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 | /* |
2db51dae NG |
15 | * Following is how we use various fields and flags of underlying |
16 | * struct page(s) to form a zspage. | |
17 | * | |
18 | * Usage of struct page fields: | |
32e7ba1e | 19 | * page->private: points to the first component (0-order) page |
2db51dae NG |
20 | * page->index (union with page->freelist): offset of the first object |
21 | * starting in this page. For the first page, this is | |
22 | * always 0, so we use this field (aka freelist) to point | |
23 | * to the first free object in zspage. | |
24 | * page->lru: links together all component pages (except the first page) | |
25 | * of a zspage | |
26 | * | |
27 | * For _first_ page only: | |
28 | * | |
32e7ba1e | 29 | * page->private: refers to the component page after the first page |
7b60a685 MK |
30 | * If the page is first_page for huge object, it stores handle. |
31 | * Look at size_class->huge. | |
2db51dae NG |
32 | * page->freelist: points to the first free object in zspage. |
33 | * Free objects are linked together using in-place | |
34 | * metadata. | |
35 | * page->objects: maximum number of objects we can store in this | |
36 | * zspage (class->zspage_order * PAGE_SIZE / class->size) | |
37 | * page->lru: links together first pages of various zspages. | |
38 | * Basically forming list of zspages in a fullness group. | |
39 | * page->mapping: class index and fullness group of the zspage | |
8f958c98 | 40 | * page->inuse: the number of objects that are used in this zspage |
2db51dae NG |
41 | * |
42 | * Usage of struct page flags: | |
43 | * PG_private: identifies the first component page | |
44 | * PG_private2: identifies the last component page | |
45 | * | |
46 | */ | |
47 | ||
61989a80 NG |
48 | #include <linux/module.h> |
49 | #include <linux/kernel.h> | |
312fcae2 | 50 | #include <linux/sched.h> |
61989a80 NG |
51 | #include <linux/bitops.h> |
52 | #include <linux/errno.h> | |
53 | #include <linux/highmem.h> | |
61989a80 NG |
54 | #include <linux/string.h> |
55 | #include <linux/slab.h> | |
56 | #include <asm/tlbflush.h> | |
57 | #include <asm/pgtable.h> | |
58 | #include <linux/cpumask.h> | |
59 | #include <linux/cpu.h> | |
0cbb613f | 60 | #include <linux/vmalloc.h> |
759b26b2 | 61 | #include <linux/preempt.h> |
0959c63f SJ |
62 | #include <linux/spinlock.h> |
63 | #include <linux/types.h> | |
0f050d99 | 64 | #include <linux/debugfs.h> |
bcf1647d | 65 | #include <linux/zsmalloc.h> |
c795779d | 66 | #include <linux/zpool.h> |
0959c63f SJ |
67 | |
68 | /* | |
69 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
70 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
71 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
72 | * to restore link_free pointer values. | |
73 | */ | |
74 | #define ZS_ALIGN 8 | |
75 | ||
76 | /* | |
77 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
78 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
79 | */ | |
80 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
81 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
82 | ||
2e40e163 MK |
83 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
84 | ||
0959c63f SJ |
85 | /* |
86 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 87 | * as single (unsigned long) handle value. |
0959c63f SJ |
88 | * |
89 | * Note that object index <obj_idx> is relative to system | |
90 | * page <PFN> it is stored in, so for each sub-page belonging | |
91 | * to a zspage, obj_idx starts with 0. | |
92 | * | |
93 | * This is made more complicated by various memory models and PAE. | |
94 | */ | |
95 | ||
96 | #ifndef MAX_PHYSMEM_BITS | |
97 | #ifdef CONFIG_HIGHMEM64G | |
98 | #define MAX_PHYSMEM_BITS 36 | |
99 | #else /* !CONFIG_HIGHMEM64G */ | |
100 | /* | |
101 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
102 | * be PAGE_SHIFT | |
103 | */ | |
104 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
105 | #endif | |
106 | #endif | |
107 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
108 | |
109 | /* | |
110 | * Memory for allocating for handle keeps object position by | |
111 | * encoding <page, obj_idx> and the encoded value has a room | |
112 | * in least bit(ie, look at obj_to_location). | |
113 | * We use the bit to synchronize between object access by | |
114 | * user and migration. | |
115 | */ | |
116 | #define HANDLE_PIN_BIT 0 | |
117 | ||
118 | /* | |
119 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
120 | * to identify the object was allocated or not. | |
121 | * It's okay to add the status bit in the least bit because | |
122 | * header keeps handle which is 4byte-aligned address so we | |
123 | * have room for two bit at least. | |
124 | */ | |
125 | #define OBJ_ALLOCATED_TAG 1 | |
126 | #define OBJ_TAG_BITS 1 | |
127 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
128 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
129 | ||
130 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
131 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
132 | #define ZS_MIN_ALLOC_SIZE \ | |
133 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 134 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 135 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
136 | |
137 | /* | |
7eb52512 | 138 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
139 | * trader-off here: |
140 | * - Large number of size classes is potentially wasteful as free page are | |
141 | * spread across these classes | |
142 | * - Small number of size classes causes large internal fragmentation | |
143 | * - Probably its better to use specific size classes (empirically | |
144 | * determined). NOTE: all those class sizes must be set as multiple of | |
145 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
146 | * | |
147 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
148 | * (reason above) | |
149 | */ | |
d662b8eb | 150 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8) |
0959c63f SJ |
151 | |
152 | /* | |
153 | * We do not maintain any list for completely empty or full pages | |
154 | */ | |
155 | enum fullness_group { | |
156 | ZS_ALMOST_FULL, | |
157 | ZS_ALMOST_EMPTY, | |
158 | _ZS_NR_FULLNESS_GROUPS, | |
159 | ||
160 | ZS_EMPTY, | |
161 | ZS_FULL | |
162 | }; | |
163 | ||
0f050d99 GM |
164 | enum zs_stat_type { |
165 | OBJ_ALLOCATED, | |
166 | OBJ_USED, | |
248ca1b0 MK |
167 | CLASS_ALMOST_FULL, |
168 | CLASS_ALMOST_EMPTY, | |
0f050d99 GM |
169 | }; |
170 | ||
6fe5186f SS |
171 | #ifdef CONFIG_ZSMALLOC_STAT |
172 | #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1) | |
173 | #else | |
174 | #define NR_ZS_STAT_TYPE (OBJ_USED + 1) | |
175 | #endif | |
176 | ||
0f050d99 GM |
177 | struct zs_size_stat { |
178 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
179 | }; | |
180 | ||
57244594 SS |
181 | #ifdef CONFIG_ZSMALLOC_STAT |
182 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
183 | #endif |
184 | ||
40f9fb8c MG |
185 | /* |
186 | * number of size_classes | |
187 | */ | |
188 | static int zs_size_classes; | |
189 | ||
0959c63f SJ |
190 | /* |
191 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
192 | * n <= N / f, where | |
193 | * n = number of allocated objects | |
194 | * N = total number of objects zspage can store | |
6dd9737e | 195 | * f = fullness_threshold_frac |
0959c63f SJ |
196 | * |
197 | * Similarly, we assign zspage to: | |
198 | * ZS_ALMOST_FULL when n > N / f | |
199 | * ZS_EMPTY when n == 0 | |
200 | * ZS_FULL when n == N | |
201 | * | |
202 | * (see: fix_fullness_group()) | |
203 | */ | |
204 | static const int fullness_threshold_frac = 4; | |
205 | ||
206 | struct size_class { | |
57244594 SS |
207 | spinlock_t lock; |
208 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; | |
0959c63f SJ |
209 | /* |
210 | * Size of objects stored in this class. Must be multiple | |
211 | * of ZS_ALIGN. | |
212 | */ | |
213 | int size; | |
214 | unsigned int index; | |
215 | ||
0f050d99 | 216 | struct zs_size_stat stats; |
0959c63f | 217 | |
7dfa4612 WY |
218 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
219 | int pages_per_zspage; | |
57244594 SS |
220 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
221 | bool huge; | |
0959c63f SJ |
222 | }; |
223 | ||
224 | /* | |
225 | * Placed within free objects to form a singly linked list. | |
226 | * For every zspage, first_page->freelist gives head of this list. | |
227 | * | |
228 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
229 | */ | |
230 | struct link_free { | |
2e40e163 MK |
231 | union { |
232 | /* | |
233 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
234 | * It's valid for non-allocated object | |
235 | */ | |
236 | void *next; | |
237 | /* | |
238 | * Handle of allocated object. | |
239 | */ | |
240 | unsigned long handle; | |
241 | }; | |
0959c63f SJ |
242 | }; |
243 | ||
244 | struct zs_pool { | |
6f3526d6 | 245 | const char *name; |
0f050d99 | 246 | |
40f9fb8c | 247 | struct size_class **size_class; |
2e40e163 | 248 | struct kmem_cache *handle_cachep; |
0959c63f SJ |
249 | |
250 | gfp_t flags; /* allocation flags used when growing pool */ | |
