mm/dmapool.c

   1 /*
   2  * DMA Pool allocator
   3  *
   4  * Copyright 2001 David Brownell
   5  * Copyright 2007 Intel Corporation
   6  *   Author: Matthew Wilcox <willy@linux.intel.com>
   7  *
   8  * This software may be redistributed and/or modified under the terms of
   9  * the GNU General Public License ("GPL") version 2 as published by the
  10  * Free Software Foundation.
  11  *
  12  * This allocator returns small blocks of a given size which are DMA-able by
  13  * the given device.  It uses the dma_alloc_coherent page allocator to get
  14  * new pages, then splits them up into blocks of the required size.
  15  * Many older drivers still have their own code to do this.
  16  *
  17  * The current design of this allocator is fairly simple.  The pool is
  18  * represented by the 'struct dma_pool' which keeps a doubly-linked list of
  19  * allocated pages.  Each page in the page_list is split into blocks of at
  20  * least 'size' bytes.  Free blocks are tracked in an unsorted singly-linked
  21  * list of free blocks within the page.  Used blocks aren't tracked, but we
  22  * keep a count of how many are currently allocated from each page.
  23  */
  24
  25 #include <linux/device.h>
  26 #include <linux/dma-mapping.h>
  27 #include <linux/dmapool.h>
  28 #include <linux/kernel.h>
  29 #include <linux/list.h>
  30 #include <linux/export.h>
  31 #include <linux/mutex.h>
  32 #include <linux/poison.h>
  33 #include <linux/sched.h>
  34 #include <linux/slab.h>
  35 #include <linux/stat.h>
  36 #include <linux/spinlock.h>
  37 #include <linux/string.h>
  38 #include <linux/types.h>
  39 #include <linux/wait.h>
  40
  41 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_SLUB_DEBUG_ON)
  42 #define DMAPOOL_DEBUG 1
  43 #endif
  44
  45 struct dma_pool {               /* the pool */
  46         struct list_head page_list;
  47         spinlock_t lock;
  48         size_t size;
  49         struct device *dev;
  50         size_t allocation;
  51         size_t boundary;
  52         char name[32];
  53         struct list_head pools;
  54 };
  55
  56 struct dma_page {               /* cacheable header for 'allocation' bytes */
  57         struct list_head page_list;
  58         void *vaddr;
  59         dma_addr_t dma;
  60         unsigned int in_use;
  61         unsigned int offset;
  62 };
  63
  64 static DEFINE_MUTEX(pools_lock);
  65 static DEFINE_MUTEX(pools_reg_lock);
  66
  67 static ssize_t
  68 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
  69 {
  70         unsigned temp;
  71         unsigned size;
  72         char *next;
  73         struct dma_page *page;
  74         struct dma_pool *pool;
  75
  76         next = buf;
  77         size = PAGE_SIZE;
  78
  79         temp = scnprintf(next, size, "poolinfo - 0.1\n");
  80         size -= temp;
  81         next += temp;
  82
  83         mutex_lock(&pools_lock);
  84         list_for_each_entry(pool, &dev->dma_pools, pools) {
  85                 unsigned pages = 0;
  86                 unsigned blocks = 0;
  87
  88                 spin_lock_irq(&pool->lock);
  89                 list_for_each_entry(page, &pool->page_list, page_list) {
  90                         pages++;
  91                         blocks += page->in_use;
  92                 }
  93                 spin_unlock_irq(&pool->lock);
  94
  95                 /* per-pool info, no real statistics yet */
  96                 temp = scnprintf(next, size, "%-16s %4u %4zu %4zu %2u\n",
  97                                  pool->name, blocks,
  98                                  pages * (pool->allocation / pool->size),
  99                                  pool->size, pages);
 100                 size -= temp;
 101                 next += temp;
 102         }
 103         mutex_unlock(&pools_lock);
 104
 105         return PAGE_SIZE - size;
 106 }
 107
 108 static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL);
 109
 110 /**
 111  * dma_pool_create - Creates a pool of consistent memory blocks, for dma.
 112  * @name: name of pool, for diagnostics
 113  * @dev: device that will be doing the DMA
 114  * @size: size of the blocks in this pool.
 115  * @align: alignment requirement for blocks; must be a power of two
 116  * @boundary: returned blocks won't cross this power of two boundary
 117  * Context: !in_interrupt()
 118  *
 119  * Returns a dma allocation pool with the requested characteristics, or
 120  * null if one can't be created.  Given one of these pools, dma_pool_alloc()
 121  * may be used to allocate memory.  Such memory will all have "consistent"
 122  * DMA mappings, accessible by the device and its driver without using
 123  * cache flushing primitives.  The actual size of blocks allocated may be
 124  * larger than requested because of alignment.
