arch/ppc/kernel/dma-mapping.c

   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/module.h>
  26 #include <linux/signal.h>
  27 #include <linux/sched.h>
  28 #include <linux/kernel.h>
  29 #include <linux/errno.h>
  30 #include <linux/string.h>
  31 #include <linux/types.h>
  32 #include <linux/ptrace.h>
  33 #include <linux/mman.h>
  34 #include <linux/mm.h>
  35 #include <linux/swap.h>
  36 #include <linux/stddef.h>
  37 #include <linux/vmalloc.h>
  38 #include <linux/init.h>
  39 #include <linux/delay.h>
  40 #include <linux/bootmem.h>
  41 #include <linux/highmem.h>
  42 #include <linux/dma-mapping.h>
  43 #include <linux/hardirq.h>
  44
  45 #include <asm/pgalloc.h>
  46 #include <asm/prom.h>
  47 #include <asm/io.h>
  48 #include <asm/mmu_context.h>
  49 #include <asm/pgtable.h>
  50 #include <asm/mmu.h>
  51 #include <asm/uaccess.h>
  52 #include <asm/smp.h>
  53 #include <asm/machdep.h>
  54
  55 int map_page(unsigned long va, phys_addr_t pa, int flags);
  56
  57 #include <asm/tlbflush.h>
  58
  59 /*
  60  * This address range defaults to a value that is safe for all
  61  * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It
  62  * can be further configured for specific applications under
  63  * the "Advanced Setup" menu. -Matt
  64  */
  65 #define CONSISTENT_BASE (CONFIG_CONSISTENT_START)
  66 #define CONSISTENT_END  (CONFIG_CONSISTENT_START + CONFIG_CONSISTENT_SIZE)
  67 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
  68
  69 /*
  70  * This is the page table (2MB) covering uncached, DMA consistent allocations
  71  */
  72 static pte_t *consistent_pte;
  73 static DEFINE_SPINLOCK(consistent_lock);
  74
  75 /*
  76  * VM region handling support.
  77  *
  78  * This should become something generic, handling VM region allocations for
  79  * vmalloc and similar (ioremap, module space, etc).
  80  *
  81  * I envisage vmalloc()'s supporting vm_struct becoming:
  82  *
  83  *  struct vm_struct {
  84  *    struct vm_region  region;
  85  *    unsigned long     flags;
  86  *    struct page       **pages;
  87  *    unsigned int      nr_pages;
  88  *    unsigned long     phys_addr;
  89  *  };
  90  *
  91  * get_vm_area() would then call vm_region_alloc with an appropriate
  92  * struct vm_region head (eg):
  93  *
  94  *  struct vm_region vmalloc_head = {
  95  *      .vm_list        = LIST_HEAD_INIT(vmalloc_head.vm_list),
  96  *      .vm_start       = VMALLOC_START,
  97  *      .vm_end         = VMALLOC_END,
  98  *  };
  99  *
 100  * However, vmalloc_head.vm_start is variable (typically, it is dependent on
 101  * the amount of RAM found at boot time.)  I would imagine that get_vm_area()
 102  * would have to initialise this each time prior to calling vm_region_alloc().
 103  */
 104 struct vm_region {
 105         struct list_head        vm_list;
 106         unsigned long           vm_start;
 107         unsigned long           vm_end;
 108 };
 109
 110 static struct vm_region consistent_head = {
 111         .vm_list        = LIST_HEAD_INIT(consistent_head.vm_list),
 112         .vm_start       = CONSISTENT_BASE,
 113         .vm_end         = CONSISTENT_END,
 114 };
 115
 116 static struct vm_region *
 117 vm_region_alloc(struct vm_region *head, size_t size, gfp_t gfp)
 118 {
 119         unsigned long addr = head->vm_start, end = head->vm_end - size;
 120         unsigned long flags;
 121         struct vm_region *c, *new;
 122
 123         new = kmalloc(sizeof(struct 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
 139  found:
 140         /*
 141          * Insert this entry _before_ the one we found.
 142          */
 143         list_add_tail(&new->vm_list, &c->vm_list);
 144         new->vm_start = addr;
 145         new->vm_end = addr + size;
 146
 147         spin_unlock_irqrestore(&consistent_lock, flags);
 148         return new;
 149
 150  nospc:
 151         spin_unlock_irqrestore(&consistent_lock, flags);
 152         kfree(new);
 153  out:
 154         return NULL;
 155 }
 156
 157 static struct vm_region *vm_region_find(struct vm_region *head, unsigned long addr)
 158 {
 159         struct vm_region *c;
 160
 161         list_for_each_entry(c, &head->vm_list, vm_list) {
 162                 if (c->vm_start == addr)
 163                         goto out;
 164         }
 165         c = NULL;
 166  out:
 167         return c;
 168 }
 169
 170 /*
 171  * Allocate DMA-coherent memory space and return both the kernel remapped
 172  * virtual and bus address for that space.
