diff options
author | Linus Torvalds <torvalds@woody.linux-foundation.org> | 2008-02-05 22:05:48 -0500 |
---|---|---|
committer | Linus Torvalds <torvalds@woody.linux-foundation.org> | 2008-02-05 22:05:48 -0500 |
commit | b297d520b9af536d5580ac505dd316be4cf5560c (patch) | |
tree | a8f292b308c166ebd383b73fa09538775352c874 | |
parent | c773633916c66f8362ca01983d97bd33e35b743f (diff) | |
parent | e34f44b3517fe545f7fd45a8c2f6ee1e5e4432d3 (diff) |
Merge branch 'dmapool' of git://git.kernel.org/pub/scm/linux/kernel/git/willy/misc
* 'dmapool' of git://git.kernel.org/pub/scm/linux/kernel/git/willy/misc:
pool: Improve memory usage for devices which can't cross boundaries
Change dmapool free block management
dmapool: Tidy up includes and add comments
dmapool: Validate parameters to dma_pool_create
Avoid taking waitqueue lock in dmapool
dmapool: Fix style problems
Move dmapool.c to mm/ directory
-rw-r--r-- | drivers/base/Makefile | 2 | ||||
-rw-r--r-- | drivers/base/dmapool.c | 481 | ||||
-rw-r--r-- | mm/Makefile | 1 | ||||
-rw-r--r-- | mm/dmapool.c | 500 |
4 files changed, 502 insertions, 482 deletions
diff --git a/drivers/base/Makefile b/drivers/base/Makefile index 63e09c015ca0..c66637392bbc 100644 --- a/drivers/base/Makefile +++ b/drivers/base/Makefile | |||
@@ -5,7 +5,7 @@ obj-y := core.o sys.o bus.o dd.o \ | |||
5 | cpu.o firmware.o init.o map.o devres.o \ | 5 | cpu.o firmware.o init.o map.o devres.o \ |
6 | attribute_container.o transport_class.o | 6 | attribute_container.o transport_class.o |
7 | obj-y += power/ | 7 | obj-y += power/ |
8 | obj-$(CONFIG_HAS_DMA) += dma-mapping.o dmapool.o | 8 | obj-$(CONFIG_HAS_DMA) += dma-mapping.o |
9 | obj-$(CONFIG_ISA) += isa.o | 9 | obj-$(CONFIG_ISA) += isa.o |
10 | obj-$(CONFIG_FW_LOADER) += firmware_class.o | 10 | obj-$(CONFIG_FW_LOADER) += firmware_class.o |
11 | obj-$(CONFIG_NUMA) += node.o | 11 | obj-$(CONFIG_NUMA) += node.o |
diff --git a/drivers/base/dmapool.c b/drivers/base/dmapool.c deleted file mode 100644 index b5034dc72a05..000000000000 --- a/drivers/base/dmapool.c +++ /dev/null | |||
@@ -1,481 +0,0 @@ | |||
1 | |||
2 | #include <linux/device.h> | ||
3 | #include <linux/mm.h> | ||
4 | #include <asm/io.h> /* Needed for i386 to build */ | ||
5 | #include <linux/dma-mapping.h> | ||
6 | #include <linux/dmapool.h> | ||
7 | #include <linux/slab.h> | ||
8 | #include <linux/module.h> | ||
9 | #include <linux/poison.h> | ||
10 | #include <linux/sched.h> | ||
11 | |||
12 | /* | ||
13 | * Pool allocator ... wraps the dma_alloc_coherent page allocator, so | ||
14 | * small blocks are easily used by drivers for bus mastering controllers. | ||
15 | * This should probably be sharing the guts of the slab allocator. | ||
16 | */ | ||
17 | |||
18 | struct dma_pool { /* the pool */ | ||
19 | struct list_head page_list; | ||
20 | spinlock_t lock; | ||
21 | size_t blocks_per_page; | ||
22 | size_t size; | ||
23 | struct device *dev; | ||
24 | size_t allocation; | ||
25 | char name [32]; | ||
26 | wait_queue_head_t waitq; | ||
27 | struct list_head pools; | ||
28 | }; | ||
29 | |||
30 | struct dma_page { /* cacheable header for 'allocation' bytes */ | ||
31 | struct list_head page_list; | ||
32 | void *vaddr; | ||
33 | dma_addr_t dma; | ||
34 | unsigned in_use; | ||
35 | unsigned long bitmap [0]; | ||
36 | }; | ||
37 | |||
38 | #define POOL_TIMEOUT_JIFFIES ((100 /* msec */ * HZ) / 1000) | ||
39 | |||
40 | static DEFINE_MUTEX (pools_lock); | ||
41 | |||
42 | static ssize_t | ||
43 | show_pools (struct device *dev, struct device_attribute *attr, char *buf) | ||
44 | { | ||
45 | unsigned temp; | ||
46 | unsigned size; | ||
47 | char *next; | ||
48 | struct dma_page *page; | ||
49 | struct dma_pool *pool; | ||
50 | |||
51 | next = buf; | ||
52 | size = PAGE_SIZE; | ||
53 | |||
54 | temp = scnprintf(next, size, "poolinfo - 0.1\n"); | ||
55 | size -= temp; | ||
56 | next += temp; | ||
57 | |||
58 | mutex_lock(&pools_lock); | ||
59 | list_for_each_entry(pool, &dev->dma_pools, pools) { | ||
60 | unsigned pages = 0; | ||
61 | unsigned blocks = 0; | ||
62 | |||
63 | list_for_each_entry(page, &pool->page_list, page_list) { | ||
64 | pages++; | ||
65 | blocks += page->in_use; | ||
66 | } | ||
67 | |||
68 | /* per-pool info, no real statistics yet */ | ||
69 | temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n", | ||
70 | pool->name, | ||
71 | blocks, pages * pool->blocks_per_page, | ||
72 | pool->size, pages); | ||
73 | size -= temp; | ||
74 | next += temp; | ||
75 | } | ||
76 | mutex_unlock(&pools_lock); | ||
77 | |||
