xref: /linux/drivers/iommu/io-pgtable-arm.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 #include "iommu-pages.h"
25 
26 #define ARM_LPAE_MAX_ADDR_BITS		52
27 #define ARM_LPAE_S2_MAX_CONCAT_PAGES	16
28 #define ARM_LPAE_MAX_LEVELS		4
29 
30 /* Struct accessors */
31 #define io_pgtable_to_data(x)						\
32 	container_of((x), struct arm_lpae_io_pgtable, iop)
33 
34 #define io_pgtable_ops_to_data(x)					\
35 	io_pgtable_to_data(io_pgtable_ops_to_pgtable(x))
36 
37 /*
38  * Calculate the right shift amount to get to the portion describing level l
39  * in a virtual address mapped by the pagetable in d.
40  */
41 #define ARM_LPAE_LVL_SHIFT(l,d)						\
42 	(((ARM_LPAE_MAX_LEVELS - (l)) * (d)->bits_per_level) +		\
43 	ilog2(sizeof(arm_lpae_iopte)))
44 
45 #define ARM_LPAE_GRANULE(d)						\
46 	(sizeof(arm_lpae_iopte) << (d)->bits_per_level)
47 #define ARM_LPAE_PGD_SIZE(d)						\
48 	(sizeof(arm_lpae_iopte) << (d)->pgd_bits)
49 
50 #define ARM_LPAE_PTES_PER_TABLE(d)					\
51 	(ARM_LPAE_GRANULE(d) >> ilog2(sizeof(arm_lpae_iopte)))
52 
53 /*
54  * Calculate the index at level l used to map virtual address a using the
55  * pagetable in d.
56  */
57 #define ARM_LPAE_PGD_IDX(l,d)						\
58 	((l) == (d)->start_level ? (d)->pgd_bits - (d)->bits_per_level : 0)
59 
60 #define ARM_LPAE_LVL_IDX(a,l,d)						\
61 	(((u64)(a) >> ARM_LPAE_LVL_SHIFT(l,d)) &			\
62 	 ((1 << ((d)->bits_per_level + ARM_LPAE_PGD_IDX(l,d))) - 1))
63 
64 /* Calculate the block/page mapping size at level l for pagetable in d. */
65 #define ARM_LPAE_BLOCK_SIZE(l,d)	(1ULL << ARM_LPAE_LVL_SHIFT(l,d))
66 
67 /* Page table bits */
68 #define ARM_LPAE_PTE_TYPE_SHIFT		0
69 #define ARM_LPAE_PTE_TYPE_MASK		0x3
70 
71 #define ARM_LPAE_PTE_TYPE_BLOCK		1
72 #define ARM_LPAE_PTE_TYPE_TABLE		3
73 #define ARM_LPAE_PTE_TYPE_PAGE		3
74 
75 #define ARM_LPAE_PTE_ADDR_MASK		GENMASK_ULL(47,12)
76 
77 #define ARM_LPAE_PTE_NSTABLE		(((arm_lpae_iopte)1) << 63)
78 #define ARM_LPAE_PTE_XN			(((arm_lpae_iopte)3) << 53)
79 #define ARM_LPAE_PTE_DBM		(((arm_lpae_iopte)1) << 51)
80 #define ARM_LPAE_PTE_AF			(((arm_lpae_iopte)1) << 10)
81 #define ARM_LPAE_PTE_SH_NS		(((arm_lpae_iopte)0) << 8)
82 #define ARM_LPAE_PTE_SH_OS		(((arm_lpae_iopte)2) << 8)
83 #define ARM_LPAE_PTE_SH_IS		(((arm_lpae_iopte)3) << 8)
84 #define ARM_LPAE_PTE_NS			(((arm_lpae_iopte)1) << 5)
85 #define ARM_LPAE_PTE_VALID		(((arm_lpae_iopte)1) << 0)
86 
87 #define ARM_LPAE_PTE_ATTR_LO_MASK	(((arm_lpae_iopte)0x3ff) << 2)
88 /* Ignore the contiguous bit for block splitting */
89 #define ARM_LPAE_PTE_ATTR_HI_MASK	(ARM_LPAE_PTE_XN | ARM_LPAE_PTE_DBM)
90 #define ARM_LPAE_PTE_ATTR_MASK		(ARM_LPAE_PTE_ATTR_LO_MASK |	\
91 					 ARM_LPAE_PTE_ATTR_HI_MASK)
92 /* Software bit for solving coherency races */
93 #define ARM_LPAE_PTE_SW_SYNC		(((arm_lpae_iopte)1) << 55)
94 
95 /* Stage-1 PTE */
96 #define ARM_LPAE_PTE_AP_UNPRIV		(((arm_lpae_iopte)1) << 6)
97 #define ARM_LPAE_PTE_AP_RDONLY_BIT	7
98 #define ARM_LPAE_PTE_AP_RDONLY		(((arm_lpae_iopte)1) << \
99 					   ARM_LPAE_PTE_AP_RDONLY_BIT)
100 #define ARM_LPAE_PTE_AP_WR_CLEAN_MASK	(ARM_LPAE_PTE_AP_RDONLY | \
101 					 ARM_LPAE_PTE_DBM)
102 #define ARM_LPAE_PTE_ATTRINDX_SHIFT	2
103 #define ARM_LPAE_PTE_nG			(((arm_lpae_iopte)1) << 11)
104 
105 /* Stage-2 PTE */
106 #define ARM_LPAE_PTE_HAP_FAULT		(((arm_lpae_iopte)0) << 6)
107 #define ARM_LPAE_PTE_HAP_READ		(((arm_lpae_iopte)1) << 6)
108 #define ARM_LPAE_PTE_HAP_WRITE		(((arm_lpae_iopte)2) << 6)
109 #define ARM_LPAE_PTE_MEMATTR_OIWB	(((arm_lpae_iopte)0xf) << 2)
110 #define ARM_LPAE_PTE_MEMATTR_NC		(((arm_lpae_iopte)0x5) << 2)
111 #define ARM_LPAE_PTE_MEMATTR_DEV	(((arm_lpae_iopte)0x1) << 2)
112 
113 /* Register bits */
114 #define ARM_LPAE_VTCR_SL0_MASK		0x3
115 
116 #define ARM_LPAE_TCR_T0SZ_SHIFT		0
117 
118 #define ARM_LPAE_VTCR_PS_SHIFT		16
119 #define ARM_LPAE_VTCR_PS_MASK		0x7
120 
121 #define ARM_LPAE_MAIR_ATTR_SHIFT(n)	((n) << 3)
122 #define ARM_LPAE_MAIR_ATTR_MASK		0xff
123 #define ARM_LPAE_MAIR_ATTR_DEVICE	0x04
124 #define ARM_LPAE_MAIR_ATTR_NC		0x44
125 #define ARM_LPAE_MAIR_ATTR_INC_OWBRWA	0xf4
126 #define ARM_LPAE_MAIR_ATTR_WBRWA	0xff
127 #define ARM_LPAE_MAIR_ATTR_IDX_NC	0
128 #define ARM_LPAE_MAIR_ATTR_IDX_CACHE	1
129 #define ARM_LPAE_MAIR_ATTR_IDX_DEV	2
130 #define ARM_LPAE_MAIR_ATTR_IDX_INC_OCACHE	3
131 
132 #define ARM_MALI_LPAE_TTBR_ADRMODE_TABLE (3u << 0)
133 #define ARM_MALI_LPAE_TTBR_READ_INNER	BIT(2)
134 #define ARM_MALI_LPAE_TTBR_SHARE_OUTER	BIT(4)
135 
136 #define ARM_MALI_LPAE_MEMATTR_IMP_DEF	0x88ULL
137 #define ARM_MALI_LPAE_MEMATTR_WRITE_ALLOC 0x8DULL
138 
139 /* IOPTE accessors */
140 #define iopte_deref(pte,d) __va(iopte_to_paddr(pte, d))
141 
142 #define iopte_type(pte)					\
143 	(((pte) >> ARM_LPAE_PTE_TYPE_SHIFT) & ARM_LPAE_PTE_TYPE_MASK)
144 
145 #define iopte_prot(pte)	((pte) & ARM_LPAE_PTE_ATTR_MASK)
146 
147 #define iopte_writeable_dirty(pte)				\
148 	(((pte) & ARM_LPAE_PTE_AP_WR_CLEAN_MASK) == ARM_LPAE_PTE_DBM)
149 
150 #define iopte_set_writeable_clean(ptep)				\
151 	set_bit(ARM_LPAE_PTE_AP_RDONLY_BIT, (unsigned long *)(ptep))
152 
153 struct arm_lpae_io_pgtable {
154 	struct io_pgtable	iop;
155 
156 	int			pgd_bits;
157 	int			start_level;