 173  */
 174 void *
 175 __dma_alloc_coherent(size_t size, dma_addr_t *handle, gfp_t gfp)
 176 {
 177         struct page *page;
 178         struct vm_region *c;
 179         unsigned long order;
 180         u64 mask = 0x00ffffff, limit; /* ISA default */
 181
 182         if (!consistent_pte) {
 183                 printk(KERN_ERR "%s: not initialised\n", __func__);
 184                 dump_stack();
 185                 return NULL;
 186         }
 187
 188         size = PAGE_ALIGN(size);
 189         limit = (mask + 1) & ~mask;
 190         if ((limit && size >= limit) || size >= (CONSISTENT_END - CONSISTENT_BASE)) {
 191                 printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n",
 192                        size, mask);
 193                 return NULL;
 194         }
 195
 196         order = get_order(size);
 197
 198         if (mask != 0xffffffff)
 199                 gfp |= GFP_DMA;
 200
 201         page = alloc_pages(gfp, order);
 202         if (!page)
 203                 goto no_page;
 204
 205         /*
 206          * Invalidate any data that might be lurking in the
 207          * kernel direct-mapped region for device DMA.
 208          */
 209         {
 210                 unsigned long kaddr = (unsigned long)page_address(page);
 211                 memset(page_address(page), 0, size);
 212                 flush_dcache_range(kaddr, kaddr + size);
 213         }
 214
 215         /*
 216          * Allocate a virtual address in the consistent mapping region.
 217          */
 218         c = vm_region_alloc(&consistent_head, size,
 219                             gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
 220         if (c) {
 221                 unsigned long vaddr = c->vm_start;
 222                 pte_t *pte = consistent_pte + CONSISTENT_OFFSET(vaddr);
 223                 struct page *end = page + (1 << order);
 224
 225                 split_page(page, order);
 226
 227                 /*
 228                  * Set the "dma handle"
 229                  */
 230                 *handle = page_to_bus(page);
 231
 232                 do {
 233                         BUG_ON(!pte_none(*pte));
 234
 235                         SetPageReserved(page);
 236                         set_pte_at(&init_mm, vaddr,
 237                                    pte, mk_pte(page, pgprot_noncached(PAGE_KERNEL)));
 238                         page++;
 239                         pte++;
 240                         vaddr += PAGE_SIZE;
 241                 } while (size -= PAGE_SIZE);
 242
 243                 /*
 244                  * Free the otherwise unused pages.
 245                  */
 246                 while (page < end) {
 247                         __free_page(page);
 248                         page++;
 249                 }
 250
 251                 return (void *)c->vm_start;
 252         }
 253
 254         if (page)
 255                 __free_pages(page, order);
 256  no_page:
 257         return NULL;
 258 }
 259 EXPORT_SYMBOL(__dma_alloc_coherent);
 260
 261 /*
 262  * free a page as defined by the above mapping.
 263  */
 264 void __dma_free_coherent(size_t size, void *vaddr)
 265 {
 266         struct vm_region *c;
 267         unsigned long flags, addr;
 268         pte_t *ptep;
 269
 270         size = PAGE_ALIGN(size);
 271
 272         spin_lock_irqsave(&consistent_lock, flags);
 273
 274         c = vm_region_find(&consistent_head, (unsigned long)vaddr);
 275         if (!c)
 276                 goto no_area;
 277
 278         if ((c->vm_end - c->vm_start) != size) {
 279                 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
 280                        __func__, c->vm_end - c->vm_start, size);
 281                 dump_stack();
 282                 size = c->vm_end - c->vm_start;
 283         }
 284
 285         ptep = consistent_pte + CONSISTENT_OFFSET(c->vm_start);
 286         addr = c->vm_start;
 287         do {
 288                 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
 289                 unsigned long pfn;
 290
 291                 ptep++;
 292                 addr += PAGE_SIZE;
 293
 294                 if (!pte_none(pte) && pte_present(pte)) {
 295                         pfn = pte_pfn(pte);
 296
 297                         if (pfn_valid(pfn)) {
 298                                 struct page *page = pfn_to_page(pfn);
 299                                 ClearPageReserved(page);
 300
 301                                 __free_page(page);
 302                                 continue;
 303                         }
 304                 }
 305
 306                 printk(KERN_CRIT "%s: bad page in kernel page table\n",
 307                        __func__);
 308         } while (size -= PAGE_SIZE);
 309
 310         flush_tlb_kernel_range(c->vm_start, c->vm_end);
 311
 312         list_del(&c->vm_list);
 313
 314         spin_unlock_irqrestore(&consistent_lock, flags);
 315
 316         kfree(c);
 317         return;
 318
 319  no_area:
 320         spin_unlock_irqrestore(&consistent_lock, flags);
 321         printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
 322                __func__, vaddr);
 323         dump_stack();
 324 }
 325 EXPORT_SYMBOL(__dma_free_coherent);
 326
 327 /*
 328  * Initialise the consistent memory allocation.
