xref: /linux/drivers/iommu/io-pgtable-arm.c (revision 34f7c6e7d4396090692a09789db231e12cb4762b)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * CPU-agnostic ARM page table allocator.
4  *
5  * Copyright (C) 2014 ARM Limited
6  *
7  * Author: Will Deacon <will.deacon@arm.com>
8  */
9 
10 #define pr_fmt(fmt)	"arm-lpae io-pgtable: " fmt
11 
12 #include <linux/atomic.h>
13 #include <linux/bitops.h>
14 #include <linux/io-pgtable.h>
15 #include <linux/kernel.h>
16 #include <linux/sizes.h>
17 #include <linux/slab.h>
18 #include <linux/types.h>
19 #include <linux/dma-mapping.h>
20 
21 #include <asm/barrier.h>
22 
23 #include "io-pgtable-arm.h"
24 
25 #define ARM_LPAE_MAX_ADDR_BITS		52
26 #define ARM_LPAE_S2_MAX_CONCAT_PAGES	16
27 #define ARM_LPAE_MAX_LEVELS		4
28 
29 /* Struct accessors */
30 #define io_pgtable_to_data(x)						\
31 	container_of((x), struct arm_lpae_io_pgtable, iop)
32 
33 #define io_pgtable_ops_to_data(x)					\
34 	io_pgtable_to_data(io_pgtable_ops_to_pgtable(x))
35 
36 /*
37  * Calculate the right shift amount to get to the portion describing level l
38  * in a virtual address mapped by the pagetable in d.
39  */
40 #define ARM_LPAE_LVL_SHIFT(l,d)						\
41 	(((ARM_LPAE_MAX_LEVELS - (l)) * (d)->bits_per_level) +		\
42 	ilog2(sizeof(arm_lpae_iopte)))
43 
44 #define ARM_LPAE_GRANULE(d)						\
45 	(sizeof(arm_lpae_iopte) << (d)->bits_per_level)
46 #define ARM_LPAE_PGD_SIZE(d)						\
47 	(sizeof(arm_lpae_iopte) << (d)->pgd_bits)
48 
49 #define ARM_LPAE_PTES_PER_TABLE(d)					\
50 	(ARM_LPAE_GRANULE(d) >> ilog2(sizeof(arm_lpae_iopte)))
51 
52 /*
53  * Calculate the index at level l used to map virtual address a using the
54  * pagetable in d.
55  */
56 #define ARM_LPAE_PGD_IDX(l,d)						\
57 	((l) == (d)->start_level ? (d)->pgd_bits - (d)->bits_per_level : 0)
58 
59 #define ARM_LPAE_LVL_IDX(a,l,d)						\
60 	(((u64)(a) >> ARM_LPAE_LVL_SHIFT(l,d)) &			\
61 	 ((1 << ((d)->bits_per_level + ARM_LPAE_PGD_IDX(l,d))) - 1))
62 
63 /* Calculate the block/page mapping size at level l for pagetable in d. */
64 #define ARM_LPAE_BLOCK_SIZE(l,d)	(1ULL << ARM_LPAE_LVL_SHIFT(l,d))
65 
66 /* Page table bits */
67 #define ARM_LPAE_PTE_TYPE_SHIFT		0
68 #define ARM_LPAE_PTE_TYPE_MASK		0x3
69 
70 #define ARM_LPAE_PTE_TYPE_BLOCK		1
71 #define ARM_LPAE_PTE_TYPE_TABLE		3
72 #define ARM_LPAE_PTE_TYPE_PAGE		3
73 
74 #define ARM_LPAE_PTE_ADDR_MASK		GENMASK_ULL(47,12)
75 
76 #define ARM_LPAE_PTE_NSTABLE		(((arm_lpae_iopte)1) << 63)
77 #define ARM_LPAE_PTE_XN			(((arm_lpae_iopte)3) << 53)
78 #define ARM_LPAE_PTE_AF			(((arm_lpae_iopte)1) << 10)
79 #define ARM_LPAE_PTE_SH_NS		(((arm_lpae_iopte)0) << 8)
80 #define ARM_LPAE_PTE_SH_OS		(((arm_lpae_iopte)2) << 8)
81 #define ARM_LPAE_PTE_SH_IS		(((arm_lpae_iopte)3) << 8)
82 #define ARM_LPAE_PTE_NS			(((arm_lpae_iopte)1) << 5)
83 #define ARM_LPAE_PTE_VALID		(((arm_lpae_iopte)1) << 0)
84 
85 #define ARM_LPAE_PTE_ATTR_LO_MASK	(((arm_lpae_iopte)0x3ff) << 2)
86 /* Ignore the contiguous bit for block splitting */
87 #define ARM_LPAE_PTE_ATTR_HI_MASK	(((arm_lpae_iopte)6) << 52)
88 #define ARM_LPAE_PTE_ATTR_MASK		(ARM_LPAE_PTE_ATTR_LO_MASK |	\
89 					 ARM_LPAE_PTE_ATTR_HI_MASK)
90 /* Software bit for solving coherency races */
91 #define ARM_LPAE_PTE_SW_SYNC		(((arm_lpae_iopte)1) << 55)
92 
93 /* Stage-1 PTE */
94 #define ARM_LPAE_PTE_AP_UNPRIV		(((arm_lpae_iopte)1) << 6)
95 #define ARM_LPAE_PTE_AP_RDONLY		(((arm_lpae_iopte)2) << 6)
96 #define ARM_LPAE_PTE_ATTRINDX_SHIFT	2
97 #define ARM_LPAE_PTE_nG			(((arm_lpae_iopte)1) << 11)
98 
99 /* Stage-2 PTE */
100 #define ARM_LPAE_PTE_HAP_FAULT		(((arm_lpae_iopte)0) << 6)
101 #define ARM_LPAE_PTE_HAP_READ		(((arm_lpae_iopte)1) << 6)
102 #define ARM_LPAE_PTE_HAP_WRITE		(((arm_lpae_iopte)2) << 6)
103 #define ARM_LPAE_PTE_MEMATTR_OIWB	(((arm_lpae_iopte)0xf) << 2)
104 #define ARM_LPAE_PTE_MEMATTR_NC		(((arm_lpae_iopte)0x5) << 2)
105 #define ARM_LPAE_PTE_MEMATTR_DEV	(((arm_lpae_iopte)0x1) << 2)
106 
107 /* Register bits */
108 #define ARM_LPAE_VTCR_SL0_MASK		0x3
109 
110 #define ARM_LPAE_TCR_T0SZ_SHIFT		0
111 
112 #define ARM_LPAE_VTCR_PS_SHIFT		16
113 #define ARM_LPAE_VTCR_PS_MASK		0x7
114 
115 #define ARM_LPAE_MAIR_ATTR_SHIFT(n)	((n) << 3)
116 #define ARM_LPAE_MAIR_ATTR_MASK		0xff
117 #define ARM_LPAE_MAIR_ATTR_DEVICE	0x04
118 #define ARM_LPAE_MAIR_ATTR_NC		0x44
119 #define ARM_LPAE_MAIR_ATTR_INC_OWBRWA	0xf4
120 #define ARM_LPAE_MAIR_ATTR_WBRWA	0xff
121 #define ARM_LPAE_MAIR_ATTR_IDX_NC	0
122 #define ARM_LPAE_MAIR_ATTR_IDX_CACHE	1
123 #define ARM_LPAE_MAIR_ATTR_IDX_DEV	2
124 #define ARM_LPAE_MAIR_ATTR_IDX_INC_OCACHE	3
125 
126 #define ARM_MALI_LPAE_TTBR_ADRMODE_TABLE (3u << 0)
127 #define ARM_MALI_LPAE_TTBR_READ_INNER	BIT(2)
128 #define ARM_MALI_LPAE_TTBR_SHARE_OUTER	BIT(4)
129 
130 #define ARM_MALI_LPAE_MEMATTR_IMP_DEF	0x88ULL
131 #define ARM_MALI_LPAE_MEMATTR_WRITE_ALLOC 0x8DULL
132 
133 #define APPLE_DART_PTE_PROT_NO_WRITE (1<<7)
134 #define APPLE_DART_PTE_PROT_NO_READ (1<<8)
135 
136 /* IOPTE accessors */
137 #define iopte_deref(pte,d) __va(iopte_to_paddr(pte, d))
