| 1 | // SPDX-License-Identifier: GPL-2.0 | 
|---|
| 2 | /* | 
|---|
| 3 | * Copyright (C) 2018-2020 Christoph Hellwig. | 
|---|
| 4 | * | 
|---|
| 5 | * DMA operations that map physical memory directly without using an IOMMU. | 
|---|
| 6 | */ | 
|---|
| 7 | #include <linux/memblock.h> /* for max_pfn */ | 
|---|
| 8 | #include <linux/export.h> | 
|---|
| 9 | #include <linux/mm.h> | 
|---|
| 10 | #include <linux/dma-map-ops.h> | 
|---|
| 11 | #include <linux/scatterlist.h> | 
|---|
| 12 | #include <linux/pfn.h> | 
|---|
| 13 | #include <linux/vmalloc.h> | 
|---|
| 14 | #include <linux/set_memory.h> | 
|---|
| 15 | #include <linux/slab.h> | 
|---|
| 16 | #include <linux/pci-p2pdma.h> | 
|---|
| 17 | #include "direct.h" | 
|---|
| 18 |  | 
|---|
| 19 | /* | 
|---|
| 20 | * Most architectures use ZONE_DMA for the first 16 Megabytes, but some use | 
|---|
| 21 | * it for entirely different regions. In that case the arch code needs to | 
|---|
| 22 | * override the variable below for dma-direct to work properly. | 
|---|
| 23 | */ | 
|---|
| 24 | u64 zone_dma_limit __ro_after_init = DMA_BIT_MASK(24); | 
|---|
| 25 |  | 
|---|
| 26 | static inline dma_addr_t phys_to_dma_direct(struct device *dev, | 
|---|
| 27 | phys_addr_t phys) | 
|---|
| 28 | { | 
|---|
| 29 | if (force_dma_unencrypted(dev)) | 
|---|
| 30 | return phys_to_dma_unencrypted(dev, paddr: phys); | 
|---|
| 31 | return phys_to_dma(dev, paddr: phys); | 
|---|
| 32 | } | 
|---|
| 33 |  | 
|---|
| 34 | static inline struct page *dma_direct_to_page(struct device *dev, | 
|---|
| 35 | dma_addr_t dma_addr) | 
|---|
| 36 | { | 
|---|
| 37 | return pfn_to_page(PHYS_PFN(dma_to_phys(dev, dma_addr))); | 
|---|
| 38 | } | 
|---|
| 39 |  | 
|---|
| 40 | u64 dma_direct_get_required_mask(struct device *dev) | 
|---|
| 41 | { | 
|---|
| 42 | phys_addr_t phys = (phys_addr_t)(max_pfn - 1) << PAGE_SHIFT; | 
|---|
| 43 | u64 max_dma = phys_to_dma_direct(dev, phys); | 
|---|
| 44 |  | 
|---|
| 45 | return (1ULL << (fls64(x: max_dma) - 1)) * 2 - 1; | 
|---|
| 46 | } | 
|---|
| 47 |  | 
|---|
| 48 | static gfp_t dma_direct_optimal_gfp_mask(struct device *dev, u64 *phys_limit) | 
|---|
| 49 | { | 
|---|
| 50 | u64 dma_limit = min_not_zero( | 
|---|
| 51 | dev->coherent_dma_mask, | 
|---|
| 52 | dev->bus_dma_limit); | 
|---|
| 53 |  | 
|---|
| 54 | /* | 
|---|
| 55 | * Optimistically try the zone that the physical address mask falls | 
|---|
| 56 | * into first.  If that returns memory that isn't actually addressable | 
|---|
| 57 | * we will fallback to the next lower zone and try again. | 
|---|
| 58 | * | 
|---|
| 59 | * Note that GFP_DMA32 and GFP_DMA are no ops without the corresponding | 
|---|
| 60 | * zones. | 
|---|
| 61 | */ | 
|---|
| 62 | *phys_limit = dma_to_phys(dev, dma_addr: dma_limit); | 
|---|
| 63 | if (*phys_limit <= zone_dma_limit) | 
|---|
| 64 | return GFP_DMA; | 
|---|
| 65 | if (*phys_limit <= DMA_BIT_MASK(32)) | 
|---|
| 66 | return GFP_DMA32; | 
|---|
| 67 | return 0; | 
|---|
| 68 | } | 
|---|
| 69 |  | 
|---|
| 70 | bool dma_coherent_ok(struct device *dev, phys_addr_t phys, size_t size) | 
|---|
| 71 | { | 
|---|
| 72 | dma_addr_t dma_addr = phys_to_dma_direct(dev, phys); | 
|---|
| 73 |  | 
|---|
| 74 | if (dma_addr == DMA_MAPPING_ERROR) | 
|---|
| 75 | return false; | 
|---|
| 76 | return dma_addr + size - 1 <= | 
|---|
| 77 | min_not_zero(dev->coherent_dma_mask, dev->bus_dma_limit); | 
|---|
| 78 | } | 
|---|
| 79 |  | 
|---|
| 80 | static int dma_set_decrypted(struct device *dev, void *vaddr, size_t size) | 
|---|
| 81 | { | 
|---|
| 82 | if (!force_dma_unencrypted(dev)) | 
|---|
| 83 | return 0; | 
|---|
