xref: /linux/arch/arm64/kvm/nested.c (revision e5843f4effaa2ffac3e789ecd4456403564961d4)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2017 - Columbia University and Linaro Ltd.
4  * Author: Jintack Lim <jintack.lim@linaro.org>
5  */
6 
7 #include <linux/bitfield.h>
8 #include <linux/kvm.h>
9 #include <linux/kvm_host.h>
10 
11 #include <asm/fixmap.h>
12 #include <asm/kvm_arm.h>
13 #include <asm/kvm_emulate.h>
14 #include <asm/kvm_mmu.h>
15 #include <asm/kvm_nested.h>
16 #include <asm/sysreg.h>
17 
18 #include "sys_regs.h"
19 #include "vgic/vgic.h"
20 
21 struct vncr_tlb {
22 	/* The guest's VNCR_EL2 */
23 	u64			gva;
24 	struct s1_walk_info	wi;
25 	struct s1_walk_result	wr;
26 
27 	u64			hpa;
28 	bool			hpa_writable;
29 
30 	/* -1 when not mapped on a CPU */
31 	atomic_t		cpu;
32 
33 	/*
34 	 * true if the TLB is valid. Can only be changed with the
35 	 * mmu_lock held.
36 	 */
37 	bool			valid;
38 };
39 
40 /*
41  * Ratio of live shadow S2 MMU per vcpu. This is a trade-off between
42  * memory usage and potential number of different sets of S2 PTs in
43  * the guests. Running out of S2 MMUs only affects performance (we
44  * will invalidate them more often).
45  */
46 #define S2_MMU_PER_VCPU		2
47 
48 int kvm_init_nested(struct kvm *kvm)
49 {
50 	kvm->arch.nested_mmus = kvmalloc_objs(struct kvm_s2_mmu *,
51 					      KVM_MAX_VCPUS * S2_MMU_PER_VCPU,
52 					      GFP_KERNEL_ACCOUNT);
53 	kvm->arch.nested_mmus_size = 0;
54 	atomic_set(&kvm->arch.vncr_tlb_count, 0);
55 
56 	return kvm->arch.nested_mmus ? 0 : -ENOMEM;
57 }
58 
59 void kvm_destroy_nested(struct kvm *kvm)
60 {
61 	for (int i = 0; i < kvm->arch.nested_mmus_size; i+= S2_MMU_PER_VCPU)
62 		kvfree(kvm->arch.nested_mmus[i]);
63 
64 	kvm->arch.nested_mmus_size = 0;
65 	kvfree(kvm->arch.nested_mmus);
66 }
67 
68 static int init_nested_s2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu)
69 {
70 	/*
71 	 * We only initialise the IPA range on the canonical MMU, which
72 	 * defines the contract between KVM and userspace on where the
73 	 * "hardware" is in the IPA space. This affects the validity of MMIO
74 	 * exits forwarded to userspace, for example.
75 	 *
76 	 * For nested S2s, we use the PARange as exposed to the guest, as it
77 	 * is allowed to use it at will to expose whatever memory map it
78 	 * wants to its own guests as it would be on real HW.
79 	 */
80 	return kvm_init_stage2_mmu(kvm, mmu, kvm_get_pa_bits(kvm));
81 }
82 
83 int kvm_vcpu_init_nested(struct kvm_vcpu *vcpu)
84 {
85 	struct kvm *kvm = vcpu->kvm;
86 	int num_mmus;
87 
88 	lockdep_assert_held(&kvm->arch.config_lock);
89 
90 	if (test_bit(KVM_ARM_VCPU_HAS_EL2_E2H0, kvm->arch.vcpu_features) &&
91 	    !cpus_have_final_cap(ARM64_HAS_HCR_NV1))
92 		return -EINVAL;
93 
94 	if (!vcpu->arch.ctxt.vncr_array)
95 		vcpu->arch.ctxt.vncr_array = (u64 *)__get_free_page(GFP_KERNEL_ACCOUNT |
96 								    __GFP_ZERO);
97 
98 	if (!vcpu->arch.ctxt.vncr_array)
99 		return -ENOMEM;
100 
101 	num_mmus = atomic_read(&kvm->online_vcpus) * S2_MMU_PER_VCPU;
102 
103 	if (num_mmus > kvm->arch.nested_mmus_size) {
104 		struct kvm_s2_mmu *tmp;
105 		int i, ret = 0;
106 
107 		tmp = kvzalloc_objs(*tmp, S2_MMU_PER_VCPU, GFP_KERNEL_ACCOUNT);
108 		if (!tmp)
109 			ret = -ENOMEM;
110 
111 		for (i = 0; !ret && i < S2_MMU_PER_VCPU; i++) {
112 			ret = init_nested_s2_mmu(kvm, &tmp[i]);
113 			if (ret)
114 				break;
115 		}
116 
117 		if (ret) {
118 			while (--i >= 0)
119 				kvm_free_stage2_pgd(&tmp[i]);
120 
121 			kvfree(tmp);
122 			free_page((unsigned long)vcpu->arch.ctxt.vncr_array);
123 			vcpu->arch.ctxt.vncr_array = NULL;
124 			return ret;
125 		}
126 
127 		guard(write_lock)(&kvm->mmu_lock);
128 
129 		for (i = 0; i < S2_MMU_PER_VCPU; i++)
130 			kvm->arch.nested_mmus[i + kvm->arch.nested_mmus_size] = &tmp[i];
131 
132 		kvm->arch.nested_mmus_size += S2_MMU_PER_VCPU;
133 	}
134 
135 	return 0;
136 }
137 
138 struct s2_walk_info {
139 	u64		baddr;
140 	unsigned int	max_oa_bits;
141 	unsigned int	pgshift;
142 	unsigned int	sl;
143 	unsigned int	t0sz;
144 	bool		be;
145 	bool		ha;
146 };
147 
148 static u32 compute_fsc(int level, u32 fsc)
149 {
150 	return fsc | (level & 0x3);
151 }
152 
153 static int esr_s2_fault(struct kvm_vcpu *vcpu, int level, u32 fsc)
154 {
155 	u32 esr;
156 
157 	esr = kvm_vcpu_get_esr(vcpu) & ~ESR_ELx_FSC;
158 	esr |= compute_fsc(level, fsc);
159 	return esr;
160 }
161 
162 static int get_ia_size(struct s2_walk_info *wi)
163 {
164 	return 64 - wi->t0sz;
165 }
166 
167 static int check_base_s2_limits(struct kvm_vcpu *vcpu, struct s2_walk_info *wi,
168 				int level, int input_size, int stride)
169 {
170 	int start_size, pa_max;
171 
172 	pa_max = kvm_get_pa_bits(vcpu->kvm);
173 
174 	/* Check translation limits */
175 	switch (BIT(wi->pgshift)) {
176 	case SZ_64K:
177 		if (level == 0 || (level == 1 && pa_max <= 42))
178 			return -EFAULT;
179 		break;
180 	case SZ_16K:
181 		if (level == 0 || (level == 1 && pa_max <= 40))
182 			return -EFAULT;
183 		break;
184 	case SZ_4K:
185 		if (level < 0 || (level == 0 && pa_max <= 42))
186 			return -EFAULT;
187 		break;
188 	}
189 
190 	/* Check input size limits */
191 	if (input_size > pa_max)
192 		return -EFAULT;
193 
194 	/* Check number of entries in starting level table */
195 	start_size = input_size - ((3 - level) * stride + wi->pgshift);
196 	if (start_size < 1 || start_size > stride + 4)
197 		return -EFAULT;
198 
199 	return 0;
200 }
201 
202 /* Check if output is within boundaries */
203 static int check_output_size(struct s2_walk_info *wi, phys_addr_t output)
204 {
205 	unsigned int output_size = wi->max_oa_bits;
206 
207 	if (output_size != 48 && (output & GENMASK_ULL(47, output_size)))
208 		return -1;
209 
210 	return 0;
211 }
212 
213 static int read_guest_s2_desc(struct kvm_vcpu *vcpu, phys_addr_t pa, u64 *desc,
214 			      struct s2_walk_info *wi)
215 {
216 	u64 val;
217 	int r;
218 
219 	r = kvm_read_guest(vcpu->kvm, pa, &val, sizeof(val));
220 	if (r)
221 		return r;
222 
223 	/*
224 	 * Handle reversedescriptors if endianness differs between the
225 	 * host and the guest hypervisor.
226 	 */
227 	if (wi->be)
228 		*desc = be64_to_cpu((__force __be64)val);
229 	else
230 		*desc = le64_to_cpu((__force __le64)val);
231 
232 	return 0;
233 }
234 
235 static int swap_guest_s2_desc(struct kvm_vcpu *vcpu, phys_addr_t pa, u64 old, u64 new,
236 			      struct s2_walk_info *wi)
237 {
238 	if (wi->be) {
239 		old = (__force u64)cpu_to_be64(old);
240 		new = (__force u64)cpu_to_be64(new);
241 	} else {
242 		old = (__force u64)cpu_to_le64(old);
243 		new = (__force u64)cpu_to_le64(new);
244 	}
245 
246 	return __kvm_at_swap_desc(vcpu->kvm, pa, old, new);
247 }
248 
249 /*
250  * This is essentially a C-version of the pseudo code from the ARM ARM
251  * AArch64.TranslationTableWalk  function.  I strongly recommend looking at
252  * that pseudocode in trying to understand this.
