On Tue, Nov 06, 2007 at 02:33:53AM -0800, akpm@linux-foundation.org wrote:
[mmotm.git] / mm / dmapool.c
blob3df063706f53e6d579ec0acabace3503a471ba29
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/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>
40 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_SLUB_DEBUG_ON)
41 #define DMAPOOL_DEBUG 1
42 #endif
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;
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 #define POOL_TIMEOUT_JIFFIES ((100 /* msec */ * HZ) / 1000)
66 static DEFINE_MUTEX(pools_lock);
68 static ssize_t
69 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
71 unsigned temp;
72 unsigned size;
73 char *next;
74 struct dma_page *page;
75 struct dma_pool *pool;
77 next = buf;
78 size = PAGE_SIZE;
80 temp = scnprintf(next, size, "poolinfo - 0.1\n");
81 size -= temp;
82 next += temp;
84 mutex_lock(&pools_lock);
85 list_for_each_entry(pool, &dev->dma_pools, pools) {
86 unsigned pages = 0;
87 unsigned blocks = 0;
89 spin_lock_irq(&pool->lock);
90 list_for_each_entry(page, &pool->page_list, page_list) {
91 pages++;
92 blocks += page->in_use;
94 spin_unlock_irq(&pool->lock);
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;
104 mutex_unlock(&pools_lock);
106 return PAGE_SIZE - size;
109 static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL);
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()
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.
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.
132 struct dma_pool *dma_pool_create(const char *name, struct device *dev,
133 size_t size, size_t align, size_t boundary)
135 struct dma_pool *retval;
136 size_t allocation;
138 if (align == 0) {
139 align = 1;
140 } else if (align & (align - 1)) {
141 return NULL;
144 if (size == 0) {
145 return NULL;
146 } else if (size < 4) {
147 size = 4;
150 if ((size % align) != 0)
151 size = ALIGN(size, align);
153 allocation = max_t(size_t, size, PAGE_SIZE);
155 if (!boundary) {
156 boundary = allocation;
157 } else if ((boundary < size) || (boundary & (boundary - 1))) {
158 return NULL;
161 retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
162 if (!retval)
163 return retval;
165 strlcpy(retval->name, name, sizeof(retval->name));
167 retval->dev = dev;
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);
176 if (dev) {
177 int ret;
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;
191 mutex_unlock(&pools_lock);
192 } else
193 INIT_LIST_HEAD(&retval->pools);
195 return retval;
197 EXPORT_SYMBOL(dma_pool_create);
199 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
201 unsigned int offset = 0;
202 unsigned int next_boundary = pool->boundary;
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;
210 *(int *)(page->vaddr + offset) = next;
211 offset = next;
212 } while (offset < pool->allocation);
215 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
217 struct dma_page *page;
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;
236 return page;
239 static inline int is_page_busy(struct dma_page *page)
241 return page->in_use != 0;
244 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
246 dma_addr_t dma = page->dma;
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);
257 * dma_pool_destroy - destroys a pool of dma memory blocks.
258 * @pool: dma pool that will be destroyed
259 * Context: !in_interrupt()
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.
264 void dma_pool_destroy(struct dma_pool *pool)
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);
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);
292 kfree(pool);
294 EXPORT_SYMBOL(dma_pool_destroy);
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
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.
306 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
307 dma_addr_t *handle)
309 unsigned long flags;
310 struct dma_page *page;
311 size_t offset;
312 void *retval;
314 spin_lock_irqsave(&pool->lock, flags);
315 restart:
316 list_for_each_entry(page, &pool->page_list, page_list) {
317 if (page->offset < pool->allocation)
318 goto ready;
320 page = pool_alloc_page(pool, GFP_ATOMIC);
321 if (!page) {
322 if (mem_flags & __GFP_WAIT) {
323 DECLARE_WAITQUEUE(wait, current);
325 __set_current_state(TASK_INTERRUPTIBLE);
326 __add_wait_queue(&pool->waitq, &wait);
327 spin_unlock_irqrestore(&pool->lock, flags);
329 schedule_timeout(POOL_TIMEOUT_JIFFIES);
331 spin_lock_irqsave(&pool->lock, flags);
332 __remove_wait_queue(&pool->waitq, &wait);
333 goto restart;
335 retval = NULL;
336 goto done;
339 ready:
340 page->in_use++;
341 offset = page->offset;
342 page->offset = *(int *)(page->vaddr + offset);
343 retval = offset + page->vaddr;
344 *handle = offset + page->dma;
345 #ifdef DMAPOOL_DEBUG
346 memset(retval, POOL_POISON_ALLOCATED, pool->size);
347 #endif
348 done:
349 spin_unlock_irqrestore(&pool->lock, flags);
350 return retval;
352 EXPORT_SYMBOL(dma_pool_alloc);
354 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
356 unsigned long flags;
357 struct dma_page *page;
359 spin_lock_irqsave(&pool->lock, flags);
360 list_for_each_entry(page, &pool->page_list, page_list) {
361 if (dma < page->dma)
362 continue;
363 if (dma < (page->dma + pool->allocation))
364 goto done;
366 page = NULL;
367 done:
368 spin_unlock_irqrestore(&pool->lock, flags);
369 return page;
373 * dma_pool_free - put block back into dma pool
374 * @pool: the dma pool holding the block
375 * @vaddr: virtual address of block
376 * @dma: dma address of block
378 * Caller promises neither device nor driver will again touch this block
379 * unless it is first re-allocated.
