blob: dddb406d07630440934ea515d7db5159bf7acafd [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Russell King0ddbccd2008-09-25 15:59:19 +01002 * linux/arch/arm/mm/dma-mapping.c
Linus Torvalds1da177e2005-04-16 15:20:36 -07003 *
4 * Copyright (C) 2000-2004 Russell King
5 *
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.
9 *
10 * DMA uncached mapping support.
11 */
12#include <linux/module.h>
13#include <linux/mm.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090014#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070015#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>
Nicolas Pitre39af22a2010-12-15 15:14:45 -050020#include <linux/highmem.h>
Jon Medhurst99d17172011-08-02 17:28:27 +010021#include <linux/slab.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070022
Lennert Buytenhek23759dc2006-04-02 00:07:39 +010023#include <asm/memory.h>
Nicolas Pitre43377452009-03-12 22:52:09 -040024#include <asm/highmem.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070025#include <asm/cacheflush.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070026#include <asm/tlbflush.h>
Kevin Hilman37134cd2006-01-12 16:12:21 +000027#include <asm/sizes.h>
Jon Medhurst99d17172011-08-02 17:28:27 +010028#include <asm/mach/arch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070029
Russell King022ae532011-07-08 21:26:59 +010030#include "mm.h"
31
Marek Szyprowski15237e12012-02-10 19:55:20 +010032/*
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.
37 *
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.
42 *
Marek Szyprowski15237e12012-02-10 19:55:20 +010043 */
Marek Szyprowski51fde3492012-02-10 19:55:20 +010044static void __dma_page_cpu_to_dev(struct page *, unsigned long,
Marek Szyprowski15237e12012-02-10 19:55:20 +010045 size_t, enum dma_data_direction);
Marek Szyprowski51fde3492012-02-10 19:55:20 +010046static void __dma_page_dev_to_cpu(struct page *, unsigned long,
Marek Szyprowski15237e12012-02-10 19:55:20 +010047 size_t, enum dma_data_direction);
48
Marek Szyprowski2dc6a012012-02-10 19:55:20 +010049/**
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
56 *
57 * Ensure that any data held in the cache is appropriately discarded
58 * or written back.
59 *
60 * The device owns this memory once this call has completed. The CPU
61 * can regain ownership by calling dma_unmap_page().
62 */
Marek Szyprowski51fde3492012-02-10 19:55:20 +010063static dma_addr_t arm_dma_map_page(struct device *dev, struct page *page,
Marek Szyprowski2dc6a012012-02-10 19:55:20 +010064 unsigned long offset, size_t size, enum dma_data_direction dir,
65 struct dma_attrs *attrs)
66{
Marek Szyprowski51fde3492012-02-10 19:55:20 +010067 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;
Marek Szyprowski2dc6a012012-02-10 19:55:20 +010070}
71
72/**
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)
78 *
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.
82 *
83 * After this call, reads by the CPU to the buffer are guaranteed to see
84 * whatever the device wrote there.
85 */
Marek Szyprowski51fde3492012-02-10 19:55:20 +010086static void arm_dma_unmap_page(struct device *dev, dma_addr_t handle,
Marek Szyprowski2dc6a012012-02-10 19:55:20 +010087 size_t size, enum dma_data_direction dir,
88 struct dma_attrs *attrs)
89{
Marek Szyprowski51fde3492012-02-10 19:55:20 +010090 if (!arch_is_coherent())
91 __dma_page_dev_to_cpu(pfn_to_page(dma_to_pfn(dev, handle)),
92 handle & ~PAGE_MASK, size, dir);
Marek Szyprowski2dc6a012012-02-10 19:55:20 +010093}
94
Marek Szyprowski51fde3492012-02-10 19:55:20 +010095static void arm_dma_sync_single_for_cpu(struct device *dev,
Marek Szyprowski2dc6a012012-02-10 19:55:20 +010096 dma_addr_t handle, size_t size, enum dma_data_direction dir)
