1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * zbud.c
  4 *
  5 * Copyright (C) 2013, Seth Jennings, IBM
  6 *
  7 * Concepts based on zcache internal zbud allocator by Dan Magenheimer.
  8 *
  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
 12 * page.
 13 *
 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.
 17 *
 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.
 25 *
 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.
 30 *
 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
 35 * zbud page.
 36 *
 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.
 43 */
 44
 45#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
 46
 47#include <linux/atomic.h>
 48#include <linux/list.h>
 49#include <linux/mm.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>
 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
 75struct zbud_pool;
 76
 77/**
 78 * struct zbud_pool - stores metadata for each zbud pool
 79 * @lock:	protects all pool fields and first|last_chunk fields of any
 80 *		zbud page in the pool
 81 * @unbuddied:	array of lists tracking zbud pages that only contain one buddy;
 82 *		the lists each zbud page is added to depends on the size of
 83 *		its free region.
 84 * @buddied:	list tracking the zbud pages that contain two buddies;
 85 *		these zbud pages are full
 86 * @lru:	list tracking the zbud pages in LRU order by most recently
 87 *		added buddy.
 88 * @pages_nr:	number of zbud pages in the pool.
 89 * @zpool:	zpool driver
 90 * @zpool_ops:	zpool operations structure with an evict callback
 91 *
 92 * This structure is allocated at pool creation time and maintains metadata
 93 * pertaining to a particular zbud pool.
 94 */
 95struct zbud_pool {
 96	spinlock_t lock;
 97	union {
 98		/*
 99		 * Reuse unbuddied[0] as buddied on the ground that
100		 * unbuddied[0] is unused.
101		 */
102		struct list_head buddied;
103		struct list_head unbuddied[NCHUNKS];
104	};
105	struct list_head lru;
106	u64 pages_nr;
107	struct zpool *zpool;
108	const struct zpool_ops *zpool_ops;
109};
110
111/*
112 * struct zbud_header - zbud page metadata occupying the first chunk of each
113 *			zbud page.
114 * @buddy:	links the zbud page into the unbuddied/buddied lists in the pool
115 * @lru:	links the zbud page into the lru list in the pool
116 * @first_chunks:	the size of the first buddy in chunks, 0 if free
117 * @last_chunks:	the size of the last buddy in chunks, 0 if free
118 */
119struct zbud_header {
120	struct list_head buddy;
121	struct list_head lru;
122	unsigned int first_chunks;
123	unsigned int last_chunks;
124	bool under_reclaim;
125};
126
127/*****************
128 * Helpers
129*****************/
130/* Just to make the code easier to read */
131enum buddy {
132	FIRST,
133	LAST
134};
135
136/* Converts an allocation size in bytes to size in zbud chunks */
137static int size_to_chunks(size_t size)
138{
139	return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
140}
141
142#define for_each_unbuddied_list(_iter, _begin) \
143	for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
144
145/* Initializes the zbud header of a newly allocated zbud page */
146static struct zbud_header *init_zbud_page(struct page *page)
147{
148	struct zbud_header *zhdr = page_address(page);
149	zhdr->first_chunks = 0;
150	zhdr->last_chunks = 0;
151	INIT_LIST_HEAD(&zhdr->buddy);
152	INIT_LIST_HEAD(&zhdr->lru);
153	zhdr->under_reclaim = false;
154	return zhdr;
155}
156
157/* Resets the struct page fields and frees the page */
158static void free_zbud_page(struct zbud_header *zhdr)
159{
160	__free_page(virt_to_page(zhdr));
161}
162
163/*
164 * Encodes the handle of a particular buddy within a zbud page
165 * Pool lock should be held as this function accesses first|last_chunks
166 */
167static unsigned long encode_handle(struct zbud_header *zhdr, enum buddy bud)
168{
169	unsigned long handle;
170
171	/*
172	 * For now, the encoded handle is actually just the pointer to the data
173	 * but this might not always be the case.  A little information hiding.
174	 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
175	 * over the zbud header in the first chunk.
176	 */
177	handle = (unsigned long)zhdr;
178	if (bud == FIRST)
179		/* skip over zbud header */
180		handle += ZHDR_SIZE_ALIGNED;
181	else /* bud == LAST */
182		handle += PAGE_SIZE - (zhdr->last_chunks  << CHUNK_SHIFT);
183	return handle;
184}
185
186/* Returns the zbud page where a given handle is stored */
187static struct zbud_header *handle_to_zbud_header(unsigned long handle)
188{
189	return (struct zbud_header *)(handle & PAGE_MASK);
190}
191
192/* Returns the number of free chunks in a zbud page */
193static int num_free_chunks(struct zbud_header *zhdr)
194{
195	/*
196	 * Rather than branch for different situations, just use the fact that
197	 * free buddies have a length of zero to simplify everything.
