blob: 26356ce4da5448165bcac172139deef54a008462 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/arch/arm/mm/consistent.c
3 *
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>
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>
20
21#include <asm/cacheflush.h>
22#include <asm/io.h>
23#include <asm/tlbflush.h>
24
25#define CONSISTENT_BASE (0xffc00000)
26#define CONSISTENT_END (0xffe00000)
27#define CONSISTENT_OFFSET(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
28
29/*
30 * This is the page table (2MB) covering uncached, DMA consistent allocations
31 */
32static pte_t *consistent_pte;
33static DEFINE_SPINLOCK(consistent_lock);
34
35/*
36 * VM region handling support.
37 *
38 * This should become something generic, handling VM region allocations for
39 * vmalloc and similar (ioremap, module space, etc).
40 *
41 * I envisage vmalloc()'s supporting vm_struct becoming:
42 *
43 * struct vm_struct {
44 * struct vm_region region;
45 * unsigned long flags;
46 * struct page **pages;
47 * unsigned int nr_pages;
48 * unsigned long phys_addr;
49 * };
50 *
51 * get_vm_area() would then call vm_region_alloc with an appropriate
52 * struct vm_region head (eg):
53 *
54 * struct vm_region vmalloc_head = {
55 * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list),
56 * .vm_start = VMALLOC_START,
57 * .vm_end = VMALLOC_END,
58 * };
59 *
60 * However, vmalloc_head.vm_start is variable (typically, it is dependent on
61 * the amount of RAM found at boot time.) I would imagine that get_vm_area()
62 * would have to initialise this each time prior to calling vm_region_alloc().
63 */
64struct vm_region {
65 struct list_head vm_list;
66 unsigned long vm_start;
67 unsigned long vm_end;
68 struct page *vm_pages;
69};
70
71static struct vm_region consistent_head = {
72 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
73 .vm_start = CONSISTENT_BASE,
74 .vm_end = CONSISTENT_END,
75};
76
77static struct vm_region *
78vm_region_alloc(struct vm_region *head, size_t size, int gfp)
79{
80 unsigned long addr = head->vm_start, end = head->vm_end - size;
81 unsigned long flags;
82 struct vm_region *c, *new;
83
84 new = kmalloc(sizeof(struct vm_region), gfp);
85 if (!new)
86 goto out;
87
88 spin_lock_irqsave(&consistent_lock, flags);
89
90 list_for_each_entry(c, &head->vm_list, vm_list) {
91 if ((addr + size) < addr)
92 goto nospc;
93 if ((addr + size) <= c->vm_start)
94 goto found;
95 addr = c->vm_end;
96 if (addr > end)
97 goto nospc;
98 }
99
100 found:
101 /*
102 * Insert this entry _before_ the one we found.
103 */
104 list_add_tail(&new->vm_list, &c->vm_list);
105 new->vm_start = addr;
106 new->vm_end = addr + size;
107
108 spin_unlock_irqrestore(&consistent_lock, flags);
109 return new;
110
111 nospc:
112 spin_unlock_irqrestore(&consistent_lock, flags);
113 kfree(new);
114 out:
115 return NULL;
116}
117
118static struct vm_region *vm_region_find(struct vm_region *head, unsigned long addr)
119{
120 struct vm_region *c;
121
122 list_for_each_entry(c, &head->vm_list, vm_list) {
123 if (c->vm_start == addr)
124 goto out;
125 }
126 c = NULL;
127 out:
128 return c;
129}
130
131#ifdef CONFIG_HUGETLB_PAGE
132#error ARM Coherent DMA allocator does not (yet) support huge TLB
133#endif
134
135static void *
136__dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, int gfp,
137 pgprot_t prot)
138{
139 struct page *page;
140 struct vm_region *c;
141 unsigned long order;
142 u64 mask = ISA_DMA_THRESHOLD, limit;
143
144 if (!consistent_pte) {
145 printk(KERN_ERR "%s: not initialised\n", __func__);
146 dump_stack();
147 return NULL;
148 }
149
150 if (dev) {
151 mask = dev->coherent_dma_mask;
152
153 /*
154 * Sanity check the DMA mask - it must be non-zero, and
155 * must be able to be satisfied by a DMA allocation.
156 */
157 if (mask == 0) {
158 dev_warn(dev, "coherent DMA mask is unset\n");
159 goto no_page;
160 }
161
162 if ((~mask) & ISA_DMA_THRESHOLD) {
163 dev_warn(dev, "coherent DMA mask %#llx is smaller "
164 "than system GFP_DMA mask %#llx\n",
165 mask, (unsigned long long)ISA_DMA_THRESHOLD);
166 goto no_page;
167 }
168 }
169
170 /*
171 * Sanity check the allocation size.
172 */
173 size = PAGE_ALIGN(size);
174 limit = (mask + 1) & ~mask;
175 if ((limit && size >= limit) ||
176 size >= (CONSISTENT_END - CONSISTENT_BASE)) {
177 printk(KERN_WARNING "coherent allocation too big "
178 "(requested %#x mask %#llx)\n", size, mask);
179 goto no_page;
180 }
181
182 order = get_order(size);
183
184 if (mask != 0xffffffff)
185 gfp |= GFP_DMA;
186
187 page = alloc_pages(gfp, order);
188 if (!page)
189 goto no_page;
190
191 /*
192 * Invalidate any data that might be lurking in the
193 * kernel direct-mapped region for device DMA.
194 */
195 {
196 unsigned long kaddr = (unsigned long)page_address(page);
197 memset(page_address(page), 0, size);
198 dmac_flush_range(kaddr, kaddr + size);
199 }
200
201 /*
202 * Allocate a virtual address in the consistent mapping region.
203 */
204 c = vm_region_alloc(&consistent_head, size,
205 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
206 if (c) {
207 pte_t *pte = consistent_pte + CONSISTENT_OFFSET(c->vm_start);
208 struct page *end = page + (1 << order);
209
210 c->vm_pages = page;
211
212 /*
213 * Set the "dma handle"
214 */
215 *handle = page_to_dma(dev, page);
216
217 do {
218 BUG_ON(!pte_none(*pte));
219
220 set_page_count(page, 1);
221 /*
222 * x86 does not mark the pages reserved...
223 */
224 SetPageReserved(page);
225 set_pte(pte, mk_pte(page, prot));
226 page++;
227 pte++;
228 } while (size -= PAGE_SIZE);
229
230 /*
231 * Free the otherwise unused pages.
232 */
233 while (page < end) {
234 set_page_count(page, 1);
235 __free_page(page);
236 page++;
237 }
238
239 return (void *)c->vm_start;
240 }
241
242 if (page)
243 __free_pages(page, order);
244 no_page:
245 *handle = ~0;
246 return NULL;
247}
248
249/*
250 * Allocate DMA-coherent memory space and return both the kernel remapped
251 * virtual and bus address for that space.
252 */
253void *
254dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, int gfp)
255{
256 return __dma_alloc(dev, size, handle, gfp,
257 pgprot_noncached(pgprot_kernel));
258}
259EXPORT_SYMBOL(dma_alloc_coherent);
260
261/*
262 * Allocate a writecombining region, in much the same way as
263 * dma_alloc_coherent above.
264 */
265void *
266dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, int gfp)
267{
268 return __dma_alloc(dev, size, handle, gfp,
269 pgprot_writecombine(pgprot_kernel));
270}
271EXPORT_SYMBOL(dma_alloc_writecombine);
272
273static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
274 void *cpu_addr, dma_addr_t dma_addr, size_t size)
275{
276 unsigned long flags, user_size, kern_size;
277 struct vm_region *c;
278 int ret = -ENXIO;
279
280 user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
281
282 spin_lock_irqsave(&consistent_lock, flags);
283 c = vm_region_find(&consistent_head, (unsigned long)cpu_addr);
284 spin_unlock_irqrestore(&consistent_lock, flags);
285
286 if (c) {
287 unsigned long off = vma->vm_pgoff;
288
289 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
290
291 if (off < kern_size &&
292 user_size <= (kern_size - off)) {
293 vma->vm_flags |= VM_RESERVED;
294 ret = remap_pfn_range(vma, vma->vm_start,
295 page_to_pfn(c->vm_pages) + off,
296 user_size << PAGE_SHIFT,
297 vma->vm_page_prot);
298 }
299 }
300
301 return ret;
302}
303
304int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
305 void *cpu_addr, dma_addr_t dma_addr, size_t size)
306{
307 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
308 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
309}
310EXPORT_SYMBOL(dma_mmap_coherent);
311
312int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
313 void *cpu_addr, dma_addr_t dma_addr, size_t size)
314{
315 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
316 return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
317}
318EXPORT_SYMBOL(dma_mmap_writecombine);
319
320/*
321 * free a page as defined by the above mapping.
322 */
323void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
324{
325 struct vm_region *c;
326 unsigned long flags, addr;
327 pte_t *ptep;
328
329 size = PAGE_ALIGN(size);
330
331 spin_lock_irqsave(&consistent_lock, flags);
332
333 c = vm_region_find(&consistent_head, (unsigned long)cpu_addr);
334 if (!c)
335 goto no_area;
336
337 if ((c->vm_end - c->vm_start) != size) {
338 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
339 __func__, c->vm_end - c->vm_start, size);
340 dump_stack();
341 size = c->vm_end - c->vm_start;
342 }
343
344 ptep = consistent_pte + CONSISTENT_OFFSET(c->vm_start);
345 addr = c->vm_start;
346 do {
347 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
348 unsigned long pfn;
349
350 ptep++;
351 addr += PAGE_SIZE;
352
353 if (!pte_none(pte) && pte_present(pte)) {
354 pfn = pte_pfn(pte);
355
356 if (pfn_valid(pfn)) {
357 struct page *page = pfn_to_page(pfn);
358
359 /*
360 * x86 does not mark the pages reserved...
361 */
362 ClearPageReserved(page);
363
364 __free_page(page);
365 continue;
366 }
367 }
368
369 printk(KERN_CRIT "%s: bad page in kernel page table\n",
370 __func__);
371 } while (size -= PAGE_SIZE);
372
373 flush_tlb_kernel_range(c->vm_start, c->vm_end);
374
375 list_del(&c->vm_list);
376
377 spin_unlock_irqrestore(&consistent_lock, flags);
378
379 kfree(c);
380 return;
381
382 no_area:
383 spin_unlock_irqrestore(&consistent_lock, flags);
384 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
385 __func__, cpu_addr);
386 dump_stack();
387}
388EXPORT_SYMBOL(dma_free_coherent);
389
390/*
391 * Initialise the consistent memory allocation.
392 */
393static int __init consistent_init(void)
394{
395 pgd_t *pgd;
396 pmd_t *pmd;
397 pte_t *pte;
398 int ret = 0;
399
400 spin_lock(&init_mm.page_table_lock);
401
402 do {
403 pgd = pgd_offset(&init_mm, CONSISTENT_BASE);
404 pmd = pmd_alloc(&init_mm, pgd, CONSISTENT_BASE);
405 if (!pmd) {
406 printk(KERN_ERR "%s: no pmd tables\n", __func__);
407 ret = -ENOMEM;
408 break;
409 }
410 WARN_ON(!pmd_none(*pmd));
411
412 pte = pte_alloc_kernel(&init_mm, pmd, CONSISTENT_BASE);
413 if (!pte) {
414 printk(KERN_ERR "%s: no pte tables\n", __func__);
415 ret = -ENOMEM;
416 break;
417 }
418
419 consistent_pte = pte;
420 } while (0);
421
422 spin_unlock(&init_mm.page_table_lock);
423
424 return ret;
425}
426
427core_initcall(consistent_init);
428
429/*
430 * Make an area consistent for devices.
431 */
432void consistent_sync(void *vaddr, size_t size, int direction)
433{
434 unsigned long start = (unsigned long)vaddr;
435 unsigned long end = start + size;
436
437 switch (direction) {
438 case DMA_FROM_DEVICE: /* invalidate only */
439 dmac_inv_range(start, end);
440 break;
441 case DMA_TO_DEVICE: /* writeback only */
442 dmac_clean_range(start, end);
443 break;
444 case DMA_BIDIRECTIONAL: /* writeback and invalidate */
445 dmac_flush_range(start, end);
446 break;
447 default:
448 BUG();
449 }
450}
451EXPORT_SYMBOL(consistent_sync);