xref: /openbmc/linux/mm/zbud.c (revision 5104d265)
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  * justifed" 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 
55 /*****************
56  * Structures
57 *****************/
58 /*
59  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
60  * adjusting internal fragmentation.  It also determines the number of
61  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
62  * allocation granularity will be in chunks of size PAGE_SIZE/64, and there
63  * will be 64 freelists per pool.
64  */
65 #define NCHUNKS_ORDER	6
66 
67 #define CHUNK_SHIFT	(PAGE_SHIFT - NCHUNKS_ORDER)
68 #define CHUNK_SIZE	(1 << CHUNK_SHIFT)
69 #define NCHUNKS		(PAGE_SIZE >> CHUNK_SHIFT)
70 #define ZHDR_SIZE_ALIGNED CHUNK_SIZE
71 
72 /**
73  * struct zbud_pool - stores metadata for each zbud pool
74  * @lock:	protects all pool fields and first|last_chunk fields of any
75  *		zbud page in the pool
76  * @unbuddied:	array of lists tracking zbud pages that only contain one buddy;
77  *		the lists each zbud page is added to depends on the size of
78  *		its free region.
79  * @buddied:	list tracking the zbud pages that contain two buddies;
80  *		these zbud pages are full
81  * @lru:	list tracking the zbud pages in LRU order by most recently
82  *		added buddy.
83  * @pages_nr:	number of zbud pages in the pool.
84  * @ops:	pointer to a structure of user defined operations specified at
85  *		pool creation time.
86  *
87  * This structure is allocated at pool creation time and maintains metadata
88  * pertaining to a particular zbud pool.
89  */
90 struct zbud_pool {
91 	spinlock_t lock;
92 	struct list_head unbuddied[NCHUNKS];
93 	struct list_head buddied;
94 	struct list_head lru;
95 	u64 pages_nr;
96 	struct zbud_ops *ops;
97 };
98 
99 /*
100  * struct zbud_header - zbud page metadata occupying the first chunk of each
101  *			zbud page.
102  * @buddy:	links the zbud page into the unbuddied/buddied lists in the pool
103  * @lru:	links the zbud page into the lru list in the pool
104  * @first_chunks:	the size of the first buddy in chunks, 0 if free
105  * @last_chunks:	the size of the last buddy in chunks, 0 if free
106  */
107 struct zbud_header {
108 	struct list_head buddy;
109 	struct list_head lru;
110 	unsigned int first_chunks;
111 	unsigned int last_chunks;
112 	bool under_reclaim;
113 };
114 
115 /*****************
116  * Helpers
117 *****************/
118 /* Just to make the code easier to read */
119 enum buddy {
120 	FIRST,
121 	LAST
122 };
123 
124 /* Converts an allocation size in bytes to size in zbud chunks */
125 static int size_to_chunks(int size)
126 {
127 	return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
128 }
129 
130 #define for_each_unbuddied_list(_iter, _begin) \
131 	for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
132 
133 /* Initializes the zbud header of a newly allocated zbud page */
134 static struct zbud_header *init_zbud_page(struct page *page)
135 {
136 	struct zbud_header *zhdr = page_address(page);
137 	zhdr->first_chunks = 0;
138 	zhdr->last_chunks = 0;
139 	INIT_LIST_HEAD(&zhdr->buddy);
140 	INIT_LIST_HEAD(&zhdr->lru);
141 	zhdr->under_reclaim = 0;
142 	return zhdr;
143 }
144 
145 /* Resets the struct page fields and frees the page */
146 static void free_zbud_page(struct zbud_header *zhdr)
147 {
148 	__free_page(virt_to_page(zhdr));
149 }
150 
151 /*
152  * Encodes the handle of a particular buddy within a zbud page
153  * Pool lock should be held as this function accesses first|last_chunks
154  */
155 static unsigned long encode_handle(struct zbud_header *zhdr, enum buddy bud)
156 {
157 	unsigned long handle;
158 
159 	/*
160 	 * For now, the encoded handle is actually just the pointer to the data
161 	 * but this might not always be the case.  A little information hiding.
162 	 * Add CHUNK_SIZE to the handle if it is the first allocation to jump
163 	 * over the zbud header in the first chunk.
