1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * CPU-agnostic ARM page table allocator.
4 *
5 * ARMv7 Short-descriptor format, supporting
6 * - Basic memory attributes
7 * - Simplified access permissions (AP[2:1] model)
8 * - Backwards-compatible TEX remap
9 * - Large pages/supersections (if indicated by the caller)
10 *
11 * Not supporting:
12 * - Legacy access permissions (AP[2:0] model)
13 *
14 * Almost certainly never supporting:
15 * - PXN
16 * - Domains
17 *
18 * Copyright (C) 2014-2015 ARM Limited
19 * Copyright (c) 2014-2015 MediaTek Inc.
20 */
21
22 #define pr_fmt(fmt) "arm-v7s io-pgtable: " fmt
23
24 #include <linux/atomic.h>
25 #include <linux/dma-mapping.h>
26 #include <linux/gfp.h>
27 #include <linux/io-pgtable.h>
28 #include <linux/iommu.h>
29 #include <linux/kernel.h>
30 #include <linux/kmemleak.h>
31 #include <linux/sizes.h>
32 #include <linux/slab.h>
33 #include <linux/spinlock.h>
34 #include <linux/types.h>
35
36 #include <asm/barrier.h>
37
38 /* Struct accessors */
39 #define io_pgtable_to_data(x) \
40 container_of((x), struct arm_v7s_io_pgtable, iop)
41
42 #define io_pgtable_ops_to_data(x) \
43 io_pgtable_to_data(io_pgtable_ops_to_pgtable(x))
44
45 /*
46 * We have 32 bits total; 12 bits resolved at level 1, 8 bits at level 2,
47 * and 12 bits in a page. With some carefully-chosen coefficients we can
48 * hide the ugly inconsistencies behind these macros and at least let the
49 * rest of the code pretend to be somewhat sane.
50 */
51 #define ARM_V7S_ADDR_BITS 32
52 #define _ARM_V7S_LVL_BITS(lvl) (16 - (lvl) * 4)
53 #define ARM_V7S_LVL_SHIFT(lvl) (ARM_V7S_ADDR_BITS - (4 + 8 * (lvl)))
54 #define ARM_V7S_TABLE_SHIFT 10
55
56 #define ARM_V7S_PTES_PER_LVL(lvl) (1 << _ARM_V7S_LVL_BITS(lvl))
57 #define ARM_V7S_TABLE_SIZE(lvl) \
58 (ARM_V7S_PTES_PER_LVL(lvl) * sizeof(arm_v7s_iopte))
59
60 #define ARM_V7S_BLOCK_SIZE(lvl) (1UL << ARM_V7S_LVL_SHIFT(lvl))
61 #define ARM_V7S_LVL_MASK(lvl) ((u32)(~0U << ARM_V7S_LVL_SHIFT(lvl)))
62 #define ARM_V7S_TABLE_MASK ((u32)(~0U << ARM_V7S_TABLE_SHIFT))
63 #define _ARM_V7S_IDX_MASK(lvl) (ARM_V7S_PTES_PER_LVL(lvl) - 1)
64 #define ARM_V7S_LVL_IDX(addr, lvl) ({ \
65 int _l = lvl; \
66 ((u32)(addr) >> ARM_V7S_LVL_SHIFT(_l)) & _ARM_V7S_IDX_MASK(_l); \
67 })
68
69 /*
70 * Large page/supersection entries are effectively a block of 16 page/section
71 * entries, along the lines of the LPAE contiguous hint, but all with the
72 * same output address. For want of a better common name we'll call them
73 * "contiguous" versions of their respective page/section entries here, but
74 * noting the distinction (WRT to TLB maintenance) that they represent *one*
75 * entry repeated 16 times, not 16 separate entries (as in the LPAE case).
76 */
77 #define ARM_V7S_CONT_PAGES 16
78
79 /* PTE type bits: these are all mixed up with XN/PXN bits in most cases */
80 #define ARM_V7S_PTE_TYPE_TABLE 0x1
81 #define ARM_V7S_PTE_TYPE_PAGE 0x2
82 #define ARM_V7S_PTE_TYPE_CONT_PAGE 0x1
83
84 #define ARM_V7S_PTE_IS_VALID(pte) (((pte) & 0x3) != 0)
85 #define ARM_V7S_PTE_IS_TABLE(pte, lvl) \
86 ((lvl) == 1 && (((pte) & 0x3) == ARM_V7S_PTE_TYPE_TABLE))
87
88 /* Page table bits */
89 #define ARM_V7S_ATTR_XN(lvl) BIT(4 * (2 - (lvl)))
90 #define ARM_V7S_ATTR_B BIT(2)
91 #define ARM_V7S_ATTR_C BIT(3)
92 #define ARM_V7S_ATTR_NS_TABLE BIT(3)
93 #define ARM_V7S_ATTR_NS_SECTION BIT(19)
94
95 #define ARM_V7S_CONT_SECTION BIT(18)
96 #define ARM_V7S_CONT_PAGE_XN_SHIFT 15
97
98 /*
99 * The attribute bits are consistently ordered*, but occupy bits [17:10] of
100 * a level 1 PTE vs. bits [11:4] at level 2. Thus we define the individual
101 * fields relative to that 8-bit block, plus a total shift relative to the PTE.
