[ARM] Convert DMA cache handling to take const void * args
[linux-2.6] / arch / arm / mm / consistent.c
1 /*
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/memory.h>
22 #include <asm/cacheflush.h>
23 #include <asm/tlbflush.h>
24 #include <asm/sizes.h>
25
26 /* Sanity check size */
27 #if (CONSISTENT_DMA_SIZE % SZ_2M)
28 #error "CONSISTENT_DMA_SIZE must be multiple of 2MiB"
29 #endif
30
31 #define CONSISTENT_END  (0xffe00000)
32 #define CONSISTENT_BASE (CONSISTENT_END - CONSISTENT_DMA_SIZE)
33
34 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
35 #define CONSISTENT_PTE_INDEX(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PGDIR_SHIFT)
36 #define NUM_CONSISTENT_PTES (CONSISTENT_DMA_SIZE >> PGDIR_SHIFT)
37
38
39 /*
40  * These are the page tables (2MB each) covering uncached, DMA consistent allocations
41  */
42 static pte_t *consistent_pte[NUM_CONSISTENT_PTES];
43 static DEFINE_SPINLOCK(consistent_lock);
44
45 /*
46  * VM region handling support.
47  *
48  * This should become something generic, handling VM region allocations for
49  * vmalloc and similar (ioremap, module space, etc).
50  *
51  * I envisage vmalloc()'s supporting vm_struct becoming:
52  *
53  *  struct vm_struct {
54  *    struct vm_region  region;
55  *    unsigned long     flags;
56  *    struct page       **pages;
57  *    unsigned int      nr_pages;
58  *    unsigned long     phys_addr;
59  *  };
60  *
61  * get_vm_area() would then call vm_region_alloc with an appropriate
62  * struct vm_region head (eg):
63  *
64  *  struct vm_region vmalloc_head = {
65  *      .vm_list        = LIST_HEAD_INIT(vmalloc_head.vm_list),
66  *      .vm_start       = VMALLOC_START,
67  *      .vm_end         = VMALLOC_END,
68  *  };
69  *
70  * However, vmalloc_head.vm_start is variable (typically, it is dependent on
71  * the amount of RAM found at boot time.)  I would imagine that get_vm_area()
72  * would have to initialise this each time prior to calling vm_region_alloc().
73  */
74 struct vm_region {
75         struct list_head        vm_list;
76         unsigned long           vm_start;
77         unsigned long           vm_end;
78         struct page             *vm_pages;
79         int                     vm_active;
80 };
81
82 static struct vm_region consistent_head = {
83         .vm_list        = LIST_HEAD_INIT(consistent_head.vm_list),
84         .vm_start       = CONSISTENT_BASE,
85         .vm_end         = CONSISTENT_END,
86 };
87
88 static struct vm_region *
89 vm_region_alloc(struct vm_region *head, size_t size, gfp_t gfp)
90 {
91         unsigned long addr = head->vm_start, end = head->vm_end - size;
92         unsigned long flags;
93         struct vm_region *c, *new;
94
95         new = kmalloc(sizeof(struct vm_region), gfp);
96         if (!new)
97                 goto out;
98
99         spin_lock_irqsave(&consistent_lock, flags);
100
101         list_for_each_entry(c, &head->vm_list, vm_list) {
102                 if ((addr + size) < addr)
103                         goto nospc;
104                 if ((addr + size) <= c->vm_start)
105                         goto found;
106                 addr = c->vm_end;
107                 if (addr > end)
108                         goto nospc;
109         }
110
111  found:
112         /*
113          * Insert this entry _before_ the one we found.
