arch/arm/mm/dma-mapping.c

   1 /*
   2  *  linux/arch/arm/mm/dma-mapping.c
   3  *
   4  *  Copyright (C) 2000-2004 Russell King
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 as
   8  * published by the Free Software Foundation.
   9  *
  10  *  DMA uncached mapping support.
  11  */
  12 #include <linux/module.h>
  13 #include <linux/mm.h>
  14 #include <linux/gfp.h>
  15 #include <linux/errno.h>
  16 #include <linux/list.h>
  17 #include <linux/init.h>
  18 #include <linux/device.h>
  19 #include <linux/dma-mapping.h>
  20
  21 #include <asm/memory.h>
  22 #include <asm/highmem.h>
  23 #include <asm/cacheflush.h>
  24 #include <asm/tlbflush.h>
  25 #include <asm/sizes.h>
  26
  27 /* Sanity check size */
  28 #if (CONSISTENT_DMA_SIZE % SZ_2M)
  29 #error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
  30 #endif
  31
  32 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
  33 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PGDIR_SHIFT)
  34 #define NUM_CONSISTENT_PTES (CONSISTENT_DMA_SIZE >> PGDIR_SHIFT)
  35
  36 static u64 get_coherent_dma_mask(struct device *dev)
  37 {
  38         u64 mask = ISA_DMA_THRESHOLD;
  39
  40         if (dev) {
  41                 mask = dev->coherent_dma_mask;
  42
  43                 /*
  44                  * Sanity check the DMA mask - it must be non-zero, and
  45                  * must be able to be satisfied by a DMA allocation.
  46                  */
  47                 if (mask == 0) {
  48                         dev_warn(dev, "coherent DMA mask is unset\n");
  49                         return 0;
  50                 }
  51
  52                 if ((~mask) & ISA_DMA_THRESHOLD) {
  53                         dev_warn(dev, "coherent DMA mask %#llx is smaller "
  54                                  "than system GFP_DMA mask %#llx\n",
  55                                  mask, (unsigned long long)ISA_DMA_THRESHOLD);
  56                         return 0;
  57                 }
  58         }
  59
  60         return mask;
  61 }
  62
  63 /*
  64  * Allocate a DMA buffer for 'dev' of size 'size' using the
  65  * specified gfp mask.  Note that 'size' must be page aligned.
  66  */
  67 static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
  68 {
  69         unsigned long order = get_order(size);
  70         struct page *page, *p, *e;
  71         void *ptr;
  72         u64 mask = get_coherent_dma_mask(dev);
  73
  74 #ifdef CONFIG_DMA_API_DEBUG
  75         u64 limit = (mask + 1) & ~mask;
  76         if (limit && size >= limit) {
  77                 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
  78                         size, mask);
  79                 return NULL;
  80         }
  81 #endif
  82
  83         if (!mask)
  84                 return NULL;
  85
  86         if (mask < 0xffffffffULL)
  87                 gfp |= GFP_DMA;
  88
  89         page = alloc_pages(gfp, order);
  90         if (!page)
  91                 return NULL;
  92
  93         /*
  94          * Now split the huge page and free the excess pages
  95          */
  96         split_page(page, order);
  97         for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
  98                 __free_page(p);
  99
 100         /*
 101          * Ensure that the allocated pages are zeroed, and that any data
 102          * lurking in the kernel direct-mapped region is invalidated.
 103          */
 104         ptr = page_address(page);
 105         memset(ptr, 0, size);
 106         dmac_flush_range(ptr, ptr + size);
 107         outer_flush_range(__pa(ptr), __pa(ptr) + size);
 108
 109         return page;
 110 }
 111
 112 /*
 113  * Free a DMA buffer.  'size' must be page aligned.
 114  */
 115 static void __dma_free_buffer(struct page *page, size_t size)
 116 {
 117         struct page *e = page + (size >> PAGE_SHIFT);
 118
 119         while (page < e) {
 120                 __free_page(page);
 121                 page++;
 122         }
 123 }
 124
 125 #ifdef CONFIG_MMU
 126 /*
 127  * These are the page tables (2MB each) covering uncached, DMA consistent allocations
 128  */
 129 static pte_t *consistent_pte[NUM_CONSISTENT_PTES];
 130
 131 #include "vmregion.h"
 132
 133 static struct arm_vmregion_head consistent_head = {
 134         .vm_lock        = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock),
 135         .vm_list        = LIST_HEAD_INIT(consistent_head.vm_list),
 136         .vm_start       = CONSISTENT_BASE,
 137         .vm_end         = CONSISTENT_END,
 138 };
 139
 140 #ifdef CONFIG_HUGETLB_PAGE
 141 #error ARM Coherent DMA allocator does not (yet) support huge TLB
 142 #endif
 143
 144 /*
 145  * Initialise the consistent memory allocation.
 146  */
 147 static int __init consistent_init(void)
 148 {
 149         int ret = 0;
 150         pgd_t *pgd;
 151         pmd_t *pmd;
 152         pte_t *pte;
 153         int i = 0;
 154         u32 base = CONSISTENT_BASE;
 155
 156         do {
 157                 pgd = pgd_offset(&init_mm, base);
 158                 pmd = pmd_alloc(&init_mm, pgd, base);
 159                 if (!pmd) {
 160                         printk(KERN_ERR "%s: no pmd tables\n", __func__);
 161                         ret = -ENOMEM;
 162                         break;
 163                 }
 164                 WARN_ON(!pmd_none(*pmd));
 165
 166                 pte = pte_alloc_kernel(pmd, base);
 167                 if (!pte) {
 168                         printk(KERN_ERR "%s: no pte tables\n", __func__);
 169                         ret = -ENOMEM;
 170                         break;
 171                 }
 172
 173                 consistent_pte[i++] = pte;
 174                 base += (1 << PGDIR_SHIFT);
 175         } while (base < CONSISTENT_END);
 176
 177         return ret;
 178 }
 179
 180 core_initcall(consistent_init);
 181
 182 static void *
 183 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot)
 184 {
 185         struct arm_vmregion *c;
 186
 187         if (!consistent_pte[0]) {
 188                 printk(KERN_ERR "%s: not initialised\n", __func__);
 189                 dump_stack();
 190                 return NULL;
 191         }
 192
 193         /*
 194          * Allocate a virtual address in the consistent mapping region.
 195          */
 196         c = arm_vmregion_alloc(&consistent_head, size,
 197                             gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
 198         if (c) {
 199                 pte_t *pte;
 200                 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
 201                 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
 202
 203                 pte = consistent_pte[idx] + off;
 204                 c->vm_pages = page;
 205
 206                 do {
 207                         BUG_ON(!pte_none(*pte));
 208
 209                         set_pte_ext(pte, mk_pte(page, prot), 0);
 210                         page++;
 211                         pte++;
 212                         off++;
 213                         if (off >= PTRS_PER_PTE) {
 214                                 off = 0;
 215                                 pte = consistent_pte[++idx];
 216                         }
 217                 } while (size -= PAGE_SIZE);
 218
 219                 return (void *)c->vm_start;
 220         }
 221         return NULL;
 222 }
 223
 224 static void __dma_free_remap(void *cpu_addr, size_t size)
 225 {
 226         struct arm_vmregion *c;
 227         unsigned long addr;
 228         pte_t *ptep;
 229         int idx;
 230         u32 off;
 231
 232         c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
 233         if (!c) {
 234                 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
 235                        __func__, cpu_addr);
 236                 dump_stack();
 237                 return;
 238         }
 239
 240         if ((c->vm_end - c->vm_start) != size) {
 241                 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
 242                        __func__, c->vm_end - c->vm_start, size);
 243                 dump_stack();
 244                 size = c->vm_end - c->vm_start;
 245         }
 246
 247         idx = CONSISTENT_PTE_INDEX(c->vm_start);
 248         off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
 249         ptep = consistent_pte[idx] + off;
 250         addr = c->vm_start;
 251         do {
 252                 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
 253
 254                 ptep++;
 255                 addr += PAGE_SIZE;
 256                 off++;
 257                 if (off >= PTRS_PER_PTE) {
 258                         off = 0;
 259                         ptep = consistent_pte[++idx];
 260                 }
 261
 262                 if (pte_none(pte) || !pte_present(pte))
 263                         printk(KERN_CRIT "%s: bad page in kernel page table\n",
 264                                __func__);
 265         } while (size -= PAGE_SIZE);
 266
 267         flush_tlb_kernel_range(c->vm_start, c->vm_end);
 268
 269         arm_vmregion_free(&consistent_head, c);
 270 }
 271
 272 #else   /* !CONFIG_MMU */
 273
 274 #define __dma_alloc_remap(page, size, gfp, prot)        page_address(page)
 275 #define __dma_free_remap(addr, size)                    do { } while (0)
 276
 277 #endif  /* CONFIG_MMU */
 278
 279 static void *
 280 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
 281             pgprot_t prot)
 282 {
 283         struct page *page;
 284         void *addr;
 285
 286         *handle = ~0;
 287         size = PAGE_ALIGN(size);
 288
 289         page = __dma_alloc_buffer(dev, size, gfp);
 290         if (!page)
 291                 return NULL;
 292
 293         if (!arch_is_coherent())
 294                 addr = __dma_alloc_remap(page, size, gfp, prot);
 295         else
 296                 addr = page_address(page);
 297
 298         if (addr)
 299                 *handle = page_to_dma(dev, page);
 300
 301         return addr;
 302 }
 303
 304 /*
 305  * Allocate DMA-coherent memory space and return both the kernel remapped
 306  * virtual and bus address for that space.
 307  */
 308 void *
 309 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
 310 {
 311         void *memory;
 312
 313         if (dma_alloc_from_coherent(dev, size, handle, &memory))
 314                 return memory;
 315
 316         return __dma_alloc(dev, size, handle, gfp,
 317                            pgprot_dmacoherent(pgprot_kernel));
 318 }
 319 EXPORT_SYMBOL(dma_alloc_coherent);
 320
 321 /*
 322  * Allocate a writecombining region, in much the same way as
 323  * dma_alloc_coherent above.
 324  */
 325 void *
 326 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
 327 {
 328         return __dma_alloc(dev, size, handle, gfp,
 329                            pgprot_writecombine(pgprot_kernel));
 330 }
 331 EXPORT_SYMBOL(dma_alloc_writecombine);
 332
 333 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
 334                     void *cpu_addr, dma_addr_t dma_addr, size_t size)
 335 {
 336         int ret = -ENXIO;
 337 #ifdef CONFIG_MMU
 338         unsigned long user_size, kern_size;
 339         struct arm_vmregion *c;
 340
 341         user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 342
 343         c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
 344         if (c) {
 345                 unsigned long off = vma->vm_pgoff;
 346
 347                 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
 348
 349                 if (off < kern_size &&
 350                     user_size <= (kern_size - off)) {
 351                         ret = remap_pfn_range(vma, vma->vm_start,
 352                                               page_to_pfn(c->vm_pages) + off,
 353                                               user_size << PAGE_SHIFT,
 354                                               vma->vm_page_prot);
 355                 }
 356         }
 357 #endif  /* CONFIG_MMU */
 358
 359         return ret;
 360 }
 361
 362 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
 363                       void *cpu_addr, dma_addr_t dma_addr, size_t size)
 364 {
 365         vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
 366         return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
 367 }
 368 EXPORT_SYMBOL(dma_mmap_coherent);
 369
 370 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
 371                           void *cpu_addr, dma_addr_t dma_addr, size_t size)
 372 {
 373         vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 374         return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
 375 }
 376 EXPORT_SYMBOL(dma_mmap_writecombine);
 377
 378 /*
 379  * free a page as defined by the above mapping.
 380  * Must not be called with IRQs disabled.
 381  */
 382 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
 383 {
 384         WARN_ON(irqs_disabled());
 385
 386         if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
 387                 return;
 388
 389         size = PAGE_ALIGN(size);
 390
 391         if (!arch_is_coherent())
 392                 __dma_free_remap(cpu_addr, size);
 393
 394         __dma_free_buffer(dma_to_page(dev, handle), size);
 395 }
 396 EXPORT_SYMBOL(dma_free_coherent);
 397
 398 /*
 399  * Make an area consistent for devices.
 400  * Note: Drivers should NOT use this function directly, as it will break
 401  * platforms with CONFIG_DMABOUNCE.
 402  * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
 403  */
 404 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
 405         enum dma_data_direction dir)
 406 {
 407         unsigned long paddr;
 408
 409         BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
 410
 411         dmac_map_area(kaddr, size, dir);
 412
 413         paddr = __pa(kaddr);
 414         if (dir == DMA_FROM_DEVICE) {
 415                 outer_inv_range(paddr, paddr + size);
 416         } else {
 417                 outer_clean_range(paddr, paddr + size);
 418         }
 419         /* FIXME: non-speculating: flush on bidirectional mappings? */
 420 }
 421 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
 422
 423 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
 424         enum dma_data_direction dir)
 425 {
 426         BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
 427
 428         /* FIXME: non-speculating: not required */
 429         /* don't bother invalidating if DMA to device */
 430         if (dir != DMA_TO_DEVICE) {
 431                 unsigned long paddr = __pa(kaddr);
 432                 outer_inv_range(paddr, paddr + size);
 433         }
 434
 435         dmac_unmap_area(kaddr, size, dir);
 436 }
 437 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
 438
 439 static void dma_cache_maint_page(struct page *page, unsigned long offset,
 440         size_t size, enum dma_data_direction dir,
 441         void (*op)(const void *, size_t, int))
 442 {
 443         /*
 444          * A single sg entry may refer to multiple physically contiguous
 445          * pages.  But we still need to process highmem pages individually.
 446          * If highmem is not configured then the bulk of this loop gets
 447          * optimized out.
 448          */
 449         size_t left = size;
 450         do {
 451                 size_t len = left;
 452                 void *vaddr;
 453
 454                 if (PageHighMem(page)) {
 455                         if (len + offset > PAGE_SIZE) {
 456                                 if (offset >= PAGE_SIZE) {
 457                                         page += offset / PAGE_SIZE;
 458                                         offset %= PAGE_SIZE;
 459                                 }
 460                                 len = PAGE_SIZE - offset;
 461                         }
 462                         vaddr = kmap_high_get(page);
 463                         if (vaddr) {
 464                                 vaddr += offset;
 465                                 op(vaddr, len, dir);
 466                                 kunmap_high(page);
 467                         } else if (cache_is_vipt()) {
 468                                 pte_t saved_pte;
 469                                 vaddr = kmap_high_l1_vipt(page, &saved_pte);
 470                                 op(vaddr + offset, len, dir);
 471                                 kunmap_high_l1_vipt(page, saved_pte);
 472                         }
 473                 } else {
 474                         vaddr = page_address(page) + offset;
 475                         op(vaddr, len, dir);
 476                 }
 477                 offset = 0;
 478                 page++;
 479                 left -= len;
 480         } while (left);
 481 }
 482
 483 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
 484         size_t size, enum dma_data_direction dir)
 485 {
 486         unsigned long paddr;
 487
 488         dma_cache_maint_page(page, off, size, dir, dmac_map_area);
 489
 490         paddr = page_to_phys(page) + off;
 491         if (dir == DMA_FROM_DEVICE) {
 492                 outer_inv_range(paddr, paddr + size);
 493         } else {
 494                 outer_clean_range(paddr, paddr + size);
 495         }
 496         /* FIXME: non-speculating: flush on bidirectional mappings? */
 497 }
 498 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
 499
 500 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
 501         size_t size, enum dma_data_direction dir)
 502 {
 503         unsigned long paddr = page_to_phys(page) + off;
 504
 505         /* FIXME: non-speculating: not required */
 506         /* don't bother invalidating if DMA to device */
 507         if (dir != DMA_TO_DEVICE)
 508                 outer_inv_range(paddr, paddr + size);
 509
 510         dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
 511 }
 512 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
 513
 514 /**
 515  * dma_map_sg - map a set of SG buffers for streaming mode DMA
 516  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 517  * @sg: list of buffers
 518  * @nents: number of buffers to map
 519  * @dir: DMA transfer direction
 520  *
 521  * Map a set of buffers described by scatterlist in streaming mode for DMA.
 522  * This is the scatter-gather version of the dma_map_single interface.
 523  * Here the scatter gather list elements are each tagged with the
 524  * appropriate dma address and length.  They are obtained via
 525  * sg_dma_{address,length}.
 526  *
 527  * Device ownership issues as mentioned for dma_map_single are the same
 528  * here.
 529  */
 530 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 531                 enum dma_data_direction dir)
 532 {
 533         struct scatterlist *s;
 534         int i, j;
 535
 536         for_each_sg(sg, s, nents, i) {
 537                 s->dma_address = dma_map_page(dev, sg_page(s), s->offset,
 538                                                 s->length, dir);
 539                 if (dma_mapping_error(dev, s->dma_address))
 540                         goto bad_mapping;
 541         }
 542         return nents;
 543
 544  bad_mapping:
 545         for_each_sg(sg, s, i, j)
 546                 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
 547         return 0;
 548 }
 549 EXPORT_SYMBOL(dma_map_sg);
 550
 551 /**
 552  * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
 553  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 554  * @sg: list of buffers
 555  * @nents: number of buffers to unmap (returned from dma_map_sg)
 556  * @dir: DMA transfer direction (same as was passed to dma_map_sg)
 557  *
 558  * Unmap a set of streaming mode DMA translations.  Again, CPU access
 559  * rules concerning calls here are the same as for dma_unmap_single().
 560  */
 561 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 562                 enum dma_data_direction dir)
 563 {
 564         struct scatterlist *s;
 565         int i;
 566
 567         for_each_sg(sg, s, nents, i)
 568                 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
 569 }
 570 EXPORT_SYMBOL(dma_unmap_sg);
 571
 572 /**
 573  * dma_sync_sg_for_cpu
 574  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 575  * @sg: list of buffers
 576  * @nents: number of buffers to map (returned from dma_map_sg)
 577  * @dir: DMA transfer direction (same as was passed to dma_map_sg)
 578  */
 579 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 580                         int nents, enum dma_data_direction dir)
 581 {
 582         struct scatterlist *s;
 583         int i;
 584
 585         for_each_sg(sg, s, nents, i) {
 586                 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
 587                                             sg_dma_len(s), dir))
 588                         continue;
 589
 590                 __dma_page_dev_to_cpu(sg_page(s), s->offset,
 591                                       s->length, dir);
 592         }
 593 }
 594 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
 595
 596 /**
 597  * dma_sync_sg_for_device
 598  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 599  * @sg: list of buffers
 600  * @nents: number of buffers to map (returned from dma_map_sg)
 601  * @dir: DMA transfer direction (same as was passed to dma_map_sg)
 602  */
 603 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 604                         int nents, enum dma_data_direction dir)
 605 {
 606         struct scatterlist *s;
 607         int i;
 608
 609         for_each_sg(sg, s, nents, i) {
 610                 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
 611                                         sg_dma_len(s), dir))
 612                         continue;
 613
 614                 __dma_page_cpu_to_dev(sg_page(s), s->offset,
 615                                       s->length, dir);
 616         }
 617 }
 618 EXPORT_SYMBOL(dma_sync_sg_for_device);