]>
Commit | Line | Data |
---|---|---|
61989a80 NG |
1 | /* |
2 | * zsmalloc memory allocator | |
3 | * | |
4 | * Copyright (C) 2011 Nitin Gupta | |
31fc00bb | 5 | * Copyright (C) 2012, 2013 Minchan Kim |
61989a80 NG |
6 | * |
7 | * This code is released using a dual license strategy: BSD/GPL | |
8 | * You can choose the license that better fits your requirements. | |
9 | * | |
10 | * Released under the terms of 3-clause BSD License | |
11 | * Released under the terms of GNU General Public License Version 2.0 | |
12 | */ | |
13 | ||
2db51dae | 14 | /* |
c3e3e88a NC |
15 | * This allocator is designed for use with zram. Thus, the allocator is |
16 | * supposed to work well under low memory conditions. In particular, it | |
17 | * never attempts higher order page allocation which is very likely to | |
18 | * fail under memory pressure. On the other hand, if we just use single | |
19 | * (0-order) pages, it would suffer from very high fragmentation -- | |
20 | * any object of size PAGE_SIZE/2 or larger would occupy an entire page. | |
21 | * This was one of the major issues with its predecessor (xvmalloc). | |
2db51dae NG |
22 | * |
23 | * To overcome these issues, zsmalloc allocates a bunch of 0-order pages | |
24 | * and links them together using various 'struct page' fields. These linked | |
25 | * pages act as a single higher-order page i.e. an object can span 0-order | |
26 | * page boundaries. The code refers to these linked pages as a single entity | |
27 | * called zspage. | |
28 | * | |
c3e3e88a NC |
29 | * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE |
30 | * since this satisfies the requirements of all its current users (in the | |
31 | * worst case, page is incompressible and is thus stored "as-is" i.e. in | |
32 | * uncompressed form). For allocation requests larger than this size, failure | |
33 | * is returned (see zs_malloc). | |
34 | * | |
35 | * Additionally, zs_malloc() does not return a dereferenceable pointer. | |
36 | * Instead, it returns an opaque handle (unsigned long) which encodes actual | |
37 | * location of the allocated object. The reason for this indirection is that | |
38 | * zsmalloc does not keep zspages permanently mapped since that would cause | |
39 | * issues on 32-bit systems where the VA region for kernel space mappings | |
40 | * is very small. So, before using the allocating memory, the object has to | |
41 | * be mapped using zs_map_object() to get a usable pointer and subsequently | |
42 | * unmapped using zs_unmap_object(). | |
43 | * | |
2db51dae NG |
44 | * Following is how we use various fields and flags of underlying |
45 | * struct page(s) to form a zspage. | |
46 | * | |
47 | * Usage of struct page fields: | |
48 | * page->first_page: points to the first component (0-order) page | |
49 | * page->index (union with page->freelist): offset of the first object | |
50 | * starting in this page. For the first page, this is | |
51 | * always 0, so we use this field (aka freelist) to point | |
52 | * to the first free object in zspage. | |
53 | * page->lru: links together all component pages (except the first page) | |
54 | * of a zspage | |
55 | * | |
56 | * For _first_ page only: | |
57 | * | |
58 | * page->private (union with page->first_page): refers to the | |
59 | * component page after the first page | |
60 | * page->freelist: points to the first free object in zspage. | |
61 | * Free objects are linked together using in-place | |
62 | * metadata. | |
63 | * page->objects: maximum number of objects we can store in this | |
64 | * zspage (class->zspage_order * PAGE_SIZE / class->size) | |
65 | * page->lru: links together first pages of various zspages. | |
66 | * Basically forming list of zspages in a fullness group. | |
67 | * page->mapping: class index and fullness group of the zspage | |
68 | * | |
69 | * Usage of struct page flags: | |
70 | * PG_private: identifies the first component page | |
71 | * PG_private2: identifies the last component page | |
72 | * | |
73 | */ | |
74 | ||
61989a80 NG |
75 | #ifdef CONFIG_ZSMALLOC_DEBUG |
76 | #define DEBUG | |
77 | #endif | |
78 | ||
79 | #include <linux/module.h> | |
80 | #include <linux/kernel.h> | |
81 | #include <linux/bitops.h> | |
82 | #include <linux/errno.h> | |
83 | #include <linux/highmem.h> | |
61989a80 NG |
84 | #include <linux/string.h> |
85 | #include <linux/slab.h> | |
86 | #include <asm/tlbflush.h> | |
87 | #include <asm/pgtable.h> | |
88 | #include <linux/cpumask.h> | |
89 | #include <linux/cpu.h> | |
0cbb613f | 90 | #include <linux/vmalloc.h> |
c60369f0 | 91 | #include <linux/hardirq.h> |
0959c63f SJ |
92 | #include <linux/spinlock.h> |
93 | #include <linux/types.h> | |
0f050d99 | 94 | #include <linux/debugfs.h> |
bcf1647d | 95 | #include <linux/zsmalloc.h> |
c795779d | 96 | #include <linux/zpool.h> |
0959c63f SJ |
97 | |
98 | /* | |
99 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
100 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
101 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
102 | * to restore link_free pointer values. | |
103 | */ | |
104 | #define ZS_ALIGN 8 | |
105 | ||
106 | /* | |
107 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
108 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
109 | */ | |
110 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
111 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
112 | ||
2e40e163 MK |
113 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
114 | ||
0959c63f SJ |
115 | /* |
116 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 117 | * as single (unsigned long) handle value. |
0959c63f SJ |
118 | * |
119 | * Note that object index <obj_idx> is relative to system | |
120 | * page <PFN> it is stored in, so for each sub-page belonging | |
121 | * to a zspage, obj_idx starts with 0. | |
122 | * | |
123 | * This is made more complicated by various memory models and PAE. | |
124 | */ | |
125 | ||
126 | #ifndef MAX_PHYSMEM_BITS | |
127 | #ifdef CONFIG_HIGHMEM64G | |
128 | #define MAX_PHYSMEM_BITS 36 | |
129 | #else /* !CONFIG_HIGHMEM64G */ | |
130 | /* | |
131 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
132 | * be PAGE_SHIFT | |
133 | */ | |
134 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
135 | #endif | |
136 | #endif | |
137 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
138 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS) | |
139 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) | |
140 | ||
141 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
142 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
143 | #define ZS_MIN_ALLOC_SIZE \ | |
144 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 MK |
145 | /* each chunk includes extra space to keep handle */ |
146 | #define ZS_MAX_ALLOC_SIZE (PAGE_SIZE + ZS_HANDLE_SIZE) | |
0959c63f SJ |
147 | |
148 | /* | |
7eb52512 | 149 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
150 | * trader-off here: |
151 | * - Large number of size classes is potentially wasteful as free page are | |
152 | * spread across these classes | |