78 | return PAGE_SIZE - size; | ||
79 | } | ||
80 | static DEVICE_ATTR (pools, S_IRUGO, show_pools, NULL); | ||
81 | |||
82 | /** | ||
83 | * dma_pool_create - Creates a pool of consistent memory blocks, for dma. | ||
84 | * @name: name of pool, for diagnostics | ||
85 | * @dev: device that will be doing the DMA | ||
86 | * @size: size of the blocks in this pool. | ||
87 | * @align: alignment requirement for blocks; must be a power of two | ||
88 | * @allocation: returned blocks won't cross this boundary (or zero) | ||
89 | * Context: !in_interrupt() | ||
90 | * | ||
91 | * Returns a dma allocation pool with the requested characteristics, or | ||
92 | * null if one can't be created. Given one of these pools, dma_pool_alloc() | ||
93 | * may be used to allocate memory. Such memory will all have "consistent" | ||
94 | * DMA mappings, accessible by the device and its driver without using | ||
95 | * cache flushing primitives. The actual size of blocks allocated may be | ||
96 | * larger than requested because of alignment. | ||
97 | * | ||
98 | * If allocation is nonzero, objects returned from dma_pool_alloc() won't | ||
99 | * cross that size boundary. This is useful for devices which have | ||
100 | * addressing restrictions on individual DMA transfers, such as not crossing | ||
101 | * boundaries of 4KBytes. | ||
102 | */ | ||
103 | struct dma_pool * | ||
104 | dma_pool_create (const char *name, struct device *dev, | ||
105 | size_t size, size_t align, size_t allocation) | ||
106 | { | ||
107 | struct dma_pool *retval; | ||
108 | |||
109 | if (align == 0) | ||
110 | align = 1; | ||
111 | if (size == 0) | ||
112 | return NULL; | ||
113 | else if (size < align) | ||
114 | size = align; | ||
115 | else if ((size % align) != 0) { | ||
116 | size += align + 1; | ||
117 | size &= ~(align - 1); | ||
118 | } | ||
119 | |||
120 | if (allocation == 0) { | ||
121 | if (PAGE_SIZE < size) | ||
122 | allocation = size; | ||
123 | else | ||
124 | allocation = PAGE_SIZE; | ||
125 | // FIXME: round up for less fragmentation | ||
126 | } else if (allocation < size) | ||
127 | return NULL; | ||
128 | |||
129 | if (!(retval = kmalloc_node (sizeof *retval, GFP_KERNEL, dev_to_node(dev)))) | ||
130 | return retval; | ||
131 | |||
132 | strlcpy (retval->name, name, sizeof retval->name); | ||
133 | |||
134 | retval->dev = dev; | ||
135 | |||
136 | INIT_LIST_HEAD (&retval->page_list); | ||
137 | spin_lock_init (&retval->lock); | ||
138 | retval->size = size; | ||
139 | retval->allocation = allocation; | ||
140 | retval->blocks_per_page = allocation / size; | ||
141 | init_waitqueue_head (&retval->waitq); | ||
142 | |||
143 | if (dev) { | ||
144 | int ret; | ||
145 | |||
146 | mutex_lock(&pools_lock); | ||
147 | if (list_empty (&dev->dma_pools)) | ||
148 | ret = device_create_file (dev, &dev_attr_pools); | ||
149 | else | ||
150 | ret = 0; | ||
151 | /* note: not currently insisting "name" be unique */ | ||
152 | if (!ret) | ||
153 | list_add (&retval->pools, &dev->dma_pools); | ||
154 | else { | ||
155 | kfree(retval); | ||
156 | retval = NULL; | ||
157 | } | ||
158 | mutex_unlock(&pools_lock); | ||
159 | } else | ||
160 | INIT_LIST_HEAD (&retval->pools); | ||
161 | |||
162 | return retval; | ||
163 | } | ||
164 | |||
165 | |||
166 | static struct dma_page * | ||
167 | pool_alloc_page (struct dma_pool *pool, gfp_t mem_flags) | ||
168 | { | ||
169 | struct dma_page *page; | ||
170 | int mapsize; | ||
171 | |||
172 | mapsize = pool->blocks_per_page; | ||
173 | mapsize = (mapsize + BITS_PER_LONG - 1) / BITS_PER_LONG; | ||
174 | mapsize *= sizeof (long); | ||
175 | |||
176 | page = kmalloc(mapsize + sizeof *page, mem_flags); | ||
177 | if (!page) | ||
178 | return NULL; | ||
179 | page->vaddr = dma_alloc_coherent (pool->dev, | ||
180 | pool->allocation, | ||
181 | &page->dma, | ||
182 | mem_flags); | ||
183 | if (page->vaddr) { | ||
184 | memset (page->bitmap, 0xff, mapsize); // bit set == free | ||
185 | #ifdef CONFIG_DEBUG_SLAB | ||
186 | memset (page->vaddr, POOL_POISON_FREED, pool->allocation); | ||
187 | #endif | ||
188 | list_add (&page->page_list, &pool->page_list); | ||
189 | page->in_use = 0; | ||
190 | } else { | ||
191 | kfree (page); | ||
192 | page = NULL; | ||
193 | } | ||
194 | return page; | ||
195 | } | ||
196 | |||
197 | |||
198 | static inline int | ||
199 | is_page_busy (int blocks, unsigned long *bitmap) | ||
200 | { | ||
201 | while (blocks > 0) { | ||
202 | if (*bitmap++ != ~0UL) | ||
203 | return 1; | ||
204 | blocks -= BITS_PER_LONG; | ||
205 | } | ||