13de8933 | 251 | atomic_long_t pages_allocated; |
0f050d99 | 252 | |
7d3f3938 | 253 | struct zs_pool_stats stats; |
ab9d306d SS |
254 | |
255 | /* Compact classes */ | |
256 | struct shrinker shrinker; | |
257 | /* | |
258 | * To signify that register_shrinker() was successful | |
259 | * and unregister_shrinker() will not Oops. | |
260 | */ | |
261 | bool shrinker_enabled; | |
0f050d99 GM |
262 | #ifdef CONFIG_ZSMALLOC_STAT |
263 | struct dentry *stat_dentry; | |
264 | #endif | |
0959c63f | 265 | }; |
61989a80 NG |
266 | |
267 | /* | |
268 | * A zspage's class index and fullness group | |
269 | * are encoded in its (first)page->mapping | |
270 | */ | |
271 | #define CLASS_IDX_BITS 28 | |
272 | #define FULLNESS_BITS 4 | |
273 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
274 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
275 | ||
f553646a | 276 | struct mapping_area { |
1b945aee | 277 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
278 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
279 | #else | |
280 | char *vm_buf; /* copy buffer for objects that span pages */ | |
281 | #endif | |
282 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
283 | enum zs_mapmode vm_mm; /* mapping mode */ | |
284 | }; | |
285 | ||
2e40e163 MK |
286 | static int create_handle_cache(struct zs_pool *pool) |
287 | { | |
288 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
289 | 0, 0, NULL); | |
290 | return pool->handle_cachep ? 0 : 1; | |
291 | } | |
292 | ||
293 | static void destroy_handle_cache(struct zs_pool *pool) | |
294 | { | |
cd10add0 | 295 | kmem_cache_destroy(pool->handle_cachep); |
2e40e163 MK |
296 | } |
297 | ||
298 | static unsigned long alloc_handle(struct zs_pool *pool) | |
299 | { | |
300 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
301 | pool->flags & ~__GFP_HIGHMEM); | |
302 | } | |
303 | ||
304 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
305 | { | |
306 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
307 | } | |
308 | ||
309 | static void record_obj(unsigned long handle, unsigned long obj) | |
310 | { | |
c102f07c JL |
311 | /* |
312 | * lsb of @obj represents handle lock while other bits | |
313 | * represent object value the handle is pointing so | |
314 | * updating shouldn't do store tearing. | |
315 | */ | |
316 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
317 | } |
318 | ||
c795779d DS |
319 | /* zpool driver */ |
320 | ||
321 | #ifdef CONFIG_ZPOOL | |
322 | ||
6f3526d6 | 323 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 324 | const struct zpool_ops *zpool_ops, |
479305fd | 325 | struct zpool *zpool) |
c795779d | 326 | { |
3eba0c6a | 327 | return zs_create_pool(name, gfp); |
c795779d DS |
328 | } |
329 | ||
330 | static void zs_zpool_destroy(void *pool) | |
331 | { | |
332 | zs_destroy_pool(pool); | |
333 | } | |
334 | ||
335 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
336 | unsigned long *handle) | |
337 | { | |
338 | *handle = zs_malloc(pool, size); | |
339 | return *handle ? 0 : -1; | |
340 | } | |
341 | static void zs_zpool_free(void *pool, unsigned long handle) | |
342 | { | |
343 | zs_free(pool, handle); | |
344 | } | |
345 | ||
346 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
347 | unsigned int *reclaimed) | |
348 | { | |
349 | return -EINVAL; | |
350 | } | |
351 | ||
352 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
353 | enum zpool_mapmode mm) | |
354 | { | |
355 | enum zs_mapmode zs_mm; | |
356 | ||
357 | switch (mm) { | |
358 | case ZPOOL_MM_RO: | |
359 | zs_mm = ZS_MM_RO; | |
360 | break; | |
361 | case ZPOOL_MM_WO: | |
362 | zs_mm = ZS_MM_WO; | |
363 | break; | |
364 | case ZPOOL_MM_RW: /* fallthru */ | |
365 | default: | |
366 | zs_mm = ZS_MM_RW; | |
367 | break; | |
368 | } | |
369 | ||
370 | return zs_map_object(pool, handle, zs_mm); | |
371 | } | |
372 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
373 | { | |
374 | zs_unmap_object(pool, handle); | |
375 | } | |
376 | ||
377 | static u64 zs_zpool_total_size(void *pool) | |
378 | { | |
722cdc17 | 379 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
380 | } |
381 | ||
382 | static struct zpool_driver zs_zpool_driver = { | |
383 | .type = "zsmalloc", | |
384 | .owner = THIS_MODULE, | |
385 | .create = zs_zpool_create, | |
386 | .destroy = zs_zpool_destroy, | |
387 | .malloc = zs_zpool_malloc, | |
388 | .free = zs_zpool_free, | |
389 | .shrink = zs_zpool_shrink, | |
390 | .map = zs_zpool_map, | |
391 | .unmap = zs_zpool_unmap, | |
392 | .total_size = zs_zpool_total_size, | |
393 | }; | |
394 | ||
137f8cff | 395 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
396 | #endif /* CONFIG_ZPOOL */ |
397 | ||
248ca1b0 MK |
398 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
399 | { | |
400 | return pages_per_zspage * PAGE_SIZE / size; | |
401 | } | |
402 | ||
61989a80 NG |
403 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
404 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
405 | ||
406 | static int is_first_page(struct page *page) | |
407 | { | |
a27545bf | 408 | return PagePrivate(page); |
61989a80 NG |
409 | } |
410 | ||
411 | static int is_last_page(struct page *page) | |
412 | { | |
a27545bf | 413 | return PagePrivate2(page); |
61989a80 NG |
414 | } |
415 | ||
a4209467 MK |
416 | static void get_zspage_mapping(struct page *first_page, |
417 | unsigned int *class_idx, | |
61989a80 NG |
418 | enum fullness_group *fullness) |
419 | { | |
420 | unsigned long m; | |
830e4bc5 | 421 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 | 422 | |
a4209467 | 423 | m = (unsigned long)first_page->mapping; |
61989a80 NG |
424 | *fullness = m & FULLNESS_MASK; |
425 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
426 | } | |
427 | ||
a4209467 MK |
428 | static void set_zspage_mapping(struct page *first_page, |
429 | unsigned int class_idx, | |
61989a80 NG |
430 | enum fullness_group fullness) |
431 | { | |
432 | unsigned long m; | |
830e4bc5 | 433 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
434 | |
435 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
436 | (fullness & FULLNESS_MASK); | |
a4209467 | 437 | first_page->mapping = (struct address_space *)m; |
61989a80 NG |
438 | } |
439 | ||
c3e3e88a NC |
440 | /* |
441 | * zsmalloc divides the pool into various size classes where each | |
442 | * class maintains a list of zspages where each zspage is divided | |
443 | * into equal sized chunks. Each allocation falls into one of these | |
444 | * classes depending on its size. This function returns index of the | |
445 | * size class which has chunk size big enough to hold the give size. | |
446 | */ | |
61989a80 NG |
447 | static int get_size_class_index(int size) |
448 | { | |
449 | int idx = 0; | |
450 | ||
451 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
452 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
453 | ZS_SIZE_CLASS_DELTA); | |
454 | ||
7b60a685 | 455 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
456 | } |
457 | ||
248ca1b0 MK |
458 | static inline void zs_stat_inc(struct size_class *class, |
459 | enum zs_stat_type type, unsigned long cnt) | |
460 | { | |
6fe5186f SS |
461 | if (type < NR_ZS_STAT_TYPE) |
462 | class->stats.objs[type] += cnt; | |
248ca1b0 MK |
463 | } |
464 | ||
465 | static inline void zs_stat_dec(struct size_class *class, | |
466 | enum zs_stat_type type, unsigned long cnt) | |
467 | { | |
6fe5186f SS |
468 | if (type < NR_ZS_STAT_TYPE) |
469 | class->stats.objs[type] -= cnt; | |
248ca1b0 MK |
470 | } |
471 | ||
472 | static inline unsigned long zs_stat_get(struct size_class *class, | |
473 | enum zs_stat_type type) | |
474 | { | |
6fe5186f SS |
475 | if (type < NR_ZS_STAT_TYPE) |
476 | return class->stats.objs[type]; | |
477 | return 0; | |
248ca1b0 MK |
478 | } |
479 | ||
57244594 SS |
480 | #ifdef CONFIG_ZSMALLOC_STAT |
481 | ||
248ca1b0 MK |
482 | static int __init zs_stat_init(void) |
483 | { | |
484 | if (!debugfs_initialized()) | |
485 | return -ENODEV; | |
486 | ||
487 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
488 | if (!zs_stat_root) | |
489 | return -ENOMEM; | |
490 | ||
491 | return 0; | |
492 | } | |
493 | ||
494 | static void __exit zs_stat_exit(void) | |
495 | { | |
496 | debugfs_remove_recursive(zs_stat_root); | |
497 | } | |
498 | ||
1120ed54 SS |
499 | static unsigned long zs_can_compact(struct size_class *class); |
500 | ||
248ca1b0 MK |
501 | static int zs_stats_size_show(struct seq_file *s, void *v) |
502 | { | |
503 | int i; | |
504 | struct zs_pool *pool = s->private; | |
505 | struct size_class *class; | |
506 | int objs_per_zspage; | |
507 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 508 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
509 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
510 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 511 | unsigned long total_freeable = 0; |
248ca1b0 | 512 | |
1120ed54 | 513 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
514 | "class", "size", "almost_full", "almost_empty", |
