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 * justified" 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>
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, and there
64 * will be 64 freelists per pool.
66 #define NCHUNKS_ORDER 6
68 #define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
69 #define CHUNK_SIZE (1 << CHUNK_SHIFT)
70 #define NCHUNKS (PAGE_SIZE >> CHUNK_SHIFT)
71 #define ZHDR_SIZE_ALIGNED CHUNK_SIZE
74 * struct zbud_pool - stores metadata for each zbud pool
75 * @lock: protects all pool fields and first|last_chunk fields of any
76 * zbud page in the pool
77 * @unbuddied: array of lists tracking zbud pages that only contain one buddy;
78 * the lists each zbud page is added to depends on the size of
80 * @buddied: list tracking the zbud pages that contain two buddies;
81 * these zbud pages are full
82 * @lru: list tracking the zbud pages in LRU order by most recently
84 * @pages_nr: number of zbud pages in the pool.
85 * @ops: pointer to a structure of user defined operations specified at
88 * This structure is allocated at pool creation time and maintains metadata
89 * pertaining to a particular zbud pool.
93 struct list_head unbuddied
[NCHUNKS
];
94 struct list_head buddied
;
101 * struct zbud_header - zbud page metadata occupying the first chunk of each
103 * @buddy: links the zbud page into the unbuddied/buddied lists in the pool
104 * @lru: links the zbud page into the lru list in the pool
105 * @first_chunks: the size of the first buddy in chunks, 0 if free
106 * @last_chunks: the size of the last buddy in chunks, 0 if free
109 struct list_head buddy
;
110 struct list_head lru
;
111 unsigned int first_chunks
;
112 unsigned int last_chunks
;
122 static int zbud_zpool_evict(struct zbud_pool
*pool
, unsigned long handle
)
124 return zpool_evict(pool
, handle
);
127 static struct zbud_ops zbud_zpool_ops
= {
128 .evict
= zbud_zpool_evict
131 static void *zbud_zpool_create(gfp_t gfp
, struct zpool_ops
*zpool_ops
)
133 return zbud_create_pool(gfp
, &zbud_zpool_ops
);
136 static void zbud_zpool_destroy(void *pool
)
138 zbud_destroy_pool(pool
);
141 static int zbud_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
142 unsigned long *handle
)
144 return zbud_alloc(pool
, size
, gfp
, handle
);
146 static void zbud_zpool_free(void *pool
, unsigned long handle
)
148 zbud_free(pool
, handle
);
151 static int zbud_zpool_shrink(void *pool
, unsigned int pages
,
152 unsigned int *reclaimed
)
154 unsigned int total
= 0;
157 while (total
< pages
) {
158 ret
= zbud_reclaim_page(pool
, 8);
170 static void *zbud_zpool_map(void *pool
, unsigned long handle
,
171 enum zpool_mapmode mm
)
173 return zbud_map(pool
, handle
);
175 static void zbud_zpool_unmap(void *pool
, unsigned long handle
)
177 zbud_unmap(pool
, handle
);
180 static u64
zbud_zpool_total_size(void *pool
)
182 return zbud_get_pool_size(pool
) * PAGE_SIZE
;
185 static struct zpool_driver zbud_zpool_driver
= {
187 .owner
= THIS_MODULE
,
188 .create
= zbud_zpool_create
,
189 .destroy
= zbud_zpool_destroy
,
190 .malloc
= zbud_zpool_malloc
,
191 .free
= zbud_zpool_free
,
192 .shrink
= zbud_zpool_shrink
,
193 .map
= zbud_zpool_map
,
194 .unmap
= zbud_zpool_unmap
,
195 .total_size
= zbud_zpool_total_size
,
198 #endif /* CONFIG_ZPOOL */
203 /* Just to make the code easier to read */
209 /* Converts an allocation size in bytes to size in zbud chunks */
210 static int size_to_chunks(size_t size
)
212 return (size
+ CHUNK_SIZE
- 1) >> CHUNK_SHIFT
;
215 #define for_each_unbuddied_list(_iter, _begin) \
216 for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
218 /* Initializes the zbud header of a newly allocated zbud page */
219 static struct zbud_header
*init_zbud_page(struct page
*page
)
221 struct zbud_header
*zhdr
= page_address(page
);
222 zhdr
->first_chunks
= 0;
223 zhdr
->last_chunks
= 0;
224 INIT_LIST_HEAD(&zhdr
->buddy
);
225 INIT_LIST_HEAD(&zhdr
->lru
);
226 zhdr
->under_reclaim
= 0;
230 /* Resets the struct page fields and frees the page */
231 static void free_zbud_page(struct zbud_header
*zhdr
)
233 __free_page(virt_to_page(zhdr
));
237 * Encodes the handle of a particular buddy within a zbud page
238 * Pool lock should be held as this function accesses first|last_chunks
240 static unsigned long encode_handle(struct zbud_header
*zhdr
, enum buddy bud
)
242 unsigned long handle
;
245 * For now, the encoded handle is actually just the pointer to the data
246 * but this might not always be the case. A little information hiding.