206 | return 0; | ||
207 | } | ||
208 | |||
209 | static void | ||
210 | pool_free_page (struct dma_pool *pool, struct dma_page *page) | ||
211 | { | ||
212 | dma_addr_t dma = page->dma; | ||
213 | |||
214 | #ifdef CONFIG_DEBUG_SLAB | ||
215 | memset (page->vaddr, POOL_POISON_FREED, pool->allocation); | ||
216 | #endif | ||
217 | dma_free_coherent (pool->dev, pool->allocation, page->vaddr, dma); | ||
218 | list_del (&page->page_list); | ||
219 | kfree (page); | ||
220 | } | ||
221 | |||
222 | |||
223 | /** | ||
224 | * dma_pool_destroy - destroys a pool of dma memory blocks. | ||
225 | * @pool: dma pool that will be destroyed | ||
226 | * Context: !in_interrupt() | ||
227 | * | ||
228 | * Caller guarantees that no more memory from the pool is in use, | ||
229 | * and that nothing will try to use the pool after this call. | ||
230 | */ | ||
231 | void | ||
232 | dma_pool_destroy (struct dma_pool *pool) | ||
233 | { | ||
234 | mutex_lock(&pools_lock); | ||
235 | list_del (&pool->pools); | ||
236 | if (pool->dev && list_empty (&pool->dev->dma_pools)) | ||
237 | device_remove_file (pool->dev, &dev_attr_pools); | ||
238 | mutex_unlock(&pools_lock); | ||
239 | |||
240 | while (!list_empty (&pool->page_list)) { | ||
241 | struct dma_page *page; | ||
242 | page = list_entry (pool->page_list.next, | ||
243 | struct dma_page, page_list); | ||
244 | if (is_page_busy (pool->blocks_per_page, page->bitmap)) { | ||
245 | if (pool->dev) | ||
246 | dev_err(pool->dev, "dma_pool_destroy %s, %p busy\n", | ||
247 | pool->name, page->vaddr); | ||
248 | else | ||
249 | printk (KERN_ERR "dma_pool_destroy %s, %p busy\n", | ||
250 | pool->name, page->vaddr); | ||
251 | /* leak the still-in-use consistent memory */ | ||
252 | list_del (&page->page_list); | ||
253 | kfree (page); | ||
254 | } else | ||
255 | pool_free_page (pool, page); | ||
256 | } | ||
257 | |||
258 | kfree (pool); | ||
259 | } | ||
260 | |||
261 | |||
262 | /** | ||
263 | * dma_pool_alloc - get a block of consistent memory | ||
264 | * @pool: dma pool that will produce the block | ||
265 | * @mem_flags: GFP_* bitmask | ||
266 | * @handle: pointer to dma address of block | ||
267 | * | ||
268 | * This returns the kernel virtual address of a currently unused block, | ||
269 | * and reports its dma address through the handle. | ||
270 | * If such a memory block can't be allocated, null is returned. | ||
271 | */ | ||
272 | void * | ||
273 | dma_pool_alloc (struct dma_pool *pool, gfp_t mem_flags, dma_addr_t *handle) | ||
274 | { | ||
275 | unsigned long flags; | ||
276 | struct dma_page *page; | ||
277 | int map, block; | ||
278 | size_t offset; | ||
279 | void *retval; | ||
280 | |||
281 | restart: | ||
282 | spin_lock_irqsave (&pool->lock, flags); | ||
283 | list_for_each_entry(page, &pool->page_list, page_list) { | ||
284 | int i; | ||
285 | /* only cachable accesses here ... */ | ||
286 | for (map = 0, i = 0; | ||
287 | i < pool->blocks_per_page; | ||
288 | i += BITS_PER_LONG, map++) { | ||
289 | if (page->bitmap [map] == 0) | ||
290 | continue; | ||
291 | block = ffz (~ page->bitmap [map]); | ||
292 | if ((i + block) < pool->blocks_per_page) { | ||
293 | clear_bit (block, &page->bitmap [map]); | ||
294 | offset = (BITS_PER_LONG * map) + block; | ||
295 | offset *= pool->size; | ||
296 | goto ready; | ||
297 | } | ||
298 | } | ||
299 | } | ||
300 | if (!(page = pool_alloc_page (pool, GFP_ATOMIC))) { | ||
301 | if (mem_flags & __GFP_WAIT) { | ||
302 | DECLARE_WAITQUEUE (wait, current); | ||
303 | |||
304 | __set_current_state(TASK_INTERRUPTIBLE); | ||
305 | add_wait_queue (&pool->waitq, &wait); | ||
306 | spin_unlock_irqrestore (&pool->lock, flags); | ||
307 | |||
308 | schedule_timeout (POOL_TIMEOUT_JIFFIES); | ||
309 | |||
310 | remove_wait_queue (&pool->waitq, &wait); | ||
311 | goto restart; | ||
312 | } | ||
313 | retval = NULL; | ||
314 | goto done; | ||
315 | } | ||
316 | |||
317 | clear_bit (0, &page->bitmap [0]); | ||
318 | offset = 0; | ||
319 | ready: | ||
320 | page->in_use++; | ||
321 | retval = offset + page->vaddr; | ||
322 | *handle = offset + page->dma; | ||
323 | #ifdef CONFIG_DEBUG_SLAB | ||
324 | memset (retval, POOL_POISON_ALLOCATED, pool->size); | ||
325 | #endif | ||
326 | done: | ||
327 | spin_unlock_irqrestore (&pool->lock, flags); | ||
328 | return retval; | ||
329 | } | ||
330 | |||
331 | |||
332 | static struct dma_page * | ||
333 | pool_find_page (struct dma_pool *pool, dma_addr_t dma) | ||
334 | { | ||
335 | unsigned long flags; | ||
336 | struct dma_page *page; | ||
