numa: convert static memory to dynamically allocated memory for per node device
[linux-2.6.git] / mm / dmapool.c
blobda1b0f0b8709d1eaf485fcd8bd8c86c0e69db6d6
1 /*
2 * DMA Pool allocator
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
43 #endif
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;
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;
64 static DEFINE_MUTEX(pools_lock);
66 static ssize_t
67 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
69 unsigned temp;
70 unsigned size;
71 char *next;
72 struct dma_page *page;
73 struct dma_pool *pool;
75 next = buf;
76 size = PAGE_SIZE;
78 temp = scnprintf(next, size, "poolinfo - 0.1\n");
79 size -= temp;
80 next += temp;
82 mutex_lock(&pools_lock);
83 list_for_each_entry(pool, &dev->dma_pools, pools) {
84 unsigned pages = 0;
85 unsigned blocks = 0;
87 spin_lock_irq(&pool->lock);
88 list_for_each_entry(page, &pool->page_list, page_list) {
89 pages++;
90 blocks += page->in_use;
92 spin_unlock_irq(&pool->lock);
94 /* per-pool info, no real statistics yet */
95 temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n",
96 pool->name, blocks,
97 pages * (pool->allocation / pool->size),
98 pool->size, pages);
99 size -= temp;
100 next += temp;
102 mutex_unlock(&pools_lock);
104 return PAGE_SIZE - size;
107 static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL);
110 * dma_pool_create - Creates a pool of consistent memory blocks, for dma.
111 * @name: name of pool, for diagnostics
112 * @dev: device that will be doing the DMA
113 * @size: size of the blocks in this pool.
114 * @align: alignment requirement for blocks; must be a power of two
115 * @boundary: returned blocks won't cross this power of two boundary
116 * Context: !in_interrupt()
118 * Returns a dma allocation pool with the requested characteristics, or
119 * null if one can't be created. Given one of these pools, dma_pool_alloc()
120 * may be used to allocate memory. Such memory will all have "consistent"
121 * DMA mappings, accessible by the device and its driver without using
122 * cache flushing primitives. The actual size of blocks allocated may be
123 * larger than requested because of alignment.
125 * If @boundary is nonzero, objects returned from dma_pool_alloc() won't
126 * cross that size boundary. This is useful for devices which have
127 * addressing restrictions on individual DMA transfers, such as not crossing
128 * boundaries of 4KBytes.
130 struct dma_pool *dma_pool_create(const char *name, struct device *dev,
131 size_t size, size_t align, size_t boundary)
133 struct dma_pool *retval;
134 size_t allocation;
136 if (align == 0) {
137 align = 1;
138 } else if (align & (align - 1)) {
139 return NULL;
142 if (size == 0) {
143 return NULL;
144 } else if (size < 4) {
145 size = 4;
148 if ((size % align) != 0)
149 size = ALIGN(size, align);
151 allocation = max_t(size_t, size, PAGE_SIZE);
153 if (!boundary) {
154 boundary = allocation;
155 } else if ((boundary < size) || (boundary & (boundary - 1))) {
156 return NULL;
159 retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
160 if (!retval)
161 return retval;
163 strlcpy(retval->name, name, sizeof(retval->name));
165 retval->dev = dev;
167 INIT_LIST_HEAD(&retval->page_list);
168 spin_lock_init(&retval->lock);
169 retval->size = size;
170 retval->boundary = boundary;
171 retval->allocation = allocation;
173 if (dev) {
174 int ret;
176 mutex_lock(&pools_lock);
177 if (list_empty(&dev->dma_pools))
178 ret = device_create_file(dev, &dev_attr_pools);
179 else
180 ret = 0;
181 /* note: not currently insisting "name" be unique */
182 if (!ret)
183 list_add(&retval->pools, &dev->dma_pools);
184 else {
185 kfree(retval);
186 retval = NULL;
188 mutex_unlock(&pools_lock);
189 } else
190 INIT_LIST_HEAD(&retval->pools);
192 return retval;
194 EXPORT_SYMBOL(dma_pool_create);
196 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
198 unsigned int offset = 0;
199 unsigned int next_boundary = pool->boundary;
201 do {
202 unsigned int next = offset + pool->size;
203 if (unlikely((next + pool->size) >= next_boundary)) {
204 next = next_boundary;
205 next_boundary += pool->boundary;
207 *(int *)(page->vaddr + offset) = next;
208 offset = next;
209 } while (offset < pool->allocation);
212 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
214 struct dma_page *page;
216 page = kmalloc(sizeof(*page), mem_flags);
217 if (!page)
218 return NULL;
219 page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
220 &page->dma, mem_flags);
221 if (page->vaddr) {
222 #ifdef DMAPOOL_DEBUG
223 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
224 #endif
225 pool_initialise_page(pool, page);
226 page->in_use = 0;
227 page->offset = 0;
228 } else {
229 kfree(page);
230 page = NULL;
232 return page;
235 static inline int is_page_busy(struct dma_page *page)
237 return page->in_use != 0;
240 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
242 dma_addr_t dma = page->dma;
244 #ifdef DMAPOOL_DEBUG
245 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
246 #endif
247 dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
248 list_del(&page->page_list);
249 kfree(page);
253 * dma_pool_destroy - destroys a pool of dma memory blocks.
