arch/powerpc/mm/dma-noncoherent.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/sched.h>
  26 #include <linux/slab.h>
  27 #include <linux/kernel.h>
  28 #include <linux/errno.h>
  29 #include <linux/string.h>
  30 #include <linux/types.h>
  31 #include <linux/highmem.h>
  32 #include <linux/dma-mapping.h>
  33
  34 #include <asm/tlbflush.h>
  35
  36 #include "mmu_decl.h"
  37
  38 /*
  39  * This address range defaults to a value that is safe for all
  40  * platforms which currently set CONFIG_NOT_COHERENT_CACHE. It
  41  * can be further configured for specific applications under
  42  * the "Advanced Setup" menu. -Matt
  43  */
  44 #define CONSISTENT_BASE         (IOREMAP_TOP)
  45 #define CONSISTENT_END          (CONSISTENT_BASE + CONFIG_CONSISTENT_SIZE)
  46 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_BASE) >> PAGE_SHIFT)
  47
  48 /*
  49  * This is the page table (2MB) covering uncached, DMA consistent allocations
  50  */
  51 static DEFINE_SPINLOCK(consistent_lock);
  52
  53 /*
  54  * VM region handling support.
  55  *
  56  * This should become something generic, handling VM region allocations for
  57  * vmalloc and similar (ioremap, module space, etc).
  58  *
  59  * I envisage vmalloc()'s supporting vm_struct becoming:
  60  *
  61  *  struct vm_struct {
  62  *    struct vm_region  region;
  63  *    unsigned long     flags;
  64  *    struct page       **pages;
  65  *    unsigned int      nr_pages;
  66  *    unsigned long     phys_addr;
  67  *  };
  68  *
  69  * get_vm_area() would then call vm_region_alloc with an appropriate
  70  * struct vm_region head (eg):
  71  *
  72  *  struct vm_region vmalloc_head = {
  73  *      .vm_list        = LIST_HEAD_INIT(vmalloc_head.vm_list),
  74  *      .vm_start       = VMALLOC_START,
  75  *      .vm_end         = VMALLOC_END,
  76  *  };
  77  *
  78  * However, vmalloc_head.vm_start is variable (typically, it is dependent on
  79  * the amount of RAM found at boot time.)  I would imagine that get_vm_area()
  80  * would have to initialise this each time prior to calling vm_region_alloc().
  81  */
  82 struct ppc_vm_region {
  83         struct list_head        vm_list;
  84         unsigned long           vm_start;
  85         unsigned long           vm_end;
  86 };
  87
  88 static struct ppc_vm_region consistent_head = {
  89         .vm_list        = LIST_HEAD_INIT(consistent_head.vm_list),
  90         .vm_start       = CONSISTENT_BASE,
  91         .vm_end         = CONSISTENT_END,
  92 };
  93
  94 static struct ppc_vm_region *
  95 ppc_vm_region_alloc(struct ppc_vm_region *head, size_t size, gfp_t gfp)
  96 {
  97         unsigned long addr = head->vm_start, end = head->vm_end - size;
  98         unsigned long flags;
  99         struct ppc_vm_region *c, *new;
 100
 101         new = kmalloc(sizeof(struct ppc_vm_region), gfp);
 102         if (!new)
 103                 goto out;
 104
 105         spin_lock_irqsave(&consistent_lock, flags);
 106
 107         list_for_each_entry(c, &head->vm_list, vm_list) {
 108                 if ((addr + size) < addr)
 109                         goto nospc;
 110                 if ((addr + size) <= c->vm_start)
 111                         goto found;
 112                 addr = c->vm_end;
 113                 if (addr > end)
 114                         goto nospc;
 115         }
 116
 117  found:
 118         /*
 119          * Insert this entry _before_ the one we found.
 120          */
 121         list_add_tail(&new->vm_list, &c->vm_list);
 122         new->vm_start = addr;
 123         new->vm_end = addr + size;
 124
 125         spin_unlock_irqrestore(&consistent_lock, flags);
 126         return new;
 127
 128  nospc:
 129         spin_unlock_irqrestore(&consistent_lock, flags);
 130         kfree(new);
 131  out:
 132         return NULL;
 133 }
 134
 135 static struct ppc_vm_region *ppc_vm_region_find(struct ppc_vm_region *head, unsigned long addr)
 136 {
 137         struct ppc_vm_region *c;
 138
 139         list_for_each_entry(c, &head->vm_list, vm_list) {
 140                 if (c->vm_start == addr)
 141                         goto out;
 142         }
 143         c = NULL;
 144  out:
 145         return c;
 146 }
 147
 148 /*
 149  * Allocate DMA-coherent memory space and return both the kernel remapped
 150  * virtual and bus address for that space.
