]>
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. | |
2db51dae NG |
35 | * page->lru: links together first pages of various zspages. |
36 | * Basically forming list of zspages in a fullness group. | |
37 | * page->mapping: class index and fullness group of the zspage | |
8f958c98 | 38 | * page->inuse: the number of objects that are used in this zspage |
2db51dae NG |
39 | * |
40 | * Usage of struct page flags: | |
41 | * PG_private: identifies the first component page | |
42 | * PG_private2: identifies the last component page | |
43 | * | |
44 | */ | |
45 | ||
4abaac9b DS |
46 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
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; | |
1fc6e27d | 214 | int objs_per_zspage; |
0959c63f SJ |
215 | unsigned int index; |
216 | ||
0f050d99 | 217 | struct zs_size_stat stats; |
0959c63f | 218 | |
7dfa4612 WY |
219 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
220 | int pages_per_zspage; | |
57244594 SS |
221 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
222 | bool huge; | |
0959c63f SJ |
223 | }; |
224 | ||
225 | /* | |
226 | * Placed within free objects to form a singly linked list. | |
227 | * For every zspage, first_page->freelist gives head of this list. | |
228 | * | |
229 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
230 | */ | |
231 | struct link_free { | |
2e40e163 MK |
232 | union { |
233 | /* | |
234 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
235 | * It's valid for non-allocated object | |
236 | */ | |
237 | void *next; | |
238 | /* | |
239 | * Handle of allocated object. | |
240 | */ | |
241 | unsigned long handle; | |
242 | }; | |
0959c63f SJ |
243 | }; |
244 | ||
245 | struct zs_pool { | |
6f3526d6 | 246 | const char *name; |
0f050d99 | 247 | |
40f9fb8c | 248 | struct size_class **size_class; |
2e40e163 | 249 | struct kmem_cache *handle_cachep; |
0959c63f | 250 | |
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 | */ | |
61989a80 | 271 | #define FULLNESS_BITS 4 |
4f42047b MK |
272 | #define CLASS_BITS 28 |
273 | ||
274 | #define FULLNESS_SHIFT 0 | |
275 | #define CLASS_SHIFT (FULLNESS_SHIFT + FULLNESS_BITS) | |
276 | ||
277 | #define FULLNESS_MASK ((1UL << FULLNESS_BITS) - 1) | |
278 | #define CLASS_MASK ((1UL << CLASS_BITS) - 1) | |
61989a80 | 279 | |
f553646a | 280 | struct mapping_area { |
1b945aee | 281 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
282 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
283 | #else | |
284 | char *vm_buf; /* copy buffer for objects that span pages */ | |
285 | #endif | |
286 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
287 | enum zs_mapmode vm_mm; /* mapping mode */ | |
288 | }; | |
289 | ||
2e40e163 MK |
290 | static int create_handle_cache(struct zs_pool *pool) |
291 | { | |
292 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
293 | 0, 0, NULL); | |
294 | return pool->handle_cachep ? 0 : 1; | |
295 | } | |
296 | ||
297 | static void destroy_handle_cache(struct zs_pool *pool) | |
298 | { | |
cd10add0 | 299 | kmem_cache_destroy(pool->handle_cachep); |
2e40e163 MK |
300 | } |
301 | ||
d0d8da2d | 302 | static unsigned long alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
303 | { |
304 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
d0d8da2d | 305 | gfp & ~__GFP_HIGHMEM); |
2e40e163 MK |
306 | } |
307 | ||
308 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
309 | { | |
310 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
311 | } | |
312 | ||
313 | static void record_obj(unsigned long handle, unsigned long obj) | |
314 | { | |
c102f07c JL |
315 | /* |
316 | * lsb of @obj represents handle lock while other bits | |
317 | * represent object value the handle is pointing so | |
318 | * updating shouldn't do store tearing. | |
319 | */ | |
320 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
321 | } |
322 | ||
c795779d DS |
323 | /* zpool driver */ |
324 | ||
325 | #ifdef CONFIG_ZPOOL | |
326 | ||
6f3526d6 | 327 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 328 | const struct zpool_ops *zpool_ops, |
479305fd | 329 | struct zpool *zpool) |
c795779d | 330 | { |
d0d8da2d SS |
331 | /* |
332 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
333 | * different contexts and its caller must provide a valid | |
334 | * gfp mask. | |
335 | */ | |
336 | return zs_create_pool(name); | |
c795779d DS |
337 | } |
338 | ||
339 | static void zs_zpool_destroy(void *pool) | |
340 | { | |
341 | zs_destroy_pool(pool); | |
342 | } | |
343 | ||
344 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
345 | unsigned long *handle) | |
346 | { | |
d0d8da2d | 347 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
348 | return *handle ? 0 : -1; |
349 | } | |
350 | static void zs_zpool_free(void *pool, unsigned long handle) | |
351 | { | |
352 | zs_free(pool, handle); | |
353 | } | |
354 | ||
355 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
356 | unsigned int *reclaimed) | |
357 | { | |
358 | return -EINVAL; | |
359 | } | |
360 | ||
361 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
362 | enum zpool_mapmode mm) | |
363 | { | |
364 | enum zs_mapmode zs_mm; | |
365 | ||
366 | switch (mm) { | |
367 | case ZPOOL_MM_RO: | |
368 | zs_mm = ZS_MM_RO; | |
369 | break; | |
370 | case ZPOOL_MM_WO: | |
371 | zs_mm = ZS_MM_WO; | |
372 | break; | |
373 | case ZPOOL_MM_RW: /* fallthru */ | |
374 | default: | |
375 | zs_mm = ZS_MM_RW; | |
376 | break; | |
377 | } | |
378 | ||
379 | return zs_map_object(pool, handle, zs_mm); | |
380 | } | |
381 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
382 | { | |
383 | zs_unmap_object(pool, handle); | |
384 | } | |
385 | ||
386 | static u64 zs_zpool_total_size(void *pool) | |
387 | { | |
722cdc17 | 388 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
389 | } |
390 | ||
391 | static struct zpool_driver zs_zpool_driver = { | |
392 | .type = "zsmalloc", | |
393 | .owner = THIS_MODULE, | |
394 | .create = zs_zpool_create, | |
395 | .destroy = zs_zpool_destroy, | |
396 | .malloc = zs_zpool_malloc, | |
397 | .free = zs_zpool_free, | |
398 | .shrink = zs_zpool_shrink, | |
399 | .map = zs_zpool_map, | |
400 | .unmap = zs_zpool_unmap, | |
401 | .total_size = zs_zpool_total_size, | |
402 | }; | |
403 | ||
137f8cff | 404 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
405 | #endif /* CONFIG_ZPOOL */ |
406 | ||
248ca1b0 MK |
407 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
408 | { | |
409 | return pages_per_zspage * PAGE_SIZE / size; | |
410 | } | |
411 | ||
61989a80 NG |
412 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
413 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
414 | ||
415 | static int is_first_page(struct page *page) | |
416 | { | |
a27545bf | 417 | return PagePrivate(page); |
61989a80 NG |
418 | } |
419 | ||
420 | static int is_last_page(struct page *page) | |
421 | { | |
a27545bf | 422 | return PagePrivate2(page); |
61989a80 NG |
423 | } |
424 | ||
4f42047b MK |
425 | static inline int get_zspage_inuse(struct page *first_page) |
426 | { | |
427 | return first_page->inuse; | |
428 | } | |
429 | ||
430 | static inline void set_zspage_inuse(struct page *first_page, int val) | |
431 | { | |
432 | first_page->inuse = val; | |
433 | } | |
434 | ||
435 | static inline void mod_zspage_inuse(struct page *first_page, int val) | |
436 | { | |
437 | first_page->inuse += val; | |
438 | } | |
439 | ||
440 | static inline int get_first_obj_offset(struct page *page) | |
441 | { | |
442 | return page->index; | |
443 | } | |
444 | ||
445 | static inline void set_first_obj_offset(struct page *page, int offset) | |
446 | { | |
447 | page->index = offset; | |
448 | } | |
449 | ||
450 | static inline unsigned long get_freeobj(struct page *first_page) | |
451 | { | |
452 | return (unsigned long)first_page->freelist; | |
453 | } | |
454 | ||
455 | static inline void set_freeobj(struct page *first_page, unsigned long obj) | |
456 | { | |
457 | first_page->freelist = (void *)obj; | |
458 | } | |
459 | ||
a4209467 MK |
460 | static void get_zspage_mapping(struct page *first_page, |
461 | unsigned int *class_idx, | |
61989a80 NG |
462 | enum fullness_group *fullness) |
463 | { | |
464 | unsigned long m; | |
830e4bc5 | 465 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 | 466 | |
a4209467 | 467 | m = (unsigned long)first_page->mapping; |
