1 /* 2 * CPU-agnostic ARM page table allocator. 3 * 4 * ARMv7 Short-descriptor format, supporting 5 * - Basic memory attributes 6 * - Simplified access permissions (AP[2:1] model) 7 * - Backwards-compatible TEX remap 8 * - Large pages/supersections (if indicated by the caller) 9 * 10 * Not supporting: 11 * - Legacy access permissions (AP[2:0] model) 12 * 13 * Almost certainly never supporting: 14 * - PXN 15 * - Domains 16 * 17 * This program is free software; you can redistribute it and/or modify 18 * it under the terms of the GNU General Public License version 2 as 19 * published by the Free Software Foundation. 20 * 21 * This program is distributed in the hope that it will be useful, 22 * but WITHOUT ANY WARRANTY; without even the implied warranty of 23 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 24 * GNU General Public License for more details. 25 * 26 * You should have received a copy of the GNU General Public License 27 * along with this program. If not, see <http://www.gnu.org/licenses/>. 28 * 29 * Copyright (C) 2014-2015 ARM Limited 30 * Copyright (c) 2014-2015 MediaTek Inc. 31 */ 32 33 #define pr_fmt(fmt) "arm-v7s io-pgtable: " fmt 34 35 #include <linux/atomic.h> 36 #include <linux/dma-mapping.h> 37 #include <linux/gfp.h> 38 #include <linux/io-pgtable.h> 39 #include <linux/iommu.h> 40 #include <linux/kernel.h> 41 #include <linux/kmemleak.h> 42 #include <linux/sizes.h> 43 #include <linux/slab.h> 44 #include <linux/spinlock.h> 45 #include <linux/types.h> 46 47 #include <asm/barrier.h> 48 49 /* Struct accessors */ 50 #define io_pgtable_to_data(x) \ 51 container_of((x), struct arm_v7s_io_pgtable, iop) 52 53 #define io_pgtable_ops_to_data(x) \ 54 io_pgtable_to_data(io_pgtable_ops_to_pgtable(x)) 55 56 /* 57 * We have 32 bits total; 12 bits resolved at level 1, 8 bits at level 2, 58 * and 12 bits in a page. With some carefully-chosen coefficients we can 59 * hide the ugly inconsistencies behind these macros and at least let the 60 * rest of the code pretend to be somewhat sane. 61 */ 62 #define ARM_V7S_ADDR_BITS 32 63 #define _ARM_V7S_LVL_BITS(lvl) (16 - (lvl) * 4) 64 #define ARM_V7S_LVL_SHIFT(lvl) (ARM_V7S_ADDR_BITS - (4 + 8 * (lvl))) 65 #define ARM_V7S_TABLE_SHIFT 10 66 67 #define ARM_V7S_PTES_PER_LVL(lvl) (1 << _ARM_V7S_LVL_BITS(lvl)) 68 #define ARM_V7S_TABLE_SIZE(lvl) \ 69 (ARM_V7S_PTES_PER_LVL(lvl) * sizeof(arm_v7s_iopte)) 70 71 #define ARM_V7S_BLOCK_SIZE(lvl) (1UL << ARM_V7S_LVL_SHIFT(lvl)) 72 #define ARM_V7S_LVL_MASK(lvl) ((u32)(~0U << ARM_V7S_LVL_SHIFT(lvl))) 73 #define ARM_V7S_TABLE_MASK ((u32)(~0U << ARM_V7S_TABLE_SHIFT)) 74 #define _ARM_V7S_IDX_MASK(lvl) (ARM_V7S_PTES_PER_LVL(lvl) - 1) 75 #define ARM_V7S_LVL_IDX(addr, lvl) ({ \ 76 int _l = lvl; \ 77 ((u32)(addr) >> ARM_V7S_LVL_SHIFT(_l)) & _ARM_V7S_IDX_MASK(_l); \ 78 }) 79 80 /* 81 * Large page/supersection entries are effectively a block of 16 page/section 82 * entries, along the lines of the LPAE contiguous hint, but all with the 83 * same output address. For want of a better common name we'll call them 84 * "contiguous" versions of their respective page/section entries here, but 85 * noting the distinction (WRT to TLB maintenance) that they represent *one* 86 * entry repeated 16 times, not 16 separate entries (as in the LPAE case). 