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/module.h>
  31 #include <linux/mutex.h>
  32 #include <linux/poison.h>
  33 #include <linux/sched.h>
  34 #include <linux/slab.h>
  35 #include <linux/spinlock.h>
  36 #include <linux/string.h>
  37 #include <linux/types.h>
  38 #include <linux/wait.h>
  39
  40 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_SLUB_DEBUG_ON)
  41 #define DMAPOOL_DEBUG 1
  42 #endif
  43
  44 struct dma_pool {               /* the pool */
  45         struct list_head page_list;
  46         spinlock_t lock;
  47         size_t size;
  48         struct device *dev;
  49         size_t allocation;
  50         size_t boundary;
  51         char name[32];
  52         wait_queue_head_t waitq;
  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 #define POOL_TIMEOUT_JIFFIES    ((100 /* msec */ * HZ) / 1000)
  65
  66 static DEFINE_MUTEX(pools_lock);
  67
  68 static ssize_t
  69 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
  70 {
  71         unsigned temp;
  72         unsigned size;
  73         char *next;
  74         struct dma_page *page;
  75         struct dma_pool *pool;
  76
  77         next = buf;
  78         size = PAGE_SIZE;
  79
  80         temp = scnprintf(next, size, "poolinfo - 0.1\n");
  81         size -= temp;
  82         next += temp;
  83
  84         mutex_lock(&pools_lock);
  85         list_for_each_entry(pool, &dev->dma_pools, pools) {
  86                 unsigned pages = 0;
  87                 unsigned blocks = 0;
  88
  89                 spin_lock_irq(&pool->lock);
  90                 list_for_each_entry(page, &pool->page_list, page_list) {
  91                         pages++;
  92                         blocks += page->in_use;
  93                 }
  94                 spin_unlock_irq(&pool->lock);
  95
  96                 /* per-pool info, no real statistics yet */
  97                 temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n",
  98                                  pool->name, blocks,
  99                                  pages * (pool->allocation / pool->size),
 100                                  pool->size, pages);
 101                 size -= temp;
 102                 next += temp;
 103         }
 104         mutex_unlock(&pools_lock);
 105
 106         return PAGE_SIZE - size;
 107 }
 108
 109 static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL);
 110
 111 /**
 112  * dma_pool_create - Creates a pool of consistent memory blocks, for dma.
 113  * @name: name of pool, for diagnostics
 114  * @dev: device that will be doing the DMA
 115  * @size: size of the blocks in this pool.
 116  * @align: alignment requirement for blocks; must be a power of two
 117  * @boundary: returned blocks won't cross this power of two boundary
 118  * Context: !in_interrupt()
 119  *
 120  * Returns a dma allocation pool with the requested characteristics, or
 121  * null if one can't be created.  Given one of these pools, dma_pool_alloc()
 122  * may be used to allocate memory.  Such memory will all have "consistent"
 123  * DMA mappings, accessible by the device and its driver without using
 124  * cache flushing primitives.  The actual size of blocks allocated may be
 125  * larger than requested because of alignment.
 126  *
 127  * If @boundary is nonzero, objects returned from dma_pool_alloc() won't
 128  * cross that size boundary.  This is useful for devices which have
 129  * addressing restrictions on individual DMA transfers, such as not crossing
 130  * boundaries of 4KBytes.
