xref: /openbmc/linux/arch/arm64/kvm/hyp/pgtable.c (revision 0eb76ba2)
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 
13 #define KVM_PGTABLE_MAX_LEVELS		4U
14 
15 #define KVM_PTE_VALID			BIT(0)
16 
17 #define KVM_PTE_TYPE			BIT(1)
18 #define KVM_PTE_TYPE_BLOCK		0
19 #define KVM_PTE_TYPE_PAGE		1
20 #define KVM_PTE_TYPE_TABLE		1
21 
22 #define KVM_PTE_ADDR_MASK		GENMASK(47, PAGE_SHIFT)
23 #define KVM_PTE_ADDR_51_48		GENMASK(15, 12)
24 
25 #define KVM_PTE_LEAF_ATTR_LO		GENMASK(11, 2)
26 
27 #define KVM_PTE_LEAF_ATTR_LO_S1_ATTRIDX	GENMASK(4, 2)
28 #define KVM_PTE_LEAF_ATTR_LO_S1_AP	GENMASK(7, 6)
29 #define KVM_PTE_LEAF_ATTR_LO_S1_AP_RO	3
30 #define KVM_PTE_LEAF_ATTR_LO_S1_AP_RW	1
31 #define KVM_PTE_LEAF_ATTR_LO_S1_SH	GENMASK(9, 8)
32 #define KVM_PTE_LEAF_ATTR_LO_S1_SH_IS	3
33 #define KVM_PTE_LEAF_ATTR_LO_S1_AF	BIT(10)
34 
35 #define KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR	GENMASK(5, 2)
36 #define KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R	BIT(6)
37 #define KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W	BIT(7)
38 #define KVM_PTE_LEAF_ATTR_LO_S2_SH	GENMASK(9, 8)
39 #define KVM_PTE_LEAF_ATTR_LO_S2_SH_IS	3
40 #define KVM_PTE_LEAF_ATTR_LO_S2_AF	BIT(10)
41 
42 #define KVM_PTE_LEAF_ATTR_HI		GENMASK(63, 51)
43 
44 #define KVM_PTE_LEAF_ATTR_HI_S1_XN	BIT(54)
45 
46 #define KVM_PTE_LEAF_ATTR_HI_S2_XN	BIT(54)
47 
48 struct kvm_pgtable_walk_data {
49 	struct kvm_pgtable		*pgt;
50 	struct kvm_pgtable_walker	*walker;
51 
52 	u64				addr;
53 	u64				end;
54 };
55 
56 static u64 kvm_granule_shift(u32 level)
57 {
58 	/* Assumes KVM_PGTABLE_MAX_LEVELS is 4 */
59 	return ARM64_HW_PGTABLE_LEVEL_SHIFT(level);
60 }
61 
62 static u64 kvm_granule_size(u32 level)
63 {
64 	return BIT(kvm_granule_shift(level));
65 }
66 
67 static bool kvm_block_mapping_supported(u64 addr, u64 end, u64 phys, u32 level)
68 {
69 	u64 granule = kvm_granule_size(level);
70 
71 	/*
72 	 * Reject invalid block mappings and don't bother with 4TB mappings for
73 	 * 52-bit PAs.
74 	 */
75 	if (level == 0 || (PAGE_SIZE != SZ_4K && level == 1))
76 		return false;
77 
78 	if (granule > (end - addr))
79 		return false;
80 
81 	return IS_ALIGNED(addr, granule) && IS_ALIGNED(phys, granule);
82 }
83 
84 static u32 kvm_pgtable_idx(struct kvm_pgtable_walk_data *data, u32 level)
85 {
86 	u64 shift = kvm_granule_shift(level);
87 	u64 mask = BIT(PAGE_SHIFT - 3) - 1;
88 
89 	return (data->addr >> shift) & mask;
90 }
91 
92 static u32 __kvm_pgd_page_idx(struct kvm_pgtable *pgt, u64 addr)
93 {
94 	u64 shift = kvm_granule_shift(pgt->start_level - 1); /* May underflow */
95 	u64 mask = BIT(pgt->ia_bits) - 1;
96 
97 	return (addr & mask) >> shift;
98 }
99 
100 static u32 kvm_pgd_page_idx(struct kvm_pgtable_walk_data *data)
101 {
102 	return __kvm_pgd_page_idx(data->pgt, data->addr);
103 }
104 
105 static u32 kvm_pgd_pages(u32 ia_bits, u32 start_level)
106 {
107 	struct kvm_pgtable pgt = {
108 		.ia_bits	= ia_bits,
109 		.start_level	= start_level,
110 	};
111 
112 	return __kvm_pgd_page_idx(&pgt, -1ULL) + 1;
113 }
114 
115 static bool kvm_pte_valid(kvm_pte_t pte)
116 {
117 	return pte & KVM_PTE_VALID;
118 }
119 
120 static bool kvm_pte_table(kvm_pte_t pte, u32 level)
121 {
122 	if (level == KVM_PGTABLE_MAX_LEVELS - 1)
123 		return false;
124 
125 	if (!kvm_pte_valid(pte))
126 		return false;
127 
128 	return FIELD_GET(KVM_PTE_TYPE, pte) == KVM_PTE_TYPE_TABLE;
129 }
130 
131 static u64 kvm_pte_to_phys(kvm_pte_t pte)
132 {
133 	u64 pa = pte & KVM_PTE_ADDR_MASK;
134 
135 	if (PAGE_SHIFT == 16)
136 		pa |= FIELD_GET(KVM_PTE_ADDR_51_48, pte) << 48;
137 
138 	return pa;
139 }
140 
141 static kvm_pte_t kvm_phys_to_pte(u64 pa)
142 {
143 	kvm_pte_t pte = pa & KVM_PTE_ADDR_MASK;
144 
145 	if (PAGE_SHIFT == 16)
146 		pte |= FIELD_PREP(KVM_PTE_ADDR_51_48, pa >> 48);
147 
148 	return pte;
149 }
150 
151 static kvm_pte_t *kvm_pte_follow(kvm_pte_t pte)
152 {
153 	return __va(kvm_pte_to_phys(pte));
154 }
155 
156 static void kvm_set_invalid_pte(kvm_pte_t *ptep)
157 {
158 	kvm_pte_t pte = *ptep;
159 	WRITE_ONCE(*ptep, pte & ~KVM_PTE_VALID);
160 }
161 
162 static void kvm_set_table_pte(kvm_pte_t *ptep, kvm_pte_t *childp)
163 {
164 	kvm_pte_t old = *ptep, pte = kvm_phys_to_pte(__pa(childp));
165 
166 	pte |= FIELD_PREP(KVM_PTE_TYPE, KVM_PTE_TYPE_TABLE);
167 	pte |= KVM_PTE_VALID;
168 
169 	WARN_ON(kvm_pte_valid(old));
170 	smp_store_release(ptep, pte);
171 }
172 
173 static bool kvm_set_valid_leaf_pte(kvm_pte_t *ptep, u64 pa, kvm_pte_t attr,
174 				   u32 level)
175 {
176 	kvm_pte_t old = *ptep, pte = kvm_phys_to_pte(pa);
177 	u64 type = (level == KVM_PGTABLE_MAX_LEVELS - 1) ? KVM_PTE_TYPE_PAGE :
178 							   KVM_PTE_TYPE_BLOCK;
179 
180 	pte |= attr & (KVM_PTE_LEAF_ATTR_LO | KVM_PTE_LEAF_ATTR_HI);
181 	pte |= FIELD_PREP(KVM_PTE_TYPE, type);
182 	pte |= KVM_PTE_VALID;
183 
184 	/* Tolerate KVM recreating the exact same mapping. */
185 	if (kvm_pte_valid(old))
186 		return old == pte;
187 
188 	smp_store_release(ptep, pte);
189 	return true;
190 }
191 
192 static int kvm_pgtable_visitor_cb(struct kvm_pgtable_walk_data *data, u64 addr,
193 				  u32 level, kvm_pte_t *ptep,
194 				  enum kvm_pgtable_walk_flags flag)
195 {
196 	struct kvm_pgtable_walker *walker = data->walker;
197 	return walker->cb(addr, data->end, level, ptep, flag, walker->arg);
198 }
199 
200 static int __kvm_pgtable_walk(struct kvm_pgtable_walk_data *data,
201 			      kvm_pte_t *pgtable, u32 level);
202 
203 static inline int __kvm_pgtable_visit(struct kvm_pgtable_walk_data *data,
204 				      kvm_pte_t *ptep, u32 level)
205 {
206 	int ret = 0;
207 	u64 addr = data->addr;
208 	kvm_pte_t *childp, pte = *ptep;
209 	bool table = kvm_pte_table(pte, level);
210 	enum kvm_pgtable_walk_flags flags = data->walker->flags;
211 
212 	if (table && (flags & KVM_PGTABLE_WALK_TABLE_PRE)) {
213 		ret = kvm_pgtable_visitor_cb(data, addr, level, ptep,
214 					     KVM_PGTABLE_WALK_TABLE_PRE);
215 	}
216 
217 	if (!table && (flags & KVM_PGTABLE_WALK_LEAF)) {
218 		ret = kvm_pgtable_visitor_cb(data, addr, level, ptep,
219 					     KVM_PGTABLE_WALK_LEAF);
220 		pte = *ptep;
221 		table = kvm_pte_table(pte, level);
222 	}
223 
224 	if (ret)
225 		goto out;
226 
227 	if (!table) {
228 		data->addr += kvm_granule_size(level);
229 		goto out;
230 	}
231 
232 	childp = kvm_pte_follow(pte);
233 	ret = __kvm_pgtable_walk(data, childp, level + 1);
234 	if (ret)
235 		goto out;
236 
237 	if (flags & KVM_PGTABLE_WALK_TABLE_POST) {
238 		ret = kvm_pgtable_visitor_cb(data, addr, level, ptep,
239 					     KVM_PGTABLE_WALK_TABLE_POST);
240 	}
241 
242 out:
243 	return ret;
244 }
245 
246 static int __kvm_pgtable_walk(struct kvm_pgtable_walk_data *data,
247 			      kvm_pte_t *pgtable, u32 level)
248 {
249 	u32 idx;
250 	int ret = 0;
251 
252 	if (WARN_ON_ONCE(level >= KVM_PGTABLE_MAX_LEVELS))
253 		return -EINVAL;
254 
255 	for (idx = kvm_pgtable_idx(data, level); idx < PTRS_PER_PTE; ++idx) {
256 		kvm_pte_t *ptep = &pgtable[idx];
257 
258 		if (data->addr >= data->end)
259 			break;
260 
261 		ret = __kvm_pgtable_visit(data, ptep, level);
262 		if (ret)
263 			break;
264 	}
265 
266 	return ret;
267 }
268 
269 static int _kvm_pgtable_walk(struct kvm_pgtable_walk_data *data)
270 {
271 	u32 idx;
272 	int ret = 0;
273 	struct kvm_pgtable *pgt = data->pgt;
274 	u64 limit = BIT(pgt->ia_bits);
275 
276 	if (data->addr > limit || data->end > limit)
277 		return -ERANGE;
278 
279 	if (!pgt->pgd)
280 		return -EINVAL;
281 
282 	for (idx = kvm_pgd_page_idx(data); data->addr < data->end; ++idx) {
283 		kvm_pte_t *ptep = &pgt->pgd[idx * PTRS_PER_PTE];
284 
285 		ret = __kvm_pgtable_walk(data, ptep, pgt->start_level);
286 		if (ret)
287 			break;
288 	}
289 
290 	return ret;
291 }
292 
293 int kvm_pgtable_walk(struct kvm_pgtable *pgt, u64 addr, u64 size,
294 		     struct kvm_pgtable_walker *walker)
295 {
296 	struct kvm_pgtable_walk_data walk_data = {
297 		.pgt	= pgt,
298 		.addr	= ALIGN_DOWN(addr, PAGE_SIZE),
299 		.end	= PAGE_ALIGN(walk_data.addr + size),
300 		.walker	= walker,
301 	};
302 
303 	return _kvm_pgtable_walk(&walk_data);
304 }
305 
306 struct hyp_map_data {
307 	u64		phys;
308 	kvm_pte_t	attr;
309 };
310 
311 static int hyp_map_set_prot_attr(enum kvm_pgtable_prot prot,
312 				 struct hyp_map_data *data)
313 {
314 	bool device = prot & KVM_PGTABLE_PROT_DEVICE;
315 	u32 mtype = device ? MT_DEVICE_nGnRE : MT_NORMAL;
316 	kvm_pte_t attr = FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S1_ATTRIDX, mtype);
317 	u32 sh = KVM_PTE_LEAF_ATTR_LO_S1_SH_IS;
318 	u32 ap = (prot & KVM_PGTABLE_PROT_W) ? KVM_PTE_LEAF_ATTR_LO_S1_AP_RW :
319 					       KVM_PTE_LEAF_ATTR_LO_S1_AP_RO;
320 
321 	if (!(prot & KVM_PGTABLE_PROT_R))
322 		return -EINVAL;
323 
324 	if (prot & KVM_PGTABLE_PROT_X) {
325 		if (prot & KVM_PGTABLE_PROT_W)
326 			return -EINVAL;
327 
328 		if (device)
329 			return -EINVAL;
330 	} else {
331 		attr |= KVM_PTE_LEAF_ATTR_HI_S1_XN;
332 	}
333 
334 	attr |= FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S1_AP, ap);
335 	attr |= FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S1_SH, sh);
336 	attr |= KVM_PTE_LEAF_ATTR_LO_S1_AF;
337 	data->attr = attr;
338 	return 0;
339 }
340 
341 static bool hyp_map_walker_try_leaf(u64 addr, u64 end, u32 level,
342 				    kvm_pte_t *ptep, struct hyp_map_data *data)
343 {
344 	u64 granule = kvm_granule_size(level), phys = data->phys;
345 