| 84 | return set_memory_decrypted(addr: (unsigned long)vaddr, PFN_UP(size)); | 
|---|
| 85 | } | 
|---|
| 86 |  | 
|---|
| 87 | static int dma_set_encrypted(struct device *dev, void *vaddr, size_t size) | 
|---|
| 88 | { | 
|---|
| 89 | int ret; | 
|---|
| 90 |  | 
|---|
| 91 | if (!force_dma_unencrypted(dev)) | 
|---|
| 92 | return 0; | 
|---|
| 93 | ret = set_memory_encrypted(addr: (unsigned long)vaddr, PFN_UP(size)); | 
|---|
| 94 | if (ret) | 
|---|
| 95 | pr_warn_ratelimited( "leaking DMA memory that can't be re-encrypted\n"); | 
|---|
| 96 | return ret; | 
|---|
| 97 | } | 
|---|
| 98 |  | 
|---|
| 99 | static void __dma_direct_free_pages(struct device *dev, struct page *page, | 
|---|
| 100 | size_t size) | 
|---|
| 101 | { | 
|---|
| 102 | if (swiotlb_free(dev, page, size)) | 
|---|
| 103 | return; | 
|---|
| 104 | dma_free_contiguous(dev, page, size); | 
|---|
| 105 | } | 
|---|
| 106 |  | 
|---|
| 107 | static struct page *dma_direct_alloc_swiotlb(struct device *dev, size_t size) | 
|---|
| 108 | { | 
|---|
| 109 | struct page *page = swiotlb_alloc(dev, size); | 
|---|
| 110 |  | 
|---|
| 111 | if (page && !dma_coherent_ok(dev, page_to_phys(page), size)) { | 
|---|
| 112 | swiotlb_free(dev, page, size); | 
|---|
| 113 | return NULL; | 
|---|
| 114 | } | 
|---|
| 115 |  | 
|---|
| 116 | return page; | 
|---|
| 117 | } | 
|---|
| 118 |  | 
|---|
| 119 | static struct page *__dma_direct_alloc_pages(struct device *dev, size_t size, | 
|---|
| 120 | gfp_t gfp, bool allow_highmem) | 
|---|
| 121 | { | 
|---|
| 122 | int node = dev_to_node(dev); | 
|---|
| 123 | struct page *page; | 
|---|
| 124 | u64 phys_limit; | 
|---|
| 125 |  | 
|---|
| 126 | WARN_ON_ONCE(!PAGE_ALIGNED(size)); | 
|---|
| 127 |  | 
|---|
| 128 | if (is_swiotlb_for_alloc(dev)) | 
|---|
| 129 | return dma_direct_alloc_swiotlb(dev, size); | 
|---|
| 130 |  | 
|---|
| 131 | gfp |= dma_direct_optimal_gfp_mask(dev, phys_limit: &phys_limit); | 
|---|
| 132 | page = dma_alloc_contiguous(dev, size, gfp); | 
|---|
| 133 | if (page) { | 
|---|
| 134 | if (dma_coherent_ok(dev, page_to_phys(page), size) && | 
|---|
| 135 | (allow_highmem || !PageHighMem(page))) | 
|---|
| 136 | return page; | 
|---|
| 137 |  | 
|---|
| 138 | dma_free_contiguous(dev, page, size); | 
|---|
| 139 | } | 
|---|
| 140 |  | 
|---|
| 141 | while ((page = alloc_pages_node(node, gfp, get_order(size))) | 
|---|
| 142 | && !dma_coherent_ok(dev, page_to_phys(page), size)) { | 
|---|
| 143 | __free_pages(page, order: get_order(size)); | 
|---|
| 144 |  | 
|---|
| 145 | if (IS_ENABLED(CONFIG_ZONE_DMA32) && | 
|---|
| 146 | phys_limit < DMA_BIT_MASK(64) && | 
|---|
| 147 | !(gfp & (GFP_DMA32 | GFP_DMA))) | 
|---|
| 148 | gfp |= GFP_DMA32; | 
|---|
| 149 | else if (IS_ENABLED(CONFIG_ZONE_DMA) && !(gfp & GFP_DMA)) | 
|---|
| 150 | gfp = (gfp & ~GFP_DMA32) | GFP_DMA; | 
|---|
| 151 | else | 
|---|
| 152 | return NULL; | 
|---|
| 153 | } | 
|---|
| 154 |  | 
|---|
| 155 | return page; | 
|---|
| 156 | } | 
|---|
| 157 |  | 
|---|
| 158 | /* | 
|---|
| 159 | * Check if a potentially blocking operations needs to dip into the atomic | 
|---|
| 160 | * pools for the given device/gfp. | 
|---|
| 161 | */ | 
|---|
| 162 | static bool dma_direct_use_pool(struct device *dev, gfp_t gfp) | 
|---|
| 163 | { | 
|---|
| 164 | return !gfpflags_allow_blocking(gfp_flags: gfp) && !is_swiotlb_for_alloc(dev); | 
|---|
| 165 | } | 
|---|
| 166 |  | 
|---|
| 167 | static void *dma_direct_alloc_from_pool(struct device *dev, size_t size, | 
|---|
| 168 | dma_addr_t *dma_handle, gfp_t gfp) | 
|---|
| 169 | { | 
|---|
| 170 | struct page *page; | 
|---|
| 171 | u64 phys_limit; | 
|---|