158 	int			bits_per_level;
159 
160 	void			*pgd;
161 };
162 
163 typedef u64 arm_lpae_iopte;
164 
165 static inline bool iopte_leaf(arm_lpae_iopte pte, int lvl,
166 			      enum io_pgtable_fmt fmt)
167 {
168 	if (lvl == (ARM_LPAE_MAX_LEVELS - 1) && fmt != ARM_MALI_LPAE)
169 		return iopte_type(pte) == ARM_LPAE_PTE_TYPE_PAGE;
170 
171 	return iopte_type(pte) == ARM_LPAE_PTE_TYPE_BLOCK;
172 }
173 
174 static inline bool iopte_table(arm_lpae_iopte pte, int lvl)
175 {
176 	if (lvl == (ARM_LPAE_MAX_LEVELS - 1))
177 		return false;
178 	return iopte_type(pte) == ARM_LPAE_PTE_TYPE_TABLE;
179 }
180 
181 static arm_lpae_iopte paddr_to_iopte(phys_addr_t paddr,
182 				     struct arm_lpae_io_pgtable *data)
183 {
184 	arm_lpae_iopte pte = paddr;
185 
186 	/* Of the bits which overlap, either 51:48 or 15:12 are always RES0 */
187 	return (pte | (pte >> (48 - 12))) & ARM_LPAE_PTE_ADDR_MASK;
188 }
189 
190 static phys_addr_t iopte_to_paddr(arm_lpae_iopte pte,
191 				  struct arm_lpae_io_pgtable *data)
192 {
193 	u64 paddr = pte & ARM_LPAE_PTE_ADDR_MASK;
194 
195 	if (ARM_LPAE_GRANULE(data) < SZ_64K)
196 		return paddr;
197 
198 	/* Rotate the packed high-order bits back to the top */
199 	return (paddr | (paddr << (48 - 12))) & (ARM_LPAE_PTE_ADDR_MASK << 4);
200 }
201 
202 static bool selftest_running = false;
203 
204 static dma_addr_t __arm_lpae_dma_addr(void *pages)
205 {
206 	return (dma_addr_t)virt_to_phys(pages);
207 }
208 
209 static void *__arm_lpae_alloc_pages(size_t size, gfp_t gfp,
210 				    struct io_pgtable_cfg *cfg,
211 				    void *cookie)
212 {
213 	struct device *dev = cfg->iommu_dev;
214 	int order = get_order(size);
215 	dma_addr_t dma;
216 	void *pages;
217 
218 	VM_BUG_ON((gfp & __GFP_HIGHMEM));
219 
220 	if (cfg->alloc)
221 		pages = cfg->alloc(cookie, size, gfp);
222 	else
223 		pages = iommu_alloc_pages_node(dev_to_node(dev), gfp, order);
224 
225 	if (!pages)
226 		return NULL;
227 
228 	if (!cfg->coherent_walk) {
229 		dma = dma_map_single(dev, pages, size, DMA_TO_DEVICE);
230 		if (dma_mapping_error(dev, dma))
231 			goto out_free;
232 		/*
233 		 * We depend on the IOMMU being able to work with any physical
234 		 * address directly, so if the DMA layer suggests otherwise by
235 		 * translating or truncating them, that bodes very badly...
236 		 */
237 		if (dma != virt_to_phys(pages))
238 			goto out_unmap;
239 	}
240 
241 	return pages;
242 
243 out_unmap:
244 	dev_err(dev, "Cannot accommodate DMA translation for IOMMU page tables\n");
245 	dma_unmap_single(dev, dma, size, DMA_TO_DEVICE);
246 
247 out_free:
248 	if (cfg->free)
249 		cfg->free(cookie, pages, size);
250 	else
251 		iommu_free_pages(pages, order);
252 
253 	return NULL;
254 }
255 
256 static void __arm_lpae_free_pages(void *pages, size_t size,
257 				  struct io_pgtable_cfg *cfg,
258 				  void *cookie)
259 {
260 	if (!cfg->coherent_walk)
261 		dma_unmap_single(cfg->iommu_dev, __arm_lpae_dma_addr(pages),
262 				 size, DMA_TO_DEVICE);
263 
264 	if (cfg->free)
265 		cfg->free(cookie, pages, size);
266 	else
267 		iommu_free_pages(pages, get_order(size));
268 }
269 
270 static void __arm_lpae_sync_pte(arm_lpae_iopte *ptep, int num_entries,
271 				struct io_pgtable_cfg *cfg)
272 {
273 	dma_sync_single_for_device(cfg->iommu_dev, __arm_lpae_dma_addr(ptep),
274 				   sizeof(*ptep) * num_entries, DMA_TO_DEVICE);
275 }
276 
277 static void __arm_lpae_clear_pte(arm_lpae_iopte *ptep, struct io_pgtable_cfg *cfg, int num_entries)
278 {
279 	for (int i = 0; i < num_entries; i++)
280 		ptep[i] = 0;
281 
282 	if (!cfg->coherent_walk && num_entries)
283 		__arm_lpae_sync_pte(ptep, num_entries, cfg);
284 }
285 
286 static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
287 			       struct iommu_iotlb_gather *gather,
288 			       unsigned long iova, size_t size, size_t pgcount,
289 			       int lvl, arm_lpae_iopte *ptep);
290 
291 static void __arm_lpae_init_pte(struct arm_lpae_io_pgtable *data,
292 				phys_addr_t paddr, arm_lpae_iopte prot,
293 				int lvl, int num_entries, arm_lpae_iopte *ptep)
294 {
295 	arm_lpae_iopte pte = prot;
296 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
297 	size_t sz = ARM_LPAE_BLOCK_SIZE(lvl, data);
298 	int i;
299 
300 	if (data->iop.fmt != ARM_MALI_LPAE && lvl == ARM_LPAE_MAX_LEVELS - 1)
301 		pte |= ARM_LPAE_PTE_TYPE_PAGE;
302 	else
303 		pte |= ARM_LPAE_PTE_TYPE_BLOCK;
304 
305 	for (i = 0; i < num_entries; i++)
306 		ptep[i] = pte | paddr_to_iopte(paddr + i * sz, data);
307 
308 	if (!cfg->coherent_walk)
309 		__arm_lpae_sync_pte(ptep, num_entries, cfg);
310 }
311 
312 static int arm_lpae_init_pte(struct arm_lpae_io_pgtable *data,
313 			     unsigned long iova, phys_addr_t paddr,
314 			     arm_lpae_iopte prot, int lvl, int num_entries,
315 			     arm_lpae_iopte *ptep)
316 {
317 	int i;
318 
319 	for (i = 0; i < num_entries; i++)
320 		if (iopte_leaf(ptep[i], lvl, data->iop.fmt)) {
321 			/* We require an unmap first */
322 			WARN_ON(!selftest_running);
323 			return -EEXIST;
324 		} else if (iopte_type(ptep[i]) == ARM_LPAE_PTE_TYPE_TABLE) {
325 			/*
326 			 * We need to unmap and free the old table before
327 			 * overwriting it with a block entry.