337 | |||
338 | spin_lock_irqsave (&pool->lock, flags); | ||
339 | list_for_each_entry(page, &pool->page_list, page_list) { | ||
340 | if (dma < page->dma) | ||
341 | continue; | ||
342 | if (dma < (page->dma + pool->allocation)) | ||
343 | goto done; | ||
344 | } | ||
345 | page = NULL; | ||
346 | done: | ||
347 | spin_unlock_irqrestore (&pool->lock, flags); | ||
348 | return page; | ||
349 | } | ||
350 | |||
351 | |||
352 | /** | ||
353 | * dma_pool_free - put block back into dma pool | ||
354 | * @pool: the dma pool holding the block | ||
355 | * @vaddr: virtual address of block | ||
356 | * @dma: dma address of block | ||
357 | * | ||
358 | * Caller promises neither device nor driver will again touch this block | ||
359 | * unless it is first re-allocated. | ||
360 | */ | ||
361 | void | ||
362 | dma_pool_free (struct dma_pool *pool, void *vaddr, dma_addr_t dma) | ||
363 | { | ||
364 | struct dma_page *page; | ||
365 | unsigned long flags; | ||
366 | int map, block; | ||
367 | |||
368 | if ((page = pool_find_page(pool, dma)) == NULL) { | ||
369 | if (pool->dev) | ||
370 | dev_err(pool->dev, "dma_pool_free %s, %p/%lx (bad dma)\n", | ||
371 | pool->name, vaddr, (unsigned long) dma); | ||
372 | else | ||
373 | printk (KERN_ERR "dma_pool_free %s, %p/%lx (bad dma)\n", | ||
374 | pool->name, vaddr, (unsigned long) dma); | ||
375 | return; | ||
376 | } | ||
377 | |||
378 | block = dma - page->dma; | ||
379 | block /= pool->size; | ||
380 | map = block / BITS_PER_LONG; | ||
381 | block %= BITS_PER_LONG; | ||
382 | |||
383 | #ifdef CONFIG_DEBUG_SLAB | ||
384 | if (((dma - page->dma) + (void *)page->vaddr) != vaddr) { | ||
385 | if (pool->dev) | ||
386 | dev_err(pool->dev, "dma_pool_free %s, %p (bad vaddr)/%Lx\n", | ||
387 | pool->name, vaddr, (unsigned long long) dma); | ||
388 | else | ||
389 | printk (KERN_ERR "dma_pool_free %s, %p (bad vaddr)/%Lx\n", | ||
390 | pool->name, vaddr, (unsigned long long) dma); | ||
391 | return; | ||
392 | } | ||
393 | if (page->bitmap [map] & (1UL << block)) { | ||
394 | if (pool->dev) | ||
395 | dev_err(pool->dev, "dma_pool_free %s, dma %Lx already free\n", | ||
396 | pool->name, (unsigned long long)dma); | ||
397 | else | ||
398 | printk (KERN_ERR "dma_pool_free %s, dma %Lx already free\n", | ||
399 | pool->name, (unsigned long long)dma); | ||
400 | return; | ||
401 | } | ||
402 | memset (vaddr, POOL_POISON_FREED, pool->size); | ||
403 | #endif | ||
404 | |||
405 | spin_lock_irqsave (&pool->lock, flags); | ||
406 | page->in_use--; | ||
407 | set_bit (block, &page->bitmap [map]); | ||
408 | if (waitqueue_active (&pool->waitq)) | ||
409 | wake_up (&pool->waitq); | ||
410 | /* | ||
411 | * Resist a temptation to do | ||
412 | * if (!is_page_busy(bpp, page->bitmap)) pool_free_page(pool, page); | ||
413 | * Better have a few empty pages hang around. | ||
414 | */ | ||
415 | spin_unlock_irqrestore (&pool->lock, flags); | ||
416 | } | ||
417 | |||
418 | /* | ||
419 | * Managed DMA pool | ||
420 | */ | ||
421 | static void dmam_pool_release(struct device *dev, void *res) | ||
422 | { | ||
423 | struct dma_pool *pool = *(struct dma_pool **)res; | ||
424 | |||
425 | dma_pool_destroy(pool); | ||
426 | } | ||
427 | |||
428 | static int dmam_pool_match(struct device *dev, void *res, void *match_data) | ||
429 | { | ||
430 | return *(struct dma_pool **)res == match_data; | ||
431 | } | ||
432 | |||
433 | /** | ||
434 | * dmam_pool_create - Managed dma_pool_create() | ||
435 | * @name: name of pool, for diagnostics | ||
436 | * @dev: device that will be doing the DMA | ||
437 | * @size: size of the blocks in this pool. | ||
438 | * @align: alignment requirement for blocks; must be a power of two | ||
439 | * @allocation: returned blocks won't cross this boundary (or zero) | ||
440 | * | ||
441 | * Managed dma_pool_create(). DMA pool created with this function is | ||
442 | * automatically destroyed on driver detach. | ||
443 | */ | ||
444 | struct dma_pool *dmam_pool_create(const char *name, struct device *dev, | ||
445 | size_t size, size_t align, size_t allocation) | ||
446 | { | ||
447 | struct dma_pool **ptr, *pool; | ||
448 | |||
449 | ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL); | ||
450 | if (!ptr) | ||
451 | return NULL; | ||
452 | |||
453 | pool = *ptr = dma_pool_create(name, dev, size, align, allocation); | ||
454 | if (pool) | ||
455 | devres_add(dev, ptr); | ||
456 | else | ||
457 | devres_free(ptr); | ||
458 | |||
459 | return pool; | ||
460 | } | ||
461 | |||
462 | /** | ||
