diff options
| -rw-r--r-- | arch/powerpc/Kconfig | 25 | ||||
| -rw-r--r-- | arch/powerpc/lib/dma-noncoherent.c | 303 |
2 files changed, 271 insertions, 57 deletions
diff --git a/arch/powerpc/Kconfig b/arch/powerpc/Kconfig index a0d1146a0578..3bb43adce44d 100644 --- a/arch/powerpc/Kconfig +++ b/arch/powerpc/Kconfig | |||
| @@ -868,6 +868,31 @@ config TASK_SIZE | |||
| 868 | default "0x80000000" if PPC_PREP || PPC_8xx | 868 | default "0x80000000" if PPC_PREP || PPC_8xx |
| 869 | default "0xc0000000" | 869 | default "0xc0000000" |
| 870 | 870 | ||
| 871 | config CONSISTENT_START_BOOL | ||
| 872 | bool "Set custom consistent memory pool address" | ||
| 873 | depends on ADVANCED_OPTIONS && NOT_COHERENT_CACHE | ||
| 874 | help | ||
| 875 | This option allows you to set the base virtual address | ||
| 876 | of the consistent memory pool. This pool of virtual | ||
| 877 | memory is used to make consistent memory allocations. | ||
| 878 | |||
| 879 | config CONSISTENT_START | ||
| 880 | hex "Base virtual address of consistent memory pool" if CONSISTENT_START_BOOL | ||
| 881 | default "0xfd000000" if (NOT_COHERENT_CACHE && 8xx) | ||
| 882 | default "0xff100000" if NOT_COHERENT_CACHE | ||
| 883 | |||
| 884 | config CONSISTENT_SIZE_BOOL | ||
| 885 | bool "Set custom consistent memory pool size" | ||
| 886 | depends on ADVANCED_OPTIONS && NOT_COHERENT_CACHE | ||
| 887 | help | ||
| 888 | This option allows you to set the size of the | ||
| 889 | consistent memory pool. This pool of virtual memory | ||
| 890 | is used to make consistent memory allocations. | ||
| 891 | |||
| 892 | config CONSISTENT_SIZE | ||
| 893 | hex "Size of consistent memory pool" if CONSISTENT_SIZE_BOOL | ||
| 894 | default "0x00200000" if NOT_COHERENT_CACHE | ||
| 895 | |||
| 871 | config PIN_TLB | 896 | config PIN_TLB |
| 872 | bool "Pinned Kernel TLBs (860 ONLY)" | 897 | bool "Pinned Kernel TLBs (860 ONLY)" |
| 873 | depends on ADVANCED_OPTIONS && 8xx | 898 | depends on ADVANCED_OPTIONS && 8xx |
diff --git a/arch/powerpc/lib/dma-noncoherent.c b/arch/powerpc/lib/dma-noncoherent.c index 005a28d380af..b7dc4c19f582 100644 --- a/arch/powerpc/lib/dma-noncoherent.c +++ b/arch/powerpc/lib/dma-noncoherent.c | |||
| @@ -29,11 +29,121 @@ | |||
| 29 | #include <linux/types.h> | 29 | #include <linux/types.h> |
| 30 | #include <linux/highmem.h> | 30 | #include <linux/highmem.h> |
| 31 | #include <linux/dma-mapping.h> | 31 | #include <linux/dma-mapping.h> |
| 32 | #include <linux/vmalloc.h> | ||
| 33 | 32 | ||
| 34 | #include <asm/tlbflush.h> | 33 | #include <asm/tlbflush.h> |
| 35 | 34 | ||
| 36 | /* | 35 | /* |
| 36 | * This address range defaults to a value that is safe for all | ||
| 37 | * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It | ||
| 38 | * can be further configured for specific applications under | ||
| 39 | * the "Advanced Setup" menu. -Matt | ||
| 40 | */ | ||
| 41 | #define CONSISTENT_BASE (CONFIG_CONSISTENT_START) | ||
| 42 | #define CONSISTENT_END (CONFIG_CONSISTENT_START + CONFIG_CONSISTENT_SIZE) | ||
| 43 | #define CONSISTENT_OFFSET(x) (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT) | ||
