arch/metag/kernel/dma.c

   1 /*
   2  *  Meta version derived from arch/powerpc/lib/dma-noncoherent.c
   3  *    Copyright (C) 2008 Imagination Technologies Ltd.
   4  *
   5  *  PowerPC version derived from arch/arm/mm/consistent.c
   6  *    Copyright (C) 2001 Dan Malek (dmalek@jlc.net)
   7  *
   8  *  Copyright (C) 2000 Russell King
   9  *
  10  * Consistent memory allocators.  Used for DMA devices that want to
  11  * share uncached memory with the processor core.  The function return
  12  * is the virtual address and 'dma_handle' is the physical address.
  13  * Mostly stolen from the ARM port, with some changes for PowerPC.
  14  *                                              -- Dan
  15  *
  16  * Reorganized to get rid of the arch-specific consistent_* functions
  17  * and provide non-coherent implementations for the DMA API. -Matt
  18  *
  19  * Added in_interrupt() safe dma_alloc_coherent()/dma_free_coherent()
  20  * implementation. This is pulled straight from ARM and barely
  21  * modified. -Matt
  22  *
  23  * This program is free software; you can redistribute it and/or modify
  24  * it under the terms of the GNU General Public License version 2 as
  25  * published by the Free Software Foundation.
  26  */
  27
  28 #include <linux/sched.h>
  29 #include <linux/kernel.h>
  30 #include <linux/errno.h>
  31 #include <linux/export.h>
  32 #include <linux/string.h>
  33 #include <linux/types.h>
  34 #include <linux/highmem.h>
  35 #include <linux/dma-mapping.h>
  36 #include <linux/slab.h>
  37
  38 #include <asm/tlbflush.h>
  39 #include <asm/mmu.h>
  40
  41 #define CONSISTENT_OFFSET(x)    (((unsigned long)(x) - CONSISTENT_START) \
  42                                         >> PAGE_SHIFT)
  43
  44 static u64 get_coherent_dma_mask(struct device *dev)
  45 {
  46         u64 mask = ~0ULL;
  47
  48         if (dev) {
  49                 mask = dev->coherent_dma_mask;
  50
  51                 /*
  52                  * Sanity check the DMA mask - it must be non-zero, and
  53                  * must be able to be satisfied by a DMA allocation.
  54                  */
  55                 if (mask == 0) {
  56                         dev_warn(dev, "coherent DMA mask is unset\n");
  57                         return 0;
  58                 }
  59         }
  60
  61         return mask;
  62 }
  63 /*
  64  * This is the page table (2MB) covering uncached, DMA consistent allocations
  65  */
  66 static pte_t *consistent_pte;
  67 static DEFINE_SPINLOCK(consistent_lock);
  68
  69 /*
  70  * VM region handling support.
  71  *
  72  * This should become something generic, handling VM region allocations for
  73  * vmalloc and similar (ioremap, module space, etc).
  74  *
  75  * I envisage vmalloc()'s supporting vm_struct becoming:
  76  *
  77  *  struct vm_struct {
  78  *    struct metag_vm_region    region;
  79  *    unsigned long     flags;
  80  *    struct page       **pages;
  81  *    unsigned int      nr_pages;
  82  *    unsigned long     phys_addr;
  83  *  };
  84  *
  85  * get_vm_area() would then call metag_vm_region_alloc with an appropriate
  86  * struct metag_vm_region head (eg):
  87  *
  88  *  struct metag_vm_region vmalloc_head = {
  89  *      .vm_list        = LIST_HEAD_INIT(vmalloc_head.vm_list),
  90  *      .vm_start       = VMALLOC_START,
  91  *      .vm_end         = VMALLOC_END,
  92  *  };
  93  *
  94  * However, vmalloc_head.vm_start is variable (typically, it is dependent on
  95  * the amount of RAM found at boot time.)  I would imagine that get_vm_area()
  96  * would have to initialise this each time prior to calling
  97  * metag_vm_region_alloc().