164 	 */
165 	handle = (unsigned long)zhdr;
166 	if (bud == FIRST)
167 		/* skip over zbud header */
168 		handle += ZHDR_SIZE_ALIGNED;
169 	else /* bud == LAST */
170 		handle += PAGE_SIZE - (zhdr->last_chunks  << CHUNK_SHIFT);
171 	return handle;
172 }
173 
174 /* Returns the zbud page where a given handle is stored */
175 static struct zbud_header *handle_to_zbud_header(unsigned long handle)
176 {
177 	return (struct zbud_header *)(handle & PAGE_MASK);
178 }
179 
180 /* Returns the number of free chunks in a zbud page */
181 static int num_free_chunks(struct zbud_header *zhdr)
182 {
183 	/*
184 	 * Rather than branch for different situations, just use the fact that
185 	 * free buddies have a length of zero to simplify everything. -1 at the
186 	 * end for the zbud header.
187 	 */
188 	return NCHUNKS - zhdr->first_chunks - zhdr->last_chunks - 1;
189 }
190 
191 /*****************
192  * API Functions
193 *****************/
194 /**
195  * zbud_create_pool() - create a new zbud pool
196  * @gfp:	gfp flags when allocating the zbud pool structure
197  * @ops:	user-defined operations for the zbud pool
198  *
199  * Return: pointer to the new zbud pool or NULL if the metadata allocation
200  * failed.
201  */
202 struct zbud_pool *zbud_create_pool(gfp_t gfp, struct zbud_ops *ops)
203 {
204 	struct zbud_pool *pool;
205 	int i;
206 
207 	pool = kmalloc(sizeof(struct zbud_pool), gfp);
208 	if (!pool)
209 		return NULL;
210 	spin_lock_init(&pool->lock);
211 	for_each_unbuddied_list(i, 0)
212 		INIT_LIST_HEAD(&pool->unbuddied[i]);
213 	INIT_LIST_HEAD(&pool->buddied);
214 	INIT_LIST_HEAD(&pool->lru);
215 	pool->pages_nr = 0;
216 	pool->ops = ops;
217 	return pool;
218 }
219 
220 /**
221  * zbud_destroy_pool() - destroys an existing zbud pool
222  * @pool:	the zbud pool to be destroyed
223  *
224  * The pool should be emptied before this function is called.
225  */
226 void zbud_destroy_pool(struct zbud_pool *pool)
227 {
228 	kfree(pool);
229 }
230 
231 /**
232  * zbud_alloc() - allocates a region of a given size
233  * @pool:	zbud pool from which to allocate
234  * @size:	size in bytes of the desired allocation
235  * @gfp:	gfp flags used if the pool needs to grow
236  * @handle:	handle of the new allocation
237  *
238  * This function will attempt to find a free region in the pool large enough to
239  * satisfy the allocation request.  A search of the unbuddied lists is
240  * performed first. If no suitable free region is found, then a new page is
241  * allocated and added to the pool to satisfy the request.
242  *
243  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
244  * as zbud pool pages.
245  *
246  * Return: 0 if success and handle is set, otherwise -EINVAL is the size or
247  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
248  * a new page.
249  */
250 int zbud_alloc(struct zbud_pool *pool, int size, gfp_t gfp,
251 			unsigned long *handle)
252 {
253 	int chunks, i, freechunks;
254 	struct zbud_header *zhdr = NULL;
255 	enum buddy bud;
256 	struct page *page;
257 
258 	if (size <= 0 || gfp & __GFP_HIGHMEM)
259 		return -EINVAL;
260 	if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED)
261 		return -ENOSPC;
262 	chunks = size_to_chunks(size);
263 	spin_lock(&pool->lock);
264 
265 	/* First, try to find an unbuddied zbud page. */
266 	zhdr = NULL;
267 	for_each_unbuddied_list(i, chunks) {
268 		if (!list_empty(&pool->unbuddied[i])) {
269 			zhdr = list_first_entry(&pool->unbuddied[i],
270 					struct zbud_header, buddy);
271 			list_del(&zhdr->buddy);
272 			if (zhdr->first_chunks == 0)
273 				bud = FIRST;
274 			else
275 				bud = LAST;
276 			goto found;
277 		}
278 	}
279 
280 	/* Couldn't find unbuddied zbud page, create new one */
281 	spin_unlock(&pool->lock);
282 	page = alloc_page(gfp);
283 	if (!page)
284 		return -ENOMEM;
285 	spin_lock(&pool->lock);
286 	pool->pages_nr++;
287 	zhdr = init_zbud_page(page);
288 	bud = FIRST;
289 
290 found:
291 	if (bud == FIRST)
292 		zhdr->first_chunks = chunks;
293 	else
294 		zhdr->last_chunks = chunks;
295 
296 	if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0) {
297 		/* Add to unbuddied list */
298 		freechunks = num_free_chunks(zhdr);
299 		list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
300 	} else {
301 		/* Add to buddied list */
302 		list_add(&zhdr->buddy, &pool->buddied);
303 	}
304 
305 	/* Add/move zbud page to beginning of LRU */
306 	if (!list_empty(&zhdr->lru))
307 		list_del(&zhdr->lru);
308 	list_add(&zhdr->lru, &pool->lru);
309 
310 	*handle = encode_handle(zhdr, bud);
311 	spin_unlock(&pool->lock);
312 
313 	return 0;
314 }
315 
316 /**
317  * zbud_free() - frees the allocation associated with the given handle
318  * @pool:	pool in which the allocation resided
319  * @handle:	handle associated with the allocation returned by zbud_alloc()
320  *
321  * In the case that the zbud page in which the allocation resides is under
322  * reclaim, as indicated by the PG_reclaim flag being set, this function
323  * only sets the first|last_chunks to 0.  The page is actually freed
324  * once both buddies are evicted (see zbud_reclaim_page() below).
325  */
326 void zbud_free(struct zbud_pool *pool, unsigned long handle)
327 {
328 	struct zbud_header *zhdr;
329 	int freechunks;
330 
331 	spin_lock(&pool->lock);
332 	zhdr = handle_to_zbud_header(handle);
333 
334 	/* If first buddy, handle will be page aligned */
335 	if ((handle - ZHDR_SIZE_ALIGNED) & ~PAGE_MASK)
336 		zhdr->last_chunks = 0;
337 	else
338 		zhdr->first_chunks = 0;
339 
340 	if (zhdr->under_reclaim) {
341 		/* zbud page is under reclaim, reclaim will free */
342 		spin_unlock(&pool->lock);
343 		return;
344 	}
345 
346 	/* Remove from existing buddy list */
347 	list_del(&zhdr->buddy);
348 
349 	if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
350 		/* zbud page is empty, free */
351 		list_del(&zhdr->lru);
352 		free_zbud_page(zhdr);
353 		pool->pages_nr--;
354 	} else {
355 		/* Add to unbuddied list */
356 		freechunks = num_free_chunks(zhdr);
357 		list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
358 	}
359 
360 	spin_unlock(&pool->lock);
361 }
362 
363 #define list_tail_entry(ptr, type, member) \
364 	list_entry((ptr)->prev, type, member)
365 
366 /**
367  * zbud_reclaim_page() - evicts allocations from a pool page and frees it
368  * @pool:	pool from which a page will attempt to be evicted
369  * @retires:	number of pages on the LRU list for which eviction will
370  *		be attempted before failing
371  *
372  * zbud reclaim is different from normal system reclaim in that the reclaim is
373  * done from the bottom, up.  This is because only the bottom layer, zbud, has
374  * information on how the allocations are organized within each zbud page. This
375  * has the potential to create interesting locking situations between zbud and
376  * the user, however.
377  *
378  * To avoid these, this is how zbud_reclaim_page() should be called:
379 
380  * The user detects a page should be reclaimed and calls zbud_reclaim_page().
381  * zbud_reclaim_page() will remove a zbud page from the pool LRU list and call
382  * the user-defined eviction handler with the pool and handle as arguments.
383  *
384  * If the handle can not be evicted, the eviction handler should return
385  * non-zero. zbud_reclaim_page() will add the zbud page back to the
386  * appropriate list and try the next zbud page on the LRU up to
387  * a user defined number of retries.
388  *
389  * If the handle is successfully evicted, the eviction handler should
390  * return 0 _and_ should have called zbud_free() on the handle. zbud_free()
391  * contains logic to delay freeing the page if the page is under reclaim,
392  * as indicated by the setting of the PG_reclaim flag on the underlying page.
393  *
394  * If all buddies in the zbud page are successfully evicted, then the
395  * zbud page can be freed.
396  *
397  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
398  * no pages to evict or an eviction handler is not registered, -EAGAIN if
399  * the retry limit was hit.
