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/kernel.h> 27#include <linux/errno.h> 28#include <linux/string.h> 29#include <linux/types.h> 30#include <linux/highmem.h> 31#include <linux/dma-mapping.h> 32 33#include <asm/tlbflush.h> 34 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 */ 48static pte_t *consistent_pte; 49static 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 */ 80struct vm_region { 81 struct list_head vm_list; 82 unsigned long vm_start; 83 unsigned long vm_end; 84}; 85 86static struct 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 92static struct vm_region * 93vm_region_alloc(struct 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 vm_region *c, *new; 98 99 new = kmalloc(sizeof(struct 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 133static struct vm_region *vm_region_find(struct vm_region *head, unsigned long addr) 134{ 135 struct 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/* 147 * Allocate DMA-coherent memory space and return both the kernel remapped 148 * virtual and bus address for that space. 149 */ 150void * 151__dma_alloc_coherent(size_t size, dma_addr_t *handle, gfp_t gfp) 152{ 153 struct page *page; 154 struct vm_region *c; 155 unsigned long order; 156 u64 mask = 0x00ffffff, limit; /* ISA default */ 157 158 if (!consistent_pte) { 159 printk(KERN_ERR "%s: not initialised\n", __func__); 160 dump_stack(); 161 return NULL; 162 } 163 164 size = PAGE_ALIGN(size); 165 limit = (mask + 1) & ~mask; 166 if ((limit && size >= limit) || size >= (CONSISTENT_END - CONSISTENT_BASE)) { 167 printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n", 168 size, mask); 169 return NULL; 170 } 171 172 order = get_order(size); 173 174 if (mask != 0xffffffff) 175 gfp |= GFP_DMA; 176 177 page = alloc_pages(gfp, order); 178 if (!page) 179 goto no_page; 180 181 /* 182 * Invalidate any data that might be lurking in the 183 * kernel direct-mapped region for device DMA. 184 */ 185 { 186 unsigned long kaddr = (unsigned long)page_address(page); 187 memset(page_address(page), 0, size); 188 flush_dcache_range(kaddr, kaddr + size); 189 } 190 191 /* 192 * Allocate a virtual address in the consistent mapping region. 193 */ 194 c = vm_region_alloc(&consistent_head, size, 195 gfp & ~(__GFP_DMA | __GFP_HIGHMEM)); 196 if (c) { 197 unsigned long vaddr = c->vm_start; 198 pte_t *pte = consistent_pte + CONSISTENT_OFFSET(vaddr); 199 struct page *end = page + (1 << order); 200 201 split_page(page, order); 202 203 /* 204 * Set the "dma handle" 205 */ 206 *handle = page_to_bus(page); 207 208 do { 209 BUG_ON(!pte_none(*pte)); 210 211 SetPageReserved(page); 212 set_pte_at(&init_mm, vaddr, 213 pte, mk_pte(page, pgprot_noncached(PAGE_KERNEL))); 214 page++; 215 pte++; 216 vaddr += PAGE_SIZE; 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; 228 } 229 230 if (page) 231 __free_pages(page, order); 232 no_page: 233 return NULL; 234} 235EXPORT_SYMBOL(__dma_alloc_coherent); 236 237/* 238 * free a page as defined by the above mapping. 239 */ 240void __dma_free_coherent(size_t size, void *vaddr) 241{ 242 struct 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 = 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; 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(); 300} 301EXPORT_SYMBOL(__dma_free_coherent); 302 303/* 304 * Initialise the consistent memory allocation. 305 */ 306static 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 WARN_ON(!pmd_none(*pmd)); 324 325 pte = pte_alloc_kernel(pmd, CONSISTENT_BASE); 326 if (!pte) { 327 printk(KERN_ERR "%s: no pte tables\n", __func__); 328 ret = -ENOMEM; 329 break; 330 } 331 332 consistent_pte = pte; 333 } while (0); 334 335 return ret; 336} 337 338core_initcall(dma_alloc_init); 339 340/* 341 * make an area consistent. 342 */ 343void __dma_sync(void *vaddr, size_t size, int direction) 344{ 345 unsigned long start = (unsigned long)vaddr; 346 unsigned long end = start + size; 347 348 switch (direction) { 349 case DMA_NONE: 350 BUG(); 351 case DMA_FROM_DEVICE: /* invalidate only */ 352 invalidate_dcache_range(start, end); 353 break; 354 case DMA_TO_DEVICE: /* writeback only */ 355 clean_dcache_range(start, end); 356 break; 357 case DMA_BIDIRECTIONAL: /* writeback and invalidate */ 358 flush_dcache_range(start, end); 359 break; 360 } 361} 362EXPORT_SYMBOL(__dma_sync); 363 364#ifdef CONFIG_HIGHMEM 365/* 366 * __dma_sync_page() implementation for systems using highmem. 367 * In this case, each page of a buffer must be kmapped/kunmapped 368 * in order to have a virtual address for __dma_sync(). This must 369 * not sleep so kmap_atomic()/kunmap_atomic() are used. 370 * 371 * Note: yes, it is possible and correct to have a buffer extend 372 * beyond the first page. 373 */ 374static inline void __dma_sync_page_highmem(struct page *page, 375 unsigned long offset, size_t size, int direction) 376{ 377 size_t seg_size = min((size_t)(PAGE_SIZE - offset), size); 378 size_t cur_size = seg_size; 379 unsigned long flags, start, seg_offset = offset; 380 int nr_segs = 1 + ((size - seg_size) + PAGE_SIZE - 1)/PAGE_SIZE; 381 int seg_nr = 0; 382 383 local_irq_save(flags); 384 385 do { 386 start = (unsigned long)kmap_atomic(page + seg_nr, 387 KM_PPC_SYNC_PAGE) + seg_offset; 388 389 /* Sync this buffer segment */ 390 __dma_sync((void *)start, seg_size, direction); 391 kunmap_atomic((void *)start, KM_PPC_SYNC_PAGE); 392 seg_nr++; 393 394 /* Calculate next buffer segment size */ 395 seg_size = min((size_t)PAGE_SIZE, size - cur_size); 396 397 /* Add the segment size to our running total */ 398 cur_size += seg_size; 399 seg_offset = 0; 400 } while (seg_nr < nr_segs); 401 402 local_irq_restore(flags); 403} 404#endif /* CONFIG_HIGHMEM */ 405 406/* 407 * __dma_sync_page makes memory consistent. identical to __dma_sync, but 408 * takes a struct page instead of a virtual address 409 */ 410void __dma_sync_page(struct page *page, unsigned long offset, 411 size_t size, int direction) 412{ 413#ifdef CONFIG_HIGHMEM 414 __dma_sync_page_highmem(page, offset, size, direction); 415#else 416 unsigned long start = (unsigned long)page_address(page) + offset; 417 __dma_sync((void *)start, size, direction); 418#endif 419} 420EXPORT_SYMBOL(__dma_sync_page); 421