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>
33 * The DMA API is built upon the notion of "buffer ownership". A buffer
34 * is either exclusively owned by the CPU (and therefore may be accessed
35 * by it) or exclusively owned by the DMA device. These helper functions
36 * represent the transitions between these two ownership states.
38 * Note, however, that on later ARMs, this notion does not work due to
39 * speculative prefetches. We model our approach on the assumption that
40 * the CPU does do speculative prefetches, which means we clean caches
41 * before transfers and delay cache invalidation until transfer completion.
44 static void __dma_page_cpu_to_dev(struct page *, unsigned long,
45 size_t, enum dma_data_direction);
46 static void __dma_page_dev_to_cpu(struct page *, unsigned long,
47 size_t, enum dma_data_direction);
50 * arm_dma_map_page - map a portion of a page for streaming DMA
51 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
52 * @page: page that buffer resides in
53 * @offset: offset into page for start of buffer
54 * @size: size of buffer to map
55 * @dir: DMA transfer direction
57 * Ensure that any data held in the cache is appropriately discarded
60 * The device owns this memory once this call has completed. The CPU
61 * can regain ownership by calling dma_unmap_page().
63 static dma_addr_t arm_dma_map_page(struct device *dev, struct page *page,
64 unsigned long offset, size_t size, enum dma_data_direction dir,
65 struct dma_attrs *attrs)
67 if (!arch_is_coherent())
68 __dma_page_cpu_to_dev(page, offset, size, dir);
69 return pfn_to_dma(dev, page_to_pfn(page)) + offset;
73 * arm_dma_unmap_page - unmap a buffer previously mapped through dma_map_page()
74 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
75 * @handle: DMA address of buffer
76 * @size: size of buffer (same as passed to dma_map_page)
77 * @dir: DMA transfer direction (same as passed to dma_map_page)
79 * Unmap a page streaming mode DMA translation. The handle and size
80 * must match what was provided in the previous dma_map_page() call.
81 * All other usages are undefined.
83 * After this call, reads by the CPU to the buffer are guaranteed to see
84 * whatever the device wrote there.
86 static void arm_dma_unmap_page(struct device *dev, dma_addr_t handle,
87 size_t size, enum dma_data_direction dir,
88 struct dma_attrs *attrs)
90 if (!arch_is_coherent())
91 __dma_page_dev_to_cpu(pfn_to_page(dma_to_pfn(dev, handle)),
92 handle & ~PAGE_MASK, size, dir);
95 static void arm_dma_sync_single_for_cpu(struct device *dev,
96 dma_addr_t handle, size_t size, enum dma_data_direction dir)
98 unsigned int offset = handle & (PAGE_SIZE - 1);
99 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
100 if (!arch_is_coherent())
101 __dma_page_dev_to_cpu(page, offset, size, dir);
104 static void arm_dma_sync_single_for_device(struct device *dev,
105 dma_addr_t handle, size_t size, enum dma_data_direction dir)
107 unsigned int offset = handle & (PAGE_SIZE - 1);
108 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
109 if (!arch_is_coherent())
110 __dma_page_cpu_to_dev(page, offset, size, dir);
113 static int arm_dma_set_mask(struct device *dev, u64 dma_mask);
115 struct dma_map_ops arm_dma_ops = {
116 .alloc = arm_dma_alloc,
117 .free = arm_dma_free,
118 .mmap = arm_dma_mmap,
119 .map_page = arm_dma_map_page,
120 .unmap_page = arm_dma_unmap_page,
121 .map_sg = arm_dma_map_sg,
122 .unmap_sg = arm_dma_unmap_sg,
123 .sync_single_for_cpu = arm_dma_sync_single_for_cpu,
124 .sync_single_for_device = arm_dma_sync_single_for_device,
125 .sync_sg_for_cpu = arm_dma_sync_sg_for_cpu,
126 .sync_sg_for_device = arm_dma_sync_sg_for_device,
127 .set_dma_mask = arm_dma_set_mask,
129 EXPORT_SYMBOL(arm_dma_ops);
131 static u64 get_coherent_dma_mask(struct device *dev)
133 u64 mask = (u64)arm_dma_limit;
136 mask = dev->coherent_dma_mask;
139 * Sanity check the DMA mask - it must be non-zero, and
140 * must be able to be satisfied by a DMA allocation.
143 dev_warn(dev, "coherent DMA mask is unset\n");
147 if ((~mask) & (u64)arm_dma_limit) {
148 dev_warn(dev, "coherent DMA mask %#llx is smaller "
149 "than system GFP_DMA mask %#llx\n",
150 mask, (u64)arm_dma_limit);
159 * Allocate a DMA buffer for 'dev' of size 'size' using the
160 * specified gfp mask. Note that 'size' must be page aligned.
162 static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
164 unsigned long order = get_order(size);
165 struct page *page, *p, *e;
167 u64 mask = get_coherent_dma_mask(dev);
169 #ifdef CONFIG_DMA_API_DEBUG
170 u64 limit = (mask + 1) & ~mask;
171 if (limit && size >= limit) {
172 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
181 if (mask < 0xffffffffULL)
184 page = alloc_pages(gfp, order);
189 * Now split the huge page and free the excess pages
191 split_page(page, order);
192 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
196 * Ensure that the allocated pages are zeroed, and that any data
197 * lurking in the kernel direct-mapped region is invalidated.
199 ptr = page_address(page);
200 memset(ptr, 0, size);
201 dmac_flush_range(ptr, ptr + size);
202 outer_flush_range(__pa(ptr), __pa(ptr) + size);
208 * Free a DMA buffer. 'size' must be page aligned.
210 static void __dma_free_buffer(struct page *page, size_t size)
212 struct page *e = page + (size >> PAGE_SHIFT);
222 #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - consistent_base) >> PAGE_SHIFT)
223 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - consistent_base) >> PMD_SHIFT)
226 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
228 static pte_t **consistent_pte;
230 #define DEFAULT_CONSISTENT_DMA_SIZE SZ_2M
232 unsigned long consistent_base = CONSISTENT_END - DEFAULT_CONSISTENT_DMA_SIZE;
234 void __init init_consistent_dma_size(unsigned long size)
236 unsigned long base = CONSISTENT_END - ALIGN(size, SZ_2M);
238 BUG_ON(consistent_pte); /* Check we're called before DMA region init */
239 BUG_ON(base < VMALLOC_END);
241 /* Grow region to accommodate specified size */
242 if (base < consistent_base)
243 consistent_base = base;
246 #include "vmregion.h"
248 static struct arm_vmregion_head consistent_head = {
249 .vm_lock = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock),
250 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
251 .vm_end = CONSISTENT_END,
254 #ifdef CONFIG_HUGETLB_PAGE
255 #error ARM Coherent DMA allocator does not (yet) support huge TLB
259 * Initialise the consistent memory allocation.
261 static int __init consistent_init(void)
269 unsigned long base = consistent_base;
270 unsigned long num_ptes = (CONSISTENT_END - base) >> PMD_SHIFT;
272 consistent_pte = kmalloc(num_ptes * sizeof(pte_t), GFP_KERNEL);
273 if (!consistent_pte) {
274 pr_err("%s: no memory\n", __func__);
278 pr_debug("DMA memory: 0x%08lx - 0x%08lx:\n", base, CONSISTENT_END);
279 consistent_head.vm_start = base;
282 pgd = pgd_offset(&init_mm, base);
284 pud = pud_alloc(&init_mm, pgd, base);
286 pr_err("%s: no pud tables\n", __func__);
291 pmd = pmd_alloc(&init_mm, pud, base);
293 pr_err("%s: no pmd tables\n", __func__);
297 WARN_ON(!pmd_none(*pmd));
299 pte = pte_alloc_kernel(pmd, base);
301 pr_err("%s: no pte tables\n", __func__);
306 consistent_pte[i++] = pte;
308 } while (base < CONSISTENT_END);
313 core_initcall(consistent_init);
316 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot,
319 struct arm_vmregion *c;
323 if (!consistent_pte) {
324 pr_err("%s: not initialised\n", __func__);
330 * Align the virtual region allocation - maximum alignment is
331 * a section size, minimum is a page size. This helps reduce
332 * fragmentation of the DMA space, and also prevents allocations
333 * smaller than a section from crossing a section boundary.
336 if (bit > SECTION_SHIFT)
341 * Allocate a virtual address in the consistent mapping region.
343 c = arm_vmregion_alloc(&consistent_head, align, size,
344 gfp & ~(__GFP_DMA | __GFP_HIGHMEM), caller);
347 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
348 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
350 pte = consistent_pte[idx] + off;
354 BUG_ON(!pte_none(*pte));
356 set_pte_ext(pte, mk_pte(page, prot), 0);
360 if (off >= PTRS_PER_PTE) {
362 pte = consistent_pte[++idx];
364 } while (size -= PAGE_SIZE);
368 return (void *)c->vm_start;
373 static void __dma_free_remap(void *cpu_addr, size_t size)
375 struct arm_vmregion *c;
381 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
383 pr_err("%s: trying to free invalid coherent area: %p\n",
389 if ((c->vm_end - c->vm_start) != size) {
390 pr_err("%s: freeing wrong coherent size (%ld != %d)\n",
391 __func__, c->vm_end - c->vm_start, size);
393 size = c->vm_end - c->vm_start;
396 idx = CONSISTENT_PTE_INDEX(c->vm_start);
397 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
398 ptep = consistent_pte[idx] + off;
401 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
406 if (off >= PTRS_PER_PTE) {
408 ptep = consistent_pte[++idx];
411 if (pte_none(pte) || !pte_present(pte))
412 pr_crit("%s: bad page in kernel page table\n",
414 } while (size -= PAGE_SIZE);
416 flush_tlb_kernel_range(c->vm_start, c->vm_end);
418 arm_vmregion_free(&consistent_head, c);
421 static inline pgprot_t __get_dma_pgprot(struct dma_attrs *attrs, pgprot_t prot)
423 prot = dma_get_attr(DMA_ATTR_WRITE_COMBINE, attrs) ?
424 pgprot_writecombine(prot) :
425 pgprot_dmacoherent(prot);
429 #else /* !CONFIG_MMU */
431 #define __dma_alloc_remap(page, size, gfp, prot, c) page_address(page)
432 #define __dma_free_remap(addr, size) do { } while (0)
433 #define __get_dma_pgprot(attrs, prot) __pgprot(0)
435 #endif /* CONFIG_MMU */
438 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
439 pgprot_t prot, const void *caller)
445 * Following is a work-around (a.k.a. hack) to prevent pages
446 * with __GFP_COMP being passed to split_page() which cannot
447 * handle them. The real problem is that this flag probably
448 * should be 0 on ARM as it is not supported on this
449 * platform; see CONFIG_HUGETLBFS.
451 gfp &= ~(__GFP_COMP);
453 *handle = DMA_ERROR_CODE;
454 size = PAGE_ALIGN(size);
456 page = __dma_alloc_buffer(dev, size, gfp);
460 if (!arch_is_coherent())
461 addr = __dma_alloc_remap(page, size, gfp, prot, caller);
463 addr = page_address(page);
466 *handle = pfn_to_dma(dev, page_to_pfn(page));
468 __dma_free_buffer(page, size);
474 * Allocate DMA-coherent memory space and return both the kernel remapped
475 * virtual and bus address for that space.
477 void *arm_dma_alloc(struct device *dev, size_t size, dma_addr_t *handle,
478 gfp_t gfp, struct dma_attrs *attrs)
480 pgprot_t prot = __get_dma_pgprot(attrs, pgprot_kernel);
483 if (dma_alloc_from_coherent(dev, size, handle, &memory))
486 return __dma_alloc(dev, size, handle, gfp, prot,
487 __builtin_return_address(0));
491 * Create userspace mapping for the DMA-coherent memory.
493 int arm_dma_mmap(struct device *dev, struct vm_area_struct *vma,
494 void *cpu_addr, dma_addr_t dma_addr, size_t size,
495 struct dma_attrs *attrs)
499 unsigned long user_size, kern_size;
500 struct arm_vmregion *c;
502 vma->vm_page_prot = __get_dma_pgprot(attrs, vma->vm_page_prot);
504 if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
507 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
509 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
511 unsigned long off = vma->vm_pgoff;
513 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
515 if (off < kern_size &&
516 user_size <= (kern_size - off)) {
517 ret = remap_pfn_range(vma, vma->vm_start,
518 page_to_pfn(c->vm_pages) + off,
519 user_size << PAGE_SHIFT,
523 #endif /* CONFIG_MMU */
529 * free a page as defined by the above mapping.
530 * Must not be called with IRQs disabled.
532 void arm_dma_free(struct device *dev, size_t size, void *cpu_addr,
533 dma_addr_t handle, struct dma_attrs *attrs)
535 WARN_ON(irqs_disabled());
537 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
540 size = PAGE_ALIGN(size);
542 if (!arch_is_coherent())
543 __dma_free_remap(cpu_addr, size);
545 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev, handle)), size);
548 static void dma_cache_maint_page(struct page *page, unsigned long offset,
549 size_t size, enum dma_data_direction dir,
550 void (*op)(const void *, size_t, int))
553 * A single sg entry may refer to multiple physically contiguous
554 * pages. But we still need to process highmem pages individually.
555 * If highmem is not configured then the bulk of this loop gets
563 if (PageHighMem(page)) {
564 if (len + offset > PAGE_SIZE) {
565 if (offset >= PAGE_SIZE) {
566 page += offset / PAGE_SIZE;
569 len = PAGE_SIZE - offset;
571 vaddr = kmap_high_get(page);
576 } else if (cache_is_vipt()) {
577 /* unmapped pages might still be cached */
578 vaddr = kmap_atomic(page);
579 op(vaddr + offset, len, dir);
580 kunmap_atomic(vaddr);
583 vaddr = page_address(page) + offset;
593 * Make an area consistent for devices.
594 * Note: Drivers should NOT use this function directly, as it will break
595 * platforms with CONFIG_DMABOUNCE.
596 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
598 static void __dma_page_cpu_to_dev(struct page *page, unsigned long off,
599 size_t size, enum dma_data_direction dir)
603 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
605 paddr = page_to_phys(page) + off;
606 if (dir == DMA_FROM_DEVICE) {
607 outer_inv_range(paddr, paddr + size);
609 outer_clean_range(paddr, paddr + size);
611 /* FIXME: non-speculating: flush on bidirectional mappings? */
614 static void __dma_page_dev_to_cpu(struct page *page, unsigned long off,
615 size_t size, enum dma_data_direction dir)
617 unsigned long paddr = page_to_phys(page) + off;
619 /* FIXME: non-speculating: not required */
620 /* don't bother invalidating if DMA to device */
621 if (dir != DMA_TO_DEVICE)
622 outer_inv_range(paddr, paddr + size);
624 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
627 * Mark the D-cache clean for this page to avoid extra flushing.
629 if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
630 set_bit(PG_dcache_clean, &page->flags);
634 * arm_dma_map_sg - map a set of SG buffers for streaming mode DMA
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
638 * @dir: DMA transfer direction
640 * Map a set of buffers described by scatterlist in streaming mode for DMA.
641 * This is the scatter-gather version of the dma_map_single interface.
642 * Here the scatter gather list elements are each tagged with the
643 * appropriate dma address and length. They are obtained via
644 * sg_dma_{address,length}.
646 * Device ownership issues as mentioned for dma_map_single are the same
649 int arm_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
650 enum dma_data_direction dir, struct dma_attrs *attrs)
652 struct dma_map_ops *ops = get_dma_ops(dev);
653 struct scatterlist *s;
656 for_each_sg(sg, s, nents, i) {
657 s->dma_address = ops->map_page(dev, sg_page(s), s->offset,
658 s->length, dir, attrs);
659 if (dma_mapping_error(dev, s->dma_address))
665 for_each_sg(sg, s, i, j)
666 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
671 * arm_dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
672 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
673 * @sg: list of buffers
674 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
675 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
677 * Unmap a set of streaming mode DMA translations. Again, CPU access
678 * rules concerning calls here are the same as for dma_unmap_single().
680 void arm_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
681 enum dma_data_direction dir, struct dma_attrs *attrs)
683 struct dma_map_ops *ops = get_dma_ops(dev);
684 struct scatterlist *s;
688 for_each_sg(sg, s, nents, i)
689 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
693 * arm_dma_sync_sg_for_cpu
694 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
695 * @sg: list of buffers
696 * @nents: number of buffers to map (returned from dma_map_sg)
697 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
699 void arm_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
700 int nents, enum dma_data_direction dir)
702 struct dma_map_ops *ops = get_dma_ops(dev);
703 struct scatterlist *s;
706 for_each_sg(sg, s, nents, i)
707 ops->sync_single_for_cpu(dev, sg_dma_address(s), s->length,
712 * arm_dma_sync_sg_for_device
713 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
714 * @sg: list of buffers
715 * @nents: number of buffers to map (returned from dma_map_sg)
716 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
718 void arm_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
719 int nents, enum dma_data_direction dir)
721 struct dma_map_ops *ops = get_dma_ops(dev);
722 struct scatterlist *s;
725 for_each_sg(sg, s, nents, i)
726 ops->sync_single_for_device(dev, sg_dma_address(s), s->length,
731 * Return whether the given device DMA address mask can be supported
732 * properly. For example, if your device can only drive the low 24-bits
733 * during bus mastering, then you would pass 0x00ffffff as the mask
736 int dma_supported(struct device *dev, u64 mask)
738 if (mask < (u64)arm_dma_limit)
742 EXPORT_SYMBOL(dma_supported);
744 static int arm_dma_set_mask(struct device *dev, u64 dma_mask)
746 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
749 *dev->dma_mask = dma_mask;
754 #define PREALLOC_DMA_DEBUG_ENTRIES 4096
756 static int __init dma_debug_do_init(void)
759 arm_vmregion_create_proc("dma-mappings", &consistent_head);
761 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
764 fs_initcall(dma_debug_do_init);