254 * @pool: dma pool that will be destroyed
255 * Context: !in_interrupt()
257 * Caller guarantees that no more memory from the pool is in use,
258 * and that nothing will try to use the pool after this call.
260 void dma_pool_destroy(struct dma_pool *pool)
262 mutex_lock(&pools_lock);
263 list_del(&pool->pools);
264 if (pool->dev && list_empty(&pool->dev->dma_pools))
265 device_remove_file(pool->dev, &dev_attr_pools);
266 mutex_unlock(&pools_lock);
268 while (!list_empty(&pool->page_list)) {
269 struct dma_page *page;
270 page = list_entry(pool->page_list.next,
271 struct dma_page, page_list);
272 if (is_page_busy(page)) {
273 if (pool->dev)
274 dev_err(pool->dev,
275 "dma_pool_destroy %s, %p busy\n",
276 pool->name, page->vaddr);
277 else
278 printk(KERN_ERR
279 "dma_pool_destroy %s, %p busy\n",
280 pool->name, page->vaddr);
281 /* leak the still-in-use consistent memory */
282 list_del(&page->page_list);
283 kfree(page);
284 } else
285 pool_free_page(pool, page);
288 kfree(pool);
290 EXPORT_SYMBOL(dma_pool_destroy);
293 * dma_pool_alloc - get a block of consistent memory
294 * @pool: dma pool that will produce the block
295 * @mem_flags: GFP_* bitmask
296 * @handle: pointer to dma address of block
298 * This returns the kernel virtual address of a currently unused block,
299 * and reports its dma address through the handle.
300 * If such a memory block can't be allocated, %NULL is returned.
302 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
303 dma_addr_t *handle)
305 unsigned long flags;
306 struct dma_page *page;
307 size_t offset;
308 void *retval;
310 might_sleep_if(mem_flags & __GFP_WAIT);
312 spin_lock_irqsave(&pool->lock, flags);
313 list_for_each_entry(page, &pool->page_list, page_list) {
314 if (page->offset < pool->allocation)
315 goto ready;
318 /* pool_alloc_page() might sleep, so temporarily drop &pool->lock */
319 spin_unlock_irqrestore(&pool->lock, flags);
321 page = pool_alloc_page(pool, mem_flags);
322 if (!page)
323 return NULL;
325 spin_lock_irqsave(&pool->lock, flags);
327 list_add(&page->page_list, &pool->page_list);
328 ready:
329 page->in_use++;
330 offset = page->offset;
331 page->offset = *(int *)(page->vaddr + offset);
332 retval = offset + page->vaddr;
333 *handle = offset + page->dma;
334 #ifdef DMAPOOL_DEBUG
335 memset(retval, POOL_POISON_ALLOCATED, pool->size);
336 #endif
337 spin_unlock_irqrestore(&pool->lock, flags);
338 return retval;
340 EXPORT_SYMBOL(dma_pool_alloc);
342 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
344 struct dma_page *page;
346 list_for_each_entry(page, &pool->page_list, page_list) {
347 if (dma < page->dma)
348 continue;
349 if (dma < (page->dma + pool->allocation))
350 return page;
352 return NULL;
356 * dma_pool_free - put block back into dma pool
357 * @pool: the dma pool holding the block
358 * @vaddr: virtual address of block
359 * @dma: dma address of block
361 * Caller promises neither device nor driver will again touch this block
362 * unless it is first re-allocated.
