1 // SPDX-License-Identifier: GPL-2.0-only
5 * Copyright (C) 2013, Seth Jennings, IBM
7 * Concepts based on zcache internal zbud allocator by Dan Magenheimer.
9 * zbud is an special purpose allocator for storing compressed pages. Contrary
10 * to what its name may suggest, zbud is not a buddy allocator, but rather an
11 * allocator that "buddies" two compressed pages together in a single memory
14 * While this design limits storage density, it has simple and deterministic
15 * reclaim properties that make it preferable to a higher density approach when
16 * reclaim will be used.
18 * zbud works by storing compressed pages, or "zpages", together in pairs in a
19 * single memory page called a "zbud page". The first buddy is "left
20 * justified" at the beginning of the zbud page, and the last buddy is "right
21 * justified" at the end of the zbud page. The benefit is that if either
22 * buddy is freed, the freed buddy space, coalesced with whatever slack space
23 * that existed between the buddies, results in the largest possible free region
24 * within the zbud page.
26 * zbud also provides an attractive lower bound on density. The ratio of zpages
27 * to zbud pages can not be less than 1. This ensures that zbud can never "do
28 * harm" by using more pages to store zpages than the uncompressed zpages would
29 * have used on their own.
31 * zbud pages are divided into "chunks". The size of the chunks is fixed at
32 * compile time and determined by NCHUNKS_ORDER below. Dividing zbud pages
33 * into chunks allows organizing unbuddied zbud pages into a manageable number
34 * of unbuddied lists according to the number of free chunks available in the
37 * The zbud API differs from that of conventional allocators in that the
38 * allocation function, zbud_alloc(), returns an opaque handle to the user,
39 * not a dereferenceable pointer. The user must map the handle using
40 * zbud_map() in order to get a usable pointer by which to access the
41 * allocation data and unmap the handle with zbud_unmap() when operations
42 * on the allocation data are complete.
45 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
47 #include <linux/atomic.h>
48 #include <linux/list.h>
50 #include <linux/module.h>
51 #include <linux/preempt.h>
52 #include <linux/slab.h>
53 #include <linux/spinlock.h>
54 #include <linux/zpool.h>
60 * NCHUNKS_ORDER determines the internal allocation granularity, effectively
61 * adjusting internal fragmentation. It also determines the number of
62 * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
63 * allocation granularity will be in chunks of size PAGE_SIZE/64. As one chunk
64 * in allocated page is occupied by zbud header, NCHUNKS will be calculated to
65 * 63 which shows the max number of free chunks in zbud page, also there will be
66 * 63 freelists per pool.
68 #define NCHUNKS_ORDER 6
70 #define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
71 #define CHUNK_SIZE (1 << CHUNK_SHIFT)
72 #define ZHDR_SIZE_ALIGNED CHUNK_SIZE
73 #define NCHUNKS ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
78 int (*evict)(struct zbud_pool *pool, unsigned long handle);
82 * struct zbud_pool - stores metadata for each zbud pool
83 * @lock: protects all pool fields and first|last_chunk fields of any
84 * zbud page in the pool
85 * @unbuddied: array of lists tracking zbud pages that only contain one buddy;
86 * the lists each zbud page is added to depends on the size of
88 * @buddied: list tracking the zbud pages that contain two buddies;
89 * these zbud pages are full
90 * @lru: list tracking the zbud pages in LRU order by most recently
92 * @pages_nr: number of zbud pages in the pool.
93 * @ops: pointer to a structure of user defined operations specified at
95 * @zpool: zpool driver
96 * @zpool_ops: zpool operations structure with an evict callback
98 * This structure is allocated at pool creation time and maintains metadata
99 * pertaining to a particular zbud pool.
105 * Reuse unbuddied[0] as buddied on the ground that
106 * unbuddied[0] is unused.
