xref: /linux/arch/arm64/kvm/nested.c (revision fab183d632628381b466a41479489541ac0e29a0)
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 
20 struct vncr_tlb {
21 	/* The guest's VNCR_EL2 */
22 	u64			gva;
23 	struct s1_walk_info	wi;
24 	struct s1_walk_result	wr;
25 
26 	u64			hpa;
27 	bool			hpa_writable;
28 
29 	/* -1 when not mapped on a CPU */
30 	int			cpu;
31 
32 	/*
33 	 * true if the TLB is valid. Can only be changed with the
34 	 * mmu_lock held.
35 	 */
36 	bool			valid;
37 };
38 
39 /*
40  * Ratio of live shadow S2 MMU per vcpu. This is a trade-off between
41  * memory usage and potential number of different sets of S2 PTs in
42  * the guests. Running out of S2 MMUs only affects performance (we
43  * will invalidate them more often).
44  */
45 #define S2_MMU_PER_VCPU		2
46 
kvm_init_nested(struct kvm * kvm)47 void kvm_init_nested(struct kvm *kvm)
48 {
49 	kvm->arch.nested_mmus = NULL;
50 	kvm->arch.nested_mmus_size = 0;
51 	atomic_set(&kvm->arch.vncr_map_count, 0);
52 }
53 
init_nested_s2_mmu(struct kvm * kvm,struct kvm_s2_mmu * mmu)54 static int init_nested_s2_mmu(struct kvm *kvm, struct kvm_s2_mmu *mmu)
55 {
56 	/*
57 	 * We only initialise the IPA range on the canonical MMU, which
58 	 * defines the contract between KVM and userspace on where the
59 	 * "hardware" is in the IPA space. This affects the validity of MMIO
60 	 * exits forwarded to userspace, for example.
61 	 *
62 	 * For nested S2s, we use the PARange as exposed to the guest, as it
63 	 * is allowed to use it at will to expose whatever memory map it
64 	 * wants to its own guests as it would be on real HW.
65 	 */
66 	return kvm_init_stage2_mmu(kvm, mmu, kvm_get_pa_bits(kvm));
67 }
68 
kvm_vcpu_init_nested(struct kvm_vcpu * vcpu)69 int kvm_vcpu_init_nested(struct kvm_vcpu *vcpu)
70 {
71 	struct kvm *kvm = vcpu->kvm;
72 	struct kvm_s2_mmu *tmp;
73 	int num_mmus, ret = 0;
74 
75 	if (test_bit(KVM_ARM_VCPU_HAS_EL2_E2H0, kvm->arch.vcpu_features) &&
76 	    !cpus_have_final_cap(ARM64_HAS_HCR_NV1))
77 		return -EINVAL;
78 
79 	if (!vcpu->arch.ctxt.vncr_array)
80 		vcpu->arch.ctxt.vncr_array = (u64 *)__get_free_page(GFP_KERNEL_ACCOUNT |
81 								    __GFP_ZERO);
82 
83 	if (!vcpu->arch.ctxt.vncr_array)
84 		return -ENOMEM;
85 
86 	/*
87 	 * Let's treat memory allocation failures as benign: If we fail to
88 	 * allocate anything, return an error and keep the allocated array
89 	 * alive. Userspace may try to recover by initializing the vcpu
90 	 * again, and there is no reason to affect the whole VM for this.
91 	 */
92 	num_mmus = atomic_read(&kvm->online_vcpus) * S2_MMU_PER_VCPU;
93 
94 	if (num_mmus > kvm->arch.nested_mmus_size) {
95 		tmp = kvcalloc(num_mmus, sizeof(*tmp), GFP_KERNEL_ACCOUNT);
96 		if (!tmp)
97 			return -ENOMEM;
98 
99 		write_lock(&kvm->mmu_lock);
100 
101 		if (kvm->arch.nested_mmus_size) {
102 			memcpy(tmp, kvm->arch.nested_mmus,
103 			       size_mul(sizeof(*tmp), kvm->arch.nested_mmus_size));
104 
105 			for (int i = 0; i < kvm->arch.nested_mmus_size; i++)
106 				tmp[i].pgt->mmu = &tmp[i];
107 		}
108 
109 		swap(kvm->arch.nested_mmus, tmp);
110 
111 		write_unlock(&kvm->mmu_lock);
112 
113 		kvfree(tmp);
114 	}
115 
116 	for (int i = kvm->arch.nested_mmus_size; !ret && i < num_mmus; i++)
117 		ret = init_nested_s2_mmu(kvm, &kvm->arch.nested_mmus[i]);
118 
119 	if (ret) {
120 		for (int i = kvm->arch.nested_mmus_size; i < num_mmus; i++)
121 			kvm_free_stage2_pgd(&kvm->arch.nested_mmus[i]);
122 
123 		free_page((unsigned long)vcpu->arch.ctxt.vncr_array);
124 		vcpu->arch.ctxt.vncr_array = NULL;
125 
126 		return ret;
127 	}
128 
129 	kvm->arch.nested_mmus_size = num_mmus;
130 
131 	return 0;
132 }
133 
134 struct s2_walk_info {
135 	u64		baddr;
136 	unsigned int	max_oa_bits;
137 	unsigned int	pgshift;
138 	unsigned int	sl;
139 	unsigned int	t0sz;
140 	bool		be;
141 	bool		ha;
142 };
143 
compute_fsc(int level,u32 fsc)144 static u32 compute_fsc(int level, u32 fsc)
145 {
146 	return fsc | (level & 0x3);
147 }
148 
esr_s2_fault(struct kvm_vcpu * vcpu,int level,u32 fsc)149 static int esr_s2_fault(struct kvm_vcpu *vcpu, int level, u32 fsc)
150 {
151 	u32 esr;
152 
153 	esr = kvm_vcpu_get_esr(vcpu) & ~ESR_ELx_FSC;
154 	esr |= compute_fsc(level, fsc);
155 	return esr;
156 }
157 
get_ia_size(struct s2_walk_info * wi)158 static int get_ia_size(struct s2_walk_info *wi)
159 {
160 	return 64 - wi->t0sz;
161 }
162 
check_base_s2_limits(struct kvm_vcpu * vcpu,struct s2_walk_info * wi,int level,int input_size,int stride)163 static int check_base_s2_limits(struct kvm_vcpu *vcpu, struct s2_walk_info *wi,
164 				int level, int input_size, int stride)
165 {
166 	int start_size, pa_max;
167 
168 	pa_max = kvm_get_pa_bits(vcpu->kvm);
169 
170 	/* Check translation limits */
171 	switch (BIT(wi->pgshift)) {
172 	case SZ_64K:
173 		if (level == 0 || (level == 1 && pa_max <= 42))
174 			return -EFAULT;
175 		break;
176 	case SZ_16K:
177 		if (level == 0 || (level == 1 && pa_max <= 40))
178 			return -EFAULT;
179 		break;
180 	case SZ_4K:
181 		if (level < 0 || (level == 0 && pa_max <= 42))
182 			return -EFAULT;
183 		break;
184 	}
185 
186 	/* Check input size limits */
187 	if (input_size > pa_max)
188 		return -EFAULT;
189 
190 	/* Check number of entries in starting level table */
191 	start_size = input_size - ((3 - level) * stride + wi->pgshift);
192 	if (start_size < 1 || start_size > stride + 4)
193 		return -EFAULT;
194 
195 	return 0;
196 }
197 
198 /* Check if output is within boundaries */
check_output_size(struct s2_walk_info * wi,phys_addr_t output)199 static int check_output_size(struct s2_walk_info *wi, phys_addr_t output)
200 {
201 	unsigned int output_size = wi->max_oa_bits;
202 
203 	if (output_size != 48 && (output & GENMASK_ULL(47, output_size)))
204 		return -1;
205 
206 	return 0;
207 }
208 
read_guest_s2_desc(struct kvm_vcpu * vcpu,phys_addr_t pa,u64 * desc,struct s2_walk_info * wi)209 static int read_guest_s2_desc(struct kvm_vcpu *vcpu, phys_addr_t pa, u64 *desc,
210 			      struct s2_walk_info *wi)
211 {
212 	u64 val;
213 	int r;
214 
215 	r = kvm_read_guest(vcpu->kvm, pa, &val, sizeof(val));
216 	if (r)
217 		return r;
218 
219 	/*
220 	 * Handle reversedescriptors if endianness differs between the
221 	 * host and the guest hypervisor.
222 	 */
223 	if (wi->be)
224 		*desc = be64_to_cpu((__force __be64)val);
225 	else
226 		*desc = le64_to_cpu((__force __le64)val);
227 
228 	return 0;
229 }
230 
swap_guest_s2_desc(struct kvm_vcpu * vcpu,phys_addr_t pa,u64 old,u64 new,struct s2_walk_info * wi)231 static int swap_guest_s2_desc(struct kvm_vcpu *vcpu, phys_addr_t pa, u64 old, u64 new,
232 			      struct s2_walk_info *wi)
233 {
234 	if (wi->be) {
235 		old = (__force u64)cpu_to_be64(old);
236 		new = (__force u64)cpu_to_be64(new);
237 	} else {
238 		old = (__force u64)cpu_to_le64(old);
239 		new = (__force u64)cpu_to_le64(new);
240 	}
241 
242 	return __kvm_at_swap_desc(vcpu->kvm, pa, old, new);
243 }
244 
245 /*
246  * This is essentially a C-version of the pseudo code from the ARM ARM
247  * AArch64.TranslationTableWalk  function.  I strongly recommend looking at
248  * that pseudocode in trying to understand this.
