blob: 2ee4e45204936e20f50deaad8b37f857d7bea630 [file] [log] [blame]
Seth Jennings4e2e2772013-07-10 16:04:55 -07001/*
2 * zbud.c
3 *
4 * Copyright (C) 2013, Seth Jennings, IBM
5 *
6 * Concepts based on zcache internal zbud allocator by Dan Magenheimer.
7 *
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
11 * page.
12 *
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.
16 *
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
Jianguo Wueee87e12013-09-11 14:21:42 -070019 * justified" at the beginning of the zbud page, and the last buddy is "right
Seth Jennings4e2e2772013-07-10 16:04:55 -070020 * 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.
24 *
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.
29 *
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
34 * zbud page.
35 *
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.
42 */
43
44#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
45
46#include <linux/atomic.h>
47#include <linux/list.h>
48#include <linux/mm.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>
Dan Streetmanc7957792014-08-06 16:08:38 -070054#include <linux/zpool.h>
Seth Jennings4e2e2772013-07-10 16:04:55 -070055
56/*****************
57 * Structures
58*****************/
59/*
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
Chao Yuf203c3b2014-10-09 15:30:04 -070063 * 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.
Seth Jennings4e2e2772013-07-10 16:04:55 -070067 */
68#define NCHUNKS_ORDER 6
69
70#define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
71#define CHUNK_SIZE (1 << CHUNK_SHIFT)
Seth Jennings4e2e2772013-07-10 16:04:55 -070072#define ZHDR_SIZE_ALIGNED CHUNK_SIZE
Chao Yuf203c3b2014-10-09 15:30:04 -070073#define NCHUNKS ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
Seth Jennings4e2e2772013-07-10 16:04:55 -070074
75/**
76 * struct zbud_pool - stores metadata for each zbud pool
77 * @lock: protects all pool fields and first|last_chunk fields of any
78 * zbud page in the pool
79 * @unbuddied: array of lists tracking zbud pages that only contain one buddy;
80 * the lists each zbud page is added to depends on the size of
81 * its free region.
82 * @buddied: list tracking the zbud pages that contain two buddies;
83 * these zbud pages are full
84 * @lru: list tracking the zbud pages in LRU order by most recently
85 * added buddy.
86 * @pages_nr: number of zbud pages in the pool.
87 * @ops: pointer to a structure of user defined operations specified at
88 * pool creation time.
89 *
90 * This structure is allocated at pool creation time and maintains metadata
91 * pertaining to a particular zbud pool.
92 */
93struct zbud_pool {
94 spinlock_t lock;
95 struct list_head unbuddied[NCHUNKS];
96 struct list_head buddied;
97 struct list_head lru;
98 u64 pages_nr;
99 struct zbud_ops *ops;
100};
101
102/*
103 * struct zbud_header - zbud page metadata occupying the first chunk of each
104 * zbud page.
105 * @buddy: links the zbud page into the unbuddied/buddied lists in the pool
106 * @lru: links the zbud page into the lru list in the pool
107 * @first_chunks: the size of the first buddy in chunks, 0 if free
108 * @last_chunks: the size of the last buddy in chunks, 0 if free
109 */
110struct zbud_header {
111 struct list_head buddy;
112 struct list_head lru;
113 unsigned int first_chunks;
114 unsigned int last_chunks;
115 bool under_reclaim;
116};
117
118/*****************
Dan Streetmanc7957792014-08-06 16:08:38 -0700119 * zpool
120 ****************/
121
122#ifdef CONFIG_ZPOOL
123
124static int zbud_zpool_evict(struct zbud_pool *pool, unsigned long handle)
125{
126 return zpool_evict(pool, handle);
127}
128
129static struct zbud_ops zbud_zpool_ops = {
130 .evict = zbud_zpool_evict
131};
132
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -0800133static void *zbud_zpool_create(char *name, gfp_t gfp,
134 struct zpool_ops *zpool_ops)
Dan Streetmanc7957792014-08-06 16:08:38 -0700135{
Heesub Shin1dd61aa2014-12-12 16:57:21 -0800136 return zbud_create_pool(gfp, zpool_ops ? &zbud_zpool_ops : NULL);
