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/acpi.h>
  22 #include <linux/bootmem.h>
  23 #include <linux/cache.h>
  24 #include <linux/export.h>
  25 #include <linux/slab.h>
  26 #include <linux/genalloc.h>
  27 #include <linux/dma-mapping.h>
  28 #include <linux/dma-contiguous.h>
  29 #include <linux/vmalloc.h>
  30 #include <linux/swiotlb.h>
  31 #include <linux/pci.h>
  32
  33 #include <asm/cacheflush.h>
  34
  35 static int swiotlb __ro_after_init;
  36
  37 static pgprot_t __get_dma_pgprot(unsigned long attrs, pgprot_t prot,
  38                                  bool coherent)
  39 {
  40         if (!coherent || (attrs & DMA_ATTR_WRITE_COMBINE))
  41                 return pgprot_writecombine(prot);
  42         return prot;
  43 }
  44
  45 static struct gen_pool *atomic_pool __ro_after_init;
  46
  47 #define DEFAULT_DMA_COHERENT_POOL_SIZE  SZ_256K
  48 static size_t atomic_pool_size __initdata = DEFAULT_DMA_COHERENT_POOL_SIZE;
  49
  50 static int __init early_coherent_pool(char *p)
  51 {
  52         atomic_pool_size = memparse(p, &p);
  53         return 0;
  54 }
  55 early_param("coherent_pool", early_coherent_pool);
  56
  57 static void *__alloc_from_pool(size_t size, struct page **ret_page, gfp_t flags)
  58 {
  59         unsigned long val;
  60         void *ptr = NULL;
  61
  62         if (!atomic_pool) {
  63                 WARN(1, "coherent pool not initialised!\n");
  64                 return NULL;
  65         }
  66
  67         val = gen_pool_alloc(atomic_pool, size);
  68         if (val) {
  69                 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
  70
  71                 *ret_page = phys_to_page(phys);
  72                 ptr = (void *)val;
  73                 memset(ptr, 0, size);
  74         }
  75
  76         return ptr;
  77 }
  78
  79 static bool __in_atomic_pool(void *start, size_t size)
  80 {
  81         return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
  82 }
  83
  84 static int __free_from_pool(void *start, size_t size)
  85 {
  86         if (!__in_atomic_pool(start, size))
  87                 return 0;
  88
  89         gen_pool_free(atomic_pool, (unsigned long)start, size);
  90
  91         return 1;
  92 }
  93
  94 static void *__dma_alloc_coherent(struct device *dev, size_t size,
  95                                   dma_addr_t *dma_handle, gfp_t flags,
  96                                   unsigned long attrs)
  97 {
  98         if (IS_ENABLED(CONFIG_ZONE_DMA) &&
  99             dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 100                 flags |= GFP_DMA;
 101         if (dev_get_cma_area(dev) && gfpflags_allow_blocking(flags)) {
 102                 struct page *page;
 103                 void *addr;
 104
 105                 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
 106                                                  get_order(size), flags);
 107                 if (!page)
 108                         return NULL;
 109
 110                 *dma_handle = phys_to_dma(dev, page_to_phys(page));
 111                 addr = page_address(page);
 112                 memset(addr, 0, size);
 113                 return addr;
 114         } else {
 115                 return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
 116         }
 117 }
 118
 119 static void __dma_free_coherent(struct device *dev, size_t size,
 120                                 void *vaddr, dma_addr_t dma_handle,
 121                                 unsigned long attrs)
 122 {
 123         bool freed;
 124         phys_addr_t paddr = dma_to_phys(dev, dma_handle);
 125
 126
 127         freed = dma_release_from_contiguous(dev,
 128                                         phys_to_page(paddr),
 129                                         size >> PAGE_SHIFT);
 130         if (!freed)
 131                 swiotlb_free_coherent(dev, size, vaddr, dma_handle);
 132 }
 133
 134 static void *__dma_alloc(struct device *dev, size_t size,
 135                          dma_addr_t *dma_handle, gfp_t flags,
 136                          unsigned long attrs)
 137 {
 138         struct page *page;
 139         void *ptr, *coherent_ptr;
 140         bool coherent = is_device_dma_coherent(dev);
 141         pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, false);
 142
 143         size = PAGE_ALIGN(size);
 144
 145         if (!coherent && !gfpflags_allow_blocking(flags)) {
