]>
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: | |
3783689a MK |
19 | * page->private: points to zspage |
20 | * page->index: offset of the first object starting in this page. | |
21 | * For the first page, this is always 0, so we use this field | |
22 | * to store handle for huge object. | |
23 | * page->next: links together all component pages of a zspage | |
2db51dae NG |
24 | * |
25 | * Usage of struct page flags: | |
26 | * PG_private: identifies the first component page | |
27 | * PG_private2: identifies the last component page | |
28 | * | |
29 | */ | |
30 | ||
4abaac9b DS |
31 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
32 | ||
61989a80 NG |
33 | #include <linux/module.h> |
34 | #include <linux/kernel.h> | |
312fcae2 | 35 | #include <linux/sched.h> |
61989a80 NG |
36 | #include <linux/bitops.h> |
37 | #include <linux/errno.h> | |
38 | #include <linux/highmem.h> | |
61989a80 NG |
39 | #include <linux/string.h> |
40 | #include <linux/slab.h> | |
41 | #include <asm/tlbflush.h> | |
42 | #include <asm/pgtable.h> | |
43 | #include <linux/cpumask.h> | |
44 | #include <linux/cpu.h> | |
0cbb613f | 45 | #include <linux/vmalloc.h> |
759b26b2 | 46 | #include <linux/preempt.h> |
0959c63f SJ |
47 | #include <linux/spinlock.h> |
48 | #include <linux/types.h> | |
0f050d99 | 49 | #include <linux/debugfs.h> |
bcf1647d | 50 | #include <linux/zsmalloc.h> |
c795779d | 51 | #include <linux/zpool.h> |
0959c63f SJ |
52 | |
53 | /* | |
54 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
55 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
56 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
57 | * to restore link_free pointer values. | |
58 | */ | |
59 | #define ZS_ALIGN 8 | |
60 | ||
61 | /* | |
62 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
63 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
64 | */ | |
65 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
66 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
67 | ||
2e40e163 MK |
68 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
69 | ||
0959c63f SJ |
70 | /* |
71 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 72 | * as single (unsigned long) handle value. |
0959c63f SJ |
73 | * |
74 | * Note that object index <obj_idx> is relative to system | |
75 | * page <PFN> it is stored in, so for each sub-page belonging | |
76 | * to a zspage, obj_idx starts with 0. | |
77 | * | |
78 | * This is made more complicated by various memory models and PAE. | |
79 | */ | |
80 | ||
81 | #ifndef MAX_PHYSMEM_BITS | |
82 | #ifdef CONFIG_HIGHMEM64G | |
83 | #define MAX_PHYSMEM_BITS 36 | |
84 | #else /* !CONFIG_HIGHMEM64G */ | |
85 | /* | |
86 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
87 | * be PAGE_SHIFT | |
88 | */ | |
89 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
90 | #endif | |
91 | #endif | |
92 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
93 | |
94 | /* | |
95 | * Memory for allocating for handle keeps object position by | |
96 | * encoding <page, obj_idx> and the encoded value has a room | |
97 | * in least bit(ie, look at obj_to_location). | |
98 | * We use the bit to synchronize between object access by | |
99 | * user and migration. | |
100 | */ | |
101 | #define HANDLE_PIN_BIT 0 | |
102 | ||
103 | /* | |
104 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
105 | * to identify the object was allocated or not. | |
106 | * It's okay to add the status bit in the least bit because | |
107 | * header keeps handle which is 4byte-aligned address so we | |
108 | * have room for two bit at least. | |
109 | */ | |
110 | #define OBJ_ALLOCATED_TAG 1 | |
111 | #define OBJ_TAG_BITS 1 | |
112 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
113 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
114 | ||
115 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
116 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
117 | #define ZS_MIN_ALLOC_SIZE \ | |
118 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 119 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 120 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
121 | |
122 | /* | |
7eb52512 | 123 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
124 | * trader-off here: |
125 | * - Large number of size classes is potentially wasteful as free page are | |
126 | * spread across these classes | |
127 | * - Small number of size classes causes large internal fragmentation | |
128 | * - Probably its better to use specific size classes (empirically | |
129 | * determined). NOTE: all those class sizes must be set as multiple of | |
130 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
131 | * | |
132 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
133 | * (reason above) | |
134 | */ | |
3783689a | 135 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
0959c63f SJ |
136 | |
137 | /* | |
138 | * We do not maintain any list for completely empty or full pages | |
139 | */ | |
140 | enum fullness_group { | |
141 | ZS_ALMOST_FULL, | |
142 | ZS_ALMOST_EMPTY, | |
0959c63f SJ |
143 | ZS_EMPTY, |
144 | ZS_FULL | |
145 | }; | |
146 | ||
0f050d99 GM |
147 | enum zs_stat_type { |
148 | OBJ_ALLOCATED, | |
149 | OBJ_USED, | |
248ca1b0 MK |
150 | CLASS_ALMOST_FULL, |
151 | CLASS_ALMOST_EMPTY, | |
0f050d99 GM |
152 | }; |
153 | ||
6fe5186f SS |
154 | #ifdef CONFIG_ZSMALLOC_STAT |
155 | #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1) | |
156 | #else | |
157 | #define NR_ZS_STAT_TYPE (OBJ_USED + 1) | |
158 | #endif | |
159 | ||
0f050d99 GM |
160 | struct zs_size_stat { |
161 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
162 | }; | |
163 | ||
57244594 SS |
164 | #ifdef CONFIG_ZSMALLOC_STAT |
165 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
166 | #endif |
167 | ||
40f9fb8c MG |
168 | /* |
169 | * number of size_classes | |
170 | */ | |
171 | static int zs_size_classes; | |
172 | ||
0959c63f SJ |
173 | /* |
174 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
175 | * n <= N / f, where | |
176 | * n = number of allocated objects | |
177 | * N = total number of objects zspage can store | |
6dd9737e | 178 | * f = fullness_threshold_frac |
0959c63f SJ |
179 | * |
180 | * Similarly, we assign zspage to: | |
181 | * ZS_ALMOST_FULL when n > N / f | |
182 | * ZS_EMPTY when n == 0 | |
183 | * ZS_FULL when n == N | |
184 | * | |
185 | * (see: fix_fullness_group()) | |
186 | */ | |
187 | static const int fullness_threshold_frac = 4; | |
188 | ||
189 | struct size_class { | |
57244594 | 190 | spinlock_t lock; |
3783689a | 191 | struct list_head fullness_list[2]; |
0959c63f SJ |
192 | /* |
193 | * Size of objects stored in this class. Must be multiple | |
194 | * of ZS_ALIGN. | |
195 | */ | |
196 | int size; | |
1fc6e27d | 197 | int objs_per_zspage; |
0959c63f SJ |
198 | unsigned int index; |
199 | ||
0f050d99 | 200 | struct zs_size_stat stats; |
0959c63f | 201 | |
7dfa4612 WY |
202 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
203 | int pages_per_zspage; | |
57244594 SS |
204 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
205 | bool huge; | |
0959c63f SJ |
206 | }; |
207 | ||
208 | /* | |
209 | * Placed within free objects to form a singly linked list. | |
3783689a | 210 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
211 | * |
212 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
213 | */ | |
214 | struct link_free { | |
2e40e163 MK |
215 | union { |
216 | /* | |
217 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
218 | * It's valid for non-allocated object | |
219 | */ | |
220 | void *next; | |
221 | /* | |
222 | * Handle of allocated object. | |
223 | */ | |
224 | unsigned long handle; | |
225 | }; | |
0959c63f SJ |
226 | }; |
227 | ||
228 | struct zs_pool { | |
6f3526d6 | 229 | const char *name; |
0f050d99 | 230 | |
40f9fb8c | 231 | struct size_class **size_class; |
2e40e163 | 232 | struct kmem_cache *handle_cachep; |
3783689a | 233 | struct kmem_cache *zspage_cachep; |
0959c63f | 234 | |
13de8933 | 235 | atomic_long_t pages_allocated; |
0f050d99 | 236 | |
7d3f3938 | 237 | struct zs_pool_stats stats; |
ab9d306d SS |
238 | |
239 | /* Compact classes */ | |
240 | struct shrinker shrinker; | |
241 | /* | |
242 | * To signify that register_shrinker() was successful | |
243 | * and unregister_shrinker() will not Oops. | |
244 | */ | |
245 | bool shrinker_enabled; | |
0f050d99 GM |
246 | #ifdef CONFIG_ZSMALLOC_STAT |
247 | struct dentry *stat_dentry; | |
248 | #endif | |
0959c63f | 249 | }; |
61989a80 NG |
250 | |
251 | /* | |
252 | * A zspage's class index and fullness group | |
253 | * are encoded in its (first)page->mapping | |
254 | */ | |
3783689a MK |
255 | #define FULLNESS_BITS 2 |
256 | #define CLASS_BITS 8 | |
4f42047b | 257 | |
3783689a MK |
258 | struct zspage { |
259 | struct { | |
260 | unsigned int fullness:FULLNESS_BITS; | |
261 | unsigned int class:CLASS_BITS; | |
262 | }; | |
263 | unsigned int inuse; | |
264 | void *freeobj; | |
265 | struct page *first_page; | |
266 | struct list_head list; /* fullness list */ | |
267 | }; | |
61989a80 | 268 | |