 151  */
 152 void *
 153 __dma_alloc_coherent(struct device *dev, size_t size, dma_addr_t *handle, gfp_t gfp)
 154 {
 155         struct page *page;
 156         struct ppc_vm_region *c;
 157         unsigned long order;
 158         u64 mask = ISA_DMA_THRESHOLD, limit;
 159
 160         if (dev) {
 161                 mask = dev->coherent_dma_mask;
 162
 163                 /*
 164                  * Sanity check the DMA mask - it must be non-zero, and
 165                  * must be able to be satisfied by a DMA allocation.
 166                  */
 167                 if (mask == 0) {
 168                         dev_warn(dev, "coherent DMA mask is unset\n");
 169                         goto no_page;
 170                 }
 171
 172                 if ((~mask) & ISA_DMA_THRESHOLD) {
 173                         dev_warn(dev, "coherent DMA mask %#llx is smaller "
 174                                  "than system GFP_DMA mask %#llx\n",
 175                                  mask, (unsigned long long)ISA_DMA_THRESHOLD);
 176                         goto no_page;
 177                 }
 178         }
 179
 180
 181         size = PAGE_ALIGN(size);
 182         limit = (mask + 1) & ~mask;
 183         if ((limit && size >= limit) ||
 184             size >= (CONSISTENT_END - CONSISTENT_BASE)) {
 185                 printk(KERN_WARNING "coherent allocation too big (requested %#x mask %#Lx)\n",
 186                        size, mask);
 187                 return NULL;
 188         }
 189
 190         order = get_order(size);
 191
 192         /* Might be useful if we ever have a real legacy DMA zone... */
 193         if (mask != 0xffffffff)
 194                 gfp |= GFP_DMA;
 195
 196         page = alloc_pages(gfp, order);
 197         if (!page)
 198                 goto no_page;
 199
 200         /*
 201          * Invalidate any data that might be lurking in the
 202          * kernel direct-mapped region for device DMA.
 203          */
 204         {
 205                 unsigned long kaddr = (unsigned long)page_address(page);
 206                 memset(page_address(page), 0, size);
 207                 flush_dcache_range(kaddr, kaddr + size);
 208         }
 209
 210         /*
 211          * Allocate a virtual address in the consistent mapping region.
 212          */
 213         c = ppc_vm_region_alloc(&consistent_head, size,
 214                             gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
 215         if (c) {
 216                 unsigned long vaddr = c->vm_start;
 217                 struct page *end = page + (1 << order);
 218
 219                 split_page(page, order);
 220
 221                 /*
 222                  * Set the "dma handle"
 223                  */
 224                 *handle = page_to_phys(page);
 225
 226                 do {
 227                         SetPageReserved(page);
 228                         map_page(vaddr, page_to_phys(page),
 229                                  pgprot_noncached(PAGE_KERNEL));
 230                         page++;
 231                         vaddr += PAGE_SIZE;
 232                 } while (size -= PAGE_SIZE);
 233
 234                 /*
 235                  * Free the otherwise unused pages.
 236                  */
 237                 while (page < end) {
 238                         __free_page(page);
 239                         page++;
 240                 }
 241
 242                 return (void *)c->vm_start;
 243         }
 244
 245         if (page)
 246                 __free_pages(page, order);
 247  no_page:
 248         return NULL;
 249 }
 250 EXPORT_SYMBOL(__dma_alloc_coherent);
 251
 252 /*
 253  * free a page as defined by the above mapping.
 254  */
 255 void __dma_free_coherent(size_t size, void *vaddr)
 256 {
 257         struct ppc_vm_region *c;
 258         unsigned long flags, addr;
 259
 260         size = PAGE_ALIGN(size);
 261
 262         spin_lock_irqsave(&consistent_lock, flags);
 263
 264         c = ppc_vm_region_find(&consistent_head, (unsigned long)vaddr);
 265         if (!c)
 266                 goto no_area;
 267
 268         if ((c->vm_end - c->vm_start) != size) {
 269                 printk(KERN_ERR "%s: freeing wrong coherent size (%ld != %d)\n",
 270                        __func__, c->vm_end - c->vm_start, size);
 271                 dump_stack();
 272                 size = c->vm_end - c->vm_start;
 273         }
 274
 275         addr = c->vm_start;
 276         do {
 277                 pte_t *ptep;
 278                 unsigned long pfn;
 279
 280                 ptep = pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(addr),
 281                                                                addr),
 282                                                     addr),
 283                                          addr);
 284                 if (!pte_none(*ptep) && pte_present(*ptep)) {
 285                         pfn = pte_pfn(*ptep);
 286                         pte_clear(&init_mm, addr, ptep);
 287                         if (pfn_valid(pfn)) {
 288                                 struct page *page = pfn_to_page(pfn);
 289
 290                                 ClearPageReserved(page);
 291                                 __free_page(page);
 292                         }
 293                 }
 294                 addr += PAGE_SIZE;
 295         } while (size -= PAGE_SIZE);
 296
 297         flush_tlb_kernel_range(c->vm_start, c->vm_end);
 298
 299         list_del(&c->vm_list);
 300
 301         spin_unlock_irqrestore(&consistent_lock, flags);
 302
 303         kfree(c);
 304         return;
 305
 306  no_area:
 307         spin_unlock_irqrestore(&consistent_lock, flags);
 308         printk(KERN_ERR "%s: trying to free invalid coherent area: %p\n",
 309                __func__, vaddr);
 310         dump_stack();
 311 }
 312 EXPORT_SYMBOL(__dma_free_coherent);
 313
 314 /*
 315  * make an area consistent.