 146                 struct page *page = NULL;
 147                 void *addr = __alloc_from_pool(size, &page, flags);
 148
 149                 if (addr)
 150                         *dma_handle = phys_to_dma(dev, page_to_phys(page));
 151
 152                 return addr;
 153         }
 154
 155         ptr = __dma_alloc_coherent(dev, size, dma_handle, flags, attrs);
 156         if (!ptr)
 157                 goto no_mem;
 158
 159         /* no need for non-cacheable mapping if coherent */
 160         if (coherent)
 161                 return ptr;
 162
 163         /* remove any dirty cache lines on the kernel alias */
 164         __dma_flush_area(ptr, size);
 165
 166         /* create a coherent mapping */
 167         page = virt_to_page(ptr);
 168         coherent_ptr = dma_common_contiguous_remap(page, size, VM_USERMAP,
 169                                                    prot, NULL);
 170         if (!coherent_ptr)
 171                 goto no_map;
 172
 173         return coherent_ptr;
 174
 175 no_map:
 176         __dma_free_coherent(dev, size, ptr, *dma_handle, attrs);
 177 no_mem:
 178         return NULL;
 179 }
 180
 181 static void __dma_free(struct device *dev, size_t size,
 182                        void *vaddr, dma_addr_t dma_handle,
 183                        unsigned long attrs)
 184 {
 185         void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle));
 186
 187         size = PAGE_ALIGN(size);
 188
 189         if (!is_device_dma_coherent(dev)) {
 190                 if (__free_from_pool(vaddr, size))
 191                         return;
 192                 vunmap(vaddr);
 193         }
 194         __dma_free_coherent(dev, size, swiotlb_addr, dma_handle, attrs);
 195 }
 196
 197 static dma_addr_t __swiotlb_map_page(struct device *dev, struct page *page,
 198                                      unsigned long offset, size_t size,
 199                                      enum dma_data_direction dir,
 200                                      unsigned long attrs)
 201 {
 202         dma_addr_t dev_addr;
 203
 204         dev_addr = swiotlb_map_page(dev, page, offset, size, dir, attrs);
 205         if (!is_device_dma_coherent(dev) &&
 206             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 207                 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 208
 209         return dev_addr;
 210 }
 211
 212
 213 static void __swiotlb_unmap_page(struct device *dev, dma_addr_t dev_addr,
 214                                  size_t size, enum dma_data_direction dir,
 215                                  unsigned long attrs)
 216 {
 217         if (!is_device_dma_coherent(dev) &&
 218             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 219                 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 220         swiotlb_unmap_page(dev, dev_addr, size, dir, attrs);
 221 }
 222
 223 static int __swiotlb_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
 224                                   int nelems, enum dma_data_direction dir,
 225                                   unsigned long attrs)
 226 {
 227         struct scatterlist *sg;
 228         int i, ret;
 229
 230         ret = swiotlb_map_sg_attrs(dev, sgl, nelems, dir, attrs);
 231         if (!is_device_dma_coherent(dev) &&
 232             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 233                 for_each_sg(sgl, sg, ret, i)
 234                         __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 235                                        sg->length, dir);
 236
 237         return ret;
 238 }
 239
 240 static void __swiotlb_unmap_sg_attrs(struct device *dev,
 241                                      struct scatterlist *sgl, int nelems,
 242                                      enum dma_data_direction dir,
 243                                      unsigned long attrs)
 244 {
 245         struct scatterlist *sg;
 246         int i;
 247
 248         if (!is_device_dma_coherent(dev) &&
 249             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 250                 for_each_sg(sgl, sg, nelems, i)
 251                         __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 252                                          sg->length, dir);
 253         swiotlb_unmap_sg_attrs(dev, sgl, nelems, dir, attrs);
 254 }
 255
 256 static void __swiotlb_sync_single_for_cpu(struct device *dev,
 257                                           dma_addr_t dev_addr, size_t size,
 258                                           enum dma_data_direction dir)
 259 {
 260         if (!is_device_dma_coherent(dev))
 261                 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 262         swiotlb_sync_single_for_cpu(dev, dev_addr, size, dir);
 263 }
 264
 265 static void __swiotlb_sync_single_for_device(struct device *dev,
 266                                              dma_addr_t dev_addr, size_t size,
 267                                              enum dma_data_direction dir)
 268 {
 269         swiotlb_sync_single_for_device(dev, dev_addr, size, dir);
 270         if (!is_device_dma_coherent(dev))
 271                 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 272 }
 273
 274 static void __swiotlb_sync_sg_for_cpu(struct device *dev,
 275                                       struct scatterlist *sgl, int nelems,
 276                                       enum dma_data_direction dir)
 277 {
 278         struct scatterlist *sg;
 279         int i;
 280
 281         if (!is_device_dma_coherent(dev))
 282                 for_each_sg(sgl, sg, nelems, i)
 283                         __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 284                                          sg->length, dir);
 285         swiotlb_sync_sg_for_cpu(dev, sgl, nelems, dir);
 286 }
 287
 288 static void __swiotlb_sync_sg_for_device(struct device *dev,
 289                                          struct scatterlist *sgl, int nelems,
 290                                          enum dma_data_direction dir)
 291 {
 292         struct scatterlist *sg;
 293         int i;
 294
 295         swiotlb_sync_sg_for_device(dev, sgl, nelems, dir);
 296         if (!is_device_dma_coherent(dev))
 297                 for_each_sg(sgl, sg, nelems, i)
 298                         __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 299                                        sg->length, dir);
 300 }
 301
 302 static int __swiotlb_mmap_pfn(struct vm_area_struct *vma,
 303                               unsigned long pfn, size_t size)
 304 {
 305         int ret = -ENXIO;
 306         unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >>
 307                                         PAGE_SHIFT;
 308         unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
 309         unsigned long off = vma->vm_pgoff;
 310
 311         if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
 312                 ret = remap_pfn_range(vma, vma->vm_start,
 313                                       pfn + off,
 314                                       vma->vm_end - vma->vm_start,
 315                                       vma->vm_page_prot);
 316         }
 317
 318         return ret;
 319 }
 320
 321 static int __swiotlb_mmap(struct device *dev,
 322                           struct vm_area_struct *vma,
 323                           void *cpu_addr, dma_addr_t dma_addr, size_t size,
 324                           unsigned long attrs)
 325 {
 326         int ret;
 327         unsigned long pfn = dma_to_phys(dev, dma_addr) >> PAGE_SHIFT;
 328
 329         vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
 330                                              is_device_dma_coherent(dev));
 331
 332         if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
 333                 return ret;
 334
 335         return __swiotlb_mmap_pfn(vma, pfn, size);
 336 }
 337
 338 static int __swiotlb_get_sgtable_page(struct sg_table *sgt,
 339                                       struct page *page, size_t size)
 340 {
 341         int ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
 342
 343         if (!ret)
 344                 sg_set_page(sgt->sgl, page, PAGE_ALIGN(size), 0);
 345
 346         return ret;
 347 }
 348
 349 static int __swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
 350                                  void *cpu_addr, dma_addr_t handle, size_t size,
 351                                  unsigned long attrs)
 352 {
 353         struct page *page = phys_to_page(dma_to_phys(dev, handle));
 354
 355         return __swiotlb_get_sgtable_page(sgt, page, size);
 356 }
 357
 358 static int __swiotlb_dma_supported(struct device *hwdev, u64 mask)
 359 {
 360         if (swiotlb)
 361                 return swiotlb_dma_supported(hwdev, mask);
 362         return 1;
 363 }
 364
 365 static int __swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t addr)
 366 {
 367         if (swiotlb)
 368                 return swiotlb_dma_mapping_error(hwdev, addr);
 369         return 0;
 370 }
 371
 372 static const struct dma_map_ops swiotlb_dma_ops = {
 373         .alloc = __dma_alloc,
 374         .free = __dma_free,
 375         .mmap = __swiotlb_mmap,
 376         .get_sgtable = __swiotlb_get_sgtable,
 377         .map_page = __swiotlb_map_page,
 378         .unmap_page = __swiotlb_unmap_page,
 379         .map_sg = __swiotlb_map_sg_attrs,
 380         .unmap_sg = __swiotlb_unmap_sg_attrs,
 381         .sync_single_for_cpu = __swiotlb_sync_single_for_cpu,
 382         .sync_single_for_device = __swiotlb_sync_single_for_device,
 383         .sync_sg_for_cpu = __swiotlb_sync_sg_for_cpu,
 384         .sync_sg_for_device = __swiotlb_sync_sg_for_device,
 385         .dma_supported = __swiotlb_dma_supported,
 386         .mapping_error = __swiotlb_dma_mapping_error,
 387 };
 388
 389 static int __init atomic_pool_init(void)
 390 {
 391         pgprot_t prot = __pgprot(PROT_NORMAL_NC);
 392         unsigned long nr_pages = atomic_pool_size >> PAGE_SHIFT;
 393         struct page *page;
 394         void *addr;
 395         unsigned int pool_size_order = get_order(atomic_pool_size);
 396
 397         if (dev_get_cma_area(NULL))
 398                 page = dma_alloc_from_contiguous(NULL, nr_pages,
 399                                                  pool_size_order, GFP_KERNEL);
 400         else
 401                 page = alloc_pages(GFP_DMA, pool_size_order);
 402
 403         if (page) {
 404                 int ret;
 405                 void *page_addr = page_address(page);
 406
 407                 memset(page_addr, 0, atomic_pool_size);
 408                 __dma_flush_area(page_addr, atomic_pool_size);
 409
 410                 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
 411                 if (!atomic_pool)
 412                         goto free_page;
 413
 414                 addr = dma_common_contiguous_remap(page, atomic_pool_size,
 415                                         VM_USERMAP, prot, atomic_pool_init);
 416
 417                 if (!addr)
 418                         goto destroy_genpool;
 419
 420                 ret = gen_pool_add_virt(atomic_pool, (unsigned long)addr,
 421                                         page_to_phys(page),
 422                                         atomic_pool_size, -1);
 423                 if (ret)
 424                         goto remove_mapping;
 425
 426                 gen_pool_set_algo(atomic_pool,
 427                                   gen_pool_first_fit_order_align,
 428                                   NULL);
 429
 430                 pr_info("DMA: preallocated %zu KiB pool for atomic allocations\n",
 431                         atomic_pool_size / 1024);
 432                 return 0;
 433         }
 434         goto out;
 435
 436 remove_mapping:
 437         dma_common_free_remap(addr, atomic_pool_size, VM_USERMAP);
 438 destroy_genpool:
 439         gen_pool_destroy(atomic_pool);
 440         atomic_pool = NULL;
 441 free_page:
 442         if (!dma_release_from_contiguous(NULL, page, nr_pages))
 443                 __free_pages(page, pool_size_order);
 444 out:
 445         pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
 446                 atomic_pool_size / 1024);
 447         return -ENOMEM;
 448 }
 449
 450 /********************************************
 451  * The following APIs are for dummy DMA ops *
 452  ********************************************/
 453
 454 static void *__dummy_alloc(struct device *dev, size_t size,
 455                            dma_addr_t *dma_handle, gfp_t flags,
 456                            unsigned long attrs)
 457 {
 458         return NULL;
 459 }
 460
 461 static void __dummy_free(struct device *dev, size_t size,
 462                          void *vaddr, dma_addr_t dma_handle,
 463                          unsigned long attrs)
 464 {
 465 }
 466
 467 static int __dummy_mmap(struct device *dev,
 468                         struct vm_area_struct *vma,
 469                         void *cpu_addr, dma_addr_t dma_addr, size_t size,
 470                         unsigned long attrs)
 471 {
 472         return -ENXIO;
 473 }
 474
 475 static dma_addr_t __dummy_map_page(struct device *dev, struct page *page,
 476                                    unsigned long offset, size_t size,
 477                                    enum dma_data_direction dir,
 478                                    unsigned long attrs)
 479 {
 480         return 0;
 481 }
 482
 483 static void __dummy_unmap_page(struct device *dev, dma_addr_t dev_addr,
 484                                size_t size, enum dma_data_direction dir,
 485                                unsigned long attrs)
 486 {
 487 }
 488
 489 static int __dummy_map_sg(struct device *dev, struct scatterlist *sgl,
 490                           int nelems, enum dma_data_direction dir,
 491                           unsigned long attrs)
 492 {
 493         return 0;
 494 }
 495
 496 static void __dummy_unmap_sg(struct device *dev,
 497                              struct scatterlist *sgl, int nelems,
 498                              enum dma_data_direction dir,
 499                              unsigned long attrs)
 500 {
 501 }
 502
 503 static void __dummy_sync_single(struct device *dev,
 504                                 dma_addr_t dev_addr, size_t size,
 505                                 enum dma_data_direction dir)
 506 {
 507 }
 508
 509 static void __dummy_sync_sg(struct device *dev,
 510                             struct scatterlist *sgl, int nelems,
 511                             enum dma_data_direction dir)
 512 {
 513 }
 514
 515 static int __dummy_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
 516 {
 517         return 1;
 518 }
 519
 520 static int __dummy_dma_supported(struct device *hwdev, u64 mask)
 521 {
 522         return 0;
 523 }
 524
 525 const struct dma_map_ops dummy_dma_ops = {
 526         .alloc                  = __dummy_alloc,
 527         .free                   = __dummy_free,
 528         .mmap                   = __dummy_mmap,
 529         .map_page               = __dummy_map_page,
 530         .unmap_page             = __dummy_unmap_page,
 531         .map_sg                 = __dummy_map_sg,
 532         .unmap_sg               = __dummy_unmap_sg,
 533         .sync_single_for_cpu    = __dummy_sync_single,
 534         .sync_single_for_device = __dummy_sync_single,
 535         .sync_sg_for_cpu        = __dummy_sync_sg,
 536         .sync_sg_for_device     = __dummy_sync_sg,
 537         .mapping_error          = __dummy_mapping_error,
 538         .dma_supported          = __dummy_dma_supported,
 539 };
 540 EXPORT_SYMBOL(dummy_dma_ops);
 541
 542 static int __init arm64_dma_init(void)
 543 {
 544         if (swiotlb_force == SWIOTLB_FORCE ||
 545             max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
 546                 swiotlb = 1;
 547
 548         return atomic_pool_init();
 549 }
 550 arch_initcall(arm64_dma_init);
 551
 552 #define PREALLOC_DMA_DEBUG_ENTRIES      4096
 553
 554 static int __init dma_debug_do_init(void)
 555 {
 556         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 557         return 0;
 558 }
 559 fs_initcall(dma_debug_do_init);
 560
 561
 562 #ifdef CONFIG_IOMMU_DMA
 563 #include <linux/dma-iommu.h>
 564 #include <linux/platform_device.h>
 565 #include <linux/amba/bus.h>
 566
 567 /* Thankfully, all cache ops are by VA so we can ignore phys here */
 568 static void flush_page(struct device *dev, const void *virt, phys_addr_t phys)
 569 {
 570         __dma_flush_area(virt, PAGE_SIZE);
 571 }
 572
 573 static void *__iommu_alloc_attrs(struct device *dev, size_t size,
 574                                  dma_addr_t *handle, gfp_t gfp,
 575                                  unsigned long attrs)
 576 {
 577         bool coherent = is_device_dma_coherent(dev);
 578         int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
 579         size_t iosize = size;
 580         void *addr;
 581
 582         if (WARN(!dev, "cannot create IOMMU mapping for unknown device\n"))
 583                 return NULL;
 584
 585         size = PAGE_ALIGN(size);
 586
 587         /*
 588          * Some drivers rely on this, and we probably don't want the
 589          * possibility of stale kernel data being read by devices anyway.
 590          */
 591         gfp |= __GFP_ZERO;
 592
 593         if (!gfpflags_allow_blocking(gfp)) {
 594                 struct page *page;
 595                 /*
 596                  * In atomic context we can't remap anything, so we'll only
 597                  * get the virtually contiguous buffer we need by way of a
 598                  * physically contiguous allocation.
 599                  */
 600                 if (coherent) {
 601                         page = alloc_pages(gfp, get_order(size));
 602                         addr = page ? page_address(page) : NULL;
 603                 } else {
 604                         addr = __alloc_from_pool(size, &page, gfp);
 605                 }
 606                 if (!addr)
 607                         return NULL;
 608
 609                 *handle = iommu_dma_map_page(dev, page, 0, iosize, ioprot);
 610                 if (iommu_dma_mapping_error(dev, *handle)) {
 611                         if (coherent)
 612                                 __free_pages(page, get_order(size));
 613                         else
 614                                 __free_from_pool(addr, size);
 615                         addr = NULL;
 616                 }
 617         } else if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
 618                 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
 619                 struct page *page;
 620
 621                 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
 622                                                  get_order(size), gfp);
 623                 if (!page)
 624                         return NULL;
 625
 626                 *handle = iommu_dma_map_page(dev, page, 0, iosize, ioprot);
 627                 if (iommu_dma_mapping_error(dev, *handle)) {
 628                         dma_release_from_contiguous(dev, page,
 629                                                     size >> PAGE_SHIFT);
 630                         return NULL;
 631                 }
 632                 addr = dma_common_contiguous_remap(page, size, VM_USERMAP,
 633                                                    prot,
 634                                                    __builtin_return_address(0));
 635                 if (addr) {
 636                         if (!coherent)
 637                                 __dma_flush_area(page_to_virt(page), iosize);
 638                         memset(addr, 0, size);
 639                 } else {
 640                         iommu_dma_unmap_page(dev, *handle, iosize, 0, attrs);
 641                         dma_release_from_contiguous(dev, page,
 642                                                     size >> PAGE_SHIFT);
 643                 }
 644         } else {
 645                 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
 646                 struct page **pages;
 647
 648                 pages = iommu_dma_alloc(dev, iosize, gfp, attrs, ioprot,
 649                                         handle, flush_page);
 650                 if (!pages)
 651                         return NULL;
 652
 653                 addr = dma_common_pages_remap(pages, size, VM_USERMAP, prot,
 654                                               __builtin_return_address(0));
 655                 if (!addr)
 656                         iommu_dma_free(dev, pages, iosize, handle);
 657         }
 658         return addr;
 659 }
 660
 661 static void __iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
 662                                dma_addr_t handle, unsigned long attrs)
 663 {
 664         size_t iosize = size;
 665
 666         size = PAGE_ALIGN(size);
 667         /*
 668          * @cpu_addr will be one of 4 things depending on how it was allocated:
 669          * - A remapped array of pages for contiguous allocations.
 670          * - A remapped array of pages from iommu_dma_alloc(), for all
 671          *   non-atomic allocations.
 672          * - A non-cacheable alias from the atomic pool, for atomic
 673          *   allocations by non-coherent devices.
 674          * - A normal lowmem address, for atomic allocations by
 675          *   coherent devices.
 676          * Hence how dodgy the below logic looks...
 677          */
 678         if (__in_atomic_pool(cpu_addr, size)) {
 679                 iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
 680                 __free_from_pool(cpu_addr, size);
 681         } else if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
 682                 struct page *page = vmalloc_to_page(cpu_addr);
 683
 684                 iommu_dma_unmap_page(dev, handle, iosize, 0, attrs);
 685                 dma_release_from_contiguous(dev, page, size >> PAGE_SHIFT);
 686                 dma_common_free_remap(cpu_addr, size, VM_USERMAP);
 687         } else if (is_vmalloc_addr(cpu_addr)){
 688                 struct vm_struct *area = find_vm_area(cpu_addr);
 689
 690                 if (WARN_ON(!area || !area->pages))
 691                         return;
 692                 iommu_dma_free(dev, area->pages, iosize, &handle);
 693                 dma_common_free_remap(cpu_addr, size, VM_USERMAP);
 694         } else {
 695                 iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
 696                 __free_pages(virt_to_page(cpu_addr), get_order(size));
 697         }
 698 }
 699
 700 static int __iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
 701                               void *cpu_addr, dma_addr_t dma_addr, size_t size,
 702                               unsigned long attrs)
 703 {
 704         struct vm_struct *area;
 705         int ret;
 706
 707         vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
 708                                              is_device_dma_coherent(dev));
 709
 710         if (dma_mmap_from_dev_coherent(dev, vma, cpu_addr, size, &ret))
 711                 return ret;
 712
 713         if (!is_vmalloc_addr(cpu_addr)) {
 714                 unsigned long pfn = page_to_pfn(virt_to_page(cpu_addr));
 715                 return __swiotlb_mmap_pfn(vma, pfn, size);
 716         }
 717
 718         if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
 719                 /*
 720                  * DMA_ATTR_FORCE_CONTIGUOUS allocations are always remapped,
 721                  * hence in the vmalloc space.
 722                  */
 723                 unsigned long pfn = vmalloc_to_pfn(cpu_addr);
 724                 return __swiotlb_mmap_pfn(vma, pfn, size);
 725         }
 726
 727         area = find_vm_area(cpu_addr);
 728         if (WARN_ON(!area || !area->pages))
 729                 return -ENXIO;
 730
 731         return iommu_dma_mmap(area->pages, size, vma);
 732 }
 733
 734 static int __iommu_get_sgtable(struct device *dev, struct sg_table *sgt,
 735                                void *cpu_addr, dma_addr_t dma_addr,
 736                                size_t size, unsigned long attrs)
 737 {
 738         unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
 739         struct vm_struct *area = find_vm_area(cpu_addr);
 740
 741         if (!is_vmalloc_addr(cpu_addr)) {
 742                 struct page *page = virt_to_page(cpu_addr);
 743                 return __swiotlb_get_sgtable_page(sgt, page, size);
 744         }
 745
 746         if (attrs & DMA_ATTR_FORCE_CONTIGUOUS) {
 747                 /*
 748                  * DMA_ATTR_FORCE_CONTIGUOUS allocations are always remapped,
 749                  * hence in the vmalloc space.
 750                  */
 751                 struct page *page = vmalloc_to_page(cpu_addr);
 752                 return __swiotlb_get_sgtable_page(sgt, page, size);
 753         }
 754
 755         if (WARN_ON(!area || !area->pages))
 756                 return -ENXIO;
 757
 758         return sg_alloc_table_from_pages(sgt, area->pages, count, 0, size,
 759                                          GFP_KERNEL);
 760 }
 761
 762 static void __iommu_sync_single_for_cpu(struct device *dev,
 763                                         dma_addr_t dev_addr, size_t size,
 764                                         enum dma_data_direction dir)
 765 {
 766         phys_addr_t phys;
 767
 768         if (is_device_dma_coherent(dev))
 769                 return;
 770
 771         phys = iommu_iova_to_phys(iommu_get_domain_for_dev(dev), dev_addr);
 772         __dma_unmap_area(phys_to_virt(phys), size, dir);
 773 }
 774
 775 static void __iommu_sync_single_for_device(struct device *dev,
 776                                            dma_addr_t dev_addr, size_t size,
 777                                            enum dma_data_direction dir)
 778 {
 779         phys_addr_t phys;
 780
 781         if (is_device_dma_coherent(dev))
 782                 return;
 783
 784         phys = iommu_iova_to_phys(iommu_get_domain_for_dev(dev), dev_addr);
 785         __dma_map_area(phys_to_virt(phys), size, dir);
 786 }
 787
 788 static dma_addr_t __iommu_map_page(struct device *dev, struct page *page,
 789                                    unsigned long offset, size_t size,
 790                                    enum dma_data_direction dir,
 791                                    unsigned long attrs)
 792 {
 793         bool coherent = is_device_dma_coherent(dev);
 794         int prot = dma_info_to_prot(dir, coherent, attrs);
 795         dma_addr_t dev_addr = iommu_dma_map_page(dev, page, offset, size, prot);
 796
 797         if (!iommu_dma_mapping_error(dev, dev_addr) &&
 798             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 799                 __iommu_sync_single_for_device(dev, dev_addr, size, dir);
 800
 801         return dev_addr;
 802 }
 803
 804 static void __iommu_unmap_page(struct device *dev, dma_addr_t dev_addr,
 805                                size_t size, enum dma_data_direction dir,
 806                                unsigned long attrs)
 807 {
 808         if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 809                 __iommu_sync_single_for_cpu(dev, dev_addr, size, dir);
 810
 811         iommu_dma_unmap_page(dev, dev_addr, size, dir, attrs);
 812 }
 813
 814 static void __iommu_sync_sg_for_cpu(struct device *dev,
 815                                     struct scatterlist *sgl, int nelems,
 816                                     enum dma_data_direction dir)
 817 {
 818         struct scatterlist *sg;
 819         int i;
 820
 821         if (is_device_dma_coherent(dev))
 822                 return;
 823
 824         for_each_sg(sgl, sg, nelems, i)
 825                 __dma_unmap_area(sg_virt(sg), sg->length, dir);
 826 }
 827
 828 static void __iommu_sync_sg_for_device(struct device *dev,