Dan Streetmanc7957792014-08-06 16:08:38 -0700137}
138
139static void zbud_zpool_destroy(void *pool)
140{
141 zbud_destroy_pool(pool);
142}
143
144static int zbud_zpool_malloc(void *pool, size_t size, gfp_t gfp,
145 unsigned long *handle)
146{
147 return zbud_alloc(pool, size, gfp, handle);
148}
149static void zbud_zpool_free(void *pool, unsigned long handle)
150{
151 zbud_free(pool, handle);
152}
153
154static int zbud_zpool_shrink(void *pool, unsigned int pages,
155 unsigned int *reclaimed)
156{
157 unsigned int total = 0;
158 int ret = -EINVAL;
159
160 while (total < pages) {
161 ret = zbud_reclaim_page(pool, 8);
162 if (ret < 0)
163 break;
164 total++;
165 }
166
167 if (reclaimed)
168 *reclaimed = total;
169
170 return ret;
171}
172
173static void *zbud_zpool_map(void *pool, unsigned long handle,
174 enum zpool_mapmode mm)
175{
176 return zbud_map(pool, handle);
177}
178static void zbud_zpool_unmap(void *pool, unsigned long handle)
179{
180 zbud_unmap(pool, handle);
181}
182
183static u64 zbud_zpool_total_size(void *pool)
184{
185 return zbud_get_pool_size(pool) * PAGE_SIZE;
186}
187
188static struct zpool_driver zbud_zpool_driver = {
189 .type = "zbud",
190 .owner = THIS_MODULE,
191 .create = zbud_zpool_create,
192 .destroy = zbud_zpool_destroy,
193 .malloc = zbud_zpool_malloc,
194 .free = zbud_zpool_free,
195 .shrink = zbud_zpool_shrink,
196 .map = zbud_zpool_map,
197 .unmap = zbud_zpool_unmap,
198 .total_size = zbud_zpool_total_size,
199};
200
Kees Cook137f8cf2014-08-29 15:18:40 -0700201MODULE_ALIAS("zpool-zbud");
Dan Streetmanc7957792014-08-06 16:08:38 -0700202#endif /* CONFIG_ZPOOL */
203
204/*****************
Seth Jennings4e2e2772013-07-10 16:04:55 -0700205 * Helpers
206*****************/
207/* Just to make the code easier to read */
208enum buddy {
209 FIRST,
210 LAST
211};
212
213/* Converts an allocation size in bytes to size in zbud chunks */
Dan Streetman99eef8e2014-08-06 16:08:33 -0700214static int size_to_chunks(size_t size)
Seth Jennings4e2e2772013-07-10 16:04:55 -0700215{
216 return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
217}
218
219#define for_each_unbuddied_list(_iter, _begin) \
220 for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
221
222/* Initializes the zbud header of a newly allocated zbud page */
223static struct zbud_header *init_zbud_page(struct page *page)
224{
225 struct zbud_header *zhdr = page_address(page);
226 zhdr->first_chunks = 0;
227 zhdr->last_chunks = 0;
228 INIT_LIST_HEAD(&zhdr->buddy);
229 INIT_LIST_HEAD(&zhdr->lru);
230 zhdr->under_reclaim = 0;
231 return zhdr;
232}
233
234/* Resets the struct page fields and frees the page */
235static void free_zbud_page(struct zbud_header *zhdr)
236{
237 __free_page(virt_to_page(zhdr));
238}
239
240/*
241 * Encodes the handle of a particular buddy within a zbud page
242 * Pool lock should be held as this function accesses first|last_chunks
243 */
244static unsigned long encode_handle(struct zbud_header *zhdr, enum buddy bud)
245{
246 unsigned long handle;
247
248 /*
249 * For now, the encoded handle is actually just the pointer to the data
250 * but this might not always be the case. A little information hiding.
251 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
252 * over the zbud header in the first chunk.
253 */
254 handle = (unsigned long)zhdr;
255 if (bud == FIRST)
256 /* skip over zbud header */
257 handle += ZHDR_SIZE_ALIGNED;
258 else /* bud == LAST */
259 handle += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
260 return handle;
261}
262
263/* Returns the zbud page where a given handle is stored */
264static struct zbud_header *handle_to_zbud_header(unsigned long handle)
265{
266 return (struct zbud_header *)(handle & PAGE_MASK);
267}
268
269/* Returns the number of free chunks in a zbud page */
270static int num_free_chunks(struct zbud_header *zhdr)
271{
272 /*
273 * Rather than branch for different situations, just use the fact that
Chao Yuf203c3b2014-10-09 15:30:04 -0700274 * free buddies have a length of zero to simplify everything.
