Change dmapool free block management

Use a list of free blocks within a page instead of using a bitmap. Update documentation to reflect this. As well as being a slight reduction in memory allocation, locked ops and lines of code, it speeds up a transaction processing benchmark by 0.4%. Signed-off-by: Matthew Wilcox <willy@linux.intel.com>

Change dmapool free block management
Use a list of free blocks within a page instead of using a bitmap. Update documentation to reflect this. As well as being a slight reduction in memory allocation, locked ops and lines of code, it speeds up a transaction processing benchmark by 0.4%. Signed-off-by: Matthew Wilcox <willy@linux.intel.com>
Matthew Wilcox
1 parent 6182a0943a
Showing 1 changed file with 58 additions and 61 deletions Side-by-side Diff
mm/dmapool.c
@@ -17,7 +17,9 @@
  * The current design of this allocator is fairly simple.  The pool is
  * represented by the 'struct dma_pool' which keeps a doubly-linked list of
  * allocated pages.  Each page in the page_list is split into blocks of at
- * least 'size' bytes.
+ * least 'size' bytes.  Free blocks are tracked in an unsorted singly-linked
+ * list of free blocks within the page.  Used blocks aren't tracked, but we
+ * keep a count of how many are currently allocated from each page.
  */
  
 #include <linux/device.h>
@@ -38,7 +40,6 @@
 struct dma_pool {		/* the pool */
 	struct list_head page_list;
 	spinlock_t lock;
-	size_t blocks_per_page;
 	size_t size;
 	struct device *dev;
 	size_t allocation;
@@ -51,8 +52,8 @@
 	struct list_head page_list;
 	void *vaddr;
 	dma_addr_t dma;
-	unsigned in_use;
-	unsigned long bitmap[0];
+	unsigned int in_use;
+	unsigned int offset;
 };
  
 #define	POOL_TIMEOUT_JIFFIES	((100 /* msec */ * HZ) / 1000)
@@ -87,8 +88,8 @@
  
 		/* per-pool info, no real statistics yet */
 		temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n",
-				 pool->name,
-				 blocks, pages * pool->blocks_per_page,
+				 pool->name, blocks,
+				 pages * (pool->allocation / pool->size),
 				 pool->size, pages);
 		size -= temp;
 		next += temp;
  
@@ -132,8 +133,11 @@
 		return NULL;
 	}
  
-	if (size == 0)
+	if (size == 0) {
 		return NULL;
+	} else if (size < 4) {
+		size = 4;
+	}
  
 	if ((size % align) != 0)
 		size = ALIGN(size, align);
@@ -160,7 +164,6 @@
 	spin_lock_init(&retval->lock);
 	retval->size = size;
 	retval->allocation = allocation;
-	retval->blocks_per_page = allocation / size;
 	init_waitqueue_head(&retval->waitq);
  
 	if (dev) {
  
  
  
  
  
  
@@ -186,28 +189,36 @@
 }
 EXPORT_SYMBOL(dma_pool_create);
  
+static void pool_initialise_page(struct dma_pool *pool, struct dma_page *page)
+{
+	unsigned int offset = 0;
+
+	do {
+		unsigned int next = offset + pool->size;
+		if (unlikely((next + pool->size) >= pool->allocation))
+			next = pool->allocation;
+		*(int *)(page->vaddr + offset) = next;
+		offset = next;
+	} while (offset < pool->allocation);
+}
+
 static struct dma_page *pool_alloc_page(struct dma_pool *pool, gfp_t mem_flags)
 {
 	struct dma_page *page;
-	int mapsize;
  
