Linux Kernel  3.7.1
 All Data Structures Namespaces Files Functions Variables Typedefs Enumerations Enumerator Macros Groups Pages
dma-noncoherent.c
Go to the documentation of this file.
1 /*
2  * PowerPC version derived from arch/arm/mm/consistent.c
3  * Copyright (C) 2001 Dan Malek ([email protected])
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 
37 #include "mmu_decl.h"
38 
39 /*
40  * This address range defaults to a value that is safe for all
41  * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It
42  * can be further configured for specific applications under
43  * the "Advanced Setup" menu. -Matt
44  */
45 #define CONSISTENT_BASE (IOREMAP_TOP)
46 #define CONSISTENT_END (CONSISTENT_BASE + CONFIG_CONSISTENT_SIZE)
47 #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
48 
49 /*
50  * This is the page table (2MB) covering uncached, DMA consistent allocations
51  */
52 static DEFINE_SPINLOCK(consistent_lock);
53 
54 /*
55  * VM region handling support.
56  *
57  * This should become something generic, handling VM region allocations for
58  * vmalloc and similar (ioremap, module space, etc).
59  *
60  * I envisage vmalloc()'s supporting vm_struct becoming:
61  *
62  * struct vm_struct {
63  * struct vm_region region;
64  * unsigned long flags;
65  * struct page **pages;
66  * unsigned int nr_pages;
67  * unsigned long phys_addr;
68  * };
69  *
70  * get_vm_area() would then call vm_region_alloc with an appropriate
71  * struct vm_region head (eg):
72  *
73  * struct vm_region vmalloc_head = {
74  * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list),
75  * .vm_start = VMALLOC_START,
76  * .vm_end = VMALLOC_END,
77  * };
78  *
79  * However, vmalloc_head.vm_start is variable (typically, it is dependent on
80  * the amount of RAM found at boot time.) I would imagine that get_vm_area()
81  * would have to initialise this each time prior to calling vm_region_alloc().
82  */
83 struct ppc_vm_region {
85  unsigned long vm_start;
86  unsigned long vm_end;
87 };
88 
89 static struct ppc_vm_region consistent_head = {
90  .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
91  .vm_start = CONSISTENT_BASE,
92  .vm_end = CONSISTENT_END,
93 };
94 
95 static struct ppc_vm_region *
96 ppc_vm_region_alloc(struct ppc_vm_region *head, size_t size, gfp_t gfp)
97 {
98  unsigned long addr = head->vm_start, end = head->vm_end - size;
99  unsigned long flags;
100  struct ppc_vm_region *c, *new;
101 
102  new = kmalloc(sizeof(struct ppc_vm_region), gfp);
103  if (!new)
104  goto out;
105 
106  spin_lock_irqsave(&consistent_lock, flags);
107 
108  list_for_each_entry(c, &head->vm_list, vm_list) {
109  if ((addr + size) < addr)
110  goto nospc;
111  if ((addr + size) <= c->vm_start)
112  goto found;
113  addr = c->vm_end;
114  if (addr > end)
115  goto nospc;
116  }
117 
118  found:
119  /*
120  * Insert this entry _before_ the one we found.
121  */
122  list_add_tail(&new->vm_list, &c->vm_list);
123  new->vm_start = addr;
124  new->vm_end = addr + size;
125 
126  spin_unlock_irqrestore(&consistent_lock, flags);
127  return new;
128 
129  nospc:
130  spin_unlock_irqrestore(&consistent_lock, flags);
131  kfree(new);
132  out:
133  return NULL;
134 }
135 
136 static struct ppc_vm_region *ppc_vm_region_find(struct ppc_vm_region *head, unsigned long addr)
137 {
138  struct ppc_vm_region *c;
139 
140  list_for_each_entry(c, &head->vm_list, vm_list) {
141  if (c->vm_start == addr)
142  goto out;
143  }
144  c = NULL;
145  out:
146  return c;
147 }
148 
149 /*
150  * Allocate DMA-coherent memory space and return both the kernel remapped
151  * virtual and bus address for that space.
