ref(__hyp_rodata_end), PAGE_HYP_RO);
2722 	if (err) {
2723 		kvm_err("Cannot map .hyp.rodata section\n");
2724 		goto out_err;
2725 	}
2726 
2727 	err = create_hyp_mappings(kvm_ksym_ref(__start_rodata),
2728 				  kvm_ksym_ref(__end_rodata), PAGE_HYP_RO);
2729 	if (err) {
2730 		kvm_err("Cannot map rodata section\n");
2731 		goto out_err;
2732 	}
2733 
2734 	/*
2735 	 * .hyp.bss is guaranteed to be placed at the beginning of the .bss
2736 	 * section thanks to an assertion in the linker script. Map it RW and
2737 	 * the rest of .bss RO.
2738 	 */
2739 	err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_start),
2740 				  kvm_ksym_ref(__hyp_bss_end), PAGE_HYP);
2741 	if (err) {
2742 		kvm_err("Cannot map hyp bss section: %d\n", err);
2743 		goto out_err;
2744 	}
2745 
2746 	err = create_hyp_mappings(kvm_ksym_ref(__hyp_bss_end),
2747 				  kvm_ksym_ref(__bss_stop), PAGE_HYP_RO);
2748 	if (err) {
2749 		kvm_err("Cannot map bss section\n");
2750 		goto out_err;
2751 	}
2752 
2753 	/*
2754 	 * Map the Hyp stack pages
2755 	 */
2756 	for_each_possible_cpu(cpu) {
2757 		struct kvm_nvhe_init_params *params = per_cpu_ptr_nvhe_sym(kvm_init_params, cpu);
2758 		char *stack_base = (char *)per_cpu(kvm_arm_hyp_stack_base, cpu);
2759 
2760 		err = create_hyp_stack(__pa(stack_base), &params->stack_hyp_va);
2761 		if (err) {
2762 			kvm_err("Cannot map hyp stack\n");
2763 			goto out_err;
2764 		}
2765 
2766 		/*
2767 		 * Save the stack PA in nvhe_init_params. This will be needed
2768 		 * to recreate the stack mapping in protected nVHE mode.
2769 		 * __hyp_pa() won't do the right thing there, since the stack
2770 		 * has been mapped in the flexible private VA space.
2771 		 */
2772 		params->stack_pa = __pa(stack_base);
2773 	}
2774 
2775 	for_each_possible_cpu(cpu) {
2776 		char *percpu_begin = (char *)kvm_nvhe_sym(kvm_arm_hyp_percpu_base)[cpu];
2777 		char *percpu_end = percpu_begin + nvhe_percpu_size();
2778 
2779 		/* Map Hyp percpu pages */
2780 		err = create_hyp_mappings(percpu_begin, percpu_end, PAGE_HYP);
2781 		if (err) {
2782 			kvm_err("Cannot map hyp percpu region\n");
2783 			goto out_err;
2784 		}
2785 
2786 		/* Prepare the CPU initialization parameters */
2787 		cpu_prepare_hyp_mode(cpu, hyp_va_bits);
2788 	}
2789 
2790 	kvm_hyp_init_symbols();
2791 
2792 	if (is_protected_kvm_enabled()) {
2793 		if (IS_ENABLED(CONFIG_ARM64_PTR_AUTH_KERNEL) &&
2794 		    cpus_have_final_cap(ARM64_HAS_ADDRESS_AUTH))
2795 			pkvm_hyp_init_ptrauth();
2796 
2797 		init_cpu_logical_map();
2798 
2799 		if (!init_psci_relay()) {
2800 			err = -ENODEV;
2801 			goto out_err;
2802 		}
2803 
2804 		err = init_pkvm_host_sve_state();
2805 		if (err)
2806 			goto out_err;
2807 
2808 		err = kvm_hyp_init_protection(hyp_va_bits);
2809 		if (err) {
2810 			kvm_err("Failed to init hyp memory protection\n");
2811 			goto out_err;
2812 		}
2813 	}
2814 
2815 	return 0;
2816 
2817 out_err:
2818 	teardown_hyp_mode();
2819 	kvm_err("error initializing Hyp mode: %d\n", err);
2820 	return err;
2821 }
2822 
2823 struct kvm_vcpu *kvm_mpidr_to_vcpu(struct kvm *kvm, unsigned long mpidr)
2824 {
2825 	struct kvm_vcpu *vcpu = NULL;
2826 	struct kvm_mpidr_data *data;
2827 	unsigned long i;
2828 
2829 	mpidr &= MPIDR_HWID_BITMASK;
2830 
2831 	rcu_read_lock();
2832 	data = rcu_dereference(kvm->arch.mpidr_data);
2833 
2834 	if (data) {
2835 		u16 idx = kvm_mpidr_index(data, mpidr);
2836 
2837 		vcpu = kvm_get_vcpu(kvm, data->cmpidr_to_idx[idx]);
2838 		if (mpidr != kvm_vcpu_get_mpidr_aff(vcpu))
2839 			vcpu = NULL;
2840 	}
2841 
2842 	rcu_read_unlock();
2843 
2844 	if (vcpu)
2845 		return vcpu;
2846 
2847 	kvm_for_each_vcpu(i, vcpu, kvm) {
2848 		if (mpidr == kvm_vcpu_get_mpidr_aff(vcpu))
2849 			return vcpu;
2850 	}
2851 	return NULL;
2852 }
2853 
2854 bool kvm_arch_irqchip_in_kernel(struct kvm *kvm)
2855 {
2856 	return irqchip_in_kernel(kvm);
2857 }
2858 
2859 int kvm_arch_irq_bypass_add_producer(struct irq_bypass_consumer *cons,
2860 				      struct irq_bypass_producer *prod)
2861 {
2862 	struct kvm_kernel_irqfd *irqfd =
2863 		container_of(cons, struct kvm_kernel_irqfd, consumer);
2864 	struct kvm_kernel_irq_routing_entry *irq_entry = &irqfd->irq_entry;
2865 
2866 	/*
2867 	 * The only thing we have a chance of directly-injecting is LPIs. Maybe
2868 	 * one day...