247 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
248 * over the zbud header in the first chunk.
250 handle
= (unsigned long)zhdr
;
252 /* skip over zbud header */
253 handle
+= ZHDR_SIZE_ALIGNED
;
254 else /* bud == LAST */
255 handle
+= PAGE_SIZE
- (zhdr
->last_chunks
<< CHUNK_SHIFT
);
259 /* Returns the zbud page where a given handle is stored */
260 static struct zbud_header
*handle_to_zbud_header(unsigned long handle
)
262 return (struct zbud_header
*)(handle
& PAGE_MASK
);
265 /* Returns the number of free chunks in a zbud page */
266 static int num_free_chunks(struct zbud_header
*zhdr
)
269 * Rather than branch for different situations, just use the fact that
270 * free buddies have a length of zero to simplify everything. -1 at the
271 * end for the zbud header.
273 return NCHUNKS
- zhdr
->first_chunks
- zhdr
->last_chunks
- 1;
280 * zbud_create_pool() - create a new zbud pool
281 * @gfp: gfp flags when allocating the zbud pool structure
282 * @ops: user-defined operations for the zbud pool
284 * Return: pointer to the new zbud pool or NULL if the metadata allocation
287 struct zbud_pool
*zbud_create_pool(gfp_t gfp
, struct zbud_ops
*ops
)
289 struct zbud_pool
*pool
;
292 pool
= kmalloc(sizeof(struct zbud_pool
), gfp
);
295 spin_lock_init(&pool
->lock
);
296 for_each_unbuddied_list(i
, 0)
297 INIT_LIST_HEAD(&pool
->unbuddied
[i
]);
298 INIT_LIST_HEAD(&pool
->buddied
);
299 INIT_LIST_HEAD(&pool
->lru
);
306 * zbud_destroy_pool() - destroys an existing zbud pool
307 * @pool: the zbud pool to be destroyed
309 * The pool should be emptied before this function is called.
311 void zbud_destroy_pool(struct zbud_pool
*pool
)
317 * zbud_alloc() - allocates a region of a given size
318 * @pool: zbud pool from which to allocate
319 * @size: size in bytes of the desired allocation
320 * @gfp: gfp flags used if the pool needs to grow
321 * @handle: handle of the new allocation
323 * This function will attempt to find a free region in the pool large enough to
324 * satisfy the allocation request. A search of the unbuddied lists is
325 * performed first. If no suitable free region is found, then a new page is
326 * allocated and added to the pool to satisfy the request.
328 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
329 * as zbud pool pages.