152  */
153 void *
154 __dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
155 {
156  struct page *page;
157  struct ppc_vm_region *c;
158  unsigned long order;
160 
161  if (dev) {
162  mask = dev->coherent_dma_mask;
163 
164  /*
165  * Sanity check the DMA mask - it must be non-zero, and
166  * must be able to be satisfied by a DMA allocation.
167  */
168  if (mask == 0) {
169  dev_warn(dev, "coherent DMA mask is unset\n");
170  goto no_page;
171  }
172 
173  if ((~mask) & ISA_DMA_THRESHOLD) {
174  dev_warn(dev, "coherent DMA mask %#llx is smaller "
175  "than system GFP_DMA mask %#llx\n",
176  mask, (unsigned long long)ISA_DMA_THRESHOLD);
177  goto no_page;
178  }
179  }
180 
181 
182  size = PAGE_ALIGN(size);
183  limit = (mask + 1) & ~mask;
184  if ((limit && size >= limit) ||
185  size >= (CONSISTENT_END - CONSISTENT_BASE)) {
186  printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n",
187  size, mask);
188  return NULL;
189  }
190 
191  order = get_order(size);
192 
193  /* Might be useful if we ever have a real legacy DMA zone... */
194  if (mask != 0xffffffff)
195  gfp |= GFP_DMA;
196 
197  page = alloc_pages(gfp, order);
198  if (!page)
199  goto no_page;
200 
201  /*
202  * Invalidate any data that might be lurking in the
203  * kernel direct-mapped region for device DMA.
204  */
205  {
206  unsigned long kaddr = (unsigned long)page_address(page);
207  memset(page_address(page), 0, size);
208  flush_dcache_range(kaddr, kaddr + size);
209  }
210 
211  /*
212  * Allocate a virtual address in the consistent mapping region.
213  */
214  c = ppc_vm_region_alloc(&consistent_head, size,
215  gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
216  if (c) {
217  unsigned long vaddr = c->vm_start;
218  struct page *end = page + (1 << order);
219 
220  split_page(page, order);
221 
222  /*
223  * Set the "dma handle"
224  */
225  *handle = page_to_phys(page);
226 
227  do {
228  SetPageReserved(page);
229  map_page(vaddr, page_to_phys(page),
231  page++;
232  vaddr += PAGE_SIZE;
233  } while (size -= PAGE_SIZE);
234 
235  /*
236  * Free the otherwise unused pages.
237  */
238  while (page < end) {
239  __free_page(page);
240  page++;
241  }
242 
243  return (void *)c->vm_start;
244  }
245 
246  if (page)
247  __free_pages(page, order);
248  no_page:
249  return NULL;
250 }
252 
253 /*
254  * free a page as defined by the above mapping.
255  */
256 void __dma_free_coherent(size_t size, void *vaddr)
257 {
258  struct ppc_vm_region *c;
259  unsigned long flags, addr;
260 
261  size = PAGE_ALIGN(size);
262 
263  spin_lock_irqsave(&consistent_lock, flags);
264 
265  c = ppc_vm_region_find(&consistent_head, (unsigned long)vaddr);
266  if (!c)
267  goto no_area;
268 
269  if ((c->vm_end - c->vm_start) != size) {
270  printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
271  __func__, c->vm_end - c->vm_start, size);
272  dump_stack();
273  size = c->vm_end - c->vm_start;
274  }
275 
276  addr = c->vm_start;
277  do {
278  pte_t *ptep;
279  unsigned long pfn;
280 
282  addr),
283  addr),
284  addr);
285  if (!pte_none(*ptep) && pte_present(*ptep)) {
286  pfn = pte_pfn(*ptep);
287  pte_clear(&init_mm, addr, ptep);
288  if (pfn_valid(pfn)) {
289  struct page *page = pfn_to_page(pfn);
290 
291  ClearPageReserved(page);
292  __free_page(page);
293  }
294  }
295  addr += PAGE_SIZE;
296  } while (size -= PAGE_SIZE);
297 
299 
300  list_del(&c->vm_list);
301 
302  spin_unlock_irqrestore(&consistent_lock, flags);
303 
304  kfree(c);
305  return;
306 
307  no_area:
308  spin_unlock_irqrestore(&consistent_lock, flags);
309  printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
310  __func__, vaddr);
311  dump_stack();
312 }
314 
315 /*
316  * make an area consistent.