  98  */
  99 struct metag_vm_region {
 100         struct list_head vm_list;
 101         unsigned long vm_start;
 102         unsigned long vm_end;
 103         struct page             *vm_pages;
 104         int                     vm_active;
 105 };
 106
 107 static struct metag_vm_region consistent_head = {
 108         .vm_list = LIST_HEAD_INIT(consistent_head.vm_list),
 109         .vm_start = CONSISTENT_START,
 110         .vm_end = CONSISTENT_END,
 111 };
 112
 113 static struct metag_vm_region *metag_vm_region_alloc(struct metag_vm_region
 114                                                      *head, size_t size,
 115                                                      gfp_t gfp)
 116 {
 117         unsigned long addr = head->vm_start, end = head->vm_end - size;
 118         unsigned long flags;
 119         struct metag_vm_region *c, *new;
 120
 121         new = kmalloc(sizeof(struct metag_vm_region), gfp);
 122         if (!new)
 123                 goto out;
 124
 125         spin_lock_irqsave(&consistent_lock, flags);
 126
 127         list_for_each_entry(c, &head->vm_list, vm_list) {
 128                 if ((addr + size) < addr)
 129                         goto nospc;
 130                 if ((addr + size) <= c->vm_start)
 131                         goto found;
 132                 addr = c->vm_end;
 133                 if (addr > end)
 134                         goto nospc;
 135         }
 136
 137 found:
 138         /*
 139          * Insert this entry _before_ the one we found.
 140          */
 141         list_add_tail(&new->vm_list, &c->vm_list);
 142         new->vm_start = addr;
 143         new->vm_end = addr + size;
 144         new->vm_active = 1;
 145
 146         spin_unlock_irqrestore(&consistent_lock, flags);
 147         return new;
 148
 149 nospc:
 150         spin_unlock_irqrestore(&consistent_lock, flags);
 151         kfree(new);
 152 out:
 153         return NULL;
 154 }
 155
 156 static struct metag_vm_region *metag_vm_region_find(struct metag_vm_region
 157                                                     *head, unsigned long addr)
 158 {
 159         struct metag_vm_region *c;
 160
 161         list_for_each_entry(c, &head->vm_list, vm_list) {
 162                 if (c->vm_active && 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 static void *metag_dma_alloc(struct device *dev, size_t size,
 175                 dma_addr_t *handle, gfp_t gfp, unsigned long attrs)
 176 {
 177         struct page *page;
 178         struct metag_vm_region *c;
 179         unsigned long order;
 180         u64 mask = get_coherent_dma_mask(dev);
 181         u64 limit;
 182
 183         if (!consistent_pte) {
 184                 pr_err("%s: not initialised\n", __func__);
 185                 dump_stack();
 186                 return NULL;
 187         }
 188
 189         if (!mask)
 190                 goto no_page;
 191         size = PAGE_ALIGN(size);
 192         limit = (mask + 1) & ~mask;
 193         if ((limit && size >= limit)
 194             || size >= (CONSISTENT_END - CONSISTENT_START)) {
 195                 pr_warn("coherent allocation too big (requested %#x mask %#Lx)\n",
 196                         size, mask);
 197                 return NULL;
 198         }
 199
 200         order = get_order(size);
 201
 202         if (mask != 0xffffffff)
 203                 gfp |= GFP_DMA;
 204
 205         page = alloc_pages(gfp, order);
 206         if (!page)
 207                 goto no_page;
 208
 209         /*
 210          * Invalidate any data that might be lurking in the
 211          * kernel direct-mapped region for device DMA.
 212          */
 213         {
 214                 void *kaddr = page_address(page);
 215                 memset(kaddr, 0, size);
 216                 flush_dcache_region(kaddr, size);
 217         }
 218
 219         /*
 220          * Allocate a virtual address in the consistent mapping region.
 221          */
 222         c = metag_vm_region_alloc(&consistent_head, size,
 223                                   gfp & ~(__GFP_DMA | __GFP_HIGHMEM));
 224         if (c) {
 225                 unsigned long vaddr = c->vm_start;
 226                 pte_t *pte = consistent_pte + CONSISTENT_OFFSET(vaddr);
 227                 struct page *end = page + (1 << order);
 228
 229                 c->vm_pages = page;
 230                 split_page(page, order);
 231
 232                 /*
 233                  * Set the "dma handle"
 234                  */
 235                 *handle = page_to_bus(page);
 236
 237                 do {
 238                         BUG_ON(!pte_none(*pte));
 239
 240                         SetPageReserved(page);
 241                         set_pte_at(&init_mm, vaddr,
 242                                    pte, mk_pte(page,
 243                                                pgprot_writecombine
 244                                                (PAGE_KERNEL)));
 245                         page++;
 246                         pte++;
 247                         vaddr += PAGE_SIZE;
 248                 } while (size -= PAGE_SIZE);
 249
 250                 /*
 251                  * Free the otherwise unused pages.
 252                  */
 253                 while (page < end) {
 254                         __free_page(page);
 255                         page++;
 256                 }
 257
 258                 return (void *)c->vm_start;
 259         }
 260
 261         if (page)
 262                 __free_pages(page, order);
 263 no_page:
 264         return NULL;
 265 }
 266
 267 /*
 268  * free a page as defined by the above mapping.
 269  */
 270 static void metag_dma_free(struct device *dev, size_t size, void *vaddr,
 271                 dma_addr_t dma_handle, unsigned long attrs)
 272 {
 273         struct metag_vm_region *c;
 274         unsigned long flags, addr;
 275         pte_t *ptep;
 276
 277         size = PAGE_ALIGN(size);
 278
 279         spin_lock_irqsave(&consistent_lock, flags);
 280
 281         c = metag_vm_region_find(&consistent_head, (unsigned long)vaddr);
 282         if (!c)
 283                 goto no_area;
 284
 285         c->vm_active = 0;
 286         if ((c->vm_end - c->vm_start) != size) {
 287                 pr_err("%s: freeing wrong coherent size (%ld != %d)\n",
 288                        __func__, c->vm_end - c->vm_start, size);
 289                 dump_stack();
 290                 size = c->vm_end - c->vm_start;
 291         }
 292
 293         ptep = consistent_pte + CONSISTENT_OFFSET(c->vm_start);
 294         addr = c->vm_start;
 295         do {
 296                 pte_t pte = ptep_get_and_clear(&init_mm, addr, ptep);
 297                 unsigned long pfn;
 298
 299                 ptep++;
 300                 addr += PAGE_SIZE;
 301
 302                 if (!pte_none(pte) && pte_present(pte)) {
 303                         pfn = pte_pfn(pte);
 304
 305                         if (pfn_valid(pfn)) {
 306                                 struct page *page = pfn_to_page(pfn);
 307                                 __free_reserved_page(page);
 308                                 continue;
 309                         }
 310                 }
 311
 312                 pr_crit("%s: bad page in kernel page table\n",
 313                         __func__);
 314         } while (size -= PAGE_SIZE);
 315
 316         flush_tlb_kernel_range(c->vm_start, c->vm_end);
 317
 318         list_del(&c->vm_list);
 319
 320         spin_unlock_irqrestore(&consistent_lock, flags);
 321
 322         kfree(c);
 323         return;
 324
 325 no_area:
 326         spin_unlock_irqrestore(&consistent_lock, flags);
 327         pr_err("%s: trying to free invalid coherent area: %p\n",
 328                __func__, vaddr);
 329         dump_stack();
 330 }
 331
 332 static int metag_dma_mmap(struct device *dev, struct vm_area_struct *vma,
 333                 void *cpu_addr, dma_addr_t dma_addr, size_t size,
 334                 unsigned long attrs)
 335 {
 336         unsigned long flags, user_size, kern_size;
 337         struct metag_vm_region *c;
 338         int ret = -ENXIO;
 339
 340         if (attrs & DMA_ATTR_WRITE_COMBINE)
 341                 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
 342         else
 343                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 344
 345         user_size = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
 346
 347         spin_lock_irqsave(&consistent_lock, flags);
 348         c = metag_vm_region_find(&consistent_head, (unsigned long)cpu_addr);
 349         spin_unlock_irqrestore(&consistent_lock, flags);
 350
 351         if (c) {
 352                 unsigned long off = vma->vm_pgoff;
 353
 354                 kern_size = (c->vm_end - c->vm_start) >> PAGE_SHIFT;
 355
 356                 if (off < kern_size &&
 357                     user_size <= (kern_size - off)) {
 358                         ret = remap_pfn_range(vma, vma->vm_start,
 359                                               page_to_pfn(c->vm_pages) + off,
 360                                               user_size << PAGE_SHIFT,
 361                                               vma->vm_page_prot);
 362                 }
 363         }
 364
 365
 366         return ret;
 367 }
 368
 369 /*
 370  * Initialise the consistent memory allocation.
 371  */
 372 static int __init dma_alloc_init(void)
 373 {
 374         pgd_t *pgd, *pgd_k;
 375         pud_t *pud, *pud_k;
 376         pmd_t *pmd, *pmd_k;
 377         pte_t *pte;
 378         int ret = 0;
 379
 380         do {
 381                 int offset = pgd_index(CONSISTENT_START);
 382                 pgd = pgd_offset(&init_mm, CONSISTENT_START);
 383                 pud = pud_alloc(&init_mm, pgd, CONSISTENT_START);
 384                 pmd = pmd_alloc(&init_mm, pud, CONSISTENT_START);
 385                 WARN_ON(!pmd_none(*pmd));
 386
 387                 pte = pte_alloc_kernel(pmd, CONSISTENT_START);
 388                 if (!pte) {
 389                         pr_err("%s: no pte tables\n", __func__);
 390                         ret = -ENOMEM;
 391                         break;
 392                 }
 393
 394                 pgd_k = ((pgd_t *) mmu_get_base()) + offset;
 395                 pud_k = pud_offset(pgd_k, CONSISTENT_START);
 396                 pmd_k = pmd_offset(pud_k, CONSISTENT_START);
 397                 set_pmd(pmd_k, *pmd);
 398
 399                 consistent_pte = pte;
 400         } while (0);
 401
 402         return ret;
 403 }
 404 early_initcall(dma_alloc_init);
 405
 406 /*
 407  * make an area consistent to devices.
 408  */
 409 static void dma_sync_for_device(void *vaddr, size_t size, int dma_direction)
 410 {
 411         /*
 412          * Ensure any writes get through the write combiner. This is necessary
 413          * even with DMA_FROM_DEVICE, or the write may dirty the cache after
 414          * we've invalidated it and get written back during the DMA.
 415          */
 416
 417         barrier();
 418
 419         switch (dma_direction) {
 420         case DMA_BIDIRECTIONAL:
 421                 /*
 422                  * Writeback to ensure the device can see our latest changes and
 423                  * so that we have no dirty lines, and invalidate the cache
 424                  * lines too in preparation for receiving the buffer back
 425                  * (dma_sync_for_cpu) later.
 426                  */
 427                 flush_dcache_region(vaddr, size);
 428                 break;
 429         case DMA_TO_DEVICE:
 430                 /*
 431                  * Writeback to ensure the device can see our latest changes.
 432                  * There's no need to invalidate as the device shouldn't write
 433                  * to the buffer.
 434                  */
 435                 writeback_dcache_region(vaddr, size);
 436                 break;
 437         case DMA_FROM_DEVICE:
 438                 /*
 439                  * Invalidate to ensure we have no dirty lines that could get
 440                  * written back during the DMA. It's also safe to flush
 441                  * (writeback) here if necessary.
 442                  */
 443                 invalidate_dcache_region(vaddr, size);
 444                 break;
 445         case DMA_NONE:
 446                 BUG();
 447         }
 448
 449         wmb();
 450 }
 451
 452 /*
 453  * make an area consistent to the core.
 454  */
 455 static void dma_sync_for_cpu(void *vaddr, size_t size, int dma_direction)
 456 {
 457         /*
 458          * Hardware L2 cache prefetch doesn't occur across 4K physical
 459          * boundaries, however according to Documentation/DMA-API-HOWTO.txt
 460          * kmalloc'd memory is DMA'able, so accesses in nearby memory could
 461          * trigger a cache fill in the DMA buffer.
 462          *
 463          * This should never cause dirty lines, so a flush or invalidate should
 464          * be safe to allow us to see data from the device.
