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