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);
123 #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - consistent_base) >> PAGE_SHIFT)
124 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - consistent_base) >> PMD_SHIFT)
127 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
129 static pte_t **consistent_pte;
131 #define DEFAULT_CONSISTENT_DMA_SIZE SZ_2M
133 unsigned long consistent_base = CONSISTENT_END - DEFAULT_CONSISTENT_DMA_SIZE;
135 void __init init_consistent_dma_size(unsigned long size)
137 unsigned long base = CONSISTENT_END - ALIGN(size, SZ_2M);
139 BUG_ON(consistent_pte); /* Check we're called before DMA region init */
140 BUG_ON(base < VMALLOC_END);
142 /* Grow region to accommodate specified size */
143 if (base < consistent_base)
144 consistent_base = base;
147 #include "vmregion.h"
149 static struct arm_vmregion_head consistent_head = {
150 .vm_lock = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock),
151 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
152 .vm_end = CONSISTENT_END,
157 * Initialise the consistent memory allocation.
159 static int __init consistent_init(void)
167 unsigned long base = consistent_base;
168 unsigned long num_ptes = (CONSISTENT_END - base) >> PMD_SHIFT;
170 consistent_pte = kmalloc(num_ptes * sizeof(pte_t), GFP_KERNEL);
171 if (!consistent_pte) {
172 pr_err("%s: no memory\n", __func__);
176 pr_debug("DMA memory: 0x%08lx - 0x%08lx:\n", base, CONSISTENT_END);
177 consistent_head.vm_start = base;
180 pgd = pgd_offset(&init_mm, base);
182 pud = pud_alloc(&init_mm, pgd, base);
184 printk(KERN_ERR "%s: no pud tables\n", __func__);
189 pmd = pmd_alloc(&init_mm, pud, base);
191 printk(KERN_ERR "%s: no pmd tables\n", __func__);
195 WARN_ON(!pmd_none(*pmd));
197 pte = pte_alloc_kernel(pmd, base);
199 printk(KERN_ERR "%s: no pte tables\n", __func__);
204 consistent_pte[i++] = pte;
206 } while (base < CONSISTENT_END);
211 core_initcall(consistent_init);
214 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot)
216 struct arm_vmregion *c;
220 if (!consistent_pte) {
221 printk(KERN_ERR "%s: not initialised\n", __func__);
227 * Align the virtual region allocation - maximum alignment is
228 * a section size, minimum is a page size. This helps reduce
229 * fragmentation of the DMA space, and also prevents allocations
230 * smaller than a section from crossing a section boundary.
233 if (bit > SECTION_SHIFT)
238 * Allocate a virtual address in the consistent mapping region.
240 c = arm_vmregion_alloc(&consistent_head, align, size,
241 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
244 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
245 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
247 pte = consistent_pte[idx] + off;
251 BUG_ON(!pte_none(*pte));
253 set_pte_ext(pte, mk_pte(page, prot), 0);
257 if (off >= PTRS_PER_PTE) {
259 pte = consistent_pte[++idx];
261 } while (size -= PAGE_SIZE);
265 return (void *)c->vm_start;
270 static void __dma_free_remap(void *cpu_addr, size_t size)
272 struct arm_vmregion *c;
278 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
280 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
286 if ((c->vm_end - c->vm_start) != size) {
287 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
288 __func__, c->vm_end - c->vm_start, size);
290 size = c->vm_end - c->vm_start;
293 idx = CONSISTENT_PTE_INDEX(c->vm_start);
294 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
295 ptep = consistent_pte[idx] + off;
298 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
303 if (off >= PTRS_PER_PTE) {
305 ptep = consistent_pte[++idx];
308 if (pte_none(pte) || !pte_present(pte))
309 printk(KERN_CRIT "%s: bad page in kernel page table\n",
311 } while (size -= PAGE_SIZE);
313 flush_tlb_kernel_range(c->vm_start, c->vm_end);
315 arm_vmregion_free(&consistent_head, c);
318 #else /* !CONFIG_MMU */
320 #define __dma_alloc_remap(page, size, gfp, prot) page_address(page)
321 #define __dma_free_remap(addr, size) do { } while (0)
323 #endif /* CONFIG_MMU */
326 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
333 * Following is a work-around (a.k.a. hack) to prevent pages
334 * with __GFP_COMP being passed to split_page() which cannot
335 * handle them. The real problem is that this flag probably
336 * should be 0 on ARM as it is not supported on this
337 * platform; see CONFIG_HUGETLBFS.
