xref: /linux/arch/arm64/kvm/hyp/pgtable.c (revision aaa44952bbd1d4db14a4d676bf9595bb5db7e7b0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Stand-alone page-table allocator for hyp stage-1 and guest stage-2.
4  * No bombay mix was harmed in the writing of this file.
5  *
6  * Copyright (C) 2020 Google LLC
7  * Author: Will Deacon <will@kernel.org>
8  */
9 
10 #include <linux/bitfield.h>
11 #include <asm/kvm_pgtable.h>
12 #include <asm/stage2_pgtable.h>
13 
14 #define KVM_PTE_VALID			BIT(0)
15 
16 #define KVM_PTE_TYPE			BIT(1)
17 #define KVM_PTE_TYPE_BLOCK		0
18 #define KVM_PTE_TYPE_PAGE		1
19 #define KVM_PTE_TYPE_TABLE		1
20 
21 #define KVM_PTE_ADDR_MASK		GENMASK(47, PAGE_SHIFT)
22 #define KVM_PTE_ADDR_51_48		GENMASK(15, 12)
23 
24 #define KVM_PTE_LEAF_ATTR_LO		GENMASK(11, 2)
25 
26 #define KVM_PTE_LEAF_ATTR_LO_S1_ATTRIDX	GENMASK(4, 2)
27 #define KVM_PTE_LEAF_ATTR_LO_S1_AP	GENMASK(7, 6)
28 #define KVM_PTE_LEAF_ATTR_LO_S1_AP_RO	3
29 #define KVM_PTE_LEAF_ATTR_LO_S1_AP_RW	1
30 #define KVM_PTE_LEAF_ATTR_LO_S1_SH	GENMASK(9, 8)
31 #define KVM_PTE_LEAF_ATTR_LO_S1_SH_IS	3
32 #define KVM_PTE_LEAF_ATTR_LO_S1_AF	BIT(10)
33 
34 #define KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR	GENMASK(5, 2)
35 #define KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R	BIT(6)
36 #define KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W	BIT(7)
37 #define KVM_PTE_LEAF_ATTR_LO_S2_SH	GENMASK(9, 8)
38 #define KVM_PTE_LEAF_ATTR_LO_S2_SH_IS	3
39 #define KVM_PTE_LEAF_ATTR_LO_S2_AF	BIT(10)
40 
41 #define KVM_PTE_LEAF_ATTR_HI		GENMASK(63, 51)
42 
43 #define KVM_PTE_LEAF_ATTR_HI_S1_XN	BIT(54)
44 
45 #define KVM_PTE_LEAF_ATTR_HI_S2_XN	BIT(54)
46 
47 #define KVM_PTE_LEAF_ATTR_S2_PERMS	(KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R | \
48 					 KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W | \
49 					 KVM_PTE_LEAF_ATTR_HI_S2_XN)
50 
51 #define KVM_PTE_LEAF_ATTR_S2_IGNORED	GENMASK(58, 55)
52 
53 #define KVM_INVALID_PTE_OWNER_MASK	GENMASK(63, 56)
54 #define KVM_MAX_OWNER_ID		1
55 
56 struct kvm_pgtable_walk_data {
57 	struct kvm_pgtable		*pgt;
58 	struct kvm_pgtable_walker	*walker;
59 
60 	u64				addr;
61 	u64				end;
62 };
63 
64 static u64 kvm_granule_shift(u32 level)
65 {
66 	/* Assumes KVM_PGTABLE_MAX_LEVELS is 4 */
67 	return ARM64_HW_PGTABLE_LEVEL_SHIFT(level);
68 }
69 
70 static u64 kvm_granule_size(u32 level)
71 {
72 	return BIT(kvm_granule_shift(level));
73 }
74 
75 #define KVM_PHYS_INVALID (-1ULL)
76 
77 static bool kvm_phys_is_valid(u64 phys)
78 {
79 	return phys < BIT(id_aa64mmfr0_parange_to_phys_shift(ID_AA64MMFR0_PARANGE_MAX));
80 }
81 
82 static bool kvm_level_supports_block_mapping(u32 level)
83 {
84 	/*
85 	 * Reject invalid block mappings and don't bother with 4TB mappings for
86 	 * 52-bit PAs.