102 */
103 #define ARM_V7S_ATTR_SHIFT(lvl) (16 - (lvl) * 6)
104
105 #define ARM_V7S_ATTR_MASK 0xff
106 #define ARM_V7S_ATTR_AP0 BIT(0)
107 #define ARM_V7S_ATTR_AP1 BIT(1)
108 #define ARM_V7S_ATTR_AP2 BIT(5)
109 #define ARM_V7S_ATTR_S BIT(6)
110 #define ARM_V7S_ATTR_NG BIT(7)
111 #define ARM_V7S_TEX_SHIFT 2
112 #define ARM_V7S_TEX_MASK 0x7
113 #define ARM_V7S_ATTR_TEX(val) (((val) & ARM_V7S_TEX_MASK) << ARM_V7S_TEX_SHIFT)
114
115 /* MediaTek extend the two bits for PA 32bit/33bit */
116 #define ARM_V7S_ATTR_MTK_PA_BIT32 BIT(9)
117 #define ARM_V7S_ATTR_MTK_PA_BIT33 BIT(4)
118
119 /* *well, except for TEX on level 2 large pages, of course :( */
120 #define ARM_V7S_CONT_PAGE_TEX_SHIFT 6
121 #define ARM_V7S_CONT_PAGE_TEX_MASK (ARM_V7S_TEX_MASK << ARM_V7S_CONT_PAGE_TEX_SHIFT)
122
123 /* Simplified access permissions */
124 #define ARM_V7S_PTE_AF ARM_V7S_ATTR_AP0
125 #define ARM_V7S_PTE_AP_UNPRIV ARM_V7S_ATTR_AP1
126 #define ARM_V7S_PTE_AP_RDONLY ARM_V7S_ATTR_AP2
127
128 /* Register bits */
129 #define ARM_V7S_RGN_NC 0
130 #define ARM_V7S_RGN_WBWA 1
131 #define ARM_V7S_RGN_WT 2
132 #define ARM_V7S_RGN_WB 3
133
134 #define ARM_V7S_PRRR_TYPE_DEVICE 1
135 #define ARM_V7S_PRRR_TYPE_NORMAL 2
136 #define ARM_V7S_PRRR_TR(n, type) (((type) & 0x3) << ((n) * 2))
137 #define ARM_V7S_PRRR_DS0 BIT(16)
138 #define ARM_V7S_PRRR_DS1 BIT(17)
139 #define ARM_V7S_PRRR_NS0 BIT(18)
140 #define ARM_V7S_PRRR_NS1 BIT(19)
141 #define ARM_V7S_PRRR_NOS(n) BIT((n) + 24)
142
143 #define ARM_V7S_NMRR_IR(n, attr) (((attr) & 0x3) << ((n) * 2))
144 #define ARM_V7S_NMRR_OR(n, attr) (((attr) & 0x3) << ((n) * 2 + 16))
145
146 #define ARM_V7S_TTBR_S BIT(1)
147 #define ARM_V7S_TTBR_NOS BIT(5)
148 #define ARM_V7S_TTBR_ORGN_ATTR(attr) (((attr) & 0x3) << 3)
149 #define ARM_V7S_TTBR_IRGN_ATTR(attr) \
150 ((((attr) & 0x1) << 6) | (((attr) & 0x2) >> 1))
151
152 #ifdef CONFIG_ZONE_DMA32
153 #define ARM_V7S_TABLE_GFP_DMA GFP_DMA32
154 #define ARM_V7S_TABLE_SLAB_FLAGS SLAB_CACHE_DMA32
155 #else
156 #define ARM_V7S_TABLE_GFP_DMA GFP_DMA
157 #define ARM_V7S_TABLE_SLAB_FLAGS SLAB_CACHE_DMA
158 #endif
159
160 typedef u32 arm_v7s_iopte;
161
162 static bool selftest_running;
163
164 struct arm_v7s_io_pgtable {
165 struct io_pgtable iop;
166
167 arm_v7s_iopte *pgd;
168 struct kmem_cache *l2_tables;
169 spinlock_t split_lock;
170 };
171
172 static bool arm_v7s_pte_is_cont(arm_v7s_iopte pte, int lvl);
173
__arm_v7s_dma_addr(void * pages)174 static dma_addr_t __arm_v7s_dma_addr(void *pages)
175 {
176 return (dma_addr_t)virt_to_phys(pages);
177 }
178
arm_v7s_is_mtk_enabled(struct io_pgtable_cfg * cfg)179 static bool arm_v7s_is_mtk_enabled(struct io_pgtable_cfg *cfg)
180 {
181 return IS_ENABLED(CONFIG_PHYS_ADDR_T_64BIT) &&
182 (cfg->quirks & IO_PGTABLE_QUIRK_ARM_MTK_EXT);
183 }
184
paddr_to_iopte(phys_addr_t paddr,int lvl,struct io_pgtable_cfg * cfg)185 static arm_v7s_iopte paddr_to_iopte(phys_addr_t paddr, int lvl,
186 struct io_pgtable_cfg *cfg)
187 {
188 arm_v7s_iopte pte = paddr & ARM_V7S_LVL_MASK(lvl);
189
190 if (!arm_v7s_is_mtk_enabled(cfg))
191 return pte;
192
193 if (paddr & BIT_ULL(32))
194 pte |= ARM_V7S_ATTR_MTK_PA_BIT32;
195 if (paddr & BIT_ULL(33))
196 pte |= ARM_V7S_ATTR_MTK_PA_BIT33;
197 return pte;
198 }
199
iopte_to_paddr(arm_v7s_iopte pte,int lvl,struct io_pgtable_cfg * cfg)200 static phys_addr_t iopte_to_paddr(arm_v7s_iopte pte, int lvl,
201 struct io_pgtable_cfg *cfg)
202 {
203 arm_v7s_iopte mask;
204 phys_addr_t paddr;
205
206 if (ARM_V7S_PTE_IS_TABLE(pte, lvl))
207 mask = ARM_V7S_TABLE_MASK;
208 else if (arm_v7s_pte_is_cont(pte, lvl))
209 mask = ARM_V7S_LVL_MASK(lvl) * ARM_V7S_CONT_PAGES;
210 else
211 mask = ARM_V7S_LVL_MASK(lvl);
212
213 paddr = pte & mask;
214 if (!arm_v7s_is_mtk_enabled(cfg))
215 return paddr;
216
217 if (pte & ARM_V7S_ATTR_MTK_PA_BIT32)
218 paddr |= BIT_ULL(32);
219 if (pte & ARM_V7S_ATTR_MTK_PA_BIT33)
220 paddr |= BIT_ULL(33);
221 return paddr;
222 }
223