108 struct list_head buddied;
109 struct list_head unbuddied[NCHUNKS];
111 struct list_head lru;
113 const struct zbud_ops *ops;
115 const struct zpool_ops *zpool_ops;
119 * struct zbud_header - zbud page metadata occupying the first chunk of each
121 * @buddy: links the zbud page into the unbuddied/buddied lists in the pool
122 * @lru: links the zbud page into the lru list in the pool
123 * @first_chunks: the size of the first buddy in chunks, 0 if free
124 * @last_chunks: the size of the last buddy in chunks, 0 if free
127 struct list_head buddy;
128 struct list_head lru;
129 unsigned int first_chunks;
130 unsigned int last_chunks;
137 /* Just to make the code easier to read */
143 /* Converts an allocation size in bytes to size in zbud chunks */
144 static int size_to_chunks(size_t size)
146 return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
149 #define for_each_unbuddied_list(_iter, _begin) \
150 for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
152 /* Initializes the zbud header of a newly allocated zbud page */
153 static struct zbud_header *init_zbud_page(struct page *page)
155 struct zbud_header *zhdr = page_address(page);
156 zhdr->first_chunks = 0;
157 zhdr->last_chunks = 0;
158 INIT_LIST_HEAD(&zhdr->buddy);
159 INIT_LIST_HEAD(&zhdr->lru);
160 zhdr->under_reclaim = false;
164 /* Resets the struct page fields and frees the page */
165 static void free_zbud_page(struct zbud_header *zhdr)
167 __free_page(virt_to_page(zhdr));
171 * Encodes the handle of a particular buddy within a zbud page
172 * Pool lock should be held as this function accesses first|last_chunks
174 static unsigned long encode_handle(struct zbud_header *zhdr, enum buddy bud)
176 unsigned long handle;
179 * For now, the encoded handle is actually just the pointer to the data
180 * but this might not always be the case. A little information hiding.
181 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
182 * over the zbud header in the first chunk.
184 handle = (unsigned long)zhdr;
186 /* skip over zbud header */
187 handle += ZHDR_SIZE_ALIGNED;
188 else /* bud == LAST */
189 handle += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
193 /* Returns the zbud page where a given handle is stored */
194 static struct zbud_header *handle_to_zbud_header(unsigned long handle)
196 return (struct zbud_header *)(handle & PAGE_MASK);
199 /* Returns the number of free chunks in a zbud page */
200 static int num_free_chunks(struct zbud_header *zhdr)
203 * Rather than branch for different situations, just use the fact that
204 * free buddies have a length of zero to simplify everything.
206 return NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
213 * zbud_create_pool() - create a new zbud pool
214 * @gfp: gfp flags when allocating the zbud pool structure
215 * @ops: user-defined operations for the zbud pool
217 * Return: pointer to the new zbud pool or NULL if the metadata allocation
220 static struct zbud_pool *zbud_create_pool(gfp_t gfp, const struct zbud_ops *ops)
222 struct zbud_pool *pool;
225 pool = kzalloc(sizeof(struct zbud_pool), gfp);
228 spin_lock_init(&pool->lock);
229 for_each_unbuddied_list(i, 0)
230 INIT_LIST_HEAD(&pool->unbuddied[i]);
231 INIT_LIST_HEAD(&pool->buddied);
232 INIT_LIST_HEAD(&pool->lru);
239 * zbud_destroy_pool() - destroys an existing zbud pool
240 * @pool: the zbud pool to be destroyed
242 * The pool should be emptied before this function is called.
244 static void zbud_destroy_pool(struct zbud_pool *pool)
250 * zbud_alloc() - allocates a region of a given size
251 * @pool: zbud pool from which to allocate
252 * @size: size in bytes of the desired allocation
253 * @gfp: gfp flags used if the pool needs to grow
254 * @handle: handle of the new allocation
256 * This function will attempt to find a free region in the pool large enough to
257 * satisfy the allocation request. A search of the unbuddied lists is
258 * performed first. If no suitable free region is found, then a new page is
259 * allocated and added to the pool to satisfy the request.
261 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
262 * as zbud pool pages.
264 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
265 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
268 static int zbud_alloc(struct zbud_pool *pool, size_t size, gfp_t gfp,
269 unsigned long *handle)
271 int chunks, i, freechunks;
272 struct zbud_header *zhdr = NULL;
276 if (!size || (gfp & __GFP_HIGHMEM))
278 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
280 chunks = size_to_chunks(size);
281 spin_lock(&pool->lock);
283 /* First, try to find an unbuddied zbud page. */
284 for_each_unbuddied_list(i, chunks) {
285 if (!list_empty(&pool->unbuddied[i])) {
286 zhdr = list_first_entry(&pool->unbuddied[i],
287 struct zbud_header, buddy);
288 list_del(&zhdr->buddy);
289 if (zhdr->first_chunks == 0)
297 /* Couldn't find unbuddied zbud page, create new one */
298 spin_unlock(&pool->lock);
299 page = alloc_page(gfp);
302 spin_lock(&pool->lock);
304 zhdr = init_zbud_page(page);
309 zhdr->first_chunks = chunks;
311 zhdr->last_chunks = chunks;
313 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0) {
314 /* Add to unbuddied list */
315 freechunks = num_free_chunks(zhdr);
316 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
318 /* Add to buddied list */
319 list_add(&zhdr->buddy, &pool->buddied);
322 /* Add/move zbud page to beginning of LRU */
323 if (!list_empty(&zhdr->lru))
324 list_del(&zhdr->lru);
325 list_add(&zhdr->lru, &pool->lru);
327 *handle = encode_handle(zhdr, bud);
328 spin_unlock(&pool->lock);
334 * zbud_free() - frees the allocation associated with the given handle
335 * @pool: pool in which the allocation resided
336 * @handle: handle associated with the allocation returned by zbud_alloc()
338 * In the case that the zbud page in which the allocation resides is under
339 * reclaim, as indicated by the PG_reclaim flag being set, this function
340 * only sets the first|last_chunks to 0. The page is actually freed
341 * once both buddies are evicted (see zbud_reclaim_page() below).
343 static void zbud_free(struct zbud_pool *pool, unsigned long handle)
345 struct zbud_header *zhdr;
348 spin_lock(&pool->lock);
349 zhdr = handle_to_zbud_header(handle);
351 /* If first buddy, handle will be page aligned */
352 if ((handle - ZHDR_SIZE_ALIGNED) & ~PAGE_MASK)
353 zhdr->last_chunks = 0;
355 zhdr->first_chunks = 0;
357 if (zhdr->under_reclaim) {
358 /* zbud page is under reclaim, reclaim will free */
359 spin_unlock(&pool->lock);
363 /* Remove from existing buddy list */
364 list_del(&zhdr->buddy);
366 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
367 /* zbud page is empty, free */
368 list_del(&zhdr->lru);
369 free_zbud_page(zhdr);
372 /* Add to unbuddied list */
373 freechunks = num_free_chunks(zhdr);
374 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
377 spin_unlock(&pool->lock);
381 * zbud_reclaim_page() - evicts allocations from a pool page and frees it
382 * @pool: pool from which a page will attempt to be evicted
383 * @retries: number of pages on the LRU list for which eviction will
384 * be attempted before failing
386 * zbud reclaim is different from normal system reclaim in that the reclaim is
387 * done from the bottom, up. This is because only the bottom layer, zbud, has
388 * information on how the allocations are organized within each zbud page. This
389 * has the potential to create interesting locking situations between zbud and
392 * To avoid these, this is how zbud_reclaim_page() should be called:
394 * The user detects a page should be reclaimed and calls zbud_reclaim_page().
395 * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
396 * the user-defined eviction handler with the pool and handle as arguments.
398 * If the handle can not be evicted, the eviction handler should return
399 * non-zero. zbud_reclaim_page() will add the zbud page back to the
400 * appropriate list and try the next zbud page on the LRU up to
401 * a user defined number of retries.
403 * If the handle is successfully evicted, the eviction handler should
404 * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
405 * contains logic to delay freeing the page if the page is under reclaim,
406 * as indicated by the setting of the PG_reclaim flag on the underlying page.
408 * If all buddies in the zbud page are successfully evicted, then the
409 * zbud page can be freed.
411 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
412 * no pages to evict or an eviction handler is not registered, -EAGAIN if
413 * the retry limit was hit.
415 static int zbud_reclaim_page(struct zbud_pool *pool, unsigned int retries)
417 int i, ret, freechunks;
418 struct zbud_header *zhdr;
419 unsigned long first_handle = 0, last_handle = 0;
421 spin_lock(&pool->lock);
422 if (!pool->ops || !pool->ops->evict || list_empty(&pool->lru) ||
424 spin_unlock(&pool->lock);
427 for (i = 0; i < retries; i++) {
428 zhdr = list_last_entry(&pool->lru, struct zbud_header, lru);
429 list_del(&zhdr->lru);
430 list_del(&zhdr->buddy);
431 /* Protect zbud page against free */
432 zhdr->under_reclaim = true;
434 * We need encode the handles before unlocking, since we can
435 * race with free that will set (first|last)_chunks to 0
439 if (zhdr->first_chunks)
440 first_handle = encode_handle(zhdr, FIRST);
441 if (zhdr->last_chunks)
442 last_handle = encode_handle(zhdr, LAST);
443 spin_unlock(&pool->lock);
445 /* Issue the eviction callback(s) */
447 ret = pool->ops->evict(pool, first_handle);
452 ret = pool->ops->evict(pool, last_handle);
457 spin_lock(&pool->lock);
458 zhdr->under_reclaim = false;
459 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
461 * Both buddies are now free, free the zbud page and
464 free_zbud_page(zhdr);
466 spin_unlock(&pool->lock);
468 } else if (zhdr->first_chunks == 0 ||
469 zhdr->last_chunks == 0) {
470 /* add to unbuddied list */
471 freechunks = num_free_chunks(zhdr);
472 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
474 /* add to buddied list */
475 list_add(&zhdr->buddy, &pool->buddied);
478 /* add to beginning of LRU */
479 list_add(&zhdr->lru, &pool->lru);
481 spin_unlock(&pool->lock);
486 * zbud_map() - maps the allocation associated with the given handle
487 * @pool: pool in which the allocation resides
488 * @handle: handle associated with the allocation to be mapped
490 * While trivial for zbud, the mapping functions for others allocators
491 * implementing this allocation API could have more complex information encoded
492 * in the handle and could create temporary mappings to make the data
493 * accessible to the user.
495 * Returns: a pointer to the mapped allocation
497 static void *zbud_map(struct zbud_pool *pool, unsigned long handle)
499 return (void *)(handle);
503 * zbud_unmap() - maps the allocation associated with the given handle
504 * @pool: pool in which the allocation resides
505 * @handle: handle associated with the allocation to be unmapped
507 static void zbud_unmap(struct zbud_pool *pool, unsigned long handle)
512 * zbud_get_pool_size() - gets the zbud pool size in pages
513 * @pool: pool whose size is being queried
515 * Returns: size in pages of the given pool. The pool lock need not be
516 * taken to access pages_nr.
518 static u64 zbud_get_pool_size(struct zbud_pool *pool)
520 return pool->pages_nr;
527 static int zbud_zpool_evict(struct zbud_pool *pool, unsigned long handle)
529 if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
530 return pool->zpool_ops->evict(pool->zpool, handle);
535 static const struct zbud_ops zbud_zpool_ops = {
536 .evict = zbud_zpool_evict
539 static void *zbud_zpool_create(const char *name, gfp_t gfp,
540 const struct zpool_ops *zpool_ops,
543 struct zbud_pool *pool;
545 pool = zbud_create_pool(gfp, zpool_ops ? &zbud_zpool_ops : NULL);
548 pool->zpool_ops = zpool_ops;
553 static void zbud_zpool_destroy(void *pool)
555 zbud_destroy_pool(pool);
558 static int zbud_zpool_malloc(void *pool, size_t size, gfp_t gfp,
559 unsigned long *handle)
561 return zbud_alloc(pool, size, gfp, handle);
563 static void zbud_zpool_free(void *pool, unsigned long handle)
565 zbud_free(pool, handle);
568 static int zbud_zpool_shrink(void *pool, unsigned int pages,
569 unsigned int *reclaimed)
571 unsigned int total = 0;
574 while (total < pages) {
575 ret = zbud_reclaim_page(pool, 8);
587 static void *zbud_zpool_map(void *pool, unsigned long handle,
588 enum zpool_mapmode mm)
590 return zbud_map(pool, handle);
592 static void zbud_zpool_unmap(void *pool, unsigned long handle)
594 zbud_unmap(pool, handle);
597 static u64 zbud_zpool_total_size(void *pool)
599 return zbud_get_pool_size(pool) * PAGE_SIZE;
602 static struct zpool_driver zbud_zpool_driver = {
604 .sleep_mapped = true,
605 .owner = THIS_MODULE,
606 .create = zbud_zpool_create,
607 .destroy = zbud_zpool_destroy,
608 .malloc = zbud_zpool_malloc,
609 .free = zbud_zpool_free,
610 .shrink = zbud_zpool_shrink,
611 .map = zbud_zpool_map,
612 .unmap = zbud_zpool_unmap,
613 .total_size = zbud_zpool_total_size,
616 MODULE_ALIAS("zpool-zbud");
618 static int __init init_zbud(void)
620 /* Make sure the zbud header will fit in one chunk */
621 BUILD_BUG_ON(sizeof(struct zbud_header) > ZHDR_SIZE_ALIGNED);
624 zpool_register_driver(&zbud_zpool_driver);
629 static void __exit exit_zbud(void)
631 zpool_unregister_driver(&zbud_zpool_driver);
632 pr_info("unloaded\n");
635 module_init(init_zbud);
636 module_exit(exit_zbud);
638 MODULE_LICENSE("GPL");
639 MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
640 MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");