114          */
115         list_add_tail(&new->vm_list, &c->vm_list);
116         new->vm_start = addr;
117         new->vm_end = addr + size;
118         new->vm_active = 1;
119
120         spin_unlock_irqrestore(&consistent_lock, flags);
121         return new;
122
123  nospc:
124         spin_unlock_irqrestore(&consistent_lock, flags);
125         kfree(new);
126  out:
127         return NULL;
128 }
129
130 static struct vm_region *vm_region_find(struct vm_region *head, unsigned long addr)
131 {
132         struct vm_region *c;
133         
134         list_for_each_entry(c, &head->vm_list, vm_list) {
135                 if (c->vm_active && c->vm_start == addr)
136                         goto out;
137         }
138         c = NULL;
139  out:
140         return c;
141 }
142
143 #ifdef CONFIG_HUGETLB_PAGE
144 #error ARM Coherent DMA allocator does not (yet) support huge TLB
145 #endif
146
147 static void *
148 __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp,
149             pgprot_t prot)
150 {
151         struct page *page;
152         struct vm_region *c;
153         unsigned long order;
154         u64 mask = ISA_DMA_THRESHOLD, limit;
155
156         if (!consistent_pte[0]) {
157                 printk(KERN_ERR "%s: not initialised\n", __func__);
158                 dump_stack();
159                 return NULL;
160         }
161
162         if (dev) {
163                 mask = dev->coherent_dma_mask;
164
165                 /*
166                  * Sanity check the DMA mask - it must be non-zero, and
167                  * must be able to be satisfied by a DMA allocation.
168                  */
169                 if (mask == 0) {
170                         dev_warn(dev, "coherent DMA mask is unset\n");
171                         goto no_page;
172                 }
173
174                 if ((~mask) & ISA_DMA_THRESHOLD) {
175                         dev_warn(dev, "coherent DMA mask %#llx is smaller "
176                                  "than system GFP_DMA mask %#llx\n",
177                                  mask, (unsigned long long)ISA_DMA_THRESHOLD);
178                         goto no_page;
179                 }
180         }
181
182         /*
183          * Sanity check the allocation size.
184          */
185         size = PAGE_ALIGN(size);
186         limit = (mask + 1) & ~mask;
187         if ((limit && size >= limit) ||
188             size >= (CONSISTENT_END - CONSISTENT_BASE)) {
189                 printk(KERN_WARNING "coherent allocation too big "
190                        "(requested %#x mask %#llx)\n", size, mask);
191                 goto no_page;
192         }
193
194         order = get_order(size);
195
196         if (mask != 0xffffffff)
197                 gfp |= GFP_DMA;
198
199         page = alloc_pages(gfp, order);
200         if (!page)
201                 goto no_page;
202
203         /*
204          * Invalidate any data that might be lurking in the
205          * kernel direct-mapped region for device DMA.
206          */
207         {
208                 void *ptr = page_address(page);
209                 memset(ptr, 0, size);
210                 dmac_flush_range(ptr, ptr + size);
211                 outer_flush_range(__pa(ptr), __pa(ptr) + size);
212         }
213
214         /*
215          * Allocate a virtual address in the consistent mapping region.
216          */
217         c = vm_region_alloc(&consistent_head, size,
218                             gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
219         if (c) {
220                 pte_t *pte;
221                 struct page *end = page + (1 << order);
222                 int idx = CONSISTENT_PTE_INDEX(c->vm_start);
223                 u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
224
225                 pte = consistent_pte[idx] + off;
226                 c->vm_pages = page;
227
228                 split_page(page, order);
229
230                 /*
231                  * Set the "dma handle"
232                  */
233                 *handle = page_to_dma(dev, page);
234
235                 do {
236                         BUG_ON(!pte_none(*pte));
237
238                         /*
239                          * x86 does not mark the pages reserved...
240                          */
241                         SetPageReserved(page);
242                         set_pte_ext(pte, mk_pte(page, prot), 0);
243                         page++;
244                         pte++;
245                         off++;
246                         if (off >= PTRS_PER_PTE) {
247                                 off = 0;
248                                 pte = consistent_pte[++idx];
249                         }
250                 } while (size -= PAGE_SIZE);
251
252                 /*
253                  * Free the otherwise unused pages.