f553646a | 269 | struct mapping_area { |
1b945aee | 270 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
271 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
272 | #else | |
273 | char *vm_buf; /* copy buffer for objects that span pages */ | |
274 | #endif | |
275 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
276 | enum zs_mapmode vm_mm; /* mapping mode */ | |
277 | }; | |
278 | ||
3783689a | 279 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
280 | { |
281 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
282 | 0, 0, NULL); | |
3783689a MK |
283 | if (!pool->handle_cachep) |
284 | return 1; | |
285 | ||
286 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
287 | 0, 0, NULL); | |
288 | if (!pool->zspage_cachep) { | |
289 | kmem_cache_destroy(pool->handle_cachep); | |
290 | pool->handle_cachep = NULL; | |
291 | return 1; | |
292 | } | |
293 | ||
294 | return 0; | |
2e40e163 MK |
295 | } |
296 | ||
3783689a | 297 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 298 | { |
cd10add0 | 299 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 300 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
301 | } |
302 | ||
3783689a | 303 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
304 | { |
305 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
d0d8da2d | 306 | gfp & ~__GFP_HIGHMEM); |
2e40e163 MK |
307 | } |
308 | ||
3783689a | 309 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
310 | { |
311 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
312 | } | |
313 | ||
3783689a MK |
314 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
315 | { | |
316 | return kmem_cache_alloc(pool->zspage_cachep, flags & ~__GFP_HIGHMEM); | |
317 | }; | |
318 | ||
319 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
320 | { | |
321 | kmem_cache_free(pool->zspage_cachep, zspage); | |
322 | } | |
323 | ||
2e40e163 MK |
324 | static void record_obj(unsigned long handle, unsigned long obj) |
325 | { | |
c102f07c JL |
326 | /* |
327 | * lsb of @obj represents handle lock while other bits | |
328 | * represent object value the handle is pointing so | |
329 | * updating shouldn't do store tearing. | |
330 | */ | |
331 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
332 | } |
333 | ||
c795779d DS |
334 | /* zpool driver */ |
335 | ||
336 | #ifdef CONFIG_ZPOOL | |
337 | ||
6f3526d6 | 338 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 339 | const struct zpool_ops *zpool_ops, |
479305fd | 340 | struct zpool *zpool) |
c795779d | 341 | { |
d0d8da2d SS |
342 | /* |
343 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
344 | * different contexts and its caller must provide a valid | |
345 | * gfp mask. | |
346 | */ | |
347 | return zs_create_pool(name); | |
c795779d DS |
348 | } |
349 | ||
350 | static void zs_zpool_destroy(void *pool) | |
351 | { | |
352 | zs_destroy_pool(pool); | |
353 | } | |
354 | ||
355 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
356 | unsigned long *handle) | |
357 | { | |
d0d8da2d | 358 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
359 | return *handle ? 0 : -1; |
360 | } | |
361 | static void zs_zpool_free(void *pool, unsigned long handle) | |
362 | { | |
363 | zs_free(pool, handle); | |
364 | } | |
365 | ||
366 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
367 | unsigned int *reclaimed) | |
368 | { | |
369 | return -EINVAL; | |
370 | } | |
371 | ||
372 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
373 | enum zpool_mapmode mm) | |
374 | { | |
375 | enum zs_mapmode zs_mm; | |
376 | ||
377 | switch (mm) { | |
378 | case ZPOOL_MM_RO: | |
379 | zs_mm = ZS_MM_RO; | |
380 | break; | |
381 | case ZPOOL_MM_WO: | |
382 | zs_mm = ZS_MM_WO; | |
383 | break; | |
384 | case ZPOOL_MM_RW: /* fallthru */ | |
385 | default: | |
386 | zs_mm = ZS_MM_RW; | |
387 | break; | |
388 | } | |
389 | ||
390 | return zs_map_object(pool, handle, zs_mm); | |
391 | } | |
392 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
393 | { | |
394 | zs_unmap_object(pool, handle); | |
395 | } | |
396 | ||
397 | static u64 zs_zpool_total_size(void *pool) | |
398 | { | |
722cdc17 | 399 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
400 | } |
401 | ||
402 | static struct zpool_driver zs_zpool_driver = { | |
403 | .type = "zsmalloc", | |
404 | .owner = THIS_MODULE, | |
405 | .create = zs_zpool_create, | |
406 | .destroy = zs_zpool_destroy, | |
407 | .malloc = zs_zpool_malloc, | |
408 | .free = zs_zpool_free, | |
409 | .shrink = zs_zpool_shrink, | |
410 | .map = zs_zpool_map, | |
411 | .unmap = zs_zpool_unmap, | |
412 | .total_size = zs_zpool_total_size, | |
413 | }; | |
414 | ||
137f8cff | 415 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
416 | #endif /* CONFIG_ZPOOL */ |
417 | ||
248ca1b0 MK |
418 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
419 | { | |
420 | return pages_per_zspage * PAGE_SIZE / size; | |
421 | } | |
422 | ||
61989a80 NG |
423 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
424 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
425 | ||
426 | static int is_first_page(struct page *page) | |
427 | { | |
a27545bf | 428 | return PagePrivate(page); |
61989a80 NG |
429 | } |
430 | ||
3783689a | 431 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 432 | { |
3783689a | 433 | return zspage->inuse; |
4f42047b MK |
434 | } |
435 | ||
3783689a | 436 | static inline void set_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 437 | { |
3783689a | 438 | zspage->inuse = val; |
4f42047b MK |
439 | } |
440 | ||
3783689a | 441 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 442 | { |
3783689a | 443 | zspage->inuse += val; |
4f42047b MK |
444 | } |
445 | ||
446 | static inline int get_first_obj_offset(struct page *page) | |
447 | { | |
3783689a MK |
448 | if (is_first_page(page)) |
449 | return 0; | |
450 | ||
4f42047b MK |
451 | return page->index; |
452 | } | |
453 | ||
454 | static inline void set_first_obj_offset(struct page *page, int offset) | |
455 | { | |
3783689a MK |
456 | if (is_first_page(page)) |
457 | return; | |
458 | ||
4f42047b MK |
459 | page->index = offset; |
460 | } | |
461 | ||
3783689a | 462 | static inline unsigned long get_freeobj(struct zspage *zspage) |
4f42047b | 463 | { |
3783689a | 464 | return (unsigned long)zspage->freeobj; |
4f42047b MK |
465 | } |
466 | ||
3783689a | 467 | static inline void set_freeobj(struct zspage *zspage, unsigned long obj) |
4f42047b | 468 | { |
3783689a | 469 | zspage->freeobj = (void *)obj; |
4f42047b MK |
470 | } |
471 | ||
3783689a | 472 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 473 | unsigned int *class_idx, |
61989a80 NG |
474 | enum fullness_group *fullness) |
475 | { | |
3783689a MK |
476 | *fullness = zspage->fullness; |
477 | *class_idx = zspage->class; | |
61989a80 NG |
478 | } |
479 | ||
3783689a | 480 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 481 | unsigned int class_idx, |
61989a80 NG |
482 | enum fullness_group fullness) |
483 | { | |
3783689a MK |
484 | zspage->class = class_idx; |
485 | zspage->fullness = fullness; | |
61989a80 NG |
486 | } |
487 | ||
c3e3e88a NC |
488 | /* |
489 | * zsmalloc divides the pool into various size classes where each | |
490 | * class maintains a list of zspages where each zspage is divided | |
491 | * into equal sized chunks. Each allocation falls into one of these | |
492 | * classes depending on its size. This function returns index of the | |
493 | * size class which has chunk size big enough to hold the give size. | |
494 | */ | |
61989a80 NG |
495 | static int get_size_class_index(int size) |
496 | { | |
497 | int idx = 0; | |
498 | ||
499 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
500 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
501 | ZS_SIZE_CLASS_DELTA); | |
502 | ||
7b60a685 | 503 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
504 | } |
505 | ||
248ca1b0 MK |
506 | static inline void zs_stat_inc(struct size_class *class, |
507 | enum zs_stat_type type, unsigned long cnt) | |
508 | { | |
6fe5186f SS |
509 | if (type < NR_ZS_STAT_TYPE) |
510 | class->stats.objs[type] += cnt; | |
248ca1b0 MK |
511 | } |
512 | ||
513 | static inline void zs_stat_dec(struct size_class *class, | |
514 | enum zs_stat_type type, unsigned long cnt) | |
515 | { | |
6fe5186f SS |
516 | if (type < NR_ZS_STAT_TYPE) |
517 | class->stats.objs[type] -= cnt; | |
248ca1b0 MK |
518 | } |
519 | ||
520 | static inline unsigned long zs_stat_get(struct size_class *class, | |
521 | enum zs_stat_type type) | |
522 | { | |
6fe5186f SS |
523 | if (type < NR_ZS_STAT_TYPE) |
524 | return class->stats.objs[type]; | |
525 | return 0; | |
248ca1b0 MK |
526 | } |
527 | ||
57244594 SS |
528 | #ifdef CONFIG_ZSMALLOC_STAT |
529 | ||
4abaac9b | 530 | static void __init zs_stat_init(void) |
248ca1b0 | 531 | { |
4abaac9b DS |
532 | if (!debugfs_initialized()) { |
533 | pr_warn("debugfs not available, stat dir not created\n"); | |
534 | return; | |
535 | } | |
248ca1b0 MK |
536 | |
537 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
538 | if (!zs_stat_root) | |