153 | * - Small number of size classes causes large internal fragmentation | |
154 | * - Probably its better to use specific size classes (empirically | |
155 | * determined). NOTE: all those class sizes must be set as multiple of | |
156 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
157 | * | |
158 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
159 | * (reason above) | |
160 | */ | |
d662b8eb | 161 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8) |
0959c63f SJ |
162 | |
163 | /* | |
164 | * We do not maintain any list for completely empty or full pages | |
165 | */ | |
166 | enum fullness_group { | |
167 | ZS_ALMOST_FULL, | |
168 | ZS_ALMOST_EMPTY, | |
169 | _ZS_NR_FULLNESS_GROUPS, | |
170 | ||
171 | ZS_EMPTY, | |
172 | ZS_FULL | |
173 | }; | |
174 | ||
0f050d99 GM |
175 | enum zs_stat_type { |
176 | OBJ_ALLOCATED, | |
177 | OBJ_USED, | |
178 | NR_ZS_STAT_TYPE, | |
179 | }; | |
180 | ||
181 | #ifdef CONFIG_ZSMALLOC_STAT | |
182 | ||
183 | static struct dentry *zs_stat_root; | |
184 | ||
185 | struct zs_size_stat { | |
186 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
187 | }; | |
188 | ||
189 | #endif | |
190 | ||
40f9fb8c MG |
191 | /* |
192 | * number of size_classes | |
193 | */ | |
194 | static int zs_size_classes; | |
195 | ||
0959c63f SJ |
196 | /* |
197 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
198 | * n <= N / f, where | |
199 | * n = number of allocated objects | |
200 | * N = total number of objects zspage can store | |
6dd9737e | 201 | * f = fullness_threshold_frac |
0959c63f SJ |
202 | * |
203 | * Similarly, we assign zspage to: | |
204 | * ZS_ALMOST_FULL when n > N / f | |
205 | * ZS_EMPTY when n == 0 | |
206 | * ZS_FULL when n == N | |
207 | * | |
208 | * (see: fix_fullness_group()) | |
209 | */ | |
210 | static const int fullness_threshold_frac = 4; | |
211 | ||
212 | struct size_class { | |
213 | /* | |
214 | * Size of objects stored in this class. Must be multiple | |
215 | * of ZS_ALIGN. | |
216 | */ | |
217 | int size; | |
218 | unsigned int index; | |
219 | ||
220 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ | |
221 | int pages_per_zspage; | |
222 | ||
0f050d99 GM |
223 | #ifdef CONFIG_ZSMALLOC_STAT |
224 | struct zs_size_stat stats; | |
225 | #endif | |
226 | ||
0959c63f SJ |
227 | spinlock_t lock; |
228 | ||
0959c63f SJ |
229 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; |
230 | }; | |
231 | ||
232 | /* | |
233 | * Placed within free objects to form a singly linked list. | |
234 | * For every zspage, first_page->freelist gives head of this list. | |
235 | * | |
236 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
237 | */ | |
238 | struct link_free { | |
2e40e163 MK |
239 | union { |
240 | /* | |
241 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
242 | * It's valid for non-allocated object | |
243 | */ | |
244 | void *next; | |
245 | /* | |
246 | * Handle of allocated object. | |
247 | */ | |
248 | unsigned long handle; | |
249 | }; | |
0959c63f SJ |
250 | }; |
251 | ||
252 | struct zs_pool { | |
0f050d99 GM |
253 | char *name; |
254 | ||
40f9fb8c | 255 | struct size_class **size_class; |
2e40e163 | 256 | struct kmem_cache *handle_cachep; |
0959c63f SJ |
257 | |
258 | gfp_t flags; /* allocation flags used when growing pool */ | |
13de8933 | 259 | atomic_long_t pages_allocated; |
0f050d99 GM |
260 | |
261 | #ifdef CONFIG_ZSMALLOC_STAT | |
262 | struct dentry *stat_dentry; | |
263 | #endif | |
0959c63f | 264 | }; |
61989a80 NG |
265 | |
266 | /* | |
267 | * A zspage's class index and fullness group | |
268 | * are encoded in its (first)page->mapping | |
269 | */ | |
270 | #define CLASS_IDX_BITS 28 | |
271 | #define FULLNESS_BITS 4 | |
272 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
273 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
274 | ||
f553646a | 275 | struct mapping_area { |
1b945aee | 276 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
277 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
278 | #else | |
279 | char *vm_buf; /* copy buffer for objects that span pages */ | |
280 | #endif | |
281 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
282 | enum zs_mapmode vm_mm; /* mapping mode */ | |
283 | }; | |
284 | ||
2e40e163 MK |
285 | static int create_handle_cache(struct zs_pool *pool) |
286 | { | |
287 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
288 | 0, 0, NULL); | |
289 | return pool->handle_cachep ? 0 : 1; | |
290 | } | |
291 | ||
292 | static void destroy_handle_cache(struct zs_pool *pool) | |
293 | { | |
294 | kmem_cache_destroy(pool->handle_cachep); | |
295 | } | |
296 | ||
297 | static unsigned long alloc_handle(struct zs_pool *pool) | |
298 | { | |
299 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
300 | pool->flags & ~__GFP_HIGHMEM); | |
301 | } | |
302 | ||
303 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
304 | { | |
305 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
306 | } | |
307 | ||
308 | static void record_obj(unsigned long handle, unsigned long obj) | |
309 | { | |
310 | *(unsigned long *)handle = obj; | |
311 | } | |
312 | ||
c795779d DS |
313 | /* zpool driver */ |
314 | ||
315 | #ifdef CONFIG_ZPOOL | |
316 | ||
3eba0c6a | 317 | static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops) |
c795779d | 318 | { |
3eba0c6a | 319 | return zs_create_pool(name, gfp); |
c795779d DS |
320 | } |
321 | ||
322 | static void zs_zpool_destroy(void *pool) | |
323 | { | |
324 | zs_destroy_pool(pool); | |
325 | } | |
326 | ||
327 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
328 | unsigned long *handle) | |
329 | { | |
330 | *handle = zs_malloc(pool, size); | |
331 | return *handle ? 0 : -1; | |
332 | } | |
333 | static void zs_zpool_free(void *pool, unsigned long handle) | |
334 | { | |
335 | zs_free(pool, handle); | |
336 | } | |
337 | ||
338 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
339 | unsigned int *reclaimed) | |
340 | { | |
341 | return -EINVAL; | |
342 | } | |
343 | ||
344 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
345 | enum zpool_mapmode mm) | |
346 | { | |
347 | enum zs_mapmode zs_mm; | |
348 | ||
349 | switch (mm) { | |
350 | case ZPOOL_MM_RO: | |
351 | zs_mm = ZS_MM_RO; | |
352 | break; | |
353 | case ZPOOL_MM_WO: | |
354 | zs_mm = ZS_MM_WO; | |
355 | break; | |
356 | case ZPOOL_MM_RW: /* fallthru */ | |
357 | default: | |
358 | zs_mm = ZS_MM_RW; | |
359 | break; | |
360 | } | |
361 | ||
362 | return zs_map_object(pool, handle, zs_mm); | |
363 | } | |
364 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
365 | { | |
366 | zs_unmap_object(pool, handle); | |
367 | } | |
368 | ||
369 | static u64 zs_zpool_total_size(void *pool) | |
370 | { | |
722cdc17 | 371 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
372 | } |
373 | ||
374 | static struct zpool_driver zs_zpool_driver = { | |
375 | .type = "zsmalloc", | |
376 | .owner = THIS_MODULE, | |
377 | .create = zs_zpool_create, | |
378 | .destroy = zs_zpool_destroy, | |
379 | .malloc = zs_zpool_malloc, | |
380 | .free = zs_zpool_free, | |
381 | .shrink = zs_zpool_shrink, | |