87 */ 88 #define ARM_V7S_CONT_PAGES 16 89 90 /* PTE type bits: these are all mixed up with XN/PXN bits in most cases */ 91 #define ARM_V7S_PTE_TYPE_TABLE 0x1 92 #define ARM_V7S_PTE_TYPE_PAGE 0x2 93 #define ARM_V7S_PTE_TYPE_CONT_PAGE 0x1 94 95 #define ARM_V7S_PTE_IS_VALID(pte) (((pte) & 0x3) != 0) 96 #define ARM_V7S_PTE_IS_TABLE(pte, lvl) \ 97 ((lvl) == 1 && (((pte) & 0x3) == ARM_V7S_PTE_TYPE_TABLE)) 98 99 /* Page table bits */ 100 #define ARM_V7S_ATTR_XN(lvl) BIT(4 * (2 - (lvl))) 101 #define ARM_V7S_ATTR_B BIT(2) 102 #define ARM_V7S_ATTR_C BIT(3) 103 #define ARM_V7S_ATTR_NS_TABLE BIT(3) 104 #define ARM_V7S_ATTR_NS_SECTION BIT(19) 105 106 #define ARM_V7S_CONT_SECTION BIT(18) 107 #define ARM_V7S_CONT_PAGE_XN_SHIFT 15 108 109 /* 110 * The attribute bits are consistently ordered*, but occupy bits [17:10] of 111 * a level 1 PTE vs. bits [11:4] at level 2. Thus we define the individual 112 * fields relative to that 8-bit block, plus a total shift relative to the PTE. 113 */ 114 #define ARM_V7S_ATTR_SHIFT(lvl) (16 - (lvl) * 6) 115 116 #define ARM_V7S_ATTR_MASK 0xff 117 #define ARM_V7S_ATTR_AP0 BIT(0) 118 #define ARM_V7S_ATTR_AP1 BIT(1) 119 #define ARM_V7S_ATTR_AP2 BIT(5) 120 #define ARM_V7S_ATTR_S BIT(6) 121 #define ARM_V7S_ATTR_NG BIT(7) 122 #define ARM_V7S_TEX_SHIFT 2 123 #define ARM_V7S_TEX_MASK 0x7 124 #define ARM_V7S_ATTR_TEX(val) (((val) & ARM_V7S_TEX_MASK) << ARM_V7S_TEX_SHIFT) 125 126 #define ARM_V7S_ATTR_MTK_4GB BIT(9) /* MTK extend it for 4GB mode */ 127 128 /* *well, except for TEX on level 2 large pages, of course :( */ 129 #define ARM_V7S_CONT_PAGE_TEX_SHIFT 6 130 #define ARM_V7S_CONT_PAGE_TEX_MASK (ARM_V7S_TEX_MASK << ARM_V7S_CONT_PAGE_TEX_SHIFT) 131 132 /* Simplified access permissions */ 133 #define ARM_V7S_PTE_AF ARM_V7S_ATTR_AP0 134 #define ARM_V7S_PTE_AP_UNPRIV ARM_V7S_ATTR_AP1 135 #define ARM_V7S_PTE_AP_RDONLY ARM_V7S_ATTR_AP2 136 137 /* Register bits */ 138 #define ARM_V7S_RGN_NC 0 139 #define ARM_V7S_RGN_WBWA 1 140 #define ARM_V7S_RGN_WT 2 141 #define ARM_V7S_RGN_WB 3 142 143 #define ARM_V7S_PRRR_TYPE_DEVICE 1 144 #define ARM_V7S_PRRR_TYPE_NORMAL 2 145 #define ARM_V7S_PRRR_TR(n, type) (((type) & 0x3) << ((n) * 2)) 146 #define ARM_V7S_PRRR_DS0 BIT(16) 147 #define ARM_V7S_PRRR_DS1 BIT(17) 148 #define ARM_V7S_PRRR_NS0 BIT(18) 149 #define ARM_V7S_PRRR_NS1 BIT(19) 150 #define ARM_V7S_PRRR_NOS(n) BIT((n) + 24) 151 152 #define ARM_V7S_NMRR_IR(n, attr) (((attr) & 0x3) << ((n) * 2)) 153 #define ARM_V7S_NMRR_OR(n, attr) (((attr) & 0x3) << ((n) * 2 + 16)) 154 155 #define ARM_V7S_TTBR_S BIT(1) 156 #define ARM_V7S_TTBR_NOS BIT(5) 157 #define ARM_V7S_TTBR_ORGN_ATTR(attr) (((attr) & 0x3) << 3) 158 #define ARM_V7S_TTBR_IRGN_ATTR(attr) \ 159 ((((attr) & 0x1) << 6) | (((attr) & 0x2) >> 1)) 160 161 #define ARM_V7S_TCR_PD1 BIT(5) 162 163 typedef u32 arm_v7s_iopte; 164 165 static bool selftest_running; 166 167 struct arm_v7s_io_pgtable { 168 struct io_pgtable iop; 169 170 arm_v7s_iopte *pgd; 171 struct kmem_cache *l2_tables; 172 spinlock_t split_lock; 173 }; 174 175 static dma_addr_t __arm_v7s_dma_addr(void *pages) 176 { 177 return (dma_addr_t)virt_to_phys(pages); 178 } 179 180 static arm_v7s_iopte *iopte_deref(arm_v7s_iopte pte, int lvl) 181 { 182 if (ARM_V7S_PTE_IS_TABLE(pte, lvl)) 183 pte &= ARM_V7S_TABLE_MASK; 184 else 185 pte &= ARM_V7S_LVL_MASK(lvl); 186 return phys_to_virt(pte); 187 } 188 189 static void *__arm_v7s_alloc_table(int lvl, gfp_t gfp, 190 struct arm_v7s_io_pgtable *data) 191 { 192 struct io_pgtable_cfg *cfg = &data->iop.cfg; 193 struct device *dev = cfg->iommu_dev; 194 phys_addr_t phys; 195 dma_addr_t dma; 196 size_t size = ARM_V7S_TABLE_SIZE(lvl); 197 void *table = NULL; 198 199 if (lvl == 1) 200 table = (void *)__get_dma_pages(__GFP_ZERO, get_order(size)); 201 else if (lvl == 2) 202 table = kmem_cache_zalloc(data->l2_tables, gfp | GFP_DMA); 203 phys = virt_to_phys(table); 204 if (phys != (arm_v7s_iopte)phys) 205 /* Doesn't fit in PTE */ 206 goto out_free; 207 if (table && !