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/config.h>
  26 #include <linux/module.h>
  27 #include <linux/signal.h>
  28 #include <linux/sched.h>
  29 #include <linux/kernel.h>
  30 #include <linux/errno.h>
  31 #include <linux/string.h>
  32 #include <linux/types.h>
  33 #include <linux/ptrace.h>
  34 #include <linux/mman.h>
  35 #include <linux/mm.h>
  36 #include <linux/swap.h>
  37 #include <linux/stddef.h>
  38 #include <linux/vmalloc.h>
  39 #include <linux/init.h>
  40 #include <linux/delay.h>
  41 #include <linux/bootmem.h>
  42 #include <linux/highmem.h>
  43 #include <linux/dma-mapping.h>
  44 #include <linux/hardirq.h>
  45
  46 #include <asm/pgalloc.h>
  47 #include <asm/prom.h>
  48 #include <asm/io.h>
  49 #include <asm/mmu_context.h>
  50 #include <asm/pgtable.h>
  51 #include <asm/mmu.h>
  52 #include <asm/uaccess.h>
  53 #include <asm/smp.h>
  54 #include <asm/machdep.h>
  55
  56 int map_page(unsigned long va, phys_addr_t pa, int flags);
  57
  58 #include <asm/tlbflush.h>
  59
  60 /*
  61  * This address range defaults to a value that is safe for all
  62  * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It
  63  * can be further configured for specific applications under
  64  * the "Advanced Setup" menu. -Matt
  65  */
  66 #define CONSISTENT_BASE (CONFIG_CONSISTENT_START)
  67 #define CONSISTENT_END  (CONFIG_CONSISTENT_START + CONFIG_CONSISTENT_SIZE)
  68 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
  69
  70 /*
  71  * This is the page table (2MB) covering uncached, DMA consistent allocations
  72  */
  73 static pte_t *consistent_pte;
  74 static spinlock_t consistent_lock = SPIN_LOCK_UNLOCKED;
  75
  76 /*
  77  * VM region handling support.
  78  *
  79  * This should become something generic, handling VM region allocations for
  80  * vmalloc and similar (ioremap, module space, etc).
  81  *
  82  * I envisage vmalloc()'s supporting vm_struct becoming:
  83  *
  84  *  struct vm_struct {
  85  *    struct vm_region  region;
  86  *    unsigned long     flags;
  87  *    struct page       **pages;
  88  *    unsigned int      nr_pages;
  89  *    unsigned long     phys_addr;
  90  *  };
  91  *
  92  * get_vm_area() would then call vm_region_alloc with an appropriate
  93  * struct vm_region head (eg):
  94  *
  95  *  struct vm_region vmalloc_head = {
  96  *      .vm_list        = LIST_HEAD_INIT(vmalloc_head.vm_list),
  97  *      .vm_start       = VMALLOC_START,
  98  *      .vm_end         = VMALLOC_END,
  99  *  };
 100  *
 101  * However, vmalloc_head.vm_start is variable (typically, it is dependent on
 102  * the amount of RAM found at boot time.)  I would imagine that get_vm_area()
 103  * would have to initialise this each time prior to calling vm_region_alloc().
 104  */
 105 struct vm_region {
 106         struct list_head        vm_list;
 107         unsigned long           vm_start;
 108         unsigned long           vm_end;
 109 };
 110
 111 static struct vm_region consistent_head = {
 112         .vm_list        = LIST_HEAD_INIT(consistent_head.vm_list),
 113         .vm_start       = CONSISTENT_BASE,
 114         .vm_end         = CONSISTENT_END,
 115 };
 116
 117 static struct vm_region *
 118 vm_region_alloc(struct vm_region *head, size_t size, int gfp)
 119 {
 120         unsigned long addr = head->vm_start, end = head->vm_end - size;
 121         unsigned long flags;
 122         struct vm_region *c, *new;
 123
 124         new = kmalloc(sizeof(struct vm_region), gfp);
 125         if (!new)
 126                 goto out;
 127
 128         spin_lock_irqsave(&consistent_lock, flags);
 129
 130         list_for_each_entry(c, &head->vm_list, vm_list) {
 131                 if ((addr + size) < addr)
 132                         goto nospc;
 133                 if ((addr + size) <= c->vm_start)
 134                         goto found;
 135                 addr = c->vm_end;
 136                 if (addr > end)
 137                         goto nospc;
 138         }
 139
 140  found:
 141         /*
 142          * Insert this entry _before_ the one we found.