| 44 | |||
| 45 | /* | ||
| 46 | * This is the page table (2MB) covering uncached, DMA consistent allocations | ||
| 47 | */ | ||
| 48 | static pte_t *consistent_pte; | ||
| 49 | static DEFINE_SPINLOCK(consistent_lock); | ||
| 50 | |||
| 51 | /* | ||
| 52 | * VM region handling support. | ||
| 53 | * | ||
| 54 | * This should become something generic, handling VM region allocations for | ||
| 55 | * vmalloc and similar (ioremap, module space, etc). | ||
| 56 | * | ||
| 57 | * I envisage vmalloc()'s supporting vm_struct becoming: | ||
| 58 | * | ||
| 59 | * struct vm_struct { | ||
| 60 | * struct vm_region region; | ||
| 61 | * unsigned long flags; | ||
| 62 | * struct page **pages; | ||
| 63 | * unsigned int nr_pages; | ||
| 64 | * unsigned long phys_addr; | ||
| 65 | * }; | ||
| 66 | * | ||
| 67 | * get_vm_area() would then call vm_region_alloc with an appropriate | ||
| 68 | * struct vm_region head (eg): | ||
| 69 | * | ||
| 70 | * struct vm_region vmalloc_head = { | ||
| 71 | * .vm_list = LIST_HEAD_INIT(vmalloc_head.vm_list), | ||
| 72 | * .vm_start = VMALLOC_START, | ||
| 73 | * .vm_end = VMALLOC_END, | ||
| 74 | * }; | ||
| 75 | * | ||
| 76 | * However, vmalloc_head.vm_start is variable (typically, it is dependent on | ||
| 77 | * the amount of RAM found at boot time.) I would imagine that get_vm_area() | ||
| 78 | * would have to initialise this each time prior to calling vm_region_alloc(). | ||
| 79 | */ | ||
| 80 | struct ppc_vm_region { | ||
| 81 | struct list_head vm_list; | ||
| 82 | unsigned long vm_start; | ||
| 83 | unsigned long vm_end; | ||
| 84 | }; | ||
| 85 | |||
| 86 | static struct ppc_vm_region consistent_head = { | ||
| 87 | .vm_list = LIST_HEAD_INIT(consistent_head.vm_list), | ||
| 88 | .vm_start = CONSISTENT_BASE, | ||
| 89 | .vm_end = CONSISTENT_END, | ||
| 90 | }; | ||
| 91 | |||
| 92 | static struct ppc_vm_region * | ||
| 93 | ppc_vm_region_alloc(struct ppc_vm_region *head, size_t size, gfp_t gfp) | ||
| 94 | { | ||
| 95 | unsigned long addr = head->vm_start, end = head->vm_end - size; | ||
| 96 | unsigned long flags; | ||
| 97 | struct ppc_vm_region *c, *new; | ||
| 98 | |||
| 99 | new = kmalloc(sizeof(struct ppc_vm_region), gfp); | ||
| 100 | if (!new) | ||
| 101 | goto out; | ||
| 102 | |||
| 103 | spin_lock_irqsave(&consistent_lock, flags); | ||
| 104 | |||
| 105 | list_for_each_entry(c, &head->vm_list, vm_list) { | ||
| 106 | if ((addr + size) < addr) | ||
| 107 | goto nospc; | ||
| 108 | if ((addr + size) <= c->vm_start) | ||
| 109 | goto found; | ||
| 110 | addr = c->vm_end; | ||
| 111 | if (addr > end) | ||
| 112 | goto nospc; | ||
| 113 | } | ||
| 114 | |||
| 115 | found: | ||
| 116 | /* | ||
| 117 | * Insert this entry _before_ the one we found. | ||
| 118 | */ | ||
| 119 | list_add_tail(&new->vm_list, &c->vm_list); | ||
| 120 | new->vm_start = addr; | ||
| 121 | new->vm_end = addr + size; | ||
| 122 | |||
| 123 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
| 124 | return new; | ||
| 125 | |||
| 126 | nospc: | ||
| 127 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
| 128 | kfree(new); | ||
