diff options
author | Russell King <rmk+kernel@arm.linux.org.uk> | 2009-12-04 10:00:00 -0500 |
---|---|---|
committer | Russell King <rmk+kernel@arm.linux.org.uk> | 2009-12-04 10:00:00 -0500 |
commit | c6baa1963c2a76ffdb157e8b9a5a55b30046b125 (patch) | |
tree | 9063398d2406042670759e207d9058ec5e7b770b /arch/arm/mm | |
parent | 5cb2faa6ede7ada9cb2bffc832c4ce60f53d6834 (diff) | |
parent | 26a26d329688ab018e068b412b03d43d7c299f0a (diff) |
Merge branch 'pending-dma-coherent' into devel
Diffstat (limited to 'arch/arm/mm')
-rw-r--r-- | arch/arm/mm/Makefile | 2 | ||||
-rw-r--r-- | arch/arm/mm/dma-mapping.c | 497 | ||||
-rw-r--r-- | arch/arm/mm/vmregion.c | 131 | ||||
-rw-r--r-- | arch/arm/mm/vmregion.h | 29 |
4 files changed, 342 insertions, 317 deletions
diff --git a/arch/arm/mm/Makefile b/arch/arm/mm/Makefile index 055cb2aa8134..42352e75742b 100644 --- a/arch/arm/mm/Makefile +++ b/arch/arm/mm/Makefile | |||
@@ -6,7 +6,7 @@ obj-y := dma-mapping.o extable.o fault.o init.o \ | |||
6 | iomap.o | 6 | iomap.o |
7 | 7 | ||
8 | obj-$(CONFIG_MMU) += fault-armv.o flush.o ioremap.o mmap.o \ | 8 | obj-$(CONFIG_MMU) += fault-armv.o flush.o ioremap.o mmap.o \ |
9 | pgd.o mmu.o | 9 | pgd.o mmu.o vmregion.o |
10 | 10 | ||
11 | ifneq ($(CONFIG_MMU),y) | 11 | ifneq ($(CONFIG_MMU),y) |
12 | obj-y += nommu.o | 12 | obj-y += nommu.o |
diff --git a/arch/arm/mm/dma-mapping.c b/arch/arm/mm/dma-mapping.c index b9590a7085ca..26325cb5d368 100644 --- a/arch/arm/mm/dma-mapping.c +++ b/arch/arm/mm/dma-mapping.c | |||
@@ -63,194 +63,152 @@ static u64 get_coherent_dma_mask(struct device *dev) | |||
63 | return mask; | 63 | return mask; |
64 | } | 64 | } |
65 | 65 | ||
66 | #ifdef CONFIG_MMU | ||
67 | /* | 66 | /* |
68 | * These are the page tables (2MB each) covering uncached, DMA consistent allocations | 67 | * Allocate a DMA buffer for 'dev' of size 'size' using the |
68 | * specified gfp mask. Note that 'size' must be page aligned. | ||
69 | */ | 69 | */ |
70 | static pte_t *consistent_pte[NUM_CONSISTENT_PTES]; | 70 | static struct page *__dma_alloc_buffer(struct device *dev, size_t size, gfp_t gfp) |
71 | static DEFINE_SPINLOCK(consistent_lock); | 71 | { |
72 | unsigned long order = get_order(size); | ||
73 | struct page *page, *p, *e; | ||
74 | void *ptr; | ||
75 | u64 mask = get_coherent_dma_mask(dev); | ||
72 | 76 | ||
73 | /* | 77 | #ifdef CONFIG_DMA_API_DEBUG |
74 | * VM region handling support. | 78 | u64 limit = (mask + 1) & ~mask; |
75 | * | 79 | if (limit && size >= limit) { |
76 | * This should become something generic, handling VM region allocations for | 80 | dev_warn(dev, "coherent allocation too big (requested %#x mask %#llx)\n", |
77 | * vmalloc and similar (ioremap, module space, etc). | 81 | size, mask); |
78 | * | 82 | return NULL; |
79 | * I envisage vmalloc()'s supporting vm_struct becoming: | 83 | } |
80 | * | 84 | #endif |
81 | * struct vm_struct { | ||
82 | * struct vm_region region; | ||
83 | * unsigned long flags; | ||
84 | * struct page **pages; | ||
85 | * unsigned int nr_pages; | ||
86 | * unsigned long phys_addr; | ||
87 | * }; | ||
88 | * | ||
89 | * get_vm_area() would then call vm_region_alloc with an appropriate | ||
90 | * struct vm_region head (eg): | ||
91 | * | ||
92 | * struct vm_region vmalloc_head = { | ||
93 | * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list), | ||
94 | * .vm_start = VMALLOC_START, | ||
95 | * .vm_end = VMALLOC_END, | ||
96 | * }; | ||
97 | * | ||
98 | * However, vmalloc_head.vm_start is variable (typically, it is dependent on | ||
99 | * the amount of RAM found at boot time.) I would imagine that get_vm_area() | ||
