Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 1 | /* Copyright (c) 2016, The Linux Foundation. All rights reserved. |
| 2 | * |
| 3 | * This program is free software; you can redistribute it and/or modify |
| 4 | * it under the terms of the GNU General Public License version 2 and |
| 5 | * only version 2 as published by the Free Software Foundation. |
| 6 | * |
| 7 | * This program is distributed in the hope that it will be useful, |
| 8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 10 | * GNU General Public License for more details. |
| 11 | */ |
| 12 | |
| 13 | #include <linux/dma-contiguous.h> |
| 14 | #include <linux/dma-mapping.h> |
| 15 | #include <linux/dma-mapping-fast.h> |
| 16 | #include <linux/io-pgtable-fast.h> |
Patrick Daly | 7bcb546 | 2016-08-03 17:27:36 -0700 | [diff] [blame] | 17 | #include <linux/vmalloc.h> |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 18 | #include <asm/cacheflush.h> |
| 19 | #include <asm/dma-iommu.h> |
| 20 | |
| 21 | |
| 22 | /* some redundant definitions... :( TODO: move to io-pgtable-fast.h */ |
| 23 | #define FAST_PAGE_SHIFT 12 |
| 24 | #define FAST_PAGE_SIZE (1UL << FAST_PAGE_SHIFT) |
| 25 | #define FAST_PAGE_MASK (~(PAGE_SIZE - 1)) |
| 26 | #define FAST_PTE_ADDR_MASK ((av8l_fast_iopte)0xfffffffff000) |
| 27 | |
| 28 | /* |
| 29 | * Checks if the allocated range (ending at @end) covered the upcoming |
| 30 | * stale bit. We don't need to know exactly where the range starts since |
| 31 | * we already know where the candidate search range started. If, starting |
| 32 | * from the beginning of the candidate search range, we had to step over |
| 33 | * (or landed directly on top of) the upcoming stale bit, then we return |
| 34 | * true. |
| 35 | * |
| 36 | * Due to wrapping, there are two scenarios we'll need to check: (1) if the |
| 37 | * range [search_start, upcoming_stale] spans 0 (i.e. search_start > |
| 38 | * upcoming_stale), and, (2) if the range: [search_start, upcoming_stale] |
| 39 | * does *not* span 0 (i.e. search_start <= upcoming_stale). And for each |
| 40 | * of those two scenarios we need to handle three cases: (1) the bit was |
| 41 | * found before wrapping or |
| 42 | */ |
| 43 | static bool __bit_covered_stale(unsigned long upcoming_stale, |
| 44 | unsigned long search_start, |
| 45 | unsigned long end) |
| 46 | { |
| 47 | if (search_start > upcoming_stale) { |
| 48 | if (end >= search_start) { |
| 49 | /* |
| 50 | * We started searching above upcoming_stale and we |
| 51 | * didn't wrap, so we couldn't have crossed |
| 52 | * upcoming_stale. |
| 53 | */ |
| 54 | return false; |
| 55 | } |
| 56 | /* |
| 57 | * We wrapped. Did we cross (or land on top of) |
| 58 | * upcoming_stale? |
| 59 | */ |
| 60 | return end >= upcoming_stale; |
| 61 | } |
| 62 | |
| 63 | if (search_start <= upcoming_stale) { |
| 64 | if (end >= search_start) { |
| 65 | /* |
| 66 | * We didn't wrap. Did we cross (or land on top |
| 67 | * of) upcoming_stale? |
| 68 | */ |
| 69 | return end >= upcoming_stale; |
| 70 | } |
| 71 | /* |
| 72 | * We wrapped. So we must have crossed upcoming_stale |
| 73 | * (since we started searching below it). |
| 74 | */ |
| 75 | return true; |
| 76 | } |
| 77 | |
| 78 | /* we should have covered all logical combinations... */ |
| 79 | WARN_ON(1); |
| 80 | return true; |
| 81 | } |
| 82 | |
| 83 | static dma_addr_t __fast_smmu_alloc_iova(struct dma_fast_smmu_mapping *mapping, |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 84 | unsigned long attrs, |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 85 | size_t size) |
| 86 | { |
| 87 | unsigned long bit, prev_search_start, nbits = size >> FAST_PAGE_SHIFT; |
| 88 | unsigned long align = (1 << get_order(size)) - 1; |
| 89 | |
| 90 | bit = bitmap_find_next_zero_area( |
