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/slab.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 /* Sanity check size */
28 #if (CONSISTENT_DMA_SIZE % SZ_2M)
29 #error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
32 #define CONSISTENT_END (0xffe00000)
33 #define CONSISTENT_BASE (CONSISTENT_END - CONSISTENT_DMA_SIZE)
35 #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
36 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PGDIR_SHIFT)
37 #define NUM_CONSISTENT_PTES (CONSISTENT_DMA_SIZE >> PGDIR_SHIFT)
41 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
43 static pte_t *consistent_pte[NUM_CONSISTENT_PTES];
44 static DEFINE_SPINLOCK(consistent_lock);
47 * VM region handling support.
49 * This should become something generic, handling VM region allocations for
50 * vmalloc and similar (ioremap, module space, etc).
52 * I envisage vmalloc()'s supporting vm_struct becoming:
55 * struct vm_region region;
56 * unsigned long flags;
57 * struct page **pages;
58 * unsigned int nr_pages;
59 * unsigned long phys_addr;
62 * get_vm_area() would then call vm_region_alloc with an appropriate
63 * struct vm_region head (eg):
65 * struct vm_region vmalloc_head = {
66 * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list),
67 * .vm_start = VMALLOC_START,
68 * .vm_end = VMALLOC_END,
71 * However, vmalloc_head.vm_start is variable (typically, it is dependent on
72 * the amount of RAM found at boot time.) I would imagine that get_vm_area()
73 * would have to initialise this each time prior to calling vm_region_alloc().
75 struct arm_vm_region {
76 struct list_head vm_list;
77 unsigned long vm_start;
79 struct page *vm_pages;
83 static struct arm_vm_region consistent_head = {
84 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
85 .vm_start = CONSISTENT_BASE,
86 .vm_end = CONSISTENT_END,
89 static struct arm_vm_region *
90 arm_vm_region_alloc(struct arm_vm_region *head, size_t size, gfp_t gfp)
92 unsigned long addr = head->vm_start, end = head->vm_end - size;
94 struct arm_vm_region *c, *new;
96 new = kmalloc(sizeof(struct arm_vm_region), gfp);
100 spin_lock_irqsave(&consistent_lock, flags);
102 list_for_each_entry(c, &head->vm_list, vm_list) {
103 if ((addr + size) < addr)
105 if ((addr + size) <= c->vm_start)
114 * Insert this entry _before_ the one we found.
116 list_add_tail(&new->vm_list, &c->vm_list);
117 new->vm_start = addr;
118 new->vm_end = addr + size;
121 spin_unlock_irqrestore(&consistent_lock, flags);
125 spin_unlock_irqrestore(&consistent_lock, flags);
131 static struct arm_vm_region *arm_vm_region_find(struct arm_vm_region *head, unsigned long addr)
133 struct arm_vm_region *c;
135 list_for_each_entry(c, &head->vm_list, vm_list) {
136 if (c->vm_active && c->vm_start == addr)
144 #ifdef CONFIG_HUGETLB_PAGE
145 #error ARM Coherent DMA allocator does not (yet) support huge TLB
149 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
153 struct arm_vm_region *c;
155 u64 mask = ISA_DMA_THRESHOLD, limit;
157 if (!consistent_pte[0]) {
158 printk(KERN_ERR "%s: not initialised\n", __func__);
164 mask = dev->coherent_dma_mask;
167 * Sanity check the DMA mask - it must be non-zero, and
168 * must be able to be satisfied by a DMA allocation.
171 dev_warn(dev, "coherent DMA mask is unset\n");
175 if ((~mask) & ISA_DMA_THRESHOLD) {
176 dev_warn(dev, "coherent DMA mask %#llx is smaller "
177 "than system GFP_DMA mask %#llx\n",
178 mask, (unsigned long long)ISA_DMA_THRESHOLD);
184 * Sanity check the allocation size.
186 size = PAGE_ALIGN(size);
187 limit = (mask + 1) & ~mask;
188 if ((limit && size >= limit) ||
189 size >= (CONSISTENT_END - CONSISTENT_BASE)) {
190 printk(KERN_WARNING "coherent allocation too big "
191 "(requested %#x mask %#llx)\n", size, mask);
195 order = get_order(size);
197 if (mask != 0xffffffff)
200 page = alloc_pages(gfp, order);
205 * Invalidate any data that might be lurking in the
206 * kernel direct-mapped region for device DMA.
209 void *ptr = page_address(page);
210 memset(ptr, 0, size);
211 dmac_flush_range(ptr, ptr + size);
212 outer_flush_range(__pa(ptr), __pa(ptr) + size);
216 * Allocate a virtual address in the consistent mapping region.
218 c = arm_vm_region_alloc(&consistent_head, size,
219 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
222 struct page *end = page + (1 << order);
223 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
224 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
226 pte = consistent_pte[idx] + off;
229 split_page(page, order);
232 * Set the "dma handle"
234 *handle = page_to_dma(dev, page);
237 BUG_ON(!pte_none(*pte));
240 * x86 does not mark the pages reserved...
242 SetPageReserved(page);
243 set_pte_ext(pte, mk_pte(page, prot), 0);
247 if (off >= PTRS_PER_PTE) {
249 pte = consistent_pte[++idx];
251 } while (size -= PAGE_SIZE);
254 * Free the otherwise unused pages.
261 return (void *)c->vm_start;
265 __free_pages(page, order);
272 * Allocate DMA-coherent memory space and return both the kernel remapped
273 * virtual and bus address for that space.
276 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
280 if (dma_alloc_from_coherent(dev, size, handle, &memory))
283 if (arch_is_coherent()) {
286 virt = kmalloc(size, gfp);
289 *handle = virt_to_dma(dev, virt);
294 return __dma_alloc(dev, size, handle, gfp,
295 pgprot_noncached(pgprot_kernel));
297 EXPORT_SYMBOL(dma_alloc_coherent);
300 * Allocate a writecombining region, in much the same way as
301 * dma_alloc_coherent above.
304 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
306 return __dma_alloc(dev, size, handle, gfp,
307 pgprot_writecombine(pgprot_kernel));
309 EXPORT_SYMBOL(dma_alloc_writecombine);
311 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
312 void *cpu_addr, dma_addr_t dma_addr, size_t size)
314 unsigned long flags, user_size, kern_size;
315 struct arm_vm_region *c;
318 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
320 spin_lock_irqsave(&consistent_lock, flags);
321 c = arm_vm_region_find(&consistent_head, (unsigned long)cpu_addr);
322 spin_unlock_irqrestore(&consistent_lock, flags);
325 unsigned long off = vma->vm_pgoff;
327 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
329 if (off < kern_size &&
330 user_size <= (kern_size - off)) {
331 ret = remap_pfn_range(vma, vma->vm_start,
332 page_to_pfn(c->vm_pages) + off,
333 user_size << PAGE_SHIFT,
341 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
342 void *cpu_addr, dma_addr_t dma_addr, size_t size)
344 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
345 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
347 EXPORT_SYMBOL(dma_mmap_coherent);
349 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
350 void *cpu_addr, dma_addr_t dma_addr, size_t size)
352 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
353 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
355 EXPORT_SYMBOL(dma_mmap_writecombine);
358 * free a page as defined by the above mapping.
359 * Must not be called with IRQs disabled.
361 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
363 struct arm_vm_region *c;
364 unsigned long flags, addr;
369 WARN_ON(irqs_disabled());
371 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
374 if (arch_is_coherent()) {
379 size = PAGE_ALIGN(size);
381 spin_lock_irqsave(&consistent_lock, flags);
382 c = arm_vm_region_find(&consistent_head, (unsigned long)cpu_addr);
387 spin_unlock_irqrestore(&consistent_lock, flags);
389 if ((c->vm_end - c->vm_start) != size) {
390 printk(KERN_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);
407 if (off >= PTRS_PER_PTE) {
409 ptep = consistent_pte[++idx];
412 if (!pte_none(pte) && pte_present(pte)) {
415 if (pfn_valid(pfn)) {
416 struct page *page = pfn_to_page(pfn);
419 * x86 does not mark the pages reserved...
421 ClearPageReserved(page);
428 printk(KERN_CRIT "%s: bad page in kernel page table\n",
430 } while (size -= PAGE_SIZE);
432 flush_tlb_kernel_range(c->vm_start, c->vm_end);
434 spin_lock_irqsave(&consistent_lock, flags);
435 list_del(&c->vm_list);
436 spin_unlock_irqrestore(&consistent_lock, flags);
442 spin_unlock_irqrestore(&consistent_lock, flags);
443 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
447 EXPORT_SYMBOL(dma_free_coherent);
450 * Initialise the consistent memory allocation.
452 static int __init consistent_init(void)
458 u32 base = CONSISTENT_BASE;
461 pgd = pgd_offset(&init_mm, base);
462 pmd = pmd_alloc(&init_mm, pgd, base);
464 printk(KERN_ERR "%s: no pmd tables\n", __func__);
468 WARN_ON(!pmd_none(*pmd));
470 pte = pte_alloc_kernel(pmd, base);
472 printk(KERN_ERR "%s: no pte tables\n", __func__);
477 consistent_pte[i++] = pte;
478 base += (1 << PGDIR_SHIFT);
479 } while (base < CONSISTENT_END);
484 core_initcall(consistent_init);
487 * Make an area consistent for devices.
488 * Note: Drivers should NOT use this function directly, as it will break
489 * platforms with CONFIG_DMABOUNCE.