Seth Jennings4e2e2772013-07-10 16:04:55 -0700275 */
Chao Yuf203c3b2014-10-09 15:30:04 -0700276 return NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
Seth Jennings4e2e2772013-07-10 16:04:55 -0700277}
278
279/*****************
280 * API Functions
281*****************/
282/**
283 * zbud_create_pool() - create a new zbud pool
284 * @gfp: gfp flags when allocating the zbud pool structure
285 * @ops: user-defined operations for the zbud pool
286 *
287 * Return: pointer to the new zbud pool or NULL if the metadata allocation
288 * failed.
289 */
290struct zbud_pool *zbud_create_pool(gfp_t gfp, struct zbud_ops *ops)
291{
292 struct zbud_pool *pool;
293 int i;
294
295 pool = kmalloc(sizeof(struct zbud_pool), gfp);
296 if (!pool)
297 return NULL;
298 spin_lock_init(&pool->lock);
299 for_each_unbuddied_list(i, 0)
300 INIT_LIST_HEAD(&pool->unbuddied[i]);
301 INIT_LIST_HEAD(&pool->buddied);
302 INIT_LIST_HEAD(&pool->lru);
303 pool->pages_nr = 0;
304 pool->ops = ops;
305 return pool;
306}
307
308/**
309 * zbud_destroy_pool() - destroys an existing zbud pool
310 * @pool: the zbud pool to be destroyed
311 *
312 * The pool should be emptied before this function is called.
313 */
314void zbud_destroy_pool(struct zbud_pool *pool)
315{
316 kfree(pool);
317}
318
319/**
320 * zbud_alloc() - allocates a region of a given size
321 * @pool: zbud pool from which to allocate
322 * @size: size in bytes of the desired allocation
323 * @gfp: gfp flags used if the pool needs to grow
324 * @handle: handle of the new allocation
325 *
326 * This function will attempt to find a free region in the pool large enough to
327 * satisfy the allocation request. A search of the unbuddied lists is
328 * performed first. If no suitable free region is found, then a new page is
329 * allocated and added to the pool to satisfy the request.
330 *
331 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
332 * as zbud pool pages.
333 *
Jianguo Wueee87e12013-09-11 14:21:42 -0700334 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
Seth Jennings4e2e2772013-07-10 16:04:55 -0700335 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
336 * a new page.
337 */
Dan Streetman99eef8e2014-08-06 16:08:33 -0700338int zbud_alloc(struct zbud_pool *pool, size_t size, gfp_t gfp,
Seth Jennings4e2e2772013-07-10 16:04:55 -0700339 unsigned long *handle)
340{
341 int chunks, i, freechunks;
342 struct zbud_header *zhdr = NULL;
343 enum buddy bud;
344 struct page *page;
345
Fabian Frederick50417c52014-06-04 16:11:07 -0700346 if (!size || (gfp & __GFP_HIGHMEM))
Seth Jennings4e2e2772013-07-10 16:04:55 -0700347 return -EINVAL;
Heesub Shin9d8c5b52013-07-31 13:53:40 -0700348 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
Seth Jennings4e2e2772013-07-10 16:04:55 -0700349 return -ENOSPC;
350 chunks = size_to_chunks(size);
351 spin_lock(&pool->lock);
352
353 /* First, try to find an unbuddied zbud page. */
354 zhdr = NULL;
355 for_each_unbuddied_list(i, chunks) {
356 if (!list_empty(&pool->unbuddied[i])) {
357 zhdr = list_first_entry(&pool->unbuddied[i],
358 struct zbud_header, buddy);
359 list_del(&zhdr->buddy);
360 if (zhdr->first_chunks == 0)
361 bud = FIRST;
362 else
363 bud = LAST;
364 goto found;
365 }
366 }
367
368 /* Couldn't find unbuddied zbud page, create new one */
369 spin_unlock(&pool->lock);
370 page = alloc_page(gfp);
371 if (!page)
372 return -ENOMEM;
373 spin_lock(&pool->lock);
374 pool->pages_nr++;
375 zhdr = init_zbud_page(page);
376 bud = FIRST;
377
378found:
379 if (bud == FIRST)
380 zhdr->first_chunks = chunks;
381 else
382 zhdr->last_chunks = chunks;
383
384 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0) {
385 /* Add to unbuddied list */
386 freechunks = num_free_chunks(zhdr);
387 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
388 } else {
389 /* Add to buddied list */
390 list_add(&zhdr->buddy, &pool->buddied);
391 }
392
393 /* Add/move zbud page to beginning of LRU */
394 if (!list_empty(&zhdr->lru))
395 list_del(&zhdr->lru);
396 list_add(&zhdr->lru, &pool->lru);
397
398 *handle = encode_handle(zhdr, bud);
399 spin_unlock(&pool->lock);
400
401 return 0;
402}
403
404/**
405 * zbud_free() - frees the allocation associated with the given handle
406 * @pool: pool in which the allocation resided
407 * @handle: handle associated with the allocation returned by zbud_alloc()
408 *
409 * In the case that the zbud page in which the allocation resides is under
410 * reclaim, as indicated by the PG_reclaim flag being set, this function
411 * only sets the first|last_chunks to 0. The page is actually freed
412 * once both buddies are evicted (see zbud_reclaim_page() below).