382 | .map = zs_zpool_map, | |
383 | .unmap = zs_zpool_unmap, | |
384 | .total_size = zs_zpool_total_size, | |
385 | }; | |
386 | ||
137f8cff | 387 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
388 | #endif /* CONFIG_ZPOOL */ |
389 | ||
61989a80 NG |
390 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
391 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
392 | ||
393 | static int is_first_page(struct page *page) | |
394 | { | |
a27545bf | 395 | return PagePrivate(page); |
61989a80 NG |
396 | } |
397 | ||
398 | static int is_last_page(struct page *page) | |
399 | { | |
a27545bf | 400 | return PagePrivate2(page); |
61989a80 NG |
401 | } |
402 | ||
403 | static void get_zspage_mapping(struct page *page, unsigned int *class_idx, | |
404 | enum fullness_group *fullness) | |
405 | { | |
406 | unsigned long m; | |
407 | BUG_ON(!is_first_page(page)); | |
408 | ||
409 | m = (unsigned long)page->mapping; | |
410 | *fullness = m & FULLNESS_MASK; | |
411 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
412 | } | |
413 | ||
414 | static void set_zspage_mapping(struct page *page, unsigned int class_idx, | |
415 | enum fullness_group fullness) | |
416 | { | |
417 | unsigned long m; | |
418 | BUG_ON(!is_first_page(page)); | |
419 | ||
420 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
421 | (fullness & FULLNESS_MASK); | |
422 | page->mapping = (struct address_space *)m; | |
423 | } | |
424 | ||
c3e3e88a NC |
425 | /* |
426 | * zsmalloc divides the pool into various size classes where each | |
427 | * class maintains a list of zspages where each zspage is divided | |
428 | * into equal sized chunks. Each allocation falls into one of these | |
429 | * classes depending on its size. This function returns index of the | |
430 | * size class which has chunk size big enough to hold the give size. | |
431 | */ | |
61989a80 NG |
432 | static int get_size_class_index(int size) |
433 | { | |
434 | int idx = 0; | |
435 | ||
436 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
437 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
438 | ZS_SIZE_CLASS_DELTA); | |
439 | ||
440 | return idx; | |
441 | } | |
442 | ||
c3e3e88a NC |
443 | /* |
444 | * For each size class, zspages are divided into different groups | |
445 | * depending on how "full" they are. This was done so that we could | |
446 | * easily find empty or nearly empty zspages when we try to shrink | |
447 | * the pool (not yet implemented). This function returns fullness | |
448 | * status of the given page. | |
449 | */ | |
61989a80 NG |
450 | static enum fullness_group get_fullness_group(struct page *page) |
451 | { | |
452 | int inuse, max_objects; | |
453 | enum fullness_group fg; | |
454 | BUG_ON(!is_first_page(page)); | |
455 | ||
456 | inuse = page->inuse; | |
457 | max_objects = page->objects; | |
458 | ||
459 | if (inuse == 0) | |
460 | fg = ZS_EMPTY; | |
461 | else if (inuse == max_objects) | |
462 | fg = ZS_FULL; | |
463 | else if (inuse <= max_objects / fullness_threshold_frac) | |
464 | fg = ZS_ALMOST_EMPTY; | |
465 | else | |
466 | fg = ZS_ALMOST_FULL; | |
467 | ||
468 | return fg; | |
469 | } | |
470 | ||
c3e3e88a NC |
471 | /* |
472 | * Each size class maintains various freelists and zspages are assigned | |
473 | * to one of these freelists based on the number of live objects they | |
474 | * have. This functions inserts the given zspage into the freelist | |
475 | * identified by <class, fullness_group>. | |
476 | */ | |
61989a80 NG |
477 | static void insert_zspage(struct page *page, struct size_class *class, |
478 | enum fullness_group fullness) | |
479 | { | |
480 | struct page **head; | |
481 | ||
482 | BUG_ON(!is_first_page(page)); | |
483 | ||
484 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
485 | return; | |
486 | ||
487 | head = &class->fullness_list[fullness]; | |
488 | if (*head) | |
489 | list_add_tail(&page->lru, &(*head)->lru); | |
490 | ||
491 | *head = page; | |
492 | } | |
493 | ||
c3e3e88a NC |
494 | /* |
495 | * This function removes the given zspage from the freelist identified | |
496 | * by <class, fullness_group>. | |
497 | */ | |
61989a80 NG |
498 | static void remove_zspage(struct page *page, struct size_class *class, |
499 | enum fullness_group fullness) | |
500 | { | |
501 | struct page **head; | |
502 | ||
503 | BUG_ON(!is_first_page(page)); | |
504 | ||
505 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
506 | return; | |
507 | ||
508 | head = &class->fullness_list[fullness]; | |
509 | BUG_ON(!*head); | |
510 | if (list_empty(&(*head)->lru)) | |
511 | *head = NULL; | |
512 | else if (*head == page) | |
513 | *head = (struct page *)list_entry((*head)->lru.next, | |
514 | struct page, lru); | |
515 | ||
516 | list_del_init(&page->lru); | |
517 | } | |
518 | ||
c3e3e88a NC |
519 | /* |
520 | * Each size class maintains zspages in different fullness groups depending | |
521 | * on the number of live objects they contain. When allocating or freeing | |
522 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
523 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
524 | * a status change has occurred for the given page and accordingly moves the | |
525 | * page from the freelist of the old fullness group to that of the new | |
526 | * fullness group. | |
527 | */ | |
c7806261 | 528 | static enum fullness_group fix_fullness_group(struct size_class *class, |
61989a80 NG |
529 | struct page *page) |
530 | { | |
531 | int class_idx; | |
61989a80 NG |
532 | enum fullness_group currfg, newfg; |
533 | ||
534 | BUG_ON(!is_first_page(page)); | |
535 | ||
536 | get_zspage_mapping(page, &class_idx, &currfg); | |
537 | newfg = get_fullness_group(page); | |
538 | if (newfg == currfg) | |
539 | goto out; | |
540 | ||
61989a80 NG |
541 | remove_zspage(page, class, currfg); |
542 | insert_zspage(page, class, newfg); | |
543 | set_zspage_mapping(page, class_idx, newfg); | |
544 | ||
545 | out: | |
546 | return newfg; | |
547 | } | |
548 | ||
549 | /* | |
550 | * We have to decide on how many pages to link together | |
551 | * to form a zspage for each size class. This is important | |
552 | * to reduce wastage due to unusable space left at end of | |
553 | * each zspage which is given as: | |
554 | * wastage = Zp - Zp % size_class | |
555 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... | |
556 | * | |
557 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
558 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
559 | * since then we can perfectly fit in 8 such objects. | |
560 | */ | |
2e3b6154 | 561 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
562 | { |
563 | int i, max_usedpc = 0; | |
564 | /* zspage order which gives maximum used size per KB */ | |
565 | int max_usedpc_order = 1; | |
566 | ||
84d4faab | 567 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
568 | int zspage_size; |
569 | int waste, usedpc; | |
570 | ||
571 | zspage_size = i * PAGE_SIZE; | |
572 | waste = zspage_size % class_size; | |
573 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
574 | ||