490 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
492 void dma_cache_maint(const void *start, size_t size, int direction)
494 void (*inner_op)(const void *, const void *);
495 void (*outer_op)(unsigned long, unsigned long);
497 BUG_ON(!virt_addr_valid(start) || !virt_addr_valid(start + size - 1));
500 case DMA_FROM_DEVICE: /* invalidate only */
501 inner_op = dmac_inv_range;
502 outer_op = outer_inv_range;
504 case DMA_TO_DEVICE: /* writeback only */
505 inner_op = dmac_clean_range;
506 outer_op = outer_clean_range;
508 case DMA_BIDIRECTIONAL: /* writeback and invalidate */
509 inner_op = dmac_flush_range;
510 outer_op = outer_flush_range;
516 inner_op(start, start + size);
517 outer_op(__pa(start), __pa(start) + size);
519 EXPORT_SYMBOL(dma_cache_maint);
521 static void dma_cache_maint_contiguous(struct page *page, unsigned long offset,
522 size_t size, int direction)
526 void (*inner_op)(const void *, const void *);
527 void (*outer_op)(unsigned long, unsigned long);
530 case DMA_FROM_DEVICE: /* invalidate only */
531 inner_op = dmac_inv_range;
532 outer_op = outer_inv_range;
534 case DMA_TO_DEVICE: /* writeback only */
535 inner_op = dmac_clean_range;
536 outer_op = outer_clean_range;
538 case DMA_BIDIRECTIONAL: /* writeback and invalidate */
539 inner_op = dmac_flush_range;
540 outer_op = outer_flush_range;
546 if (!PageHighMem(page)) {
547 vaddr = page_address(page) + offset;
548 inner_op(vaddr, vaddr + size);
550 vaddr = kmap_high_get(page);
553 inner_op(vaddr, vaddr + size);
558 paddr = page_to_phys(page) + offset;
559 outer_op(paddr, paddr + size);
562 void dma_cache_maint_page(struct page *page, unsigned long offset,
563 size_t size, int dir)
566 * A single sg entry may refer to multiple physically contiguous
567 * pages. But we still need to process highmem pages individually.
568 * If highmem is not configured then the bulk of this loop gets
574 if (PageHighMem(page) && len + offset > PAGE_SIZE) {
575 if (offset >= PAGE_SIZE) {
576 page += offset / PAGE_SIZE;
579 len = PAGE_SIZE - offset;
581 dma_cache_maint_contiguous(page, offset, len, dir);
587 EXPORT_SYMBOL(dma_cache_maint_page);
590 * dma_map_sg - map a set of SG buffers for streaming mode DMA
591 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
592 * @sg: list of buffers
593 * @nents: number of buffers to map
594 * @dir: DMA transfer direction
596 * Map a set of buffers described by scatterlist in streaming mode for DMA.
597 * This is the scatter-gather version of the dma_map_single interface.
598 * Here the scatter gather list elements are each tagged with the
599 * appropriate dma address and length. They are obtained via
600 * sg_dma_{address,length}.
602 * Device ownership issues as mentioned for dma_map_single are the same
605 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
606 enum dma_data_direction dir)
608 struct scatterlist *s;
611 for_each_sg(sg, s, nents, i) {
612 s->dma_address = dma_map_page(dev, sg_page(s), s->offset,
614 if (dma_mapping_error(dev, s->dma_address))
620 for_each_sg(sg, s, i, j)
621 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
624 EXPORT_SYMBOL(dma_map_sg);
627 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
628 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
629 * @sg: list of buffers
630 * @nents: number of buffers to unmap (returned from dma_map_sg)
631 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
633 * Unmap a set of streaming mode DMA translations. Again, CPU access
634 * rules concerning calls here are the same as for dma_unmap_single().
636 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
637 enum dma_data_direction dir)
639 struct scatterlist *s;
642 for_each_sg(sg, s, nents, i)
643 dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
645 EXPORT_SYMBOL(dma_unmap_sg);
648 * dma_sync_sg_for_cpu
649 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
650 * @sg: list of buffers
651 * @nents: number of buffers to map (returned from dma_map_sg)
652 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
654 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
655 int nents, enum dma_data_direction dir)
657 struct scatterlist *s;
660 for_each_sg(sg, s, nents, i) {
661 dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
665 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
668 * dma_sync_sg_for_device
669 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
670 * @sg: list of buffers
671 * @nents: number of buffers to map (returned from dma_map_sg)
672 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
674 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
675 int nents, enum dma_data_direction dir)
677 struct scatterlist *s;
680 for_each_sg(sg, s, nents, i) {
681 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
685 if (!arch_is_coherent())
686 dma_cache_maint_page(sg_page(s), s->offset,
690 EXPORT_SYMBOL(dma_sync_sg_for_device);