ipmr/ip6mr: advertise mfc stats via rtnetlink
[linux-2.6.git] / mm / dmapool.c
blobc5ab33bca0a8183bffc1b00e92583b5514d45c96
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 wait_queue_head_t waitq;
54 struct list_head pools;
57 struct dma_page { /* cacheable header for 'allocation' bytes */
58 struct list_head page_list;
59 void *vaddr;
60 dma_addr_t dma;
61 unsigned int in_use;
62 unsigned int offset;
65 #define POOL_TIMEOUT_JIFFIES ((100 /* msec */ * HZ) / 1000)
67 static DEFINE_MUTEX(pools_lock);
69 static ssize_t
70 show_pools(struct device *dev, struct device_attribute *attr, char *buf)
72 unsigned temp;
73 unsigned size;
74 char *next;
75 struct dma_page *page;
76 struct dma_pool *pool;
78 next = buf;
79 size = PAGE_SIZE;
81 temp = scnprintf(next, size, "poolinfo - 0.1\n");
82 size -= temp;
83 next += temp;
85 mutex_lock(&pools_lock);
86 list_for_each_entry(pool, &dev->dma_pools, pools) {
87 unsigned pages = 0;
88 unsigned blocks = 0;
90 spin_lock_irq(&pool->lock);
91 list_for_each_entry(page, &pool->page_list, page_list) {
92 pages++;
93 blocks += page->in_use;
95 spin_unlock_irq(&pool->lock);
97 /* per-pool info, no real statistics yet */
98 temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n",
99 pool->name, blocks,
100 pages * (pool->allocation / pool->size),
101 pool->size, pages);
102 size -= temp;
103 next += temp;
105 mutex_unlock(&pools_lock);
107 return PAGE_SIZE - size;
110 static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL);
113 * dma_pool_create - Creates a pool of consistent memory blocks, for dma.
114 * @name: name of pool, for diagnostics
115 * @dev: device that will be doing the DMA
116 * @size: size of the blocks in this pool.
117 * @align: alignment requirement for blocks; must be a power of two
118 * @boundary: returned blocks won't cross this power of two boundary
119 * Context: !in_interrupt()
121 * Returns a dma allocation pool with the requested characteristics, or
122 * null if one can't be created. Given one of these pools, dma_pool_alloc()
123 * may be used to allocate memory. Such memory will all have "consistent"
124 * DMA mappings, accessible by the device and its driver without using
125 * cache flushing primitives. The actual size of blocks allocated may be
126 * larger than requested because of alignment.
128 * If @boundary is nonzero, objects returned from dma_pool_alloc() won't
129 * cross that size boundary. This is useful for devices which have
130 * addressing restrictions on individual DMA transfers, such as not crossing
131 * boundaries of 4KBytes.
133 struct dma_pool *dma_pool_create(const char *name, struct device *dev,
134 size_t size, size_t align, size_t boundary)
136 struct dma_pool *retval;
137 size_t allocation;
139 if (align == 0) {
140 align = 1;
141 } else if (align & (align - 1)) {
142 return NULL;
145 if (size == 0) {
146 return NULL;
147 } else if (size < 4) {
148 size = 4;
151 if ((size % align) != 0)
152 size = ALIGN(size, align);
154 allocation = max_t(size_t, size, PAGE_SIZE);
156 if (!boundary) {
157 boundary = allocation;
158 } else if ((boundary < size) || (boundary & (boundary - 1))) {
159 return NULL;
162 retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev));
163 if (!retval)
164 return retval;
166 strlcpy(retval->name, name, sizeof(retval->name));
168 retval->dev = dev;
170 INIT_LIST_HEAD(&retval->page_list);
171 spin_lock_init(&retval->lock);
172 retval->size = size;
173 retval->boundary = boundary;
174 retval->allocation = allocation;
175 init_waitqueue_head(&retval->waitq);
177 if (dev) {
178 int ret;
180 mutex_lock(&pools_lock);
181 if (list_empty(&dev->dma_pools))
182 ret = device_create_file(dev, &dev_attr_pools);
183 else
184 ret = 0;
185 /* note: not currently insisting "name" be unique */
186 if (!ret)
187 list_add(&retval->pools, &dev->dma_pools);
188 else {
189 kfree(retval);
190 retval = NULL;
192 mutex_unlock(&pools_lock);
193 } else
194 INIT_LIST_HEAD(&retval->pools);
196 return retval;
198 EXPORT_SYMBOL(dma_pool_create);
200 static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
202 unsigned int offset = 0;
203 unsigned int next_boundary = pool->boundary;
205 do {
206 unsigned int next = offset + pool->size;
207 if (unlikely((next + pool->size) >= next_boundary)) {
208 next = next_boundary;
209 next_boundary += pool->boundary;