2869 	 */
2870 	if (irq_entry->type != KVM_IRQ_ROUTING_MSI)
2871 		return 0;
2872 
2873 	return kvm_vgic_v4_set_forwarding(irqfd->kvm, prod->irq,
2874 					  &irqfd->irq_entry);
2875 }
2876 
2877 void kvm_arch_irq_bypass_del_producer(struct irq_bypass_consumer *cons,
2878 				      struct irq_bypass_producer *prod)
2879 {
2880 	struct kvm_kernel_irqfd *irqfd =
2881 		container_of(cons, struct kvm_kernel_irqfd, consumer);
2882 	struct kvm_kernel_irq_routing_entry *irq_entry = &irqfd->irq_entry;
2883 
2884 	if (irq_entry->type != KVM_IRQ_ROUTING_MSI)
2885 		return;
2886 
2887 	kvm_vgic_v4_unset_forwarding(irqfd->kvm, prod->irq);
2888 }
2889 
2890 void kvm_arch_update_irqfd_routing(struct kvm_kernel_irqfd *irqfd,
2891 				   struct kvm_kernel_irq_routing_entry *old,
2892 				   struct kvm_kernel_irq_routing_entry *new)
2893 {
2894 	if (old->type == KVM_IRQ_ROUTING_MSI &&
2895 	    new->type == KVM_IRQ_ROUTING_MSI &&
2896 	    !memcmp(&old->msi, &new->msi, sizeof(new->msi)))
2897 		return;
2898 
2899 	/*
2900 	 * Remapping the vLPI requires taking the its_lock mutex to resolve
2901 	 * the new translation. We're in spinlock land at this point, so no
2902 	 * chance of resolving the translation.
2903 	 *
2904 	 * Unmap the vLPI and fall back to software LPI injection.
2905 	 */
2906 	return kvm_vgic_v4_unset_forwarding(irqfd->kvm, irqfd->producer->irq);
2907 }
2908 
2909 void kvm_arch_irq_bypass_stop(struct irq_bypass_consumer *cons)
2910 {
2911 	struct kvm_kernel_irqfd *irqfd =
2912 		container_of(cons, struct kvm_kernel_irqfd, consumer);
2913 
2914 	kvm_arm_halt_guest(irqfd->kvm);
2915 }
2916 
2917 void kvm_arch_irq_bypass_start(struct irq_bypass_consumer *cons)
2918 {
2919 	struct kvm_kernel_irqfd *irqfd =
2920 		container_of(cons, struct kvm_kernel_irqfd, consumer);
2921 
2922 	kvm_arm_resume_guest(irqfd->kvm);
2923 }
2924 
2925 /* Initialize Hyp-mode and memory mappings on all CPUs */
2926 static __init int kvm_arm_init(void)
2927 {
2928 	int err;
2929 	bool in_hyp_mode;
2930 
2931 	if (!is_hyp_mode_available()) {
2932 		kvm_info("HYP mode not available\n");
2933 		return -ENODEV;
2934 	}
2935 
2936 	if (kvm_get_mode() == KVM_MODE_NONE) {
2937 		kvm_info("KVM disabled from command line\n");
2938 		return -ENODEV;
2939 	}
2940 
2941 	err = kvm_sys_reg_table_init();
2942 	if (err) {
2943 		kvm_info("Error initializing system register tables");
2944 		return err;
2945 	}
2946 
2947 	in_hyp_mode = is_kernel_in_hyp_mode();
2948 
2949 	if (cpus_have_final_cap(ARM64_WORKAROUND_DEVICE_LOAD_ACQUIRE) ||
2950 	    cpus_have_final_cap(ARM64_WORKAROUND_1508412))
2951 		kvm_info("Guests without required CPU erratum workarounds can deadlock system!\n" \
2952 			 "Only trusted guests should be used on this system.\n");
2953 
2954 	err = kvm_set_ipa_limit();
2955 	if (err)
2956 		return err;
2957 
2958 	err = kvm_arm_init_sve();
2959 	if (err)
2960 		return err;
2961 
2962 	err = kvm_arm_vmid_alloc_init();
2963 	if (err) {
2964 		kvm_err("Failed to initialize VMID allocator.\n");
2965 		return err;
2966 	}
2967 
2968 	if (!in_hyp_mode) {
2969 		err = init_hyp_mode();
2970 		if (err)
2971 			goto out_err;
2972 	}
2973 
2974 	err = kvm_init_vector_slots();
2975 	if (err) {
2976 		kvm_err("Cannot initialise vector slots\n");
2977 		goto out_hyp;
2978 	}
2979 
2980 	err = init_subsystems();
2981 	if (err)
2982 		goto out_hyp;
2983 
2984 	kvm_info("%s%sVHE%s mode initialized successfully\n",
2985 		 in_hyp_mode ? "" : (is_protected_kvm_enabled() ?
