4 * Copyright (C) 2013, Seth Jennings, IBM
6 * Concepts based on zcache internal zbud allocator by Dan Magenheimer.
8 * zbud is an special purpose allocator for storing compressed pages. Contrary
9 * to what its name may suggest, zbud is not a buddy allocator, but rather an
10 * allocator that "buddies" two compressed pages together in a single memory
13 * While this design limits storage density, it has simple and deterministic
14 * reclaim properties that make it preferable to a higher density approach when
15 * reclaim will be used.
17 * zbud works by storing compressed pages, or "zpages", together in pairs in a
18 * single memory page called a "zbud page". The first buddy is "left
19 * justifed" at the beginning of the zbud page, and the last buddy is "right
20 * justified" at the end of the zbud page. The benefit is that if either
21 * buddy is freed, the freed buddy space, coalesced with whatever slack space
22 * that existed between the buddies, results in the largest possible free region
23 * within the zbud page.
25 * zbud also provides an attractive lower bound on density. The ratio of zpages
26 * to zbud pages can not be less than 1. This ensures that zbud can never "do
27 * harm" by using more pages to store zpages than the uncompressed zpages would
28 * have used on their own.
30 * zbud pages are divided into "chunks". The size of the chunks is fixed at
31 * compile time and determined by NCHUNKS_ORDER below. Dividing zbud pages
32 * into chunks allows organizing unbuddied zbud pages into a manageable number
33 * of unbuddied lists according to the number of free chunks available in the
36 * The zbud API differs from that of conventional allocators in that the
37 * allocation function, zbud_alloc(), returns an opaque handle to the user,
38 * not a dereferenceable pointer. The user must map the handle using
39 * zbud_map() in order to get a usable pointer by which to access the
40 * allocation data and unmap the handle with zbud_unmap() when operations
41 * on the allocation data are complete.
44 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
46 #include <linux/atomic.h>
47 #include <linux/list.h>
49 #include <linux/module.h>
50 #include <linux/preempt.h>
51 #include <linux/slab.h>
52 #include <linux/spinlock.h>
53 #include <linux/zbud.h>
59 * NCHUNKS_ORDER determines the internal allocation granularity, effectively
60 * adjusting internal fragmentation. It also determines the number of
61 * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
62 * allocation granularity will be in chunks of size PAGE_SIZE/64, and there
63 * will be 64 freelists per pool.
65 #define NCHUNKS_ORDER 6
67 #define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
68 #define CHUNK_SIZE (1 << CHUNK_SHIFT)
69 #define NCHUNKS (PAGE_SIZE >> CHUNK_SHIFT)
70 #define ZHDR_SIZE_ALIGNED CHUNK_SIZE
73 * struct zbud_pool - stores metadata for each zbud pool
74 * @lock: protects all pool fields and first|last_chunk fields of any
75 * zbud page in the pool
76 * @unbuddied: array of lists tracking zbud pages that only contain one buddy;
77 * the lists each zbud page is added to depends on the size of
79 * @buddied: list tracking the zbud pages that contain two buddies;
80 * these zbud pages are full
81 * @lru: list tracking the zbud pages in LRU order by most recently
83 * @pages_nr: number of zbud pages in the pool.
84 * @ops: pointer to a structure of user defined operations specified at
87 * This structure is allocated at pool creation time and maintains metadata
88 * pertaining to a particular zbud pool.
92 struct list_head unbuddied[NCHUNKS];
93 struct list_head buddied;
100 * struct zbud_header - zbud page metadata occupying the first chunk of each
102 * @buddy: links the zbud page into the unbuddied/buddied lists in the pool
103 * @lru: links the zbud page into the lru list in the pool
104 * @first_chunks: the size of the first buddy in chunks, 0 if free
105 * @last_chunks: the size of the last buddy in chunks, 0 if free
108 struct list_head buddy;
109 struct list_head lru;
110 unsigned int first_chunks;
111 unsigned int last_chunks;
118 /* Just to make the code easier to read */
124 /* Converts an allocation size in bytes to size in zbud chunks */
125 static int size_to_chunks(int size)
127 return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
130 #define for_each_unbuddied_list(_iter, _begin) \
131 for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
133 /* Initializes the zbud header of a newly allocated zbud page */
134 static struct zbud_header *init_zbud_page(struct page *page)
136 struct zbud_header *zhdr = page_address(page);
137 zhdr->first_chunks = 0;
138 zhdr->last_chunks = 0;
139 INIT_LIST_HEAD(&zhdr->buddy);
140 INIT_LIST_HEAD(&zhdr->lru);
141 zhdr->under_reclaim = 0;
145 /* Resets the struct page fields and frees the page */
146 static void free_zbud_page(struct zbud_header *zhdr)
148 __free_page(virt_to_page(zhdr));
152 * Encodes the handle of a particular buddy within a zbud page
153 * Pool lock should be held as this function accesses first|last_chunks
155 static unsigned long encode_handle(struct zbud_header *zhdr, enum buddy bud)
157 unsigned long handle;
160 * For now, the encoded handle is actually just the pointer to the data
161 * but this might not always be the case. A little information hiding.