515 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 516 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
517 | |
518 | for (i = 0; i < zs_size_classes; i++) { | |
519 | class = pool->size_class[i]; | |
520 | ||
521 | if (class->index != i) | |
522 | continue; | |
523 | ||
524 | spin_lock(&class->lock); | |
525 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
526 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
527 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
528 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 529 | freeable = zs_can_compact(class); |
248ca1b0 MK |
530 | spin_unlock(&class->lock); |
531 | ||
532 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
533 | class->pages_per_zspage); | |
534 | pages_used = obj_allocated / objs_per_zspage * | |
535 | class->pages_per_zspage; | |
536 | ||
1120ed54 SS |
537 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
538 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
539 | i, class->size, class_almost_full, class_almost_empty, |
540 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 541 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
542 | |
543 | total_class_almost_full += class_almost_full; | |
544 | total_class_almost_empty += class_almost_empty; | |
545 | total_objs += obj_allocated; | |
546 | total_used_objs += obj_used; | |
547 | total_pages += pages_used; | |
1120ed54 | 548 | total_freeable += freeable; |
248ca1b0 MK |
549 | } |
550 | ||
551 | seq_puts(s, "\n"); | |
1120ed54 | 552 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
553 | "Total", "", total_class_almost_full, |
554 | total_class_almost_empty, total_objs, | |
1120ed54 | 555 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
556 | |
557 | return 0; | |
558 | } | |
559 | ||
560 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
561 | { | |
562 | return single_open(file, zs_stats_size_show, inode->i_private); | |
563 | } | |
564 | ||
565 | static const struct file_operations zs_stat_size_ops = { | |
566 | .open = zs_stats_size_open, | |
567 | .read = seq_read, | |
568 | .llseek = seq_lseek, | |
569 | .release = single_release, | |
570 | }; | |
571 | ||
6f3526d6 | 572 | static int zs_pool_stat_create(const char *name, struct zs_pool *pool) |
248ca1b0 MK |
573 | { |
574 | struct dentry *entry; | |
575 | ||
576 | if (!zs_stat_root) | |
577 | return -ENODEV; | |
578 | ||
579 | entry = debugfs_create_dir(name, zs_stat_root); | |
580 | if (!entry) { | |
581 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
582 | return -ENOMEM; | |
583 | } | |
584 | pool->stat_dentry = entry; | |
585 | ||
586 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
587 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
588 | if (!entry) { | |
589 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
590 | name, "classes"); | |
591 | return -ENOMEM; | |
592 | } | |
593 | ||
594 | return 0; | |
595 | } | |
596 | ||
597 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
598 | { | |
599 | debugfs_remove_recursive(pool->stat_dentry); | |
600 | } | |
601 | ||
602 | #else /* CONFIG_ZSMALLOC_STAT */ | |
248ca1b0 MK |
603 | static int __init zs_stat_init(void) |
604 | { | |
605 | return 0; | |
606 | } | |
607 | ||
608 | static void __exit zs_stat_exit(void) | |
609 | { | |
610 | } | |
611 | ||
6f3526d6 | 612 | static inline int zs_pool_stat_create(const char *name, struct zs_pool *pool) |
248ca1b0 MK |
613 | { |
614 | return 0; | |
615 | } | |
616 | ||
617 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
618 | { | |
619 | } | |
248ca1b0 MK |
620 | #endif |
621 | ||
622 | ||
c3e3e88a NC |
623 | /* |
624 | * For each size class, zspages are divided into different groups | |
625 | * depending on how "full" they are. This was done so that we could | |
626 | * easily find empty or nearly empty zspages when we try to shrink | |
627 | * the pool (not yet implemented). This function returns fullness | |
628 | * status of the given page. | |
629 | */ | |
a4209467 | 630 | static enum fullness_group get_fullness_group(struct page *first_page) |
61989a80 NG |
631 | { |
632 | int inuse, max_objects; | |
633 | enum fullness_group fg; | |
830e4bc5 MK |
634 | |
635 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); | |
61989a80 | 636 | |
a4209467 MK |
637 | inuse = first_page->inuse; |
638 | max_objects = first_page->objects; | |
61989a80 NG |
639 | |
640 | if (inuse == 0) | |
641 | fg = ZS_EMPTY; | |
642 | else if (inuse == max_objects) | |
643 | fg = ZS_FULL; | |
d3d07c92 | 644 | else if (inuse <= 3 * max_objects / fullness_threshold_frac) |
61989a80 NG |
645 | fg = ZS_ALMOST_EMPTY; |
646 | else | |
647 | fg = ZS_ALMOST_FULL; | |
648 | ||
649 | return fg; | |
650 | } | |
651 | ||
c3e3e88a NC |
652 | /* |
653 | * Each size class maintains various freelists and zspages are assigned | |
654 | * to one of these freelists based on the number of live objects they | |
655 | * have. This functions inserts the given zspage into the freelist | |
656 | * identified by <class, fullness_group>. | |
657 | */ | |
a4209467 | 658 | static void insert_zspage(struct page *first_page, struct size_class *class, |
61989a80 NG |
659 | enum fullness_group fullness) |
660 | { | |
661 | struct page **head; | |
662 | ||
830e4bc5 | 663 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
664 | |
665 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
666 | return; | |
667 | ||
248ca1b0 MK |
668 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
669 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
58f17117 SS |
670 | |
671 | head = &class->fullness_list[fullness]; | |
672 | if (!*head) { | |
a4209467 | 673 | *head = first_page; |
58f17117 SS |
674 | return; |
675 | } | |
676 | ||
677 | /* | |
678 | * We want to see more ZS_FULL pages and less almost | |
679 | * empty/full. Put pages with higher ->inuse first. | |
680 | */ | |
a4209467 MK |
681 | list_add_tail(&first_page->lru, &(*head)->lru); |
682 | if (first_page->inuse >= (*head)->inuse) | |
683 | *head = first_page; | |
61989a80 NG |
684 | } |
685 | ||
c3e3e88a NC |
686 | /* |
687 | * This function removes the given zspage from the freelist identified | |
688 | * by <class, fullness_group>. | |
689 | */ | |
a4209467 | 690 | static void remove_zspage(struct page *first_page, struct size_class *class, |
61989a80 NG |
691 | enum fullness_group fullness) |
692 | { | |
693 | struct page **head; | |
694 | ||
830e4bc5 | 695 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
696 | |
697 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
698 | return; | |
699 | ||
700 | head = &class->fullness_list[fullness]; | |
830e4bc5 | 701 | VM_BUG_ON_PAGE(!*head, first_page); |
61989a80 NG |
702 | if (list_empty(&(*head)->lru)) |
703 | *head = NULL; | |
a4209467 | 704 | else if (*head == first_page) |
61989a80 NG |
705 | *head = (struct page *)list_entry((*head)->lru.next, |
706 | struct page, lru); | |
707 | ||
a4209467 | 708 | list_del_init(&first_page->lru); |
248ca1b0 MK |
709 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
710 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
711 | } |
712 | ||
c3e3e88a NC |
713 | /* |
714 | * Each size class maintains zspages in different fullness groups depending | |
715 | * on the number of live objects they contain. When allocating or freeing | |
716 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
717 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
718 | * a status change has occurred for the given page and accordingly moves the | |
719 | * page from the freelist of the old fullness group to that of the new | |
720 | * fullness group. | |
721 | */ | |
c7806261 | 722 | static enum fullness_group fix_fullness_group(struct size_class *class, |
a4209467 | 723 | struct page *first_page) |
61989a80 NG |
724 | { |
725 | int class_idx; | |
61989a80 NG |
726 | enum fullness_group currfg, newfg; |
727 | ||
a4209467 MK |
728 | get_zspage_mapping(first_page, &class_idx, &currfg); |
729 | newfg = get_fullness_group(first_page); | |
61989a80 NG |
730 | if (newfg == currfg) |
731 | goto out; | |
732 | ||
a4209467 MK |
733 | remove_zspage(first_page, class, currfg); |
734 | insert_zspage(first_page, class, newfg); | |
735 | set_zspage_mapping(first_page, class_idx, newfg); | |
61989a80 NG |
736 | |
737 | out: | |
738 | return newfg; | |
739 | } | |
740 | ||
741 | /* | |
742 | * We have to decide on how many pages to link together | |
743 | * to form a zspage for each size class. This is important | |
744 | * to reduce wastage due to unusable space left at end of | |
745 | * each zspage which is given as: | |
888fa374 YX |
746 | * wastage = Zp % class_size |
747 | * usage = Zp - wastage | |
61989a80 NG |
748 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
749 | * | |
750 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
751 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
752 | * since then we can perfectly fit in 8 such objects. | |
753 | */ | |
2e3b6154 | 754 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
755 | { |
756 | int i, max_usedpc = 0; | |