 125  *
 126  * If @boundary is nonzero, objects returned from dma_pool_alloc() won't
 127  * cross that size boundary.  This is useful for devices which have
 128  * addressing restrictions on individual DMA transfers, such as not crossing
 129  * boundaries of 4KBytes.
 130  */
 131 struct dma_pool *dma_pool_create(const char *name, struct device *dev,
 132                                  size_t size, size_t align, size_t boundary)
 133 {
 134         struct dma_pool *retval;
 135         size_t allocation;
 136         bool empty = false;
 137
 138         if (align == 0)
 139                 align = 1;
 140         else if (align & (align - 1))
 141                 return NULL;
 142
 143         if (size == 0)
 144                 return NULL;
 145         else if (size < 4)
 146                 size = 4;
 147
 148         if ((size % align) != 0)
 149                 size = ALIGN(size, align);
 150
 151         allocation = max_t(size_t, size, PAGE_SIZE);
 152
 153         if (!boundary)
 154                 boundary = allocation;
 155         else if ((boundary < size) || (boundary & (boundary - 1)))
 156                 return NULL;
 157
 158         retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
 159         if (!retval)
 160                 return retval;
 161
 162         strlcpy(retval->name, name, sizeof(retval->name));
 163
 164         retval->dev = dev;
 165
 166         INIT_LIST_HEAD(&retval->page_list);
 167         spin_lock_init(&retval->lock);
 168         retval->size = size;
 169         retval->boundary = boundary;
 170         retval->allocation = allocation;
 171
 172         INIT_LIST_HEAD(&retval->pools);
 173
 174         /*
 175          * pools_lock ensures that the ->dma_pools list does not get corrupted.
 176          * pools_reg_lock ensures that there is not a race between
 177          * dma_pool_create() and dma_pool_destroy() or within dma_pool_create()
 178          * when the first invocation of dma_pool_create() failed on
 179          * device_create_file() and the second assumes that it has been done (I
 180          * know it is a short window).
 181          */
 182         mutex_lock(&pools_reg_lock);
 183         mutex_lock(&pools_lock);
 184         if (list_empty(&dev->dma_pools))
 185                 empty = true;
 186         list_add(&retval->pools, &dev->dma_pools);
 187         mutex_unlock(&pools_lock);
 188         if (empty) {
 189                 int err;
 190
 191                 err = device_create_file(dev, &dev_attr_pools);
 192                 if (err) {
 193                         mutex_lock(&pools_lock);
 194                         list_del(&retval->pools);
 195                         mutex_unlock(&pools_lock);
 196                         mutex_unlock(&pools_reg_lock);
 197                         kfree(retval);
 198                         return NULL;
 199                 }
 200         }
 201         mutex_unlock(&pools_reg_lock);
 202         return retval;
 203 }
 204 EXPORT_SYMBOL(dma_pool_create);
 205
 206 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
 207 {
 208         unsigned int offset = 0;
 209         unsigned int next_boundary = pool->boundary;
 210
 211         do {
 212                 unsigned int next = offset + pool->size;
 213                 if (unlikely((next + pool->size) >= next_boundary)) {
 214                         next = next_boundary;
 215                         next_boundary += pool->boundary;
 216                 }
 217                 *(int *)(page->vaddr + offset) = next;
 218                 offset = next;
 219         } while (offset < pool->allocation);
 220 }
 221
 222 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
 223 {
 224         struct dma_page *page;
 225
 226         page = kmalloc(sizeof(*page), mem_flags);
 227         if (!page)
 228                 return NULL;
 229         page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
 230                                          &page->dma, mem_flags);
 231         if (page->vaddr) {
 232 #ifdef  DMAPOOL_DEBUG
 233                 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 234 #endif
 235                 pool_initialise_page(pool, page);
 236                 page->in_use = 0;
 237                 page->offset = 0;
 238         } else {
 239                 kfree(page);
 240                 page = NULL;
 241         }
 242         return page;
 243 }
 244
 245 static inline bool is_page_busy(struct dma_page *page)
 246 {
 247         return page->in_use != 0;
 248 }
 249
 250 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
 251 {
 252         dma_addr_t dma = page->dma;
 253
 254 #ifdef  DMAPOOL_DEBUG
 255         memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 256 #endif
 257         dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
 258         list_del(&page->page_list);
 259         kfree(page);
 260 }
 261
 262 /**
 263  * dma_pool_destroy - destroys a pool of dma memory blocks.
 264  * @pool: dma pool that will be destroyed
 265  * Context: !in_interrupt()
 266  *
 267  * Caller guarantees that no more memory from the pool is in use,
 268  * and that nothing will try to use the pool after this call.