97{
98 unsigned int offset = handle & (PAGE_SIZE - 1);
99 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
Marek Szyprowski51fde3492012-02-10 19:55:20 +0100100 if (!arch_is_coherent())
101 __dma_page_dev_to_cpu(page, offset, size, dir);
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100102}
103
Marek Szyprowski51fde3492012-02-10 19:55:20 +0100104static void arm_dma_sync_single_for_device(struct device *dev,
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100105 dma_addr_t handle, size_t size, enum dma_data_direction dir)
106{
107 unsigned int offset = handle & (PAGE_SIZE - 1);
108 struct page *page = pfn_to_page(dma_to_pfn(dev, handle-offset));
Marek Szyprowski51fde3492012-02-10 19:55:20 +0100109 if (!arch_is_coherent())
110 __dma_page_cpu_to_dev(page, offset, size, dir);
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100111}
112
113static int arm_dma_set_mask(struct device *dev, u64 dma_mask);
114
115struct dma_map_ops arm_dma_ops = {
116 .map_page = arm_dma_map_page,
117 .unmap_page = arm_dma_unmap_page,
118 .map_sg = arm_dma_map_sg,
119 .unmap_sg = arm_dma_unmap_sg,
120 .sync_single_for_cpu = arm_dma_sync_single_for_cpu,
121 .sync_single_for_device = arm_dma_sync_single_for_device,
122 .sync_sg_for_cpu = arm_dma_sync_sg_for_cpu,
123 .sync_sg_for_device = arm_dma_sync_sg_for_device,
124 .set_dma_mask = arm_dma_set_mask,
125};
126EXPORT_SYMBOL(arm_dma_ops);
127
Catalin Marinasab6494f2009-07-24 12:35:02 +0100128static u64 get_coherent_dma_mask(struct device *dev)
129{
Russell King022ae532011-07-08 21:26:59 +0100130 u64 mask = (u64)arm_dma_limit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131
Catalin Marinasab6494f2009-07-24 12:35:02 +0100132 if (dev) {
133 mask = dev->coherent_dma_mask;
134
135 /*
136 * Sanity check the DMA mask - it must be non-zero, and
137 * must be able to be satisfied by a DMA allocation.
138 */
139 if (mask == 0) {
140 dev_warn(dev, "coherent DMA mask is unset\n");
141 return 0;
142 }
143
Russell King022ae532011-07-08 21:26:59 +0100144 if ((~mask) & (u64)arm_dma_limit) {
Catalin Marinasab6494f2009-07-24 12:35:02 +0100145 dev_warn(dev, "coherent DMA mask %#llx is smaller "
146 "than system GFP_DMA mask %#llx\n",
Russell King022ae532011-07-08 21:26:59 +0100147 mask, (u64)arm_dma_limit);
Catalin Marinasab6494f2009-07-24 12:35:02 +0100148 return 0;
149 }
150 }
151
152 return mask;
153}
154
Russell King7a9a32a2009-11-19 15:31:07 +0000155/*
156 * Allocate a DMA buffer for 'dev' of size 'size' using the
157 * specified gfp mask. Note that 'size' must be page aligned.
158 */
159static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp)
160{
161 unsigned long order = get_order(size);
162 struct page *page, *p, *e;
163 void *ptr;
164 u64 mask = get_coherent_dma_mask(dev);
165
166#ifdef CONFIG_DMA_API_DEBUG
167 u64 limit = (mask + 1) & ~mask;
168 if (limit && size >= limit) {
169 dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n",
170 size, mask);
171 return NULL;
172 }
173#endif
174
175 if (!mask)
176 return NULL;
177
178 if (mask < 0xffffffffULL)
179 gfp |= GFP_DMA;
180
181 page = alloc_pages(gfp, order);
182 if (!page)
183 return NULL;
184
185 /*
186 * Now split the huge page and free the excess pages
187 */
188 split_page(page, order);
189 for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++)
190 __free_page(p);
191
192 /*
193 * Ensure that the allocated pages are zeroed, and that any data
194 * lurking in the kernel direct-mapped region is invalidated.
195 */
196 ptr = page_address(page);
197 memset(ptr, 0, size);
198 dmac_flush_range(ptr, ptr + size);
199 outer_flush_range(__pa(ptr), __pa(ptr) + size);
200
201 return page;
202}
203
204/*
205 * Free a DMA buffer. 'size' must be page aligned.