 131  */
 132 struct dma_pool *dma_pool_create(const char *name, struct device *dev,
 133                                  size_t size, size_t align, size_t boundary)
 134 {
 135         struct dma_pool *retval;
 136         size_t allocation;
 137
 138         if (align == 0) {
 139                 align = 1;
 140         } else if (align & (align - 1)) {
 141                 return NULL;
 142         }
 143
 144         if (size == 0) {
 145                 return NULL;
 146         } else if (size < 4) {
 147                 size = 4;
 148         }
 149
 150         if ((size % align) != 0)
 151                 size = ALIGN(size, align);
 152
 153         allocation = max_t(size_t, size, PAGE_SIZE);
 154
 155         if (!boundary) {
 156                 boundary = allocation;
 157         } else if ((boundary < size) || (boundary & (boundary - 1))) {
 158                 return NULL;
 159         }
 160
 161         retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
 162         if (!retval)
 163                 return retval;
 164
 165         strlcpy(retval->name, name, sizeof(retval->name));
 166
 167         retval->dev = dev;
 168
 169         INIT_LIST_HEAD(&retval->page_list);
 170         spin_lock_init(&retval->lock);
 171         retval->size = size;
 172         retval->boundary = boundary;
 173         retval->allocation = allocation;
 174         init_waitqueue_head(&retval->waitq);
 175
 176         if (dev) {
 177                 int ret;
 178
 179                 mutex_lock(&pools_lock);
 180                 if (list_empty(&dev->dma_pools))
 181                         ret = device_create_file(dev, &dev_attr_pools);
 182                 else
 183                         ret = 0;
 184                 /* note:  not currently insisting "name" be unique */
 185                 if (!ret)
 186                         list_add(&retval->pools, &dev->dma_pools);
 187                 else {
 188                         kfree(retval);
 189                         retval = NULL;
 190                 }
 191                 mutex_unlock(&pools_lock);
 192         } else
 193                 INIT_LIST_HEAD(&retval->pools);
 194
 195         return retval;
 196 }
 197 EXPORT_SYMBOL(dma_pool_create);
 198
 199 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
 200 {
 201         unsigned int offset = 0;
 202         unsigned int next_boundary = pool->boundary;
 203
 204         do {
 205                 unsigned int next = offset + pool->size;
 206                 if (unlikely((next + pool->size) >= next_boundary)) {
 207                         next = next_boundary;
 208                         next_boundary += pool->boundary;
 209                 }
 210                 *(int *)(page->vaddr + offset) = next;
 211                 offset = next;
 212         } while (offset < pool->allocation);
 213 }
 214
 215 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
 216 {
 217         struct dma_page *page;
 218
 219         page = kmalloc(sizeof(*page), mem_flags);
 220         if (!page)
 221                 return NULL;
 222         page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
 223                                          &page->dma, mem_flags);
 224         if (page->vaddr) {
 225 #ifdef  DMAPOOL_DEBUG
 226                 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 227 #endif
 228                 pool_initialise_page(pool, page);
 229                 list_add(&page->page_list, &pool->page_list);
 230                 page->in_use = 0;
 231                 page->offset = 0;
 232         } else {
 233                 kfree(page);
 234                 page = NULL;
 235         }
 236         return page;
 237 }
 238
 239 static inline int is_page_busy(struct dma_page *page)
 240 {
 241         return page->in_use != 0;
 242 }
 243
 244 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
 245 {
 246         dma_addr_t dma = page->dma;
 247
 248 #ifdef  DMAPOOL_DEBUG
 249         memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 250 #endif
 251         dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
 252         list_del(&page->page_list);
 253         kfree(page);
 254 }
 255
 256 /**
 257  * dma_pool_destroy - destroys a pool of dma memory blocks.
 258  * @pool: dma pool that will be destroyed
 259  * Context: !in_interrupt()
 260  *
 261  * Caller guarantees that no more memory from the pool is in use,
 262  * and that nothing will try to use the pool after this call.