575 | if (usedpc > max_usedpc) { | |
576 | max_usedpc = usedpc; | |
577 | max_usedpc_order = i; | |
578 | } | |
579 | } | |
580 | ||
581 | return max_usedpc_order; | |
582 | } | |
583 | ||
584 | /* | |
585 | * A single 'zspage' is composed of many system pages which are | |
586 | * linked together using fields in struct page. This function finds | |
587 | * the first/head page, given any component page of a zspage. | |
588 | */ | |
589 | static struct page *get_first_page(struct page *page) | |
590 | { | |
591 | if (is_first_page(page)) | |
592 | return page; | |
593 | else | |
594 | return page->first_page; | |
595 | } | |
596 | ||
597 | static struct page *get_next_page(struct page *page) | |
598 | { | |
599 | struct page *next; | |
600 | ||
601 | if (is_last_page(page)) | |
602 | next = NULL; | |
603 | else if (is_first_page(page)) | |
e842b976 | 604 | next = (struct page *)page_private(page); |
61989a80 NG |
605 | else |
606 | next = list_entry(page->lru.next, struct page, lru); | |
607 | ||
608 | return next; | |
609 | } | |
610 | ||
67296874 OH |
611 | /* |
612 | * Encode <page, obj_idx> as a single handle value. | |
613 | * On hardware platforms with physical memory starting at 0x0 the pfn | |
614 | * could be 0 so we ensure that the handle will never be 0 by adjusting the | |
615 | * encoded obj_idx value before encoding. | |
616 | */ | |
61989a80 NG |
617 | static void *obj_location_to_handle(struct page *page, unsigned long obj_idx) |
618 | { | |
619 | unsigned long handle; | |
620 | ||
621 | if (!page) { | |
622 | BUG_ON(obj_idx); | |
623 | return NULL; | |
624 | } | |
625 | ||
626 | handle = page_to_pfn(page) << OBJ_INDEX_BITS; | |
67296874 | 627 | handle |= ((obj_idx + 1) & OBJ_INDEX_MASK); |
61989a80 NG |
628 | |
629 | return (void *)handle; | |
630 | } | |
631 | ||
67296874 OH |
632 | /* |
633 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
634 | * decoded obj_idx back to its original value since it was adjusted in | |
635 | * obj_location_to_handle(). | |
636 | */ | |
2e40e163 | 637 | static void obj_to_location(unsigned long handle, struct page **page, |
61989a80 NG |
638 | unsigned long *obj_idx) |
639 | { | |
c2344348 | 640 | *page = pfn_to_page(handle >> OBJ_INDEX_BITS); |
67296874 | 641 | *obj_idx = (handle & OBJ_INDEX_MASK) - 1; |
61989a80 NG |
642 | } |
643 | ||
2e40e163 MK |
644 | static unsigned long handle_to_obj(unsigned long handle) |
645 | { | |
646 | return *(unsigned long *)handle; | |
647 | } | |
648 | ||
61989a80 NG |
649 | static unsigned long obj_idx_to_offset(struct page *page, |
650 | unsigned long obj_idx, int class_size) | |
651 | { | |
652 | unsigned long off = 0; | |
653 | ||
654 | if (!is_first_page(page)) | |
655 | off = page->index; | |
656 | ||
657 | return off + obj_idx * class_size; | |
658 | } | |
659 | ||
f4477e90 NG |
660 | static void reset_page(struct page *page) |
661 | { | |
662 | clear_bit(PG_private, &page->flags); | |
663 | clear_bit(PG_private_2, &page->flags); | |
664 | set_page_private(page, 0); | |
665 | page->mapping = NULL; | |
666 | page->freelist = NULL; | |
22b751c3 | 667 | page_mapcount_reset(page); |
f4477e90 NG |
668 | } |
669 | ||
61989a80 NG |
670 | static void free_zspage(struct page *first_page) |
671 | { | |
f4477e90 | 672 | struct page *nextp, *tmp, *head_extra; |
61989a80 NG |
673 | |
674 | BUG_ON(!is_first_page(first_page)); | |
675 | BUG_ON(first_page->inuse); | |
676 | ||
f4477e90 | 677 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 678 | |
f4477e90 | 679 | reset_page(first_page); |
61989a80 NG |
680 | __free_page(first_page); |
681 | ||
682 | /* zspage with only 1 system page */ | |
f4477e90 | 683 | if (!head_extra) |
61989a80 NG |
684 | return; |
685 | ||
f4477e90 | 686 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 687 | list_del(&nextp->lru); |
f4477e90 | 688 | reset_page(nextp); |
61989a80 NG |
689 | __free_page(nextp); |
690 | } | |
f4477e90 NG |
691 | reset_page(head_extra); |
692 | __free_page(head_extra); | |
61989a80 NG |
693 | } |
694 | ||
695 | /* Initialize a newly allocated zspage */ | |
696 | static void init_zspage(struct page *first_page, struct size_class *class) | |
697 | { | |
698 | unsigned long off = 0; | |
699 | struct page *page = first_page; | |
700 | ||
701 | BUG_ON(!is_first_page(first_page)); | |
702 | while (page) { | |
703 | struct page *next_page; | |
704 | struct link_free *link; | |
5538c562 | 705 | unsigned int i = 1; |
af4ee5e9 | 706 | void *vaddr; |
61989a80 NG |
707 | |
708 | /* | |
709 | * page->index stores offset of first object starting | |
710 | * in the page. For the first page, this is always 0, | |
711 | * so we use first_page->index (aka ->freelist) to store | |
712 | * head of corresponding zspage's freelist. | |
713 | */ | |
714 | if (page != first_page) | |
715 | page->index = off; | |
716 | ||
af4ee5e9 MK |
717 | vaddr = kmap_atomic(page); |
718 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
719 | |
720 | while ((off += class->size) < PAGE_SIZE) { | |
721 | link->next = obj_location_to_handle(page, i++); | |
722 | link += class->size / sizeof(*link); | |
61989a80 NG |
723 | } |
724 | ||
725 | /* | |
726 | * We now come to the last (full or partial) object on this | |
727 | * page, which must point to the first object on the next | |
728 | * page (if present) | |
729 | */ | |
730 | next_page = get_next_page(page); | |
731 | link->next = obj_location_to_handle(next_page, 0); | |
af4ee5e9 | 732 | kunmap_atomic(vaddr); |
61989a80 | 733 | page = next_page; |
5538c562 | 734 | off %= PAGE_SIZE; |
61989a80 NG |
735 | } |
736 | } | |
737 | ||
738 | /* | |
739 | * Allocate a zspage for the given size class | |
740 | */ | |
741 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
742 | { | |
743 | int i, error; | |
b4b700c5 | 744 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
745 | |
746 | /* | |
747 | * Allocate individual pages and link them together as: | |
748 | * 1. first page->private = first sub-page | |
749 | * 2. all sub-pages are linked together using page->lru | |
750 | * 3. each sub-page is linked to the first page using page->first_page | |
751 | * | |
752 | * For each size class, First/Head pages are linked together using | |
753 | * page->lru. Also, we set PG_private to identify the first page | |
754 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
755 | * identify the last page. | |
756 | */ | |
757 | error = -ENOMEM; | |
2e3b6154 | 758 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 759 | struct page *page; |
61989a80 NG |
760 | |
761 | page = alloc_page(flags); | |
762 | if (!page) | |
763 | goto cleanup; | |
764 | ||
765 | INIT_LIST_HEAD(&page->lru); | |
766 | if (i == 0) { /* first page */ | |
a27545bf | 767 | SetPagePrivate(page); |
61989a80 NG |
768 | set_page_private(page, 0); |
769 | first_page = page; | |
770 | first_page->inuse = 0; | |
771 | } | |
772 | if (i == 1) | |
e842b976 | 773 | set_page_private(first_page, (unsigned long)page); |
61989a80 NG |
774 | if (i >= 1) |
775 | page->first_page = first_page; | |
776 | if (i >= 2) | |