 465          */
 466         if (_meta_l2c_pf_is_enabled()) {
 467                 switch (dma_direction) {
 468                 case DMA_BIDIRECTIONAL:
 469                 case DMA_FROM_DEVICE:
 470                         invalidate_dcache_region(vaddr, size);
 471                         break;
 472                 case DMA_TO_DEVICE:
 473                         /* The device shouldn't have written to the buffer */
 474                         break;
 475                 case DMA_NONE:
 476                         BUG();
 477                 }
 478         }
 479
 480         rmb();
 481 }
 482
 483 static dma_addr_t metag_dma_map_page(struct device *dev, struct page *page,
 484                 unsigned long offset, size_t size,
 485                 enum dma_data_direction direction, unsigned long attrs)
 486 {
 487         dma_sync_for_device((void *)(page_to_phys(page) + offset), size,
 488                             direction);
 489         return page_to_phys(page) + offset;
 490 }
 491
 492 static void metag_dma_unmap_page(struct device *dev, dma_addr_t dma_address,
 493                 size_t size, enum dma_data_direction direction,
 494                 unsigned long attrs)
 495 {
 496         dma_sync_for_cpu(phys_to_virt(dma_address), size, direction);
 497 }
 498
 499 static int metag_dma_map_sg(struct device *dev, struct scatterlist *sglist,
 500                 int nents, enum dma_data_direction direction,
 501                 unsigned long attrs)
 502 {
 503         struct scatterlist *sg;
 504         int i;
 505
 506         for_each_sg(sglist, sg, nents, i) {
 507                 BUG_ON(!sg_page(sg));
 508
 509                 sg->dma_address = sg_phys(sg);
 510                 dma_sync_for_device(sg_virt(sg), sg->length, direction);
 511         }
 512
 513         return nents;
 514 }
 515
 516
 517 static void metag_dma_unmap_sg(struct device *dev, struct scatterlist *sglist,
 518                 int nhwentries, enum dma_data_direction direction,
 519                 unsigned long attrs)
 520 {
 521         struct scatterlist *sg;
 522         int i;
 523
 524         for_each_sg(sglist, sg, nhwentries, i) {
 525                 BUG_ON(!sg_page(sg));
 526
 527                 sg->dma_address = sg_phys(sg);
 528                 dma_sync_for_cpu(sg_virt(sg), sg->length, direction);
 529         }
 530 }
 531
 532 static void metag_dma_sync_single_for_cpu(struct device *dev,
 533                 dma_addr_t dma_handle, size_t size,
 534                 enum dma_data_direction direction)
 535 {
 536         dma_sync_for_cpu(phys_to_virt(dma_handle), size, direction);
 537 }
 538
 539 static void metag_dma_sync_single_for_device(struct device *dev,
 540                 dma_addr_t dma_handle, size_t size,
 541                 enum dma_data_direction direction)
 542 {
 543         dma_sync_for_device(phys_to_virt(dma_handle), size, direction);
 544 }
 545
 546 static void metag_dma_sync_sg_for_cpu(struct device *dev,
 547                 struct scatterlist *sglist, int nelems,
 548                 enum dma_data_direction direction)
 549 {
 550         int i;
 551         struct scatterlist *sg;
 552
 553         for_each_sg(sglist, sg, nelems, i)
 554                 dma_sync_for_cpu(sg_virt(sg), sg->length, direction);
 555 }
 556
 557 static void metag_dma_sync_sg_for_device(struct device *dev,
 558                 struct scatterlist *sglist, int nelems,
 559                 enum dma_data_direction direction)
 560 {
 561         int i;
 562         struct scatterlist *sg;
 563
 564         for_each_sg(sglist, sg, nelems, i)
 565                 dma_sync_for_device(sg_virt(sg), sg->length, direction);
 566 }
 567
 568 struct dma_map_ops metag_dma_ops = {
 569         .alloc                  = metag_dma_alloc,
 570         .free                   = metag_dma_free,
 571         .map_page               = metag_dma_map_page,
 572         .map_sg                 = metag_dma_map_sg,
 573         .sync_single_for_device = metag_dma_sync_single_for_device,
 574         .sync_single_for_cpu    = metag_dma_sync_single_for_cpu,
 575         .sync_sg_for_cpu        = metag_dma_sync_sg_for_cpu,
 576         .mmap                   = metag_dma_mmap,
 577 };
 578 EXPORT_SYMBOL(metag_dma_ops);