206 */
207static void __dma_free_buffer(struct page *page, size_t size)
208{
209 struct page *e = page + (size >> PAGE_SHIFT);
210
211 while (page < e) {
212 __free_page(page);
213 page++;
214 }
215}
216
Catalin Marinasab6494f2009-07-24 12:35:02 +0100217#ifdef CONFIG_MMU
Catalin Marinasa5e9d382010-06-21 15:09:06 +0100218
Jon Medhurst99d17172011-08-02 17:28:27 +0100219#define CONSISTENT_OFFSET(x) (((unsigned long)(x) - consistent_base) >> PAGE_SHIFT)
Linus Torvalds1fdb24e2011-10-28 12:02:27 -0700220#define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - consistent_base) >> PMD_SHIFT)
Catalin Marinasa5e9d382010-06-21 15:09:06 +0100221
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222/*
Kevin Hilman37134cd2006-01-12 16:12:21 +0000223 * These are the page tables (2MB each) covering uncached, DMA consistent allocations
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224 */
Jon Medhurst99d17172011-08-02 17:28:27 +0100225static pte_t **consistent_pte;
226
Jon Medhurst99d17172011-08-02 17:28:27 +0100227#define DEFAULT_CONSISTENT_DMA_SIZE SZ_2M
Jon Medhurst99d17172011-08-02 17:28:27 +0100228
229unsigned long consistent_base = CONSISTENT_END - DEFAULT_CONSISTENT_DMA_SIZE;
230
231void __init init_consistent_dma_size(unsigned long size)
232{
233 unsigned long base = CONSISTENT_END - ALIGN(size, SZ_2M);
234
235 BUG_ON(consistent_pte); /* Check we're called before DMA region init */
236 BUG_ON(base < VMALLOC_END);
237
238 /* Grow region to accommodate specified size */
239 if (base < consistent_base)
240 consistent_base = base;
241}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700242
Russell King13ccf3a2009-11-19 15:07:04 +0000243#include "vmregion.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Russell King13ccf3a2009-11-19 15:07:04 +0000245static struct arm_vmregion_head consistent_head = {
246 .vm_lock = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248 .vm_end = CONSISTENT_END,
249};
250
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251#ifdef CONFIG_HUGETLB_PAGE
252#error ARM Coherent DMA allocator does not (yet) support huge TLB
253#endif
254
Russell King88c58f32009-11-19 16:46:02 +0000255/*
256 * Initialise the consistent memory allocation.
257 */
258static int __init consistent_init(void)
259{
260 int ret = 0;
261 pgd_t *pgd;
Russell King516295e2010-11-21 16:27:49 +0000262 pud_t *pud;
Russell King88c58f32009-11-19 16:46:02 +0000263 pmd_t *pmd;
264 pte_t *pte;
265 int i = 0;
Jon Medhurst99d17172011-08-02 17:28:27 +0100266 unsigned long base = consistent_base;
Catalin Marinas53cbcbc2011-11-17 13:11:21 +0100267 unsigned long num_ptes = (CONSISTENT_END - base) >> PMD_SHIFT;
Jon Medhurst99d17172011-08-02 17:28:27 +0100268
269 consistent_pte = kmalloc(num_ptes * sizeof(pte_t), GFP_KERNEL);
270 if (!consistent_pte) {
271 pr_err("%s: no memory\n", __func__);
272 return -ENOMEM;
273 }
274
275 pr_debug("DMA memory: 0x%08lx - 0x%08lx:\n", base, CONSISTENT_END);
276 consistent_head.vm_start = base;
Russell King88c58f32009-11-19 16:46:02 +0000277
278 do {
279 pgd = pgd_offset(&init_mm, base);
Russell King516295e2010-11-21 16:27:49 +0000280
281 pud = pud_alloc(&init_mm, pgd, base);
282 if (!pud) {
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100283 pr_err("%s: no pud tables\n", __func__);
Russell King516295e2010-11-21 16:27:49 +0000284 ret = -ENOMEM;
285 break;
286 }
287
288 pmd = pmd_alloc(&init_mm, pud, base);
Russell King88c58f32009-11-19 16:46:02 +0000289 if (!pmd) {
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100290 pr_err("%s: no pmd tables\n", __func__);
Russell King88c58f32009-11-19 16:46:02 +0000291 ret = -ENOMEM;
292 break;
293 }
294 WARN_ON(!pmd_none(*pmd));
295
296 pte = pte_alloc_kernel(pmd, base);
297 if (!pte) {
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100298 pr_err("%s: no pte tables\n", __func__);
Russell King88c58f32009-11-19 16:46:02 +0000299 ret = -ENOMEM;
300 break;
301 }
302
303 consistent_pte[i++] = pte;
Catalin Marinase73fc882011-08-23 14:07:23 +0100304 base += PMD_SIZE;
Russell King88c58f32009-11-19 16:46:02 +0000305 } while (base < CONSISTENT_END);
306
307 return ret;
308}
309
310core_initcall(consistent_init);
311
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312static void *
Russell King45cd5292012-01-12 23:08:07 +0000313__dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot,
314 const void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315{
Russell King13ccf3a2009-11-19 15:07:04 +0000316 struct arm_vmregion *c;
Russell King5bc23d32010-07-25 08:57:02 +0100317 size_t align;
318 int bit;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319
Jon Medhurst99d17172011-08-02 17:28:27 +0100320 if (!consistent_pte) {
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100321 pr_err("%s: not initialised\n", __func__);
Russell Kingebd7a842009-11-19 20:58:31 +0000322 dump_stack();
Russell Kingebd7a842009-11-19 20:58:31 +0000323 return NULL;
324 }
325
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 /*
Russell King5bc23d32010-07-25 08:57:02 +0100327 * Align the virtual region allocation - maximum alignment is
328 * a section size, minimum is a page size. This helps reduce
329 * fragmentation of the DMA space, and also prevents allocations
330 * smaller than a section from crossing a section boundary.