 263  */
 264 void dma_pool_destroy(struct dma_pool *pool)
 265 {
 266         mutex_lock(&pools_lock);
 267         list_del(&pool->pools);
 268         if (pool->dev && list_empty(&pool->dev->dma_pools))
 269                 device_remove_file(pool->dev, &dev_attr_pools);
 270         mutex_unlock(&pools_lock);
 271
 272         while (!list_empty(&pool->page_list)) {
 273                 struct dma_page *page;
 274                 page = list_entry(pool->page_list.next,
 275                                   struct dma_page, page_list);
 276                 if (is_page_busy(page)) {
 277                         if (pool->dev)
 278                                 dev_err(pool->dev,
 279                                         "dma_pool_destroy %s, %p busy\n",
 280                                         pool->name, page->vaddr);
 281                         else
 282                                 printk(KERN_ERR
 283                                        "dma_pool_destroy %s, %p busy\n",
 284                                        pool->name, page->vaddr);
 285                         /* leak the still-in-use consistent memory */
 286                         list_del(&page->page_list);
 287                         kfree(page);
 288                 } else
 289                         pool_free_page(pool, page);
 290         }
 291
 292         kfree(pool);
 293 }
 294 EXPORT_SYMBOL(dma_pool_destroy);
 295
 296 /**
 297  * dma_pool_alloc - get a block of consistent memory
 298  * @pool: dma pool that will produce the block
 299  * @mem_flags: GFP_* bitmask
 300  * @handle: pointer to dma address of block
 301  *
 302  * This returns the kernel virtual address of a currently unused block,
 303  * and reports its dma address through the handle.
 304  * If such a memory block can't be allocated, %NULL is returned.
 305  */
 306 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
 307                      dma_addr_t *handle)
 308 {
 309         unsigned long flags;
 310         struct dma_page *page;
 311         size_t offset;
 312         void *retval;
 313
 314         might_sleep_if(mem_flags & __GFP_WAIT);
 315
 316         spin_lock_irqsave(&pool->lock, flags);
 317  restart:
 318         list_for_each_entry(page, &pool->page_list, page_list) {
 319                 if (page->offset < pool->allocation)
 320                         goto ready;
 321         }
 322         page = pool_alloc_page(pool, GFP_ATOMIC);
 323         if (!page) {
 324                 if (mem_flags & __GFP_WAIT) {
 325                         DECLARE_WAITQUEUE(wait, current);
 326
 327                         __set_current_state(TASK_INTERRUPTIBLE);
 328                         __add_wait_queue(&pool->waitq, &wait);
 329                         spin_unlock_irqrestore(&pool->lock, flags);
 330
 331                         schedule_timeout(POOL_TIMEOUT_JIFFIES);
 332
 333                         spin_lock_irqsave(&pool->lock, flags);
 334                         __remove_wait_queue(&pool->waitq, &wait);
 335                         goto restart;
 336                 }
 337                 retval = NULL;
 338                 goto done;
 339         }
 340
 341  ready:
 342         page->in_use++;
 343         offset = page->offset;
 344         page->offset = *(int *)(page->vaddr + offset);
 345         retval = offset + page->vaddr;
 346         *handle = offset + page->dma;
 347 #ifdef  DMAPOOL_DEBUG
 348         memset(retval, POOL_POISON_ALLOCATED, pool->size);
 349 #endif
 350  done:
 351         spin_unlock_irqrestore(&pool->lock, flags);
 352         return retval;
 353 }
 354 EXPORT_SYMBOL(dma_pool_alloc);
 355
 356 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
 357 {
 358         unsigned long flags;
 359         struct dma_page *page;
 360
 361         spin_lock_irqsave(&pool->lock, flags);
 362         list_for_each_entry(page, &pool->page_list, page_list) {
 363                 if (dma < page->dma)
 364                         continue;
 365                 if (dma < (page->dma + pool->allocation))
 366                         goto done;
 367         }
 368         page = NULL;
 369  done:
 370         spin_unlock_irqrestore(&pool->lock, flags);
 371         return page;
 372 }
 373
 374 /**
 375  * dma_pool_free - put block back into dma pool
 376  * @pool: the dma pool holding the block
 377  * @vaddr: virtual address of block
 378  * @dma: dma address of block
 379  *
 380  * Caller promises neither device nor driver will again touch this block
 381  * unless it is first re-allocated.
 382  */
 383 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
 384 {
 385         struct dma_page *page;
 386         unsigned long flags;
 387         unsigned int offset;
 388
 389         page = pool_find_page(pool, dma);
 390         if (!page) {
 391                 if (pool->dev)
 392                         dev_err(pool->dev,
 393                                 "dma_pool_free %s, %p/%lx (bad dma)\n",
 394                                 pool->name, vaddr, (unsigned long)dma);
 395                 else
 396                         printk(KERN_ERR "dma_pool_free %s, %p/%lx (bad dma)\n",
 397                                pool->name, vaddr, (unsigned long)dma);
 398                 return;
 399         }
 400
 401         offset = vaddr - page->vaddr;
 402 #ifdef  DMAPOOL_DEBUG
 403         if ((dma - page->dma) != offset) {
 404                 if (pool->dev)
 405                         dev_err(pool->dev,
 406                                 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
 407                                 pool->name, vaddr, (unsigned long long)dma);
 408                 else
 409                         printk(KERN_ERR
 410                                "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
 411                                pool->name, vaddr, (unsigned long long)dma);
 412                 return;
 413         }
 414         {
 415                 unsigned int chain = page->offset;
 416                 while (chain < pool->allocation) {
 417                         if (chain != offset) {
 418                                 chain = *(int *)(page->vaddr + chain);
 419                                 continue;
 420                         }
 421                         if (pool->dev)
 422                                 dev_err(pool->dev, "dma_pool_free %s, dma %Lx "
 423                                         "already free\n", pool->name,
 424                                         (unsigned long long)dma);
 425                         else
 426                                 printk(KERN_ERR "dma_pool_free %s, dma %Lx "
 427                                         "already free\n", pool->name,
 428                                         (unsigned long long)dma);
 429                         return;
 430                 }
 431         }
 432         memset(vaddr, POOL_POISON_FREED, pool->size);
 433 #endif
 434
 435         spin_lock_irqsave(&pool->lock, flags);
 436         page->in_use--;
 437         *(int *)vaddr = page->offset;
 438         page->offset = offset;
 439         if (waitqueue_active(&pool->waitq))
 440                 wake_up_locked(&pool->waitq);
 441         /*
 442          * Resist a temptation to do
 443          *    if (!is_page_busy(page)) pool_free_page(pool, page);
 444          * Better have a few empty pages hang around.
 445          */
 446         spin_unlock_irqrestore(&pool->lock, flags);
 447 }
 448 EXPORT_SYMBOL(dma_pool_free);
 449
 450 /*
 451  * Managed DMA pool
 452  */
 453 static void dmam_pool_release(struct device *dev, void *res)
 454 {
 455         struct dma_pool *pool = *(struct dma_pool **)res;
 456
 457         dma_pool_destroy(pool);
 458 }
 459
 460 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
 461 {
 462         return *(struct dma_pool **)res == match_data;
 463 }
 464
 465 /**
 466  * dmam_pool_create - Managed dma_pool_create()
 467  * @name: name of pool, for diagnostics
 468  * @dev: device that will be doing the DMA
 469  * @size: size of the blocks in this pool.
 470  * @align: alignment requirement for blocks; must be a power of two
 471  * @allocation: returned blocks won't cross this boundary (or zero)
 472  *
 473  * Managed dma_pool_create().  DMA pool created with this function is
 474  * automatically destroyed on driver detach.
 475  */
 476 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
 477                                   size_t size, size_t align, size_t allocation)
 478 {
 479         struct dma_pool **ptr, *pool;
 480
 481         ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
 482         if (!ptr)
 483                 return NULL;
 484
 485         pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
 486         if (pool)
 487                 devres_add(dev, ptr);
 488         else
 489                 devres_free(ptr);
 490
 491         return pool;
 492 }
 493 EXPORT_SYMBOL(dmam_pool_create);
 494
 495 /**
 496  * dmam_pool_destroy - Managed dma_pool_destroy()
 497  * @pool: dma pool that will be destroyed
 498  *
 499  * Managed dma_pool_destroy().
 500  */
 501 void dmam_pool_destroy(struct dma_pool *pool)
 502 {
 503         struct device *dev = pool->dev;
 504
 505         dma_pool_destroy(pool);
 506         WARN_ON(devres_destroy(dev, dmam_pool_release, dmam_pool_match, pool));
 507 }
 508 EXPORT_SYMBOL(dmam_pool_destroy);