253  *
254  * Must be called with the kvm->srcu read lock held
255  */
256 static int walk_nested_s2_pgd(struct kvm_vcpu *vcpu, phys_addr_t ipa,
257 			      struct s2_walk_info *wi, struct kvm_s2_trans *out)
258 {
259 	int first_block_level, level, stride, input_size, base_lower_bound;
260 	phys_addr_t base_addr;
261 	unsigned int addr_top, addr_bottom;
262 	u64 desc, new_desc;  /* page table entry */
263 	int ret;
264 	phys_addr_t paddr;
265 
266 	switch (BIT(wi->pgshift)) {
267 	default:
268 	case SZ_64K:
269 	case SZ_16K:
270 		level = 3 - wi->sl;
271 		first_block_level = 2;
272 		break;
273 	case SZ_4K:
274 		level = 2 - wi->sl;
275 		first_block_level = 1;
276 		break;
277 	}
278 
279 	stride = wi->pgshift - 3;
280 	input_size = get_ia_size(wi);
281 	if (input_size > 48 || input_size < 25)
282 		return -EFAULT;
283 
284 	ret = check_base_s2_limits(vcpu, wi, level, input_size, stride);
285 	if (WARN_ON(ret)) {
286 		out->esr = compute_fsc(0, ESR_ELx_FSC_FAULT);
287 		return ret;
288 	}
289 
290 	base_lower_bound = 3 + input_size - ((3 - level) * stride +
291 			   wi->pgshift);
292 	base_addr = wi->baddr & GENMASK_ULL(47, base_lower_bound);
293 
294 	if (check_output_size(wi, base_addr)) {
295 		/* R_BFHQH */
296 		out->esr = compute_fsc(0, ESR_ELx_FSC_ADDRSZ);
297 		return 1;
298 	}
299 
300 	addr_top = input_size - 1;
301 
302 	while (1) {
303 		phys_addr_t index;
304 
305 		addr_bottom = (3 - level) * stride + wi->pgshift;
306 		index = (ipa & GENMASK_ULL(addr_top, addr_bottom))
307 			>> (addr_bottom - 3);
308 
309 		paddr = base_addr | index;
310 		ret = read_guest_s2_desc(vcpu, paddr, &desc, wi);
311 		if (ret < 0) {
312 			out->esr = ESR_ELx_FSC_SEA_TTW(level);
313 			return ret;
314 		}
315 
316 		new_desc = desc;
317 
318 		/* Check for valid descriptor at this point */
319 		if (!(desc & KVM_PTE_VALID)) {
320 			out->esr = compute_fsc(level, ESR_ELx_FSC_FAULT);
321 			out->desc = desc;
322 			return 1;
323 		}
324 
325 		if (FIELD_GET(KVM_PTE_TYPE, desc) == KVM_PTE_TYPE_BLOCK) {
326 			if (level < 3)
327 				break;
328 
329 			out->esr = compute_fsc(level, ESR_ELx_FSC_FAULT);
330 			out->desc = desc;
331 			return 1;
332 		}
333 
334 		/* We're at the final level */
335 		if (level == 3)
336 			break;
337 
338 		if (check_output_size(wi, desc)) {
339 			out->esr = compute_fsc(level, ESR_ELx_FSC_ADDRSZ);
340 			out->desc = desc;
341 			return 1;
342 		}
343 
344 		base_addr = desc & GENMASK_ULL(47, wi->pgshift);
345 
346 		level += 1;
347 		addr_top = addr_bottom - 1;
348 	}
349 
350 	if (level < first_block_level) {
351 		out->esr = compute_fsc(level, ESR_ELx_FSC_FAULT);
352 		out->desc = desc;
353 		return 1;
354 	}
355 
356 	if (check_output_size(wi, desc)) {
357 		out->esr = compute_fsc(level, ESR_ELx_FSC_ADDRSZ);
358 		out->desc = desc;
359 		return 1;
360 	}
361 
362 	if (wi->ha)
363 		new_desc |= KVM_PTE_LEAF_ATTR_LO_S2_AF;
364 
365 	if (new_desc != desc) {
366 		ret = swap_guest_s2_desc(vcpu, paddr, desc, new_desc, wi);
367 		if (ret == -EAGAIN)
368 			return ret;
369 		if (ret) {
370 			out->esr = ESR_ELx_FSC_SEA_TTW(level);
371 			out->desc = desc;
372 			return 1;
373 		}
374 
375 		desc = new_desc;
376 	}
377 
378 	if (!(desc & KVM_PTE_LEAF_ATTR_LO_S2_AF)) {
379 		out->esr = compute_fsc(level, ESR_ELx_FSC_ACCESS);
380 		out->desc = desc;
381 		return 1;
382 	}
383 
384 	addr_bottom += contiguous_bit_shift(desc, wi, level);
385 
386 	/* Calculate and return the result */
387 	paddr = (desc & GENMASK_ULL(47, addr_bottom)) |
388 		(ipa & GENMASK_ULL(addr_bottom - 1, 0));
389 	out->output = paddr;
390 	out->block_size = 1UL << ((3 - level) * stride + wi->pgshift);
391 	out->readable = desc & KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
392 	out->writable = desc & KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
393 	out->level = level;
394 	out->desc = desc;
395 	return 0;
396 }
397 
398 #define _has_tgran_2(__r, __sz)						\
399 	({								\
400 		u64 _s1, _s2, _mmfr0 = __r;				\
401 									\
402 		_s2 = SYS_FIELD_GET(ID_AA64MMFR0_EL1,			\
403 				    TGRAN##__sz##_2, _mmfr0);		\
404 									\
405 		_s1 = SYS_FIELD_GET(ID_AA64MMFR0_EL1,			\
406 				    TGRAN##__sz, _mmfr0);		\
407 									\
408 		((_s2 != ID_AA64MMFR0_EL1_TGRAN##__sz##_2_NI &&		\
409 		  _s2 != ID_AA64MMFR0_EL1_TGRAN##__sz##_2_TGRAN##__sz) || \
410 		 (_s2 == ID_AA64MMFR0_EL1_TGRAN##__sz##_2_TGRAN##__sz && \
411 		  _s1 != ID_AA64MMFR0_EL1_TGRAN##__sz##_NI));		\
412 	})
413 
414 static bool has_tgran_2(u64 mmfr0, unsigned int shift)
415 {
416 	switch (shift) {
417 	case 12:
418 		return _has_tgran_2(mmfr0, 4);
419 	case 14:
420 		return _has_tgran_2(mmfr0, 16);
421 	case 16:
422 		return _has_tgran_2(mmfr0, 64);
423 	default:
424 		BUG();
425 	}
426 }
427 
428 static unsigned int fallback_tgran2_shift(u64 mmfr0)
429 {
430 	if (has_tgran_2(mmfr0, PAGE_SHIFT))
431 		return PAGE_SHIFT;
432 	else if (has_tgran_2(mmfr0, 12))
433 		return 12;
434 	else if (has_tgran_2(mmfr0, 14))
435 		return 14;
436 	else if (has_tgran_2(mmfr0, 16))
437 		return 16;
438 	else
439 		return PAGE_SHIFT;
440 }
441 
442 static unsigned int vtcr_to_tg0_pgshift(struct kvm *kvm, u64 vtcr)
443 {
444 	u64 tg0 = FIELD_GET(VTCR_EL2_TG0_MASK, vtcr);
445 	u64 mmfr0 = kvm_read_vm_id_reg(kvm, SYS_ID_AA64MMFR0_EL1);
446 	unsigned int shift;
447 
448 	switch (tg0) {
449 	case VTCR_EL2_TG0_4K:
450 		shift = 12;
451 		break;
452 	case VTCR_EL2_TG0_16K:
453 		shift = 14;
454 		break;
455 	case VTCR_EL2_TG0_64K:
456 	/* IMPDEF: treat any other value as 64k, subject to fallback */
457 	default:
458 		shift = 16;
459 	}
460 
461 	/*
462 	 * If TGx is programmed to an unimplemented value (not advertised in
463 	 * ID_AA64MMFR0_EL1), we should treat it as if an implemented value is
464 	 * written, as per the architecture. Choose an available one while
465 	 * prioritizing PAGE_SIZE.
466 	 */
467 	if (!has_tgran_2(mmfr0, shift))
468 		return fallback_tgran2_shift(mmfr0);
469 
470 	return shift;
471 }
472 
473 static size_t vtcr_to_tg0_pgsize(struct kvm *kvm, u64 vtcr)
474 {
475 	return BIT(vtcr_to_tg0_pgshift(kvm, vtcr));
476 }
477 
478 static void setup_s2_walk(struct kvm_vcpu *vcpu, struct s2_walk_info *wi)
479 {
480 	u64 vtcr = vcpu_read_sys_reg(vcpu, VTCR_EL2);
481 
482 	wi->baddr = vcpu_read_sys_reg(vcpu, VTTBR_EL2);
483 	wi->t0sz = vtcr & VTCR_EL2_T0SZ_MASK;
484 	wi->pgshift = vtcr_to_tg0_pgshift(vcpu->kvm, vtcr);
485 	wi->sl = FIELD_GET(VTCR_EL2_SL0_MASK, vtcr);
486 	/* Global limit for now, should eventually be per-VM */
487 	wi->max_oa_bits = min(get_kvm_ipa_limit(),
488 			      ps_to_output_size(FIELD_GET(VTCR_EL2_PS_MASK, vtcr), false));
489 	wi->ha = vtcr & VTCR_EL2_HA;
490 	wi->be = vcpu_read_sys_reg(vcpu, SCTLR_EL2) & SCTLR_ELx_EE;
491 }
492 
493 int kvm_walk_nested_s2(struct kvm_vcpu *vcpu, phys_addr_t gipa,
494 		       struct kvm_s2_trans *result)
495 {
496 	struct s2_walk_info wi;
497 	int ret;
498 
499 	result->esr = 0;
500 
501 	if (!vcpu_has_nv(vcpu))
502 		return 0;
503 
504 	setup_s2_walk(vcpu, &wi);
505 
506 	ret = walk_nested_s2_pgd(vcpu, gipa, &wi, result);
507 	if (ret)
508 		result->esr |= (kvm_vcpu_get_esr(vcpu) & ~ESR_ELx_FSC);
509 
510 	return ret;
511 }
512 
513 static unsigned int __ttl_to_size(u8 ttl)
514 {
515 	int level = ttl & 3;
516 	int gran = (ttl >> 2) & 3;
517 	unsigned int max_size = 0;
518 
519 	switch (gran) {
520 	case TLBI_TTL_TG_4K:
521 		switch (level) {
522 		case 0:
523 			break;
524 		case 1:
525 			max_size = SZ_1G;
526 			break;
527 		case 2:
528 			max_size = SZ_2M;
529 			break;
530 		case 3:
531 			max_size = SZ_4K;
532 			break;
533 		}
534 		break;
535 	case TLBI_TTL_TG_16K:
536 		switch (level) {
537 		case 0:
538 		case 1:
539 			break;
540 		case 2:
541 			max_size = SZ_32M;
542 			break;
543 		case 3:
544 			max_size = SZ_16K;
545 			break;
546 		}
547 		break;
548 	case TLBI_TTL_TG_64K:
549 		switch (level) {
550 		case 0:
551 		case 1:
552 			/* No 52bit IPA support */
553 			break;
554 		case 2:
555 			max_size = SZ_512M;
556 			break;
557 		case 3:
558 			max_size = SZ_64K;
559 			break;
560 		}
561 		break;
562 	default:			/* No size information */
563 		break;
564 	}
565 
566 	return max_size;
567 }
568 
569 static unsigned int ttl_to_size(u8 ttl)
570 {
571 	return __ttl_to_size(ttl) ?: SZ_1G;
572 }
573 
574 static u8 pgshift_level_to_ttl(u16 shift, s8 level)
575 {
576 	u8 ttl;
577 
578 	/*
579 	 * If we don't have a proper level, fallback to the maximum
580 	 * size.
581 	 */
582 	if (level < 0)
583 		return 0;
584 
585 	switch(shift) {
586 	case 12:
587 		ttl = TLBI_TTL_TG_4K;
588 		break;
589 	case 14:
590 		ttl = TLBI_TTL_TG_16K;
591 		break;
592 	case 16:
593 		ttl = TLBI_TTL_TG_64K;
594 		break;
595 	default:
596 		BUG();
597 	}
598 
599 	ttl <<= 2;
600 	ttl |= level & 3;
601 
602 	return ttl;
603 }
604 
605 /*
606  * Compute the equivalent of the TTL field by parsing the shadow PT.  The
607  * granule size is extracted from the cached VTCR_EL2.TG0 while the level is
608  * retrieved from first entry carrying the level as a tag.