4abaac9b | 539 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
540 | } |
541 | ||
542 | static void __exit zs_stat_exit(void) | |
543 | { | |
544 | debugfs_remove_recursive(zs_stat_root); | |
545 | } | |
546 | ||
1120ed54 SS |
547 | static unsigned long zs_can_compact(struct size_class *class); |
548 | ||
248ca1b0 MK |
549 | static int zs_stats_size_show(struct seq_file *s, void *v) |
550 | { | |
551 | int i; | |
552 | struct zs_pool *pool = s->private; | |
553 | struct size_class *class; | |
554 | int objs_per_zspage; | |
555 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 556 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
557 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
558 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 559 | unsigned long total_freeable = 0; |
248ca1b0 | 560 | |
1120ed54 | 561 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
562 | "class", "size", "almost_full", "almost_empty", |
563 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 564 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
565 | |
566 | for (i = 0; i < zs_size_classes; i++) { | |
567 | class = pool->size_class[i]; | |
568 | ||
569 | if (class->index != i) | |
570 | continue; | |
571 | ||
572 | spin_lock(&class->lock); | |
573 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
574 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
575 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
576 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 577 | freeable = zs_can_compact(class); |
248ca1b0 MK |
578 | spin_unlock(&class->lock); |
579 | ||
580 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
581 | class->pages_per_zspage); | |
582 | pages_used = obj_allocated / objs_per_zspage * | |
583 | class->pages_per_zspage; | |
584 | ||
1120ed54 SS |
585 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
586 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
587 | i, class->size, class_almost_full, class_almost_empty, |
588 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 589 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
590 | |
591 | total_class_almost_full += class_almost_full; | |
592 | total_class_almost_empty += class_almost_empty; | |
593 | total_objs += obj_allocated; | |
594 | total_used_objs += obj_used; | |
595 | total_pages += pages_used; | |
1120ed54 | 596 | total_freeable += freeable; |
248ca1b0 MK |
597 | } |
598 | ||
599 | seq_puts(s, "\n"); | |
1120ed54 | 600 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
601 | "Total", "", total_class_almost_full, |
602 | total_class_almost_empty, total_objs, | |
1120ed54 | 603 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
604 | |
605 | return 0; | |
606 | } | |
607 | ||
608 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
609 | { | |
610 | return single_open(file, zs_stats_size_show, inode->i_private); | |
611 | } | |
612 | ||
613 | static const struct file_operations zs_stat_size_ops = { | |
614 | .open = zs_stats_size_open, | |
615 | .read = seq_read, | |
616 | .llseek = seq_lseek, | |
617 | .release = single_release, | |
618 | }; | |
619 | ||
d34f6157 | 620 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
621 | { |
622 | struct dentry *entry; | |
623 | ||
4abaac9b DS |
624 | if (!zs_stat_root) { |
625 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 626 | return; |
4abaac9b | 627 | } |
248ca1b0 MK |
628 | |
629 | entry = debugfs_create_dir(name, zs_stat_root); | |
630 | if (!entry) { | |
631 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 632 | return; |
248ca1b0 MK |
633 | } |
634 | pool->stat_dentry = entry; | |
635 | ||
636 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
637 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
638 | if (!entry) { | |
639 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
640 | name, "classes"); | |
4abaac9b DS |
641 | debugfs_remove_recursive(pool->stat_dentry); |
642 | pool->stat_dentry = NULL; | |
248ca1b0 | 643 | } |
248ca1b0 MK |
644 | } |
645 | ||
646 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
647 | { | |
648 | debugfs_remove_recursive(pool->stat_dentry); | |
649 | } | |
650 | ||
651 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 652 | static void __init zs_stat_init(void) |
248ca1b0 | 653 | { |
248ca1b0 MK |
654 | } |
655 | ||
656 | static void __exit zs_stat_exit(void) | |
657 | { | |
658 | } | |
659 | ||
d34f6157 | 660 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 661 | { |
248ca1b0 MK |
662 | } |
663 | ||
664 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
665 | { | |
666 | } | |
248ca1b0 MK |
667 | #endif |
668 | ||
c3e3e88a NC |
669 | /* |
670 | * For each size class, zspages are divided into different groups | |
671 | * depending on how "full" they are. This was done so that we could | |
672 | * easily find empty or nearly empty zspages when we try to shrink | |
673 | * the pool (not yet implemented). This function returns fullness | |
674 | * status of the given page. | |
675 | */ | |
1fc6e27d | 676 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 677 | struct zspage *zspage) |
61989a80 | 678 | { |
1fc6e27d | 679 | int inuse, objs_per_zspage; |
61989a80 | 680 | enum fullness_group fg; |
830e4bc5 | 681 | |
3783689a | 682 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 683 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
684 | |
685 | if (inuse == 0) | |
686 | fg = ZS_EMPTY; | |
1fc6e27d | 687 | else if (inuse == objs_per_zspage) |
61989a80 | 688 | fg = ZS_FULL; |
1fc6e27d | 689 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
690 | fg = ZS_ALMOST_EMPTY; |
691 | else | |
692 | fg = ZS_ALMOST_FULL; | |
693 | ||
694 | return fg; | |
695 | } | |
696 | ||
c3e3e88a NC |
697 | /* |
698 | * Each size class maintains various freelists and zspages are assigned | |
699 | * to one of these freelists based on the number of live objects they | |
700 | * have. This functions inserts the given zspage into the freelist | |
701 | * identified by <class, fullness_group>. | |
702 | */ | |
251cbb95 | 703 | static void insert_zspage(struct size_class *class, |
3783689a MK |
704 | struct zspage *zspage, |
705 | enum fullness_group fullness) | |
61989a80 | 706 | { |
3783689a | 707 | struct zspage *head; |
61989a80 | 708 | |
3783689a | 709 | if (fullness >= ZS_EMPTY) |
61989a80 NG |
710 | return; |
711 | ||
3783689a MK |
712 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
713 | struct zspage, list); | |
714 | ||
248ca1b0 MK |
715 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
716 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
58f17117 | 717 | |
58f17117 | 718 | /* |
3783689a MK |
719 | * We want to see more ZS_FULL pages and less almost empty/full. |
720 | * Put pages with higher ->inuse first. | |
58f17117 | 721 | */ |
3783689a MK |
722 | if (head) { |
723 | if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) { | |
724 | list_add(&zspage->list, &head->list); | |
725 | return; | |
726 | } | |
727 | } | |
728 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
61989a80 NG |
729 | } |
730 | ||
c3e3e88a NC |
731 | /* |
732 | * This function removes the given zspage from the freelist identified | |
733 | * by <class, fullness_group>. | |
734 | */ | |
251cbb95 | 735 | static void remove_zspage(struct size_class *class, |
3783689a MK |
736 | struct zspage *zspage, |
737 | enum fullness_group fullness) | |
61989a80 | 738 | { |
3783689a | 739 | if (fullness >= ZS_EMPTY) |
61989a80 NG |
740 | return; |
741 | ||
3783689a | 742 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
61989a80 | 743 | |
3783689a | 744 | list_del_init(&zspage->list); |
248ca1b0 MK |
745 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
746 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
747 | } |
748 | ||
c3e3e88a NC |
749 | /* |
750 | * Each size class maintains zspages in different fullness groups depending | |
751 | * on the number of live objects they contain. When allocating or freeing | |
752 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
753 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
754 | * a status change has occurred for the given page and accordingly moves the | |
755 | * page from the freelist of the old fullness group to that of the new | |
756 | * fullness group. | |
757 | */ | |
c7806261 | 758 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 759 | struct zspage *zspage) |
61989a80 NG |
760 | { |
761 | int class_idx; | |
61989a80 NG |
762 | enum fullness_group currfg, newfg; |
763 | ||
3783689a MK |
764 | get_zspage_mapping(zspage, &class_idx, &currfg); |
765 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
766 | if (newfg == currfg) |
767 | goto out; | |
768 | ||
3783689a MK |
769 | remove_zspage(class, zspage, currfg); |
770 | insert_zspage(class, zspage, newfg); | |
771 | set_zspage_mapping(zspage, class_idx, newfg); | |
61989a80 NG |
772 | |
773 | out: | |
774 | return newfg; | |
775 | } | |
776 | ||
777 | /* | |
778 | * We have to decide on how many pages to link together | |
779 | * to form a zspage for each size class. This is important | |