328 			 */
329 			arm_lpae_iopte *tblp;
330 			size_t sz = ARM_LPAE_BLOCK_SIZE(lvl, data);
331 
332 			tblp = ptep - ARM_LPAE_LVL_IDX(iova, lvl, data);
333 			if (__arm_lpae_unmap(data, NULL, iova + i * sz, sz, 1,
334 					     lvl, tblp) != sz) {
335 				WARN_ON(1);
336 				return -EINVAL;
337 			}
338 		}
339 
340 	__arm_lpae_init_pte(data, paddr, prot, lvl, num_entries, ptep);
341 	return 0;
342 }
343 
344 static arm_lpae_iopte arm_lpae_install_table(arm_lpae_iopte *table,
345 					     arm_lpae_iopte *ptep,
346 					     arm_lpae_iopte curr,
347 					     struct arm_lpae_io_pgtable *data)
348 {
349 	arm_lpae_iopte old, new;
350 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
351 
352 	new = paddr_to_iopte(__pa(table), data) | ARM_LPAE_PTE_TYPE_TABLE;
353 	if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_NS)
354 		new |= ARM_LPAE_PTE_NSTABLE;
355 
356 	/*
357 	 * Ensure the table itself is visible before its PTE can be.
358 	 * Whilst we could get away with cmpxchg64_release below, this
359 	 * doesn't have any ordering semantics when !CONFIG_SMP.
360 	 */
361 	dma_wmb();
362 
363 	old = cmpxchg64_relaxed(ptep, curr, new);
364 
365 	if (cfg->coherent_walk || (old & ARM_LPAE_PTE_SW_SYNC))
366 		return old;
367 
368 	/* Even if it's not ours, there's no point waiting; just kick it */
369 	__arm_lpae_sync_pte(ptep, 1, cfg);
370 	if (old == curr)
371 		WRITE_ONCE(*ptep, new | ARM_LPAE_PTE_SW_SYNC);
372 
373 	return old;
374 }
375 
376 static int __arm_lpae_map(struct arm_lpae_io_pgtable *data, unsigned long iova,
377 			  phys_addr_t paddr, size_t size, size_t pgcount,
378 			  arm_lpae_iopte prot, int lvl, arm_lpae_iopte *ptep,
379 			  gfp_t gfp, size_t *mapped)
380 {
381 	arm_lpae_iopte *cptep, pte;
382 	size_t block_size = ARM_LPAE_BLOCK_SIZE(lvl, data);
383 	size_t tblsz = ARM_LPAE_GRANULE(data);
384 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
385 	int ret = 0, num_entries, max_entries, map_idx_start;
386 
387 	/* Find our entry at the current level */
388 	map_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
389 	ptep += map_idx_start;
390 
391 	/* If we can install a leaf entry at this level, then do so */
392 	if (size == block_size) {
393 		max_entries = ARM_LPAE_PTES_PER_TABLE(data) - map_idx_start;
394 		num_entries = min_t(int, pgcount, max_entries);
395 		ret = arm_lpae_init_pte(data, iova, paddr, prot, lvl, num_entries, ptep);
396 		if (!ret)
397 			*mapped += num_entries * size;
398 
399 		return ret;
400 	}
401 
402 	/* We can't allocate tables at the final level */
403 	if (WARN_ON(lvl >= ARM_LPAE_MAX_LEVELS - 1))
404 		return -EINVAL;
405 
406 	/* Grab a pointer to the next level */
407 	pte = READ_ONCE(*ptep);
408 	if (!pte) {
409 		cptep = __arm_lpae_alloc_pages(tblsz, gfp, cfg, data->iop.cookie);
410 		if (!cptep)
411 			return -ENOMEM;
412 
413 		pte = arm_lpae_install_table(cptep, ptep, 0, data);
414 		if (pte)
415 			__arm_lpae_free_pages(cptep, tblsz, cfg, data->iop.cookie);
416 	} else if (!cfg->coherent_walk && !(pte & ARM_LPAE_PTE_SW_SYNC)) {
417 		__arm_lpae_sync_pte(ptep, 1, cfg);
418 	}
419 
420 	if (pte && !iopte_leaf(pte, lvl, data->iop.fmt)) {
421 		cptep = iopte_deref(pte, data);
422 	} else if (pte) {
423 		/* We require an unmap first */
424 		WARN_ON(!selftest_running);
425 		return -EEXIST;
426 	}
427 
428 	/* Rinse, repeat */
429 	return __arm_lpae_map(data, iova, paddr, size, pgcount, prot, lvl + 1,
430 			      cptep, gfp, mapped);
431 }
432 
433 static arm_lpae_iopte arm_lpae_prot_to_pte(struct arm_lpae_io_pgtable *data,
434 					   int prot)
435 {
436 	arm_lpae_iopte pte;
437 
438 	if (data->iop.fmt == ARM_64_LPAE_S1 ||
439 	    data->iop.fmt == ARM_32_LPAE_S1) {
440 		pte = ARM_LPAE_PTE_nG;
441 		if (!(prot & IOMMU_WRITE) && (prot & IOMMU_READ))
442 			pte |= ARM_LPAE_PTE_AP_RDONLY;
443 		else if (data->iop.cfg.quirks & IO_PGTABLE_QUIRK_ARM_HD)
444 			pte |= ARM_LPAE_PTE_DBM;
445 		if (!(prot & IOMMU_PRIV))
446 			pte |= ARM_LPAE_PTE_AP_UNPRIV;
447 	} else {
448 		pte = ARM_LPAE_PTE_HAP_FAULT;
449 		if (prot & IOMMU_READ)
450 			pte |= ARM_LPAE_PTE_HAP_READ;
451 		if (prot & IOMMU_WRITE)
452 			pte |= ARM_LPAE_PTE_HAP_WRITE;
453 	}
454 
455 	/*
456 	 * Note that this logic is structured to accommodate Mali LPAE
457 	 * having stage-1-like attributes but stage-2-like permissions.
458 	 */
459 	if (data->iop.fmt == ARM_64_LPAE_S2 ||
460 	    data->iop.fmt == ARM_32_LPAE_S2) {
461 		if (prot & IOMMU_MMIO)
462 			pte |= ARM_LPAE_PTE_MEMATTR_DEV;
463 		else if (prot & IOMMU_CACHE)
464 			pte |= ARM_LPAE_PTE_MEMATTR_OIWB;
465 		else
466 			pte |= ARM_LPAE_PTE_MEMATTR_NC;
467 	} else {
468 		if (prot & IOMMU_MMIO)
469 			pte |= (ARM_LPAE_MAIR_ATTR_IDX_DEV
470 				<< ARM_LPAE_PTE_ATTRINDX_SHIFT);
471 		else if (prot & IOMMU_CACHE)
472 			pte |= (ARM_LPAE_MAIR_ATTR_IDX_CACHE
473 				<< ARM_LPAE_PTE_ATTRINDX_SHIFT);
474 	}
475 
476 	/*
477 	 * Also Mali has its own notions of shareability wherein its Inner
478 	 * domain covers the cores within the GPU, and its Outer domain is
479 	 * "outside the GPU" (i.e. either the Inner or System domain in CPU
480 	 * terms, depending on coherency).