346 	if (!kvm_block_mapping_supported(addr, end, phys, level))
347 		return false;
348 
349 	WARN_ON(!kvm_set_valid_leaf_pte(ptep, phys, data->attr, level));
350 	data->phys += granule;
351 	return true;
352 }
353 
354 static int hyp_map_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
355 			  enum kvm_pgtable_walk_flags flag, void * const arg)
356 {
357 	kvm_pte_t *childp;
358 
359 	if (hyp_map_walker_try_leaf(addr, end, level, ptep, arg))
360 		return 0;
361 
362 	if (WARN_ON(level == KVM_PGTABLE_MAX_LEVELS - 1))
363 		return -EINVAL;
364 
365 	childp = (kvm_pte_t *)get_zeroed_page(GFP_KERNEL);
366 	if (!childp)
367 		return -ENOMEM;
368 
369 	kvm_set_table_pte(ptep, childp);
370 	return 0;
371 }
372 
373 int kvm_pgtable_hyp_map(struct kvm_pgtable *pgt, u64 addr, u64 size, u64 phys,
374 			enum kvm_pgtable_prot prot)
375 {
376 	int ret;
377 	struct hyp_map_data map_data = {
378 		.phys	= ALIGN_DOWN(phys, PAGE_SIZE),
379 	};
380 	struct kvm_pgtable_walker walker = {
381 		.cb	= hyp_map_walker,
382 		.flags	= KVM_PGTABLE_WALK_LEAF,
383 		.arg	= &map_data,
384 	};
385 
386 	ret = hyp_map_set_prot_attr(prot, &map_data);
387 	if (ret)
388 		return ret;
389 
390 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
391 	dsb(ishst);
392 	isb();
393 	return ret;
394 }
395 
396 int kvm_pgtable_hyp_init(struct kvm_pgtable *pgt, u32 va_bits)
397 {
398 	u64 levels = ARM64_HW_PGTABLE_LEVELS(va_bits);
399 
400 	pgt->pgd = (kvm_pte_t *)get_zeroed_page(GFP_KERNEL);
401 	if (!pgt->pgd)
402 		return -ENOMEM;
403 
404 	pgt->ia_bits		= va_bits;
405 	pgt->start_level	= KVM_PGTABLE_MAX_LEVELS - levels;
406 	pgt->mmu		= NULL;
407 	return 0;
408 }
409 
410 static int hyp_free_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
411 			   enum kvm_pgtable_walk_flags flag, void * const arg)
412 {
413 	free_page((unsigned long)kvm_pte_follow(*ptep));
414 	return 0;
415 }
416 
417 void kvm_pgtable_hyp_destroy(struct kvm_pgtable *pgt)
418 {
419 	struct kvm_pgtable_walker walker = {
420 		.cb	= hyp_free_walker,
421 		.flags	= KVM_PGTABLE_WALK_TABLE_POST,
422 	};
423 
424 	WARN_ON(kvm_pgtable_walk(pgt, 0, BIT(pgt->ia_bits), &walker));
425 	free_page((unsigned long)pgt->pgd);
426 	pgt->pgd = NULL;
427 }
428 
429 struct stage2_map_data {
430 	u64				phys;
431 	kvm_pte_t			attr;
432 
433 	kvm_pte_t			*anchor;
434 
435 	struct kvm_s2_mmu		*mmu;
436 	struct kvm_mmu_memory_cache	*memcache;
437 };
438 
439 static int stage2_map_set_prot_attr(enum kvm_pgtable_prot prot,
440 				    struct stage2_map_data *data)
441 {
442 	bool device = prot & KVM_PGTABLE_PROT_DEVICE;
443 	kvm_pte_t attr = device ? PAGE_S2_MEMATTR(DEVICE_nGnRE) :
444 			    PAGE_S2_MEMATTR(NORMAL);
445 	u32 sh = KVM_PTE_LEAF_ATTR_LO_S2_SH_IS;
446 
447 	if (!(prot & KVM_PGTABLE_PROT_X))
448 		attr |= KVM_PTE_LEAF_ATTR_HI_S2_XN;
449 	else if (device)
450 		return -EINVAL;
451 
452 	if (prot & KVM_PGTABLE_PROT_R)
453 		attr |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
454 
455 	if (prot & KVM_PGTABLE_PROT_W)
456 		attr |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
457 
458 	attr |= FIELD_PREP(KVM_PTE_LEAF_ATTR_LO_S2_SH, sh);
459 	attr |= KVM_PTE_LEAF_ATTR_LO_S2_AF;
460 	data->attr = attr;
461 	return 0;
462 }
463 
464 static bool stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
465 				       kvm_pte_t *ptep,
466 				       struct stage2_map_data *data)
467 {
468 	u64 granule = kvm_granule_size(level), phys = data->phys;
469 
470 	if (!kvm_block_mapping_supported(addr, end, phys, level))
471 		return false;
472 
473 	/*
474 	 * If the PTE was already valid, drop the refcount on the table
475 	 * early, as it will be bumped-up again in stage2_map_walk_leaf().
476 	 * This ensures that the refcount stays constant across a valid to
477 	 * valid PTE update.
478 	 */
479 	if (kvm_pte_valid(*ptep))
480 		put_page(virt_to_page(ptep));
481 
482 	if (kvm_set_valid_leaf_pte(ptep, phys, data->attr, level))
483 		goto out;
484 
485 	/* There's an existing valid leaf entry, so perform break-before-make */
486 	kvm_set_invalid_pte(ptep);
487 	kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, data->mmu, addr, level);
488 	kvm_set_valid_leaf_pte(ptep, phys, data->attr, level);
489 out:
490 	data->phys += granule;
491 	return true;
492 }
493 
494 static int stage2_map_walk_table_pre(u64 addr, u64 end, u32 level,
495 				     kvm_pte_t *ptep,
496 				     struct stage2_map_data *data)
497 {
498 	if (data->anchor)
499 		return 0;
500 
501 	if (!kvm_block_mapping_supported(addr, end, data->phys, level))
502 		return 0;
503 
504 	kvm_set_invalid_pte(ptep);
505 
506 	/*
507 	 * Invalidate the whole stage-2, as we may have numerous leaf
508 	 * entries below us which would otherwise need invalidating
509 	 * individually.
510 	 */
511 	kvm_call_hyp(__kvm_tlb_flush_vmid, data->mmu);
512 	data->anchor = ptep;
513 	return 0;
514 }
515 
516 static int stage2_map_walk_leaf(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
517 				struct stage2_map_data *data)
518 {
519 	kvm_pte_t *childp, pte = *ptep;
520 	struct page *page = virt_to_page(ptep);
521 
522 	if (data->anchor) {
523 		if (kvm_pte_valid(pte))
524 			put_page(page);
525 
526 		return 0;
527 	}
528 
529 	if (stage2_map_walker_try_leaf(addr, end, level, ptep, data))
530 		goto out_get_page;
531 
532 	if (WARN_ON(level == KVM_PGTABLE_MAX_LEVELS - 1))
533 		return -EINVAL;
534 
535 	if (!data->memcache)
536 		return -ENOMEM;
537 
538 	childp = kvm_mmu_memory_cache_alloc(data->memcache);
539 	if (!childp)
540 		return -ENOMEM;
541 
542 	/*
543 	 * If we've run into an existing block mapping then replace it with
544 	 * a table. Accesses beyond 'end' that fall within the new table
545 	 * will be mapped lazily.
546 	 */
547 	if (kvm_pte_valid(pte)) {
548 		kvm_set_invalid_pte(ptep);
549 		kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, data->mmu, addr, level);
550 		put_page(page);
551 	}
552 
553 	kvm_set_table_pte(ptep, childp);
554 
555 out_get_page:
556 	get_page(page);
557 	return 0;
558 }
559 
560 static int stage2_map_walk_table_post(u64 addr, u64 end, u32 level,
561 				      kvm_pte_t *ptep,
562 				      struct stage2_map_data *data)
563 {
564 	int ret = 0;
565 
566 	if (!data->anchor)
567 		return 0;
568 
569 	free_page((unsigned long)kvm_pte_follow(*ptep));
570 	put_page(virt_to_page(ptep));
571 
572 	if (data->anchor == ptep) {
573 		data->anchor = NULL;
574 		ret = stage2_map_walk_leaf(addr, end, level, ptep, data);
575 	}
576 
577 	return ret;
578 }
579 
580 /*
581  * This is a little fiddly, as we use all three of the walk flags. The idea
582  * is that the TABLE_PRE callback runs for table entries on the way down,
583  * looking for table entries which we could conceivably replace with a
584  * block entry for this mapping. If it finds one, then it sets the 'anchor'
585  * field in 'struct stage2_map_data' to point at the table entry, before
586  * clearing the entry to zero and descending into the now detached table.
587  *
588  * The behaviour of the LEAF callback then depends on whether or not the
589  * anchor has been set. If not, then we're not using a block mapping higher
590  * up the table and we perform the mapping at the existing leaves instead.
591  * If, on the other hand, the anchor _is_ set, then we drop references to
592  * all valid leaves so that the pages beneath the anchor can be freed.
593  *
594  * Finally, the TABLE_POST callback does nothing if the anchor has not
595  * been set, but otherwise frees the page-table pages while walking back up
596  * the page-table, installing the block entry when it revisits the anchor
597  * pointer and clearing the anchor to NULL.
598  */
599 static int stage2_map_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
600 			     enum kvm_pgtable_walk_flags flag, void * const arg)
601 {
602 	struct stage2_map_data *data = arg;
603 
604 	switch (flag) {
605 	case KVM_PGTABLE_WALK_TABLE_PRE:
606 		return stage2_map_walk_table_pre(addr, end, level, ptep, data);
607 	case KVM_PGTABLE_WALK_LEAF:
608 		return stage2_map_walk_leaf(addr, end, level, ptep, data);
609 	case KVM_PGTABLE_WALK_TABLE_POST:
610 		return stage2_map_walk_table_post(addr, end, level, ptep, data);
611 	}
612 
613 	return -EINVAL;
614 }
615 
616 int kvm_pgtable_stage2_map(struct kvm_pgtable *pgt, u64 addr, u64 size,
617 			   u64 phys, enum kvm_pgtable_prot prot,
618 			   struct kvm_mmu_memory_cache *mc)
619 {
620 	int ret;
621 	struct stage2_map_data map_data = {
622 		.phys		= ALIGN_DOWN(phys, PAGE_SIZE),
623 		.mmu		= pgt->mmu,
624 		.memcache	= mc,
625 	};
626 	struct kvm_pgtable_walker walker = {
627 		.cb		= stage2_map_walker,
628 		.flags		= KVM_PGTABLE_WALK_TABLE_PRE |
629 				  KVM_PGTABLE_WALK_LEAF |
630 				  KVM_PGTABLE_WALK_TABLE_POST,
631 		.arg		= &map_data,
632 	};
633 
634 	ret = stage2_map_set_prot_attr(prot, &map_data);
635 	if (ret)
636 		return ret;
637 
638 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
639 	dsb(ishst);
640 	return ret;
641 }
642 
643 static void stage2_flush_dcache(void *addr, u64 size)
644 {
645 	if (cpus_have_const_cap(ARM64_HAS_STAGE2_FWB))
646 		return;
647 
648 	__flush_dcache_area(addr, size);
649 }
650 
651 static bool stage2_pte_cacheable(kvm_pte_t pte)
652 {
653 	u64 memattr = pte & KVM_PTE_LEAF_ATTR_LO_S2_MEMATTR;
654 	return memattr == PAGE_S2_MEMATTR(NORMAL);
655 }
656 
657 static int stage2_unmap_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
658 			       enum kvm_pgtable_walk_flags flag,
659 			       void * const arg)
660 {
661 	struct kvm_s2_mmu *mmu = arg;
662 	kvm_pte_t pte = *ptep, *childp = NULL;
663 	bool need_flush = false;
664 
665 	if (!kvm_pte_valid(pte))
666 		return 0;
667 
668 	if (kvm_pte_table(pte, level)) {
669 		childp = kvm_pte_follow(pte);
670 
671 		if (page_count(virt_to_page(childp)) != 1)
672 			return 0;
673 	} else if (stage2_pte_cacheable(pte)) {
674 		need_flush = true;
675 	}
676 
677 	/*
678 	 * This is similar to the map() path in that we unmap the entire
679 	 * block entry and rely on the remaining portions being faulted
680 	 * back lazily.
681 	 */
682 	kvm_set_invalid_pte(ptep);
683 	kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, mmu, addr, level);
684 	put_page(virt_to_page(ptep));
685 
686 	if (need_flush) {
687 		stage2_flush_dcache(kvm_pte_follow(pte),
688 				    kvm_granule_size(level));
689 	}
690 
691 	if (childp)
692 		free_page((unsigned long)childp);
693 
694 	return 0;
695 }
696 
697 int kvm_pgtable_stage2_unmap(struct kvm_pgtable *pgt, u64 addr, u64 size)
698 {
699 	struct kvm_pgtable_walker walker = {
700 		.cb	= stage2_unmap_walker,
701 		.arg	= pgt->mmu,
702 		.flags	= KVM_PGTABLE_WALK_LEAF | KVM_PGTABLE_WALK_TABLE_POST,
703 	};
704 
705 	return kvm_pgtable_walk(pgt, addr, size, &walker);
706 }
707 
708 struct stage2_attr_data {
709 	kvm_pte_t	attr_set;
710 	kvm_pte_t	attr_clr;
711 	kvm_pte_t	pte;
712 	u32		level;
713 };
714 
715 static int stage2_attr_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
716 			      enum kvm_pgtable_walk_flags flag,
717 			      void * const arg)
718 {
719 	kvm_pte_t pte = *ptep;
720 	struct stage2_attr_data *data = arg;
721 
722 	if (!kvm_pte_valid(pte))
723 		return 0;
724 
725 	data->level = level;
726 	data->pte = pte;
727 	pte &= ~data->attr_clr;
728 	pte |= data->attr_set;
729 
730 	/*
731 	 * We may race with the CPU trying to set the access flag here,
732 	 * but worst-case the access flag update gets lost and will be
733 	 * set on the next access instead.
734 	 */
735 	if (data->pte != pte)
736 		WRITE_ONCE(*ptep, pte);
737 
738 	return 0;
739 }
740 
741 static int stage2_update_leaf_attrs(struct kvm_pgtable *pgt, u64 addr,
742 				    u64 size, kvm_pte_t attr_set,
743 				    kvm_pte_t attr_clr, kvm_pte_t *orig_pte,
744 				    u32 *level)
745 {
746 	int ret;
747 	kvm_pte_t attr_mask = KVM_PTE_LEAF_ATTR_LO | KVM_PTE_LEAF_ATTR_HI;
748 	struct stage2_attr_data data = {
749 		.attr_set	= attr_set & attr_mask,
750 		.attr_clr	= attr_clr & attr_mask,
751 	};
752 	struct kvm_pgtable_walker walker = {
753 		.cb		= stage2_attr_walker,
754 		.arg		= &data,
755 		.flags		= KVM_PGTABLE_WALK_LEAF,
756 	};
757 
758 	ret = kvm_pgtable_walk(pgt, addr, size, &walker);
759 	if (ret)
760 		return ret;
761 
762 	if (orig_pte)
763 		*orig_pte = data.pte;
764 
765 	if (level)
766 		*level = data.level;
767 	return 0;
768 }
769 
770 int kvm_pgtable_stage2_wrprotect(struct kvm_pgtable *pgt, u64 addr, u64 size)
771 {
772 	return stage2_update_leaf_attrs(pgt, addr, size, 0,
773 					KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W,
774 					NULL, NULL);
775 }
776 
777 kvm_pte_t kvm_pgtable_stage2_mkyoung(struct kvm_pgtable *pgt, u64 addr)
778 {
779 	kvm_pte_t pte = 0;
780 	stage2_update_leaf_attrs(pgt, addr, 1, KVM_PTE_LEAF_ATTR_LO_S2_AF, 0,
781 				 &pte, NULL);
782 	dsb(ishst);
783 	return pte;
784 }
785 
786 kvm_pte_t kvm_pgtable_stage2_mkold(struct kvm_pgtable *pgt, u64 addr)
787 {
788 	kvm_pte_t pte = 0;
789 	stage2_update_leaf_attrs(pgt, addr, 1, 0, KVM_PTE_LEAF_ATTR_LO_S2_AF,
790 				 &pte, NULL);
791 	/*
792 	 * "But where's the TLBI?!", you scream.
793 	 * "Over in the core code", I sigh.
794 	 *
795 	 * See the '->clear_flush_young()' callback on the KVM mmu notifier.
796 	 */
797 	return pte;
798 }
799 
800 bool kvm_pgtable_stage2_is_young(struct kvm_pgtable *pgt, u64 addr)
801 {
802 	kvm_pte_t pte = 0;
803 	stage2_update_leaf_attrs(pgt, addr, 1, 0, 0, &pte, NULL);
804 	return pte & KVM_PTE_LEAF_ATTR_LO_S2_AF;
805 }
806 
807 int kvm_pgtable_stage2_relax_perms(struct kvm_pgtable *pgt, u64 addr,
808 				   enum kvm_pgtable_prot prot)
809 {
810 	int ret;
811 	u32 level;
812 	kvm_pte_t set = 0, clr = 0;
813 
814 	if (prot & KVM_PGTABLE_PROT_R)
815 		set |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_R;
816 
817 	if (prot & KVM_PGTABLE_PROT_W)
818 		set |= KVM_PTE_LEAF_ATTR_LO_S2_S2AP_W;
819 
820 	if (prot & KVM_PGTABLE_PROT_X)
821 		clr |= KVM_PTE_LEAF_ATTR_HI_S2_XN;
822 
823 	ret = stage2_update_leaf_attrs(pgt, addr, 1, set, clr, NULL, &level);
824 	if (!ret)
825 		kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, pgt->mmu, addr, level);
826 	return ret;
827 }
828 
829 static int stage2_flush_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
830 			       enum kvm_pgtable_walk_flags flag,
831 			       void * const arg)
832 {
833 	kvm_pte_t pte = *ptep;
834 
835 	if (!kvm_pte_valid(pte) || !stage2_pte_cacheable(pte))
836 		return 0;
837 
838 	stage2_flush_dcache(kvm_pte_follow(pte), kvm_granule_size(level));
839 	return 0;
840 }
841 
842 int kvm_pgtable_stage2_flush(struct kvm_pgtable *pgt, u64 addr, u64 size)
843 {
844 	struct kvm_pgtable_walker walker = {
845 		.cb	= stage2_flush_walker,
846 		.flags	= KVM_PGTABLE_WALK_LEAF,
847 	};
848 
849 	if (cpus_have_const_cap(ARM64_HAS_STAGE2_FWB))
850 		return 0;
851 
852 	return kvm_pgtable_walk(pgt, addr, size, &walker);
853 }
854 
855 int kvm_pgtable_stage2_init(struct kvm_pgtable *pgt, struct kvm *kvm)
856 {
857 	size_t pgd_sz;
858 	u64 vtcr = kvm->arch.vtcr;
859 	u32 ia_bits = VTCR_EL2_IPA(vtcr);
860 	u32 sl0 = FIELD_GET(VTCR_EL2_SL0_MASK, vtcr);
861 	u32 start_level = VTCR_EL2_TGRAN_SL0_BASE - sl0;
862 
863 	pgd_sz = kvm_pgd_pages(ia_bits, start_level) * PAGE_SIZE;
864 	pgt->pgd = alloc_pages_exact(pgd_sz, GFP_KERNEL_ACCOUNT | __GFP_ZERO);
865 	if (!pgt->pgd)
866 		return -ENOMEM;
867 
868 	pgt->ia_bits		= ia_bits;
869 	pgt->start_level	= start_level;
870 	pgt->mmu		= &kvm->arch.mmu;
871 
872 	/* Ensure zeroed PGD pages are visible to the hardware walker */
873 	dsb(ishst);
874 	return 0;
875 }
876 
877 static int stage2_free_walker(u64 addr, u64 end, u32 level, kvm_pte_t *ptep,
878 			      enum kvm_pgtable_walk_flags flag,
879 			      void * const arg)
880 {
881 	kvm_pte_t pte = *ptep;
882 
883 	if (!kvm_pte_valid(pte))
884 		return 0;
885 
886 	put_page(virt_to_page(ptep));
887 
888 	if (kvm_pte_table(pte, level))
889 		free_page((unsigned long)kvm_pte_follow(pte));
890 
891 	return 0;
892 }
893 
894 void kvm_pgtable_stage2_destroy(struct kvm_pgtable *pgt)
895 {
896 	size_t pgd_sz;
897 	struct kvm_pgtable_walker walker = {
898 		.cb	= stage2_free_walker,
899 		.flags	= KVM_PGTABLE_WALK_LEAF |
900 			  KVM_PGTABLE_WALK_TABLE_POST,
901 	};
902 
903 	WARN_ON(kvm_pgtable_walk(pgt, 0, BIT(pgt->ia_bits), &walker));
904 	pgd_sz = kvm_pgd_pages(pgt->ia_bits, pgt->start_level) * PAGE_SIZE;
905 	free_pages_exact(pgt->pgd, pgd_sz);
906 	pgt->pgd = NULL;
907 }
908