364 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
366 struct dma_page *page;
367 unsigned long flags;
368 unsigned int offset;
370 spin_lock_irqsave(&pool->lock, flags);
371 page = pool_find_page(pool, dma);
372 if (!page) {
373 spin_unlock_irqrestore(&pool->lock, flags);
374 if (pool->dev)
375 dev_err(pool->dev,
376 "dma_pool_free %s, %p/%lx (bad dma)\n",
377 pool->name, vaddr, (unsigned long)dma);
378 else
379 printk(KERN_ERR "dma_pool_free %s, %p/%lx (bad dma)\n",
380 pool->name, vaddr, (unsigned long)dma);
381 return;
384 offset = vaddr - page->vaddr;
385 #ifdef DMAPOOL_DEBUG
386 if ((dma - page->dma) != offset) {
387 spin_unlock_irqrestore(&pool->lock, flags);
388 if (pool->dev)
389 dev_err(pool->dev,
390 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
391 pool->name, vaddr, (unsigned long long)dma);
392 else
393 printk(KERN_ERR
394 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
395 pool->name, vaddr, (unsigned long long)dma);
396 return;
399 unsigned int chain = page->offset;
400 while (chain < pool->allocation) {
401 if (chain != offset) {
402 chain = *(int *)(page->vaddr + chain);
403 continue;
405 spin_unlock_irqrestore(&pool->lock, flags);
406 if (pool->dev)
407 dev_err(pool->dev, "dma_pool_free %s, dma %Lx "
408 "already free\n", pool->name,
409 (unsigned long long)dma);
410 else
411 printk(KERN_ERR "dma_pool_free %s, dma %Lx "
412 "already free\n", pool->name,
413 (unsigned long long)dma);
414 return;
417 memset(vaddr, POOL_POISON_FREED, pool->size);
418 #endif
420 page->in_use--;
421 *(int *)vaddr = page->offset;
422 page->offset = offset;
424 * Resist a temptation to do
425 * if (!is_page_busy(page)) pool_free_page(pool, page);
426 * Better have a few empty pages hang around.
428 spin_unlock_irqrestore(&pool->lock, flags);
430 EXPORT_SYMBOL(dma_pool_free);
433 * Managed DMA pool
435 static void dmam_pool_release(struct device *dev, void *res)
437 struct dma_pool *pool = *(struct dma_pool **)res;
439 dma_pool_destroy(pool);
442 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
444 return *(struct dma_pool **)res == match_data;
448 * dmam_pool_create - Managed dma_pool_create()
449 * @name: name of pool, for diagnostics
450 * @dev: device that will be doing the DMA
451 * @size: size of the blocks in this pool.
452 * @align: alignment requirement for blocks; must be a power of two
453 * @allocation: returned blocks won't cross this boundary (or zero)
455 * Managed dma_pool_create(). DMA pool created with this function is
456 * automatically destroyed on driver detach.
458 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
459 size_t size, size_t align, size_t allocation)
461 struct dma_pool **ptr, *pool;
463 ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
464 if (!ptr)
465 return NULL;
467 pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
468 if (pool)
469 devres_add(dev, ptr);
470 else
471 devres_free(ptr);
473 return pool;
475 EXPORT_SYMBOL(dmam_pool_create);
478 * dmam_pool_destroy - Managed dma_pool_destroy()
479 * @pool: dma pool that will be destroyed
481 * Managed dma_pool_destroy().
483 void dmam_pool_destroy(struct dma_pool *pool)
485 struct device *dev = pool->dev;
487 WARN_ON(devres_destroy(dev, dmam_pool_release, dmam_pool_match, pool));
488 dma_pool_destroy(pool);
490 EXPORT_SYMBOL(dmam_pool_destroy);