481 	 */
482 	if (prot & IOMMU_CACHE && data->iop.fmt != ARM_MALI_LPAE)
483 		pte |= ARM_LPAE_PTE_SH_IS;
484 	else
485 		pte |= ARM_LPAE_PTE_SH_OS;
486 
487 	if (prot & IOMMU_NOEXEC)
488 		pte |= ARM_LPAE_PTE_XN;
489 
490 	if (data->iop.cfg.quirks & IO_PGTABLE_QUIRK_ARM_NS)
491 		pte |= ARM_LPAE_PTE_NS;
492 
493 	if (data->iop.fmt != ARM_MALI_LPAE)
494 		pte |= ARM_LPAE_PTE_AF;
495 
496 	return pte;
497 }
498 
499 static int arm_lpae_map_pages(struct io_pgtable_ops *ops, unsigned long iova,
500 			      phys_addr_t paddr, size_t pgsize, size_t pgcount,
501 			      int iommu_prot, gfp_t gfp, size_t *mapped)
502 {
503 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
504 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
505 	arm_lpae_iopte *ptep = data->pgd;
506 	int ret, lvl = data->start_level;
507 	arm_lpae_iopte prot;
508 	long iaext = (s64)iova >> cfg->ias;
509 
510 	if (WARN_ON(!pgsize || (pgsize & cfg->pgsize_bitmap) != pgsize))
511 		return -EINVAL;
512 
513 	if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_TTBR1)
514 		iaext = ~iaext;
515 	if (WARN_ON(iaext || paddr >> cfg->oas))
516 		return -ERANGE;
517 
518 	if (!(iommu_prot & (IOMMU_READ | IOMMU_WRITE)))
519 		return -EINVAL;
520 
521 	prot = arm_lpae_prot_to_pte(data, iommu_prot);
522 	ret = __arm_lpae_map(data, iova, paddr, pgsize, pgcount, prot, lvl,
523 			     ptep, gfp, mapped);
524 	/*
525 	 * Synchronise all PTE updates for the new mapping before there's
526 	 * a chance for anything to kick off a table walk for the new iova.
527 	 */
528 	wmb();
529 
530 	return ret;
531 }
532 
533 static void __arm_lpae_free_pgtable(struct arm_lpae_io_pgtable *data, int lvl,
534 				    arm_lpae_iopte *ptep)
535 {
536 	arm_lpae_iopte *start, *end;
537 	unsigned long table_size;
538 
539 	if (lvl == data->start_level)
540 		table_size = ARM_LPAE_PGD_SIZE(data);
541 	else
542 		table_size = ARM_LPAE_GRANULE(data);
543 
544 	start = ptep;
545 
546 	/* Only leaf entries at the last level */
547 	if (lvl == ARM_LPAE_MAX_LEVELS - 1)
548 		end = ptep;
549 	else
550 		end = (void *)ptep + table_size;
551 
552 	while (ptep != end) {
553 		arm_lpae_iopte pte = *ptep++;
554 
555 		if (!pte || iopte_leaf(pte, lvl, data->iop.fmt))
556 			continue;
557 
558 		__arm_lpae_free_pgtable(data, lvl + 1, iopte_deref(pte, data));
559 	}
560 
561 	__arm_lpae_free_pages(start, table_size, &data->iop.cfg, data->iop.cookie);
562 }
563 
564 static void arm_lpae_free_pgtable(struct io_pgtable *iop)
565 {
566 	struct arm_lpae_io_pgtable *data = io_pgtable_to_data(iop);
567 
568 	__arm_lpae_free_pgtable(data, data->start_level, data->pgd);
569 	kfree(data);
570 }
571 
572 static size_t arm_lpae_split_blk_unmap(struct arm_lpae_io_pgtable *data,
573 				       struct iommu_iotlb_gather *gather,
574 				       unsigned long iova, size_t size,
575 				       arm_lpae_iopte blk_pte, int lvl,
576 				       arm_lpae_iopte *ptep, size_t pgcount)
577 {
578 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
579 	arm_lpae_iopte pte, *tablep;
580 	phys_addr_t blk_paddr;
581 	size_t tablesz = ARM_LPAE_GRANULE(data);
582 	size_t split_sz = ARM_LPAE_BLOCK_SIZE(lvl, data);
583 	int ptes_per_table = ARM_LPAE_PTES_PER_TABLE(data);
584 	int i, unmap_idx_start = -1, num_entries = 0, max_entries;
585 
586 	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
587 		return 0;
588 
589 	tablep = __arm_lpae_alloc_pages(tablesz, GFP_ATOMIC, cfg, data->iop.cookie);
590 	if (!tablep)
591 		return 0; /* Bytes unmapped */
592 
593 	if (size == split_sz) {
594 		unmap_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
595 		max_entries = ptes_per_table - unmap_idx_start;
596 		num_entries = min_t(int, pgcount, max_entries);
597 	}
598 
599 	blk_paddr = iopte_to_paddr(blk_pte, data);
600 	pte = iopte_prot(blk_pte);
601 
602 	for (i = 0; i < ptes_per_table; i++, blk_paddr += split_sz) {
603 		/* Unmap! */
604 		if (i >= unmap_idx_start && i < (unmap_idx_start + num_entries))
605 			continue;
606 
607 		__arm_lpae_init_pte(data, blk_paddr, pte, lvl, 1, &tablep[i]);
608 	}
609 
610 	pte = arm_lpae_install_table(tablep, ptep, blk_pte, data);
611 	if (pte != blk_pte) {
612 		__arm_lpae_free_pages(tablep, tablesz, cfg, data->iop.cookie);
613 		/*
614 		 * We may race against someone unmapping another part of this
615 		 * block, but anything else is invalid. We can't misinterpret
616 		 * a page entry here since we're never at the last level.
617 		 */
618 		if (iopte_type(pte) != ARM_LPAE_PTE_TYPE_TABLE)
619 			return 0;
620 
621 		tablep = iopte_deref(pte, data);
622 	} else if (unmap_idx_start >= 0) {
623 		for (i = 0; i < num_entries; i++)
624 			io_pgtable_tlb_add_page(&data->iop, gather, iova + i * size, size);
625 
626 		return num_entries * size;
627 	}
628 
629 	return __arm_lpae_unmap(data, gather, iova, size, pgcount, lvl, tablep);
630 }
631 
632 static size_t __arm_lpae_unmap(struct arm_lpae_io_pgtable *data,
633 			       struct iommu_iotlb_gather *gather,
634 			       unsigned long iova, size_t size, size_t pgcount,
635 			       int lvl, arm_lpae_iopte *ptep)
636 {
637 	arm_lpae_iopte pte;
638 	struct io_pgtable *iop = &data->iop;
639 	int i = 0, num_entries, max_entries, unmap_idx_start;
640 
641 	/* Something went horribly wrong and we ran out of page table */
642 	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
643 		return 0;
644 
645 	unmap_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
646 	ptep += unmap_idx_start;
647 	pte = READ_ONCE(*ptep);
648 	if (WARN_ON(!pte))
649 		return 0;
650 
651 	/* If the size matches this level, we're in the right place */
652 	if (size == ARM_LPAE_BLOCK_SIZE(lvl, data)) {
653 		max_entries = ARM_LPAE_PTES_PER_TABLE(data) - unmap_idx_start;
654 		num_entries = min_t(int, pgcount, max_entries);
655 
656 		/* Find and handle non-leaf entries */
657 		for (i = 0; i < num_entries; i++) {
658 			pte = READ_ONCE(ptep[i]);
659 			if (WARN_ON(!pte))
660 				break;
661 
662 			if (!iopte_leaf(pte, lvl, iop->fmt)) {
663 				__arm_lpae_clear_pte(&ptep[i], &iop->cfg, 1);
664 
665 				/* Also flush any partial walks */
666 				io_pgtable_tlb_flush_walk(iop, iova + i * size, size,
667 							  ARM_LPAE_GRANULE(data));
668 				__arm_lpae_free_pgtable(data, lvl + 1, iopte_deref(pte, data));
669 			}
670 		}
671 
672 		/* Clear the remaining entries */
673 		__arm_lpae_clear_pte(ptep, &iop->cfg, i);
674 
675 		if (gather && !iommu_iotlb_gather_queued(gather))
676 			for (int j = 0; j < i; j++)
677 				io_pgtable_tlb_add_page(iop, gather, iova + j * size, size);
678 
679 		return i * size;
680 	} else if (iopte_leaf(pte, lvl, iop->fmt)) {
681 		/*
682 		 * Insert a table at the next level to map the old region,
683 		 * minus the part we want to unmap
684 		 */
685 		return arm_lpae_split_blk_unmap(data, gather, iova, size, pte,
686 						lvl + 1, ptep, pgcount);
687 	}
688 
689 	/* Keep on walkin' */
690 	ptep = iopte_deref(pte, data);
691 	return __arm_lpae_unmap(data, gather, iova, size, pgcount, lvl + 1, ptep);
692 }
693 
694 static size_t arm_lpae_unmap_pages(struct io_pgtable_ops *ops, unsigned long iova,
695 				   size_t pgsize, size_t pgcount,
696 				   struct iommu_iotlb_gather *gather)
697 {
698 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
699 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
700 	arm_lpae_iopte *ptep = data->pgd;
701 	long iaext = (s64)iova >> cfg->ias;
702 
703 	if (WARN_ON(!pgsize || (pgsize & cfg->pgsize_bitmap) != pgsize || !pgcount))
704 		return 0;
705 
706 	if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_TTBR1)
707 		iaext = ~iaext;
708 	if (WARN_ON(iaext))
709 		return 0;
710 
711 	return __arm_lpae_unmap(data, gather, iova, pgsize, pgcount,
712 				data->start_level, ptep);
713 }
714 
715 static phys_addr_t arm_lpae_iova_to_phys(struct io_pgtable_ops *ops,
716 					 unsigned long iova)
717 {
718 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
719 	arm_lpae_iopte pte, *ptep = data->pgd;
720 	int lvl = data->start_level;
721 
722 	do {
723 		/* Valid IOPTE pointer? */
724 		if (!ptep)
725 			return 0;
726 
727 		/* Grab the IOPTE we're interested in */
728 		ptep += ARM_LPAE_LVL_IDX(iova, lvl, data);
729 		pte = READ_ONCE(*ptep);
730 
731 		/* Valid entry? */
732 		if (!pte)
733 			return 0;
734 
735 		/* Leaf entry? */
736 		if (iopte_leaf(pte, lvl, data->iop.fmt))
737 			goto found_translation;
738 
739 		/* Take it to the next level */
740 		ptep = iopte_deref(pte, data);
741 	} while (++lvl < ARM_LPAE_MAX_LEVELS);
742 
743 	/* Ran out of page tables to walk */
744 	return 0;
745 
746 found_translation:
747 	iova &= (ARM_LPAE_BLOCK_SIZE(lvl, data) - 1);
748 	return iopte_to_paddr(pte, data) | iova;
749 }
750 
751 struct io_pgtable_walk_data {
752 	struct iommu_dirty_bitmap	*dirty;
753 	unsigned long			flags;
754 	u64				addr;
755 	const u64			end;
756 };
757 
758 static int __arm_lpae_iopte_walk_dirty(struct arm_lpae_io_pgtable *data,
759 				       struct io_pgtable_walk_data *walk_data,
760 				       arm_lpae_iopte *ptep,
761 				       int lvl);
762 
763 static int io_pgtable_visit_dirty(struct arm_lpae_io_pgtable *data,
764 				  struct io_pgtable_walk_data *walk_data,
765 				  arm_lpae_iopte *ptep, int lvl)
766 {
767 	struct io_pgtable *iop = &data->iop;
768 	arm_lpae_iopte pte = READ_ONCE(*ptep);
769 
770 	if (iopte_leaf(pte, lvl, iop->fmt)) {
771 		size_t size = ARM_LPAE_BLOCK_SIZE(lvl, data);
772 
773 		if (iopte_writeable_dirty(pte)) {
774 			iommu_dirty_bitmap_record(walk_data->dirty,
775 						  walk_data->addr, size);
776 			if (!(walk_data->flags & IOMMU_DIRTY_NO_CLEAR))
777 				iopte_set_writeable_clean(ptep);
778 		}
779 		walk_data->addr += size;
780 		return 0;
781 	}
782 
783 	if (WARN_ON(!iopte_table(pte, lvl)))
784 		return -EINVAL;
785 
786 	ptep = iopte_deref(pte, data);
787 	return __arm_lpae_iopte_walk_dirty(data, walk_data, ptep, lvl + 1);
788 }
789 
790 static int __arm_lpae_iopte_walk_dirty(struct arm_lpae_io_pgtable *data,
791 				       struct io_pgtable_walk_data *walk_data,
792 				       arm_lpae_iopte *ptep,
793 				       int lvl)
794 {
795 	u32 idx;
796 	int max_entries, ret;
797 
798 	if (WARN_ON(lvl == ARM_LPAE_MAX_LEVELS))
799 		return -EINVAL;
800 
801 	if (lvl == data->start_level)
802 		max_entries = ARM_LPAE_PGD_SIZE(data) / sizeof(arm_lpae_iopte);
803 	else
804 		max_entries = ARM_LPAE_PTES_PER_TABLE(data);
805 
806 	for (idx = ARM_LPAE_LVL_IDX(walk_data->addr, lvl, data);
807 	     (idx < max_entries) && (walk_data->addr < walk_data->end); ++idx) {
808 		ret = io_pgtable_visit_dirty(data, walk_data, ptep + idx, lvl);
809 		if (ret)
810 			return ret;
811 	}
812 
813 	return 0;
814 }
815 
816 static int arm_lpae_read_and_clear_dirty(struct io_pgtable_ops *ops,
817 					 unsigned long iova, size_t size,
818 					 unsigned long flags,
819 					 struct iommu_dirty_bitmap *dirty)
820 {
821 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
822 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
823 	struct io_pgtable_walk_data walk_data = {
824 		.dirty = dirty,
825 		.flags = flags,
826 		.addr = iova,
827 		.end = iova + size,
828 	};
829 	arm_lpae_iopte *ptep = data->pgd;
830 	int lvl = data->start_level;
831 
832 	if (WARN_ON(!size))
833 		return -EINVAL;
834 	if (WARN_ON((iova + size - 1) & ~(BIT(cfg->ias) - 1)))
835 		return -EINVAL;
836 	if (data->iop.fmt != ARM_64_LPAE_S1)
837 		return -EINVAL;
838 
839 	return __arm_lpae_iopte_walk_dirty(data, &walk_data, ptep, lvl);
840 }
841 
842 static void arm_lpae_restrict_pgsizes(struct io_pgtable_cfg *cfg)
843 {
844 	unsigned long granule, page_sizes;
845 	unsigned int max_addr_bits = 48;
846 
847 	/*
848 	 * We need to restrict the supported page sizes to match the
849 	 * translation regime for a particular granule. Aim to match
850 	 * the CPU page size if possible, otherwise prefer smaller sizes.
851 	 * While we're at it, restrict the block sizes to match the
852 	 * chosen granule.
853 	 */
854 	if (cfg->pgsize_bitmap & PAGE_SIZE)
855 		granule = PAGE_SIZE;
856 	else if (cfg->pgsize_bitmap & ~PAGE_MASK)
857 		granule = 1UL << __fls(cfg->pgsize_bitmap & ~PAGE_MASK);
858 	else if (cfg->pgsize_bitmap & PAGE_MASK)
859 		granule = 1UL << __ffs(cfg->pgsize_bitmap & PAGE_MASK);
860 	else
861 		granule = 0;
862 
863 	switch (granule) {
864 	case SZ_4K:
865 		page_sizes = (SZ_4K | SZ_2M | SZ_1G);
866 		break;
867 	case SZ_16K:
868 		page_sizes = (SZ_16K | SZ_32M);
869 		break;
870 	case SZ_64K:
871 		max_addr_bits = 52;
872 		page_sizes = (SZ_64K | SZ_512M);
873 		if (cfg->oas > 48)
874 			page_sizes |= 1ULL << 42; /* 4TB */
875 		break;
876 	default:
877 		page_sizes = 0;
878 	}
879 
880 	cfg->pgsize_bitmap &= page_sizes;
881 	cfg->ias = min(cfg->ias, max_addr_bits);
882 	cfg->oas = min(cfg->oas, max_addr_bits);
883 }
884 
885 static struct arm_lpae_io_pgtable *
886 arm_lpae_alloc_pgtable(struct io_pgtable_cfg *cfg)
887 {
888 	struct arm_lpae_io_pgtable *data;
889 	int levels, va_bits, pg_shift;
890 
891 	arm_lpae_restrict_pgsizes(cfg);
892 
893 	if (!(cfg->pgsize_bitmap & (SZ_4K | SZ_16K | SZ_64K)))
894 		return NULL;
895 
896 	if (cfg->ias > ARM_LPAE_MAX_ADDR_BITS)
897 		return NULL;
898 
899 	if (cfg->oas > ARM_LPAE_MAX_ADDR_BITS)
900 		return NULL;
901 
902 	data = kmalloc(sizeof(*data), GFP_KERNEL);
903 	if (!data)
904 		return NULL;
905 
906 	pg_shift = __ffs(cfg->pgsize_bitmap);
907 	data->bits_per_level = pg_shift - ilog2(sizeof(arm_lpae_iopte));
908 
909 	va_bits = cfg->ias - pg_shift;
910 	levels = DIV_ROUND_UP(va_bits, data->bits_per_level);
911 	data->start_level = ARM_LPAE_MAX_LEVELS - levels;
912 
913 	/* Calculate the actual size of our pgd (without concatenation) */
914 	data->pgd_bits = va_bits - (data->bits_per_level * (levels - 1));
915 
916 	data->iop.ops = (struct io_pgtable_ops) {
917 		.map_pages	= arm_lpae_map_pages,
918 		.unmap_pages	= arm_lpae_unmap_pages,
919 		.iova_to_phys	= arm_lpae_iova_to_phys,
920 		.read_and_clear_dirty = arm_lpae_read_and_clear_dirty,
921 	};
922 
923 	return data;
924 }
925 
926 static struct io_pgtable *
927 arm_64_lpae_alloc_pgtable_s1(struct io_pgtable_cfg *cfg, void *cookie)
928 {
929 	u64 reg;
930 	struct arm_lpae_io_pgtable *data;
931 	typeof(&cfg->arm_lpae_s1_cfg.tcr) tcr = &cfg->arm_lpae_s1_cfg.tcr;
932 	bool tg1;
933 
934 	if (cfg->quirks & ~(IO_PGTABLE_QUIRK_ARM_NS |
935 			    IO_PGTABLE_QUIRK_ARM_TTBR1 |
936 			    IO_PGTABLE_QUIRK_ARM_OUTER_WBWA |
937 			    IO_PGTABLE_QUIRK_ARM_HD))
938 		return NULL;
939 
940 	data = arm_lpae_alloc_pgtable(cfg);
941 	if (!data)
942 		return NULL;
943 
944 	/* TCR */
945 	if (cfg->coherent_walk) {
946 		tcr->sh = ARM_LPAE_TCR_SH_IS;
947 		tcr->irgn = ARM_LPAE_TCR_RGN_WBWA;
948 		tcr->orgn = ARM_LPAE_TCR_RGN_WBWA;
949 		if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_OUTER_WBWA)
950 			goto out_free_data;
951 	} else {
952 		tcr->sh = ARM_LPAE_TCR_SH_OS;
953 		tcr->irgn = ARM_LPAE_TCR_RGN_NC;
954 		if (!(cfg->quirks & IO_PGTABLE_QUIRK_ARM_OUTER_WBWA))
955 			tcr->orgn = ARM_LPAE_TCR_RGN_NC;
956 		else
957 			tcr->orgn = ARM_LPAE_TCR_RGN_WBWA;
958 	}
959 
960 	tg1 = cfg->quirks & IO_PGTABLE_QUIRK_ARM_TTBR1;
961 	switch (ARM_LPAE_GRANULE(data)) {
962 	case SZ_4K:
963 		tcr->tg = tg1 ? ARM_LPAE_TCR_TG1_4K : ARM_LPAE_TCR_TG0_4K;
964 		break;
965 	case SZ_16K:
966 		tcr->tg = tg1 ? ARM_LPAE_TCR_TG1_16K : ARM_LPAE_TCR_TG0_16K;
967 		break;
968 	case SZ_64K:
969 		tcr->tg = tg1 ? ARM_LPAE_TCR_TG1_64K : ARM_LPAE_TCR_TG0_64K;
970 		break;
971 	}
972 
973 	switch (cfg->oas) {
974 	case 32:
975 		tcr->ips = ARM_LPAE_TCR_PS_32_BIT;
976 		break;
977 	case 36:
978 		tcr->ips = ARM_LPAE_TCR_PS_36_BIT;
979 		break;
980 	case 40:
981 		tcr->ips = ARM_LPAE_TCR_PS_40_BIT;
982 		break;
983 	case 42:
984 		tcr->ips = ARM_LPAE_TCR_PS_42_BIT;
985 		break;
986 	case 44:
987 		tcr->ips = ARM_LPAE_TCR_PS_44_BIT;
988 		break;
989 	case 48:
990 		tcr->ips = ARM_LPAE_TCR_PS_48_BIT;
991 		break;
992 	case 52:
993 		tcr->ips = ARM_LPAE_TCR_PS_52_BIT;
994 		break;
995 	default:
996 		goto out_free_data;
997 	}
998 
999 	tcr->tsz = 64ULL - cfg->ias;
1000 
1001 	/* MAIRs */
1002 	reg = (ARM_LPAE_MAIR_ATTR_NC
1003 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_NC)) |
1004 	      (ARM_LPAE_MAIR_ATTR_WBRWA
1005 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_CACHE)) |
1006 	      (ARM_LPAE_MAIR_ATTR_DEVICE
1007 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_DEV)) |
1008 	      (ARM_LPAE_MAIR_ATTR_INC_OWBRWA
1009 	       << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_INC_OCACHE));
1010 
1011 	cfg->arm_lpae_s1_cfg.mair = reg;
1012 
1013 	/* Looking good; allocate a pgd */
1014 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data),
1015 					   GFP_KERNEL, cfg, cookie);
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 	/* TTBR */
1023 	cfg->arm_lpae_s1_cfg.ttbr = 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_64_lpae_alloc_pgtable_s2(struct io_pgtable_cfg *cfg, void *cookie)
1033 {
1034 	u64 sl;
1035 	struct arm_lpae_io_pgtable *data;
1036 	typeof(&cfg->arm_lpae_s2_cfg.vtcr) vtcr = &cfg->arm_lpae_s2_cfg.vtcr;
1037 
1038 	/* The NS quirk doesn't apply at stage 2 */
1039 	if (cfg->quirks)
1040 		return NULL;
1041 
1042 	data = arm_lpae_alloc_pgtable(cfg);
1043 	if (!data)
1044 		return NULL;
1045 
1046 	/*
1047 	 * Concatenate PGDs at level 1 if possible in order to reduce
1048 	 * the depth of the stage-2 walk.
1049 	 */
1050 	if (data->start_level == 0) {
1051 		unsigned long pgd_pages;
1052 
1053 		pgd_pages = ARM_LPAE_PGD_SIZE(data) / sizeof(arm_lpae_iopte);
1054 		if (pgd_pages <= ARM_LPAE_S2_MAX_CONCAT_PAGES) {
1055 			data->pgd_bits += data->bits_per_level;
1056 			data->start_level++;
1057 		}
1058 	}
1059 
1060 	/* VTCR */
1061 	if (cfg->coherent_walk) {
1062 		vtcr->sh = ARM_LPAE_TCR_SH_IS;
1063 		vtcr->irgn = ARM_LPAE_TCR_RGN_WBWA;
1064 		vtcr->orgn = ARM_LPAE_TCR_RGN_WBWA;
1065 	} else {
1066 		vtcr->sh = ARM_LPAE_TCR_SH_OS;
1067 		vtcr->irgn = ARM_LPAE_TCR_RGN_NC;
1068 		vtcr->orgn = ARM_LPAE_TCR_RGN_NC;
1069 	}
1070 
1071 	sl = data->start_level;
1072 
1073 	switch (ARM_LPAE_GRANULE(data)) {
1074 	case SZ_4K:
1075 		vtcr->tg = ARM_LPAE_TCR_TG0_4K;
1076 		sl++; /* SL0 format is different for 4K granule size */
1077 		break;
1078 	case SZ_16K:
1079 		vtcr->tg = ARM_LPAE_TCR_TG0_16K;
1080 		break;
1081 	case SZ_64K:
1082 		vtcr->tg = ARM_LPAE_TCR_TG0_64K;
1083 		break;
1084 	}
1085 
1086 	switch (cfg->oas) {
1087 	case 32:
1088 		vtcr->ps = ARM_LPAE_TCR_PS_32_BIT;
1089 		break;
1090 	case 36:
1091 		vtcr->ps = ARM_LPAE_TCR_PS_36_BIT;
1092 		break;
1093 	case 40:
1094 		vtcr->ps = ARM_LPAE_TCR_PS_40_BIT;
1095 		break;
1096 	case 42:
1097 		vtcr->ps = ARM_LPAE_TCR_PS_42_BIT;
1098 		break;
1099 	case 44:
1100 		vtcr->ps = ARM_LPAE_TCR_PS_44_BIT;
1101 		break;
1102 	case 48:
1103 		vtcr->ps = ARM_LPAE_TCR_PS_48_BIT;
1104 		break;
1105 	case 52:
1106 		vtcr->ps = ARM_LPAE_TCR_PS_52_BIT;
1107 		break;
1108 	default:
1109 		goto out_free_data;
1110 	}
1111 
1112 	vtcr->tsz = 64ULL - cfg->ias;
1113 	vtcr->sl = ~sl & ARM_LPAE_VTCR_SL0_MASK;
1114 
1115 	/* Allocate pgd pages */
1116 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data),
1117 					   GFP_KERNEL, cfg, cookie);
1118 	if (!data->pgd)
1119 		goto out_free_data;
1120 
1121 	/* Ensure the empty pgd is visible before any actual TTBR write */
1122 	wmb();
1123 
1124 	/* VTTBR */
1125 	cfg->arm_lpae_s2_cfg.vttbr = virt_to_phys(data->pgd);
1126 	return &data->iop;
1127 
1128 out_free_data:
1129 	kfree(data);
1130 	return NULL;
1131 }
1132 
1133 static struct io_pgtable *
1134 arm_32_lpae_alloc_pgtable_s1(struct io_pgtable_cfg *cfg, void *cookie)
1135 {
1136 	if (cfg->ias > 32 || cfg->oas > 40)
1137 		return NULL;
1138 
1139 	cfg->pgsize_bitmap &= (SZ_4K | SZ_2M | SZ_1G);
1140 	return arm_64_lpae_alloc_pgtable_s1(cfg, cookie);
1141 }
1142 
1143 static struct io_pgtable *
1144 arm_32_lpae_alloc_pgtable_s2(struct io_pgtable_cfg *cfg, void *cookie)
1145 {
1146 	if (cfg->ias > 40 || cfg->oas > 40)
1147 		return NULL;
1148 
1149 	cfg->pgsize_bitmap &= (SZ_4K | SZ_2M | SZ_1G);
1150 	return arm_64_lpae_alloc_pgtable_s2(cfg, cookie);
1151 }
1152 
1153 static struct io_pgtable *
1154 arm_mali_lpae_alloc_pgtable(struct io_pgtable_cfg *cfg, void *cookie)
1155 {
1156 	struct arm_lpae_io_pgtable *data;
1157 
1158 	/* No quirks for Mali (hopefully) */
1159 	if (cfg->quirks)
1160 		return NULL;
1161 
1162 	if (cfg->ias > 48 || cfg->oas > 40)
1163 		return NULL;
1164 
1165 	cfg->pgsize_bitmap &= (SZ_4K | SZ_2M | SZ_1G);
1166 
1167 	data = arm_lpae_alloc_pgtable(cfg);
1168 	if (!data)
1169 		return NULL;
1170 
1171 	/* Mali seems to need a full 4-level table regardless of IAS */
1172 	if (data->start_level > 0) {
1173 		data->start_level = 0;
1174 		data->pgd_bits = 0;
1175 	}
1176 	/*
1177 	 * MEMATTR: Mali has no actual notion of a non-cacheable type, so the
1178 	 * best we can do is mimic the out-of-tree driver and hope that the
1179 	 * "implementation-defined caching policy" is good enough. Similarly,
1180 	 * we'll use it for the sake of a valid attribute for our 'device'
1181 	 * index, although callers should never request that in practice.
1182 	 */
1183 	cfg->arm_mali_lpae_cfg.memattr =
1184 		(ARM_MALI_LPAE_MEMATTR_IMP_DEF
1185 		 << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_NC)) |
1186 		(ARM_MALI_LPAE_MEMATTR_WRITE_ALLOC
1187 		 << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_CACHE)) |
1188 		(ARM_MALI_LPAE_MEMATTR_IMP_DEF
1189 		 << ARM_LPAE_MAIR_ATTR_SHIFT(ARM_LPAE_MAIR_ATTR_IDX_DEV));
1190 
1191 	data->pgd = __arm_lpae_alloc_pages(ARM_LPAE_PGD_SIZE(data), GFP_KERNEL,
1192 					   cfg, cookie);
1193 	if (!data->pgd)
1194 		goto out_free_data;
1195 
1196 	/* Ensure the empty pgd is visible before TRANSTAB can be written */
1197 	wmb();
1198 
1199 	cfg->arm_mali_lpae_cfg.transtab = virt_to_phys(data->pgd) |
1200 					  ARM_MALI_LPAE_TTBR_READ_INNER |
1201 					  ARM_MALI_LPAE_TTBR_ADRMODE_TABLE;
1202 	if (cfg->coherent_walk)
1203 		cfg->arm_mali_lpae_cfg.transtab |= ARM_MALI_LPAE_TTBR_SHARE_OUTER;
1204 
1205 	return &data->iop;
1206 
1207 out_free_data:
1208 	kfree(data);
1209 	return NULL;
1210 }
1211 
1212 struct io_pgtable_init_fns io_pgtable_arm_64_lpae_s1_init_fns = {
1213 	.caps	= IO_PGTABLE_CAP_CUSTOM_ALLOCATOR,
1214 	.alloc	= arm_64_lpae_alloc_pgtable_s1,
1215 	.free	= arm_lpae_free_pgtable,
1216 };
1217 
1218 struct io_pgtable_init_fns io_pgtable_arm_64_lpae_s2_init_fns = {
1219 	.caps	= IO_PGTABLE_CAP_CUSTOM_ALLOCATOR,
1220 	.alloc	= arm_64_lpae_alloc_pgtable_s2,
1221 	.free	= arm_lpae_free_pgtable,
1222 };
1223 
1224 struct io_pgtable_init_fns io_pgtable_arm_32_lpae_s1_init_fns = {
1225 	.caps	= IO_PGTABLE_CAP_CUSTOM_ALLOCATOR,
1226 	.alloc	= arm_32_lpae_alloc_pgtable_s1,
1227 	.free	= arm_lpae_free_pgtable,
1228 };
1229 
1230 struct io_pgtable_init_fns io_pgtable_arm_32_lpae_s2_init_fns = {
1231 	.caps	= IO_PGTABLE_CAP_CUSTOM_ALLOCATOR,
1232 	.alloc	= arm_32_lpae_alloc_pgtable_s2,
1233 	.free	= arm_lpae_free_pgtable,
1234 };
1235 
1236 struct io_pgtable_init_fns io_pgtable_arm_mali_lpae_init_fns = {
1237 	.caps	= IO_PGTABLE_CAP_CUSTOM_ALLOCATOR,
1238 	.alloc	= arm_mali_lpae_alloc_pgtable,
1239 	.free	= arm_lpae_free_pgtable,
1240 };
1241 
1242 #ifdef CONFIG_IOMMU_IO_PGTABLE_LPAE_SELFTEST
1243 
1244 static struct io_pgtable_cfg *cfg_cookie __initdata;
1245 
1246 static void __init dummy_tlb_flush_all(void *cookie)
1247 {
1248 	WARN_ON(cookie != cfg_cookie);
1249 }
1250 
1251 static void __init dummy_tlb_flush(unsigned long iova, size_t size,
1252 				   size_t granule, void *cookie)
1253 {
1254 	WARN_ON(cookie != cfg_cookie);
1255 	WARN_ON(!(size & cfg_cookie->pgsize_bitmap));
1256 }
1257 
1258 static void __init dummy_tlb_add_page(struct iommu_iotlb_gather *gather,
1259 				      unsigned long iova, size_t granule,
1260 				      void *cookie)
1261 {
1262 	dummy_tlb_flush(iova, granule, granule, cookie);
1263 }
1264 
1265 static const struct iommu_flush_ops dummy_tlb_ops __initconst = {
1266 	.tlb_flush_all	= dummy_tlb_flush_all,
1267 	.tlb_flush_walk	= dummy_tlb_flush,
1268 	.tlb_add_page	= dummy_tlb_add_page,
1269 };
1270 
1271 static void __init arm_lpae_dump_ops(struct io_pgtable_ops *ops)
1272 {
1273 	struct arm_lpae_io_pgtable *data = io_pgtable_ops_to_data(ops);
1274 	struct io_pgtable_cfg *cfg = &data->iop.cfg;
1275 
1276 	pr_err("cfg: pgsize_bitmap 0x%lx, ias %u-bit\n",
1277 		cfg->pgsize_bitmap, cfg->ias);
1278 	pr_err("data: %d levels, 0x%zx pgd_size, %u pg_shift, %u bits_per_level, pgd @ %p\n",
1279 		ARM_LPAE_MAX_LEVELS - data->start_level, ARM_LPAE_PGD_SIZE(data),
1280 		ilog2(ARM_LPAE_GRANULE(data)), data->bits_per_level, data->pgd);
1281 }
1282 
1283 #define __FAIL(ops, i)	({						\
1284 		WARN(1, "selftest: test failed for fmt idx %d\n", (i));	\
1285 		arm_lpae_dump_ops(ops);					\
1286 		selftest_running = false;				\
1287 		-EFAULT;						\
1288 })
1289 
1290 static int __init arm_lpae_run_tests(struct io_pgtable_cfg *cfg)
1291 {
1292 	static const enum io_pgtable_fmt fmts[] __initconst = {
1293 		ARM_64_LPAE_S1,
1294 		ARM_64_LPAE_S2,
1295 	};
1296 
1297 	int i, j;
1298 	unsigned long iova;
1299 	size_t size, mapped;
1300 	struct io_pgtable_ops *ops;
1301 
1302 	selftest_running = true;
1303 
1304 	for (i = 0; i < ARRAY_SIZE(fmts); ++i) {
1305 		cfg_cookie = cfg;
1306 		ops = alloc_io_pgtable_ops(fmts[i], cfg, cfg);
1307 		if (!ops) {
1308 			pr_err("selftest: failed to allocate io pgtable ops\n");
1309 			return -ENOMEM;
1310 		}
1311 
1312 		/*
1313 		 * Initial sanity checks.
1314 		 * Empty page tables shouldn't provide any translations.
1315 		 */
1316 		if (ops->iova_to_phys(ops, 42))
1317 			return __FAIL(ops, i);
1318 
1319 		if (ops->iova_to_phys(ops, SZ_1G + 42))
1320 			return __FAIL(ops, i);
1321 
1322 		if (ops->iova_to_phys(ops, SZ_2G + 42))
1323 			return __FAIL(ops, i);
1324 
1325 		/*
1326 		 * Distinct mappings of different granule sizes.
1327 		 */
1328 		iova = 0;
1329 		for_each_set_bit(j, &cfg->pgsize_bitmap, BITS_PER_LONG) {
1330 			size = 1UL << j;
1331 
1332 			if (ops->map_pages(ops, iova, iova, size, 1,
1333 					   IOMMU_READ | IOMMU_WRITE |
1334 					   IOMMU_NOEXEC | IOMMU_CACHE,
1335 					   GFP_KERNEL, &mapped))
1336 				return __FAIL(ops, i);
1337 
1338 			/* Overlapping mappings */
1339 			if (!ops->map_pages(ops, iova, iova + size, size, 1,
1340 					    IOMMU_READ | IOMMU_NOEXEC,
1341 					    GFP_KERNEL, &mapped))
1342 				return __FAIL(ops, i);
1343 
1344 			if (ops->iova_to_phys(ops, iova + 42) != (iova + 42))
1345 				return __FAIL(ops, i);
1346 
1347 			iova += SZ_1G;
1348 		}
1349 
1350 		/* Partial unmap */
1351 		size = 1UL << __ffs(cfg->pgsize_bitmap);
1352 		if (ops->unmap_pages(ops, SZ_1G + size, size, 1, NULL) != size)
1353 			return __FAIL(ops, i);
1354 
1355 		/* Remap of partial unmap */
1356 		if (ops->map_pages(ops, SZ_1G + size, size, size, 1,
1357 				   IOMMU_READ, GFP_KERNEL, &mapped))
1358 			return __FAIL(ops, i);
1359 
1360 		if (ops->iova_to_phys(ops, SZ_1G + size + 42) != (size + 42))
1361 			return __FAIL(ops, i);
1362 
1363 		/* Full unmap */
1364 		iova = 0;
1365 		for_each_set_bit(j, &cfg->pgsize_bitmap, BITS_PER_LONG) {
1366 			size = 1UL << j;
1367 
1368 			if (ops->unmap_pages(ops, iova, size, 1, NULL) != size)
1369 				return __FAIL(ops, i);
1370 
1371 			if (ops->iova_to_phys(ops, iova + 42))
1372 				return __FAIL(ops, i);
1373 
1374 			/* Remap full block */
1375 			if (ops->map_pages(ops, iova, iova, size, 1,
1376 					   IOMMU_WRITE, GFP_KERNEL, &mapped))
1377 				return __FAIL(ops, i);
1378 
1379 			if (ops->iova_to_phys(ops, iova + 42) != (iova + 42))
1380 				return __FAIL(ops, i);
1381 
1382 			iova += SZ_1G;
1383 		}
1384 
1385 		free_io_pgtable_ops(ops);
1386 	}
1387 
1388 	selftest_running = false;
1389 	return 0;
1390 }
1391 
1392 static int __init arm_lpae_do_selftests(void)
1393 {
1394 	static const unsigned long pgsize[] __initconst = {
1395 		SZ_4K | SZ_2M | SZ_1G,
1396 		SZ_16K | SZ_32M,
1397 		SZ_64K | SZ_512M,
1398 	};
1399 
1400 	static const unsigned int ias[] __initconst = {
1401 		32, 36, 40, 42, 44, 48,
1402 	};
1403 
1404 	int i, j, pass = 0, fail = 0;
1405 	struct device dev;
1406 	struct io_pgtable_cfg cfg = {
1407 		.tlb = &dummy_tlb_ops,
1408 		.oas = 48,
1409 		.coherent_walk = true,
1410 		.iommu_dev = &dev,
1411 	};
1412 
1413 	/* __arm_lpae_alloc_pages() merely needs dev_to_node() to work */
1414 	set_dev_node(&dev, NUMA_NO_NODE);
1415 
1416 	for (i = 0; i < ARRAY_SIZE(pgsize); ++i) {
1417 		for (j = 0; j < ARRAY_SIZE(ias); ++j) {
1418 			cfg.pgsize_bitmap = pgsize[i];
1419 			cfg.ias = ias[j];
1420 			pr_info("selftest: pgsize_bitmap 0x%08lx, IAS %u\n",
1421 				pgsize[i], ias[j]);
1422 			if (arm_lpae_run_tests(&cfg))
1423 				fail++;
1424 			else
1425 				pass++;
1426 		}
1427 	}
1428 
1429 	pr_info("selftest: completed with %d PASS %d FAIL\n", pass, fail);
1430 	return fail ? -EFAULT : 0;
1431 }
1432 subsys_initcall(arm_lpae_do_selftests);
1433 #endif
1434