xref: /linux/arch/arm64/kvm/hyp/pgtable.c (revision bdd1a21b52557ea8f61d0a5dc2f77151b576eb70)
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 bool stage2_pte_cacheable(struct kvm_pgtable *pgt, kvm_pte_t pte)
581 {
582 	u64 memattr = pte & KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR;
583 	return memattr == KVM_S2_MEMATTR(pgt, NORMAL);
584 }
585 
586 static bool stage2_pte_executable(kvm_pte_t pte)
587 {
588 	return !(pte & KVM_PTE_LEAF_ATTR_HI_S2_XN);
589 }
590 
591 static int stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
592 				      kvm_pte_t *ptep,
593 				      struct stage2_map_data *data)
594 {
595 	kvm_pte_t new, old = *ptep;
596 	u64 granule = kvm_granule_size(level), phys = data->phys;
597 	struct kvm_pgtable *pgt = data->mmu->pgt;
598 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
599 
600 	if (!kvm_block_mapping_supported(addr, end, phys, level))
601 		return -E2BIG;
602 
603 	if (kvm_phys_is_valid(phys))
604 		new = kvm_init_valid_leaf_pte(phys, data->attr, level);
605 	else
606 		new = kvm_init_invalid_leaf_owner(data->owner_id);
607 
608 	if (stage2_pte_is_counted(old)) {
609 		/*
610 		 * Skip updating the PTE if we are trying to recreate the exact
611 		 * same mapping or only change the access permissions. Instead,
612 		 * the vCPU will exit one more time from guest if still needed
613 		 * and then go through the path of relaxing permissions.
614 		 */
615 		if (!stage2_pte_needs_update(old, new))
616 			return -EAGAIN;
617 
618 		stage2_put_pte(ptep, data->mmu, addr, level, mm_ops);
619 	}
620 
621 	/* Perform CMOs before installation of the guest stage-2 PTE */
622 	if (mm_ops->dcache_clean_inval_poc && stage2_pte_cacheable(pgt, new))
623 		mm_ops->dcache_clean_inval_poc(kvm_pte_follow(new, mm_ops),
624 						granule);
625 
626 	if (mm_ops->icache_inval_pou && stage2_pte_executable(new))
627 		mm_ops->icache_inval_pou(kvm_pte_follow(new, mm_ops), granule);
628 
629 	smp_store_release(ptep, new);
630 	if (stage2_pte_is_counted(new))
631 		mm_ops->get_page(ptep);
632 	if (kvm_phys_is_valid(phys))
633 		data->phys += granule;
634 	return 0;
635 }
636 
637 static int stage2_map_walk_table_pre(u64 addr, u64 end, u32 level,
638 				     kvm_pte_t *ptep,
639 				     struct stage2_map_data *data)
640 {
641 	if (data->anchor)
642 		return 0;
643 
644 	if (!kvm_block_mapping_supported(addr, end, data->phys, level))
645 		return 0;
646 
647 	data->childp = kvm_pte_follow(*ptep, data->mm_ops);
648 	kvm_clear_pte(ptep);
649 
650 	/*
651 	 * Invalidate the whole stage-2, as we may have numerous leaf
652 	 * entries below us which would otherwise need invalidating
653 	 * individually.
654 	 */
655 	kvm_call_hyp(__kvm_tlb_flush_vmid, data->mmu);
656 	data->anchor = ptep;
657 	return 0;
658 }
659 
660 static int stage2_map_walk_leaf(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
661 				struct stage2_map_data *data)
662 {
663 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
664 	kvm_pte_t *childp, pte = *ptep;
665 	int ret;
666 
667 	if (data->anchor) {
668 		if (stage2_pte_is_counted(pte))
669 			mm_ops->put_page(ptep);
670 
671 		return 0;
672 	}
673 
674 	ret = stage2_map_walker_try_leaf(addr, end, level, ptep, data);
675 	if (ret != -E2BIG)
676 		return ret;
677 
678 	if (WARN_ON(level == KVM_PGTABLE_MAX_LEVELS - 1))
679 		return -EINVAL;
680 
681 	if (!data->memcache)
682 		return -ENOMEM;
683 
684 	childp = mm_ops->zalloc_page(data->memcache);
685 	if (!childp)
686 		return -ENOMEM;
687 
688 	/*
689 	 * If we've run into an existing block mapping then replace it with
690 	 * a table. Accesses beyond 'end' that fall within the new table
691 	 * will be mapped lazily.