463 | * dmam_pool_destroy - Managed dma_pool_destroy() | ||
464 | * @pool: dma pool that will be destroyed | ||
465 | * | ||
466 | * Managed dma_pool_destroy(). | ||
467 | */ | ||
468 | void dmam_pool_destroy(struct dma_pool *pool) | ||
469 | { | ||
470 | struct device *dev = pool->dev; | ||
471 | |||
472 | dma_pool_destroy(pool); | ||
473 | WARN_ON(devres_destroy(dev, dmam_pool_release, dmam_pool_match, pool)); | ||
474 | } | ||
475 | |||
476 | EXPORT_SYMBOL (dma_pool_create); | ||
477 | EXPORT_SYMBOL (dma_pool_destroy); | ||
478 | EXPORT_SYMBOL (dma_pool_alloc); | ||
479 | EXPORT_SYMBOL (dma_pool_free); | ||
480 | EXPORT_SYMBOL (dmam_pool_create); | ||
481 | EXPORT_SYMBOL (dmam_pool_destroy); | ||
diff --git a/mm/Makefile b/mm/Makefile index 44e2528af70c..4af5dff37277 100644 --- a/mm/Makefile +++ b/mm/Makefile | |||
@@ -16,6 +16,7 @@ obj-y := bootmem.o filemap.o mempool.o oom_kill.o fadvise.o \ | |||
16 | obj-$(CONFIG_PROC_PAGE_MONITOR) += pagewalk.o | 16 | obj-$(CONFIG_PROC_PAGE_MONITOR) += pagewalk.o |
17 | obj-$(CONFIG_BOUNCE) += bounce.o | 17 | obj-$(CONFIG_BOUNCE) += bounce.o |
18 | obj-$(CONFIG_SWAP) += page_io.o swap_state.o swapfile.o thrash.o | 18 | obj-$(CONFIG_SWAP) += page_io.o swap_state.o swapfile.o thrash.o |
19 | obj-$(CONFIG_HAS_DMA) += dmapool.o | ||
19 | obj-$(CONFIG_HUGETLBFS) += hugetlb.o | 20 | obj-$(CONFIG_HUGETLBFS) += hugetlb.o |
20 | obj-$(CONFIG_NUMA) += mempolicy.o | 21 | obj-$(CONFIG_NUMA) += mempolicy.o |
21 | obj-$(CONFIG_SPARSEMEM) += sparse.o | 22 | obj-$(CONFIG_SPARSEMEM) += sparse.o |
diff --git a/mm/dmapool.c b/mm/dmapool.c new file mode 100644 index 000000000000..34aaac451a96 --- /dev/null +++ b/mm/dmapool.c | |||
@@ -0,0 +1,500 @@ | |||
1 | /* | ||
2 | * DMA Pool allocator | ||
3 | * | ||
4 | * Copyright 2001 David Brownell | ||
5 | * Copyright 2007 Intel Corporation | ||
6 | * Author: Matthew Wilcox <willy@linux.intel.com> | ||
7 | * | ||
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. | ||
11 | * | ||
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. | ||
16 | * | ||
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. | ||
23 | */ | ||
24 | |||
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> | ||
39 | |||
40 | struct dma_pool { /* the pool */ | ||
41 | struct list_head page_list; | ||
42 | spinlock_t lock; | ||
43 | size_t size; | ||
44 | struct device *dev; | ||
45 | size_t allocation; | ||
46 | size_t boundary; | ||
47 | char name[32]; | ||
48 | wait_queue_head_t waitq; | ||
49 | struct list_head pools; | ||
50 | }; | ||
51 | |||
52 | struct dma_page { /* cacheable header for 'allocation' bytes */ | ||
53 | struct list_head page_list; | ||
54 | void *vaddr; | ||
55 | dma_addr_t dma; | ||
56 | unsigned int in_use; | ||
57 | unsigned int offset; | ||
58 | }; | ||
59 | |||
60 | #define POOL_TIMEOUT_JIFFIES ((100 /* msec */ * HZ) / 1000) | ||
61 | |||
62 | static DEFINE_MUTEX(pools_lock); | ||
63 | |||
64 | static ssize_t | ||
65 | show_pools(struct device *dev, struct device_attribute *attr, char *buf) | ||
66 | { | ||
67 | unsigned temp; | ||
68 | unsigned size; | ||
69 | char *next; | ||
70 | struct dma_page *page; | ||
71 | struct dma_pool *pool; | ||
72 | |||
73 | next = buf; | ||
74 | size = PAGE_SIZE; | ||
75 | |||
76 | temp = scnprintf(next, size, "poolinfo - 0.1\n"); | ||
77 | size -= temp; | ||
78 | next += temp; | ||
79 | |||
80 | mutex_lock(&pools_lock); | ||
81 | list_for_each_entry(pool, &dev->dma_pools, pools) { | ||
82 | unsigned pages = 0; | ||
83 | unsigned blocks = 0; | ||
84 | |||
85 | list_for_each_entry(page, &pool->page_list, page_list) { | ||
86 | pages++; | ||
87 | blocks += page->in_use; | ||
88 | } | ||
89 | |||
90 | /* per-pool info, no real statistics yet */ | ||
91 | temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n", | ||
92 | pool->name, blocks, | ||
93 | pages * (pool->allocation / pool->size), | ||
94 | pool->size, pages); | ||
95 | size -= temp; | ||
96 | next += temp; | ||
97 | } | ||
98 | mutex_unlock(&pools_lock); | ||
99 | |||
100 | return PAGE_SIZE - size; | ||
101 | } | ||
102 | |||
103 | static DEVICE_ATTR(pools, S_IRUGO, show_pools, NULL); | ||
104 | |||
105 | /** | ||
106 | * dma_pool_create - Creates a pool of consistent memory blocks, for dma. | ||
107 | * @name: name of pool, for diagnostics | ||
108 | * @dev: device that will be doing the DMA | ||
109 | * @size: size of the blocks in this pool. | ||
110 | * @align: alignment requirement for blocks; must be a power of two | ||
111 | * @boundary: returned blocks won't cross this power of two boundary | ||
112 | * Context: !in_interrupt() | ||
113 | * | ||
114 | * Returns a dma allocation pool with the requested characteristics, or | ||
115 | * null if one can't be created. Given one of these pools, dma_pool_alloc() | ||
116 | * may be used to allocate memory. Such memory will all have "consistent" | ||
117 | * DMA mappings, accessible by the device and its driver without using | ||
118 | * cache flushing primitives. The actual size of blocks allocated may be | ||
119 | * larger than requested because of alignment. | ||
120 | * | ||
121 | * If @boundary is nonzero, objects returned from dma_pool_alloc() won't | ||
122 | * cross that size boundary. This is useful for devices which have | ||
123 | * addressing restrictions on individual DMA transfers, such as not crossing | ||
124 | * boundaries of 4KBytes. | ||
125 | */ | ||
126 | struct dma_pool *dma_pool_create(const char *name, struct device *dev, | ||
127 | size_t size, size_t align, size_t boundary) | ||
128 | { | ||
129 | struct dma_pool *retval; | ||
130 | size_t allocation; | ||
131 | |||
132 | if (align == 0) { | ||
133 | align = 1; | ||
134 | } else if (align & (align - 1)) { | ||
135 | return NULL; | ||
136 | } | ||
137 | |||
138 | if (size == 0) { | ||
139 | return NULL; | ||
140 | } else if (size < 4) { | ||
141 | size = 4; | ||
142 | } | ||
143 | |||
144 | if ((size % align) != 0) | ||
145 | size = ALIGN(size, align); | ||
146 | |||
147 | allocation = max_t(size_t, size, PAGE_SIZE); | ||
148 | |||
149 | if (!boundary) { | ||
150 | boundary = allocation; | ||
151 | } else if ((boundary < size) || (boundary & (boundary - 1))) { | ||
152 | return NULL; | ||
153 | } | ||
154 | |||
155 | retval = kmalloc_node(sizeof(*retval), GFP_KERNEL, dev_to_node(dev)); | ||
156 | if (!retval) | ||
157 | return retval; | ||
158 | |||
159 | strlcpy(retval->name, name, sizeof(retval->name)); | ||
160 | |||
161 | retval->dev = dev; | ||
162 | |||
163 | INIT_LIST_HEAD(&retval->page_list); | ||
164 | spin_lock_init(&retval->lock); | ||
165 | retval->size = size; | ||
166 | retval->boundary = boundary; | ||
167 | retval->allocation = allocation; | ||
168 | init_waitqueue_head(&retval->waitq); | ||
169 | |||
170 | if (dev) { | ||
171 | int ret; | ||
172 | |||
173 | mutex_lock(&pools_lock); | ||
174 | if (list_empty(&dev->dma_pools)) | ||
175 | ret = device_create_file(dev, &dev_attr_pools); | ||
176 | else | ||
177 | ret = 0; | ||
178 | /* note: not currently insisting "name" be unique */ | ||
179 | if (!ret) | ||
180 | list_add(&retval->pools, &dev->dma_pools); | ||
181 | else { | ||
182 | kfree(retval); | ||
183 | retval = NULL; | ||
184 | } | ||
185 | mutex_unlock(&pools_lock); | ||
186 | } else | ||
187 | INIT_LIST_HEAD(&retval->pools); | ||
188 | |||
189 | return retval; | ||
190 | } | ||
191 | EXPORT_SYMBOL(dma_pool_create); | ||
192 | |||
193 | static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page) | ||
194 | { | ||
195 | unsigned int offset = 0; | ||
196 | unsigned int next_boundary = pool->boundary; | ||
197 | |||
198 | do { | ||
199 | unsigned int next = offset + pool->size; | ||
200 | if (unlikely((next + pool->size) >= next_boundary)) { | ||
201 | next = next_boundary; | ||
202 | next_boundary += pool->boundary; | ||
203 | } | ||
204 | *(int *)(page->vaddr + offset) = next; | ||
205 | offset = next; | ||
206 | } while (offset < pool->allocation); | ||
207 | } | ||
208 | |||
209 | static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags) | ||
210 | { | ||
211 | struct dma_page *page; | ||
212 | |||
213 | page = kmalloc(sizeof(*page), mem_flags); | ||
214 | if (!page) | ||
215 | return NULL; | ||
216 | page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation, | ||
217 | &page->dma, mem_flags); | ||
218 | if (page->vaddr) { | ||
219 | #ifdef CONFIG_DEBUG_SLAB | ||
220 | memset(page->vaddr, POOL_POISON_FREED, pool->allocation); | ||
221 | #endif | ||
222 | pool_initialise_page(pool, page); | ||
223 | list_add(&page->page_list, &pool->page_list); | ||
224 | page->in_use = 0; | ||
225 | page->offset = 0; | ||
226 | } else { | ||
227 | kfree(page); | ||
228 | page = NULL; | ||
229 | } | ||
230 | return page; | ||
231 | } | ||
232 | |||
233 | static inline int is_page_busy(struct dma_page *page) | ||
234 | { | ||
235 | return page->in_use != 0; | ||
236 | } | ||
237 | |||
238 | static void pool_free_page(struct dma_pool *pool, struct dma_page *page) | ||
239 | { | ||
240 | dma_addr_t dma = page->dma; | ||
241 | |||
242 | #ifdef CONFIG_DEBUG_SLAB | ||