 329  */
 330 static int __init dma_alloc_init(void)
 331 {
 332         pgd_t *pgd;
 333         pmd_t *pmd;
 334         pte_t *pte;
 335         int ret = 0;
 336
 337         do {
 338                 pgd = pgd_offset(&init_mm, CONSISTENT_BASE);
 339                 pmd = pmd_alloc(&init_mm, pgd, CONSISTENT_BASE);
 340                 if (!pmd) {
 341                         printk(KERN_ERR "%s: no pmd tables\n", __func__);
 342                         ret = -ENOMEM;
 343                         break;
 344                 }
 345                 WARN_ON(!pmd_none(*pmd));
 346
 347                 pte = pte_alloc_kernel(pmd, CONSISTENT_BASE);
 348                 if (!pte) {
 349                         printk(KERN_ERR "%s: no pte tables\n", __func__);
 350                         ret = -ENOMEM;
 351                         break;
 352                 }
 353
 354                 consistent_pte = pte;
 355         } while (0);
 356
 357         return ret;
 358 }
 359
 360 core_initcall(dma_alloc_init);
 361
 362 /*
 363  * make an area consistent.
 364  */
 365 void __dma_sync(void *vaddr, size_t size, int direction)
 366 {
 367         unsigned long start = (unsigned long)vaddr;
 368         unsigned long end   = start + size;
 369
 370         switch (direction) {
 371         case DMA_NONE:
 372                 BUG();
 373         case DMA_FROM_DEVICE:   /* invalidate only */
 374                 invalidate_dcache_range(start, end);
 375                 break;
 376         case DMA_TO_DEVICE:             /* writeback only */
 377                 clean_dcache_range(start, end);
 378                 break;
 379         case DMA_BIDIRECTIONAL: /* writeback and invalidate */
 380                 flush_dcache_range(start, end);
 381                 break;
 382         }
 383 }
 384 EXPORT_SYMBOL(__dma_sync);
 385
 386 #ifdef CONFIG_HIGHMEM
 387 /*
 388  * __dma_sync_page() implementation for systems using highmem.
 389  * In this case, each page of a buffer must be kmapped/kunmapped
 390  * in order to have a virtual address for __dma_sync(). This must
 391  * not sleep so kmap_atomic()/kunmap_atomic() are used.
 392  *
 393  * Note: yes, it is possible and correct to have a buffer extend
 394  * beyond the first page.
 395  */
 396 static inline void __dma_sync_page_highmem(struct page *page,
 397                 unsigned long offset, size_t size, int direction)
 398 {
 399         size_t seg_size = min((size_t)(PAGE_SIZE - offset), size);
 400         size_t cur_size = seg_size;
 401         unsigned long flags, start, seg_offset = offset;
 402         int nr_segs = 1 + ((size - seg_size) + PAGE_SIZE - 1)/PAGE_SIZE;
 403         int seg_nr = 0;
 404
 405         local_irq_save(flags);
 406
 407         do {
 408                 start = (unsigned long)kmap_atomic(page + seg_nr,
 409                                 KM_PPC_SYNC_PAGE) + seg_offset;
 410
 411                 /* Sync this buffer segment */
 412                 __dma_sync((void *)start, seg_size, direction);
 413                 kunmap_atomic((void *)start, KM_PPC_SYNC_PAGE);
 414                 seg_nr++;
 415
 416                 /* Calculate next buffer segment size */
 417                 seg_size = min((size_t)PAGE_SIZE, size - cur_size);
 418
 419                 /* Add the segment size to our running total */
 420                 cur_size += seg_size;
 421                 seg_offset = 0;
 422         } while (seg_nr < nr_segs);
 423
 424         local_irq_restore(flags);
 425 }
 426 #endif /* CONFIG_HIGHMEM */
 427
 428 /*
 429  * __dma_sync_page makes memory consistent. identical to __dma_sync, but
 430  * takes a struct page instead of a virtual address
 431  */
 432 void __dma_sync_page(struct page *page, unsigned long offset,
 433         size_t size, int direction)
 434 {
 435 #ifdef CONFIG_HIGHMEM
 436         __dma_sync_page_highmem(page, offset, size, direction);
 437 #else
 438         unsigned long start = (unsigned long)page_address(page) + offset;
 439         __dma_sync((void *)start, size, direction);
 440 #endif
 441 }
 442 EXPORT_SYMBOL(__dma_sync_page);