100 | * would have to initialise this each time prior to calling vm_region_alloc(). | ||
101 | */ | ||
102 | struct arm_vm_region { | ||
103 | struct list_head vm_list; | ||
104 | unsigned long vm_start; | ||
105 | unsigned long vm_end; | ||
106 | struct page *vm_pages; | ||
107 | int vm_active; | ||
108 | }; | ||
109 | 85 | ||
110 | static struct arm_vm_region consistent_head = { | 86 | if (!mask) |
111 | .vm_list = LIST_HEAD_INIT(consistent_head.vm_list), | 87 | return NULL; |
112 | .vm_start = CONSISTENT_BASE, | ||
113 | .vm_end = CONSISTENT_END, | ||
114 | }; | ||
115 | 88 | ||
116 | static struct arm_vm_region * | 89 | if (mask < 0xffffffffULL) |
117 | arm_vm_region_alloc(struct arm_vm_region *head, size_t size, gfp_t gfp) | 90 | gfp |= GFP_DMA; |
118 | { | 91 | |
119 | unsigned long addr = head->vm_start, end = head->vm_end - size; | 92 | page = alloc_pages(gfp, order); |
120 | unsigned long flags; | 93 | if (!page) |
121 | struct arm_vm_region *c, *new; | 94 | return NULL; |
122 | |||
123 | new = kmalloc(sizeof(struct arm_vm_region), gfp); | ||
124 | if (!new) | ||
125 | goto out; | ||
126 | |||
127 | spin_lock_irqsave(&consistent_lock, flags); | ||
128 | |||
129 | list_for_each_entry(c, &head->vm_list, vm_list) { | ||
130 | if ((addr + size) < addr) | ||
131 | goto nospc; | ||
132 | if ((addr + size) <= c->vm_start) | ||
133 | goto found; | ||
134 | addr = c->vm_end; | ||
135 | if (addr > end) | ||
136 | goto nospc; | ||
137 | } | ||
138 | 95 | ||
139 | found: | ||
140 | /* | 96 | /* |
141 | * Insert this entry _before_ the one we found. | 97 | * Now split the huge page and free the excess pages |
142 | */ | 98 | */ |
143 | list_add_tail(&new->vm_list, &c->vm_list); | 99 | split_page(page, order); |
144 | new->vm_start = addr; | 100 | for (p = page + (size >> PAGE_SHIFT), e = page + (1 << order); p < e; p++) |
145 | new->vm_end = addr + size; | 101 | __free_page(p); |
146 | new->vm_active = 1; | 102 | |
147 | 103 | /* | |
148 | spin_unlock_irqrestore(&consistent_lock, flags); | 104 | * Ensure that the allocated pages are zeroed, and that any data |
149 | return new; | 105 | * lurking in the kernel direct-mapped region is invalidated. |
150 | 106 | */ | |
151 | nospc: | 107 | ptr = page_address(page); |
152 | spin_unlock_irqrestore(&consistent_lock, flags); | 108 | memset(ptr, 0, size); |
153 | kfree(new); | 109 | dmac_flush_range(ptr, ptr + size); |
154 | out: | 110 | outer_flush_range(__pa(ptr), __pa(ptr) + size); |
155 | return NULL; | 111 | |
112 | return page; | ||
156 | } | 113 | } |
157 | 114 | ||
158 | static struct arm_vm_region *arm_vm_region_find(struct arm_vm_region *head, unsigned long addr) | 115 | /* |
116 | * Free a DMA buffer. 'size' must be page aligned. | ||
117 | */ | ||
118 | static void __dma_free_buffer(struct page *page, size_t size) | ||
159 | { | 119 | { |
160 | struct arm_vm_region *c; | 120 | struct page *e = page + (size >> PAGE_SHIFT); |
161 | 121 | ||
162 | list_for_each_entry(c, &head->vm_list, vm_list) { | 122 | while (page < e) { |
163 | if (c->vm_active && c->vm_start == addr) | 123 | __free_page(page); |
164 | goto out; | 124 | page++; |
165 | } | 125 | } |
166 | c = NULL; | ||
167 | out: | ||
168 | return c; | ||
169 | } | 126 | } |
170 | 127 | ||
128 | #ifdef CONFIG_MMU | ||
129 | /* | ||
130 | * These are the page tables (2MB each) covering uncached, DMA consistent allocations | ||
131 | */ | ||
132 | static pte_t *consistent_pte[NUM_CONSISTENT_PTES]; | ||
133 | |||
134 | #include "vmregion.h" | ||
135 | |||
136 | static struct arm_vmregion_head consistent_head = { | ||
137 | .vm_lock = __SPIN_LOCK_UNLOCKED(&consistent_head.vm_lock), | ||