2986 				     "Protected " : "Hyp "),
2987 		 in_hyp_mode ? "" : (cpus_have_final_cap(ARM64_KVM_HVHE) ?
2988 				     "h" : "n"),
2989 		 cpus_have_final_cap(ARM64_HAS_NESTED_VIRT) ? "+NV2": "");
2990 
2991 	/*
2992 	 * FIXME: Do something reasonable if kvm_init() fails after pKVM
2993 	 * hypervisor protection is finalized.
2994 	 */
2995 	err = kvm_init(sizeof(struct kvm_vcpu), 0, THIS_MODULE);
2996 	if (err)
2997 		goto out_subs;
2998 
2999 	/*
3000 	 * This should be called after initialization is done and failure isn't
3001 	 * possible anymore.
3002 	 */
3003 	if (!in_hyp_mode)
3004 		finalize_init_hyp_mode();
3005 
3006 	kvm_arm_initialised = true;
3007 
3008 	return 0;
3009 
3010 out_subs:
3011 	teardown_subsystems();
3012 out_hyp:
3013 	if (!in_hyp_mode)
3014 		teardown_hyp_mode();
3015 out_err:
3016 	kvm_arm_vmid_alloc_free();
3017 	return err;
3018 }
3019 
3020 static int __init early_kvm_mode_cfg(char *arg)
3021 {
3022 	if (!arg)
3023 		return -EINVAL;
3024 
3025 	if (strcmp(arg, "none") == 0) {
3026 		kvm_mode = KVM_MODE_NONE;
3027 		return 0;
3028 	}
3029 
3030 	if (!is_hyp_mode_available()) {
3031 		pr_warn_once("KVM is not available. Ignoring kvm-arm.mode\n");
3032 		return 0;
3033 	}
3034 
3035 	if (strcmp(arg, "protected") == 0) {
3036 		if (!is_kernel_in_hyp_mode())
3037 			kvm_mode = KVM_MODE_PROTECTED;
3038 		else
3039 			pr_warn_once("Protected KVM not available with VHE\n");
3040 
3041 		return 0;
3042 	}
3043 
3044 	if (strcmp(arg, "nvhe") == 0 && !WARN_ON(is_kernel_in_hyp_mode())) {
3045 		kvm_mode = KVM_MODE_DEFAULT;
3046 		return 0;
3047 	}
3048 
3049 	if (strcmp(arg, "nested") == 0 && !WARN_ON(!is_kernel_in_hyp_mode())) {
3050 		kvm_mode = KVM_MODE_NV;
3051 		return 0;
3052 	}
3053 
3054 	return -EINVAL;
3055 }
3056 early_param("kvm-arm.mode", early_kvm_mode_cfg);
3057 
3058 static int __init early_kvm_wfx_trap_policy_cfg(char *arg, enum kvm_wfx_trap_policy *p)
3059 {
3060 	if (!arg)
3061 		return -EINVAL;
3062 
3063 	if (strcmp(arg, "trap") == 0) {
3064 		*p = KVM_WFX_TRAP;
3065 		return 0;
3066 	}
3067 
3068 	if (strcmp(arg, "notrap") == 0) {
3069 		*p = KVM_WFX_NOTRAP;
3070 		return 0;
3071 	}
3072 
3073 	return -EINVAL;
3074 }
3075 
3076 static int __init early_kvm_wfi_trap_policy_cfg(char *arg)
3077 {
3078 	return early_kvm_wfx_trap_policy_cfg(arg, &kvm_wfi_trap_policy);
3079 }
3080 early_param("kvm-arm.wfi_trap_policy", early_kvm_wfi_trap_policy_cfg);
3081 
3082 static int __init early_kvm_wfe_trap_policy_cfg(char *arg)
3083 {
3084 	return early_kvm_wfx_trap_policy_cfg(arg, &kvm_wfe_trap_policy);
3085 }
3086 early_param("kvm-arm.wfe_trap_policy", early_kvm_wfe_trap_policy_cfg);
3087 
3088 enum kvm_mode kvm_get_mode(void)
3089 {
3090 	return kvm_mode;
3091 }
3092 
3093 module_init(kvm_arm_init);
3094