138 
139 #define iopte_type(pte)					\
140 	(((pte) >> ARM_LPAE_PTE_TYPE_SHIFT) & ARM_LPAE_PTE_TYPE_MASK)
141 
142 #define iopte_prot(pte)	((pte) & ARM_LPAE_PTE_ATTR_MASK)
143 
144 struct arm_lpae_io_pgtable {
145 	struct io_pgtable	iop;
146 
147 	int			pgd_bits;
148 	int			start_level;
149 	int			bits_per_level;
150 
151 	void			*pgd;
152 };
153 
154 typedef u64 arm_lpae_iopte;
155 
156 static inline bool iopte_leaf(arm_lpae_iopte pte, int lvl,
157 			      enum io_pgtable_fmt fmt)
158 {
159 	if (lvl == (ARM_LPAE_MAX_LEVELS - 1) && fmt != ARM_MALI_LPAE)
160 		return iopte_type(pte) == ARM_LPAE_PTE_TYPE_PAGE;
161 
162 	return iopte_type(pte) == ARM_LPAE_PTE_TYPE_BLOCK;
163 }
164 
165 static arm_lpae_iopte paddr_to_iopte(phys_addr_t paddr,
166 				     struct arm_lpae_io_pgtable *data)
167 {
168 	arm_lpae_iopte pte = paddr;
169 
170 	/* Of the bits which overlap, either 51:48 or 15:12 are always RES0 */
171 	return (pte | (pte >> (48 - 12))) & ARM_LPAE_PTE_ADDR_MASK;
172 }
173 
174 static phys_addr_t iopte_to_paddr(arm_lpae_iopte pte,
175 				  struct arm_lpae_io_pgtable *data)
176 {
177 	u64 paddr = pte & ARM_LPAE_PTE_ADDR_MASK;
178 
179 	if (ARM_LPAE_GRANULE(data) < SZ_64K)
180 		return paddr;
181 
182 	/* Rotate the packed high-order bits back to the top */
183 	return (paddr | (paddr << (48 - 12))) & (ARM_LPAE_PTE_ADDR_MASK << 4);
184 }
185 
186 static bool selftest_running = false;
187 
188 static dma_addr_t __arm_lpae_dma_addr(void *pages)
189 {
190 	return (dma_addr_t)virt_to_phys(pages);
191 }
192 
193 static void *__arm_lpae_alloc_pages(size_t size, gfp_t gfp,
194 				    struct io_pgtable_cfg *cfg)
195 {
196 	struct device *dev = cfg->iommu_dev;
197 	int order = get_order(size);
198 	struct page *p;
199 	dma_addr_t dma;
200 	void *pages;
201 
202 	VM_BUG_ON((gfp & __GFP_HIGHMEM));
203 	p = alloc_pages_node(dev ? dev_to_node(dev) : NUMA_NO_NODE,
204 			     gfp | __GFP_ZERO, order);
205 	if (!p)
206 		return NULL;
207 
208 	pages = page_address(p);
209 	if (!cfg->coherent_walk) {
210 		dma = dma_map_single(dev, pages, size, DMA_TO_DEVICE);
211 		if (dma_mapping_error(dev, dma))
212 			goto out_free;
213 		/*
214 		 * We depend on the IOMMU being able to work with any physical
215 		 * address directly, so if the DMA layer suggests otherwise by
216 		 * translating or truncating them, that bodes very badly...
217 		 */
218 		if (dma != virt_to_phys(pages))
219 			goto out_unmap;
220 	}
221 
222 	return pages;
223 
224 out_unmap:
225 	dev_err(dev, "Cannot accommodate DMA translation for IOMMU page tables\n");
226 	dma_unmap_single(dev, dma, size, DMA_TO_DEVICE);
227 out_free:
228 	__free_pages(p, order);
229 	return NULL;
230 }
231 
232 static void __arm_lpae_free_pages(void *pages, size_t size,
233 				  struct io_pgtable_cfg *cfg)
234 {
235 	if (!cfg->coherent_walk)
236 		dma_unmap_single(cfg->iommu_dev, __arm_lpae_dma_addr(pages),
237 				 size, DMA_TO_DEVICE);
238 	free_pages((unsigned long)pages, get_order(size));
239 }
240 
241 static void __arm_lpae_sync_pte(arm_lpae_iopte *ptep, int num_entries,
242 				struct io_pgtable_cfg *cfg)
243 {
244 	dma_sync_single_for_device(cfg->iommu_dev, __arm_lpae_dma_addr(ptep),
245 				   sizeof(*ptep) * num_entries, DMA_TO_DEVICE);
246 }
247 
248 static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct io_pgtable_cfg *cfg)
249 {
250 
251 	*ptep = 0;
252 
253 	if (!cfg->coherent_walk)
254 		__arm_lpae_sync_pte(ptep, 1, cfg);
255 }
256 
257 static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
258 			       struct iommu_iotlb_gather *gather,
259 			       unsigned long iova, size_t size, size_t pgcount,
260 			       int lvl, arm_lpae_iopte *ptep);
261 
262 static void __arm_lpae_init_pte(struct arm_lpae_io_pgtable *data,
263 				phys_addr_t paddr, arm_lpae_iopte prot,
264 				int lvl, int num_entries, arm_lpae_iopte *ptep)
265 {
266 	arm_lpae_iopte pte = prot;
267 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
268 	size_t sz = ARM_LPAE_BLOCK_SIZE(lvl, data);
269 	int i;
270 
271 	if (data->iop.fmt != ARM_MALI_LPAE && lvl == ARM_LPAE_MAX_LEVELS - 1)
272 		pte |= ARM_LPAE_PTE_TYPE_PAGE;
273 	else
274 		pte |= ARM_LPAE_PTE_TYPE_BLOCK;
275 
276 	for (i = 0; i < num_entries; i++)
277 		ptep[i] = pte | paddr_to_iopte(paddr + i * sz, data);
278 
279 	if (!cfg->coherent_walk)
280 		__arm_lpae_sync_pte(ptep, num_entries, cfg);
281 }
282 
283 static int arm_lpae_init_pte(struct arm_lpae_io_pgtable *data,
284 			     unsigned long iova, phys_addr_t paddr,
285 			     arm_lpae_iopte prot, int lvl, int num_entries,
286 			     arm_lpae_iopte *ptep)
287 {
288 	int i;
289 
290 	for (i = 0; i < num_entries; i++)
291 		if (iopte_leaf(ptep[i], lvl, data->iop.fmt)) {
292 			/* We require an unmap first */
293 			WARN_ON(!selftest_running);
294 			return -EEXIST;
295 		} else if (iopte_type(ptep[i]) == ARM_LPAE_PTE_TYPE_TABLE) {
296 			/*
297 			 * We need to unmap and free the old table before
298 			 * overwriting it with a block entry.