| 172 | void *ret; | 
|---|
| 173 |  | 
|---|
| 174 | if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_DMA_COHERENT_POOL))) | 
|---|
| 175 | return NULL; | 
|---|
| 176 |  | 
|---|
| 177 | gfp |= dma_direct_optimal_gfp_mask(dev, phys_limit: &phys_limit); | 
|---|
| 178 | page = dma_alloc_from_pool(dev, size, cpu_addr: &ret, flags: gfp, phys_addr_ok: dma_coherent_ok); | 
|---|
| 179 | if (!page) | 
|---|
| 180 | return NULL; | 
|---|
| 181 | *dma_handle = phys_to_dma_direct(dev, page_to_phys(page)); | 
|---|
| 182 | return ret; | 
|---|
| 183 | } | 
|---|
| 184 |  | 
|---|
| 185 | static void *dma_direct_alloc_no_mapping(struct device *dev, size_t size, | 
|---|
| 186 | dma_addr_t *dma_handle, gfp_t gfp) | 
|---|
| 187 | { | 
|---|
| 188 | struct page *page; | 
|---|
| 189 |  | 
|---|
| 190 | page = __dma_direct_alloc_pages(dev, size, gfp: gfp & ~__GFP_ZERO, allow_highmem: true); | 
|---|
| 191 | if (!page) | 
|---|
| 192 | return NULL; | 
|---|
| 193 |  | 
|---|
| 194 | /* remove any dirty cache lines on the kernel alias */ | 
|---|
| 195 | if (!PageHighMem(page)) | 
|---|
| 196 | arch_dma_prep_coherent(page, size); | 
|---|
| 197 |  | 
|---|
| 198 | /* return the page pointer as the opaque cookie */ | 
|---|
| 199 | *dma_handle = phys_to_dma_direct(dev, page_to_phys(page)); | 
|---|
| 200 | return page; | 
|---|
| 201 | } | 
|---|
| 202 |  | 
|---|
| 203 | void *dma_direct_alloc(struct device *dev, size_t size, | 
|---|
| 204 | dma_addr_t *dma_handle, gfp_t gfp, unsigned long attrs) | 
|---|
| 205 | { | 
|---|
| 206 | bool remap = false, set_uncached = false; | 
|---|
| 207 | struct page *page; | 
|---|
| 208 | void *ret; | 
|---|
| 209 |  | 
|---|
| 210 | size = PAGE_ALIGN(size); | 
|---|
| 211 | if (attrs & DMA_ATTR_NO_WARN) | 
|---|
| 212 | gfp |= __GFP_NOWARN; | 
|---|
| 213 |  | 
|---|
| 214 | if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) && | 
|---|
| 215 | !force_dma_unencrypted(dev) && !is_swiotlb_for_alloc(dev)) | 
|---|
| 216 | return dma_direct_alloc_no_mapping(dev, size, dma_handle, gfp); | 
|---|
| 217 |  | 
|---|
| 218 | if (!dev_is_dma_coherent(dev)) { | 
|---|
| 219 | if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALLOC) && | 
|---|
| 220 | !is_swiotlb_for_alloc(dev)) | 
|---|
| 221 | return arch_dma_alloc(dev, size, dma_handle, gfp, | 
|---|
| 222 | attrs); | 
|---|
| 223 |  | 
|---|
| 224 | /* | 
|---|
| 225 | * If there is a global pool, always allocate from it for | 
|---|
| 226 | * non-coherent devices. | 
|---|
| 227 | */ | 
|---|
| 228 | if (IS_ENABLED(CONFIG_DMA_GLOBAL_POOL)) | 
|---|
| 229 | return dma_alloc_from_global_coherent(dev, size, | 
|---|
| 230 | dma_handle); | 
|---|
| 231 |  | 
|---|
| 232 | /* | 
|---|
| 233 | * Otherwise we require the architecture to either be able to | 
|---|
| 234 | * mark arbitrary parts of the kernel direct mapping uncached, | 
|---|
| 235 | * or remapped it uncached. | 
|---|
| 236 | */ | 
|---|
| 237 | set_uncached = IS_ENABLED(CONFIG_ARCH_HAS_DMA_SET_UNCACHED); | 
|---|
| 238 | remap = IS_ENABLED(CONFIG_DMA_DIRECT_REMAP); | 
|---|
| 239 | if (!set_uncached && !remap) { | 
|---|
| 240 | pr_warn_once( "coherent DMA allocations not supported on this platform.\n"); | 
|---|
| 241 | return NULL; | 
|---|
| 242 | } | 
|---|
| 243 | } | 
|---|
| 244 |  | 
|---|
| 245 | /* | 
|---|
| 246 | * Remapping or decrypting memory may block, allocate the memory from | 
|---|
| 247 | * the atomic pools instead if we aren't allowed block. | 
|---|
| 248 | */ | 
|---|
| 249 | if ((remap || force_dma_unencrypted(dev)) && | 
|---|
| 250 | dma_direct_use_pool(dev, gfp)) | 
|---|
| 251 | return dma_direct_alloc_from_pool(dev, size, dma_handle, gfp); | 
|---|
| 252 |  | 
|---|
| 253 | /* we always manually zero the memory once we are done */ | 
|---|
| 254 | page = __dma_direct_alloc_pages(dev, size, gfp: gfp & ~__GFP_ZERO, allow_highmem: true); | 
|---|
| 255 | if (!page) | 
|---|
| 256 | return NULL; | 
|---|
| 257 |  | 
|---|
| 258 | /* | 
|---|
| 259 | * dma_alloc_contiguous can return highmem pages depending on a | 
|---|
| 260 | * combination the cma= arguments and per-arch setup.  These need to be | 
|---|
| 261 | * remapped to return a kernel virtual address. | 
|---|
| 262 | */ | 
|---|
| 263 | if (PageHighMem(page)) { | 
|---|
| 264 | remap = true; | 
|---|
| 265 | set_uncached = false; | 
|---|
| 266 | } | 
|---|
| 267 |  | 
|---|
| 268 | if (remap) { | 
|---|
| 269 | pgprot_t prot = dma_pgprot(dev, PAGE_KERNEL, attrs); | 
|---|
| 270 |  | 
|---|
| 271 | if (force_dma_unencrypted(dev)) | 
|---|
| 272 | prot = pgprot_decrypted(prot); | 
|---|
| 273 |  | 
|---|
| 274 | /* remove any dirty cache lines on the kernel alias */ | 
|---|
| 275 | arch_dma_prep_coherent(page, size); | 
|---|
| 276 |  | 
|---|
| 277 | /* create a coherent mapping */ | 
|---|
| 278 | ret = dma_common_contiguous_remap(page, size, prot, | 
|---|
| 279 | caller: __builtin_return_address(0)); | 
|---|
| 280 | if (!ret) | 
|---|
| 281 | goto out_free_pages; | 
|---|
| 282 | } else { | 
|---|
| 283 | ret = page_address(page); | 
|---|
| 284 | if (dma_set_decrypted(dev, vaddr: ret, size)) | 
|---|
| 285 | goto out_leak_pages; | 
|---|
| 286 | } | 
|---|
| 287 |  | 
|---|
| 288 | memset(s: ret, c: 0, n: size); | 
|---|
| 289 |  | 
|---|
| 290 | if (set_uncached) { | 
|---|
| 291 | arch_dma_prep_coherent(page, size); | 
|---|
| 292 | ret = arch_dma_set_uncached(addr: ret, size); | 
|---|
| 293 | if (IS_ERR(ptr: ret)) | 
|---|
| 294 | goto out_encrypt_pages; | 
|---|
| 295 | } | 
|---|
| 296 |  | 
|---|
| 297 | *dma_handle = phys_to_dma_direct(dev, page_to_phys(page)); | 
|---|
| 298 | return ret; | 
|---|
| 299 |  | 
|---|
| 300 | out_encrypt_pages: | 
|---|
| 301 | if (dma_set_encrypted(dev, page_address(page), size)) | 
|---|
| 302 | return NULL; | 
|---|
| 303 | out_free_pages: | 
|---|
| 304 | __dma_direct_free_pages(dev, page, size); | 
|---|
| 305 | return NULL; | 
|---|
| 306 | out_leak_pages: | 
|---|
| 307 | return NULL; | 
|---|
| 308 | } | 
|---|
| 309 |  | 
|---|
| 310 | void dma_direct_free(struct device *dev, size_t size, | 
|---|
| 311 | void *cpu_addr, dma_addr_t dma_addr, unsigned long attrs) | 
|---|
| 312 | { | 
|---|
| 313 | unsigned int page_order = get_order(size); | 
|---|
| 314 |  | 
|---|
| 315 | if ((attrs & DMA_ATTR_NO_KERNEL_MAPPING) && | 
|---|
| 316 | !force_dma_unencrypted(dev) && !is_swiotlb_for_alloc(dev)) { | 
|---|
| 317 | /* cpu_addr is a struct page cookie, not a kernel address */ | 
|---|
| 318 | dma_free_contiguous(dev, page: cpu_addr, size); | 
|---|
| 319 | return; | 
|---|
| 320 | } | 
|---|
| 321 |  | 
|---|
| 322 | if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_ALLOC) && | 
|---|
| 323 | !dev_is_dma_coherent(dev) && | 
|---|
| 324 | !is_swiotlb_for_alloc(dev)) { | 
|---|
| 325 | arch_dma_free(dev, size, cpu_addr, dma_addr, attrs); | 
|---|
| 326 | return; | 
|---|
| 327 | } | 
|---|
| 328 |  | 
|---|
| 329 | if (IS_ENABLED(CONFIG_DMA_GLOBAL_POOL) && | 
|---|
| 330 | !dev_is_dma_coherent(dev)) { | 
|---|
| 331 | if (!dma_release_from_global_coherent(order: page_order, vaddr: cpu_addr)) | 
|---|
| 332 | WARN_ON_ONCE(1); | 
|---|
| 333 | return; | 
|---|
| 334 | } | 
|---|
| 335 |  | 
|---|
| 336 | /* If cpu_addr is not from an atomic pool, dma_free_from_pool() fails */ | 
|---|
| 337 | if (IS_ENABLED(CONFIG_DMA_COHERENT_POOL) && | 
|---|
| 338 | dma_free_from_pool(dev, start: cpu_addr, PAGE_ALIGN(size))) | 
|---|
| 339 | return; | 
|---|
| 340 |  | 
|---|
| 341 | if (is_vmalloc_addr(x: cpu_addr)) { | 
|---|
| 342 | vunmap(addr: cpu_addr); | 
|---|
| 343 | } else { | 
|---|
| 344 | if (IS_ENABLED(CONFIG_ARCH_HAS_DMA_CLEAR_UNCACHED)) | 
|---|
| 345 | arch_dma_clear_uncached(addr: cpu_addr, size); | 
|---|
| 346 | if (dma_set_encrypted(dev, vaddr: cpu_addr, size)) | 
|---|
| 347 | return; | 
|---|
| 348 | } | 
|---|
| 349 |  | 
|---|
| 350 | __dma_direct_free_pages(dev, page: dma_direct_to_page(dev, dma_addr), size); | 
|---|
| 351 | } | 
|---|
| 352 |  | 
|---|
| 353 | struct page *dma_direct_alloc_pages(struct device *dev, size_t size, | 
|---|
| 354 | dma_addr_t *dma_handle, enum dma_data_direction dir, gfp_t gfp) | 
|---|
| 355 | { | 
|---|
| 356 | struct page *page; | 
|---|
| 357 | void *ret; | 
|---|
| 358 |  | 
|---|
| 359 | if (force_dma_unencrypted(dev) && dma_direct_use_pool(dev, gfp)) | 
|---|
| 360 | return dma_direct_alloc_from_pool(dev, size, dma_handle, gfp); | 
|---|
| 361 |  | 
|---|
| 362 | page = __dma_direct_alloc_pages(dev, size, gfp, allow_highmem: false); | 
|---|
| 363 | if (!page) | 
|---|
| 364 | return NULL; | 
|---|
| 365 |  | 
|---|
| 366 | ret = page_address(page); | 
|---|
| 367 | if (dma_set_decrypted(dev, vaddr: ret, size)) | 
|---|
| 368 | goto out_leak_pages; | 
|---|
| 369 | memset(s: ret, c: 0, n: size); | 
|---|
| 370 | *dma_handle = phys_to_dma_direct(dev, page_to_phys(page)); | 
|---|
| 371 | return page; | 
|---|
| 372 | out_leak_pages: | 
|---|
| 373 | return NULL; | 
|---|
| 374 | } | 
|---|
| 375 |  | 
|---|
| 376 | void dma_direct_free_pages(struct device *dev, size_t size, | 
|---|
| 377 | struct page *page, dma_addr_t dma_addr, | 
|---|
| 378 | enum dma_data_direction dir) | 
|---|
| 379 | { | 
|---|
| 380 | void *vaddr = page_address(page); | 
|---|
| 381 |  | 
|---|
| 382 | /* If cpu_addr is not from an atomic pool, dma_free_from_pool() fails */ | 
|---|
| 383 | if (IS_ENABLED(CONFIG_DMA_COHERENT_POOL) && | 
|---|
| 384 | dma_free_from_pool(dev, start: vaddr, size)) | 
|---|
| 385 | return; | 
|---|
| 386 |  | 
|---|
| 387 | if (dma_set_encrypted(dev, vaddr, size)) | 
|---|
| 388 | return; | 
|---|
| 389 | __dma_direct_free_pages(dev, page, size); | 
|---|
| 390 | } | 
|---|
| 391 |  | 
|---|
| 392 | #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \ | 
|---|
| 393 | defined(CONFIG_SWIOTLB) | 
|---|
| 394 | void dma_direct_sync_sg_for_device(struct device *dev, | 
|---|
| 395 | struct scatterlist *sgl, int nents, enum dma_data_direction dir) | 
|---|
| 396 | { | 
|---|
| 397 | struct scatterlist *sg; | 
|---|
| 398 | int i; | 
|---|
| 399 |  | 
|---|
| 400 | for_each_sg(sgl, sg, nents, i) { | 
|---|
| 401 | phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg)); | 
|---|
| 402 |  | 
|---|
| 403 | swiotlb_sync_single_for_device(dev, addr: paddr, size: sg->length, dir); | 
|---|
| 404 |  | 
|---|
| 405 | if (!dev_is_dma_coherent(dev)) | 
|---|
| 406 | arch_sync_dma_for_device(paddr, size: sg->length, | 
|---|
| 407 | dir); | 
|---|
| 408 | } | 
|---|
| 409 | } | 
|---|
| 410 | #endif | 
|---|
| 411 |  | 
|---|
| 412 | #if defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \ | 
|---|
| 413 | defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL) || \ | 
|---|
| 414 | defined(CONFIG_SWIOTLB) | 
|---|
| 415 | void dma_direct_sync_sg_for_cpu(struct device *dev, | 
|---|
| 416 | struct scatterlist *sgl, int nents, enum dma_data_direction dir) | 
|---|
| 417 | { | 
|---|