692 	 */
693 	if (stage2_pte_is_counted(pte))
694 		stage2_put_pte(ptep, data->mmu, addr, level, mm_ops);
695 
696 	kvm_set_table_pte(ptep, childp, mm_ops);
697 	mm_ops->get_page(ptep);
698 
699 	return 0;
700 }
701 
702 static int stage2_map_walk_table_post(u64 addr, u64 end, u32 level,
703 				      kvm_pte_t *ptep,
704 				      struct stage2_map_data *data)
705 {
706 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
707 	kvm_pte_t *childp;
708 	int ret = 0;
709 
710 	if (!data->anchor)
711 		return 0;
712 
713 	if (data->anchor == ptep) {
714 		childp = data->childp;
715 		data->anchor = NULL;
716 		data->childp = NULL;
717 		ret = stage2_map_walk_leaf(addr, end, level, ptep, data);
718 	} else {
719 		childp = kvm_pte_follow(*ptep, mm_ops);
720 	}
721 
722 	mm_ops->put_page(childp);
723 	mm_ops->put_page(ptep);
724 
725 	return ret;
726 }
727 
728 /*
729  * This is a little fiddly, as we use all three of the walk flags. The idea
730  * is that the TABLE_PRE callback runs for table entries on the way down,
731  * looking for table entries which we could conceivably replace with a
732  * block entry for this mapping. If it finds one, then it sets the 'anchor'
733  * field in 'struct stage2_map_data' to point at the table entry, before
734  * clearing the entry to zero and descending into the now detached table.
735  *
736  * The behaviour of the LEAF callback then depends on whether or not the
737  * anchor has been set. If not, then we're not using a block mapping higher
738  * up the table and we perform the mapping at the existing leaves instead.
739  * If, on the other hand, the anchor _is_ set, then we drop references to
740  * all valid leaves so that the pages beneath the anchor can be freed.
741  *
742  * Finally, the TABLE_POST callback does nothing if the anchor has not
743  * been set, but otherwise frees the page-table pages while walking back up
744  * the page-table, installing the block entry when it revisits the anchor
745  * pointer and clearing the anchor to NULL.
746  */
747 static int stage2_map_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
748 			     enum kvm_pgtable_walk_flags flag, void * const arg)
749 {
750 	struct stage2_map_data *data = arg;
751 
752 	switch (flag) {
753 	case KVM_PGTABLE_WALK_TABLE_PRE:
754 		return stage2_map_walk_table_pre(addr, end, level, ptep, data);
755 	case KVM_PGTABLE_WALK_LEAF:
756 		return stage2_map_walk_leaf(addr, end, level, ptep, data);
757 	case KVM_PGTABLE_WALK_TABLE_POST:
758 		return stage2_map_walk_table_post(addr, end, level, ptep, data);
759 	}
760 
761 	return -EINVAL;
762 }
763 
764 int kvm_pgtable_stage2_map(struct kvm_pgtable *pgt, u64 addr, u64 size,
765 			   u64 phys, enum kvm_pgtable_prot prot,
766 			   void *mc)
767 {
768 	int ret;
769 	struct stage2_map_data map_data = {
770 		.phys		= ALIGN_DOWN(phys, PAGE_SIZE),
771 		.mmu		= pgt->mmu,
772 		.memcache	= mc,
773 		.mm_ops		= pgt->mm_ops,
774 	};
775 	struct kvm_pgtable_walker walker = {
776 		.cb		= stage2_map_walker,
777 		.flags		= KVM_PGTABLE_WALK_TABLE_PRE |
778 				  KVM_PGTABLE_WALK_LEAF |
779 				  KVM_PGTABLE_WALK_TABLE_POST,
780 		.arg		= &map_data,
781 	};
782 
783 	if (WARN_ON((pgt->flags & KVM_PGTABLE_S2_IDMAP) && (addr != phys)))
784 		return -EINVAL;
785 
786 	ret = stage2_set_prot_attr(pgt, prot, &map_data.attr);