400  */
401 int zbud_reclaim_page(struct zbud_pool *pool, unsigned int retries)
402 {
403 	int i, ret, freechunks;
404 	struct zbud_header *zhdr;
405 	unsigned long first_handle = 0, last_handle = 0;
406 
407 	spin_lock(&pool->lock);
408 	if (!pool->ops || !pool->ops->evict || list_empty(&pool->lru) ||
409 			retries == 0) {
410 		spin_unlock(&pool->lock);
411 		return -EINVAL;
412 	}
413 	for (i = 0; i < retries; i++) {
414 		zhdr = list_tail_entry(&pool->lru, struct zbud_header, lru);
415 		list_del(&zhdr->lru);
416 		list_del(&zhdr->buddy);
417 		/* Protect zbud page against free */
418 		zhdr->under_reclaim = true;
419 		/*
420 		 * We need encode the handles before unlocking, since we can
421 		 * race with free that will set (first|last)_chunks to 0
422 		 */
423 		first_handle = 0;
424 		last_handle = 0;
425 		if (zhdr->first_chunks)
426 			first_handle = encode_handle(zhdr, FIRST);
427 		if (zhdr->last_chunks)
428 			last_handle = encode_handle(zhdr, LAST);
429 		spin_unlock(&pool->lock);
430 
431 		/* Issue the eviction callback(s) */
432 		if (first_handle) {
433 			ret = pool->ops->evict(pool, first_handle);
434 			if (ret)
435 				goto next;
436 		}
437 		if (last_handle) {
438 			ret = pool->ops->evict(pool, last_handle);
439 			if (ret)
440 				goto next;
441 		}
442 next:
443 		spin_lock(&pool->lock);
444 		zhdr->under_reclaim = false;
445 		if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
446 			/*
447 			 * Both buddies are now free, free the zbud page and
448 			 * return success.
449 			 */
450 			free_zbud_page(zhdr);
451 			pool->pages_nr--;
452 			spin_unlock(&pool->lock);
453 			return 0;
454 		} else if (zhdr->first_chunks == 0 ||
455 				zhdr->last_chunks == 0) {
456 			/* add to unbuddied list */
457 			freechunks = num_free_chunks(zhdr);
458 			list_add(&zhdr->buddy, &pool->unbuddied[freechunks]);
459 		} else {
460 			/* add to buddied list */
461 			list_add(&zhdr->buddy, &pool->buddied);
462 		}
463 
464 		/* add to beginning of LRU */
465 		list_add(&zhdr->lru, &pool->lru);
466 	}
467 	spin_unlock(&pool->lock);
468 	return -EAGAIN;
469 }
470 
471 /**
472  * zbud_map() - maps the allocation associated with the given handle
473  * @pool:	pool in which the allocation resides
474  * @handle:	handle associated with the allocation to be mapped
475  *
476  * While trivial for zbud, the mapping functions for others allocators
477  * implementing this allocation API could have more complex information encoded
478  * in the handle and could create temporary mappings to make the data
479  * accessible to the user.
480  *
481  * Returns: a pointer to the mapped allocation
482  */
483 void *zbud_map(struct zbud_pool *pool, unsigned long handle)
484 {
485 	return (void *)(handle);
486 }
487 
488 /**
489  * zbud_unmap() - maps the allocation associated with the given handle
490  * @pool:	pool in which the allocation resides
491  * @handle:	handle associated with the allocation to be unmapped
492  */
493 void zbud_unmap(struct zbud_pool *pool, unsigned long handle)
494 {
495 }
496 
497 /**
498  * zbud_get_pool_size() - gets the zbud pool size in pages
499  * @pool:	pool whose size is being queried
500  *
501  * Returns: size in pages of the given pool.  The pool lock need not be
502  * taken to access pages_nr.
503  */
504 u64 zbud_get_pool_size(struct zbud_pool *pool)
505 {
506 	return pool->pages_nr;
507 }
508 
509 static int __init init_zbud(void)
510 {
511 	/* Make sure the zbud header will fit in one chunk */
512 	BUILD_BUG_ON(sizeof(struct zbud_header) > ZHDR_SIZE_ALIGNED);
513 	pr_info("loaded\n");
514 	return 0;
515 }
516 
517 static void __exit exit_zbud(void)
518 {
519 	pr_info("unloaded\n");
520 }
521 
522 module_init(init_zbud);
523 module_exit(exit_zbud);
524 
525 MODULE_LICENSE("GPL");
526 MODULE_AUTHOR("Seth Jennings <sjenning@linux.vnet.ibm.com>");
527 MODULE_DESCRIPTION("Buddy Allocator for Compressed Pages");
528