254                  */
255                 while (page < end) {
256                         __free_page(page);
257                         page++;
258                 }
259
260                 return (void *)c->vm_start;
261         }
262
263         if (page)
264                 __free_pages(page, order);
265  no_page:
266         *handle = ~0;
267         return NULL;
268 }
269
270 /*
271  * Allocate DMA-coherent memory space and return both the kernel remapped
272  * virtual and bus address for that space.
273  */
274 void *
275 dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
276 {
277         if (arch_is_coherent()) {
278                 void *virt;
279
280                 virt = kmalloc(size, gfp);
281                 if (!virt)
282                         return NULL;
283                 *handle =  virt_to_dma(dev, virt);
284
285                 return virt;
286         }
287
288         return __dma_alloc(dev, size, handle, gfp,
289                            pgprot_noncached(pgprot_kernel));
290 }
291 EXPORT_SYMBOL(dma_alloc_coherent);
292
293 /*
294  * Allocate a writecombining region, in much the same way as
295  * dma_alloc_coherent above.
296  */
297 void *
298 dma_alloc_writecombine(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
299 {
300         return __dma_alloc(dev, size, handle, gfp,
301                            pgprot_writecombine(pgprot_kernel));
302 }
303 EXPORT_SYMBOL(dma_alloc_writecombine);
304
305 static int dma_mmap(struct device *dev, struct vm_area_struct *vma,
306                     void *cpu_addr, dma_addr_t dma_addr, size_t size)
307 {
308         unsigned long flags, user_size, kern_size;
309         struct vm_region *c;
310         int ret = -ENXIO;
311
312         user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
313
314         spin_lock_irqsave(&consistent_lock, flags);
315         c = vm_region_find(&consistent_head, (unsigned long)cpu_addr);
316         spin_unlock_irqrestore(&consistent_lock, flags);
317
318         if (c) {
319                 unsigned long off = vma->vm_pgoff;
320
321                 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
322
323                 if (off < kern_size &&
324                     user_size <= (kern_size - off)) {
325                         vma->vm_flags |= VM_RESERVED;
326                         ret = remap_pfn_range(vma, vma->vm_start,
327                                               page_to_pfn(c->vm_pages) + off,
328                                               user_size << PAGE_SHIFT,
329                                               vma->vm_page_prot);
330                 }
331         }
332
333         return ret;
334 }
335
336 int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma,
337                       void *cpu_addr, dma_addr_t dma_addr, size_t size)
338 {
339         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
340         return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
341 }
342 EXPORT_SYMBOL(dma_mmap_coherent);
343
344 int dma_mmap_writecombine(struct device *dev, struct vm_area_struct *vma,
345                           void *cpu_addr, dma_addr_t dma_addr, size_t size)
346 {
347         vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
348         return dma_mmap(dev, vma, cpu_addr, dma_addr, size);
349 }
350 EXPORT_SYMBOL(dma_mmap_writecombine);
351
352 /*
353  * free a page as defined by the above mapping.
354  * Must not be called with IRQs disabled.
355  */
356 void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle)
357 {
358         struct vm_region *c;
359         unsigned long flags, addr;
360         pte_t *ptep;
361         int idx;
362         u32 off;
363
364         WARN_ON(irqs_disabled());
365
366         if (arch_is_coherent()) {
367                 kfree(cpu_addr);
368                 return;
369         }
370
371         size = PAGE_ALIGN(size);
372
373         spin_lock_irqsave(&consistent_lock, flags);
374         c = vm_region_find(&consistent_head, (unsigned long)cpu_addr);
375         if (!c)
376                 goto no_area;
377
378         c->vm_active = 0;
379         spin_unlock_irqrestore(&consistent_lock, flags);
380
381         if ((c->vm_end - c->vm_start) != size) {
382                 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
383                        __func__, c->vm_end - c->vm_start, size);
384                 dump_stack();
385                 size = c->vm_end - c->vm_start;
386         }
387
388         idx = CONSISTENT_PTE_INDEX(c->vm_start);
389         off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1);
390         ptep = consistent_pte[idx] + off;
391         addr = c->vm_start;
392         do {
393                 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
394                 unsigned long pfn;
395
396                 ptep++;
397                 addr += PAGE_SIZE;
398                 off++;
399                 if (off >= PTRS_PER_PTE) {
400                         off = 0;
401                         ptep = consistent_pte[++idx];
402                 }
403
404                 if (!pte_none(pte) && pte_present(pte)) {
405                         pfn = pte_pfn(pte);
406
407                         if (pfn_valid(pfn)) {
408                                 struct page *page = pfn_to_page(pfn);
409
410                                 /*
411                                  * x86 does not mark the pages reserved...
412                                  */
413                                 ClearPageReserved(page);
414
415                                 __free_page(page);
416                                 continue;
417                         }
418                 }
419
420                 printk(KERN_CRIT "%s: bad page in kernel page table\n",
421                        __func__);
422         } while (size -= PAGE_SIZE);
423
424         flush_tlb_kernel_range(c->vm_start, c->vm_end);
425
426         spin_lock_irqsave(&consistent_lock, flags);
427         list_del(&c->vm_list);
428         spin_unlock_irqrestore(&consistent_lock, flags);
429
430         kfree(c);
431         return;
432
433  no_area:
434         spin_unlock_irqrestore(&consistent_lock, flags);
435         printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
436                __func__, cpu_addr);
437         dump_stack();
438 }
439 EXPORT_SYMBOL(dma_free_coherent);
440
441 /*
442  * Initialise the consistent memory allocation.
443  */
444 static int __init consistent_init(void)
445 {
446         pgd_t *pgd;
447         pmd_t *pmd;
448         pte_t *pte;
449         int ret = 0, i = 0;
450         u32 base = CONSISTENT_BASE;
451
452         do {
453                 pgd = pgd_offset(&init_mm, base);
454                 pmd = pmd_alloc(&init_mm, pgd, base);
455                 if (!pmd) {
456                         printk(KERN_ERR "%s: no pmd tables\n", __func__);
457                         ret = -ENOMEM;
458                         break;
459                 }
460                 WARN_ON(!pmd_none(*pmd));
461
462                 pte = pte_alloc_kernel(pmd, base);
463                 if (!pte) {
464                         printk(KERN_ERR "%s: no pte tables\n", __func__);
465                         ret = -ENOMEM;
466                         break;
467                 }
468
469                 consistent_pte[i++] = pte;
470                 base += (1 << PGDIR_SHIFT);
471         } while (base < CONSISTENT_END);
472
473         return ret;
474 }
475
476 core_initcall(consistent_init);
477
478 /*
479  * Make an area consistent for devices.
480  * Note: Drivers should NOT use this function directly, as it will break
481  * platforms with CONFIG_DMABOUNCE.
482  * Use the driver DMA support - see dma-mapping.h (dma_sync_*)
483  */
484 void consistent_sync(const void *start, size_t size, int direction)
485 {
486         const void *end = start + size;
487
488         BUG_ON(!virt_addr_valid(start) || !virt_addr_valid(end));
489
490         switch (direction) {
491         case DMA_FROM_DEVICE:           /* invalidate only */
492                 dmac_inv_range(start, end);
493                 outer_inv_range(__pa(start), __pa(end));
494                 break;
495         case DMA_TO_DEVICE:             /* writeback only */
496                 dmac_clean_range(start, end);
497                 outer_clean_range(__pa(start), __pa(end));
498                 break;
499         case DMA_BIDIRECTIONAL:         /* writeback and invalidate */
500                 dmac_flush_range(start, end);
501                 outer_flush_range(__pa(start), __pa(end));
502                 break;
503         default:
504                 BUG();
505         }
506 }
507 EXPORT_SYMBOL(consistent_sync);