609  */
610 static u8 get_guest_mapping_ttl(struct kvm_s2_mmu *mmu, u64 addr)
611 {
612 	size_t tg0_size = vtcr_to_tg0_pgsize(kvm_s2_mmu_to_kvm(mmu), mmu->tlb_vtcr);
613 	u64 tmp, sz = 0;
614 	kvm_pte_t pte;
615 	u8 ttl, level;
616 
617 	lockdep_assert_held_write(&kvm_s2_mmu_to_kvm(mmu)->mmu_lock);
618 
619 	switch (tg0_size) {
620 	case SZ_4K:
621 		ttl = (TLBI_TTL_TG_4K << 2);
622 		break;
623 	case SZ_16K:
624 		ttl = (TLBI_TTL_TG_16K << 2);
625 		break;
626 	case SZ_64K:
627 	default:	    /* IMPDEF: treat any other value as 64k */
628 		ttl = (TLBI_TTL_TG_64K << 2);
629 		break;
630 	}
631 
632 	tmp = addr;
633 
634 again:
635 	/* Iteratively compute the block sizes for a particular granule size */
636 	switch (tg0_size) {
637 	case SZ_4K:
638 		if	(sz < SZ_4K)	sz = SZ_4K;
639 		else if (sz < SZ_2M)	sz = SZ_2M;
640 		else if (sz < SZ_1G)	sz = SZ_1G;
641 		else			sz = 0;
642 		break;
643 	case SZ_16K:
644 		if	(sz < SZ_16K)	sz = SZ_16K;
645 		else if (sz < SZ_32M)	sz = SZ_32M;
646 		else			sz = 0;
647 		break;
648 	case SZ_64K:
649 	default:	    /* IMPDEF: treat any other value as 64k */
650 		if	(sz < SZ_64K)	sz = SZ_64K;
651 		else if (sz < SZ_512M)	sz = SZ_512M;
652 		else			sz = 0;
653 		break;
654 	}
655 
656 	if (sz == 0)
657 		return 0;
658 
659 	tmp &= ~(sz - 1);
660 	if (kvm_pgtable_get_leaf(mmu->pgt, tmp, &pte, NULL))
661 		goto again;
662 	if (!(pte & PTE_VALID))
663 		goto again;
664 	level = FIELD_GET(KVM_NV_GUEST_MAP_SZ, pte);
665 	if (!level)
666 		goto again;
667 
668 	ttl |= level;
669 
670 	/*
671 	 * We now have found some level information in the shadow S2. Check
672 	 * that the resulting range is actually including the original IPA.
673 	 */
674 	sz = ttl_to_size(ttl);
675 	if (addr < (tmp + sz))
676 		return ttl;
677 
678 	return 0;
679 }
680 
681 unsigned long compute_tlb_inval_range(struct kvm_s2_mmu *mmu, u64 val)
682 {
683 	struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
684 	unsigned long max_size;
685 	u8 ttl;
686 
687 	ttl = FIELD_GET(TLBI_TTL_MASK, val);
688 
689 	if (!ttl || !kvm_has_feat(kvm, ID_AA64MMFR2_EL1, TTL, IMP)) {
690 		/* No TTL, check the shadow S2 for a hint */
691 		u64 addr = (val & GENMASK_ULL(35, 0)) << 12;
692 		ttl = get_guest_mapping_ttl(mmu, addr);
693 	}
694 
695 	/*
696 	 * Don't use the default 1GB fallback, as we can adapt to the
697 	 * max mapping size we allow at S2.
698 	 */
699 	max_size = __ttl_to_size(ttl);
700 
701 	if (!max_size) {
702 		/* Compute the maximum extent of the invalidation */
703 		switch (vtcr_to_tg0_pgsize(kvm, mmu->tlb_vtcr)) {
704 		case SZ_4K:
705 			max_size = SZ_1G;
706 			break;
707 		case SZ_16K:
708 			max_size = SZ_32M;
709 			break;
710 		case SZ_64K:
711 		default:    /* IMPDEF: treat any other value as 64k */
712 			/*
713 			 * No, we do not support 52bit IPA in nested yet. Once
714 			 * we do, this should be 4TB.
715 			 */
716 			max_size = SZ_512M;
717 			break;
718 		}
719 	}
720 
721 	WARN_ON(!max_size);
722 	return max_size;
723 }
724 
725 /*
726  * We can have multiple *different* MMU contexts with the same VMID:
727  *
728  * - S2 being enabled or not, hence differing by the HCR_EL2.VM bit
729  *
730  * - Multiple vcpus using private S2s (huh huh...), hence differing by the
731  *   VBBTR_EL2.BADDR address
732  *
733  * - A combination of the above...
734  *
735  * We can always identify which MMU context to pick at run-time.  However,
736  * TLB invalidation involving a VMID must take action on all the TLBs using
737  * this particular VMID. This translates into applying the same invalidation
738  * operation to all the contexts that are using this VMID. Moar phun!
739  */
740 void kvm_s2_mmu_iterate_by_vmid(struct kvm *kvm, u16 vmid,
741 				const union tlbi_info *info,
742 				void (*tlbi_callback)(struct kvm_s2_mmu *,
743 						      const union tlbi_info *))
744 {
745 	write_lock(&kvm->mmu_lock);
746 
747 	for (int i = 0; i < kvm->arch.nested_mmus_size; i++) {
748 		struct kvm_s2_mmu *mmu = kvm->arch.nested_mmus[i];
749 
750 		if (!kvm_s2_mmu_valid(mmu))
751 			continue;
752 
753 		if (vmid == get_vmid(mmu->tlb_vttbr))
754 			tlbi_callback(mmu, info);
755 	}
756 
757 	write_unlock(&kvm->mmu_lock);
758 }
759 
760 struct kvm_s2_mmu *lookup_s2_mmu(struct kvm_vcpu *vcpu)
761 {
762 	struct kvm *kvm = vcpu->kvm;
763 	bool nested_stage2_enabled;
764 	u64 vttbr, vtcr, hcr;
765 
766 	lockdep_assert_held_write(&kvm->mmu_lock);
767 
768 	vttbr = vcpu_read_sys_reg(vcpu, VTTBR_EL2);
769 	vtcr = vcpu_read_sys_reg(vcpu, VTCR_EL2);
770 	hcr = vcpu_read_sys_reg(vcpu, HCR_EL2);
771 
772 	nested_stage2_enabled = hcr & HCR_VM;
773 
774 	/* Don't consider the CnP bit for the vttbr match */
775 	vttbr &= ~VTTBR_CNP_BIT;
776 
777 	/*
778 	 * Two possibilities when looking up a S2 MMU context:
779 	 *
780 	 * - either S2 is enabled in the guest, and we need a context that is
781 	 *   S2-enabled and matches the full VTTBR (VMID+BADDR) and VTCR,
782 	 *   which makes it safe from a TLB conflict perspective (a broken
783 	 *   guest won't be able to generate them),
784 	 *
785 	 * - or S2 is disabled, and we need a context that is S2-disabled
786 	 *   and matches the VMID only, as all TLBs are tagged by VMID even
787 	 *   if S2 translation is disabled.
788 	 */
789 	for (int i = 0; i < kvm->arch.nested_mmus_size; i++) {
790 		struct kvm_s2_mmu *mmu = kvm->arch.nested_mmus[i];
791 
792 		if (!kvm_s2_mmu_valid(mmu))
793 			continue;
794 
795 		if (nested_stage2_enabled &&
796 		    mmu->nested_stage2_enabled &&
797 		    vttbr == mmu->tlb_vttbr &&
798 		    vtcr == mmu->tlb_vtcr)
799 			return mmu;
800 
801 		if (!nested_stage2_enabled &&
802 		    !mmu->nested_stage2_enabled &&
803 		    get_vmid(vttbr) == get_vmid(mmu->tlb_vttbr))
804 			return mmu;
805 	}
806 	return NULL;
807 }
808 
809 static struct kvm_s2_mmu *get_s2_mmu_nested(struct kvm_vcpu *vcpu)
810 {
811 	struct kvm *kvm = vcpu->kvm;
812 	struct kvm_s2_mmu *s2_mmu;
813 	int i;
814 
815 	lockdep_assert_held_write(&vcpu->kvm->mmu_lock);
816 
817 	s2_mmu = lookup_s2_mmu(vcpu);
818 	if (s2_mmu)
819 		goto out;
820 
821 	/*
822 	 * Make sure we don't always search from the same point, or we
823 	 * will always reuse a potentially active context, leaving
824 	 * free contexts unused.
825 	 */
826 	for (i = kvm->arch.nested_mmus_next;
827 	     i < (kvm->arch.nested_mmus_size + kvm->arch.nested_mmus_next);
828 	     i++) {
829 		s2_mmu = kvm->arch.nested_mmus[i % kvm->arch.nested_mmus_size];
830 
831 		if (atomic_read(&s2_mmu->refcnt) == 0)
832 			break;
833 	}
834 	BUG_ON(atomic_read(&s2_mmu->refcnt)); /* We have struct MMUs to spare */
835 
836 	/* Set the scene for the next search */
837 	kvm->arch.nested_mmus_next = (i + 1) % kvm->arch.nested_mmus_size;
838 
839 	/* Make sure we don't forget to do the laundry */
840 	if (kvm_s2_mmu_valid(s2_mmu)) {
841 		kvm_nested_s2_ptdump_remove_debugfs(s2_mmu);
842 		s2_mmu->pending_unmap = true;
843 	}
844 
845 	/*
846 	 * The virtual VMID (modulo CnP) will be used as a key when matching
847 	 * an existing kvm_s2_mmu.
848 	 *
849 	 * We cache VTCR at allocation time, once and for all. It'd be great
850 	 * if the guest didn't screw that one up, as this is not very
851 	 * forgiving...
852 	 */
853 	s2_mmu->tlb_vttbr = vcpu_read_sys_reg(vcpu, VTTBR_EL2) & ~VTTBR_CNP_BIT;
854 	s2_mmu->tlb_vtcr = vcpu_read_sys_reg(vcpu, VTCR_EL2);
855 	s2_mmu->nested_stage2_enabled = vcpu_read_sys_reg(vcpu, HCR_EL2) & HCR_VM;
856 
857 	kvm_nested_s2_ptdump_create_debugfs(s2_mmu);
858 
859 out:
860 	atomic_inc(&s2_mmu->refcnt);
861 
862 	/*
863 	 * Set the vCPU request to perform an unmap, even if the pending unmap
864 	 * originates from another vCPU. This guarantees that the MMU has been
865 	 * completely unmapped before any vCPU actually uses it, and allows
866 	 * multiple vCPUs to lend a hand with completing the unmap.
867 	 */
868 	if (s2_mmu->pending_unmap)
869 		kvm_make_request(KVM_REQ_NESTED_S2_UNMAP, vcpu);
870 
871 	return s2_mmu;
872 }
873 
874 void kvm_init_nested_s2_mmu(struct kvm_s2_mmu *mmu)
875 {
876 	/* CnP being set denotes an invalid entry */
877 	mmu->tlb_vttbr = VTTBR_CNP_BIT;
878 	mmu->nested_stage2_enabled = false;
879 	atomic_set(&mmu->refcnt, 0);
880 }
881 
882 void kvm_vcpu_load_hw_mmu(struct kvm_vcpu *vcpu)
883 {
884 	/*
885 	 * If the vCPU kept its reference on the MMU after the last put,
886 	 * keep rolling with it.
887 	 */
888 	if (is_hyp_ctxt(vcpu)) {
889 		if (!vcpu->arch.hw_mmu)
890 			vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
891 	} else {
892 		if (!vcpu->arch.hw_mmu) {
893 			scoped_guard(write_lock, &vcpu->kvm->mmu_lock)
894 				vcpu->arch.hw_mmu = get_s2_mmu_nested(vcpu);
895 		}
896 
897 		if (__vcpu_sys_reg(vcpu, HCR_EL2) & HCR_NV)
898 			kvm_make_request(KVM_REQ_MAP_L1_VNCR_EL2, vcpu);
899 	}
900 }
901 
902 /*
903  * Unmapping an L1 VNCR can happen concurrently without the mmu lock being
904  * effective (vcpu_put() vs TLBI handling). The atomic_xchg below ensures
905  * that only one CPU sets it to -1 while getting a valid CPU number back.