138 | .vm_list = LIST_HEAD_INIT(consistent_head.vm_list), | ||
139 | .vm_start = CONSISTENT_BASE, | ||
140 | .vm_end = CONSISTENT_END, | ||
141 | }; | ||
142 | |||
171 | #ifdef CONFIG_HUGETLB_PAGE | 143 | #ifdef CONFIG_HUGETLB_PAGE |
172 | #error ARM Coherent DMA allocator does not (yet) support huge TLB | 144 | #error ARM Coherent DMA allocator does not (yet) support huge TLB |
173 | #endif | 145 | #endif |
174 | 146 | ||
175 | static void * | 147 | /* |
176 | __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp, | 148 | * Initialise the consistent memory allocation. |
177 | pgprot_t prot) | 149 | */ |
150 | static int __init consistent_init(void) | ||
178 | { | 151 | { |
179 | struct page *page; | 152 | int ret = 0; |
180 | struct arm_vm_region *c; | 153 | pgd_t *pgd; |
181 | unsigned long order; | 154 | pmd_t *pmd; |
182 | u64 mask = get_coherent_dma_mask(dev); | 155 | pte_t *pte; |
183 | u64 limit; | 156 | int i = 0; |
157 | u32 base = CONSISTENT_BASE; | ||
184 | 158 | ||
185 | if (!consistent_pte[0]) { | 159 | do { |
186 | printk(KERN_ERR "%s: not initialised\n", __func__); | 160 | pgd = pgd_offset(&init_mm, base); |
187 | dump_stack(); | 161 | pmd = pmd_alloc(&init_mm, pgd, base); |
188 | return NULL; | 162 | if (!pmd) { |
189 | } | 163 | printk(KERN_ERR "%s: no pmd tables\n", __func__); |
164 | ret = -ENOMEM; | ||
165 | break; | ||
166 | } | ||
167 | WARN_ON(!pmd_none(*pmd)); | ||
190 | 168 | ||
191 | if (!mask) | 169 | pte = pte_alloc_kernel(pmd, base); |
192 | goto no_page; | 170 | if (!pte) { |
171 | printk(KERN_ERR "%s: no pte tables\n", __func__); | ||
172 | ret = -ENOMEM; | ||
173 | break; | ||
174 | } | ||
193 | 175 | ||
194 | /* | 176 | consistent_pte[i++] = pte; |
195 | * Sanity check the allocation size. | 177 | base += (1 << PGDIR_SHIFT); |
196 | */ | 178 | } while (base < CONSISTENT_END); |
197 | size = PAGE_ALIGN(size); | ||
198 | limit = (mask + 1) & ~mask; | ||
199 | if ((limit && size >= limit) || | ||
200 | size >= (CONSISTENT_END - CONSISTENT_BASE)) { | ||
201 | printk(KERN_WARNING "coherent allocation too big " | ||
202 | "(requested %#x mask %#llx)\n", size, mask); | ||
203 | goto no_page; | ||
204 | } | ||
205 | 179 | ||
206 | order = get_order(size); | 180 | return ret; |
181 | } | ||
207 | 182 | ||
208 | if (mask < 0xffffffffULL) | 183 | core_initcall(consistent_init); |
209 | gfp |= GFP_DMA; | ||
210 | 184 | ||
211 | page = alloc_pages(gfp, order); | 185 | static void * |
212 | if (!page) | 186 | __dma_alloc_remap(struct page *page, size_t size, gfp_t gfp, pgprot_t prot) |
213 | goto no_page; | 187 | { |
188 | struct arm_vmregion *c; | ||
214 | 189 | ||
215 | /* | 190 | if (!consistent_pte[0]) { |
216 | * Invalidate any data that might be lurking in the | 191 | printk(KERN_ERR "%s: not initialised\n", __func__); |
217 | * kernel direct-mapped region for device DMA. | 192 | dump_stack(); |
218 | */ | 193 | return NULL; |
219 | { | ||
220 | void *ptr = page_address(page); | ||
221 | memset(ptr, 0, size); | ||
222 | dmac_flush_range(ptr, ptr + size); | ||
223 | outer_flush_range(__pa(ptr), __pa(ptr) + size); | ||
224 | } | 194 | } |
225 | 195 | ||
226 | /* | 196 | /* |
227 | * Allocate a virtual address in the consistent mapping region. | 197 | * Allocate a virtual address in the consistent mapping region. |
228 | */ | 198 | */ |
229 | c = arm_vm_region_alloc(&consistent_head, size, | 199 | c = arm_vmregion_alloc(&consistent_head, size, |
230 | gfp & ~(__GFP_DMA | __GFP_HIGHMEM)); | 200 | gfp & ~(__GFP_DMA | __GFP_HIGHMEM)); |
231 | if (c) { | 201 | if (c) { |
232 | pte_t *pte; | 202 | pte_t *pte; |
233 | struct page *end = page + (1 << order); | ||