 269  */
 270 void dma_pool_destroy(struct dma_pool *pool)
 271 {
 272         bool empty = false;
 273
 274         if (unlikely(!pool))
 275                 return;
 276
 277         mutex_lock(&pools_reg_lock);
 278         mutex_lock(&pools_lock);
 279         list_del(&pool->pools);
 280         if (pool->dev && list_empty(&pool->dev->dma_pools))
 281                 empty = true;
 282         mutex_unlock(&pools_lock);
 283         if (empty)
 284                 device_remove_file(pool->dev, &dev_attr_pools);
 285         mutex_unlock(&pools_reg_lock);
 286
 287         while (!list_empty(&pool->page_list)) {
 288                 struct dma_page *page;
 289                 page = list_entry(pool->page_list.next,
 290                                   struct dma_page, page_list);
 291                 if (is_page_busy(page)) {
 292                         if (pool->dev)
 293                                 dev_err(pool->dev,
 294                                         "dma_pool_destroy %s, %p busy\n",
 295                                         pool->name, page->vaddr);
 296                         else
 297                                 pr_err("dma_pool_destroy %s, %p busy\n",
 298                                        pool->name, page->vaddr);
 299                         /* leak the still-in-use consistent memory */
 300                         list_del(&page->page_list);
 301                         kfree(page);
 302                 } else
 303                         pool_free_page(pool, page);
 304         }
 305
 306         kfree(pool);
 307 }
 308 EXPORT_SYMBOL(dma_pool_destroy);
 309
 310 /**
 311  * dma_pool_alloc - get a block of consistent memory
 312  * @pool: dma pool that will produce the block
 313  * @mem_flags: GFP_* bitmask
 314  * @handle: pointer to dma address of block
 315  *
 316  * This returns the kernel virtual address of a currently unused block,
 317  * and reports its dma address through the handle.
 318  * If such a memory block can't be allocated, %NULL is returned.
 319  */
 320 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
 321                      dma_addr_t *handle)
 322 {
 323         unsigned long flags;
 324         struct dma_page *page;
 325         size_t offset;
 326         void *retval;
 327
 328         might_sleep_if(gfpflags_allow_blocking(mem_flags));
 329
 330         spin_lock_irqsave(&pool->lock, flags);
 331         list_for_each_entry(page, &pool->page_list, page_list) {
 332                 if (page->offset < pool->allocation)
 333                         goto ready;
 334         }
 335
 336         /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */
 337         spin_unlock_irqrestore(&pool->lock, flags);
 338
 339         page = pool_alloc_page(pool, mem_flags & (~__GFP_ZERO));
 340         if (!page)
 341                 return NULL;
 342
 343         spin_lock_irqsave(&pool->lock, flags);
 344
 345         list_add(&page->page_list, &pool->page_list);
 346  ready:
 347         page->in_use++;
 348         offset = page->offset;
 349         page->offset = *(int *)(page->vaddr + offset);
 350         retval = offset + page->vaddr;
 351         *handle = offset + page->dma;
 352 #ifdef  DMAPOOL_DEBUG
 353         {
 354                 int i;
 355                 u8 *data = retval;
 356                 /* page->offset is stored in first 4 bytes */
 357                 for (i = sizeof(page->offset); i < pool->size; i++) {
 358                         if (data[i] == POOL_POISON_FREED)
 359                                 continue;
 360                         if (pool->dev)
 361                                 dev_err(pool->dev,
 362                                         "dma_pool_alloc %s, %p (corrupted)\n",
 363                                         pool->name, retval);
 364                         else
 365                                 pr_err("dma_pool_alloc %s, %p (corrupted)\n",
 366                                         pool->name, retval);
 367
 368                         /*
 369                          * Dump the first 4 bytes even if they are not
 370                          * POOL_POISON_FREED
 371                          */
 372                         print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1,
 373                                         data, pool->size, 1);
 374                         break;
 375                 }
 376         }
 377         if (!(mem_flags & __GFP_ZERO))
 378                 memset(retval, POOL_POISON_ALLOCATED, pool->size);
 379 #endif
 380         spin_unlock_irqrestore(&pool->lock, flags);
 381
 382         if (mem_flags & __GFP_ZERO)
 383                 memset(retval, 0, pool->size);
 384
 385         return retval;
 386 }
 387 EXPORT_SYMBOL(dma_pool_alloc);
 388
 389 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
 390 {
 391         struct dma_page *page;
 392
 393         list_for_each_entry(page, &pool->page_list, page_list) {
 394                 if (dma < page->dma)
 395                         continue;
 396                 if ((dma - page->dma) < pool->allocation)
 397                         return page;
 398         }
 399         return NULL;
 400 }
 401
 402 /**
 403  * dma_pool_free - put block back into dma pool
 404  * @pool: the dma pool holding the block
 405  * @vaddr: virtual address of block
 406  * @dma: dma address of block
 407  *
 408  * Caller promises neither device nor driver will again touch this block
 409  * unless it is first re-allocated.