331 */
Russell Kingc947f692010-11-03 16:00:15 +0000332 bit = fls(size - 1);
Russell King5bc23d32010-07-25 08:57:02 +0100333 if (bit > SECTION_SHIFT)
334 bit = SECTION_SHIFT;
335 align = 1 << bit;
336
337 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 * Allocate a virtual address in the consistent mapping region.
339 */
Russell King5bc23d32010-07-25 08:57:02 +0100340 c = arm_vmregion_alloc(&consistent_head, align, size,
Russell King45cd5292012-01-12 23:08:07 +0000341 gfp & ~(__GFP_DMA | __GFP_HIGHMEM), caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 if (c) {
Kevin Hilman37134cd2006-01-12 16:12:21 +0000343 pte_t *pte;
Kevin Hilman37134cd2006-01-12 16:12:21 +0000344 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
345 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346
Kevin Hilman37134cd2006-01-12 16:12:21 +0000347 pte = consistent_pte[idx] + off;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 c->vm_pages = page;
349
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 do {
351 BUG_ON(!pte_none(*pte));
352
Russell Kingad1ae2f2006-12-13 14:34:43 +0000353 set_pte_ext(pte, mk_pte(page, prot), 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 page++;
355 pte++;
Kevin Hilman37134cd2006-01-12 16:12:21 +0000356 off++;
357 if (off >= PTRS_PER_PTE) {
358 off = 0;
359 pte = consistent_pte[++idx];
360 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 } while (size -= PAGE_SIZE);
362
Russell King2be23c42010-09-08 16:27:56 +0100363 dsb();
364
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 return (void *)c->vm_start;
366 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 return NULL;
368}
Russell King695ae0a2009-11-19 16:31:39 +0000369
370static void __dma_free_remap(void *cpu_addr, size_t size)
371{
372 struct arm_vmregion *c;
373 unsigned long addr;
374 pte_t *ptep;
375 int idx;
376 u32 off;
377
378 c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr);
379 if (!c) {
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100380 pr_err("%s: trying to free invalid coherent area: %p\n",
Russell King695ae0a2009-11-19 16:31:39 +0000381 __func__, cpu_addr);
382 dump_stack();
383 return;
384 }
385
386 if ((c->vm_end - c->vm_start) != size) {
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100387 pr_err("%s: freeing wrong coherent size (%ld != %d)\n",
Russell King695ae0a2009-11-19 16:31:39 +0000388 __func__, c->vm_end - c->vm_start, size);
389 dump_stack();
390 size = c->vm_end - c->vm_start;
391 }
392
393 idx = CONSISTENT_PTE_INDEX(c->vm_start);
394 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
395 ptep = consistent_pte[idx] + off;
396 addr = c->vm_start;
397 do {
398 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
Russell King695ae0a2009-11-19 16:31:39 +0000399
400 ptep++;
401 addr += PAGE_SIZE;
402 off++;
403 if (off >= PTRS_PER_PTE) {
404 off = 0;
405 ptep = consistent_pte[++idx];
406 }
407
Russell Kingacaac252009-11-20 18:19:52 +0000408 if (pte_none(pte) || !pte_present(pte))
Marek Szyprowski6b6f7702012-02-28 10:19:14 +0100409 pr_crit("%s: bad page in kernel page table\n",
410 __func__);
Russell King695ae0a2009-11-19 16:31:39 +0000411 } while (size -= PAGE_SIZE);
412
413 flush_tlb_kernel_range(c->vm_start, c->vm_end);
414
415 arm_vmregion_free(&consistent_head, c);
416}
417
Catalin Marinasab6494f2009-07-24 12:35:02 +0100418#else /* !CONFIG_MMU */
Russell King695ae0a2009-11-19 16:31:39 +0000419
Russell King45cd5292012-01-12 23:08:07 +0000420#define __dma_alloc_remap(page, size, gfp, prot, c) page_address(page)
Russell King31ebf942009-11-19 21:12:17 +0000421#define __dma_free_remap(addr, size) do { } while (0)
422
423#endif /* CONFIG_MMU */
424
Catalin Marinasab6494f2009-07-24 12:35:02 +0100425static void *
426__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
Russell King45cd5292012-01-12 23:08:07 +0000427 pgprot_t prot, const void *caller)
Catalin Marinasab6494f2009-07-24 12:35:02 +0100428{
Russell King04da5692009-11-19 15:54:45 +0000429 struct page *page;
Russell King31ebf942009-11-19 21:12:17 +0000430 void *addr;
Catalin Marinasab6494f2009-07-24 12:35:02 +0100431
Sumit Bhattacharyaea2e7052011-11-24 00:47:12 +0100432 /*
433 * Following is a work-around (a.k.a. hack) to prevent pages
434 * with __GFP_COMP being passed to split_page() which cannot
435 * handle them. The real problem is that this flag probably
436 * should be 0 on ARM as it is not supported on this
437 * platform; see CONFIG_HUGETLBFS.