4f42047b MK |
468 | *fullness = (m >> FULLNESS_SHIFT) & FULLNESS_MASK; |
469 | *class_idx = (m >> CLASS_SHIFT) & CLASS_MASK; | |
61989a80 NG |
470 | } |
471 | ||
a4209467 MK |
472 | static void set_zspage_mapping(struct page *first_page, |
473 | unsigned int class_idx, | |
61989a80 NG |
474 | enum fullness_group fullness) |
475 | { | |
476 | unsigned long m; | |
830e4bc5 | 477 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 | 478 | |
4f42047b | 479 | m = (class_idx << CLASS_SHIFT) | (fullness << FULLNESS_SHIFT); |
a4209467 | 480 | first_page->mapping = (struct address_space *)m; |
61989a80 NG |
481 | } |
482 | ||
c3e3e88a NC |
483 | /* |
484 | * zsmalloc divides the pool into various size classes where each | |
485 | * class maintains a list of zspages where each zspage is divided | |
486 | * into equal sized chunks. Each allocation falls into one of these | |
487 | * classes depending on its size. This function returns index of the | |
488 | * size class which has chunk size big enough to hold the give size. | |
489 | */ | |
61989a80 NG |
490 | static int get_size_class_index(int size) |
491 | { | |
492 | int idx = 0; | |
493 | ||
494 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
495 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
496 | ZS_SIZE_CLASS_DELTA); | |
497 | ||
7b60a685 | 498 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
499 | } |
500 | ||
248ca1b0 MK |
501 | static inline void zs_stat_inc(struct size_class *class, |
502 | enum zs_stat_type type, unsigned long cnt) | |
503 | { | |
6fe5186f SS |
504 | if (type < NR_ZS_STAT_TYPE) |
505 | class->stats.objs[type] += cnt; | |
248ca1b0 MK |
506 | } |
507 | ||
508 | static inline void zs_stat_dec(struct size_class *class, | |
509 | enum zs_stat_type type, unsigned long cnt) | |
510 | { | |
6fe5186f SS |
511 | if (type < NR_ZS_STAT_TYPE) |
512 | class->stats.objs[type] -= cnt; | |
248ca1b0 MK |
513 | } |
514 | ||
515 | static inline unsigned long zs_stat_get(struct size_class *class, | |
516 | enum zs_stat_type type) | |
517 | { | |
6fe5186f SS |
518 | if (type < NR_ZS_STAT_TYPE) |
519 | return class->stats.objs[type]; | |
520 | return 0; | |
248ca1b0 MK |
521 | } |
522 | ||
57244594 SS |
523 | #ifdef CONFIG_ZSMALLOC_STAT |
524 | ||
4abaac9b | 525 | static void __init zs_stat_init(void) |
248ca1b0 | 526 | { |
4abaac9b DS |
527 | if (!debugfs_initialized()) { |
528 | pr_warn("debugfs not available, stat dir not created\n"); | |
529 | return; | |
530 | } | |
248ca1b0 MK |
531 | |
532 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
533 | if (!zs_stat_root) | |
4abaac9b | 534 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
535 | } |
536 | ||
537 | static void __exit zs_stat_exit(void) | |
538 | { | |
539 | debugfs_remove_recursive(zs_stat_root); | |
540 | } | |
541 | ||
1120ed54 SS |
542 | static unsigned long zs_can_compact(struct size_class *class); |
543 | ||
248ca1b0 MK |
544 | static int zs_stats_size_show(struct seq_file *s, void *v) |
545 | { | |
546 | int i; | |
547 | struct zs_pool *pool = s->private; | |
548 | struct size_class *class; | |
549 | int objs_per_zspage; | |
550 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 551 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
552 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
553 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 554 | unsigned long total_freeable = 0; |
248ca1b0 | 555 | |
1120ed54 | 556 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
557 | "class", "size", "almost_full", "almost_empty", |
558 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 559 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
560 | |
561 | for (i = 0; i < zs_size_classes; i++) { | |
562 | class = pool->size_class[i]; | |
563 | ||
564 | if (class->index != i) | |
565 | continue; | |
566 | ||
567 | spin_lock(&class->lock); | |
568 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
569 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
570 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
571 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 572 | freeable = zs_can_compact(class); |
248ca1b0 MK |
573 | spin_unlock(&class->lock); |
574 | ||
575 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
576 | class->pages_per_zspage); | |
577 | pages_used = obj_allocated / objs_per_zspage * | |
578 | class->pages_per_zspage; | |
579 | ||
1120ed54 SS |
580 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
581 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
582 | i, class->size, class_almost_full, class_almost_empty, |
583 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 584 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
585 | |
586 | total_class_almost_full += class_almost_full; | |
587 | total_class_almost_empty += class_almost_empty; | |
588 | total_objs += obj_allocated; | |
589 | total_used_objs += obj_used; | |
590 | total_pages += pages_used; | |
1120ed54 | 591 | total_freeable += freeable; |
248ca1b0 MK |
592 | } |
593 | ||
594 | seq_puts(s, "\n"); | |
1120ed54 | 595 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
596 | "Total", "", total_class_almost_full, |
597 | total_class_almost_empty, total_objs, | |
1120ed54 | 598 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
599 | |
600 | return 0; | |
601 | } | |
602 | ||
603 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
604 | { | |
605 | return single_open(file, zs_stats_size_show, inode->i_private); | |
606 | } | |
607 | ||
608 | static const struct file_operations zs_stat_size_ops = { | |
609 | .open = zs_stats_size_open, | |
610 | .read = seq_read, | |
611 | .llseek = seq_lseek, | |
612 | .release = single_release, | |
613 | }; | |
614 | ||
d34f6157 | 615 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
616 | { |
617 | struct dentry *entry; | |
618 | ||
4abaac9b DS |
619 | if (!zs_stat_root) { |
620 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 621 | return; |
4abaac9b | 622 | } |
248ca1b0 MK |
623 | |
624 | entry = debugfs_create_dir(name, zs_stat_root); | |
625 | if (!entry) { | |
626 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 627 | return; |
248ca1b0 MK |
628 | } |
629 | pool->stat_dentry = entry; | |
630 | ||
631 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
632 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
633 | if (!entry) { | |
634 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
635 | name, "classes"); | |
4abaac9b DS |
636 | debugfs_remove_recursive(pool->stat_dentry); |
637 | pool->stat_dentry = NULL; | |
248ca1b0 | 638 | } |
248ca1b0 MK |
639 | } |
640 | ||
641 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
642 | { | |
643 | debugfs_remove_recursive(pool->stat_dentry); | |
644 | } | |
645 | ||
646 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 647 | static void __init zs_stat_init(void) |
248ca1b0 | 648 | { |
248ca1b0 MK |
649 | } |
650 | ||
651 | static void __exit zs_stat_exit(void) | |
652 | { | |
653 | } | |
654 | ||
d34f6157 | 655 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 656 | { |
248ca1b0 MK |
657 | } |
658 | ||
659 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
660 | { | |
661 | } | |
248ca1b0 MK |
662 | #endif |
663 | ||
c3e3e88a NC |
664 | /* |
665 | * For each size class, zspages are divided into different groups | |
666 | * depending on how "full" they are. This was done so that we could | |
667 | * easily find empty or nearly empty zspages when we try to shrink | |
668 | * the pool (not yet implemented). This function returns fullness | |
669 | * status of the given page. | |
670 | */ | |
1fc6e27d MK |
671 | static enum fullness_group get_fullness_group(struct size_class *class, |
672 | struct page *first_page) | |
61989a80 | 673 | { |
1fc6e27d | 674 | int inuse, objs_per_zspage; |
61989a80 | 675 | enum fullness_group fg; |
830e4bc5 MK |
676 | |
677 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); | |
61989a80 | 678 | |
4f42047b | 679 | inuse = get_zspage_inuse(first_page); |
1fc6e27d | 680 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
681 | |
682 | if (inuse == 0) | |
683 | fg = ZS_EMPTY; | |
1fc6e27d | 684 | else if (inuse == objs_per_zspage) |
61989a80 | 685 | fg = ZS_FULL; |
1fc6e27d | 686 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
687 | fg = ZS_ALMOST_EMPTY; |