iopte_deref(arm_v7s_iopte pte,int lvl,struct arm_v7s_io_pgtable * data)224 static arm_v7s_iopte *iopte_deref(arm_v7s_iopte pte, int lvl,
225 struct arm_v7s_io_pgtable *data)
226 {
227 return phys_to_virt(iopte_to_paddr(pte, lvl, &data->iop.cfg));
228 }
229
__arm_v7s_alloc_table(int lvl,gfp_t gfp,struct arm_v7s_io_pgtable * data)230 static void *__arm_v7s_alloc_table(int lvl, gfp_t gfp,
231 struct arm_v7s_io_pgtable *data)
232 {
233 struct io_pgtable_cfg *cfg = &data->iop.cfg;
234 struct device *dev = cfg->iommu_dev;
235 phys_addr_t phys;
236 dma_addr_t dma;
237 size_t size = ARM_V7S_TABLE_SIZE(lvl);
238 void *table = NULL;
239
240 if (lvl == 1)
241 table = (void *)__get_free_pages(
242 __GFP_ZERO | ARM_V7S_TABLE_GFP_DMA, get_order(size));
243 else if (lvl == 2)
244 table = kmem_cache_zalloc(data->l2_tables, gfp);
245
246 if (!table)
247 return NULL;
248
249 phys = virt_to_phys(table);
250 if (phys != (arm_v7s_iopte)phys) {
251 /* Doesn't fit in PTE */
252 dev_err(dev, "Page table does not fit in PTE: %pa", &phys);
253 goto out_free;
254 }
255 if (!cfg->coherent_walk) {
256 dma = dma_map_single(dev, table, size, DMA_TO_DEVICE);
257 if (dma_mapping_error(dev, dma))
258 goto out_free;
259 /*
260 * We depend on the IOMMU being able to work with any physical
261 * address directly, so if the DMA layer suggests otherwise by
262 * translating or truncating them, that bodes very badly...
263 */
264 if (dma != phys)
265 goto out_unmap;
266 }
267 if (lvl == 2)
268 kmemleak_ignore(table);
269 return table;
270
271 out_unmap:
272 dev_err(dev, "Cannot accommodate DMA translation for IOMMU page tables\n");
273 dma_unmap_single(dev, dma, size, DMA_TO_DEVICE);
274 out_free:
275 if (lvl == 1)
276 free_pages((unsigned long)table, get_order(size));
277 else
278 kmem_cache_free(data->l2_tables, table);
279 return NULL;
280 }
281
__arm_v7s_free_table(void * table,int lvl,struct arm_v7s_io_pgtable * data)282 static void __arm_v7s_free_table(void *table, int lvl,
283 struct arm_v7s_io_pgtable *data)
284 {
285 struct io_pgtable_cfg *cfg = &data->iop.cfg;
286 struct device *dev = cfg->iommu_dev;
287 size_t size = ARM_V7S_TABLE_SIZE(lvl);
288
289 if (!cfg->coherent_walk)
290 dma_unmap_single(dev, __arm_v7s_dma_addr(table), size,
291 DMA_TO_DEVICE);
292 if (lvl == 1)
293 free_pages((unsigned long)table, get_order(size));
294 else
295 kmem_cache_free(data->l2_tables, table);
296 }
297
__arm_v7s_pte_sync(arm_v7s_iopte * ptep,int num_entries,struct io_pgtable_cfg * cfg)298 static void __arm_v7s_pte_sync(arm_v7s_iopte *ptep, int num_entries,
299 struct io_pgtable_cfg *cfg)
300 {
301 if (cfg->coherent_walk)
302 return;
303
304 dma_sync_single_for_device(cfg->iommu_dev, __arm_v7s_dma_addr(ptep),
305 num_entries * sizeof(*ptep), DMA_TO_DEVICE);
306 }
__arm_v7s_set_pte(arm_v7s_iopte * ptep,arm_v7s_iopte pte,int num_entries,struct io_pgtable_cfg * cfg)307 static void __arm_v7s_set_pte(arm_v7s_iopte *ptep, arm_v7s_iopte pte,
308 int num_entries, struct io_pgtable_cfg *cfg)
309 {
310 int i;
311
312 for (i = 0; i < num_entries; i++)
313 ptep[i] = pte;
314
315 __arm_v7s_pte_sync(ptep, num_entries, cfg);
316 }
317
arm_v7s_prot_to_pte(int prot,int lvl,struct io_pgtable_cfg * cfg)318 static arm_v7s_iopte arm_v7s_prot_to_pte(int prot, int lvl,
319 struct io_pgtable_cfg *cfg)
320 {
321 bool ap = !(cfg->quirks & IO_PGTABLE_QUIRK_NO_PERMS);
322 arm_v7s_iopte pte = ARM_V7S_ATTR_NG | ARM_V7S_ATTR_S;
323
324 if (!(prot & IOMMU_MMIO))
325 pte |= ARM_V7S_ATTR_TEX(1);
326 if (ap) {
327 pte |= ARM_V7S_PTE_AF;
328 if (!(prot & IOMMU_PRIV))
329 pte |= ARM_V7S_PTE_AP_UNPRIV;
330 if (!(prot & IOMMU_WRITE))
331 pte |= ARM_V7S_PTE_AP_RDONLY;
332 }
333 pte <<= ARM_V7S_ATTR_SHIFT(lvl);
334
335 if ((prot & IOMMU_NOEXEC) && ap)
336 pte |= ARM_V7S_ATTR_XN(lvl);
337 if (prot & IOMMU_MMIO)
338 pte |= ARM_V7S_ATTR_B;
339 else if (prot & IOMMU_CACHE)
340 pte |= ARM_V7S_ATTR_B | ARM_V7S_ATTR_C;
341
342 pte |= ARM_V7S_PTE_TYPE_PAGE;