| 418 | struct scatterlist *sg; | 
|---|
| 419 | int i; | 
|---|
| 420 |  | 
|---|
| 421 | for_each_sg(sgl, sg, nents, i) { | 
|---|
| 422 | phys_addr_t paddr = dma_to_phys(dev, sg_dma_address(sg)); | 
|---|
| 423 |  | 
|---|
| 424 | if (!dev_is_dma_coherent(dev)) | 
|---|
| 425 | arch_sync_dma_for_cpu(paddr, size: sg->length, dir); | 
|---|
| 426 |  | 
|---|
| 427 | swiotlb_sync_single_for_cpu(dev, addr: paddr, size: sg->length, dir); | 
|---|
| 428 |  | 
|---|
| 429 | if (dir == DMA_FROM_DEVICE) | 
|---|
| 430 | arch_dma_mark_clean(paddr, size: sg->length); | 
|---|
| 431 | } | 
|---|
| 432 |  | 
|---|
| 433 | if (!dev_is_dma_coherent(dev)) | 
|---|
| 434 | arch_sync_dma_for_cpu_all(); | 
|---|
| 435 | } | 
|---|
| 436 |  | 
|---|
| 437 | /* | 
|---|
| 438 | * Unmaps segments, except for ones marked as pci_p2pdma which do not | 
|---|
| 439 | * require any further action as they contain a bus address. | 
|---|
| 440 | */ | 
|---|
| 441 | void dma_direct_unmap_sg(struct device *dev, struct scatterlist *sgl, | 
|---|
| 442 | int nents, enum dma_data_direction dir, unsigned long attrs) | 
|---|
| 443 | { | 
|---|
| 444 | struct scatterlist *sg; | 
|---|
| 445 | int i; | 
|---|
| 446 |  | 
|---|
| 447 | for_each_sg(sgl,  sg, nents, i) { | 
|---|
| 448 | if (sg_dma_is_bus_address(sg)) | 
|---|
| 449 | sg_dma_unmark_bus_address(sg); | 
|---|
| 450 | else | 
|---|
| 451 | dma_direct_unmap_phys(dev, addr: sg->dma_address, | 
|---|
| 452 | sg_dma_len(sg), dir, attrs); | 
|---|
| 453 | } | 
|---|
| 454 | } | 
|---|
| 455 | #endif | 
|---|
| 456 |  | 
|---|
| 457 | int dma_direct_map_sg(struct device *dev, struct scatterlist *sgl, int nents, | 
|---|
| 458 | enum dma_data_direction dir, unsigned long attrs) | 
|---|
| 459 | { | 
|---|
| 460 | struct pci_p2pdma_map_state p2pdma_state = {}; | 
|---|
| 461 | struct scatterlist *sg; | 
|---|
| 462 | int i, ret; | 
|---|
| 463 |  | 
|---|
| 464 | for_each_sg(sgl, sg, nents, i) { | 
|---|
| 465 | switch (pci_p2pdma_state(state: &p2pdma_state, dev, page: sg_page(sg))) { | 
|---|
| 466 | case PCI_P2PDMA_MAP_THRU_HOST_BRIDGE: | 
|---|
| 467 | /* | 
|---|
| 468 | * Any P2P mapping that traverses the PCI host bridge | 
|---|
| 469 | * must be mapped with CPU physical address and not PCI | 
|---|
| 470 | * bus addresses. | 
|---|
| 471 | */ | 
|---|
| 472 | break; | 
|---|
| 473 | case PCI_P2PDMA_MAP_NONE: | 
|---|
| 474 | sg->dma_address = dma_direct_map_phys(dev, phys: sg_phys(sg), | 
|---|
| 475 | size: sg->length, dir, attrs); | 
|---|
| 476 | if (sg->dma_address == DMA_MAPPING_ERROR) { | 
|---|
| 477 | ret = -EIO; | 
|---|
| 478 | goto out_unmap; | 
|---|
| 479 | } | 
|---|
| 480 | break; | 
|---|
| 481 | case PCI_P2PDMA_MAP_BUS_ADDR: | 
|---|
| 482 | sg->dma_address = pci_p2pdma_bus_addr_map(state: &p2pdma_state, | 
|---|
| 483 | paddr: sg_phys(sg)); | 
|---|
| 484 | sg_dma_mark_bus_address(sg); | 
|---|
| 485 | continue; | 
|---|
| 486 | default: | 
|---|
| 487 | ret = -EREMOTEIO; | 
|---|
| 488 | goto out_unmap; | 
|---|
| 489 | } | 
|---|
| 490 | sg_dma_len(sg) = sg->length; | 
|---|
| 491 | } | 
|---|
| 492 |  | 
|---|
| 493 | return nents; | 
|---|
| 494 |  | 
|---|
| 495 | out_unmap: | 
|---|
| 496 | dma_direct_unmap_sg(dev, sgl, nents: i, dir, attrs: attrs | DMA_ATTR_SKIP_CPU_SYNC); | 
|---|
| 497 | return ret; | 
|---|
| 498 | } | 
|---|
| 499 |  | 
|---|
| 500 | int dma_direct_get_sgtable(struct device *dev, struct sg_table *sgt, | 
|---|
| 501 | void *cpu_addr, dma_addr_t dma_addr, size_t size, | 
|---|
| 502 | unsigned long attrs) | 
|---|
| 503 | { | 
|---|
| 504 | struct page *page = dma_direct_to_page(dev, dma_addr); | 
|---|
| 505 | int ret; | 
|---|
| 506 |  | 
|---|
| 507 | ret = sg_alloc_table(sgt, 1, GFP_KERNEL); | 
|---|
| 508 | if (!ret) | 
|---|
| 509 | sg_set_page(sg: sgt->sgl, page, PAGE_ALIGN(size), offset: 0); | 
|---|
| 510 | return ret; | 
|---|
| 511 | } | 
|---|
| 512 |  | 
|---|
| 513 | bool dma_direct_can_mmap(struct device *dev) | 
|---|
| 514 | { | 
|---|
| 515 | return dev_is_dma_coherent(dev) || | 
|---|
| 516 | IS_ENABLED(CONFIG_DMA_NONCOHERENT_MMAP); | 
|---|
| 517 | } | 
|---|
| 518 |  | 
|---|
| 519 | int dma_direct_mmap(struct device *dev, struct vm_area_struct *vma, | 
|---|
| 520 | void *cpu_addr, dma_addr_t dma_addr, size_t size, | 
|---|
| 521 | unsigned long attrs) | 
|---|
| 522 | { | 
|---|
| 523 | unsigned long user_count = vma_pages(vma); | 
|---|
| 524 | unsigned long count = PAGE_ALIGN(size) >> PAGE_SHIFT; | 
|---|
| 525 | unsigned long pfn = PHYS_PFN(dma_to_phys(dev, dma_addr)); | 
|---|
| 526 | int ret = -ENXIO; | 
|---|
| 527 |  | 
|---|
| 528 | vma->vm_page_prot = dma_pgprot(dev, prot: vma->vm_page_prot, attrs); | 
|---|
| 529 | if (force_dma_unencrypted(dev)) | 
|---|
| 530 | vma->vm_page_prot = pgprot_decrypted(vma->vm_page_prot); | 
|---|
| 531 |  | 
|---|
| 532 | if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret)) | 
|---|
| 533 | return ret; | 
|---|
| 534 | if (dma_mmap_from_global_coherent(vma, cpu_addr, size, ret: &ret)) | 
|---|
| 535 | return ret; | 
|---|
| 536 |  | 
|---|
| 537 | if (vma->vm_pgoff >= count || user_count > count - vma->vm_pgoff) | 
|---|
| 538 | return -ENXIO; | 
|---|
| 539 | return remap_pfn_range(vma, addr: vma->vm_start, pfn: pfn + vma->vm_pgoff, | 
|---|
| 540 | size: user_count << PAGE_SHIFT, vma->vm_page_prot); | 
|---|
| 541 | } | 
|---|
| 542 |  | 
|---|
| 543 | int dma_direct_supported(struct device *dev, u64 mask) | 
|---|
| 544 | { | 
|---|
| 545 | u64 min_mask = (max_pfn - 1) << PAGE_SHIFT; | 
|---|
| 546 |  | 
|---|
| 547 | /* | 
|---|
| 548 | * Because 32-bit DMA masks are so common we expect every architecture | 
|---|
| 549 | * to be able to satisfy them - either by not supporting more physical | 
|---|
| 550 | * memory, or by providing a ZONE_DMA32.  If neither is the case, the | 
|---|
| 551 | * architecture needs to use an IOMMU instead of the direct mapping. | 
|---|
| 552 | */ | 
|---|
| 553 | if (mask >= DMA_BIT_MASK(32)) | 
|---|
| 554 | return 1; | 
|---|
| 555 |  | 
|---|
| 556 | /* | 
|---|
| 557 | * This check needs to be against the actual bit mask value, so use | 
|---|
| 558 | * phys_to_dma_unencrypted() here so that the SME encryption mask isn't | 
|---|
| 559 | * part of the check. | 
|---|
| 560 | */ | 
|---|
| 561 | if (IS_ENABLED(CONFIG_ZONE_DMA)) | 
|---|
| 562 | min_mask = min_t(u64, min_mask, zone_dma_limit); | 
|---|
| 563 | return mask >= phys_to_dma_unencrypted(dev, paddr: min_mask); | 
|---|
| 564 | } | 
|---|
| 565 |  | 
|---|
| 566 | static const struct bus_dma_region *dma_find_range(struct device *dev, | 
|---|
| 567 | unsigned long start_pfn) | 
|---|
| 568 | { | 
|---|
| 569 | const struct bus_dma_region *m; | 
|---|
| 570 |  | 
|---|
| 571 | for (m = dev->dma_range_map; PFN_DOWN(m->size); m++) { | 
|---|
| 572 | unsigned long cpu_start_pfn = PFN_DOWN(m->cpu_start); | 
|---|
| 573 |  | 
|---|
| 574 | if (start_pfn >= cpu_start_pfn && | 
|---|
| 575 | start_pfn - cpu_start_pfn < PFN_DOWN(m->size)) | 
|---|
| 576 | return m; | 
|---|
| 577 | } | 
|---|
| 578 |  | 
|---|
| 579 | return NULL; | 
|---|
| 580 | } | 
|---|
| 581 |  | 
|---|
| 582 | /* | 
|---|
| 583 | * To check whether all ram resource ranges are covered by dma range map | 
|---|
| 584 | * Returns 0 when further check is needed | 
|---|
| 585 | * Returns 1 if there is some RAM range can't be covered by dma_range_map | 
|---|
| 586 | */ | 
|---|
| 587 | static int check_ram_in_range_map(unsigned long start_pfn, | 
|---|
| 588 | unsigned long nr_pages, void *data) | 
|---|
| 589 | { | 
|---|
| 590 | unsigned long end_pfn = start_pfn + nr_pages; | 
|---|
| 591 | struct device *dev = data; | 
|---|
| 592 |  | 
|---|
| 593 | while (start_pfn < end_pfn) { | 
|---|
| 594 | const struct bus_dma_region *bdr; | 
|---|
| 595 |  | 
|---|
| 596 | bdr = dma_find_range(dev, start_pfn); | 
|---|
| 597 | if (!bdr) | 
|---|
| 598 | return 1; | 
|---|
| 599 |  | 
|---|
| 600 | start_pfn = PFN_DOWN(bdr->cpu_start) + PFN_DOWN(bdr->size); | 
|---|
| 601 | } | 
|---|
| 602 |  | 
|---|
| 603 | return 0; | 
|---|
| 604 | } | 
|---|
| 605 |  | 
|---|
| 606 | bool dma_direct_all_ram_mapped(struct device *dev) | 
|---|
| 607 | { | 
|---|
| 608 | if (!dev->dma_range_map) | 
|---|
| 609 | return true; | 
|---|
| 610 | return !walk_system_ram_range(start_pfn: 0, PFN_DOWN(ULONG_MAX) + 1, arg: dev, | 
|---|
| 611 | func: check_ram_in_range_map); | 
|---|
| 612 | } | 
|---|
| 613 |  | 
|---|
| 614 | size_t dma_direct_max_mapping_size(struct device *dev) | 
|---|
| 615 | { | 
|---|
| 616 | /* If SWIOTLB is active, use its maximum mapping size */ | 
|---|
| 617 | if (is_swiotlb_active(dev) && | 
|---|
| 618 | (dma_addressing_limited(dev) || is_swiotlb_force_bounce(dev))) | 
|---|
| 619 | return swiotlb_max_mapping_size(dev); | 
|---|
| 620 | return SIZE_MAX; | 
|---|
| 621 | } | 
|---|
| 622 |  | 
|---|
| 623 | bool dma_direct_need_sync(struct device *dev, dma_addr_t dma_addr) | 
|---|
| 624 | { | 
|---|
| 625 | return !dev_is_dma_coherent(dev) || | 
|---|
| 626 | swiotlb_find_pool(dev, paddr: dma_to_phys(dev, dma_addr)); | 
|---|
| 627 | } | 
|---|
| 628 |  | 
|---|
| 629 | /** | 
|---|
| 630 | * dma_direct_set_offset - Assign scalar offset for a single DMA range. | 
|---|
| 631 | * @dev:	device pointer; needed to "own" the alloced memory. | 
|---|
| 632 | * @cpu_start:  beginning of memory region covered by this offset. | 
|---|
| 633 | * @dma_start:  beginning of DMA/PCI region covered by this offset. | 
|---|
| 634 | * @size:	size of the region. | 
|---|
| 635 | * | 
|---|
| 636 | * This is for the simple case of a uniform offset which cannot | 
|---|
| 637 | * be discovered by "dma-ranges". | 
|---|
| 638 | * | 
|---|
| 639 | * It returns -ENOMEM if out of memory, -EINVAL if a map | 
|---|
| 640 | * already exists, 0 otherwise. | 
|---|
| 641 | * | 
|---|
| 642 | * Note: any call to this from a driver is a bug.  The mapping needs | 
|---|
| 643 | * to be described by the device tree or other firmware interfaces. | 
|---|
| 644 | */ | 
|---|
| 645 | int dma_direct_set_offset(struct device *dev, phys_addr_t cpu_start, | 
|---|
| 646 | dma_addr_t dma_start, u64 size) | 
|---|
| 647 | { | 
|---|
| 648 | struct bus_dma_region *map; | 
|---|
| 649 | u64 offset = (u64)cpu_start - (u64)dma_start; | 
|---|
| 650 |  | 
|---|
| 651 | if (dev->dma_range_map) { | 
|---|
| 652 | dev_err(dev, "attempt to add DMA range to existing map\n"); | 
|---|
| 653 | return -EINVAL; | 
|---|
| 654 | } | 
|---|
| 655 |  | 
|---|
| 656 | if (!offset) | 
|---|
| 657 | return 0; | 
|---|
| 658 |  | 
|---|
| 659 | map = kcalloc(2, sizeof(*map), GFP_KERNEL); | 
|---|
| 660 | if (!map) | 
|---|
| 661 | return -ENOMEM; | 
|---|
| 662 | map[0].cpu_start = cpu_start; | 
|---|
| 663 | map[0].dma_start = dma_start; | 
|---|
| 664 | map[0].size = size; | 
|---|
| 665 | dev->dma_range_map = map; | 
|---|
| 666 | return 0; | 
|---|
| 667 | } | 
|---|
| 668 |  | 
|---|