688 | else | |
689 | fg = ZS_ALMOST_FULL; | |
690 | ||
691 | return fg; | |
692 | } | |
693 | ||
c3e3e88a NC |
694 | /* |
695 | * Each size class maintains various freelists and zspages are assigned | |
696 | * to one of these freelists based on the number of live objects they | |
697 | * have. This functions inserts the given zspage into the freelist | |
698 | * identified by <class, fullness_group>. | |
699 | */ | |
251cbb95 MK |
700 | static void insert_zspage(struct size_class *class, |
701 | enum fullness_group fullness, | |
702 | struct page *first_page) | |
61989a80 NG |
703 | { |
704 | struct page **head; | |
705 | ||
830e4bc5 | 706 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
707 | |
708 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
709 | return; | |
710 | ||
248ca1b0 MK |
711 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
712 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
58f17117 SS |
713 | |
714 | head = &class->fullness_list[fullness]; | |
715 | if (!*head) { | |
a4209467 | 716 | *head = first_page; |
58f17117 SS |
717 | return; |
718 | } | |
719 | ||
720 | /* | |
721 | * We want to see more ZS_FULL pages and less almost | |
722 | * empty/full. Put pages with higher ->inuse first. | |
723 | */ | |
a4209467 | 724 | list_add_tail(&first_page->lru, &(*head)->lru); |
4f42047b | 725 | if (get_zspage_inuse(first_page) >= get_zspage_inuse(*head)) |
a4209467 | 726 | *head = first_page; |
61989a80 NG |
727 | } |
728 | ||
c3e3e88a NC |
729 | /* |
730 | * This function removes the given zspage from the freelist identified | |
731 | * by <class, fullness_group>. | |
732 | */ | |
251cbb95 MK |
733 | static void remove_zspage(struct size_class *class, |
734 | enum fullness_group fullness, | |
735 | struct page *first_page) | |
61989a80 NG |
736 | { |
737 | struct page **head; | |
738 | ||
830e4bc5 | 739 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
740 | |
741 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
742 | return; | |
743 | ||
744 | head = &class->fullness_list[fullness]; | |
830e4bc5 | 745 | VM_BUG_ON_PAGE(!*head, first_page); |
61989a80 NG |
746 | if (list_empty(&(*head)->lru)) |
747 | *head = NULL; | |
a4209467 | 748 | else if (*head == first_page) |
61989a80 NG |
749 | *head = (struct page *)list_entry((*head)->lru.next, |
750 | struct page, lru); | |
751 | ||
a4209467 | 752 | list_del_init(&first_page->lru); |
248ca1b0 MK |
753 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
754 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
755 | } |
756 | ||
c3e3e88a NC |
757 | /* |
758 | * Each size class maintains zspages in different fullness groups depending | |
759 | * on the number of live objects they contain. When allocating or freeing | |
760 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
761 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
762 | * a status change has occurred for the given page and accordingly moves the | |
763 | * page from the freelist of the old fullness group to that of the new | |
764 | * fullness group. | |
765 | */ | |
c7806261 | 766 | static enum fullness_group fix_fullness_group(struct size_class *class, |
a4209467 | 767 | struct page *first_page) |
61989a80 NG |
768 | { |
769 | int class_idx; | |
61989a80 NG |
770 | enum fullness_group currfg, newfg; |
771 | ||
a4209467 | 772 | get_zspage_mapping(first_page, &class_idx, &currfg); |
1fc6e27d | 773 | newfg = get_fullness_group(class, first_page); |
61989a80 NG |
774 | if (newfg == currfg) |
775 | goto out; | |
776 | ||
251cbb95 MK |
777 | remove_zspage(class, currfg, first_page); |
778 | insert_zspage(class, newfg, first_page); | |
a4209467 | 779 | set_zspage_mapping(first_page, class_idx, newfg); |
61989a80 NG |
780 | |
781 | out: | |
782 | return newfg; | |
783 | } | |
784 | ||
785 | /* | |
786 | * We have to decide on how many pages to link together | |
787 | * to form a zspage for each size class. This is important | |
788 | * to reduce wastage due to unusable space left at end of | |
789 | * each zspage which is given as: | |
888fa374 YX |
790 | * wastage = Zp % class_size |
791 | * usage = Zp - wastage | |
61989a80 NG |
792 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
793 | * | |
794 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
795 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
796 | * since then we can perfectly fit in 8 such objects. | |
797 | */ | |
2e3b6154 | 798 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
799 | { |
800 | int i, max_usedpc = 0; | |
801 | /* zspage order which gives maximum used size per KB */ | |
802 | int max_usedpc_order = 1; | |
803 | ||
84d4faab | 804 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
805 | int zspage_size; |
806 | int waste, usedpc; | |
807 | ||
808 | zspage_size = i * PAGE_SIZE; | |
809 | waste = zspage_size % class_size; | |
810 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
811 | ||
812 | if (usedpc > max_usedpc) { | |
813 | max_usedpc = usedpc; | |
814 | max_usedpc_order = i; | |
815 | } | |
816 | } | |
817 | ||
818 | return max_usedpc_order; | |
819 | } | |
820 | ||
821 | /* | |
822 | * A single 'zspage' is composed of many system pages which are | |
823 | * linked together using fields in struct page. This function finds | |
824 | * the first/head page, given any component page of a zspage. | |
825 | */ | |
826 | static struct page *get_first_page(struct page *page) | |
827 | { | |
828 | if (is_first_page(page)) | |
829 | return page; | |
830 | else | |
32e7ba1e | 831 | return (struct page *)page_private(page); |
61989a80 NG |
832 | } |
833 | ||
834 | static struct page *get_next_page(struct page *page) | |
835 | { | |
836 | struct page *next; | |
837 | ||
838 | if (is_last_page(page)) | |
839 | next = NULL; | |
840 | else if (is_first_page(page)) | |
e842b976 | 841 | next = (struct page *)page_private(page); |
61989a80 NG |
842 | else |
843 | next = list_entry(page->lru.next, struct page, lru); | |
844 | ||
845 | return next; | |
846 | } | |
847 | ||
67296874 OH |
848 | /* |
849 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 850 | * We use the least bit of handle for tagging. |
67296874 | 851 | */ |
312fcae2 | 852 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 853 | { |
312fcae2 | 854 | unsigned long obj; |
61989a80 NG |
855 | |
856 | if (!page) { | |
830e4bc5 | 857 | VM_BUG_ON(obj_idx); |
61989a80 NG |
858 | return NULL; |
859 | } | |
860 | ||
312fcae2 MK |
861 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
862 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
863 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 864 | |
312fcae2 | 865 | return (void *)obj; |
61989a80 NG |
866 | } |
867 | ||
67296874 OH |
868 | /* |
869 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
870 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 871 | * location_to_obj(). |
67296874 | 872 | */ |
312fcae2 | 873 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
874 | unsigned long *obj_idx) |
875 | { | |
312fcae2 MK |
876 | obj >>= OBJ_TAG_BITS; |
877 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
878 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
879 | } |
880 | ||
2e40e163 MK |
881 | static unsigned long handle_to_obj(unsigned long handle) |
882 | { | |
883 | return *(unsigned long *)handle; | |
884 | } | |
885 | ||
7b60a685 MK |
886 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
887 | void *obj) | |
312fcae2 | 888 | { |
7b60a685 | 889 | if (class->huge) { |
830e4bc5 | 890 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
12a7bfad | 891 | return page_private(page); |
7b60a685 MK |
892 | } else |
893 | return *(unsigned long *)obj; | |
312fcae2 MK |
894 | } |
895 | ||
61989a80 NG |
896 | static unsigned long obj_idx_to_offset(struct page *page, |
897 | unsigned long obj_idx, int class_size) | |
898 | { | |
899 | unsigned long off = 0; | |
900 | ||
901 | if (!is_first_page(page)) | |
4f42047b | 902 | off = get_first_obj_offset(page); |
61989a80 NG |
903 | |
904 | return off + obj_idx * class_size; | |
905 | } | |
906 | ||
312fcae2 MK |
907 | static inline int trypin_tag(unsigned long handle) |
908 | { | |
1b8320b6 | 909 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
910 | } |
911 | ||
912 | static void pin_tag(unsigned long handle) | |
913 | { | |
