arch/arm64/mm/dma-mapping.c

   1 /*
   2  * SWIOTLB-based DMA API implementation
   3  *
   4  * Copyright (C) 2012 ARM Ltd.
   5  * Author: Catalin Marinas <[email protected]>
   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
  32 #include <asm/cacheflush.h>
  33
  34 static int swiotlb __ro_after_init;
  35
  36 static pgprot_t __get_dma_pgprot(unsigned long attrs, pgprot_t prot,
  37                                  bool coherent)
  38 {
  39         if (!coherent || (attrs & DMA_ATTR_WRITE_COMBINE))
  40                 return pgprot_writecombine(prot);
  41         return prot;
  42 }
  43
  44 static struct gen_pool *atomic_pool;
  45
  46 #define DEFAULT_DMA_COHERENT_POOL_SIZE  SZ_256K
  47 static size_t atomic_pool_size __initdata = DEFAULT_DMA_COHERENT_POOL_SIZE;
  48
  49 static int __init early_coherent_pool(char *p)
  50 {
  51         atomic_pool_size = memparse(p, &p);
  52         return 0;
  53 }
  54 early_param("coherent_pool", early_coherent_pool);
  55
  56 static void *__alloc_from_pool(size_t size, struct page **ret_page, gfp_t flags)
  57 {
  58         unsigned long val;
  59         void *ptr = NULL;
  60
  61         if (!atomic_pool) {
  62                 WARN(1, "coherent pool not initialised!\n");
  63                 return NULL;
  64         }
  65
  66         val = gen_pool_alloc(atomic_pool, size);
  67         if (val) {
  68                 phys_addr_t phys = gen_pool_virt_to_phys(atomic_pool, val);
  69
  70                 *ret_page = phys_to_page(phys);
  71                 ptr = (void *)val;
  72                 memset(ptr, 0, size);
  73         }
  74
  75         return ptr;
  76 }
  77
  78 static bool __in_atomic_pool(void *start, size_t size)
  79 {
  80         return addr_in_gen_pool(atomic_pool, (unsigned long)start, size);
  81 }
  82
  83 static int __free_from_pool(void *start, size_t size)
  84 {
  85         if (!__in_atomic_pool(start, size))
  86                 return 0;
  87
  88         gen_pool_free(atomic_pool, (unsigned long)start, size);
  89
  90         return 1;
  91 }
  92
  93 static void *__dma_alloc_coherent(struct device *dev, size_t size,
  94                                   dma_addr_t *dma_handle, gfp_t flags,
  95                                   unsigned long attrs)
  96 {
  97         if (dev == NULL) {
  98                 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
  99                 return NULL;
 100         }
 101
 102         if (IS_ENABLED(CONFIG_ZONE_DMA) &&
 103             dev->coherent_dma_mask <= DMA_BIT_MASK(32))
 104                 flags |= GFP_DMA;
 105         if (dev_get_cma_area(dev) && gfpflags_allow_blocking(flags)) {
 106                 struct page *page;
 107                 void *addr;
 108
 109                 page = dma_alloc_from_contiguous(dev, size >> PAGE_SHIFT,
 110                                                  get_order(size), flags);
 111                 if (!page)
 112                         return NULL;
 113
 114                 *dma_handle = phys_to_dma(dev, page_to_phys(page));
 115                 addr = page_address(page);
 116                 memset(addr, 0, size);
 117                 return addr;
 118         } else {
 119                 return swiotlb_alloc_coherent(dev, size, dma_handle, flags);
 120         }
 121 }
 122
 123 static void __dma_free_coherent(struct device *dev, size_t size,
 124                                 void *vaddr, dma_addr_t dma_handle,
 125                                 unsigned long attrs)
 126 {
 127         bool freed;
 128         phys_addr_t paddr = dma_to_phys(dev, dma_handle);
 129
 130         if (dev == NULL) {
 131                 WARN_ONCE(1, "Use an actual device structure for DMA allocation\n");
 132                 return;
 133         }
 134
 135         freed = dma_release_from_contiguous(dev,
 136                                         phys_to_page(paddr),
 137                                         size >> PAGE_SHIFT);
 138         if (!freed)
 139                 swiotlb_free_coherent(dev, size, vaddr, dma_handle);
 140 }
 141
 142 static void *__dma_alloc(struct device *dev, size_t size,
 143                          dma_addr_t *dma_handle, gfp_t flags,
 144                          unsigned long attrs)
 145 {
 146         struct page *page;
 147         void *ptr, *coherent_ptr;
 148         bool coherent = is_device_dma_coherent(dev);
 149         pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, false);
 150
 151         size = PAGE_ALIGN(size);