757 | /* zspage order which gives maximum used size per KB */ | |
758 | int max_usedpc_order = 1; | |
759 | ||
84d4faab | 760 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
761 | int zspage_size; |
762 | int waste, usedpc; | |
763 | ||
764 | zspage_size = i * PAGE_SIZE; | |
765 | waste = zspage_size % class_size; | |
766 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
767 | ||
768 | if (usedpc > max_usedpc) { | |
769 | max_usedpc = usedpc; | |
770 | max_usedpc_order = i; | |
771 | } | |
772 | } | |
773 | ||
774 | return max_usedpc_order; | |
775 | } | |
776 | ||
777 | /* | |
778 | * A single 'zspage' is composed of many system pages which are | |
779 | * linked together using fields in struct page. This function finds | |
780 | * the first/head page, given any component page of a zspage. | |
781 | */ | |
782 | static struct page *get_first_page(struct page *page) | |
783 | { | |
784 | if (is_first_page(page)) | |
785 | return page; | |
786 | else | |
32e7ba1e | 787 | return (struct page *)page_private(page); |
61989a80 NG |
788 | } |
789 | ||
790 | static struct page *get_next_page(struct page *page) | |
791 | { | |
792 | struct page *next; | |
793 | ||
794 | if (is_last_page(page)) | |
795 | next = NULL; | |
796 | else if (is_first_page(page)) | |
e842b976 | 797 | next = (struct page *)page_private(page); |
61989a80 NG |
798 | else |
799 | next = list_entry(page->lru.next, struct page, lru); | |
800 | ||
801 | return next; | |
802 | } | |
803 | ||
67296874 OH |
804 | /* |
805 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 806 | * We use the least bit of handle for tagging. |
67296874 | 807 | */ |
312fcae2 | 808 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 809 | { |
312fcae2 | 810 | unsigned long obj; |
61989a80 NG |
811 | |
812 | if (!page) { | |
830e4bc5 | 813 | VM_BUG_ON(obj_idx); |
61989a80 NG |
814 | return NULL; |
815 | } | |
816 | ||
312fcae2 MK |
817 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
818 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
819 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 820 | |
312fcae2 | 821 | return (void *)obj; |
61989a80 NG |
822 | } |
823 | ||
67296874 OH |
824 | /* |
825 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
826 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 827 | * location_to_obj(). |
67296874 | 828 | */ |
312fcae2 | 829 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
830 | unsigned long *obj_idx) |
831 | { | |
312fcae2 MK |
832 | obj >>= OBJ_TAG_BITS; |
833 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
834 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
835 | } |
836 | ||
2e40e163 MK |
837 | static unsigned long handle_to_obj(unsigned long handle) |
838 | { | |
839 | return *(unsigned long *)handle; | |
840 | } | |
841 | ||
7b60a685 MK |
842 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
843 | void *obj) | |
312fcae2 | 844 | { |
7b60a685 | 845 | if (class->huge) { |
830e4bc5 | 846 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
12a7bfad | 847 | return page_private(page); |
7b60a685 MK |
848 | } else |
849 | return *(unsigned long *)obj; | |
312fcae2 MK |
850 | } |
851 | ||
61989a80 NG |
852 | static unsigned long obj_idx_to_offset(struct page *page, |
853 | unsigned long obj_idx, int class_size) | |
854 | { | |
855 | unsigned long off = 0; | |
856 | ||
857 | if (!is_first_page(page)) | |
858 | off = page->index; | |
859 | ||
860 | return off + obj_idx * class_size; | |
861 | } | |
862 | ||
312fcae2 MK |
863 | static inline int trypin_tag(unsigned long handle) |
864 | { | |
865 | unsigned long *ptr = (unsigned long *)handle; | |
866 | ||
867 | return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr); | |
868 | } | |
869 | ||
870 | static void pin_tag(unsigned long handle) | |
871 | { | |
872 | while (!trypin_tag(handle)); | |
873 | } | |
874 | ||
875 | static void unpin_tag(unsigned long handle) | |
876 | { | |
877 | unsigned long *ptr = (unsigned long *)handle; | |
878 | ||
879 | clear_bit_unlock(HANDLE_PIN_BIT, ptr); | |
880 | } | |
881 | ||
f4477e90 NG |
882 | static void reset_page(struct page *page) |
883 | { | |
884 | clear_bit(PG_private, &page->flags); | |
885 | clear_bit(PG_private_2, &page->flags); | |
886 | set_page_private(page, 0); | |
887 | page->mapping = NULL; | |
888 | page->freelist = NULL; | |
22b751c3 | 889 | page_mapcount_reset(page); |
f4477e90 NG |
890 | } |
891 | ||
61989a80 NG |
892 | static void free_zspage(struct page *first_page) |
893 | { | |
f4477e90 | 894 | struct page *nextp, *tmp, *head_extra; |
61989a80 | 895 | |
830e4bc5 MK |
896 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
897 | VM_BUG_ON_PAGE(first_page->inuse, first_page); | |
61989a80 | 898 | |
f4477e90 | 899 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 900 | |
f4477e90 | 901 | reset_page(first_page); |
61989a80 NG |
902 | __free_page(first_page); |
903 | ||
904 | /* zspage with only 1 system page */ | |
f4477e90 | 905 | if (!head_extra) |
61989a80 NG |
906 | return; |
907 | ||
f4477e90 | 908 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 909 | list_del(&nextp->lru); |
f4477e90 | 910 | reset_page(nextp); |
61989a80 NG |
911 | __free_page(nextp); |
912 | } | |
f4477e90 NG |
913 | reset_page(head_extra); |
914 | __free_page(head_extra); | |
61989a80 NG |
915 | } |
916 | ||
917 | /* Initialize a newly allocated zspage */ | |
918 | static void init_zspage(struct page *first_page, struct size_class *class) | |
919 | { | |
920 | unsigned long off = 0; | |
921 | struct page *page = first_page; | |
922 | ||
830e4bc5 MK |
923 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
924 | ||
61989a80 NG |
925 | while (page) { |
926 | struct page *next_page; | |
927 | struct link_free *link; | |
5538c562 | 928 | unsigned int i = 1; |
af4ee5e9 | 929 | void *vaddr; |
61989a80 NG |
930 | |
931 | /* | |
932 | * page->index stores offset of first object starting | |
933 | * in the page. For the first page, this is always 0, | |
934 | * so we use first_page->index (aka ->freelist) to store | |
935 | * head of corresponding zspage's freelist. | |
936 | */ | |
937 | if (page != first_page) | |
938 | page->index = off; | |
939 | ||
af4ee5e9 MK |
940 | vaddr = kmap_atomic(page); |
941 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
942 | |
943 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 944 | link->next = location_to_obj(page, i++); |
5538c562 | 945 | link += class->size / sizeof(*link); |
61989a80 NG |
946 | } |
947 | ||
948 | /* | |
949 | * We now come to the last (full or partial) object on this | |
950 | * page, which must point to the first object on the next | |
951 | * page (if present) | |
952 | */ | |
953 | next_page = get_next_page(page); | |
312fcae2 | 954 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 955 | kunmap_atomic(vaddr); |
61989a80 | 956 | page = next_page; |
5538c562 | 957 | off %= PAGE_SIZE; |
61989a80 NG |
958 | } |
959 | } | |
960 | ||
961 | /* | |
962 | * Allocate a zspage for the given size class | |
963 | */ | |
964 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
965 | { | |
966 | int i, error; | |
b4b700c5 | 967 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
968 | |
969 | /* | |
970 | * Allocate individual pages and link them together as: | |
971 | * 1. first page->private = first sub-page | |
972 | * 2. all sub-pages are linked together using page->lru | |
32e7ba1e | 973 | * 3. each sub-page is linked to the first page using page->private |
61989a80 NG |
974 | * |
975 | * For each size class, First/Head pages are linked together using | |
976 | * page->lru. Also, we set PG_private to identify the first page | |
977 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
978 | * identify the last page. | |
979 | */ | |
980 | error = -ENOMEM; | |
2e3b6154 | 981 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 982 | struct page *page; |
61989a80 NG |
983 | |
984 | page = alloc_page(flags); | |
985 | if (!page) | |
986 | goto cleanup; | |
987 | ||
988 | INIT_LIST_HEAD(&page->lru); | |
989 | if (i == 0) { /* first page */ | |
a27545bf | 990 | SetPagePrivate(page); |
61989a80 NG |
991 | set_page_private(page, 0); |
992 | first_page = page; | |
993 | first_page->inuse = 0; | |
994 | } | |
995 | if (i == 1) | |
e842b976 | 996 | set_page_private(first_page, (unsigned long)page); |
61989a80 | 997 | if (i >= 1) |
32e7ba1e | 998 | set_page_private(page, (unsigned long)first_page); |
61989a80 NG |
999 | if (i >= 2) |
1000 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 1001 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 1002 | SetPagePrivate2(page); |
61989a80 NG |
1003 | prev_page = page; |
1004 | } | |
1005 | ||
1006 | init_zspage(first_page, class); | |
1007 | ||
312fcae2 | 1008 | first_page->freelist = location_to_obj(first_page, 0); |
61989a80 | 1009 | /* Maximum number of objects we can store in this zspage */ |
2e3b6154 | 1010 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
1011 | |
1012 | error = 0; /* Success */ | |
1013 | ||
1014 | cleanup: | |
1015 | if (unlikely(error) && first_page) { | |