1b8320b6 | 914 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
915 | } |
916 | ||
917 | static void unpin_tag(unsigned long handle) | |
918 | { | |
1b8320b6 | 919 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
920 | } |
921 | ||
f4477e90 NG |
922 | static void reset_page(struct page *page) |
923 | { | |
924 | clear_bit(PG_private, &page->flags); | |
925 | clear_bit(PG_private_2, &page->flags); | |
926 | set_page_private(page, 0); | |
927 | page->mapping = NULL; | |
928 | page->freelist = NULL; | |
22b751c3 | 929 | page_mapcount_reset(page); |
f4477e90 NG |
930 | } |
931 | ||
61989a80 NG |
932 | static void free_zspage(struct page *first_page) |
933 | { | |
f4477e90 | 934 | struct page *nextp, *tmp, *head_extra; |
61989a80 | 935 | |
830e4bc5 | 936 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
4f42047b | 937 | VM_BUG_ON_PAGE(get_zspage_inuse(first_page), first_page); |
61989a80 | 938 | |
f4477e90 | 939 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 940 | |
f4477e90 | 941 | reset_page(first_page); |
61989a80 NG |
942 | __free_page(first_page); |
943 | ||
944 | /* zspage with only 1 system page */ | |
f4477e90 | 945 | if (!head_extra) |
61989a80 NG |
946 | return; |
947 | ||
f4477e90 | 948 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 949 | list_del(&nextp->lru); |
f4477e90 | 950 | reset_page(nextp); |
61989a80 NG |
951 | __free_page(nextp); |
952 | } | |
f4477e90 NG |
953 | reset_page(head_extra); |
954 | __free_page(head_extra); | |
61989a80 NG |
955 | } |
956 | ||
957 | /* Initialize a newly allocated zspage */ | |
251cbb95 | 958 | static void init_zspage(struct size_class *class, struct page *first_page) |
61989a80 NG |
959 | { |
960 | unsigned long off = 0; | |
961 | struct page *page = first_page; | |
962 | ||
830e4bc5 MK |
963 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
964 | ||
61989a80 NG |
965 | while (page) { |
966 | struct page *next_page; | |
967 | struct link_free *link; | |
5538c562 | 968 | unsigned int i = 1; |
af4ee5e9 | 969 | void *vaddr; |
61989a80 NG |
970 | |
971 | /* | |
972 | * page->index stores offset of first object starting | |
973 | * in the page. For the first page, this is always 0, | |
974 | * so we use first_page->index (aka ->freelist) to store | |
975 | * head of corresponding zspage's freelist. | |
976 | */ | |
977 | if (page != first_page) | |
4f42047b | 978 | set_first_obj_offset(page, off); |
61989a80 | 979 | |
af4ee5e9 MK |
980 | vaddr = kmap_atomic(page); |
981 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
982 | |
983 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 984 | link->next = location_to_obj(page, i++); |
5538c562 | 985 | link += class->size / sizeof(*link); |
61989a80 NG |
986 | } |
987 | ||
988 | /* | |
989 | * We now come to the last (full or partial) object on this | |
990 | * page, which must point to the first object on the next | |
991 | * page (if present) | |
992 | */ | |
993 | next_page = get_next_page(page); | |
312fcae2 | 994 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 995 | kunmap_atomic(vaddr); |
61989a80 | 996 | page = next_page; |
5538c562 | 997 | off %= PAGE_SIZE; |
61989a80 NG |
998 | } |
999 | } | |
1000 | ||
1001 | /* | |
1002 | * Allocate a zspage for the given size class | |
1003 | */ | |
1004 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
1005 | { | |
1006 | int i, error; | |
b4b700c5 | 1007 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
1008 | |
1009 | /* | |
1010 | * Allocate individual pages and link them together as: | |
1011 | * 1. first page->private = first sub-page | |
1012 | * 2. all sub-pages are linked together using page->lru | |
32e7ba1e | 1013 | * 3. each sub-page is linked to the first page using page->private |
61989a80 NG |
1014 | * |
1015 | * For each size class, First/Head pages are linked together using | |
1016 | * page->lru. Also, we set PG_private to identify the first page | |
1017 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
1018 | * identify the last page. | |
1019 | */ | |
1020 | error = -ENOMEM; | |
2e3b6154 | 1021 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 1022 | struct page *page; |
61989a80 NG |
1023 | |
1024 | page = alloc_page(flags); | |
1025 | if (!page) | |
1026 | goto cleanup; | |
1027 | ||
1028 | INIT_LIST_HEAD(&page->lru); | |
1029 | if (i == 0) { /* first page */ | |
a27545bf | 1030 | SetPagePrivate(page); |
61989a80 NG |
1031 | set_page_private(page, 0); |
1032 | first_page = page; | |
4f42047b | 1033 | set_zspage_inuse(first_page, 0); |
61989a80 NG |
1034 | } |
1035 | if (i == 1) | |
e842b976 | 1036 | set_page_private(first_page, (unsigned long)page); |
61989a80 | 1037 | if (i >= 1) |
32e7ba1e | 1038 | set_page_private(page, (unsigned long)first_page); |
61989a80 NG |
1039 | if (i >= 2) |
1040 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 1041 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 1042 | SetPagePrivate2(page); |
61989a80 NG |
1043 | prev_page = page; |
1044 | } | |
1045 | ||
251cbb95 | 1046 | init_zspage(class, first_page); |
61989a80 | 1047 | |
4f42047b | 1048 | set_freeobj(first_page, (unsigned long)location_to_obj(first_page, 0)); |
61989a80 NG |
1049 | error = 0; /* Success */ |
1050 | ||
1051 | cleanup: | |
1052 | if (unlikely(error) && first_page) { | |
1053 | free_zspage(first_page); | |
1054 | first_page = NULL; | |
1055 | } | |
1056 | ||
1057 | return first_page; | |
1058 | } | |
1059 | ||
1060 | static struct page *find_get_zspage(struct size_class *class) | |
1061 | { | |
1062 | int i; | |
1063 | struct page *page; | |
1064 | ||
1065 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1066 | page = class->fullness_list[i]; | |
1067 | if (page) | |
1068 | break; | |
1069 | } | |
1070 | ||
1071 | return page; | |
1072 | } | |
1073 | ||
1b945aee | 1074 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1075 | static inline int __zs_cpu_up(struct mapping_area *area) |
1076 | { | |
1077 | /* | |
1078 | * Make sure we don't leak memory if a cpu UP notification | |
1079 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1080 | */ | |
1081 | if (area->vm) | |
1082 | return 0; | |
1083 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1084 | if (!area->vm) | |
1085 | return -ENOMEM; | |
1086 | return 0; | |
1087 | } | |
1088 | ||
1089 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1090 | { | |
1091 | if (area->vm) | |
1092 | free_vm_area(area->vm); | |
1093 | area->vm = NULL; | |
1094 | } | |
1095 | ||
1096 | static inline void *__zs_map_object(struct mapping_area *area, | |
1097 | struct page *pages[2], int off, int size) | |
1098 | { | |
f6f8ed47 | 1099 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1100 | area->vm_addr = area->vm->addr; |
1101 | return area->vm_addr + off; | |
1102 | } | |
1103 | ||
1104 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1105 | struct page *pages[2], int off, int size) | |
1106 | { | |
1107 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1108 | |
d95abbbb | 1109 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1110 | } |
1111 | ||
1b945aee | 1112 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1113 | |
1114 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1115 | { | |
1116 | /* | |
1117 | * Make sure we don't leak memory if a cpu UP notification | |
1118 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1119 | */ | |
1120 | if (area->vm_buf) | |
1121 | return 0; | |
40f9fb8c | 1122 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1123 | if (!area->vm_buf) |
1124 | return -ENOMEM; | |
1125 | return 0; | |
1126 | } | |
1127 | ||
1128 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1129 | { | |
40f9fb8c | 1130 | kfree(area->vm_buf); |
f553646a SJ |
1131 | area->vm_buf = NULL; |
1132 | } | |
1133 | ||
1134 | static void *__zs_map_object(struct mapping_area *area, | |
1135 | struct page *pages[2], int off, int size) | |
5f601902 | 1136 | { |
5f601902 SJ |
1137 | int sizes[2]; |
1138 | void *addr; | |
f553646a | 1139 | char *buf = area->vm_buf; |
5f601902 | 1140 | |
f553646a SJ |
1141 | /* disable page faults to match kmap_atomic() return conditions */ |
1142 | pagefault_disable(); | |
1143 | ||
1144 | /* no read fastpath */ | |
1145 | if (area->vm_mm == ZS_MM_WO) | |
1146 | goto out; | |
5f601902 SJ |
1147 | |
1148 | sizes[0] = PAGE_SIZE - off; | |