| 91 | mapping->bitmap, mapping->num_4k_pages, mapping->next_start, |
| 92 | nbits, align); |
| 93 | if (unlikely(bit > mapping->num_4k_pages)) { |
| 94 | /* try wrapping */ |
| 95 | mapping->next_start = 0; /* TODO: SHOULD I REALLY DO THIS?!? */ |
| 96 | bit = bitmap_find_next_zero_area( |
| 97 | mapping->bitmap, mapping->num_4k_pages, 0, nbits, |
| 98 | align); |
| 99 | if (unlikely(bit > mapping->num_4k_pages)) |
| 100 | return DMA_ERROR_CODE; |
| 101 | } |
| 102 | |
| 103 | bitmap_set(mapping->bitmap, bit, nbits); |
| 104 | prev_search_start = mapping->next_start; |
| 105 | mapping->next_start = bit + nbits; |
| 106 | if (unlikely(mapping->next_start >= mapping->num_4k_pages)) |
| 107 | mapping->next_start = 0; |
| 108 | |
| 109 | /* |
| 110 | * If we just re-allocated a VA whose TLB hasn't been invalidated |
| 111 | * since it was last used and unmapped, we need to invalidate it |
| 112 | * here. We actually invalidate the entire TLB so that we don't |
| 113 | * have to invalidate the TLB again until we wrap back around. |
| 114 | */ |
| 115 | if (mapping->have_stale_tlbs && |
| 116 | __bit_covered_stale(mapping->upcoming_stale_bit, |
| 117 | prev_search_start, |
| 118 | bit + nbits - 1)) { |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 119 | bool skip_sync = (attrs & DMA_ATTR_SKIP_CPU_SYNC); |
| 120 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 121 | iommu_tlbiall(mapping->domain); |
| 122 | mapping->have_stale_tlbs = false; |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 123 | av8l_fast_clear_stale_ptes(mapping->pgtbl_pmds, skip_sync); |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 124 | } |
| 125 | |
| 126 | return (bit << FAST_PAGE_SHIFT) + mapping->base; |
| 127 | } |
| 128 | |
| 129 | /* |
| 130 | * Checks whether the candidate bit will be allocated sooner than the |
| 131 | * current upcoming stale bit. We can say candidate will be upcoming |
| 132 | * sooner than the current upcoming stale bit if it lies between the |
| 133 | * starting bit of the next search range and the upcoming stale bit |
| 134 | * (allowing for wrap-around). |
| 135 | * |
| 136 | * Stated differently, we're checking the relative ordering of three |
| 137 | * unsigned numbers. So we need to check all 6 (i.e. 3!) permutations, |
| 138 | * namely: |
| 139 | * |
| 140 | * 0 |---A---B---C---| TOP (Case 1) |
| 141 | * 0 |---A---C---B---| TOP (Case 2) |
| 142 | * 0 |---B---A---C---| TOP (Case 3) |
| 143 | * 0 |---B---C---A---| TOP (Case 4) |
| 144 | * 0 |---C---A---B---| TOP (Case 5) |
| 145 | * 0 |---C---B---A---| TOP (Case 6) |
| 146 | * |
| 147 | * Note that since we're allowing numbers to wrap, the following three |
| 148 | * scenarios are all equivalent for Case 1: |
| 149 | * |
| 150 | * 0 |---A---B---C---| TOP |
| 151 | * 0 |---C---A---B---| TOP (C has wrapped. This is Case 5.) |
| 152 | * 0 |---B---C---A---| TOP (C and B have wrapped. This is Case 4.) |
| 153 | * |
| 154 | * In any of these cases, if we start searching from A, we will find B |
| 155 | * before we find C. |
| 156 | * |
| 157 | * We can also find two equivalent cases for Case 2: |
| 158 | * |
| 159 | * 0 |---A---C---B---| TOP |
| 160 | * 0 |---B---A---C---| TOP (B has wrapped. This is Case 3.) |
| 161 | * 0 |---C---B---A---| TOP (B and C have wrapped. This is Case 6.) |
| 162 | * |
| 163 | * In any of these cases, if we start searching from A, we will find C |
| 164 | * before we find B. |
| 165 | */ |
| 166 | static bool __bit_is_sooner(unsigned long candidate, |
| 167 | struct dma_fast_smmu_mapping *mapping) |
| 168 | { |
| 169 | unsigned long A = mapping->next_start; |
| 170 | unsigned long B = candidate; |
| 171 | unsigned long C = mapping->upcoming_stale_bit; |
| 172 | |
| 173 | if ((A < B && B < C) || /* Case 1 */ |