249  *
250  * Must be called with the kvm->srcu read lock held
251  */
walk_nested_s2_pgd(struct kvm_vcpu * vcpu,phys_addr_t ipa,struct s2_walk_info * wi,struct kvm_s2_trans * out)252 static int walk_nested_s2_pgd(struct kvm_vcpu *vcpu, phys_addr_t ipa,
253 			      struct s2_walk_info *wi, struct kvm_s2_trans *out)
254 {
255 	int first_block_level, level, stride, input_size, base_lower_bound;
256 	phys_addr_t base_addr;
257 	unsigned int addr_top, addr_bottom;
258 	u64 desc, new_desc;  /* page table entry */
259 	int ret;
260 	phys_addr_t paddr;
261 
262 	switch (BIT(wi->pgshift)) {
263 	default:
264 	case SZ_64K:
265 	case SZ_16K:
266 		level = 3 - wi->sl;
267 		first_block_level = 2;
268 		break;
269 	case SZ_4K:
270 		level = 2 - wi->sl;
271 		first_block_level = 1;
272 		break;
273 	}
274 
275 	stride = wi->pgshift - 3;
276 	input_size = get_ia_size(wi);
277 	if (input_size > 48 || input_size < 25)
278 		return -EFAULT;
279 
280 	ret = check_base_s2_limits(vcpu, wi, level, input_size, stride);
281 	if (WARN_ON(ret)) {
282 		out->esr = compute_fsc(0, ESR_ELx_FSC_FAULT);
283 		return ret;
284 	}
285 
286 	base_lower_bound = 3 + input_size - ((3 - level) * stride +
287 			   wi->pgshift);
288 	base_addr = wi->baddr & GENMASK_ULL(47, base_lower_bound);
289 
290 	if (check_output_size(wi, base_addr)) {
291 		/* R_BFHQH */
292 		out->esr = compute_fsc(0, ESR_ELx_FSC_ADDRSZ);
293 		return 1;
294 	}
295 
296 	addr_top = input_size - 1;
297 
298 	while (1) {
299 		phys_addr_t index;
300 
301 		addr_bottom = (3 - level) * stride + wi->pgshift;
302 		index = (ipa & GENMASK_ULL(addr_top, addr_bottom))
303 			>> (addr_bottom - 3);
304 
305 		paddr = base_addr | index;
306 		ret = read_guest_s2_desc(vcpu, paddr, &desc, wi);
307 		if (ret < 0) {
308 			out->esr = ESR_ELx_FSC_SEA_TTW(level);
309 			return ret;
310 		}
311 
312 		new_desc = desc;
313 
314 		/* Check for valid descriptor at this point */
315 		if (!(desc & KVM_PTE_VALID)) {
316 			out->esr = compute_fsc(level, ESR_ELx_FSC_FAULT);
317 			out->desc = desc;
318 			return 1;
319 		}
320 
321 		if (FIELD_GET(KVM_PTE_TYPE, desc) == KVM_PTE_TYPE_BLOCK) {
322 			if (level < 3)
323 				break;
324 
325 			out->esr = compute_fsc(level, ESR_ELx_FSC_FAULT);
326 			out->desc = desc;
327 			return 1;
328 		}
329 
330 		/* We're at the final level */
331 		if (level == 3)
332 			break;
333 
334 		if (check_output_size(wi, desc)) {
335 			out->esr = compute_fsc(level, ESR_ELx_FSC_ADDRSZ);
336 			out->desc = desc;
337 			return 1;
338 		}
339 
340 		base_addr = desc & GENMASK_ULL(47, wi->pgshift);
341 
342 		level += 1;
343 		addr_top = addr_bottom - 1;
344 	}
345 
346 	if (level < first_block_level) {
347 		out->esr = compute_fsc(level, ESR_ELx_FSC_FAULT);
348 		out->desc = desc;
349 		return 1;
350 	}
351 
352 	if (check_output_size(wi, desc)) {
353 		out->esr = compute_fsc(level, ESR_ELx_FSC_ADDRSZ);
354 		out->desc = desc;
355 		return 1;
356 	}
357 
358 	if (wi->ha)
359 		new_desc |= KVM_PTE_LEAF_ATTR_LO_S2_AF;
360 
361 	if (new_desc != desc) {
362 		ret = swap_guest_s2_desc(vcpu, paddr, desc, new_desc, wi);
363 		if (ret == -EAGAIN)
364 			return ret;
365 		if (ret) {
366 			out->esr = ESR_ELx_FSC_SEA_TTW(level);
367 			out->desc = desc;
368 			return 1;
369 		}
370 
371 		desc = new_desc;
372 	}
373 
374 	if (!(desc & KVM_PTE_LEAF_ATTR_LO_S2_AF)) {
375 		out->esr = compute_fsc(level, ESR_ELx_FSC_ACCESS);
376 		out->desc = desc;
377 		return 1;
378 	}
379 
380 	addr_bottom += contiguous_bit_shift(desc, wi, level);
381 
382 	/* Calculate and return the result */
383 	paddr = (desc & GENMASK_ULL(47, addr_bottom)) |
384 		(ipa & GENMASK_ULL(addr_bottom - 1, 0));
385 	out->output = paddr;
386 	out->block_size = 1UL << ((3 - level) * stride + wi->pgshift);
387 	out->readable = desc & KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
388 	out->writable = desc & KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
389 	out->level = level;
390 	out->desc = desc;
391 	return 0;
392 }
393 
394 #define _has_tgran_2(__r, __sz)						\
395 	({								\
396 		u64 _s1, _s2, _mmfr0 = __r;				\
397 									\
398 		_s2 = SYS_FIELD_GET(ID_AA64MMFR0_EL1,			\
399 				    TGRAN##__sz##_2, _mmfr0);		\
400 									\
401 		_s1 = SYS_FIELD_GET(ID_AA64MMFR0_EL1,			\
402 				    TGRAN##__sz, _mmfr0);		\
403 									\
404 		((_s2 != ID_AA64MMFR0_EL1_TGRAN##__sz##_2_NI &&		\
405 		  _s2 != ID_AA64MMFR0_EL1_TGRAN##__sz##_2_TGRAN##__sz) || \
406 		 (_s2 == ID_AA64MMFR0_EL1_TGRAN##__sz##_2_TGRAN##__sz && \
407 		  _s1 != ID_AA64MMFR0_EL1_TGRAN##__sz##_NI));		\
408 	})
409 
has_tgran_2(u64 mmfr0,unsigned int shift)410 static bool has_tgran_2(u64 mmfr0, unsigned int shift)
411 {
412 	switch (shift) {
413 	case 12:
414 		return _has_tgran_2(mmfr0, 4);
415 	case 14:
416 		return _has_tgran_2(mmfr0, 16);
417 	case 16:
418 		return _has_tgran_2(mmfr0, 64);
419 	default:
420 		BUG();
421 	}
422 }
423 
fallback_tgran2_shift(u64 mmfr0)424 static unsigned int fallback_tgran2_shift(u64 mmfr0)
425 {
426 	if (has_tgran_2(mmfr0, PAGE_SHIFT))
427 		return PAGE_SHIFT;
428 	else if (has_tgran_2(mmfr0, 12))
429 		return 12;
430 	else if (has_tgran_2(mmfr0, 14))
431 		return 14;
432 	else if (has_tgran_2(mmfr0, 16))
433 		return 16;
434 	else
435 		return PAGE_SHIFT;
436 }
437 
vtcr_to_tg0_pgshift(struct kvm * kvm,u64 vtcr)438 static unsigned int vtcr_to_tg0_pgshift(struct kvm *kvm, u64 vtcr)
439 {
440 	u64 tg0 = FIELD_GET(VTCR_EL2_TG0_MASK, vtcr);
441 	u64 mmfr0 = kvm_read_vm_id_reg(kvm, SYS_ID_AA64MMFR0_EL1);
442 	unsigned int shift;
443 
444 	switch (tg0) {
445 	case VTCR_EL2_TG0_4K:
446 		shift = 12;
447 		break;
448 	case VTCR_EL2_TG0_16K:
449 		shift = 14;
450 		break;
451 	case VTCR_EL2_TG0_64K:
452 	/* IMPDEF: treat any other value as 64k, subject to fallback */
453 	default:
454 		shift = 16;
455 	}
456 
457 	/*
458 	 * If TGx is programmed to an unimplemented value (not advertised in
459 	 * ID_AA64MMFR0_EL1), we should treat it as if an implemented value is
460 	 * written, as per the architecture. Choose an available one while
461 	 * prioritizing PAGE_SIZE.
462 	 */
463 	if (!has_tgran_2(mmfr0, shift))
464 		return fallback_tgran2_shift(mmfr0);
465 
466 	return shift;
467 }
468 
vtcr_to_tg0_pgsize(struct kvm * kvm,u64 vtcr)469 static size_t vtcr_to_tg0_pgsize(struct kvm *kvm, u64 vtcr)
470 {
471 	return BIT(vtcr_to_tg0_pgshift(kvm, vtcr));
472 }
473 
setup_s2_walk(struct kvm_vcpu * vcpu,struct s2_walk_info * wi)474 static void setup_s2_walk(struct kvm_vcpu *vcpu, struct s2_walk_info *wi)
475 {
476 	u64 vtcr = vcpu_read_sys_reg(vcpu, VTCR_EL2);
477 
478 	wi->baddr = vcpu_read_sys_reg(vcpu, VTTBR_EL2);
479 	wi->t0sz = vtcr & VTCR_EL2_T0SZ_MASK;
480 	wi->pgshift = vtcr_to_tg0_pgshift(vcpu->kvm, vtcr);
481 	wi->sl = FIELD_GET(VTCR_EL2_SL0_MASK, vtcr);
482 	/* Global limit for now, should eventually be per-VM */
483 	wi->max_oa_bits = min(get_kvm_ipa_limit(),
484 			      ps_to_output_size(FIELD_GET(VTCR_EL2_PS_MASK, vtcr), false));
485 	wi->ha = vtcr & VTCR_EL2_HA;
486 	wi->be = vcpu_read_sys_reg(vcpu, SCTLR_EL2) & SCTLR_ELx_EE;
487 }
488 
kvm_walk_nested_s2(struct kvm_vcpu * vcpu,phys_addr_t gipa,struct kvm_s2_trans * result)489 int kvm_walk_nested_s2(struct kvm_vcpu *vcpu, phys_addr_t gipa,
490 		       struct kvm_s2_trans *result)
491 {
492 	struct s2_walk_info wi;
493 	int ret;
494 
495 	result->esr = 0;
496 
497 	if (!vcpu_has_nv(vcpu))
498 		return 0;
499 
500 	setup_s2_walk(vcpu, &wi);
501 
502 	ret = walk_nested_s2_pgd(vcpu, gipa, &wi, result);
503 	if (ret)
504 		result->esr |= (kvm_vcpu_get_esr(vcpu) & ~ESR_ELx_FSC);
505 
506 	return ret;
507 }
508 
ttl_to_size(u8 ttl)509 static unsigned int ttl_to_size(u8 ttl)
510 {
511 	int level = ttl & 3;
512 	int gran = (ttl >> 2) & 3;
513 	unsigned int max_size = 0;
514 
515 	switch (gran) {
516 	case TLBI_TTL_TG_4K:
517 		switch (level) {
518 		case 0:
519 			break;
520 		case 1:
521 			max_size = SZ_1G;
522 			break;
523 		case 2:
524 			max_size = SZ_2M;
525 			break;
526 		case 3:
527 			max_size = SZ_4K;
528 			break;
529 		}
530 		break;
531 	case TLBI_TTL_TG_16K:
532 		switch (level) {
533 		case 0:
534 		case 1:
535 			break;
536 		case 2:
537 			max_size = SZ_32M;
538 			break;
539 		case 3:
540 			max_size = SZ_16K;
541 			break;
542 		}
543 		break;
544 	case TLBI_TTL_TG_64K:
545 		switch (level) {
546 		case 0:
547 		case 1:
548 			/* No 52bit IPA support */
549 			break;
550 		case 2:
551 			max_size = SZ_512M;
552 			break;
553 		case 3:
554 			max_size = SZ_64K;
555 			break;
556 		}
557 		break;
558 	default:			/* No size information */
559 		break;
560 	}
561 
562 	return max_size;
563 }
564 
pgshift_level_to_ttl(u16 shift,u8 level)565 static u8 pgshift_level_to_ttl(u16 shift, u8 level)
566 {
567 	u8 ttl;
568 
569 	switch(shift) {
570 	case 12:
571 		ttl = TLBI_TTL_TG_4K;
572 		break;
573 	case 14:
574 		ttl = TLBI_TTL_TG_16K;
575 		break;
576 	case 16:
577 		ttl = TLBI_TTL_TG_64K;
578 		break;
579 	default:
580 		BUG();
581 	}
582 
583 	ttl <<= 2;
584 	ttl |= level & 3;
585 
586 	return ttl;
587 }
588 
589 /*
590  * Compute the equivalent of the TTL field by parsing the shadow PT.  The
591  * granule size is extracted from the cached VTCR_EL2.TG0 while the level is
592  * retrieved from first entry carrying the level as a tag.