| 129 | out: | ||
| 130 | return NULL; | ||
| 131 | } | ||
| 132 | |||
| 133 | static struct ppc_vm_region *ppc_vm_region_find(struct ppc_vm_region *head, unsigned long addr) | ||
| 134 | { | ||
| 135 | struct ppc_vm_region *c; | ||
| 136 | |||
| 137 | list_for_each_entry(c, &head->vm_list, vm_list) { | ||
| 138 | if (c->vm_start == addr) | ||
| 139 | goto out; | ||
| 140 | } | ||
| 141 | c = NULL; | ||
| 142 | out: | ||
| 143 | return c; | ||
| 144 | } | ||
| 145 | |||
| 146 | /* | ||
| 37 | * Allocate DMA-coherent memory space and return both the kernel remapped | 147 | * Allocate DMA-coherent memory space and return both the kernel remapped |
| 38 | * virtual and bus address for that space. | 148 | * virtual and bus address for that space. |
| 39 | */ | 149 | */ |
| @@ -41,21 +151,21 @@ void * | |||
| 41 | __dma_alloc_coherent(size_t size, dma_addr_t *handle, gfp_t gfp) | 151 | __dma_alloc_coherent(size_t size, dma_addr_t *handle, gfp_t gfp) |
| 42 | { | 152 | { |
| 43 | struct page *page; | 153 | struct page *page; |
| 154 | struct ppc_vm_region *c; | ||
| 44 | unsigned long order; | 155 | unsigned long order; |
| 45 | int i; | ||
| 46 | unsigned int nr_pages = PAGE_ALIGN(size)>>PAGE_SHIFT; | ||
| 47 | unsigned int array_size = nr_pages * sizeof(struct page *); | ||
| 48 | struct page **pages; | ||
| 49 | struct page *end; | ||
| 50 | u64 mask = 0x00ffffff, limit; /* ISA default */ | 156 | u64 mask = 0x00ffffff, limit; /* ISA default */ |
| 51 | struct vm_struct *area; | ||
| 52 | 157 | ||
| 53 | BUG_ON(!mem_init_done); | 158 | if (!consistent_pte) { |
| 159 | printk(KERN_ERR "%s: not initialised\n", __func__); | ||
| 160 | dump_stack(); | ||
| 161 | return NULL; | ||
| 162 | } | ||
| 163 | |||
| 54 | size = PAGE_ALIGN(size); | 164 | size = PAGE_ALIGN(size); |
| 55 | limit = (mask + 1) & ~mask; | 165 | limit = (mask + 1) & ~mask; |
| 56 | if (limit && size >= limit) { | 166 | if ((limit && size >= limit) || size >= (CONSISTENT_END - CONSISTENT_BASE)) { |
| 57 | printk(KERN_WARNING "coherent allocation too big (requested " | 167 | printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n", |
| 58 | "%#x mask %#Lx)\n", size, mask); | 168 | size, mask); |
| 59 | return NULL; | 169 | return NULL; |
| 60 | } | 170 | } |
| 61 | 171 | ||
| @@ -68,8 +178,6 @@ __dma_alloc_coherent(size_t size, dma_addr_t *handle, gfp_t gfp) | |||
| 68 | if (!page) | 178 | if (!page) |
| 69 | goto no_page; | 179 | goto no_page; |
| 70 | 180 | ||
| 71 | end = page + (1 << order); | ||
| 72 | |||
| 73 | /* | 181 | /* |
| 74 | * Invalidate any data that might be lurking in the | 182 | * Invalidate any data that might be lurking in the |
| 75 | * kernel direct-mapped region for device DMA. | 183 | * kernel direct-mapped region for device DMA. |
| @@ -80,59 +188,48 @@ __dma_alloc_coherent(size_t size, dma_addr_t *handle, gfp_t gfp) | |||
| 80 | flush_dcache_range(kaddr, kaddr + size); | 188 | flush_dcache_range(kaddr, kaddr + size); |
| 81 | } | 189 | } |
| 82 | 190 | ||
| 83 | split_page(page, order); | ||
| 84 | |||
| 85 | /* | 191 | /* |
