1 /*
2  * xvmalloc memory allocator
3  *
4  * Copyright (C) 2008, 2009, 2010  Nitin Gupta
5  *
6  * This code is released using a dual license strategy: BSD/GPL
7  * You can choose the licence that better fits your requirements.
8  *
9  * Released under the terms of 3-clause BSD License
10  * Released under the terms of GNU General Public License Version 2.0
11  */
12 
13 #ifdef CONFIG_ZRAM_DEBUG
14 #define DEBUG
15 #endif
16 
17 #include <linux/module.h>
18 #include <linux/kernel.h>
19 #include <linux/bitops.h>
20 #include <linux/errno.h>
21 #include <linux/highmem.h>
22 #include <linux/init.h>
23 #include <linux/string.h>
24 #include <linux/slab.h>
25 
26 #include "xvmalloc.h"
27 #include "xvmalloc_int.h"
28 
stat_inc(u64 * value)29 static void stat_inc(u64 *value)
30 {
31 	*value = *value + 1;
32 }
33 
stat_dec(u64 * value)34 static void stat_dec(u64 *value)
35 {
36 	*value = *value - 1;
37 }
38 
test_flag(struct block_header * block,enum blockflags flag)39 static int test_flag(struct block_header *block, enum blockflags flag)
40 {
41 	return block->prev & BIT(flag);
42 }
43 
set_flag(struct block_header * block,enum blockflags flag)44 static void set_flag(struct block_header *block, enum blockflags flag)
45 {
46 	block->prev |= BIT(flag);
47 }
48 
clear_flag(struct block_header * block,enum blockflags flag)49 static void clear_flag(struct block_header *block, enum blockflags flag)
50 {
51 	block->prev &= ~BIT(flag);
52 }
53 
54 /*
55  * Given <page, offset> pair, provide a dereferencable pointer.
56  * This is called from xv_malloc/xv_free path, so it
57  * needs to be fast.
58  */
get_ptr_atomic(struct page * page,u16 offset,enum km_type type)59 static void *get_ptr_atomic(struct page *page, u16 offset, enum km_type type)
60 {
61 	unsigned char *base;
62 
63 	base = kmap_atomic(page, type);
64 	return base + offset;
65 }
66 
put_ptr_atomic(void * ptr,enum km_type type)67 static void put_ptr_atomic(void *ptr, enum km_type type)
68 {
69 	kunmap_atomic(ptr, type);
70 }
71 
get_blockprev(struct block_header * block)72 static u32 get_blockprev(struct block_header *block)
73 {
74 	return block->prev & PREV_MASK;
75 }
76 
set_blockprev(struct block_header * block,u16 new_offset)77 static void set_blockprev(struct block_header *block, u16 new_offset)
78 {
79 	block->prev = new_offset | (block->prev & FLAGS_MASK);
80 }
81 
BLOCK_NEXT(struct block_header * block)82 static struct block_header *BLOCK_NEXT(struct block_header *block)
83 {
84 	return (struct block_header *)
85 		((char *)block + block->size + XV_ALIGN);
86 }
87 
88 /*
89  * Get index of free list containing blocks of maximum size
90  * which is less than or equal to given size.
91  */
get_index_for_insert(u32 size)92 static u32 get_index_for_insert(u32 size)
93 {
94 	if (unlikely(size > XV_MAX_ALLOC_SIZE))
95 		size = XV_MAX_ALLOC_SIZE;
96 	size &= ~FL_DELTA_MASK;
97 	return (size - XV_MIN_ALLOC_SIZE) >> FL_DELTA_SHIFT;
98 }
99 
100 /*
101  * Get index of free list having blocks of size greater than
102  * or equal to requested size.
103  */
get_index(u32 size)104 static u32 get_index(u32 size)
105 {
106 	if (unlikely(size < XV_MIN_ALLOC_SIZE))
107 		size = XV_MIN_ALLOC_SIZE;
108 	size = ALIGN(size, FL_DELTA);
109 	return (size - XV_MIN_ALLOC_SIZE) >> FL_DELTA_SHIFT;
110 }
111 
112 /**
113  * find_block - find block of at least given size
114  * @pool: memory pool to search from
115  * @size: size of block required
116  * @page: page containing required block
117  * @offset: offset within the page where block is located.