| 174 | (C < A && A < B) || /* Case 5 */ |
| 175 | (B < C && C < A)) /* Case 4 */ |
| 176 | return true; |
| 177 | |
| 178 | if ((A < C && C < B) || /* Case 2 */ |
| 179 | (B < A && A < C) || /* Case 3 */ |
| 180 | (C < B && B < A)) /* Case 6 */ |
| 181 | return false; |
| 182 | |
| 183 | /* |
| 184 | * For simplicity, we've been ignoring the possibility of any of |
| 185 | * our three numbers being equal. Handle those cases here (they |
| 186 | * shouldn't happen very often, (I think?)). |
| 187 | */ |
| 188 | |
| 189 | /* |
| 190 | * If candidate is the next bit to be searched then it's definitely |
| 191 | * sooner. |
| 192 | */ |
| 193 | if (A == B) |
| 194 | return true; |
| 195 | |
| 196 | /* |
| 197 | * If candidate is the next upcoming stale bit we'll return false |
| 198 | * to avoid doing `upcoming = candidate' in the caller (which would |
| 199 | * be useless since they're already equal) |
| 200 | */ |
| 201 | if (B == C) |
| 202 | return false; |
| 203 | |
| 204 | /* |
| 205 | * If next start is the upcoming stale bit then candidate can't |
| 206 | * possibly be sooner. The "soonest" bit is already selected. |
| 207 | */ |
| 208 | if (A == C) |
| 209 | return false; |
| 210 | |
| 211 | /* We should have covered all logical combinations. */ |
| 212 | WARN(1, "Well, that's awkward. A=%ld, B=%ld, C=%ld\n", A, B, C); |
| 213 | return true; |
| 214 | } |
| 215 | |
| 216 | static void __fast_smmu_free_iova(struct dma_fast_smmu_mapping *mapping, |
| 217 | dma_addr_t iova, size_t size) |
| 218 | { |
| 219 | unsigned long start_bit = (iova - mapping->base) >> FAST_PAGE_SHIFT; |
| 220 | unsigned long nbits = size >> FAST_PAGE_SHIFT; |
| 221 | |
| 222 | /* |
| 223 | * We don't invalidate TLBs on unmap. We invalidate TLBs on map |
| 224 | * when we're about to re-allocate a VA that was previously |
| 225 | * unmapped but hasn't yet been invalidated. So we need to keep |
| 226 | * track of which bit is the closest to being re-allocated here. |
| 227 | */ |
| 228 | if (__bit_is_sooner(start_bit, mapping)) |
| 229 | mapping->upcoming_stale_bit = start_bit; |
| 230 | |
| 231 | bitmap_clear(mapping->bitmap, start_bit, nbits); |
| 232 | mapping->have_stale_tlbs = true; |
| 233 | } |
| 234 | |
| 235 | |
| 236 | static void __fast_dma_page_cpu_to_dev(struct page *page, unsigned long off, |
| 237 | size_t size, enum dma_data_direction dir) |
| 238 | { |
| 239 | __dma_map_area(page_address(page) + off, size, dir); |
| 240 | } |
| 241 | |
| 242 | static void __fast_dma_page_dev_to_cpu(struct page *page, unsigned long off, |
| 243 | size_t size, enum dma_data_direction dir) |
| 244 | { |
| 245 | __dma_unmap_area(page_address(page) + off, size, dir); |
| 246 | |
| 247 | /* TODO: WHAT IS THIS? */ |
| 248 | /* |
| 249 | * Mark the D-cache clean for this page to avoid extra flushing. |
| 250 | */ |
| 251 | if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE) |
| 252 | set_bit(PG_dcache_clean, &page->flags); |
| 253 | } |
| 254 | |
| 255 | static int __fast_dma_direction_to_prot(enum dma_data_direction dir) |
| 256 | { |
| 257 | switch (dir) { |
| 258 | case DMA_BIDIRECTIONAL: |
| 259 | return IOMMU_READ | IOMMU_WRITE; |
| 260 | case DMA_TO_DEVICE: |
| 261 | return IOMMU_READ; |
| 262 | case DMA_FROM_DEVICE: |
| 263 | return IOMMU_WRITE; |
| 264 | default: |
| 265 | return 0; |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | static dma_addr_t fast_smmu_map_page(struct device *dev, struct page *page, |
| 270 | unsigned long offset, size_t size, |
| 271 | enum dma_data_direction dir, |
| 272 | unsigned long attrs) |
| 273 | { |
| 274 | struct dma_fast_smmu_mapping *mapping = dev->archdata.mapping->fast; |
| 275 | dma_addr_t iova; |
| 276 | unsigned long flags; |
| 277 | av8l_fast_iopte *pmd; |
| 278 | phys_addr_t phys_plus_off = page_to_phys(page) + offset; |
| 279 | phys_addr_t phys_to_map = round_down(phys_plus_off, FAST_PAGE_SIZE); |
| 280 | unsigned long offset_from_phys_to_map = phys_plus_off & ~FAST_PAGE_MASK; |
| 281 | size_t len = ALIGN(size + offset_from_phys_to_map, FAST_PAGE_SIZE); |
| 282 | int nptes = len >> FAST_PAGE_SHIFT; |
| 283 | bool skip_sync = (attrs & DMA_ATTR_SKIP_CPU_SYNC); |
| 284 | int prot = __fast_dma_direction_to_prot(dir); |
| 285 | |
| 286 | if (attrs & DMA_ATTR_STRONGLY_ORDERED) |
| 287 | prot |= IOMMU_MMIO; |
| 288 | |
| 289 | if (!skip_sync) |
| 290 | __fast_dma_page_cpu_to_dev(phys_to_page(phys_to_map), |
| 291 | offset_from_phys_to_map, size, dir); |
| 292 | |
| 293 | spin_lock_irqsave(&mapping->lock, flags); |
| 294 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 295 | iova = __fast_smmu_alloc_iova(mapping, attrs, len); |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 296 | |
| 297 | if (unlikely(iova == DMA_ERROR_CODE)) |
| 298 | goto fail; |
| 299 | |
| 300 | pmd = iopte_pmd_offset(mapping->pgtbl_pmds, iova); |
| 301 | |
| 302 | if (unlikely(av8l_fast_map_public(pmd, phys_to_map, len, prot))) |
| 303 | goto fail_free_iova; |
| 304 | |
| 305 | if (!skip_sync) /* TODO: should ask SMMU if coherent */ |
| 306 | dmac_clean_range(pmd, pmd + nptes); |
| 307 | |
| 308 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 309 | return iova + offset_from_phys_to_map; |
| 310 | |
| 311 | fail_free_iova: |
| 312 | __fast_smmu_free_iova(mapping, iova, size); |
| 313 | fail: |
| 314 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 315 | return DMA_ERROR_CODE; |
| 316 | } |
| 317 | |
| 318 | static void fast_smmu_unmap_page(struct device *dev, dma_addr_t iova, |
| 319 | size_t size, enum dma_data_direction dir, |
| 320 | unsigned long attrs) |
| 321 | { |
| 322 | struct dma_fast_smmu_mapping *mapping = dev->archdata.mapping->fast; |
| 323 | unsigned long flags; |
| 324 | av8l_fast_iopte *pmd = iopte_pmd_offset(mapping->pgtbl_pmds, iova); |
| 325 | unsigned long offset = iova & ~FAST_PAGE_MASK; |
| 326 | size_t len = ALIGN(size + offset, FAST_PAGE_SIZE); |
| 327 | int nptes = len >> FAST_PAGE_SHIFT; |
| 328 | struct page *page = phys_to_page((*pmd & FAST_PTE_ADDR_MASK)); |
| 329 | bool skip_sync = (attrs & DMA_ATTR_SKIP_CPU_SYNC); |
| 330 | |
| 331 | if (!skip_sync) |
| 332 | __fast_dma_page_dev_to_cpu(page, offset, size, dir); |
| 333 | |
| 334 | spin_lock_irqsave(&mapping->lock, flags); |
| 335 | av8l_fast_unmap_public(pmd, len); |
| 336 | if (!skip_sync) /* TODO: should ask SMMU if coherent */ |
| 337 | dmac_clean_range(pmd, pmd + nptes); |
| 338 | __fast_smmu_free_iova(mapping, iova, len); |
| 339 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 340 | } |
| 341 | |
Liam Mark | 78d7fb5 | 2016-12-01 13:05:31 -0800 | [diff] [blame^] | 342 | static void fast_smmu_sync_single_for_cpu(struct device *dev, |
| 343 | dma_addr_t iova, size_t size, enum dma_data_direction dir) |
| 344 | { |
| 345 | struct dma_fast_smmu_mapping *mapping = dev->archdata.mapping->fast; |
| 346 | av8l_fast_iopte *pmd = iopte_pmd_offset(mapping->pgtbl_pmds, iova); |
| 347 | unsigned long offset = iova & ~FAST_PAGE_MASK; |
| 348 | struct page *page = phys_to_page((*pmd & FAST_PTE_ADDR_MASK)); |
| 349 | |
| 350 | if (!is_device_dma_coherent(dev)) |
| 351 | __fast_dma_page_dev_to_cpu(page, offset, size, dir); |
| 352 | } |
| 353 | |
| 354 | static void fast_smmu_sync_single_for_device(struct device *dev, |
| 355 | dma_addr_t iova, size_t size, enum dma_data_direction dir) |
| 356 | { |
| 357 | struct dma_fast_smmu_mapping *mapping = dev->archdata.mapping->fast; |
| 358 | av8l_fast_iopte *pmd = iopte_pmd_offset(mapping->pgtbl_pmds, iova); |
| 359 | unsigned long offset = iova & ~FAST_PAGE_MASK; |