87 	 */
88 	return !(level == 0 || (PAGE_SIZE != SZ_4K && level == 1));
89 }
90 
91 static bool kvm_block_mapping_supported(u64 addr, u64 end, u64 phys, u32 level)
92 {
93 	u64 granule = kvm_granule_size(level);
94 
95 	if (!kvm_level_supports_block_mapping(level))
96 		return false;
97 
98 	if (granule > (end - addr))
99 		return false;
100 
101 	if (kvm_phys_is_valid(phys) && !IS_ALIGNED(phys, granule))
102 		return false;
103 
104 	return IS_ALIGNED(addr, granule);
105 }
106 
107 static u32 kvm_pgtable_idx(struct kvm_pgtable_walk_data *data, u32 level)
108 {
109 	u64 shift = kvm_granule_shift(level);
110 	u64 mask = BIT(PAGE_SHIFT - 3) - 1;
111 
112 	return (data->addr >> shift) & mask;
113 }
114 
115 static u32 __kvm_pgd_page_idx(struct kvm_pgtable *pgt, u64 addr)
116 {
117 	u64 shift = kvm_granule_shift(pgt->start_level - 1); /* May underflow */
118 	u64 mask = BIT(pgt->ia_bits) - 1;
119 
120 	return (addr & mask) >> shift;
121 }
122 
123 static u32 kvm_pgd_page_idx(struct kvm_pgtable_walk_data *data)
124 {
125 	return __kvm_pgd_page_idx(data->pgt, data->addr);
126 }
127 
128 static u32 kvm_pgd_pages(u32 ia_bits, u32 start_level)
129 {
130 	struct kvm_pgtable pgt = {
131 		.ia_bits	= ia_bits,
132 		.start_level	= start_level,
133 	};
134 
135 	return __kvm_pgd_page_idx(&pgt, -1ULL) + 1;
136 }
137 
138 static bool kvm_pte_valid(kvm_pte_t pte)
139 {
140 	return pte & KVM_PTE_VALID;
141 }
142 
143 static bool kvm_pte_table(kvm_pte_t pte, u32 level)
144 {
145 	if (level == KVM_PGTABLE_MAX_LEVELS - 1)
146 		return false;
147 
148 	if (!kvm_pte_valid(pte))
149 		return false;
150 
151 	return FIELD_GET(KVM_PTE_TYPE, pte) == KVM_PTE_TYPE_TABLE;
152 }
153 
154 static u64 kvm_pte_to_phys(kvm_pte_t pte)
155 {
156 	u64 pa = pte & KVM_PTE_ADDR_MASK;
157 
158 	if (PAGE_SHIFT == 16)
159 		pa |= FIELD_GET(KVM_PTE_ADDR_51_48, pte) << 48;
160 
161 	return pa;
162 }
163 
164 static kvm_pte_t kvm_phys_to_pte(u64 pa)
165 {
166 	kvm_pte_t pte = pa & KVM_PTE_ADDR_MASK;
167 
168 	if (PAGE_SHIFT == 16)
169 		pte |= FIELD_PREP(KVM_PTE_ADDR_51_48, pa >> 48);
170 
171 	return pte;
172 }
173 
174 static kvm_pte_t *kvm_pte_follow(kvm_pte_t pte, struct kvm_pgtable_mm_ops *mm_ops)
175 {
176 	return mm_ops->phys_to_virt(kvm_pte_to_phys(pte));
177 }
178 
179 static void kvm_clear_pte(kvm_pte_t *ptep)
180 {
181 	WRITE_ONCE(*ptep, 0);
182 }
183 
184 static void kvm_set_table_pte(kvm_pte_t *ptep, kvm_pte_t *childp,
185 			      struct kvm_pgtable_mm_ops *mm_ops)
186 {
187 	kvm_pte_t old = *ptep, pte = kvm_phys_to_pte(mm_ops->virt_to_phys(childp));
188 
189 	pte |= FIELD_PREP(KVM_PTE_TYPE, KVM_PTE_TYPE_TABLE);
190 	pte |= KVM_PTE_VALID;
191 
192 	WARN_ON(kvm_pte_valid(old));
193 	smp_store_release(ptep, pte);
194 }
195 
196 static kvm_pte_t kvm_init_valid_leaf_pte(u64 pa, kvm_pte_t attr, u32 level)
197 {
198 	kvm_pte_t pte = kvm_phys_to_pte(pa);
199 	u64 type = (level == KVM_PGTABLE_MAX_LEVELS - 1) ? KVM_PTE_TYPE_PAGE :
200 							   KVM_PTE_TYPE_BLOCK;
201 
202 	pte |= attr & (KVM_PTE_LEAF_ATTR_LO | KVM_PTE_LEAF_ATTR_HI);
203 	pte |= FIELD_PREP(KVM_PTE_TYPE, type);
204 	pte |= KVM_PTE_VALID;
205 
206 	return pte;
207 }
208 
209 static kvm_pte_t kvm_init_invalid_leaf_owner(u8 owner_id)
210 {
211 	return FIELD_PREP(KVM_INVALID_PTE_OWNER_MASK, owner_id);
212 }
213 
214 static int kvm_pgtable_visitor_cb(struct kvm_pgtable_walk_data *data, u64 addr,
215 				  u32 level, kvm_pte_t *ptep,
216 				  enum kvm_pgtable_walk_flags flag)
217 {
218 	struct kvm_pgtable_walker *walker = data->walker;
219 	return walker->cb(addr, data->end, level, ptep, flag, walker->arg);
220 }
221 
222 static int __kvm_pgtable_walk(struct kvm_pgtable_walk_data *data,
223 			      kvm_pte_t *pgtable, u32 level);
224 
225 static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data,
226 				      kvm_pte_t *ptep, u32 level)
227 {
228 	int ret = 0;
229 	u64 addr = data->addr;
230 	kvm_pte_t *childp, pte = *ptep;
231 	bool table = kvm_pte_table(pte, level);
232 	enum kvm_pgtable_walk_flags flags = data->walker->flags;
233 
234 	if (table && (flags & KVM_PGTABLE_WALK_TABLE_PRE)) {
235 		ret = kvm_pgtable_visitor_cb(data, addr, level, ptep,
236 					     KVM_PGTABLE_WALK_TABLE_PRE);
237 	}
238 
239 	if (!table && (flags & KVM_PGTABLE_WALK_LEAF)) {
240 		ret = kvm_pgtable_visitor_cb(data, addr, level, ptep,
241 					     KVM_PGTABLE_WALK_LEAF);
242 		pte = *ptep;