 316  */
 317 void __dma_sync(void *vaddr, size_t size, int direction)
 318 {
 319         unsigned long start = (unsigned long)vaddr;
 320         unsigned long end   = start + size;
 321
 322         switch (direction) {
 323         case DMA_NONE:
 324                 BUG();
 325         case DMA_FROM_DEVICE:
 326                 /*
 327                  * invalidate only when cache-line aligned otherwise there is
 328                  * the potential for discarding uncommitted data from the cache
 329                  */
 330                 if ((start & (L1_CACHE_BYTES - 1)) || (size & (L1_CACHE_BYTES - 1)))
 331                         flush_dcache_range(start, end);
 332                 else
 333                         invalidate_dcache_range(start, end);
 334                 break;
 335         case DMA_TO_DEVICE:             /* writeback only */
 336                 clean_dcache_range(start, end);
 337                 break;
 338         case DMA_BIDIRECTIONAL: /* writeback and invalidate */
 339                 flush_dcache_range(start, end);
 340                 break;
 341         }
 342 }
 343 EXPORT_SYMBOL(__dma_sync);
 344
 345 #ifdef CONFIG_HIGHMEM
 346 /*
 347  * __dma_sync_page() implementation for systems using highmem.
 348  * In this case, each page of a buffer must be kmapped/kunmapped
 349  * in order to have a virtual address for __dma_sync(). This must
 350  * not sleep so kmap_atomic()/kunmap_atomic() are used.
 351  *
 352  * Note: yes, it is possible and correct to have a buffer extend
 353  * beyond the first page.
 354  */
 355 static inline void __dma_sync_page_highmem(struct page *page,
 356                 unsigned long offset, size_t size, int direction)
 357 {
 358         size_t seg_size = min((size_t)(PAGE_SIZE - offset), size);
 359         size_t cur_size = seg_size;
 360         unsigned long flags, start, seg_offset = offset;
 361         int nr_segs = 1 + ((size - seg_size) + PAGE_SIZE - 1)/PAGE_SIZE;
 362         int seg_nr = 0;
 363
 364         local_irq_save(flags);
 365
 366         do {
 367                 start = (unsigned long)kmap_atomic(page + seg_nr,
 368                                 KM_PPC_SYNC_PAGE) + seg_offset;
 369
 370                 /* Sync this buffer segment */
 371                 __dma_sync((void *)start, seg_size, direction);
 372                 kunmap_atomic((void *)start, KM_PPC_SYNC_PAGE);
 373                 seg_nr++;
 374
 375                 /* Calculate next buffer segment size */
 376                 seg_size = min((size_t)PAGE_SIZE, size - cur_size);
 377
 378                 /* Add the segment size to our running total */
 379                 cur_size += seg_size;
 380                 seg_offset = 0;
 381         } while (seg_nr < nr_segs);
 382
 383         local_irq_restore(flags);
 384 }
 385 #endif /* CONFIG_HIGHMEM */
 386
 387 /*
 388  * __dma_sync_page makes memory consistent. identical to __dma_sync, but
 389  * takes a struct page instead of a virtual address
 390  */
 391 void __dma_sync_page(struct page *page, unsigned long offset,
 392         size_t size, int direction)
 393 {
 394 #ifdef CONFIG_HIGHMEM
 395         __dma_sync_page_highmem(page, offset, size, direction);
 396 #else
 397         unsigned long start = (unsigned long)page_address(page) + offset;
 398         __dma_sync((void *)start, size, direction);
 399 #endif
 400 }
 401 EXPORT_SYMBOL(__dma_sync_page);