339 gfp &= ~(__GFP_COMP);
342 size = PAGE_ALIGN(size);
344 page = __dma_alloc_buffer(dev, size, gfp);
348 if (!arch_is_coherent())
349 addr = __dma_alloc_remap(page, size, gfp, prot);
351 addr = page_address(page);
354 *handle = pfn_to_dma(dev, page_to_pfn(page));
356 __dma_free_buffer(page, size);
362 * Allocate DMA-coherent memory space and return both the kernel remapped
363 * virtual and bus address for that space.
366 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
370 if (dma_alloc_from_coherent(dev, size, handle, &memory))
373 return __dma_alloc(dev, size, handle, gfp,
374 pgprot_dmacoherent(pgprot_kernel));
376 EXPORT_SYMBOL(dma_alloc_coherent);
379 * Allocate a writecombining region, in much the same way as
380 * dma_alloc_coherent above.
383 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
385 return __dma_alloc(dev, size, handle, gfp,
386 pgprot_writecombine(pgprot_kernel));
388 EXPORT_SYMBOL(dma_alloc_writecombine);
390 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
391 void *cpu_addr, dma_addr_t dma_addr, size_t size)
395 unsigned long user_size, kern_size;
396 struct arm_vmregion *c;
398 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
400 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
402 unsigned long off = vma->vm_pgoff;
404 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
406 if (off < kern_size &&
407 user_size <= (kern_size - off)) {
408 ret = remap_pfn_range(vma, vma->vm_start,
409 page_to_pfn(c->vm_pages) + off,
410 user_size << PAGE_SHIFT,
414 #endif /* CONFIG_MMU */
419 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
420 void *cpu_addr, dma_addr_t dma_addr, size_t size)
422 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
423 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
425 EXPORT_SYMBOL(dma_mmap_coherent);
427 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
428 void *cpu_addr, dma_addr_t dma_addr, size_t size)
430 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
431 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
433 EXPORT_SYMBOL(dma_mmap_writecombine);
436 * free a page as defined by the above mapping.
437 * Must not be called with IRQs disabled.
439 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
441 WARN_ON(irqs_disabled());
443 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
446 size = PAGE_ALIGN(size);
448 if (!arch_is_coherent())
449 __dma_free_remap(cpu_addr, size);
451 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev, handle)), size);
453 EXPORT_SYMBOL(dma_free_coherent);
456 * Make an area consistent for devices.
457 * Note: Drivers should NOT use this function directly, as it will break
458 * platforms with CONFIG_DMABOUNCE.
459 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
461 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
462 enum dma_data_direction dir)
466 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
468 dmac_map_area(kaddr, size, dir);
471 if (dir == DMA_FROM_DEVICE) {
472 outer_inv_range(paddr, paddr + size);
474 outer_clean_range(paddr, paddr + size);
476 /* FIXME: non-speculating: flush on bidirectional mappings? */
478 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
480 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
481 enum dma_data_direction dir)
483 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
485 /* FIXME: non-speculating: not required */
486 /* don't bother invalidating if DMA to device */
487 if (dir != DMA_TO_DEVICE) {
488 unsigned long paddr = __pa(kaddr);
489 outer_inv_range(paddr, paddr + size);
492 dmac_unmap_area(kaddr, size, dir);
494 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
496 static void dma_cache_maint_page(struct page *page, unsigned long offset,
497 size_t size, enum dma_data_direction dir,
498 void (*op)(const void *, size_t, int))
503 pfn = page_to_pfn(page) + offset / PAGE_SIZE;
507 * A single sg entry may refer to multiple physically contiguous
508 * pages. But we still need to process highmem pages individually.
509 * If highmem is not configured then the bulk of this loop gets
516 page = pfn_to_page(pfn);
518 if (PageHighMem(page)) {
519 if (len + offset > PAGE_SIZE)
520 len = PAGE_SIZE - offset;
521 vaddr = kmap_high_get(page);
526 } else if (cache_is_vipt()) {
527 /* unmapped pages might still be cached */
528 vaddr = kmap_atomic(page);
529 op(vaddr + offset, len, dir);
530 kunmap_atomic(vaddr);
533 vaddr = page_address(page) + offset;
542 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
543 size_t size, enum dma_data_direction dir)
547 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
549 paddr = page_to_phys(page) + off;
550 if (dir == DMA_FROM_DEVICE) {
551 outer_inv_range(paddr, paddr + size);
553 outer_clean_range(paddr, paddr + size);
555 /* FIXME: non-speculating: flush on bidirectional mappings? */
557 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
559 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
560 size_t size, enum dma_data_direction dir)
562 unsigned long paddr = page_to_phys(page) + off;
564 /* FIXME: non-speculating: not required */
565 /* don't bother invalidating if DMA to device */
566 if (dir != DMA_TO_DEVICE)
567 outer_inv_range(paddr, paddr + size);
569 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
572 * Mark the D-cache clean for this page to avoid extra flushing.
574 if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
575 set_bit(PG_dcache_clean, &page->flags);
577 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
580 * dma_map_sg - map a set of SG buffers for streaming mode DMA
581 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
582 * @sg: list of buffers
583 * @nents: number of buffers to map
584 * @dir: DMA transfer direction
586 * Map a set of buffers described by scatterlist in streaming mode for DMA.
587 * This is the scatter-gather version of the dma_map_single interface.
588 * Here the scatter gather list elements are each tagged with the
589 * appropriate dma address and length. They are obtained via
590 * sg_dma_{address,length}.
592 * Device ownership issues as mentioned for dma_map_single are the same
595 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
596 enum dma_data_direction dir)
598 struct scatterlist *s;
601 BUG_ON(!valid_dma_direction(dir));
603 for_each_sg(sg, s, nents, i) {
604 s->dma_address = __dma_map_page(dev, sg_page(s), s->offset,
606 if (dma_mapping_error(dev, s->dma_address))
609 debug_dma_map_sg(dev, sg, nents, nents, dir);
613 for_each_sg(sg, s, i, j)
614 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
617 EXPORT_SYMBOL(dma_map_sg);
620 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
621 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
622 * @sg: list of buffers
623 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
624 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
626 * Unmap a set of streaming mode DMA translations. Again, CPU access
627 * rules concerning calls here are the same as for dma_unmap_single().
629 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
630 enum dma_data_direction dir)
632 struct scatterlist *s;
635 debug_dma_unmap_sg(dev, sg, nents, dir);
637 for_each_sg(sg, s, nents, i)
638 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
640 EXPORT_SYMBOL(dma_unmap_sg);
643 * dma_sync_sg_for_cpu
644 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
645 * @sg: list of buffers
646 * @nents: number of buffers to map (returned from dma_map_sg)
647 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
649 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
650 int nents, enum dma_data_direction dir)
652 struct scatterlist *s;
655 for_each_sg(sg, s, nents, i) {
656 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
660 __dma_page_dev_to_cpu(sg_page(s), s->offset,
664 debug_dma_sync_sg_for_cpu(dev, sg, nents, dir);
666 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
669 * dma_sync_sg_for_device
670 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
671 * @sg: list of buffers
672 * @nents: number of buffers to map (returned from dma_map_sg)
673 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
675 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
676 int nents, enum dma_data_direction dir)
678 struct scatterlist *s;
681 for_each_sg(sg, s, nents, i) {
682 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
686 __dma_page_cpu_to_dev(sg_page(s), s->offset,
690 debug_dma_sync_sg_for_device(dev, sg, nents, dir);
692 EXPORT_SYMBOL(dma_sync_sg_for_device);
695 * Return whether the given device DMA address mask can be supported
696 * properly. For example, if your device can only drive the low 24-bits
697 * during bus mastering, then you would pass 0x00ffffff as the mask
700 int dma_supported(struct device *dev, u64 mask)
702 if (mask < (u64)arm_dma_limit)
706 EXPORT_SYMBOL(dma_supported);
708 int dma_set_mask(struct device *dev, u64 dma_mask)
710 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
713 #ifndef CONFIG_DMABOUNCE
714 *dev->dma_mask = dma_mask;
719 EXPORT_SYMBOL(dma_set_mask);
721 #define PREALLOC_DMA_DEBUG_ENTRIES 4096
723 static int __init dma_debug_do_init(void)
725 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
728 fs_initcall(dma_debug_do_init);