1/*
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
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.
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>
54#include <linux/zpool.h>
55
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
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.
67 */
68#define NCHUNKS_ORDER 6
69
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)
74
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 const struct zbud_ops *ops;
100#ifdef CONFIG_ZPOOL
101 struct zpool *zpool;
102 const struct zpool_ops *zpool_ops;
103#endif
104};
105
106/*
107 * struct zbud_header - zbud page metadata occupying the first chunk of each
108 * zbud page.
109 * @buddy: links the zbud page into the unbuddied/buddied lists in the pool
110 * @lru: links the zbud page into the lru list in the pool
111 * @first_chunks: the size of the first buddy in chunks, 0 if free
112 * @last_chunks: the size of the last buddy in chunks, 0 if free
113 */
114struct zbud_header {
115 struct list_head buddy;
116 struct list_head lru;
117 unsigned int first_chunks;
118 unsigned int last_chunks;
119 bool under_reclaim;
120};
121
122/*****************
123 * zpool
124 ****************/
125
126#ifdef CONFIG_ZPOOL
127
128static int zbud_zpool_evict(struct zbud_pool *pool, unsigned long handle)
129{
130 if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
131 return pool->zpool_ops->evict(pool->zpool, handle);
132 else
133 return -ENOENT;
134}
135
136static const struct zbud_ops zbud_zpool_ops = {
137 .evict = zbud_zpool_evict
138};
139
140static void *zbud_zpool_create(const char *name, gfp_t gfp,
141 const struct zpool_ops *zpool_ops,
142 struct zpool *zpool)
143{
144 struct zbud_pool *pool;
145
146 pool = zbud_create_pool(gfp, zpool_ops ? &zbud_zpool_ops : NULL);
147 if (pool) {
148 pool->zpool = zpool;
149 pool->zpool_ops = zpool_ops;
150 }
151 return pool;
152}
153
154static void zbud_zpool_destroy(void *pool)
155{
156 zbud_destroy_pool(pool);
157}
158
159static int zbud_zpool_malloc(void *pool, size_t size, gfp_t gfp,
160 unsigned long *handle)
161{
162 return zbud_alloc(pool, size, gfp, handle);
163}
164static void zbud_zpool_free(void *pool, unsigned long handle)
165{
166 zbud_free(pool, handle);
167}
168
169static int zbud_zpool_shrink(void *pool, unsigned int pages,
170 unsigned int *reclaimed)
171{
172 unsigned int total = 0;
173 int ret = -EINVAL;
174
175 while (total < pages) {
176 ret = zbud_reclaim_page(pool, 8);
177 if (ret < 0)
178 break;
179 total++;
180 }
181
182 if (reclaimed)
183 *reclaimed = total;
184
185 return ret;
186}
187
188static void *zbud_zpool_map(void *pool, unsigned long handle,
189 enum zpool_mapmode mm)
190{
191 return zbud_map(pool, handle);
192}
193static void zbud_zpool_unmap(void *pool, unsigned long handle)
194{
195 zbud_unmap(pool, handle);
196}
197
198static u64 zbud_zpool_total_size(void *pool)
199{
200 return zbud_get_pool_size(pool) * PAGE_SIZE;
201}
202
203static struct zpool_driver zbud_zpool_driver = {
204 .type = "zbud",
205 .owner = THIS_MODULE,
206 .create = zbud_zpool_create,
207 .destroy = zbud_zpool_destroy,
208 .malloc = zbud_zpool_malloc,
209 .free = zbud_zpool_free,
210 .shrink = zbud_zpool_shrink,
211 .map = zbud_zpool_map,
212 .unmap = zbud_zpool_unmap,
213 .total_size = zbud_zpool_total_size,
214};
215
216MODULE_ALIAS("zpool-zbud");
217#endif /* CONFIG_ZPOOL */
218
219/*****************
220 * Helpers
221*****************/
222/* Just to make the code easier to read */
223enum buddy {
224 FIRST,
225 LAST
226};
227
228/* Converts an allocation size in bytes to size in zbud chunks */
229static int size_to_chunks(size_t size)
230{
231 return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
232}
233
234#define for_each_unbuddied_list(_iter, _begin) \
235 for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
236
237/* Initializes the zbud header of a newly allocated zbud page */
238static struct zbud_header *init_zbud_page(struct page *page)
239{
240 struct zbud_header *zhdr = page_address(page);
241 zhdr->first_chunks = 0;
242 zhdr->last_chunks = 0;
243 INIT_LIST_HEAD(&zhdr->buddy);
244 INIT_LIST_HEAD(&zhdr->lru);
245 zhdr->under_reclaim = 0;
246 return zhdr;
247}
248
249/* Resets the struct page fields and frees the page */
250static void free_zbud_page(struct zbud_header *zhdr)
251{
252 __free_page(virt_to_page(zhdr));
253}
254
255/*
256 * Encodes the handle of a particular buddy within a zbud page
257 * Pool lock should be held as this function accesses first|last_chunks
258 */
259static unsigned long encode_handle(struct zbud_header *zhdr, enum buddy bud)
260{
261 unsigned long handle;
262
263 /*
264 * For now, the encoded handle is actually just the pointer to the data
265 * but this might not always be the case. A little information hiding.