234 | int idx = CONSISTENT_PTE_INDEX(c->vm_start); | 203 | int idx = CONSISTENT_PTE_INDEX(c->vm_start); |
235 | u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1); | 204 | u32 off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1); |
236 | 205 | ||
237 | pte = consistent_pte[idx] + off; | 206 | pte = consistent_pte[idx] + off; |
238 | c->vm_pages = page; | 207 | c->vm_pages = page; |
239 | 208 | ||
240 | split_page(page, order); | ||
241 | |||
242 | /* | ||
243 | * Set the "dma handle" | ||
244 | */ | ||
245 | *handle = page_to_dma(dev, page); | ||
246 | |||
247 | do { | 209 | do { |
248 | BUG_ON(!pte_none(*pte)); | 210 | BUG_ON(!pte_none(*pte)); |
249 | 211 | ||
250 | /* | ||
251 | * x86 does not mark the pages reserved... | ||
252 | */ | ||
253 | SetPageReserved(page); | ||
254 | set_pte_ext(pte, mk_pte(page, prot), 0); | 212 | set_pte_ext(pte, mk_pte(page, prot), 0); |
255 | page++; | 213 | page++; |
256 | pte++; | 214 | pte++; |
@@ -261,48 +219,90 @@ __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp, | |||
261 | } | 219 | } |
262 | } while (size -= PAGE_SIZE); | 220 | } while (size -= PAGE_SIZE); |
263 | 221 | ||
264 | /* | ||
265 | * Free the otherwise unused pages. | ||
266 | */ | ||
267 | while (page < end) { | ||
268 | __free_page(page); | ||
269 | page++; | ||
270 | } | ||
271 | |||
272 | return (void *)c->vm_start; | 222 | return (void *)c->vm_start; |
273 | } | 223 | } |
274 | |||
275 | if (page) | ||
276 | __free_pages(page, order); | ||
277 | no_page: | ||
278 | *handle = ~0; | ||
279 | return NULL; | 224 | return NULL; |
280 | } | 225 | } |
226 | |||
227 | static void __dma_free_remap(void *cpu_addr, size_t size) | ||
228 | { | ||
229 | struct arm_vmregion *c; | ||
230 | unsigned long addr; | ||
231 | pte_t *ptep; | ||
232 | int idx; | ||
233 | u32 off; | ||
234 | |||
235 | c = arm_vmregion_find_remove(&consistent_head, (unsigned long)cpu_addr); | ||
236 | if (!c) { | ||
237 | printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n", | ||
238 | __func__, cpu_addr); | ||
239 | dump_stack(); | ||
240 | return; | ||
241 | } | ||
242 | |||
243 | if ((c->vm_end - c->vm_start) != size) { | ||
244 | printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n", | ||
245 | __func__, c->vm_end - c->vm_start, size); | ||
246 | dump_stack(); | ||
247 | size = c->vm_end - c->vm_start; | ||
248 | } | ||
249 | |||
250 | idx = CONSISTENT_PTE_INDEX(c->vm_start); | ||
251 | off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1); | ||
252 | ptep = consistent_pte[idx] + off; | ||
253 | addr = c->vm_start; | ||
254 | do { | ||
255 | pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep); | ||
256 | |||
257 | ptep++; | ||
258 | addr += PAGE_SIZE; | ||
259 | off++; | ||
260 | if (off >= PTRS_PER_PTE) { | ||
261 | off = 0; | ||
262 | ptep = consistent_pte[++idx]; | ||
263 | } | ||
264 | |||
265 | if (pte_none(pte) || !pte_present(pte)) | ||
266 | printk(KERN_CRIT "%s: bad page in kernel page table\n", | ||
267 | __func__); | ||
268 | } while (size -= PAGE_SIZE); | ||
269 | |||
270 | flush_tlb_kernel_range(c->vm_start, c->vm_end); | ||
271 | |||
272 | arm_vmregion_free(&consistent_head, c); | ||
273 | } | ||
274 | |||
281 | #else /* !CONFIG_MMU */ | 275 | #else /* !CONFIG_MMU */ |
276 | |||
277 | #define __dma_alloc_remap(page, size, gfp, prot) page_address(page) | ||
278 | #define __dma_free_remap(addr, size) do { } while (0) | ||
279 | |||
280 | #endif /* CONFIG_MMU */ | ||
281 | |||
282 | static void * | 282 | static void * |
283 | __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp, | 283 | __dma_alloc(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp, |