118  *
119  * Searches two level bitmap to locate block of at least
120  * the given size. If such a block is found, it provides
121  * <page, offset> to identify this block and returns index
122  * in freelist where we found this block.
123  * Otherwise, returns 0 and <page, offset> params are not touched.
124  */
find_block(struct xv_pool * pool,u32 size,struct page ** page,u32 * offset)125 static u32 find_block(struct xv_pool *pool, u32 size,
126 			struct page **page, u32 *offset)
127 {
128 	ulong flbitmap, slbitmap;
129 	u32 flindex, slindex, slbitstart;
130 
131 	/* There are no free blocks in this pool */
132 	if (!pool->flbitmap)
133 		return 0;
134 
135 	/* Get freelist index correspoding to this size */
136 	slindex = get_index(size);
137 	slbitmap = pool->slbitmap[slindex / BITS_PER_LONG];
138 	slbitstart = slindex % BITS_PER_LONG;
139 
140 	/*
141 	 * If freelist is not empty at this index, we found the
142 	 * block - head of this list. This is approximate best-fit match.
143 	 */
144 	if (test_bit(slbitstart, &slbitmap)) {
145 		*page = pool->freelist[slindex].page;
146 		*offset = pool->freelist[slindex].offset;
147 		return slindex;
148 	}
149 
150 	/*
151 	 * No best-fit found. Search a bit further in bitmap for a free block.
152 	 * Second level bitmap consists of series of 32-bit chunks. Search
153 	 * further in the chunk where we expected a best-fit, starting from
154 	 * index location found above.
155 	 */
156 	slbitstart++;
157 	slbitmap >>= slbitstart;
158 
159 	/* Skip this search if we were already at end of this bitmap chunk */
160 	if ((slbitstart != BITS_PER_LONG) && slbitmap) {
161 		slindex += __ffs(slbitmap) + 1;
162 		*page = pool->freelist[slindex].page;
163 		*offset = pool->freelist[slindex].offset;
164 		return slindex;
165 	}
166 
167 	/* Now do a full two-level bitmap search to find next nearest fit */
168 	flindex = slindex / BITS_PER_LONG;
169 
170 	flbitmap = (pool->flbitmap) >> (flindex + 1);
171 	if (!flbitmap)
172 		return 0;
173 
174 	flindex += __ffs(flbitmap) + 1;
175 	slbitmap = pool->slbitmap[flindex];
176 	slindex = (flindex * BITS_PER_LONG) + __ffs(slbitmap);
177 	*page = pool->freelist[slindex].page;
178 	*offset = pool->freelist[slindex].offset;
179 
180 	return slindex;
181 }
182 
183 /*
184  * Insert block at <page, offset> in freelist of given pool.
185  * freelist used depends on block size.
186  */
insert_block(struct xv_pool * pool,struct page * page,u32 offset,struct block_header * block)187 static void insert_block(struct xv_pool *pool, struct page *page, u32 offset,
188 			struct block_header *block)
189 {
190 	u32 flindex, slindex;
191 	struct block_header *nextblock;
192 
193 	slindex = get_index_for_insert(block->size);
194 	flindex = slindex / BITS_PER_LONG;
195 
196 	block->link.prev_page = NULL;
197 	block->link.prev_offset = 0;
198 	block->link.next_page = pool->freelist[slindex].page;
199 	block->link.next_offset = pool->freelist[slindex].offset;
200 	pool->freelist[slindex].page = page;
201 	pool->freelist[slindex].offset = offset;
202 
203 	if (block->link.next_page) {
204 		nextblock = get_ptr_atomic(block->link.next_page,
205 					block->link.next_offset, KM_USER1);
206 		nextblock->link.prev_page = page;
207 		nextblock->link.prev_offset = offset;
208 		put_ptr_atomic(nextblock, KM_USER1);
209 		/* If there was a next page then the free bits are set. */
210 		return;
211 	}
212 
213 	__set_bit(slindex % BITS_PER_LONG, &pool->slbitmap[flindex]);
214 	__set_bit(flindex, &pool->flbitmap);
215 }
216 
217 /*
218  * Remove block from freelist. Index 'slindex' identifies the freelist.