| 360 | struct page *page = phys_to_page((*pmd & FAST_PTE_ADDR_MASK)); |
| 361 | |
| 362 | if (!is_device_dma_coherent(dev)) |
| 363 | __fast_dma_page_cpu_to_dev(page, offset, size, dir); |
| 364 | } |
| 365 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 366 | static int fast_smmu_map_sg(struct device *dev, struct scatterlist *sg, |
| 367 | int nents, enum dma_data_direction dir, |
| 368 | unsigned long attrs) |
| 369 | { |
| 370 | return -EINVAL; |
| 371 | } |
| 372 | |
| 373 | static void fast_smmu_unmap_sg(struct device *dev, |
| 374 | struct scatterlist *sg, int nents, |
| 375 | enum dma_data_direction dir, |
| 376 | unsigned long attrs) |
| 377 | { |
| 378 | WARN_ON_ONCE(1); |
| 379 | } |
| 380 | |
Liam Mark | 78d7fb5 | 2016-12-01 13:05:31 -0800 | [diff] [blame^] | 381 | static void fast_smmu_sync_sg_for_cpu(struct device *dev, |
| 382 | struct scatterlist *sg, int nents, enum dma_data_direction dir) |
| 383 | { |
| 384 | WARN_ON_ONCE(1); |
| 385 | } |
| 386 | |
| 387 | static void fast_smmu_sync_sg_for_device(struct device *dev, |
| 388 | struct scatterlist *sg, int nents, enum dma_data_direction dir) |
| 389 | { |
| 390 | WARN_ON_ONCE(1); |
| 391 | } |
| 392 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 393 | static void __fast_smmu_free_pages(struct page **pages, int count) |
| 394 | { |
| 395 | int i; |
| 396 | |
| 397 | for (i = 0; i < count; i++) |
| 398 | __free_page(pages[i]); |
| 399 | kvfree(pages); |
| 400 | } |
| 401 | |
| 402 | static struct page **__fast_smmu_alloc_pages(unsigned int count, gfp_t gfp) |
| 403 | { |
| 404 | struct page **pages; |
| 405 | unsigned int i = 0, array_size = count * sizeof(*pages); |
| 406 | |
| 407 | if (array_size <= PAGE_SIZE) |
| 408 | pages = kzalloc(array_size, GFP_KERNEL); |
| 409 | else |
| 410 | pages = vzalloc(array_size); |
| 411 | if (!pages) |
| 412 | return NULL; |
| 413 | |
| 414 | /* IOMMU can map any pages, so himem can also be used here */ |
| 415 | gfp |= __GFP_NOWARN | __GFP_HIGHMEM; |
| 416 | |
| 417 | for (i = 0; i < count; ++i) { |
| 418 | struct page *page = alloc_page(gfp); |
| 419 | |
| 420 | if (!page) { |
| 421 | __fast_smmu_free_pages(pages, i); |
| 422 | return NULL; |
| 423 | } |
| 424 | pages[i] = page; |
| 425 | } |
| 426 | return pages; |
| 427 | } |
| 428 | |
| 429 | static void *fast_smmu_alloc(struct device *dev, size_t size, |
| 430 | dma_addr_t *handle, gfp_t gfp, |
| 431 | unsigned long attrs) |
| 432 | { |
| 433 | struct dma_fast_smmu_mapping *mapping = dev->archdata.mapping->fast; |
| 434 | struct sg_table sgt; |
| 435 | dma_addr_t dma_addr, iova_iter; |
| 436 | void *addr; |
| 437 | av8l_fast_iopte *ptep; |
| 438 | unsigned long flags; |
| 439 | struct sg_mapping_iter miter; |
| 440 | unsigned int count = ALIGN(size, SZ_4K) >> PAGE_SHIFT; |
| 441 | int prot = IOMMU_READ | IOMMU_WRITE; /* TODO: extract from attrs */ |
| 442 | pgprot_t remap_prot = pgprot_writecombine(PAGE_KERNEL); |
| 443 | struct page **pages; |
| 444 | |
| 445 | *handle = DMA_ERROR_CODE; |
| 446 | |
| 447 | pages = __fast_smmu_alloc_pages(count, gfp); |
| 448 | if (!pages) { |
| 449 | dev_err(dev, "no pages\n"); |
| 450 | return NULL; |
| 451 | } |
| 452 | |
| 453 | size = ALIGN(size, SZ_4K); |
| 454 | if (sg_alloc_table_from_pages(&sgt, pages, count, 0, size, gfp)) { |
| 455 | dev_err(dev, "no sg tablen\n"); |
| 456 | goto out_free_pages; |
| 457 | } |
| 458 | |
| 459 | if (!(prot & IOMMU_CACHE)) { |
| 460 | /* |
| 461 | * The CPU-centric flushing implied by SG_MITER_TO_SG isn't |
| 462 | * sufficient here, so skip it by using the "wrong" direction. |
| 463 | */ |
| 464 | sg_miter_start(&miter, sgt.sgl, sgt.orig_nents, |
| 465 | SG_MITER_FROM_SG); |
| 466 | while (sg_miter_next(&miter)) |
Kyle Yan | 65be4a5 | 2016-10-31 15:05:00 -0700 | [diff] [blame] | 467 | __dma_flush_area(miter.addr, miter.length); |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 468 | sg_miter_stop(&miter); |
| 469 | } |
| 470 | |
| 471 | spin_lock_irqsave(&mapping->lock, flags); |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 472 | dma_addr = __fast_smmu_alloc_iova(mapping, attrs, size); |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 473 | if (dma_addr == DMA_ERROR_CODE) { |
| 474 | dev_err(dev, "no iova\n"); |
| 475 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 476 | goto out_free_sg; |
| 477 | } |
| 478 | iova_iter = dma_addr; |
| 479 | sg_miter_start(&miter, sgt.sgl, sgt.orig_nents, |
| 480 | SG_MITER_FROM_SG | SG_MITER_ATOMIC); |
| 481 | while (sg_miter_next(&miter)) { |
| 482 | int nptes = miter.length >> FAST_PAGE_SHIFT; |
| 483 | |
| 484 | ptep = iopte_pmd_offset(mapping->pgtbl_pmds, iova_iter); |
| 485 | if (unlikely(av8l_fast_map_public( |
| 486 | ptep, page_to_phys(miter.page), |
| 487 | miter.length, prot))) { |
| 488 | dev_err(dev, "no map public\n"); |
| 489 | /* TODO: unwind previously successful mappings */ |
| 490 | goto out_free_iova; |
| 491 | } |
| 492 | dmac_clean_range(ptep, ptep + nptes); |
| 493 | iova_iter += miter.length; |
| 494 | } |
| 495 | sg_miter_stop(&miter); |
| 496 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 497 | |
| 498 | addr = dma_common_pages_remap(pages, size, VM_USERMAP, remap_prot, |
| 499 | __builtin_return_address(0)); |
| 500 | if (!addr) { |
| 501 | dev_err(dev, "no common pages\n"); |
| 502 | goto out_unmap; |
| 503 | } |
| 504 | |
| 505 | *handle = dma_addr; |
| 506 | sg_free_table(&sgt); |
| 507 | return addr; |
| 508 | |
| 509 | out_unmap: |
| 510 | /* need to take the lock again for page tables and iova */ |
| 511 | spin_lock_irqsave(&mapping->lock, flags); |
| 512 | ptep = iopte_pmd_offset(mapping->pgtbl_pmds, dma_addr); |
| 513 | av8l_fast_unmap_public(ptep, size); |
| 514 | dmac_clean_range(ptep, ptep + count); |
| 515 | out_free_iova: |
| 516 | __fast_smmu_free_iova(mapping, dma_addr, size); |
| 517 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 518 | out_free_sg: |
| 519 | sg_free_table(&sgt); |
| 520 | out_free_pages: |
| 521 | __fast_smmu_free_pages(pages, count); |
| 522 | return NULL; |
| 523 | } |
| 524 | |
| 525 | static void fast_smmu_free(struct device *dev, size_t size, |
| 526 | void *vaddr, dma_addr_t dma_handle, |
| 527 | unsigned long attrs) |
| 528 | { |
| 529 | struct dma_fast_smmu_mapping *mapping = dev->archdata.mapping->fast; |
| 530 | struct vm_struct *area; |
| 531 | struct page **pages; |
| 532 | size_t count = ALIGN(size, SZ_4K) >> FAST_PAGE_SHIFT; |
| 533 | av8l_fast_iopte *ptep; |
| 534 | unsigned long flags; |
| 535 | |
| 536 | size = ALIGN(size, SZ_4K); |
| 537 | |
| 538 | area = find_vm_area(vaddr); |
| 539 | if (WARN_ON_ONCE(!area)) |
| 540 | return; |
| 541 | |
| 542 | pages = area->pages; |
| 543 | dma_common_free_remap(vaddr, size, VM_USERMAP, false); |
| 544 | ptep = iopte_pmd_offset(mapping->pgtbl_pmds, dma_handle); |
| 545 | spin_lock_irqsave(&mapping->lock, flags); |
| 546 | av8l_fast_unmap_public(ptep, size); |
| 547 | dmac_clean_range(ptep, ptep + count); |
| 548 | __fast_smmu_free_iova(mapping, dma_handle, size); |
| 549 | spin_unlock_irqrestore(&mapping->lock, flags); |
| 550 | __fast_smmu_free_pages(pages, count); |
| 551 | } |
| 552 | |
Patrick Daly | 7bcb546 | 2016-08-03 17:27:36 -0700 | [diff] [blame] | 553 | static int fast_smmu_mmap_attrs(struct device *dev, struct vm_area_struct *vma, |
| 554 | void *cpu_addr, dma_addr_t dma_addr, |
| 555 | size_t size, unsigned long attrs) |
| 556 | { |
| 557 | struct vm_struct *area; |
| 558 | unsigned long uaddr = vma->vm_start; |
| 559 | struct page **pages; |
| 560 | int i, nr_pages, ret = 0; |