777 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 778 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 779 | SetPagePrivate2(page); |
61989a80 NG |
780 | prev_page = page; |
781 | } | |
782 | ||
783 | init_zspage(first_page, class); | |
784 | ||
785 | first_page->freelist = obj_location_to_handle(first_page, 0); | |
786 | /* Maximum number of objects we can store in this zspage */ | |
2e3b6154 | 787 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
788 | |
789 | error = 0; /* Success */ | |
790 | ||
791 | cleanup: | |
792 | if (unlikely(error) && first_page) { | |
793 | free_zspage(first_page); | |
794 | first_page = NULL; | |
795 | } | |
796 | ||
797 | return first_page; | |
798 | } | |
799 | ||
800 | static struct page *find_get_zspage(struct size_class *class) | |
801 | { | |
802 | int i; | |
803 | struct page *page; | |
804 | ||
805 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
806 | page = class->fullness_list[i]; | |
807 | if (page) | |
808 | break; | |
809 | } | |
810 | ||
811 | return page; | |
812 | } | |
813 | ||
1b945aee | 814 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
815 | static inline int __zs_cpu_up(struct mapping_area *area) |
816 | { | |
817 | /* | |
818 | * Make sure we don't leak memory if a cpu UP notification | |
819 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
820 | */ | |
821 | if (area->vm) | |
822 | return 0; | |
823 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
824 | if (!area->vm) | |
825 | return -ENOMEM; | |
826 | return 0; | |
827 | } | |
828 | ||
829 | static inline void __zs_cpu_down(struct mapping_area *area) | |
830 | { | |
831 | if (area->vm) | |
832 | free_vm_area(area->vm); | |
833 | area->vm = NULL; | |
834 | } | |
835 | ||
836 | static inline void *__zs_map_object(struct mapping_area *area, | |
837 | struct page *pages[2], int off, int size) | |
838 | { | |
f6f8ed47 | 839 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
840 | area->vm_addr = area->vm->addr; |
841 | return area->vm_addr + off; | |
842 | } | |
843 | ||
844 | static inline void __zs_unmap_object(struct mapping_area *area, | |
845 | struct page *pages[2], int off, int size) | |
846 | { | |
847 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 848 | |
d95abbbb | 849 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
850 | } |
851 | ||
1b945aee | 852 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
853 | |
854 | static inline int __zs_cpu_up(struct mapping_area *area) | |
855 | { | |
856 | /* | |
857 | * Make sure we don't leak memory if a cpu UP notification | |
858 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
859 | */ | |
860 | if (area->vm_buf) | |
861 | return 0; | |
40f9fb8c | 862 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
863 | if (!area->vm_buf) |
864 | return -ENOMEM; | |
865 | return 0; | |
866 | } | |
867 | ||
868 | static inline void __zs_cpu_down(struct mapping_area *area) | |
869 | { | |
40f9fb8c | 870 | kfree(area->vm_buf); |
f553646a SJ |
871 | area->vm_buf = NULL; |
872 | } | |
873 | ||
874 | static void *__zs_map_object(struct mapping_area *area, | |
875 | struct page *pages[2], int off, int size) | |
5f601902 | 876 | { |
5f601902 SJ |
877 | int sizes[2]; |
878 | void *addr; | |
f553646a | 879 | char *buf = area->vm_buf; |
5f601902 | 880 | |
f553646a SJ |
881 | /* disable page faults to match kmap_atomic() return conditions */ |
882 | pagefault_disable(); | |
883 | ||
884 | /* no read fastpath */ | |
885 | if (area->vm_mm == ZS_MM_WO) | |
886 | goto out; | |
5f601902 SJ |
887 | |
888 | sizes[0] = PAGE_SIZE - off; | |
889 | sizes[1] = size - sizes[0]; | |
890 | ||
5f601902 SJ |
891 | /* copy object to per-cpu buffer */ |
892 | addr = kmap_atomic(pages[0]); | |
893 | memcpy(buf, addr + off, sizes[0]); | |
894 | kunmap_atomic(addr); | |
895 | addr = kmap_atomic(pages[1]); | |
896 | memcpy(buf + sizes[0], addr, sizes[1]); | |
897 | kunmap_atomic(addr); | |
f553646a SJ |
898 | out: |
899 | return area->vm_buf; | |
5f601902 SJ |
900 | } |
901 | ||
f553646a SJ |
902 | static void __zs_unmap_object(struct mapping_area *area, |
903 | struct page *pages[2], int off, int size) | |
5f601902 | 904 | { |
5f601902 SJ |
905 | int sizes[2]; |
906 | void *addr; | |
2e40e163 | 907 | char *buf; |
5f601902 | 908 | |
f553646a SJ |
909 | /* no write fastpath */ |
910 | if (area->vm_mm == ZS_MM_RO) | |
911 | goto out; | |
5f601902 | 912 | |
2e40e163 MK |
913 | buf = area->vm_buf + ZS_HANDLE_SIZE; |
914 | size -= ZS_HANDLE_SIZE; | |
915 | off += ZS_HANDLE_SIZE; | |
916 | ||
5f601902 SJ |
917 | sizes[0] = PAGE_SIZE - off; |
918 | sizes[1] = size - sizes[0]; | |
919 | ||
920 | /* copy per-cpu buffer to object */ | |
921 | addr = kmap_atomic(pages[0]); | |
922 | memcpy(addr + off, buf, sizes[0]); | |
923 | kunmap_atomic(addr); | |
924 | addr = kmap_atomic(pages[1]); | |
925 | memcpy(addr, buf + sizes[0], sizes[1]); | |
926 | kunmap_atomic(addr); | |
f553646a SJ |
927 | |
928 | out: | |
929 | /* enable page faults to match kunmap_atomic() return conditions */ | |
930 | pagefault_enable(); | |
5f601902 | 931 | } |
61989a80 | 932 | |
1b945aee | 933 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 934 | |
61989a80 NG |
935 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
936 | void *pcpu) | |
937 | { | |
f553646a | 938 | int ret, cpu = (long)pcpu; |
61989a80 NG |
939 | struct mapping_area *area; |
940 | ||
941 | switch (action) { | |
942 | case CPU_UP_PREPARE: | |
943 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
944 | ret = __zs_cpu_up(area); |
945 | if (ret) | |
946 | return notifier_from_errno(ret); | |
61989a80 NG |
947 | break; |
948 | case CPU_DEAD: | |
949 | case CPU_UP_CANCELED: | |
950 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 951 | __zs_cpu_down(area); |
61989a80 NG |
952 | break; |
953 | } | |
954 | ||
955 | return NOTIFY_OK; | |
956 | } | |
957 | ||
958 | static struct notifier_block zs_cpu_nb = { | |
959 | .notifier_call = zs_cpu_notifier | |
960 | }; | |
961 | ||
b1b00a5b | 962 | static int zs_register_cpu_notifier(void) |
61989a80 | 963 | { |
b1b00a5b | 964 | int cpu, uninitialized_var(ret); |
61989a80 | 965 | |
f0e71fcd SB |
966 | cpu_notifier_register_begin(); |
967 | ||
968 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
969 | for_each_online_cpu(cpu) { |
970 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
971 | if (notifier_to_errno(ret)) |
972 | break; | |
61989a80 | 973 | } |
f0e71fcd SB |
974 | |
975 | cpu_notifier_register_done(); | |
b1b00a5b SS |
976 | return notifier_to_errno(ret); |
977 | } | |
f0e71fcd | 978 | |
66cdef66 | 979 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 980 | { |
66cdef66 | 981 | int cpu; |
40f9fb8c | 982 | |
66cdef66 | 983 | cpu_notifier_register_begin(); |
40f9fb8c | 984 | |
66cdef66 GM |
985 | for_each_online_cpu(cpu) |
986 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
987 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 988 | |
66cdef66 | 989 | cpu_notifier_register_done(); |
b1b00a5b SS |
990 | } |
991 | ||
66cdef66 | 992 | static void init_zs_size_classes(void) |