243 | memset(page->vaddr, POOL_POISON_FREED, pool->allocation); | ||
244 | #endif | ||
245 | dma_free_coherent(pool->dev, pool->allocation, page->vaddr, dma); | ||
246 | list_del(&page->page_list); | ||
247 | kfree(page); | ||
248 | } | ||
249 | |||
250 | /** | ||
251 | * dma_pool_destroy - destroys a pool of dma memory blocks. | ||
252 | * @pool: dma pool that will be destroyed | ||
253 | * Context: !in_interrupt() | ||
254 | * | ||
255 | * Caller guarantees that no more memory from the pool is in use, | ||
256 | * and that nothing will try to use the pool after this call. | ||
257 | */ | ||
258 | void dma_pool_destroy(struct dma_pool *pool) | ||
259 | { | ||
260 | mutex_lock(&pools_lock); | ||
261 | list_del(&pool->pools); | ||
262 | if (pool->dev && list_empty(&pool->dev->dma_pools)) | ||
263 | device_remove_file(pool->dev, &dev_attr_pools); | ||
264 | mutex_unlock(&pools_lock); | ||
265 | |||
266 | while (!list_empty(&pool->page_list)) { | ||
267 | struct dma_page *page; | ||
268 | page = list_entry(pool->page_list.next, | ||
269 | struct dma_page, page_list); | ||
270 | if (is_page_busy(page)) { | ||
271 | if (pool->dev) | ||
272 | dev_err(pool->dev, | ||
273 | "dma_pool_destroy %s, %p busy\n", | ||
274 | pool->name, page->vaddr); | ||
275 | else | ||
276 | printk(KERN_ERR | ||
277 | "dma_pool_destroy %s, %p busy\n", | ||
278 | pool->name, page->vaddr); | ||
279 | /* leak the still-in-use consistent memory */ | ||
280 | list_del(&page->page_list); | ||
281 | kfree(page); | ||
282 | } else | ||
283 | pool_free_page(pool, page); | ||
284 | } | ||
285 | |||
286 | kfree(pool); | ||
287 | } | ||
288 | EXPORT_SYMBOL(dma_pool_destroy); | ||
289 | |||
290 | /** | ||
291 | * dma_pool_alloc - get a block of consistent memory | ||
292 | * @pool: dma pool that will produce the block | ||
293 | * @mem_flags: GFP_* bitmask | ||
294 | * @handle: pointer to dma address of block | ||
295 | * | ||
296 | * This returns the kernel virtual address of a currently unused block, | ||
297 | * and reports its dma address through the handle. | ||
298 | * If such a memory block can't be allocated, %NULL is returned. | ||
299 | */ | ||
300 | void *dma_pool_alloc(struct dma_pool *pool, gfp_t mem_flags, | ||
301 | dma_addr_t *handle) | ||
302 | { | ||
303 | unsigned long flags; | ||
304 | struct dma_page *page; | ||
305 | size_t offset; | ||
306 | void *retval; | ||
307 | |||
308 | spin_lock_irqsave(&pool->lock, flags); | ||
309 | restart: | ||
310 | list_for_each_entry(page, &pool->page_list, page_list) { | ||
311 | if (page->offset < pool->allocation) | ||
312 | goto ready; | ||
313 | } | ||
314 | page = pool_alloc_page(pool, GFP_ATOMIC); | ||
315 | if (!page) { | ||
316 | if (mem_flags & __GFP_WAIT) { | ||
317 | DECLARE_WAITQUEUE(wait, current); | ||
318 | |||
319 | __set_current_state(TASK_INTERRUPTIBLE); | ||
320 | __add_wait_queue(&pool->waitq, &wait); | ||
321 | spin_unlock_irqrestore(&pool->lock, flags); | ||
322 | |||
323 | schedule_timeout(POOL_TIMEOUT_JIFFIES); | ||
324 | |||
325 | spin_lock_irqsave(&pool->lock, flags); | ||
326 | __remove_wait_queue(&pool->waitq, &wait); | ||
327 | goto restart; | ||
328 | } | ||
329 | retval = NULL; | ||
330 | goto done; | ||
331 | } | ||
332 | |||
333 | ready: | ||
334 | page->in_use++; | ||
335 | offset = page->offset; | ||
336 | page->offset = *(int *)(page->vaddr + offset); | ||
337 | retval = offset + page->vaddr; | ||
338 | *handle = offset + page->dma; | ||
339 | #ifdef CONFIG_DEBUG_SLAB | ||
340 | memset(retval, POOL_POISON_ALLOCATED, pool->size); | ||
341 | #endif | ||
342 | done: | ||
343 | spin_unlock_irqrestore(&pool->lock, flags); | ||
344 | return retval; | ||
345 | } | ||
346 | EXPORT_SYMBOL(dma_pool_alloc); | ||
347 | |||
348 | static struct dma_page *pool_find_page(struct dma_pool *pool, dma_addr_t dma) | ||
349 | { | ||
350 | unsigned long flags; | ||
351 | struct dma_page *page; | ||
352 | |||
353 | spin_lock_irqsave(&pool->lock, flags); | ||
354 | list_for_each_entry(page, &pool->page_list, page_list) { | ||
355 | if (dma < page->dma) | ||
356 | continue; | ||
357 | if (dma < (page->dma + pool->allocation)) | ||
358 | goto done; | ||
359 | } | ||
360 | page = NULL; | ||
361 | done: | ||
362 | spin_unlock_irqrestore(&pool->lock, flags); | ||
363 | return page; | ||
364 | } | ||
365 | |||
366 | /** | ||
367 | * dma_pool_free - put block back into dma pool | ||
368 | * @pool: the dma pool holding the block | ||
369 | * @vaddr: virtual address of block | ||
370 | * @dma: dma address of block | ||
371 | * | ||