 143          */
 144         list_add_tail(&new->vm_list, &c->vm_list);
 145         new->vm_start = addr;
 146         new->vm_end = addr + size;
 147
 148         spin_unlock_irqrestore(&consistent_lock, flags);
 149         return new;
 150
 151  nospc:
 152         spin_unlock_irqrestore(&consistent_lock, flags);
 153         kfree(new);
 154  out:
 155         return NULL;
 156 }
 157
 158 static struct vm_region *vm_region_find(struct vm_region *head, unsigned long addr)
 159 {
 160         struct vm_region *c;
 161
 162         list_for_each_entry(c, &head->vm_list, vm_list) {
 163                 if (c->vm_start == addr)
 164                         goto out;
 165         }
 166         c = NULL;
 167  out:
 168         return c;
 169 }
 170
 171 /*
 172  * Allocate DMA-coherent memory space and return both the kernel remapped
 173  * virtual and bus address for that space.
 174  */
 175 void *
 176 __dma_alloc_coherent(size_t size, dma_addr_t *handle, int gfp)
 177 {
 178         struct page *page;
 179         struct vm_region *c;
 180         unsigned long order;
 181         u64 mask = 0x00ffffff, limit; /* ISA default */
 182
 183         if (!consistent_pte) {
 184                 printk(KERN_ERR "%s: not initialised\n", __func__);
 185                 dump_stack();
 186                 return NULL;
 187         }
 188
 189         size = PAGE_ALIGN(size);
 190         limit = (mask + 1) & ~mask;
 191         if ((limit && size >= limit) || size >= (CONSISTENT_END - CONSISTENT_BASE)) {
 192                 printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n",
 193                        size, mask);
 194                 return NULL;
 195         }
 196
 197         order = get_order(size);
 198
 199         if (mask != 0xffffffff)
 200                 gfp |= GFP_DMA;
 201
 202         page = alloc_pages(gfp, order);
 203         if (!page)
 204                 goto no_page;
 205
 206         /*
 207          * Invalidate any data that might be lurking in the
 208          * kernel direct-mapped region for device DMA.
 209          */
 210         {
 211                 unsigned long kaddr = (unsigned long)page_address(page);
 212                 memset(page_address(page), 0, size);
 213                 flush_dcache_range(kaddr, kaddr + size);
 214         }
 215
 216         /*
 217          * Allocate a virtual address in the consistent mapping region.
 218          */
 219         c = vm_region_alloc(&consistent_head, size,
 220                             gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
 221         if (c) {
 222                 pte_t *pte = consistent_pte + CONSISTENT_OFFSET(c->vm_start);
 223                 struct page *end = page + (1 << order);
 224
 225                 /*
 226                  * Set the "dma handle"
 227                  */
 228                 *handle = page_to_bus(page);
 229
 230                 do {
 231                         BUG_ON(!pte_none(*pte));
 232
 233                         set_page_count(page, 1);
 234                         SetPageReserved(page);
 235                         set_pte(pte, mk_pte(page, pgprot_noncached(PAGE_KERNEL)));
 236                         page++;
 237                         pte++;
 238                 } while (size -= PAGE_SIZE);
 239
 240                 /*
 241                  * Free the otherwise unused pages.
 242                  */
 243                 while (page < end) {
 244                         set_page_count(page, 1);
 245                         __free_page(page);
 246                         page++;
 247                 }
 248
 249                 return (void *)c->vm_start;
 250         }
 251
 252         if (page)
 253                 __free_pages(page, order);
 254  no_page:
 255         return NULL;
 256 }
 257 EXPORT_SYMBOL(__dma_alloc_coherent);
 258
 259 /*
 260  * free a page as defined by the above mapping.
 261  */
 262 void __dma_free_coherent(size_t size, void *vaddr)
 263 {
 264         struct vm_region *c;
 265         unsigned long flags;
 266         pte_t *ptep;
 267
 268         size = PAGE_ALIGN(size);
 269
 270         spin_lock_irqsave(&consistent_lock, flags);
 271
 272         c = vm_region_find(&consistent_head, (unsigned long)vaddr);
 273         if (!c)
 274                 goto no_area;
 275
 276         if ((c->vm_end - c->vm_start) != size) {
 277                 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
 278                        __func__, c->vm_end - c->vm_start, size);
 279                 dump_stack();
 280                 size = c->vm_end - c->vm_start;
 281         }
 282
 283         ptep = consistent_pte + CONSISTENT_OFFSET(c->vm_start);
 284         do {
 285                 pte_t pte = ptep_get_and_clear(ptep);
 286                 unsigned long pfn;
 287
 288                 ptep++;
 289
 290                 if (!pte_none(pte) && pte_present(pte)) {
 291                         pfn = pte_pfn(pte);
 292
 293                         if (pfn_valid(pfn)) {
 294                                 struct page *page = pfn_to_page(pfn);
 295                                 ClearPageReserved(page);
 296
 297                                 __free_page(page);
 298                                 continue;
 299                         }
 300                 }
 301
 302                 printk(KERN_CRIT "%s: bad page in kernel page table\n",
 303                        __func__);
 304         } while (size -= PAGE_SIZE);
 305
 306         flush_tlb_kernel_range(c->vm_start, c->vm_end);
 307
 308         list_del(&c->vm_list);
 309
 310         spin_unlock_irqrestore(&consistent_lock, flags);
 311
 312         kfree(c);
 313         return;
 314
 315  no_area:
 316         spin_unlock_irqrestore(&consistent_lock, flags);
 317         printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
 318                __func__, vaddr);
 319         dump_stack();
 320 }
 321 EXPORT_SYMBOL(__dma_free_coherent);
 322
 323 /*
 324  * Initialise the consistent memory allocation.
 325  */
 326 static int __init dma_alloc_init(void)
 327 {
 328         pgd_t *pgd;
 329         pmd_t *pmd;
 330         pte_t *pte;
 331         int ret = 0;
 332
 333         spin_lock(&init_mm.page_table_lock);
 334
 335         do {
 336                 pgd = pgd_offset(&init_mm, CONSISTENT_BASE);
 337                 pmd = pmd_alloc(&init_mm, pgd, CONSISTENT_BASE);
 338                 if (!pmd) {
 339                         printk(KERN_ERR "%s: no pmd tables\n", __func__);
 340                         ret = -ENOMEM;
 341                         break;
 342                 }
 343                 WARN_ON(!pmd_none(*pmd));
 344
 345                 pte = pte_alloc_kernel(&init_mm, pmd, CONSISTENT_BASE);
 346                 if (!pte) {
 347                         printk(KERN_ERR "%s: no pte tables\n", __func__);
 348                         ret = -ENOMEM;
 349                         break;
 350                 }
 351
 352                 consistent_pte = pte;
 353         } while (0);
 354
 355         spin_unlock(&init_mm.page_table_lock);
 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_atmomic()/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, size) - offset;
 400         size_t cur_size = seg_size;
 401         unsigned long flags, start, seg_offset = offset;
 402         int nr_segs = PAGE_ALIGN(size + (PAGE_SIZE - offset))/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);