b1b00a5b | 993 | { |
66cdef66 | 994 | int nr; |
c795779d | 995 | |
66cdef66 GM |
996 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
997 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
998 | nr += 1; | |
40f9fb8c | 999 | |
66cdef66 | 1000 | zs_size_classes = nr; |
61989a80 NG |
1001 | } |
1002 | ||
9eec4cd5 JK |
1003 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
1004 | { | |
1005 | return pages_per_zspage * PAGE_SIZE / size; | |
1006 | } | |
1007 | ||
1008 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) | |
1009 | { | |
1010 | if (prev->pages_per_zspage != pages_per_zspage) | |
1011 | return false; | |
1012 | ||
1013 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1014 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1015 | return false; | |
1016 | ||
1017 | return true; | |
1018 | } | |
1019 | ||
0f050d99 GM |
1020 | #ifdef CONFIG_ZSMALLOC_STAT |
1021 | ||
1022 | static inline void zs_stat_inc(struct size_class *class, | |
1023 | enum zs_stat_type type, unsigned long cnt) | |
1024 | { | |
1025 | class->stats.objs[type] += cnt; | |
1026 | } | |
1027 | ||
1028 | static inline void zs_stat_dec(struct size_class *class, | |
1029 | enum zs_stat_type type, unsigned long cnt) | |
1030 | { | |
1031 | class->stats.objs[type] -= cnt; | |
1032 | } | |
1033 | ||
1034 | static inline unsigned long zs_stat_get(struct size_class *class, | |
1035 | enum zs_stat_type type) | |
1036 | { | |
1037 | return class->stats.objs[type]; | |
1038 | } | |
1039 | ||
1040 | static int __init zs_stat_init(void) | |
1041 | { | |
1042 | if (!debugfs_initialized()) | |
1043 | return -ENODEV; | |
1044 | ||
1045 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
1046 | if (!zs_stat_root) | |
1047 | return -ENOMEM; | |
1048 | ||
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | static void __exit zs_stat_exit(void) | |
1053 | { | |
1054 | debugfs_remove_recursive(zs_stat_root); | |
1055 | } | |
1056 | ||
1057 | static int zs_stats_size_show(struct seq_file *s, void *v) | |
1058 | { | |
1059 | int i; | |
1060 | struct zs_pool *pool = s->private; | |
1061 | struct size_class *class; | |
1062 | int objs_per_zspage; | |
1063 | unsigned long obj_allocated, obj_used, pages_used; | |
1064 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1065 | ||
1066 | seq_printf(s, " %5s %5s %13s %10s %10s\n", "class", "size", | |
1067 | "obj_allocated", "obj_used", "pages_used"); | |
1068 | ||
1069 | for (i = 0; i < zs_size_classes; i++) { | |
1070 | class = pool->size_class[i]; | |
1071 | ||
1072 | if (class->index != i) | |
1073 | continue; | |
1074 | ||
1075 | spin_lock(&class->lock); | |
1076 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
1077 | obj_used = zs_stat_get(class, OBJ_USED); | |
1078 | spin_unlock(&class->lock); | |
1079 | ||
1080 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
1081 | class->pages_per_zspage); | |
1082 | pages_used = obj_allocated / objs_per_zspage * | |
1083 | class->pages_per_zspage; | |
1084 | ||
1085 | seq_printf(s, " %5u %5u %10lu %10lu %10lu\n", i, | |
1086 | class->size, obj_allocated, obj_used, pages_used); | |
1087 | ||
1088 | total_objs += obj_allocated; | |
1089 | total_used_objs += obj_used; | |
1090 | total_pages += pages_used; | |
1091 | } | |
1092 | ||
1093 | seq_puts(s, "\n"); | |
1094 | seq_printf(s, " %5s %5s %10lu %10lu %10lu\n", "Total", "", | |
1095 | total_objs, total_used_objs, total_pages); | |
1096 | ||
1097 | return 0; | |
1098 | } | |
1099 | ||
1100 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
1101 | { | |
1102 | return single_open(file, zs_stats_size_show, inode->i_private); | |
1103 | } | |
1104 | ||
1105 | static const struct file_operations zs_stat_size_ops = { | |
1106 | .open = zs_stats_size_open, | |
1107 | .read = seq_read, | |
1108 | .llseek = seq_lseek, | |
1109 | .release = single_release, | |
1110 | }; | |
1111 | ||
1112 | static int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
1113 | { | |
1114 | struct dentry *entry; | |
1115 | ||
1116 | if (!zs_stat_root) | |
1117 | return -ENODEV; | |
1118 | ||
1119 | entry = debugfs_create_dir(name, zs_stat_root); | |
1120 | if (!entry) { | |
1121 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
1122 | return -ENOMEM; | |
1123 | } | |
1124 | pool->stat_dentry = entry; | |
1125 | ||
1126 | entry = debugfs_create_file("obj_in_classes", S_IFREG | S_IRUGO, | |
1127 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
1128 | if (!entry) { | |
1129 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
1130 | name, "obj_in_classes"); | |
1131 | return -ENOMEM; | |
1132 | } | |
1133 | ||
1134 | return 0; | |
1135 | } | |
1136 | ||
1137 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
1138 | { | |
1139 | debugfs_remove_recursive(pool->stat_dentry); | |
1140 | } | |
1141 | ||
1142 | #else /* CONFIG_ZSMALLOC_STAT */ | |
1143 | ||
1144 | static inline void zs_stat_inc(struct size_class *class, | |
1145 | enum zs_stat_type type, unsigned long cnt) | |
1146 | { | |
1147 | } | |
1148 | ||
1149 | static inline void zs_stat_dec(struct size_class *class, | |
1150 | enum zs_stat_type type, unsigned long cnt) | |
1151 | { | |
1152 | } | |
1153 | ||
1154 | static inline unsigned long zs_stat_get(struct size_class *class, | |
1155 | enum zs_stat_type type) | |
1156 | { | |
1157 | return 0; | |
1158 | } | |
1159 | ||
1160 | static int __init zs_stat_init(void) | |
1161 | { | |
1162 | return 0; | |
1163 | } | |
1164 | ||
1165 | static void __exit zs_stat_exit(void) | |
1166 | { | |
1167 | } | |
1168 | ||
1169 | static inline int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
1170 | { | |
1171 | return 0; | |
1172 | } | |
1173 | ||
1174 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
1175 | { | |
1176 | } | |
1177 | ||
1178 | #endif | |
1179 | ||
66cdef66 GM |
1180 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1181 | { | |
1182 | return atomic_long_read(&pool->pages_allocated); | |
1183 | } | |
1184 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1185 | ||
4bbc0bc0 | 1186 | /** |
66cdef66 GM |
1187 | * zs_map_object - get address of allocated object from handle. |
1188 | * @pool: pool from which the object was allocated | |
1189 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1190 | * |
66cdef66 GM |
1191 | * Before using an object allocated from zs_malloc, it must be mapped using |
1192 | * this function. When done with the object, it must be unmapped using | |
1193 | * zs_unmap_object. | |
4bbc0bc0 | 1194 | * |
66cdef66 GM |
1195 | * Only one object can be mapped per cpu at a time. There is no protection |
1196 | * against nested mappings. | |
1197 | * | |
1198 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1199 | */ |
66cdef66 GM |
1200 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1201 | enum zs_mapmode mm) | |
61989a80 | 1202 | { |
66cdef66 | 1203 | struct page *page; |
2e40e163 | 1204 | unsigned long obj, obj_idx, off; |
61989a80 | 1205 | |
66cdef66 GM |
1206 | unsigned int class_idx; |
1207 | enum fullness_group fg; | |
1208 | struct size_class *class; | |
1209 | struct mapping_area *area; | |
1210 | struct page *pages[2]; | |
2e40e163 | 1211 | void *ret; |
61989a80 | 1212 | |
66cdef66 | 1213 | BUG_ON(!handle); |
40f9fb8c | 1214 | |
9eec4cd5 | 1215 | /* |
66cdef66 GM |
1216 | * Because we use per-cpu mapping areas shared among the |
1217 | * pools/users, we can't allow mapping in interrupt context | |
1218 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1219 | */ |
66cdef66 | 1220 | BUG_ON(in_interrupt()); |
61989a80 | 1221 | |
2e40e163 MK |
1222 | obj = handle_to_obj(handle); |
1223 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1224 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1225 | class = pool->size_class[class_idx]; | |
1226 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1227 | |
66cdef66 GM |
1228 | area = &get_cpu_var(zs_map_area); |
1229 | area->vm_mm = mm; | |
1230 | if (off + class->size <= PAGE_SIZE) { | |
1231 | /* this object is contained entirely within a page */ | |
1232 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1233 | ret = area->vm_addr + off; |
1234 | goto out; | |
61989a80 NG |
1235 | } |
1236 | ||
66cdef66 GM |
1237 | /* this object spans two pages */ |
1238 | pages[0] = page; | |
1239 | pages[1] = get_next_page(page); | |
1240 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1241 | |
2e40e163 MK |
1242 | ret = __zs_map_object(area, pages, off, class->size); |
1243 | out: | |
1244 | return ret + ZS_HANDLE_SIZE; | |
61989a80 | 1245 | } |
66cdef66 | 1246 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1247 | |
66cdef66 | 1248 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1249 | { |
66cdef66 | 1250 | struct page *page; |
2e40e163 | 1251 | unsigned long obj, obj_idx, off; |
61989a80 | 1252 | |
66cdef66 GM |
1253 | unsigned int class_idx; |
1254 | enum fullness_group fg; | |
1255 | struct size_class *class; | |
1256 | struct mapping_area *area; | |
9eec4cd5 | 1257 | |
66cdef66 | 1258 | BUG_ON(!handle); |
9eec4cd5 | 1259 | |
2e40e163 MK |
1260 | obj = handle_to_obj(handle); |
1261 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1262 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1263 | class = pool->size_class[class_idx]; | |
1264 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1265 | |
66cdef66 GM |
1266 | area = this_cpu_ptr(&zs_map_area); |
1267 | if (off + class->size <= PAGE_SIZE) | |
1268 | kunmap_atomic(area->vm_addr); | |
1269 | else { | |
1270 | struct page *pages[2]; | |
40f9fb8c | 1271 | |
66cdef66 GM |
1272 | pages[0] = page; |
1273 | pages[1] = get_next_page(page); | |
1274 | BUG_ON(!pages[1]); | |
1275 | ||
1276 | __zs_unmap_object(area, pages, off, class->size); | |
1277 | } | |
1278 | put_cpu_var(zs_map_area); | |
61989a80 | 1279 | } |
66cdef66 | 1280 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1281 | |
c7806261 MK |
1282 | static unsigned long obj_malloc(struct page *first_page, |
1283 | struct size_class *class, unsigned long handle) | |
1284 | { | |
1285 | unsigned long obj; | |
1286 | struct link_free *link; | |
1287 | ||
1288 | struct page *m_page; | |
1289 | unsigned long m_objidx, m_offset; | |
1290 | void *vaddr; | |
1291 | ||
1292 | obj = (unsigned long)first_page->freelist; | |
1293 | obj_to_location(obj, &m_page, &m_objidx); | |
1294 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1295 | ||
1296 | vaddr = kmap_atomic(m_page); | |
1297 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1298 | first_page->freelist = link->next; | |
1299 | /* record handle in the header of allocated chunk */ | |
1300 | link->handle = handle; | |
1301 | kunmap_atomic(vaddr); | |
1302 | first_page->inuse++; | |
1303 | zs_stat_inc(class, OBJ_USED, 1); | |
1304 | ||
1305 | return obj; | |
1306 | } | |
1307 | ||
1308 | ||
61989a80 NG |
1309 | /** |
1310 | * zs_malloc - Allocate block of given size from pool. | |
1311 | * @pool: pool to allocate from | |
1312 | * @size: size of block to allocate | |
61989a80 | 1313 | * |
00a61d86 | 1314 | * On success, handle to the allocated object is returned, |
c2344348 | 1315 | * otherwise 0. |
61989a80 NG |
1316 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1317 | */ | |
c2344348 | 1318 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
61989a80 | 1319 | { |
2e40e163 | 1320 | unsigned long handle, obj; |
61989a80 | 1321 | struct size_class *class; |
c7806261 | 1322 | struct page *first_page; |
61989a80 | 1323 | |
2e40e163 MK |
1324 | if (unlikely(!size || (size + ZS_HANDLE_SIZE) > ZS_MAX_ALLOC_SIZE)) |
1325 | return 0; | |
1326 | ||
1327 | handle = alloc_handle(pool); | |
1328 | if (!handle) | |
c2344348 | 1329 | return 0; |
61989a80 | 1330 | |
2e40e163 MK |
1331 | /* extra space in chunk to keep the handle */ |
1332 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1333 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1334 | |
1335 | spin_lock(&class->lock); | |
1336 | first_page = find_get_zspage(class); | |
1337 | ||
1338 | if (!first_page) { | |
1339 | spin_unlock(&class->lock); | |
1340 | first_page = alloc_zspage(class, pool->flags); | |
2e40e163 MK |
1341 | if (unlikely(!first_page)) { |
1342 | free_handle(pool, handle); | |
c2344348 | 1343 | return 0; |
2e40e163 | 1344 | } |
61989a80 NG |
1345 | |
1346 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1347 | atomic_long_add(class->pages_per_zspage, |
1348 | &pool->pages_allocated); | |
0f050d99 | 1349 | |
61989a80 | 1350 | spin_lock(&class->lock); |
0f050d99 GM |
1351 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1352 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1353 | } |
1354 | ||
c7806261 | 1355 | obj = obj_malloc(first_page, class, handle); |
61989a80 | 1356 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1357 | fix_fullness_group(class, first_page); |
2e40e163 | 1358 | record_obj(handle, obj); |
61989a80 NG |
1359 | spin_unlock(&class->lock); |
1360 | ||
2e40e163 | 1361 | return handle; |
61989a80 NG |
1362 | } |
1363 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1364 | ||
c7806261 MK |
1365 | static void obj_free(struct zs_pool *pool, struct size_class *class, |
1366 | unsigned long obj) | |
61989a80 NG |
1367 | { |
1368 | struct link_free *link; | |
1369 | struct page *first_page, *f_page; | |
c7806261 | 1370 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1371 | void *vaddr; |
61989a80 | 1372 | int class_idx; |
61989a80 NG |
1373 | enum fullness_group fullness; |
1374 | ||
c7806261 | 1375 | BUG_ON(!obj); |
61989a80 | 1376 | |
2e40e163 | 1377 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1378 | first_page = get_first_page(f_page); |
1379 | ||
1380 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
61989a80 NG |
1381 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1382 | ||
c7806261 | 1383 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1384 | |
1385 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1386 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1387 | link->next = first_page->freelist; |
af4ee5e9 | 1388 | kunmap_atomic(vaddr); |
c2344348 | 1389 | first_page->freelist = (void *)obj; |