 152
 153         if (!coherent && !gfpflags_allow_blocking(flags)) {
 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_area(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                                                    prot, 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                        unsigned long attrs)
 193 {
 194         void *swiotlb_addr = phys_to_virt(dma_to_phys(dev, dma_handle));
 195
 196         size = PAGE_ALIGN(size);
 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                                      unsigned long 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             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 216                 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 217
 218         return dev_addr;
 219 }
 220
 221
 222 static void __swiotlb_unmap_page(struct device *dev, dma_addr_t dev_addr,
 223                                  size_t size, enum dma_data_direction dir,
 224                                  unsigned long attrs)
 225 {
 226         if (!is_device_dma_coherent(dev) &&
 227             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 228                 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 229         swiotlb_unmap_page(dev, dev_addr, size, dir, attrs);
 230 }
 231
 232 static int __swiotlb_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
 233                                   int nelems, enum dma_data_direction dir,
 234                                   unsigned long attrs)
 235 {
 236         struct scatterlist *sg;
 237         int i, ret;
 238
 239         ret = swiotlb_map_sg_attrs(dev, sgl, nelems, dir, attrs);
 240         if (!is_device_dma_coherent(dev) &&
 241             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 242                 for_each_sg(sgl, sg, ret, i)
 243                         __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 244                                        sg->length, dir);
 245
 246         return ret;
 247 }
 248
 249 static void __swiotlb_unmap_sg_attrs(struct device *dev,
 250                                      struct scatterlist *sgl, int nelems,
 251                                      enum dma_data_direction dir,
 252                                      unsigned long attrs)
 253 {
 254         struct scatterlist *sg;
 255         int i;
 256
 257         if (!is_device_dma_coherent(dev) &&
 258             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 259                 for_each_sg(sgl, sg, nelems, i)
 260                         __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 261                                          sg->length, dir);
 262         swiotlb_unmap_sg_attrs(dev, sgl, nelems, dir, attrs);
 263 }
 264
 265 static void __swiotlb_sync_single_for_cpu(struct device *dev,
 266                                           dma_addr_t dev_addr, size_t size,
 267                                           enum dma_data_direction dir)
 268 {
 269         if (!is_device_dma_coherent(dev))
 270                 __dma_unmap_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 271         swiotlb_sync_single_for_cpu(dev, dev_addr, size, dir);
 272 }
 273
 274 static void __swiotlb_sync_single_for_device(struct device *dev,
 275                                              dma_addr_t dev_addr, size_t size,
 276                                              enum dma_data_direction dir)
 277 {
 278         swiotlb_sync_single_for_device(dev, dev_addr, size, dir);
 279         if (!is_device_dma_coherent(dev))
 280                 __dma_map_area(phys_to_virt(dma_to_phys(dev, dev_addr)), size, dir);
 281 }
 282
 283 static void __swiotlb_sync_sg_for_cpu(struct device *dev,
 284                                       struct scatterlist *sgl, int nelems,
 285                                       enum dma_data_direction dir)
 286 {
 287         struct scatterlist *sg;
 288         int i;
 289
 290         if (!is_device_dma_coherent(dev))
 291                 for_each_sg(sgl, sg, nelems, i)
 292                         __dma_unmap_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 293                                          sg->length, dir);
 294         swiotlb_sync_sg_for_cpu(dev, sgl, nelems, dir);
 295 }
 296
 297 static void __swiotlb_sync_sg_for_device(struct device *dev,
 298                                          struct scatterlist *sgl, int nelems,
 299                                          enum dma_data_direction dir)
 300 {