-	mapsize = pool->blocks_per_page;
-	mapsize = (mapsize + BITS_PER_LONG - 1) / BITS_PER_LONG;
-	mapsize *= sizeof(long);
-
-	page = kmalloc(mapsize + sizeof *page, mem_flags);
+	page = kmalloc(sizeof(*page), mem_flags);
 	if (!page)
 		return NULL;
-	page->vaddr = dma_alloc_coherent(pool->dev,
-					 pool->allocation,
+	page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
 					 &page->dma, mem_flags);
 	if (page->vaddr) {
-		memset(page->bitmap, 0xff, mapsize);	/* bit set == free */
 #ifdef	CONFIG_DEBUG_SLAB
 		memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
 #endif
+		pool_initialise_page(pool, page);
 		list_add(&page->page_list, &pool->page_list);
 		page->in_use = 0;
+		page->offset = 0;
 	} else {
 		kfree(page);
 		page = NULL;
  
@@ -215,14 +226,9 @@
 	return page;
 }
  
-static inline int is_page_busy(int blocks, unsigned long *bitmap)
+static inline int is_page_busy(struct dma_page *page)
 {
-	while (blocks > 0) {
-		if (*bitmap++ != ~0UL)
-			return 1;
-		blocks -= BITS_PER_LONG;
-	}
-	return 0;
+	return page->in_use != 0;
 }
  
 static void pool_free_page(struct dma_pool *pool, struct dma_page *page)
@@ -257,7 +263,7 @@
 		struct dma_page *page;
 		page = list_entry(pool->page_list.next,
 				  struct dma_page, page_list);
-		if (is_page_busy(pool->blocks_per_page, page->bitmap)) {
+		if (is_page_busy(page)) {
 			if (pool->dev)
 				dev_err(pool->dev,
 					"dma_pool_destroy %s, %p busy\n",
  
@@ -292,27 +298,14 @@
 {
 	unsigned long flags;
 	struct dma_page *page;
-	int map, block;
 	size_t offset;
 	void *retval;
  
 	spin_lock_irqsave(&pool->lock, flags);
  restart:
 	list_for_each_entry(page, &pool->page_list, page_list) {
-		int i;
-		/* only cachable accesses here ... */
-		for (map = 0, i = 0;
-		     i < pool->blocks_per_page; i += BITS_PER_LONG, map++) {
-			if (page->bitmap[map] == 0)
-				continue;
-			block = ffz(~page->bitmap[map]);
-			if ((i + block) < pool->blocks_per_page) {
-				clear_bit(block, &page->bitmap[map]);
-				offset = (BITS_PER_LONG * map) + block;
-				offset *= pool->size;
-				goto ready;
-			}
-		}
+		if (page->offset < pool->allocation)
+			goto ready;
 	}
 	page = pool_alloc_page(pool, GFP_ATOMIC);
 	if (!page) {
  
@@ -333,10 +326,10 @@
 		goto done;
 	}
  
-	clear_bit(0, &page->bitmap[0]);
-	offset = 0;
  ready:
 	page->in_use++;
+	offset = page->offset;
+	page->offset = *(int *)(page->vaddr + offset);
 	retval = offset + page->vaddr;
 	*handle = offset + page->dma;
 #ifdef	CONFIG_DEBUG_SLAB
@@ -379,7 +372,7 @@
 {
 	struct dma_page *page;
 	unsigned long flags;
-	int map, block;
+	unsigned int offset;
  
 	page = pool_find_page(pool, dma);
 	if (!page) {
  
@@ -393,13 +386,9 @@
 		return;
 	}
  
-	block = dma - page->dma;
-	block /= pool->size;
-	map = block / BITS_PER_LONG;
-	block %= BITS_PER_LONG;
-
+	offset = vaddr - page->vaddr;
 #ifdef	CONFIG_DEBUG_SLAB
-	if (((dma - page->dma) + (void *)page->vaddr) != vaddr) {
+	if ((dma - page->dma) != offset) {
 		if (pool->dev)
 			dev_err(pool->dev,
 				"dma_pool_free %s, %p (bad vaddr)/%Lx\n",
  
