arch/arm64/mm/dma-mapping.c

   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                 memset(ptr, 0, size);
  71         }
  72
  73         return ptr;
  74 }
  75
  76 static bool __in_atomic_pool(void *start, size_t size)
  77 {
  78         return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
  79 }
  80
  81 static int __free_from_pool(void *start, size_t size)
  82 {
  83         if (!__in_atomic_pool(start, size))
  84                 return 0;
  85
  86         gen_pool_free(atomic_pool, (unsigned long)start, size);
  87
  88         return 1;
  89 }
  90
  91 static void *__dma_alloc_coherent(struct device *dev, size_t size,
  92                                   dma_addr_t *dma_handle, gfp_t flags,
  93                                   struct dma_attrs *attrs)
  94 {
  95         if (dev == NULL) {
  96                 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
  97                 return NULL;
  98         }
  99
 100         if (IS_ENABLED(CONFIG_ZONE_DMA) &&
 101             dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 102                 flags |= GFP_DMA;
 103         if (IS_ENABLED(CONFIG_DMA_CMA) && (flags & __GFP_WAIT)) {
 104                 struct page *page;
 105                 void *addr;
 106
 107                 size = PAGE_ALIGN(size);
 108                 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
 109                                                         get_order(size));
 110                 if (!page)
 111                         return NULL;
 112
 113                 *dma_handle = phys_to_dma(dev, page_to_phys(page));
 114                 addr = page_address(page);
 115                 memset(addr, 0, size);
 116                 return addr;
 117         } else {
 118                 return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
 119         }
 120 }
 121
 122 static void __dma_free_coherent(struct device *dev, size_t size,
 123                                 void *vaddr, dma_addr_t dma_handle,
 124                                 struct dma_attrs *attrs)
 125 {
 126         bool freed;
 127         phys_addr_t paddr = dma_to_phys(dev, dma_handle);
 128
 129         if (dev == NULL) {
 130                 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
 131                 return;
 132         }
 133
 134         freed = dma_release_from_contiguous(dev,
 135                                         phys_to_page(paddr),
 136                                         size >> PAGE_SHIFT);
 137         if (!freed)
 138                 swiotlb_free_coherent(dev, size, vaddr, dma_handle);
 139 }
 140
 141 static void *__dma_alloc(struct device *dev, size_t size,
 142                          dma_addr_t *dma_handle, gfp_t flags,
 143                          struct dma_attrs *attrs)
 144 {
 145         struct page *page;
 146         void *ptr, *coherent_ptr;
 147         bool coherent = is_device_dma_coherent(dev);
 148
 149         size = PAGE_ALIGN(size);
 150
 151         if (!coherent && !(flags & __GFP_WAIT)) {
 152                 struct page *page = NULL;
 153                 void *addr = __alloc_from_pool(size, &page, flags);
 154
 155                 if (addr)
 156                         *dma_handle = phys_to_dma(dev, page_to_phys(page));
 157
 158                 return addr;
 159         }
 160
 161         ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
 162         if (!ptr)
 163                 goto no_mem;
 164
 165         /* no need for non-cacheable mapping if coherent */
 166         if (coherent)
 167                 return ptr;
 168
 169         /* remove any dirty cache lines on the kernel alias */
 170         __dma_flush_range(ptr, ptr + size);
 171
 172         /* create a coherent mapping */
 173         page = virt_to_page(ptr);
 174         coherent_ptr = dma_common_contiguous_remap(page, size, VM_USERMAP,
 175                                 __get_dma_pgprot(attrs,
 176                                         __pgprot(PROT_NORMAL_NC), false),
 177                                         NULL);
 178         if (!coherent_ptr)
 179                 goto no_map;
 180
 181         return coherent_ptr;
 182
 183 no_map:
 184         __dma_free_coherent(dev, size, ptr, *dma_handle, attrs);
 185 no_mem:
 186         *dma_handle = DMA_ERROR_CODE;
 187         return NULL;
 188 }
 189
 190 static void __dma_free(struct device *dev, size_t size,
 191                        void *vaddr, dma_addr_t dma_handle,
 192                        struct dma_attrs *attrs)
 193 {
 194         void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle));
 195
 196         if (!is_device_dma_coherent(dev)) {
 197                 if (__free_from_pool(vaddr, size))
 198                         return;
 199                 vunmap(vaddr);
 200         }
 201         __dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
 202 }
 203
 204 static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,
 205                                      unsigned long offset, size_t size,
 206                                      enum dma_data_direction dir,
 207                                      struct dma_attrs *attrs)
 208 {
 209         dma_addr_t dev_addr;
 210
 211         dev_addr = swiotlb_map_page(dev, page, offset, size, dir, attrs);
 212         if (!is_device_dma_coherent(dev))
 213                 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 214
 215         return dev_addr;
 216 }
 217
 218
 219 static void __swiotlb_unmap_page(struct device *dev, dma_addr_t dev_addr,
 220                                  size_t size, enum dma_data_direction dir,
 221                                  struct dma_attrs *attrs)
 222 {
 223         if (!is_device_dma_coherent(dev))
 224                 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 225         swiotlb_unmap_page(dev, dev_addr, size, dir, attrs);
 226 }
 227
 228 static int __swiotlb_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
 229                                   int nelems, enum dma_data_direction dir,
 230                                   struct dma_attrs *attrs)
 231 {
 232         struct scatterlist *sg;
 233         int i, ret;
 234
 235         ret = swiotlb_map_sg_attrs(dev, sgl, nelems, dir, attrs);
 236         if (!is_device_dma_coherent(dev))
 237                 for_each_sg(sgl, sg, ret, i)
 238                         __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 239                                        sg->length, dir);
 240
 241         return ret;
 242 }
 243
 244 static void __swiotlb_unmap_sg_attrs(struct device *dev,
 245                                      struct scatterlist *sgl, int nelems,
 246                                      enum dma_data_direction dir,
 247                                      struct dma_attrs *attrs)
 248 {
 249         struct scatterlist *sg;
 250         int i;
 251
 252         if (!is_device_dma_coherent(dev))
 253                 for_each_sg(sgl, sg, nelems, i)
 254                         __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 255                                          sg->length, dir);
 256         swiotlb_unmap_sg_attrs(dev, sgl, nelems, dir, attrs);
 257 }
 258
 259 static void __swiotlb_sync_single_for_cpu(struct device *dev,
 260                                           dma_addr_t dev_addr, size_t size,
 261                                           enum dma_data_direction dir)
 262 {
 263         if (!is_device_dma_coherent(dev))
 264                 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 265         swiotlb_sync_single_for_cpu(dev, dev_addr, size, dir);
 266 }
 267
 268 static void __swiotlb_sync_single_for_device(struct device *dev,
 269                                              dma_addr_t dev_addr, size_t size,
 270                                              enum dma_data_direction dir)
 271 {
 272         swiotlb_sync_single_for_device(dev, dev_addr, size, dir);
 273         if (!is_device_dma_coherent(dev))
 274                 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 275 }
 276
 277 static void __swiotlb_sync_sg_for_cpu(struct device *dev,
 278                                       struct scatterlist *sgl, int nelems,
 279                                       enum dma_data_direction dir)
 280 {
 281         struct scatterlist *sg;
 282         int i;
 283
 284         if (!is_device_dma_coherent(dev))
 285                 for_each_sg(sgl, sg, nelems, i)
 286                         __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 287                                          sg->length, dir);
 288         swiotlb_sync_sg_for_cpu(dev, sgl, nelems, dir);
 289 }
 290
 291 static void __swiotlb_sync_sg_for_device(struct device *dev,
 292                                          struct scatterlist *sgl, int nelems,
 293                                          enum dma_data_direction dir)
 294 {
 295         struct scatterlist *sg;
 296         int i;
 297
 298         swiotlb_sync_sg_for_device(dev, sgl, nelems, dir);
 299         if (!is_device_dma_coherent(dev))
 300                 for_each_sg(sgl, sg, nelems, i)
 301                         __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 302                                        sg->length, dir);
 303 }
 304
 305 /* vma->vm_page_prot must be set appropriately before calling this function */
 306 static int __dma_common_mmap(struct device *dev, struct vm_area_struct *vma,
 307                              void *cpu_addr, dma_addr_t dma_addr, size_t size)
 308 {
 309         int ret = -ENXIO;
 310         unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >>
 311                                         PAGE_SHIFT;
 312         unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
 313         unsigned long pfn = dma_to_phys(dev, dma_addr) >> PAGE_SHIFT;
 314         unsigned long off = vma->vm_pgoff;
 315
 316         if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
 317                 return ret;
 318
 319         if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
 320                 ret = remap_pfn_range(vma, vma->vm_start,
 321                                       pfn + off,
 322                                       vma->vm_end - vma->vm_start,
 323                                       vma->vm_page_prot);
 324         }
 325
 326         return ret;
 327 }
 328
 329 static int __swiotlb_mmap(struct device *dev,
 330                           struct vm_area_struct *vma,
 331                           void *cpu_addr, dma_addr_t dma_addr, size_t size,
 332                           struct dma_attrs *attrs)
 333 {
 334         vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
 335                                              is_device_dma_coherent(dev));
 336         return __dma_common_mmap(dev, vma, cpu_addr, dma_addr, size);
 337 }
 338
 339 static struct dma_map_ops swiotlb_dma_ops = {
 340         .alloc = __dma_alloc,
 341         .free = __dma_free,
 342         .mmap = __swiotlb_mmap,
 343         .map_page = __swiotlb_map_page,
 344         .unmap_page = __swiotlb_unmap_page,
 345         .map_sg = __swiotlb_map_sg_attrs,
 346         .unmap_sg = __swiotlb_unmap_sg_attrs,
 347         .sync_single_for_cpu = __swiotlb_sync_single_for_cpu,
 348         .sync_single_for_device = __swiotlb_sync_single_for_device,
 349         .sync_sg_for_cpu = __swiotlb_sync_sg_for_cpu,
 350         .sync_sg_for_device = __swiotlb_sync_sg_for_device,
 351         .dma_supported = swiotlb_dma_supported,
 352         .mapping_error = swiotlb_dma_mapping_error,
 353 };
 354
 355 static int __init atomic_pool_init(void)
 356 {
 357         pgprot_t prot = __pgprot(PROT_NORMAL_NC);
 358         unsigned long nr_pages = atomic_pool_size >> PAGE_SHIFT;
 359         struct page *page;
 360         void *addr;
 361         unsigned int pool_size_order = get_order(atomic_pool_size);
 362
 363         if (dev_get_cma_area(NULL))
 364                 page = dma_alloc_from_contiguous(NULL, nr_pages,
 365                                                         pool_size_order);
 366         else
 367                 page = alloc_pages(GFP_DMA, pool_size_order);
 368
 369         if (page) {
 370                 int ret;
 371                 void *page_addr = page_address(page);
 372
 373                 memset(page_addr, 0, atomic_pool_size);
 374                 __dma_flush_range(page_addr, page_addr + atomic_pool_size);
 375
 376                 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
 377                 if (!atomic_pool)
 378                         goto free_page;
 379
 380                 addr = dma_common_contiguous_remap(page, atomic_pool_size,
 381                                         VM_USERMAP, prot, atomic_pool_init);
 382
 383                 if (!addr)
 384                         goto destroy_genpool;
 385
 386                 ret = gen_pool_add_virt(atomic_pool, (unsigned long)addr,
 387                                         page_to_phys(page),
 388                                         atomic_pool_size, -1);
 389                 if (ret)
 390                         goto remove_mapping;
 391
 392                 gen_pool_set_algo(atomic_pool,
 393                                   gen_pool_first_fit_order_align,
 394                                   (void *)PAGE_SHIFT);
 395
 396                 pr_info("DMA: preallocated %zu KiB pool for atomic allocations\n",
 397                         atomic_pool_size / 1024);
 398                 return 0;
 399         }
 400         goto out;
 401
 402 remove_mapping:
 403         dma_common_free_remap(addr, atomic_pool_size, VM_USERMAP);
 404 destroy_genpool:
 405         gen_pool_destroy(atomic_pool);
 406         atomic_pool = NULL;
 407 free_page:
 408         if (!dma_release_from_contiguous(NULL, page, nr_pages))
 409                 __free_pages(page, pool_size_order);
 410 out:
 411         pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
 412                 atomic_pool_size / 1024);
 413         return -ENOMEM;
 414 }
 415
 416 static int __init arm64_dma_init(void)
 417 {
 418         int ret;
 419
 420         dma_ops = &swiotlb_dma_ops;
 421
 422         ret = atomic_pool_init();
 423
 424         return ret;
 425 }
 426 arch_initcall(arm64_dma_init);
 427
 428 #define PREALLOC_DMA_DEBUG_ENTRIES      4096
 429
 430 static int __init dma_debug_do_init(void)
 431 {
 432         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 433         return 0;
 434 }
 435 fs_initcall(dma_debug_do_init);