 829                                        struct scatterlist *sgl, int nelems,
 830                                        enum dma_data_direction dir)
 831 {
 832         struct scatterlist *sg;
 833         int i;
 834
 835         if (is_device_dma_coherent(dev))
 836                 return;
 837
 838         for_each_sg(sgl, sg, nelems, i)
 839                 __dma_map_area(sg_virt(sg), sg->length, dir);
 840 }
 841
 842 static int __iommu_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
 843                                 int nelems, enum dma_data_direction dir,
 844                                 unsigned long attrs)
 845 {
 846         bool coherent = is_device_dma_coherent(dev);
 847
 848         if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 849                 __iommu_sync_sg_for_device(dev, sgl, nelems, dir);
 850
 851         return iommu_dma_map_sg(dev, sgl, nelems,
 852                                 dma_info_to_prot(dir, coherent, attrs));
 853 }
 854
 855 static void __iommu_unmap_sg_attrs(struct device *dev,
 856                                    struct scatterlist *sgl, int nelems,
 857                                    enum dma_data_direction dir,
 858                                    unsigned long attrs)
 859 {
 860         if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 861                 __iommu_sync_sg_for_cpu(dev, sgl, nelems, dir);
 862
 863         iommu_dma_unmap_sg(dev, sgl, nelems, dir, attrs);
 864 }
 865
 866 static const struct dma_map_ops iommu_dma_ops = {
 867         .alloc = __iommu_alloc_attrs,
 868         .free = __iommu_free_attrs,
 869         .mmap = __iommu_mmap_attrs,
 870         .get_sgtable = __iommu_get_sgtable,
 871         .map_page = __iommu_map_page,
 872         .unmap_page = __iommu_unmap_page,
 873         .map_sg = __iommu_map_sg_attrs,
 874         .unmap_sg = __iommu_unmap_sg_attrs,
 875         .sync_single_for_cpu = __iommu_sync_single_for_cpu,
 876         .sync_single_for_device = __iommu_sync_single_for_device,
 877         .sync_sg_for_cpu = __iommu_sync_sg_for_cpu,
 878         .sync_sg_for_device = __iommu_sync_sg_for_device,
 879         .map_resource = iommu_dma_map_resource,
 880         .unmap_resource = iommu_dma_unmap_resource,
 881         .mapping_error = iommu_dma_mapping_error,
 882 };
 883
 884 static int __init __iommu_dma_init(void)
 885 {
 886         return iommu_dma_init();
 887 }
 888 arch_initcall(__iommu_dma_init);
 889
 890 static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
 891                                   const struct iommu_ops *ops)
 892 {
 893         struct iommu_domain *domain;
 894
 895         if (!ops)
 896                 return;
 897
 898         /*
 899          * The IOMMU core code allocates the default DMA domain, which the
 900          * underlying IOMMU driver needs to support via the dma-iommu layer.
 901          */
 902         domain = iommu_get_domain_for_dev(dev);
 903
 904         if (!domain)
 905                 goto out_err;
 906
 907         if (domain->type == IOMMU_DOMAIN_DMA) {
 908                 if (iommu_dma_init_domain(domain, dma_base, size, dev))
 909                         goto out_err;
 910
 911                 dev->dma_ops = &iommu_dma_ops;
 912         }
 913
 914         return;
 915
 916 out_err:
 917          pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
 918                  dev_name(dev));
 919 }
 920
 921 void arch_teardown_dma_ops(struct device *dev)
 922 {
 923         dev->dma_ops = NULL;
 924 }
 925
 926 #else
 927
 928 static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
 929                                   const struct iommu_ops *iommu)
 930 { }
 931
 932 #endif  /* CONFIG_IOMMU_DMA */
 933
 934 void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
 935                         const struct iommu_ops *iommu, bool coherent)
 936 {
 937         if (!dev->dma_ops)
 938                 dev->dma_ops = &swiotlb_dma_ops;
 939
 940         dev->archdata.dma_coherent = coherent;
 941         __iommu_setup_dma_ops(dev, dma_base, size, iommu);
 942
 943 #ifdef CONFIG_XEN
 944         if (xen_initial_domain()) {
 945                 dev->archdata.dev_dma_ops = dev->dma_ops;
 946                 dev->dma_ops = xen_dma_ops;
 947         }
 948 #endif
 949 }