780 | * to reduce wastage due to unusable space left at end of | |
781 | * each zspage which is given as: | |
888fa374 YX |
782 | * wastage = Zp % class_size |
783 | * usage = Zp - wastage | |
61989a80 NG |
784 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
785 | * | |
786 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
787 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
788 | * since then we can perfectly fit in 8 such objects. | |
789 | */ | |
2e3b6154 | 790 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
791 | { |
792 | int i, max_usedpc = 0; | |
793 | /* zspage order which gives maximum used size per KB */ | |
794 | int max_usedpc_order = 1; | |
795 | ||
84d4faab | 796 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
797 | int zspage_size; |
798 | int waste, usedpc; | |
799 | ||
800 | zspage_size = i * PAGE_SIZE; | |
801 | waste = zspage_size % class_size; | |
802 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
803 | ||
804 | if (usedpc > max_usedpc) { | |
805 | max_usedpc = usedpc; | |
806 | max_usedpc_order = i; | |
807 | } | |
808 | } | |
809 | ||
810 | return max_usedpc_order; | |
811 | } | |
812 | ||
3783689a MK |
813 | |
814 | static struct zspage *get_zspage(struct page *page) | |
61989a80 | 815 | { |
3783689a | 816 | return (struct zspage *)page->private; |
61989a80 NG |
817 | } |
818 | ||
819 | static struct page *get_next_page(struct page *page) | |
820 | { | |
3783689a | 821 | return page->next; |
61989a80 NG |
822 | } |
823 | ||
67296874 OH |
824 | /* |
825 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 826 | * We use the least bit of handle for tagging. |
67296874 | 827 | */ |
312fcae2 | 828 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 829 | { |
312fcae2 | 830 | unsigned long obj; |
61989a80 NG |
831 | |
832 | if (!page) { | |
830e4bc5 | 833 | VM_BUG_ON(obj_idx); |
61989a80 NG |
834 | return NULL; |
835 | } | |
836 | ||
312fcae2 MK |
837 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
838 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
839 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 840 | |
312fcae2 | 841 | return (void *)obj; |
61989a80 NG |
842 | } |
843 | ||
67296874 OH |
844 | /* |
845 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
846 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 847 | * location_to_obj(). |
67296874 | 848 | */ |
312fcae2 | 849 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
850 | unsigned long *obj_idx) |
851 | { | |
312fcae2 MK |
852 | obj >>= OBJ_TAG_BITS; |
853 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
854 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
855 | } |
856 | ||
2e40e163 MK |
857 | static unsigned long handle_to_obj(unsigned long handle) |
858 | { | |
859 | return *(unsigned long *)handle; | |
860 | } | |
861 | ||
7b60a685 MK |
862 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
863 | void *obj) | |
312fcae2 | 864 | { |
7b60a685 | 865 | if (class->huge) { |
830e4bc5 | 866 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3783689a | 867 | return page->index; |
7b60a685 MK |
868 | } else |
869 | return *(unsigned long *)obj; | |
312fcae2 MK |
870 | } |
871 | ||
61989a80 NG |
872 | static unsigned long obj_idx_to_offset(struct page *page, |
873 | unsigned long obj_idx, int class_size) | |
874 | { | |
3783689a | 875 | unsigned long off; |
61989a80 | 876 | |
3783689a | 877 | off = get_first_obj_offset(page); |
61989a80 NG |
878 | |
879 | return off + obj_idx * class_size; | |
880 | } | |
881 | ||
312fcae2 MK |
882 | static inline int trypin_tag(unsigned long handle) |
883 | { | |
1b8320b6 | 884 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
885 | } |
886 | ||
887 | static void pin_tag(unsigned long handle) | |
888 | { | |
1b8320b6 | 889 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
890 | } |
891 | ||
892 | static void unpin_tag(unsigned long handle) | |
893 | { | |
1b8320b6 | 894 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
895 | } |
896 | ||
f4477e90 NG |
897 | static void reset_page(struct page *page) |
898 | { | |
899 | clear_bit(PG_private, &page->flags); | |
900 | clear_bit(PG_private_2, &page->flags); | |
901 | set_page_private(page, 0); | |
3783689a | 902 | page->index = 0; |
f4477e90 NG |
903 | } |
904 | ||
3783689a | 905 | static void free_zspage(struct zs_pool *pool, struct zspage *zspage) |
61989a80 | 906 | { |
3783689a | 907 | struct page *page, *next; |
61989a80 | 908 | |
3783689a | 909 | VM_BUG_ON(get_zspage_inuse(zspage)); |
61989a80 | 910 | |
3783689a MK |
911 | next = page = zspage->first_page; |
912 | do { | |
913 | next = page->next; | |
914 | reset_page(page); | |
915 | put_page(page); | |
916 | page = next; | |
917 | } while (page != NULL); | |
61989a80 | 918 | |
3783689a | 919 | cache_free_zspage(pool, zspage); |
61989a80 NG |
920 | } |
921 | ||
922 | /* Initialize a newly allocated zspage */ | |
3783689a | 923 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 NG |
924 | { |
925 | unsigned long off = 0; | |
3783689a | 926 | struct page *page = zspage->first_page; |
830e4bc5 | 927 | |
61989a80 NG |
928 | while (page) { |
929 | struct page *next_page; | |
930 | struct link_free *link; | |
5538c562 | 931 | unsigned int i = 1; |
af4ee5e9 | 932 | void *vaddr; |
61989a80 | 933 | |
3783689a | 934 | set_first_obj_offset(page, off); |
61989a80 | 935 | |
af4ee5e9 MK |
936 | vaddr = kmap_atomic(page); |
937 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
938 | |
939 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 940 | link->next = location_to_obj(page, i++); |
5538c562 | 941 | link += class->size / sizeof(*link); |
61989a80 NG |
942 | } |
943 | ||
944 | /* | |
945 | * We now come to the last (full or partial) object on this | |
946 | * page, which must point to the first object on the next | |
947 | * page (if present) | |
948 | */ | |
949 | next_page = get_next_page(page); | |
312fcae2 | 950 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 951 | kunmap_atomic(vaddr); |
61989a80 | 952 | page = next_page; |
5538c562 | 953 | off %= PAGE_SIZE; |
61989a80 | 954 | } |
bdb0af7c | 955 | |
3783689a MK |
956 | set_freeobj(zspage, |
957 | (unsigned long)location_to_obj(zspage->first_page, 0)); | |
61989a80 NG |
958 | } |
959 | ||
3783689a MK |
960 | static void create_page_chain(struct zspage *zspage, struct page *pages[], |
961 | int nr_pages) | |
61989a80 | 962 | { |
bdb0af7c MK |
963 | int i; |
964 | struct page *page; | |
965 | struct page *prev_page = NULL; | |
61989a80 NG |
966 | |
967 | /* | |
968 | * Allocate individual pages and link them together as: | |
3783689a MK |
969 | * 1. all pages are linked together using page->next |
970 | * 2. each sub-page point to zspage using page->private | |
61989a80 | 971 | * |
3783689a MK |
972 | * we set PG_private to identify the first page (i.e. no other sub-page |
973 | * has this flag set) and PG_private_2 to identify the last page. | |
61989a80 | 974 | */ |
bdb0af7c MK |
975 | for (i = 0; i < nr_pages; i++) { |
976 | page = pages[i]; | |
3783689a | 977 | set_page_private(page, (unsigned long)zspage); |
bdb0af7c | 978 | if (i == 0) { |
3783689a | 979 | zspage->first_page = page; |
a27545bf | 980 | SetPagePrivate(page); |
3783689a MK |
981 | } else { |
982 | prev_page->next = page; | |
61989a80 | 983 | } |
3783689a | 984 | if (i == nr_pages - 1) { |
a27545bf | 985 | SetPagePrivate2(page); |
3783689a MK |
986 | page->next = NULL; |
987 | } | |
61989a80 NG |
988 | prev_page = page; |
989 | } | |
bdb0af7c | 990 | } |
61989a80 | 991 | |
bdb0af7c MK |
992 | /* |
993 | * Allocate a zspage for the given size class | |
994 | */ | |
3783689a MK |
995 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
996 | struct size_class *class, | |
997 | gfp_t gfp) | |
bdb0af7c MK |
998 | { |
999 | int i; | |
bdb0af7c | 1000 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1001 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1002 | ||
1003 | if (!zspage) | |
1004 | return NULL; | |
1005 | ||
1006 | memset(zspage, 0, sizeof(struct zspage)); | |
61989a80 | 1007 | |
bdb0af7c MK |
1008 | for (i = 0; i < class->pages_per_zspage; i++) { |
1009 | struct page *page; | |
61989a80 | 1010 | |
3783689a | 1011 | page = alloc_page(gfp); |
bdb0af7c MK |
1012 | if (!page) { |
1013 | while (--i >= 0) | |
1014 | __free_page(pages[i]); | |
3783689a | 1015 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1016 | return NULL; |
1017 | } | |
1018 | pages[i] = page; | |
61989a80 NG |
1019 | } |
1020 | ||
3783689a MK |
1021 | create_page_chain(zspage, pages, class->pages_per_zspage); |
1022 | init_zspage(class, zspage); | |
bdb0af7c | 1023 | |
3783689a | 1024 | return zspage; |
61989a80 NG |
1025 | } |
1026 | ||
3783689a | 1027 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1028 | { |
1029 | int i; | |
3783689a | 1030 | struct zspage *zspage; |
61989a80 | 1031 | |
3783689a MK |
1032 | for (i = ZS_ALMOST_FULL; i <= ZS_ALMOST_EMPTY; i++) { |
1033 | zspage = list_first_entry_or_null(&class->fullness_list[i], | |
1034 | struct zspage, list); | |
1035 | if (zspage) | |
61989a80 NG |
1036 | break; |
1037 | } | |
1038 | ||
3783689a | 1039 | return zspage; |
61989a80 NG |
1040 | } |
1041 | ||
1b945aee | 1042 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1043 | static inline int __zs_cpu_up(struct mapping_area *area) |
1044 | { | |
1045 | /* | |
1046 | * Make sure we don't leak memory if a cpu UP notification | |
1047 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1048 | */ | |
1049 | if (area->vm) | |
1050 | return 0; | |
1051 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1052 | if (!area->vm) | |
1053 | return -ENOMEM; | |
1054 | return 0; | |
1055 | } | |
1056 | ||
1057 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1058 | { | |
1059 | if (area->vm) | |
1060 | free_vm_area(area->vm); | |
1061 | area->vm = NULL; | |
1062 | } | |
1063 | ||
1064 | static inline void *__zs_map_object(struct mapping_area *area, | |
1065 | struct page *pages[2], int off, int size) | |
1066 | { | |
f6f8ed47 | 1067 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1068 | area->vm_addr = area->vm->addr; |
1069 | return area->vm_addr + off; | |
1070 | } | |
1071 | ||
1072 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1073 | struct page *pages[2], int off, int size) | |
1074 | { | |
1075 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1076 | |
d95abbbb | 1077 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1078 | } |
1079 | ||
1b945aee | 1080 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1081 | |
1082 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1083 | { | |
1084 | /* | |
1085 | * Make sure we don't leak memory if a cpu UP notification | |
1086 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1087 | */ | |
1088 | if (area->vm_buf) | |
1089 | return 0; | |
40f9fb8c | 1090 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1091 | if (!area->vm_buf) |
1092 | return -ENOMEM; | |
1093 | return 0; | |
1094 | } | |
1095 | ||
1096 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1097 | { | |
40f9fb8c | 1098 | kfree(area->vm_buf); |
f553646a SJ |
1099 | area->vm_buf = NULL; |
1100 | } | |
1101 | ||
1102 | static void *__zs_map_object(struct mapping_area *area, | |
1103 | struct page *pages[2], int off, int size) | |
5f601902 | 1104 | { |
5f601902 SJ |
1105 | int sizes[2]; |
1106 | void *addr; | |
f553646a | 1107 | char *buf = area->vm_buf; |
5f601902 | 1108 | |
f553646a SJ |
1109 | /* disable page faults to match kmap_atomic() return conditions */ |
1110 | pagefault_disable(); | |
1111 | ||
1112 | /* no read fastpath */ | |
1113 | if (area->vm_mm == ZS_MM_WO) | |
1114 | goto out; | |
5f601902 SJ |
1115 | |
1116 | sizes[0] = PAGE_SIZE - off; | |
1117 | sizes[1] = size - sizes[0]; | |
1118 | ||
5f601902 SJ |
1119 | /* copy object to per-cpu buffer */ |
1120 | addr = kmap_atomic(pages[0]); | |
1121 | memcpy(buf, addr + off, sizes[0]); | |
1122 | kunmap_atomic(addr); | |
1123 | addr = kmap_atomic(pages[1]); | |
1124 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1125 | kunmap_atomic(addr); | |
f553646a SJ |
1126 | out: |
1127 | return area->vm_buf; | |
5f601902 SJ |
1128 | } |
1129 | ||
f553646a SJ |
1130 | static void __zs_unmap_object(struct mapping_area *area, |
1131 | struct page *pages[2], int off, int size) | |
5f601902 | 1132 | { |
5f601902 SJ |
1133 | int sizes[2]; |
1134 | void *addr; | |
2e40e163 | 1135 | char *buf; |
5f601902 | 1136 | |
f553646a SJ |
1137 | /* no write fastpath */ |
1138 | if (area->vm_mm == ZS_MM_RO) | |
1139 | goto out; | |
5f601902 | 1140 | |
7b60a685 | 1141 | buf = area->vm_buf; |
a82cbf07 YX |
1142 | buf = buf + ZS_HANDLE_SIZE; |
1143 | size -= ZS_HANDLE_SIZE; | |
1144 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1145 | |
5f601902 SJ |
1146 | sizes[0] = PAGE_SIZE - off; |
1147 | sizes[1] = size - sizes[0]; | |
1148 | ||
1149 | /* copy per-cpu buffer to object */ | |
1150 | addr = kmap_atomic(pages[0]); | |
1151 | memcpy(addr + off, buf, sizes[0]); | |
1152 | kunmap_atomic(addr); | |
1153 | addr = kmap_atomic(pages[1]); | |
1154 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1155 | kunmap_atomic(addr); | |
f553646a SJ |
1156 | |
1157 | out: | |
1158 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1159 | pagefault_enable(); | |
5f601902 | 1160 | } |
61989a80 | 1161 | |
1b945aee | 1162 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1163 | |
61989a80 NG |
1164 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1165 | void *pcpu) | |
1166 | { | |
f553646a | 1167 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1168 | struct mapping_area *area; |
1169 | ||
1170 | switch (action) { | |
1171 | case CPU_UP_PREPARE: | |
1172 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1173 | ret = __zs_cpu_up(area); |
1174 | if (ret) | |
1175 | return notifier_from_errno(ret); | |
61989a80 NG |
1176 | break; |
1177 | case CPU_DEAD: | |
1178 | case CPU_UP_CANCELED: | |
1179 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1180 | __zs_cpu_down(area); |
61989a80 NG |
1181 | break; |
1182 | } | |
1183 | ||
1184 | return NOTIFY_OK; | |
1185 | } | |
1186 | ||
1187 | static struct notifier_block zs_cpu_nb = { | |
1188 | .notifier_call = zs_cpu_notifier | |
1189 | }; | |
1190 | ||
b1b00a5b | 1191 | static int zs_register_cpu_notifier(void) |
61989a80 | 1192 | { |
b1b00a5b | 1193 | int cpu, uninitialized_var(ret); |
61989a80 | 1194 | |
f0e71fcd SB |
1195 | cpu_notifier_register_begin(); |
1196 | ||
1197 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1198 | for_each_online_cpu(cpu) { |
1199 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1200 | if (notifier_to_errno(ret)) |
1201 | break; | |
61989a80 | 1202 | } |
f0e71fcd SB |
1203 | |
1204 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1205 | return notifier_to_errno(ret); |
1206 | } | |
f0e71fcd | 1207 | |
66cdef66 | 1208 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1209 | { |
66cdef66 | 1210 | int cpu; |
40f9fb8c | 1211 | |
66cdef66 | 1212 | cpu_notifier_register_begin(); |
40f9fb8c | 1213 | |
66cdef66 GM |
1214 | for_each_online_cpu(cpu) |
1215 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1216 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1217 | |
66cdef66 | 1218 | cpu_notifier_register_done(); |
b1b00a5b SS |
1219 | } |
1220 | ||
66cdef66 | 1221 | static void init_zs_size_classes(void) |
b1b00a5b | 1222 | { |
66cdef66 | 1223 | int nr; |
c795779d | 1224 | |
66cdef66 GM |
1225 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1226 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1227 | nr += 1; | |
40f9fb8c | 1228 | |
66cdef66 | 1229 | zs_size_classes = nr; |
61989a80 NG |
1230 | } |
1231 | ||
9eec4cd5 JK |
1232 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1233 | { | |
1234 | if (prev->pages_per_zspage != pages_per_zspage) | |
1235 | return false; | |
1236 | ||
1237 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1238 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1239 | return false; | |
1240 | ||
1241 | return true; | |
1242 | } | |
1243 | ||
3783689a | 1244 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1245 | { |
3783689a | 1246 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1247 | } |
1248 | ||
66cdef66 GM |
1249 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1250 | { | |
1251 | return atomic_long_read(&pool->pages_allocated); | |
1252 | } | |
1253 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1254 | ||
4bbc0bc0 | 1255 | /** |
66cdef66 GM |
1256 | * zs_map_object - get address of allocated object from handle. |
1257 | * @pool: pool from which the object was allocated | |
1258 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1259 | * |
66cdef66 GM |
1260 | * Before using an object allocated from zs_malloc, it must be mapped using |
1261 | * this function. When done with the object, it must be unmapped using | |
1262 | * zs_unmap_object. | |
4bbc0bc0 | 1263 | * |
66cdef66 GM |
1264 | * Only one object can be mapped per cpu at a time. There is no protection |
1265 | * against nested mappings. | |
1266 | * | |
1267 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1268 | */ |
66cdef66 GM |
1269 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1270 | enum zs_mapmode mm) | |
61989a80 | 1271 | { |
3783689a | 1272 | struct zspage *zspage; |
66cdef66 | 1273 | struct page *page; |
2e40e163 | 1274 | unsigned long obj, obj_idx, off; |
61989a80 | 1275 | |
66cdef66 GM |
1276 | unsigned int class_idx; |
1277 | enum fullness_group fg; | |
1278 | struct size_class *class; | |
1279 | struct mapping_area *area; | |
1280 | struct page *pages[2]; | |
2e40e163 | 1281 | void *ret; |
61989a80 | 1282 | |
9eec4cd5 | 1283 | /* |
66cdef66 GM |
1284 | * Because we use per-cpu mapping areas shared among the |
1285 | * pools/users, we can't allow mapping in interrupt context | |
1286 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1287 | */ |
830e4bc5 | 1288 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1289 | |
312fcae2 MK |
1290 | /* From now on, migration cannot move the object */ |