243 		table = kvm_pte_table(pte, level);
244 	}
245 
246 	if (ret)
247 		goto out;
248 
249 	if (!table) {
250 		data->addr = ALIGN_DOWN(data->addr, kvm_granule_size(level));
251 		data->addr += kvm_granule_size(level);
252 		goto out;
253 	}
254 
255 	childp = kvm_pte_follow(pte, data->pgt->mm_ops);
256 	ret = __kvm_pgtable_walk(data, childp, level + 1);
257 	if (ret)
258 		goto out;
259 
260 	if (flags & KVM_PGTABLE_WALK_TABLE_POST) {
261 		ret = kvm_pgtable_visitor_cb(data, addr, level, ptep,
262 					     KVM_PGTABLE_WALK_TABLE_POST);
263 	}
264 
265 out:
266 	return ret;
267 }
268 
269 static int __kvm_pgtable_walk(struct kvm_pgtable_walk_data *data,
270 			      kvm_pte_t *pgtable, u32 level)
271 {
272 	u32 idx;
273 	int ret = 0;
274 
275 	if (WARN_ON_ONCE(level >= KVM_PGTABLE_MAX_LEVELS))
276 		return -EINVAL;
277 
278 	for (idx = kvm_pgtable_idx(data, level); idx < PTRS_PER_PTE; ++idx) {
279 		kvm_pte_t *ptep = &pgtable[idx];
280 
281 		if (data->addr >= data->end)
282 			break;
283 
284 		ret = __kvm_pgtable_visit(data, ptep, level);
285 		if (ret)
286 			break;
287 	}
288 
289 	return ret;
290 }
291 
292 static int _kvm_pgtable_walk(struct kvm_pgtable_walk_data *data)
293 {
294 	u32 idx;
295 	int ret = 0;
296 	struct kvm_pgtable *pgt = data->pgt;
297 	u64 limit = BIT(pgt->ia_bits);
298 
299 	if (data->addr > limit || data->end > limit)
300 		return -ERANGE;
301 
302 	if (!pgt->pgd)
303 		return -EINVAL;
304 
305 	for (idx = kvm_pgd_page_idx(data); data->addr < data->end; ++idx) {
306 		kvm_pte_t *ptep = &pgt->pgd[idx * PTRS_PER_PTE];
307 
308 		ret = __kvm_pgtable_walk(data, ptep, pgt->start_level);
309 		if (ret)
310 			break;
311 	}
312 
313 	return ret;
314 }
315 
316 int kvm_pgtable_walk(struct kvm_pgtable *pgt, u64 addr, u64 size,
317 		     struct kvm_pgtable_walker *walker)
318 {
319 	struct kvm_pgtable_walk_data walk_data = {
320 		.pgt	= pgt,
321 		.addr	= ALIGN_DOWN(addr, PAGE_SIZE),
322 		.end	= PAGE_ALIGN(walk_data.addr + size),
323 		.walker	= walker,
324 	};
325 
326 	return _kvm_pgtable_walk(&walk_data);
327 }
328 
329 struct hyp_map_data {
330 	u64				phys;
331 	kvm_pte_t			attr;
332 	struct kvm_pgtable_mm_ops	*mm_ops;
333 };
334 
335 static int hyp_set_prot_attr(enum kvm_pgtable_prot prot, kvm_pte_t *ptep)
336 {
337 	bool device = prot & KVM_PGTABLE_PROT_DEVICE;
338 	u32 mtype = device ? MT_DEVICE_nGnRE : MT_NORMAL;
339 	kvm_pte_t attr = FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S1_ATTRIDX, mtype);
340 	u32 sh = KVM_PTE_LEAF_ATTR_LO_S1_SH_IS;
341 	u32 ap = (prot & KVM_PGTABLE_PROT_W) ? KVM_PTE_LEAF_ATTR_LO_S1_AP_RW :
342 					       KVM_PTE_LEAF_ATTR_LO_S1_AP_RO;
343 
344 	if (!(prot & KVM_PGTABLE_PROT_R))
345 		return -EINVAL;
346 
347 	if (prot & KVM_PGTABLE_PROT_X) {
348 		if (prot & KVM_PGTABLE_PROT_W)
349 			return -EINVAL;
350 
351 		if (device)
352 			return -EINVAL;
353 	} else {
354 		attr |= KVM_PTE_LEAF_ATTR_HI_S1_XN;
355 	}
356 
357 	attr |= FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S1_AP, ap);
358 	attr |= FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S1_SH, sh);
359 	attr |= KVM_PTE_LEAF_ATTR_LO_S1_AF;
360 	*ptep = attr;
361 
362 	return 0;
363 }
364 
365 static bool hyp_map_walker_try_leaf(u64 addr, u64 end, u32 level,
366 				    kvm_pte_t *ptep, struct hyp_map_data *data)
367 {
368 	kvm_pte_t new, old = *ptep;
369 	u64 granule = kvm_granule_size(level), phys = data->phys;
370 
371 	if (!kvm_block_mapping_supported(addr, end, phys, level))
372 		return false;
373 
374 	/* Tolerate KVM recreating the exact same mapping */
375 	new = kvm_init_valid_leaf_pte(phys, data->attr, level);
376 	if (old != new && !WARN_ON(kvm_pte_valid(old)))
377 		smp_store_release(ptep, new);
378 
379 	data->phys += granule;
380 	return true;
381 }
382 
383 static int hyp_map_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
384 			  enum kvm_pgtable_walk_flags flag, void * const arg)
385 {
386 	kvm_pte_t *childp;
387 	struct hyp_map_data *data = arg;
388 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
389 
390 	if (hyp_map_walker_try_leaf(addr, end, level, ptep, arg))
391 		return 0;
392 
393 	if (WARN_ON(level == KVM_PGTABLE_MAX_LEVELS - 1))
394 		return -EINVAL;
395 
396 	childp = (kvm_pte_t *)mm_ops->zalloc_page(NULL);
397 	if (!childp)
398 		return -ENOMEM;
399 
400 	kvm_set_table_pte(ptep, childp, mm_ops);
401 	return 0;
402 }
403 