(cfg->quirks & IO_PGTABLE_QUIRK_NO_DMA)) { 208 dma = dma_map_single(dev, table, size, DMA_TO_DEVICE); 209 if (dma_mapping_error(dev, dma)) 210 goto out_free; 211 /* 212 * We depend on the IOMMU being able to work with any physical 213 * address directly, so if the DMA layer suggests otherwise by 214 * translating or truncating them, that bodes very badly... 215 */ 216 if (dma != phys) 217 goto out_unmap; 218 } 219 if (lvl == 2) 220 kmemleak_ignore(table); 221 return table; 222 223 out_unmap: 224 dev_err(dev, "Cannot accommodate DMA translation for IOMMU page tables\n"); 225 dma_unmap_single(dev, dma, size, DMA_TO_DEVICE); 226 out_free: 227 if (lvl == 1) 228 free_pages((unsigned long)table, get_order(size)); 229 else 230 kmem_cache_free(data->l2_tables, table); 231 return NULL; 232 } 233 234 static void __arm_v7s_free_table(void *table, int lvl, 235 struct arm_v7s_io_pgtable *data) 236 { 237 struct io_pgtable_cfg *cfg = &data->iop.cfg; 238 struct device *dev = cfg->iommu_dev; 239 size_t size = ARM_V7S_TABLE_SIZE(lvl); 240 241 if (!(cfg->quirks & IO_PGTABLE_QUIRK_NO_DMA)) 242 dma_unmap_single(dev, __arm_v7s_dma_addr(table), size, 243 DMA_TO_DEVICE); 244 if (lvl == 1) 245 free_pages((unsigned long)table, get_order(size)); 246 else 247 kmem_cache_free(data->l2_tables, table); 248 } 249 250 static void __arm_v7s_pte_sync(arm_v7s_iopte *ptep, int num_entries, 251 struct io_pgtable_cfg *cfg) 252 { 253 if (cfg->quirks & IO_PGTABLE_QUIRK_NO_DMA) 254 return; 255 256 dma_sync_single_for_device(cfg->iommu_dev, __arm_v7s_dma_addr(ptep), 257 num_entries * sizeof(*ptep), DMA_TO_DEVICE); 258 } 259 static void __arm_v7s_set_pte(arm_v7s_iopte *ptep, arm_v7s_iopte pte, 260 int num_entries, struct io_pgtable_cfg *cfg) 261 { 262 int i; 263 264 for (i = 0; i < num_entries; i++) 265 ptep[i] = pte; 266 267 __arm_v7s_pte_sync(ptep, num_entries, cfg); 268 } 269 270 static arm_v7s_iopte arm_v7s_prot_to_pte(int prot, int lvl, 271 struct io_pgtable_cfg *cfg) 272 { 273 bool ap = !(cfg->quirks & IO_PGTABLE_QUIRK_NO_PERMS); 274 arm_v7s_iopte pte = ARM_V7S_ATTR_NG | ARM_V7S_ATTR_S; 275 276 if (!(prot & IOMMU_MMIO)) 277 pte |= ARM_V7S_ATTR_TEX(1); 278 if (ap) { 279 pte |= ARM_V7S_PTE_AF; 280 if (!(prot & IOMMU_PRIV)) 281 pte |= ARM_V7S_PTE_AP_UNPRIV; 282 if (!(prot & IOMMU_WRITE)) 283 pte |= ARM_V7S_PTE_AP_RDONLY; 284 } 285 pte <<= ARM_V7S_ATTR_SHIFT(lvl); 286 287 if ((prot & IOMMU_NOEXEC) && ap) 288 pte |= ARM_V7S_ATTR_XN(lvl); 289 if (prot & IOMMU_MMIO) 290 pte |= ARM_V7S_ATTR_B; 291 else if (prot & IOMMU_CACHE) 292 pte |= ARM_V7S_ATTR_B | ARM_V7S_ATTR_C; 293 294 pte |= ARM_V7S_PTE_TYPE_PAGE; 295 if (lvl == 1 && (cfg->quirks & IO_PGTABLE_QUIRK_ARM_NS)) 296 pte |= ARM_V7S_ATTR_NS_SECTION; 297 298 if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_MTK_4GB) 299 pte |= ARM_V7S_ATTR_MTK_4GB; 300 301 return pte; 302 } 303 304 static int arm_v7s_pte_to_prot(arm_v7s_iopte pte, int lvl) 305 { 306 int prot = IOMMU_READ; 307 arm_v7s_iopte attr = pte >> ARM_V7S_ATTR_SHIFT(lvl); 308 309 if (!(attr & ARM_V7S_PTE_AP_RDONLY)) 310 prot |= IOMMU_WRITE; 311 if (!