211 *(int *)(page->vaddr + offset) = next;
212 offset = next;
213 } while (offset < pool->allocation);
216 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
218 struct dma_page *page;
220 page = kmalloc(sizeof(*page), mem_flags);
221 if (!page)
222 return NULL;
223 page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
224 &page->dma, mem_flags);
225 if (page->vaddr) {
226 #ifdef DMAPOOL_DEBUG
227 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
228 #endif
229 pool_initialise_page(pool, page);
230 list_add(&page->page_list, &pool->page_list);
231 page->in_use = 0;
232 page->offset = 0;
233 } else {
234 kfree(page);
235 page = NULL;
237 return page;
240 static inline int is_page_busy(struct dma_page *page)
242 return page->in_use != 0;
245 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
247 dma_addr_t dma = page->dma;
249 #ifdef DMAPOOL_DEBUG
250 memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
251 #endif
252 dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma);
253 list_del(&page->page_list);
254 kfree(page);
258 * dma_pool_destroy - destroys a pool of dma memory blocks.
259 * @pool: dma pool that will be destroyed
260 * Context: !in_interrupt()
262 * Caller guarantees that no more memory from the pool is in use,
263 * and that nothing will try to use the pool after this call.
265 void dma_pool_destroy(struct dma_pool *pool)
267 mutex_lock(&pools_lock);
268 list_del(&pool->pools);
269 if (pool->dev && list_empty(&pool->dev->dma_pools))
270 device_remove_file(pool->dev, &dev_attr_pools);
271 mutex_unlock(&pools_lock);
273 while (!list_empty(&pool->page_list)) {
274 struct dma_page *page;
275 page = list_entry(pool->page_list.next,
276 struct dma_page, page_list);
277 if (is_page_busy(page)) {
278 if (pool->dev)
279 dev_err(pool->dev,
280 "dma_pool_destroy %s, %p busy\n",
281 pool->name, page->vaddr);
282 else
283 printk(KERN_ERR
284 "dma_pool_destroy %s, %p busy\n",
285 pool->name, page->vaddr);
286 /* leak the still-in-use consistent memory */
287 list_del(&page->page_list);
288 kfree(page);
289 } else
290 pool_free_page(pool, page);
293 kfree(pool);
295 EXPORT_SYMBOL(dma_pool_destroy);
298 * dma_pool_alloc - get a block of consistent memory
299 * @pool: dma pool that will produce the block
300 * @mem_flags: GFP_* bitmask
301 * @handle: pointer to dma address of block
303 * This returns the kernel virtual address of a currently unused block,
304 * and reports its dma address through the handle.
305 * If such a memory block can't be allocated, %NULL is returned.
307 void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags,
308 dma_addr_t *handle)
310 unsigned long flags;
311 struct dma_page *page;
312 size_t offset;
313 void *retval;
315 might_sleep_if(mem_flags & __GFP_WAIT);
317 spin_lock_irqsave(&pool->lock, flags);
318 restart:
319 list_for_each_entry(page, &pool->page_list, page_list) {
320 if (page->offset < pool->allocation)
321 goto ready;
323 page = pool_alloc_page(pool, GFP_ATOMIC);
324 if (!page) {
325 if (mem_flags & __GFP_WAIT) {
326 DECLARE_WAITQUEUE(wait, current);
328 __set_current_state(TASK_UNINTERRUPTIBLE);
329 __add_wait_queue(&pool->waitq, &wait);
330 spin_unlock_irqrestore(&pool->lock, flags);
332 schedule_timeout(POOL_TIMEOUT_JIFFIES);
334 spin_lock_irqsave(&pool->lock, flags);
335 __remove_wait_queue(&pool->waitq, &wait);
336 goto restart;
338 retval = NULL;
339 goto done;
342 ready:
343 page->in_use++;
344 offset = page->offset;
345 page->offset = *(int *)(page->vaddr + offset);
346 retval = offset + page->vaddr;
347 *handle = offset + page->dma;
348 #ifdef DMAPOOL_DEBUG
349 memset(retval, POOL_POISON_ALLOCATED, pool->size);
350 #endif
351 done:
352 spin_unlock_irqrestore(&pool->lock, flags);
353 return retval;
355 EXPORT_SYMBOL(dma_pool_alloc);
357 static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma)
359 struct dma_page *page;
361 list_for_each_entry(page, &pool->page_list, page_list) {
362 if (dma < page->dma)
363 continue;
364 if (dma < (page->dma + pool->allocation))
365 return page;
367 return NULL;
371 * dma_pool_free - put block back into dma pool
372 * @pool: the dma pool holding the block
373 * @vaddr: virtual address of block
374 * @dma: dma address of block
376 * Caller promises neither device nor driver will again touch this block
377 * unless it is first re-allocated.