1149 | sizes[1] = size - sizes[0]; | |
1150 | ||
5f601902 SJ |
1151 | /* copy object to per-cpu buffer */ |
1152 | addr = kmap_atomic(pages[0]); | |
1153 | memcpy(buf, addr + off, sizes[0]); | |
1154 | kunmap_atomic(addr); | |
1155 | addr = kmap_atomic(pages[1]); | |
1156 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1157 | kunmap_atomic(addr); | |
f553646a SJ |
1158 | out: |
1159 | return area->vm_buf; | |
5f601902 SJ |
1160 | } |
1161 | ||
f553646a SJ |
1162 | static void __zs_unmap_object(struct mapping_area *area, |
1163 | struct page *pages[2], int off, int size) | |
5f601902 | 1164 | { |
5f601902 SJ |
1165 | int sizes[2]; |
1166 | void *addr; | |
2e40e163 | 1167 | char *buf; |
5f601902 | 1168 | |
f553646a SJ |
1169 | /* no write fastpath */ |
1170 | if (area->vm_mm == ZS_MM_RO) | |
1171 | goto out; | |
5f601902 | 1172 | |
7b60a685 | 1173 | buf = area->vm_buf; |
a82cbf07 YX |
1174 | buf = buf + ZS_HANDLE_SIZE; |
1175 | size -= ZS_HANDLE_SIZE; | |
1176 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1177 | |
5f601902 SJ |
1178 | sizes[0] = PAGE_SIZE - off; |
1179 | sizes[1] = size - sizes[0]; | |
1180 | ||
1181 | /* copy per-cpu buffer to object */ | |
1182 | addr = kmap_atomic(pages[0]); | |
1183 | memcpy(addr + off, buf, sizes[0]); | |
1184 | kunmap_atomic(addr); | |
1185 | addr = kmap_atomic(pages[1]); | |
1186 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1187 | kunmap_atomic(addr); | |
f553646a SJ |
1188 | |
1189 | out: | |
1190 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1191 | pagefault_enable(); | |
5f601902 | 1192 | } |
61989a80 | 1193 | |
1b945aee | 1194 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1195 | |
61989a80 NG |
1196 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1197 | void *pcpu) | |
1198 | { | |
f553646a | 1199 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1200 | struct mapping_area *area; |
1201 | ||
1202 | switch (action) { | |
1203 | case CPU_UP_PREPARE: | |
1204 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1205 | ret = __zs_cpu_up(area); |
1206 | if (ret) | |
1207 | return notifier_from_errno(ret); | |
61989a80 NG |
1208 | break; |
1209 | case CPU_DEAD: | |
1210 | case CPU_UP_CANCELED: | |
1211 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1212 | __zs_cpu_down(area); |
61989a80 NG |
1213 | break; |
1214 | } | |
1215 | ||
1216 | return NOTIFY_OK; | |
1217 | } | |
1218 | ||
1219 | static struct notifier_block zs_cpu_nb = { | |
1220 | .notifier_call = zs_cpu_notifier | |
1221 | }; | |
1222 | ||
b1b00a5b | 1223 | static int zs_register_cpu_notifier(void) |
61989a80 | 1224 | { |
b1b00a5b | 1225 | int cpu, uninitialized_var(ret); |
61989a80 | 1226 | |
f0e71fcd SB |
1227 | cpu_notifier_register_begin(); |
1228 | ||
1229 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1230 | for_each_online_cpu(cpu) { |
1231 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1232 | if (notifier_to_errno(ret)) |
1233 | break; | |
61989a80 | 1234 | } |
f0e71fcd SB |
1235 | |
1236 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1237 | return notifier_to_errno(ret); |
1238 | } | |
f0e71fcd | 1239 | |
66cdef66 | 1240 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1241 | { |
66cdef66 | 1242 | int cpu; |
40f9fb8c | 1243 | |
66cdef66 | 1244 | cpu_notifier_register_begin(); |
40f9fb8c | 1245 | |
66cdef66 GM |
1246 | for_each_online_cpu(cpu) |
1247 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1248 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1249 | |
66cdef66 | 1250 | cpu_notifier_register_done(); |
b1b00a5b SS |
1251 | } |
1252 | ||
66cdef66 | 1253 | static void init_zs_size_classes(void) |
b1b00a5b | 1254 | { |
66cdef66 | 1255 | int nr; |
c795779d | 1256 | |
66cdef66 GM |
1257 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1258 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1259 | nr += 1; | |
40f9fb8c | 1260 | |
66cdef66 | 1261 | zs_size_classes = nr; |
61989a80 NG |
1262 | } |
1263 | ||
9eec4cd5 JK |
1264 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1265 | { | |
1266 | if (prev->pages_per_zspage != pages_per_zspage) | |
1267 | return false; | |
1268 | ||
1269 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1270 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1271 | return false; | |
1272 | ||
1273 | return true; | |
1274 | } | |
1275 | ||
1fc6e27d | 1276 | static bool zspage_full(struct size_class *class, struct page *first_page) |
312fcae2 | 1277 | { |
830e4bc5 | 1278 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
312fcae2 | 1279 | |
4f42047b | 1280 | return get_zspage_inuse(first_page) == class->objs_per_zspage; |
312fcae2 MK |
1281 | } |
1282 | ||
66cdef66 GM |
1283 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1284 | { | |
1285 | return atomic_long_read(&pool->pages_allocated); | |
1286 | } | |
1287 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1288 | ||
4bbc0bc0 | 1289 | /** |
66cdef66 GM |
1290 | * zs_map_object - get address of allocated object from handle. |
1291 | * @pool: pool from which the object was allocated | |
1292 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1293 | * |
66cdef66 GM |
1294 | * Before using an object allocated from zs_malloc, it must be mapped using |
1295 | * this function. When done with the object, it must be unmapped using | |
1296 | * zs_unmap_object. | |
4bbc0bc0 | 1297 | * |
66cdef66 GM |
1298 | * Only one object can be mapped per cpu at a time. There is no protection |
1299 | * against nested mappings. | |
1300 | * | |
1301 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1302 | */ |
66cdef66 GM |
1303 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1304 | enum zs_mapmode mm) | |
61989a80 | 1305 | { |
66cdef66 | 1306 | struct page *page; |
2e40e163 | 1307 | unsigned long obj, obj_idx, off; |
61989a80 | 1308 | |
66cdef66 GM |
1309 | unsigned int class_idx; |
1310 | enum fullness_group fg; | |
1311 | struct size_class *class; | |
1312 | struct mapping_area *area; | |
1313 | struct page *pages[2]; | |
2e40e163 | 1314 | void *ret; |
61989a80 | 1315 | |
9eec4cd5 | 1316 | /* |
66cdef66 GM |
1317 | * Because we use per-cpu mapping areas shared among the |
1318 | * pools/users, we can't allow mapping in interrupt context | |
1319 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1320 | */ |
830e4bc5 | 1321 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1322 | |
312fcae2 MK |
1323 | /* From now on, migration cannot move the object */ |
1324 | pin_tag(handle); | |
1325 | ||
2e40e163 MK |
1326 | obj = handle_to_obj(handle); |
1327 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1328 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1329 | class = pool->size_class[class_idx]; | |
1330 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1331 | |
66cdef66 GM |
1332 | area = &get_cpu_var(zs_map_area); |
1333 | area->vm_mm = mm; | |
1334 | if (off + class->size <= PAGE_SIZE) { | |
1335 | /* this object is contained entirely within a page */ | |
1336 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1337 | ret = area->vm_addr + off; |
1338 | goto out; | |
61989a80 NG |
1339 | } |
1340 | ||
66cdef66 GM |
1341 | /* this object spans two pages */ |
1342 | pages[0] = page; | |
1343 | pages[1] = get_next_page(page); | |
1344 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1345 | |
2e40e163 MK |
1346 | ret = __zs_map_object(area, pages, off, class->size); |
1347 | out: | |
7b60a685 MK |
1348 | if (!class->huge) |
1349 | ret += ZS_HANDLE_SIZE; | |
1350 | ||
1351 | return ret; | |
61989a80 | 1352 | } |
66cdef66 | 1353 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1354 | |
66cdef66 | 1355 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1356 | { |
66cdef66 | 1357 | struct page *page; |
2e40e163 | 1358 | unsigned long obj, obj_idx, off; |
61989a80 | 1359 | |
66cdef66 GM |
1360 | unsigned int class_idx; |
1361 | enum fullness_group fg; | |
1362 | struct size_class *class; | |
1363 | struct mapping_area *area; | |
9eec4cd5 | 1364 | |
2e40e163 MK |
1365 | obj = handle_to_obj(handle); |
1366 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1367 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1368 | class = pool->size_class[class_idx]; | |
1369 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1370 | |
66cdef66 GM |
1371 | area = this_cpu_ptr(&zs_map_area); |