266 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
267 * over the zbud header in the first chunk.
268 */
269 handle = (unsigned long)zhdr;
270 if (bud == FIRST)
271 /* skip over zbud header */
272 handle += ZHDR_SIZE_ALIGNED;
273 else /* bud == LAST */
274 handle += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
275 return handle;
276}
277
278/* Returns the zbud page where a given handle is stored */
279static struct zbud_header *handle_to_zbud_header(unsigned long handle)
280{
281 return (struct zbud_header *)(handle & PAGE_MASK);
282}
283
284/* Returns the number of free chunks in a zbud page */
285static int num_free_chunks(struct zbud_header *zhdr)
286{
287 /*
288 * Rather than branch for different situations, just use the fact that
289 * free buddies have a length of zero to simplify everything.
290 */
291 return NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
292}
293
294/*****************
295 * API Functions
296*****************/
297/**
298 * zbud_create_pool() - create a new zbud pool
299 * @gfp: gfp flags when allocating the zbud pool structure
300 * @ops: user-defined operations for the zbud pool
301 *
302 * Return: pointer to the new zbud pool or NULL if the metadata allocation
303 * failed.
304 */
305struct zbud_pool *zbud_create_pool(gfp_t gfp, const struct zbud_ops *ops)
306{
307 struct zbud_pool *pool;
308 int i;
309
310 pool = kzalloc(sizeof(struct zbud_pool), gfp);
311 if (!pool)
312 return NULL;
313 spin_lock_init(&pool->lock);
314 for_each_unbuddied_list(i, 0)
315 INIT_LIST_HEAD(&pool->unbuddied[i]);
316 INIT_LIST_HEAD(&pool->buddied);
317 INIT_LIST_HEAD(&pool->lru);
318 pool->pages_nr = 0;
319 pool->ops = ops;
320 return pool;
321}
322
323/**
324 * zbud_destroy_pool() - destroys an existing zbud pool
325 * @pool: the zbud pool to be destroyed
326 *
327 * The pool should be emptied before this function is called.
328 */
329void zbud_destroy_pool(struct zbud_pool *pool)
330{
331 kfree(pool);
332}
333
334/**
335 * zbud_alloc() - allocates a region of a given size
336 * @pool: zbud pool from which to allocate
337 * @size: size in bytes of the desired allocation
338 * @gfp: gfp flags used if the pool needs to grow
339 * @handle: handle of the new allocation
340 *
341 * This function will attempt to find a free region in the pool large enough to
342 * satisfy the allocation request. A search of the unbuddied lists is
343 * performed first. If no suitable free region is found, then a new page is
344 * allocated and added to the pool to satisfy the request.
345 *
346 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
347 * as zbud pool pages.
348 *
349 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
350 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
351 * a new page.
352 */
353int zbud_alloc(struct zbud_pool *pool, size_t size, gfp_t gfp,
354 unsigned long *handle)
355{
356 int chunks, i, freechunks;
357 struct zbud_header *zhdr = NULL;
358 enum buddy bud;
359 struct page *page;
360
361 if (!size || (gfp & __GFP_HIGHMEM))
362 return -EINVAL;
363 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
364 return -ENOSPC;
365 chunks = size_to_chunks(size);
366 spin_lock(&pool->lock);
367
368 /* First, try to find an unbuddied zbud page. */
369 zhdr = NULL;
370 for_each_unbuddied_list(i, chunks) {
371 if (!list_empty(&pool->unbuddied[i])) {
372 zhdr = list_first_entry(&pool->unbuddied[i],
373 struct zbud_header, buddy);
374 list_del(&zhdr->buddy);
375 if (zhdr->first_chunks == 0)
376 bud = FIRST;
377 else
378 bud = LAST;
379 goto found;
380 }
381 }
382
383 /* Couldn't find unbuddied zbud page, create new one */
384 spin_unlock(&pool->lock);
385 page = alloc_page(gfp);
386 if (!page)
387 return -ENOMEM;
388 spin_lock(&pool->lock);
389 pool->pages_nr++;
390 zhdr = init_zbud_page(page);
391 bud = FIRST;
392
393found:
394 if (bud == FIRST)
395 zhdr->first_chunks = chunks;
396 else
397 zhdr->last_chunks = chunks;
398
399 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0) {
400 /* Add to unbuddied list */
401 freechunks = num_free_chunks(zhdr);
402 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
403 } else {
404 /* Add to buddied list */
405 list_add(&zhdr->buddy, &pool->buddied);
406 }
407
408 /* Add/move zbud page to beginning of LRU */
409 if (!list_empty(&zhdr->lru))
410 list_del(&zhdr->lru);
411 list_add(&zhdr->lru, &pool->lru);
412
413 *handle = encode_handle(zhdr, bud);
414 spin_unlock(&pool->lock);
415
416 return 0;
417}
418
419/**
420 * zbud_free() - frees the allocation associated with the given handle
421 * @pool: pool in which the allocation resided
422 * @handle: handle associated with the allocation returned by zbud_alloc()
423 *
424 * In the case that the zbud page in which the allocation resides is under
425 * reclaim, as indicated by the PG_reclaim flag being set, this function
426 * only sets the first|last_chunks to 0. The page is actually freed
427 * once both buddies are evicted (see zbud_reclaim_page() below).
428 */
429void zbud_free(struct zbud_pool *pool, unsigned long handle)
430{
431 struct zbud_header *zhdr;
432 int freechunks;
433
434 spin_lock(&pool->lock);
435 zhdr = handle_to_zbud_header(handle);
436
437 /* If first buddy, handle will be page aligned */
438 if ((handle - ZHDR_SIZE_ALIGNED) & ~PAGE_MASK)
439 zhdr->last_chunks = 0;
440 else
441 zhdr->first_chunks = 0;
442
443 if (zhdr->under_reclaim) {
444 /* zbud page is under reclaim, reclaim will free */
445 spin_unlock(&pool->lock);
446 return;
447 }
448
449 /* Remove from existing buddy list */
450 list_del(&zhdr->buddy);
451
452 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
453 /* zbud page is empty, free */
454 list_del(&zhdr->lru);
455 free_zbud_page(zhdr);
456 pool->pages_nr--;
457 } else {
458 /* Add to unbuddied list */
459 freechunks = num_free_chunks(zhdr);
460 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
461 }
462
463 spin_unlock(&pool->lock);
464}
465
466/**
467 * zbud_reclaim_page() - evicts allocations from a pool page and frees it
468 * @pool: pool from which a page will attempt to be evicted
469 * @retries: number of pages on the LRU list for which eviction will
470 * be attempted before failing
471 *
472 * zbud reclaim is different from normal system reclaim in that the reclaim is
473 * done from the bottom, up. This is because only the bottom layer, zbud, has
474 * information on how the allocations are organized within each zbud page. This
475 * has the potential to create interesting locking situations between zbud and
476 * the user, however.
477 *
478 * To avoid these, this is how zbud_reclaim_page() should be called:
479 *
480 * The user detects a page should be reclaimed and calls zbud_reclaim_page().
481 * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
482 * the user-defined eviction handler with the pool and handle as arguments.
483 *
484 * If the handle can not be evicted, the eviction handler should return
485 * non-zero. zbud_reclaim_page() will add the zbud page back to the
486 * appropriate list and try the next zbud page on the LRU up to
487 * a user defined number of retries.
488 *
489 * If the handle is successfully evicted, the eviction handler should
490 * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
491 * contains logic to delay freeing the page if the page is under reclaim,
492 * as indicated by the setting of the PG_reclaim flag on the underlying page.
493 *
494 * If all buddies in the zbud page are successfully evicted, then the
495 * zbud page can be freed.
496 *
497 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
498 * no pages to evict or an eviction handler is not registered, -EAGAIN if
499 * the retry limit was hit.
