spi-topcliff-pch: supports a spi mode setup and bit order setup by IO control
[zen-stable.git] / arch / arm / mm / dma-mapping.c
blob1aa664a1999fce45c2548726b50b6fa924608ec2
1 /*
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>
13 #include <linux/mm.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>
30 #include "mm.h"
32 static u64 get_coherent_dma_mask(struct device *dev)
34 u64 mask = (u64)arm_dma_limit;
36 if (dev) {
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.
43 if (mask == 0) {
44 dev_warn(dev, "coherent DMA mask is unset\n");
45 return 0;
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);
52 return 0;
56 return mask;
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;
67 void *ptr;
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",
74 size, mask);
75 return NULL;
77 #endif
79 if (!mask)
80 return NULL;
82 if (mask < 0xffffffffULL)
83 gfp |= GFP_DMA;
85 page = alloc_pages(gfp, order);
86 if (!page)
87 return NULL;
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++)
94 __free_page(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);
105 return page;
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);
115 while (page < e) {
116 __free_page(page);
117 page++;
121 #ifdef CONFIG_MMU
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
157 #endif
160 * Initialise the consistent memory allocation.
162 static int __init consistent_init(void)
164 int ret = 0;
165 pgd_t *pgd;
166 pud_t *pud;
167 pmd_t *pmd;
168 pte_t *pte;
169 int i = 0;
170 unsigned long base = consistent_base;
171 unsigned long num_ptes = (CONSISTENT_END - base) >> PMD_SHIFT;
173 consistent_pte = kmalloc(num_ptes * sizeof(pte_t), GFP_KERNEL);
174 if (!consistent_pte) {
175 pr_err("%s: no memory\n", __func__);
176 return -ENOMEM;
179 pr_debug("DMA memory: 0x%08lx - 0x%08lx:\n", base, CONSISTENT_END);
180 consistent_head.vm_start = base;
182 do {
183 pgd = pgd_offset(&init_mm, base);
185 pud = pud_alloc(&init_mm, pgd, base);
186 if (!pud) {
187 printk(KERN_ERR "%s: no pud tables\n", __func__);
188 ret = -ENOMEM;
189 break;
192 pmd = pmd_alloc(&init_mm, pud, base);
193 if (!pmd) {
194 printk(KERN_ERR "%s: no pmd tables\n", __func__);
195 ret = -ENOMEM;
196 break;
198 WARN_ON(!pmd_none(*pmd));
200 pte = pte_alloc_kernel(pmd, base);
201 if (!pte) {
202 printk(KERN_ERR "%s: no pte tables\n", __func__);
203 ret = -ENOMEM;
204 break;
207 consistent_pte[i++] = pte;
208 base += PMD_SIZE;
209 } while (base < CONSISTENT_END);
211 return ret;
214 core_initcall(consistent_init);
216 static void *
217 __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot)
219 struct arm_vmregion *c;
220 size_t align;
221 int bit;
223 if (!consistent_pte) {
224 printk(KERN_ERR "%s: not initialised\n", __func__);
225 dump_stack();
226 return NULL;
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.
235 bit = fls(size - 1);
236 if (bit > SECTION_SHIFT)
237 bit = SECTION_SHIFT;
238 align = 1 << bit;
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));
245 if (c) {
246 pte_t *pte;
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;
251 c->vm_pages = page;
253 do {
254 BUG_ON(!pte_none(*pte));
256 set_pte_ext(pte, mk_pte(page, prot), 0);
257 page++;
258 pte++;
259 off++;
260 if (off >= PTRS_PER_PTE) {
261 off = 0;
262 pte = consistent_pte[++idx];
264 } while (size -= PAGE_SIZE);
266 dsb();
268 return (void *)c->vm_start;
270 return NULL;
273 static void __dma_free_remap(void *cpu_addr, size_t size)
275 struct arm_vmregion *c;
276 unsigned long addr;
277 pte_t *ptep;
278 int idx;
279 u32 off;
281 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
282 if (!c) {
283 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
284 __func__, cpu_addr);
285 dump_stack();
286 return;
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);
292 dump_stack();
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;
299 addr = c->vm_start;
300 do {
301 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
303 ptep++;
304 addr += PAGE_SIZE;
305 off++;
306 if (off >= PTRS_PER_PTE) {
307 off = 0;
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",
313 __func__);
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 */
328 static void *
329 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
330 pgprot_t prot)
332 struct page *page;
333 void *addr;
336 * Following is a work-around (a.k.a. hack) to prevent pages
337 * with __GFP_COMP being passed to split_page() which cannot
338 * handle them. The real problem is that this flag probably
339 * should be 0 on ARM as it is not supported on this
340 * platform; see CONFIG_HUGETLBFS.
342 gfp &= ~(__GFP_COMP);
344 *handle = ~0;
345 size = PAGE_ALIGN(size);
347 page = __dma_alloc_buffer(dev, size, gfp);
348 if (!page)
349 return NULL;
351 if (!arch_is_coherent())
352 addr = __dma_alloc_remap(page, size, gfp, prot);
353 else
354 addr = page_address(page);
356 if (addr)
357 *handle = pfn_to_dma(dev, page_to_pfn(page));
358 else
359 __dma_free_buffer(page, size);
361 return addr;
365 * Allocate DMA-coherent memory space and return both the kernel remapped
366 * virtual and bus address for that space.
368 void *
369 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
371 void *memory;
373 if (dma_alloc_from_coherent(dev, size, handle, &memory))
374 return memory;
376 return __dma_alloc(dev, size, handle, gfp,
377 pgprot_dmacoherent(pgprot_kernel));
379 EXPORT_SYMBOL(dma_alloc_coherent);
382 * Allocate a writecombining region, in much the same way as
383 * dma_alloc_coherent above.
385 void *
386 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
388 return __dma_alloc(dev, size, handle, gfp,
389 pgprot_writecombine(pgprot_kernel));
391 EXPORT_SYMBOL(dma_alloc_writecombine);
393 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
394 void *cpu_addr, dma_addr_t dma_addr, size_t size)
396 int ret = -ENXIO;
397 #ifdef CONFIG_MMU
398 unsigned long user_size, kern_size;
399 struct arm_vmregion *c;
401 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
403 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
404 if (c) {
405 unsigned long off = vma->vm_pgoff;
407 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
409 if (off < kern_size &&
410 user_size <= (kern_size - off)) {
411 ret = remap_pfn_range(vma, vma->vm_start,
412 page_to_pfn(c->vm_pages) + off,
413 user_size << PAGE_SHIFT,
414 vma->vm_page_prot);
417 #endif /* CONFIG_MMU */
419 return ret;
422 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
423 void *cpu_addr, dma_addr_t dma_addr, size_t size)
425 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
426 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
428 EXPORT_SYMBOL(dma_mmap_coherent);
430 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
431 void *cpu_addr, dma_addr_t dma_addr, size_t size)
433 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
434 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
436 EXPORT_SYMBOL(dma_mmap_writecombine);
439 * free a page as defined by the above mapping.
440 * Must not be called with IRQs disabled.
442 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
444 WARN_ON(irqs_disabled());
446 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
447 return;
449 size = PAGE_ALIGN(size);
451 if (!arch_is_coherent())
452 __dma_free_remap(cpu_addr, size);
454 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev, handle)), size);
456 EXPORT_SYMBOL(dma_free_coherent);
459 * Make an area consistent for devices.
460 * Note: Drivers should NOT use this function directly, as it will break
461 * platforms with CONFIG_DMABOUNCE.
462 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
464 void ___dma_single_cpu_to_dev(const void *kaddr, size_t size,
465 enum dma_data_direction dir)
467 unsigned long paddr;
469 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
471 dmac_map_area(kaddr, size, dir);
473 paddr = __pa(kaddr);
474 if (dir == DMA_FROM_DEVICE) {
475 outer_inv_range(paddr, paddr + size);
476 } else {
477 outer_clean_range(paddr, paddr + size);
479 /* FIXME: non-speculating: flush on bidirectional mappings? */
481 EXPORT_SYMBOL(___dma_single_cpu_to_dev);
483 void ___dma_single_dev_to_cpu(const void *kaddr, size_t size,
484 enum dma_data_direction dir)
486 BUG_ON(!virt_addr_valid(kaddr) || !virt_addr_valid(kaddr + size - 1));
488 /* FIXME: non-speculating: not required */
489 /* don't bother invalidating if DMA to device */
490 if (dir != DMA_TO_DEVICE) {
491 unsigned long paddr = __pa(kaddr);
492 outer_inv_range(paddr, paddr + size);
495 dmac_unmap_area(kaddr, size, dir);
497 EXPORT_SYMBOL(___dma_single_dev_to_cpu);
499 static void dma_cache_maint_page(struct page *page, unsigned long offset,
500 size_t size, enum dma_data_direction dir,
501 void (*op)(const void *, size_t, int))
504 * A single sg entry may refer to multiple physically contiguous
505 * pages. But we still need to process highmem pages individually.
506 * If highmem is not configured then the bulk of this loop gets
507 * optimized out.
509 size_t left = size;
510 do {
511 size_t len = left;
512 void *vaddr;
514 if (PageHighMem(page)) {
515 if (len + offset > PAGE_SIZE) {
516 if (offset >= PAGE_SIZE) {
517 page += offset / PAGE_SIZE;
518 offset %= PAGE_SIZE;
520 len = PAGE_SIZE - offset;
522 vaddr = kmap_high_get(page);
523 if (vaddr) {
524 vaddr += offset;
525 op(vaddr, len, dir);
526 kunmap_high(page);
527 } else if (cache_is_vipt()) {
528 /* unmapped pages might still be cached */
529 vaddr = kmap_atomic(page);
530 op(vaddr + offset, len, dir);
531 kunmap_atomic(vaddr);
533 } else {
534 vaddr = page_address(page) + offset;
535 op(vaddr, len, dir);
537 offset = 0;
538 page++;
539 left -= len;
540 } while (left);
543 void ___dma_page_cpu_to_dev(struct page *page, unsigned long off,
544 size_t size, enum dma_data_direction dir)
546 unsigned long paddr;
548 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
550 paddr = page_to_phys(page) + off;
551 if (dir == DMA_FROM_DEVICE) {
552 outer_inv_range(paddr, paddr + size);
553 } else {
554 outer_clean_range(paddr, paddr + size);
556 /* FIXME: non-speculating: flush on bidirectional mappings? */
558 EXPORT_SYMBOL(___dma_page_cpu_to_dev);
560 void ___dma_page_dev_to_cpu(struct page *page, unsigned long off,
561 size_t size, enum dma_data_direction dir)
563 unsigned long paddr = page_to_phys(page) + off;
565 /* FIXME: non-speculating: not required */
566 /* don't bother invalidating if DMA to device */
567 if (dir != DMA_TO_DEVICE)
568 outer_inv_range(paddr, paddr + size);
570 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
573 * Mark the D-cache clean for this page to avoid extra flushing.
575 if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
576 set_bit(PG_dcache_clean, &page->flags);
578 EXPORT_SYMBOL(___dma_page_dev_to_cpu);
581 * dma_map_sg - map a set of SG buffers for streaming mode DMA
582 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
583 * @sg: list of buffers
584 * @nents: number of buffers to map
585 * @dir: DMA transfer direction
587 * Map a set of buffers described by scatterlist in streaming mode for DMA.
588 * This is the scatter-gather version of the dma_map_single interface.
589 * Here the scatter gather list elements are each tagged with the
590 * appropriate dma address and length. They are obtained via
591 * sg_dma_{address,length}.
593 * Device ownership issues as mentioned for dma_map_single are the same
594 * here.
596 int dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
597 enum dma_data_direction dir)
599 struct scatterlist *s;
600 int i, j;
602 BUG_ON(!valid_dma_direction(dir));
604 for_each_sg(sg, s, nents, i) {
605 s->dma_address = __dma_map_page(dev, sg_page(s), s->offset,
606 s->length, dir);
607 if (dma_mapping_error(dev, s->dma_address))
608 goto bad_mapping;
610 debug_dma_map_sg(dev, sg, nents, nents, dir);
611 return nents;
613 bad_mapping:
614 for_each_sg(sg, s, i, j)
615 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
616 return 0;
618 EXPORT_SYMBOL(dma_map_sg);
621 * dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
622 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
623 * @sg: list of buffers
624 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
625 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
627 * Unmap a set of streaming mode DMA translations. Again, CPU access
628 * rules concerning calls here are the same as for dma_unmap_single().
630 void dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
631 enum dma_data_direction dir)
633 struct scatterlist *s;
634 int i;
636 debug_dma_unmap_sg(dev, sg, nents, dir);
638 for_each_sg(sg, s, nents, i)
639 __dma_unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir);
641 EXPORT_SYMBOL(dma_unmap_sg);
644 * dma_sync_sg_for_cpu
645 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
646 * @sg: list of buffers
647 * @nents: number of buffers to map (returned from dma_map_sg)
648 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
650 void dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
651 int nents, enum dma_data_direction dir)
653 struct scatterlist *s;
654 int i;
656 for_each_sg(sg, s, nents, i) {
657 if (!dmabounce_sync_for_cpu(dev, sg_dma_address(s), 0,
658 sg_dma_len(s), dir))
659 continue;
661 __dma_page_dev_to_cpu(sg_page(s), s->offset,
662 s->length, dir);
665 debug_dma_sync_sg_for_cpu(dev, sg, nents, dir);
667 EXPORT_SYMBOL(dma_sync_sg_for_cpu);
670 * dma_sync_sg_for_device
671 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
672 * @sg: list of buffers
673 * @nents: number of buffers to map (returned from dma_map_sg)
674 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
676 void dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
677 int nents, enum dma_data_direction dir)
679 struct scatterlist *s;
680 int i;
682 for_each_sg(sg, s, nents, i) {
683 if (!dmabounce_sync_for_device(dev, sg_dma_address(s), 0,
684 sg_dma_len(s), dir))
685 continue;
687 __dma_page_cpu_to_dev(sg_page(s), s->offset,
688 s->length, dir);
691 debug_dma_sync_sg_for_device(dev, sg, nents, dir);
693 EXPORT_SYMBOL(dma_sync_sg_for_device);
696 * Return whether the given device DMA address mask can be supported
697 * properly. For example, if your device can only drive the low 24-bits
698 * during bus mastering, then you would pass 0x00ffffff as the mask
699 * to this function.
701 int dma_supported(struct device *dev, u64 mask)
703 if (mask < (u64)arm_dma_limit)
704 return 0;
705 return 1;
707 EXPORT_SYMBOL(dma_supported);
709 int dma_set_mask(struct device *dev, u64 dma_mask)
711 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
712 return -EIO;
714 #ifndef CONFIG_DMABOUNCE
715 *dev->dma_mask = dma_mask;
716 #endif
718 return 0;
720 EXPORT_SYMBOL(dma_set_mask);
722 #define PREALLOC_DMA_DEBUG_ENTRIES 4096
724 static int __init dma_debug_do_init(void)
726 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
727 return 0;
729 fs_initcall(dma_debug_do_init);