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
,
155 #ifdef CONFIG_HUGETLB_PAGE
156 #error ARM Coherent DMA allocator does not (yet) support huge TLB
160 * Initialise the consistent memory allocation.
162 static int __init
consistent_init(void)
170 unsigned long base
= consistent_base
;
171 unsigned long num_ptes
= (CONSISTENT_END
- base
) >> PGDIR_SHIFT
;
173 consistent_pte
= kmalloc(num_ptes
* sizeof(pte_t
), GFP_KERNEL
);
174 if (!consistent_pte
) {
175 pr_err("%s: no memory\n", __func__
);
179 pr_debug("DMA memory: 0x%08lx - 0x%08lx:\n", base
, CONSISTENT_END
);
180 consistent_head
.vm_start
= base
;
183 pgd
= pgd_offset(&init_mm
, base
);
185 pud
= pud_alloc(&init_mm
, pgd
, base
);
187 printk(KERN_ERR
"%s: no pud tables\n", __func__
);
192 pmd
= pmd_alloc(&init_mm
, pud
, base
);
194 printk(KERN_ERR
"%s: no pmd tables\n", __func__
);
198 WARN_ON(!pmd_none(*pmd
));
200 pte
= pte_alloc_kernel(pmd
, base
);
202 printk(KERN_ERR
"%s: no pte tables\n", __func__
);
207 consistent_pte
[i
++] = pte
;
209 } while (base
< CONSISTENT_END
);
214 core_initcall(consistent_init
);
217 __dma_alloc_remap(struct page
*page
, size_t size
, gfp_t gfp
, pgprot_t prot
)
219 struct arm_vmregion
*c
;
223 if (!consistent_pte
) {
224 printk(KERN_ERR
"%s: not initialised\n", __func__
);
230 * Align the virtual region allocation - maximum alignment is
231 * a section size, minimum is a page size. This helps reduce
232 * fragmentation of the DMA space, and also prevents allocations
233 * smaller than a section from crossing a section boundary.
236 if (bit
> SECTION_SHIFT
)
241 * Allocate a virtual address in the consistent mapping region.
243 c
= arm_vmregion_alloc(&consistent_head
, align
, size
,
244 gfp
& ~(__GFP_DMA
| __GFP_HIGHMEM
));
247 int idx
= CONSISTENT_PTE_INDEX(c
->vm_start
);
248 u32 off
= CONSISTENT_OFFSET(c
->vm_start
) & (PTRS_PER_PTE
-1);
250 pte
= consistent_pte
[idx
] + off
;
254 BUG_ON(!pte_none(*pte
));
256 set_pte_ext(pte
, mk_pte(page
, prot
), 0);
260 if (off
>= PTRS_PER_PTE
) {
262 pte
= consistent_pte
[++idx
];
264 } while (size
-= PAGE_SIZE
);
268 return (void *)c
->vm_start
;
273 static void __dma_free_remap(void *cpu_addr
, size_t size
)
275 struct arm_vmregion
*c
;
281 c
= arm_vmregion_find_remove(&consistent_head
, (unsigned long)cpu_addr
);
283 printk(KERN_ERR
"%s: trying to free invalid coherent area: %p\n",
289 if ((c
->vm_end
- c
->vm_start
) != size
) {
290 printk(KERN_ERR
"%s: freeing wrong coherent size (%ld != %d)\n",
291 __func__
, c
->vm_end
- c
->vm_start
, size
);
293 size
= c
->vm_end
- c
->vm_start
;
296 idx
= CONSISTENT_PTE_INDEX(c
->vm_start
);
297 off
= CONSISTENT_OFFSET(c
->vm_start
) & (PTRS_PER_PTE
-1);
298 ptep
= consistent_pte
[idx
] + off
;
301 pte_t pte
= ptep_get_and_clear(&init_mm
, addr
, ptep
);
306 if (off
>= PTRS_PER_PTE
) {
308 ptep
= consistent_pte
[++idx
];
311 if (pte_none(pte
) || !pte_present(pte
))
312 printk(KERN_CRIT
"%s: bad page in kernel page table\n",
314 } while (size
-= PAGE_SIZE
);
316 flush_tlb_kernel_range(c
->vm_start
, c
->vm_end
);
318 arm_vmregion_free(&consistent_head
, c
);
321 #else /* !CONFIG_MMU */
323 #define __dma_alloc_remap(page, size, gfp, prot) page_address(page)
324 #define __dma_free_remap(addr, size) do { } while (0)
326 #endif /* CONFIG_MMU */
329 __dma_alloc(struct device
*dev
, size_t size
, dma_addr_t
*handle
, gfp_t gfp
,
336 size
= PAGE_ALIGN(size
);
338 page
= __dma_alloc_buffer(dev
, size
, gfp
);
342 if (!arch_is_coherent())
343 addr
= __dma_alloc_remap(page
, size
, gfp
, prot
);
345 addr
= page_address(page
);
348 *handle
= pfn_to_dma(dev
, page_to_pfn(page
));
350 __dma_free_buffer(page
, size
);
356 * Allocate DMA-coherent memory space and return both the kernel remapped
357 * virtual and bus address for that space.
