1 /* 2 * SWIOTLB-based DMA API implementation 3 * 4 * Copyright (C) 2012 ARM Ltd. 5 * Author: Catalin Marinas <catalin.marinas@arm.com> 6 * 7 * This program is free software; you can redistribute it and/or modify 8 * it under the terms of the GNU General Public License version 2 as 9 * published by the Free Software Foundation. 10 * 11 * This program is distributed in the hope that it will be useful, 12 * but WITHOUT ANY WARRANTY; without even the implied warranty of 13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 14 * GNU General Public License for more details. 15 * 16 * You should have received a copy of the GNU General Public License 17 * along with this program. If not, see <http://www.gnu.org/licenses/>. 18 */ 19 20 #include <linux/gfp.h> 21 #include <linux/export.h> 22 #include <linux/slab.h> 23 #include <linux/genalloc.h> 24 #include <linux/dma-mapping.h> 25 #include <linux/dma-contiguous.h> 26 #include <linux/vmalloc.h> 27 #include <linux/swiotlb.h> 28 29 #include <asm/cacheflush.h> 30 31 struct dma_map_ops *dma_ops; 32 EXPORT_SYMBOL(dma_ops); 33 34 static pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot, 35 bool coherent) 36 { 37 if (!coherent || dma_get_attr(DMA_ATTR_WRITE_COMBINE, attrs)) 38 return pgprot_writecombine(prot); 39 return prot; 40 } 41 42 static struct gen_pool *atomic_pool; 43 44 #define DEFAULT_DMA_COHERENT_POOL_SIZE SZ_256K 45 static size_t atomic_pool_size = DEFAULT_DMA_COHERENT_POOL_SIZE; 46 47 static int __init early_coherent_pool(char *p) 48 { 49 atomic_pool_size = memparse(p, &p); 50 return 0; 51 } 52 early_param("coherent_pool", early_coherent_pool); 53 54 static void *__alloc_from_pool(size_t size, struct page **ret_page, gfp_t flags) 55 { 56 unsigned long val; 57 void *ptr = NULL; 58 59 if (!atomic_pool) { 60 WARN(1, "coherent pool not initialised!\n"); 61 return NULL; 62 } 63 64 val = gen_pool_alloc(atomic_pool, size); 65 if (val) { 66 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val); 67 68 *ret_page = phys_to_page(phys); 69 ptr = (void *)val; 70 if (flags & __GFP_ZERO) 71 memset(ptr, 0, size); 72 } 73 74 return ptr; 75 } 76 77 static bool __in_atomic_pool(void *start, size_t size) 78 { 79 return addr_in_gen_pool(atomic_pool, (unsigned long)start, size); 80 } 81 82 static int __free_from_pool(void *start, size_t size) 83 { 84 if (!__in_atomic_pool(start, size)) 85 return 0; 86 87 gen_pool_free(atomic_pool, (unsigned long)start, size); 88 89 return 1; 90 } 91 92 static void *__dma_alloc_coherent(struct device *dev, size_t size, 93 dma_addr_t *dma_handle, gfp_t flags, 94 struct dma_attrs *attrs) 95 { 96 if (dev == NULL) { 97 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n"); 98 return NULL; 99 } 100 101 if (IS_ENABLED(CONFIG_ZONE_DMA) && 102 dev->coherent_dma_mask <= DMA_BIT_MASK(32)) 103 flags |= GFP_DMA; 104 if (IS_ENABLED(CONFIG_DMA_CMA) && (flags & __GFP_WAIT)) { 105 struct page *page; 106 void *addr; 107 108 size = PAGE_ALIGN(size); 109 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT, 110 get_order(size)); 111 if (!page) 112 return NULL; 113 114 *dma_handle = phys_to_dma(dev, page_to_phys(page)); 115 addr = page_address(page); 116 if (flags & __GFP_ZERO) 117 memset(addr, 0, size); 118 return addr; 119 } else { 120 return swiotlb_alloc_coherent(dev, size, dma_handle, flags); 121 } 122 } 123 124 static void __dma_free_coherent(struct device *dev, size_t size, 125 void *vaddr, dma_addr_t dma_handle, 126 struct dma_attrs *attrs) 127 { 128 bool freed; 129 phys_addr_t paddr = dma_to_phys(dev, dma_handle); 130 131 if (dev == NULL) { 132 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n"); 133 return; 134 } 135 136 freed = dma_release_from_contiguous(dev, 137 phys_to_page(paddr), 138 size >> PAGE_SHIFT); 139 if (!freed) 140 swiotlb_free_coherent(dev, size, vaddr, dma_handle); 141 } 142 143 static void *__dma_alloc(struct device *dev, size_t size, 144 dma_addr_t *dma_handle, gfp_t flags, 145 struct dma_attrs *attrs) 146 { 147 struct page *page; 148 void *ptr, *coherent_ptr; 149 bool coherent = is_device_dma_coherent(dev); 150 151 size = PAGE_ALIGN(size); 152 153 if (!coherent && !(flags & __GFP_WAIT)) { 154 struct page *page = NULL; 155 void *addr = __alloc_from_pool(size, &page, flags); 156 157 if (addr) 158 *dma_handle = phys_to_dma(dev, page_to_phys(page)); 159 160 return addr; 161 } 162 163 ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs); 164 if (!ptr) 165 goto no_mem; 166 167 /* no need for non-cacheable mapping if coherent */ 168 if (coherent) 169 return ptr; 170 171 /* remove any dirty cache lines on the kernel alias */ 172 __dma_flush_range(ptr, ptr + size); 173 174 /* create a coherent mapping */ 175 page = virt_to_page(ptr); 176 coherent_ptr = dma_common_contiguous_remap(page, size, VM_USERMAP, 177 __get_dma_pgprot(attrs, 178 __pgprot(PROT_NORMAL_NC), false), 179 NULL); 180 if (!coherent_ptr) 181 goto no_map; 182 183 return coherent_ptr; 184 185 no_map: 186 __dma_free_coherent(dev, size, ptr, *dma_handle, attrs); 187 no_mem: 188 *dma_handle = DMA_ERROR_CODE; 189 return NULL; 190 } 191 192 static void __dma_free(struct device *dev, size_t size, 193 void *vaddr, dma_addr_t dma_handle, 194 struct dma_attrs *attrs) 195 { 196 void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle)); 197 198 if (!is_device_dma_coherent(dev)) { 199 if (__free_from_pool(vaddr, size)) 200 return; 201 vunmap(vaddr); 202 } 203 __dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs); 204 } 205 206 static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page, 207 unsigned long offset, size_t size, 208 enum dma_data_direction dir, 209 struct dma_attrs *attrs) 210 { 211 dma_addr_t dev_addr; 212 213 dev_addr = swiotlb_map_page(dev, page, offset, size, dir, attrs); 214 if (!is_device_dma_coherent(dev)) 215 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir); 216 217 return dev_addr; 218 } 219 220 221 static void __swiotlb_unmap_page(struct device *dev, dma_addr_t dev_addr, 222 size_t size, enum dma_data_direction dir, 223 struct dma_attrs *attrs) 224 { 225 if (!is_device_dma_coherent(dev)) 226 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir); 227 swiotlb_unmap_page(dev, dev_addr, size, dir, attrs); 228 } 229 230 static int __swiotlb_map_sg_attrs(struct device *dev, struct scatterlist *sgl, 231 int nelems, enum dma_data_direction dir, 232 struct dma_attrs *attrs) 233 { 234 struct scatterlist *sg; 235 int i, ret; 236 237 ret = swiotlb_map_sg_attrs(dev, sgl, nelems, dir, attrs); 238 if (!is_device_dma_coherent(dev)) 239 for_each_sg(sgl, sg, ret, i) 240 __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)), 241 sg->length, dir); 242 243 return ret; 244 } 245 246 static void __swiotlb_unmap_sg_attrs(struct device *dev, 247 struct scatterlist *sgl, int nelems, 248 enum dma_data_direction dir, 249 struct dma_attrs *attrs) 250 { 251 struct scatterlist *sg; 252 int i; 253 254 if (!is_device_dma_coherent(dev)) 255 for_each_sg(sgl, sg, nelems, i) 256 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)), 257 sg->length, dir); 258 swiotlb_unmap_sg_attrs(dev, sgl, nelems, dir, attrs); 259 } 260 261 static void __swiotlb_sync_single_for_cpu(struct device *dev, 262 dma_addr_t dev_addr, size_t size, 263 enum dma_data_direction dir) 264 { 265 if (!is_device_dma_coherent(dev)) 266 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir); 267 swiotlb_sync_single_for_cpu(dev, dev_addr, size, dir); 268 } 269 270 static void __swiotlb_sync_single_for_device(struct device *dev, 271 dma_addr_t dev_addr, size_t size, 272 enum dma_data_direction dir) 273 { 274 swiotlb_sync_single_for_device(dev, dev_addr, size, dir); 275 if (!is_device_dma_coherent(dev)) 276 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir); 277 } 278 279 static void __swiotlb_sync_sg_for_cpu(struct device *dev, 280 struct scatterlist *sgl, int nelems, 281 enum dma_data_direction dir) 282 { 283 struct scatterlist *sg; 284 int i; 285 286 if (!is_device_dma_coherent(dev)) 287 for_each_sg(sgl, sg, nelems, i) 288 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)), 289 sg->length, dir); 290 swiotlb_sync_sg_for_cpu(dev, sgl, nelems, dir); 291 } 292 293 static void __swiotlb_sync_sg_for_device(struct device *dev, 294 struct scatterlist *sgl, int nelems, 295 enum dma_data_direction dir) 296 { 297 struct scatterlist *sg; 298 int i; 299 300 swiotlb_sync_sg_for_device(dev, sgl, nelems, dir); 301 if (!is_device_dma_coherent(dev)) 302 for_each_sg(sgl, sg, nelems, i) 303 __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)), 304 sg->length, dir); 305 } 306 307 /* vma->vm_page_prot must be set appropriately before calling this function */ 308 static int __dma_common_mmap(struct device *dev, struct vm_area_struct *vma, 309 void *cpu_addr, dma_addr_t dma_addr, size_t size) 310 { 311 int ret = -ENXIO; 312 unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >> 313 PAGE_SHIFT; 314 unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT; 315 unsigned long pfn = dma_to_phys(dev, dma_addr) >> PAGE_SHIFT; 316 unsigned long off = vma->vm_pgoff; 317 318 if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret)) 319 return ret; 320 321 if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) { 322 ret = remap_pfn_range(vma, vma->vm_start, 323 pfn + off, 324 vma->vm_end - vma->vm_start, 325 vma->vm_page_prot); 326 } 327 328 return ret; 329 } 330 331 static int __swiotlb_mmap(struct device *dev, 332 struct vm_area_struct *vma, 333 void *cpu_addr, dma_addr_t dma_addr, size_t size, 334 struct dma_attrs *attrs) 335 { 336 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot, 337 is_device_dma_coherent(dev)); 338 return __dma_common_mmap(dev, vma, cpu_addr, dma_addr, size); 339 } 340 341 static struct dma_map_ops swiotlb_dma_ops = { 342 .alloc = __dma_alloc, 343 .free = __dma_free, 344 .mmap = __swiotlb_mmap, 345 .map_page = __swiotlb_map_page, 346 .unmap_page = __swiotlb_unmap_page, 347 .map_sg = __swiotlb_map_sg_attrs, 348 .unmap_sg = __swiotlb_unmap_sg_attrs, 349 .sync_single_for_cpu = __swiotlb_sync_single_for_cpu, 350 .sync_single_for_device = __swiotlb_sync_single_for_device, 351 .sync_sg_for_cpu = __swiotlb_sync_sg_for_cpu, 352 .sync_sg_for_device = __swiotlb_sync_sg_for_device, 353 .dma_supported = swiotlb_dma_supported, 354 .mapping_error = swiotlb_dma_mapping_error, 355 }; 356 357 static int __init atomic_pool_init(void) 358 { 359 pgprot_t prot = __pgprot(PROT_NORMAL_NC); 360 unsigned long nr_pages = atomic_pool_size >> PAGE_SHIFT; 361 struct page *page; 362 void *addr; 363 unsigned int pool_size_order = get_order(atomic_pool_size); 364 365 if (dev_get_cma_area(NULL)) 366 page = dma_alloc_from_contiguous(NULL, nr_pages, 367 pool_size_order); 368 else 369 page = alloc_pages(GFP_DMA, pool_size_order); 370 371 if (page) { 372 int ret; 373 void *page_addr = page_address(page); 374 375 memset(page_addr, 0, atomic_pool_size); 376 __dma_flush_range(page_addr, page_addr + atomic_pool_size); 377 378 atomic_pool = gen_pool_create(PAGE_SHIFT, -1); 379 if (!atomic_pool) 380 goto free_page; 381 382 addr = dma_common_contiguous_remap(page, atomic_pool_size, 383 VM_USERMAP, prot, atomic_pool_init); 384 385 if (!addr) 386 goto destroy_genpool; 387 388 ret = gen_pool_add_virt(atomic_pool, (unsigned long)addr, 389 page_to_phys(page), 390 atomic_pool_size, -1); 391 if (ret) 392 goto remove_mapping; 393 394 gen_pool_set_algo(atomic_pool, 395 gen_pool_first_fit_order_align, 396 (void *)PAGE_SHIFT); 397 398 pr_info("DMA: preallocated %zu KiB pool for atomic allocations\n", 399 atomic_pool_size / 1024); 400 return 0; 401 } 402 goto out; 403 404 remove_mapping: 405 dma_common_free_remap(addr, atomic_pool_size, VM_USERMAP); 406 destroy_genpool: 407 gen_pool_destroy(atomic_pool); 408 atomic_pool = NULL; 409 free_page: 410 if (!dma_release_from_contiguous(NULL, page, nr_pages)) 411 __free_pages(page, pool_size_order); 412 out: 413 pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n", 414 atomic_pool_size / 1024); 415 return -ENOMEM; 416 } 417 418 static int __init arm64_dma_init(void) 419 { 420 int ret; 421 422 dma_ops = &swiotlb_dma_ops; 423 424 ret = atomic_pool_init(); 425 426 return ret; 427 } 428 arch_initcall(arm64_dma_init); 429 430 #define PREALLOC_DMA_DEBUG_ENTRIES 4096 431 432 static int __init dma_debug_do_init(void) 433 { 434 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES); 435 return 0; 436 } 437 fs_initcall(dma_debug_do_init); 438