500 */
501int zbud_reclaim_page(struct zbud_pool *pool, unsigned int retries)
502{
503 int i, ret, freechunks;
504 struct zbud_header *zhdr;
505 unsigned long first_handle = 0, last_handle = 0;
506
507 spin_lock(&pool->lock);
508 if (!pool->ops || !pool->ops->evict || list_empty(&pool->lru) ||
509 retries == 0) {
510 spin_unlock(&pool->lock);
511 return -EINVAL;
512 }
513 for (i = 0; i < retries; i++) {
514 zhdr = list_last_entry(&pool->lru, struct zbud_header, lru);
515 list_del(&zhdr->lru);
516 list_del(&zhdr->buddy);
517 /* Protect zbud page against free */
518 zhdr->under_reclaim = true;
519 /*
520 * We need encode the handles before unlocking, since we can
521 * race with free that will set (first|last)_chunks to 0
522 */
523 first_handle = 0;
524 last_handle = 0;
525 if (zhdr->first_chunks)
526 first_handle = encode_handle(zhdr, FIRST);
527 if (zhdr->last_chunks)
528 last_handle = encode_handle(zhdr, LAST);
529 spin_unlock(&pool->lock);
530
531 /* Issue the eviction callback(s) */
532 if (first_handle) {
533 ret = pool->ops->evict(pool, first_handle);
534 if (ret)
535 goto next;
536 }
537 if (last_handle) {
538 ret = pool->ops->evict(pool, last_handle);
539 if (ret)
540 goto next;
541 }
542next:
543 spin_lock(&pool->lock);
544 zhdr->under_reclaim = false;
545 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
546 /*
547 * Both buddies are now free, free the zbud page and
548 * return success.
549 */
550 free_zbud_page(zhdr);
551 pool->pages_nr--;
552 spin_unlock(&pool->lock);
553 return 0;
554 } else if (zhdr->first_chunks == 0 ||
555 zhdr->last_chunks == 0) {
556 /* add to unbuddied list */
557 freechunks = num_free_chunks(zhdr);
558 list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
559 } else {
560 /* add to buddied list */
561 list_add(&zhdr->buddy, &pool->buddied);
562 }
563
564 /* add to beginning of LRU */
565 list_add(&zhdr->lru, &pool->lru);
566 }
567 spin_unlock(&pool->lock);
568 return -EAGAIN;
569}
570
571/**
572 * zbud_map() - maps the allocation associated with the given handle
573 * @pool: pool in which the allocation resides
574 * @handle: handle associated with the allocation to be mapped
575 *
576 * While trivial for zbud, the mapping functions for others allocators
577 * implementing this allocation API could have more complex information encoded
578 * in the handle and could create temporary mappings to make the data
579 * accessible to the user.
580 *
581 * Returns: a pointer to the mapped allocation
582 */
583void *zbud_map(struct zbud_pool *pool, unsigned long handle)
584{
585 return (void *)(handle);
586}
587
588/**
589 * zbud_unmap() - maps the allocation associated with the given handle
590 * @pool: pool in which the allocation resides
591 * @handle: handle associated with the allocation to be unmapped
592 */
593void zbud_unmap(struct zbud_pool *pool, unsigned long handle)
594{
595}
596
597/**
598 * zbud_get_pool_size() - gets the zbud pool size in pages
599 * @pool: pool whose size is being queried
600 *
601 * Returns: size in pages of the given pool. The pool lock need not be
602 * taken to access pages_nr.
603 */
604u64 zbud_get_pool_size(struct zbud_pool *pool)
605{
606 return pool->pages_nr;
607}
608
609static int __init init_zbud(void)
610{
611 /* Make sure the zbud header will fit in one chunk */
612 BUILD_BUG_ON(sizeof(struct zbud_header) > ZHDR_SIZE_ALIGNED);
613 pr_info("loaded\n");
614
615#ifdef CONFIG_ZPOOL
616 zpool_register_driver(&zbud_zpool_driver);
617#endif
618
619 return 0;
620}
621
622static void __exit exit_zbud(void)
623{
624#ifdef CONFIG_ZPOOL
625 zpool_unregister_driver(&zbud_zpool_driver);
626#endif
627
628 pr_info("unloaded\n");
629}
630
631module_init(init_zbud);
632module_exit(exit_zbud);
633
634MODULE_LICENSE("GPL");
635MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
636MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");
637