198	 */
199	return NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
200}
201
202/*****************
203 * API Functions
204*****************/
205/**
206 * zbud_create_pool() - create a new zbud pool
207 * @gfp:	gfp flags when allocating the zbud pool structure
208 *
209 * Return: pointer to the new zbud pool or NULL if the metadata allocation
210 * failed.
211 */
212static struct zbud_pool *zbud_create_pool(gfp_t gfp)
213{
214	struct zbud_pool *pool;
215	int i;
216
217	pool = kzalloc(sizeof(struct zbud_pool), gfp);
218	if (!pool)
219		return NULL;
220	spin_lock_init(&pool->lock);
221	for_each_unbuddied_list(i, 0)
222		INIT_LIST_HEAD(&pool->unbuddied[i]);
223	INIT_LIST_HEAD(&pool->buddied);
224	INIT_LIST_HEAD(&pool->lru);
225	pool->pages_nr = 0;
226	return pool;
227}
228
229/**
230 * zbud_destroy_pool() - destroys an existing zbud pool
231 * @pool:	the zbud pool to be destroyed
232 *
233 * The pool should be emptied before this function is called.
234 */
235static void zbud_destroy_pool(struct zbud_pool *pool)
236{
237	kfree(pool);
238}
239
240/**
241 * zbud_alloc() - allocates a region of a given size
242 * @pool:	zbud pool from which to allocate
243 * @size:	size in bytes of the desired allocation
244 * @gfp:	gfp flags used if the pool needs to grow
245 * @handle:	handle of the new allocation
246 *
247 * This function will attempt to find a free region in the pool large enough to
248 * satisfy the allocation request.  A search of the unbuddied lists is
249 * performed first. If no suitable free region is found, then a new page is
250 * allocated and added to the pool to satisfy the request.
251 *
252 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
253 * as zbud pool pages.
254 *
255 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
256 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
257 * a new page.
258 */
259static int zbud_alloc(struct zbud_pool *pool, size_t size, gfp_t gfp,
260			unsigned long *handle)
261{
262	int chunks, i, freechunks;
263	struct zbud_header *zhdr = NULL;
264	enum buddy bud;
265	struct page *page;
266
267	if (!size || (gfp & __GFP_HIGHMEM))
268		return -EINVAL;
269	if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
270		return -ENOSPC;
271	chunks = size_to_chunks(size);
272	spin_lock(&pool->lock);
273
274	/* First, try to find an unbuddied zbud page. */
275	for_each_unbuddied_list(i, chunks) {
276		if (!list_empty(&pool->unbuddied[i])) {
277			zhdr = list_first_entry(&pool->unbuddied[i],
278					struct zbud_header, buddy);
279			list_del(&zhdr->buddy);
280			if (zhdr->first_chunks == 0)
281				bud = FIRST;
282			else
283				bud = LAST;
284			goto found;
285		}
286	}
287
288	/* Couldn't find unbuddied zbud page, create new one */
289	spin_unlock(&pool->lock);
290	page = alloc_page(gfp);
291	if (!page)
292		return -ENOMEM;
293	spin_lock(&pool->lock);
294	pool->pages_nr++;
295	zhdr = init_zbud_page(page);
296	bud = FIRST;
297
298found:
299	if (bud == FIRST)
300		zhdr->first_chunks = chunks;
301	else
302		zhdr->last_chunks = chunks;
303
304	if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0) {
305		/* Add to unbuddied list */
306		freechunks = num_free_chunks(zhdr);
307		list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
308	} else {
309		/* Add to buddied list */
310		list_add(&zhdr->buddy, &pool->buddied);
311	}
312
313	/* Add/move zbud page to beginning of LRU */
314	if (!list_empty(&zhdr->lru))
315		list_del(&zhdr->lru);
316	list_add(&zhdr->lru, &pool->lru);
317
318	*handle = encode_handle(zhdr, bud);
319	spin_unlock(&pool->lock);
320
321	return 0;
322}
323
324/**
325 * zbud_free() - frees the allocation associated with the given handle
326 * @pool:	pool in which the allocation resided
327 * @handle:	handle associated with the allocation returned by zbud_alloc()
328 *
329 * In the case that the zbud page in which the allocation resides is under
330 * reclaim, as indicated by the PG_reclaim flag being set, this function
331 * only sets the first|last_chunks to 0.  The page is actually freed
332 * once both buddies are evicted (see zbud_reclaim_page() below).
333 */
334static void zbud_free(struct zbud_pool *pool, unsigned long handle)
335{
336	struct zbud_header *zhdr;
337	int freechunks;
338
339	spin_lock(&pool->lock);
340	zhdr = handle_to_zbud_header(handle);
341
342	/* If first buddy, handle will be page aligned */
343	if ((handle - ZHDR_SIZE_ALIGNED) & ~PAGE_MASK)
344		zhdr->last_chunks = 0;
345	else
346		zhdr->first_chunks = 0;
347
348	if (zhdr->under_reclaim) {
349		/* zbud page is under reclaim, reclaim will free */
350		spin_unlock(&pool->lock);
351		return;
352	}
353
354	/* Remove from existing buddy list */
355	list_del(&zhdr->buddy);
356
357	if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
358		/* zbud page is empty, free */
359		list_del(&zhdr->lru);
360		free_zbud_page(zhdr);
361		pool->pages_nr--;
362	} else {
363		/* Add to unbuddied list */
364		freechunks = num_free_chunks(zhdr);
365		list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
366	}
367
368	spin_unlock(&pool->lock);
369}
370
371/**
372 * zbud_reclaim_page() - evicts allocations from a pool page and frees it
373 * @pool:	pool from which a page will attempt to be evicted
374 * @retries:	number of pages on the LRU list for which eviction will
375 *		be attempted before failing
376 *
377 * zbud reclaim is different from normal system reclaim in that the reclaim is
378 * done from the bottom, up.  This is because only the bottom layer, zbud, has
379 * information on how the allocations are organized within each zbud page. This
380 * has the potential to create interesting locking situations between zbud and
381 * the user, however.
382 *
383 * To avoid these, this is how zbud_reclaim_page() should be called:
384 *
385 * The user detects a page should be reclaimed and calls zbud_reclaim_page().
386 * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
387 * the user-defined eviction handler with the pool and handle as arguments.
388 *
389 * If the handle can not be evicted, the eviction handler should return
390 * non-zero. zbud_reclaim_page() will add the zbud page back to the
391 * appropriate list and try the next zbud page on the LRU up to
392 * a user defined number of retries.
393 *
394 * If the handle is successfully evicted, the eviction handler should
395 * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
396 * contains logic to delay freeing the page if the page is under reclaim,
397 * as indicated by the setting of the PG_reclaim flag on the underlying page.
398 *
399 * If all buddies in the zbud page are successfully evicted, then the
400 * zbud page can be freed.
401 *
402 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
403 * no pages to evict or an eviction handler is not registered, -EAGAIN if
404 * the retry limit was hit.
405 */
406static int zbud_reclaim_page(struct zbud_pool *pool, unsigned int retries)
407{
408	int i, ret, freechunks;
409	struct zbud_header *zhdr;
410	unsigned long first_handle = 0, last_handle = 0;
411
412	spin_lock(&pool->lock);
413	if (list_empty(&pool->lru)) {
414		spin_unlock(&pool->lock);
415		return -EINVAL;
416	}
417	for (i = 0; i < retries; i++) {
418		zhdr = list_last_entry(&pool->lru, struct zbud_header, lru);
419		list_del(&zhdr->lru);
420		list_del(&zhdr->buddy);
421		/* Protect zbud page against free */
422		zhdr->under_reclaim = true;
423		/*
424		 * We need encode the handles before unlocking, since we can
425		 * race with free that will set (first|last)_chunks to 0
426		 */
427		first_handle = 0;
428		last_handle = 0;
429		if (zhdr->first_chunks)
430			first_handle = encode_handle(zhdr, FIRST);
431		if (zhdr->last_chunks)
432			last_handle = encode_handle(zhdr, LAST);
433		spin_unlock(&pool->lock);
434
435		/* Issue the eviction callback(s) */
436		if (first_handle) {
437			ret = pool->zpool_ops->evict(pool->zpool, first_handle);
438			if (ret)
439				goto next;
440		}
441		if (last_handle) {
442			ret = pool->zpool_ops->evict(pool->zpool, last_handle);
443			if (ret)
444				goto next;
445		}
446next:
447		spin_lock(&pool->lock);
448		zhdr->under_reclaim = false;
449		if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
450			/*
451			 * Both buddies are now free, free the zbud page and
452			 * return success.