| 86 | * Set the "dma handle" | 192 | * Allocate a virtual address in the consistent mapping region. |
| 87 | */ | 193 | */ |
| 88 | *handle = page_to_phys(page); | 194 | c = ppc_vm_region_alloc(&consistent_head, size, |
| 89 | 195 | gfp & ~(__GFP_DMA | __GFP_HIGHMEM)); | |
| 90 | area = get_vm_area_caller(size, VM_IOREMAP, | 196 | if (c) { |
| 91 | __builtin_return_address(1)); | 197 | unsigned long vaddr = c->vm_start; |
| 92 | if (!area) | 198 | pte_t *pte = consistent_pte + CONSISTENT_OFFSET(vaddr); |
| 93 | goto out_free_pages; | 199 | struct page *end = page + (1 << order); |
| 94 | |||
| 95 | if (array_size > PAGE_SIZE) { | ||
| 96 | pages = vmalloc(array_size); | ||
| 97 | area->flags |= VM_VPAGES; | ||
| 98 | } else { | ||
| 99 | pages = kmalloc(array_size, GFP_KERNEL); | ||
| 100 | } | ||
| 101 | if (!pages) | ||
| 102 | goto out_free_area; | ||
| 103 | 200 | ||
| 104 | area->pages = pages; | 201 | split_page(page, order); |
| 105 | area->nr_pages = nr_pages; | ||
| 106 | 202 | ||
| 107 | for (i = 0; i < nr_pages; i++) | 203 | /* |
| 108 | pages[i] = page + i; | 204 | * Set the "dma handle" |
| 205 | */ | ||
| 206 | *handle = page_to_phys(page); | ||
| 109 | 207 | ||
| 110 | if (map_vm_area(area, pgprot_noncached(PAGE_KERNEL), &pages)) | 208 | do { |
| 111 | goto out_unmap; | 209 | BUG_ON(!pte_none(*pte)); |
| 112 | 210 | ||
| 113 | /* | 211 | SetPageReserved(page); |
| 114 | * Free the otherwise unused pages. | 212 | set_pte_at(&init_mm, vaddr, |
| 115 | */ | 213 | pte, mk_pte(page, pgprot_noncached(PAGE_KERNEL))); |
| 116 | page += nr_pages; | 214 | page++; |
| 117 | while (page < end) { | 215 | pte++; |
| 118 | __free_page(page); | 216 | vaddr += PAGE_SIZE; |
| 119 | page++; | 217 | } while (size -= PAGE_SIZE); |
| 218 | |||
| 219 | /* | ||
| 220 | * Free the otherwise unused pages. | ||
| 221 | */ | ||
| 222 | while (page < end) { | ||
| 223 | __free_page(page); | ||
| 224 | page++; | ||
| 225 | } | ||
| 226 | |||
| 227 | return (void *)c->vm_start; | ||
| 120 | } | 228 | } |
| 121 | 229 | ||
| 122 | return area->addr; | ||
| 123 | out_unmap: | ||
| 124 | vunmap(area->addr); | ||
| 125 | if (array_size > PAGE_SIZE) | ||
| 126 | vfree(pages); | ||
| 127 | else | ||
| 128 | kfree(pages); | ||
| 129 | goto out_free_pages; | ||
| 130 | out_free_area: | ||
| 131 | free_vm_area(area); | ||
| 132 | out_free_pages: | ||
| 133 | if (page) | 230 | if (page) |
| 134 | __free_pages(page, order); | 231 | __free_pages(page, order); |
| 135 | no_page: | 232 | no_page: |
| 136 | return NULL; | 233 | return NULL; |
| 137 | } | 234 | } |
| 138 | EXPORT_SYMBOL(__dma_alloc_coherent); | 235 | EXPORT_SYMBOL(__dma_alloc_coherent); |
| @@ -142,12 +239,104 @@ EXPORT_SYMBOL(__dma_alloc_coherent); | |||
| 142 | */ | 239 | */ |
| 143 | void __dma_free_coherent(size_t size, void *vaddr) | 240 | void __dma_free_coherent(size_t size, void *vaddr) |
| 144 | { | 241 | { |
| 145 | vfree(vaddr); | 242 | struct ppc_vm_region *c; |
| 243 | unsigned long flags, addr; | ||
| 244 | pte_t *ptep; | ||
| 245 | |||
| 246 | size = PAGE_ALIGN(size); | ||
| 247 | |||
| 248 | spin_lock_irqsave(&consistent_lock, flags); | ||
| 249 | |||