 301         struct scatterlist *sg;
 302         int i;
 303
 304         swiotlb_sync_sg_for_device(dev, sgl, nelems, dir);
 305         if (!is_device_dma_coherent(dev))
 306                 for_each_sg(sgl, sg, nelems, i)
 307                         __dma_map_area(phys_to_virt(dma_to_phys(dev, sg->dma_address)),
 308                                        sg->length, dir);
 309 }
 310
 311 static int __swiotlb_mmap(struct device *dev,
 312                           struct vm_area_struct *vma,
 313                           void *cpu_addr, dma_addr_t dma_addr, size_t size,
 314                           unsigned long attrs)
 315 {
 316         int ret = -ENXIO;
 317         unsigned long nr_vma_pages = (vma->vm_end - vma->vm_start) >>
 318                                         PAGE_SHIFT;
 319         unsigned long nr_pages = PAGE_ALIGN(size) >> PAGE_SHIFT;
 320         unsigned long pfn = dma_to_phys(dev, dma_addr) >> PAGE_SHIFT;
 321         unsigned long off = vma->vm_pgoff;
 322
 323         vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
 324                                              is_device_dma_coherent(dev));
 325
 326         if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
 327                 return ret;
 328
 329         if (off < nr_pages && nr_vma_pages <= (nr_pages - off)) {
 330                 ret = remap_pfn_range(vma, vma->vm_start,
 331                                       pfn + off,
 332                                       vma->vm_end - vma->vm_start,
 333                                       vma->vm_page_prot);
 334         }
 335
 336         return ret;
 337 }
 338
 339 static int __swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
 340                                  void *cpu_addr, dma_addr_t handle, size_t size,
 341                                  unsigned long attrs)
 342 {
 343         int ret = sg_alloc_table(sgt, 1, GFP_KERNEL);
 344
 345         if (!ret)
 346                 sg_set_page(sgt->sgl, phys_to_page(dma_to_phys(dev, handle)),
 347                             PAGE_ALIGN(size), 0);
 348
 349         return ret;
 350 }
 351
 352 static int __swiotlb_dma_supported(struct device *hwdev, u64 mask)
 353 {
 354         if (swiotlb)
 355                 return swiotlb_dma_supported(hwdev, mask);
 356         return 1;
 357 }
 358
 359 static int __swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t addr)
 360 {
 361         if (swiotlb)
 362                 return swiotlb_dma_mapping_error(hwdev, addr);
 363         return 0;
 364 }
 365
 366 static const struct dma_map_ops swiotlb_dma_ops = {
 367         .alloc = __dma_alloc,
 368         .free = __dma_free,
 369         .mmap = __swiotlb_mmap,
 370         .get_sgtable = __swiotlb_get_sgtable,
 371         .map_page = __swiotlb_map_page,
 372         .unmap_page = __swiotlb_unmap_page,
 373         .map_sg = __swiotlb_map_sg_attrs,
 374         .unmap_sg = __swiotlb_unmap_sg_attrs,
 375         .sync_single_for_cpu = __swiotlb_sync_single_for_cpu,
 376         .sync_single_for_device = __swiotlb_sync_single_for_device,
 377         .sync_sg_for_cpu = __swiotlb_sync_sg_for_cpu,
 378         .sync_sg_for_device = __swiotlb_sync_sg_for_device,
 379         .dma_supported = __swiotlb_dma_supported,
 380         .mapping_error = __swiotlb_dma_mapping_error,
 381 };
 382
 383 static int __init atomic_pool_init(void)
 384 {
 385         pgprot_t prot = __pgprot(PROT_NORMAL_NC);
 386         unsigned long nr_pages = atomic_pool_size >> PAGE_SHIFT;
 387         struct page *page;
 388         void *addr;
 389         unsigned int pool_size_order = get_order(atomic_pool_size);
 390
 391         if (dev_get_cma_area(NULL))
 392                 page = dma_alloc_from_contiguous(NULL, nr_pages,
 393                                                  pool_size_order, GFP_KERNEL);
 394         else
 395                 page = alloc_pages(GFP_DMA, pool_size_order);
 396
 397         if (page) {
 398                 int ret;
 399                 void *page_addr = page_address(page);
 400
 401                 memset(page_addr, 0, atomic_pool_size);
 402                 __dma_flush_area(page_addr, atomic_pool_size);
 403
 404                 atomic_pool = gen_pool_create(PAGE_SHIFT, -1);
 405                 if (!atomic_pool)
 406                         goto free_page;
 407
 408                 addr = dma_common_contiguous_remap(page, atomic_pool_size,
 409                                         VM_USERMAP, prot, atomic_pool_init);
 410
 411                 if (!addr)
 412                         goto destroy_genpool;
 413
 414                 ret = gen_pool_add_virt(atomic_pool, (unsigned long)addr,
 415                                         page_to_phys(page),
 416                                         atomic_pool_size, -1);
 417                 if (ret)
 418                         goto remove_mapping;
 419
 420                 gen_pool_set_algo(atomic_pool,
 421                                   gen_pool_first_fit_order_align,
 422                                   (void *)PAGE_SHIFT);
 423
 424                 pr_info("DMA: preallocated %zu KiB pool for atomic allocations\n",
 425                         atomic_pool_size / 1024);
 426                 return 0;
 427         }
 428         goto out;
 429
 430 remove_mapping:
 431         dma_common_free_remap(addr, atomic_pool_size, VM_USERMAP);
 432 destroy_genpool:
 433         gen_pool_destroy(atomic_pool);
 434         atomic_pool = NULL;
 435 free_page:
 436         if (!dma_release_from_contiguous(NULL, page, nr_pages))
 437                 __free_pages(page, pool_size_order);
 438 out:
 439         pr_err("DMA: failed to allocate %zu KiB pool for atomic coherent allocation\n",
 440                 atomic_pool_size / 1024);
 441         return -ENOMEM;
 442 }
 443
 444 /********************************************
 445  * The following APIs are for dummy DMA ops *
 446  ********************************************/
 447
 448 static void *__dummy_alloc(struct device *dev, size_t size,
 449                            dma_addr_t *dma_handle, gfp_t flags,
 450                            unsigned long attrs)
 451 {
 452         return NULL;
 453 }
 454
 455 static void __dummy_free(struct device *dev, size_t size,
 456                          void *vaddr, dma_addr_t dma_handle,
 457                          unsigned long attrs)
 458 {
 459 }
 460
 461 static int __dummy_mmap(struct device *dev,
 462                         struct vm_area_struct *vma,
 463                         void *cpu_addr, dma_addr_t dma_addr, size_t size,
 464                         unsigned long attrs)
 465 {
 466         return -ENXIO;
 467 }
 468
 469 static dma_addr_t __dummy_map_page(struct device *dev, struct page *page,
 470                                    unsigned long offset, size_t size,
 471                                    enum dma_data_direction dir,
 472                                    unsigned long attrs)
 473 {
 474         return DMA_ERROR_CODE;
 475 }
 476
 477 static void __dummy_unmap_page(struct device *dev, dma_addr_t dev_addr,
 478                                size_t size, enum dma_data_direction dir,
 479                                unsigned long attrs)
 480 {
 481 }
 482
 483 static int __dummy_map_sg(struct device *dev, struct scatterlist *sgl,
 484                           int nelems, enum dma_data_direction dir,
 485                           unsigned long attrs)
 486 {
 487         return 0;
 488 }
 489
 490 static void __dummy_unmap_sg(struct device *dev,
 491                              struct scatterlist *sgl, int nelems,
 492                              enum dma_data_direction dir,
 493                              unsigned long attrs)
 494 {
 495 }
 496
 497 static void __dummy_sync_single(struct device *dev,
 498                                 dma_addr_t dev_addr, size_t size,
 499                                 enum dma_data_direction dir)
 500 {
 501 }
 502
 503 static void __dummy_sync_sg(struct device *dev,
 504                             struct scatterlist *sgl, int nelems,
 505                             enum dma_data_direction dir)
 506 {
 507 }
 508
 509 static int __dummy_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
 510 {
 511         return 1;
 512 }
 513
 514 static int __dummy_dma_supported(struct device *hwdev, u64 mask)
 515 {
 516         return 0;
 517 }
 518
 519 const struct dma_map_ops dummy_dma_ops = {
 520         .alloc                  = __dummy_alloc,
 521         .free                   = __dummy_free,
 522         .mmap                   = __dummy_mmap,
 523         .map_page               = __dummy_map_page,
 524         .unmap_page             = __dummy_unmap_page,
 525         .map_sg                 = __dummy_map_sg,
 526         .unmap_sg               = __dummy_unmap_sg,
 527         .sync_single_for_cpu    = __dummy_sync_single,
 528         .sync_single_for_device = __dummy_sync_single,
 529         .sync_sg_for_cpu        = __dummy_sync_sg,
 530         .sync_sg_for_device     = __dummy_sync_sg,
 531         .mapping_error          = __dummy_mapping_error,
 532         .dma_supported          = __dummy_dma_supported,
 533 };
 534 EXPORT_SYMBOL(dummy_dma_ops);
 535
 536 static int __init arm64_dma_init(void)
 537 {
 538         if (swiotlb_force == SWIOTLB_FORCE ||
 539             max_pfn > (arm64_dma_phys_limit >> PAGE_SHIFT))
 540                 swiotlb = 1;
 541
 542         return atomic_pool_init();
 543 }
 544 arch_initcall(arm64_dma_init);
 545
 546 #define PREALLOC_DMA_DEBUG_ENTRIES      4096
 547
 548 static int __init dma_debug_do_init(void)
 549 {
 550         dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
 551         return 0;
 552 }
 553 fs_initcall(dma_debug_do_init);
 554
 555
 556 #ifdef CONFIG_IOMMU_DMA
 557 #include <linux/dma-iommu.h>
 558 #include <linux/platform_device.h>
 559 #include <linux/amba/bus.h>
 560
 561 /* Thankfully, all cache ops are by VA so we can ignore phys here */
 562 static void flush_page(struct device *dev, const void *virt, phys_addr_t phys)
 563 {
 564         __dma_flush_area(virt, PAGE_SIZE);
 565 }
 566
 567 static void *__iommu_alloc_attrs(struct device *dev, size_t size,
 568                                  dma_addr_t *handle, gfp_t gfp,
 569                                  unsigned long attrs)
 570 {
 571         bool coherent = is_device_dma_coherent(dev);
 572         int ioprot = dma_info_to_prot(DMA_BIDIRECTIONAL, coherent, attrs);
 573         size_t iosize = size;
 574         void *addr;
 575
 576         if (WARN(!dev, "cannot create IOMMU mapping for unknown device\n"))
 577                 return NULL;
 578
 579         size = PAGE_ALIGN(size);
 580
 581         /*
 582          * Some drivers rely on this, and we probably don't want the
 583          * possibility of stale kernel data being read by devices anyway.
 584          */
 585         gfp |= __GFP_ZERO;
 586
 587         if (gfpflags_allow_blocking(gfp)) {
 588                 struct page **pages;
 589                 pgprot_t prot = __get_dma_pgprot(attrs, PAGE_KERNEL, coherent);
 590
 591                 pages = iommu_dma_alloc(dev, iosize, gfp, attrs, ioprot,
 592                                         handle, flush_page);
 593                 if (!pages)
 594                         return NULL;
 595
 596                 addr = dma_common_pages_remap(pages, size, VM_USERMAP, prot,
 597                                               __builtin_return_address(0));
 598                 if (!addr)
 599                         iommu_dma_free(dev, pages, iosize, handle);
 600         } else {
 601                 struct page *page;
 602                 /*
 603                  * In atomic context we can't remap anything, so we'll only
 604                  * get the virtually contiguous buffer we need by way of a
 605                  * physically contiguous allocation.
 606                  */
 607                 if (coherent) {
 608                         page = alloc_pages(gfp, get_order(size));
 609                         addr = page ? page_address(page) : NULL;
 610                 } else {
 611                         addr = __alloc_from_pool(size, &page, gfp);
 612                 }
 613                 if (!addr)
 614                         return NULL;
 615
 616                 *handle = iommu_dma_map_page(dev, page, 0, iosize, ioprot);
 617                 if (iommu_dma_mapping_error(dev, *handle)) {
 618                         if (coherent)
 619                                 __free_pages(page, get_order(size));
 620                         else
 621                                 __free_from_pool(addr, size);
 622                         addr = NULL;
 623                 }
 624         }
 625         return addr;
 626 }
 627
 628 static void __iommu_free_attrs(struct device *dev, size_t size, void *cpu_addr,
 629                                dma_addr_t handle, unsigned long attrs)
 630 {
 631         size_t iosize = size;
 632
 633         size = PAGE_ALIGN(size);
 634         /*
 635          * @cpu_addr will be one of 3 things depending on how it was allocated:
 636          * - A remapped array of pages from iommu_dma_alloc(), for all
 637          *   non-atomic allocations.
 638          * - A non-cacheable alias from the atomic pool, for atomic
 639          *   allocations by non-coherent devices.
 640          * - A normal lowmem address, for atomic allocations by
 641          *   coherent devices.
 642          * Hence how dodgy the below logic looks...
 643          */
 644         if (__in_atomic_pool(cpu_addr, size)) {
 645                 iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
 646                 __free_from_pool(cpu_addr, size);
 647         } else if (is_vmalloc_addr(cpu_addr)){
 648                 struct vm_struct *area = find_vm_area(cpu_addr);
 649
 650                 if (WARN_ON(!area || !area->pages))
 651                         return;
 652                 iommu_dma_free(dev, area->pages, iosize, &handle);
 653                 dma_common_free_remap(cpu_addr, size, VM_USERMAP);
 654         } else {
 655                 iommu_dma_unmap_page(dev, handle, iosize, 0, 0);
 656                 __free_pages(virt_to_page(cpu_addr), get_order(size));
 657         }
 658 }
 659
 660 static int __iommu_mmap_attrs(struct device *dev, struct vm_area_struct *vma,
 661                               void *cpu_addr, dma_addr_t dma_addr, size_t size,
 662                               unsigned long attrs)
 663 {
 664         struct vm_struct *area;
 665         int ret;
 666
 667         vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot,
 668                                              is_device_dma_coherent(dev));
 669
 670         if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
 671                 return ret;
 672
 673         area = find_vm_area(cpu_addr);
 674         if (WARN_ON(!area || !area->pages))
 675                 return -ENXIO;
 676
 677         return iommu_dma_mmap(area->pages, size, vma);
 678 }
 679
 680 static int __iommu_get_sgtable(struct device *dev, struct sg_table *sgt,
 681                                void *cpu_addr, dma_addr_t dma_addr,
 682                                size_t size, unsigned long attrs)
 683 {
 684         unsigned int count = PAGE_ALIGN(size) >> PAGE_SHIFT;
 685         struct vm_struct *area = find_vm_area(cpu_addr);
 686
 687         if (WARN_ON(!area || !area->pages))
 688                 return -ENXIO;
 689
 690         return sg_alloc_table_from_pages(sgt, area->pages, count, 0, size,
 691                                          GFP_KERNEL);
 692 }
 693
 694 static void __iommu_sync_single_for_cpu(struct device *dev,
 695                                         dma_addr_t dev_addr, size_t size,
 696                                         enum dma_data_direction dir)
 697 {
 698         phys_addr_t phys;
 699
 700         if (is_device_dma_coherent(dev))
 701                 return;
 702
 703         phys = iommu_iova_to_phys(iommu_get_domain_for_dev(dev), dev_addr);
 704         __dma_unmap_area(phys_to_virt(phys), size, dir);
 705 }
 706
 707 static void __iommu_sync_single_for_device(struct device *dev,
 708                                            dma_addr_t dev_addr, size_t size,
 709                                            enum dma_data_direction dir)
 710 {
 711         phys_addr_t phys;
 712
 713         if (is_device_dma_coherent(dev))
 714                 return;
 715
 716         phys = iommu_iova_to_phys(iommu_get_domain_for_dev(dev), dev_addr);
 717         __dma_map_area(phys_to_virt(phys), size, dir);
 718 }
 719
 720 static dma_addr_t __iommu_map_page(struct device *dev, struct page *page,
 721                                    unsigned long offset, size_t size,
 722                                    enum dma_data_direction dir,
 723                                    unsigned long attrs)
 724 {
 725         bool coherent = is_device_dma_coherent(dev);
 726         int prot = dma_info_to_prot(dir, coherent, attrs);
 727         dma_addr_t dev_addr = iommu_dma_map_page(dev, page, offset, size, prot);
 728
 729         if (!iommu_dma_mapping_error(dev, dev_addr) &&
 730             (attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 731                 __iommu_sync_single_for_device(dev, dev_addr, size, dir);
 732
 733         return dev_addr;
 734 }
 735
 736 static void __iommu_unmap_page(struct device *dev, dma_addr_t dev_addr,
 737                                size_t size, enum dma_data_direction dir,
 738                                unsigned long attrs)
 739 {
 740         if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 741                 __iommu_sync_single_for_cpu(dev, dev_addr, size, dir);