787 	if (ret)
788 		return ret;
789 
790 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
791 	dsb(ishst);
792 	return ret;
793 }
794 
795 int kvm_pgtable_stage2_set_owner(struct kvm_pgtable *pgt, u64 addr, u64 size,
796 				 void *mc, u8 owner_id)
797 {
798 	int ret;
799 	struct stage2_map_data map_data = {
800 		.phys		= KVM_PHYS_INVALID,
801 		.mmu		= pgt->mmu,
802 		.memcache	= mc,
803 		.mm_ops		= pgt->mm_ops,
804 		.owner_id	= owner_id,
805 	};
806 	struct kvm_pgtable_walker walker = {
807 		.cb		= stage2_map_walker,
808 		.flags		= KVM_PGTABLE_WALK_TABLE_PRE |
809 				  KVM_PGTABLE_WALK_LEAF |
810 				  KVM_PGTABLE_WALK_TABLE_POST,
811 		.arg		= &map_data,
812 	};
813 
814 	if (owner_id > KVM_MAX_OWNER_ID)
815 		return -EINVAL;
816 
817 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
818 	return ret;
819 }
820 
821 static int stage2_unmap_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
822 			       enum kvm_pgtable_walk_flags flag,
823 			       void * const arg)
824 {
825 	struct kvm_pgtable *pgt = arg;
826 	struct kvm_s2_mmu *mmu = pgt->mmu;
827 	struct kvm_pgtable_mm_ops *mm_ops = pgt->mm_ops;
828 	kvm_pte_t pte = *ptep, *childp = NULL;
829 	bool need_flush = false;
830 
831 	if (!kvm_pte_valid(pte)) {
832 		if (stage2_pte_is_counted(pte)) {
833 			kvm_clear_pte(ptep);
834 			mm_ops->put_page(ptep);
835 		}
836 		return 0;
837 	}
838 
839 	if (kvm_pte_table(pte, level)) {
840 		childp = kvm_pte_follow(pte, mm_ops);
841 
842 		if (mm_ops->page_count(childp) != 1)
843 			return 0;
844 	} else if (stage2_pte_cacheable(pgt, pte)) {
845 		need_flush = !stage2_has_fwb(pgt);
846 	}
847 
848 	/*
849 	 * This is similar to the map() path in that we unmap the entire
850 	 * block entry and rely on the remaining portions being faulted
851 	 * back lazily.
852 	 */
853 	stage2_put_pte(ptep, mmu, addr, level, mm_ops);
854 
855 	if (need_flush) {
856 		kvm_pte_t *pte_follow = kvm_pte_follow(pte, mm_ops);
857 
858 		dcache_clean_inval_poc((unsigned long)pte_follow,
859 				    (unsigned long)pte_follow +
860 					    kvm_granule_size(level));
861 	}
862 
863 	if (childp)
864 		mm_ops->put_page(childp);
865 
866 	return 0;
867 }
868 
869 int kvm_pgtable_stage2_unmap(struct kvm_pgtable *pgt, u64 addr, u64 size)
870 {
871 	struct kvm_pgtable_walker walker = {
872 		.cb	= stage2_unmap_walker,
873 		.arg	= pgt,
874 		.flags	= KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
875 	};
876 
877 	return kvm_pgtable_walk(pgt, addr, size, &walker);
878 }
879 
880 struct stage2_attr_data {
881 	kvm_pte_t			attr_set;
882 	kvm_pte_t			attr_clr;
883 	kvm_pte_t			pte;
884 	u32				level;
885 	struct kvm_pgtable_mm_ops	*mm_ops;
886 };
887 
888 static int stage2_attr_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
889 			      enum kvm_pgtable_walk_flags flag,
890 			      void * const arg)
891 {
892 	kvm_pte_t pte = *ptep;
893 	struct stage2_attr_data *data = arg;
894 	struct kvm_pgtable_mm_ops *mm_ops = data->mm_ops;
895 
896 	if (!kvm_pte_valid(pte))
897 		return 0;
898 
899 	data->level = level;
900 	data->pte = pte;
901 	pte &= ~data->attr_clr;
902 	pte |= data->attr_set;
903 
904 	/*
905 	 * We may race with the CPU trying to set the access flag here,
906 	 * but worst-case the access flag update gets lost and will be
907 	 * set on the next access instead.
908 	 */
909 	if (data->pte != pte) {
910 		/*
911 		 * Invalidate instruction cache before updating the guest
912 		 * stage-2 PTE if we are going to add executable permission.
913 		 */
914 		if (mm_ops->icache_inval_pou &&
915 		    stage2_pte_executable(pte) && !stage2_pte_executable(*ptep))
916 			mm_ops->icache_inval_pou(kvm_pte_follow(pte, mm_ops),
917 						  kvm_granule_size(level));
918 		WRITE_ONCE(*ptep, pte);
919 	}
920 
921 	return 0;
922 }
923 
924 static int stage2_update_leaf_attrs(struct kvm_pgtable *pgt, u64 addr,
925 				    u64 size, kvm_pte_t attr_set,
926 				    kvm_pte_t attr_clr, kvm_pte_t *orig_pte,
927 				    u32 *level)
928 {
929 	int ret;
930 	kvm_pte_t attr_mask = KVM_PTE_LEAF_ATTR_LO | KVM_PTE_LEAF_ATTR_HI;
931 	struct stage2_attr_data data = {
932 		.attr_set	= attr_set & attr_mask,
933 		.attr_clr	= attr_clr & attr_mask,
934 		.mm_ops		= pgt->mm_ops,
935 	};
936 	struct kvm_pgtable_walker walker = {
937 		.cb		= stage2_attr_walker,
938 		.arg		= &data,
939 		.flags		= KVM_PGTABLE_WALK_LEAF,
940 	};
941 
942 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
943 	if (ret)
944 		return ret;
945 
946 	if (orig_pte)
947 		*orig_pte = data.pte;
948 
949 	if (level)
950 		*level = data.level;
951 	return 0;
952 }
953 
954 int kvm_pgtable_stage2_wrprotect(struct kvm_pgtable *pgt, u64 addr, u64 size)
955 {
956 	return stage2_update_leaf_attrs(pgt, addr, size, 0,
957 					KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W,
958 					NULL, NULL);
959 }
960 
961 kvm_pte_t kvm_pgtable_stage2_mkyoung(struct kvm_pgtable *pgt, u64 addr)
962 {
963 	kvm_pte_t pte = 0;
964 	stage2_update_leaf_attrs(pgt, addr, 1, KVM_PTE_LEAF_ATTR_LO_S2_AF, 0,
965 				 &pte, NULL);
966 	dsb(ishst);
967 	return pte;
968 }
969 
970 kvm_pte_t kvm_pgtable_stage2_mkold(struct kvm_pgtable *pgt, u64 addr)
971 {
972 	kvm_pte_t pte = 0;
973 	stage2_update_leaf_attrs(pgt, addr, 1, 0, KVM_PTE_LEAF_ATTR_LO_S2_AF,
974 				 &pte, NULL);
975 	/*
976 	 * "But where's the TLBI?!", you scream.
977 	 * "Over in the core code", I sigh.
978 	 *
979 	 * See the '->clear_flush_young()' callback on the KVM mmu notifier.
980 	 */
981 	return pte;
982 }
983 
984 bool kvm_pgtable_stage2_is_young(struct kvm_pgtable *pgt, u64 addr)
985 {
986 	kvm_pte_t pte = 0;
987 	stage2_update_leaf_attrs(pgt, addr, 1, 0, 0, &pte, NULL);
988 	return pte & KVM_PTE_LEAF_ATTR_LO_S2_AF;
989 }
990 
991 int kvm_pgtable_stage2_relax_perms(struct kvm_pgtable *pgt, u64 addr,
992 				   enum kvm_pgtable_prot prot)
993 {
994 	int ret;
995 	u32 level;
996 	kvm_pte_t set = 0, clr = 0;
997 
998 	if (prot & KVM_PGTABLE_PROT_R)
999 		set |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
1000 
1001 	if (prot & KVM_PGTABLE_PROT_W)
1002 		set |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
1003 
1004 	if (prot & KVM_PGTABLE_PROT_X)
1005 		clr |= KVM_PTE_LEAF_ATTR_HI_S2_XN;
1006 
1007 	ret = stage2_update_leaf_attrs(pgt, addr, 1, set, clr, NULL, &level);
1008 	if (!ret)
1009 		kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, pgt->mmu, addr, level);
1010 	return ret;
1011 }
1012 
1013 static int stage2_flush_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
1014 			       enum kvm_pgtable_walk_flags flag,
1015 			       void * const arg)
1016 {
1017 	struct kvm_pgtable *pgt = arg;
1018 	struct kvm_pgtable_mm_ops *mm_ops = pgt->mm_ops;
1019 	kvm_pte_t pte = *ptep;
1020 	kvm_pte_t *pte_follow;
1021 
1022 	if (!kvm_pte_valid(pte) || !stage2_pte_cacheable(pgt, pte))
1023 		return 0;
1024 
1025 	pte_follow = kvm_pte_follow(pte, mm_ops);
1026 	dcache_clean_inval_poc((unsigned long)pte_follow,
1027 			    (unsigned long)pte_follow +
1028 				    kvm_granule_size(level));
1029 	return 0;
1030 }
1031 
1032 int kvm_pgtable_stage2_flush(struct kvm_pgtable *pgt, u64 addr, u64 size)
1033 {
1034 	struct kvm_pgtable_walker walker = {
1035 		.cb	= stage2_flush_walker,
1036 		.flags	= KVM_PGTABLE_WALK_LEAF,
1037 		.arg	= pgt,
1038 	};
1039 
1040 	if (stage2_has_fwb(pgt))
1041 		return 0;
1042 
1043 	return kvm_pgtable_walk(pgt, addr, size, &walker);
1044 }
1045 
1046 int kvm_pgtable_stage2_init_flags(struct kvm_pgtable *pgt, struct kvm_arch *arch,
1047 				  struct kvm_pgtable_mm_ops *mm_ops,
1048 				  enum kvm_pgtable_stage2_flags flags)
1049 {
1050 	size_t pgd_sz;
1051 	u64 vtcr = arch->vtcr;
1052 	u32 ia_bits = VTCR_EL2_IPA(vtcr);
1053 	u32 sl0 = FIELD_GET(VTCR_EL2_SL0_MASK, vtcr);
1054 	u32 start_level = VTCR_EL2_TGRAN_SL0_BASE - sl0;
1055 
1056 	pgd_sz = kvm_pgd_pages(ia_bits, start_level) * PAGE_SIZE;
1057 	pgt->pgd = mm_ops->zalloc_pages_exact(pgd_sz);
1058 	if (!pgt->pgd)
1059 		return -ENOMEM;
1060 
1061 	pgt->ia_bits		= ia_bits;
1062 	pgt->start_level	= start_level;
1063 	pgt->mm_ops		= mm_ops;
1064 	pgt->mmu		= &arch->mmu;
1065 	pgt->flags		= flags;
1066 
1067 	/* Ensure zeroed PGD pages are visible to the hardware walker */
1068 	dsb(ishst);
1069 	return 0;
1070 }
1071 
1072 static int stage2_free_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
1073 			      enum kvm_pgtable_walk_flags flag,
1074 			      void * const arg)
1075 {
1076 	struct kvm_pgtable_mm_ops *mm_ops = arg;
1077 	kvm_pte_t pte = *ptep;
1078 
1079 	if (!stage2_pte_is_counted(pte))
1080 		return 0;
1081 
1082 	mm_ops->put_page(ptep);
1083 
1084 	if (kvm_pte_table(pte, level))
1085 		mm_ops->put_page(kvm_pte_follow(pte, mm_ops));
1086 
1087 	return 0;
1088 }
1089 
1090 void kvm_pgtable_stage2_destroy(struct kvm_pgtable *pgt)
1091 {
1092 	size_t pgd_sz;
1093 	struct kvm_pgtable_walker walker = {
1094 		.cb	= stage2_free_walker,
1095 		.flags	= KVM_PGTABLE_WALK_LEAF |
1096 			  KVM_PGTABLE_WALK_TABLE_POST,
1097 		.arg	= pgt->mm_ops,
1098 	};
1099 
1100 	WARN_ON(kvm_pgtable_walk(pgt, 0, BIT(pgt->ia_bits), &walker));
1101 	pgd_sz = kvm_pgd_pages(pgt->ia_bits, pgt->start_level) * PAGE_SIZE;
1102 	pgt->mm_ops->free_pages_exact(pgt->pgd, pgd_sz);
1103 	pgt->pgd = NULL;
1104 }
1105 
1106 #define KVM_PTE_LEAF_S2_COMPAT_MASK	(KVM_PTE_LEAF_ATTR_S2_PERMS | \
1107 					 KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR | \
1108 					 KVM_PTE_LEAF_ATTR_S2_IGNORED)
1109 
1110 static int stage2_check_permission_walker(u64 addr, u64 end, u32 level,
1111 					  kvm_pte_t *ptep,
1112 					  enum kvm_pgtable_walk_flags flag,
1113 					  void * const arg)
1114 {
1115 	kvm_pte_t old_attr, pte = *ptep, *new_attr = arg;
1116 
1117 	/*
1118 	 * Compatible mappings are either invalid and owned by the page-table
1119 	 * owner (whose id is 0), or valid with matching permission attributes.
1120 	 */
1121 	if (kvm_pte_valid(pte)) {
1122 		old_attr = pte & KVM_PTE_LEAF_S2_COMPAT_MASK;
1123 		if (old_attr != *new_attr)
1124 			return -EEXIST;
1125 	} else if (pte) {
1126 		return -EEXIST;
1127 	}
1128 
1129 	return 0;
1130 }
1131 
1132 int kvm_pgtable_stage2_find_range(struct kvm_pgtable *pgt, u64 addr,
1133 				  enum kvm_pgtable_prot prot,
1134 				  struct kvm_mem_range *range)
1135 {
1136 	kvm_pte_t attr;
1137 	struct kvm_pgtable_walker check_perm_walker = {
1138 		.cb		= stage2_check_permission_walker,
1139 		.flags		= KVM_PGTABLE_WALK_LEAF,
1140 		.arg		= &attr,
1141 	};
1142 	u64 granule, start, end;
1143 	u32 level;
1144 	int ret;
1145 
1146 	ret = stage2_set_prot_attr(pgt, prot, &attr);
1147 	if (ret)
1148 		return ret;
1149 	attr &= KVM_PTE_LEAF_S2_COMPAT_MASK;
1150 
1151 	for (level = pgt->start_level; level < KVM_PGTABLE_MAX_LEVELS; level++) {
1152 		granule = kvm_granule_size(level);
1153 		start = ALIGN_DOWN(addr, granule);
1154 		end = start + granule;
1155 
1156 		if (!kvm_level_supports_block_mapping(level))
1157 			continue;
1158 
1159 		if (start < range->start || range->end < end)
1160 			continue;
1161 
1162 		/*
1163 		 * Check the presence of existing mappings with incompatible
1164 		 * permissions within the current block range, and try one level
1165 		 * deeper if one is found.
1166 		 */
1167 		ret = kvm_pgtable_walk(pgt, start, granule, &check_perm_walker);
1168 		if (ret != -EEXIST)
1169 			break;
1170 	}
1171 
1172 	if (!ret) {
1173 		range->start = start;
1174 		range->end = end;
1175 	}
1176 
1177 	return ret;
1178 }
1179