(attr & ARM_V7S_PTE_AP_UNPRIV)) 312 prot |= IOMMU_PRIV; 313 if ((attr & (ARM_V7S_TEX_MASK << ARM_V7S_TEX_SHIFT)) == 0) 314 prot |= IOMMU_MMIO; 315 else if (pte & ARM_V7S_ATTR_C) 316 prot |= IOMMU_CACHE; 317 if (pte & ARM_V7S_ATTR_XN(lvl)) 318 prot |= IOMMU_NOEXEC; 319 320 return prot; 321 } 322 323 static arm_v7s_iopte arm_v7s_pte_to_cont(arm_v7s_iopte pte, int lvl) 324 { 325 if (lvl == 1) { 326 pte |= ARM_V7S_CONT_SECTION; 327 } else if (lvl == 2) { 328 arm_v7s_iopte xn = pte & ARM_V7S_ATTR_XN(lvl); 329 arm_v7s_iopte tex = pte & ARM_V7S_CONT_PAGE_TEX_MASK; 330 331 pte ^= xn | tex | ARM_V7S_PTE_TYPE_PAGE; 332 pte |= (xn << ARM_V7S_CONT_PAGE_XN_SHIFT) | 333 (tex << ARM_V7S_CONT_PAGE_TEX_SHIFT) | 334 ARM_V7S_PTE_TYPE_CONT_PAGE; 335 } 336 return pte; 337 } 338 339 static arm_v7s_iopte arm_v7s_cont_to_pte(arm_v7s_iopte pte, int lvl) 340 { 341 if (lvl == 1) { 342 pte &= ~ARM_V7S_CONT_SECTION; 343 } else if (lvl == 2) { 344 arm_v7s_iopte xn = pte & BIT(ARM_V7S_CONT_PAGE_XN_SHIFT); 345 arm_v7s_iopte tex = pte & (ARM_V7S_CONT_PAGE_TEX_MASK << 346 ARM_V7S_CONT_PAGE_TEX_SHIFT); 347 348 pte ^= xn | tex | ARM_V7S_PTE_TYPE_CONT_PAGE; 349 pte |= (xn >> ARM_V7S_CONT_PAGE_XN_SHIFT) | 350 (tex >> ARM_V7S_CONT_PAGE_TEX_SHIFT) | 351 ARM_V7S_PTE_TYPE_PAGE; 352 } 353 return pte; 354 } 355 356 static bool arm_v7s_pte_is_cont(arm_v7s_iopte pte, int lvl) 357 { 358 if (lvl == 1 && !ARM_V7S_PTE_IS_TABLE(pte, lvl)) 359 return pte & ARM_V7S_CONT_SECTION; 360 else if (lvl == 2) 361 return !(pte & ARM_V7S_PTE_TYPE_PAGE); 362 return false; 363 } 364 365 static size_t __arm_v7s_unmap(struct arm_v7s_io_pgtable *, unsigned long, 366 size_t, int, arm_v7s_iopte *); 367 368 static int arm_v7s_init_pte(struct arm_v7s_io_pgtable *data, 369 unsigned long iova, phys_addr_t paddr, int prot, 370 int lvl, int num_entries, arm_v7s_iopte *ptep) 371 { 372 struct io_pgtable_cfg *cfg = &data->iop.cfg; 373 arm_v7s_iopte pte; 374 int i; 375 376 for (i = 0; i < num_entries; i++) 377 if (ARM_V7S_PTE_IS_TABLE(ptep[i], lvl)) { 378 /* 379 * We need to unmap and free the old table before 380 * overwriting it with a block entry. 381 */ 382 arm_v7s_iopte *tblp; 383 size_t sz = ARM_V7S_BLOCK_SIZE(lvl); 384 385 tblp = ptep - ARM_V7S_LVL_IDX(iova, lvl); 386 if (WARN_ON(__arm_v7s_unmap(data, iova + i * sz, 387 sz, lvl, tblp) != sz)) 388 return -EINVAL; 389 } else if (ptep[i]) { 390 /* We require an unmap first */ 391 WARN_ON(!selftest_running); 392 return -EEXIST; 393 } 394 395 pte = arm_v7s_prot_to_pte(prot, lvl, cfg); 396 if (num_entries > 1) 397 pte = arm_v7s_pte_to_cont(pte, lvl); 398 399 pte |= paddr & ARM_V7S_LVL_MASK(lvl); 400 401 __arm_v7s_set_pte(ptep, pte, num_entries, cfg); 402 return 0; 403 } 404 405 static arm_v7s_iopte arm_v7s_install_table(arm_v7s_iopte *table, 406 arm_v7s_iopte *ptep, 407 arm_v7s_iopte curr, 408 struct io_pgtable_cfg *cfg) 409 { 410 arm_v7s_iopte old, new; 411 412 new = virt_to_phys(table) | ARM_V7S_PTE_TYPE_TABLE; 413 if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_NS) 414 new |= ARM_V7S_ATTR_NS_TABLE; 415 416 /* 417 * Ensure the table itself is visible before its PTE can be. 418 * Whilst we could get away with cmpxchg64_release below, this 419 * doesn't have any ordering semantics when !CONFIG_SMP. 420 */ 421 dma_wmb(); 422 423 old = cmpxchg_relaxed(ptep, curr, new); 424 __arm_v7s_pte_sync(ptep, 1, cfg); 425 426 return old; 427 } 428 429 static int __arm_v7s_map(struct arm_v7s_io_pgtable *data, unsigned long iova, 430 phys_addr_t paddr, size_t size, int prot, 431 int lvl, arm_v7s_iopte *ptep) 432 { 433 struct io_pgtable_cfg *cfg = &data->iop.cfg; 434 arm_v7s_iopte pte, *cptep; 435 int num_entries = size >> ARM_V7S_LVL_SHIFT(lvl); 436 437 /* Find our entry at the current level */ 438 ptep += ARM_V7S_LVL_IDX(iova, lvl); 439 440 /* If we can install a leaf entry at this level, then do so */ 441 if (num_entries) 442 return arm_v7s_init_pte(data, iova, paddr, prot, 443 lvl, num_entries, ptep); 