219  */
remove_block(struct xv_pool * pool,struct page * page,u32 offset,struct block_header * block,u32 slindex)220 static void remove_block(struct xv_pool *pool, struct page *page, u32 offset,
221 			struct block_header *block, u32 slindex)
222 {
223 	u32 flindex = slindex / BITS_PER_LONG;
224 	struct block_header *tmpblock;
225 
226 	if (block->link.prev_page) {
227 		tmpblock = get_ptr_atomic(block->link.prev_page,
228 				block->link.prev_offset, KM_USER1);
229 		tmpblock->link.next_page = block->link.next_page;
230 		tmpblock->link.next_offset = block->link.next_offset;
231 		put_ptr_atomic(tmpblock, KM_USER1);
232 	}
233 
234 	if (block->link.next_page) {
235 		tmpblock = get_ptr_atomic(block->link.next_page,
236 				block->link.next_offset, KM_USER1);
237 		tmpblock->link.prev_page = block->link.prev_page;
238 		tmpblock->link.prev_offset = block->link.prev_offset;
239 		put_ptr_atomic(tmpblock, KM_USER1);
240 	}
241 
242 	/* Is this block is at the head of the freelist? */
243 	if (pool->freelist[slindex].page == page
244 	   && pool->freelist[slindex].offset == offset) {
245 
246 		pool->freelist[slindex].page = block->link.next_page;
247 		pool->freelist[slindex].offset = block->link.next_offset;
248 
249 		if (pool->freelist[slindex].page) {
250 			struct block_header *tmpblock;
251 			tmpblock = get_ptr_atomic(pool->freelist[slindex].page,
252 					pool->freelist[slindex].offset,
253 					KM_USER1);
254 			tmpblock->link.prev_page = NULL;
255 			tmpblock->link.prev_offset = 0;
256 			put_ptr_atomic(tmpblock, KM_USER1);
257 		} else {
258 			/* This freelist bucket is empty */
259 			__clear_bit(slindex % BITS_PER_LONG,
260 				    &pool->slbitmap[flindex]);
261 			if (!pool->slbitmap[flindex])
262 				__clear_bit(flindex, &pool->flbitmap);
263 		}
264 	}
265 
266 	block->link.prev_page = NULL;
267 	block->link.prev_offset = 0;
268 	block->link.next_page = NULL;
269 	block->link.next_offset = 0;
270 }
271 
272 /*
273  * Allocate a page and add it to freelist of given pool.
274  */
grow_pool(struct xv_pool * pool,gfp_t flags)275 static int grow_pool(struct xv_pool *pool, gfp_t flags)
276 {
277 	struct page *page;
278 	struct block_header *block;
279 
280 	page = alloc_page(flags);
281 	if (unlikely(!page))
282 		return -ENOMEM;
283 
284 	stat_inc(&pool->total_pages);
285 
286 	spin_lock(&pool->lock);
287 	block = get_ptr_atomic(page, 0, KM_USER0);
288 
289 	block->size = PAGE_SIZE - XV_ALIGN;
290 	set_flag(block, BLOCK_FREE);
291 	clear_flag(block, PREV_FREE);
292 	set_blockprev(block, 0);
293 
294 	insert_block(pool, page, 0, block);
295 
296 	put_ptr_atomic(block, KM_USER0);
297 	spin_unlock(&pool->lock);
298 
299 	return 0;
300 }
301 
302 /*
303  * Create a memory pool. Allocates freelist, bitmaps and other
304  * per-pool metadata.