343 if (lvl == 1 && (cfg->quirks & IO_PGTABLE_QUIRK_ARM_NS))
344 pte |= ARM_V7S_ATTR_NS_SECTION;
345
346 return pte;
347 }
348
arm_v7s_pte_to_prot(arm_v7s_iopte pte,int lvl)349 static int arm_v7s_pte_to_prot(arm_v7s_iopte pte, int lvl)
350 {
351 int prot = IOMMU_READ;
352 arm_v7s_iopte attr = pte >> ARM_V7S_ATTR_SHIFT(lvl);
353
354 if (!(attr & ARM_V7S_PTE_AP_RDONLY))
355 prot |= IOMMU_WRITE;
356 if (!(attr & ARM_V7S_PTE_AP_UNPRIV))
357 prot |= IOMMU_PRIV;
358 if ((attr & (ARM_V7S_TEX_MASK << ARM_V7S_TEX_SHIFT)) == 0)
359 prot |= IOMMU_MMIO;
360 else if (pte & ARM_V7S_ATTR_C)
361 prot |= IOMMU_CACHE;
362 if (pte & ARM_V7S_ATTR_XN(lvl))
363 prot |= IOMMU_NOEXEC;
364
365 return prot;
366 }
367
arm_v7s_pte_to_cont(arm_v7s_iopte pte,int lvl)368 static arm_v7s_iopte arm_v7s_pte_to_cont(arm_v7s_iopte pte, int lvl)
369 {
370 if (lvl == 1) {
371 pte |= ARM_V7S_CONT_SECTION;
372 } else if (lvl == 2) {
373 arm_v7s_iopte xn = pte & ARM_V7S_ATTR_XN(lvl);
374 arm_v7s_iopte tex = pte & ARM_V7S_CONT_PAGE_TEX_MASK;
375
376 pte ^= xn | tex | ARM_V7S_PTE_TYPE_PAGE;
377 pte |= (xn << ARM_V7S_CONT_PAGE_XN_SHIFT) |
378 (tex << ARM_V7S_CONT_PAGE_TEX_SHIFT) |
379 ARM_V7S_PTE_TYPE_CONT_PAGE;
380 }
381 return pte;
382 }
383
arm_v7s_cont_to_pte(arm_v7s_iopte pte,int lvl)384 static arm_v7s_iopte arm_v7s_cont_to_pte(arm_v7s_iopte pte, int lvl)
385 {
386 if (lvl == 1) {
387 pte &= ~ARM_V7S_CONT_SECTION;
388 } else if (lvl == 2) {
389 arm_v7s_iopte xn = pte & BIT(ARM_V7S_CONT_PAGE_XN_SHIFT);
390 arm_v7s_iopte tex = pte & (ARM_V7S_CONT_PAGE_TEX_MASK <<
391 ARM_V7S_CONT_PAGE_TEX_SHIFT);
392
393 pte ^= xn | tex | ARM_V7S_PTE_TYPE_CONT_PAGE;
394 pte |= (xn >> ARM_V7S_CONT_PAGE_XN_SHIFT) |
395 (tex >> ARM_V7S_CONT_PAGE_TEX_SHIFT) |
396 ARM_V7S_PTE_TYPE_PAGE;
397 }
398 return pte;
399 }
400
arm_v7s_pte_is_cont(arm_v7s_iopte pte,int lvl)401 static bool arm_v7s_pte_is_cont(arm_v7s_iopte pte, int lvl)
402 {
403 if (lvl == 1 && !ARM_V7S_PTE_IS_TABLE(pte, lvl))
404 return pte & ARM_V7S_CONT_SECTION;
405 else if (lvl == 2)
406 return !(pte & ARM_V7S_PTE_TYPE_PAGE);
407 return false;
408 }
409
410 static size_t __arm_v7s_unmap(struct arm_v7s_io_pgtable *,
411 struct iommu_iotlb_gather *, unsigned long,
412 size_t, int, arm_v7s_iopte *);
413
arm_v7s_init_pte(struct arm_v7s_io_pgtable * data,unsigned long iova,phys_addr_t paddr,int prot,int lvl,int num_entries,arm_v7s_iopte * ptep)414 static int arm_v7s_init_pte(struct arm_v7s_io_pgtable *data,
415 unsigned long iova, phys_addr_t paddr, int prot,
416 int lvl, int num_entries, arm_v7s_iopte *ptep)
417 {
418 struct io_pgtable_cfg *cfg = &data->iop.cfg;
419 arm_v7s_iopte pte;
420 int i;
421
422 for (i = 0; i < num_entries; i++)
423 if (ARM_V7S_PTE_IS_TABLE(ptep[i], lvl)) {
424 /*
425 * We need to unmap and free the old table before
426 * overwriting it with a block entry.
427 */
428 arm_v7s_iopte *tblp;
429 size_t sz = ARM_V7S_BLOCK_SIZE(lvl);
430
431 tblp = ptep - ARM_V7S_LVL_IDX(iova, lvl);
432 if (WARN_ON(__arm_v7s_unmap(data, NULL, iova + i * sz,
433 sz, lvl, tblp) != sz))
434 return -EINVAL;
435 } else if (ptep[i]) {
436 /* We require an unmap first */
437 WARN_ON(!selftest_running);
438 return -EEXIST;
439 }
440
441 pte = arm_v7s_prot_to_pte(prot, lvl, cfg);
442 if (num_entries > 1)
443 pte = arm_v7s_pte_to_cont(pte, lvl);
444
445 pte |= paddr_to_iopte(paddr, lvl, cfg);
446
447 __arm_v7s_set_pte(ptep, pte, num_entries, cfg);
448 return 0;
449 }
450
arm_v7s_install_table(arm_v7s_iopte * table,arm_v7s_iopte * ptep,arm_v7s_iopte curr,struct io_pgtable_cfg * cfg)451 static arm_v7s_iopte arm_v7s_install_table(arm_v7s_iopte *table,
452 arm_v7s_iopte *ptep,
453 arm_v7s_iopte curr,
454 struct io_pgtable_cfg *cfg)
455 {
456 arm_v7s_iopte old, new;
457
458 new = virt_to_phys(table) | ARM_V7S_PTE_TYPE_TABLE;
459 if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_NS)
460 new |= ARM_V7S_ATTR_NS_TABLE;
461
462 /*
463 * Ensure the table itself is visible before its PTE can be.