438 */
439 gfp &= ~(__GFP_COMP);
440
Marek Szyprowski553ac782012-02-29 14:45:28 +0100441 *handle = DMA_ERROR_CODE;
Russell King04da5692009-11-19 15:54:45 +0000442 size = PAGE_ALIGN(size);
443
444 page = __dma_alloc_buffer(dev, size, gfp);
445 if (!page)
446 return NULL;
447
Russell King31ebf942009-11-19 21:12:17 +0000448 if (!arch_is_coherent())
Russell King45cd5292012-01-12 23:08:07 +0000449 addr = __dma_alloc_remap(page, size, gfp, prot, caller);
Russell King31ebf942009-11-19 21:12:17 +0000450 else
451 addr = page_address(page);
452
453 if (addr)
Russell King9eedd962011-01-03 00:00:17 +0000454 *handle = pfn_to_dma(dev, page_to_pfn(page));
Russell Kingd8e89b42011-09-22 10:32:25 +0100455 else
456 __dma_free_buffer(page, size);
Russell King31ebf942009-11-19 21:12:17 +0000457
458 return addr;
Catalin Marinasab6494f2009-07-24 12:35:02 +0100459}
Russell King695ae0a2009-11-19 16:31:39 +0000460
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461/*
462 * Allocate DMA-coherent memory space and return both the kernel remapped
463 * virtual and bus address for that space.
464 */
465void *
Al Virof9e32142005-10-21 03:20:58 -0400466dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
Dmitry Baryshkov1fe53262008-07-18 13:30:14 +0400468 void *memory;
469
470 if (dma_alloc_from_coherent(dev, size, handle, &memory))
471 return memory;
472
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 return __dma_alloc(dev, size, handle, gfp,
Russell King45cd5292012-01-12 23:08:07 +0000474 pgprot_dmacoherent(pgprot_kernel),
475 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476}
477EXPORT_SYMBOL(dma_alloc_coherent);
478
479/*
480 * Allocate a writecombining region, in much the same way as
481 * dma_alloc_coherent above.
482 */
483void *
Al Virof9e32142005-10-21 03:20:58 -0400484dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700485{
486 return __dma_alloc(dev, size, handle, gfp,
Russell King45cd5292012-01-12 23:08:07 +0000487 pgprot_writecombine(pgprot_kernel),
488 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489}
490EXPORT_SYMBOL(dma_alloc_writecombine);
491
492static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
493 void *cpu_addr, dma_addr_t dma_addr, size_t size)
494{
Catalin Marinasab6494f2009-07-24 12:35:02 +0100495 int ret = -ENXIO;
496#ifdef CONFIG_MMU
Russell King13ccf3a2009-11-19 15:07:04 +0000497 unsigned long user_size, kern_size;
498 struct arm_vmregion *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499
Marek Szyprowski47142f02012-05-15 19:04:13 +0200500 if (dma_mmap_from_coherent(dev, vma, cpu_addr, size, &ret))
501 return ret;
502
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
504
Russell King13ccf3a2009-11-19 15:07:04 +0000505 c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 if (c) {
507 unsigned long off = vma->vm_pgoff;
508
509 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
510
511 if (off < kern_size &&
512 user_size <= (kern_size - off)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 ret = remap_pfn_range(vma, vma->vm_start,
514 page_to_pfn(c->vm_pages) + off,
515 user_size << PAGE_SHIFT,
516 vma->vm_page_prot);
517 }
518 }
Catalin Marinasab6494f2009-07-24 12:35:02 +0100519#endif /* CONFIG_MMU */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520
521 return ret;
522}
523
524int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
525 void *cpu_addr, dma_addr_t dma_addr, size_t size)
526{
Russell King26a26d32009-11-20 21:06:43 +0000527 vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
529}
530EXPORT_SYMBOL(dma_mmap_coherent);
531
532int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
533 void *cpu_addr, dma_addr_t dma_addr, size_t size)
534{
535 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
536 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
537}
538EXPORT_SYMBOL(dma_mmap_writecombine);
539
540/*
541 * free a page as defined by the above mapping.