  
@@ -410,28 +399,36 @@
 			       pool->name, vaddr, (unsigned long long)dma);
 		return;
 	}
-	if (page->bitmap[map] & (1UL << block)) {
-		if (pool->dev)
-			dev_err(pool->dev,
-				"dma_pool_free %s, dma %Lx already free\n",
-				pool->name, (unsigned long long)dma);
-		else
-			printk(KERN_ERR
-			       "dma_pool_free %s, dma %Lx already free\n",
-			       pool->name, (unsigned long long)dma);
-		return;
+	{
+		unsigned int chain = page->offset;
+		while (chain < pool->allocation) {
+			if (chain != offset) {
+				chain = *(int *)(page->vaddr + chain);
+				continue;
+			}
+			if (pool->dev)
+				dev_err(pool->dev, "dma_pool_free %s, dma %Lx "
+					"already free\n", pool->name,
+					(unsigned long long)dma);
+			else
+				printk(KERN_ERR "dma_pool_free %s, dma %Lx "
+					"already free\n", pool->name,
+					(unsigned long long)dma);
+			return;
+		}
 	}
 	memset(vaddr, POOL_POISON_FREED, pool->size);
 #endif
  
 	spin_lock_irqsave(&pool->lock, flags);
 	page->in_use--;
-	set_bit(block, &page->bitmap[map]);
+	*(int *)vaddr = page->offset;
+	page->offset = offset;
 	if (waitqueue_active(&pool->waitq))
 		wake_up_locked(&pool->waitq);
 	/*
 	 * Resist a temptation to do
-	 *    if (!is_page_busy(bpp, page->bitmap)) pool_free_page(pool, page);
+	 *    if (!is_page_busy(page)) pool_free_page(pool, page);
 	 * Better have a few empty pages hang around.
 	 */
 	spin_unlock_irqrestore(&pool->lock, flags);
...	...	@@ -17,7 +17,9 @@
17	17	* The current design of this allocator is fairly simple. The pool is
18	18	* represented by the 'struct dma_pool' which keeps a doubly-linked list of
19	19	* allocated pages. Each page in the page_list is split into blocks of at
20		- * least 'size' bytes.
	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.
21	23	*/
22	24
23	25	#include <linux/device.h>
...	...	@@ -38,7 +40,6 @@
38	40	struct dma_pool { /* the pool */
39	41	struct list_head page_list;
40	42	spinlock_t lock;
41		- size_t blocks_per_page;
42	43	size_t size;
43	44	struct device *dev;
44	45	size_t allocation;
...	...	@@ -51,8 +52,8 @@
51	52	struct list_head page_list;
52	53	void *vaddr;
53	54	dma_addr_t dma;
54		- unsigned in_use;
55		- unsigned long bitmap[0];
	55	+ unsigned int in_use;
	56	+ unsigned int offset;
56	57	};
57	58
58	59	#define POOL_TIMEOUT_JIFFIES ((100 /* msec / HZ) / 1000)
...	...	@@ -87,8 +88,8 @@
87	88
88	89	/* per-pool info, no real statistics yet */
89	90	temp = scnprintf(next, size, "%-16s %4u %4Zu %4Zu %2u\n",
90		- pool->name,
91		- blocks, pages * pool->blocks_per_page,
	91	+ pool->name, blocks,
	92	+ pages * (pool->allocation / pool->size),
92	93	pool->size, pages);
93	94	size -= temp;
94	95	next += temp;
95	96
...	...	@@ -132,8 +133,11 @@
132	133	return NULL;
133	134	}
134	135
135		- if (size == 0)
	136	+ if (size == 0) {
136	137	return NULL;
	138	+ } else if (size < 4) {
	139	+ size = 4;
	140	+ }
137	141
138	142	if ((size % align) != 0)
139	143	size = ALIGN(size, align);
...	...	@@ -160,7 +164,6 @@
160	164	spin_lock_init(&retval->lock);
161	165	retval->size = size;
162	166	retval->allocation = allocation;
163		- retval->blocks_per_page = allocation / size;
164	167	init_waitqueue_head(&retval->waitq);
165	168
166	169	if (dev) {
167	170
168	171
169	172
170	173
171	174
172	175
...	...	@@ -186,28 +189,36 @@
186	189	}
187	190	EXPORT_SYMBOL(dma_pool_create);
188	191
	192	+static void pool_initialise_page(struct dma_pool pool, struct dma_page page)
	193	+{
	194	+ unsigned int offset = 0;
	195	+
	196	+ do {
	197	+ unsigned int next = offset + pool->size;
	198	+ if (unlikely((next + pool->size) >= pool->allocation))
	199	+ next = pool->allocation;
	200	+ (int )(page->vaddr + offset) = next;
	201	+ offset = next;
	202	+ } while (offset < pool->allocation);
	203	+}
	204	+
189	205	static struct dma_page pool_alloc_page(struct dma_pool pool, gfp_t mem_flags)
190	206	{
191	207	struct dma_page *page;
192		- int mapsize;
193	208
194		- mapsize = pool->blocks_per_page;
195		- mapsize = (mapsize + BITS_PER_LONG - 1) / BITS_PER_LONG;
196		- mapsize *= sizeof(long);
197		-
198		- page = kmalloc(mapsize + sizeof *page, mem_flags);
	209	+ page = kmalloc(sizeof(*page), mem_flags);
199	210	if (!page)
200	211	return NULL;
201		- page->vaddr = dma_alloc_coherent(pool->dev,
202		- pool->allocation,
	212	+ page->vaddr = dma_alloc_coherent(pool->dev, pool->allocation,
203	213	&page->dma, mem_flags);
204	214	if (page->vaddr) {
205		- memset(page->bitmap, 0xff, mapsize); /* bit set == free */
206	215	#ifdef CONFIG_DEBUG_SLAB
207	216	memset(page->vaddr, POOL_POISON_FREED, pool->allocation);
208	217	#endif
	218	+ pool_initialise_page(pool, page);
209	219	list_add(&page->page_list, &pool->page_list);
210	220	page->in_use = 0;
	221	+ page->offset = 0;
211	222	} else {
212	223	kfree(page);
213	224	page = NULL;
214	225
...	...	@@ -215,14 +226,9 @@
215	226	return page;
216	227	}
217	228
218		-static inline int is_page_busy(int blocks, unsigned long *bitmap)
	229	+static inline int is_page_busy(struct dma_page *page)
219	230	{
220		- while (blocks > 0) {
221		- if (*bitmap++ != ~0UL)
222		- return 1;
223		- blocks -= BITS_PER_LONG;
224		- }
225		- return 0;
	231	+ return page->in_use != 0;
226	232	}
227	233
228	234	static void pool_free_page(struct dma_pool pool, struct dma_page page)
...	...	@@ -257,7 +263,7 @@
257	263	struct dma_page *page;
258	264	page = list_entry(pool->page_list.next,
259	265	struct dma_page, page_list);
260		- if (is_page_busy(pool->blocks_per_page, page->bitmap)) {
	266	+ if (is_page_busy(page)) {
261	267	if (pool->dev)
262	268	dev_err(pool->dev,
263	269	"dma_pool_destroy %s, %p busy\n",
264	270
...	...	@@ -292,27 +298,14 @@
292	298	{
293	299	unsigned long flags;
294	300	struct dma_page *page;
295		- int map, block;
296	301	size_t offset;
297	302	void *retval;
298	303
299	304	spin_lock_irqsave(&pool->lock, flags);
300	305	restart:
301	306	list_for_each_entry(page, &pool->page_list, page_list) {
302		- int i;
303		- /* only cachable accesses here ... */
304		- for (map = 0, i = 0;
305		- i < pool->blocks_per_page; i += BITS_PER_LONG, map++) {
306		- if (page->bitmap[map] == 0)
307		- continue;
308		- block = ffz(~page->bitmap[map]);
309		- if ((i + block) < pool->blocks_per_page) {
310		- clear_bit(block, &page->bitmap[map]);
311		- offset = (BITS_PER_LONG * map) + block;
312		- offset *= pool->size;
313		- goto ready;
314		- }
315		- }
	307	+ if (page->offset < pool->allocation)
	308	+ goto ready;
316	309	}
317	310	page = pool_alloc_page(pool, GFP_ATOMIC);
318	311	if (!page) {
319	312
...	...	@@ -333,10 +326,10 @@
333	326	goto done;
334	327	}
335	328
336		- clear_bit(0, &page->bitmap[0]);
337		- offset = 0;
338	329	ready:
339	330	page->in_use++;
	331	+ offset = page->offset;
	332	+ page->offset = (int )(page->vaddr + offset);
340	333	retval = offset + page->vaddr;
341	334	*handle = offset + page->dma;
342	335	#ifdef CONFIG_DEBUG_SLAB
...	...	@@ -379,7 +372,7 @@
379	372	{
380	373	struct dma_page *page;
381	374	unsigned long flags;
382		- int map, block;
	375	+ unsigned int offset;
383	376
384	377	page = pool_find_page(pool, dma);
385	378	if (!page) {
386	379
...	...	@@ -393,13 +386,9 @@
393	386	return;
394	387	}
395	388
396		- block = dma - page->dma;
397		- block /= pool->size;
398		- map = block / BITS_PER_LONG;
399		- block %= BITS_PER_LONG;
400		-
	389	+ offset = vaddr - page->vaddr;
401	390	#ifdef CONFIG_DEBUG_SLAB
402		- if (((dma - page->dma) + (void *)page->vaddr) != vaddr) {
	391	+ if ((dma - page->dma) != offset) {
403	392	if (pool->dev)
404	393	dev_err(pool->dev,
405	394	"dma_pool_free %s, %p (bad vaddr)/%Lx\n",
406	395
407	396
...	...	@@ -410,28 +399,36 @@
410	399	pool->name, vaddr, (unsigned long long)dma);
411	400	return;
412	401	}
413		- if (page->bitmap[map] & (1UL << block)) {
414		- if (pool->dev)
415		- dev_err(pool->dev,
416		- "dma_pool_free %s, dma %Lx already free\n",
417		- pool->name, (unsigned long long)dma);
418		- else
419		- printk(KERN_ERR
420		- "dma_pool_free %s, dma %Lx already free\n",
421		- pool->name, (unsigned long long)dma);
422		- return;
	402	+ {
	403	+ unsigned int chain = page->offset;
	404	+ while (chain < pool->allocation) {
	405	+ if (chain != offset) {
	406	+ chain = (int )(page->vaddr + chain);
	407	+ continue;
	408	+ }
	409	+ if (pool->dev)
	410	+ dev_err(pool->dev, "dma_pool_free %s, dma %Lx "
	411	+ "already free\n", pool->name,
	412	+ (unsigned long long)dma);
	413	+ else
	414	+ printk(KERN_ERR "dma_pool_free %s, dma %Lx "
	415	+ "already free\n", pool->name,
	416	+ (unsigned long long)dma);
	417	+ return;
	418	+ }
423	419	}
424	420	memset(vaddr, POOL_POISON_FREED, pool->size);
425	421	#endif
426	422
427	423	spin_lock_irqsave(&pool->lock, flags);
428	424	page->in_use--;
429		- set_bit(block, &page->bitmap[map]);
	425	+ (int )vaddr = page->offset;
	426	+ page->offset = offset;
430	427	if (waitqueue_active(&pool->waitq))
431	428	wake_up_locked(&pool->waitq);
432	429	/*
433	430	* Resist a temptation to do
434		- * if (!is_page_busy(bpp, page->bitmap)) pool_free_page(pool, page);
	431	+ * if (!is_page_busy(page)) pool_free_page(pool, page);
435	432	* Better have a few empty pages hang around.
436	433	*/
437	434	spin_unlock_irqrestore(&pool->lock, flags);