1016 | free_zspage(first_page); | |
1017 | first_page = NULL; | |
1018 | } | |
1019 | ||
1020 | return first_page; | |
1021 | } | |
1022 | ||
1023 | static struct page *find_get_zspage(struct size_class *class) | |
1024 | { | |
1025 | int i; | |
1026 | struct page *page; | |
1027 | ||
1028 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1029 | page = class->fullness_list[i]; | |
1030 | if (page) | |
1031 | break; | |
1032 | } | |
1033 | ||
1034 | return page; | |
1035 | } | |
1036 | ||
1b945aee | 1037 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1038 | static inline int __zs_cpu_up(struct mapping_area *area) |
1039 | { | |
1040 | /* | |
1041 | * Make sure we don't leak memory if a cpu UP notification | |
1042 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1043 | */ | |
1044 | if (area->vm) | |
1045 | return 0; | |
1046 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1047 | if (!area->vm) | |
1048 | return -ENOMEM; | |
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1053 | { | |
1054 | if (area->vm) | |
1055 | free_vm_area(area->vm); | |
1056 | area->vm = NULL; | |
1057 | } | |
1058 | ||
1059 | static inline void *__zs_map_object(struct mapping_area *area, | |
1060 | struct page *pages[2], int off, int size) | |
1061 | { | |
f6f8ed47 | 1062 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1063 | area->vm_addr = area->vm->addr; |
1064 | return area->vm_addr + off; | |
1065 | } | |
1066 | ||
1067 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1068 | struct page *pages[2], int off, int size) | |
1069 | { | |
1070 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1071 | |
d95abbbb | 1072 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1073 | } |
1074 | ||
1b945aee | 1075 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1076 | |
1077 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1078 | { | |
1079 | /* | |
1080 | * Make sure we don't leak memory if a cpu UP notification | |
1081 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1082 | */ | |
1083 | if (area->vm_buf) | |
1084 | return 0; | |
40f9fb8c | 1085 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1086 | if (!area->vm_buf) |
1087 | return -ENOMEM; | |
1088 | return 0; | |
1089 | } | |
1090 | ||
1091 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1092 | { | |
40f9fb8c | 1093 | kfree(area->vm_buf); |
f553646a SJ |
1094 | area->vm_buf = NULL; |
1095 | } | |
1096 | ||
1097 | static void *__zs_map_object(struct mapping_area *area, | |
1098 | struct page *pages[2], int off, int size) | |
5f601902 | 1099 | { |
5f601902 SJ |
1100 | int sizes[2]; |
1101 | void *addr; | |
f553646a | 1102 | char *buf = area->vm_buf; |
5f601902 | 1103 | |
f553646a SJ |
1104 | /* disable page faults to match kmap_atomic() return conditions */ |
1105 | pagefault_disable(); | |
1106 | ||
1107 | /* no read fastpath */ | |
1108 | if (area->vm_mm == ZS_MM_WO) | |
1109 | goto out; | |
5f601902 SJ |
1110 | |
1111 | sizes[0] = PAGE_SIZE - off; | |
1112 | sizes[1] = size - sizes[0]; | |
1113 | ||
5f601902 SJ |
1114 | /* copy object to per-cpu buffer */ |
1115 | addr = kmap_atomic(pages[0]); | |
1116 | memcpy(buf, addr + off, sizes[0]); | |
1117 | kunmap_atomic(addr); | |
1118 | addr = kmap_atomic(pages[1]); | |
1119 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1120 | kunmap_atomic(addr); | |
f553646a SJ |
1121 | out: |
1122 | return area->vm_buf; | |
5f601902 SJ |
1123 | } |
1124 | ||
f553646a SJ |
1125 | static void __zs_unmap_object(struct mapping_area *area, |
1126 | struct page *pages[2], int off, int size) | |
5f601902 | 1127 | { |
5f601902 SJ |
1128 | int sizes[2]; |
1129 | void *addr; | |
2e40e163 | 1130 | char *buf; |
5f601902 | 1131 | |
f553646a SJ |
1132 | /* no write fastpath */ |
1133 | if (area->vm_mm == ZS_MM_RO) | |
1134 | goto out; | |
5f601902 | 1135 | |
7b60a685 | 1136 | buf = area->vm_buf; |
a82cbf07 YX |
1137 | buf = buf + ZS_HANDLE_SIZE; |
1138 | size -= ZS_HANDLE_SIZE; | |
1139 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1140 | |
5f601902 SJ |
1141 | sizes[0] = PAGE_SIZE - off; |
1142 | sizes[1] = size - sizes[0]; | |
1143 | ||
1144 | /* copy per-cpu buffer to object */ | |
1145 | addr = kmap_atomic(pages[0]); | |
1146 | memcpy(addr + off, buf, sizes[0]); | |
1147 | kunmap_atomic(addr); | |
1148 | addr = kmap_atomic(pages[1]); | |
1149 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1150 | kunmap_atomic(addr); | |
f553646a SJ |
1151 | |
1152 | out: | |
1153 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1154 | pagefault_enable(); | |
5f601902 | 1155 | } |
61989a80 | 1156 | |
1b945aee | 1157 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1158 | |
61989a80 NG |
1159 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1160 | void *pcpu) | |
1161 | { | |
f553646a | 1162 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1163 | struct mapping_area *area; |
1164 | ||
1165 | switch (action) { | |
1166 | case CPU_UP_PREPARE: | |
1167 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1168 | ret = __zs_cpu_up(area); |
1169 | if (ret) | |
1170 | return notifier_from_errno(ret); | |
61989a80 NG |
1171 | break; |
1172 | case CPU_DEAD: | |
1173 | case CPU_UP_CANCELED: | |
1174 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1175 | __zs_cpu_down(area); |
61989a80 NG |
1176 | break; |
1177 | } | |
1178 | ||
1179 | return NOTIFY_OK; | |
1180 | } | |
1181 | ||
1182 | static struct notifier_block zs_cpu_nb = { | |
1183 | .notifier_call = zs_cpu_notifier | |
1184 | }; | |
1185 | ||
b1b00a5b | 1186 | static int zs_register_cpu_notifier(void) |
61989a80 | 1187 | { |
b1b00a5b | 1188 | int cpu, uninitialized_var(ret); |
61989a80 | 1189 | |
f0e71fcd SB |
1190 | cpu_notifier_register_begin(); |
1191 | ||
1192 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1193 | for_each_online_cpu(cpu) { |
1194 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1195 | if (notifier_to_errno(ret)) |
1196 | break; | |
61989a80 | 1197 | } |
f0e71fcd SB |
1198 | |
1199 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1200 | return notifier_to_errno(ret); |
1201 | } | |
f0e71fcd | 1202 | |
66cdef66 | 1203 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1204 | { |
66cdef66 | 1205 | int cpu; |
40f9fb8c | 1206 | |
66cdef66 | 1207 | cpu_notifier_register_begin(); |
40f9fb8c | 1208 | |
66cdef66 GM |
1209 | for_each_online_cpu(cpu) |
1210 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1211 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1212 | |
66cdef66 | 1213 | cpu_notifier_register_done(); |
b1b00a5b SS |
1214 | } |
1215 | ||
66cdef66 | 1216 | static void init_zs_size_classes(void) |
b1b00a5b | 1217 | { |
66cdef66 | 1218 | int nr; |
c795779d | 1219 | |
66cdef66 GM |
1220 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1221 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1222 | nr += 1; | |
40f9fb8c | 1223 | |
66cdef66 | 1224 | zs_size_classes = nr; |
61989a80 NG |
1225 | } |
1226 | ||
9eec4cd5 JK |
1227 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1228 | { | |
1229 | if (prev->pages_per_zspage != pages_per_zspage) | |
1230 | return false; | |
1231 | ||
1232 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1233 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1234 | return false; | |
1235 | ||
1236 | return true; | |
1237 | } | |
1238 | ||
a4209467 | 1239 | static bool zspage_full(struct page *first_page) |
312fcae2 | 1240 | { |
830e4bc5 | 1241 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
312fcae2 | 1242 | |
a4209467 | 1243 | return first_page->inuse == first_page->objects; |
312fcae2 MK |
1244 | } |
1245 | ||
66cdef66 GM |
1246 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1247 | { | |
1248 | return atomic_long_read(&pool->pages_allocated); | |
1249 | } | |
1250 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1251 | ||
4bbc0bc0 | 1252 | /** |
66cdef66 GM |
1253 | * zs_map_object - get address of allocated object from handle. |
1254 | * @pool: pool from which the object was allocated | |
1255 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1256 | * |
66cdef66 GM |
1257 | * Before using an object allocated from zs_malloc, it must be mapped using |
1258 | * this function. When done with the object, it must be unmapped using | |
1259 | * zs_unmap_object. | |
4bbc0bc0 | 1260 | * |
66cdef66 GM |
1261 | * Only one object can be mapped per cpu at a time. There is no protection |
1262 | * against nested mappings. | |
1263 | * | |
1264 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1265 | */ |
66cdef66 GM |
1266 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1267 | enum zs_mapmode mm) | |
61989a80 | 1268 | { |
66cdef66 | 1269 | struct page *page; |
2e40e163 | 1270 | unsigned long obj, obj_idx, off; |
61989a80 | 1271 | |
66cdef66 GM |
1272 | unsigned int class_idx; |
1273 | enum fullness_group fg; | |
1274 | struct size_class *class; | |
1275 | struct mapping_area *area; | |
1276 | struct page *pages[2]; | |
2e40e163 | 1277 | void *ret; |
61989a80 | 1278 | |
9eec4cd5 | 1279 | /* |
66cdef66 GM |
