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 static u64 get_coherent_dma_mask(struct device *dev)
29 u64 mask = ISA_DMA_THRESHOLD;
32 mask = dev->coherent_dma_mask;
35 * Sanity check the DMA mask - it must be non-zero, and
36 * must be able to be satisfied by a DMA allocation.
39 dev_warn(dev, "coherent DMA mask is unset\n");
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);
55 * Allocate a DMA buffer for 'dev' of size 'size' using the
56 * specified gfp mask. Note that 'size' must be page aligned.
58 static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
60 unsigned long order = get_order(size);
61 struct page *page, *p, *e;
63 u64 mask = get_coherent_dma_mask(dev);
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",
77 if (mask < 0xffffffffULL)
80 page = alloc_pages(gfp, order);
85 * Now split the huge page and free the excess pages
87 split_page(page, order);
88 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
92 * Ensure that the allocated pages are zeroed, and that any data
93 * lurking in the kernel direct-mapped region is invalidated.
95 ptr = page_address(page);
97 dmac_flush_range(ptr, ptr + size);
98 outer_flush_range(__pa(ptr), __pa(ptr) + size);
104 * Free a DMA buffer. 'size' must be page aligned.
106 static void __dma_free_buffer(struct page *page, size_t size)
108 struct page *e = page + (size >> PAGE_SHIFT);
117 /* Sanity check size */
118 #if (CONSISTENT_DMA_SIZE % SZ_2M)
119 #error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
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)
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;
189 if (!consistent_pte[0]) {
190 printk(KERN_ERR "%s: not initialised\n", __func__);
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.
201 bit = fls(size - 1) + 1;
202 if (bit > SECTION_SHIFT)
207 * Allocate a virtual address in the consistent mapping region.
209 c = arm_vmregion_alloc(&consistent_head, align, size,
210 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
213 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
214 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
216 pte = consistent_pte[idx] + off;
220 BUG_ON(!pte_none(*pte));
222 set_pte_ext(pte, mk_pte(page, prot), 0);
226 if (off >= PTRS_PER_PTE) {
228 pte = consistent_pte[++idx];
230 } while (size -= PAGE_SIZE);
232 return (void *)c->vm_start;
237 static void __dma_free_remap(void *cpu_addr, size_t size)
239 struct arm_vmregion *c;
245 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
247 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
253 if ((c->vm_end - c->vm_start) != size) {
254 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
255 __func__, c->vm_end - c->vm_start, size);
257 size = c->vm_end - c->vm_start;
260 idx = CONSISTENT_PTE_INDEX(c->vm_start);
261 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
262 ptep = consistent_pte[idx] + off;
265 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
270 if (off >= PTRS_PER_PTE) {
272 ptep = consistent_pte[++idx];
275 if (pte_none(pte) || !pte_present(pte))
276 printk(KERN_CRIT "%s: bad page in kernel page table\n",
278 } while (size -= PAGE_SIZE);
280 flush_tlb_kernel_range(c->vm_start, c->vm_end);
282 arm_vmregion_free(&consistent_head, c);
285 #else /* !CONFIG_MMU */
287 #define __dma_alloc_remap(page, size, gfp, prot) page_address(page)
288 #define __dma_free_remap(addr, size) do { } while (0)
290 #endif /* CONFIG_MMU */
293 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
300 size = PAGE_ALIGN(size);
302 page = __dma_alloc_buffer(dev, size, gfp);
306 if (!arch_is_coherent())
307 addr = __dma_alloc_remap(page, size, gfp, prot);
309 addr = page_address(page);
312 *handle = page_to_dma(dev, page);
318 * Allocate DMA-coherent memory space and return both the kernel remapped
319 * virtual and bus address for that space.
322 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
326 if (dma_alloc_from_coherent(dev, size, handle, &memory))
329 return __dma_alloc(dev, size, handle, gfp,
330 pgprot_dmacoherent(pgprot_kernel));
332 EXPORT_SYMBOL(dma_alloc_coherent);
335 * Allocate a writecombining region, in much the same way as
336 * dma_alloc_coherent above.
339 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
341 return __dma_alloc(dev, size, handle, gfp,
342 pgprot_writecombine(pgprot_kernel));
344 EXPORT_SYMBOL(dma_alloc_writecombine);
346 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
347 void *cpu_addr, dma_addr_t dma_addr, size_t size)
351 unsigned long user_size, kern_size;
352 struct arm_vmregion *c;
354 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
356 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
358 unsigned long off = vma->vm_pgoff;
360 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
362 if (off < kern_size &&
363 user_size <= (kern_size - off)) {
364 ret = remap_pfn_range(vma, vma->vm_start,
365 page_to_pfn(c->vm_pages) + off,
366 user_size << PAGE_SHIFT,
370 #endif /* CONFIG_MMU */
375 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
376 void *cpu_addr, dma_addr_t dma_addr, size_t size)
378 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
379 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
381 EXPORT_SYMBOL(dma_mmap_coherent);
383 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
384 void *cpu_addr, dma_addr_t dma_addr, size_t size)
386 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
387 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
389 EXPORT_SYMBOL(dma_mmap_writecombine);
392 * free a page as defined by the above mapping.
393 * Must not be called with IRQs disabled.
395 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
397 WARN_ON(irqs_disabled());
399 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
402 size = PAGE_ALIGN(size);
404 if (!arch_is_coherent())
405 __dma_free_remap(cpu_addr, size);
407 __dma_free_buffer(dma_to_page(dev, handle), size);
409 EXPORT_SYMBOL(dma_free_coherent);
412 * Make an area consistent for devices.
413 * Note: Drivers should NOT use this function directly, as it will break
414 * platforms with CONFIG_DMABOUNCE.
415 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
417 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
418 enum dma_data_direction dir)
422 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
424 dmac_map_area(kaddr, size, dir);
427 if (dir == DMA_FROM_DEVICE) {
428 outer_inv_range(paddr, paddr + size);
430 outer_clean_range(paddr, paddr + size);
432 /* FIXME: non-speculating: flush on bidirectional mappings? */
434 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
436 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
437 enum dma_data_direction dir)
439 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
441 /* FIXME: non-speculating: not required */
442 /* don't bother invalidating if DMA to device */
443 if (dir != DMA_TO_DEVICE) {
444 unsigned long paddr = __pa(kaddr);
445 outer_inv_range(paddr, paddr + size);
448 dmac_unmap_area(kaddr, size, dir);
450 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
452 static void dma_cache_maint_page(struct page *page, unsigned long offset,
453 size_t size, enum dma_data_direction dir,
454 void (*op)(const void *, size_t, int))
457 * A single sg entry may refer to multiple physically contiguous
458 * pages. But we still need to process highmem pages individually.
459 * If highmem is not configured then the bulk of this loop gets
467 if (PageHighMem(page)) {
468 if (len + offset > PAGE_SIZE) {
469 if (offset >= PAGE_SIZE) {
470 page += offset / PAGE_SIZE;
473 len = PAGE_SIZE - offset;
475 vaddr = kmap_high_get(page);
480 } else if (cache_is_vipt()) {
482 vaddr = kmap_high_l1_vipt(page, &saved_pte);
483 op(vaddr + offset, len, dir);
484 kunmap_high_l1_vipt(page, saved_pte);
487 vaddr = page_address(page) + offset;
496 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
497 size_t size, enum dma_data_direction dir)
501 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
503 paddr = page_to_phys(page) + off;
504 if (dir == DMA_FROM_DEVICE) {
505 outer_inv_range(paddr, paddr + size);
507 outer_clean_range(paddr, paddr + size);
509 /* FIXME: non-speculating: flush on bidirectional mappings? */
511 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
513 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
514 size_t size, enum dma_data_direction dir)
516 unsigned long paddr = page_to_phys(page) + off;
518 /* FIXME: non-speculating: not required */
519 /* don't bother invalidating if DMA to device */
520 if (dir != DMA_TO_DEVICE)
521 outer_inv_range(paddr, paddr + size);
523 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
525 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
528 * dma_map_sg - map a set of SG buffers for streaming mode DMA
529 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
530 * @sg: list of buffers
531 * @nents: number of buffers to map
532 * @dir: DMA transfer direction
534 * Map a set of buffers described by scatterlist in streaming mode for DMA.
535 * This is the scatter-gather version of the dma_map_single interface.
536 * Here the scatter gather list elements are each tagged with the
537 * appropriate dma address and length. They are obtained via
538 * sg_dma_{address,length}.
540 * Device ownership issues as mentioned for dma_map_single are the same
543 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
544 enum dma_data_direction dir)
546 struct scatterlist *s;
549 for_each_sg(sg, s, nents, i) {
550 s->dma_address = dma_map_page(dev, sg_page(s), s->offset,
552 if (dma_mapping_error(dev, s->dma_address))
558 for_each_sg(sg, s, i, j)
559 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
562 EXPORT_SYMBOL(dma_map_sg);
565 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
566 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
567 * @sg: list of buffers
568 * @nents: number of buffers to unmap (returned from dma_map_sg)
569 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
571 * Unmap a set of streaming mode DMA translations. Again, CPU access
572 * rules concerning calls here are the same as for dma_unmap_single().
574 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
575 enum dma_data_direction dir)
577 struct scatterlist *s;
580 for_each_sg(sg, s, nents, i)
581 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
583 EXPORT_SYMBOL(dma_unmap_sg);
586 * dma_sync_sg_for_cpu
587 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
588 * @sg: list of buffers
589 * @nents: number of buffers to map (returned from dma_map_sg)
590 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
592 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
593 int nents, enum dma_data_direction dir)
595 struct scatterlist *s;
598 for_each_sg(sg, s, nents, i) {
599 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
603 __dma_page_dev_to_cpu(sg_page(s), s->offset,
607 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
610 * dma_sync_sg_for_device
611 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
612 * @sg: list of buffers
613 * @nents: number of buffers to map (returned from dma_map_sg)
614 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
616 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
617 int nents, enum dma_data_direction dir)
619 struct scatterlist *s;
622 for_each_sg(sg, s, nents, i) {
623 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
627 __dma_page_cpu_to_dev(sg_page(s), s->offset,
631 EXPORT_SYMBOL(dma_sync_sg_for_device);