464 * Whilst we could get away with cmpxchg64_release below, this
465 * doesn't have any ordering semantics when !CONFIG_SMP.
466 */
467 dma_wmb();
468
469 old = cmpxchg_relaxed(ptep, curr, new);
470 __arm_v7s_pte_sync(ptep, 1, cfg);
471
472 return old;
473 }
474
__arm_v7s_map(struct arm_v7s_io_pgtable * data,unsigned long iova,phys_addr_t paddr,size_t size,int prot,int lvl,arm_v7s_iopte * ptep,gfp_t gfp)475 static int __arm_v7s_map(struct arm_v7s_io_pgtable *data, unsigned long iova,
476 phys_addr_t paddr, size_t size, int prot,
477 int lvl, arm_v7s_iopte *ptep, gfp_t gfp)
478 {
479 struct io_pgtable_cfg *cfg = &data->iop.cfg;
480 arm_v7s_iopte pte, *cptep;
481 int num_entries = size >> ARM_V7S_LVL_SHIFT(lvl);
482
483 /* Find our entry at the current level */
484 ptep += ARM_V7S_LVL_IDX(iova, lvl);
485
486 /* If we can install a leaf entry at this level, then do so */
487 if (num_entries)
488 return arm_v7s_init_pte(data, iova, paddr, prot,
489 lvl, num_entries, ptep);
490
491 /* We can't allocate tables at the final level */
492 if (WARN_ON(lvl == 2))
493 return -EINVAL;
494
495 /* Grab a pointer to the next level */
496 pte = READ_ONCE(*ptep);
497 if (!pte) {
498 cptep = __arm_v7s_alloc_table(lvl + 1, gfp, data);
499 if (!cptep)
500 return -ENOMEM;
501
502 pte = arm_v7s_install_table(cptep, ptep, 0, cfg);
503 if (pte)
504 __arm_v7s_free_table(cptep, lvl + 1, data);
505 } else {
506 /* We've no easy way of knowing if it's synced yet, so... */
507 __arm_v7s_pte_sync(ptep, 1, cfg);
508 }
509
510 if (ARM_V7S_PTE_IS_TABLE(pte, lvl)) {
511 cptep = iopte_deref(pte, lvl, data);
512 } else if (pte) {
513 /* We require an unmap first */
514 WARN_ON(!selftest_running);
515 return -EEXIST;
516 }
517
518 /* Rinse, repeat */
519 return __arm_v7s_map(data, iova, paddr, size, prot, lvl + 1, cptep, gfp);
520 }
521
arm_v7s_map(struct io_pgtable_ops * ops,unsigned long iova,phys_addr_t paddr,size_t size,int prot,gfp_t gfp)522 static int arm_v7s_map(struct io_pgtable_ops *ops, unsigned long iova,
523 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
524 {
525 struct arm_v7s_io_pgtable *data = io_pgtable_ops_to_data(ops);
526 struct io_pgtable *iop = &data->iop;
527 int ret;
528
529 /* If no access, then nothing to do */
530 if (!(prot & (IOMMU_READ | IOMMU_WRITE)))
531 return 0;
532
533 if (WARN_ON(iova >= (1ULL << data->iop.cfg.ias) ||
534 paddr >= (1ULL << data->iop.cfg.oas)))
535 return -ERANGE;
536
537 ret = __arm_v7s_map(data, iova, paddr, size, prot, 1, data->pgd, gfp);
538 /*
539 * Synchronise all PTE updates for the new mapping before there's
540 * a chance for anything to kick off a table walk for the new iova.
541 */
542 if (iop->cfg.quirks & IO_PGTABLE_QUIRK_TLBI_ON_MAP) {
543 io_pgtable_tlb_flush_walk(iop, iova, size,
544 ARM_V7S_BLOCK_SIZE(2));
545 } else {
546 wmb();
547 }
548
549 return ret;
550 }
551
arm_v7s_free_pgtable(struct io_pgtable * iop)552 static void arm_v7s_free_pgtable(struct io_pgtable *iop)
553 {
554 struct arm_v7s_io_pgtable *data = io_pgtable_to_data(iop);
555 int i;
556
557 for (i = 0; i < ARM_V7S_PTES_PER_LVL(1); i++) {
558 arm_v7s_iopte pte = data->pgd[i];
559
560 if (ARM_V7S_PTE_IS_TABLE(pte, 1))
561 __arm_v7s_free_table(iopte_deref(pte, 1, data),
562 2, data);
563 }
564 __arm_v7s_free_table(data->pgd, 1, data);
565 kmem_cache_destroy(data->l2_tables);
566 kfree(data);
567 }
568
arm_v7s_split_cont(struct arm_v7s_io_pgtable * data,unsigned long iova,int idx,int lvl,arm_v7s_iopte * ptep)569 static arm_v7s_iopte arm_v7s_split_cont(struct arm_v7s_io_pgtable *data,
570 unsigned long iova, int idx, int lvl,
571 arm_v7s_iopte *ptep)
572 {
573 struct io_pgtable *iop = &data->iop;
574 arm_v7s_iopte pte;
575 size_t size = ARM_V7S_BLOCK_SIZE(lvl);
576 int i;
577
578 /* Check that we didn't lose a race to get the lock */
579 pte = *ptep;
580 if (!arm_v7s_pte_is_cont(pte, lvl))
581 return pte;
582
583 ptep -= idx & (ARM_V7S_CONT_PAGES - 1);
584 pte = arm_v7s_cont_to_pte(pte, lvl);
585 for (i = 0; i < ARM_V7S_CONT_PAGES; i++)