284 | pgprot_t prot) | 284 | pgprot_t prot) |
285 | { | 285 | { |
286 | void *virt; | 286 | struct page *page; |
287 | u64 mask = get_coherent_dma_mask(dev); | 287 | void *addr; |
288 | 288 | ||
289 | if (!mask) | 289 | *handle = ~0; |
290 | goto error; | 290 | size = PAGE_ALIGN(size); |
291 | 291 | ||
292 | if (mask < 0xffffffffULL) | 292 | page = __dma_alloc_buffer(dev, size, gfp); |
293 | gfp |= GFP_DMA; | 293 | if (!page) |
294 | virt = kmalloc(size, gfp); | 294 | return NULL; |
295 | if (!virt) | ||
296 | goto error; | ||
297 | 295 | ||
298 | *handle = virt_to_dma(dev, virt); | 296 | if (!arch_is_coherent()) |
299 | return virt; | 297 | addr = __dma_alloc_remap(page, size, gfp, prot); |
298 | else | ||
299 | addr = page_address(page); | ||
300 | 300 | ||
301 | error: | 301 | if (addr) |
302 | *handle = ~0; | 302 | *handle = page_to_dma(dev, page); |
303 | return NULL; | 303 | |
304 | return addr; | ||
304 | } | 305 | } |
305 | #endif /* CONFIG_MMU */ | ||
306 | 306 | ||
307 | /* | 307 | /* |
308 | * Allocate DMA-coherent memory space and return both the kernel remapped | 308 | * Allocate DMA-coherent memory space and return both the kernel remapped |
@@ -316,19 +316,8 @@ dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gf | |||
316 | if (dma_alloc_from_coherent(dev, size, handle, &memory)) | 316 | if (dma_alloc_from_coherent(dev, size, handle, &memory)) |
317 | return memory; | 317 | return memory; |
318 | 318 | ||
319 | if (arch_is_coherent()) { | ||
320 | void *virt; | ||
321 | |||
322 | virt = kmalloc(size, gfp); | ||
323 | if (!virt) | ||
324 | return NULL; | ||
325 | *handle = virt_to_dma(dev, virt); | ||
326 | |||
327 | return virt; | ||
328 | } | ||
329 | |||
330 | return __dma_alloc(dev, size, handle, gfp, | 319 | return __dma_alloc(dev, size, handle, gfp, |
331 | pgprot_noncached(pgprot_kernel)); | 320 | pgprot_dmacoherent(pgprot_kernel)); |
332 | } | 321 | } |
333 | EXPORT_SYMBOL(dma_alloc_coherent); | 322 | EXPORT_SYMBOL(dma_alloc_coherent); |
334 | 323 | ||
@@ -349,15 +338,12 @@ static int dma_mmap(struct device *dev, struct vm_area_struct *vma, | |||
349 | { | 338 | { |
350 | int ret = -ENXIO; | 339 | int ret = -ENXIO; |
351 | #ifdef CONFIG_MMU | 340 | #ifdef CONFIG_MMU |
352 | unsigned long flags, user_size, kern_size; | 341 | unsigned long user_size, kern_size; |
353 | struct arm_vm_region *c; | 342 | struct arm_vmregion *c; |
354 | 343 | ||
355 | user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; | 344 | user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT; |
356 | 345 | ||
357 | spin_lock_irqsave(&consistent_lock, flags); | 346 | c = arm_vmregion_find(&consistent_head, (unsigned long)cpu_addr); |
358 | c = arm_vm_region_find(&consistent_head, (unsigned long)cpu_addr); | ||
359 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
360 | |||
361 | if (c) { | 347 | if (c) { |
362 | unsigned long off = vma->vm_pgoff; | 348 | unsigned long off = vma->vm_pgoff; |
363 | 349 | ||
@@ -379,7 +365,7 @@ static int dma_mmap(struct device *dev, struct vm_area_struct *vma, | |||
379 | int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma, | 365 | int dma_mmap_coherent(struct device *dev, struct vm_area_struct *vma, |
380 | void *cpu_addr, dma_addr_t dma_addr, size_t size) | 366 | void *cpu_addr, dma_addr_t dma_addr, size_t size) |
381 | { | 367 | { |
382 | vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); | 368 | vma->vm_page_prot = pgprot_dmacoherent(vma->vm_page_prot); |
383 | return dma_mmap(dev, vma, cpu_addr, dma_addr, size); | 369 | return dma_mmap(dev, vma, cpu_addr, dma_addr, size); |
384 | } | 370 | } |