299 			 */
300 			arm_lpae_iopte *tblp;
301 			size_t sz = ARM_LPAE_BLOCK_SIZE(lvl, data);
302 
303 			tblp = ptep - ARM_LPAE_LVL_IDX(iova, lvl, data);
304 			if (__arm_lpae_unmap(data, NULL, iova + i * sz, sz, 1,
305 					     lvl, tblp) != sz) {
306 				WARN_ON(1);
307 				return -EINVAL;
308 			}
309 		}
310 
311 	__arm_lpae_init_pte(data, paddr, prot, lvl, num_entries, ptep);
312 	return 0;
313 }
314 
315 static arm_lpae_iopte arm_lpae_install_table(arm_lpae_iopte *table,
316 					     arm_lpae_iopte *ptep,
317 					     arm_lpae_iopte curr,
318 					     struct arm_lpae_io_pgtable *data)
319 {
320 	arm_lpae_iopte old, new;
321 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
322 
323 	new = paddr_to_iopte(__pa(table), data) | ARM_LPAE_PTE_TYPE_TABLE;
324 	if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_NS)
325 		new |= ARM_LPAE_PTE_NSTABLE;
326 
327 	/*
328 	 * Ensure the table itself is visible before its PTE can be.
329 	 * Whilst we could get away with cmpxchg64_release below, this
330 	 * doesn't have any ordering semantics when !CONFIG_SMP.
331 	 */
332 	dma_wmb();
333 
334 	old = cmpxchg64_relaxed(ptep, curr, new);
335 
336 	if (cfg->coherent_walk || (old & ARM_LPAE_PTE_SW_SYNC))
337 		return old;
338 
339 	/* Even if it's not ours, there's no point waiting; just kick it */
340 	__arm_lpae_sync_pte(ptep, 1, cfg);
341 	if (old == curr)
342 		WRITE_ONCE(*ptep, new | ARM_LPAE_PTE_SW_SYNC);
343 
344 	return old;
345 }
346 
347 static int __arm_lpae_map(struct arm_lpae_io_pgtable *data, unsigned long iova,
348 			  phys_addr_t paddr, size_t size, size_t pgcount,
349 			  arm_lpae_iopte prot, int lvl, arm_lpae_iopte *ptep,
350 			  gfp_t gfp, size_t *mapped)
351 {
352 	arm_lpae_iopte *cptep, pte;
353 	size_t block_size = ARM_LPAE_BLOCK_SIZE(lvl, data);
354 	size_t tblsz = ARM_LPAE_GRANULE(data);
355 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
356 	int ret = 0, num_entries, max_entries, map_idx_start;
357 
358 	/* Find our entry at the current level */
359 	map_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
360 	ptep += map_idx_start;
361 
362 	/* If we can install a leaf entry at this level, then do so */
363 	if (size == block_size) {
364 		max_entries = ARM_LPAE_PTES_PER_TABLE(data) - map_idx_start;
365 		num_entries = min_t(int, pgcount, max_entries);
366 		ret = arm_lpae_init_pte(data, iova, paddr, prot, lvl, num_entries, ptep);
367 		if (!ret && mapped)
368 			*mapped += num_entries * size;
369 
370 		return ret;
371 	}
372 
373 	/* We can't allocate tables at the final level */
374 	if (WARN_ON(lvl >= ARM_LPAE_MAX_LEVELS - 1))
375 		return -EINVAL;
376 
377 	/* Grab a pointer to the next level */
378 	pte = READ_ONCE(*ptep);
379 	if (!pte) {
380 		cptep = __arm_lpae_alloc_pages(tblsz, gfp, cfg);
381 		if (!cptep)
382 			return -ENOMEM;
383 
384 		pte = arm_lpae_install_table(cptep, ptep, 0, data);
385 		if (pte)
386 			__arm_lpae_free_pages(cptep, tblsz, cfg);
387 	} else if (!cfg->coherent_walk && !(pte & ARM_LPAE_PTE_SW_SYNC)) {
388 		__arm_lpae_sync_pte(ptep, 1, cfg);
389 	}
390 
391 	if (pte && !iopte_leaf(pte, lvl, data->iop.fmt)) {
392 		cptep = iopte_deref(pte, data);
393 	} else if (pte) {
394 		/* We require an unmap first */
395 		WARN_ON(!selftest_running);
396 		return -EEXIST;
397 	}
398 
399 	/* Rinse, repeat */
400 	return __arm_lpae_map(data, iova, paddr, size, pgcount, prot, lvl + 1,
401 			      cptep, gfp, mapped);
402 }
403 
404 static arm_lpae_iopte arm_lpae_prot_to_pte(struct arm_lpae_io_pgtable *data,
405 					   int prot)
406 {
407 	arm_lpae_iopte pte;
408 
409 	if (data->iop.fmt == APPLE_DART) {
410 		pte = 0;
411 		if (!(prot & IOMMU_WRITE))
412 			pte |= APPLE_DART_PTE_PROT_NO_WRITE;
413 		if (!(prot & IOMMU_READ))
414 			pte |= APPLE_DART_PTE_PROT_NO_READ;
415 		return pte;
416 	}
417 
418 	if (data->iop.fmt == ARM_64_LPAE_S1 ||
419 	    data->iop.fmt == ARM_32_LPAE_S1) {
420 		pte = ARM_LPAE_PTE_nG;
421 		if (!(prot & IOMMU_WRITE) && (prot & IOMMU_READ))
422 			pte |= ARM_LPAE_PTE_AP_RDONLY;
423 		if (!(prot & IOMMU_PRIV))
424 			pte |= ARM_LPAE_PTE_AP_UNPRIV;
425 	} else {
426 		pte = ARM_LPAE_PTE_HAP_FAULT;
427 		if (prot & IOMMU_READ)
428 			pte |= ARM_LPAE_PTE_HAP_READ;
429 		if (prot & IOMMU_WRITE)
430 			pte |= ARM_LPAE_PTE_HAP_WRITE;
431 	}
432 
433 	/*
434 	 * Note that this logic is structured to accommodate Mali LPAE
435 	 * having stage-1-like attributes but stage-2-like permissions.
436 	 */
437 	if (data->iop.fmt == ARM_64_LPAE_S2 ||
438 	    data->iop.fmt == ARM_32_LPAE_S2) {
439 		if (prot & IOMMU_MMIO)
440 			pte |= ARM_LPAE_PTE_MEMATTR_DEV;
441 		else if (prot & IOMMU_CACHE)
442 			pte |= ARM_LPAE_PTE_MEMATTR_OIWB;
443 		else
444 			pte |= ARM_LPAE_PTE_MEMATTR_NC;
445 	} else {
446 		if (prot & IOMMU_MMIO)
447 			pte |= (ARM_LPAE_MAIR_ATTR_IDX_DEV
448 				<< ARM_LPAE_PTE_ATTRINDX_SHIFT);
449 		else if (prot & IOMMU_CACHE)
450 			pte |= (ARM_LPAE_MAIR_ATTR_IDX_CACHE
451 				<< ARM_LPAE_PTE_ATTRINDX_SHIFT);
452 	}
453 
454 	/*
455 	 * Also Mali has its own notions of shareability wherein its Inner
456 	 * domain covers the cores within the GPU, and its Outer domain is
457 	 * "outside the GPU" (i.e. either the Inner or System domain in CPU
458 	 * terms, depending on coherency).