906  */
907 static int unmap_l1_vncr(struct vncr_tlb *vt)
908 {
909 	int cpu = atomic_xchg_relaxed(&vt->cpu, -1);
910 
911 	if (cpu != -1)
912 		clear_fixmap(vncr_fixmap(cpu));
913 
914 	return cpu;
915 }
916 
917 static void this_cpu_reset_vncr_fixmap(struct kvm_vcpu *vcpu)
918 {
919 	if (!host_data_test_flag(L1_VNCR_MAPPED))
920 		return;
921 
922 	BUG_ON(is_hyp_ctxt(vcpu));
923 
924 	/*
925 	 * Unconditionally unmap the local VNCR if we have lost the race
926 	 * against a concurrent TLBI. Otherwise we could end-up running
927 	 * another vcpu with VNCR still mapped if the TLBI thread is
928 	 * preempted between the exchange and the clear_fixmap().
929 	 *
930 	 * Note that we do not care about the TLBI nuking the fixmap behind
931 	 * the back of an running vcpu. This will only generate a fault and
932 	 * possibly a retranslation.
933 	 */
934 	if (unmap_l1_vncr(vcpu->arch.vncr_tlb) == -1)
935 		clear_fixmap(vncr_fixmap(smp_processor_id()));
936 	host_data_clear_flag(L1_VNCR_MAPPED);
937 }
938 
939 void kvm_vcpu_put_hw_mmu(struct kvm_vcpu *vcpu)
940 {
941 	/* Unconditionally drop the VNCR mapping if we have one */
942 	this_cpu_reset_vncr_fixmap(vcpu);
943 
944 	/*
945 	 * Keep a reference on the associated stage-2 MMU if the vCPU is
946 	 * scheduling out and not in WFI emulation, suggesting it is likely to
947 	 * reuse the MMU sometime soon.
948 	 */
949 	if (vcpu->scheduled_out && !vcpu_get_flag(vcpu, IN_WFI))
950 		return;
951 
952 	if (kvm_is_nested_s2_mmu(vcpu->kvm, vcpu->arch.hw_mmu))
953 		atomic_dec(&vcpu->arch.hw_mmu->refcnt);
954 
955 	vcpu->arch.hw_mmu = NULL;
956 }
957 
958 /*
959  * Returns non-zero if permission fault is handled by injecting it to the next
960  * level hypervisor.
961  */
962 int kvm_s2_handle_perm_fault(struct kvm_vcpu *vcpu, struct kvm_s2_trans *trans)
963 {
964 	bool forward_fault = false;
965 
966 	trans->esr = 0;
967 
968 	if (!kvm_vcpu_trap_is_permission_fault(vcpu))
969 		return 0;
970 
971 	if (kvm_vcpu_trap_is_iabt(vcpu)) {
972 		if (vcpu_mode_priv(vcpu))
973 			forward_fault = !kvm_s2_trans_exec_el1(vcpu->kvm, trans);
974 		else
975 			forward_fault = !kvm_s2_trans_exec_el0(vcpu->kvm, trans);
976 	} else {
977 		bool write_fault = kvm_is_write_fault(vcpu);
978 
979 		forward_fault = ((write_fault && !trans->writable) ||
980 				 (!write_fault && !trans->readable));
981 	}
982 
983 	if (forward_fault)
984 		trans->esr = esr_s2_fault(vcpu, trans->level, ESR_ELx_FSC_PERM);
985 
986 	return forward_fault;
987 }
988 
989 int kvm_inject_s2_fault(struct kvm_vcpu *vcpu, u64 esr_el2)
990 {
991 	vcpu_write_sys_reg(vcpu, vcpu->arch.fault.far_el2, FAR_EL2);
992 	vcpu_write_sys_reg(vcpu, vcpu->arch.fault.hpfar_el2, HPFAR_EL2);
993 
994 	return kvm_inject_nested_sync(vcpu, esr_el2);
995 }
996 
997 u16 get_asid_by_regime(struct kvm_vcpu *vcpu, enum trans_regime regime)
998 {
999 	enum vcpu_sysreg ttbr_elx;
1000 	u64 tcr;
1001 	u16 asid;
1002 
1003 	switch (regime) {
1004 	case TR_EL10:
1005 		tcr = vcpu_read_sys_reg(vcpu, TCR_EL1);
1006 		ttbr_elx = (tcr & TCR_A1) ? TTBR1_EL1 : TTBR0_EL1;
1007 		break;
1008 	case TR_EL20:
1009 		tcr = vcpu_read_sys_reg(vcpu, TCR_EL2);
1010 		ttbr_elx = (tcr & TCR_A1) ? TTBR1_EL2 : TTBR0_EL2;
1011 		break;
1012 	default:
1013 		BUG();
1014 	}
1015 
1016 	asid = FIELD_GET(TTBRx_EL1_ASID, vcpu_read_sys_reg(vcpu, ttbr_elx));
1017 	if (!kvm_has_feat_enum(vcpu->kvm, ID_AA64MMFR0_EL1, ASIDBITS, 16) ||
1018 	    !(tcr & TCR_ASID16))
1019 		asid &= GENMASK(7, 0);
1020 
1021 	return asid;
1022 }
1023 
1024 static void invalidate_vncr(struct kvm *kvm, struct vncr_tlb *vt)
1025 {
1026 	BUG_ON(!vt->valid);
1027 	vt->valid = false;
1028 	unmap_l1_vncr(vt);
1029 	atomic_dec(&kvm->arch.vncr_tlb_count);
1030 }
1031 
1032 static bool vncr_tlb_intersects(struct vncr_tlb *vt, u64 addr,
1033 				u64 scope_start, u64 scope_size)
1034 {
1035 	u64 tlb_size, tlb_start, tlb_end, scope_end;
1036 
1037 	tlb_size = ttl_to_size(pgshift_level_to_ttl(vt->wi.pgshift, vt->wr.level));
1038 
1039 	tlb_start = addr & ~(tlb_size - 1);
1040 	tlb_end = tlb_start + tlb_size - 1;
1041 	scope_end = scope_start + scope_size - 1;
1042 
1043 	return !(tlb_end < scope_start || tlb_start > scope_end);
1044 }
1045 
1046 /*
1047  * VNCR TLB invalidation occurs from MMU notifiers or TLBI instructions, and
1048  * either can race against a vcpu not being onlined yet (no pseudo-TLB
1049  * allocated). Similarly, the TLB might be invalid.  Skip those, as they
1050  * obviously don't participate in the invalidation at this stage.
1051  */
1052 #define kvm_for_each_vncr_tlb(idx, vcpup, tlbp, kvm)	\
1053 	kvm_for_each_vcpu(idx, vcpup, kvm)		\
1054 		if (((tlbp) = vcpup->arch.vncr_tlb) &&	\
1055 		    (tlbp)->valid)
1056 
1057 static void kvm_invalidate_vncr_ipa(struct kvm *kvm, u64 start, u64 end)
1058 {
1059 	struct kvm_vcpu *vcpu;
1060 	struct vncr_tlb *vt;
1061 	unsigned long i;
1062 
1063 	lockdep_assert_held_write(&kvm->mmu_lock);
1064 
1065 	if (!kvm_has_feat(kvm, ID_AA64MMFR4_EL1, NV_frac, NV2_ONLY))
1066 		return;
1067 
1068 	/*
1069 	 * Note that invalidating the VNCR on the back of an MMU notifier
1070 	 * doesn't require messing with the invalidation counter for a
1071 	 * parallel walk. The notifier itself will have bumped the counter,
1072 	 * making sure we rewalk.
1073 	 */
1074 	kvm_for_each_vncr_tlb(i, vcpu, vt, kvm)
1075 		if (vncr_tlb_intersects(vt, vt->wr.pa, start, end - start))
1076 			invalidate_vncr(kvm, vt);
1077 }
1078 
1079 struct s1e2_tlbi_scope {
1080 	enum {
1081 		TLBI_ALL,
1082 		TLBI_VA,
1083 		TLBI_VAA,
1084 		TLBI_ASID,
1085 	} type;
1086 
1087 	u16 asid;
1088 	u64 va;
1089 	u64 size;
1090 };
1091 
1092 static void invalidate_vncr_va(struct kvm *kvm,
1093 			       struct s1e2_tlbi_scope *scope)
1094 {
1095 	struct kvm_vcpu *vcpu;
1096 	struct vncr_tlb *vt;
1097 	unsigned long i;
1098 
1099 	lockdep_assert_held_write(&kvm->mmu_lock);
1100 
1101 	/*
1102 	 * We might be performing a parallel S1 walk, so bump up the
1103 	 * invalidation counter even in the absence of an actual VNCR TLB
1104 	 * invalidation, as this could indicate that the guest has gone
1105 	 * through a BBM sequence.