| 561 | |
| 562 | vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot); |
| 563 | |
| 564 | area = find_vm_area(cpu_addr); |
| 565 | if (!area) |
| 566 | return -EINVAL; |
| 567 | |
| 568 | pages = area->pages; |
| 569 | nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; |
| 570 | for (i = vma->vm_pgoff; i < nr_pages && uaddr < vma->vm_end; i++) { |
| 571 | ret = vm_insert_page(vma, uaddr, pages[i]); |
| 572 | if (ret) |
| 573 | break; |
| 574 | uaddr += PAGE_SIZE; |
| 575 | } |
| 576 | |
| 577 | return ret; |
| 578 | } |
| 579 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 580 | static int fast_smmu_dma_supported(struct device *dev, u64 mask) |
| 581 | { |
| 582 | return mask <= 0xffffffff; |
| 583 | } |
| 584 | |
| 585 | static int fast_smmu_mapping_error(struct device *dev, |
| 586 | dma_addr_t dma_addr) |
| 587 | { |
| 588 | return dma_addr == DMA_ERROR_CODE; |
| 589 | } |
| 590 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 591 | static void __fast_smmu_mapped_over_stale(struct dma_fast_smmu_mapping *fast, |
| 592 | void *data) |
| 593 | { |
| 594 | av8l_fast_iopte *ptep = data; |
| 595 | dma_addr_t iova; |
| 596 | unsigned long bitmap_idx; |
| 597 | |
| 598 | bitmap_idx = (unsigned long)(ptep - fast->pgtbl_pmds); |
| 599 | iova = bitmap_idx << FAST_PAGE_SHIFT; |
| 600 | dev_err(fast->dev, "Mapped over stale tlb at %pa\n", &iova); |
| 601 | dev_err(fast->dev, "bitmap (failure at idx %lu):\n", bitmap_idx); |
| 602 | dev_err(fast->dev, "ptep: %p pmds: %p diff: %lu\n", ptep, |
| 603 | fast->pgtbl_pmds, ptep - fast->pgtbl_pmds); |
| 604 | print_hex_dump(KERN_ERR, "bmap: ", DUMP_PREFIX_ADDRESS, |
| 605 | 32, 8, fast->bitmap, fast->bitmap_size, false); |
| 606 | } |
| 607 | |
| 608 | static int fast_smmu_notify(struct notifier_block *self, |
| 609 | unsigned long action, void *data) |
| 610 | { |
| 611 | struct dma_fast_smmu_mapping *fast = container_of( |
| 612 | self, struct dma_fast_smmu_mapping, notifier); |
| 613 | |
| 614 | switch (action) { |
| 615 | case MAPPED_OVER_STALE_TLB: |
| 616 | __fast_smmu_mapped_over_stale(fast, data); |
| 617 | return NOTIFY_OK; |
| 618 | default: |
| 619 | WARN(1, "Unhandled notifier action"); |
| 620 | return NOTIFY_DONE; |
| 621 | } |
| 622 | } |
| 623 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 624 | static const struct dma_map_ops fast_smmu_dma_ops = { |
| 625 | .alloc = fast_smmu_alloc, |
| 626 | .free = fast_smmu_free, |
Patrick Daly | 7bcb546 | 2016-08-03 17:27:36 -0700 | [diff] [blame] | 627 | .mmap = fast_smmu_mmap_attrs, |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 628 | .map_page = fast_smmu_map_page, |
| 629 | .unmap_page = fast_smmu_unmap_page, |
Liam Mark | 78d7fb5 | 2016-12-01 13:05:31 -0800 | [diff] [blame^] | 630 | .sync_single_for_cpu = fast_smmu_sync_single_for_cpu, |
| 631 | .sync_single_for_device = fast_smmu_sync_single_for_device, |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 632 | .map_sg = fast_smmu_map_sg, |
| 633 | .unmap_sg = fast_smmu_unmap_sg, |
Liam Mark | 78d7fb5 | 2016-12-01 13:05:31 -0800 | [diff] [blame^] | 634 | .sync_sg_for_cpu = fast_smmu_sync_sg_for_cpu, |
| 635 | .sync_sg_for_device = fast_smmu_sync_sg_for_device, |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 636 | .dma_supported = fast_smmu_dma_supported, |
| 637 | .mapping_error = fast_smmu_mapping_error, |
| 638 | }; |
| 639 | |
| 640 | /** |
| 641 | * __fast_smmu_create_mapping_sized |
| 642 | * @base: bottom of the VA range |
| 643 | * @size: size of the VA range in bytes |
| 644 | * |
| 645 | * Creates a mapping structure which holds information about used/unused IO |
| 646 | * address ranges, which is required to perform mapping with IOMMU aware |
| 647 | * functions. The only VA range supported is [0, 4GB). |
| 648 | * |
| 649 | * The client device need to be attached to the mapping with |
| 650 | * fast_smmu_attach_device function. |
| 651 | */ |
| 652 | static struct dma_fast_smmu_mapping *__fast_smmu_create_mapping_sized( |
| 653 | dma_addr_t base, size_t size) |
| 654 | { |
| 655 | struct dma_fast_smmu_mapping *fast; |
| 656 | |
| 657 | fast = kzalloc(sizeof(struct dma_fast_smmu_mapping), GFP_KERNEL); |
| 658 | if (!fast) |
| 659 | goto err; |
| 660 | |
| 661 | fast->base = base; |
| 662 | fast->size = size; |
| 663 | fast->num_4k_pages = size >> FAST_PAGE_SHIFT; |
| 664 | fast->bitmap_size = BITS_TO_LONGS(fast->num_4k_pages) * sizeof(long); |
| 665 | |
| 666 | fast->bitmap = kzalloc(fast->bitmap_size, GFP_KERNEL); |
| 667 | if (!fast->bitmap) |
| 668 | goto err2; |
| 669 | |
| 670 | spin_lock_init(&fast->lock); |
| 671 | |
| 672 | return fast; |
| 673 | err2: |
| 674 | kfree(fast); |
| 675 | err: |
| 676 | return ERR_PTR(-ENOMEM); |
| 677 | } |
| 678 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 679 | /** |
| 680 | * fast_smmu_attach_device |
| 681 | * @dev: valid struct device pointer |
| 682 | * @mapping: io address space mapping structure (returned from |
| 683 | * fast_smmu_create_mapping) |
| 684 | * |
| 685 | * Attaches specified io address space mapping to the provided device, |
| 686 | * this replaces the dma operations (dma_map_ops pointer) with the |
| 687 | * IOMMU aware version. More than one client might be attached to |
| 688 | * the same io address space mapping. |
| 689 | */ |
| 690 | int fast_smmu_attach_device(struct device *dev, |
| 691 | struct dma_iommu_mapping *mapping) |
| 692 | { |
| 693 | int atomic_domain = 1; |
| 694 | struct iommu_domain *domain = mapping->domain; |
| 695 | struct iommu_pgtbl_info info; |
| 696 | size_t size = mapping->bits << PAGE_SHIFT; |
| 697 | |
| 698 | if (mapping->base + size > (SZ_1G * 4ULL)) |
| 699 | return -EINVAL; |
| 700 | |
| 701 | if (iommu_domain_set_attr(domain, DOMAIN_ATTR_ATOMIC, |
| 702 | &atomic_domain)) |
| 703 | return -EINVAL; |
| 704 | |
| 705 | mapping->fast = __fast_smmu_create_mapping_sized(mapping->base, size); |
| 706 | if (IS_ERR(mapping->fast)) |
| 707 | return -ENOMEM; |
| 708 | mapping->fast->domain = domain; |
| 709 | mapping->fast->dev = dev; |
| 710 | |
| 711 | if (iommu_attach_device(domain, dev)) |
| 712 | return -EINVAL; |
| 713 | |
| 714 | if (iommu_domain_get_attr(domain, DOMAIN_ATTR_PGTBL_INFO, |
| 715 | &info)) { |
| 716 | dev_err(dev, "Couldn't get page table info\n"); |
| 717 | fast_smmu_detach_device(dev, mapping); |
| 718 | return -EINVAL; |
| 719 | } |
| 720 | mapping->fast->pgtbl_pmds = info.pmds; |
| 721 | |
Mitchel Humpherys | 5c704e0 | 2015-12-21 15:06:34 -0800 | [diff] [blame] | 722 | mapping->fast->notifier.notifier_call = fast_smmu_notify; |
| 723 | av8l_register_notify(&mapping->fast->notifier); |
| 724 | |
Mitchel Humpherys | 0e43f0a | 2015-10-08 15:03:09 -0700 | [diff] [blame] | 725 | dev->archdata.mapping = mapping; |
| 726 | set_dma_ops(dev, &fast_smmu_dma_ops); |
| 727 | |
| 728 | return 0; |
| 729 | } |
| 730 | EXPORT_SYMBOL(fast_smmu_attach_device); |
| 731 | |
| 732 | /** |
| 733 | * fast_smmu_detach_device |
| 734 | * @dev: valid struct device pointer |
| 735 | * |
| 736 | * Detaches the provided device from a previously attached map. |
| 737 | * This voids the dma operations (dma_map_ops pointer) |
| 738 | */ |
| 739 | void fast_smmu_detach_device(struct device *dev, |
| 740 | struct dma_iommu_mapping *mapping) |
| 741 | { |
| 742 | iommu_detach_device(mapping->domain, dev); |
| 743 | dev->archdata.mapping = NULL; |
| 744 | set_dma_ops(dev, NULL); |
| 745 | |
| 746 | kfree(mapping->fast->bitmap); |
| 747 | kfree(mapping->fast); |
| 748 | } |
| 749 | EXPORT_SYMBOL(fast_smmu_detach_device); |