1372 | if (off + class->size <= PAGE_SIZE) | |
1373 | kunmap_atomic(area->vm_addr); | |
1374 | else { | |
1375 | struct page *pages[2]; | |
40f9fb8c | 1376 | |
66cdef66 GM |
1377 | pages[0] = page; |
1378 | pages[1] = get_next_page(page); | |
1379 | BUG_ON(!pages[1]); | |
1380 | ||
1381 | __zs_unmap_object(area, pages, off, class->size); | |
1382 | } | |
1383 | put_cpu_var(zs_map_area); | |
312fcae2 | 1384 | unpin_tag(handle); |
61989a80 | 1385 | } |
66cdef66 | 1386 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1387 | |
251cbb95 MK |
1388 | static unsigned long obj_malloc(struct size_class *class, |
1389 | struct page *first_page, unsigned long handle) | |
c7806261 MK |
1390 | { |
1391 | unsigned long obj; | |
1392 | struct link_free *link; | |
1393 | ||
1394 | struct page *m_page; | |
1395 | unsigned long m_objidx, m_offset; | |
1396 | void *vaddr; | |
1397 | ||
312fcae2 | 1398 | handle |= OBJ_ALLOCATED_TAG; |
4f42047b | 1399 | obj = get_freeobj(first_page); |
c7806261 MK |
1400 | obj_to_location(obj, &m_page, &m_objidx); |
1401 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1402 | ||
1403 | vaddr = kmap_atomic(m_page); | |
1404 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
4f42047b | 1405 | set_freeobj(first_page, (unsigned long)link->next); |
7b60a685 MK |
1406 | if (!class->huge) |
1407 | /* record handle in the header of allocated chunk */ | |
1408 | link->handle = handle; | |
1409 | else | |
1410 | /* record handle in first_page->private */ | |
1411 | set_page_private(first_page, handle); | |
c7806261 | 1412 | kunmap_atomic(vaddr); |
4f42047b | 1413 | mod_zspage_inuse(first_page, 1); |
c7806261 MK |
1414 | zs_stat_inc(class, OBJ_USED, 1); |
1415 | ||
1416 | return obj; | |
1417 | } | |
1418 | ||
1419 | ||
61989a80 NG |
1420 | /** |
1421 | * zs_malloc - Allocate block of given size from pool. | |
1422 | * @pool: pool to allocate from | |
1423 | * @size: size of block to allocate | |
61989a80 | 1424 | * |
00a61d86 | 1425 | * On success, handle to the allocated object is returned, |
c2344348 | 1426 | * otherwise 0. |
61989a80 NG |
1427 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1428 | */ | |
d0d8da2d | 1429 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1430 | { |
2e40e163 | 1431 | unsigned long handle, obj; |
61989a80 | 1432 | struct size_class *class; |
c7806261 | 1433 | struct page *first_page; |
61989a80 | 1434 | |
7b60a685 | 1435 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1436 | return 0; |
1437 | ||
d0d8da2d | 1438 | handle = alloc_handle(pool, gfp); |
2e40e163 | 1439 | if (!handle) |
c2344348 | 1440 | return 0; |
61989a80 | 1441 | |
2e40e163 MK |
1442 | /* extra space in chunk to keep the handle */ |
1443 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1444 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1445 | |
1446 | spin_lock(&class->lock); | |
1447 | first_page = find_get_zspage(class); | |
1448 | ||
1449 | if (!first_page) { | |
1450 | spin_unlock(&class->lock); | |
d0d8da2d | 1451 | first_page = alloc_zspage(class, gfp); |
2e40e163 MK |
1452 | if (unlikely(!first_page)) { |
1453 | free_handle(pool, handle); | |
c2344348 | 1454 | return 0; |
2e40e163 | 1455 | } |
61989a80 NG |
1456 | |
1457 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1458 | atomic_long_add(class->pages_per_zspage, |
1459 | &pool->pages_allocated); | |
0f050d99 | 1460 | |
61989a80 | 1461 | spin_lock(&class->lock); |
0f050d99 GM |
1462 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1463 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1464 | } |
1465 | ||
251cbb95 | 1466 | obj = obj_malloc(class, first_page, handle); |
61989a80 | 1467 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1468 | fix_fullness_group(class, first_page); |
2e40e163 | 1469 | record_obj(handle, obj); |
61989a80 NG |
1470 | spin_unlock(&class->lock); |
1471 | ||
2e40e163 | 1472 | return handle; |
61989a80 NG |
1473 | } |
1474 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1475 | ||
1ee47165 | 1476 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1477 | { |
1478 | struct link_free *link; | |
1479 | struct page *first_page, *f_page; | |
c7806261 | 1480 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1481 | void *vaddr; |
61989a80 | 1482 | |
312fcae2 | 1483 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1484 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1485 | first_page = get_first_page(f_page); |
1486 | ||
61989a80 NG |
1487 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1488 | ||
c7806261 | 1489 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1490 | |
1491 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1492 | link = (struct link_free *)(vaddr + f_offset); |
4f42047b | 1493 | link->next = (void *)get_freeobj(first_page); |
7b60a685 MK |
1494 | if (class->huge) |
1495 | set_page_private(first_page, 0); | |
af4ee5e9 | 1496 | kunmap_atomic(vaddr); |
4f42047b MK |
1497 | set_freeobj(first_page, obj); |
1498 | mod_zspage_inuse(first_page, -1); | |
0f050d99 | 1499 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1500 | } |
1501 | ||
1502 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1503 | { | |
1504 | struct page *first_page, *f_page; | |
1505 | unsigned long obj, f_objidx; | |
1506 | int class_idx; | |
1507 | struct size_class *class; | |
1508 | enum fullness_group fullness; | |
1509 | ||
1510 | if (unlikely(!handle)) | |
1511 | return; | |
1512 | ||
312fcae2 | 1513 | pin_tag(handle); |
c7806261 | 1514 | obj = handle_to_obj(handle); |
c7806261 MK |
1515 | obj_to_location(obj, &f_page, &f_objidx); |
1516 | first_page = get_first_page(f_page); | |
1517 | ||
1518 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1519 | class = pool->size_class[class_idx]; | |
1520 | ||
1521 | spin_lock(&class->lock); | |
1ee47165 | 1522 | obj_free(class, obj); |
c7806261 | 1523 | fullness = fix_fullness_group(class, first_page); |
312fcae2 | 1524 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1525 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1526 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1527 | atomic_long_sub(class->pages_per_zspage, |
1528 | &pool->pages_allocated); | |
1529 | free_zspage(first_page); | |
1530 | } | |
61989a80 | 1531 | spin_unlock(&class->lock); |
312fcae2 | 1532 | unpin_tag(handle); |
61989a80 | 1533 | |
312fcae2 MK |
1534 | free_handle(pool, handle); |
1535 | } | |
1536 | EXPORT_SYMBOL_GPL(zs_free); | |
1537 | ||
251cbb95 MK |
1538 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1539 | unsigned long src) | |
312fcae2 MK |
1540 | { |
1541 | struct page *s_page, *d_page; | |
1542 | unsigned long s_objidx, d_objidx; | |
1543 | unsigned long s_off, d_off; | |
1544 | void *s_addr, *d_addr; | |
1545 | int s_size, d_size, size; | |
1546 | int written = 0; | |
1547 | ||
1548 | s_size = d_size = class->size; | |
1549 | ||
1550 | obj_to_location(src, &s_page, &s_objidx); | |
1551 | obj_to_location(dst, &d_page, &d_objidx); | |
1552 | ||
1553 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1554 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1555 | ||
1556 | if (s_off + class->size > PAGE_SIZE) | |
1557 | s_size = PAGE_SIZE - s_off; | |
1558 | ||
1559 | if (d_off + class->size > PAGE_SIZE) | |
1560 | d_size = PAGE_SIZE - d_off; | |
1561 | ||
1562 | s_addr = kmap_atomic(s_page); | |
1563 | d_addr = kmap_atomic(d_page); | |
1564 | ||
1565 | while (1) { | |
1566 | size = min(s_size, d_size); | |
1567 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1568 | written += size; | |
1569 | ||
1570 | if (written == class->size) | |
1571 | break; | |
1572 | ||
495819ea SS |
1573 | s_off += size; |
1574 | s_size -= size; | |
1575 | d_off += size; | |
1576 | d_size -= size; | |
1577 | ||
1578 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1579 | kunmap_atomic(d_addr); |
1580 | kunmap_atomic(s_addr); | |
1581 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1582 | s_addr = kmap_atomic(s_page); |
1583 | d_addr = kmap_atomic(d_page); | |
1584 | s_size = class->size - written; | |
1585 | s_off = 0; | |
312fcae2 MK |
1586 | } |
1587 | ||
495819ea | 1588 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1589 | kunmap_atomic(d_addr); |
1590 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1591 | d_addr = kmap_atomic(d_page); |