593  */
get_guest_mapping_ttl(struct kvm_s2_mmu * mmu,u64 addr)594 static u8 get_guest_mapping_ttl(struct kvm_s2_mmu *mmu, u64 addr)
595 {
596 	size_t tg0_size = vtcr_to_tg0_pgsize(kvm_s2_mmu_to_kvm(mmu), mmu->tlb_vtcr);
597 	u64 tmp, sz = 0;
598 	kvm_pte_t pte;
599 	u8 ttl, level;
600 
601 	lockdep_assert_held_write(&kvm_s2_mmu_to_kvm(mmu)->mmu_lock);
602 
603 	switch (tg0_size) {
604 	case SZ_4K:
605 		ttl = (TLBI_TTL_TG_4K << 2);
606 		break;
607 	case SZ_16K:
608 		ttl = (TLBI_TTL_TG_16K << 2);
609 		break;
610 	case SZ_64K:
611 	default:	    /* IMPDEF: treat any other value as 64k */
612 		ttl = (TLBI_TTL_TG_64K << 2);
613 		break;
614 	}
615 
616 	tmp = addr;
617 
618 again:
619 	/* Iteratively compute the block sizes for a particular granule size */
620 	switch (tg0_size) {
621 	case SZ_4K:
622 		if	(sz < SZ_4K)	sz = SZ_4K;
623 		else if (sz < SZ_2M)	sz = SZ_2M;
624 		else if (sz < SZ_1G)	sz = SZ_1G;
625 		else			sz = 0;
626 		break;
627 	case SZ_16K:
628 		if	(sz < SZ_16K)	sz = SZ_16K;
629 		else if (sz < SZ_32M)	sz = SZ_32M;
630 		else			sz = 0;
631 		break;
632 	case SZ_64K:
633 	default:	    /* IMPDEF: treat any other value as 64k */
634 		if	(sz < SZ_64K)	sz = SZ_64K;
635 		else if (sz < SZ_512M)	sz = SZ_512M;
636 		else			sz = 0;
637 		break;
638 	}
639 
640 	if (sz == 0)
641 		return 0;
642 
643 	tmp &= ~(sz - 1);
644 	if (kvm_pgtable_get_leaf(mmu->pgt, tmp, &pte, NULL))
645 		goto again;
646 	if (!(pte & PTE_VALID))
647 		goto again;
648 	level = FIELD_GET(KVM_NV_GUEST_MAP_SZ, pte);
649 	if (!level)
650 		goto again;
651 
652 	ttl |= level;
653 
654 	/*
655 	 * We now have found some level information in the shadow S2. Check
656 	 * that the resulting range is actually including the original IPA.
657 	 */
658 	sz = ttl_to_size(ttl);
659 	if (addr < (tmp + sz))
660 		return ttl;
661 
662 	return 0;
663 }
664 
compute_tlb_inval_range(struct kvm_s2_mmu * mmu,u64 val)665 unsigned long compute_tlb_inval_range(struct kvm_s2_mmu *mmu, u64 val)
666 {
667 	struct kvm *kvm = kvm_s2_mmu_to_kvm(mmu);
668 	unsigned long max_size;
669 	u8 ttl;
670 
671 	ttl = FIELD_GET(TLBI_TTL_MASK, val);
672 
673 	if (!ttl || !kvm_has_feat(kvm, ID_AA64MMFR2_EL1, TTL, IMP)) {
674 		/* No TTL, check the shadow S2 for a hint */
675 		u64 addr = (val & GENMASK_ULL(35, 0)) << 12;
676 		ttl = get_guest_mapping_ttl(mmu, addr);
677 	}
678 
679 	max_size = ttl_to_size(ttl);
680 
681 	if (!max_size) {
682 		/* Compute the maximum extent of the invalidation */
683 		switch (vtcr_to_tg0_pgsize(kvm, mmu->tlb_vtcr)) {
684 		case SZ_4K:
685 			max_size = SZ_1G;
686 			break;
687 		case SZ_16K:
688 			max_size = SZ_32M;
689 			break;
690 		case SZ_64K:
691 		default:    /* IMPDEF: treat any other value as 64k */
692 			/*
693 			 * No, we do not support 52bit IPA in nested yet. Once
694 			 * we do, this should be 4TB.
695 			 */
696 			max_size = SZ_512M;
697 			break;
698 		}
699 	}
700 
701 	WARN_ON(!max_size);
702 	return max_size;
703 }
704 
705 /*
706  * We can have multiple *different* MMU contexts with the same VMID:
707  *
708  * - S2 being enabled or not, hence differing by the HCR_EL2.VM bit
709  *
710  * - Multiple vcpus using private S2s (huh huh...), hence differing by the
711  *   VBBTR_EL2.BADDR address
712  *
713  * - A combination of the above...
714  *
715  * We can always identify which MMU context to pick at run-time.  However,
716  * TLB invalidation involving a VMID must take action on all the TLBs using
717  * this particular VMID. This translates into applying the same invalidation
718  * operation to all the contexts that are using this VMID. Moar phun!
719  */
kvm_s2_mmu_iterate_by_vmid(struct kvm * kvm,u16 vmid,const union tlbi_info * info,void (* tlbi_callback)(struct kvm_s2_mmu *,const union tlbi_info *))720 void kvm_s2_mmu_iterate_by_vmid(struct kvm *kvm, u16 vmid,
721 				const union tlbi_info *info,
722 				void (*tlbi_callback)(struct kvm_s2_mmu *,
723 						      const union tlbi_info *))
724 {
725 	write_lock(&kvm->mmu_lock);
726 
727 	for (int i = 0; i < kvm->arch.nested_mmus_size; i++) {
728 		struct kvm_s2_mmu *mmu = &kvm->arch.nested_mmus[i];
729 
730 		if (!kvm_s2_mmu_valid(mmu))
731 			continue;
732 
733 		if (vmid == get_vmid(mmu->tlb_vttbr))
734 			tlbi_callback(mmu, info);
735 	}
736 
737 	write_unlock(&kvm->mmu_lock);
738 }
739 
lookup_s2_mmu(struct kvm_vcpu * vcpu)740 struct kvm_s2_mmu *lookup_s2_mmu(struct kvm_vcpu *vcpu)
741 {
742 	struct kvm *kvm = vcpu->kvm;
743 	bool nested_stage2_enabled;
744 	u64 vttbr, vtcr, hcr;
745 
746 	lockdep_assert_held_write(&kvm->mmu_lock);
747 
748 	vttbr = vcpu_read_sys_reg(vcpu, VTTBR_EL2);
749 	vtcr = vcpu_read_sys_reg(vcpu, VTCR_EL2);
750 	hcr = vcpu_read_sys_reg(vcpu, HCR_EL2);
751 
752 	nested_stage2_enabled = hcr & HCR_VM;
753 
754 	/* Don't consider the CnP bit for the vttbr match */
755 	vttbr &= ~VTTBR_CNP_BIT;
756 
757 	/*
758 	 * Two possibilities when looking up a S2 MMU context:
759 	 *
760 	 * - either S2 is enabled in the guest, and we need a context that is
761 	 *   S2-enabled and matches the full VTTBR (VMID+BADDR) and VTCR,
762 	 *   which makes it safe from a TLB conflict perspective (a broken
763 	 *   guest won't be able to generate them),
764 	 *
765 	 * - or S2 is disabled, and we need a context that is S2-disabled
766 	 *   and matches the VMID only, as all TLBs are tagged by VMID even
767 	 *   if S2 translation is disabled.
768 	 */
769 	for (int i = 0; i < kvm->arch.nested_mmus_size; i++) {
770 		struct kvm_s2_mmu *mmu = &kvm->arch.nested_mmus[i];
771 
772 		if (!kvm_s2_mmu_valid(mmu))
773 			continue;
774 
775 		if (nested_stage2_enabled &&
776 		    mmu->nested_stage2_enabled &&
777 		    vttbr == mmu->tlb_vttbr &&
778 		    vtcr == mmu->tlb_vtcr)
779 			return mmu;
780 
781 		if (!nested_stage2_enabled &&
782 		    !mmu->nested_stage2_enabled &&
783 		    get_vmid(vttbr) == get_vmid(mmu->tlb_vttbr))
784 			return mmu;
785 	}
786 	return NULL;
787 }
788 
get_s2_mmu_nested(struct kvm_vcpu * vcpu)789 static struct kvm_s2_mmu *get_s2_mmu_nested(struct kvm_vcpu *vcpu)
790 {
791 	struct kvm *kvm = vcpu->kvm;
792 	struct kvm_s2_mmu *s2_mmu;
793 	int i;
794 
795 	lockdep_assert_held_write(&vcpu->kvm->mmu_lock);
796 
797 	s2_mmu = lookup_s2_mmu(vcpu);
798 	if (s2_mmu)
799 		goto out;
800 
801 	/*
802 	 * Make sure we don't always search from the same point, or we
803 	 * will always reuse a potentially active context, leaving
804 	 * free contexts unused.
805 	 */
806 	for (i = kvm->arch.nested_mmus_next;
807 	     i < (kvm->arch.nested_mmus_size + kvm->arch.nested_mmus_next);
808 	     i++) {
809 		s2_mmu = &kvm->arch.nested_mmus[i % kvm->arch.nested_mmus_size];
810 
811 		if (atomic_read(&s2_mmu->refcnt) == 0)
812 			break;
813 	}
814 	BUG_ON(atomic_read(&s2_mmu->refcnt)); /* We have struct MMUs to spare */
815 
816 	/* Set the scene for the next search */
817 	kvm->arch.nested_mmus_next = (i + 1) % kvm->arch.nested_mmus_size;
818 
819 	/* Make sure we don't forget to do the laundry */
820 	if (kvm_s2_mmu_valid(s2_mmu)) {
821 		kvm_nested_s2_ptdump_remove_debugfs(s2_mmu);
822 		s2_mmu->pending_unmap = true;
823 	}
824 
825 	/*
826 	 * The virtual VMID (modulo CnP) will be used as a key when matching
827 	 * an existing kvm_s2_mmu.
828 	 *
829 	 * We cache VTCR at allocation time, once and for all. It'd be great
830 	 * if the guest didn't screw that one up, as this is not very
831 	 * forgiving...
832 	 */
833 	s2_mmu->tlb_vttbr = vcpu_read_sys_reg(vcpu, VTTBR_EL2) & ~VTTBR_CNP_BIT;
834 	s2_mmu->tlb_vtcr = vcpu_read_sys_reg(vcpu, VTCR_EL2);
835 	s2_mmu->nested_stage2_enabled = vcpu_read_sys_reg(vcpu, HCR_EL2) & HCR_VM;
836 
837 	kvm_nested_s2_ptdump_create_debugfs(s2_mmu);
838 
839 out:
840 	atomic_inc(&s2_mmu->refcnt);
841 
842 	/*
843 	 * Set the vCPU request to perform an unmap, even if the pending unmap
844 	 * originates from another vCPU. This guarantees that the MMU has been
845 	 * completely unmapped before any vCPU actually uses it, and allows
846 	 * multiple vCPUs to lend a hand with completing the unmap.
847 	 */
848 	if (s2_mmu->pending_unmap)
849 		kvm_make_request(KVM_REQ_NESTED_S2_UNMAP, vcpu);
850 
851 	return s2_mmu;
852 }
853 
kvm_init_nested_s2_mmu(struct kvm_s2_mmu * mmu)854 void kvm_init_nested_s2_mmu(struct kvm_s2_mmu *mmu)
855 {
856 	/* CnP being set denotes an invalid entry */
857 	mmu->tlb_vttbr = VTTBR_CNP_BIT;
858 	mmu->nested_stage2_enabled = false;
859 	atomic_set(&mmu->refcnt, 0);
860 }
861 
kvm_vcpu_load_hw_mmu(struct kvm_vcpu * vcpu)862 void kvm_vcpu_load_hw_mmu(struct kvm_vcpu *vcpu)
863 {
864 	/*
865 	 * If the vCPU kept its reference on the MMU after the last put,
866 	 * keep rolling with it.
867 	 */
868 	if (is_hyp_ctxt(vcpu)) {
869 		if (!vcpu->arch.hw_mmu)
870 			vcpu->arch.hw_mmu = &vcpu->kvm->arch.mmu;
871 	} else {
872 		if (!vcpu->arch.hw_mmu) {
873 			scoped_guard(write_lock, &vcpu->kvm->mmu_lock)
874 				vcpu->arch.hw_mmu = get_s2_mmu_nested(vcpu);
875 		}
876 
877 		if (__vcpu_sys_reg(vcpu, HCR_EL2) & HCR_NV)
878 			kvm_make_request(KVM_REQ_MAP_L1_VNCR_EL2, vcpu);
879 	}
880 }
881 
this_cpu_reset_vncr_fixmap(struct kvm_vcpu * vcpu)882 static void this_cpu_reset_vncr_fixmap(struct kvm_vcpu *vcpu)
883 {
884 	if (!host_data_test_flag(L1_VNCR_MAPPED))
885 		return;
886 
887 	BUG_ON(vcpu->arch.vncr_tlb->cpu != smp_processor_id());
888 	BUG_ON(is_hyp_ctxt(vcpu));
889 
890 	clear_fixmap(vncr_fixmap(vcpu->arch.vncr_tlb->cpu));
891 	vcpu->arch.vncr_tlb->cpu = -1;
892 	host_data_clear_flag(L1_VNCR_MAPPED);
893 	atomic_dec(&vcpu->kvm->arch.vncr_map_count);
894 }
895 
kvm_vcpu_put_hw_mmu(struct kvm_vcpu * vcpu)896 void kvm_vcpu_put_hw_mmu(struct kvm_vcpu *vcpu)
897 {
898 	/* Unconditionally drop the VNCR mapping if we have one */
899 	this_cpu_reset_vncr_fixmap(vcpu);
900 
901 	/*
902 	 * Keep a reference on the associated stage-2 MMU if the vCPU is
903 	 * scheduling out and not in WFI emulation, suggesting it is likely to
904 	 * reuse the MMU sometime soon.
905 	 */
906 	if (vcpu->scheduled_out && !vcpu_get_flag(vcpu, IN_WFI))
907 		return;
908 
909 	if (kvm_is_nested_s2_mmu(vcpu->kvm, vcpu->arch.hw_mmu))
910 		atomic_dec(&vcpu->arch.hw_mmu->refcnt);
911 
912 	vcpu->arch.hw_mmu = NULL;
913 }
914 
915 /*
916  * Returns non-zero if permission fault is handled by injecting it to the next
917  * level hypervisor.
918  */
kvm_s2_handle_perm_fault(struct kvm_vcpu * vcpu,struct kvm_s2_trans * trans)919 int kvm_s2_handle_perm_fault(struct kvm_vcpu *vcpu, struct kvm_s2_trans *trans)
920 {
921 	bool forward_fault = false;
922 
923 	trans->esr = 0;
924 
925 	if (!kvm_vcpu_trap_is_permission_fault(vcpu))
926 		return 0;
927 
928 	if (kvm_vcpu_trap_is_iabt(vcpu)) {
929 		if (vcpu_mode_priv(vcpu))
930 			forward_fault = !kvm_s2_trans_exec_el1(vcpu->kvm, trans);
931 		else
932 			forward_fault = !kvm_s2_trans_exec_el0(vcpu->kvm, trans);
933 	} else {
934 		bool write_fault = kvm_is_write_fault(vcpu);
935 
936 		forward_fault = ((write_fault && !trans->writable) ||
937 				 (!write_fault && !trans->readable));
938 	}
939 
940 	if (forward_fault)
941 		trans->esr = esr_s2_fault(vcpu, trans->level, ESR_ELx_FSC_PERM);
942 
943 	return forward_fault;
944 }
945 
kvm_inject_s2_fault(struct kvm_vcpu * vcpu,u64 esr_el2)946 int kvm_inject_s2_fault(struct kvm_vcpu *vcpu, u64 esr_el2)
947 {
948 	vcpu_write_sys_reg(vcpu, vcpu->arch.fault.far_el2, FAR_EL2);
949 	vcpu_write_sys_reg(vcpu, vcpu->arch.fault.hpfar_el2, HPFAR_EL2);
950 
951 	return kvm_inject_nested_sync(vcpu, esr_el2);
952 }
953 
get_asid_by_regime(struct kvm_vcpu * vcpu,enum trans_regime regime)954 u16 get_asid_by_regime(struct kvm_vcpu *vcpu, enum trans_regime regime)
955 {
956 	enum vcpu_sysreg ttbr_elx;
957 	u64 tcr;
958 	u16 asid;
959 
960 	switch (regime) {
961 	case TR_EL10:
962 		tcr = vcpu_read_sys_reg(vcpu, TCR_EL1);
963 		ttbr_elx = (tcr & TCR_A1) ? TTBR1_EL1 : TTBR0_EL1;
964 		break;
965 	case TR_EL20:
966 		tcr = vcpu_read_sys_reg(vcpu, TCR_EL2);
967 		ttbr_elx = (tcr & TCR_A1) ? TTBR1_EL2 : TTBR0_EL2;
968 		break;
969 	default:
970 		BUG();
971 	}
972 
973 	asid = FIELD_GET(TTBRx_EL1_ASID, vcpu_read_sys_reg(vcpu, ttbr_elx));
974 	if (!kvm_has_feat_enum(vcpu->kvm, ID_AA64MMFR0_EL1, ASIDBITS, 16) ||
975 	    !(tcr & TCR_ASID16))
976 		asid &= GENMASK(7, 0);
977 
978 	return asid;
979 }
980 
invalidate_vncr(struct vncr_tlb * vt)981 static void invalidate_vncr(struct vncr_tlb *vt)
982 {
983 	vt->valid = false;
984 	if (vt->cpu != -1)
985 		clear_fixmap(vncr_fixmap(vt->cpu));
986 }
987 
988 /*
989  * VNCR TLB invalidation occurs from MMU notifiers or TLBI instructions, and
990  * either can race against a vcpu not being onlined yet (no pseudo-TLB
991  * allocated). Similarly, the TLB might be invalid.  Skip those, as they
992  * obviously don't participate in the invalidation at this stage.
993  */
994 #define kvm_for_each_vncr_tlb(idx, vcpup, tlbp, kvm)	\
995 	kvm_for_each_vcpu(idx, vcpup, kvm)		\
996 		if (((tlbp) = vcpup->arch.vncr_tlb) &&	\
997 		    (tlbp)->valid)
998 
kvm_invalidate_vncr_ipa(struct kvm * kvm,u64 start,u64 end)999 static void kvm_invalidate_vncr_ipa(struct kvm *kvm, u64 start, u64 end)
1000 {
1001 	struct kvm_vcpu *vcpu;
1002 	struct vncr_tlb *vt;
1003 	unsigned long i;
1004 
1005 	lockdep_assert_held_write(&kvm->mmu_lock);
1006 
1007 	if (!kvm_has_feat(kvm, ID_AA64MMFR4_EL1, NV_frac, NV2_ONLY))
1008 		return;
1009 
1010 	kvm_for_each_vncr_tlb(i, vcpu, vt, kvm) {
1011 		u64 ipa_start, ipa_end, ipa_size;
1012 
1013 		ipa_size = ttl_to_size(pgshift_level_to_ttl(vt->wi.pgshift,
1014 							    vt->wr.level));
1015 		ipa_start = vt->wr.pa & ~(ipa_size - 1);
1016 		ipa_end = ipa_start + ipa_size;
1017 
1018 		if (ipa_end <= start || ipa_start >= end)
1019 			continue;
1020 
1021 		invalidate_vncr(vt);
1022 	}
1023 }
1024 
1025 struct s1e2_tlbi_scope {
1026 	enum {
1027 		TLBI_ALL,
1028 		TLBI_VA,
1029 		TLBI_VAA,
1030 		TLBI_ASID,
1031 	} type;
1032 
1033 	u16 asid;
1034 	u64 va;
1035 	u64 size;
1036 };
1037 
invalidate_vncr_va(struct kvm * kvm,struct s1e2_tlbi_scope * scope)1038 static void invalidate_vncr_va(struct kvm *kvm,
1039 			       struct s1e2_tlbi_scope *scope)
1040 {
1041 	struct kvm_vcpu *vcpu;
1042 	struct vncr_tlb *vt;
1043 	unsigned long i;
1044 
1045 	lockdep_assert_held_write(&kvm->mmu_lock);
1046 
1047 	kvm_for_each_vncr_tlb(i, vcpu, vt, kvm) {
1048 		u64 va_start, va_end, va_size;
1049 
1050 		va_size = ttl_to_size(pgshift_level_to_ttl(vt->wi.pgshift,
1051 							   vt->wr.level));
1052 		va_start = vt->gva & ~(va_size - 1);
1053 		va_end = va_start + va_size;
1054 
1055 		switch (scope->type) {
1056 		case TLBI_ALL:
1057 			break;
1058 
1059 		case TLBI_VA:
1060 			if (va_end <= scope->va ||
1061 			    va_start >= (scope->va + scope->size))
1062 				continue;
1063 			if (vt->wr.nG && vt->wr.asid != scope->asid)
1064 				continue;
1065 			break;
1066 
1067 		case TLBI_VAA:
1068 			if (va_end <= scope->va ||
1069 			    va_start >= (scope->va + scope->size))
1070 				continue;
1071 			break;
1072 
1073 		case TLBI_ASID:
1074 			if (!vt->wr.nG || vt->wr.asid != scope->asid)
1075 				continue;
1076 			break;
1077 		}
1078 
1079 		invalidate_vncr(vt);
1080 	}
1081 }
1082 
1083 #define tlbi_va_s1_to_va(v)	(u64)sign_extend64((v) << 12, 48)
1084 
compute_s1_tlbi_range(struct kvm_vcpu * vcpu,u32 inst,u64 val,struct s1e2_tlbi_scope * scope)1085 static void compute_s1_tlbi_range(struct kvm_vcpu *vcpu, u32 inst, u64 val,
1086 				  struct s1e2_tlbi_scope *scope)
1087 {
1088 	switch (inst) {
1089 	case OP_TLBI_ALLE2:
1090 	case OP_TLBI_ALLE2IS:
1091 	case OP_TLBI_ALLE2OS:
1092 	case OP_TLBI_VMALLE1:
1093 	case OP_TLBI_VMALLE1IS:
1094 	case OP_TLBI_VMALLE1OS:
1095 	case OP_TLBI_ALLE2NXS:
1096 	case OP_TLBI_ALLE2ISNXS:
1097 	case OP_TLBI_ALLE2OSNXS:
1098 	case OP_TLBI_VMALLE1NXS:
1099 	case OP_TLBI_VMALLE1ISNXS:
1100 	case OP_TLBI_VMALLE1OSNXS:
1101 		scope->type = TLBI_ALL;
1102 		break;
1103 	case OP_TLBI_VAE2:
1104 	case OP_TLBI_VAE2IS:
1105 	case OP_TLBI_VAE2OS:
1106 	case OP_TLBI_VAE1:
1107 	case OP_TLBI_VAE1IS:
1108 	case OP_TLBI_VAE1OS:
1109 	case OP_TLBI_VAE2NXS:
1110 	case OP_TLBI_VAE2ISNXS:
1111 	case OP_TLBI_VAE2OSNXS:
1112 	case OP_TLBI_VAE1NXS:
1113 	case OP_TLBI_VAE1ISNXS:
1114 	case OP_TLBI_VAE1OSNXS:
1115 	case OP_TLBI_VALE2:
1116 	case OP_TLBI_VALE2IS:
1117 	case OP_TLBI_VALE2OS:
1118 	case OP_TLBI_VALE1:
1119 	case OP_TLBI_VALE1IS:
1120 	case OP_TLBI_VALE1OS:
1121 	case OP_TLBI_VALE2NXS:
1122 	case OP_TLBI_VALE2ISNXS:
1123 	case OP_TLBI_VALE2OSNXS:
1124 	case OP_TLBI_VALE1NXS:
1125 	case OP_TLBI_VALE1ISNXS:
1126 	case OP_TLBI_VALE1OSNXS:
1127 		scope->type = TLBI_VA;
1128 		scope->size = ttl_to_size(FIELD_GET(TLBI_TTL_MASK, val));
1129 		if (!scope->size)
1130 			scope->size = SZ_1G;
1131 		scope->va = tlbi_va_s1_to_va(val) & ~(scope->size - 1);
1132 		scope->asid = FIELD_GET(TLBIR_ASID_MASK, val);
1133 		break;
1134 	case OP_TLBI_ASIDE1:
1135 	case OP_TLBI_ASIDE1IS:
1136 	case OP_TLBI_ASIDE1OS:
1137 	case OP_TLBI_ASIDE1NXS:
1138 	case OP_TLBI_ASIDE1ISNXS:
1139 	case OP_TLBI_ASIDE1OSNXS:
1140 		scope->type = TLBI_ASID;
1141 		scope->asid = FIELD_GET(TLBIR_ASID_MASK, val);
1142 		break;
1143 	case OP_TLBI_VAAE1:
1144 	case OP_TLBI_VAAE1IS:
1145 	case OP_TLBI_VAAE1OS:
1146 	case OP_TLBI_VAAE1NXS:
1147 	case OP_TLBI_VAAE1ISNXS:
1148 	case OP_TLBI_VAAE1OSNXS:
1149 	case OP_TLBI_VAALE1:
1150 	case OP_TLBI_VAALE1IS:
1151 	case OP_TLBI_VAALE1OS:
1152 	case OP_TLBI_VAALE1NXS:
1153 	case OP_TLBI_VAALE1ISNXS:
1154 	case OP_TLBI_VAALE1OSNXS:
1155 		scope->type = TLBI_VAA;
1156 		scope->size = ttl_to_size(FIELD_GET(TLBI_TTL_MASK, val));
1157 		if (!scope->size)
1158 			scope->size = SZ_1G;
1159 		scope->va = tlbi_va_s1_to_va(val) & ~(scope->size - 1);
1160 		break;
1161 	case OP_TLBI_RVAE2:
1162 	case OP_TLBI_RVAE2IS:
1163 	case OP_TLBI_RVAE2OS:
1164 	case OP_TLBI_RVAE1:
1165 	case OP_TLBI_RVAE1IS:
1166 	case OP_TLBI_RVAE1OS:
1167 	case OP_TLBI_RVAE2NXS:
1168 	case OP_TLBI_RVAE2ISNXS:
1169 	case OP_TLBI_RVAE2OSNXS:
1170 	case OP_TLBI_RVAE1NXS:
1171 	case OP_TLBI_RVAE1ISNXS:
1172 	case OP_TLBI_RVAE1OSNXS:
1173 	case OP_TLBI_RVALE2:
1174 	case OP_TLBI_RVALE2IS:
1175 	case OP_TLBI_RVALE2OS:
1176 	case OP_TLBI_RVALE1:
1177 	case OP_TLBI_RVALE1IS:
1178 	case OP_TLBI_RVALE1OS:
1179 	case OP_TLBI_RVALE2NXS:
1180 	case OP_TLBI_RVALE2ISNXS:
1181 	case OP_TLBI_RVALE2OSNXS:
1182 	case OP_TLBI_RVALE1NXS:
1183 	case OP_TLBI_RVALE1ISNXS:
1184 	case OP_TLBI_RVALE1OSNXS:
1185 		scope->type = TLBI_VA;
1186 		scope->va = decode_range_tlbi(val, &scope->size, &scope->asid);
1187 		break;
1188 	case OP_TLBI_RVAAE1:
1189 	case OP_TLBI_RVAAE1IS:
1190 	case OP_TLBI_RVAAE1OS:
1191 	case OP_TLBI_RVAAE1NXS:
1192 	case OP_TLBI_RVAAE1ISNXS:
1193 	case OP_TLBI_RVAAE1OSNXS:
1194 	case OP_TLBI_RVAALE1:
1195 	case OP_TLBI_RVAALE1IS:
1196 	case OP_TLBI_RVAALE1OS:
1197 	case OP_TLBI_RVAALE1NXS:
1198 	case OP_TLBI_RVAALE1ISNXS:
1199 	case OP_TLBI_RVAALE1OSNXS:
1200 		scope->type = TLBI_VAA;
1201 		scope->va = decode_range_tlbi(val, &scope->size, NULL);
1202 		break;
1203 	}
1204 }
1205 
kvm_handle_s1e2_tlbi(struct kvm_vcpu * vcpu,u32 inst,u64 val)1206 void kvm_handle_s1e2_tlbi(struct kvm_vcpu *vcpu, u32 inst, u64 val)
1207 {
1208 	struct s1e2_tlbi_scope scope = {};
1209 
1210 	compute_s1_tlbi_range(vcpu, inst, val, &scope);
1211 
1212 	guard(write_lock)(&vcpu->kvm->mmu_lock);
1213 	invalidate_vncr_va(vcpu->kvm, &scope);
1214 }
1215 
kvm_nested_s2_wp(struct kvm * kvm)1216 void kvm_nested_s2_wp(struct kvm *kvm)
1217 {
1218 	int i;
1219 
1220 	lockdep_assert_held_write(&kvm->mmu_lock);
1221 
1222 	if (!kvm->arch.nested_mmus_size)
1223 		return;
1224 
1225 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1226 		struct kvm_s2_mmu *mmu = &kvm->arch.nested_mmus[i];
1227 
1228 		if (kvm_s2_mmu_valid(mmu))
1229 			kvm_stage2_wp_range(mmu, 0, kvm_phys_size(mmu));
1230 	}
1231 
1232 	kvm_invalidate_vncr_ipa(kvm, 0, BIT(kvm->arch.mmu.pgt->ia_bits));
1233 }
1234 
kvm_nested_s2_unmap(struct kvm * kvm,bool may_block)1235 void kvm_nested_s2_unmap(struct kvm *kvm, bool may_block)
1236 {
1237 	int i;
1238 
1239 	lockdep_assert_held_write(&kvm->mmu_lock);
1240 
1241 	if (!kvm->arch.nested_mmus_size)
1242 		return;
1243 
1244 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1245 		struct kvm_s2_mmu *mmu = &kvm->arch.nested_mmus[i];
1246 
1247 		if (kvm_s2_mmu_valid(mmu))
1248 			kvm_stage2_unmap_range(mmu, 0, kvm_phys_size(mmu), may_block);
1249 	}
1250 
1251 	kvm_invalidate_vncr_ipa(kvm, 0, BIT(kvm->arch.mmu.pgt->ia_bits));
1252 }
1253 
kvm_nested_s2_flush(struct kvm * kvm)1254 void kvm_nested_s2_flush(struct kvm *kvm)
1255 {
1256 	int i;
1257 
1258 	lockdep_assert_held_write(&kvm->mmu_lock);
1259 
1260 	if (!kvm->arch.nested_mmus_size)
1261 		return;
1262 
1263 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1264 		struct kvm_s2_mmu *mmu = &kvm->arch.nested_mmus[i];
1265 
1266 		if (kvm_s2_mmu_valid(mmu))
1267 			kvm_stage2_flush_range(mmu, 0, kvm_phys_size(mmu));
1268 	}
1269 }
1270 
kvm_arch_flush_shadow_all(struct kvm * kvm)1271 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1272 {
1273 	int i;
1274 
1275 	for (i = 0; i < kvm->arch.nested_mmus_size; i++) {
1276 		struct kvm_s2_mmu *mmu = &kvm->arch.nested_mmus[i];
1277 
1278 		if (!WARN_ON(atomic_read(&mmu->refcnt)))
1279 			kvm_free_stage2_pgd(mmu);
1280 	}
1281 	kvfree(kvm->arch.nested_mmus);
1282 	kvm->arch.nested_mmus = NULL;
1283 	kvm->arch.nested_mmus_size = 0;
1284 	kvm_uninit_stage2_mmu(kvm);
1285 }
1286 
1287 /*
1288  * Dealing with VNCR_EL2 exposed by the *guest* is a complicated matter:
1289  *
1290  * - We introduce an internal representation of a vcpu-private TLB,
1291  *   representing the mapping between the guest VA contained in VNCR_EL2,
1292  *   the IPA the guest's EL2 PTs point to, and the actual PA this lives at.
1293  *
1294  * - On translation fault from a nested VNCR access, we create such a TLB.
1295  *   If there is no mapping to describe, the guest inherits the fault.
1296  *   Crucially, no actual mapping is done at this stage.
1297  *
1298  * - On vcpu_load() in a non-HYP context with HCR_EL2.NV==1, if the above
1299  *   TLB exists, we map it in the fixmap for this CPU, and run with it. We
1300  *   have to respect the permissions dictated by the guest, but not the
1301  *   memory type (FWB is a must).
1302  *
1303  * - Note that we usually don't do a vcpu_load() on the back of a fault
1304  *   (unless we are preempted), so the resolution of a translation fault
1305  *   must go via a request that will map the VNCR page in the fixmap.
1306  *   vcpu_load() might as well use the same mechanism.
1307  *
1308  * - On vcpu_put() in a non-HYP context with HCR_EL2.NV==1, if the TLB was
1309  *   mapped, we unmap it. Yes it is that simple. The TLB still exists
1310  *   though, and may be reused at a later load.
1311  *
1312  * - On permission fault, we simply forward the fault to the guest's EL2.
1313  *   Get out of my way.
1314  *
1315  * - On any TLBI for the EL2&0 translation regime, we must find any TLB that
1316  *   intersects with the TLBI request, invalidate it, and unmap the page
1317  *   from the fixmap. Because we need to look at all the vcpu-private TLBs,
1318  *   this requires some wide-ranging locking to ensure that nothing races
1319  *   against it. This may require some refcounting to avoid the search when
1320  *   no such TLB is present.
1321  *
1322  * - On MMU notifiers, we must invalidate our TLB in a similar way, but
1323  *   looking at the IPA instead. The funny part is that there may not be a
1324  *   stage-2 mapping for this page if L1 hasn't accessed it using LD/ST
1325  *   instructions.
1326  */
1327 
kvm_vcpu_allocate_vncr_tlb(struct kvm_vcpu * vcpu)1328 int kvm_vcpu_allocate_vncr_tlb(struct kvm_vcpu *vcpu)
1329 {
1330 	if (!kvm_has_feat(vcpu->kvm, ID_AA64MMFR4_EL1, NV_frac, NV2_ONLY))
1331 		return 0;
1332 
1333 	if (!vcpu->arch.vncr_tlb) {
1334 		struct vncr_tlb *vt = kzalloc_obj(*vcpu->arch.vncr_tlb,
1335 						  GFP_KERNEL_ACCOUNT);
1336 
1337 		/*
1338 		 * Taking the lock on assignment ensures that the TLB is
1339 		 * seen as initialised when following the pointer (release
1340 		 * semantics of the unlock), and avoids having acquires on
1341 		 * each user which already take the lock.
1342 		 */
1343 		scoped_guard(write_lock, &vcpu->kvm->mmu_lock)
1344 			vcpu->arch.vncr_tlb = vt;
1345 	}
1346 
1347 	if (!vcpu->arch.vncr_tlb)
1348 		return -ENOMEM;
1349 
1350 	return 0;
1351 }
1352 
read_vncr_el2(struct kvm_vcpu * vcpu)1353 static u64 read_vncr_el2(struct kvm_vcpu *vcpu)
1354 {
1355 	return (u64)sign_extend64(__vcpu_sys_reg(vcpu, VNCR_EL2), 48);
1356 }
1357 
kvm_translate_vncr(struct kvm_vcpu * vcpu,bool * is_gmem)1358 static int kvm_translate_vncr(struct kvm_vcpu *vcpu, bool *is_gmem)
1359 {
1360 	struct kvm_memory_slot *memslot;
1361 	bool write_fault, writable;
1362 	unsigned long mmu_seq;
1363 	struct vncr_tlb *vt;
1364 	struct page *page;
1365 	u64 va, pfn, gfn;
1366 	int ret;
1367 
1368 	vt = vcpu->arch.vncr_tlb;
1369 
1370 	/*
1371 	 * If we're about to walk the EL2 S1 PTs, we must invalidate the
1372 	 * current TLB, as it could be sampled from another vcpu doing a
1373 	 * TLBI *IS. A real CPU wouldn't do that, but we only keep a single
1374 	 * translation, so not much of a choice.
1375 	 *
1376 	 * We also prepare the next walk wilst we're at it.
1377 	 */
1378 	scoped_guard(write_lock, &vcpu->kvm->mmu_lock) {
1379 		this_cpu_reset_vncr_fixmap(vcpu);
1380 		vt->valid = false;
1381 
1382 		vt->wi = (struct s1_walk_info) {
1383 			.regime	= TR_EL20,
1384 			.as_el0	= false,
1385 			.pan	= false,
1386 		};
1387 		vt->wr = (struct s1_walk_result){};
1388 	}
1389 
1390 	guard(srcu)(&vcpu->kvm->srcu);
1391 
1392 	va =  read_vncr_el2(vcpu);
1393 
1394 	ret = __kvm_translate_va(vcpu, &vt->wi, &vt->wr, va);
1395 	if (ret)
1396 		return ret;
1397 
1398 	write_fault = kvm_is_write_fault(vcpu);
1399 
1400 	mmu_seq = vcpu->kvm->mmu_invalidate_seq;
1401 	smp_rmb();
1402 
1403 	gfn = vt->wr.pa >> PAGE_SHIFT;
1404 	memslot = gfn_to_memslot(vcpu->kvm, gfn);
1405 	if (!memslot) {
1406 		fail_s1_walk(&vt->wr, ESR_ELx_FSC_EXTABT, false);
1407 		return -EFAULT;
1408 	}
1409 
1410 	*is_gmem = kvm_slot_has_gmem(memslot);
1411 	if (!*is_gmem) {
1412 		pfn = __kvm_faultin_pfn(memslot, gfn, write_fault ? FOLL_WRITE : 0,
1413 					&writable, &page);
1414 		if (is_error_noslot_pfn(pfn)) {
1415 			fail_s1_walk(&vt->wr, ESR_ELx_FSC_EXTABT, false);
1416 			return -EFAULT;
1417 		}
1418 	} else {
1419 		ret = kvm_gmem_get_pfn(vcpu->kvm, memslot, gfn, &pfn, &page, NULL);
1420 		if (ret) {
1421 			kvm_prepare_memory_fault_exit(vcpu, vt->wr.pa, PAGE_SIZE,
1422 					      write_fault, false, false);
1423 			return ret;
1424 		}
1425 
1426 		writable = !(memslot->flags & KVM_MEM_READONLY);
1427 	}
1428 
1429 	/*
1430 	 * FIXME: This check is too restrictive as KVM allows cacheable memory
1431 	 * attributes for PFNMAP VMAs that have cacheable attributes in host
1432 	 * stage-1.
1433 	 */
1434 	if (!pfn_is_map_memory(pfn)) {
1435 		kvm_release_faultin_page(vcpu->kvm, page, true, false);
1436 		fail_s1_walk(&vt->wr, ESR_ELx_FSC_EXTABT, false);
1437 		return -EINVAL;
1438 	}
1439 
1440 	scoped_guard(write_lock, &vcpu->kvm->mmu_lock) {
1441 		if (mmu_invalidate_retry(vcpu->kvm, mmu_seq)) {
1442 			kvm_release_faultin_page(vcpu->kvm, page, true, false);
1443 			return -EAGAIN;
1444 		}
1445 
1446 		vt->gva = va;
1447 		vt->hpa = pfn << PAGE_SHIFT;
1448 		vt->hpa_writable = writable;
1449 		vt->valid = true;
1450 		vt->cpu = -1;
1451 
1452 		kvm_make_request(KVM_REQ_MAP_L1_VNCR_EL2, vcpu);
1453 		kvm_release_faultin_page(vcpu->kvm, page, false, vt->wr.pw && vt->hpa_writable);
1454 	}
1455 
1456 	if (vt->wr.pw && vt->hpa_writable)
1457 		mark_page_dirty(vcpu->kvm, gfn);
1458 
1459 	return 0;
1460 }
1461 
handle_vncr_perm(struct kvm_vcpu * vcpu)1462 static void handle_vncr_perm(struct kvm_vcpu *vcpu)
1463 {
1464 	struct vncr_tlb *vt = vcpu->arch.vncr_tlb;
1465 	u64 esr = kvm_vcpu_get_esr(vcpu);
1466 	u64 fsc;
1467 
1468 	/*
1469 	 * Promote to an external abort if the stage-1 permits writes but the
1470 	 * HPA is read-only (e.g. RO memslot).
1471 	 */
1472 	if (kvm_is_write_fault(vcpu) && vt->wr.pw && !vt->hpa_writable)
1473 		fsc = ESR_ELx_FSC_EXTABT;
1474 	/*
1475 	 * Otherwise, inject a permission fault using the guest's translation
1476 	 * level rather than the host's.
1477 	 */
1478 	else
1479 		fsc = ESR_ELx_FSC_PERM_L(vt->wr.level);
1480 
1481 	esr &= ~ESR_ELx_FSC;
1482 	esr |= FIELD_PREP(ESR_ELx_FSC, fsc);
1483 
1484 	kvm_inject_nested_sync(vcpu, esr);
1485 }
1486 
kvm_handle_vncr_abort(struct kvm_vcpu * vcpu)1487 int kvm_handle_vncr_abort(struct kvm_vcpu *vcpu)
1488 {
1489 	struct vncr_tlb *vt = vcpu->arch.vncr_tlb;
1490 	u64 esr = kvm_vcpu_get_esr(vcpu);
1491 	bool is_gmem = false;
1492 	bool perm;
1493 	int ret;
1494 
1495 	WARN_ON_ONCE(!(esr & ESR_ELx_VNCR));
1496 
1497 	if (kvm_vcpu_abt_issea(vcpu))
1498 		return kvm_handle_guest_sea(vcpu);
1499 
1500 	if (!esr_fsc_is_translation_fault(esr) && !esr_fsc_is_permission_fault(esr)) {
1501 		KVM_BUG(1, vcpu->kvm, "Unhandled VNCR abort, ESR=%llx\n", esr);
1502 		return -EIO;
1503 	}
1504 
1505 	ret = kvm_translate_vncr(vcpu, &is_gmem);
1506 	switch (ret) {
1507 	case -EAGAIN:
1508 		/* Let's try again... */
1509 		return 1;
1510 	case -ENOMEM:
1511 		/*
1512 		 * For guest_memfd, this indicates that it failed to
1513 		 * create a folio to back the memory. Inform userspace.
1514 		 */
1515 		if (is_gmem)
1516 			return 0;
1517 		/* Otherwise, let's try again... */
1518 		break;
1519 	case -EFAULT:
1520 	case -EIO:
1521 	case -EHWPOISON:
1522 		if (is_gmem)
1523 			return 0;
1524 		fallthrough;
1525 	case -EINVAL:
1526 	case -ENOENT:
1527 	case -EACCES:
1528 		/*
1529 		 * Translation failed, inject the corresponding
1530 		 * exception back to EL2.
1531 		 */
1532 		esr &= ~ESR_ELx_FSC;
1533 		esr |= FIELD_PREP(ESR_ELx_FSC, vt->wr.fst);
1534 
1535 		kvm_inject_nested_sync(vcpu, esr);
1536 		break;
1537 	case 0:
1538 		perm = kvm_is_write_fault(vcpu) ? vt->wr.pw && vt->hpa_writable : vt->wr.pr;
1539 		if (!perm)
1540 			handle_vncr_perm(vcpu);
1541 		break;
1542 	}
1543 
1544 	return 1;
1545 }
1546 
kvm_map_l1_vncr(struct kvm_vcpu * vcpu)1547 static void kvm_map_l1_vncr(struct kvm_vcpu *vcpu)
1548 {
1549 	struct vncr_tlb *vt = vcpu->arch.vncr_tlb;
1550 	pgprot_t prot;
1551 
1552 	guard(preempt)();
1553 	guard(read_lock)(&vcpu->kvm->mmu_lock);
1554 
1555 	/*
1556 	 * The request to map VNCR may have raced against some other
1557 	 * event, such as an interrupt, and may not be valid anymore.
1558 	 */
1559 	if (is_hyp_ctxt(vcpu))
1560 		return;
1561 
1562 	/*
1563 	 * Check that the pseudo-TLB is valid and that VNCR_EL2 still
1564 	 * contains the expected value. If it doesn't, we simply bail out
1565 	 * without a mapping -- a transformed MSR/MRS will generate the
1566 	 * fault and allows us to populate the pseudo-TLB.
1567 	 */
1568 	if (!vt->valid)
1569 		return;
1570 
1571 	if (read_vncr_el2(vcpu) != vt->gva)
1572 		return;
1573 
1574 	if (vt->wr.nG && get_asid_by_regime(vcpu, TR_EL20) != vt->wr.asid)
1575 		return;
1576 
1577 	vt->cpu = smp_processor_id();
1578 
1579 	if (vt->hpa_writable && vt->wr.pw && vt->wr.pr)
1580 		prot = PAGE_KERNEL;
1581 	else if (vt->wr.pr)
1582 		prot = PAGE_KERNEL_RO;
1583 	else
1584 		prot = PAGE_NONE;
1585 
1586 	/*
1587 	 * We can't map write-only (or no permission at all) in the kernel,
1588 	 * but the guest can do it if using POE, so we'll have to turn a
1589 	 * translation fault into a permission fault at runtime.
1590 	 * FIXME: WO doesn't work at all, need POE support in the kernel.
1591 	 */
1592 	if (pgprot_val(prot) != pgprot_val(PAGE_NONE)) {
1593 		__set_fixmap(vncr_fixmap(vt->cpu), vt->hpa, prot);
1594 		host_data_set_flag(L1_VNCR_MAPPED);
1595 		atomic_inc(&vcpu->kvm->arch.vncr_map_count);
1596 	}
1597 }
1598 
1599 /*
1600  * Our emulated CPU doesn't support all the possible features. For the
1601  * sake of simplicity (and probably mental sanity), wipe out a number
1602  * of feature bits we don't intend to support for the time being.
1603  * This list should get updated as new features get added to the NV
1604  * support, and new extension to the architecture.
1605  */
limit_nv_id_reg(struct kvm * kvm,u32 reg,u64 val)1606 u64 limit_nv_id_reg(struct kvm *kvm, u32 reg, u64 val)
1607 {
1608 	u64 orig_val = val;
1609 
1610 	switch (reg) {
1611 	case SYS_ID_AA64ISAR1_EL1:
1612 		/* Support everything but LS64 and Spec Invalidation */
1613 		val &= ~(ID_AA64ISAR1_EL1_LS64	|
1614 			 ID_AA64ISAR1_EL1_SPECRES);
1615 		break;
1616 
1617 	case SYS_ID_AA64PFR0_EL1:
1618 		/* No RME, AMU, MPAM, or S-EL2 */
1619 		val &= ~(ID_AA64PFR0_EL1_RME	|
1620 			 ID_AA64PFR0_EL1_AMU	|
1621 			 ID_AA64PFR0_EL1_MPAM	|
1622 			 ID_AA64PFR0_EL1_SEL2	|
1623 			 ID_AA64PFR0_EL1_EL3	|
1624 			 ID_AA64PFR0_EL1_EL2	|
1625 			 ID_AA64PFR0_EL1_EL1	|
1626 			 ID_AA64PFR0_EL1_EL0);
1627 		/* 64bit only at any EL */
1628 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL0, IMP);
1629 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL1, IMP);
1630 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL2, IMP);
1631 		val |= SYS_FIELD_PREP_ENUM(ID_AA64PFR0_EL1, EL3, IMP);
1632 		break;
1633 
1634 	case SYS_ID_AA64PFR1_EL1:
1635 		/* Only support BTI, SSBS, CSV2_frac */
1636 		val &= ~(ID_AA64PFR1_EL1_PFAR		|
1637 			 ID_AA64PFR1_EL1_MTEX		|
1638 			 ID_AA64PFR1_EL1_THE		|
1639 			 ID_AA64PFR1_EL1_GCS		|
1640 			 ID_AA64PFR1_EL1_MTE_frac	|
1641 			 ID_AA64PFR1_EL1_NMI		|
1642 			 ID_AA64PFR1_EL1_SME		|
1643 			 ID_AA64PFR1_EL1_RES0		|
1644 			 ID_AA64PFR1_EL1_MPAM_frac	|
1645 			 ID_AA64PFR1_EL1_MTE);
1646 		break;
1647 
1648 	case SYS_ID_AA64PFR2_EL1:
1649 		/* GICv5 is not yet supported for NV */
1650 		val &= ~ID_AA64PFR2_EL1_GCIE;
1651 		break;
1652 
1653 	case SYS_ID_AA64MMFR0_EL1:
1654 		/* Hide ExS, Secure Memory */
1655 		val &= ~(ID_AA64MMFR0_EL1_EXS		|
1656 			 ID_AA64MMFR0_EL1_TGRAN4_2	|
1657 			 ID_AA64MMFR0_EL1_TGRAN16_2	|
1658 			 ID_AA64MMFR0_EL1_TGRAN64_2	|
1659 			 ID_AA64MMFR0_EL1_SNSMEM);
1660 
1661 		/* Hide CNTPOFF if present */
1662 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR0_EL1, ECV, IMP);
1663 
1664 		/* Disallow unsupported S2 page sizes */
1665 		switch (PAGE_SIZE) {
1666 		case SZ_64K:
1667 			val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN16_2, NI);
1668 			fallthrough;
1669 		case SZ_16K:
1670 			val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN4_2, NI);
1671 			fallthrough;
1672 		case SZ_4K:
1673 			/* Support everything */
1674 			break;
1675 		}
1676 
1677 		/*
1678 		 * Since we can't support a guest S2 page size smaller
1679 		 * than the host's own page size (due to KVM only
1680 		 * populating its own S2 using the kernel's page
1681 		 * size), advertise the limitation using FEAT_GTG.
1682 		 */
1683 		switch (PAGE_SIZE) {
1684 		case SZ_4K:
1685 			if (_has_tgran_2(orig_val, 4))
1686 				val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN4_2, IMP);
1687 			fallthrough;
1688 		case SZ_16K:
1689 			if (_has_tgran_2(orig_val, 16))
1690 				val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN16_2, IMP);
1691 			fallthrough;
1692 		case SZ_64K:
1693 			if (_has_tgran_2(orig_val, 64))
1694 				val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR0_EL1, TGRAN64_2, IMP);
1695 			break;
1696 		}
1697 
1698 		/* Cap PARange to 48bits */
1699 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR0_EL1, PARANGE, 48);
1700 		break;
1701 
1702 	case SYS_ID_AA64MMFR1_EL1:
1703 		val &= ~(ID_AA64MMFR1_EL1_CMOW		|
1704 			 ID_AA64MMFR1_EL1_nTLBPA	|
1705 			 ID_AA64MMFR1_EL1_ETS);
1706 
1707 		/* FEAT_E2H0 implies no VHE */
1708 		if (test_bit(KVM_ARM_VCPU_HAS_EL2_E2H0, kvm->arch.vcpu_features))
1709 			val &= ~ID_AA64MMFR1_EL1_VH;
1710 
1711 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64MMFR1_EL1, HAFDBS, AF);
1712 		break;
1713 
1714 	case SYS_ID_AA64MMFR2_EL1:
1715 		val &= ~(ID_AA64MMFR2_EL1_BBM	|
1716 			 ID_AA64MMFR2_EL1_TTL	|
1717 			 GENMASK_ULL(47, 44)	|
1718 			 ID_AA64MMFR2_EL1_ST	|
1719 			 ID_AA64MMFR2_EL1_CCIDX	|
1720 			 ID_AA64MMFR2_EL1_VARange);
1721 
1722 		/* Force TTL support */
1723 		val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR2_EL1, TTL, IMP);
1724 		break;
1725 
1726 	case SYS_ID_AA64MMFR4_EL1:
1727 		/*
1728 		 * You get EITHER
1729 		 *
1730 		 * - FEAT_VHE without FEAT_E2H0
1731 		 * - FEAT_NV limited to FEAT_NV2
1732 		 * - HCR_EL2.NV1 being RES0
1733 		 *
1734 		 * OR
1735 		 *
1736 		 * - FEAT_E2H0 without FEAT_VHE nor FEAT_NV
1737 		 *
1738 		 * Life is too short for anything else.
1739 		 */
1740 		if (test_bit(KVM_ARM_VCPU_HAS_EL2_E2H0, kvm->arch.vcpu_features)) {
1741 			val = 0;
1742 		} else {
1743 			val = SYS_FIELD_PREP_ENUM(ID_AA64MMFR4_EL1, NV_frac, NV2_ONLY);
1744 			val |= SYS_FIELD_PREP_ENUM(ID_AA64MMFR4_EL1, E2H0, NI_NV1);
1745 		}
1746 		break;
1747 
1748 	case SYS_ID_AA64DFR0_EL1:
1749 		/* Only limited support for PMU, Debug, BPs, WPs, and HPMN0 */
1750 		val &= ~(ID_AA64DFR0_EL1_ExtTrcBuff	|
1751 			 ID_AA64DFR0_EL1_BRBE		|
1752 			 ID_AA64DFR0_EL1_MTPMU		|
1753 			 ID_AA64DFR0_EL1_TraceBuffer	|
1754 			 ID_AA64DFR0_EL1_TraceFilt	|
1755 			 ID_AA64DFR0_EL1_PMSVer		|
1756 			 ID_AA64DFR0_EL1_CTX_CMPs	|
1757 			 ID_AA64DFR0_EL1_SEBEP		|
1758 			 ID_AA64DFR0_EL1_PMSS		|
1759 			 ID_AA64DFR0_EL1_TraceVer);
1760 
1761 		/*
1762 		 * FEAT_Debugv8p9 requires support for extended breakpoints /
1763 		 * watchpoints.
1764 		 */
1765 		val = ID_REG_LIMIT_FIELD_ENUM(val, ID_AA64DFR0_EL1, DebugVer, V8P8);
1766 		break;
1767 	}
1768 
1769 	return val;
1770 }
1771 
kvm_vcpu_apply_reg_masks(const struct kvm_vcpu * vcpu,enum vcpu_sysreg sr,u64 v)1772 u64 kvm_vcpu_apply_reg_masks(const struct kvm_vcpu *vcpu,
1773 			     enum vcpu_sysreg sr, u64 v)
1774 {
1775 	struct resx resx;
1776 
1777 	resx = kvm_get_sysreg_resx(vcpu->kvm, sr);
1778 	v &= ~resx.res0;
1779 	v |= resx.res1;
1780 
1781 	return v;
1782 }
1783 
set_sysreg_masks(struct kvm * kvm,int sr,struct resx resx)1784 static __always_inline void set_sysreg_masks(struct kvm *kvm, int sr, struct resx resx)
1785 {
1786 	BUILD_BUG_ON(!__builtin_constant_p(sr));
1787 	BUILD_BUG_ON(sr < __SANITISED_REG_START__);
1788 	BUILD_BUG_ON(sr >= NR_SYS_REGS);
1789 
1790 	kvm_set_sysreg_resx(kvm, sr, resx);
1791 }
1792 
kvm_init_nv_sysregs(struct kvm_vcpu * vcpu)1793 int kvm_init_nv_sysregs(struct kvm_vcpu *vcpu)
1794 {
1795 	struct kvm *kvm = vcpu->kvm;
1796 	struct resx resx;
1797 
1798 	lockdep_assert_held(&kvm->arch.config_lock);
1799 
1800 	if (kvm->arch.sysreg_masks)
1801 		goto out;
1802 
1803 	kvm->arch.sysreg_masks = kzalloc_obj(*(kvm->arch.sysreg_masks),
1804 					     GFP_KERNEL_ACCOUNT);
1805 	if (!kvm->arch.sysreg_masks)
1806 		return -ENOMEM;
1807 
1808 	/* VTTBR_EL2 */
1809 	resx = (typeof(resx)){};
1810 	if (!kvm_has_feat_enum(kvm, ID_AA64MMFR1_EL1, VMIDBits, 16))
1811 		resx.res0 |= GENMASK(63, 56);
1812 	if (!kvm_has_feat(kvm, ID_AA64MMFR2_EL1, CnP, IMP))
1813 		resx.res0 |= VTTBR_CNP_BIT;
1814 	set_sysreg_masks(kvm, VTTBR_EL2, resx);
1815 
1816 	/* VTCR_EL2 */
1817 	resx = get_reg_fixed_bits(kvm, VTCR_EL2);
1818 	set_sysreg_masks(kvm, VTCR_EL2, resx);
1819 
1820 	/* VMPIDR_EL2 */
1821 	resx.res0 = GENMASK(63, 40) | GENMASK(30, 24);
1822 	resx.res1 = BIT(31);
1823 	set_sysreg_masks(kvm, VMPIDR_EL2, resx);
1824 
1825 	/* HCR_EL2 */
1826 	resx = get_reg_fixed_bits(kvm, HCR_EL2);
1827 	set_sysreg_masks(kvm, HCR_EL2, resx);
1828 
1829 	/* HCRX_EL2 */
1830 	resx = get_reg_fixed_bits(kvm, HCRX_EL2);
1831 	set_sysreg_masks(kvm, HCRX_EL2, resx);
1832 
1833 	/* HFG[RW]TR_EL2 */
1834 	resx = get_reg_fixed_bits(kvm, HFGRTR_EL2);
1835 	set_sysreg_masks(kvm, HFGRTR_EL2, resx);
1836 	resx = get_reg_fixed_bits(kvm, HFGWTR_EL2);
1837 	set_sysreg_masks(kvm, HFGWTR_EL2, resx);
1838 
1839 	/* HDFG[RW]TR_EL2 */
1840 	resx = get_reg_fixed_bits(kvm, HDFGRTR_EL2);
1841 	set_sysreg_masks(kvm, HDFGRTR_EL2, resx);
1842 	resx = get_reg_fixed_bits(kvm, HDFGWTR_EL2);
1843 	set_sysreg_masks(kvm, HDFGWTR_EL2, resx);
1844 
1845 	/* HFGITR_EL2 */
1846 	resx = get_reg_fixed_bits(kvm, HFGITR_EL2);
1847 	set_sysreg_masks(kvm, HFGITR_EL2, resx);
1848 
1849 	/* HAFGRTR_EL2 - not a lot to see here */
1850 	resx = get_reg_fixed_bits(kvm, HAFGRTR_EL2);
1851 	set_sysreg_masks(kvm, HAFGRTR_EL2, resx);
1852 
1853 	/* HFG[RW]TR2_EL2 */
1854 	resx = get_reg_fixed_bits(kvm, HFGRTR2_EL2);
1855 	set_sysreg_masks(kvm, HFGRTR2_EL2, resx);
1856 	resx = get_reg_fixed_bits(kvm, HFGWTR2_EL2);
1857 	set_sysreg_masks(kvm, HFGWTR2_EL2, resx);
1858 
1859 	/* HDFG[RW]TR2_EL2 */
1860 	resx = get_reg_fixed_bits(kvm, HDFGRTR2_EL2);
1861 	set_sysreg_masks(kvm, HDFGRTR2_EL2, resx);
1862 	resx = get_reg_fixed_bits(kvm, HDFGWTR2_EL2);
1863 	set_sysreg_masks(kvm, HDFGWTR2_EL2, resx);
1864 
1865 	/* HFGITR2_EL2 */
1866 	resx = get_reg_fixed_bits(kvm, HFGITR2_EL2);
1867 	set_sysreg_masks(kvm, HFGITR2_EL2, resx);
1868 
1869 	/* TCR2_EL2 */
1870 	resx = get_reg_fixed_bits(kvm, TCR2_EL2);
1871 	set_sysreg_masks(kvm, TCR2_EL2, resx);
1872 
1873 	/* SCTLR_EL1 */
1874 	resx = get_reg_fixed_bits(kvm, SCTLR_EL1);
1875 	set_sysreg_masks(kvm, SCTLR_EL1, resx);
1876 
1877 	/* SCTLR_EL2 */
1878 	resx = get_reg_fixed_bits(kvm, SCTLR_EL2);
1879 	set_sysreg_masks(kvm, SCTLR_EL2, resx);
1880 
1881 	/* SCTLR2_ELx */
1882 	resx = get_reg_fixed_bits(kvm, SCTLR2_EL1);
1883 	set_sysreg_masks(kvm, SCTLR2_EL1, resx);
1884 	resx = get_reg_fixed_bits(kvm, SCTLR2_EL2);
1885 	set_sysreg_masks(kvm, SCTLR2_EL2, resx);
1886 
1887 	/* MDCR_EL2 */
1888 	resx = get_reg_fixed_bits(kvm, MDCR_EL2);
1889 	set_sysreg_masks(kvm, MDCR_EL2, resx);
1890 
1891 	/* CNTHCTL_EL2 */
1892 	resx.res0 = GENMASK(63, 20);
1893 	resx.res1 = 0;
1894 	if (!kvm_has_feat(kvm, ID_AA64PFR0_EL1, RME, IMP))
1895 		resx.res0 |= CNTHCTL_CNTPMASK | CNTHCTL_CNTVMASK;
1896 	if (!kvm_has_feat(kvm, ID_AA64MMFR0_EL1, ECV, CNTPOFF)) {
1897 		resx.res0 |= CNTHCTL_ECV;
1898 		if (!kvm_has_feat(kvm, ID_AA64MMFR0_EL1, ECV, IMP))
1899 			resx.res0 |= (CNTHCTL_EL1TVT | CNTHCTL_EL1TVCT |
1900 				      CNTHCTL_EL1NVPCT | CNTHCTL_EL1NVVCT);
1901 	}
1902 	if (!kvm_has_feat(kvm, ID_AA64MMFR1_EL1, VH, IMP))
1903 		resx.res0 |= GENMASK(11, 8);
1904 	set_sysreg_masks(kvm, CNTHCTL_EL2, resx);
1905 
1906 	/* ICH_HCR_EL2 */
1907 	resx.res0 = ICH_HCR_EL2_RES0;
1908 	resx.res1 = ICH_HCR_EL2_RES1;
1909 	if (!(kvm_vgic_global_state.ich_vtr_el2 & ICH_VTR_EL2_TDS))
1910 		resx.res0 |= ICH_HCR_EL2_TDIR;
1911 	/* No GICv4 is presented to the guest */
1912 	resx.res0 |= ICH_HCR_EL2_DVIM | ICH_HCR_EL2_vSGIEOICount;
1913 	set_sysreg_masks(kvm, ICH_HCR_EL2, resx);
1914 
1915 	/* VNCR_EL2 */
1916 	resx.res0 = VNCR_EL2_RES0;
1917 	resx.res1 = VNCR_EL2_RES1;
1918 	set_sysreg_masks(kvm, VNCR_EL2, resx);
1919 
1920 	/* ZCR_EL2 - bits 8:4 are RAZ/WI so treat them as RES0 */
1921 	resx.res0 = ZCR_ELx_RES0 | GENMASK_ULL(8, 4);
1922 	resx.res1 = ZCR_ELx_RES1;
1923 	set_sysreg_masks(kvm, ZCR_EL2, resx);
1924 
1925 out:
1926 	for (enum vcpu_sysreg sr = __SANITISED_REG_START__; sr < NR_SYS_REGS; sr++)
1927 		__vcpu_rmw_sys_reg(vcpu, sr, |=, 0);
1928 
1929 	return 0;
1930 }
1931 
check_nested_vcpu_requests(struct kvm_vcpu * vcpu)1932 void check_nested_vcpu_requests(struct kvm_vcpu *vcpu)
1933 {
1934 	if (kvm_check_request(KVM_REQ_NESTED_S2_UNMAP, vcpu)) {
1935 		struct kvm_s2_mmu *mmu = vcpu->arch.hw_mmu;
1936 
1937 		write_lock(&vcpu->kvm->mmu_lock);
1938 		if (mmu->pending_unmap) {
1939 			kvm_stage2_unmap_range(mmu, 0, kvm_phys_size(mmu), true);
1940 			mmu->pending_unmap = false;
1941 		}
1942 		write_unlock(&vcpu->kvm->mmu_lock);
1943 	}
1944 
1945 	if (kvm_check_request(KVM_REQ_MAP_L1_VNCR_EL2, vcpu))
1946 		kvm_map_l1_vncr(vcpu);
1947 
1948 	/* Must be last, as may switch context! */
1949 	if (kvm_check_request(KVM_REQ_GUEST_HYP_IRQ_PENDING, vcpu))
1950 		kvm_inject_nested_irq(vcpu);
1951 }
1952 
1953 /*
1954  * One of the many architectural bugs in FEAT_NV2 is that the guest hypervisor
1955  * can write to HCR_EL2 behind our back, potentially changing the exception
1956  * routing / masking for even the host context.
1957  *
1958  * What follows is some slop to (1) react to exception routing / masking and (2)
1959  * preserve the pending SError state across translation regimes.
1960  */
kvm_nested_flush_hwstate(struct kvm_vcpu * vcpu)1961 void kvm_nested_flush_hwstate(struct kvm_vcpu *vcpu)
1962 {
1963 	if (!vcpu_has_nv(vcpu))
1964 		return;
1965 
1966 	if (unlikely(vcpu_test_and_clear_flag(vcpu, NESTED_SERROR_PENDING)))
1967 		kvm_inject_serror_esr(vcpu, vcpu_get_vsesr(vcpu));
1968 }
1969 
kvm_nested_sync_hwstate(struct kvm_vcpu * vcpu)1970 void kvm_nested_sync_hwstate(struct kvm_vcpu *vcpu)
1971 {
1972 	unsigned long *hcr = vcpu_hcr(vcpu);
1973 
1974 	if (!vcpu_has_nv(vcpu))
1975 		return;
1976 
1977 	/*
1978 	 * We previously decided that an SError was deliverable to the guest.
1979 	 * Reap the pending state from HCR_EL2 and...
1980 	 */
1981 	if (unlikely(__test_and_clear_bit(__ffs(HCR_VSE), hcr)))
1982 		vcpu_set_flag(vcpu, NESTED_SERROR_PENDING);
1983 
1984 	/*
1985 	 * Re-attempt SError injection in case the deliverability has changed,
1986 	 * which is necessary to faithfully emulate WFI the case of a pending
1987 	 * SError being a wakeup condition.
1988 	 */
1989 	if (unlikely(vcpu_test_and_clear_flag(vcpu, NESTED_SERROR_PENDING)))
1990 		kvm_inject_serror_esr(vcpu, vcpu_get_vsesr(vcpu));
1991 }
1992 
1993 /*
1994  * KVM unconditionally sets most of these traps anyway but use an allowlist
1995  * to document the guest hypervisor traps that may take precedence and guard
1996  * against future changes to the non-nested trap configuration.
1997  */
1998 #define NV_MDCR_GUEST_INCLUDE	(MDCR_EL2_TDE	|	\
1999 				 MDCR_EL2_TDA	|	\
2000 				 MDCR_EL2_TDRA	|	\
2001 				 MDCR_EL2_TTRF	|	\
2002 				 MDCR_EL2_TPMS	|	\
2003 				 MDCR_EL2_TPM	|	\
2004 				 MDCR_EL2_TPMCR	|	\
2005 				 MDCR_EL2_TDCC	|	\
2006 				 MDCR_EL2_TDOSA)
2007 
kvm_nested_setup_mdcr_el2(struct kvm_vcpu * vcpu)2008 void kvm_nested_setup_mdcr_el2(struct kvm_vcpu *vcpu)
2009 {
2010 	u64 guest_mdcr = __vcpu_sys_reg(vcpu, MDCR_EL2);
2011 
2012 	if (is_nested_ctxt(vcpu))
2013 		vcpu->arch.mdcr_el2 |= (guest_mdcr & NV_MDCR_GUEST_INCLUDE);
2014 	/*
2015 	 * In yet another example where FEAT_NV2 is fscking broken, accesses
2016 	 * to MDSCR_EL1 are redirected to the VNCR despite having an effect
2017 	 * at EL2. Use a big hammer to apply sanity.
2018 	 *
2019 	 * Unless of course we have FEAT_FGT, in which case we can precisely
2020 	 * trap MDSCR_EL1.
2021 	 */
2022 	else if (!cpus_have_final_cap(ARM64_HAS_FGT))
2023 		vcpu->arch.mdcr_el2 |= MDCR_EL2_TDA;
2024 }
2025