1291 | pin_tag(handle); | |
1292 | ||
2e40e163 MK |
1293 | obj = handle_to_obj(handle); |
1294 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1295 | zspage = get_zspage(page); |
1296 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 GM |
1297 | class = pool->size_class[class_idx]; |
1298 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1299 | |
66cdef66 GM |
1300 | area = &get_cpu_var(zs_map_area); |
1301 | area->vm_mm = mm; | |
1302 | if (off + class->size <= PAGE_SIZE) { | |
1303 | /* this object is contained entirely within a page */ | |
1304 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1305 | ret = area->vm_addr + off; |
1306 | goto out; | |
61989a80 NG |
1307 | } |
1308 | ||
66cdef66 GM |
1309 | /* this object spans two pages */ |
1310 | pages[0] = page; | |
1311 | pages[1] = get_next_page(page); | |
1312 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1313 | |
2e40e163 MK |
1314 | ret = __zs_map_object(area, pages, off, class->size); |
1315 | out: | |
7b60a685 MK |
1316 | if (!class->huge) |
1317 | ret += ZS_HANDLE_SIZE; | |
1318 | ||
1319 | return ret; | |
61989a80 | 1320 | } |
66cdef66 | 1321 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1322 | |
66cdef66 | 1323 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1324 | { |
3783689a | 1325 | struct zspage *zspage; |
66cdef66 | 1326 | struct page *page; |
2e40e163 | 1327 | unsigned long obj, obj_idx, off; |
61989a80 | 1328 | |
66cdef66 GM |
1329 | unsigned int class_idx; |
1330 | enum fullness_group fg; | |
1331 | struct size_class *class; | |
1332 | struct mapping_area *area; | |
9eec4cd5 | 1333 | |
2e40e163 MK |
1334 | obj = handle_to_obj(handle); |
1335 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1336 | zspage = get_zspage(page); |
1337 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 GM |
1338 | class = pool->size_class[class_idx]; |
1339 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1340 | |
66cdef66 GM |
1341 | area = this_cpu_ptr(&zs_map_area); |
1342 | if (off + class->size <= PAGE_SIZE) | |
1343 | kunmap_atomic(area->vm_addr); | |
1344 | else { | |
1345 | struct page *pages[2]; | |
40f9fb8c | 1346 | |
66cdef66 GM |
1347 | pages[0] = page; |
1348 | pages[1] = get_next_page(page); | |
1349 | BUG_ON(!pages[1]); | |
1350 | ||
1351 | __zs_unmap_object(area, pages, off, class->size); | |
1352 | } | |
1353 | put_cpu_var(zs_map_area); | |
312fcae2 | 1354 | unpin_tag(handle); |
61989a80 | 1355 | } |
66cdef66 | 1356 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1357 | |
251cbb95 | 1358 | static unsigned long obj_malloc(struct size_class *class, |
3783689a | 1359 | struct zspage *zspage, unsigned long handle) |
c7806261 MK |
1360 | { |
1361 | unsigned long obj; | |
1362 | struct link_free *link; | |
1363 | ||
1364 | struct page *m_page; | |
1365 | unsigned long m_objidx, m_offset; | |
1366 | void *vaddr; | |
1367 | ||
312fcae2 | 1368 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1369 | obj = get_freeobj(zspage); |
c7806261 MK |
1370 | obj_to_location(obj, &m_page, &m_objidx); |
1371 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1372 | ||
1373 | vaddr = kmap_atomic(m_page); | |
1374 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3783689a | 1375 | set_freeobj(zspage, (unsigned long)link->next); |
7b60a685 MK |
1376 | if (!class->huge) |
1377 | /* record handle in the header of allocated chunk */ | |
1378 | link->handle = handle; | |
1379 | else | |
3783689a MK |
1380 | /* record handle to page->index */ |
1381 | zspage->first_page->index = handle; | |
1382 | ||
c7806261 | 1383 | kunmap_atomic(vaddr); |
3783689a | 1384 | mod_zspage_inuse(zspage, 1); |
c7806261 MK |
1385 | zs_stat_inc(class, OBJ_USED, 1); |
1386 | ||
1387 | return obj; | |
1388 | } | |
1389 | ||
1390 | ||
61989a80 NG |
1391 | /** |
1392 | * zs_malloc - Allocate block of given size from pool. | |
1393 | * @pool: pool to allocate from | |
1394 | * @size: size of block to allocate | |
61989a80 | 1395 | * |
00a61d86 | 1396 | * On success, handle to the allocated object is returned, |
c2344348 | 1397 | * otherwise 0. |
61989a80 NG |
1398 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1399 | */ | |
d0d8da2d | 1400 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1401 | { |
2e40e163 | 1402 | unsigned long handle, obj; |
61989a80 | 1403 | struct size_class *class; |
3783689a | 1404 | struct zspage *zspage; |
61989a80 | 1405 | |
7b60a685 | 1406 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1407 | return 0; |
1408 | ||
3783689a | 1409 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1410 | if (!handle) |
c2344348 | 1411 | return 0; |
61989a80 | 1412 | |
2e40e163 MK |
1413 | /* extra space in chunk to keep the handle */ |
1414 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1415 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1416 | |
1417 | spin_lock(&class->lock); | |
3783689a | 1418 | zspage = find_get_zspage(class); |
61989a80 | 1419 | |
3783689a | 1420 | if (!zspage) { |
61989a80 | 1421 | spin_unlock(&class->lock); |
3783689a MK |
1422 | zspage = alloc_zspage(pool, class, gfp); |
1423 | if (unlikely(!zspage)) { | |
1424 | cache_free_handle(pool, handle); | |
c2344348 | 1425 | return 0; |
2e40e163 | 1426 | } |
61989a80 | 1427 | |
3783689a | 1428 | set_zspage_mapping(zspage, class->index, ZS_EMPTY); |
13de8933 MK |
1429 | atomic_long_add(class->pages_per_zspage, |
1430 | &pool->pages_allocated); | |
0f050d99 | 1431 | |
61989a80 | 1432 | spin_lock(&class->lock); |
0f050d99 GM |
1433 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1434 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1435 | } |
1436 | ||
3783689a | 1437 | obj = obj_malloc(class, zspage, handle); |
61989a80 | 1438 | /* Now move the zspage to another fullness group, if required */ |
3783689a | 1439 | fix_fullness_group(class, zspage); |
2e40e163 | 1440 | record_obj(handle, obj); |
61989a80 NG |
1441 | spin_unlock(&class->lock); |
1442 | ||
2e40e163 | 1443 | return handle; |
61989a80 NG |
1444 | } |
1445 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1446 | ||
1ee47165 | 1447 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1448 | { |
1449 | struct link_free *link; | |
3783689a MK |
1450 | struct zspage *zspage; |
1451 | struct page *f_page; | |
c7806261 | 1452 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1453 | void *vaddr; |
61989a80 | 1454 | |
312fcae2 | 1455 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1456 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1457 | zspage = get_zspage(f_page); |
61989a80 | 1458 | |
61989a80 NG |
1459 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1460 | ||
c7806261 | 1461 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1462 | |
1463 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1464 | link = (struct link_free *)(vaddr + f_offset); |
3783689a | 1465 | link->next = (void *)get_freeobj(zspage); |
af4ee5e9 | 1466 | kunmap_atomic(vaddr); |
3783689a MK |
1467 | set_freeobj(zspage, obj); |
1468 | mod_zspage_inuse(zspage, -1); | |
0f050d99 | 1469 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1470 | } |
1471 | ||
1472 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1473 | { | |
3783689a MK |
1474 | struct zspage *zspage; |
1475 | struct page *f_page; | |
c7806261 MK |
1476 | unsigned long obj, f_objidx; |
1477 | int class_idx; | |
1478 | struct size_class *class; | |
1479 | enum fullness_group fullness; | |
1480 | ||
1481 | if (unlikely(!handle)) | |
1482 | return; | |
1483 | ||
312fcae2 | 1484 | pin_tag(handle); |
c7806261 | 1485 | obj = handle_to_obj(handle); |
c7806261 | 1486 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1487 | zspage = get_zspage(f_page); |
c7806261 | 1488 | |
3783689a | 1489 | get_zspage_mapping(zspage, &class_idx, &fullness); |
c7806261 MK |
1490 | class = pool->size_class[class_idx]; |
1491 | ||
1492 | spin_lock(&class->lock); | |
1ee47165 | 1493 | obj_free(class, obj); |
3783689a | 1494 | fullness = fix_fullness_group(class, zspage); |
312fcae2 | 1495 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1496 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1497 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1498 | atomic_long_sub(class->pages_per_zspage, |
1499 | &pool->pages_allocated); | |
3783689a | 1500 | free_zspage(pool, zspage); |
312fcae2 | 1501 | } |
61989a80 | 1502 | spin_unlock(&class->lock); |
312fcae2 | 1503 | unpin_tag(handle); |
61989a80 | 1504 | |
3783689a | 1505 | cache_free_handle(pool, handle); |
312fcae2 MK |
1506 | } |
1507 | EXPORT_SYMBOL_GPL(zs_free); | |
1508 | ||
251cbb95 MK |
1509 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1510 | unsigned long src) | |
312fcae2 MK |
1511 | { |
1512 | struct page *s_page, *d_page; | |
1513 | unsigned long s_objidx, d_objidx; | |
1514 | unsigned long s_off, d_off; | |
1515 | void *s_addr, *d_addr; | |
1516 | int s_size, d_size, size; | |
1517 | int written = 0; | |
1518 | ||
1519 | s_size = d_size = class->size; | |
1520 | ||
1521 | obj_to_location(src, &s_page, &s_objidx); | |