| 250 | c = ppc_vm_region_find(&consistent_head, (unsigned long)vaddr); | ||
| 251 | if (!c) | ||
| 252 | goto no_area; | ||
| 253 | |||
| 254 | if ((c->vm_end - c->vm_start) != size) { | ||
| 255 | printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n", | ||
| 256 | __func__, c->vm_end - c->vm_start, size); | ||
| 257 | dump_stack(); | ||
| 258 | size = c->vm_end - c->vm_start; | ||
| 259 | } | ||
| 260 | |||
| 261 | ptep = consistent_pte + CONSISTENT_OFFSET(c->vm_start); | ||
| 262 | addr = c->vm_start; | ||
| 263 | do { | ||
| 264 | pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep); | ||
| 265 | unsigned long pfn; | ||
| 266 | |||
| 267 | ptep++; | ||
| 268 | addr += PAGE_SIZE; | ||
| 146 | 269 | ||
| 270 | if (!pte_none(pte) && pte_present(pte)) { | ||
| 271 | pfn = pte_pfn(pte); | ||
| 272 | |||
| 273 | if (pfn_valid(pfn)) { | ||
| 274 | struct page *page = pfn_to_page(pfn); | ||
| 275 | ClearPageReserved(page); | ||
| 276 | |||
| 277 | __free_page(page); | ||
| 278 | continue; | ||
| 279 | } | ||
| 280 | } | ||
| 281 | |||
| 282 | printk(KERN_CRIT "%s: bad page in kernel page table\n", | ||
| 283 | __func__); | ||
| 284 | } while (size -= PAGE_SIZE); | ||
| 285 | |||
| 286 | flush_tlb_kernel_range(c->vm_start, c->vm_end); | ||
| 287 | |||
| 288 | list_del(&c->vm_list); | ||
| 289 | |||
| 290 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
| 291 | |||
| 292 | kfree(c); | ||
| 293 | return; | ||
| 294 | |||
| 295 | no_area: | ||
| 296 | spin_unlock_irqrestore(&consistent_lock, flags); | ||
| 297 | printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n", | ||
| 298 | __func__, vaddr); | ||
| 299 | dump_stack(); | ||
| 147 | } | 300 | } |
| 148 | EXPORT_SYMBOL(__dma_free_coherent); | 301 | EXPORT_SYMBOL(__dma_free_coherent); |
| 149 | 302 | ||
| 150 | /* | 303 | /* |
| 304 | * Initialise the consistent memory allocation. | ||
| 305 | */ | ||
| 306 | static int __init dma_alloc_init(void) | ||
| 307 | { | ||
| 308 | pgd_t *pgd; | ||
| 309 | pud_t *pud; | ||
| 310 | pmd_t *pmd; | ||
| 311 | pte_t *pte; | ||
| 312 | int ret = 0; | ||
| 313 | |||
| 314 | do { | ||
| 315 | pgd = pgd_offset(&init_mm, CONSISTENT_BASE); | ||
| 316 | pud = pud_alloc(&init_mm, pgd, CONSISTENT_BASE); | ||
| 317 | pmd = pmd_alloc(&init_mm, pud, CONSISTENT_BASE); | ||
| 318 | if (!pmd) { | ||
| 319 | printk(KERN_ERR "%s: no pmd tables\n", __func__); | ||
| 320 | ret = -ENOMEM; | ||
| 321 | break; | ||
| 322 | } | ||
| 323 | |||
| 324 | pte = pte_alloc_kernel(pmd, CONSISTENT_BASE); | ||
| 325 | if (!pte) { | ||
| 326 | printk(KERN_ERR "%s: no pte tables\n", __func__); | ||
| 327 | ret = -ENOMEM; | ||
| 328 | break; | ||
| 329 | } | ||
| 330 | |||
| 331 | consistent_pte = pte; | ||
| 332 | } while (0); | ||
| 333 | |||
| 334 | return ret; | ||
| 335 | } | ||
| 336 | |||
| 337 | core_initcall(dma_alloc_init); | ||
| 338 | |||
| 339 | /* | ||
| 151 | * make an area consistent. | 340 | * make an area consistent. |
| 152 | */ | 341 | */ |
| 153 | void __dma_sync(void *vaddr, size_t size, int direction) | 342 | void __dma_sync(void *vaddr, size_t size, int direction) |