1592 | d_size = class->size - written; | |
1593 | d_off = 0; | |
312fcae2 MK |
1594 | } |
1595 | } | |
1596 | ||
1597 | kunmap_atomic(d_addr); | |
1598 | kunmap_atomic(s_addr); | |
1599 | } | |
1600 | ||
1601 | /* | |
1602 | * Find alloced object in zspage from index object and | |
1603 | * return handle. | |
1604 | */ | |
251cbb95 MK |
1605 | static unsigned long find_alloced_obj(struct size_class *class, |
1606 | struct page *page, int index) | |
312fcae2 MK |
1607 | { |
1608 | unsigned long head; | |
1609 | int offset = 0; | |
1610 | unsigned long handle = 0; | |
1611 | void *addr = kmap_atomic(page); | |
1612 | ||
1613 | if (!is_first_page(page)) | |
4f42047b | 1614 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1615 | offset += class->size * index; |
1616 | ||
1617 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1618 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1619 | if (head & OBJ_ALLOCATED_TAG) { |
1620 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1621 | if (trypin_tag(handle)) | |
1622 | break; | |
1623 | handle = 0; | |
1624 | } | |
1625 | ||
1626 | offset += class->size; | |
1627 | index++; | |
1628 | } | |
1629 | ||
1630 | kunmap_atomic(addr); | |
1631 | return handle; | |
1632 | } | |
1633 | ||
1634 | struct zs_compact_control { | |
1635 | /* Source page for migration which could be a subpage of zspage. */ | |
1636 | struct page *s_page; | |
1637 | /* Destination page for migration which should be a first page | |
1638 | * of zspage. */ | |
1639 | struct page *d_page; | |
1640 | /* Starting object index within @s_page which used for live object | |
1641 | * in the subpage. */ | |
1642 | int index; | |
312fcae2 MK |
1643 | }; |
1644 | ||
1645 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1646 | struct zs_compact_control *cc) | |
1647 | { | |
1648 | unsigned long used_obj, free_obj; | |
1649 | unsigned long handle; | |
1650 | struct page *s_page = cc->s_page; | |
1651 | struct page *d_page = cc->d_page; | |
1652 | unsigned long index = cc->index; | |
312fcae2 MK |
1653 | int ret = 0; |
1654 | ||
1655 | while (1) { | |
251cbb95 | 1656 | handle = find_alloced_obj(class, s_page, index); |
312fcae2 MK |
1657 | if (!handle) { |
1658 | s_page = get_next_page(s_page); | |
1659 | if (!s_page) | |
1660 | break; | |
1661 | index = 0; | |
1662 | continue; | |
1663 | } | |
1664 | ||
1665 | /* Stop if there is no more space */ | |
1fc6e27d | 1666 | if (zspage_full(class, d_page)) { |
312fcae2 MK |
1667 | unpin_tag(handle); |
1668 | ret = -ENOMEM; | |
1669 | break; | |
1670 | } | |
1671 | ||
1672 | used_obj = handle_to_obj(handle); | |
251cbb95 MK |
1673 | free_obj = obj_malloc(class, d_page, handle); |
1674 | zs_object_copy(class, free_obj, used_obj); | |
312fcae2 | 1675 | index++; |
c102f07c JL |
1676 | /* |
1677 | * record_obj updates handle's value to free_obj and it will | |
1678 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1679 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1680 | * let's keep the lock bit. | |
1681 | */ | |
1682 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1683 | record_obj(handle, free_obj); |
1684 | unpin_tag(handle); | |
1ee47165 | 1685 | obj_free(class, used_obj); |
312fcae2 MK |
1686 | } |
1687 | ||
1688 | /* Remember last position in this iteration */ | |
1689 | cc->s_page = s_page; | |
1690 | cc->index = index; | |
312fcae2 MK |
1691 | |
1692 | return ret; | |
1693 | } | |
1694 | ||
0dc63d48 | 1695 | static struct page *isolate_target_page(struct size_class *class) |
312fcae2 MK |
1696 | { |
1697 | int i; | |
1698 | struct page *page; | |
1699 | ||
1700 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1701 | page = class->fullness_list[i]; | |
1702 | if (page) { | |
251cbb95 | 1703 | remove_zspage(class, i, page); |
312fcae2 MK |
1704 | break; |
1705 | } | |
1706 | } | |
1707 | ||
1708 | return page; | |
1709 | } | |
1710 | ||
860c707d SS |
1711 | /* |
1712 | * putback_zspage - add @first_page into right class's fullness list | |
1713 | * @pool: target pool | |
1714 | * @class: destination class | |
1715 | * @first_page: target page | |
1716 | * | |
1717 | * Return @fist_page's fullness_group | |
1718 | */ | |
1719 | static enum fullness_group putback_zspage(struct zs_pool *pool, | |
1720 | struct size_class *class, | |
1721 | struct page *first_page) | |
312fcae2 | 1722 | { |
312fcae2 MK |
1723 | enum fullness_group fullness; |
1724 | ||
1fc6e27d | 1725 | fullness = get_fullness_group(class, first_page); |
251cbb95 | 1726 | insert_zspage(class, fullness, first_page); |
839373e6 MK |
1727 | set_zspage_mapping(first_page, class->index, fullness); |
1728 | ||
13de8933 | 1729 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1730 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1731 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1732 | atomic_long_sub(class->pages_per_zspage, |
1733 | &pool->pages_allocated); | |
312fcae2 | 1734 | |
61989a80 | 1735 | free_zspage(first_page); |
13de8933 | 1736 | } |
860c707d SS |
1737 | |
1738 | return fullness; | |
61989a80 | 1739 | } |
312fcae2 MK |
1740 | |
1741 | static struct page *isolate_source_page(struct size_class *class) | |
1742 | { | |
ad9d5e17 MK |
1743 | int i; |
1744 | struct page *page = NULL; | |
1745 | ||
1746 | for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) { | |
1747 | page = class->fullness_list[i]; | |
1748 | if (!page) | |
1749 | continue; | |
312fcae2 | 1750 | |
251cbb95 | 1751 | remove_zspage(class, i, page); |
ad9d5e17 MK |
1752 | break; |
1753 | } | |
312fcae2 MK |
1754 | |
1755 | return page; | |
1756 | } | |
1757 | ||
04f05909 SS |
1758 | /* |
1759 | * | |
1760 | * Based on the number of unused allocated objects calculate | |
1761 | * and return the number of pages that we can free. | |
04f05909 SS |
1762 | */ |
1763 | static unsigned long zs_can_compact(struct size_class *class) | |
1764 | { | |
1765 | unsigned long obj_wasted; | |
44f43e99 SS |
1766 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
1767 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 1768 | |
44f43e99 SS |
1769 | if (obj_allocated <= obj_used) |
1770 | return 0; | |
04f05909 | 1771 | |
44f43e99 | 1772 | obj_wasted = obj_allocated - obj_used; |
04f05909 SS |
1773 | obj_wasted /= get_maxobj_per_zspage(class->size, |
1774 | class->pages_per_zspage); | |
1775 | ||
6cbf16b3 | 1776 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
1777 | } |
1778 | ||
7d3f3938 | 1779 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1780 | { |
312fcae2 MK |
1781 | struct zs_compact_control cc; |
1782 | struct page *src_page; | |
1783 | struct page *dst_page = NULL; | |
312fcae2 | 1784 | |
312fcae2 MK |
1785 | spin_lock(&class->lock); |
1786 | while ((src_page = isolate_source_page(class))) { | |
1787 | ||
04f05909 SS |
1788 | if (!zs_can_compact(class)) |
1789 | break; | |
1790 | ||
312fcae2 MK |
1791 | cc.index = 0; |
1792 | cc.s_page = src_page; | |
1793 | ||
0dc63d48 | 1794 | while ((dst_page = isolate_target_page(class))) { |
312fcae2 MK |
1795 | cc.d_page = dst_page; |
1796 | /* | |
0dc63d48 SS |
1797 | * If there is no more space in dst_page, resched |
1798 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1799 | */ |
1800 | if (!migrate_zspage(pool, class, &cc)) | |
1801 | break; | |
1802 | ||
1803 | putback_zspage(pool, class, dst_page); | |
312fcae2 MK |
1804 | } |
1805 | ||
1806 | /* Stop if we couldn't find slot */ | |
1807 | if (dst_page == NULL) | |
1808 | break; | |
1809 | ||
1810 | putback_zspage(pool, class, dst_page); | |
860c707d | 1811 | if (putback_zspage(pool, class, src_page) == ZS_EMPTY) |
6cbf16b3 | 1812 | pool->stats.pages_compacted += class->pages_per_zspage; |
312fcae2 | 1813 | spin_unlock(&class->lock); |
312fcae2 MK |
1814 | cond_resched(); |
1815 | spin_lock(&class->lock); | |
1816 | } | |
1817 | ||
1818 | if (src_page) | |
1819 | putback_zspage(pool, class, src_page); | |
1820 | ||
7d3f3938 | 1821 | spin_unlock(&class->lock); |
312fcae2 MK |
1822 | } |
1823 | ||
1824 | unsigned long zs_compact(struct zs_pool *pool) | |
1825 | { | |
1826 | int i; | |
312fcae2 MK |
1827 | struct size_class *class; |
1828 | ||
1829 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1830 | class = pool->size_class[i]; | |
1831 | if (!class) | |
1832 | continue; | |
1833 | if (class->index != i) | |
1834 | continue; | |
7d3f3938 | 1835 | __zs_compact(pool, class); |
312fcae2 MK |
1836 | } |
1837 | ||
860c707d | 1838 | return pool->stats.pages_compacted; |
312fcae2 MK |
1839 | } |
1840 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1841 | |
7d3f3938 SS |
1842 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1843 | { | |
1844 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1845 | } | |
1846 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1847 | ||
ab9d306d SS |
1848 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
1849 | struct shrink_control *sc) | |
1850 | { | |
1851 | unsigned long pages_freed; | |
1852 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1853 | shrinker); | |
1854 | ||
1855 | pages_freed = pool->stats.pages_compacted; | |
1856 | /* | |
1857 | * Compact classes and calculate compaction delta. | |
1858 | * Can run concurrently with a manually triggered | |
1859 | * (by user) compaction. | |
1860 | */ | |
1861 | pages_freed = zs_compact(pool) - pages_freed; | |
1862 | ||
1863 | return pages_freed ? pages_freed : SHRINK_STOP; | |
1864 | } | |
1865 | ||
1866 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
1867 | struct shrink_control *sc) | |
1868 | { | |
1869 | int i; | |
1870 | struct size_class *class; | |
1871 | unsigned long pages_to_free = 0; | |
1872 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1873 | shrinker); | |
1874 | ||
ab9d306d SS |
1875 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1876 | class = pool->size_class[i]; | |
1877 | if (!class) | |
1878 | continue; | |
1879 | if (class->index != i) | |
1880 | continue; | |
1881 | ||
ab9d306d | 1882 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
1883 | } |
1884 | ||
1885 | return pages_to_free; | |
1886 | } | |
1887 | ||
1888 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
1889 | { | |
1890 | if (pool->shrinker_enabled) { | |
1891 | unregister_shrinker(&pool->shrinker); | |
1892 | pool->shrinker_enabled = false; | |
1893 | } | |
1894 | } | |
1895 | ||
1896 | static int zs_register_shrinker(struct zs_pool *pool) | |
1897 | { | |
1898 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
1899 | pool->shrinker.count_objects = zs_shrinker_count; | |
1900 | pool->shrinker.batch = 0; | |
1901 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
1902 | ||
1903 | return register_shrinker(&pool->shrinker); | |
1904 | } | |
1905 | ||
00a61d86 | 1906 | /** |
66cdef66 GM |
1907 | * zs_create_pool - Creates an allocation pool to work from. |
1908 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1909 | * |
66cdef66 GM |
1910 | * This function must be called before anything when using |
1911 | * the zsmalloc allocator. | |
166cfda7 | 1912 | * |
66cdef66 GM |
1913 | * On success, a pointer to the newly created pool is returned, |
1914 | * otherwise NULL. | |
396b7fd6 | 1915 | */ |
d0d8da2d | 1916 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 1917 | { |
66cdef66 GM |
1918 | int i; |
1919 | struct zs_pool *pool; | |
1920 | struct size_class *prev_class = NULL; | |
61989a80 | 1921 | |
66cdef66 GM |
1922 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1923 | if (!pool) | |
1924 | return NULL; | |
61989a80 | 1925 | |
66cdef66 GM |
1926 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1927 | GFP_KERNEL); | |
1928 | if (!pool->size_class) { | |
1929 | kfree(pool); | |
1930 | return NULL; | |
1931 | } | |
61989a80 | 1932 | |
2e40e163 MK |
1933 | pool->name = kstrdup(name, GFP_KERNEL); |
1934 | if (!pool->name) | |
1935 | goto err; | |
1936 | ||
1937 | if (create_handle_cache(pool)) | |
1938 | goto err; | |
1939 | ||
c60369f0 | 1940 | /* |
66cdef66 GM |
1941 | * Iterate reversly, because, size of size_class that we want to use |
1942 | * for merging should be larger or equal to current size. | |
c60369f0 | 1943 | */ |
66cdef66 GM |
1944 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1945 | int size; | |
1946 | int pages_per_zspage; | |
1947 | struct size_class *class; | |
c60369f0 | 1948 | |
66cdef66 GM |
1949 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1950 | if (size > ZS_MAX_ALLOC_SIZE) | |
1951 | size = ZS_MAX_ALLOC_SIZE; | |
1952 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1953 | |
66cdef66 GM |
1954 | /* |
1955 | * size_class is used for normal zsmalloc operation such | |
1956 | * as alloc/free for that size. Although it is natural that we | |
1957 | * have one size_class for each size, there is a chance that we | |
1958 | * can get more memory utilization if we use one size_class for | |
1959 | * many different sizes whose size_class have same | |
1960 | * characteristics. So, we makes size_class point to | |
1961 | * previous size_class if possible. | |
1962 | */ | |
1963 | if (prev_class) { | |
1964 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1965 | pool->size_class[i] = prev_class; | |
1966 | continue; | |
1967 | } | |
1968 | } | |
1969 | ||
1970 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1971 | if (!class) | |
1972 | goto err; | |
1973 | ||
1974 | class->size = size; | |
1975 | class->index = i; | |
1976 | class->pages_per_zspage = pages_per_zspage; | |
1fc6e27d MK |
1977 | class->objs_per_zspage = class->pages_per_zspage * |
1978 | PAGE_SIZE / class->size; | |
1979 | if (pages_per_zspage == 1 && class->objs_per_zspage == 1) | |
7b60a685 | 1980 | class->huge = true; |
66cdef66 GM |
1981 | spin_lock_init(&class->lock); |
1982 | pool->size_class[i] = class; | |
1983 | ||
1984 | prev_class = class; | |
61989a80 NG |
1985 | } |
1986 | ||
d34f6157 DS |
1987 | /* debug only, don't abort if it fails */ |
1988 | zs_pool_stat_create(pool, name); | |
0f050d99 | 1989 | |
ab9d306d SS |
1990 | /* |
1991 | * Not critical, we still can use the pool | |
1992 | * and user can trigger compaction manually. | |
1993 | */ | |
1994 | if (zs_register_shrinker(pool) == 0) | |
1995 | pool->shrinker_enabled = true; | |
66cdef66 GM |
1996 | return pool; |
1997 | ||
1998 | err: | |
1999 | zs_destroy_pool(pool); | |
2000 | return NULL; | |
61989a80 | 2001 | } |
66cdef66 | 2002 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2003 | |
66cdef66 | 2004 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2005 | { |
66cdef66 | 2006 | int i; |
61989a80 | 2007 | |
ab9d306d | 2008 | zs_unregister_shrinker(pool); |
0f050d99 GM |
2009 | zs_pool_stat_destroy(pool); |
2010 | ||
66cdef66 GM |
2011 | for (i = 0; i < zs_size_classes; i++) { |
2012 | int fg; | |
2013 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2014 | |
66cdef66 GM |
2015 | if (!class) |
2016 | continue; | |
61989a80 | 2017 | |
66cdef66 GM |
2018 | if (class->index != i) |
2019 | continue; | |
61989a80 | 2020 | |
66cdef66 GM |
2021 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
2022 | if (class->fullness_list[fg]) { | |
2023 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
2024 | class->size, fg); | |
2025 | } | |
2026 | } | |
2027 | kfree(class); | |
2028 | } | |
f553646a | 2029 | |
2e40e163 | 2030 | destroy_handle_cache(pool); |
66cdef66 | 2031 | kfree(pool->size_class); |
0f050d99 | 2032 | kfree(pool->name); |
66cdef66 GM |
2033 | kfree(pool); |
2034 | } | |
2035 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2036 | |
66cdef66 GM |
2037 | static int __init zs_init(void) |
2038 | { | |
2039 | int ret = zs_register_cpu_notifier(); | |
2040 | ||
0f050d99 GM |
2041 | if (ret) |
2042 | goto notifier_fail; | |
66cdef66 GM |
2043 | |
2044 | init_zs_size_classes(); | |
2045 | ||
2046 | #ifdef CONFIG_ZPOOL | |
2047 | zpool_register_driver(&zs_zpool_driver); | |
2048 | #endif | |
0f050d99 | 2049 | |
4abaac9b DS |
2050 | zs_stat_init(); |
2051 | ||
66cdef66 | 2052 | return 0; |
0f050d99 | 2053 | |
0f050d99 GM |
2054 | notifier_fail: |
2055 | zs_unregister_cpu_notifier(); | |
2056 | ||
2057 | return ret; | |
61989a80 | 2058 | } |
61989a80 | 2059 | |
66cdef66 | 2060 | static void __exit zs_exit(void) |
61989a80 | 2061 | { |
66cdef66 GM |
2062 | #ifdef CONFIG_ZPOOL |
2063 | zpool_unregister_driver(&zs_zpool_driver); | |
2064 | #endif | |
2065 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
2066 | |
2067 | zs_stat_exit(); | |
61989a80 | 2068 | } |
069f101f BH |
2069 | |
2070 | module_init(zs_init); | |
2071 | module_exit(zs_exit); | |
2072 | ||
2073 | MODULE_LICENSE("Dual BSD/GPL"); | |
2074 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |