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 static u64 get_coherent_dma_mask(struct device *dev)
  28 {
  29         u64 mask = ISA_DMA_THRESHOLD;
  30
  31         if (dev) {
  32                 mask = dev->coherent_dma_mask;
  33
  34                 /*
  35                  * Sanity check the DMA mask - it must be non-zero, and
  36                  * must be able to be satisfied by a DMA allocation.
  37                  */
  38                 if (mask == 0) {
  39                         dev_warn(dev, "coherent DMA mask is unset\n");
  40                         return 0;
  41                 }
  42
  43                 if ((~mask) & ISA_DMA_THRESHOLD) {
  44                         dev_warn(dev, "coherent DMA mask %#llx is smaller "
  45                                  "than system GFP_DMA mask %#llx\n",
  46                                  mask, (unsigned long long)ISA_DMA_THRESHOLD);
  47                         return 0;
  48                 }
  49         }
  50
  51         return mask;
  52 }
  53
  54 /*
  55  * Allocate a DMA buffer for 'dev' of size 'size' using the
  56  * specified gfp mask.  Note that 'size' must be page aligned.
  57  */
  58 static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
  59 {
  60         unsigned long order = get_order(size);
  61         struct page *page, *p, *e;
  62         void *ptr;
  63         u64 mask = get_coherent_dma_mask(dev);
  64
  65 #ifdef CONFIG_DMA_API_DEBUG
  66         u64 limit = (mask + 1) & ~mask;
  67         if (limit && size >= limit) {
  68                 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
  69                         size, mask);
  70                 return NULL;
  71         }
  72 #endif
  73
  74         if (!mask)
  75                 return NULL;
  76
  77         if (mask < 0xffffffffULL)
  78                 gfp |= GFP_DMA;
  79
  80         page = alloc_pages(gfp, order);
  81         if (!page)
  82                 return NULL;
  83
  84         /*
  85          * Now split the huge page and free the excess pages
  86          */
  87         split_page(page, order);
  88         for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
  89                 __free_page(p);
  90
  91         /*
  92          * Ensure that the allocated pages are zeroed, and that any data
  93          * lurking in the kernel direct-mapped region is invalidated.
  94          */
  95         ptr = page_address(page);
  96         memset(ptr, 0, size);
  97         dmac_flush_range(ptr, ptr + size);
  98         outer_flush_range(__pa(ptr), __pa(ptr) + size);
  99
 100         return page;
 101 }
 102
 103 /*
 104  * Free a DMA buffer.  'size' must be page aligned.
 105  */
 106 static void __dma_free_buffer(struct page *page, size_t size)
 107 {
 108         struct page *e = page + (size >> PAGE_SHIFT);
 109
 110         while (page < e) {
 111                 __free_page(page);
 112                 page++;
 113         }
 114 }
 115
 116 #ifdef CONFIG_MMU
 117 /* Sanity check size */
 118 #if (CONSISTENT_DMA_SIZE % SZ_2M)
 119 #error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
 120 #endif
 121
 122 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
 123 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PGDIR_SHIFT)
 124 #define NUM_CONSISTENT_PTES (CONSISTENT_DMA_SIZE >> PGDIR_SHIFT)
 125
 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         size_t align;
 187         int bit;
 188
 189         if (!consistent_pte[0]) {
 190                 printk(KERN_ERR "%s: not initialised\n", __func__);
 191                 dump_stack();
 192                 return NULL;
 193         }
 194
 195         /*
 196          * Align the virtual region allocation - maximum alignment is
 197          * a section size, minimum is a page size.  This helps reduce
 198          * fragmentation of the DMA space, and also prevents allocations
 199          * smaller than a section from crossing a section boundary.
 200          */
 201         bit = fls(size - 1);
 202         if (bit > SECTION_SHIFT)
 203                 bit = SECTION_SHIFT;
 204         align = 1 << bit;
 205
 206         /*
 207          * Allocate a virtual address in the consistent mapping region.
 208          */
 209         c = arm_vmregion_alloc(&consistent_head, align, size,
 210                             gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
 211         if (c) {
 212                 pte_t *pte;
 213                 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
 214                 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
 215
 216                 pte = consistent_pte[idx] + off;
 217                 c->vm_pages = page;
 218
 219                 do {
 220                         BUG_ON(!pte_none(*pte));
 221
 222                         set_pte_ext(pte, mk_pte(page, prot), 0);
 223                         page++;
 224                         pte++;
 225                         off++;
 226                         if (off >= PTRS_PER_PTE) {
 227                                 off = 0;
 228                                 pte = consistent_pte[++idx];
 229                         }
 230                 } while (size -= PAGE_SIZE);
 231
 232                 dsb();
 233
 234                 return (void *)c->vm_start;
 235         }
 236         return NULL;
 237 }
 238
 239 static void __dma_free_remap(void *cpu_addr, size_t size)
 240 {
 241         struct arm_vmregion *c;
 242         unsigned long addr;
 243         pte_t *ptep;
 244         int idx;
 245         u32 off;
 246
 247         c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
 248         if (!c) {
 249                 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
 250                        __func__, cpu_addr);
 251                 dump_stack();
 252                 return;
 253         }
 254
 255         if ((c->vm_end - c->vm_start) != size) {
 256                 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
 257                        __func__, c->vm_end - c->vm_start, size);
 258                 dump_stack();
 259                 size = c->vm_end - c->vm_start;
 260         }
 261
 262         idx = CONSISTENT_PTE_INDEX(c->vm_start);
 263         off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
 264         ptep = consistent_pte[idx] + off;
 265         addr = c->vm_start;
 266         do {
 267                 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
 268
 269                 ptep++;
 270                 addr += PAGE_SIZE;
 271                 off++;
 272                 if (off >= PTRS_PER_PTE) {
 273                         off = 0;
 274                         ptep = consistent_pte[++idx];
 275                 }
 276
 277                 if (pte_none(pte) || !pte_present(pte))
 278                         printk(KERN_CRIT "%s: bad page in kernel page table\n",
 279                                __func__);
 280         } while (size -= PAGE_SIZE);
 281
 282         flush_tlb_kernel_range(c->vm_start, c->vm_end);
 283
 284         arm_vmregion_free(&consistent_head, c);
 285 }
 286
 287 #else   /* !CONFIG_MMU */
 288
 289 #define __dma_alloc_remap(page, size, gfp, prot)        page_address(page)
 290 #define __dma_free_remap(addr, size)                    do { } while (0)
 291
 292 #endif  /* CONFIG_MMU */
 293
 294 static void *
 295 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
 296             pgprot_t prot)
 297 {
 298         struct page *page;
 299         void *addr;
 300
 301         *handle = ~0;
 302         size = PAGE_ALIGN(size);
 303
 304         page = __dma_alloc_buffer(dev, size, gfp);
 305         if (!page)
 306                 return NULL;
 307
 308         if (!arch_is_coherent())
 309                 addr = __dma_alloc_remap(page, size, gfp, prot);
 310         else
 311                 addr = page_address(page);
 312
 313         if (addr)
 314                 *handle = page_to_dma(dev, page);
 315
 316         return addr;
 317 }
 318
 319 /*
 320  * Allocate DMA-coherent memory space and return both the kernel remapped
 321  * virtual and bus address for that space.
 322  */
 323 void *
 324 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
 325 {
 326         void *memory;
 327
 328         if (dma_alloc_from_coherent(dev, size, handle, &memory))
 329                 return memory;
 330
 331         return __dma_alloc(dev, size, handle, gfp,
 332                            pgprot_dmacoherent(pgprot_kernel));
 333 }
 334 EXPORT_SYMBOL(dma_alloc_coherent);
 335
 336 /*
 337  * Allocate a writecombining region, in much the same way as
 338  * dma_alloc_coherent above.
 339  */
 340 void *
 341 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
 342 {
 343         return __dma_alloc(dev, size, handle, gfp,
 344                            pgprot_writecombine(pgprot_kernel));
 345 }
 346 EXPORT_SYMBOL(dma_alloc_writecombine);
 347
 348 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
 349                     void *cpu_addr, dma_addr_t dma_addr, size_t size)
 350 {
 351         int ret = -ENXIO;
 352 #ifdef CONFIG_MMU
 353         unsigned long user_size, kern_size;
 354         struct arm_vmregion *c;
 355
 356         user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 357
 358         c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
 359         if (c) {
 360                 unsigned long off = vma->vm_pgoff;
 361
 362                 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
 363
 364                 if (off < kern_size &&
 365                     user_size <= (kern_size - off)) {
 366                         ret = remap_pfn_range(vma, vma->vm_start,
 367                                               page_to_pfn(c->vm_pages) + off,
 368                                               user_size << PAGE_SHIFT,
 369                                               vma->vm_page_prot);
 370                 }
 371         }
 372 #endif  /* CONFIG_MMU */
 373
 374         return ret;
 375 }
 376
 377 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
 378                       void *cpu_addr, dma_addr_t dma_addr, size_t size)
 379 {
 380         vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
 381         return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
 382 }
 383 EXPORT_SYMBOL(dma_mmap_coherent);
 384
 385 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
 386                           void *cpu_addr, dma_addr_t dma_addr, size_t size)
 387 {
 388         vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 389         return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
 390 }
 391 EXPORT_SYMBOL(dma_mmap_writecombine);
 392
 393 /*
 394  * free a page as defined by the above mapping.
 395  * Must not be called with IRQs disabled.
 396  */
 397 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
 398 {
 399         WARN_ON(irqs_disabled());
 400
 401         if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
 402                 return;
 403
 404         size = PAGE_ALIGN(size);
 405
 406         if (!arch_is_coherent())
 407                 __dma_free_remap(cpu_addr, size);
 408
 409         __dma_free_buffer(dma_to_page(dev, handle), size);
 410 }
 411 EXPORT_SYMBOL(dma_free_coherent);
 412
 413 /*
 414  * Make an area consistent for devices.
 415  * Note: Drivers should NOT use this function directly, as it will break
 416  * platforms with CONFIG_DMABOUNCE.
 417  * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
 418  */
 419 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
 420         enum dma_data_direction dir)
 421 {
 422         unsigned long paddr;
 423
 424         BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
 425
 426         dmac_map_area(kaddr, size, dir);
 427
 428         paddr = __pa(kaddr);
 429         if (dir == DMA_FROM_DEVICE) {
 430                 outer_inv_range(paddr, paddr + size);
 431         } else {
 432                 outer_clean_range(paddr, paddr + size);
 433         }
 434         /* FIXME: non-speculating: flush on bidirectional mappings? */
 435 }
 436 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
 437
 438 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
 439         enum dma_data_direction dir)
 440 {
 441         BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
 442
 443         /* FIXME: non-speculating: not required */
 444         /* don't bother invalidating if DMA to device */
 445         if (dir != DMA_TO_DEVICE) {
 446                 unsigned long paddr = __pa(kaddr);
 447                 outer_inv_range(paddr, paddr + size);
 448         }
 449
 450         dmac_unmap_area(kaddr, size, dir);
 451 }
 452 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
 453
 454 static void dma_cache_maint_page(struct page *page, unsigned long offset,
 455         size_t size, enum dma_data_direction dir,
 456         void (*op)(const void *, size_t, int))
 457 {
 458         /*
 459          * A single sg entry may refer to multiple physically contiguous
 460          * pages.  But we still need to process highmem pages individually.
 461          * If highmem is not configured then the bulk of this loop gets
 462          * optimized out.
 463          */
 464         size_t left = size;
 465         do {
 466                 size_t len = left;
 467                 void *vaddr;
 468
 469                 if (PageHighMem(page)) {
 470                         if (len + offset > PAGE_SIZE) {
 471                                 if (offset >= PAGE_SIZE) {
 472                                         page += offset / PAGE_SIZE;
 473                                         offset %= PAGE_SIZE;
 474                                 }
 475                                 len = PAGE_SIZE - offset;
 476                         }
 477                         vaddr = kmap_high_get(page);
 478                         if (vaddr) {
 479                                 vaddr += offset;
 480                                 op(vaddr, len, dir);
 481                                 kunmap_high(page);
 482                         } else if (cache_is_vipt()) {
 483                                 pte_t saved_pte;
 484                                 vaddr = kmap_high_l1_vipt(page, &saved_pte);
 485                                 op(vaddr + offset, len, dir);
 486                                 kunmap_high_l1_vipt(page, saved_pte);
 487                         }
 488                 } else {
 489                         vaddr = page_address(page) + offset;
 490                         op(vaddr, len, dir);
 491                 }
 492                 offset = 0;
 493                 page++;
 494                 left -= len;
 495         } while (left);
 496 }
 497
 498 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
 499         size_t size, enum dma_data_direction dir)
 500 {
 501         unsigned long paddr;
 502
 503         dma_cache_maint_page(page, off, size, dir, dmac_map_area);
 504
 505         paddr = page_to_phys(page) + off;
 506         if (dir == DMA_FROM_DEVICE) {
 507                 outer_inv_range(paddr, paddr + size);
 508         } else {
 509                 outer_clean_range(paddr, paddr + size);
 510         }
 511         /* FIXME: non-speculating: flush on bidirectional mappings? */
 512 }
 513 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
 514
 515 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
 516         size_t size, enum dma_data_direction dir)
 517 {
 518         unsigned long paddr = page_to_phys(page) + off;
 519
 520         /* FIXME: non-speculating: not required */
 521         /* don't bother invalidating if DMA to device */
 522         if (dir != DMA_TO_DEVICE)
 523                 outer_inv_range(paddr, paddr + size);
 524
 525         dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
 526
 527         /*
 528          * Mark the D-cache clean for this page to avoid extra flushing.
 529          */
 530         if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
 531                 set_bit(PG_dcache_clean, &page->flags);
 532 }
 533 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
 534
 535 /**
 536  * dma_map_sg - map a set of SG buffers for streaming mode DMA
 537  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 538  * @sg: list of buffers
 539  * @nents: number of buffers to map
 540  * @dir: DMA transfer direction
 541  *
 542  * Map a set of buffers described by scatterlist in streaming mode for DMA.
 543  * This is the scatter-gather version of the dma_map_single interface.
 544  * Here the scatter gather list elements are each tagged with the
 545  * appropriate dma address and length.  They are obtained via
 546  * sg_dma_{address,length}.
 547  *
 548  * Device ownership issues as mentioned for dma_map_single are the same
 549  * here.
 550  */
 551 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
 552                 enum dma_data_direction dir)
 553 {
 554         struct scatterlist *s;
 555         int i, j;
 556
 557         for_each_sg(sg, s, nents, i) {
 558                 s->dma_address = dma_map_page(dev, sg_page(s), s->offset,
 559                                                 s->length, dir);
 560                 if (dma_mapping_error(dev, s->dma_address))
 561                         goto bad_mapping;
 562         }
 563         return nents;
 564
 565  bad_mapping:
 566         for_each_sg(sg, s, i, j)
 567                 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
 568         return 0;
 569 }
 570 EXPORT_SYMBOL(dma_map_sg);
 571
 572 /**
 573  * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
 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 unmap (returned from dma_map_sg)
 577  * @dir: DMA transfer direction (same as was passed to dma_map_sg)
 578  *
 579  * Unmap a set of streaming mode DMA translations.  Again, CPU access
 580  * rules concerning calls here are the same as for dma_unmap_single().
 581  */
 582 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
 583                 enum dma_data_direction dir)
 584 {
 585         struct scatterlist *s;
 586         int i;
 587
 588         for_each_sg(sg, s, nents, i)
 589                 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
 590 }
 591 EXPORT_SYMBOL(dma_unmap_sg);
 592
 593 /**
 594  * dma_sync_sg_for_cpu
 595  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 596  * @sg: list of buffers
 597  * @nents: number of buffers to map (returned from dma_map_sg)
 598  * @dir: DMA transfer direction (same as was passed to dma_map_sg)
 599  */
 600 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
 601                         int nents, enum dma_data_direction dir)
 602 {
 603         struct scatterlist *s;
 604         int i;
 605
 606         for_each_sg(sg, s, nents, i) {
 607                 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
 608                                             sg_dma_len(s), dir))
 609                         continue;
 610
 611                 __dma_page_dev_to_cpu(sg_page(s), s->offset,
 612                                       s->length, dir);
 613         }
 614 }
 615 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
 616
 617 /**
 618  * dma_sync_sg_for_device
 619  * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
 620  * @sg: list of buffers
 621  * @nents: number of buffers to map (returned from dma_map_sg)
 622  * @dir: DMA transfer direction (same as was passed to dma_map_sg)
 623  */
 624 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
 625                         int nents, enum dma_data_direction dir)
 626 {
 627         struct scatterlist *s;
 628         int i;
 629
 630         for_each_sg(sg, s, nents, i) {
 631                 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
 632                                         sg_dma_len(s), dir))
 633                         continue;
 634
 635                 __dma_page_cpu_to_dev(sg_page(s), s->offset,
 636                                       s->length, dir);
 637         }
 638 }
 639 EXPORT_SYMBOL(dma_sync_sg_for_device);