1522 | obj_to_location(dst, &d_page, &d_objidx); | |
1523 | ||
1524 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1525 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1526 | ||
1527 | if (s_off + class->size > PAGE_SIZE) | |
1528 | s_size = PAGE_SIZE - s_off; | |
1529 | ||
1530 | if (d_off + class->size > PAGE_SIZE) | |
1531 | d_size = PAGE_SIZE - d_off; | |
1532 | ||
1533 | s_addr = kmap_atomic(s_page); | |
1534 | d_addr = kmap_atomic(d_page); | |
1535 | ||
1536 | while (1) { | |
1537 | size = min(s_size, d_size); | |
1538 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1539 | written += size; | |
1540 | ||
1541 | if (written == class->size) | |
1542 | break; | |
1543 | ||
495819ea SS |
1544 | s_off += size; |
1545 | s_size -= size; | |
1546 | d_off += size; | |
1547 | d_size -= size; | |
1548 | ||
1549 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1550 | kunmap_atomic(d_addr); |
1551 | kunmap_atomic(s_addr); | |
1552 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1553 | s_addr = kmap_atomic(s_page); |
1554 | d_addr = kmap_atomic(d_page); | |
1555 | s_size = class->size - written; | |
1556 | s_off = 0; | |
312fcae2 MK |
1557 | } |
1558 | ||
495819ea | 1559 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1560 | kunmap_atomic(d_addr); |
1561 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1562 | d_addr = kmap_atomic(d_page); |
1563 | d_size = class->size - written; | |
1564 | d_off = 0; | |
312fcae2 MK |
1565 | } |
1566 | } | |
1567 | ||
1568 | kunmap_atomic(d_addr); | |
1569 | kunmap_atomic(s_addr); | |
1570 | } | |
1571 | ||
1572 | /* | |
1573 | * Find alloced object in zspage from index object and | |
1574 | * return handle. | |
1575 | */ | |
251cbb95 MK |
1576 | static unsigned long find_alloced_obj(struct size_class *class, |
1577 | struct page *page, int index) | |
312fcae2 MK |
1578 | { |
1579 | unsigned long head; | |
1580 | int offset = 0; | |
1581 | unsigned long handle = 0; | |
1582 | void *addr = kmap_atomic(page); | |
1583 | ||
3783689a | 1584 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1585 | offset += class->size * index; |
1586 | ||
1587 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1588 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1589 | if (head & OBJ_ALLOCATED_TAG) { |
1590 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1591 | if (trypin_tag(handle)) | |
1592 | break; | |
1593 | handle = 0; | |
1594 | } | |
1595 | ||
1596 | offset += class->size; | |
1597 | index++; | |
1598 | } | |
1599 | ||
1600 | kunmap_atomic(addr); | |
1601 | return handle; | |
1602 | } | |
1603 | ||
1604 | struct zs_compact_control { | |
3783689a | 1605 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1606 | struct page *s_page; |
1607 | /* Destination page for migration which should be a first page | |
1608 | * of zspage. */ | |
1609 | struct page *d_page; | |
1610 | /* Starting object index within @s_page which used for live object | |
1611 | * in the subpage. */ | |
1612 | int index; | |
312fcae2 MK |
1613 | }; |
1614 | ||
1615 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1616 | struct zs_compact_control *cc) | |
1617 | { | |
1618 | unsigned long used_obj, free_obj; | |
1619 | unsigned long handle; | |
1620 | struct page *s_page = cc->s_page; | |
1621 | struct page *d_page = cc->d_page; | |
1622 | unsigned long index = cc->index; | |
312fcae2 MK |
1623 | int ret = 0; |
1624 | ||
1625 | while (1) { | |
251cbb95 | 1626 | handle = find_alloced_obj(class, s_page, index); |
312fcae2 MK |
1627 | if (!handle) { |
1628 | s_page = get_next_page(s_page); | |
1629 | if (!s_page) | |
1630 | break; | |
1631 | index = 0; | |
1632 | continue; | |
1633 | } | |
1634 | ||
1635 | /* Stop if there is no more space */ | |
3783689a | 1636 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1637 | unpin_tag(handle); |
1638 | ret = -ENOMEM; | |
1639 | break; | |
1640 | } | |
1641 | ||
1642 | used_obj = handle_to_obj(handle); | |
3783689a | 1643 | free_obj = obj_malloc(class, get_zspage(d_page), handle); |
251cbb95 | 1644 | zs_object_copy(class, free_obj, used_obj); |
312fcae2 | 1645 | index++; |
c102f07c JL |
1646 | /* |
1647 | * record_obj updates handle's value to free_obj and it will | |
1648 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1649 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1650 | * let's keep the lock bit. | |
1651 | */ | |
1652 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1653 | record_obj(handle, free_obj); |
1654 | unpin_tag(handle); | |
1ee47165 | 1655 | obj_free(class, used_obj); |
312fcae2 MK |
1656 | } |
1657 | ||
1658 | /* Remember last position in this iteration */ | |
1659 | cc->s_page = s_page; | |
1660 | cc->index = index; | |
312fcae2 MK |
1661 | |
1662 | return ret; | |
1663 | } | |
1664 | ||
3783689a | 1665 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1666 | { |
1667 | int i; | |
3783689a MK |
1668 | struct zspage *zspage; |
1669 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1670 | |
3783689a MK |
1671 | if (!source) { |
1672 | fg[0] = ZS_ALMOST_FULL; | |
1673 | fg[1] = ZS_ALMOST_EMPTY; | |
1674 | } | |
1675 | ||
1676 | for (i = 0; i < 2; i++) { | |
1677 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1678 | struct zspage, list); | |
1679 | if (zspage) { | |
1680 | remove_zspage(class, zspage, fg[i]); | |
1681 | return zspage; | |
312fcae2 MK |
1682 | } |
1683 | } | |
1684 | ||
3783689a | 1685 | return zspage; |
312fcae2 MK |
1686 | } |
1687 | ||
860c707d | 1688 | /* |
3783689a | 1689 | * putback_zspage - add @zspage into right class's fullness list |
860c707d SS |
1690 | * @pool: target pool |
1691 | * @class: destination class | |
3783689a | 1692 | * @zspage: target page |
860c707d | 1693 | * |
3783689a | 1694 | * Return @zspage's fullness_group |
860c707d SS |
1695 | */ |
1696 | static enum fullness_group putback_zspage(struct zs_pool *pool, | |
1697 | struct size_class *class, | |
3783689a | 1698 | struct zspage *zspage) |
312fcae2 | 1699 | { |
312fcae2 MK |
1700 | enum fullness_group fullness; |
1701 | ||
3783689a MK |
1702 | fullness = get_fullness_group(class, zspage); |
1703 | insert_zspage(class, zspage, fullness); | |
1704 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1705 | |
13de8933 | 1706 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1707 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1708 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1709 | atomic_long_sub(class->pages_per_zspage, |
1710 | &pool->pages_allocated); | |
312fcae2 | 1711 | |
3783689a | 1712 | free_zspage(pool, zspage); |
13de8933 | 1713 | } |
860c707d SS |
1714 | |
1715 | return fullness; | |
61989a80 | 1716 | } |
312fcae2 | 1717 | |
04f05909 SS |
1718 | /* |
1719 | * | |
1720 | * Based on the number of unused allocated objects calculate | |
1721 | * and return the number of pages that we can free. | |
04f05909 SS |
1722 | */ |
1723 | static unsigned long zs_can_compact(struct size_class *class) | |
1724 | { | |
1725 | unsigned long obj_wasted; | |
44f43e99 SS |
1726 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
1727 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 1728 | |
44f43e99 SS |
1729 | if (obj_allocated <= obj_used) |
1730 | return 0; | |
04f05909 | 1731 | |
44f43e99 | 1732 | obj_wasted = obj_allocated - obj_used; |
04f05909 SS |
1733 | obj_wasted /= get_maxobj_per_zspage(class->size, |
1734 | class->pages_per_zspage); | |
1735 | ||
6cbf16b3 | 1736 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
1737 | } |
1738 | ||
7d3f3938 | 1739 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1740 | { |
312fcae2 | 1741 | struct zs_compact_control cc; |
3783689a MK |
1742 | struct zspage *src_zspage; |
1743 | struct zspage *dst_zspage = NULL; | |
312fcae2 | 1744 | |
312fcae2 | 1745 | spin_lock(&class->lock); |
3783689a | 1746 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 1747 | |
04f05909 SS |
1748 | if (!zs_can_compact(class)) |
1749 | break; | |
1750 | ||
312fcae2 | 1751 | cc.index = 0; |
3783689a | 1752 | cc.s_page = src_zspage->first_page; |
312fcae2 | 1753 | |
3783689a MK |
1754 | while ((dst_zspage = isolate_zspage(class, false))) { |
1755 | cc.d_page = dst_zspage->first_page; | |
312fcae2 | 1756 | /* |
0dc63d48 SS |
1757 | * If there is no more space in dst_page, resched |
1758 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1759 | */ |
1760 | if (!migrate_zspage(pool, class, &cc)) | |
1761 | break; | |
1762 | ||
3783689a | 1763 | putback_zspage(pool, class, dst_zspage); |
312fcae2 MK |
1764 | } |
1765 | ||
1766 | /* Stop if we couldn't find slot */ | |
3783689a | 1767 | if (dst_zspage == NULL) |
312fcae2 MK |
1768 | break; |
1769 | ||
3783689a MK |
1770 | putback_zspage(pool, class, dst_zspage); |
1771 | if (putback_zspage(pool, class, src_zspage) == ZS_EMPTY) | |
6cbf16b3 | 1772 | pool->stats.pages_compacted += class->pages_per_zspage; |
312fcae2 | 1773 | spin_unlock(&class->lock); |
312fcae2 MK |
1774 | cond_resched(); |
1775 | spin_lock(&class->lock); | |
1776 | } | |
1777 | ||
3783689a MK |
1778 | if (src_zspage) |