1280 | * Because we use per-cpu mapping areas shared among the |
1281 | * pools/users, we can't allow mapping in interrupt context | |
1282 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1283 | */ |
830e4bc5 | 1284 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1285 | |
312fcae2 MK |
1286 | /* From now on, migration cannot move the object */ |
1287 | pin_tag(handle); | |
1288 | ||
2e40e163 MK |
1289 | obj = handle_to_obj(handle); |
1290 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1291 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1292 | class = pool->size_class[class_idx]; | |
1293 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1294 | |
66cdef66 GM |
1295 | area = &get_cpu_var(zs_map_area); |
1296 | area->vm_mm = mm; | |
1297 | if (off + class->size <= PAGE_SIZE) { | |
1298 | /* this object is contained entirely within a page */ | |
1299 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1300 | ret = area->vm_addr + off; |
1301 | goto out; | |
61989a80 NG |
1302 | } |
1303 | ||
66cdef66 GM |
1304 | /* this object spans two pages */ |
1305 | pages[0] = page; | |
1306 | pages[1] = get_next_page(page); | |
1307 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1308 | |
2e40e163 MK |
1309 | ret = __zs_map_object(area, pages, off, class->size); |
1310 | out: | |
7b60a685 MK |
1311 | if (!class->huge) |
1312 | ret += ZS_HANDLE_SIZE; | |
1313 | ||
1314 | return ret; | |
61989a80 | 1315 | } |
66cdef66 | 1316 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1317 | |
66cdef66 | 1318 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1319 | { |
66cdef66 | 1320 | struct page *page; |
2e40e163 | 1321 | unsigned long obj, obj_idx, off; |
61989a80 | 1322 | |
66cdef66 GM |
1323 | unsigned int class_idx; |
1324 | enum fullness_group fg; | |
1325 | struct size_class *class; | |
1326 | struct mapping_area *area; | |
9eec4cd5 | 1327 | |
2e40e163 MK |
1328 | obj = handle_to_obj(handle); |
1329 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1330 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1331 | class = pool->size_class[class_idx]; | |
1332 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1333 | |
66cdef66 GM |
1334 | area = this_cpu_ptr(&zs_map_area); |
1335 | if (off + class->size <= PAGE_SIZE) | |
1336 | kunmap_atomic(area->vm_addr); | |
1337 | else { | |
1338 | struct page *pages[2]; | |
40f9fb8c | 1339 | |
66cdef66 GM |
1340 | pages[0] = page; |
1341 | pages[1] = get_next_page(page); | |
1342 | BUG_ON(!pages[1]); | |
1343 | ||
1344 | __zs_unmap_object(area, pages, off, class->size); | |
1345 | } | |
1346 | put_cpu_var(zs_map_area); | |
312fcae2 | 1347 | unpin_tag(handle); |
61989a80 | 1348 | } |
66cdef66 | 1349 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1350 | |
c7806261 MK |
1351 | static unsigned long obj_malloc(struct page *first_page, |
1352 | struct size_class *class, unsigned long handle) | |
1353 | { | |
1354 | unsigned long obj; | |
1355 | struct link_free *link; | |
1356 | ||
1357 | struct page *m_page; | |
1358 | unsigned long m_objidx, m_offset; | |
1359 | void *vaddr; | |
1360 | ||
312fcae2 | 1361 | handle |= OBJ_ALLOCATED_TAG; |
c7806261 MK |
1362 | obj = (unsigned long)first_page->freelist; |
1363 | obj_to_location(obj, &m_page, &m_objidx); | |
1364 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1365 | ||
1366 | vaddr = kmap_atomic(m_page); | |
1367 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1368 | first_page->freelist = link->next; | |
7b60a685 MK |
1369 | if (!class->huge) |
1370 | /* record handle in the header of allocated chunk */ | |
1371 | link->handle = handle; | |
1372 | else | |
1373 | /* record handle in first_page->private */ | |
1374 | set_page_private(first_page, handle); | |
c7806261 MK |
1375 | kunmap_atomic(vaddr); |
1376 | first_page->inuse++; | |
1377 | zs_stat_inc(class, OBJ_USED, 1); | |
1378 | ||
1379 | return obj; | |
1380 | } | |
1381 | ||
1382 | ||
61989a80 NG |
1383 | /** |
1384 | * zs_malloc - Allocate block of given size from pool. | |
1385 | * @pool: pool to allocate from | |
1386 | * @size: size of block to allocate | |
61989a80 | 1387 | * |
00a61d86 | 1388 | * On success, handle to the allocated object is returned, |
c2344348 | 1389 | * otherwise 0. |
61989a80 NG |
1390 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1391 | */ | |
c2344348 | 1392 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
61989a80 | 1393 | { |
2e40e163 | 1394 | unsigned long handle, obj; |
61989a80 | 1395 | struct size_class *class; |
c7806261 | 1396 | struct page *first_page; |
61989a80 | 1397 | |
7b60a685 | 1398 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1399 | return 0; |
1400 | ||
1401 | handle = alloc_handle(pool); | |
1402 | if (!handle) | |
c2344348 | 1403 | return 0; |
61989a80 | 1404 | |
2e40e163 MK |
1405 | /* extra space in chunk to keep the handle */ |
1406 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1407 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1408 | |
1409 | spin_lock(&class->lock); | |
1410 | first_page = find_get_zspage(class); | |
1411 | ||
1412 | if (!first_page) { | |
1413 | spin_unlock(&class->lock); | |
1414 | first_page = alloc_zspage(class, pool->flags); | |
2e40e163 MK |
1415 | if (unlikely(!first_page)) { |
1416 | free_handle(pool, handle); | |
c2344348 | 1417 | return 0; |
2e40e163 | 1418 | } |
61989a80 NG |
1419 | |
1420 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1421 | atomic_long_add(class->pages_per_zspage, |
1422 | &pool->pages_allocated); | |
0f050d99 | 1423 | |
61989a80 | 1424 | spin_lock(&class->lock); |
0f050d99 GM |
1425 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1426 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1427 | } |
1428 | ||
c7806261 | 1429 | obj = obj_malloc(first_page, class, handle); |
61989a80 | 1430 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1431 | fix_fullness_group(class, first_page); |
2e40e163 | 1432 | record_obj(handle, obj); |
61989a80 NG |
1433 | spin_unlock(&class->lock); |
1434 | ||
2e40e163 | 1435 | return handle; |
61989a80 NG |
1436 | } |
1437 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1438 | ||
c7806261 MK |
1439 | static void obj_free(struct zs_pool *pool, struct size_class *class, |
1440 | unsigned long obj) | |
61989a80 NG |
1441 | { |
1442 | struct link_free *link; | |
1443 | struct page *first_page, *f_page; | |
c7806261 | 1444 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1445 | void *vaddr; |
61989a80 | 1446 | |
312fcae2 | 1447 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1448 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1449 | first_page = get_first_page(f_page); |
1450 | ||
61989a80 NG |
1451 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1452 | ||
c7806261 | 1453 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1454 | |
1455 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1456 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1457 | link->next = first_page->freelist; |
7b60a685 MK |
1458 | if (class->huge) |
1459 | set_page_private(first_page, 0); | |
af4ee5e9 | 1460 | kunmap_atomic(vaddr); |
c2344348 | 1461 | first_page->freelist = (void *)obj; |
61989a80 | 1462 | first_page->inuse--; |
0f050d99 | 1463 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1464 | } |
1465 | ||
1466 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1467 | { | |
1468 | struct page *first_page, *f_page; | |
1469 | unsigned long obj, f_objidx; | |
1470 | int class_idx; | |
1471 | struct size_class *class; | |
1472 | enum fullness_group fullness; | |
1473 | ||
1474 | if (unlikely(!handle)) | |
1475 | return; | |
1476 | ||
312fcae2 | 1477 | pin_tag(handle); |
c7806261 | 1478 | obj = handle_to_obj(handle); |
c7806261 MK |
1479 | obj_to_location(obj, &f_page, &f_objidx); |
1480 | first_page = get_first_page(f_page); | |
1481 | ||
1482 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1483 | class = pool->size_class[class_idx]; | |
1484 | ||
1485 | spin_lock(&class->lock); | |
1486 | obj_free(pool, class, obj); | |
1487 | fullness = fix_fullness_group(class, first_page); | |
312fcae2 | 1488 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1489 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1490 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1491 | atomic_long_sub(class->pages_per_zspage, |
1492 | &pool->pages_allocated); | |
1493 | free_zspage(first_page); | |
1494 | } | |
61989a80 | 1495 | spin_unlock(&class->lock); |
312fcae2 | 1496 | unpin_tag(handle); |
61989a80 | 1497 | |
312fcae2 MK |
1498 | free_handle(pool, handle); |
1499 | } | |
1500 | EXPORT_SYMBOL_GPL(zs_free); | |
1501 | ||
0dc63d48 | 1502 | static void zs_object_copy(unsigned long dst, unsigned long src, |
312fcae2 MK |
1503 | struct size_class *class) |
1504 | { | |
1505 | struct page *s_page, *d_page; | |
1506 | unsigned long s_objidx, d_objidx; | |
1507 | unsigned long s_off, d_off; | |
1508 | void *s_addr, *d_addr; | |
1509 | int s_size, d_size, size; | |
1510 | int written = 0; | |
1511 | ||
1512 | s_size = d_size = class->size; | |
1513 | ||