404 int kvm_pgtable_hyp_map(struct kvm_pgtable *pgt, u64 addr, u64 size, u64 phys,
405 			enum kvm_pgtable_prot prot)
406 {
407 	int ret;
408 	struct hyp_map_data map_data = {
409 		.phys	= ALIGN_DOWN(phys, PAGE_SIZE),
410 		.mm_ops	= pgt->mm_ops,
411 	};
412 	struct kvm_pgtable_walker walker = {
413 		.cb	= hyp_map_walker,
414 		.flags	= KVM_PGTABLE_WALK_LEAF,
415 		.arg	= &map_data,
416 	};
417 
418 	ret = hyp_set_prot_attr(prot, &map_data.attr);
419 	if (ret)
420 		return ret;
421 
422 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
423 	dsb(ishst);
424 	isb();
425 	return ret;
426 }
427 
428 int kvm_pgtable_hyp_init(struct kvm_pgtable *pgt, u32 va_bits,
429 			 struct kvm_pgtable_mm_ops *mm_ops)
430 {
431 	u64 levels = ARM64_HW_PGTABLE_LEVELS(va_bits);
432 
433 	pgt->pgd = (kvm_pte_t *)mm_ops->zalloc_page(NULL);
434 	if (!pgt->pgd)
435 		return -ENOMEM;
436 
437 	pgt->ia_bits		= va_bits;
438 	pgt->start_level	= KVM_PGTABLE_MAX_LEVELS - levels;
439 	pgt->mm_ops		= mm_ops;
440 	pgt->mmu		= NULL;
441 	return 0;
442 }
443 
444 static int hyp_free_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
445 			   enum kvm_pgtable_walk_flags flag, void * const arg)
446 {
447 	struct kvm_pgtable_mm_ops *mm_ops = arg;
448 
449 	mm_ops->put_page((void *)kvm_pte_follow(*ptep, mm_ops));
450 	return 0;
451 }
452 
453 void kvm_pgtable_hyp_destroy(struct kvm_pgtable *pgt)
454 {
455 	struct kvm_pgtable_walker walker = {
456 		.cb	= hyp_free_walker,
457 		.flags	= KVM_PGTABLE_WALK_TABLE_POST,
458 		.arg	= pgt->mm_ops,
459 	};
460 
461 	WARN_ON(kvm_pgtable_walk(pgt, 0, BIT(pgt->ia_bits), &walker));
462 	pgt->mm_ops->put_page(pgt->pgd);
463 	pgt->pgd = NULL;
464 }
465 
466 struct stage2_map_data {
467 	u64				phys;
468 	kvm_pte_t			attr;
469 	u8				owner_id;
470 
471 	kvm_pte_t			*anchor;
472 	kvm_pte_t			*childp;
473 
474 	struct kvm_s2_mmu		*mmu;
475 	void				*memcache;
476 
477 	struct kvm_pgtable_mm_ops	*mm_ops;
478 };
479 
480 u64 kvm_get_vtcr(u64 mmfr0, u64 mmfr1, u32 phys_shift)
481 {
482 	u64 vtcr = VTCR_EL2_FLAGS;
483 	u8 lvls;
484 
485 	vtcr |= kvm_get_parange(mmfr0) << VTCR_EL2_PS_SHIFT;
486 	vtcr |= VTCR_EL2_T0SZ(phys_shift);
487 	/*
488 	 * Use a minimum 2 level page table to prevent splitting
489 	 * host PMD huge pages at stage2.
490 	 */
491 	lvls = stage2_pgtable_levels(phys_shift);
492 	if (lvls < 2)
493 		lvls = 2;
494 	vtcr |= VTCR_EL2_LVLS_TO_SL0(lvls);
495 
496 	/*
497 	 * Enable the Hardware Access Flag management, unconditionally
498 	 * on all CPUs. The features is RES0 on CPUs without the support
499 	 * and must be ignored by the CPUs.
500 	 */
501 	vtcr |= VTCR_EL2_HA;
502 
503 	/* Set the vmid bits */
504 	vtcr |= (get_vmid_bits(mmfr1) == 16) ?
505 		VTCR_EL2_VS_16BIT :
506 		VTCR_EL2_VS_8BIT;
507 
508 	return vtcr;
509 }
510 
511 static bool stage2_has_fwb(struct kvm_pgtable *pgt)
512 {
513 	if (!cpus_have_const_cap(ARM64_HAS_STAGE2_FWB))
514 		return false;
515 
516 	return !(pgt->flags & KVM_PGTABLE_S2_NOFWB);
517 }
518 
519 #define KVM_S2_MEMATTR(pgt, attr) PAGE_S2_MEMATTR(attr, stage2_has_fwb(pgt))
520 
521 static int stage2_set_prot_attr(struct kvm_pgtable *pgt, enum kvm_pgtable_prot prot,
522 				kvm_pte_t *ptep)
523 {
524 	bool device = prot & KVM_PGTABLE_PROT_DEVICE;
525 	kvm_pte_t attr = device ? KVM_S2_MEMATTR(pgt, DEVICE_nGnRE) :
526 			    KVM_S2_MEMATTR(pgt, NORMAL);
527 	u32 sh = KVM_PTE_LEAF_ATTR_LO_S2_SH_IS;
528 
529 	if (!(prot & KVM_PGTABLE_PROT_X))
530 		attr |= KVM_PTE_LEAF_ATTR_HI_S2_XN;
531 	else if (device)
532 		return -EINVAL;
533 
534 	if (prot & KVM_PGTABLE_PROT_R)
535 		attr |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
536 
537 	if (prot & KVM_PGTABLE_PROT_W)
538 		attr |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
539 
540 	attr |= FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S2_SH, sh);
541 	attr |= KVM_PTE_LEAF_ATTR_LO_S2_AF;
542 	*ptep = attr;
543 
544 	return 0;
545 }
546 
547 static bool stage2_pte_needs_update(kvm_pte_t old, kvm_pte_t new)
548 {
549 	if (!kvm_pte_valid(old) || !kvm_pte_valid(new))
550 		return true;
551 
552 	return ((old ^ new) & (~KVM_PTE_LEAF_ATTR_S2_PERMS));
553 }
554 
555 static bool stage2_pte_is_counted(kvm_pte_t pte)
556 {
557 	/*
558 	 * The refcount tracks valid entries as well as invalid entries if they
559 	 * encode ownership of a page to another entity than the page-table
560 	 * owner, whose id is 0.