444 445 /* We can't allocate tables at the final level */ 446 if (WARN_ON(lvl == 2)) 447 return -EINVAL; 448 449 /* Grab a pointer to the next level */ 450 pte = READ_ONCE(*ptep); 451 if (!pte) { 452 cptep = __arm_v7s_alloc_table(lvl + 1, GFP_ATOMIC, data); 453 if (!cptep) 454 return -ENOMEM; 455 456 pte = arm_v7s_install_table(cptep, ptep, 0, cfg); 457 if (pte) 458 __arm_v7s_free_table(cptep, lvl + 1, data); 459 } else { 460 /* We've no easy way of knowing if it's synced yet, so... */ 461 __arm_v7s_pte_sync(ptep, 1, cfg); 462 } 463 464 if (ARM_V7S_PTE_IS_TABLE(pte, lvl)) { 465 cptep = iopte_deref(pte, lvl); 466 } else if (pte) { 467 /* We require an unmap first */ 468 WARN_ON(!selftest_running); 469 return -EEXIST; 470 } 471 472 /* Rinse, repeat */ 473 return __arm_v7s_map(data, iova, paddr, size, prot, lvl + 1, cptep); 474 } 475 476 static int arm_v7s_map(struct io_pgtable_ops *ops, unsigned long iova, 477 phys_addr_t paddr, size_t size, int prot) 478 { 479 struct arm_v7s_io_pgtable *data = io_pgtable_ops_to_data(ops); 480 struct io_pgtable *iop = &data->iop; 481 int ret; 482 483 /* If no access, then nothing to do */ 484 if (!(prot & (IOMMU_READ | IOMMU_WRITE))) 485 return 0; 486 487 if (WARN_ON(upper_32_bits(iova) || upper_32_bits(paddr))) 488 return -ERANGE; 489 490 ret = __arm_v7s_map(data, iova, paddr, size, prot, 1, data->pgd); 491 /* 492 * Synchronise all PTE updates for the new mapping before there's 493 * a chance for anything to kick off a table walk for the new iova. 494 */ 495 if (iop->cfg.quirks & IO_PGTABLE_QUIRK_TLBI_ON_MAP) { 496 io_pgtable_tlb_add_flush(iop, iova, size, 497 ARM_V7S_BLOCK_SIZE(2), false); 498 io_pgtable_tlb_sync(iop); 499 } else { 500 wmb(); 501 } 502 503 return ret; 504 } 505 506 static void arm_v7s_free_pgtable(struct io_pgtable *iop) 507 { 508 struct arm_v7s_io_pgtable *data = io_pgtable_to_data(iop); 509 int i; 510 511 for (i = 0; i < ARM_V7S_PTES_PER_LVL(1); i++) { 512 arm_v7s_iopte pte = data->pgd[i]; 513 514 if (ARM_V7S_PTE_IS_TABLE(pte, 1)) 515 __arm_v7s_free_table(iopte_deref(pte, 1), 2, data); 516 } 517 __arm_v7s_free_table(data->pgd, 1, data); 518 kmem_cache_destroy(data->l2_tables); 519 kfree(data); 520 } 521 522 static arm_v7s_iopte arm_v7s_split_cont(struct arm_v7s_io_pgtable *data, 523 unsigned long iova, int idx, int lvl, 524 arm_v7s_iopte *ptep) 525 { 526 struct io_pgtable *iop = &data->iop; 527 arm_v7s_iopte pte; 528 size_t size = ARM_V7S_BLOCK_SIZE(lvl); 529 int i; 530 531 /* Check that we didn't lose a race to get the lock */ 532 pte = *ptep; 533 if (!arm_v7s_pte_is_cont(pte, lvl)) 534 return pte; 535 536 ptep -= idx & (ARM_V7S_CONT_PAGES - 1); 537 pte = arm_v7s_cont_to_pte(pte, lvl); 538 for (i = 0; i < ARM_V7S_CONT_PAGES; i++) 539 ptep[i] = pte + i * size; 540 541 __arm_v7s_pte_sync(ptep, ARM_V7S_CONT_PAGES, &iop->cfg); 542 543 size *= ARM_V7S_CONT_PAGES; 544 io_pgtable_tlb_add_flush(iop, iova, size, size, true); 545 io_pgtable_tlb_sync(iop); 546 return pte; 547 } 548 549 static size_t arm_v7s_split_blk_unmap(struct arm_v7s_io_pgtable *data, 550 unsigned long iova, size_t size, 551 arm_v7s_iopte blk_pte, 552 arm_v7s_iopte *ptep) 553 { 554 struct io_pgtable_cfg *cfg = &data->iop.cfg; 555 arm_v7s_iopte pte, *tablep; 556 int i, unmap_idx, num_entries, num_ptes; 557 558 tablep = __arm_v7s_alloc_table(2, GFP_ATOMIC, data); 559 if (!tablep) 560 return 0; /* Bytes unmapped */ 561 562 num_ptes = ARM_V7S_PTES_PER_LVL(2); 563 num_entries = size >> ARM_V7S_LVL_SHIFT(2); 564 unmap_idx = ARM_V7S_LVL_IDX(iova, 2); 565 566 pte = arm_v7s_prot_to_pte(arm_v7s_pte_to_prot(blk_pte, 1), 2, cfg); 567 if (num_entries > 1) 568 pte = arm_v7s_pte_to_cont(pte, 2); 569 570 for (i = 0; i < num_ptes; i += num_entries, pte += size) { 571 /* Unmap! */ 572 if (i == unmap_idx) 573 continue; 574 575 __arm_v7s_set_pte(&tablep[i], pte, num_entries, cfg); 576 } 577 578 pte = arm_v7s_install_table(tablep, ptep, blk_pte, cfg); 579 if (pte != blk_pte) { 580 __arm_v7s_free_table(tablep, 2, data); 581 582 if (!ARM_V7S_PTE_IS_TABLE(pte, 1)) 583 return 0; 584 585 tablep = iopte_deref(pte, 1); 586 return __arm_v7s_unmap(data, iova, size, 2, tablep); 587 } 588 589 io_pgtable_tlb_add_flush(&data->iop, iova, size, size, true); 590 io_pgtable_tlb_sync(&data->iop); 591 return size; 592 } 593 594 static size_t __arm_v7s_unmap(struct arm_v7s_io_pgtable *data, 595 unsigned long iova, size_t size, int lvl, 596 arm_v7s_iopte *ptep) 597 { 598 arm_v7s_iopte pte[ARM_V7S_CONT_PAGES]; 599 struct io_pgtable *iop = &data->iop; 600 int idx, i = 0, num_entries = size >> ARM_V7S_LVL_SHIFT(lvl); 601 602 /* Something went horribly wrong and we ran out of page table */ 603 if (WARN_ON(lvl > 2)) 604 return 0; 605 606 idx = ARM_V7S_LVL_IDX(iova, lvl); 607 ptep += idx; 608 do { 609 pte[i] = READ_ONCE(ptep[i]); 610 if (WARN_ON(!ARM_V7S_PTE_IS_VALID(pte[i]))) 611 return 0; 612 } while (++i < num_entries); 613 614 /* 615 * If we've hit a contiguous 'large page' entry at this level, it 616 * needs splitting first, unless we're unmapping the whole lot. 617 * 618 * For splitting, we can't rewrite 16 PTEs atomically, and since we 619 * can't necessarily assume TEX remap we don't have a software bit to 620 * mark live entries being split. In practice (i.e. DMA API code), we 621 * will never be splitting large pages anyway, so just wrap this edge 622 * case in a lock for the sake of correctness and be done with it. 623 */ 624 if (num_entries <= 1 && arm_v7s_pte_is_cont(pte[0], lvl)) { 625 unsigned long flags; 626 627 spin_lock_irqsave(&data->split_lock, flags); 628 pte[0] = arm_v7s_split_cont(data, iova, idx, lvl, ptep); 629 spin_unlock_irqrestore(&data->split_lock, flags); 630 } 631 632 /* If the size matches this level, we're in the right place */ 633 if (num_entries) { 634 size_t blk_size = ARM_V7S_BLOCK_SIZE(lvl); 635 636 __arm_v7s_set_pte(ptep, 0, num_entries, &iop->cfg); 637 638 for (i = 0; i < num_entries; i++) { 639 if (ARM_V7S_PTE_IS_TABLE(pte[i], lvl)) { 640 /* Also flush any partial walks */ 641 io_pgtable_tlb_add_flush(iop, iova, blk_size, 642 ARM_V7S_BLOCK_SIZE(lvl + 1), false); 643 io_pgtable_tlb_sync(iop); 644 ptep = iopte_deref(pte[i], lvl); 645 __arm_v7s_free_table(ptep, lvl + 1, data); 646 } else if (iop->cfg.quirks & IO_PGTABLE_QUIRK_NON_STRICT) { 647 /* 648 * Order the PTE update against queueing the IOVA, to 649 * guarantee that a flush callback from a different CPU 650 * has observed it before the TLBIALL can be issued. 651 */ 652 smp_wmb(); 653 } else { 654 io_pgtable_tlb_add_flush(iop, iova, blk_size, 655 blk_size, true); 656 } 657 iova += blk_size; 658 } 659 return size; 660 } else if (lvl == 1 && !ARM_V7S_PTE_IS_TABLE(pte[0], lvl)) { 661 /* 662 * Insert a table at the next level to map the old region, 663 * minus the part we want to unmap 664 */ 665 return arm_v7s_split_blk_unmap(data, iova, size, pte[0], ptep); 666 } 667 668 /* Keep on walkin' */ 669 ptep = iopte_deref(pte[0], lvl); 670 return __arm_v7s_unmap(data, iova, size, lvl + 1, ptep); 671 } 672 673 static size_t arm_v7s_unmap(struct io_pgtable_ops *ops, unsigned long iova, 674 size_t size) 675 { 676 struct arm_v7s_io_pgtable *data = io_pgtable_ops_to_data(ops); 677 678 if (WARN_ON(upper_32_bits(iova))) 679 return 0; 680 681 return __arm_v7s_unmap(data, iova, size, 1, data->pgd); 682 } 683 684 static phys_addr_t arm_v7s_iova_to_phys(struct io_pgtable_ops *ops, 685 unsigned long iova) 686 { 687 struct arm_v7s_io_pgtable *data = io_pgtable_ops_to_data(ops); 688 arm_v7s_iopte *ptep = data->pgd, pte; 689 int lvl = 0; 690 u32 mask; 691 692 do { 693 ptep += ARM_V7S_LVL_IDX(iova, ++lvl); 694 pte = READ_ONCE(*ptep); 695 ptep = iopte_deref(pte, lvl); 696 } while (ARM_V7S_PTE_IS_TABLE(pte, lvl)); 697 698 if (!ARM_V7S_PTE_IS_VALID(pte)) 699 return 0; 700 701 mask = ARM_V7S_LVL_MASK(lvl); 702 if (arm_v7s_pte_is_cont(pte, lvl)) 703 mask *= ARM_V7S_CONT_PAGES; 704 return (pte & mask) | (iova & ~mask); 705 } 706 707 static struct io_pgtable *arm_v7s_alloc_pgtable(struct io_pgtable_cfg *cfg, 708 void *cookie) 709 { 710 struct arm_v7s_io_pgtable *data; 711 712 if (cfg->ias > ARM_V7S_ADDR_BITS || cfg->oas > ARM_V7S_ADDR_BITS) 713 return NULL; 714 715 if (cfg->quirks & ~(IO_PGTABLE_QUIRK_ARM_NS | 716 IO_PGTABLE_QUIRK_NO_PERMS | 717 IO_PGTABLE_QUIRK_TLBI_ON_MAP | 718 IO_PGTABLE_QUIRK_ARM_MTK_4GB | 719 IO_PGTABLE_QUIRK_NO_DMA | 720 IO_PGTABLE_QUIRK_NON_STRICT)) 721 return NULL; 722 723 /* If ARM_MTK_4GB is enabled, the NO_PERMS is also expected. */ 724 if (cfg->quirks & IO_PGTABLE_QUIRK_ARM_MTK_4GB && 725 !(cfg->quirks & IO_PGTABLE_QUIRK_NO_PERMS)) 726 return NULL; 727 728 data = kmalloc(sizeof(*data), GFP_KERNEL); 729 if (!data) 730 return NULL; 731 732 spin_lock_init(&data->split_lock); 733 data->l2_tables = kmem_cache_create("io-pgtable_armv7s_l2", 734 ARM_V7S_TABLE_SIZE(2), 735 ARM_V7S_TABLE_SIZE(2), 736 SLAB_CACHE_DMA, NULL); 737 if (!data->l2_tables) 738 goto out_free_data; 739 740 data->iop.ops = (struct io_pgtable_ops) { 741 .map = arm_v7s_map, 742 .unmap = arm_v7s_unmap, 743 .iova_to_phys = arm_v7s_iova_to_phys, 744 }; 745 746 /* We have to do this early for __arm_v7s_alloc_table to work... */ 747 data->iop.cfg = *cfg; 748 749 /* 750 * Unless the IOMMU driver indicates supersection support by 751 * having SZ_16M set in the initial bitmap, they won't be used. 752 */ 753 cfg->pgsize_bitmap &= SZ_4K | SZ_64K | SZ_1M | SZ_16M; 754 755 /* TCR: T0SZ=0, disable TTBR1 */ 756 cfg->arm_v7s_cfg.tcr = ARM_V7S_TCR_PD1; 757 758 /* 759 * TEX remap: the indices used map to the closest equivalent types 760 * under the non-TEX-remap interpretation of those attribute bits, 761 * excepting various implementation-defined aspects of shareability. 762 */ 763 cfg->arm_v7s_cfg.prrr = ARM_V7S_PRRR_TR(1, ARM_V7S_PRRR_TYPE_DEVICE) | 764 ARM_V7S_PRRR_TR(4, ARM_V7S_PRRR_TYPE_NORMAL) | 765 ARM_V7S_PRRR_TR(7, ARM_V7S_PRRR_TYPE_NORMAL) | 766 ARM_V7S_PRRR_DS0 | ARM_V7S_PRRR_DS1 | 767 ARM_V7S_PRRR_NS1 | ARM_V7S_PRRR_NOS(7); 768 cfg->arm_v7s_cfg.nmrr = ARM_V7S_NMRR_IR(7, ARM_V7S_RGN_WBWA) | 769 ARM_V7S_NMRR_OR(7, ARM_V7S_RGN_WBWA); 770 771 /* Looking good; allocate a pgd */ 772 data->pgd = __arm_v7s_alloc_table(1, GFP_KERNEL, data); 773 if (!data->pgd) 774 goto out_free_data; 775 776 /* Ensure the empty pgd is visible before any actual TTBR write */ 777 wmb(); 778 779 /* TTBRs */ 780 cfg->arm_v7s_cfg.ttbr[0] = virt_to_phys(data->pgd) | 781 ARM_V7S_TTBR_S | ARM_V7S_TTBR_NOS | 782 ARM_V7S_TTBR_IRGN_ATTR(ARM_V7S_RGN_WBWA) | 783 ARM_V7S_TTBR_ORGN_ATTR(ARM_V7S_RGN_WBWA); 784 cfg->arm_v7s_cfg.ttbr[1] = 0; 785 return &data->iop; 786 787 out_free_data: 788 kmem_cache_destroy(data->l2_tables); 789 kfree(data); 790 return NULL; 791 } 792 793 struct io_pgtable_init_fns io_pgtable_arm_v7s_init_fns = { 794 .alloc = arm_v7s_alloc_pgtable, 795 .free = arm_v7s_free_pgtable, 796 }; 797 798 #ifdef CONFIG_IOMMU_IO_PGTABLE_ARMV7S_SELFTEST 799 800 static struct io_pgtable_cfg *cfg_cookie; 801 802 static void dummy_tlb_flush_all(void *cookie) 803 { 804 WARN_ON(cookie != cfg_cookie); 805 } 806 807 static void dummy_tlb_add_flush(unsigned long iova, size_t size, 808 size_t granule, bool leaf, void *cookie) 809 { 810 WARN_ON(cookie != cfg_cookie); 811 WARN_ON(!