381 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
383 struct dma_page *page;
384 unsigned long flags;
385 unsigned int offset;
387 page = pool_find_page(pool, dma);
388 if (!page) {
389 if (pool->dev)
390 dev_err(pool->dev,
391 "dma_pool_free %s, %p/%lx (bad dma)\n",
392 pool->name, vaddr, (unsigned long)dma);
393 else
394 printk(KERN_ERR "dma_pool_free %s, %p/%lx (bad dma)\n",
395 pool->name, vaddr, (unsigned long)dma);
396 return;
399 offset = vaddr - page->vaddr;
400 #ifdef DMAPOOL_DEBUG
401 if ((dma - page->dma) != offset) {
402 if (pool->dev)
403 dev_err(pool->dev,
404 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
405 pool->name, vaddr, (unsigned long long)dma);
406 else
407 printk(KERN_ERR
408 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
409 pool->name, vaddr, (unsigned long long)dma);
410 return;
413 unsigned int chain = page->offset;
414 while (chain < pool->allocation) {
415 if (chain != offset) {
416 chain = *(int *)(page->vaddr + chain);
417 continue;
419 if (pool->dev)
420 dev_err(pool->dev, "dma_pool_free %s, dma %Lx "
421 "already free\n", pool->name,
422 (unsigned long long)dma);
423 else
424 printk(KERN_ERR "dma_pool_free %s, dma %Lx "
425 "already free\n", pool->name,
426 (unsigned long long)dma);
427 return;
430 memset(vaddr, POOL_POISON_FREED, pool->size);
431 #endif
433 spin_lock_irqsave(&pool->lock, flags);
434 page->in_use--;
435 *(int *)vaddr = page->offset;
436 page->offset = offset;
437 if (waitqueue_active(&pool->waitq))
438 wake_up_locked(&pool->waitq);
440 * Resist a temptation to do
441 * if (!is_page_busy(page)) pool_free_page(pool, page);
442 * Better have a few empty pages hang around.
444 spin_unlock_irqrestore(&pool->lock, flags);
446 EXPORT_SYMBOL(dma_pool_free);
449 * Managed DMA pool
451 static void dmam_pool_release(struct device *dev, void *res)
453 struct dma_pool *pool = *(struct dma_pool **)res;
455 dma_pool_destroy(pool);
458 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
460 return *(struct dma_pool **)res == match_data;
464 * dmam_pool_create - Managed dma_pool_create()
465 * @name: name of pool, for diagnostics
466 * @dev: device that will be doing the DMA
467 * @size: size of the blocks in this pool.
468 * @align: alignment requirement for blocks; must be a power of two
469 * @allocation: returned blocks won't cross this boundary (or zero)
471 * Managed dma_pool_create(). DMA pool created with this function is
472 * automatically destroyed on driver detach.
474 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
475 size_t size, size_t align, size_t allocation)
477 struct dma_pool **ptr, *pool;
479 ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
480 if (!ptr)
481 return NULL;
483 pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
484 if (pool)
485 devres_add(dev, ptr);
486 else
487 devres_free(ptr);
489 return pool;
491 EXPORT_SYMBOL(dmam_pool_create);
494 * dmam_pool_destroy - Managed dma_pool_destroy()
495 * @pool: dma pool that will be destroyed
497 * Managed dma_pool_destroy().
499 void dmam_pool_destroy(struct dma_pool *pool)
501 struct device *dev = pool->dev;
503 dma_pool_destroy(pool);
504 WARN_ON(devres_destroy(dev, dmam_pool_release, dmam_pool_match, pool));
506 EXPORT_SYMBOL(dmam_pool_destroy);