413 */
414void zbud_free(struct zbud_pool *pool, unsigned long handle)
415{
416 struct zbud_header *zhdr;
417 int freechunks;
418
419 spin_lock(&pool->lock);
420 zhdr = handle_to_zbud_header(handle);
421
422 /* If first buddy, handle will be page aligned */
423 if ((handle - ZHDR_SIZE_ALIGNED) & ~PAGE_MASK)
424 zhdr->last_chunks = 0;
425 else
426 zhdr->first_chunks = 0;
427
428 if (zhdr->under_reclaim) {
429 /* zbud page is under reclaim, reclaim will free */
430 spin_unlock(&pool->lock);
431 return;
432 }
433
434 /* Remove from existing buddy list */
435 list_del(&zhdr->buddy);
436
437 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
438 /* zbud page is empty, free */
439 list_del(&zhdr->lru);
440 free_zbud_page(zhdr);
441 pool->pages_nr--;
442 } else {
443 /* Add to unbuddied list */
444 freechunks = num_free_chunks(zhdr);
445 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
446 }
447
448 spin_unlock(&pool->lock);
449}
450
451#define list_tail_entry(ptr, type, member) \
452 list_entry((ptr)->prev, type, member)
453
454/**
455 * zbud_reclaim_page() - evicts allocations from a pool page and frees it
456 * @pool: pool from which a page will attempt to be evicted
457 * @retires: number of pages on the LRU list for which eviction will
458 * be attempted before failing
459 *
460 * zbud reclaim is different from normal system reclaim in that the reclaim is
461 * done from the bottom, up. This is because only the bottom layer, zbud, has
462 * information on how the allocations are organized within each zbud page. This
463 * has the potential to create interesting locking situations between zbud and
464 * the user, however.
465 *
466 * To avoid these, this is how zbud_reclaim_page() should be called:
467
468 * The user detects a page should be reclaimed and calls zbud_reclaim_page().
469 * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
470 * the user-defined eviction handler with the pool and handle as arguments.
471 *
472 * If the handle can not be evicted, the eviction handler should return
473 * non-zero. zbud_reclaim_page() will add the zbud page back to the
474 * appropriate list and try the next zbud page on the LRU up to
475 * a user defined number of retries.
476 *
477 * If the handle is successfully evicted, the eviction handler should
478 * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
479 * contains logic to delay freeing the page if the page is under reclaim,
480 * as indicated by the setting of the PG_reclaim flag on the underlying page.
481 *
482 * If all buddies in the zbud page are successfully evicted, then the
483 * zbud page can be freed.
484 *
485 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
486 * no pages to evict or an eviction handler is not registered, -EAGAIN if
487 * the retry limit was hit.