1106 	 */
1107 	kvm->mmu_invalidate_seq++;
1108 	smp_wmb();
1109 
1110 	kvm_for_each_vncr_tlb(i, vcpu, vt, kvm) {
1111 		switch (scope->type) {
1112 		case TLBI_ALL:
1113 			break;
1114 
1115 		case TLBI_VA:
1116 			if (!vncr_tlb_intersects(vt, vt->gva, scope->va, scope->size))
1117 				continue;
1118 			if (vt->wr.nG && vt->wr.asid != scope->asid)
1119 				continue;
1120 			break;
1121 
1122 		case TLBI_VAA:
1123 			if (!vncr_tlb_intersects(vt, vt->gva, scope->va, scope->size))
1124 				continue;
1125 			break;
1126 
1127 		case TLBI_ASID:
1128 			if (!vt->wr.nG || vt->wr.asid != scope->asid)
1129 				continue;
1130 			break;
1131 		}
1132 
1133 		invalidate_vncr(kvm, vt);
1134 	}
1135 }
1136 
1137 #define tlbi_va_s1_to_va(v)	(u64)sign_extend64((v) << 12, 48)
1138 
1139 static void compute_s1_tlbi_range(struct kvm_vcpu *vcpu, u32 inst, u64 val,
1140 				  struct s1e2_tlbi_scope *scope)
1141 {
1142 	switch (inst) {
1143 	case OP_TLBI_ALLE2:
1144 	case OP_TLBI_ALLE2IS:
1145 	case OP_TLBI_ALLE2OS:
1146 	case OP_TLBI_VMALLE1:
1147 	case OP_TLBI_VMALLE1IS:
1148 	case OP_TLBI_VMALLE1OS:
1149 	case OP_TLBI_ALLE2NXS:
1150 	case OP_TLBI_ALLE2ISNXS:
1151 	case OP_TLBI_ALLE2OSNXS:
1152 	case OP_TLBI_VMALLE1NXS:
1153 	case OP_TLBI_VMALLE1ISNXS:
1154 	case OP_TLBI_VMALLE1OSNXS:
1155 		scope->type = TLBI_ALL;
1156 		break;
1157 	case OP_TLBI_VAE2:
1158 	case OP_TLBI_VAE2IS:
1159 	case OP_TLBI_VAE2OS:
1160 	case OP_TLBI_VAE1:
1161 	case OP_TLBI_VAE1IS:
1162 	case OP_TLBI_VAE1OS:
1163 	case OP_TLBI_VAE2NXS:
1164 	case OP_TLBI_VAE2ISNXS:
1165 	case OP_TLBI_VAE2OSNXS:
1166 	case OP_TLBI_VAE1NXS:
1167 	case OP_TLBI_VAE1ISNXS:
1168 	case OP_TLBI_VAE1OSNXS:
1169 	case OP_TLBI_VALE2:
1170 	case OP_TLBI_VALE2IS:
1171 	case OP_TLBI_VALE2OS:
1172 	case OP_TLBI_VALE1:
1173 	case OP_TLBI_VALE1IS:
1174 	case OP_TLBI_VALE1OS:
1175 	case OP_TLBI_VALE2NXS:
1176 	case OP_TLBI_VALE2ISNXS:
1177 	case OP_TLBI_VALE2OSNXS:
1178 	case OP_TLBI_VALE1NXS:
1179 	case OP_TLBI_VALE1ISNXS:
1180 	case OP_TLBI_VALE1OSNXS:
1181 		scope->type = TLBI_VA;
1182 		scope->size = ttl_to_size(FIELD_GET(TLBI_TTL_MASK, val));
1183 		scope->va = tlbi_va_s1_to_va(val) & ~(scope->size - 1);
1184 		scope->asid = FIELD_GET(TLBIR_ASID_MASK, val);
1185 		break;
1186 	case OP_TLBI_ASIDE1:
1187 	case OP_TLBI_ASIDE1IS:
1188 	case OP_TLBI_ASIDE1OS:
1189 	case OP_TLBI_ASIDE1NXS:
1190 	case OP_TLBI_ASIDE1ISNXS:
1191 	case OP_TLBI_ASIDE1OSNXS:
1192 		scope->type = TLBI_ASID;
1193 		scope->asid = FIELD_GET(TLBIR_ASID_MASK, val);
1194 		break;
1195 	case OP_TLBI_VAAE1:
1196 	case OP_TLBI_VAAE1IS:
1197 	case OP_TLBI_VAAE1OS:
1198 	case OP_TLBI_VAAE1NXS:
1199 	case OP_TLBI_VAAE1ISNXS:
1200 	case OP_TLBI_VAAE1OSNXS:
1201 	case OP_TLBI_VAALE1:
1202 	case OP_TLBI_VAALE1IS:
1203 	case OP_TLBI_VAALE1OS:
1204 	case OP_TLBI_VAALE1NXS:
1205 	case OP_TLBI_VAALE1ISNXS:
1206 	case OP_TLBI_VAALE1OSNXS:
1207 		scope->type = TLBI_VAA;
1208 		scope->size = ttl_to_size(FIELD_GET(TLBI_TTL_MASK, val));
1209 		scope->va = tlbi_va_s1_to_va(val) & ~(scope->size - 1);
1210 		break;
1211 	case OP_TLBI_RVAE2:
1212 	case OP_TLBI_RVAE2IS:
1213 	case OP_TLBI_RVAE2OS:
1214 	case OP_TLBI_RVAE1:
1215 	case OP_TLBI_RVAE1IS:
1216 	case OP_TLBI_RVAE1OS:
1217 	case OP_TLBI_RVAE2NXS:
1218 	case OP_TLBI_RVAE2ISNXS:
1219 	case OP_TLBI_RVAE2OSNXS:
1220 	case OP_TLBI_RVAE1NXS:
1221 	case OP_TLBI_RVAE1ISNXS:
1222 	case OP_TLBI_RVAE1OSNXS:
1223 	case OP_TLBI_RVALE2:
1224 	case OP_TLBI_RVALE2IS:
1225 	case OP_TLBI_RVALE2OS:
1226 	case OP_TLBI_RVALE1:
1227 	case OP_TLBI_RVALE1IS:
1228 	case OP_TLBI_RVALE1OS:
1229 	case OP_TLBI_RVALE2NXS:
1230 	case OP_TLBI_RVALE2ISNXS:
1231 	case OP_TLBI_RVALE2OSNXS:
1232 	case OP_TLBI_RVALE1NXS:
1233 	case OP_TLBI_RVALE1ISNXS:
1234 	case OP_TLBI_RVALE1OSNXS:
1235 		scope->type = TLBI_VA;
1236 		scope->va = decode_range_tlbi(val, &scope->size, &scope->asid);
1237 		break;
1238 	case OP_TLBI_RVAAE1:
1239 	case OP_TLBI_RVAAE1IS:
1240 	case OP_TLBI_RVAAE1OS:
1241 	case OP_TLBI_RVAAE1NXS:
1242 	case OP_TLBI_RVAAE1ISNXS:
1243 	case OP_TLBI_RVAAE1OSNXS:
1244 	case OP_TLBI_RVAALE1:
1245 	case OP_TLBI_RVAALE1IS:
1246 	case OP_TLBI_RVAALE1OS:
1247 	case OP_TLBI_RVAALE1NXS:
1248 	case OP_TLBI_RVAALE1ISNXS:
1249 	case OP_TLBI_RVAALE1OSNXS:
1250 		scope->type = TLBI_VAA;
1251 		scope->va = decode_range_tlbi(val, &scope->size, NULL);
1252 		break;
1253 	}
1254 }
1255 
1256 void kvm_handle_s1e2_tlbi(struct kvm_vcpu *vcpu, u32 inst, u64 val)
1257 {
1258 	struct s1e2_tlbi_scope scope = {};
1259 
1260 	compute_s1_tlbi_range(vcpu, inst, val, &scope);
1261 
1262 	guard(write_lock)(&vcpu->kvm->mmu_lock);
1263 	invalidate_vncr_va(vcpu->kvm, &scope);
1264 }
1265 
1266 static void kvm_invalidate_vncr_ipa_all(struct kvm *kvm)
1267 {
1268 	struct kvm_pgtable *pgt = kvm->arch.mmu.pgt;
1269 
1270 	lockdep_assert_held_write(&kvm->mmu_lock);
1271 
1272 	/* if the mmu lock was dropped, pgt teardown may have raced. */
1273 	if (pgt)
1274 		kvm_invalidate_vncr_ipa(kvm, 0, BIT(pgt->ia_bits));
1275 }
1276 
1277 void kvm_nested_s2_wp(struct kvm *kvm)
1278 {
1279 	int i;
1280 
1281 	lockdep_assert_held_write(&kvm->mmu_lock);
1282 
1283 	if (!kvm->arch.nested_mmus_size)
1284 		return;
1285 
1286 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1287 		struct kvm_s2_mmu *mmu = kvm->arch.nested_mmus[i];
1288 
1289 		if (kvm_s2_mmu_valid(mmu))
1290 			kvm_stage2_wp_range(mmu, 0, kvm_phys_size(mmu));
1291 	}
1292 
1293 	kvm_invalidate_vncr_ipa_all(kvm);
1294 }
1295 
1296 void kvm_nested_s2_unmap(struct kvm *kvm, bool may_block)
1297 {
1298 	int i;
1299 
1300 	lockdep_assert_held_write(&kvm->mmu_lock);
1301 
1302 	if (!kvm->arch.nested_mmus_size)
1303 		return;
1304 
1305 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1306 		struct kvm_s2_mmu *mmu = kvm->arch.nested_mmus[i];
1307 
1308 		if (kvm_s2_mmu_valid(mmu))
1309 			kvm_stage2_unmap_range(mmu, 0, kvm_phys_size(mmu), may_block);
1310 	}
1311 
1312 	kvm_invalidate_vncr_ipa_all(kvm);
1313 }
1314 
1315 void kvm_nested_s2_flush(struct kvm *kvm)
1316 {
1317 	int i;
1318 
1319 	lockdep_assert_held_write(&kvm->mmu_lock);
1320 
1321 	if (!kvm->arch.nested_mmus_size)
1322 		return;
1323 
1324 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1325 		struct kvm_s2_mmu *mmu = kvm->arch.nested_mmus[i];
1326 
1327 		if (kvm_s2_mmu_valid(mmu))
1328 			kvm_stage2_flush_range(mmu, 0, kvm_phys_size(mmu));
1329 	}
1330 }
1331 
1332 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1333 {
1334 	for (int i = 0; i < kvm->arch.nested_mmus_size; i++) {
1335 		struct kvm_s2_mmu *mmu = kvm->arch.nested_mmus[i];
1336 
1337 		if (!WARN_ON(atomic_read(&mmu->refcnt)))
1338 			kvm_free_stage2_pgd(mmu);
1339 	}
1340 	kvm_uninit_stage2_mmu(kvm);
1341 }
1342 
1343 /*
1344  * Dealing with VNCR_EL2 exposed by the *guest* is a complicated matter:
1345  *
1346  * - We introduce an internal representation of a vcpu-private TLB,
1347  *   representing the mapping between the guest VA contained in VNCR_EL2,
1348  *   the IPA the guest's EL2 PTs point to, and the actual PA this lives at.
1349  *
1350  * - On translation fault from a nested VNCR access, we create such a TLB.
1351  *   If there is no mapping to describe, the guest inherits the fault.
1352  *   Crucially, no actual mapping is done at this stage.
1353  *
1354  * - On vcpu_load() in a non-HYP context with HCR_EL2.NV==1, if the above
1355  *   TLB exists, we map it in the fixmap for this CPU, and run with it. We
1356  *   have to respect the permissions dictated by the guest, but not the
1357  *   memory type (FWB is a must).
1358  *
1359  * - Note that we usually don't do a vcpu_load() on the back of a fault
1360  *   (unless we are preempted), so the resolution of a translation fault
1361  *   must go via a request that will map the VNCR page in the fixmap.
1362  *   vcpu_load() might as well use the same mechanism.
1363  *
1364  * - On vcpu_put() in a non-HYP context with HCR_EL2.NV==1, if the TLB was
1365  *   mapped, we unmap it. Yes it is that simple. The TLB still exists
1366  *   though, and may be reused at a later load.
1367  *
1368  * - On permission fault, we simply forward the fault to the guest's EL2.
1369  *   Get out of my way.
1370  *
1371  * - On any TLBI for the EL2&0 translation regime, we must find any TLB that
1372  *   intersects with the TLBI request, invalidate it, and unmap the page
1373  *   from the fixmap. Because we need to look at all the vcpu-private TLBs,
1374  *   this requires some wide-ranging locking to ensure that nothing races
1375  *   against it. This requires some refcounting to avoid the search when
1376  *   no such TLB is present (see below).
1377  *
1378  * - On MMU notifiers, we must invalidate our TLB in a similar way, but
1379  *   looking at the IPA instead. The funny part is that there may not be a
1380  *   stage-2 mapping for this page if L1 hasn't accessed it using LD/ST
1381  *   instructions.
1382  *
1383  * - vncr_tlb_count tracks the number of valid VNCR TLBs VM-wide. This isn't
1384  *   the number of *mapped* L1 VNCR pages, which is likely be a subset (and
1385  *   by definition, a TLBI handled from L1 runs with the canonical VNCR
1386  *   page, not the L1's). The innermost trap handling code checks this to
1387  *   find out whether to return to the guest ASAP (no L1 TLBs) or to visit
1388  *   this part of the world for some extra invalidation work.
1389  */
1390 
1391 int kvm_vcpu_allocate_vncr_tlb(struct kvm_vcpu *vcpu)
1392 {
1393 	if (!kvm_has_feat(vcpu->kvm, ID_AA64MMFR4_EL1, NV_frac, NV2_ONLY))
1394 		return 0;
1395 
1396 	if (!vcpu->arch.vncr_tlb) {
1397 		struct vncr_tlb *vt = kzalloc_obj(*vcpu->arch.vncr_tlb,
1398 						  GFP_KERNEL_ACCOUNT);
1399 
1400 		/*
1401 		 * Taking the lock on assignment ensures that the TLB is
1402 		 * seen as initialised when following the pointer (release
1403 		 * semantics of the unlock), and avoids having acquires on
1404 		 * each user which already take the lock.