Russell King5edf71a2005-11-25 15:52:51 +0000542 * Must not be called with IRQs disabled.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543 */
544void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
545{
Russell King5edf71a2005-11-25 15:52:51 +0000546 WARN_ON(irqs_disabled());
547
Dmitry Baryshkov1fe53262008-07-18 13:30:14 +0400548 if (dma_release_from_coherent(dev, get_order(size), cpu_addr))
549 return;
550
Russell King3e82d012009-11-19 15:38:12 +0000551 size = PAGE_ALIGN(size);
552
Russell King695ae0a2009-11-19 16:31:39 +0000553 if (!arch_is_coherent())
554 __dma_free_remap(cpu_addr, size);
Russell King7a9a32a2009-11-19 15:31:07 +0000555
Russell King9eedd962011-01-03 00:00:17 +0000556 __dma_free_buffer(pfn_to_page(dma_to_pfn(dev, handle)), size);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557}
558EXPORT_SYMBOL(dma_free_coherent);
559
Russell King65af1912009-11-24 17:53:33 +0000560static void dma_cache_maint_page(struct page *page, unsigned long offset,
Russell Kinga9c91472009-11-26 16:19:58 +0000561 size_t size, enum dma_data_direction dir,
562 void (*op)(const void *, size_t, int))
Russell King65af1912009-11-24 17:53:33 +0000563{
564 /*
565 * A single sg entry may refer to multiple physically contiguous
566 * pages. But we still need to process highmem pages individually.
567 * If highmem is not configured then the bulk of this loop gets
568 * optimized out.
569 */
570 size_t left = size;
571 do {
572 size_t len = left;
Russell King93f1d622009-11-24 14:41:01 +0000573 void *vaddr;
574
575 if (PageHighMem(page)) {
576 if (len + offset > PAGE_SIZE) {
577 if (offset >= PAGE_SIZE) {
578 page += offset / PAGE_SIZE;
579 offset %= PAGE_SIZE;
580 }
581 len = PAGE_SIZE - offset;
Russell King65af1912009-11-24 17:53:33 +0000582 }
Russell King93f1d622009-11-24 14:41:01 +0000583 vaddr = kmap_high_get(page);
584 if (vaddr) {
585 vaddr += offset;
Russell Kinga9c91472009-11-26 16:19:58 +0000586 op(vaddr, len, dir);
Russell King93f1d622009-11-24 14:41:01 +0000587 kunmap_high(page);
Nicolas Pitre7e5a69e2010-03-29 21:46:02 +0100588 } else if (cache_is_vipt()) {
Nicolas Pitre39af22a2010-12-15 15:14:45 -0500589 /* unmapped pages might still be cached */
590 vaddr = kmap_atomic(page);
Nicolas Pitre7e5a69e2010-03-29 21:46:02 +0100591 op(vaddr + offset, len, dir);
Nicolas Pitre39af22a2010-12-15 15:14:45 -0500592 kunmap_atomic(vaddr);
Russell King93f1d622009-11-24 14:41:01 +0000593 }
594 } else {
595 vaddr = page_address(page) + offset;
Russell Kinga9c91472009-11-26 16:19:58 +0000596 op(vaddr, len, dir);
Russell King65af1912009-11-24 17:53:33 +0000597 }
Russell King65af1912009-11-24 17:53:33 +0000598 offset = 0;
599 page++;
600 left -= len;
601 } while (left);
602}
603
Marek Szyprowski51fde3492012-02-10 19:55:20 +0100604/*
605 * Make an area consistent for devices.
606 * Note: Drivers should NOT use this function directly, as it will break
607 * platforms with CONFIG_DMABOUNCE.
608 * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
609 */
610static void __dma_page_cpu_to_dev(struct page *page, unsigned long off,
Russell King65af1912009-11-24 17:53:33 +0000611 size_t size, enum dma_data_direction dir)
612{
Nicolas Pitre43377452009-03-12 22:52:09 -0400613 unsigned long paddr;
Nicolas Pitre43377452009-03-12 22:52:09 -0400614
Russell Kinga9c91472009-11-26 16:19:58 +0000615 dma_cache_maint_page(page, off, size, dir, dmac_map_area);
Nicolas Pitre43377452009-03-12 22:52:09 -0400616
Russell King65af1912009-11-24 17:53:33 +0000617 paddr = page_to_phys(page) + off;
Russell King2ffe2da2009-10-31 16:52:16 +0000618 if (dir == DMA_FROM_DEVICE) {
619 outer_inv_range(paddr, paddr + size);
620 } else {
621 outer_clean_range(paddr, paddr + size);
622 }
623 /* FIXME: non-speculating: flush on bidirectional mappings? */
Nicolas Pitre43377452009-03-12 22:52:09 -0400624}
Russell King4ea0d732009-11-24 16:27:17 +0000625
Marek Szyprowski51fde3492012-02-10 19:55:20 +0100626static void __dma_page_dev_to_cpu(struct page *page, unsigned long off,
Russell King4ea0d732009-11-24 16:27:17 +0000627 size_t size, enum dma_data_direction dir)
628{
Russell King2ffe2da2009-10-31 16:52:16 +0000629 unsigned long paddr = page_to_phys(page) + off;
630
631 /* FIXME: non-speculating: not required */
632 /* don't bother invalidating if DMA to device */
633 if (dir != DMA_TO_DEVICE)
634 outer_inv_range(paddr, paddr + size);
635
Russell Kinga9c91472009-11-26 16:19:58 +0000636 dma_cache_maint_page(page, off, size, dir, dmac_unmap_area);
Catalin Marinasc0177802010-09-13 15:57:36 +0100637
638 /*
639 * Mark the D-cache clean for this page to avoid extra flushing.
640 */
641 if (dir != DMA_TO_DEVICE && off == 0 && size >= PAGE_SIZE)
642 set_bit(PG_dcache_clean, &page->flags);
Russell King4ea0d732009-11-24 16:27:17 +0000643}
Nicolas Pitre43377452009-03-12 22:52:09 -0400644
Russell Kingafd1a322008-09-25 16:30:57 +0100645/**
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100646 * arm_dma_map_sg - map a set of SG buffers for streaming mode DMA
Russell Kingafd1a322008-09-25 16:30:57 +0100647 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
648 * @sg: list of buffers
649 * @nents: number of buffers to map
650 * @dir: DMA transfer direction
651 *
652 * Map a set of buffers described by scatterlist in streaming mode for DMA.
653 * This is the scatter-gather version of the dma_map_single interface.
654 * Here the scatter gather list elements are each tagged with the
655 * appropriate dma address and length. They are obtained via
656 * sg_dma_{address,length}.
657 *
658 * Device ownership issues as mentioned for dma_map_single are the same
659 * here.
660 */
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100661int arm_dma_map_sg(struct device *dev, struct scatterlist *sg, int nents,
662 enum dma_data_direction dir, struct dma_attrs *attrs)
Russell Kingafd1a322008-09-25 16:30:57 +0100663{
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100664 struct dma_map_ops *ops = get_dma_ops(dev);
Russell Kingafd1a322008-09-25 16:30:57 +0100665 struct scatterlist *s;
Russell King01135d922008-09-25 21:05:02 +0100666 int i, j;
Russell Kingafd1a322008-09-25 16:30:57 +0100667
668 for_each_sg(sg, s, nents, i) {
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100669 s->dma_address = ops->map_page(dev, sg_page(s), s->offset,
670 s->length, dir, attrs);
Russell King01135d922008-09-25 21:05:02 +0100671 if (dma_mapping_error(dev, s->dma_address))
672 goto bad_mapping;
Russell Kingafd1a322008-09-25 16:30:57 +0100673 }
Russell Kingafd1a322008-09-25 16:30:57 +0100674 return nents;
Russell King01135d922008-09-25 21:05:02 +0100675
676 bad_mapping:
677 for_each_sg(sg, s, i, j)
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100678 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
Russell King01135d922008-09-25 21:05:02 +0100679 return 0;
Russell Kingafd1a322008-09-25 16:30:57 +0100680}
Russell Kingafd1a322008-09-25 16:30:57 +0100681
682/**
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100683 * arm_dma_unmap_sg - unmap a set of SG buffers mapped by dma_map_sg
Russell Kingafd1a322008-09-25 16:30:57 +0100684 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
685 * @sg: list of buffers
Linus Walleij0adfca62011-01-12 18:50:37 +0100686 * @nents: number of buffers to unmap (same as was passed to dma_map_sg)
Russell Kingafd1a322008-09-25 16:30:57 +0100687 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
688 *
689 * Unmap a set of streaming mode DMA translations. Again, CPU access
690 * rules concerning calls here are the same as for dma_unmap_single().
691 */
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100692void arm_dma_unmap_sg(struct device *dev, struct scatterlist *sg, int nents,
693 enum dma_data_direction dir, struct dma_attrs *attrs)
Russell Kingafd1a322008-09-25 16:30:57 +0100694{
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100695 struct dma_map_ops *ops = get_dma_ops(dev);
Russell King01135d922008-09-25 21:05:02 +0100696 struct scatterlist *s;
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100697
Russell King01135d922008-09-25 21:05:02 +0100698 int i;
699
700 for_each_sg(sg, s, nents, i)
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100701 ops->unmap_page(dev, sg_dma_address(s), sg_dma_len(s), dir, attrs);
Russell Kingafd1a322008-09-25 16:30:57 +0100702}
Russell Kingafd1a322008-09-25 16:30:57 +0100703
704/**
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100705 * arm_dma_sync_sg_for_cpu
Russell Kingafd1a322008-09-25 16:30:57 +0100706 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
707 * @sg: list of buffers
708 * @nents: number of buffers to map (returned from dma_map_sg)
709 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
710 */
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100711void arm_dma_sync_sg_for_cpu(struct device *dev, struct scatterlist *sg,
Russell Kingafd1a322008-09-25 16:30:57 +0100712 int nents, enum dma_data_direction dir)
713{
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100714 struct dma_map_ops *ops = get_dma_ops(dev);
Russell Kingafd1a322008-09-25 16:30:57 +0100715 struct scatterlist *s;
716 int i;
717
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100718 for_each_sg(sg, s, nents, i)
719 ops->sync_single_for_cpu(dev, sg_dma_address(s), s->length,
720 dir);
Russell Kingafd1a322008-09-25 16:30:57 +0100721}
Russell Kingafd1a322008-09-25 16:30:57 +0100722
723/**
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100724 * arm_dma_sync_sg_for_device
Russell Kingafd1a322008-09-25 16:30:57 +0100725 * @dev: valid struct device pointer, or NULL for ISA and EISA-like devices
726 * @sg: list of buffers
727 * @nents: number of buffers to map (returned from dma_map_sg)
728 * @dir: DMA transfer direction (same as was passed to dma_map_sg)
729 */
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100730void arm_dma_sync_sg_for_device(struct device *dev, struct scatterlist *sg,
Russell Kingafd1a322008-09-25 16:30:57 +0100731 int nents, enum dma_data_direction dir)
732{
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100733 struct dma_map_ops *ops = get_dma_ops(dev);
Russell Kingafd1a322008-09-25 16:30:57 +0100734 struct scatterlist *s;
735 int i;
736
Marek Szyprowski2a550e72012-02-10 19:55:20 +0100737 for_each_sg(sg, s, nents, i)
738 ops->sync_single_for_device(dev, sg_dma_address(s), s->length,
739 dir);
Russell Kingafd1a322008-09-25 16:30:57 +0100740}
Russell King24056f52011-01-03 11:29:28 +0000741
Russell King022ae532011-07-08 21:26:59 +0100742/*
743 * Return whether the given device DMA address mask can be supported
744 * properly. For example, if your device can only drive the low 24-bits
745 * during bus mastering, then you would pass 0x00ffffff as the mask
746 * to this function.
747 */
748int dma_supported(struct device *dev, u64 mask)
749{
750 if (mask < (u64)arm_dma_limit)
751 return 0;
752 return 1;
753}
754EXPORT_SYMBOL(dma_supported);
755
Marek Szyprowski2dc6a012012-02-10 19:55:20 +0100756static int arm_dma_set_mask(struct device *dev, u64 dma_mask)
Russell King022ae532011-07-08 21:26:59 +0100757{
758 if (!dev->dma_mask || !dma_supported(dev, dma_mask))
759 return -EIO;
760
Russell King022ae532011-07-08 21:26:59 +0100761 *dev->dma_mask = dma_mask;
Russell King022ae532011-07-08 21:26:59 +0100762
763 return 0;
764}
Russell King022ae532011-07-08 21:26:59 +0100765
Russell King24056f52011-01-03 11:29:28 +0000766#define PREALLOC_DMA_DEBUG_ENTRIES 4096
767
768static int __init dma_debug_do_init(void)
769{
Russell King45cd5292012-01-12 23:08:07 +0000770#ifdef CONFIG_MMU
771 arm_vmregion_create_proc("dma-mappings", &consistent_head);
772#endif
Russell King24056f52011-01-03 11:29:28 +0000773 dma_debug_init(PREALLOC_DMA_DEBUG_ENTRIES);
774 return 0;
775}
776fs_initcall(dma_debug_do_init);