488 */
489int zbud_reclaim_page(struct zbud_pool *pool, unsigned int retries)
490{
491 int i, ret, freechunks;
492 struct zbud_header *zhdr;
493 unsigned long first_handle = 0, last_handle = 0;
494
495 spin_lock(&pool->lock);
496 if (!pool->ops || !pool->ops->evict || list_empty(&pool->lru) ||
497 retries == 0) {
498 spin_unlock(&pool->lock);
499 return -EINVAL;
500 }
501 for (i = 0; i < retries; i++) {
502 zhdr = list_tail_entry(&pool->lru, struct zbud_header, lru);
503 list_del(&zhdr->lru);
504 list_del(&zhdr->buddy);
505 /* Protect zbud page against free */
506 zhdr->under_reclaim = true;
507 /*
508 * We need encode the handles before unlocking, since we can
509 * race with free that will set (first|last)_chunks to 0
510 */
511 first_handle = 0;
512 last_handle = 0;
513 if (zhdr->first_chunks)
514 first_handle = encode_handle(zhdr, FIRST);
515 if (zhdr->last_chunks)
516 last_handle = encode_handle(zhdr, LAST);
517 spin_unlock(&pool->lock);
518
519 /* Issue the eviction callback(s) */
520 if (first_handle) {
521 ret = pool->ops->evict(pool, first_handle);
522 if (ret)
523 goto next;
524 }
525 if (last_handle) {
526 ret = pool->ops->evict(pool, last_handle);
527 if (ret)
528 goto next;
529 }
530next:
531 spin_lock(&pool->lock);
532 zhdr->under_reclaim = false;
533 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
534 /*
535 * Both buddies are now free, free the zbud page and
536 * return success.
537 */
538 free_zbud_page(zhdr);
539 pool->pages_nr--;
540 spin_unlock(&pool->lock);
541 return 0;
542 } else if (zhdr->first_chunks == 0 ||
543 zhdr->last_chunks == 0) {
544 /* add to unbuddied list */
545 freechunks = num_free_chunks(zhdr);
546 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
547 } else {
548 /* add to buddied list */
549 list_add(&zhdr->buddy, &pool->buddied);
550 }
551
552 /* add to beginning of LRU */
553 list_add(&zhdr->lru, &pool->lru);
554 }
555 spin_unlock(&pool->lock);
556 return -EAGAIN;
557}
558
559/**
560 * zbud_map() - maps the allocation associated with the given handle
561 * @pool: pool in which the allocation resides
562 * @handle: handle associated with the allocation to be mapped
563 *
564 * While trivial for zbud, the mapping functions for others allocators
565 * implementing this allocation API could have more complex information encoded
566 * in the handle and could create temporary mappings to make the data
567 * accessible to the user.
568 *
569 * Returns: a pointer to the mapped allocation
570 */
571void *zbud_map(struct zbud_pool *pool, unsigned long handle)
572{
573 return (void *)(handle);
574}
575
576/**
577 * zbud_unmap() - maps the allocation associated with the given handle
578 * @pool: pool in which the allocation resides
579 * @handle: handle associated with the allocation to be unmapped
580 */
581void zbud_unmap(struct zbud_pool *pool, unsigned long handle)
582{
583}
584
585/**
586 * zbud_get_pool_size() - gets the zbud pool size in pages
587 * @pool: pool whose size is being queried
588 *
589 * Returns: size in pages of the given pool. The pool lock need not be
590 * taken to access pages_nr.
591 */
592u64 zbud_get_pool_size(struct zbud_pool *pool)
593{
594 return pool->pages_nr;
595}
596
597static int __init init_zbud(void)
598{
599 /* Make sure the zbud header will fit in one chunk */
600 BUILD_BUG_ON(sizeof(struct zbud_header) > ZHDR_SIZE_ALIGNED);
601 pr_info("loaded\n");
Dan Streetmanc7957792014-08-06 16:08:38 -0700602
603#ifdef CONFIG_ZPOOL
604 zpool_register_driver(&zbud_zpool_driver);
605#endif
606
Seth Jennings4e2e2772013-07-10 16:04:55 -0700607 return 0;
608}
609
610static void __exit exit_zbud(void)
611{
Dan Streetmanc7957792014-08-06 16:08:38 -0700612#ifdef CONFIG_ZPOOL
613 zpool_unregister_driver(&zbud_zpool_driver);
614#endif
615
Seth Jennings4e2e2772013-07-10 16:04:55 -0700616 pr_info("unloaded\n");
617}
618
619module_init(init_zbud);
620module_exit(exit_zbud);
621
622MODULE_LICENSE("GPL");
Seth Jennings68386da2014-11-12 21:08:46 -0600623MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
Seth Jennings4e2e2772013-07-10 16:04:55 -0700624MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");