561 	 */
562 	return !!pte;
563 }
564 
565 static void stage2_put_pte(kvm_pte_t *ptep, struct kvm_s2_mmu *mmu, u64 addr,
566 			   u32 level, struct kvm_pgtable_mm_ops *mm_ops)
567 {
568 	/*
569 	 * Clear the existing PTE, and perform break-before-make with
570 	 * TLB maintenance if it was valid.
571 	 */
572 	if (kvm_pte_valid(*ptep)) {
573 		kvm_clear_pte(ptep);
574 		kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu, addr, level);
575 	}
576 
577 	mm_ops->put_page(ptep);
578 }
579 
580 static int stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
581 				      kvm_pte_t *ptep,
582 				      struct stage2_map_data *data)
583 {
584 	kvm_pte_t new, old = *ptep;
585 	u64 granule = kvm_granule_size(level), phys = data->phys;
586 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
587 
588 	if (!kvm_block_mapping_supported(addr, end, phys, level))
589 		return -E2BIG;
590 
591 	if (kvm_phys_is_valid(phys))
592 		new = kvm_init_valid_leaf_pte(phys, data->attr, level);
593 	else
594 		new = kvm_init_invalid_leaf_owner(data->owner_id);
595 
596 	if (stage2_pte_is_counted(old)) {
597 		/*
598 		 * Skip updating the PTE if we are trying to recreate the exact
599 		 * same mapping or only change the access permissions. Instead,
600 		 * the vCPU will exit one more time from guest if still needed
601 		 * and then go through the path of relaxing permissions.
602 		 */
603 		if (!stage2_pte_needs_update(old, new))
604 			return -EAGAIN;
605 
606 		stage2_put_pte(ptep, data->mmu, addr, level, mm_ops);
607 	}
608 
609 	smp_store_release(ptep, new);
610 	if (stage2_pte_is_counted(new))
611 		mm_ops->get_page(ptep);
612 	if (kvm_phys_is_valid(phys))
613 		data->phys += granule;
614 	return 0;
615 }
616 
617 static int stage2_map_walk_table_pre(u64 addr, u64 end, u32 level,
618 				     kvm_pte_t *ptep,
619 				     struct stage2_map_data *data)
620 {
621 	if (data->anchor)
622 		return 0;
623 
624 	if (!kvm_block_mapping_supported(addr, end, data->phys, level))
625 		return 0;
626 
627 	data->childp = kvm_pte_follow(*ptep, data->mm_ops);
628 	kvm_clear_pte(ptep);
629 
630 	/*
631 	 * Invalidate the whole stage-2, as we may have numerous leaf
632 	 * entries below us which would otherwise need invalidating
633 	 * individually.
634 	 */
635 	kvm_call_hyp(__kvm_tlb_flush_vmid, data->mmu);
636 	data->anchor = ptep;
637 	return 0;
638 }
639 
640 static int stage2_map_walk_leaf(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
641 				struct stage2_map_data *data)
642 {
643 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
644 	kvm_pte_t *childp, pte = *ptep;
645 	int ret;
646 
647 	if (data->anchor) {
648 		if (stage2_pte_is_counted(pte))
649 			mm_ops->put_page(ptep);
650 
651 		return 0;
652 	}
653 
654 	ret = stage2_map_walker_try_leaf(addr, end, level, ptep, data);
655 	if (ret != -E2BIG)
656 		return ret;
657 
658 	if (WARN_ON(level == KVM_PGTABLE_MAX_LEVELS - 1))
659 		return -EINVAL;
660 
661 	if (!data->memcache)
662 		return -ENOMEM;
663 
664 	childp = mm_ops->zalloc_page(data->memcache);
665 	if (!childp)
666 		return -ENOMEM;
667 
668 	/*
669 	 * If we've run into an existing block mapping then replace it with
670 	 * a table. Accesses beyond 'end' that fall within the new table
671 	 * will be mapped lazily.
672 	 */
673 	if (stage2_pte_is_counted(pte))
674 		stage2_put_pte(ptep, data->mmu, addr, level, mm_ops);
675 
676 	kvm_set_table_pte(ptep, childp, mm_ops);
677 	mm_ops->get_page(ptep);
678 
679 	return 0;
680 }
681 
682 static int stage2_map_walk_table_post(u64 addr, u64 end, u32 level,
683 				      kvm_pte_t *ptep,
684 				      struct stage2_map_data *data)
685 {
686 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
687 	kvm_pte_t *childp;
688 	int ret = 0;
689 
690 	if (!data->anchor)
691 		return 0;
692 
693 	if (data->anchor == ptep) {
694 		childp = data->childp;
695 		data->anchor = NULL;
696 		data->childp = NULL;
697 		ret = stage2_map_walk_leaf(addr, end, level, ptep, data);
698 	} else {
699 		childp = kvm_pte_follow(*ptep, mm_ops);
700 	}
701 
702 	mm_ops->put_page(childp);
703 	mm_ops->put_page(ptep);
704 
705 	return ret;
706 }
707 
708 /*
709  * This is a little fiddly, as we use all three of the walk flags. The idea
710  * is that the TABLE_PRE callback runs for table entries on the way down,
711  * looking for table entries which we could conceivably replace with a
712  * block entry for this mapping. If it finds one, then it sets the 'anchor'
713  * field in 'struct stage2_map_data' to point at the table entry, before
714  * clearing the entry to zero and descending into the now detached table.