162 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
163 * over the zbud header in the first chunk.
165 handle = (unsigned long)zhdr;
167 /* skip over zbud header */
168 handle += ZHDR_SIZE_ALIGNED;
169 else /* bud == LAST */
170 handle += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
174 /* Returns the zbud page where a given handle is stored */
175 static struct zbud_header *handle_to_zbud_header(unsigned long handle)
177 return (struct zbud_header *)(handle & PAGE_MASK);
180 /* Returns the number of free chunks in a zbud page */
181 static int num_free_chunks(struct zbud_header *zhdr)
184 * Rather than branch for different situations, just use the fact that
185 * free buddies have a length of zero to simplify everything. -1 at the
186 * end for the zbud header.
188 return NCHUNKS - zhdr->first_chunks - zhdr->last_chunks - 1;
195 * zbud_create_pool() - create a new zbud pool
196 * @gfp: gfp flags when allocating the zbud pool structure
197 * @ops: user-defined operations for the zbud pool
199 * Return: pointer to the new zbud pool or NULL if the metadata allocation
202 struct zbud_pool *zbud_create_pool(gfp_t gfp, struct zbud_ops *ops)
204 struct zbud_pool *pool;
207 pool = kmalloc(sizeof(struct zbud_pool), gfp);
210 spin_lock_init(&pool->lock);
211 for_each_unbuddied_list(i, 0)
212 INIT_LIST_HEAD(&pool->unbuddied[i]);
213 INIT_LIST_HEAD(&pool->buddied);
214 INIT_LIST_HEAD(&pool->lru);
221 * zbud_destroy_pool() - destroys an existing zbud pool
222 * @pool: the zbud pool to be destroyed
224 * The pool should be emptied before this function is called.
226 void zbud_destroy_pool(struct zbud_pool *pool)
232 * zbud_alloc() - allocates a region of a given size
233 * @pool: zbud pool from which to allocate
234 * @size: size in bytes of the desired allocation
235 * @gfp: gfp flags used if the pool needs to grow
236 * @handle: handle of the new allocation
238 * This function will attempt to find a free region in the pool large enough to
239 * satisfy the allocation request. A search of the unbuddied lists is
240 * performed first. If no suitable free region is found, then a new page is
241 * allocated and added to the pool to satisfy the request.
243 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
244 * as zbud pool pages.
246 * Return: 0 if success and handle is set, otherwise -EINVAL is the size or
247 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
250 int zbud_alloc(struct zbud_pool *pool, int size, gfp_t gfp,
251 unsigned long *handle)
253 int chunks, i, freechunks;
254 struct zbud_header *zhdr = NULL;
258 if (size <= 0 || gfp & __GFP_HIGHMEM)
260 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
262 chunks = size_to_chunks(size);
263 spin_lock(&pool->lock);
265 /* First, try to find an unbuddied zbud page. */
267 for_each_unbuddied_list(i, chunks) {
268 if (!list_empty(&pool->unbuddied[i])) {
269 zhdr = list_first_entry(&pool->unbuddied[i],
270 struct zbud_header, buddy);
271 list_del(&zhdr->buddy);
272 if (zhdr->first_chunks == 0)
280 /* Couldn't find unbuddied zbud page, create new one */
281 spin_unlock(&pool->lock);
282 page = alloc_page(gfp);
285 spin_lock(&pool->lock);
287 zhdr = init_zbud_page(page);
292 zhdr->first_chunks = chunks;
294 zhdr->last_chunks = chunks;
296 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0) {
297 /* Add to unbuddied list */
298 freechunks = num_free_chunks(zhdr);
299 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
301 /* Add to buddied list */
302 list_add(&zhdr->buddy, &pool->buddied);
305 /* Add/move zbud page to beginning of LRU */
306 if (!list_empty(&zhdr->lru))
307 list_del(&zhdr->lru);
308 list_add(&zhdr->lru, &pool->lru);
310 *handle = encode_handle(zhdr, bud);
311 spin_unlock(&pool->lock);
317 * zbud_free() - frees the allocation associated with the given handle
318 * @pool: pool in which the allocation resided
319 * @handle: handle associated with the allocation returned by zbud_alloc()
321 * In the case that the zbud page in which the allocation resides is under
322 * reclaim, as indicated by the PG_reclaim flag being set, this function
323 * only sets the first|last_chunks to 0. The page is actually freed
324 * once both buddies are evicted (see zbud_reclaim_page() below).