379 void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma)
381 struct dma_page *page;
382 unsigned long flags;
383 unsigned int offset;
385 spin_lock_irqsave(&pool->lock, flags);
386 page = pool_find_page(pool, dma);
387 if (!page) {
388 spin_unlock_irqrestore(&pool->lock, flags);
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 spin_unlock_irqrestore(&pool->lock, flags);
403 if (pool->dev)
404 dev_err(pool->dev,
405 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
406 pool->name, vaddr, (unsigned long long)dma);
407 else
408 printk(KERN_ERR
409 "dma_pool_free %s, %p (bad vaddr)/%Lx\n",
410 pool->name, vaddr, (unsigned long long)dma);
411 return;
414 unsigned int chain = page->offset;
415 while (chain < pool->allocation) {
416 if (chain != offset) {
417 chain = *(int *)(page->vaddr + chain);
418 continue;
420 spin_unlock_irqrestore(&pool->lock, flags);
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;
432 memset(vaddr, POOL_POISON_FREED, pool->size);
433 #endif
435 page->in_use--;
436 *(int *)vaddr = page->offset;
437 page->offset = offset;
438 if (waitqueue_active(&pool->waitq))
439 wake_up_locked(&pool->waitq);
441 * Resist a temptation to do
442 * if (!is_page_busy(page)) pool_free_page(pool, page);
443 * Better have a few empty pages hang around.
445 spin_unlock_irqrestore(&pool->lock, flags);
447 EXPORT_SYMBOL(dma_pool_free);
450 * Managed DMA pool
452 static void dmam_pool_release(struct device *dev, void *res)
454 struct dma_pool *pool = *(struct dma_pool **)res;
456 dma_pool_destroy(pool);
459 static int dmam_pool_match(struct device *dev, void *res, void *match_data)
461 return *(struct dma_pool **)res == match_data;
465 * dmam_pool_create - Managed dma_pool_create()
466 * @name: name of pool, for diagnostics
467 * @dev: device that will be doing the DMA
468 * @size: size of the blocks in this pool.
469 * @align: alignment requirement for blocks; must be a power of two
470 * @allocation: returned blocks won't cross this boundary (or zero)
472 * Managed dma_pool_create(). DMA pool created with this function is
473 * automatically destroyed on driver detach.
475 struct dma_pool *dmam_pool_create(const char *name, struct device *dev,
476 size_t size, size_t align, size_t allocation)
478 struct dma_pool **ptr, *pool;
480 ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL);
481 if (!ptr)
482 return NULL;
484 pool = *ptr = dma_pool_create(name, dev, size, align, allocation);
485 if (pool)
486 devres_add(dev, ptr);
487 else
488 devres_free(ptr);
490 return pool;
492 EXPORT_SYMBOL(dmam_pool_create);
495 * dmam_pool_destroy - Managed dma_pool_destroy()
496 * @pool: dma pool that will be destroyed
498 * Managed dma_pool_destroy().
500 void dmam_pool_destroy(struct dma_pool *pool)
502 struct device *dev = pool->dev;
504 WARN_ON(devres_destroy(dev, dmam_pool_release, dmam_pool_match, pool));
505 dma_pool_destroy(pool);
507 EXPORT_SYMBOL(dmam_pool_destroy);