317  */
318 void __dma_sync(void *vaddr, size_t size, int direction)
319 {
320  unsigned long start = (unsigned long)vaddr;
321  unsigned long end = start + size;
322 
323  switch (direction) {
324  case DMA_NONE:
325  BUG();
326  case DMA_FROM_DEVICE:
327  /*
328  * invalidate only when cache-line aligned otherwise there is
329  * the potential for discarding uncommitted data from the cache
330  */
331  if ((start & (L1_CACHE_BYTES - 1)) || (size & (L1_CACHE_BYTES - 1)))
332  flush_dcache_range(start, end);
333  else
334  invalidate_dcache_range(start, end);
335  break;
336  case DMA_TO_DEVICE: /* writeback only */
337  clean_dcache_range(start, end);
338  break;
339  case DMA_BIDIRECTIONAL: /* writeback and invalidate */
340  flush_dcache_range(start, end);
341  break;
342  }
343 }
345 
346 #ifdef CONFIG_HIGHMEM
347 /*
348  * __dma_sync_page() implementation for systems using highmem.
349  * In this case, each page of a buffer must be kmapped/kunmapped
350  * in order to have a virtual address for __dma_sync(). This must
351  * not sleep so kmap_atomic()/kunmap_atomic() are used.
352  *
353  * Note: yes, it is possible and correct to have a buffer extend
354  * beyond the first page.
355  */
356 static inline void __dma_sync_page_highmem(struct page *page,
357  unsigned long offset, size_t size, int direction)
358 {
359  size_t seg_size = min((size_t)(PAGE_SIZE - offset), size);
360  size_t cur_size = seg_size;
361  unsigned long flags, start, seg_offset = offset;
362  int nr_segs = 1 + ((size - seg_size) + PAGE_SIZE - 1)/PAGE_SIZE;
363  int seg_nr = 0;
364 
365  local_irq_save(flags);
366 
367  do {
368  start = (unsigned long)kmap_atomic(page + seg_nr) + seg_offset;
369 
370  /* Sync this buffer segment */
371  __dma_sync((void *)start, seg_size, direction);
372  kunmap_atomic((void *)start);
373  seg_nr++;
374 
375  /* Calculate next buffer segment size */
376  seg_size = min((size_t)PAGE_SIZE, size - cur_size);
377 
378  /* Add the segment size to our running total */
379  cur_size += seg_size;
380  seg_offset = 0;
381  } while (seg_nr < nr_segs);
382 
383  local_irq_restore(flags);
384 }
385 #endif /* CONFIG_HIGHMEM */
386 
387 /*
388  * __dma_sync_page makes memory consistent. identical to __dma_sync, but
389  * takes a struct page instead of a virtual address
390  */
391 void __dma_sync_page(struct page *page, unsigned long offset,
392  size_t size, int direction)
393 {
394 #ifdef CONFIG_HIGHMEM
395  __dma_sync_page_highmem(page, offset, size, direction);
396 #else
397  unsigned long start = (unsigned long)page_address(page) + offset;
398  __dma_sync((void *)start, size, direction);
399 #endif
400 }
402 
403 /*
404  * Return the PFN for a given cpu virtual address returned by
405  * __dma_alloc_coherent. This is used by dma_mmap_coherent()
406  */
407 unsigned long __dma_get_coherent_pfn(unsigned long cpu_addr)
408 {
409  /* This should always be populated, so we don't test every
410  * level. If that fails, we'll have a nice crash which
411  * will be as good as a BUG_ON()
412  */
413  pgd_t *pgd = pgd_offset_k(cpu_addr);
414  pud_t *pud = pud_offset(pgd, cpu_addr);
415  pmd_t *pmd = pmd_offset(pud, cpu_addr);
416  pte_t *ptep = pte_offset_kernel(pmd, cpu_addr);
417 
418  if (pte_none(*ptep) || !pte_present(*ptep))
419  return 0;
420  return pte_pfn(*ptep);
421 }