326 void zbud_free(struct zbud_pool *pool, unsigned long handle)
328 struct zbud_header *zhdr;
331 spin_lock(&pool->lock);
332 zhdr = handle_to_zbud_header(handle);
334 /* If first buddy, handle will be page aligned */
335 if ((handle - ZHDR_SIZE_ALIGNED) & ~PAGE_MASK)
336 zhdr->last_chunks = 0;
338 zhdr->first_chunks = 0;
340 if (zhdr->under_reclaim) {
341 /* zbud page is under reclaim, reclaim will free */
342 spin_unlock(&pool->lock);
346 /* Remove from existing buddy list */
347 list_del(&zhdr->buddy);
349 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
350 /* zbud page is empty, free */
351 list_del(&zhdr->lru);
352 free_zbud_page(zhdr);
355 /* Add to unbuddied list */
356 freechunks = num_free_chunks(zhdr);
357 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
360 spin_unlock(&pool->lock);
363 #define list_tail_entry(ptr, type, member) \
364 list_entry((ptr)->prev, type, member)
367 * zbud_reclaim_page() - evicts allocations from a pool page and frees it
368 * @pool: pool from which a page will attempt to be evicted
369 * @retires: number of pages on the LRU list for which eviction will
370 * be attempted before failing
372 * zbud reclaim is different from normal system reclaim in that the reclaim is
373 * done from the bottom, up. This is because only the bottom layer, zbud, has
374 * information on how the allocations are organized within each zbud page. This
375 * has the potential to create interesting locking situations between zbud and
378 * To avoid these, this is how zbud_reclaim_page() should be called:
380 * The user detects a page should be reclaimed and calls zbud_reclaim_page().
381 * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
382 * the user-defined eviction handler with the pool and handle as arguments.
384 * If the handle can not be evicted, the eviction handler should return
385 * non-zero. zbud_reclaim_page() will add the zbud page back to the
386 * appropriate list and try the next zbud page on the LRU up to
387 * a user defined number of retries.
389 * If the handle is successfully evicted, the eviction handler should
390 * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
391 * contains logic to delay freeing the page if the page is under reclaim,
392 * as indicated by the setting of the PG_reclaim flag on the underlying page.
394 * If all buddies in the zbud page are successfully evicted, then the
395 * zbud page can be freed.
397 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
398 * no pages to evict or an eviction handler is not registered, -EAGAIN if
399 * the retry limit was hit.
401 int zbud_reclaim_page(struct zbud_pool *pool, unsigned int retries)
403 int i, ret, freechunks;
404 struct zbud_header *zhdr;
405 unsigned long first_handle = 0, last_handle = 0;
407 spin_lock(&pool->lock);
408 if (!pool->ops || !pool->ops->evict || list_empty(&pool->lru) ||
410 spin_unlock(&pool->lock);
413 for (i = 0; i < retries; i++) {
414 zhdr = list_tail_entry(&pool->lru, struct zbud_header, lru);
415 list_del(&zhdr->lru);
416 list_del(&zhdr->buddy);
417 /* Protect zbud page against free */
418 zhdr->under_reclaim = true;
420 * We need encode the handles before unlocking, since we can
421 * race with free that will set (first|last)_chunks to 0
425 if (zhdr->first_chunks)
426 first_handle = encode_handle(zhdr, FIRST);
427 if (zhdr->last_chunks)
428 last_handle = encode_handle(zhdr, LAST);
429 spin_unlock(&pool->lock);
431 /* Issue the eviction callback(s) */
433 ret = pool->ops->evict(pool, first_handle);
438 ret = pool->ops->evict(pool, last_handle);
443 spin_lock(&pool->lock);
444 zhdr->under_reclaim = false;
445 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
447 * Both buddies are now free, free the zbud page and
450 free_zbud_page(zhdr);
452 spin_unlock(&pool->lock);
454 } else if (zhdr->first_chunks == 0 ||
455 zhdr->last_chunks == 0) {
456 /* add to unbuddied list */
457 freechunks = num_free_chunks(zhdr);
458 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
460 /* add to buddied list */
461 list_add(&zhdr->buddy, &pool->buddied);
464 /* add to beginning of LRU */
465 list_add(&zhdr->lru, &pool->lru);
467 spin_unlock(&pool->lock);
472 * zbud_map() - maps the allocation associated with the given handle
473 * @pool: pool in which the allocation resides
474 * @handle: handle associated with the allocation to be mapped
476 * While trivial for zbud, the mapping functions for others allocators
477 * implementing this allocation API could have more complex information encoded
478 * in the handle and could create temporary mappings to make the data
479 * accessible to the user.
481 * Returns: a pointer to the mapped allocation
483 void *zbud_map(struct zbud_pool *pool, unsigned long handle)
485 return (void *)(handle);
489 * zbud_unmap() - maps the allocation associated with the given handle
490 * @pool: pool in which the allocation resides
491 * @handle: handle associated with the allocation to be unmapped
493 void zbud_unmap(struct zbud_pool *pool, unsigned long handle)
498 * zbud_get_pool_size() - gets the zbud pool size in pages
499 * @pool: pool whose size is being queried
501 * Returns: size in pages of the given pool. The pool lock need not be
502 * taken to access pages_nr.
504 u64 zbud_get_pool_size(struct zbud_pool *pool)
506 return pool->pages_nr;
509 static int __init init_zbud(void)
511 /* Make sure the zbud header will fit in one chunk */
512 BUILD_BUG_ON(sizeof(struct zbud_header) > ZHDR_SIZE_ALIGNED);
517 static void __exit exit_zbud(void)
519 pr_info("unloaded\n");
522 module_init(init_zbud);
523 module_exit(exit_zbud);
525 MODULE_LICENSE("GPL");
526 MODULE_AUTHOR("Seth Jennings <sjenning@linux.vnet.ibm.com>");
527 MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");