715  *
716  * The behaviour of the LEAF callback then depends on whether or not the
717  * anchor has been set. If not, then we're not using a block mapping higher
718  * up the table and we perform the mapping at the existing leaves instead.
719  * If, on the other hand, the anchor _is_ set, then we drop references to
720  * all valid leaves so that the pages beneath the anchor can be freed.
721  *
722  * Finally, the TABLE_POST callback does nothing if the anchor has not
723  * been set, but otherwise frees the page-table pages while walking back up
724  * the page-table, installing the block entry when it revisits the anchor
725  * pointer and clearing the anchor to NULL.
726  */
727 static int stage2_map_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
728 			     enum kvm_pgtable_walk_flags flag, void * const arg)
729 {
730 	struct stage2_map_data *data = arg;
731 
732 	switch (flag) {
733 	case KVM_PGTABLE_WALK_TABLE_PRE:
734 		return stage2_map_walk_table_pre(addr, end, level, ptep, data);
735 	case KVM_PGTABLE_WALK_LEAF:
736 		return stage2_map_walk_leaf(addr, end, level, ptep, data);
737 	case KVM_PGTABLE_WALK_TABLE_POST:
738 		return stage2_map_walk_table_post(addr, end, level, ptep, data);
739 	}
740 
741 	return -EINVAL;
742 }
743 
744 int kvm_pgtable_stage2_map(struct kvm_pgtable *pgt, u64 addr, u64 size,
745 			   u64 phys, enum kvm_pgtable_prot prot,
746 			   void *mc)
747 {
748 	int ret;
749 	struct stage2_map_data map_data = {
750 		.phys		= ALIGN_DOWN(phys, PAGE_SIZE),
751 		.mmu		= pgt->mmu,
752 		.memcache	= mc,
753 		.mm_ops		= pgt->mm_ops,
754 	};
755 	struct kvm_pgtable_walker walker = {
756 		.cb		= stage2_map_walker,
757 		.flags		= KVM_PGTABLE_WALK_TABLE_PRE |
758 				  KVM_PGTABLE_WALK_LEAF |
759 				  KVM_PGTABLE_WALK_TABLE_POST,
760 		.arg		= &map_data,
761 	};
762 
763 	if (WARN_ON((pgt->flags & KVM_PGTABLE_S2_IDMAP) && (addr != phys)))
764 		return -EINVAL;
765 
766 	ret = stage2_set_prot_attr(pgt, prot, &map_data.attr);
767 	if (ret)
768 		return ret;
769 
770 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
771 	dsb(ishst);
772 	return ret;
773 }
774 
775 int kvm_pgtable_stage2_set_owner(struct kvm_pgtable *pgt, u64 addr, u64 size,
776 				 void *mc, u8 owner_id)
777 {
778 	int ret;
779 	struct stage2_map_data map_data = {
780 		.phys		= KVM_PHYS_INVALID,
781 		.mmu		= pgt->mmu,
782 		.memcache	= mc,
783 		.mm_ops		= pgt->mm_ops,
784 		.owner_id	= owner_id,
785 	};
786 	struct kvm_pgtable_walker walker = {
787 		.cb		= stage2_map_walker,
788 		.flags		= KVM_PGTABLE_WALK_TABLE_PRE |
789 				  KVM_PGTABLE_WALK_LEAF |
790 				  KVM_PGTABLE_WALK_TABLE_POST,
791 		.arg		= &map_data,
792 	};
793 
794 	if (owner_id > KVM_MAX_OWNER_ID)
795 		return -EINVAL;
796 
797 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
798 	return ret;
799 }
800 
801 static bool stage2_pte_cacheable(struct kvm_pgtable *pgt, kvm_pte_t pte)
802 {
803 	u64 memattr = pte & KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR;
804 	return memattr == KVM_S2_MEMATTR(pgt, NORMAL);
805 }
806 
807 static int stage2_unmap_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
808 			       enum kvm_pgtable_walk_flags flag,
809 			       void * const arg)
810 {
811 	struct kvm_pgtable *pgt = arg;
812 	struct kvm_s2_mmu *mmu = pgt->mmu;
813 	struct kvm_pgtable_mm_ops *mm_ops = pgt->mm_ops;
814 	kvm_pte_t pte = *ptep, *childp = NULL;
815 	bool need_flush = false;
816 
817 	if (!kvm_pte_valid(pte)) {
818 		if (stage2_pte_is_counted(pte)) {
819 			kvm_clear_pte(ptep);
820 			mm_ops->put_page(ptep);
821 		}
822 		return 0;
823 	}
824 
825 	if (kvm_pte_table(pte, level)) {
826 		childp = kvm_pte_follow(pte, mm_ops);
827 
828 		if (mm_ops->page_count(childp) != 1)
829 			return 0;
830 	} else if (stage2_pte_cacheable(pgt, pte)) {
831 		need_flush = !stage2_has_fwb(pgt);
832 	}
833 
834 	/*
835 	 * This is similar to the map() path in that we unmap the entire
836 	 * block entry and rely on the remaining portions being faulted
837 	 * back lazily.
838 	 */
839 	stage2_put_pte(ptep, mmu, addr, level, mm_ops);
840 
841 	if (need_flush) {
842 		__flush_dcache_area(kvm_pte_follow(pte, mm_ops),
843 				    kvm_granule_size(level));
844 	}
845 
846 	if (childp)
847 		mm_ops->put_page(childp);
848 
849 	return 0;
850 }
851 
852 int kvm_pgtable_stage2_unmap(struct kvm_pgtable *pgt, u64 addr, u64 size)
853 {
854 	struct kvm_pgtable_walker walker = {
855 		.cb	= stage2_unmap_walker,
856 		.arg	= pgt,
857 		.flags	= KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
858 	};
859 
860 	return kvm_pgtable_walk(pgt, addr, size, &walker);
861 }
862 
863 struct stage2_attr_data {
864 	kvm_pte_t	attr_set;
865 	kvm_pte_t	attr_clr;
866 	kvm_pte_t	pte;
867 	u32		level;
868 };
869 
870 static int stage2_attr_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
871 			      enum kvm_pgtable_walk_flags flag,
872 			      void * const arg)
873 {
874 	kvm_pte_t pte = *ptep;
875 	struct stage2_attr_data *data = arg;
876 
877 	if (!kvm_pte_valid(pte))
878 		return 0;
879 
880 	data->level = level;
881 	data->pte = pte;
882 	pte &= ~data->attr_clr;
883 	pte |= data->attr_set;
884 
885 	/*
886 	 * We may race with the CPU trying to set the access flag here,
887 	 * but worst-case the access flag update gets lost and will be
888 	 * set on the next access instead.
889 	 */
890 	if (data->pte != pte)
891 		WRITE_ONCE(*ptep, pte);
892 
893 	return 0;
894 }
895 
896 static int stage2_update_leaf_attrs(struct kvm_pgtable *pgt, u64 addr,
897 				    u64 size, kvm_pte_t attr_set,
898 				    kvm_pte_t attr_clr, kvm_pte_t *orig_pte,
899 				    u32 *level)
900 {
901 	int ret;
902 	kvm_pte_t attr_mask = KVM_PTE_LEAF_ATTR_LO | KVM_PTE_LEAF_ATTR_HI;
903 	struct stage2_attr_data data = {
904 		.attr_set	= attr_set & attr_mask,
905 		.attr_clr	= attr_clr & attr_mask,
906 	};
907 	struct kvm_pgtable_walker walker = {
908 		.cb		= stage2_attr_walker,
909 		.arg		= &data,
910 		.flags		= KVM_PGTABLE_WALK_LEAF,
911 	};
912 
913 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
914 	if (ret)
915 		return ret;
916 
917 	if (orig_pte)
918 		*orig_pte = data.pte;
919 
920 	if (level)
921 		*level = data.level;
922 	return 0;
923 }
924 
925 int kvm_pgtable_stage2_wrprotect(struct kvm_pgtable *pgt, u64 addr, u64 size)
926 {
927 	return stage2_update_leaf_attrs(pgt, addr, size, 0,
928 					KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W,
929 					NULL, NULL);
930 }
931 
932 kvm_pte_t kvm_pgtable_stage2_mkyoung(struct kvm_pgtable *pgt, u64 addr)
933 {
934 	kvm_pte_t pte = 0;
935 	stage2_update_leaf_attrs(pgt, addr, 1, KVM_PTE_LEAF_ATTR_LO_S2_AF, 0,
936 				 &pte, NULL);
937 	dsb(ishst);
938 	return pte;
939 }
940 
941 kvm_pte_t kvm_pgtable_stage2_mkold(struct kvm_pgtable *pgt, u64 addr)
942 {
943 	kvm_pte_t pte = 0;
944 	stage2_update_leaf_attrs(pgt, addr, 1, 0, KVM_PTE_LEAF_ATTR_LO_S2_AF,
945 				 &pte, NULL);
946 	/*
947 	 * "But where's the TLBI?!", you scream.
948 	 * "Over in the core code", I sigh.
949 	 *
950 	 * See the '->clear_flush_young()' callback on the KVM mmu notifier.
951 	 */
952 	return pte;
953 }
954 
955 bool kvm_pgtable_stage2_is_young(struct kvm_pgtable *pgt, u64 addr)
956 {
957 	kvm_pte_t pte = 0;
958 	stage2_update_leaf_attrs(pgt, addr, 1, 0, 0, &pte, NULL);
959 	return pte & KVM_PTE_LEAF_ATTR_LO_S2_AF;
960 }
961 
962 int kvm_pgtable_stage2_relax_perms(struct kvm_pgtable *pgt, u64 addr,
963 				   enum kvm_pgtable_prot prot)
964 {
965 	int ret;
966 	u32 level;
967 	kvm_pte_t set = 0, clr = 0;
968 
969 	if (prot & KVM_PGTABLE_PROT_R)
970 		set |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
971 
972 	if (prot & KVM_PGTABLE_PROT_W)
973 		set |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
974 
975 	if (prot & KVM_PGTABLE_PROT_X)
976 		clr |= KVM_PTE_LEAF_ATTR_HI_S2_XN;
977 
978 	ret = stage2_update_leaf_attrs(pgt, addr, 1, set, clr, NULL, &level);
979 	if (!ret)
980 		kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, pgt->mmu, addr, level);
981 	return ret;
982 }
983 
984 static int stage2_flush_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
985 			       enum kvm_pgtable_walk_flags flag,
986 			       void * const arg)
987 {
988 	struct kvm_pgtable *pgt = arg;
989 	struct kvm_pgtable_mm_ops *mm_ops = pgt->mm_ops;
990 	kvm_pte_t pte = *ptep;
991 
992 	if (!kvm_pte_valid(pte) || !stage2_pte_cacheable(pgt, pte))
993 		return 0;
994 
995 	__flush_dcache_area(kvm_pte_follow(pte, mm_ops), kvm_granule_size(level));
996 	return 0;
997 }
998 
999 int kvm_pgtable_stage2_flush(struct kvm_pgtable *pgt, u64 addr, u64 size)
1000 {
1001 	struct kvm_pgtable_walker walker = {
1002 		.cb	= stage2_flush_walker,
1003 		.flags	= KVM_PGTABLE_WALK_LEAF,
1004 		.arg	= pgt,
1005 	};
1006 
1007 	if (stage2_has_fwb(pgt))
1008 		return 0;
1009 
1010 	return kvm_pgtable_walk(pgt, addr, size, &walker);
1011 }
1012 
1013 int kvm_pgtable_stage2_init_flags(struct kvm_pgtable *pgt, struct kvm_arch *arch,
1014 				  struct kvm_pgtable_mm_ops *mm_ops,
1015 				  enum kvm_pgtable_stage2_flags flags)
1016 {
1017 	size_t pgd_sz;
1018 	u64 vtcr = arch->vtcr;
1019 	u32 ia_bits = VTCR_EL2_IPA(vtcr);
1020 	u32 sl0 = FIELD_GET(VTCR_EL2_SL0_MASK, vtcr);
1021 	u32 start_level = VTCR_EL2_TGRAN_SL0_BASE - sl0;
1022 
1023 	pgd_sz = kvm_pgd_pages(ia_bits, start_level) * PAGE_SIZE;
1024 	pgt->pgd = mm_ops->zalloc_pages_exact(pgd_sz);
1025 	if (!pgt->pgd)
1026 		return -ENOMEM;
1027 
1028 	pgt->ia_bits		= ia_bits;
1029 	pgt->start_level	= start_level;
1030 	pgt->mm_ops		= mm_ops;
1031 	pgt->mmu		= &arch->mmu;
1032 	pgt->flags		= flags;
1033 
1034 	/* Ensure zeroed PGD pages are visible to the hardware walker */
1035 	dsb(ishst);
1036 	return 0;
1037 }
1038 
1039 static int stage2_free_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
1040 			      enum kvm_pgtable_walk_flags flag,
1041 			      void * const arg)
1042 {
1043 	struct kvm_pgtable_mm_ops *mm_ops = arg;
1044 	kvm_pte_t pte = *ptep;
1045 
1046 	if (!stage2_pte_is_counted(pte))
1047 		return 0;
1048 
1049 	mm_ops->put_page(ptep);
1050 
1051 	if (kvm_pte_table(pte, level))
1052 		mm_ops->put_page(kvm_pte_follow(pte, mm_ops));
1053 
1054 	return 0;
1055 }
1056 
1057 void kvm_pgtable_stage2_destroy(struct kvm_pgtable *pgt)
1058 {
1059 	size_t pgd_sz;
1060 	struct kvm_pgtable_walker walker = {
1061 		.cb	= stage2_free_walker,
1062 		.flags	= KVM_PGTABLE_WALK_LEAF |
1063 			  KVM_PGTABLE_WALK_TABLE_POST,
1064 		.arg	= pgt->mm_ops,
1065 	};
1066 
1067 	WARN_ON(kvm_pgtable_walk(pgt, 0, BIT(pgt->ia_bits), &walker));
1068 	pgd_sz = kvm_pgd_pages(pgt->ia_bits, pgt->start_level) * PAGE_SIZE;
1069 	pgt->mm_ops->free_pages_exact(pgt->pgd, pgd_sz);
1070 	pgt->pgd = NULL;
1071 }
1072 
1073 #define KVM_PTE_LEAF_S2_COMPAT_MASK	(KVM_PTE_LEAF_ATTR_S2_PERMS | \
1074 					 KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR | \
1075 					 KVM_PTE_LEAF_ATTR_S2_IGNORED)
1076 
1077 static int stage2_check_permission_walker(u64 addr, u64 end, u32 level,
1078 					  kvm_pte_t *ptep,
1079 					  enum kvm_pgtable_walk_flags flag,
1080 					  void * const arg)
1081 {
1082 	kvm_pte_t old_attr, pte = *ptep, *new_attr = arg;
1083 
1084 	/*
1085 	 * Compatible mappings are either invalid and owned by the page-table
1086 	 * owner (whose id is 0), or valid with matching permission attributes.
1087 	 */
1088 	if (kvm_pte_valid(pte)) {
1089 		old_attr = pte & KVM_PTE_LEAF_S2_COMPAT_MASK;
1090 		if (old_attr != *new_attr)
1091 			return -EEXIST;
1092 	} else if (pte) {
1093 		return -EEXIST;
1094 	}
1095 
1096 	return 0;
1097 }
1098 
1099 int kvm_pgtable_stage2_find_range(struct kvm_pgtable *pgt, u64 addr,
1100 				  enum kvm_pgtable_prot prot,
1101 				  struct kvm_mem_range *range)
1102 {
1103 	kvm_pte_t attr;
1104 	struct kvm_pgtable_walker check_perm_walker = {
1105 		.cb		= stage2_check_permission_walker,
1106 		.flags		= KVM_PGTABLE_WALK_LEAF,
1107 		.arg		= &attr,
1108 	};
1109 	u64 granule, start, end;
1110 	u32 level;
1111 	int ret;
1112 
1113 	ret = stage2_set_prot_attr(pgt, prot, &attr);
1114 	if (ret)
1115 		return ret;
1116 	attr &= KVM_PTE_LEAF_S2_COMPAT_MASK;
1117 
1118 	for (level = pgt->start_level; level < KVM_PGTABLE_MAX_LEVELS; level++) {
1119 		granule = kvm_granule_size(level);
1120 		start = ALIGN_DOWN(addr, granule);
1121 		end = start + granule;
1122 
1123 		if (!kvm_level_supports_block_mapping(level))
1124 			continue;
1125 
1126 		if (start < range->start || range->end < end)
1127 			continue;
1128 
1129 		/*
1130 		 * Check the presence of existing mappings with incompatible
1131 		 * permissions within the current block range, and try one level
1132 		 * deeper if one is found.
1133 		 */
1134 		ret = kvm_pgtable_walk(pgt, start, granule, &check_perm_walker);
1135 		if (ret != -EEXIST)
1136 			break;
1137 	}
1138 
1139 	if (!ret) {
1140 		range->start = start;
1141 		range->end = end;
1142 	}
1143 
1144 	return ret;
1145 }
1146