1779 | putback_zspage(pool, class, src_zspage); | |
312fcae2 | 1780 | |
7d3f3938 | 1781 | spin_unlock(&class->lock); |
312fcae2 MK |
1782 | } |
1783 | ||
1784 | unsigned long zs_compact(struct zs_pool *pool) | |
1785 | { | |
1786 | int i; | |
312fcae2 MK |
1787 | struct size_class *class; |
1788 | ||
1789 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1790 | class = pool->size_class[i]; | |
1791 | if (!class) | |
1792 | continue; | |
1793 | if (class->index != i) | |
1794 | continue; | |
7d3f3938 | 1795 | __zs_compact(pool, class); |
312fcae2 MK |
1796 | } |
1797 | ||
860c707d | 1798 | return pool->stats.pages_compacted; |
312fcae2 MK |
1799 | } |
1800 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1801 | |
7d3f3938 SS |
1802 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1803 | { | |
1804 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1805 | } | |
1806 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1807 | ||
ab9d306d SS |
1808 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
1809 | struct shrink_control *sc) | |
1810 | { | |
1811 | unsigned long pages_freed; | |
1812 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1813 | shrinker); | |
1814 | ||
1815 | pages_freed = pool->stats.pages_compacted; | |
1816 | /* | |
1817 | * Compact classes and calculate compaction delta. | |
1818 | * Can run concurrently with a manually triggered | |
1819 | * (by user) compaction. | |
1820 | */ | |
1821 | pages_freed = zs_compact(pool) - pages_freed; | |
1822 | ||
1823 | return pages_freed ? pages_freed : SHRINK_STOP; | |
1824 | } | |
1825 | ||
1826 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
1827 | struct shrink_control *sc) | |
1828 | { | |
1829 | int i; | |
1830 | struct size_class *class; | |
1831 | unsigned long pages_to_free = 0; | |
1832 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1833 | shrinker); | |
1834 | ||
ab9d306d SS |
1835 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1836 | class = pool->size_class[i]; | |
1837 | if (!class) | |
1838 | continue; | |
1839 | if (class->index != i) | |
1840 | continue; | |
1841 | ||
ab9d306d | 1842 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
1843 | } |
1844 | ||
1845 | return pages_to_free; | |
1846 | } | |
1847 | ||
1848 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
1849 | { | |
1850 | if (pool->shrinker_enabled) { | |
1851 | unregister_shrinker(&pool->shrinker); | |
1852 | pool->shrinker_enabled = false; | |
1853 | } | |
1854 | } | |
1855 | ||
1856 | static int zs_register_shrinker(struct zs_pool *pool) | |
1857 | { | |
1858 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
1859 | pool->shrinker.count_objects = zs_shrinker_count; | |
1860 | pool->shrinker.batch = 0; | |
1861 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
1862 | ||
1863 | return register_shrinker(&pool->shrinker); | |
1864 | } | |
1865 | ||
00a61d86 | 1866 | /** |
66cdef66 GM |
1867 | * zs_create_pool - Creates an allocation pool to work from. |
1868 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1869 | * |
66cdef66 GM |
1870 | * This function must be called before anything when using |
1871 | * the zsmalloc allocator. | |
166cfda7 | 1872 | * |
66cdef66 GM |
1873 | * On success, a pointer to the newly created pool is returned, |
1874 | * otherwise NULL. | |
396b7fd6 | 1875 | */ |
d0d8da2d | 1876 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 1877 | { |
66cdef66 GM |
1878 | int i; |
1879 | struct zs_pool *pool; | |
1880 | struct size_class *prev_class = NULL; | |
61989a80 | 1881 | |
66cdef66 GM |
1882 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1883 | if (!pool) | |
1884 | return NULL; | |
61989a80 | 1885 | |
66cdef66 GM |
1886 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1887 | GFP_KERNEL); | |
1888 | if (!pool->size_class) { | |
1889 | kfree(pool); | |
1890 | return NULL; | |
1891 | } | |
61989a80 | 1892 | |
2e40e163 MK |
1893 | pool->name = kstrdup(name, GFP_KERNEL); |
1894 | if (!pool->name) | |
1895 | goto err; | |
1896 | ||
3783689a | 1897 | if (create_cache(pool)) |
2e40e163 MK |
1898 | goto err; |
1899 | ||
c60369f0 | 1900 | /* |
66cdef66 GM |
1901 | * Iterate reversly, because, size of size_class that we want to use |
1902 | * for merging should be larger or equal to current size. | |
c60369f0 | 1903 | */ |
66cdef66 GM |
1904 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1905 | int size; | |
1906 | int pages_per_zspage; | |
1907 | struct size_class *class; | |
3783689a | 1908 | int fullness = 0; |
c60369f0 | 1909 | |
66cdef66 GM |
1910 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1911 | if (size > ZS_MAX_ALLOC_SIZE) | |
1912 | size = ZS_MAX_ALLOC_SIZE; | |
1913 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1914 | |
66cdef66 GM |
1915 | /* |
1916 | * size_class is used for normal zsmalloc operation such | |
1917 | * as alloc/free for that size. Although it is natural that we | |
1918 | * have one size_class for each size, there is a chance that we | |
1919 | * can get more memory utilization if we use one size_class for | |
1920 | * many different sizes whose size_class have same | |
1921 | * characteristics. So, we makes size_class point to | |
1922 | * previous size_class if possible. | |
1923 | */ | |
1924 | if (prev_class) { | |
1925 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1926 | pool->size_class[i] = prev_class; | |
1927 | continue; | |
1928 | } | |
1929 | } | |
1930 | ||
1931 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1932 | if (!class) | |
1933 | goto err; | |
1934 | ||
1935 | class->size = size; | |
1936 | class->index = i; | |
1937 | class->pages_per_zspage = pages_per_zspage; | |
1fc6e27d MK |
1938 | class->objs_per_zspage = class->pages_per_zspage * |
1939 | PAGE_SIZE / class->size; | |
1940 | if (pages_per_zspage == 1 && class->objs_per_zspage == 1) | |
7b60a685 | 1941 | class->huge = true; |
66cdef66 GM |
1942 | spin_lock_init(&class->lock); |
1943 | pool->size_class[i] = class; | |
3783689a MK |
1944 | for (fullness = ZS_ALMOST_FULL; fullness <= ZS_ALMOST_EMPTY; |
1945 | fullness++) | |
1946 | INIT_LIST_HEAD(&class->fullness_list[fullness]); | |
66cdef66 GM |
1947 | |
1948 | prev_class = class; | |
61989a80 NG |
1949 | } |
1950 | ||
d34f6157 DS |
1951 | /* debug only, don't abort if it fails */ |
1952 | zs_pool_stat_create(pool, name); | |
0f050d99 | 1953 | |
ab9d306d SS |
1954 | /* |
1955 | * Not critical, we still can use the pool | |
1956 | * and user can trigger compaction manually. | |
1957 | */ | |
1958 | if (zs_register_shrinker(pool) == 0) | |
1959 | pool->shrinker_enabled = true; | |
66cdef66 GM |
1960 | return pool; |
1961 | ||
1962 | err: | |
1963 | zs_destroy_pool(pool); | |
1964 | return NULL; | |
61989a80 | 1965 | } |
66cdef66 | 1966 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1967 | |
66cdef66 | 1968 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1969 | { |
66cdef66 | 1970 | int i; |
61989a80 | 1971 | |
ab9d306d | 1972 | zs_unregister_shrinker(pool); |
0f050d99 GM |
1973 | zs_pool_stat_destroy(pool); |
1974 | ||
66cdef66 GM |
1975 | for (i = 0; i < zs_size_classes; i++) { |
1976 | int fg; | |
1977 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1978 | |
66cdef66 GM |
1979 | if (!class) |
1980 | continue; | |
61989a80 | 1981 | |
66cdef66 GM |
1982 | if (class->index != i) |
1983 | continue; | |
61989a80 | 1984 | |
3783689a MK |
1985 | for (fg = ZS_ALMOST_FULL; fg <= ZS_ALMOST_EMPTY; fg++) { |
1986 | if (!list_empty(&class->fullness_list[fg])) { | |
66cdef66 GM |
1987 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
1988 | class->size, fg); | |
1989 | } | |
1990 | } | |
1991 | kfree(class); | |
1992 | } | |
f553646a | 1993 | |
3783689a | 1994 | destroy_cache(pool); |
66cdef66 | 1995 | kfree(pool->size_class); |
0f050d99 | 1996 | kfree(pool->name); |
66cdef66 GM |
1997 | kfree(pool); |
1998 | } | |
1999 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2000 | |
66cdef66 GM |
2001 | static int __init zs_init(void) |
2002 | { | |
2003 | int ret = zs_register_cpu_notifier(); | |
2004 | ||
0f050d99 GM |
2005 | if (ret) |
2006 | goto notifier_fail; | |
66cdef66 GM |
2007 | |
2008 | init_zs_size_classes(); | |
2009 | ||
2010 | #ifdef CONFIG_ZPOOL | |
2011 | zpool_register_driver(&zs_zpool_driver); | |
2012 | #endif | |
0f050d99 | 2013 | |
4abaac9b DS |
2014 | zs_stat_init(); |
2015 | ||
66cdef66 | 2016 | return 0; |
0f050d99 | 2017 | |
0f050d99 GM |
2018 | notifier_fail: |
2019 | zs_unregister_cpu_notifier(); | |
2020 | ||
2021 | return ret; | |
61989a80 | 2022 | } |
61989a80 | 2023 | |
66cdef66 | 2024 | static void __exit zs_exit(void) |
61989a80 | 2025 | { |
66cdef66 GM |
2026 | #ifdef CONFIG_ZPOOL |
2027 | zpool_unregister_driver(&zs_zpool_driver); | |
2028 | #endif | |
2029 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
2030 | |
2031 | zs_stat_exit(); | |
61989a80 | 2032 | } |
069f101f BH |
2033 | |
2034 | module_init(zs_init); | |
2035 | module_exit(zs_exit); | |
2036 | ||
2037 | MODULE_LICENSE("Dual BSD/GPL"); | |
2038 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |