1 /*
2 * PowerPC version derived from arch/arm/mm/consistent.c
3 * Copyright (C) 2001 Dan Malek (dmalek@jlc.net)
4 *
5 * Copyright (C) 2000 Russell King
6 *
7 * Consistent memory allocators. Used for DMA devices that want to
8 * share uncached memory with the processor core. The function return
9 * is the virtual address and 'dma_handle' is the physical address.
10 * Mostly stolen from the ARM port, with some changes for PowerPC.
11 * -- Dan
12 *
13 * Reorganized to get rid of the arch-specific consistent_* functions
14 * and provide non-coherent implementations for the DMA API. -Matt
15 *
16 * Added in_interrupt() safe dma_alloc_coherent()/dma_free_coherent()
17 * implementation. This is pulled straight from ARM and barely
18 * modified. -Matt
19 *
20 * This program is free software; you can redistribute it and/or modify
21 * it under the terms of the GNU General Public License version 2 as
22 * published by the Free Software Foundation.
23 */
24
25 #include <linux/sched.h>
26 #include <linux/slab.h>
27 #include <linux/kernel.h>
28 #include <linux/errno.h>
29 #include <linux/string.h>
30 #include <linux/types.h>
31 #include <linux/highmem.h>
32 #include <linux/dma-mapping.h>
33 #include <linux/export.h>
34
35 #include <asm/tlbflush.h>
36 #include <asm/dma.h>
37
38 #include "mmu_decl.h"
39
40 /*
41 * This address range defaults to a value that is safe for all
42 * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It
43 * can be further configured for specific applications under
44 * the "Advanced Setup" menu. -Matt
45 */
46 #define CONSISTENT_BASE (IOREMAP_TOP)
47 #define CONSISTENT_END (CONSISTENT_BASE + CONFIG_CONSISTENT_SIZE)
48 #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
49
50 /*
51 * This is the page table (2MB) covering uncached, DMA consistent allocations
52 */
53 static DEFINE_SPINLOCK(consistent_lock);
54
55 /*
56 * VM region handling support.
57 *
58 * This should become something generic, handling VM region allocations for
59 * vmalloc and similar (ioremap, module space, etc).
60 *
61 * I envisage vmalloc()'s supporting vm_struct becoming:
62 *
63 * struct vm_struct {
64 * struct vm_region region;
65 * unsigned long flags;
66 * struct page **pages;
67 * unsigned int nr_pages;
68 * unsigned long phys_addr;
69 * };
70 *
71 * get_vm_area() would then call vm_region_alloc with an appropriate
72 * struct vm_region head (eg):
73 *
74 * struct vm_region vmalloc_head = {
75 * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list),
76 * .vm_start = VMALLOC_START,
77 * .vm_end = VMALLOC_END,
78 * };
79 *
80 * However, vmalloc_head.vm_start is variable (typically, it is dependent on
81 * the amount of RAM found at boot time.) I would imagine that get_vm_area()
82 * would have to initialise this each time prior to calling vm_region_alloc().
83 */
84 struct ppc_vm_region {
85 struct list_head vm_list;
86 unsigned long vm_start;
87 unsigned long vm_end;
88 };
89
90 static struct ppc_vm_region consistent_head = {
91 .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
92 .vm_start = CONSISTENT_BASE,
93 .vm_end = CONSISTENT_END,
94 };
95
96 static struct ppc_vm_region *
ppc_vm_region_alloc(struct ppc_vm_region * head,size_t size,gfp_t gfp)97 ppc_vm_region_alloc(struct ppc_vm_region *head, size_t size, gfp_t gfp)
98 {
99 unsigned long addr = head->vm_start, end = head->vm_end - size;
100 unsigned long flags;
101 struct ppc_vm_region *c, *new;
102
103 new = kmalloc(sizeof(struct ppc_vm_region), gfp);
104 if (!new)
105 goto out;
106
107 spin_lock_irqsave(&consistent_lock, flags);
108
109 list_for_each_entry(c, &head->vm_list, vm_list) {
110 if ((addr + size) < addr)
111 goto nospc;
112 if ((addr + size) <= c->vm_start)
113 goto found;
114 addr = c->vm_end;
115 if (addr > end)
116 goto nospc;
117 }
118
119 found:
120 /*
121 * Insert this entry _before_ the one we found.
122 */
123 list_add_tail(&new->vm_list, &c->vm_list);
124 new->vm_start = addr;
125 new->vm_end = addr + size;
126
127 spin_unlock_irqrestore(&consistent_lock, flags);
128 return new;
129
130 nospc:
131 spin_unlock_irqrestore(&consistent_lock, flags);
132 kfree(new);
133 out:
134 return NULL;
135 }
136
ppc_vm_region_find(struct ppc_vm_region * head,unsigned long addr)137 static struct ppc_vm_region *ppc_vm_region_find(struct ppc_vm_region *head, unsigned long addr)
138 {
139 struct ppc_vm_region *c;
140
141 list_for_each_entry(c, &head->vm_list, vm_list) {
142 if (c->vm_start == addr)
143 goto out;
144 }
145 c = NULL;
146 out:
147 return c;
148 }
149
150 /*
151 * Allocate DMA-coherent memory space and return both the kernel remapped
152 * virtual and bus address for that space.
153 */
154 void *
__dma_alloc_coherent(struct device * dev,size_t size,dma_addr_t * handle,gfp_t gfp)155 __dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
156 {
157 struct page *page;
158 struct ppc_vm_region *c;
159 unsigned long order;
160 u64 mask = ISA_DMA_THRESHOLD, limit;
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 size = PAGE_ALIGN(size);
184 limit = (mask + 1) & ~mask;
185 if ((limit && size >= limit) ||
186 size >= (CONSISTENT_END - CONSISTENT_BASE)) {
187 printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n",
188 size, mask);
189 return NULL;
190 }
191
192 order = get_order(size);
193
194 /* Might be useful if we ever have a real legacy DMA zone... */
195 if (mask != 0xffffffff)
196 gfp |= GFP_DMA;
197
198 page = alloc_pages(gfp, order);
199 if (!page)
200 goto no_page;
201
202 /*
203 * Invalidate any data that might be lurking in the
204 * kernel direct-mapped region for device DMA.
205 */
206 {
207 unsigned long kaddr = (unsigned long)page_address(page);
208 memset(page_address(page), 0, size);
209 flush_dcache_range(kaddr, kaddr + size);
210 }
211
212 /*
213 * Allocate a virtual address in the consistent mapping region.
214 */
215 c = ppc_vm_region_alloc(&consistent_head, size,
216 gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
217 if (c) {
218 unsigned long vaddr = c->vm_start;
219 struct page *end = page + (1 << order);
220
221 split_page(page, order);
222
223 /*
224 * Set the "dma handle"
225 */
226 *handle = page_to_phys(page);
227
228 do {
229 SetPageReserved(page);
230 map_page(vaddr, page_to_phys(page),
231 pgprot_noncached(PAGE_KERNEL));
232 page++;
233 vaddr += PAGE_SIZE;
234 } while (size -= PAGE_SIZE);
235
236 /*
237 * Free the otherwise unused pages.
238 */
239 while (page < end) {
240 __free_page(page);
241 page++;
242 }
243
244 return (void *)c->vm_start;
245 }
246
247 if (page)
248 __free_pages(page, order);
249 no_page:
250 return NULL;
251 }
252 EXPORT_SYMBOL(__dma_alloc_coherent);
253
254 /*
255 * free a page as defined by the above mapping.
256 */
__dma_free_coherent(size_t size,void * vaddr)257 void __dma_free_coherent(size_t size, void *vaddr)
258 {
259 struct ppc_vm_region *c;
260 unsigned long flags, addr;
261
262 size = PAGE_ALIGN(size);
263
264 spin_lock_irqsave(&consistent_lock, flags);
265
266 c = ppc_vm_region_find(&consistent_head, (unsigned long)vaddr);
267 if (!c)
268 goto no_area;
269
270 if ((c->vm_end - c->vm_start) != size) {
271 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
272 __func__, c->vm_end - c->vm_start, size);
273 dump_stack();
274 size = c->vm_end - c->vm_start;
275 }
276
277 addr = c->vm_start;
278 do {
279 pte_t *ptep;
280 unsigned long pfn;
281
282 ptep = pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(addr),
283 addr),
284 addr),
285 addr);
286 if (!pte_none(*ptep) && pte_present(*ptep)) {
287 pfn = pte_pfn(*ptep);
288 pte_clear(&init_mm, addr, ptep);
289 if (pfn_valid(pfn)) {
290 struct page *page = pfn_to_page(pfn);
291 __free_reserved_page(page);
292 }
293 }
294 addr += PAGE_SIZE;
295 } while (size -= PAGE_SIZE);
296
297 flush_tlb_kernel_range(c->vm_start, c->vm_end);
298
299 list_del(&c->vm_list);
300
301 spin_unlock_irqrestore(&consistent_lock, flags);
302
303 kfree(c);
304 return;
305
306 no_area:
307 spin_unlock_irqrestore(&consistent_lock, flags);
308 printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
309 __func__, vaddr);
310 dump_stack();
311 }
312 EXPORT_SYMBOL(__dma_free_coherent);
313
314 /*
315 * make an area consistent.
316 */
__dma_sync(void * vaddr,size_t size,int direction)317 void __dma_sync(void *vaddr, size_t size, int direction)
318 {
319 unsigned long start = (unsigned long)vaddr;
320 unsigned long end = start + size;
321
322 switch (direction) {
323 case DMA_NONE:
324 BUG();
325 case DMA_FROM_DEVICE:
326 /*
327 * invalidate only when cache-line aligned otherwise there is
328 * the potential for discarding uncommitted data from the cache
329 */
330 if ((start & (L1_CACHE_BYTES - 1)) || (size & (L1_CACHE_BYTES - 1)))
331 flush_dcache_range(start, end);
332 else
333 invalidate_dcache_range(start, end);
334 break;
335 case DMA_TO_DEVICE: /* writeback only */
336 clean_dcache_range(start, end);
337 break;
338 case DMA_BIDIRECTIONAL: /* writeback and invalidate */
339 flush_dcache_range(start, end);
340 break;
341 }
342 }
343 EXPORT_SYMBOL(__dma_sync);
344
345 #ifdef CONFIG_HIGHMEM
346 /*
347 * __dma_sync_page() implementation for systems using highmem.
348 * In this case, each page of a buffer must be kmapped/kunmapped
349 * in order to have a virtual address for __dma_sync(). This must
350 * not sleep so kmap_atomic()/kunmap_atomic() are used.
351 *
352 * Note: yes, it is possible and correct to have a buffer extend
353 * beyond the first page.
354 */
__dma_sync_page_highmem(struct page * page,unsigned long offset,size_t size,int direction)355 static inline void __dma_sync_page_highmem(struct page *page,
356 unsigned long offset, size_t size, int direction)
357 {
358 size_t seg_size = min((size_t)(PAGE_SIZE - offset), size);
359 size_t cur_size = seg_size;
360 unsigned long flags, start, seg_offset = offset;
361 int nr_segs = 1 + ((size - seg_size) + PAGE_SIZE - 1)/PAGE_SIZE;
362 int seg_nr = 0;
363
364 local_irq_save(flags);
365
366 do {
367 start = (unsigned long)kmap_atomic(page + seg_nr) + seg_offset;
368
369 /* Sync this buffer segment */
370 __dma_sync((void *)start, seg_size, direction);
371 kunmap_atomic((void *)start);
372 seg_nr++;
373
374 /* Calculate next buffer segment size */
375 seg_size = min((size_t)PAGE_SIZE, size - cur_size);
376
377 /* Add the segment size to our running total */
378 cur_size += seg_size;
379 seg_offset = 0;
380 } while (seg_nr < nr_segs);
381
382 local_irq_restore(flags);
383 }
384 #endif /* CONFIG_HIGHMEM */
385
386 /*
387 * __dma_sync_page makes memory consistent. identical to __dma_sync, but
388 * takes a struct page instead of a virtual address
389 */
__dma_sync_page(struct page * page,unsigned long offset,size_t size,int direction)390 void __dma_sync_page(struct page *page, unsigned long offset,
391 size_t size, int direction)
392 {
393 #ifdef CONFIG_HIGHMEM
394 __dma_sync_page_highmem(page, offset, size, direction);
395 #else
396 unsigned long start = (unsigned long)page_address(page) + offset;
397 __dma_sync((void *)start, size, direction);
398 #endif
399 }
400 EXPORT_SYMBOL(__dma_sync_page);
401
402 /*
403 * Return the PFN for a given cpu virtual address returned by
404 * __dma_alloc_coherent. This is used by dma_mmap_coherent()
405 */
__dma_get_coherent_pfn(unsigned long cpu_addr)406 unsigned long __dma_get_coherent_pfn(unsigned long cpu_addr)
407 {
408 /* This should always be populated, so we don't test every
409 * level. If that fails, we'll have a nice crash which
410 * will be as good as a BUG_ON()
411 */
412 pgd_t *pgd = pgd_offset_k(cpu_addr);
413 pud_t *pud = pud_offset(pgd, cpu_addr);
414 pmd_t *pmd = pmd_offset(pud, cpu_addr);
415 pte_t *ptep = pte_offset_kernel(pmd, cpu_addr);
416
417 if (pte_none(*ptep) || !pte_present(*ptep))
418 return 0;
419 return pte_pfn(*ptep);
420 }
421