305  */
xv_create_pool(void)306 struct xv_pool *xv_create_pool(void)
307 {
308 	u32 ovhd_size;
309 	struct xv_pool *pool;
310 
311 	ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
312 	pool = kzalloc(ovhd_size, GFP_KERNEL);
313 	if (!pool)
314 		return NULL;
315 
316 	spin_lock_init(&pool->lock);
317 
318 	return pool;
319 }
320 EXPORT_SYMBOL_GPL(xv_create_pool);
321 
xv_destroy_pool(struct xv_pool * pool)322 void xv_destroy_pool(struct xv_pool *pool)
323 {
324 	kfree(pool);
325 }
326 EXPORT_SYMBOL_GPL(xv_destroy_pool);
327 
328 /**
329  * xv_malloc - Allocate block of given size from pool.
330  * @pool: pool to allocate from
331  * @size: size of block to allocate
332  * @page: page no. that holds the object
333  * @offset: location of object within page
334  *
335  * On success, <page, offset> identifies block allocated
336  * and 0 is returned. On failure, <page, offset> is set to
337  * 0 and -ENOMEM is returned.
338  *
339  * Allocation requests with size > XV_MAX_ALLOC_SIZE will fail.
340  */
xv_malloc(struct xv_pool * pool,u32 size,struct page ** page,u32 * offset,gfp_t flags)341 int xv_malloc(struct xv_pool *pool, u32 size, struct page **page,
342 		u32 *offset, gfp_t flags)
343 {
344 	int error;
345 	u32 index, tmpsize, origsize, tmpoffset;
346 	struct block_header *block, *tmpblock;
347 
348 	*page = NULL;
349 	*offset = 0;
350 	origsize = size;
351 
352 	if (unlikely(!size || size > XV_MAX_ALLOC_SIZE))
353 		return -ENOMEM;
354 
355 	size = ALIGN(size, XV_ALIGN);
356 
357 	spin_lock(&pool->lock);
358 
359 	index = find_block(pool, size, page, offset);
360 
361 	if (!*page) {
362 		spin_unlock(&pool->lock);
363 		if (flags & GFP_NOWAIT)
364 			return -ENOMEM;
365 		error = grow_pool(pool, flags);
366 		if (unlikely(error))
367 			return error;
368 
369 		spin_lock(&pool->lock);
370 		index = find_block(pool, size, page, offset);
371 	}
372 
373 	if (!*page) {
374 		spin_unlock(&pool->lock);
375 		return -ENOMEM;
376 	}
377 
378 	block = get_ptr_atomic(*page, *offset, KM_USER0);
379 
380 	remove_block(pool, *page, *offset, block, index);
381 
382 	/* Split the block if required */
383 	tmpoffset = *offset + size + XV_ALIGN;
384 	tmpsize = block->size - size;
385 	tmpblock = (struct block_header *)((char *)block + size + XV_ALIGN);
386 	if (tmpsize) {
387 		tmpblock->size = tmpsize - XV_ALIGN;
388 		set_flag(tmpblock, BLOCK_FREE);
389 		clear_flag(tmpblock, PREV_FREE);
390 
391 		set_blockprev(tmpblock, *offset);
392 		if (tmpblock->size >= XV_MIN_ALLOC_SIZE)
393 			insert_block(pool, *page, tmpoffset, tmpblock);
394 
395 		if (tmpoffset + XV_ALIGN + tmpblock->size != PAGE_SIZE) {
396 			tmpblock = BLOCK_NEXT(tmpblock);
397 			set_blockprev(tmpblock, tmpoffset);
398 		}
399 	} else {
400 		/* This block is exact fit */
401 		if (tmpoffset != PAGE_SIZE)
402 			clear_flag(tmpblock, PREV_FREE);
403 	}
404 
405 	block->size = origsize;
406 	clear_flag(block, BLOCK_FREE);
407 
408 	put_ptr_atomic(block, KM_USER0);
409 	spin_unlock(&pool->lock);
410 
411 	*offset += XV_ALIGN;
412 
413 	return 0;
414 }
415 EXPORT_SYMBOL_GPL(xv_malloc);
416 
417 /*
418  * Free block identified with <page, offset>
419  */
xv_free(struct xv_pool * pool,struct page * page,u32 offset)420 void xv_free(struct xv_pool *pool, struct page *page, u32 offset)
421 {
422 	void *page_start;
423 	struct block_header *block, *tmpblock;
424 
425 	offset -= XV_ALIGN;
426 
427 	spin_lock(&pool->lock);
428 
429 	page_start = get_ptr_atomic(page, 0, KM_USER0);
430 	block = (struct block_header *)((char *)page_start + offset);
431 
432 	/* Catch double free bugs */
433 	BUG_ON(test_flag(block, BLOCK_FREE));
434 
435 	block->size = ALIGN(block->size, XV_ALIGN);
436 
437 	tmpblock = BLOCK_NEXT(block);
438 	if (offset + block->size + XV_ALIGN == PAGE_SIZE)
439 		tmpblock = NULL;
440 
441 	/* Merge next block if its free */
442 	if (tmpblock && test_flag(tmpblock, BLOCK_FREE)) {
443 		/*
444 		 * Blocks smaller than XV_MIN_ALLOC_SIZE
445 		 * are not inserted in any free list.
446 		 */
447 		if (tmpblock->size >= XV_MIN_ALLOC_SIZE) {
448 			remove_block(pool, page,
449 				    offset + block->size + XV_ALIGN, tmpblock,
450 				    get_index_for_insert(tmpblock->size));
451 		}
452 		block->size += tmpblock->size + XV_ALIGN;
453 	}
454 
455 	/* Merge previous block if its free */
456 	if (test_flag(block, PREV_FREE)) {
457 		tmpblock = (struct block_header *)((char *)(page_start) +
458 						get_blockprev(block));
459 		offset = offset - tmpblock->size - XV_ALIGN;
460 
461 		if (tmpblock->size >= XV_MIN_ALLOC_SIZE)
462 			remove_block(pool, page, offset, tmpblock,
463 				    get_index_for_insert(tmpblock->size));
464 
465 		tmpblock->size += block->size + XV_ALIGN;
466 		block = tmpblock;
467 	}
468 
469 	/* No used objects in this page. Free it. */
470 	if (block->size == PAGE_SIZE - XV_ALIGN) {
471 		put_ptr_atomic(page_start, KM_USER0);
472 		spin_unlock(&pool->lock);
473 
474 		__free_page(page);
475 		stat_dec(&pool->total_pages);
476 		return;
477 	}
478 
479 	set_flag(block, BLOCK_FREE);
480 	if (block->size >= XV_MIN_ALLOC_SIZE)
481 		insert_block(pool, page, offset, block);
482 
483 	if (offset + block->size + XV_ALIGN != PAGE_SIZE) {
484 		tmpblock = BLOCK_NEXT(block);
485 		set_flag(tmpblock, PREV_FREE);
486 		set_blockprev(tmpblock, offset);
487 	}
488 
489 	put_ptr_atomic(page_start, KM_USER0);
490 	spin_unlock(&pool->lock);
491 }
492 EXPORT_SYMBOL_GPL(xv_free);
493 
xv_get_object_size(void * obj)494 u32 xv_get_object_size(void *obj)
495 {
496 	struct block_header *blk;
497 
498 	blk = (struct block_header *)((char *)(obj) - XV_ALIGN);
499 	return blk->size;
500 }
501 EXPORT_SYMBOL_GPL(xv_get_object_size);
502 
503 /*
504  * Returns total memory used by allocator (userdata + metadata)
505  */
xv_get_total_size_bytes(struct xv_pool * pool)506 u64 xv_get_total_size_bytes(struct xv_pool *pool)
507 {
508 	return pool->total_pages << PAGE_SHIFT;
509 }
510 EXPORT_SYMBOL_GPL(xv_get_total_size_bytes);
511