1405 		 */
1406 		scoped_guard(write_lock, &vcpu->kvm->mmu_lock)
1407 			vcpu->arch.vncr_tlb = vt;
1408 	}
1409 
1410 	if (!vcpu->arch.vncr_tlb)
1411 		return -ENOMEM;
1412 
1413 	return 0;
1414 }
1415 
1416 static u64 read_vncr_el2(struct kvm_vcpu *vcpu)
1417 {
1418 	return (u64)sign_extend64(__vcpu_sys_reg(vcpu, VNCR_EL2), 48);
1419 }
1420 
1421 static int kvm_translate_vncr(struct kvm_vcpu *vcpu, bool *is_gmem)
1422 {
1423 	struct kvm_memory_slot *memslot;
1424 	bool write_fault, writable;
1425 	unsigned long mmu_seq;
1426 	struct vncr_tlb *vt;
1427 	struct page *page;
1428 	u64 va, pfn, gfn;
1429 	int ret;
1430 
1431 	vt = vcpu->arch.vncr_tlb;
1432 
1433 	/*
1434 	 * If we're about to walk the EL2 S1 PTs, we must invalidate the
1435 	 * current TLB, as it could be sampled from another vcpu doing a
1436 	 * TLBI *IS. A real CPU wouldn't do that, but we only keep a single
1437 	 * translation, so not much of a choice.
1438 	 *
1439 	 * We also prepare the next walk wilst we're at it.
1440 	 */
1441 	scoped_guard(write_lock, &vcpu->kvm->mmu_lock) {
1442 		this_cpu_reset_vncr_fixmap(vcpu);
1443 		if (vt->valid)
1444 			invalidate_vncr(vcpu->kvm, vt);
1445 
1446 		vt->wi = (struct s1_walk_info) {
1447 			.regime	= TR_EL20,
1448 			.as_el0	= false,
1449 			.pan	= false,
1450 		};
1451 		vt->wr = (struct s1_walk_result){};
1452 	}
1453 
1454 	guard(srcu)(&vcpu->kvm->srcu);
1455 
1456 	va =  read_vncr_el2(vcpu);
1457 
1458 	mmu_seq = vcpu->kvm->mmu_invalidate_seq;
1459 	smp_rmb();
1460 
1461 	ret = __kvm_translate_va(vcpu, &vt->wi, &vt->wr, va);
1462 	if (ret)
1463 		return ret;
1464 
1465 	write_fault = kvm_is_write_fault(vcpu);
1466 
1467 	gfn = vt->wr.pa >> PAGE_SHIFT;
1468 	memslot = gfn_to_memslot(vcpu->kvm, gfn);
1469 	if (!memslot) {
1470 		fail_s1_walk(&vt->wr, ESR_ELx_FSC_EXTABT, false);
1471 		return -EFAULT;
1472 	}
1473 
1474 	*is_gmem = kvm_slot_has_gmem(memslot);
1475 	if (!*is_gmem) {
1476 		pfn = __kvm_faultin_pfn(memslot, gfn, write_fault ? FOLL_WRITE : 0,
1477 					&writable, &page);
1478 		if (is_error_noslot_pfn(pfn)) {
1479 			fail_s1_walk(&vt->wr, ESR_ELx_FSC_EXTABT, false);
1480 			return -EFAULT;
1481 		}
1482 	} else {
1483 		ret = kvm_gmem_get_pfn(vcpu->kvm, memslot, gfn, &pfn, &page, NULL);
1484 		if (ret) {
1485 			kvm_prepare_memory_fault_exit(vcpu, vt->wr.pa, PAGE_SIZE,
1486 					      write_fault, false, false);
1487 			return ret;
1488 		}
1489 
1490 		writable = !(memslot->flags & KVM_MEM_READONLY);
1491 	}
1492 
1493 	/*
1494 	 * FIXME: This check is too restrictive as KVM allows cacheable memory
1495 	 * attributes for PFNMAP VMAs that have cacheable attributes in host
1496 	 * stage-1.
1497 	 */
1498 	if (!pfn_is_map_memory(pfn)) {
1499 		kvm_release_faultin_page(vcpu->kvm, page, true, false);
1500 		fail_s1_walk(&vt->wr, ESR_ELx_FSC_EXTABT, false);
1501 		return -EINVAL;
1502 	}
1503 
1504 	scoped_guard(write_lock, &vcpu->kvm->mmu_lock) {
1505 		if (mmu_invalidate_retry(vcpu->kvm, mmu_seq)) {
1506 			kvm_release_faultin_page(vcpu->kvm, page, true, false);
1507 			return -EAGAIN;
1508 		}
1509 
1510 		vt->gva = va;
1511 		vt->hpa = pfn << PAGE_SHIFT;
1512 		vt->hpa_writable = writable;
1513 		vt->valid = true;
1514 		atomic_set(&vt->cpu, -1);
1515 
1516 		kvm_make_request(KVM_REQ_MAP_L1_VNCR_EL2, vcpu);
1517 		kvm_release_faultin_page(vcpu->kvm, page, false, vt->wr.pw && vt->hpa_writable);
1518 	}
1519 
1520 	if (vt->wr.pw && vt->hpa_writable)
1521 		mark_page_dirty(vcpu->kvm, gfn);
1522 
1523 	return 0;
1524 }
1525 
1526 static void handle_vncr_perm(struct kvm_vcpu *vcpu)
1527 {
1528 	struct vncr_tlb *vt = vcpu->arch.vncr_tlb;
1529 	u64 esr = kvm_vcpu_get_esr(vcpu);
1530 	u64 fsc;
1531 
1532 	/*
1533 	 * Promote to an external abort if the stage-1 permits writes but the
1534 	 * HPA is read-only (e.g. RO memslot).
1535 	 */
1536 	if (kvm_is_write_fault(vcpu) && vt->wr.pw && !vt->hpa_writable)
1537 		fsc = ESR_ELx_FSC_EXTABT;
1538 	/*
1539 	 * Otherwise, inject a permission fault using the guest's translation
1540 	 * level rather than the host's.
1541 	 */
1542 	else
1543 		fsc = ESR_ELx_FSC_PERM_L(vt->wr.level);
1544 
1545 	esr &= ~ESR_ELx_FSC;
1546 	esr |= FIELD_PREP(ESR_ELx_FSC, fsc);
1547 
1548 	kvm_inject_nested_sync(vcpu, esr);
1549 }
1550 
1551 int kvm_handle_vncr_abort(struct kvm_vcpu *vcpu)
1552 {
1553 	struct vncr_tlb *vt = vcpu->arch.vncr_tlb;
1554 	u64 esr = kvm_vcpu_get_esr(vcpu);
1555 	bool is_gmem = false;
1556 	bool perm;
1557 	int ret;
1558 
1559 	WARN_ON_ONCE(!(esr & ESR_ELx_VNCR));
1560 
1561 	if (kvm_vcpu_abt_issea(vcpu))
1562 		return kvm_handle_guest_sea(vcpu);
1563 
1564 	if (!esr_fsc_is_translation_fault(esr) && !esr_fsc_is_permission_fault(esr)) {
1565 		KVM_BUG(1, vcpu->kvm, "Unhandled VNCR abort, ESR=%llx\n", esr);
1566 		return -EIO;
1567 	}
1568 
1569 	/*
1570 	 * Speculatively increment the TLB count to make sure concurrent
1571 	 * TLBIs will take the slow path, and will interact with the retry
1572 	 * mechanism. Drop it again on error.
1573 	 */
1574 	atomic_inc(&vcpu->kvm->arch.vncr_tlb_count);
1575 	smp_mb__after_atomic();
1576 
1577 	ret = kvm_translate_vncr(vcpu, &is_gmem);
1578 	if (ret) {
1579 		smp_mb__before_atomic();
1580 		atomic_dec(&vcpu->kvm->arch.vncr_tlb_count);
1581 	}
1582 
1583 	switch (ret) {
1584 	case -EAGAIN:
1585 		/* Let's try again... */
1586 		return 1;
1587 	case -ENOMEM:
1588 		/*
1589 		 * For guest_memfd, this indicates that it failed to
1590 		 * create a folio to back the memory. Inform userspace.
1591 		 */
1592 		if (is_gmem)
1593 			return 0;
1594 		/* Otherwise, let's try again... */
1595 		break;
1596 	case -EFAULT:
1597 	case -EIO:
1598 	case -EHWPOISON:
1599 		if (is_gmem)
1600 			return 0;
1601 		fallthrough;
1602 	case -EINVAL:
1603 	case -ENOENT:
1604 	case -EACCES:
1605 		/*
1606 		 * Translation failed, inject the corresponding
1607 		 * exception back to EL2.
1608 		 */
1609 		esr &= ~ESR_ELx_FSC;
1610 		esr |= FIELD_PREP(ESR_ELx_FSC, vt->wr.fst);
1611 
1612 		kvm_inject_nested_sync(vcpu, esr);
1613 		break;
1614 	case 0:
1615 		perm = kvm_is_write_fault(vcpu) ? vt->wr.pw && vt->hpa_writable : vt->wr.pr;
1616 		if (!perm)
1617 			handle_vncr_perm(vcpu);
1618 		break;
1619 	}
1620 
1621 	return 1;
1622 }
1623 
1624 static void kvm_map_l1_vncr(struct kvm_vcpu *vcpu)
1625 {
1626 	struct vncr_tlb *vt = vcpu->arch.vncr_tlb;
1627 	pgprot_t prot;
1628 
1629 	guard(preempt)();
1630 	guard(read_lock)(&vcpu->kvm->mmu_lock);
1631 
1632 	/*
1633 	 * The request to map VNCR may have raced against some other
1634 	 * event, such as an interrupt, and may not be valid anymore.
1635 	 */
1636 	if (is_hyp_ctxt(vcpu))
1637 		return;
1638 
1639 	/*
1640 	 * Check that the pseudo-TLB is valid and that VNCR_EL2 still
1641 	 * contains the expected value. If it doesn't, we simply bail out
1642 	 * without a mapping -- a transformed MSR/MRS will generate the
1643 	 * fault and allows us to populate the pseudo-TLB.
1644 	 */
1645 	if (!vt->valid)
1646 		return;
1647 
1648 	/* We cache the MMU state in the TLB. Check that it matches. */
1649 	if (!!(vcpu_read_sys_reg(vcpu, SCTLR_EL2) & SCTLR_ELx_M) != s1_walk_translated(&vt->wr))
1650 		return;
1651 
1652 	if (read_vncr_el2(vcpu) != vt->gva)
1653 		return;
1654 
1655 	if (vt->wr.nG && get_asid_by_regime(vcpu, TR_EL20) != vt->wr.asid)
1656 		return;
1657 
1658 	if (vt->hpa_writable && vt->wr.pw && vt->wr.pr)
1659 		prot = PAGE_KERNEL;
1660 	else if (vt->wr.pr)
1661 		prot = PAGE_KERNEL_RO;
1662 	else
1663 		prot = PAGE_NONE;
1664 
1665 	/*
1666 	 * We can't map write-only (or no permission at all) in the kernel,
1667 	 * but the guest can do it if using POE, so we'll have to turn a
1668 	 * translation fault into a permission fault at runtime.
1669 	 * FIXME: WO doesn't work at all, need POE support in the kernel.
1670 	 */
1671 	if (pgprot_val(prot) != pgprot_val(PAGE_NONE)) {
1672 		atomic_set(&vt->cpu, smp_processor_id());
1673 		__set_fixmap(vncr_fixmap(atomic_read(&vt->cpu)), vt->hpa, prot);
1674 		host_data_set_flag(L1_VNCR_MAPPED);
1675 	}
1676 }
1677 
1678 /*
1679  * Our emulated CPU doesn't support all the possible features. For the
1680  * sake of simplicity (and probably mental sanity), wipe out a number
1681  * of feature bits we don't intend to support for the time being.
1682  * This list should get updated as new features get added to the NV
1683  * support, and new extension to the architecture.
1684  */
1685 u64 limit_nv_id_reg(struct kvm *kvm, u32 reg, u64 val)
1686 {
1687 	u64 orig_val = val;
1688 
1689 	switch (reg) {
1690 	case SYS_ID_AA64ISAR1_EL1:
1691 		/* Support everything but LS64 and Spec Invalidation */
1692 		val &= ~(ID_AA64ISAR1_EL1_LS64	|
1693 			 ID_AA64ISAR1_EL1_SPECRES);
1694 		break;
1695 
1696 	case SYS_ID_AA64PFR0_EL1:
1697 		/* No RME, AMU, MPAM, or S-EL2 */
1698 		val &= ~(ID_AA64PFR0_EL1_RME	|
1699 			 ID_AA64PFR0_EL1_AMU	|
1700 			 ID_AA64PFR0_EL1_MPAM	|
1701 			 ID_AA64PFR0_EL1_SEL2	|
1702 			 ID_AA64PFR0_EL1_EL3	|
1703 			 ID_AA64PFR0_EL1_EL2	|
1704 			 ID_AA64PFR0_EL1_EL1	|
1705 			 ID_AA64PFR0_EL1_EL0);
1706 		/* 64bit only at any EL */
1707 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL0, IMP);
1708 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL1, IMP);
1709 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL2, IMP);
1710 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL3, IMP);
1711 		break;
1712 
1713 	case SYS_ID_AA64PFR1_EL1:
1714 		/* Only support BTI, SSBS, CSV2_frac */
1715 		val &= ~(ID_AA64PFR1_EL1_PFAR		|
1716 			 ID_AA64PFR1_EL1_MTEX		|
1717 			 ID_AA64PFR1_EL1_THE		|
1718 			 ID_AA64PFR1_EL1_GCS		|
1719 			 ID_AA64PFR1_EL1_MTE_frac	|
1720 			 ID_AA64PFR1_EL1_NMI		|
1721 			 ID_AA64PFR1_EL1_SME		|
1722 			 ID_AA64PFR1_EL1_RES0		|
1723 			 ID_AA64PFR1_EL1_MPAM_frac	|
1724 			 ID_AA64PFR1_EL1_MTE);
1725 		break;
1726 
1727 	case SYS_ID_AA64PFR2_EL1:
1728 		/* GICv5 is not yet supported for NV */
1729 		val &= ~ID_AA64PFR2_EL1_GCIE;
1730 		break;
1731 
1732 	case SYS_ID_AA64MMFR0_EL1:
1733 		/* Hide ExS, Secure Memory */
1734 		val &= ~(ID_AA64MMFR0_EL1_EXS		|
1735 			 ID_AA64MMFR0_EL1_TGRAN4_2	|
1736 			 ID_AA64MMFR0_EL1_TGRAN16_2	|
1737 			 ID_AA64MMFR0_EL1_TGRAN64_2	|
1738 			 ID_AA64MMFR0_EL1_SNSMEM);
1739 
1740 		/* Hide CNTPOFF if present */
1741 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR0_EL1, ECV, IMP);
1742 
1743 		/* Disallow unsupported S2 page sizes */
1744 		switch (PAGE_SIZE) {
1745 		case SZ_64K:
1746 			val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN16_2, NI);
1747 			fallthrough;
1748 		case SZ_16K:
1749 			val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN4_2, NI);
1750 			fallthrough;
1751 		case SZ_4K:
1752 			/* Support everything */
1753 			break;
1754 		}
1755 
1756 		/*
1757 		 * Since we can't support a guest S2 page size smaller
1758 		 * than the host's own page size (due to KVM only
1759 		 * populating its own S2 using the kernel's page
1760 		 * size), advertise the limitation using FEAT_GTG.
1761 		 */
1762 		switch (PAGE_SIZE) {
1763 		case SZ_4K:
1764 			if (_has_tgran_2(orig_val, 4))
1765 				val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN4_2, IMP);
1766 			fallthrough;
1767 		case SZ_16K:
1768 			if (_has_tgran_2(orig_val, 16))
1769 				val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN16_2, IMP);
1770 			fallthrough;
1771 		case SZ_64K:
1772 			if (_has_tgran_2(orig_val, 64))
1773 				val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN64_2, IMP);
1774 			break;
1775 		}
1776 
1777 		/* Cap PARange to 48bits */
1778 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR0_EL1, PARANGE, 48);
1779 		break;
1780 
1781 	case SYS_ID_AA64MMFR1_EL1:
1782 		val &= ~(ID_AA64MMFR1_EL1_CMOW		|
1783 			 ID_AA64MMFR1_EL1_nTLBPA	|
1784 			 ID_AA64MMFR1_EL1_ETS);
1785 
1786 		/* FEAT_E2H0 implies no VHE */
1787 		if (test_bit(KVM_ARM_VCPU_HAS_EL2_E2H0, kvm->arch.vcpu_features))
1788 			val &= ~ID_AA64MMFR1_EL1_VH;
1789 
1790 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR1_EL1, HAFDBS, AF);
1791 		break;
1792 
1793 	case SYS_ID_AA64MMFR2_EL1:
1794 		val &= ~(ID_AA64MMFR2_EL1_BBM	|
1795 			 ID_AA64MMFR2_EL1_TTL	|
1796 			 GENMASK_ULL(47, 44)	|
1797 			 ID_AA64MMFR2_EL1_ST	|
1798 			 ID_AA64MMFR2_EL1_CCIDX	|
1799 			 ID_AA64MMFR2_EL1_VARange);
1800 
1801 		/* Force TTL support */
1802 		val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR2_EL1, TTL, IMP);
1803 		break;
1804 
1805 	case SYS_ID_AA64MMFR4_EL1:
1806 		/*
1807 		 * You get EITHER
1808 		 *
1809 		 * - FEAT_VHE without FEAT_E2H0
1810 		 * - FEAT_NV limited to FEAT_NV2(p1)/NV3
1811 		 * - HCR_EL2.NV1 being RES0
1812 		 *
1813 		 * OR
1814 		 *
1815 		 * - FEAT_E2H0 without FEAT_VHE nor FEAT_NV
1816 		 *
1817 		 * Life is too short for anything else.
1818 		 */
1819 		if (test_bit(KVM_ARM_VCPU_HAS_EL2_E2H0, kvm->arch.vcpu_features)) {
1820 			val = 0;
1821 		} else {
1822 			val &= ID_AA64MMFR4_EL1_NV_frac;
1823 			if (cpus_have_final_cap(ARM64_HAS_NV3))
1824 				val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR4_EL1, NV_frac, NV3);
1825 			else if (cpus_have_final_cap(ARM64_HAS_NV2P1))
1826 				val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR4_EL1, NV_frac, NV2P1);
1827 			else
1828 				val = SYS_FIELD_PREP_ENUM(ID_AA64MMFR4_EL1, NV_frac, NV2_ONLY);
1829 			val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR4_EL1, E2H0, NI_NV1);
1830 		}
1831 		break;
1832 
1833 	case SYS_ID_AA64DFR0_EL1:
1834 		/* Only limited support for PMU, Debug, BPs, WPs, and HPMN0 */
1835 		val &= ~(ID_AA64DFR0_EL1_ExtTrcBuff	|
1836 			 ID_AA64DFR0_EL1_BRBE		|
1837 			 ID_AA64DFR0_EL1_MTPMU		|
1838 			 ID_AA64DFR0_EL1_TraceBuffer	|
1839 			 ID_AA64DFR0_EL1_TraceFilt	|
1840 			 ID_AA64DFR0_EL1_PMSVer		|
1841 			 ID_AA64DFR0_EL1_CTX_CMPs	|
1842 			 ID_AA64DFR0_EL1_SEBEP		|
1843 			 ID_AA64DFR0_EL1_PMSS		|
1844 			 ID_AA64DFR0_EL1_TraceVer);
1845 
1846 		/*
1847 		 * FEAT_Debugv8p9 requires support for extended breakpoints /
1848 		 * watchpoints.
1849 		 */
1850 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64DFR0_EL1, DebugVer, V8P8);
1851 		break;
1852 	}
1853 
1854 	return val;
1855 }
1856 
1857 u64 kvm_vcpu_apply_reg_masks(const struct kvm_vcpu *vcpu,
1858 			     enum vcpu_sysreg sr, u64 v)
1859 {
1860 	struct resx resx;
1861 
1862 	resx = kvm_get_sysreg_resx(vcpu->kvm, sr);
1863 	v &= ~resx.res0;
1864 	v |= resx.res1;
1865 
1866 	return v;
1867 }
1868 
1869 static __always_inline void set_sysreg_masks(struct kvm *kvm, int sr, struct resx resx)
1870 {
1871 	BUILD_BUG_ON(!__builtin_constant_p(sr));
1872 	BUILD_BUG_ON(sr < __SANITISED_REG_START__);
1873 	BUILD_BUG_ON(sr >= NR_SYS_REGS);
1874 
1875 	kvm_set_sysreg_resx(kvm, sr, resx);
1876 }
1877 
1878 int kvm_init_nv_sysregs(struct kvm_vcpu *vcpu)
1879 {
1880 	struct kvm *kvm = vcpu->kvm;
1881 	struct resx resx;
1882 
1883 	lockdep_assert_held(&kvm->arch.config_lock);
1884 
1885 	if (kvm->arch.sysreg_masks)
1886 		goto out;
1887 
1888 	kvm->arch.sysreg_masks = kzalloc_obj(*(kvm->arch.sysreg_masks),
1889 					     GFP_KERNEL_ACCOUNT);
1890 	if (!kvm->arch.sysreg_masks)
1891 		return -ENOMEM;
1892 
1893 	/* VTTBR_EL2 */
1894 	resx = (typeof(resx)){};
1895 	if (!kvm_has_feat_enum(kvm, ID_AA64MMFR1_EL1, VMIDBits, 16))
1896 		resx.res0 |= GENMASK(63, 56);
1897 	if (!kvm_has_feat(kvm, ID_AA64MMFR2_EL1, CnP, IMP))
1898 		resx.res0 |= VTTBR_CNP_BIT;
1899 	set_sysreg_masks(kvm, VTTBR_EL2, resx);
1900 
1901 	/* VTCR_EL2 */
1902 	resx = get_reg_fixed_bits(kvm, VTCR_EL2);
1903 	set_sysreg_masks(kvm, VTCR_EL2, resx);
1904 
1905 	/* VMPIDR_EL2 */
1906 	resx.res0 = GENMASK(63, 40) | GENMASK(30, 24);
1907 	resx.res1 = BIT(31);
1908 	set_sysreg_masks(kvm, VMPIDR_EL2, resx);
1909 
1910 	/* HCR_EL2 */
1911 	resx = get_reg_fixed_bits(kvm, HCR_EL2);
1912 	set_sysreg_masks(kvm, HCR_EL2, resx);
1913 
1914 	/* NVHCR_EL2 */
1915 	resx = get_reg_fixed_bits(kvm, NVHCR_EL2);
1916 	set_sysreg_masks(kvm, NVHCR_EL2, resx);
1917 
1918 	/* HCRX_EL2 */
1919 	resx = get_reg_fixed_bits(kvm, HCRX_EL2);
1920 	set_sysreg_masks(kvm, HCRX_EL2, resx);
1921 
1922 	/* HFG[RW]TR_EL2 */
1923 	resx = get_reg_fixed_bits(kvm, HFGRTR_EL2);
1924 	set_sysreg_masks(kvm, HFGRTR_EL2, resx);
1925 	resx = get_reg_fixed_bits(kvm, HFGWTR_EL2);
1926 	set_sysreg_masks(kvm, HFGWTR_EL2, resx);
1927 
1928 	/* HDFG[RW]TR_EL2 */
1929 	resx = get_reg_fixed_bits(kvm, HDFGRTR_EL2);
1930 	set_sysreg_masks(kvm, HDFGRTR_EL2, resx);
1931 	resx = get_reg_fixed_bits(kvm, HDFGWTR_EL2);
1932 	set_sysreg_masks(kvm, HDFGWTR_EL2, resx);
1933 
1934 	/* HFGITR_EL2 */
1935 	resx = get_reg_fixed_bits(kvm, HFGITR_EL2);
1936 	set_sysreg_masks(kvm, HFGITR_EL2, resx);
1937 
1938 	/* HAFGRTR_EL2 - not a lot to see here */
1939 	resx = get_reg_fixed_bits(kvm, HAFGRTR_EL2);
1940 	set_sysreg_masks(kvm, HAFGRTR_EL2, resx);
1941 
1942 	/* HFG[RW]TR2_EL2 */
1943 	resx = get_reg_fixed_bits(kvm, HFGRTR2_EL2);
1944 	set_sysreg_masks(kvm, HFGRTR2_EL2, resx);
1945 	resx = get_reg_fixed_bits(kvm, HFGWTR2_EL2);
1946 	set_sysreg_masks(kvm, HFGWTR2_EL2, resx);
1947 
1948 	/* HDFG[RW]TR2_EL2 */
1949 	resx = get_reg_fixed_bits(kvm, HDFGRTR2_EL2);
1950 	set_sysreg_masks(kvm, HDFGRTR2_EL2, resx);
1951 	resx = get_reg_fixed_bits(kvm, HDFGWTR2_EL2);
1952 	set_sysreg_masks(kvm, HDFGWTR2_EL2, resx);
1953 
1954 	/* HFGITR2_EL2 */
1955 	resx = get_reg_fixed_bits(kvm, HFGITR2_EL2);
1956 	set_sysreg_masks(kvm, HFGITR2_EL2, resx);
1957 
1958 	/* TCR2_EL2 */
1959 	resx = get_reg_fixed_bits(kvm, TCR2_EL2);
1960 	set_sysreg_masks(kvm, TCR2_EL2, resx);
1961 
1962 	/* SCTLR_EL1 */
1963 	resx = get_reg_fixed_bits(kvm, SCTLR_EL1);
1964 	set_sysreg_masks(kvm, SCTLR_EL1, resx);
1965 
1966 	/* SCTLR_EL2 */
1967 	resx = get_reg_fixed_bits(kvm, SCTLR_EL2);
1968 	set_sysreg_masks(kvm, SCTLR_EL2, resx);
1969 
1970 	/* SCTLR2_ELx */
1971 	resx = get_reg_fixed_bits(kvm, SCTLR2_EL1);
1972 	set_sysreg_masks(kvm, SCTLR2_EL1, resx);
1973 	resx = get_reg_fixed_bits(kvm, SCTLR2_EL2);
1974 	set_sysreg_masks(kvm, SCTLR2_EL2, resx);
1975 
1976 	/* MDCR_EL2 */
1977 	resx = get_reg_fixed_bits(kvm, MDCR_EL2);
1978 	set_sysreg_masks(kvm, MDCR_EL2, resx);
1979 
1980 	/* CNTHCTL_EL2 */
1981 	resx.res0 = GENMASK(63, 20);
1982 	resx.res1 = 0;
1983 	if (!kvm_has_feat(kvm, ID_AA64PFR0_EL1, RME, IMP))
1984 		resx.res0 |= CNTHCTL_CNTPMASK | CNTHCTL_CNTVMASK;
1985 	if (!kvm_has_feat(kvm, ID_AA64MMFR0_EL1, ECV, CNTPOFF)) {
1986 		resx.res0 |= CNTHCTL_ECV;
1987 		if (!kvm_has_feat(kvm, ID_AA64MMFR0_EL1, ECV, IMP))
1988 			resx.res0 |= (CNTHCTL_EL1TVT | CNTHCTL_EL1TVCT |
1989 				      CNTHCTL_EL1NVPCT | CNTHCTL_EL1NVVCT);
1990 	}
1991 	if (!kvm_has_feat(kvm, ID_AA64MMFR1_EL1, VH, IMP))
1992 		resx.res0 |= GENMASK(11, 8);
1993 	set_sysreg_masks(kvm, CNTHCTL_EL2, resx);
1994 
1995 	/* ICH_HCR_EL2 */
1996 	resx.res0 = ICH_HCR_EL2_RES0;
1997 	resx.res1 = ICH_HCR_EL2_RES1;
1998 	if (!(vgic_ich_vtr() & ICH_VTR_EL2_TDS))
1999 		resx.res0 |= ICH_HCR_EL2_TDIR;
2000 	/* No GICv4 is presented to the guest */
2001 	resx.res0 |= ICH_HCR_EL2_DVIM | ICH_HCR_EL2_vSGIEOICount;
2002 	set_sysreg_masks(kvm, ICH_HCR_EL2, resx);
2003 
2004 	/* VNCR_EL2 */
2005 	resx.res0 = VNCR_EL2_RES0;
2006 	resx.res1 = VNCR_EL2_RES1;
2007 	set_sysreg_masks(kvm, VNCR_EL2, resx);
2008 
2009 	/* ZCR_EL2 - bits 8:4 are RAZ/WI so treat them as RES0 */
2010 	resx.res0 = ZCR_ELx_RES0 | GENMASK_ULL(8, 4);
2011 	resx.res1 = ZCR_ELx_RES1;
2012 	set_sysreg_masks(kvm, ZCR_EL2, resx);
2013 
2014 out:
2015 	for (enum vcpu_sysreg sr = __SANITISED_REG_START__; sr < NR_SYS_REGS; sr++)
2016 		__vcpu_rmw_sys_reg(vcpu, sr, |=, 0);
2017 
2018 	return 0;
2019 }
2020 
2021 void check_nested_vcpu_requests(struct kvm_vcpu *vcpu)
2022 {
2023 	if (kvm_check_request(KVM_REQ_NESTED_S2_UNMAP, vcpu)) {
2024 		struct kvm_s2_mmu *mmu = vcpu->arch.hw_mmu;
2025 
2026 		write_lock(&vcpu->kvm->mmu_lock);
2027 		if (mmu->pending_unmap) {
2028 			kvm_stage2_unmap_range(mmu, 0, kvm_phys_size(mmu), true);
2029 			mmu->pending_unmap = false;
2030 		}
2031 		write_unlock(&vcpu->kvm->mmu_lock);
2032 	}
2033 
2034 	if (kvm_check_request(KVM_REQ_MAP_L1_VNCR_EL2, vcpu))
2035 		kvm_map_l1_vncr(vcpu);
2036 
2037 	/* Must be last, as may switch context! */
2038 	if (kvm_check_request(KVM_REQ_GUEST_HYP_IRQ_PENDING, vcpu))
2039 		kvm_inject_nested_irq(vcpu);
2040 }
2041 
2042 /*
2043  * One of the many architectural bugs in FEAT_NV2 is that the guest hypervisor
2044  * can write to HCR_EL2 behind our back, potentially changing the exception
2045  * routing / masking for even the host context.
2046  *
2047  * What follows is some slop to (1) react to exception routing / masking and (2)
2048  * preserve the pending SError state across translation regimes.
2049  */
2050 void kvm_nested_flush_hwstate(struct kvm_vcpu *vcpu)
2051 {
2052 	if (!vcpu_has_nv(vcpu))
2053 		return;
2054 
2055 	if (unlikely(vcpu_test_and_clear_flag(vcpu, NESTED_SERROR_PENDING)))
2056 		kvm_inject_serror_esr(vcpu, vcpu_get_vsesr(vcpu));
2057 }
2058 
2059 void kvm_nested_sync_hwstate(struct kvm_vcpu *vcpu)
2060 {
2061 	unsigned long *hcr = vcpu_hcr(vcpu);
2062 
2063 	if (!vcpu_has_nv(vcpu))
2064 		return;
2065 
2066 	/*
2067 	 * We previously decided that an SError was deliverable to the guest.
2068 	 * Reap the pending state from HCR_EL2 and...
2069 	 */
2070 	if (unlikely(__test_and_clear_bit(__ffs(HCR_VSE), hcr)))
2071 		vcpu_set_flag(vcpu, NESTED_SERROR_PENDING);
2072 
2073 	/*
2074 	 * Re-attempt SError injection in case the deliverability has changed,
2075 	 * which is necessary to faithfully emulate WFI the case of a pending
2076 	 * SError being a wakeup condition.
2077 	 */
2078 	if (unlikely(vcpu_test_and_clear_flag(vcpu, NESTED_SERROR_PENDING)))
2079 		kvm_inject_serror_esr(vcpu, vcpu_get_vsesr(vcpu));
2080 }
2081 
2082 /*
2083  * KVM unconditionally sets most of these traps anyway but use an allowlist
2084  * to document the guest hypervisor traps that may take precedence and guard
2085  * against future changes to the non-nested trap configuration.
2086  */
2087 #define NV_MDCR_GUEST_INCLUDE	(MDCR_EL2_TDE	|	\
2088 				 MDCR_EL2_TDA	|	\
2089 				 MDCR_EL2_TDRA	|	\
2090 				 MDCR_EL2_TTRF	|	\
2091 				 MDCR_EL2_TPMS	|	\
2092 				 MDCR_EL2_TPM	|	\
2093 				 MDCR_EL2_TPMCR	|	\
2094 				 MDCR_EL2_TDCC	|	\
2095 				 MDCR_EL2_TDOSA)
2096 
2097 void kvm_nested_setup_mdcr_el2(struct kvm_vcpu *vcpu)
2098 {
2099 	u64 guest_mdcr = __vcpu_sys_reg(vcpu, MDCR_EL2);
2100 
2101 	if (is_nested_ctxt(vcpu))
2102 		vcpu->arch.mdcr_el2 |= (guest_mdcr & NV_MDCR_GUEST_INCLUDE);
2103 	/*
2104 	 * In yet another example where FEAT_NV2 is fscking broken, accesses
2105 	 * to MDSCR_EL1 are redirected to the VNCR despite having an effect
2106 	 * at EL2. Use a big hammer to apply sanity.
2107 	 *
2108 	 * Unless of course we have FEAT_FGT, in which case we can precisely
2109 	 * trap MDSCR_EL1.
2110 	 */
2111 	else if (!cpus_have_final_cap(ARM64_HAS_FGT))
2112 		vcpu->arch.mdcr_el2 |= MDCR_EL2_TDA;
2113 }
2114