1514 | obj_to_location(src, &s_page, &s_objidx); | |
1515 | obj_to_location(dst, &d_page, &d_objidx); | |
1516 | ||
1517 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1518 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1519 | ||
1520 | if (s_off + class->size > PAGE_SIZE) | |
1521 | s_size = PAGE_SIZE - s_off; | |
1522 | ||
1523 | if (d_off + class->size > PAGE_SIZE) | |
1524 | d_size = PAGE_SIZE - d_off; | |
1525 | ||
1526 | s_addr = kmap_atomic(s_page); | |
1527 | d_addr = kmap_atomic(d_page); | |
1528 | ||
1529 | while (1) { | |
1530 | size = min(s_size, d_size); | |
1531 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1532 | written += size; | |
1533 | ||
1534 | if (written == class->size) | |
1535 | break; | |
1536 | ||
495819ea SS |
1537 | s_off += size; |
1538 | s_size -= size; | |
1539 | d_off += size; | |
1540 | d_size -= size; | |
1541 | ||
1542 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1543 | kunmap_atomic(d_addr); |
1544 | kunmap_atomic(s_addr); | |
1545 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1546 | s_addr = kmap_atomic(s_page); |
1547 | d_addr = kmap_atomic(d_page); | |
1548 | s_size = class->size - written; | |
1549 | s_off = 0; | |
312fcae2 MK |
1550 | } |
1551 | ||
495819ea | 1552 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1553 | kunmap_atomic(d_addr); |
1554 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1555 | d_addr = kmap_atomic(d_page); |
1556 | d_size = class->size - written; | |
1557 | d_off = 0; | |
312fcae2 MK |
1558 | } |
1559 | } | |
1560 | ||
1561 | kunmap_atomic(d_addr); | |
1562 | kunmap_atomic(s_addr); | |
1563 | } | |
1564 | ||
1565 | /* | |
1566 | * Find alloced object in zspage from index object and | |
1567 | * return handle. | |
1568 | */ | |
1569 | static unsigned long find_alloced_obj(struct page *page, int index, | |
1570 | struct size_class *class) | |
1571 | { | |
1572 | unsigned long head; | |
1573 | int offset = 0; | |
1574 | unsigned long handle = 0; | |
1575 | void *addr = kmap_atomic(page); | |
1576 | ||
1577 | if (!is_first_page(page)) | |
1578 | offset = page->index; | |
1579 | offset += class->size * index; | |
1580 | ||
1581 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1582 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1583 | if (head & OBJ_ALLOCATED_TAG) { |
1584 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1585 | if (trypin_tag(handle)) | |
1586 | break; | |
1587 | handle = 0; | |
1588 | } | |
1589 | ||
1590 | offset += class->size; | |
1591 | index++; | |
1592 | } | |
1593 | ||
1594 | kunmap_atomic(addr); | |
1595 | return handle; | |
1596 | } | |
1597 | ||
1598 | struct zs_compact_control { | |
1599 | /* Source page for migration which could be a subpage of zspage. */ | |
1600 | struct page *s_page; | |
1601 | /* Destination page for migration which should be a first page | |
1602 | * of zspage. */ | |
1603 | struct page *d_page; | |
1604 | /* Starting object index within @s_page which used for live object | |
1605 | * in the subpage. */ | |
1606 | int index; | |
312fcae2 MK |
1607 | }; |
1608 | ||
1609 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1610 | struct zs_compact_control *cc) | |
1611 | { | |
1612 | unsigned long used_obj, free_obj; | |
1613 | unsigned long handle; | |
1614 | struct page *s_page = cc->s_page; | |
1615 | struct page *d_page = cc->d_page; | |
1616 | unsigned long index = cc->index; | |
312fcae2 MK |
1617 | int ret = 0; |
1618 | ||
1619 | while (1) { | |
1620 | handle = find_alloced_obj(s_page, index, class); | |
1621 | if (!handle) { | |
1622 | s_page = get_next_page(s_page); | |
1623 | if (!s_page) | |
1624 | break; | |
1625 | index = 0; | |
1626 | continue; | |
1627 | } | |
1628 | ||
1629 | /* Stop if there is no more space */ | |
1630 | if (zspage_full(d_page)) { | |
1631 | unpin_tag(handle); | |
1632 | ret = -ENOMEM; | |
1633 | break; | |
1634 | } | |
1635 | ||
1636 | used_obj = handle_to_obj(handle); | |
1637 | free_obj = obj_malloc(d_page, class, handle); | |
0dc63d48 | 1638 | zs_object_copy(free_obj, used_obj, class); |
312fcae2 | 1639 | index++; |
c102f07c JL |
1640 | /* |
1641 | * record_obj updates handle's value to free_obj and it will | |
1642 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1643 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1644 | * let's keep the lock bit. | |
1645 | */ | |
1646 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1647 | record_obj(handle, free_obj); |
1648 | unpin_tag(handle); | |
1649 | obj_free(pool, class, used_obj); | |
312fcae2 MK |
1650 | } |
1651 | ||
1652 | /* Remember last position in this iteration */ | |
1653 | cc->s_page = s_page; | |
1654 | cc->index = index; | |
312fcae2 MK |
1655 | |
1656 | return ret; | |
1657 | } | |
1658 | ||
0dc63d48 | 1659 | static struct page *isolate_target_page(struct size_class *class) |
312fcae2 MK |
1660 | { |
1661 | int i; | |
1662 | struct page *page; | |
1663 | ||
1664 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1665 | page = class->fullness_list[i]; | |
1666 | if (page) { | |
1667 | remove_zspage(page, class, i); | |
1668 | break; | |
1669 | } | |
1670 | } | |
1671 | ||
1672 | return page; | |
1673 | } | |
1674 | ||
860c707d SS |
1675 | /* |
1676 | * putback_zspage - add @first_page into right class's fullness list | |
1677 | * @pool: target pool | |
1678 | * @class: destination class | |
1679 | * @first_page: target page | |
1680 | * | |
1681 | * Return @fist_page's fullness_group | |
1682 | */ | |
1683 | static enum fullness_group putback_zspage(struct zs_pool *pool, | |
1684 | struct size_class *class, | |
1685 | struct page *first_page) | |
312fcae2 | 1686 | { |
312fcae2 MK |
1687 | enum fullness_group fullness; |
1688 | ||
839373e6 | 1689 | fullness = get_fullness_group(first_page); |
312fcae2 | 1690 | insert_zspage(first_page, class, fullness); |
839373e6 MK |
1691 | set_zspage_mapping(first_page, class->index, fullness); |
1692 | ||
13de8933 | 1693 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1694 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1695 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1696 | atomic_long_sub(class->pages_per_zspage, |
1697 | &pool->pages_allocated); | |
312fcae2 | 1698 | |
61989a80 | 1699 | free_zspage(first_page); |
13de8933 | 1700 | } |
860c707d SS |
1701 | |
1702 | return fullness; | |
61989a80 | 1703 | } |
312fcae2 MK |
1704 | |
1705 | static struct page *isolate_source_page(struct size_class *class) | |
1706 | { | |
ad9d5e17 MK |
1707 | int i; |
1708 | struct page *page = NULL; | |
1709 | ||
1710 | for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) { | |
1711 | page = class->fullness_list[i]; | |
1712 | if (!page) | |
1713 | continue; | |
312fcae2 | 1714 | |
ad9d5e17 MK |
1715 | remove_zspage(page, class, i); |
1716 | break; | |
1717 | } | |
312fcae2 MK |
1718 | |
1719 | return page; | |
1720 | } | |
1721 | ||
04f05909 SS |
1722 | /* |
1723 | * | |
1724 | * Based on the number of unused allocated objects calculate | |
1725 | * and return the number of pages that we can free. | |
04f05909 SS |
1726 | */ |
1727 | static unsigned long zs_can_compact(struct size_class *class) | |
1728 | { | |
1729 | unsigned long obj_wasted; | |
44f43e99 SS |
1730 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
1731 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 1732 | |
44f43e99 SS |
1733 | if (obj_allocated <= obj_used) |
1734 | return 0; | |
04f05909 | 1735 | |
44f43e99 | 1736 | obj_wasted = obj_allocated - obj_used; |
04f05909 SS |
1737 | obj_wasted /= get_maxobj_per_zspage(class->size, |
1738 | class->pages_per_zspage); | |
1739 | ||
6cbf16b3 | 1740 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
1741 | } |
1742 | ||
7d3f3938 | 1743 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1744 | { |
312fcae2 MK |
1745 | struct zs_compact_control cc; |
1746 | struct page *src_page; | |
1747 | struct page *dst_page = NULL; | |
312fcae2 | 1748 | |
312fcae2 MK |
1749 | spin_lock(&class->lock); |
1750 | while ((src_page = isolate_source_page(class))) { | |
1751 | ||
04f05909 SS |
1752 | if (!zs_can_compact(class)) |
1753 | break; | |
1754 | ||
312fcae2 MK |
1755 | cc.index = 0; |
1756 | cc.s_page = src_page; | |
1757 | ||
0dc63d48 | 1758 | while ((dst_page = isolate_target_page(class))) { |
312fcae2 MK |
1759 | cc.d_page = dst_page; |
1760 | /* | |
0dc63d48 SS |
1761 | * If there is no more space in dst_page, resched |
1762 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1763 | */ |
1764 | if (!migrate_zspage(pool, class, &cc)) | |
1765 | break; | |
1766 | ||
1767 | putback_zspage(pool, class, dst_page); | |
312fcae2 MK |
1768 | } |
1769 | ||
1770 | /* Stop if we couldn't find slot */ | |
1771 | if (dst_page == NULL) | |
1772 | break; | |
1773 | ||
1774 | putback_zspage(pool, class, dst_page); | |
860c707d | 1775 | if (putback_zspage(pool, class, src_page) == ZS_EMPTY) |
6cbf16b3 | 1776 | pool->stats.pages_compacted += class->pages_per_zspage; |
312fcae2 | 1777 | spin_unlock(&class->lock); |
312fcae2 MK |
1778 | cond_resched(); |
1779 | spin_lock(&class->lock); | |
1780 | } | |
1781 | ||
1782 | if (src_page) | |
1783 | putback_zspage(pool, class, src_page); | |
1784 | ||
7d3f3938 | 1785 | spin_unlock(&class->lock); |
312fcae2 MK |
1786 | } |
1787 | ||
1788 | unsigned long zs_compact(struct zs_pool *pool) | |
1789 | { | |
1790 | int i; | |
312fcae2 MK |
1791 | struct size_class *class; |
1792 | ||
1793 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1794 | class = pool->size_class[i]; | |
1795 | if (!class) | |
1796 | continue; | |
1797 | if (class->index != i) | |
1798 | continue; | |
7d3f3938 | 1799 | __zs_compact(pool, class); |
312fcae2 MK |
1800 | } |
1801 | ||
860c707d | 1802 | return pool->stats.pages_compacted; |
312fcae2 MK |
1803 | } |
1804 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1805 | |
7d3f3938 SS |
1806 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1807 | { | |
1808 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1809 | } | |
1810 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1811 | ||
ab9d306d SS |
1812 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
1813 | struct shrink_control *sc) | |
1814 | { | |
1815 | unsigned long pages_freed; | |
1816 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1817 | shrinker); | |
1818 | ||
1819 | pages_freed = pool->stats.pages_compacted; | |
1820 | /* | |
1821 | * Compact classes and calculate compaction delta. | |
1822 | * Can run concurrently with a manually triggered | |
1823 | * (by user) compaction. | |
1824 | */ | |
1825 | pages_freed = zs_compact(pool) - pages_freed; | |
1826 | ||
1827 | return pages_freed ? pages_freed : SHRINK_STOP; | |
1828 | } | |
1829 | ||
1830 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
1831 | struct shrink_control *sc) | |
1832 | { | |
1833 | int i; | |
1834 | struct size_class *class; | |
1835 | unsigned long pages_to_free = 0; | |
1836 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1837 | shrinker); | |
1838 | ||
ab9d306d SS |
1839 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1840 | class = pool->size_class[i]; | |
1841 | if (!class) | |
1842 | continue; | |
1843 | if (class->index != i) | |
1844 | continue; | |
1845 | ||
ab9d306d | 1846 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
1847 | } |
1848 | ||
1849 | return pages_to_free; | |
1850 | } | |
1851 | ||
1852 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
1853 | { | |
1854 | if (pool->shrinker_enabled) { | |
1855 | unregister_shrinker(&pool->shrinker); | |
1856 | pool->shrinker_enabled = false; | |
1857 | } | |
1858 | } | |
1859 | ||
1860 | static int zs_register_shrinker(struct zs_pool *pool) | |
1861 | { | |
1862 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
1863 | pool->shrinker.count_objects = zs_shrinker_count; | |
1864 | pool->shrinker.batch = 0; | |
1865 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
1866 | ||
1867 | return register_shrinker(&pool->shrinker); | |
1868 | } | |
1869 | ||
00a61d86 | 1870 | /** |
66cdef66 GM |
1871 | * zs_create_pool - Creates an allocation pool to work from. |
1872 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1873 | * |
66cdef66 GM |
1874 | * This function must be called before anything when using |
1875 | * the zsmalloc allocator. | |
166cfda7 | 1876 | * |
66cdef66 GM |
1877 | * On success, a pointer to the newly created pool is returned, |
1878 | * otherwise NULL. | |
396b7fd6 | 1879 | */ |
6f3526d6 | 1880 | struct zs_pool *zs_create_pool(const char *name, gfp_t flags) |
61989a80 | 1881 | { |
66cdef66 GM |
1882 | int i; |
1883 | struct zs_pool *pool; | |
1884 | struct size_class *prev_class = NULL; | |
61989a80 | 1885 | |
66cdef66 GM |
1886 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1887 | if (!pool) | |
1888 | return NULL; | |
61989a80 | 1889 | |
66cdef66 GM |
1890 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1891 | GFP_KERNEL); | |
1892 | if (!pool->size_class) { | |
1893 | kfree(pool); | |
1894 | return NULL; | |
1895 | } | |
61989a80 | 1896 | |
2e40e163 MK |
1897 | pool->name = kstrdup(name, GFP_KERNEL); |
1898 | if (!pool->name) | |
1899 | goto err; | |
1900 | ||
1901 | if (create_handle_cache(pool)) | |
1902 | goto err; | |
1903 | ||
c60369f0 | 1904 | /* |
66cdef66 GM |
1905 | * Iterate reversly, because, size of size_class that we want to use |
1906 | * for merging should be larger or equal to current size. | |
c60369f0 | 1907 | */ |
66cdef66 GM |
1908 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1909 | int size; | |
1910 | int pages_per_zspage; | |
1911 | struct size_class *class; | |
c60369f0 | 1912 | |
66cdef66 GM |
1913 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1914 | if (size > ZS_MAX_ALLOC_SIZE) | |
1915 | size = ZS_MAX_ALLOC_SIZE; | |
1916 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1917 | |
66cdef66 GM |
1918 | /* |
1919 | * size_class is used for normal zsmalloc operation such | |
1920 | * as alloc/free for that size. Although it is natural that we | |
1921 | * have one size_class for each size, there is a chance that we | |
1922 | * can get more memory utilization if we use one size_class for | |
1923 | * many different sizes whose size_class have same | |
1924 | * characteristics. So, we makes size_class point to | |
1925 | * previous size_class if possible. | |
1926 | */ | |
1927 | if (prev_class) { | |
1928 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1929 | pool->size_class[i] = prev_class; | |
1930 | continue; | |
1931 | } | |
1932 | } | |
1933 | ||
1934 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1935 | if (!class) | |
1936 | goto err; | |
1937 | ||
1938 | class->size = size; | |
1939 | class->index = i; | |
1940 | class->pages_per_zspage = pages_per_zspage; | |
7b60a685 MK |
1941 | if (pages_per_zspage == 1 && |
1942 | get_maxobj_per_zspage(size, pages_per_zspage) == 1) | |
1943 | class->huge = true; | |
66cdef66 GM |
1944 | spin_lock_init(&class->lock); |
1945 | pool->size_class[i] = class; | |
1946 | ||
1947 | prev_class = class; | |
61989a80 NG |
1948 | } |
1949 | ||
66cdef66 | 1950 | pool->flags = flags; |
b7418510 | 1951 | |
0f050d99 GM |
1952 | if (zs_pool_stat_create(name, pool)) |
1953 | goto err; | |
1954 | ||
ab9d306d SS |
1955 | /* |
1956 | * Not critical, we still can use the pool | |
1957 | * and user can trigger compaction manually. | |
1958 | */ | |
1959 | if (zs_register_shrinker(pool) == 0) | |
1960 | pool->shrinker_enabled = true; | |
66cdef66 GM |
1961 | return pool; |
1962 | ||
1963 | err: | |
1964 | zs_destroy_pool(pool); | |
1965 | return NULL; | |
61989a80 | 1966 | } |
66cdef66 | 1967 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1968 | |
66cdef66 | 1969 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1970 | { |
66cdef66 | 1971 | int i; |
61989a80 | 1972 | |
ab9d306d | 1973 | zs_unregister_shrinker(pool); |
0f050d99 GM |
1974 | zs_pool_stat_destroy(pool); |
1975 | ||
66cdef66 GM |
1976 | for (i = 0; i < zs_size_classes; i++) { |
1977 | int fg; | |
1978 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1979 | |
66cdef66 GM |
1980 | if (!class) |
1981 | continue; | |
61989a80 | 1982 | |
66cdef66 GM |
1983 | if (class->index != i) |
1984 | continue; | |
61989a80 | 1985 | |
66cdef66 GM |
1986 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1987 | if (class->fullness_list[fg]) { | |
1988 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1989 | class->size, fg); | |
1990 | } | |
1991 | } | |
1992 | kfree(class); | |
1993 | } | |
f553646a | 1994 | |
2e40e163 | 1995 | destroy_handle_cache(pool); |
66cdef66 | 1996 | kfree(pool->size_class); |
0f050d99 | 1997 | kfree(pool->name); |
66cdef66 GM |
1998 | kfree(pool); |
1999 | } | |
2000 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2001 | |
66cdef66 GM |
2002 | static int __init zs_init(void) |
2003 | { | |
2004 | int ret = zs_register_cpu_notifier(); | |
2005 | ||
0f050d99 GM |
2006 | if (ret) |
2007 | goto notifier_fail; | |
66cdef66 GM |
2008 | |
2009 | init_zs_size_classes(); | |
2010 | ||
2011 | #ifdef CONFIG_ZPOOL | |
2012 | zpool_register_driver(&zs_zpool_driver); | |
2013 | #endif | |
0f050d99 GM |
2014 | |
2015 | ret = zs_stat_init(); | |
2016 | if (ret) { | |
2017 | pr_err("zs stat initialization failed\n"); | |
2018 | goto stat_fail; | |
2019 | } | |
66cdef66 | 2020 | return 0; |
0f050d99 GM |
2021 | |
2022 | stat_fail: | |
2023 | #ifdef CONFIG_ZPOOL | |
2024 | zpool_unregister_driver(&zs_zpool_driver); | |
2025 | #endif | |
2026 | notifier_fail: | |
2027 | zs_unregister_cpu_notifier(); | |
2028 | ||
2029 | return ret; | |
61989a80 | 2030 | } |
61989a80 | 2031 | |
66cdef66 | 2032 | static void __exit zs_exit(void) |
61989a80 | 2033 | { |
66cdef66 GM |
2034 | #ifdef CONFIG_ZPOOL |
2035 | zpool_unregister_driver(&zs_zpool_driver); | |
2036 | #endif | |
2037 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
2038 | |
2039 | zs_stat_exit(); | |
61989a80 | 2040 | } |
069f101f BH |
2041 | |
2042 | module_init(zs_init); | |
2043 | module_exit(zs_exit); | |
2044 | ||
2045 | MODULE_LICENSE("Dual BSD/GPL"); | |
2046 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |