2 * linux/arch/arm/mm/dma-mapping.c
4 * Copyright (C) 2000-2004 Russell King
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.
10 * DMA uncached mapping support.
12 #include <linux/module.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>
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>
27 /* Sanity check size */
28 #if (CONSISTENT_DMA_SIZE % SZ_2M)
29 #error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
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)
36 static u64 get_coherent_dma_mask(struct device *dev)
38 u64 mask = ISA_DMA_THRESHOLD;
41 mask = dev->coherent_dma_mask;
44 * Sanity check the DMA mask - it must be non-zero, and
45 * must be able to be satisfied by a DMA allocation.
48 dev_warn(dev, "coherent DMA mask is unset\n");
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);
64 * Allocate a DMA buffer for 'dev' of size 'size' using the
65 * specified gfp mask. Note that 'size' must be page aligned.
67 static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
69 unsigned long order = get_order(size);
70 struct page *page, *p, *e;
72 u64 mask = get_coherent_dma_mask(dev);
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",
86 if (mask < 0xffffffffULL)
89 page = alloc_pages(gfp, order);
94 * Now split the huge page and free the excess pages
96 split_page(page, order);
97 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
101 * Ensure that the allocated pages are zeroed, and that any data
102 * lurking in the kernel direct-mapped region is invalidated.
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);
113 * Free a DMA buffer. 'size' must be page aligned.
115 static void __dma_free_buffer(struct page *page, size_t size)
117 struct page *e = page + (size >> PAGE_SHIFT);
127 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
129 static pte_t *consistent_pte[NUM_CONSISTENT_PTES];
131 #include "vmregion.h"
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,
140 #ifdef CONFIG_HUGETLB_PAGE
141 #error ARM Coherent DMA allocator does not (yet) support huge TLB
145 * Initialise the consistent memory allocation.
147 static int __init consistent_init(void)
154 u32 base = CONSISTENT_BASE;
157 pgd = pgd_offset(&init_mm, base);
158 pmd = pmd_alloc(&init_mm, pgd, base);
160 printk(KERN_ERR "%s: no pmd tables\n", __func__);
164 WARN_ON(!pmd_none(*pmd));
166 pte = pte_alloc_kernel(pmd, base);
168 printk(KERN_ERR "%s: no pte tables\n", __func__);
173 consistent_pte[i++] = pte;
174 base += (1 << PGDIR_SHIFT);
175 } while (base < CONSISTENT_END);
180 core_initcall(consistent_init);
183 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot)
185 struct arm_vmregion *c;
187 if (!consistent_pte[0]) {
188 printk(KERN_ERR "%s: not initialised\n", __func__);
194 * Allocate a virtual address in the consistent mapping region.
196 c = arm_vmregion_alloc(&consistent_head, size,
197 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
200 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
201 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
203 pte = consistent_pte[idx] + off;
207 BUG_ON(!pte_none(*pte));
209 set_pte_ext(pte, mk_pte(page, prot), 0);
213 if (off >= PTRS_PER_PTE) {
215 pte = consistent_pte[++idx];
217 } while (size -= PAGE_SIZE);
219 return (void *)c->vm_start;
224 static void __dma_free_remap(void *cpu_addr, size_t size)
226 struct arm_vmregion *c;
232 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
234 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
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);
244 size = c->vm_end - c->vm_start;
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;
252 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
257 if (off >= PTRS_PER_PTE) {
259 ptep = consistent_pte[++idx];
262 if (pte_none(pte) || !pte_present(pte))
263 printk(KERN_CRIT "%s: bad page in kernel page table\n",
265 } while (size -= PAGE_SIZE);
267 flush_tlb_kernel_range(c->vm_start, c->vm_end);
269 arm_vmregion_free(&consistent_head, c);
272 #else /* !CONFIG_MMU */
274 #define __dma_alloc_remap(page, size, gfp, prot) page_address(page)
275 #define __dma_free_remap(addr, size) do { } while (0)
277 #endif /* CONFIG_MMU */
280 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
287 size = PAGE_ALIGN(size);
289 page = __dma_alloc_buffer(dev, size, gfp);
293 if (!arch_is_coherent())
294 addr = __dma_alloc_remap(page, size, gfp, prot);
296 addr = page_address(page);
299 *handle = page_to_dma(dev, page);
305 * Allocate DMA-coherent memory space and return both the kernel remapped
306 * virtual and bus address for that space.
309 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
313 if (dma_alloc_from_coherent(dev, size, handle, &memory))
316 return __dma_alloc(dev, size, handle, gfp,
317 pgprot_dmacoherent(pgprot_kernel));
319 EXPORT_SYMBOL(dma_alloc_coherent);
322 * Allocate a writecombining region, in much the same way as
323 * dma_alloc_coherent above.