385 | EXPORT_SYMBOL(dma_mmap_coherent); | 371 | EXPORT_SYMBOL(dma_mmap_coherent); |
@@ -396,144 +382,23 @@ EXPORT_SYMBOL(dma_mmap_writecombine); | |||
396 | * free a page as defined by the above mapping. | 382 | * free a page as defined by the above mapping. |
397 | * Must not be called with IRQs disabled. | 383 | * Must not be called with IRQs disabled. |
398 | */ | 384 | */ |
399 | #ifdef CONFIG_MMU | ||
400 | void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle) | 385 | void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle) |
401 | { | 386 | { |
402 | struct arm_vm_region *c; | ||
403 | unsigned long flags, addr; | ||
404 | pte_t *ptep; | ||
405 | int idx; | ||
406 | u32 off; | ||
407 | |||
408 | WARN_ON(irqs_disabled()); | 387 | WARN_ON(irqs_disabled()); |
409 | 388 | ||
410 | if (dma_release_from_coherent(dev, get_order(size), cpu_addr)) | 389 | if (dma_release_from_coherent(dev, get_order(size), cpu_addr)) |
411 | return; | 390 | return; |
412 | 391 | ||
413 | if (arch_is_coherent()) { | ||
414 | kfree(cpu_addr); | ||
415 | return; | ||
416 | } | ||
417 | |||
418 | size = PAGE_ALIGN(size); | 392 | size = PAGE_ALIGN(size); |
419 | 393 | ||
420 | spin_lock_irqsave(&consistent_lock, flags); | 394 | if (!arch_is_coherent()) |
421 | c = arm_vm_region_find(&consistent_head, (unsigned long)cpu_addr); | 395 | __dma_free_remap(cpu_addr, size); |
422 | if (!c) | ||
423 | goto no_area; | ||
424 | |||
425 | c->vm_active = 0; | ||
426 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
427 | |||
428 | if ((c->vm_end - c->vm_start) != size) { | ||
429 | printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n", | ||
430 | __func__, c->vm_end - c->vm_start, size); | ||
431 | dump_stack(); | ||
432 | size = c->vm_end - c->vm_start; | ||
433 | } | ||
434 | |||
435 | idx = CONSISTENT_PTE_INDEX(c->vm_start); | ||
436 | off = CONSISTENT_OFFSET(c->vm_start) & (PTRS_PER_PTE-1); | ||
437 | ptep = consistent_pte[idx] + off; | ||
438 | addr = c->vm_start; | ||
439 | do { | ||
440 | pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep); | ||
441 | unsigned long pfn; | ||
442 | |||
443 | ptep++; | ||
444 | addr += PAGE_SIZE; | ||
445 | off++; | ||
446 | if (off >= PTRS_PER_PTE) { | ||
447 | off = 0; | ||
448 | ptep = consistent_pte[++idx]; | ||
449 | } | ||
450 | |||
451 | if (!pte_none(pte) && pte_present(pte)) { | ||
452 | pfn = pte_pfn(pte); | ||
453 | |||
454 | if (pfn_valid(pfn)) { | ||
455 | struct page *page = pfn_to_page(pfn); | ||
456 | |||
457 | /* | ||
458 | * x86 does not mark the pages reserved... | ||
459 | */ | ||
460 | ClearPageReserved(page); | ||
461 | |||
462 | __free_page(page); | ||
463 | continue; | ||
464 | } | ||
465 | } | ||
466 | |||
467 | printk(KERN_CRIT "%s: bad page in kernel page table\n", | ||
468 | __func__); | ||
469 | } while (size -= PAGE_SIZE); | ||
470 | |||
471 | flush_tlb_kernel_range(c->vm_start, c->vm_end); | ||
472 | |||
473 | spin_lock_irqsave(&consistent_lock, flags); | ||
474 | list_del(&c->vm_list); | ||
475 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
476 | |||
477 | kfree(c); | ||
478 | return; | ||
479 | 396 | ||
480 | no_area: | 397 | __dma_free_buffer(dma_to_page(dev, handle), size); |
481 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
482 | printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n", | ||
483 | __func__, cpu_addr); | ||
484 | dump_stack(); | ||
485 | } | 398 | } |
486 | #else /* !CONFIG_MMU */ | ||
487 | void dma_free_coherent(struct device *dev, size_t size, void *cpu_addr, dma_addr_t handle) | ||
488 | { | ||
489 | if (dma_release_from_coherent(dev, get_order(size), cpu_addr)) | ||
490 | return; | ||