453			 */
454			free_zbud_page(zhdr);
455			pool->pages_nr--;
456			spin_unlock(&pool->lock);
457			return 0;
458		} else if (zhdr->first_chunks == 0 ||
459				zhdr->last_chunks == 0) {
460			/* add to unbuddied list */
461			freechunks = num_free_chunks(zhdr);
462			list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
463		} else {
464			/* add to buddied list */
465			list_add(&zhdr->buddy, &pool->buddied);
466		}
467
468		/* add to beginning of LRU */
469		list_add(&zhdr->lru, &pool->lru);
470	}
471	spin_unlock(&pool->lock);
472	return -EAGAIN;
473}
474
475/**
476 * zbud_map() - maps the allocation associated with the given handle
477 * @pool:	pool in which the allocation resides
478 * @handle:	handle associated with the allocation to be mapped
479 *
480 * While trivial for zbud, the mapping functions for others allocators
481 * implementing this allocation API could have more complex information encoded
482 * in the handle and could create temporary mappings to make the data
483 * accessible to the user.
484 *
485 * Returns: a pointer to the mapped allocation
486 */
487static void *zbud_map(struct zbud_pool *pool, unsigned long handle)
488{
489	return (void *)(handle);
490}
491
492/**
493 * zbud_unmap() - maps the allocation associated with the given handle
494 * @pool:	pool in which the allocation resides
495 * @handle:	handle associated with the allocation to be unmapped
496 */
497static void zbud_unmap(struct zbud_pool *pool, unsigned long handle)
498{
499}
500
501/**
502 * zbud_get_pool_size() - gets the zbud pool size in pages
503 * @pool:	pool whose size is being queried
504 *
505 * Returns: size in pages of the given pool.  The pool lock need not be
506 * taken to access pages_nr.
507 */
508static u64 zbud_get_pool_size(struct zbud_pool *pool)
509{
510	return pool->pages_nr;
511}
512
513/*****************
514 * zpool
515 ****************/
516
517static void *zbud_zpool_create(const char *name, gfp_t gfp,
518			       const struct zpool_ops *zpool_ops,
519			       struct zpool *zpool)
520{
521	struct zbud_pool *pool;
522
523	pool = zbud_create_pool(gfp);
524	if (pool) {
525		pool->zpool = zpool;
526		pool->zpool_ops = zpool_ops;
527	}
528	return pool;
529}
530
531static void zbud_zpool_destroy(void *pool)
532{
533	zbud_destroy_pool(pool);
534}
535
536static int zbud_zpool_malloc(void *pool, size_t size, gfp_t gfp,
537			unsigned long *handle)
538{
539	return zbud_alloc(pool, size, gfp, handle);
540}
541static void zbud_zpool_free(void *pool, unsigned long handle)
542{
543	zbud_free(pool, handle);
544}
545
546static int zbud_zpool_shrink(void *pool, unsigned int pages,
547			unsigned int *reclaimed)
548{
549	unsigned int total = 0;
550	int ret = -EINVAL;
551
552	while (total < pages) {
553		ret = zbud_reclaim_page(pool, 8);
554		if (ret < 0)
555			break;
556		total++;
557	}
558
559	if (reclaimed)
560		*reclaimed = total;
561
562	return ret;
563}
564
565static void *zbud_zpool_map(void *pool, unsigned long handle,
566			enum zpool_mapmode mm)
567{
568	return zbud_map(pool, handle);
569}
570static void zbud_zpool_unmap(void *pool, unsigned long handle)
571{
572	zbud_unmap(pool, handle);
573}
574
575static u64 zbud_zpool_total_size(void *pool)
576{
577	return zbud_get_pool_size(pool) * PAGE_SIZE;
578}
579
580static struct zpool_driver zbud_zpool_driver = {
581	.type =		"zbud",
582	.sleep_mapped = true,
583	.owner =	THIS_MODULE,
584	.create =	zbud_zpool_create,
585	.destroy =	zbud_zpool_destroy,
586	.malloc =	zbud_zpool_malloc,
587	.free =		zbud_zpool_free,
588	.shrink =	zbud_zpool_shrink,
589	.map =		zbud_zpool_map,
590	.unmap =	zbud_zpool_unmap,
591	.total_size =	zbud_zpool_total_size,
592};
593
594MODULE_ALIAS("zpool-zbud");
595
596static int __init init_zbud(void)
597{
598	/* Make sure the zbud header will fit in one chunk */
599	BUILD_BUG_ON(sizeof(struct zbud_header) > ZHDR_SIZE_ALIGNED);
600	pr_info("loaded\n");
601
602	zpool_register_driver(&zbud_zpool_driver);
603
604	return 0;
605}
606
607static void __exit exit_zbud(void)
608{
609	zpool_unregister_driver(&zbud_zpool_driver);
610	pr_info("unloaded\n");
611}
612
613module_init(init_zbud);
614module_exit(exit_zbud);
615
616MODULE_LICENSE("GPL");
617MODULE_AUTHOR("Seth Jennings <sjennings@variantweb.net>");
618MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");