61989a80 | 1390 | first_page->inuse--; |
0f050d99 | 1391 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1392 | } |
1393 | ||
1394 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1395 | { | |
1396 | struct page *first_page, *f_page; | |
1397 | unsigned long obj, f_objidx; | |
1398 | int class_idx; | |
1399 | struct size_class *class; | |
1400 | enum fullness_group fullness; | |
1401 | ||
1402 | if (unlikely(!handle)) | |
1403 | return; | |
1404 | ||
1405 | obj = handle_to_obj(handle); | |
1406 | free_handle(pool, handle); | |
1407 | obj_to_location(obj, &f_page, &f_objidx); | |
1408 | first_page = get_first_page(f_page); | |
1409 | ||
1410 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1411 | class = pool->size_class[class_idx]; | |
1412 | ||
1413 | spin_lock(&class->lock); | |
1414 | obj_free(pool, class, obj); | |
1415 | fullness = fix_fullness_group(class, first_page); | |
0f050d99 GM |
1416 | if (fullness == ZS_EMPTY) |
1417 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( | |
1418 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1419 | spin_unlock(&class->lock); |
1420 | ||
13de8933 MK |
1421 | if (fullness == ZS_EMPTY) { |
1422 | atomic_long_sub(class->pages_per_zspage, | |
1423 | &pool->pages_allocated); | |
61989a80 | 1424 | free_zspage(first_page); |
13de8933 | 1425 | } |
61989a80 NG |
1426 | } |
1427 | EXPORT_SYMBOL_GPL(zs_free); | |
1428 | ||
00a61d86 | 1429 | /** |
66cdef66 GM |
1430 | * zs_create_pool - Creates an allocation pool to work from. |
1431 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1432 | * |
66cdef66 GM |
1433 | * This function must be called before anything when using |
1434 | * the zsmalloc allocator. | |
166cfda7 | 1435 | * |
66cdef66 GM |
1436 | * On success, a pointer to the newly created pool is returned, |
1437 | * otherwise NULL. | |
396b7fd6 | 1438 | */ |
3eba0c6a | 1439 | struct zs_pool *zs_create_pool(char *name, gfp_t flags) |
61989a80 | 1440 | { |
66cdef66 GM |
1441 | int i; |
1442 | struct zs_pool *pool; | |
1443 | struct size_class *prev_class = NULL; | |
61989a80 | 1444 | |
66cdef66 GM |
1445 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1446 | if (!pool) | |
1447 | return NULL; | |
61989a80 | 1448 | |
66cdef66 GM |
1449 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1450 | GFP_KERNEL); | |
1451 | if (!pool->size_class) { | |
1452 | kfree(pool); | |
1453 | return NULL; | |
1454 | } | |
61989a80 | 1455 | |
2e40e163 MK |
1456 | pool->name = kstrdup(name, GFP_KERNEL); |
1457 | if (!pool->name) | |
1458 | goto err; | |
1459 | ||
1460 | if (create_handle_cache(pool)) | |
1461 | goto err; | |
1462 | ||
c60369f0 | 1463 | /* |
66cdef66 GM |
1464 | * Iterate reversly, because, size of size_class that we want to use |
1465 | * for merging should be larger or equal to current size. | |
c60369f0 | 1466 | */ |
66cdef66 GM |
1467 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1468 | int size; | |
1469 | int pages_per_zspage; | |
1470 | struct size_class *class; | |
c60369f0 | 1471 | |
66cdef66 GM |
1472 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1473 | if (size > ZS_MAX_ALLOC_SIZE) | |
1474 | size = ZS_MAX_ALLOC_SIZE; | |
1475 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1476 | |
66cdef66 GM |
1477 | /* |
1478 | * size_class is used for normal zsmalloc operation such | |
1479 | * as alloc/free for that size. Although it is natural that we | |
1480 | * have one size_class for each size, there is a chance that we | |
1481 | * can get more memory utilization if we use one size_class for | |
1482 | * many different sizes whose size_class have same | |
1483 | * characteristics. So, we makes size_class point to | |
1484 | * previous size_class if possible. | |
1485 | */ | |
1486 | if (prev_class) { | |
1487 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1488 | pool->size_class[i] = prev_class; | |
1489 | continue; | |
1490 | } | |
1491 | } | |
1492 | ||
1493 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1494 | if (!class) | |
1495 | goto err; | |
1496 | ||
1497 | class->size = size; | |
1498 | class->index = i; | |
1499 | class->pages_per_zspage = pages_per_zspage; | |
1500 | spin_lock_init(&class->lock); | |
1501 | pool->size_class[i] = class; | |
1502 | ||
1503 | prev_class = class; | |
61989a80 NG |
1504 | } |
1505 | ||
66cdef66 | 1506 | pool->flags = flags; |
b7418510 | 1507 | |
0f050d99 GM |
1508 | if (zs_pool_stat_create(name, pool)) |
1509 | goto err; | |
1510 | ||
66cdef66 GM |
1511 | return pool; |
1512 | ||
1513 | err: | |
1514 | zs_destroy_pool(pool); | |
1515 | return NULL; | |
61989a80 | 1516 | } |
66cdef66 | 1517 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1518 | |
66cdef66 | 1519 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1520 | { |
66cdef66 | 1521 | int i; |
61989a80 | 1522 | |
0f050d99 GM |
1523 | zs_pool_stat_destroy(pool); |
1524 | ||
66cdef66 GM |
1525 | for (i = 0; i < zs_size_classes; i++) { |
1526 | int fg; | |
1527 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1528 | |
66cdef66 GM |
1529 | if (!class) |
1530 | continue; | |
61989a80 | 1531 | |
66cdef66 GM |
1532 | if (class->index != i) |
1533 | continue; | |
61989a80 | 1534 | |
66cdef66 GM |
1535 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1536 | if (class->fullness_list[fg]) { | |
1537 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1538 | class->size, fg); | |
1539 | } | |
1540 | } | |
1541 | kfree(class); | |
1542 | } | |
f553646a | 1543 | |
2e40e163 | 1544 | destroy_handle_cache(pool); |
66cdef66 | 1545 | kfree(pool->size_class); |
0f050d99 | 1546 | kfree(pool->name); |
66cdef66 GM |
1547 | kfree(pool); |
1548 | } | |
1549 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 1550 | |
66cdef66 GM |
1551 | static int __init zs_init(void) |
1552 | { | |
1553 | int ret = zs_register_cpu_notifier(); | |
1554 | ||
0f050d99 GM |
1555 | if (ret) |
1556 | goto notifier_fail; | |
66cdef66 GM |
1557 | |
1558 | init_zs_size_classes(); | |
1559 | ||
1560 | #ifdef CONFIG_ZPOOL | |
1561 | zpool_register_driver(&zs_zpool_driver); | |
1562 | #endif | |
0f050d99 GM |
1563 | |
1564 | ret = zs_stat_init(); | |
1565 | if (ret) { | |
1566 | pr_err("zs stat initialization failed\n"); | |
1567 | goto stat_fail; | |
1568 | } | |
66cdef66 | 1569 | return 0; |
0f050d99 GM |
1570 | |
1571 | stat_fail: | |
1572 | #ifdef CONFIG_ZPOOL | |
1573 | zpool_unregister_driver(&zs_zpool_driver); | |
1574 | #endif | |
1575 | notifier_fail: | |
1576 | zs_unregister_cpu_notifier(); | |
1577 | ||
1578 | return ret; | |
61989a80 | 1579 | } |
61989a80 | 1580 | |
66cdef66 | 1581 | static void __exit zs_exit(void) |
61989a80 | 1582 | { |
66cdef66 GM |
1583 | #ifdef CONFIG_ZPOOL |
1584 | zpool_unregister_driver(&zs_zpool_driver); | |
1585 | #endif | |
1586 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
1587 | |
1588 | zs_stat_exit(); | |
61989a80 | 1589 | } |
069f101f BH |
1590 | |
1591 | module_init(zs_init); | |
1592 | module_exit(zs_exit); | |
1593 | ||
1594 | MODULE_LICENSE("Dual BSD/GPL"); | |
1595 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |