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)) {
146 } else if (size < 4) {
150 if ((size % align) != 0)
151 size = ALIGN(size, align);
153 allocation = max_t(size_t, size, PAGE_SIZE);
156 boundary = allocation;
157 } else if ((boundary < size) || (boundary & (boundary - 1))) {
161 retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
165 strlcpy(retval->name, name, sizeof(retval->name));
169 INIT_LIST_HEAD(&retval->page_list);
170 spin_lock_init(&retval->lock);
172 retval->boundary = boundary;
173 retval->allocation = allocation;
175 INIT_LIST_HEAD(&retval->pools);
178 * pools_lock ensures that the ->dma_pools list does not get corrupted.
179 * pools_reg_lock ensures that there is not a race between
180 * dma_pool_create() and dma_pool_destroy() or within dma_pool_create()
181 * when the first invocation of dma_pool_create() failed on
182 * device_create_file() and the second assumes that it has been done (I
183 * know it is a short window).
185 mutex_lock(&pools_reg_lock);
186 mutex_lock(&pools_lock);
187 if (list_empty(&dev->dma_pools))
189 list_add(&retval->pools, &dev->dma_pools);
190 mutex_unlock(&pools_lock);
194 err = device_create_file(dev, &dev_attr_pools);
196 mutex_lock(&pools_lock);
197 list_del(&retval->pools);
198 mutex_unlock(&pools_lock);
199 mutex_unlock(&pools_reg_lock);
204 mutex_unlock(&pools_reg_lock);
207 EXPORT_SYMBOL(dma_pool_create);
209 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
211 unsigned int offset = 0;
212 unsigned int next_boundary = pool->boundary;
215 unsigned int next = offset + pool->size;
216 if (unlikely((next + pool->size) >= next_boundary)) {
217 next = next_boundary;
218 next_boundary += pool->boundary;
220 *(int *)(page->vaddr + offset) = next;
222 } while (offset < pool->allocation);
225 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
227 struct dma_page *page;
229 page = kmalloc(sizeof(*page), mem_flags);
232 page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
233 &page->dma, mem_flags);
236 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
238 pool_initialise_page(pool, page);
248 static inline int is_page_busy(struct dma_page *page)
250 return page->in_use != 0;
253 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
255 dma_addr_t dma = page->dma;
258 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
260 dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
261 list_del(&page->page_list);
266 * dma_pool_destroy - destroys a pool of dma memory blocks.
267 * @pool: dma pool that will be destroyed
268 * Context: !in_interrupt()
270 * Caller guarantees that no more memory from the pool is in use,
271 * and that nothing will try to use the pool after this call.
273 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);
298 "dma_pool_destroy %s, %p busy\n",
299 pool->name, page->vaddr);
300 /* leak the still-in-use consistent memory */
301 list_del(&page->page_list);
304 pool_free_page(pool, page);
309 EXPORT_SYMBOL(dma_pool_destroy);
312 * dma_pool_alloc - get a block of consistent memory
313 * @pool: dma pool that will produce the block
314 * @mem_flags: GFP_* bitmask
315 * @handle: pointer to dma address of block
317 * This returns the kernel virtual address of a currently unused block,
318 * and reports its dma address through the handle.
319 * If such a memory block can't be allocated, %NULL is returned.
321 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
325 struct dma_page *page;
329 might_sleep_if(mem_flags & __GFP_WAIT);
331 spin_lock_irqsave(&pool->lock, flags);
332 list_for_each_entry(page, &pool->page_list, page_list) {
333 if (page->offset < pool->allocation)
337 /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */
338 spin_unlock_irqrestore(&pool->lock, flags);
340 page = pool_alloc_page(pool, mem_flags);
344 spin_lock_irqsave(&pool->lock, flags);
346 list_add(&page->page_list, &pool->page_list);
349 offset = page->offset;
350 page->offset = *(int *)(page->vaddr + offset);
351 retval = offset + page->vaddr;
352 *handle = offset + page->dma;
357 /* page->offset is stored in first 4 bytes */
358 for (i = sizeof(page->offset); i < pool->size; i++) {
359 if (data[i] == POOL_POISON_FREED)
363 "dma_pool_alloc %s, %p (corrupted)\n",
366 pr_err("dma_pool_alloc %s, %p (corrupted)\n",
370 * Dump the first 4 bytes even if they are not
373 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1,
374 data, pool->size, 1);
378 memset(retval, POOL_POISON_ALLOCATED, pool->size);
380 spin_unlock_irqrestore(&pool->lock, flags);
383 EXPORT_SYMBOL(dma_pool_alloc);
385 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
387 struct dma_page *page;
389 list_for_each_entry(page, &pool->page_list, page_list) {
392 if (dma < (page->dma + pool->allocation))
399 * dma_pool_free - put block back into dma pool
400 * @pool: the dma pool holding the block
401 * @vaddr: virtual address of block
402 * @dma: dma address of block
404 * Caller promises neither device nor driver will again touch this block
405 * unless it is first re-allocated.
407 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
409 struct dma_page *page;
413 spin_lock_irqsave(&pool->lock, flags);
414 page = pool_find_page(pool, dma);
416 spin_unlock_irqrestore(&pool->lock, flags);
419 "dma_pool_free %s, %p/%lx (bad dma)\n",
420 pool->name, vaddr, (unsigned long)dma);
422 printk(KERN_ERR "dma_pool_free %s, %p/%lx (bad dma)\n",
423 pool->name, vaddr, (unsigned long)dma);
427 offset = vaddr - page->vaddr;
429 if ((dma - page->dma) != offset) {
430 spin_unlock_irqrestore(&pool->lock, flags);
433 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
434 pool->name, vaddr, (unsigned long long)dma);
437 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
438 pool->name, vaddr, (unsigned long long)dma);
442 unsigned int chain = page->offset;
443 while (chain < pool->allocation) {
444 if (chain != offset) {
445 chain = *(int *)(page->vaddr + chain);
448 spin_unlock_irqrestore(&pool->lock, flags);
450 dev_err(pool->dev, "dma_pool_free %s, dma %Lx "
451 "already free\n", pool->name,
452 (unsigned long long)dma);
454 printk(KERN_ERR "dma_pool_free %s, dma %Lx "
455 "already free\n", pool->name,
456 (unsigned long long)dma);
460 memset(vaddr, POOL_POISON_FREED, pool->size);
464 *(int *)vaddr = page->offset;
465 page->offset = offset;
467 * Resist a temptation to do
468 * if (!is_page_busy(page)) pool_free_page(pool, page);
469 * Better have a few empty pages hang around.
471 spin_unlock_irqrestore(&pool->lock, flags);
473 EXPORT_SYMBOL(dma_pool_free);
478 static void dmam_pool_release(struct device *dev, void *res)
480 struct dma_pool *pool = *(struct dma_pool **)res;
482 dma_pool_destroy(pool);
485 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
487 return *(struct dma_pool **)res == match_data;
491 * dmam_pool_create - Managed dma_pool_create()
492 * @name: name of pool, for diagnostics
493 * @dev: device that will be doing the DMA
494 * @size: size of the blocks in this pool.
495 * @align: alignment requirement for blocks; must be a power of two
496 * @allocation: returned blocks won't cross this boundary (or zero)
498 * Managed dma_pool_create(). DMA pool created with this function is
499 * automatically destroyed on driver detach.
501 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
502 size_t size, size_t align, size_t allocation)
504 struct dma_pool **ptr, *pool;
506 ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
510 pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
512 devres_add(dev, ptr);
518 EXPORT_SYMBOL(dmam_pool_create);
521 * dmam_pool_destroy - Managed dma_pool_destroy()
522 * @pool: dma pool that will be destroyed
524 * Managed dma_pool_destroy().
526 void dmam_pool_destroy(struct dma_pool *pool)
528 struct device *dev = pool->dev;
530 WARN_ON(devres_release(dev, dmam_pool_release, dmam_pool_match, pool));
532 EXPORT_SYMBOL(dmam_pool_destroy);