360 dma_alloc_coherent(struct device
*dev
, size_t size
, dma_addr_t
*handle
, gfp_t gfp
)
364 if (dma_alloc_from_coherent(dev
, size
, handle
, &memory
))
367 return __dma_alloc(dev
, size
, handle
, gfp
,
368 pgprot_dmacoherent(pgprot_kernel
));
370 EXPORT_SYMBOL(dma_alloc_coherent
);
373 * Allocate a writecombining region, in much the same way as
374 * dma_alloc_coherent above.
377 dma_alloc_writecombine(struct device
*dev
, size_t size
, dma_addr_t
*handle
, gfp_t gfp
)
379 return __dma_alloc(dev
, size
, handle
, gfp
,
380 pgprot_writecombine(pgprot_kernel
));
382 EXPORT_SYMBOL(dma_alloc_writecombine
);
384 static int dma_mmap(struct device
*dev
, struct vm_area_struct
*vma
,
385 void *cpu_addr
, dma_addr_t dma_addr
, size_t size
)
389 unsigned long user_size
, kern_size
;
390 struct arm_vmregion
*c
;
392 user_size
= (vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
;
394 c
= arm_vmregion_find(&consistent_head
, (unsigned long)cpu_addr
);
396 unsigned long off
= vma
->vm_pgoff
;
398 kern_size
= (c
->vm_end
- c
->vm_start
) >> PAGE_SHIFT
;
400 if (off
< kern_size
&&
401 user_size
<= (kern_size
- off
)) {
402 ret
= remap_pfn_range(vma
, vma
->vm_start
,
403 page_to_pfn(c
->vm_pages
) + off
,
404 user_size
<< PAGE_SHIFT
,
408 #endif /* CONFIG_MMU */
413 int dma_mmap_coherent(struct device
*dev
, struct vm_area_struct
*vma
,
414 void *cpu_addr
, dma_addr_t dma_addr
, size_t size
)
416 vma
->vm_page_prot
= pgprot_dmacoherent(vma
->vm_page_prot
);
417 return dma_mmap(dev
, vma
, cpu_addr
, dma_addr
, size
);
419 EXPORT_SYMBOL(dma_mmap_coherent
);
421 int dma_mmap_writecombine(struct device
*dev
, struct vm_area_struct
*vma
,
422 void *cpu_addr
, dma_addr_t dma_addr
, size_t size
)
424 vma
->vm_page_prot
= pgprot_writecombine(vma
->vm_page_prot
);
425 return dma_mmap(dev
, vma
, cpu_addr
, dma_addr
, size
);
427 EXPORT_SYMBOL(dma_mmap_writecombine
);
430 * free a page as defined by the above mapping.
431 * Must not be called with IRQs disabled.
433 void dma_free_coherent(struct device
*dev
, size_t size
, void *cpu_addr
, dma_addr_t handle
)
435 WARN_ON(irqs_disabled());
437 if (dma_release_from_coherent(dev
, get_order(size
), cpu_addr
))
440 size
= PAGE_ALIGN(size
);
442 if (!arch_is_coherent())
443 __dma_free_remap(cpu_addr
, size
);
445 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev
, handle
)), size
);
447 EXPORT_SYMBOL(dma_free_coherent
);
450 * Make an area consistent for devices.
