4 * Copyright 2001 David Brownell
5 * Copyright 2007 Intel Corporation
6 * Author: Matthew Wilcox <willy@linux.intel.com>
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.
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.
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.
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>
41 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_SLUB_DEBUG_ON)
42 #define DMAPOOL_DEBUG 1
45 struct dma_pool { /* the pool */
46 struct list_head page_list;
53 struct list_head pools;
56 struct dma_page { /* cacheable header for 'allocation' bytes */
57 struct list_head page_list;
64 static DEFINE_MUTEX(pools_lock);
65 static DEFINE_MUTEX(pools_reg_lock);
68 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
73 struct dma_page *page;
74 struct dma_pool *pool;
79 temp = scnprintf(next, size, "poolinfo - 0.1\n");
83 mutex_lock(&pools_lock);
84 list_for_each_entry(pool, &dev->dma_pools, pools) {
88 spin_lock_irq(&pool->lock);
89 list_for_each_entry(page, &pool->page_list, page_list) {
91 blocks += page->in_use;
93 spin_unlock_irq(&pool->lock);
95 /* per-pool info, no real statistics yet */
96 temp = scnprintf(next, size, "%-16s %4u %4zu %4zu %2u\n",
98 pages * (pool->allocation / pool->size),
103 mutex_unlock(&pools_lock);
105 return PAGE_SIZE - size;
108 static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL);
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()
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.
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.
131 struct dma_pool *dma_pool_create(const char *name, struct device *dev,
132 size_t size, size_t align, size_t boundary)
134 struct dma_pool *retval;
140 else if (align & (align - 1))
148 if ((size % align) != 0)
149 size = ALIGN(size, align);
151 allocation = max_t(size_t, size, PAGE_SIZE);
154 boundary = allocation;
155 else if ((boundary < size) || (boundary & (boundary - 1)))
158 retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
162 strlcpy(retval->name, name, sizeof(retval->name));
166 INIT_LIST_HEAD(&retval->page_list);
167 spin_lock_init(&retval->lock);
169 retval->boundary = boundary;
170 retval->allocation = allocation;
172 INIT_LIST_HEAD(&retval->pools);
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).
182 mutex_lock(&pools_reg_lock);
183 mutex_lock(&pools_lock);
184 if (list_empty(&dev->dma_pools))
186 list_add(&retval->pools, &dev->dma_pools);
187 mutex_unlock(&pools_lock);
191 err = device_create_file(dev, &dev_attr_pools);
193 mutex_lock(&pools_lock);
194 list_del(&retval->pools);
195 mutex_unlock(&pools_lock);
196 mutex_unlock(&pools_reg_lock);
201 mutex_unlock(&pools_reg_lock);
204 EXPORT_SYMBOL(dma_pool_create);
206 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
208 unsigned int offset = 0;
209 unsigned int next_boundary = pool->boundary;
212 unsigned int next = offset + pool->size;
213 if (unlikely((next + pool->size) >= next_boundary)) {
214 next = next_boundary;
215 next_boundary += pool->boundary;
217 *(int *)(page->vaddr + offset) = next;
219 } while (offset < pool->allocation);
222 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
224 struct dma_page *page;
226 page = kmalloc(sizeof(*page), mem_flags);
229 page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
230 &page->dma, mem_flags);
233 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
235 pool_initialise_page(pool, page);
245 static inline bool is_page_busy(struct dma_page *page)
247 return page->in_use != 0;
250 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
252 dma_addr_t dma = page->dma;
255 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
257 dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
258 list_del(&page->page_list);
263 * dma_pool_destroy - destroys a pool of dma memory blocks.
264 * @pool: dma pool that will be destroyed
265 * Context: !in_interrupt()
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.
270 void dma_pool_destroy(struct dma_pool *pool)
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))
282 mutex_unlock(&pools_lock);
284 device_remove_file(pool->dev, &dev_attr_pools);
285 mutex_unlock(&pools_reg_lock);
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)) {
294 "dma_pool_destroy %s, %p busy\n",
295 pool->name, page->vaddr);
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);
303 pool_free_page(pool, page);
308 EXPORT_SYMBOL(dma_pool_destroy);
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
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.
320 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
324 struct dma_page *page;
328 might_sleep_if(gfpflags_allow_blocking(mem_flags));
330 spin_lock_irqsave(&pool->lock, flags);
331 list_for_each_entry(page, &pool->page_list, page_list) {
332 if (page->offset < pool->allocation)
336 /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */
337 spin_unlock_irqrestore(&pool->lock, flags);
339 page = pool_alloc_page(pool, mem_flags & (~__GFP_ZERO));
343 spin_lock_irqsave(&pool->lock, flags);
345 list_add(&page->page_list, &pool->page_list);
348 offset = page->offset;
349 page->offset = *(int *)(page->vaddr + offset);
350 retval = offset + page->vaddr;
351 *handle = offset + page->dma;
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)
362 "dma_pool_alloc %s, %p (corrupted)\n",
365 pr_err("dma_pool_alloc %s, %p (corrupted)\n",
369 * Dump the first 4 bytes even if they are not
372 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1,
373 data, pool->size, 1);
377 if (!(mem_flags & __GFP_ZERO))
378 memset(retval, POOL_POISON_ALLOCATED, pool->size);
380 spin_unlock_irqrestore(&pool->lock, flags);
382 if (mem_flags & __GFP_ZERO)
383 memset(retval, 0, pool->size);
387 EXPORT_SYMBOL(dma_pool_alloc);
389 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
391 struct dma_page *page;
393 list_for_each_entry(page, &pool->page_list, page_list) {
396 if ((dma - page->dma) < pool->allocation)
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
408 * Caller promises neither device nor driver will again touch this block
409 * unless it is first re-allocated.
411 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
413 struct dma_page *page;
417 spin_lock_irqsave(&pool->lock, flags);
418 page = pool_find_page(pool, dma);
420 spin_unlock_irqrestore(&pool->lock, flags);
423 "dma_pool_free %s, %p/%lx (bad dma)\n",
424 pool->name, vaddr, (unsigned long)dma);
426 pr_err("dma_pool_free %s, %p/%lx (bad dma)\n",
427 pool->name, vaddr, (unsigned long)dma);
431 offset = vaddr - page->vaddr;
433 if ((dma - page->dma) != offset) {
434 spin_unlock_irqrestore(&pool->lock, flags);
437 "dma_pool_free %s, %p (bad vaddr)/%pad\n",
438 pool->name, vaddr, &dma);
440 pr_err("dma_pool_free %s, %p (bad vaddr)/%pad\n",
441 pool->name, vaddr, &dma);
445 unsigned int chain = page->offset;
446 while (chain < pool->allocation) {
447 if (chain != offset) {
448 chain = *(int *)(page->vaddr + chain);
451 spin_unlock_irqrestore(&pool->lock, flags);
453 dev_err(pool->dev, "dma_pool_free %s, dma %pad already free\n",
456 pr_err("dma_pool_free %s, dma %pad already free\n",
461 memset(vaddr, POOL_POISON_FREED, pool->size);
465 *(int *)vaddr = page->offset;
466 page->offset = offset;
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.
472 spin_unlock_irqrestore(&pool->lock, flags);
474 EXPORT_SYMBOL(dma_pool_free);
479 static void dmam_pool_release(struct device *dev, void *res)
481 struct dma_pool *pool = *(struct dma_pool **)res;
483 dma_pool_destroy(pool);
486 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
488 return *(struct dma_pool **)res == match_data;
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)
499 * Managed dma_pool_create(). DMA pool created with this function is
500 * automatically destroyed on driver detach.
502 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
503 size_t size, size_t align, size_t allocation)
505 struct dma_pool **ptr, *pool;
507 ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
511 pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
513 devres_add(dev, ptr);
519 EXPORT_SYMBOL(dmam_pool_create);
522 * dmam_pool_destroy - Managed dma_pool_destroy()
523 * @pool: dma pool that will be destroyed
525 * Managed dma_pool_destroy().
527 void dmam_pool_destroy(struct dma_pool *pool)
529 struct device *dev = pool->dev;
531 WARN_ON(devres_release(dev, dmam_pool_release, dmam_pool_match, pool));
533 EXPORT_SYMBOL(dmam_pool_destroy);