 742
 743         iommu_dma_unmap_page(dev, dev_addr, size, dir, attrs);
 744 }
 745
 746 static void __iommu_sync_sg_for_cpu(struct device *dev,
 747                                     struct scatterlist *sgl, int nelems,
 748                                     enum dma_data_direction dir)
 749 {
 750         struct scatterlist *sg;
 751         int i;
 752
 753         if (is_device_dma_coherent(dev))
 754                 return;
 755
 756         for_each_sg(sgl, sg, nelems, i)
 757                 __dma_unmap_area(sg_virt(sg), sg->length, dir);
 758 }
 759
 760 static void __iommu_sync_sg_for_device(struct device *dev,
 761                                        struct scatterlist *sgl, int nelems,
 762                                        enum dma_data_direction dir)
 763 {
 764         struct scatterlist *sg;
 765         int i;
 766
 767         if (is_device_dma_coherent(dev))
 768                 return;
 769
 770         for_each_sg(sgl, sg, nelems, i)
 771                 __dma_map_area(sg_virt(sg), sg->length, dir);
 772 }
 773
 774 static int __iommu_map_sg_attrs(struct device *dev, struct scatterlist *sgl,
 775                                 int nelems, enum dma_data_direction dir,
 776                                 unsigned long attrs)
 777 {
 778         bool coherent = is_device_dma_coherent(dev);
 779
 780         if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 781                 __iommu_sync_sg_for_device(dev, sgl, nelems, dir);
 782
 783         return iommu_dma_map_sg(dev, sgl, nelems,
 784                                 dma_info_to_prot(dir, coherent, attrs));
 785 }
 786
 787 static void __iommu_unmap_sg_attrs(struct device *dev,
 788                                    struct scatterlist *sgl, int nelems,
 789                                    enum dma_data_direction dir,
 790                                    unsigned long attrs)
 791 {
 792         if ((attrs & DMA_ATTR_SKIP_CPU_SYNC) == 0)
 793                 __iommu_sync_sg_for_cpu(dev, sgl, nelems, dir);
 794
 795         iommu_dma_unmap_sg(dev, sgl, nelems, dir, attrs);
 796 }
 797
 798 static const struct dma_map_ops iommu_dma_ops = {
 799         .alloc = __iommu_alloc_attrs,
 800         .free = __iommu_free_attrs,
 801         .mmap = __iommu_mmap_attrs,
 802         .get_sgtable = __iommu_get_sgtable,
 803         .map_page = __iommu_map_page,
 804         .unmap_page = __iommu_unmap_page,
 805         .map_sg = __iommu_map_sg_attrs,
 806         .unmap_sg = __iommu_unmap_sg_attrs,
 807         .sync_single_for_cpu = __iommu_sync_single_for_cpu,
 808         .sync_single_for_device = __iommu_sync_single_for_device,
 809         .sync_sg_for_cpu = __iommu_sync_sg_for_cpu,
 810         .sync_sg_for_device = __iommu_sync_sg_for_device,
 811         .map_resource = iommu_dma_map_resource,
 812         .unmap_resource = iommu_dma_unmap_resource,
 813         .mapping_error = iommu_dma_mapping_error,
 814 };
 815
 816 /*
 817  * TODO: Right now __iommu_setup_dma_ops() gets called too early to do
 818  * everything it needs to - the device is only partially created and the
 819  * IOMMU driver hasn't seen it yet, so it can't have a group. Thus we
 820  * need this delayed attachment dance. Once IOMMU probe ordering is sorted
 821  * to move the arch_setup_dma_ops() call later, all the notifier bits below
 822  * become unnecessary, and will go away.
 823  */
 824 struct iommu_dma_notifier_data {
 825         struct list_head list;
 826         struct device *dev;
 827         const struct iommu_ops *ops;
 828         u64 dma_base;
 829         u64 size;
 830 };
 831 static LIST_HEAD(iommu_dma_masters);
 832 static DEFINE_MUTEX(iommu_dma_notifier_lock);
 833
 834 static bool do_iommu_attach(struct device *dev, const struct iommu_ops *ops,
 835                            u64 dma_base, u64 size)
 836 {
 837         struct iommu_domain *domain = iommu_get_domain_for_dev(dev);
 838
 839         /*
 840          * If the IOMMU driver has the DMA domain support that we require,
 841          * then the IOMMU core will have already configured a group for this
 842          * device, and allocated the default domain for that group.