586 ptep[i] = pte + i * size;
587
588 __arm_v7s_pte_sync(ptep, ARM_V7S_CONT_PAGES, &iop->cfg);
589
590 size *= ARM_V7S_CONT_PAGES;
591 io_pgtable_tlb_flush_leaf(iop, iova, size, size);
592 return pte;
593 }
594
arm_v7s_split_blk_unmap(struct arm_v7s_io_pgtable * data,struct iommu_iotlb_gather * gather,unsigned long iova,size_t size,arm_v7s_iopte blk_pte,arm_v7s_iopte * ptep)595 static size_t arm_v7s_split_blk_unmap(struct arm_v7s_io_pgtable *data,
596 struct iommu_iotlb_gather *gather,
597 unsigned long iova, size_t size,
598 arm_v7s_iopte blk_pte,
599 arm_v7s_iopte *ptep)
600 {
601 struct io_pgtable_cfg *cfg = &data->iop.cfg;
602 arm_v7s_iopte pte, *tablep;
603 int i, unmap_idx, num_entries, num_ptes;
604
605 tablep = __arm_v7s_alloc_table(2, GFP_ATOMIC, data);
606 if (!tablep)
607 return 0; /* Bytes unmapped */
608
609 num_ptes = ARM_V7S_PTES_PER_LVL(2);
610 num_entries = size >> ARM_V7S_LVL_SHIFT(2);
611 unmap_idx = ARM_V7S_LVL_IDX(iova, 2);
612
613 pte = arm_v7s_prot_to_pte(arm_v7s_pte_to_prot(blk_pte, 1), 2, cfg);
614 if (num_entries > 1)
615 pte = arm_v7s_pte_to_cont(pte, 2);
616
617 for (i = 0; i < num_ptes; i += num_entries, pte += size) {
618 /* Unmap! */
619 if (i == unmap_idx)
620 continue;
621
622 __arm_v7s_set_pte(&tablep[i], pte, num_entries, cfg);
623 }
624
625 pte = arm_v7s_install_table(tablep, ptep, blk_pte, cfg);
626 if (pte != blk_pte) {
627 __arm_v7s_free_table(tablep, 2, data);
628
629 if (!ARM_V7S_PTE_IS_TABLE(pte, 1))
630 return 0;
631
632 tablep = iopte_deref(pte, 1, data);
633 return __arm_v7s_unmap(data, gather, iova, size, 2, tablep);
634 }
635
636 io_pgtable_tlb_add_page(&data->iop, gather, iova, size);
637 return size;
638 }
639
__arm_v7s_unmap(struct arm_v7s_io_pgtable * data,struct iommu_iotlb_gather * gather,unsigned long iova,size_t size,int lvl,arm_v7s_iopte * ptep)640 static size_t __arm_v7s_unmap(struct arm_v7s_io_pgtable *data,
641 struct iommu_iotlb_gather *gather,
642 unsigned long iova, size_t size, int lvl,
643 arm_v7s_iopte *ptep)
644 {
645 arm_v7s_iopte pte[ARM_V7S_CONT_PAGES];
646 struct io_pgtable *iop = &data->iop;
647 int idx, i = 0, num_entries = size >> ARM_V7S_LVL_SHIFT(lvl);
648
649 /* Something went horribly wrong and we ran out of page table */
650 if (WARN_ON(lvl > 2))
651 return 0;
652
653 idx = ARM_V7S_LVL_IDX(iova, lvl);
654 ptep += idx;
655 do {
656 pte[i] = READ_ONCE(ptep[i]);
657 if (WARN_ON(!ARM_V7S_PTE_IS_VALID(pte[i])))
658 return 0;
659 } while (++i < num_entries);
660
661 /*
662 * If we've hit a contiguous 'large page' entry at this level, it
663 * needs splitting first, unless we're unmapping the whole lot.
664 *
665 * For splitting, we can't rewrite 16 PTEs atomically, and since we
666 * can't necessarily assume TEX remap we don't have a software bit to
667 * mark live entries being split. In practice (i.e. DMA API code), we
668 * will never be splitting large pages anyway, so just wrap this edge
669 * case in a lock for the sake of correctness and be done with it.
670 */
671 if (num_entries <= 1 && arm_v7s_pte_is_cont(pte[0], lvl)) {
672 unsigned long flags;
673
674 spin_lock_irqsave(&data->split_lock, flags);
675 pte[0] = arm_v7s_split_cont(data, iova, idx, lvl, ptep);
676 spin_unlock_irqrestore(&data->split_lock, flags);
677 }
678
679 /* If the size matches this level, we're in the right place */
680 if (num_entries) {
681 size_t blk_size = ARM_V7S_BLOCK_SIZE(lvl);
682
683 __arm_v7s_set_pte(ptep, 0, num_entries, &iop->cfg);
684
685 for (i = 0; i < num_entries; i++) {
686 if (ARM_V7S_PTE_IS_TABLE(pte[i], lvl)) {
687 /* Also flush any partial walks */
688 io_pgtable_tlb_flush_walk(iop, iova, blk_size,
689 ARM_V7S_BLOCK_SIZE(lvl + 1));
690 ptep = iopte_deref(pte[i], lvl, data);
691 __arm_v7s_free_table(ptep, lvl + 1, data);
692 } else if (iop->cfg.quirks & IO_PGTABLE_QUIRK_NON_STRICT) {
693 /*
694 * Order the PTE update against queueing the IOVA, to
695 * guarantee that a flush callback from a different CPU
696 * has observed it before the TLBIALL can be issued.