459 	 */
460 	if (prot & IOMMU_CACHE && data->iop.fmt != ARM_MALI_LPAE)
461 		pte |= ARM_LPAE_PTE_SH_IS;
462 	else
463 		pte |= ARM_LPAE_PTE_SH_OS;
464 
465 	if (prot & IOMMU_NOEXEC)
466 		pte |= ARM_LPAE_PTE_XN;
467 
468 	if (data->iop.cfg.quirks & IO_PGTABLE_QUIRK_ARM_NS)
469 		pte |= ARM_LPAE_PTE_NS;
470 
471 	if (data->iop.fmt != ARM_MALI_LPAE)
472 		pte |= ARM_LPAE_PTE_AF;
473 
474 	return pte;
475 }
476 
477 static int arm_lpae_map_pages(struct io_pgtable_ops *ops, unsigned long iova,
478 			      phys_addr_t paddr, size_t pgsize, size_t pgcount,
479 			      int iommu_prot, gfp_t gfp, size_t *mapped)
480 {
481 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
482 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
483 	arm_lpae_iopte *ptep = data->pgd;
484 	int ret, lvl = data->start_level;
485 	arm_lpae_iopte prot;
486 	long iaext = (s64)iova >> cfg->ias;
487 
488 	if (WARN_ON(!pgsize || (pgsize & cfg->pgsize_bitmap) != pgsize))
489 		return -EINVAL;
490 
491 	if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_TTBR1)
492 		iaext = ~iaext;
493 	if (WARN_ON(iaext || paddr >> cfg->oas))
494 		return -ERANGE;
495 
496 	/* If no access, then nothing to do */
497 	if (!(iommu_prot & (IOMMU_READ | IOMMU_WRITE)))
498 		return 0;
499 
500 	prot = arm_lpae_prot_to_pte(data, iommu_prot);
501 	ret = __arm_lpae_map(data, iova, paddr, pgsize, pgcount, prot, lvl,
502 			     ptep, gfp, mapped);
503 	/*
504 	 * Synchronise all PTE updates for the new mapping before there's
505 	 * a chance for anything to kick off a table walk for the new iova.
506 	 */
507 	wmb();
508 
509 	return ret;
510 }
511 
512 static int arm_lpae_map(struct io_pgtable_ops *ops, unsigned long iova,
513 			phys_addr_t paddr, size_t size, int iommu_prot, gfp_t gfp)
514 {
515 	return arm_lpae_map_pages(ops, iova, paddr, size, 1, iommu_prot, gfp,
516 				  NULL);
517 }
518 
519 static void __arm_lpae_free_pgtable(struct arm_lpae_io_pgtable *data, int lvl,
520 				    arm_lpae_iopte *ptep)
521 {
522 	arm_lpae_iopte *start, *end;
523 	unsigned long table_size;
524 
525 	if (lvl == data->start_level)
526 		table_size = ARM_LPAE_PGD_SIZE(data);
527 	else
528 		table_size = ARM_LPAE_GRANULE(data);
529 
530 	start = ptep;
531 
532 	/* Only leaf entries at the last level */
533 	if (lvl == ARM_LPAE_MAX_LEVELS - 1)
534 		end = ptep;
535 	else
536 		end = (void *)ptep + table_size;
537 
538 	while (ptep != end) {
539 		arm_lpae_iopte pte = *ptep++;
540 
541 		if (!pte || iopte_leaf(pte, lvl, data->iop.fmt))
542 			continue;
543 
544 		__arm_lpae_free_pgtable(data, lvl + 1, iopte_deref(pte, data));
545 	}
546 
547 	__arm_lpae_free_pages(start, table_size, &data->iop.cfg);
548 }
549 
550 static void arm_lpae_free_pgtable(struct io_pgtable *iop)
551 {
552 	struct arm_lpae_io_pgtable *data = io_pgtable_to_data(iop);
553 
554 	__arm_lpae_free_pgtable(data, data->start_level, data->pgd);
555 	kfree(data);
556 }
557 
558 static size_t arm_lpae_split_blk_unmap(struct arm_lpae_io_pgtable *data,
559 				       struct iommu_iotlb_gather *gather,
560 				       unsigned long iova, size_t size,
561 				       arm_lpae_iopte blk_pte, int lvl,
562 				       arm_lpae_iopte *ptep, size_t pgcount)
563 {
564 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
565 	arm_lpae_iopte pte, *tablep;
566 	phys_addr_t blk_paddr;
567 	size_t tablesz = ARM_LPAE_GRANULE(data);
568 	size_t split_sz = ARM_LPAE_BLOCK_SIZE(lvl, data);
569 	int ptes_per_table = ARM_LPAE_PTES_PER_TABLE(data);
570 	int i, unmap_idx_start = -1, num_entries = 0, max_entries;
571 
572 	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
573 		return 0;
574 
575 	tablep = __arm_lpae_alloc_pages(tablesz, GFP_ATOMIC, cfg);
576 	if (!tablep)
577 		return 0; /* Bytes unmapped */
578 
579 	if (size == split_sz) {
580 		unmap_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
581 		max_entries = ptes_per_table - unmap_idx_start;
582 		num_entries = min_t(int, pgcount, max_entries);
583 	}
584 
585 	blk_paddr = iopte_to_paddr(blk_pte, data);
586 	pte = iopte_prot(blk_pte);
587 
588 	for (i = 0; i < ptes_per_table; i++, blk_paddr += split_sz) {
589 		/* Unmap! */
590 		if (i >= unmap_idx_start && i < (unmap_idx_start + num_entries))
591 			continue;
592 
593 		__arm_lpae_init_pte(data, blk_paddr, pte, lvl, 1, &tablep[i]);
594 	}
595 
596 	pte = arm_lpae_install_table(tablep, ptep, blk_pte, data);
597 	if (pte != blk_pte) {
598 		__arm_lpae_free_pages(tablep, tablesz, cfg);
599 		/*
600 		 * We may race against someone unmapping another part of this
601 		 * block, but anything else is invalid. We can't misinterpret
602 		 * a page entry here since we're never at the last level.
603 		 */
604 		if (iopte_type(pte) != ARM_LPAE_PTE_TYPE_TABLE)
605 			return 0;
606 
607 		tablep = iopte_deref(pte, data);
608 	} else if (unmap_idx_start >= 0) {
609 		for (i = 0; i < num_entries; i++)
610 			io_pgtable_tlb_add_page(&data->iop, gather, iova + i * size, size);
611 
612 		return num_entries * size;
613 	}
614 
615 	return __arm_lpae_unmap(data, gather, iova, size, pgcount, lvl, tablep);
616 }
617 
618 static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
619 			       struct iommu_iotlb_gather *gather,
620 			       unsigned long iova, size_t size, size_t pgcount,
621 			       int lvl, arm_lpae_iopte *ptep)
622 {
623 	arm_lpae_iopte pte;
624 	struct io_pgtable *iop = &data->iop;
625 	int i = 0, num_entries, max_entries, unmap_idx_start;
626 
627 	/* Something went horribly wrong and we ran out of page table */
628 	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
629 		return 0;
630 
631 	unmap_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
632 	ptep += unmap_idx_start;
633 	pte = READ_ONCE(*ptep);
634 	if (WARN_ON(!pte))
635 		return 0;
636 
637 	/* If the size matches this level, we're in the right place */
638 	if (size == ARM_LPAE_BLOCK_SIZE(lvl, data)) {
639 		max_entries = ARM_LPAE_PTES_PER_TABLE(data) - unmap_idx_start;
640 		num_entries = min_t(int, pgcount, max_entries);
641 
642 		while (i < num_entries) {
643 			pte = READ_ONCE(*ptep);
644 			if (WARN_ON(!pte))
645 				break;
646 
647 			__arm_lpae_clear_pte(ptep, &iop->cfg);
648 
649 			if (!iopte_leaf(pte, lvl, iop->fmt)) {
650 				/* Also flush any partial walks */
651 				io_pgtable_tlb_flush_walk(iop, iova + i * size, size,
652 							  ARM_LPAE_GRANULE(data));
653 				__arm_lpae_free_pgtable(data, lvl + 1, iopte_deref(pte, data));
654 			} else if (!iommu_iotlb_gather_queued(gather)) {
655 				io_pgtable_tlb_add_page(iop, gather, iova + i * size, size);
656 			}
657 
658 			ptep++;
659 			i++;
660 		}
661 
662 		return i * size;
663 	} else if (iopte_leaf(pte, lvl, iop->fmt)) {
664 		/*
665 		 * Insert a table at the next level to map the old region,
666 		 * minus the part we want to unmap
667 		 */
668 		return arm_lpae_split_blk_unmap(data, gather, iova, size, pte,
669 						lvl + 1, ptep, pgcount);
670 	}
671 
672 	/* Keep on walkin' */
673 	ptep = iopte_deref(pte, data);
674 	return __arm_lpae_unmap(data, gather, iova, size, pgcount, lvl + 1, ptep);
675 }
676 
677 static size_t arm_lpae_unmap_pages(struct io_pgtable_ops *ops, unsigned long iova,
678 				   size_t pgsize, size_t pgcount,
679 				   struct iommu_iotlb_gather *gather)
680 {
681 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
682 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
683 	arm_lpae_iopte *ptep = data->pgd;
684 	long iaext = (s64)iova >> cfg->ias;
685 
686 	if (WARN_ON(!pgsize || (pgsize & cfg->pgsize_bitmap) != pgsize || !pgcount))
687 		return 0;
688 
689 	if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_TTBR1)
690 		iaext = ~iaext;
691 	if (WARN_ON(iaext))
692 		return 0;
693 
694 	return __arm_lpae_unmap(data, gather, iova, pgsize, pgcount,
695 				data->start_level, ptep);
696 }
697 
698 static size_t arm_lpae_unmap(struct io_pgtable_ops *ops, unsigned long iova,
699 			     size_t size, struct iommu_iotlb_gather *gather)
700 {
701 	return arm_lpae_unmap_pages(ops, iova, size, 1, gather);
702 }
703 
704 static phys_addr_t arm_lpae_iova_to_phys(struct io_pgtable_ops *ops,
705 					 unsigned long iova)
706 {
707 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
708 	arm_lpae_iopte pte, *ptep = data->pgd;
709 	int lvl = data->start_level;
710 
711 	do {
712 		/* Valid IOPTE pointer? */
713 		if (!ptep)
714 			return 0;
715 
716 		/* Grab the IOPTE we're interested in */
717 		ptep += ARM_LPAE_LVL_IDX(iova, lvl, data);
718 		pte = READ_ONCE(*ptep);
719 
720 		/* Valid entry? */
721 		if (!pte)
722 			return 0;
723 
724 		/* Leaf entry? */
725 		if (iopte_leaf(pte, lvl, data->iop.fmt))
726 			goto found_translation;
727 
728 		/* Take it to the next level */
729 		ptep = iopte_deref(pte, data);
730 	} while (++lvl < ARM_LPAE_MAX_LEVELS);
731 
732 	/* Ran out of page tables to walk */
733 	return 0;
734 
735 found_translation:
736 	iova &= (ARM_LPAE_BLOCK_SIZE(lvl, data) - 1);
737 	return iopte_to_paddr(pte, data) | iova;
738 }
739 
740 static void arm_lpae_restrict_pgsizes(struct io_pgtable_cfg *cfg)
741 {
742 	unsigned long granule, page_sizes;
743 	unsigned int max_addr_bits = 48;
744 
745 	/*
746 	 * We need to restrict the supported page sizes to match the
747 	 * translation regime for a particular granule. Aim to match
748 	 * the CPU page size if possible, otherwise prefer smaller sizes.
749 	 * While we're at it, restrict the block sizes to match the
750 	 * chosen granule.
751 	 */
752 	if (cfg->pgsize_bitmap & PAGE_SIZE)
753 		granule = PAGE_SIZE;
754 	else if (cfg->pgsize_bitmap & ~PAGE_MASK)
755 		granule = 1UL << __fls(cfg->pgsize_bitmap & ~PAGE_MASK);
756 	else if (cfg->pgsize_bitmap & PAGE_MASK)
757 		granule = 1UL << __ffs(cfg->pgsize_bitmap & PAGE_MASK);
758 	else
759 		granule = 0;
760 
761 	switch (granule) {
762 	case SZ_4K:
763 		page_sizes = (SZ_4K | SZ_2M | SZ_1G);
764 		break;
765 	case SZ_16K:
766 		page_sizes = (SZ_16K | SZ_32M);
767 		break;
768 	case SZ_64K:
769 		max_addr_bits = 52;
770 		page_sizes = (SZ_64K | SZ_512M);
771 		if (cfg->oas > 48)
772 			page_sizes |= 1ULL << 42; /* 4TB */
773 		break;
774 	default:
775 		page_sizes = 0;
776 	}
777 
778 	cfg->pgsize_bitmap &= page_sizes;
779 	cfg->ias = min(cfg->ias, max_addr_bits);
780 	cfg->oas = min(cfg->oas, max_addr_bits);
781 }
782 
783 static struct arm_lpae_io_pgtable *
784 arm_lpae_alloc_pgtable(struct io_pgtable_cfg *cfg)
785 {