326 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
328 return __dma_alloc(dev, size, handle, gfp,
329 pgprot_writecombine(pgprot_kernel));
331 EXPORT_SYMBOL(dma_alloc_writecombine);
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)
338 unsigned long user_size, kern_size;
339 struct arm_vmregion *c;
341 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
343 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
345 unsigned long off = vma->vm_pgoff;
347 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
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,
357 #endif /* CONFIG_MMU */
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)
365 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
366 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
368 EXPORT_SYMBOL(dma_mmap_coherent);
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)
373 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
374 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
376 EXPORT_SYMBOL(dma_mmap_writecombine);
379 * free a page as defined by the above mapping.
380 * Must not be called with IRQs disabled.
382 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
384 WARN_ON(irqs_disabled());
386 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
389 size = PAGE_ALIGN(size);
391 if (!arch_is_coherent())
392 __dma_free_remap(cpu_addr, size);
394 __dma_free_buffer(dma_to_page(dev, handle), size);
396 EXPORT_SYMBOL(dma_free_coherent);
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_*)
404 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
405 enum dma_data_direction dir)
409 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
411 dmac_map_area(kaddr, size, dir);
414 if (dir == DMA_FROM_DEVICE) {
415 outer_inv_range(paddr, paddr + size);
417 outer_clean_range(paddr, paddr + size);
419 /* FIXME: non-speculating: flush on bidirectional mappings? */
421 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
423 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
424 enum dma_data_direction dir)
426 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
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);
435 dmac_unmap_area(kaddr, size, dir);
437 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
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))
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
454 if (PageHighMem(page)) {
455 if (len + offset > PAGE_SIZE) {
456 if (offset >= PAGE_SIZE) {
457 page += offset / PAGE_SIZE;
460 len = PAGE_SIZE - offset;
462 vaddr = kmap_high_get(page);
469 vaddr = page_address(page) + offset;
478 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
479 size_t size, enum dma_data_direction dir)
483 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
485 paddr = page_to_phys(page) + off;
486 if (dir == DMA_FROM_DEVICE) {
487 outer_inv_range(paddr, paddr + size);
489 outer_clean_range(paddr, paddr + size);
491 /* FIXME: non-speculating: flush on bidirectional mappings? */
493 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
495 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
496 size_t size, enum dma_data_direction dir)
498 unsigned long paddr = page_to_phys(page) + off;
500 /* FIXME: non-speculating: not required */
501 /* don't bother invalidating if DMA to device */
502 if (dir != DMA_TO_DEVICE)
503 outer_inv_range(paddr, paddr + size);
505 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
507 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
510 * dma_map_sg - map a set of SG buffers for streaming mode DMA
511 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
512 * @sg: list of buffers
513 * @nents: number of buffers to map
514 * @dir: DMA transfer direction
516 * Map a set of buffers described by scatterlist in streaming mode for DMA.
517 * This is the scatter-gather version of the dma_map_single interface.
518 * Here the scatter gather list elements are each tagged with the
519 * appropriate dma address and length. They are obtained via
520 * sg_dma_{address,length}.
522 * Device ownership issues as mentioned for dma_map_single are the same
525 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
526 enum dma_data_direction dir)
528 struct scatterlist *s;
531 for_each_sg(sg, s, nents, i) {
532 s->dma_address = dma_map_page(dev, sg_page(s), s->offset,
534 if (dma_mapping_error(dev, s->dma_address))
540 for_each_sg(sg, s, i, j)
541 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
544 EXPORT_SYMBOL(dma_map_sg);
547 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
548 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
549 * @sg: list of buffers
550 * @nents: number of buffers to unmap (returned from dma_map_sg)
551 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
553 * Unmap a set of streaming mode DMA translations. Again, CPU access
554 * rules concerning calls here are the same as for dma_unmap_single().
556 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
557 enum dma_data_direction dir)
559 struct scatterlist *s;
562 for_each_sg(sg, s, nents, i)
563 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
565 EXPORT_SYMBOL(dma_unmap_sg);
568 * dma_sync_sg_for_cpu
569 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
570 * @sg: list of buffers
571 * @nents: number of buffers to map (returned from dma_map_sg)
572 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
574 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
575 int nents, enum dma_data_direction dir)
577 struct scatterlist *s;
580 for_each_sg(sg, s, nents, i) {
581 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
585 __dma_page_dev_to_cpu(sg_page(s), s->offset,
589 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
592 * dma_sync_sg_for_device
593 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
594 * @sg: list of buffers
595 * @nents: number of buffers to map (returned from dma_map_sg)
596 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
598 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
599 int nents, enum dma_data_direction dir)
601 struct scatterlist *s;
604 for_each_sg(sg, s, nents, i) {
605 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
609 __dma_page_cpu_to_dev(sg_page(s), s->offset,
613 EXPORT_SYMBOL(dma_sync_sg_for_device);