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>
20 #include <linux/highmem.h>
21 #include <linux/slab.h>
23 #include <asm/memory.h>
24 #include <asm/highmem.h>
25 #include <asm/cacheflush.h>
26 #include <asm/tlbflush.h>
27 #include <asm/sizes.h>
28 #include <asm/mach/arch.h>
32 static u64 get_coherent_dma_mask(struct device *dev)
34 u64 mask = (u64)arm_dma_limit;
37 mask = dev->coherent_dma_mask;
40 * Sanity check the DMA mask - it must be non-zero, and
41 * must be able to be satisfied by a DMA allocation.
44 dev_warn(dev, "coherent DMA mask is unset\n");
48 if ((~mask) & (u64)arm_dma_limit) {
49 dev_warn(dev, "coherent DMA mask %#llx is smaller "
50 "than system GFP_DMA mask %#llx\n",
51 mask, (u64)arm_dma_limit);
60 * Allocate a DMA buffer for 'dev' of size 'size' using the
61 * specified gfp mask. Note that 'size' must be page aligned.
63 static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
65 unsigned long order = get_order(size);
66 struct page *page, *p, *e;
68 u64 mask = get_coherent_dma_mask(dev);
70 #ifdef CONFIG_DMA_API_DEBUG
71 u64 limit = (mask + 1) & ~mask;
72 if (limit && size >= limit) {
73 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
82 if (mask < 0xffffffffULL)
85 page = alloc_pages(gfp, order);
90 * Now split the huge page and free the excess pages
92 split_page(page, order);
93 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
97 * Ensure that the allocated pages are zeroed, and that any data
98 * lurking in the kernel direct-mapped region is invalidated.
100 ptr = page_address(page);
101 memset(ptr, 0, size);
102 dmac_flush_range(ptr, ptr + size);
103 outer_flush_range(__pa(ptr), __pa(ptr) + size);
109 * Free a DMA buffer. 'size' must be page aligned.
111 static void __dma_free_buffer(struct page *page, size_t size)
113 struct page *e = page + (size >> PAGE_SHIFT);
124 #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - consistent_base) >> PAGE_SHIFT)
125 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - consistent_base) >> PGDIR_SHIFT)
128 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
130 static pte_t **consistent_pte;
132 #define DEFAULT_CONSISTENT_DMA_SIZE SZ_2M
134 unsigned long consistent_base = CONSISTENT_END - DEFAULT_CONSISTENT_DMA_SIZE;
136 void __init init_consistent_dma_size(unsigned long size)
138 unsigned long base = CONSISTENT_END - ALIGN(size, SZ_2M);
140 BUG_ON(consistent_pte); /* Check we're called before DMA region init */
141 BUG_ON(base < VMALLOC_END);
143 /* Grow region to accommodate specified size */
144 if (base < consistent_base)
145 consistent_base = base;
148 #include "vmregion.h"
150 static struct arm_vmregion_head consistent_head = {
151 .vm_lock = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock),
152 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
153 .vm_end = CONSISTENT_END,
156 #ifdef CONFIG_HUGETLB_PAGE
157 #error ARM Coherent DMA allocator does not (yet) support huge TLB
161 * Initialise the consistent memory allocation.
163 static int __init consistent_init(void)
171 unsigned long base = consistent_base;
172 unsigned long num_ptes = (CONSISTENT_END - base) >> PGDIR_SHIFT;
174 consistent_pte = kmalloc(num_ptes * sizeof(pte_t), GFP_KERNEL);
175 if (!consistent_pte) {
176 pr_err("%s: no memory\n", __func__);
180 pr_debug("DMA memory: 0x%08lx - 0x%08lx:\n", base, CONSISTENT_END);
181 consistent_head.vm_start = base;
184 pgd = pgd_offset(&init_mm, base);
186 pud = pud_alloc(&init_mm, pgd, base);
188 printk(KERN_ERR "%s: no pud tables\n", __func__);
193 pmd = pmd_alloc(&init_mm, pud, base);
195 printk(KERN_ERR "%s: no pmd tables\n", __func__);
199 WARN_ON(!pmd_none(*pmd));
201 pte = pte_alloc_kernel(pmd, base);
203 printk(KERN_ERR "%s: no pte tables\n", __func__);
208 consistent_pte[i++] = pte;
209 base += (1 << PGDIR_SHIFT);
210 } while (base < CONSISTENT_END);
215 core_initcall(consistent_init);
218 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot)
220 struct arm_vmregion *c;
224 if (!consistent_pte) {
225 printk(KERN_ERR "%s: not initialised\n", __func__);
231 * Align the virtual region allocation - maximum alignment is
232 * a section size, minimum is a page size. This helps reduce
233 * fragmentation of the DMA space, and also prevents allocations
234 * smaller than a section from crossing a section boundary.
237 if (bit > SECTION_SHIFT)
242 * Allocate a virtual address in the consistent mapping region.