697 */
698 smp_wmb();
699 } else {
700 io_pgtable_tlb_add_page(iop, gather, iova, blk_size);
701 }
702 iova += blk_size;
703 }
704 return size;
705 } else if (lvl == 1 && !ARM_V7S_PTE_IS_TABLE(pte[0], lvl)) {
706 /*
707 * Insert a table at the next level to map the old region,
708 * minus the part we want to unmap
709 */
710 return arm_v7s_split_blk_unmap(data, gather, iova, size, pte[0],
711 ptep);
712 }
713
714 /* Keep on walkin' */
715 ptep = iopte_deref(pte[0], lvl, data);
716 return __arm_v7s_unmap(data, gather, iova, size, lvl + 1, ptep);
717 }
718
arm_v7s_unmap(struct io_pgtable_ops * ops,unsigned long iova,size_t size,struct iommu_iotlb_gather * gather)719 static size_t arm_v7s_unmap(struct io_pgtable_ops *ops, unsigned long iova,
720 size_t size, struct iommu_iotlb_gather *gather)
721 {
722 struct arm_v7s_io_pgtable *data = io_pgtable_ops_to_data(ops);
723
724 if (WARN_ON(upper_32_bits(iova)))
725 return 0;
726
727 return __arm_v7s_unmap(data, gather, iova, size, 1, data->pgd);
728 }
729
arm_v7s_iova_to_phys(struct io_pgtable_ops * ops,unsigned long iova)730 static phys_addr_t arm_v7s_iova_to_phys(struct io_pgtable_ops *ops,
731 unsigned long iova)
732 {
733 struct arm_v7s_io_pgtable *data = io_pgtable_ops_to_data(ops);
734 arm_v7s_iopte *ptep = data->pgd, pte;
735 int lvl = 0;
736 u32 mask;
737
738 do {
739 ptep += ARM_V7S_LVL_IDX(iova, ++lvl);
740 pte = READ_ONCE(*ptep);
741 ptep = iopte_deref(pte, lvl, data);
742 } while (ARM_V7S_PTE_IS_TABLE(pte, lvl));
743
744 if (!ARM_V7S_PTE_IS_VALID(pte))
745 return 0;
746
747 mask = ARM_V7S_LVL_MASK(lvl);
748 if (arm_v7s_pte_is_cont(pte, lvl))
749 mask *= ARM_V7S_CONT_PAGES;
750 return iopte_to_paddr(pte, lvl, &data->iop.cfg) | (iova & ~mask);
751 }
752
arm_v7s_alloc_pgtable(struct io_pgtable_cfg * cfg,void * cookie)753 static struct io_pgtable *arm_v7s_alloc_pgtable(struct io_pgtable_cfg *cfg,
754 void *cookie)
755 {
756 struct arm_v7s_io_pgtable *data;
757
758 if (cfg->ias > ARM_V7S_ADDR_BITS)
759 return NULL;
760
761 if (cfg->oas > (arm_v7s_is_mtk_enabled(cfg) ? 34 : ARM_V7S_ADDR_BITS))
762 return NULL;
763
764 if (cfg->quirks & ~(IO_PGTABLE_QUIRK_ARM_NS |
765 IO_PGTABLE_QUIRK_NO_PERMS |
766 IO_PGTABLE_QUIRK_TLBI_ON_MAP |
767 IO_PGTABLE_QUIRK_ARM_MTK_EXT |
768 IO_PGTABLE_QUIRK_NON_STRICT))
769 return NULL;
770
771 /* If ARM_MTK_4GB is enabled, the NO_PERMS is also expected. */
772 if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_MTK_EXT &&
773 !(cfg->quirks & IO_PGTABLE_QUIRK_NO_PERMS))
774 return NULL;
775
776 data = kmalloc(sizeof(*data), GFP_KERNEL);
777 if (!data)
778 return NULL;
779
780 spin_lock_init(&data->split_lock);
781 data->l2_tables = kmem_cache_create("io-pgtable_armv7s_l2",
782 ARM_V7S_TABLE_SIZE(2),
783 ARM_V7S_TABLE_SIZE(2),
784 ARM_V7S_TABLE_SLAB_FLAGS, NULL);
785 if (!data->l2_tables)
786 goto out_free_data;
787
788 data->iop.ops = (struct io_pgtable_ops) {
789 .map = arm_v7s_map,
790 .unmap = arm_v7s_unmap,
791 .iova_to_phys = arm_v7s_iova_to_phys,
792 };
793
794 /* We have to do this early for __arm_v7s_alloc_table to work... */
795 data->iop.cfg = *cfg;
796
797 /*
798 * Unless the IOMMU driver indicates supersection support by
799 * having SZ_16M set in the initial bitmap, they won't be used.
800 */
801 cfg->pgsize_bitmap &= SZ_4K | SZ_64K | SZ_1M | SZ_16M;
802
803 /* TCR: T0SZ=0, EAE=0 (if applicable) */
804 cfg->arm_v7s_cfg.tcr = 0;
805
806 /*
807 * TEX remap: the indices used map to the closest equivalent types
808 * under the non-TEX-remap interpretation of those attribute bits,
809 * excepting various implementation-defined aspects of shareability.