(size & cfg_cookie->pgsize_bitmap)); 812 } 813 814 static void dummy_tlb_sync(void *cookie) 815 { 816 WARN_ON(cookie != cfg_cookie); 817 } 818 819 static const struct iommu_gather_ops dummy_tlb_ops = { 820 .tlb_flush_all = dummy_tlb_flush_all, 821 .tlb_add_flush = dummy_tlb_add_flush, 822 .tlb_sync = dummy_tlb_sync, 823 }; 824 825 #define __FAIL(ops) ({ \ 826 WARN(1, "selftest: test failed\n"); \ 827 selftest_running = false; \ 828 -EFAULT; \ 829 }) 830 831 static int __init arm_v7s_do_selftests(void) 832 { 833 struct io_pgtable_ops *ops; 834 struct io_pgtable_cfg cfg = { 835 .tlb = &dummy_tlb_ops, 836 .oas = 32, 837 .ias = 32, 838 .quirks = IO_PGTABLE_QUIRK_ARM_NS | IO_PGTABLE_QUIRK_NO_DMA, 839 .pgsize_bitmap = SZ_4K | SZ_64K | SZ_1M | SZ_16M, 840 }; 841 unsigned int iova, size, iova_start; 842 unsigned int i, loopnr = 0; 843 844 selftest_running = true; 845 846 cfg_cookie = &cfg; 847 848 ops = alloc_io_pgtable_ops(ARM_V7S, &cfg, &cfg); 849 if (!ops) { 850 pr_err("selftest: failed to allocate io pgtable ops\n"); 851 return -EINVAL; 852 } 853 854 /* 855 * Initial sanity checks. 856 * Empty page tables shouldn't provide any translations. 857 */ 858 if (ops->iova_to_phys(ops, 42)) 859 return __FAIL(ops); 860 861 if (ops->iova_to_phys(ops, SZ_1G + 42)) 862 return __FAIL(ops); 863 864 if (ops->iova_to_phys(ops, SZ_2G + 42)) 865 return __FAIL(ops); 866 867 /* 868 * Distinct mappings of different granule sizes. 869 */ 870 iova = 0; 871 for_each_set_bit(i, &cfg.pgsize_bitmap, BITS_PER_LONG) { 872 size = 1UL << i; 873 if (ops->map(ops, iova, iova, size, IOMMU_READ | 874 IOMMU_WRITE | 875 IOMMU_NOEXEC | 876 IOMMU_CACHE)) 877 return __FAIL(ops); 878 879 /* Overlapping mappings */ 880 if (!ops->map(ops, iova, iova + size, size, 881 IOMMU_READ | IOMMU_NOEXEC)) 882 return __FAIL(ops); 883 884 if (ops->iova_to_phys(ops, iova + 42) != (iova + 42)) 885 return __FAIL(ops); 886 887 iova += SZ_16M; 888 loopnr++; 889 } 890 891 /* Partial unmap */ 892 i = 1; 893 size = 1UL << __ffs(cfg.pgsize_bitmap); 894 while (i < loopnr) { 895 iova_start = i * SZ_16M; 896 if (ops->unmap(ops, iova_start + size, size) != size) 897 return __FAIL(ops); 898 899 /* Remap of partial unmap */ 900 if (ops->map(ops, iova_start + size, size, size, IOMMU_READ)) 901 return __FAIL(ops); 902 903 if (ops->iova_to_phys(ops, iova_start + size + 42) 904 != (size + 42)) 905 return __FAIL(ops); 906 i++; 907 } 908 909 /* Full unmap */ 910 iova = 0; 911 for_each_set_bit(i, &cfg.pgsize_bitmap, BITS_PER_LONG) { 912 size = 1UL << i; 913 914 if (ops->unmap(ops, iova, size) != size) 915 return __FAIL(ops); 916 917 if (ops->iova_to_phys(ops, iova + 42)) 918 return __FAIL(ops); 919 920 /* Remap full block */ 921 if (ops->map(ops, iova, iova, size, IOMMU_WRITE)) 922 return __FAIL(ops); 923 924 if (ops->iova_to_phys(ops, iova + 42) != (iova + 42)) 925 return __FAIL(ops); 926 927 iova += SZ_16M; 928 } 929 930 free_io_pgtable_ops(ops); 931 932 selftest_running = false; 933 934 pr_info("self test ok\n"); 935 return 0; 936 } 937 subsys_initcall(arm_v7s_do_selftests); 938 #endif 939