 843          */
 844         if (!domain)
 845                 goto out_err;
 846
 847         if (domain->type == IOMMU_DOMAIN_DMA) {
 848                 if (iommu_dma_init_domain(domain, dma_base, size, dev))
 849                         goto out_err;
 850
 851                 dev->dma_ops = &iommu_dma_ops;
 852         }
 853
 854         return true;
 855 out_err:
 856         pr_warn("Failed to set up IOMMU for device %s; retaining platform DMA ops\n",
 857                  dev_name(dev));
 858         return false;
 859 }
 860
 861 static void queue_iommu_attach(struct device *dev, const struct iommu_ops *ops,
 862                               u64 dma_base, u64 size)
 863 {
 864         struct iommu_dma_notifier_data *iommudata;
 865
 866         iommudata = kzalloc(sizeof(*iommudata), GFP_KERNEL);
 867         if (!iommudata)
 868                 return;
 869
 870         iommudata->dev = dev;
 871         iommudata->ops = ops;
 872         iommudata->dma_base = dma_base;
 873         iommudata->size = size;
 874
 875         mutex_lock(&iommu_dma_notifier_lock);
 876         list_add(&iommudata->list, &iommu_dma_masters);
 877         mutex_unlock(&iommu_dma_notifier_lock);
 878 }
 879
 880 static int __iommu_attach_notifier(struct notifier_block *nb,
 881                                    unsigned long action, void *data)
 882 {
 883         struct iommu_dma_notifier_data *master, *tmp;
 884
 885         if (action != BUS_NOTIFY_BIND_DRIVER)
 886                 return 0;
 887
 888         mutex_lock(&iommu_dma_notifier_lock);
 889         list_for_each_entry_safe(master, tmp, &iommu_dma_masters, list) {
 890                 if (data == master->dev && do_iommu_attach(master->dev,
 891                                 master->ops, master->dma_base, master->size)) {
 892                         list_del(&master->list);
 893                         kfree(master);
 894                         break;
 895                 }
 896         }
 897         mutex_unlock(&iommu_dma_notifier_lock);
 898         return 0;
 899 }
 900
 901 static int __init register_iommu_dma_ops_notifier(struct bus_type *bus)
 902 {
 903         struct notifier_block *nb = kzalloc(sizeof(*nb), GFP_KERNEL);
 904         int ret;
 905
 906         if (!nb)
 907                 return -ENOMEM;
 908
 909         nb->notifier_call = __iommu_attach_notifier;
 910
 911         ret = bus_register_notifier(bus, nb);
 912         if (ret) {
 913                 pr_warn("Failed to register DMA domain notifier; IOMMU DMA ops unavailable on bus '%s'\n",
 914                         bus->name);
 915                 kfree(nb);
 916         }
 917         return ret;
 918 }
 919
 920 static int __init __iommu_dma_init(void)
 921 {
 922         int ret;
 923
 924         ret = iommu_dma_init();
 925         if (!ret)
 926                 ret = register_iommu_dma_ops_notifier(&platform_bus_type);
 927         if (!ret)
 928                 ret = register_iommu_dma_ops_notifier(&amba_bustype);
 929 #ifdef CONFIG_PCI
 930         if (!ret)
 931                 ret = register_iommu_dma_ops_notifier(&pci_bus_type);
 932 #endif
 933         return ret;
 934 }
 935 arch_initcall(__iommu_dma_init);
 936
 937 static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
 938                                   const struct iommu_ops *ops)
 939 {
 940         struct iommu_group *group;
 941
 942         if (!ops)
 943                 return;
 944         /*
 945          * TODO: As a concession to the future, we're ready to handle being
 946          * called both early and late (i.e. after bus_add_device). Once all
 947          * the platform bus code is reworked to call us late and the notifier
 948          * junk above goes away, move the body of do_iommu_attach here.
 949          */
 950         group = iommu_group_get(dev);
 951         if (group) {
 952                 do_iommu_attach(dev, ops, dma_base, size);
 953                 iommu_group_put(group);
 954         } else {
 955                 queue_iommu_attach(dev, ops, dma_base, size);
 956         }
 957 }
 958
 959 void arch_teardown_dma_ops(struct device *dev)
 960 {
 961         dev->dma_ops = NULL;
 962 }
 963
 964 #else
 965
 966 static void __iommu_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
 967                                   const struct iommu_ops *iommu)
 968 { }
 969
 970 #endif  /* CONFIG_IOMMU_DMA */
 971
 972 void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
 973                         const struct iommu_ops *iommu, bool coherent)
 974 {
 975         if (!dev->dma_ops)
 976                 dev->dma_ops = &swiotlb_dma_ops;
 977
 978         dev->archdata.dma_coherent = coherent;
 979         __iommu_setup_dma_ops(dev, dma_base, size, iommu);
 980 }