331 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
332 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
335 int zbud_alloc(struct zbud_pool
*pool
, size_t size
, gfp_t gfp
,
336 unsigned long *handle
)
338 int chunks
, i
, freechunks
;
339 struct zbud_header
*zhdr
= NULL
;
343 if (!size
|| (gfp
& __GFP_HIGHMEM
))
345 if (size
> PAGE_SIZE
- ZHDR_SIZE_ALIGNED
- CHUNK_SIZE
)
347 chunks
= size_to_chunks(size
);
348 spin_lock(&pool
->lock
);
350 /* First, try to find an unbuddied zbud page. */
352 for_each_unbuddied_list(i
, chunks
) {
353 if (!list_empty(&pool
->unbuddied
[i
])) {
354 zhdr
= list_first_entry(&pool
->unbuddied
[i
],
355 struct zbud_header
, buddy
);
356 list_del(&zhdr
->buddy
);
357 if (zhdr
->first_chunks
== 0)
365 /* Couldn't find unbuddied zbud page, create new one */
366 spin_unlock(&pool
->lock
);
367 page
= alloc_page(gfp
);
370 spin_lock(&pool
->lock
);
372 zhdr
= init_zbud_page(page
);
377 zhdr
->first_chunks
= chunks
;
379 zhdr
->last_chunks
= chunks
;
381 if (zhdr
->first_chunks
== 0 || zhdr
->last_chunks
== 0) {
382 /* Add to unbuddied list */
383 freechunks
= num_free_chunks(zhdr
);
384 list_add(&zhdr
->buddy
, &pool
->unbuddied
[freechunks
]);
386 /* Add to buddied list */
387 list_add(&zhdr
->buddy
, &pool
->buddied
);
390 /* Add/move zbud page to beginning of LRU */
391 if (!list_empty(&zhdr
->lru
))
392 list_del(&zhdr
->lru
);
393 list_add(&zhdr
->lru
, &pool
->lru
);
395 *handle
= encode_handle(zhdr
, bud
);
396 spin_unlock(&pool
->lock
);
402 * zbud_free() - frees the allocation associated with the given handle
403 * @pool: pool in which the allocation resided
404 * @handle: handle associated with the allocation returned by zbud_alloc()
406 * In the case that the zbud page in which the allocation resides is under
407 * reclaim, as indicated by the PG_reclaim flag being set, this function
408 * only sets the first|last_chunks to 0. The page is actually freed
409 * once both buddies are evicted (see zbud_reclaim_page() below).
411 void zbud_free(struct zbud_pool
*pool
, unsigned long handle
)
413 struct zbud_header
*zhdr
;
416 spin_lock(&pool
->lock
);
417 zhdr
= handle_to_zbud_header(handle
);
419 /* If first buddy, handle will be page aligned */
420 if ((handle
- ZHDR_SIZE_ALIGNED
) & ~PAGE_MASK
)
421 zhdr
->last_chunks
= 0;
423 zhdr
->first_chunks
= 0;
425 if (zhdr
->under_reclaim
) {
426 /* zbud page is under reclaim, reclaim will free */
427 spin_unlock(&pool
->lock
);
431 /* Remove from existing buddy list */
432 list_del(&zhdr
->buddy
);
434 if (zhdr
->first_chunks
== 0 && zhdr
->last_chunks
== 0) {
435 /* zbud page is empty, free */
436 list_del(&zhdr
->lru
);
437 free_zbud_page(zhdr
);
440 /* Add to unbuddied list */
441 freechunks
= num_free_chunks(zhdr
);
442 list_add(&zhdr
->buddy
, &pool
->unbuddied
[freechunks
]);
445 spin_unlock(&pool
->lock
);
448 #define list_tail_entry(ptr, type, member) \
449 list_entry((ptr)->prev, type, member)
452 * zbud_reclaim_page() - evicts allocations from a pool page and frees it
453 * @pool: pool from which a page will attempt to be evicted
454 * @retires: number of pages on the LRU list for which eviction will
455 * be attempted before failing
457 * zbud reclaim is different from normal system reclaim in that the reclaim is
458 * done from the bottom, up. This is because only the bottom layer, zbud, has
459 * information on how the allocations are organized within each zbud page. This
460 * has the potential to create interesting locking situations between zbud and
463 * To avoid these, this is how zbud_reclaim_page() should be called:
465 * The user detects a page should be reclaimed and calls zbud_reclaim_page().
466 * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
467 * the user-defined eviction handler with the pool and handle as arguments.
469 * If the handle can not be evicted, the eviction handler should return
470 * non-zero. zbud_reclaim_page() will add the zbud page back to the
471 * appropriate list and try the next zbud page on the LRU up to
472 * a user defined number of retries.
474 * If the handle is successfully evicted, the eviction handler should
475 * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
476 * contains logic to delay freeing the page if the page is under reclaim,
477 * as indicated by the setting of the PG_reclaim flag on the underlying page.
479 * If all buddies in the zbud page are successfully evicted, then the
480 * zbud page can be freed.
482 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
483 * no pages to evict or an eviction handler is not registered, -EAGAIN if
484 * the retry limit was hit.
486 int zbud_reclaim_page(struct zbud_pool
*pool
, unsigned int retries
)
488 int i
, ret
, freechunks
;
489 struct zbud_header
*zhdr
;
490 unsigned long first_handle
= 0, last_handle
= 0;
492 spin_lock(&pool
->lock
);
493 if (!pool
->ops
|| !pool
->ops
->evict
|| list_empty(&pool
->lru
) ||
495 spin_unlock(&pool
->lock
);
498 for (i
= 0; i
< retries
; i
++) {
499 zhdr
= list_tail_entry(&pool
->lru
, struct zbud_header
, lru
);
500 list_del(&zhdr
->lru
);
501 list_del(&zhdr
->buddy
);
502 /* Protect zbud page against free */
503 zhdr
->under_reclaim
= true;
505 * We need encode the handles before unlocking, since we can
506 * race with free that will set (first|last)_chunks to 0
510 if (zhdr
->first_chunks
)
511 first_handle
= encode_handle(zhdr
, FIRST
);
512 if (zhdr
->last_chunks
)
513 last_handle
= encode_handle(zhdr
, LAST
);
514 spin_unlock(&pool
->lock
);
516 /* Issue the eviction callback(s) */
518 ret
= pool
->ops
->evict(pool
, first_handle
);
523 ret
= pool
->ops
->evict(pool
, last_handle
);
528 spin_lock(&pool
->lock
);
529 zhdr
->under_reclaim
= false;
530 if (zhdr
->first_chunks
== 0 && zhdr
->last_chunks
== 0) {
532 * Both buddies are now free, free the zbud page and
535 free_zbud_page(zhdr
);
537 spin_unlock(&pool
->lock
);
539 } else if (zhdr
->first_chunks
== 0 ||
540 zhdr
->last_chunks
== 0) {
541 /* add to unbuddied list */
542 freechunks
= num_free_chunks(zhdr
);
543 list_add(&zhdr
->buddy
, &pool
->unbuddied
[freechunks
]);
545 /* add to buddied list */
546 list_add(&zhdr
->buddy
, &pool
->buddied
);
549 /* add to beginning of LRU */
550 list_add(&zhdr
->lru
, &pool
->lru
);
552 spin_unlock(&pool
->lock
);
557 * zbud_map() - maps the allocation associated with the given handle
558 * @pool: pool in which the allocation resides
559 * @handle: handle associated with the allocation to be mapped
561 * While trivial for zbud, the mapping functions for others allocators
562 * implementing this allocation API could have more complex information encoded
563 * in the handle and could create temporary mappings to make the data
564 * accessible to the user.
566 * Returns: a pointer to the mapped allocation
568 void *zbud_map(struct zbud_pool
*pool
, unsigned long handle
)
570 return (void *)(handle
);
574 * zbud_unmap() - maps the allocation associated with the given handle
575 * @pool: pool in which the allocation resides
576 * @handle: handle associated with the allocation to be unmapped
578 void zbud_unmap(struct zbud_pool
*pool
, unsigned long handle
)
583 * zbud_get_pool_size() - gets the zbud pool size in pages
584 * @pool: pool whose size is being queried
586 * Returns: size in pages of the given pool. The pool lock need not be
587 * taken to access pages_nr.
589 u64
zbud_get_pool_size(struct zbud_pool
*pool
)
591 return pool
->pages_nr
;
594 static int __init
init_zbud(void)
596 /* Make sure the zbud header will fit in one chunk */
597 BUILD_BUG_ON(sizeof(struct zbud_header
) > ZHDR_SIZE_ALIGNED
);
601 zpool_register_driver(&zbud_zpool_driver
);
607 static void __exit
exit_zbud(void)
610 zpool_unregister_driver(&zbud_zpool_driver
);
613 pr_info("unloaded\n");
616 module_init(init_zbud
);
617 module_exit(exit_zbud
);
619 MODULE_LICENSE("GPL");
620 MODULE_AUTHOR("Seth Jennings <sjenning@linux.vnet.ibm.com>");
621 MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");