810 */
811 cfg->arm_v7s_cfg.prrr = ARM_V7S_PRRR_TR(1, ARM_V7S_PRRR_TYPE_DEVICE) |
812 ARM_V7S_PRRR_TR(4, ARM_V7S_PRRR_TYPE_NORMAL) |
813 ARM_V7S_PRRR_TR(7, ARM_V7S_PRRR_TYPE_NORMAL) |
814 ARM_V7S_PRRR_DS0 | ARM_V7S_PRRR_DS1 |
815 ARM_V7S_PRRR_NS1 | ARM_V7S_PRRR_NOS(7);
816 cfg->arm_v7s_cfg.nmrr = ARM_V7S_NMRR_IR(7, ARM_V7S_RGN_WBWA) |
817 ARM_V7S_NMRR_OR(7, ARM_V7S_RGN_WBWA);
818
819 /* Looking good; allocate a pgd */
820 data->pgd = __arm_v7s_alloc_table(1, GFP_KERNEL, data);
821 if (!data->pgd)
822 goto out_free_data;
823
824 /* Ensure the empty pgd is visible before any actual TTBR write */
825 wmb();
826
827 /* TTBR */
828 cfg->arm_v7s_cfg.ttbr = virt_to_phys(data->pgd) | ARM_V7S_TTBR_S |
829 (cfg->coherent_walk ? (ARM_V7S_TTBR_NOS |
830 ARM_V7S_TTBR_IRGN_ATTR(ARM_V7S_RGN_WBWA) |
831 ARM_V7S_TTBR_ORGN_ATTR(ARM_V7S_RGN_WBWA)) :
832 (ARM_V7S_TTBR_IRGN_ATTR(ARM_V7S_RGN_NC) |
833 ARM_V7S_TTBR_ORGN_ATTR(ARM_V7S_RGN_NC)));
834 return &data->iop;
835
836 out_free_data:
837 kmem_cache_destroy(data->l2_tables);
838 kfree(data);
839 return NULL;
840 }
841
842 struct io_pgtable_init_fns io_pgtable_arm_v7s_init_fns = {
843 .alloc = arm_v7s_alloc_pgtable,
844 .free = arm_v7s_free_pgtable,
845 };
846
847 #ifdef CONFIG_IOMMU_IO_PGTABLE_ARMV7S_SELFTEST
848
849 static struct io_pgtable_cfg *cfg_cookie __initdata;
850
dummy_tlb_flush_all(void * cookie)851 static void __init dummy_tlb_flush_all(void *cookie)
852 {
853 WARN_ON(cookie != cfg_cookie);
854 }
855
dummy_tlb_flush(unsigned long iova,size_t size,size_t granule,void * cookie)856 static void __init dummy_tlb_flush(unsigned long iova, size_t size,
857 size_t granule, void *cookie)
858 {
859 WARN_ON(cookie != cfg_cookie);
860 WARN_ON(!(size & cfg_cookie->pgsize_bitmap));
861 }
862
dummy_tlb_add_page(struct iommu_iotlb_gather * gather,unsigned long iova,size_t granule,void * cookie)863 static void __init dummy_tlb_add_page(struct iommu_iotlb_gather *gather,
864 unsigned long iova, size_t granule,
865 void *cookie)
866 {
867 dummy_tlb_flush(iova, granule, granule, cookie);
868 }
869
870 static const struct iommu_flush_ops dummy_tlb_ops __initconst = {
871 .tlb_flush_all = dummy_tlb_flush_all,
872 .tlb_flush_walk = dummy_tlb_flush,
873 .tlb_flush_leaf = dummy_tlb_flush,
874 .tlb_add_page = dummy_tlb_add_page,
875 };
876
877 #define __FAIL(ops) ({ \
878 WARN(1, "selftest: test failed\n"); \
879 selftest_running = false; \
880 -EFAULT; \
881 })
882
arm_v7s_do_selftests(void)883 static int __init arm_v7s_do_selftests(void)
884 {
885 struct io_pgtable_ops *ops;
886 struct io_pgtable_cfg cfg = {
887 .tlb = &dummy_tlb_ops,
888 .oas = 32,
889 .ias = 32,
890 .coherent_walk = true,
891 .quirks = IO_PGTABLE_QUIRK_ARM_NS,
892 .pgsize_bitmap = SZ_4K | SZ_64K | SZ_1M | SZ_16M,
893 };
894 unsigned int iova, size, iova_start;
895 unsigned int i, loopnr = 0;
896
897 selftest_running = true;
898
899 cfg_cookie = &cfg;
900
901 ops = alloc_io_pgtable_ops(ARM_V7S, &cfg, &cfg);
902 if (!ops) {
903 pr_err("selftest: failed to allocate io pgtable ops\n");
904 return -EINVAL;
905 }
906
907 /*
908 * Initial sanity checks.
909 * Empty page tables shouldn't provide any translations.
910 */
911 if (ops->iova_to_phys(ops, 42))
912 return __FAIL(ops);
913
914 if (ops->iova_to_phys(ops, SZ_1G + 42))
915 return __FAIL(ops);
916
917 if (ops->iova_to_phys(ops, SZ_2G + 42))
918 return __FAIL(ops);
919
920 /*
921 * Distinct mappings of different granule sizes.
922 */
923 iova = 0;
924 for_each_set_bit(i, &cfg.pgsize_bitmap, BITS_PER_LONG) {
925 size = 1UL << i;
926 if (ops->map(ops, iova, iova, size, IOMMU_READ |
927 IOMMU_WRITE |
928 IOMMU_NOEXEC |
929 IOMMU_CACHE, GFP_KERNEL))
930 return __FAIL(ops);
931
932 /* Overlapping mappings */
933 if (!ops->map(ops, iova, iova + size, size,
934 IOMMU_READ | IOMMU_NOEXEC, GFP_KERNEL))
935 return __FAIL(ops);
936
937 if (ops->iova_to_phys(ops, iova + 42) != (iova + 42))
938 return __FAIL(ops);
939
940 iova += SZ_16M;
941 loopnr++;
942 }
943
944 /* Partial unmap */
945 i = 1;
946 size = 1UL << __ffs(cfg.pgsize_bitmap);
947 while (i < loopnr) {
948 iova_start = i * SZ_16M;
949 if (ops->unmap(ops, iova_start + size, size, NULL) != size)
950 return __FAIL(ops);
951
952 /* Remap of partial unmap */
953 if (ops->map(ops, iova_start + size, size, size, IOMMU_READ, GFP_KERNEL))
954 return __FAIL(ops);
955
956 if (ops->iova_to_phys(ops, iova_start + size + 42)
957 != (size + 42))
958 return __FAIL(ops);
959 i++;
960 }
961
962 /* Full unmap */
963 iova = 0;
964 for_each_set_bit(i, &cfg.pgsize_bitmap, BITS_PER_LONG) {
965 size = 1UL << i;
966
967 if (ops->unmap(ops, iova, size, NULL) != size)
968 return __FAIL(ops);
969
970 if (ops->iova_to_phys(ops, iova + 42))
971 return __FAIL(ops);
972
973 /* Remap full block */
974 if (ops->map(ops, iova, iova, size, IOMMU_WRITE, GFP_KERNEL))
975 return __FAIL(ops);
976
977 if (ops->iova_to_phys(ops, iova + 42) != (iova + 42))
978 return __FAIL(ops);
979
980 iova += SZ_16M;
981 }
982
983 free_io_pgtable_ops(ops);
984
985 selftest_running = false;
986
987 pr_info("self test ok\n");
988 return 0;
989 }
990 subsys_initcall(arm_v7s_do_selftests);
991 #endif
992