786 	struct arm_lpae_io_pgtable *data;
787 	int levels, va_bits, pg_shift;
788 
789 	arm_lpae_restrict_pgsizes(cfg);
790 
791 	if (!(cfg->pgsize_bitmap & (SZ_4K | SZ_16K | SZ_64K)))
792 		return NULL;
793 
794 	if (cfg->ias > ARM_LPAE_MAX_ADDR_BITS)
795 		return NULL;
796 
797 	if (cfg->oas > ARM_LPAE_MAX_ADDR_BITS)
798 		return NULL;
799 
800 	data = kmalloc(sizeof(*data), GFP_KERNEL);
801 	if (!data)
802 		return NULL;
803 
804 	pg_shift = __ffs(cfg->pgsize_bitmap);
805 	data->bits_per_level = pg_shift - ilog2(sizeof(arm_lpae_iopte));
806 
807 	va_bits = cfg->ias - pg_shift;
808 	levels = DIV_ROUND_UP(va_bits, data->bits_per_level);
809 	data->start_level = ARM_LPAE_MAX_LEVELS - levels;
810 
811 	/* Calculate the actual size of our pgd (without concatenation) */
812 	data->pgd_bits = va_bits - (data->bits_per_level * (levels - 1));
813 
814 	data->iop.ops = (struct io_pgtable_ops) {
815 		.map		= arm_lpae_map,
816 		.map_pages	= arm_lpae_map_pages,
817 		.unmap		= arm_lpae_unmap,
818 		.unmap_pages	= arm_lpae_unmap_pages,
819 		.iova_to_phys	= arm_lpae_iova_to_phys,
820 	};
821 
822 	return data;
823 }
824 
825 static struct io_pgtable *
826 arm_64_lpae_alloc_pgtable_s1(struct io_pgtable_cfg *cfg, void *cookie)
827 {
828 	u64 reg;
829 	struct arm_lpae_io_pgtable *data;
830 	typeof(&cfg->arm_lpae_s1_cfg.tcr) tcr = &cfg->arm_lpae_s1_cfg.tcr;
831 	bool tg1;
832 
833 	if (cfg->quirks & ~(IO_PGTABLE_QUIRK_ARM_NS |
834 			    IO_PGTABLE_QUIRK_ARM_TTBR1 |
835 			    IO_PGTABLE_QUIRK_ARM_OUTER_WBWA))
836 		return NULL;
837 
838 	data = arm_lpae_alloc_pgtable(cfg);
839 	if (!data)
840 		return NULL;
841 
842 	/* TCR */
843 	if (cfg->coherent_walk) {
844 		tcr->sh = ARM_LPAE_TCR_SH_IS;
845 		tcr->irgn = ARM_LPAE_TCR_RGN_WBWA;
846 		tcr->orgn = ARM_LPAE_TCR_RGN_WBWA;
847 		if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_OUTER_WBWA)
848 			goto out_free_data;
849 	} else {
850 		tcr->sh = ARM_LPAE_TCR_SH_OS;
851 		tcr->irgn = ARM_LPAE_TCR_RGN_NC;
852 		if (!(cfg->quirks & IO_PGTABLE_QUIRK_ARM_OUTER_WBWA))
853 			tcr->orgn = ARM_LPAE_TCR_RGN_NC;
854 		else
855 			tcr->orgn = ARM_LPAE_TCR_RGN_WBWA;
856 	}
857 
858 	tg1 = cfg->quirks & IO_PGTABLE_QUIRK_ARM_TTBR1;
859 	switch (ARM_LPAE_GRANULE(data)) {
860 	case SZ_4K:
861 		tcr->tg = tg1 ? ARM_LPAE_TCR_TG1_4K : ARM_LPAE_TCR_TG0_4K;
862 		break;
863 	case SZ_16K:
864 		tcr->tg = tg1 ? ARM_LPAE_TCR_TG1_16K : ARM_LPAE_TCR_TG0_16K;
865 		break;
866 	case SZ_64K:
867 		tcr->tg = tg1 ? ARM_LPAE_TCR_TG1_64K : ARM_LPAE_TCR_TG0_64K;
868 		break;
869 	}
870 
871 	switch (cfg->oas) {
872 	case 32:
873 		tcr->ips = ARM_LPAE_TCR_PS_32_BIT;
874 		break;
875 	case 36:
876 		tcr->ips = ARM_LPAE_TCR_PS_36_BIT;
877 		break;
878 	case 40:
879 		tcr->ips = ARM_LPAE_TCR_PS_40_BIT;
880 		break;
881 	case 42:
882 		tcr->ips = ARM_LPAE_TCR_PS_42_BIT;
883 		break;
884 	case 44:
885 		tcr->ips = ARM_LPAE_TCR_PS_44_BIT;
886 		break;
887 	case 48:
888 		tcr->ips = ARM_LPAE_TCR_PS_48_BIT;
889 		break;
890 	case 52:
891 		tcr->ips = ARM_LPAE_TCR_PS_52_BIT;
892 		break;
893 	default:
894 		goto out_free_data;
895 	}
896 
897 	tcr->tsz = 64ULL - cfg->ias;
898 
899 	/* MAIRs */
900 	reg = (ARM_LPAE_MAIR_ATTR_NC
901 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_NC)) |
902 	      (ARM_LPAE_MAIR_ATTR_WBRWA
903 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_CACHE)) |
904 	      (ARM_LPAE_MAIR_ATTR_DEVICE
905 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_DEV)) |
906 	      (ARM_LPAE_MAIR_ATTR_INC_OWBRWA
907 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_INC_OCACHE));
908 
909 	cfg->arm_lpae_s1_cfg.mair = reg;
910 
911 	/* Looking good; allocate a pgd */
912 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data),
913 					   GFP_KERNEL, cfg);
914 	if (!data->pgd)
915 		goto out_free_data;
916 
917 	/* Ensure the empty pgd is visible before any actual TTBR write */
918 	wmb();
919 
920 	/* TTBR */
921 	cfg->arm_lpae_s1_cfg.ttbr = virt_to_phys(data->pgd);
922 	return &data->iop;
923 
924 out_free_data:
925 	kfree(data);
926 	return NULL;
927 }
928 
929 static struct io_pgtable *
930 arm_64_lpae_alloc_pgtable_s2(struct io_pgtable_cfg *cfg, void *cookie)
931 {
932 	u64 sl;
933 	struct arm_lpae_io_pgtable *data;
934 	typeof(&cfg->arm_lpae_s2_cfg.vtcr) vtcr = &cfg->arm_lpae_s2_cfg.vtcr;
935 
936 	/* The NS quirk doesn't apply at stage 2 */
937 	if (cfg->quirks)
938 		return NULL;
939 
940 	data = arm_lpae_alloc_pgtable(cfg);
941 	if (!data)
942 		return NULL;
943 
944 	/*
945 	 * Concatenate PGDs at level 1 if possible in order to reduce
946 	 * the depth of the stage-2 walk.
947 	 */
948 	if (data->start_level == 0) {
949 		unsigned long pgd_pages;
950 
951 		pgd_pages = ARM_LPAE_PGD_SIZE(data) / sizeof(arm_lpae_iopte);
952 		if (pgd_pages <= ARM_LPAE_S2_MAX_CONCAT_PAGES) {
953 			data->pgd_bits += data->bits_per_level;
954 			data->start_level++;
955 		}
956 	}
957 
958 	/* VTCR */
959 	if (cfg->coherent_walk) {
960 		vtcr->sh = ARM_LPAE_TCR_SH_IS;
961 		vtcr->irgn = ARM_LPAE_TCR_RGN_WBWA;
962 		vtcr->orgn = ARM_LPAE_TCR_RGN_WBWA;
963 	} else {
964 		vtcr->sh = ARM_LPAE_TCR_SH_OS;
965 		vtcr->irgn = ARM_LPAE_TCR_RGN_NC;
966 		vtcr->orgn = ARM_LPAE_TCR_RGN_NC;
967 	}
968 
969 	sl = data->start_level;
970 
971 	switch (ARM_LPAE_GRANULE(data)) {
972 	case SZ_4K:
973 		vtcr->tg = ARM_LPAE_TCR_TG0_4K;
974 		sl++; /* SL0 format is different for 4K granule size */
975 		break;
976 	case SZ_16K:
977 		vtcr->tg = ARM_LPAE_TCR_TG0_16K;
978 		break;
979 	case SZ_64K:
980 		vtcr->tg = ARM_LPAE_TCR_TG0_64K;
981 		break;
982 	}
983 
984 	switch (cfg->oas) {
985 	case 32:
986 		vtcr->ps = ARM_LPAE_TCR_PS_32_BIT;
987 		break;
988 	case 36:
989 		vtcr->ps = ARM_LPAE_TCR_PS_36_BIT;
990 		break;
991 	case 40:
992 		vtcr->ps = ARM_LPAE_TCR_PS_40_BIT;
993 		break;
994 	case 42:
995 		vtcr->ps = ARM_LPAE_TCR_PS_42_BIT;
996 		break;
997 	case 44:
998 		vtcr->ps = ARM_LPAE_TCR_PS_44_BIT;
999 		break;
1000 	case 48:
1001 		vtcr->ps = ARM_LPAE_TCR_PS_48_BIT;
1002 		break;
1003 	case 52:
1004 		vtcr->ps = ARM_LPAE_TCR_PS_52_BIT;
1005 		break;
1006 	default:
1007 		goto out_free_data;
1008 	}
1009 
1010 	vtcr->tsz = 64ULL - cfg->ias;
1011 	vtcr->sl = ~sl & ARM_LPAE_VTCR_SL0_MASK;
1012 
1013 	/* Allocate pgd pages */
1014 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data),
1015 					   GFP_KERNEL, cfg);
1016 	if (!data->pgd)
1017 		goto out_free_data;
1018 
1019 	/* Ensure the empty pgd is visible before any actual TTBR write */
1020 	wmb();
1021 
1022 	/* VTTBR */
1023 	cfg->arm_lpae_s2_cfg.vttbr = virt_to_phys(data->pgd);
1024 	return &data->iop;
1025 
1026 out_free_data:
1027 	kfree(data);
1028 	return NULL;
1029 }
1030 
1031 static struct io_pgtable *
1032 arm_32_lpae_alloc_pgtable_s1(struct io_pgtable_cfg *cfg, void *cookie)
1033 {
1034 	if (cfg->ias > 32 || cfg->oas > 40)
1035 		return NULL;
1036 
1037 	cfg->pgsize_bitmap &= (SZ_4K | SZ_2M | SZ_1G);
1038 	return arm_64_lpae_alloc_pgtable_s1(cfg, cookie);
1039 }
1040 
1041 static struct io_pgtable *
1042 arm_32_lpae_alloc_pgtable_s2(struct io_pgtable_cfg *cfg, void *cookie)
1043 {
1044 	if (cfg->ias > 40 || cfg->oas > 40)
1045 		return NULL;
1046 
1047 	cfg->pgsize_bitmap &= (SZ_4K | SZ_2M | SZ_1G);
1048 	return arm_64_lpae_alloc_pgtable_s2(cfg, cookie);
1049 }
1050 
1051 static struct io_pgtable *
1052 arm_mali_lpae_alloc_pgtable(struct io_pgtable_cfg *cfg, void *cookie)
1053 {
1054 	struct arm_lpae_io_pgtable *data;
1055 
1056 	/* No quirks for Mali (hopefully) */
1057 	if (cfg->quirks)
1058 		return NULL;
1059 
1060 	if (cfg->ias > 48 || cfg->oas > 40)
1061 		return NULL;
1062 
1063 	cfg->pgsize_bitmap &= (SZ_4K | SZ_2M | SZ_1G);
1064 
1065 	data = arm_lpae_alloc_pgtable(cfg);
1066 	if (!data)
1067 		return NULL;
1068 
1069 	/* Mali seems to need a full 4-level table regardless of IAS */
1070 	if (data->start_level > 0) {
1071 		data->start_level = 0;
1072 		data->pgd_bits = 0;
1073 	}
1074 	/*
1075 	 * MEMATTR: Mali has no actual notion of a non-cacheable type, so the
1076 	 * best we can do is mimic the out-of-tree driver and hope that the
1077 	 * "implementation-defined caching policy" is good enough. Similarly,
1078 	 * we'll use it for the sake of a valid attribute for our 'device'
1079 	 * index, although callers should never request that in practice.
1080 	 */
1081 	cfg->arm_mali_lpae_cfg.memattr =
1082 		(ARM_MALI_LPAE_MEMATTR_IMP_DEF
1083 		 << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_NC)) |
1084 		(ARM_MALI_LPAE_MEMATTR_WRITE_ALLOC
1085 		 << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_CACHE)) |
1086 		(ARM_MALI_LPAE_MEMATTR_IMP_DEF
1087 		 << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_DEV));
1088 
1089 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data), GFP_KERNEL,
1090 					   cfg);
1091 	if (!data->pgd)
1092 		goto out_free_data;
1093 
1094 	/* Ensure the empty pgd is visible before TRANSTAB can be written */
1095 	wmb();
1096 
1097 	cfg->arm_mali_lpae_cfg.transtab = virt_to_phys(data->pgd) |
1098 					  ARM_MALI_LPAE_TTBR_READ_INNER |
1099 					  ARM_MALI_LPAE_TTBR_ADRMODE_TABLE;
1100 	if (cfg->coherent_walk)
1101 		cfg->arm_mali_lpae_cfg.transtab |= ARM_MALI_LPAE_TTBR_SHARE_OUTER;
1102 
1103 	return &data->iop;
1104 
1105 out_free_data:
1106 	kfree(data);
1107 	return NULL;
1108 }
1109 
1110 static struct io_pgtable *
1111 apple_dart_alloc_pgtable(struct io_pgtable_cfg *cfg, void *cookie)
1112 {
1113 	struct arm_lpae_io_pgtable *data;
1114 	int i;
1115 
1116 	if (cfg->oas > 36)
1117 		return NULL;
1118 
1119 	data = arm_lpae_alloc_pgtable(cfg);
1120 	if (!data)
1121 		return NULL;
1122 
1123 	/*
1124 	 * The table format itself always uses two levels, but the total VA
1125 	 * space is mapped by four separate tables, making the MMIO registers
1126 	 * an effective "level 1". For simplicity, though, we treat this
1127 	 * equivalently to LPAE stage 2 concatenation at level 2, with the
1128 	 * additional TTBRs each just pointing at consecutive pages.
1129 	 */
1130 	if (data->start_level < 1)
1131 		goto out_free_data;
1132 	if (data->start_level == 1 && data->pgd_bits > 2)
1133 		goto out_free_data;
1134 	if (data->start_level > 1)
1135 		data->pgd_bits = 0;
1136 	data->start_level = 2;
1137 	cfg->apple_dart_cfg.n_ttbrs = 1 << data->pgd_bits;
1138 	data->pgd_bits += data->bits_per_level;
1139 
1140 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data), GFP_KERNEL,
1141 					   cfg);
1142 	if (!data->pgd)
1143 		goto out_free_data;
1144 
1145 	for (i = 0; i < cfg->apple_dart_cfg.n_ttbrs; ++i)
1146 		cfg->apple_dart_cfg.ttbr[i] =
1147 			virt_to_phys(data->pgd + i * ARM_LPAE_GRANULE(data));
1148 
1149 	return &data->iop;
1150 
1151 out_free_data:
1152 	kfree(data);
1153 	return NULL;
1154 }
1155 
1156 struct io_pgtable_init_fns io_pgtable_arm_64_lpae_s1_init_fns = {
1157 	.alloc	= arm_64_lpae_alloc_pgtable_s1,
1158 	.free	= arm_lpae_free_pgtable,
1159 };
1160 
1161 struct io_pgtable_init_fns io_pgtable_arm_64_lpae_s2_init_fns = {
1162 	.alloc	= arm_64_lpae_alloc_pgtable_s2,
1163 	.free	= arm_lpae_free_pgtable,
1164 };
1165 
1166 struct io_pgtable_init_fns io_pgtable_arm_32_lpae_s1_init_fns = {
1167 	.alloc	= arm_32_lpae_alloc_pgtable_s1,
1168 	.free	= arm_lpae_free_pgtable,
1169 };
1170 
1171 struct io_pgtable_init_fns io_pgtable_arm_32_lpae_s2_init_fns = {
1172 	.alloc	= arm_32_lpae_alloc_pgtable_s2,
1173 	.free	= arm_lpae_free_pgtable,
1174 };
1175 
1176 struct io_pgtable_init_fns io_pgtable_arm_mali_lpae_init_fns = {
1177 	.alloc	= arm_mali_lpae_alloc_pgtable,
1178 	.free	= arm_lpae_free_pgtable,
1179 };
1180 
1181 struct io_pgtable_init_fns io_pgtable_apple_dart_init_fns = {
1182 	.alloc	= apple_dart_alloc_pgtable,
1183 	.free	= arm_lpae_free_pgtable,
1184 };
1185 
1186 #ifdef CONFIG_IOMMU_IO_PGTABLE_LPAE_SELFTEST
1187 
1188 static struct io_pgtable_cfg *cfg_cookie __initdata;
1189 
1190 static void __init dummy_tlb_flush_all(void *cookie)
1191 {
1192 	WARN_ON(cookie != cfg_cookie);
1193 }
1194 
1195 static void __init dummy_tlb_flush(unsigned long iova, size_t size,
1196 				   size_t granule, void *cookie)
1197 {
1198 	WARN_ON(cookie != cfg_cookie);
1199 	WARN_ON(!(size & cfg_cookie->pgsize_bitmap));
1200 }
1201 
1202 static void __init dummy_tlb_add_page(struct iommu_iotlb_gather *gather,
1203 				      unsigned long iova, size_t granule,
1204 				      void *cookie)
1205 {
1206 	dummy_tlb_flush(iova, granule, granule, cookie);
1207 }
1208 
1209 static const struct iommu_flush_ops dummy_tlb_ops __initconst = {
1210 	.tlb_flush_all	= dummy_tlb_flush_all,
1211 	.tlb_flush_walk	= dummy_tlb_flush,
1212 	.tlb_add_page	= dummy_tlb_add_page,
1213 };
1214 
1215 static void __init arm_lpae_dump_ops(struct io_pgtable_ops *ops)
1216 {
1217 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
1218 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
1219 
1220 	pr_err("cfg: pgsize_bitmap 0x%lx, ias %u-bit\n",
1221 		cfg->pgsize_bitmap, cfg->ias);
1222 	pr_err("data: %d levels, 0x%zx pgd_size, %u pg_shift, %u bits_per_level, pgd @ %p\n",
1223 		ARM_LPAE_MAX_LEVELS - data->start_level, ARM_LPAE_PGD_SIZE(data),
1224 		ilog2(ARM_LPAE_GRANULE(data)), data->bits_per_level, data->pgd);
1225 }
1226 
1227 #define __FAIL(ops, i)	({						\
1228 		WARN(1, "selftest: test failed for fmt idx %d\n", (i));	\
1229 		arm_lpae_dump_ops(ops);					\
1230 		selftest_running = false;				\
1231 		-EFAULT;						\
1232 })
1233 
1234 static int __init arm_lpae_run_tests(struct io_pgtable_cfg *cfg)
1235 {
1236 	static const enum io_pgtable_fmt fmts[] __initconst = {
1237 		ARM_64_LPAE_S1,
1238 		ARM_64_LPAE_S2,
1239 	};
1240 
1241 	int i, j;
1242 	unsigned long iova;
1243 	size_t size;
1244 	struct io_pgtable_ops *ops;
1245 
1246 	selftest_running = true;
1247 
1248 	for (i = 0; i < ARRAY_SIZE(fmts); ++i) {
1249 		cfg_cookie = cfg;
1250 		ops = alloc_io_pgtable_ops(fmts[i], cfg, cfg);
1251 		if (!ops) {
1252 			pr_err("selftest: failed to allocate io pgtable ops\n");
1253 			return -ENOMEM;
1254 		}
1255 
1256 		/*
1257 		 * Initial sanity checks.
1258 		 * Empty page tables shouldn't provide any translations.
1259 		 */
1260 		if (ops->iova_to_phys(ops, 42))
1261 			return __FAIL(ops, i);
1262 
1263 		if (ops->iova_to_phys(ops, SZ_1G + 42))
1264 			return __FAIL(ops, i);
1265 
1266 		if (ops->iova_to_phys(ops, SZ_2G + 42))
1267 			return __FAIL(ops, i);
1268 
1269 		/*
1270 		 * Distinct mappings of different granule sizes.
1271 		 */
1272 		iova = 0;
1273 		for_each_set_bit(j, &cfg->pgsize_bitmap, BITS_PER_LONG) {
1274 			size = 1UL << j;
1275 
1276 			if (ops->map(ops, iova, iova, size, IOMMU_READ |
1277 							    IOMMU_WRITE |
1278 							    IOMMU_NOEXEC |
1279 							    IOMMU_CACHE, GFP_KERNEL))
1280 				return __FAIL(ops, i);
1281 
1282 			/* Overlapping mappings */
1283 			if (!ops->map(ops, iova, iova + size, size,
1284 				      IOMMU_READ | IOMMU_NOEXEC, GFP_KERNEL))
1285 				return __FAIL(ops, i);
1286 
1287 			if (ops->iova_to_phys(ops, iova + 42) != (iova + 42))
1288 				return __FAIL(ops, i);
1289 
1290 			iova += SZ_1G;
1291 		}
1292 
1293 		/* Partial unmap */
1294 		size = 1UL << __ffs(cfg->pgsize_bitmap);
1295 		if (ops->unmap(ops, SZ_1G + size, size, NULL) != size)
1296 			return __FAIL(ops, i);
1297 
1298 		/* Remap of partial unmap */
1299 		if (ops->map(ops, SZ_1G + size, size, size, IOMMU_READ, GFP_KERNEL))
1300 			return __FAIL(ops, i);
1301 
1302 		if (ops->iova_to_phys(ops, SZ_1G + size + 42) != (size + 42))
1303 			return __FAIL(ops, i);
1304 
1305 		/* Full unmap */
1306 		iova = 0;
1307 		for_each_set_bit(j, &cfg->pgsize_bitmap, BITS_PER_LONG) {
1308 			size = 1UL << j;
1309 
1310 			if (ops->unmap(ops, iova, size, NULL) != size)
1311 				return __FAIL(ops, i);
1312 
1313 			if (ops->iova_to_phys(ops, iova + 42))
1314 				return __FAIL(ops, i);
1315 
1316 			/* Remap full block */
1317 			if (ops->map(ops, iova, iova, size, IOMMU_WRITE, GFP_KERNEL))
1318 				return __FAIL(ops, i);
1319 
1320 			if (ops->iova_to_phys(ops, iova + 42) != (iova + 42))
1321 				return __FAIL(ops, i);
1322 
1323 			iova += SZ_1G;
1324 		}
1325 
1326 		free_io_pgtable_ops(ops);
1327 	}
1328 
1329 	selftest_running = false;
1330 	return 0;
1331 }
1332 
1333 static int __init arm_lpae_do_selftests(void)
1334 {
1335 	static const unsigned long pgsize[] __initconst = {
1336 		SZ_4K | SZ_2M | SZ_1G,
1337 		SZ_16K | SZ_32M,
1338 		SZ_64K | SZ_512M,
1339 	};
1340 
1341 	static const unsigned int ias[] __initconst = {
1342 		32, 36, 40, 42, 44, 48,
1343 	};
1344 
1345 	int i, j, pass = 0, fail = 0;
1346 	struct io_pgtable_cfg cfg = {
1347 		.tlb = &dummy_tlb_ops,
1348 		.oas = 48,
1349 		.coherent_walk = true,
1350 	};
1351 
1352 	for (i = 0; i < ARRAY_SIZE(pgsize); ++i) {
1353 		for (j = 0; j < ARRAY_SIZE(ias); ++j) {
1354 			cfg.pgsize_bitmap = pgsize[i];
1355 			cfg.ias = ias[j];
1356 			pr_info("selftest: pgsize_bitmap 0x%08lx, IAS %u\n",
1357 				pgsize[i], ias[j]);
1358 			if (arm_lpae_run_tests(&cfg))
1359 				fail++;
1360 			else
1361 				pass++;
1362 		}
1363 	}
1364 
1365 	pr_info("selftest: completed with %d PASS %d FAIL\n", pass, fail);
1366 	return fail ? -EFAULT : 0;
1367 }
1368 subsys_initcall(arm_lpae_do_selftests);
1369 #endif
1370