244 c = arm_vmregion_alloc(&consistent_head, align, size,
245 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
248 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
249 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
251 pte = consistent_pte[idx] + off;
255 BUG_ON(!pte_none(*pte));
257 set_pte_ext(pte, mk_pte(page, prot), 0);
261 if (off >= PTRS_PER_PTE) {
263 pte = consistent_pte[++idx];
265 } while (size -= PAGE_SIZE);
269 return (void *)c->vm_start;
274 static void __dma_free_remap(void *cpu_addr, size_t size)
276 struct arm_vmregion *c;
282 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
284 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
290 if ((c->vm_end - c->vm_start) != size) {
291 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
292 __func__, c->vm_end - c->vm_start, size);
294 size = c->vm_end - c->vm_start;
297 idx = CONSISTENT_PTE_INDEX(c->vm_start);
298 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
299 ptep = consistent_pte[idx] + off;
302 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
307 if (off >= PTRS_PER_PTE) {
309 ptep = consistent_pte[++idx];
312 if (pte_none(pte) || !pte_present(pte))
313 printk(KERN_CRIT "%s: bad page in kernel page table\n",
315 } while (size -= PAGE_SIZE);
317 flush_tlb_kernel_range(c->vm_start, c->vm_end);
319 arm_vmregion_free(&consistent_head, c);
322 #else /* !CONFIG_MMU */
324 #define __dma_alloc_remap(page, size, gfp, prot) page_address(page)
325 #define __dma_free_remap(addr, size) do { } while (0)
327 #endif /* CONFIG_MMU */
330 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
337 size = PAGE_ALIGN(size);
339 page = __dma_alloc_buffer(dev, size, gfp);
343 if (!arch_is_coherent())
344 addr = __dma_alloc_remap(page, size, gfp, prot);
346 addr = page_address(page);
349 *handle = pfn_to_dma(dev, page_to_pfn(page));
355 * Allocate DMA-coherent memory space and return both the kernel remapped
356 * virtual and bus address for that space.
359 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
363 if (dma_alloc_from_coherent(dev, size, handle, &memory))
366 return __dma_alloc(dev, size, handle, gfp,
367 pgprot_dmacoherent(pgprot_kernel));
369 EXPORT_SYMBOL(dma_alloc_coherent);
372 * Allocate a writecombining region, in much the same way as
373 * dma_alloc_coherent above.
376 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
378 return __dma_alloc(dev, size, handle, gfp,
379 pgprot_writecombine(pgprot_kernel));
381 EXPORT_SYMBOL(dma_alloc_writecombine);
383 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
384 void *cpu_addr, dma_addr_t dma_addr, size_t size)
388 unsigned long user_size, kern_size;
389 struct arm_vmregion *c;
391 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
393 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
395 unsigned long off = vma->vm_pgoff;
397 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
399 if (off < kern_size &&
400 user_size <= (kern_size - off)) {
401 ret = remap_pfn_range(vma, vma->vm_start,
402 page_to_pfn(c->vm_pages) + off,
403 user_size << PAGE_SHIFT,
407 #endif /* CONFIG_MMU */
412 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
413 void *cpu_addr, dma_addr_t dma_addr, size_t size)
415 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
416 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
418 EXPORT_SYMBOL(dma_mmap_coherent);
420 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
421 void *cpu_addr, dma_addr_t dma_addr, size_t size)
423 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
424 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
426 EXPORT_SYMBOL(dma_mmap_writecombine);
429 * free a page as defined by the above mapping.
430 * Must not be called with IRQs disabled.
432 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
434 WARN_ON(irqs_disabled());
436 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
439 size = PAGE_ALIGN(size);
441 if (!arch_is_coherent())
442 __dma_free_remap(cpu_addr, size);
444 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev, handle)), size);
446 EXPORT_SYMBOL(dma_free_coherent);
449 * Make an area consistent for devices.
450 * Note: Drivers should NOT use this function directly, as it will break
451 * platforms with CONFIG_DMABOUNCE.
452 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
454 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
455 enum dma_data_direction dir)
459 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
461 dmac_map_area(kaddr, size, dir);
464 if (dir == DMA_FROM_DEVICE) {
465 outer_inv_range(paddr, paddr + size);
467 outer_clean_range(paddr, paddr + size);
469 /* FIXME: non-speculating: flush on bidirectional mappings? */
471 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
473 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
474 enum dma_data_direction dir)
476 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
478 /* FIXME: non-speculating: not required */
479 /* don't bother invalidating if DMA to device */
480 if (dir != DMA_TO_DEVICE) {
481 unsigned long paddr = __pa(kaddr);
482 outer_inv_range(paddr, paddr + size);
485 dmac_unmap_area(kaddr, size, dir);
487 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
489 static void dma_cache_maint_page(struct page *page, unsigned long offset,
490 size_t size, enum dma_data_direction dir,
491 void (*op)(const void *, size_t, int))
494 * A single sg entry may refer to multiple physically contiguous
495 * pages. But we still need to process highmem pages individually.
496 * If highmem is not configured then the bulk of this loop gets
504 if (PageHighMem(page)) {
505 if (len + offset > PAGE_SIZE) {
506 if (offset >= PAGE_SIZE) {
507 page += offset / PAGE_SIZE;
510 len = PAGE_SIZE - offset;
512 vaddr = kmap_high_get(page);
517 } else if (cache_is_vipt()) {
518 /* unmapped pages might still be cached */
519 vaddr = kmap_atomic(page);
520 op(vaddr + offset, len, dir);
521 kunmap_atomic(vaddr);
524 vaddr = page_address(page) + offset;
533 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
534 size_t size, enum dma_data_direction dir)
538 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
540 paddr = page_to_phys(page) + off;
541 if (dir == DMA_FROM_DEVICE) {
542 outer_inv_range(paddr, paddr + size);
544 outer_clean_range(paddr, paddr + size);
546 /* FIXME: non-speculating: flush on bidirectional mappings? */
548 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
550 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
551 size_t size, enum dma_data_direction dir)
553 unsigned long paddr = page_to_phys(page) + off;
555 /* FIXME: non-speculating: not required */
556 /* don't bother invalidating if DMA to device */
557 if (dir != DMA_TO_DEVICE)
558 outer_inv_range(paddr, paddr + size);
560 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
563 * Mark the D-cache clean for this page to avoid extra flushing.
565 if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
566 set_bit(PG_dcache_clean, &page->flags);
568 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
571 * dma_map_sg - map a set of SG buffers for streaming mode DMA
572 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
573 * @sg: list of buffers
574 * @nents: number of buffers to map
575 * @dir: DMA transfer direction
577 * Map a set of buffers described by scatterlist in streaming mode for DMA.
578 * This is the scatter-gather version of the dma_map_single interface.
579 * Here the scatter gather list elements are each tagged with the
580 * appropriate dma address and length. They are obtained via
581 * sg_dma_{address,length}.
583 * Device ownership issues as mentioned for dma_map_single are the same
586 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
587 enum dma_data_direction dir)
589 struct scatterlist *s;
592 BUG_ON(!valid_dma_direction(dir));
594 for_each_sg(sg, s, nents, i) {
595 s->dma_address = __dma_map_page(dev, sg_page(s), s->offset,
597 if (dma_mapping_error(dev, s->dma_address))
600 debug_dma_map_sg(dev, sg, nents, nents, dir);
604 for_each_sg(sg, s, i, j)
605 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
608 EXPORT_SYMBOL(dma_map_sg);
611 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
612 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
613 * @sg: list of buffers
614 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
615 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
617 * Unmap a set of streaming mode DMA translations. Again, CPU access
618 * rules concerning calls here are the same as for dma_unmap_single().
620 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
621 enum dma_data_direction dir)
623 struct scatterlist *s;
626 debug_dma_unmap_sg(dev, sg, nents, dir);
628 for_each_sg(sg, s, nents, i)
629 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
631 EXPORT_SYMBOL(dma_unmap_sg);
634 * dma_sync_sg_for_cpu
635 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
636 * @sg: list of buffers
637 * @nents: number of buffers to map (returned from dma_map_sg)
638 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
640 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
641 int nents, enum dma_data_direction dir)
643 struct scatterlist *s;
646 for_each_sg(sg, s, nents, i) {
647 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
651 __dma_page_dev_to_cpu(sg_page(s), s->offset,
655 debug_dma_sync_sg_for_cpu(dev, sg, nents, dir);
657 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
660 * dma_sync_sg_for_device
661 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
662 * @sg: list of buffers
663 * @nents: number of buffers to map (returned from dma_map_sg)
664 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
666 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
667 int nents, enum dma_data_direction dir)
669 struct scatterlist *s;
672 for_each_sg(sg, s, nents, i) {
673 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
677 __dma_page_cpu_to_dev(sg_page(s), s->offset,
681 debug_dma_sync_sg_for_device(dev, sg, nents, dir);
683 EXPORT_SYMBOL(dma_sync_sg_for_device);
686 * Return whether the given device DMA address mask can be supported
687 * properly. For example, if your device can only drive the low 24-bits
688 * during bus mastering, then you would pass 0x00ffffff as the mask
691 int dma_supported(struct device *dev, u64 mask)
693 if (mask < (u64)arm_dma_limit)
697 EXPORT_SYMBOL(dma_supported);
699 int dma_set_mask(struct device *dev, u64 dma_mask)
701 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
704 #ifndef CONFIG_DMABOUNCE
705 *dev->dma_mask = dma_mask;
710 EXPORT_SYMBOL(dma_set_mask);
712 #define PREALLOC_DMA_DEBUG_ENTRIES 4096
714 static int __init dma_debug_do_init(void)
716 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
719 fs_initcall(dma_debug_do_init);