372 | * Caller promises neither device nor driver will again touch this block | ||
373 | * unless it is first re-allocated. | ||
374 | */ | ||
375 | void dma_pool_free(struct dma_pool *pool, void *vaddr, dma_addr_t dma) | ||
376 | { | ||
377 | struct dma_page *page; | ||
378 | unsigned long flags; | ||
379 | unsigned int offset; | ||
380 | |||
381 | page = pool_find_page(pool, dma); | ||
382 | if (!page) { | ||
383 | if (pool->dev) | ||
384 | dev_err(pool->dev, | ||
385 | "dma_pool_free %s, %p/%lx (bad dma)\n", | ||
386 | pool->name, vaddr, (unsigned long)dma); | ||
387 | else | ||
388 | printk(KERN_ERR "dma_pool_free %s, %p/%lx (bad dma)\n", | ||
389 | pool->name, vaddr, (unsigned long)dma); | ||
390 | return; | ||
391 | } | ||
392 | |||
393 | offset = vaddr - page->vaddr; | ||
394 | #ifdef CONFIG_DEBUG_SLAB | ||
395 | if ((dma - page->dma) != offset) { | ||
396 | if (pool->dev) | ||
397 | dev_err(pool->dev, | ||
398 | "dma_pool_free %s, %p (bad vaddr)/%Lx\n", | ||
399 | pool->name, vaddr, (unsigned long long)dma); | ||
400 | else | ||
401 | printk(KERN_ERR | ||
402 | "dma_pool_free %s, %p (bad vaddr)/%Lx\n", | ||
403 | pool->name, vaddr, (unsigned long long)dma); | ||
404 | return; | ||
405 | } | ||
406 | { | ||
407 | unsigned int chain = page->offset; | ||
408 | while (chain < pool->allocation) { | ||
409 | if (chain != offset) { | ||
410 | chain = *(int *)(page->vaddr + chain); | ||
411 | continue; | ||
412 | } | ||
413 | if (pool->dev) | ||
414 | dev_err(pool->dev, "dma_pool_free %s, dma %Lx " | ||
415 | "already free\n", pool->name, | ||
416 | (unsigned long long)dma); | ||
417 | else | ||
418 | printk(KERN_ERR "dma_pool_free %s, dma %Lx " | ||
419 | "already free\n", pool->name, | ||
420 | (unsigned long long)dma); | ||
421 | return; | ||
422 | } | ||
423 | } | ||
424 | memset(vaddr, POOL_POISON_FREED, pool->size); | ||
425 | #endif | ||
426 | |||
427 | spin_lock_irqsave(&pool->lock, flags); | ||
428 | page->in_use--; | ||
429 | *(int *)vaddr = page->offset; | ||
430 | page->offset = offset; | ||
431 | if (waitqueue_active(&pool->waitq)) | ||
432 | wake_up_locked(&pool->waitq); | ||
433 | /* | ||
434 | * Resist a temptation to do | ||
435 | * if (!is_page_busy(page)) pool_free_page(pool, page); | ||
436 | * Better have a few empty pages hang around. | ||
437 | */ | ||
438 | spin_unlock_irqrestore(&pool->lock, flags); | ||
439 | } | ||
440 | EXPORT_SYMBOL(dma_pool_free); | ||
441 | |||
442 | /* | ||
443 | * Managed DMA pool | ||
444 | */ | ||
445 | static void dmam_pool_release(struct device *dev, void *res) | ||
446 | { | ||
447 | struct dma_pool *pool = *(struct dma_pool **)res; | ||
448 | |||
449 | dma_pool_destroy(pool); | ||
450 | } | ||
451 | |||
452 | static int dmam_pool_match(struct device *dev, void *res, void *match_data) | ||
453 | { | ||
454 | return *(struct dma_pool **)res == match_data; | ||
455 | } | ||
456 | |||
457 | /** | ||
458 | * dmam_pool_create - Managed dma_pool_create() | ||
459 | * @name: name of pool, for diagnostics | ||
460 | * @dev: device that will be doing the DMA | ||
461 | * @size: size of the blocks in this pool. | ||
462 | * @align: alignment requirement for blocks; must be a power of two | ||
463 | * @allocation: returned blocks won't cross this boundary (or zero) | ||
464 | * | ||
465 | * Managed dma_pool_create(). DMA pool created with this function is | ||
466 | * automatically destroyed on driver detach. | ||
467 | */ | ||
468 | struct dma_pool *dmam_pool_create(const char *name, struct device *dev, | ||
469 | size_t size, size_t align, size_t allocation) | ||
470 | { | ||
471 | struct dma_pool **ptr, *pool; | ||
472 | |||
473 | ptr = devres_alloc(dmam_pool_release, sizeof(*ptr), GFP_KERNEL); | ||
474 | if (!ptr) | ||
475 | return NULL; | ||
476 | |||
477 | pool = *ptr = dma_pool_create(name, dev, size, align, allocation); | ||
478 | if (pool) | ||
479 | devres_add(dev, ptr); | ||
480 | else | ||
481 | devres_free(ptr); | ||
482 | |||
483 | return pool; | ||
484 | } | ||
485 | EXPORT_SYMBOL(dmam_pool_create); | ||
486 | |||
487 | /** | ||
488 | * dmam_pool_destroy - Managed dma_pool_destroy() | ||
489 | * @pool: dma pool that will be destroyed | ||
490 | * | ||
491 | * Managed dma_pool_destroy(). | ||
492 | */ | ||
493 | void dmam_pool_destroy(struct dma_pool *pool) | ||
494 | { | ||
495 | struct device *dev = pool->dev; | ||
496 | |||
497 | dma_pool_destroy(pool); | ||
498 | WARN_ON(devres_destroy(dev, dmam_pool_release, dmam_pool_match, pool)); | ||
499 | } | ||
500 | EXPORT_SYMBOL(dmam_pool_destroy); | ||