491 | kfree(cpu_addr); | ||
492 | } | ||
493 | #endif /* CONFIG_MMU */ | ||
494 | EXPORT_SYMBOL(dma_free_coherent); | 399 | EXPORT_SYMBOL(dma_free_coherent); |
495 | 400 | ||
496 | /* | 401 | /* |
497 | * Initialise the consistent memory allocation. | ||
498 | */ | ||
499 | static int __init consistent_init(void) | ||
500 | { | ||
501 | int ret = 0; | ||
502 | #ifdef CONFIG_MMU | ||
503 | pgd_t *pgd; | ||
504 | pmd_t *pmd; | ||
505 | pte_t *pte; | ||
506 | int i = 0; | ||
507 | u32 base = CONSISTENT_BASE; | ||
508 | |||
509 | do { | ||
510 | pgd = pgd_offset(&init_mm, base); | ||
511 | pmd = pmd_alloc(&init_mm, pgd, base); | ||
512 | if (!pmd) { | ||
513 | printk(KERN_ERR "%s: no pmd tables\n", __func__); | ||
514 | ret = -ENOMEM; | ||
515 | break; | ||
516 | } | ||
517 | WARN_ON(!pmd_none(*pmd)); | ||
518 | |||
519 | pte = pte_alloc_kernel(pmd, base); | ||
520 | if (!pte) { | ||
521 | printk(KERN_ERR "%s: no pte tables\n", __func__); | ||
522 | ret = -ENOMEM; | ||
523 | break; | ||
524 | } | ||
525 | |||
526 | consistent_pte[i++] = pte; | ||
527 | base += (1 << PGDIR_SHIFT); | ||
528 | } while (base < CONSISTENT_END); | ||
529 | #endif /* !CONFIG_MMU */ | ||
530 | |||
531 | return ret; | ||
532 | } | ||
533 | |||
534 | core_initcall(consistent_init); | ||
535 | |||
536 | /* | ||
537 | * Make an area consistent for devices. | 402 | * Make an area consistent for devices. |
538 | * Note: Drivers should NOT use this function directly, as it will break | 403 | * Note: Drivers should NOT use this function directly, as it will break |
539 | * platforms with CONFIG_DMABOUNCE. | 404 | * platforms with CONFIG_DMABOUNCE. |
diff --git a/arch/arm/mm/vmregion.c b/arch/arm/mm/vmregion.c new file mode 100644 index 000000000000..19e09bdb1b8a --- /dev/null +++ b/arch/arm/mm/vmregion.c | |||
@@ -0,0 +1,131 @@ | |||
1 | #include <linux/spinlock.h> | ||
2 | #include <linux/list.h> | ||
3 | #include <linux/slab.h> | ||
4 | |||
5 | #include "vmregion.h" | ||
6 | |||
7 | /* | ||
8 | * VM region handling support. | ||
9 | * | ||
10 | * This should become something generic, handling VM region allocations for | ||
11 | * vmalloc and similar (ioremap, module space, etc). | ||
12 | * | ||
13 | * I envisage vmalloc()'s supporting vm_struct becoming: | ||
14 | * | ||
15 | * struct vm_struct { | ||
16 | * struct vmregion region; | ||
17 | * unsigned long flags; | ||
18 | * struct page **pages; | ||
19 | * unsigned int nr_pages; | ||
20 | * unsigned long phys_addr; | ||
21 | * }; | ||
22 | * | ||
23 | * get_vm_area() would then call vmregion_alloc with an appropriate | ||
24 | * struct vmregion head (eg): | ||
25 | * | ||
26 | * struct vmregion vmalloc_head = { | ||
27 | * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list), | ||
28 | * .vm_start = VMALLOC_START, | ||
29 | * .vm_end = VMALLOC_END, | ||
30 | * }; | ||
31 | * | ||
32 | * However, vmalloc_head.vm_start is variable (typically, it is dependent on | ||
33 | * the amount of RAM found at boot time.) I would imagine that get_vm_area() | ||
34 | * would have to initialise this each time prior to calling vmregion_alloc(). | ||
35 | */ | ||
36 | |||
37 | struct arm_vmregion * | ||
38 | arm_vmregion_alloc(struct arm_vmregion_head *head, size_t size, gfp_t gfp) | ||
39 | { | ||
40 | unsigned long addr = head->vm_start, end = head->vm_end - size; | ||
41 | unsigned long flags; | ||
42 | struct arm_vmregion *c, *new; | ||
43 | |||
44 | if (head->vm_end - head->vm_start < size) { | ||
45 | printk(KERN_WARNING "%s: allocation too big (requested %#x)\n", | ||
46 | __func__, size); | ||
47 | goto out; | ||
48 | } | ||
49 | |||
50 | new = kmalloc(sizeof(struct arm_vmregion), gfp); | ||
51 | if (!new) | ||
52 | goto out; | ||
53 | |||
54 | spin_lock_irqsave(&head->vm_lock, flags); | ||
55 | |||
56 | list_for_each_entry(c, &head->vm_list, vm_list) { | ||
57 | if ((addr + size) < addr) | ||
58 | goto nospc; | ||
59 | if ((addr + size) <= c->vm_start) | ||
60 | goto found; | ||
61 | addr = c->vm_end; | ||
62 | if (addr > end) | ||
63 | goto nospc; | ||
64 | } | ||
65 | |||
66 | found: | ||
67 | /* | ||
68 | * Insert this entry _before_ the one we found. | ||
69 | */ | ||
70 | list_add_tail(&new->vm_list, &c->vm_list); | ||
71 | new->vm_start = addr; | ||
72 | new->vm_end = addr + size; | ||
73 | new->vm_active = 1; | ||
74 | |||
75 | spin_unlock_irqrestore(&head->vm_lock, flags); | ||
76 | return new; | ||
77 | |||
78 | nospc: | ||
79 | spin_unlock_irqrestore(&head->vm_lock, flags); | ||
80 | kfree(new); | ||
81 | out: | ||
82 | return NULL; | ||
83 | } | ||
84 | |||
85 | static struct arm_vmregion *__arm_vmregion_find(struct arm_vmregion_head *head, unsigned long addr) | ||
86 | { | ||
87 | struct arm_vmregion *c; | ||
88 | |||
89 | list_for_each_entry(c, &head->vm_list, vm_list) { | ||
90 | if (c->vm_active && c->vm_start == addr) | ||
91 | goto out; | ||
92 | } | ||
93 | c = NULL; | ||
94 | out: | ||
95 | return c; | ||
96 | } | ||
97 | |||
98 | struct arm_vmregion *arm_vmregion_find(struct arm_vmregion_head *head, unsigned long addr) | ||
99 | { | ||
100 | struct arm_vmregion *c; | ||
101 | unsigned long flags; | ||
102 | |||
103 | spin_lock_irqsave(&head->vm_lock, flags); | ||
104 | c = __arm_vmregion_find(head, addr); | ||
105 | spin_unlock_irqrestore(&head->vm_lock, flags); | ||
106 | return c; | ||
107 | } | ||
108 | |||
109 | struct arm_vmregion *arm_vmregion_find_remove(struct arm_vmregion_head *head, unsigned long addr) | ||
110 | { | ||
111 | struct arm_vmregion *c; | ||
112 | unsigned long flags; | ||
113 | |||
114 | spin_lock_irqsave(&head->vm_lock, flags); | ||
115 | c = __arm_vmregion_find(head, addr); | ||
116 | if (c) | ||
117 | c->vm_active = 0; | ||
118 | spin_unlock_irqrestore(&head->vm_lock, flags); | ||
119 | return c; | ||
120 | } | ||
121 | |||
122 | void arm_vmregion_free(struct arm_vmregion_head *head, struct arm_vmregion *c) | ||
123 | { | ||
124 | unsigned long flags; | ||
125 | |||
126 | spin_lock_irqsave(&head->vm_lock, flags); | ||
127 | list_del(&c->vm_list); | ||
128 | spin_unlock_irqrestore(&head->vm_lock, flags); | ||
129 | |||
130 | kfree(c); | ||
131 | } | ||
diff --git a/arch/arm/mm/vmregion.h b/arch/arm/mm/vmregion.h new file mode 100644 index 000000000000..6b2cdbdf3a85 --- /dev/null +++ b/arch/arm/mm/vmregion.h | |||
@@ -0,0 +1,29 @@ | |||
1 | #ifndef VMREGION_H | ||
2 | #define VMREGION_H | ||
3 | |||
4 | #include <linux/spinlock.h> | ||
5 | #include <linux/list.h> | ||
6 | |||
7 | struct page; | ||
8 | |||
9 | struct arm_vmregion_head { | ||
10 | spinlock_t vm_lock; | ||
11 | struct list_head vm_list; | ||
12 | unsigned long vm_start; | ||
13 | unsigned long vm_end; | ||
14 | }; | ||
15 | |||
16 | struct arm_vmregion { | ||
17 | struct list_head vm_list; | ||
18 | unsigned long vm_start; | ||
19 | unsigned long vm_end; | ||
20 | struct page *vm_pages; | ||
21 | int vm_active; | ||
22 | }; | ||
23 | |||
24 | struct arm_vmregion *arm_vmregion_alloc(struct arm_vmregion_head *, size_t, gfp_t); | ||
25 | struct arm_vmregion *arm_vmregion_find(struct arm_vmregion_head *, unsigned long); | ||
26 | struct arm_vmregion *arm_vmregion_find_remove(struct arm_vmregion_head *, unsigned long); | ||
27 | void arm_vmregion_free(struct arm_vmregion_head *, struct arm_vmregion *); | ||
28 | |||
29 | #endif | ||