451 * Note: Drivers should NOT use this function directly, as it will break
452 * platforms with CONFIG_DMABOUNCE.
453 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
455 void ___dma_single_cpu_to_dev(const void *kaddr
, size_t size
,
456 enum dma_data_direction dir
)
460 BUG_ON(!virt_addr_valid(kaddr
) || !virt_addr_valid(kaddr
+ size
- 1));
462 dmac_map_area(kaddr
, size
, dir
);
465 if (dir
== DMA_FROM_DEVICE
) {
466 outer_inv_range(paddr
, paddr
+ size
);
468 outer_clean_range(paddr
, paddr
+ size
);
470 /* FIXME: non-speculating: flush on bidirectional mappings? */
472 EXPORT_SYMBOL(___dma_single_cpu_to_dev
);
474 void ___dma_single_dev_to_cpu(const void *kaddr
, size_t size
,
475 enum dma_data_direction dir
)
477 BUG_ON(!virt_addr_valid(kaddr
) || !virt_addr_valid(kaddr
+ size
- 1));
479 /* FIXME: non-speculating: not required */
480 /* don't bother invalidating if DMA to device */
481 if (dir
!= DMA_TO_DEVICE
) {
482 unsigned long paddr
= __pa(kaddr
);
483 outer_inv_range(paddr
, paddr
+ size
);
486 dmac_unmap_area(kaddr
, size
, dir
);
488 EXPORT_SYMBOL(___dma_single_dev_to_cpu
);
490 static void dma_cache_maint_page(struct page
*page
, unsigned long offset
,
491 size_t size
, enum dma_data_direction dir
,
492 void (*op
)(const void *, size_t, int))
495 * A single sg entry may refer to multiple physically contiguous
496 * pages. But we still need to process highmem pages individually.
497 * If highmem is not configured then the bulk of this loop gets
505 if (PageHighMem(page
)) {
506 if (len
+ offset
> PAGE_SIZE
) {
507 if (offset
>= PAGE_SIZE
) {
508 page
+= offset
/ PAGE_SIZE
;
511 len
= PAGE_SIZE
- offset
;
513 vaddr
= kmap_high_get(page
);
518 } else if (cache_is_vipt()) {
519 /* unmapped pages might still be cached */
520 vaddr
= kmap_atomic(page
);
521 op(vaddr
+ offset
, len
, dir
);
522 kunmap_atomic(vaddr
);
525 vaddr
= page_address(page
) + offset
;
534 void ___dma_page_cpu_to_dev(struct page
*page
, unsigned long off
,
535 size_t size
, enum dma_data_direction dir
)
539 dma_cache_maint_page(page
, off
, size
, dir
, dmac_map_area
);
541 paddr
= page_to_phys(page
) + off
;
542 if (dir
== DMA_FROM_DEVICE
) {
543 outer_inv_range(paddr
, paddr
+ size
);
545 outer_clean_range(paddr
, paddr
+ size
);
547 /* FIXME: non-speculating: flush on bidirectional mappings? */
549 EXPORT_SYMBOL(___dma_page_cpu_to_dev
);
551 void ___dma_page_dev_to_cpu(struct page
*page
, unsigned long off
,
552 size_t size
, enum dma_data_direction dir
)
554 unsigned long paddr
= page_to_phys(page
) + off
;
556 /* FIXME: non-speculating: not required */
557 /* don't bother invalidating if DMA to device */
558 if (dir
!= DMA_TO_DEVICE
)
559 outer_inv_range(paddr
, paddr
+ size
);
561 dma_cache_maint_page(page
, off
, size
, dir
, dmac_unmap_area
);
564 * Mark the D-cache clean for this page to avoid extra flushing.
566 if (dir
!= DMA_TO_DEVICE
&& off
== 0 && size
>= PAGE_SIZE
)
567 set_bit(PG_dcache_clean
, &page
->flags
);
569 EXPORT_SYMBOL(___dma_page_dev_to_cpu
);
572 * dma_map_sg - map a set of SG buffers for streaming mode DMA
573 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
574 * @sg: list of buffers
575 * @nents: number of buffers to map
576 * @dir: DMA transfer direction
578 * Map a set of buffers described by scatterlist in streaming mode for DMA.
579 * This is the scatter-gather version of the dma_map_single interface.
580 * Here the scatter gather list elements are each tagged with the
581 * appropriate dma address and length. They are obtained via
582 * sg_dma_{address,length}.
584 * Device ownership issues as mentioned for dma_map_single are the same
587 int dma_map_sg(struct device
*dev
, struct scatterlist
*sg
, int nents
,
588 enum dma_data_direction dir
)
590 struct scatterlist
*s
;
593 BUG_ON(!valid_dma_direction(dir
));
595 for_each_sg(sg
, s
, nents
, i
) {
596 s
->dma_address
= __dma_map_page(dev
, sg_page(s
), s
->offset
,
598 if (dma_mapping_error(dev
, s
->dma_address
))
601 debug_dma_map_sg(dev
, sg
, nents
, nents
, dir
);
605 for_each_sg(sg
, s
, i
, j
)
606 __dma_unmap_page(dev
, sg_dma_address(s
), sg_dma_len(s
), dir
);
609 EXPORT_SYMBOL(dma_map_sg
);
612 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
613 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
614 * @sg: list of buffers
615 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
616 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
618 * Unmap a set of streaming mode DMA translations. Again, CPU access
619 * rules concerning calls here are the same as for dma_unmap_single().
621 void dma_unmap_sg(struct device
*dev
, struct scatterlist
*sg
, int nents
,
622 enum dma_data_direction dir
)
624 struct scatterlist
*s
;
627 debug_dma_unmap_sg(dev
, sg
, nents
, dir
);
629 for_each_sg(sg
, s
, nents
, i
)
630 __dma_unmap_page(dev
, sg_dma_address(s
), sg_dma_len(s
), dir
);
632 EXPORT_SYMBOL(dma_unmap_sg
);
635 * dma_sync_sg_for_cpu
636 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
637 * @sg: list of buffers
638 * @nents: number of buffers to map (returned from dma_map_sg)
639 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
641 void dma_sync_sg_for_cpu(struct device
*dev
, struct scatterlist
*sg
,
642 int nents
, enum dma_data_direction dir
)
644 struct scatterlist
*s
;
647 for_each_sg(sg
, s
, nents
, i
) {
648 if (!dmabounce_sync_for_cpu(dev
, sg_dma_address(s
), 0,
652 __dma_page_dev_to_cpu(sg_page(s
), s
->offset
,
656 debug_dma_sync_sg_for_cpu(dev
, sg
, nents
, dir
);
658 EXPORT_SYMBOL(dma_sync_sg_for_cpu
);
661 * dma_sync_sg_for_device
662 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
663 * @sg: list of buffers
664 * @nents: number of buffers to map (returned from dma_map_sg)
665 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
667 void dma_sync_sg_for_device(struct device
*dev
, struct scatterlist
*sg
,
668 int nents
, enum dma_data_direction dir
)
670 struct scatterlist
*s
;
673 for_each_sg(sg
, s
, nents
, i
) {
674 if (!dmabounce_sync_for_device(dev
, sg_dma_address(s
), 0,
678 __dma_page_cpu_to_dev(sg_page(s
), s
->offset
,
682 debug_dma_sync_sg_for_device(dev
, sg
, nents
, dir
);
684 EXPORT_SYMBOL(dma_sync_sg_for_device
);
687 * Return whether the given device DMA address mask can be supported
688 * properly. For example, if your device can only drive the low 24-bits
689 * during bus mastering, then you would pass 0x00ffffff as the mask
692 int dma_supported(struct device
*dev
, u64 mask
)
694 if (mask
< (u64
)arm_dma_limit
)
698 EXPORT_SYMBOL(dma_supported
);
700 int dma_set_mask(struct device
*dev
, u64 dma_mask
)
702 if (!dev
->dma_mask
|| !dma_supported(dev
, dma_mask
))
705 #ifndef CONFIG_DMABOUNCE
706 *dev
->dma_mask
= dma_mask
;
711 EXPORT_SYMBOL(dma_set_mask
);
713 #define PREALLOC_DMA_DEBUG_ENTRIES 4096
715 static int __init
dma_debug_do_init(void)
717 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES
);
720 fs_initcall(dma_debug_do_init
);