 410  */
 411 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
 412 {
 413         struct dma_page *page;
 414         unsigned long flags;
 415         unsigned int offset;
 416
 417         spin_lock_irqsave(&pool->lock, flags);
 418         page = pool_find_page(pool, dma);
 419         if (!page) {
 420                 spin_unlock_irqrestore(&pool->lock, flags);
 421                 if (pool->dev)
 422                         dev_err(pool->dev,
 423                                 "dma_pool_free %s, %p/%lx (bad dma)\n",
 424                                 pool->name, vaddr, (unsigned long)dma);
 425                 else
 426                         pr_err("dma_pool_free %s, %p/%lx (bad dma)\n",
 427                                pool->name, vaddr, (unsigned long)dma);
 428                 return;
 429         }
 430
 431         offset = vaddr - page->vaddr;
 432 #ifdef  DMAPOOL_DEBUG
 433         if ((dma - page->dma) != offset) {
 434                 spin_unlock_irqrestore(&pool->lock, flags);
 435                 if (pool->dev)
 436                         dev_err(pool->dev,
 437                                 "dma_pool_free %s, %p (bad vaddr)/%pad\n",
 438                                 pool->name, vaddr, &dma);
 439                 else
 440                         pr_err("dma_pool_free %s, %p (bad vaddr)/%pad\n",
 441                                pool->name, vaddr, &dma);
 442                 return;
 443         }
 444         {
 445                 unsigned int chain = page->offset;
 446                 while (chain < pool->allocation) {
 447                         if (chain != offset) {
 448                                 chain = *(int *)(page->vaddr + chain);
 449                                 continue;
 450                         }
 451                         spin_unlock_irqrestore(&pool->lock, flags);
 452                         if (pool->dev)
 453                                 dev_err(pool->dev, "dma_pool_free %s, dma %pad already free\n",
 454                                         pool->name, &dma);
 455                         else
 456                                 pr_err("dma_pool_free %s, dma %pad already free\n",
 457                                        pool->name, &dma);
 458                         return;
 459                 }
 460         }
 461         memset(vaddr, POOL_POISON_FREED, pool->size);
 462 #endif
 463
 464         page->in_use--;
 465         *(int *)vaddr = page->offset;
 466         page->offset = offset;
 467         /*
 468          * Resist a temptation to do
 469          *    if (!is_page_busy(page)) pool_free_page(pool, page);
 470          * Better have a few empty pages hang around.
 471          */
 472         spin_unlock_irqrestore(&pool->lock, flags);
 473 }
 474 EXPORT_SYMBOL(dma_pool_free);
 475
 476 /*
 477  * Managed DMA pool
 478  */
 479 static void dmam_pool_release(struct device *dev, void *res)
 480 {
 481         struct dma_pool *pool = *(struct dma_pool **)res;
 482
 483         dma_pool_destroy(pool);
 484 }
 485
 486 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
 487 {
 488         return *(struct dma_pool **)res == match_data;
 489 }
 490
 491 /**
 492  * dmam_pool_create - Managed dma_pool_create()
 493  * @name: name of pool, for diagnostics
 494  * @dev: device that will be doing the DMA
 495  * @size: size of the blocks in this pool.
 496  * @align: alignment requirement for blocks; must be a power of two
 497  * @allocation: returned blocks won't cross this boundary (or zero)
 498  *
 499  * Managed dma_pool_create().  DMA pool created with this function is
 500  * automatically destroyed on driver detach.
 501  */
 502 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
 503                                   size_t size, size_t align, size_t allocation)
 504 {
 505         struct dma_pool **ptr, *pool;
 506
 507         ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
 508         if (!ptr)
 509                 return NULL;
 510
 511         pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
 512         if (pool)
 513                 devres_add(dev, ptr);
 514         else
 515                 devres_free(ptr);
 516
 517         return pool;
 518 }
 519 EXPORT_SYMBOL(dmam_pool_create);
 520
 521 /**
 522  * dmam_pool_destroy - Managed dma_pool_destroy()
 523  * @pool: dma pool that will be destroyed
 524  *
 525  * Managed dma_pool_destroy().
 526  */
 527 void dmam_pool_destroy(struct dma_pool *pool)
 528 {
 529         struct device *dev = pool->dev;
 530
 531         WARN_ON(devres_release(dev, dmam_pool_release, dmam_pool_match, pool));
 532 }
 533 EXPORT_SYMBOL(dmam_pool_destroy);