]>
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 | 73 | * |
bfd093f5 | 74 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
75 | * |
76 | * This is made more complicated by various memory models and PAE. | |
77 | */ | |
78 | ||
79 | #ifndef MAX_PHYSMEM_BITS | |
80 | #ifdef CONFIG_HIGHMEM64G | |
81 | #define MAX_PHYSMEM_BITS 36 | |
82 | #else /* !CONFIG_HIGHMEM64G */ | |
83 | /* | |
84 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
85 | * be PAGE_SHIFT | |
86 | */ | |
87 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
88 | #endif | |
89 | #endif | |
90 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
91 | |
92 | /* | |
93 | * Memory for allocating for handle keeps object position by | |
94 | * encoding <page, obj_idx> and the encoded value has a room | |
95 | * in least bit(ie, look at obj_to_location). | |
96 | * We use the bit to synchronize between object access by | |
97 | * user and migration. | |
98 | */ | |
99 | #define HANDLE_PIN_BIT 0 | |
100 | ||
101 | /* | |
102 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
103 | * to identify the object was allocated or not. | |
104 | * It's okay to add the status bit in the least bit because | |
105 | * header keeps handle which is 4byte-aligned address so we | |
106 | * have room for two bit at least. | |
107 | */ | |
108 | #define OBJ_ALLOCATED_TAG 1 | |
109 | #define OBJ_TAG_BITS 1 | |
110 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
111 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
112 | ||
113 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
114 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
115 | #define ZS_MIN_ALLOC_SIZE \ | |
116 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 117 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 118 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
119 | |
120 | /* | |
7eb52512 | 121 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
122 | * trader-off here: |
123 | * - Large number of size classes is potentially wasteful as free page are | |
124 | * spread across these classes | |
125 | * - Small number of size classes causes large internal fragmentation | |
126 | * - Probably its better to use specific size classes (empirically | |
127 | * determined). NOTE: all those class sizes must be set as multiple of | |
128 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
129 | * | |
130 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
131 | * (reason above) | |
132 | */ | |
3783689a | 133 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
0959c63f SJ |
134 | |
135 | /* | |
136 | * We do not maintain any list for completely empty or full pages | |
137 | */ | |
138 | enum fullness_group { | |
139 | ZS_ALMOST_FULL, | |
140 | ZS_ALMOST_EMPTY, | |
0959c63f SJ |
141 | ZS_EMPTY, |
142 | ZS_FULL | |
143 | }; | |
144 | ||
0f050d99 GM |
145 | enum zs_stat_type { |
146 | OBJ_ALLOCATED, | |
147 | OBJ_USED, | |
248ca1b0 MK |
148 | CLASS_ALMOST_FULL, |
149 | CLASS_ALMOST_EMPTY, | |
0f050d99 GM |
150 | }; |
151 | ||
6fe5186f SS |
152 | #ifdef CONFIG_ZSMALLOC_STAT |
153 | #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1) | |
154 | #else | |
155 | #define NR_ZS_STAT_TYPE (OBJ_USED + 1) | |
156 | #endif | |
157 | ||
0f050d99 GM |
158 | struct zs_size_stat { |
159 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
160 | }; | |
161 | ||
57244594 SS |
162 | #ifdef CONFIG_ZSMALLOC_STAT |
163 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
164 | #endif |
165 | ||
40f9fb8c MG |
166 | /* |
167 | * number of size_classes | |
168 | */ | |
169 | static int zs_size_classes; | |
170 | ||
0959c63f SJ |
171 | /* |
172 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
173 | * n <= N / f, where | |
174 | * n = number of allocated objects | |
175 | * N = total number of objects zspage can store | |
6dd9737e | 176 | * f = fullness_threshold_frac |
0959c63f SJ |
177 | * |
178 | * Similarly, we assign zspage to: | |
179 | * ZS_ALMOST_FULL when n > N / f | |
180 | * ZS_EMPTY when n == 0 | |
181 | * ZS_FULL when n == N | |
182 | * | |
183 | * (see: fix_fullness_group()) | |
184 | */ | |
185 | static const int fullness_threshold_frac = 4; | |
186 | ||
187 | struct size_class { | |
57244594 | 188 | spinlock_t lock; |
3783689a | 189 | struct list_head fullness_list[2]; |
0959c63f SJ |
190 | /* |
191 | * Size of objects stored in this class. Must be multiple | |
192 | * of ZS_ALIGN. | |
193 | */ | |
194 | int size; | |
1fc6e27d | 195 | int objs_per_zspage; |
0959c63f SJ |
196 | unsigned int index; |
197 | ||
0f050d99 | 198 | struct zs_size_stat stats; |
0959c63f | 199 | |
7dfa4612 WY |
200 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
201 | int pages_per_zspage; | |
57244594 SS |
202 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
203 | bool huge; | |
0959c63f SJ |
204 | }; |
205 | ||
206 | /* | |
207 | * Placed within free objects to form a singly linked list. | |
3783689a | 208 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
209 | * |
210 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
211 | */ | |
212 | struct link_free { | |
2e40e163 MK |
213 | union { |
214 | /* | |
bfd093f5 | 215 | * Free object index; |
2e40e163 MK |
216 | * It's valid for non-allocated object |
217 | */ | |
bfd093f5 | 218 | unsigned long next; |
2e40e163 MK |
219 | /* |
220 | * Handle of allocated object. | |
221 | */ | |
222 | unsigned long handle; | |
223 | }; | |
0959c63f SJ |
224 | }; |
225 | ||
226 | struct zs_pool { | |
6f3526d6 | 227 | const char *name; |
0f050d99 | 228 | |
40f9fb8c | 229 | struct size_class **size_class; |
2e40e163 | 230 | struct kmem_cache *handle_cachep; |
3783689a | 231 | struct kmem_cache *zspage_cachep; |
0959c63f | 232 | |
13de8933 | 233 | atomic_long_t pages_allocated; |
0f050d99 | 234 | |
7d3f3938 | 235 | struct zs_pool_stats stats; |
ab9d306d SS |
236 | |
237 | /* Compact classes */ | |
238 | struct shrinker shrinker; | |
239 | /* | |
240 | * To signify that register_shrinker() was successful | |
241 | * and unregister_shrinker() will not Oops. | |
242 | */ | |
243 | bool shrinker_enabled; | |
0f050d99 GM |
244 | #ifdef CONFIG_ZSMALLOC_STAT |
245 | struct dentry *stat_dentry; | |
246 | #endif | |
0959c63f | 247 | }; |
61989a80 NG |
248 | |
249 | /* | |
250 | * A zspage's class index and fullness group | |
251 | * are encoded in its (first)page->mapping | |
252 | */ | |
3783689a MK |
253 | #define FULLNESS_BITS 2 |
254 | #define CLASS_BITS 8 | |
4f42047b | 255 | |
3783689a MK |
256 | struct zspage { |
257 | struct { | |
258 | unsigned int fullness:FULLNESS_BITS; | |
259 | unsigned int class:CLASS_BITS; | |
260 | }; | |
261 | unsigned int inuse; | |
bfd093f5 | 262 | unsigned int freeobj; |
3783689a MK |
263 | struct page *first_page; |
264 | struct list_head list; /* fullness list */ | |
265 | }; | |
61989a80 | 266 | |
f553646a | 267 | struct mapping_area { |
1b945aee | 268 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
269 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
270 | #else | |
271 | char *vm_buf; /* copy buffer for objects that span pages */ | |
272 | #endif | |
273 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
274 | enum zs_mapmode vm_mm; /* mapping mode */ | |
275 | }; | |
276 | ||
3783689a | 277 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
278 | { |
279 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
280 | 0, 0, NULL); | |
3783689a MK |
281 | if (!pool->handle_cachep) |
282 | return 1; | |
283 | ||
284 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
285 | 0, 0, NULL); | |
286 | if (!pool->zspage_cachep) { | |
287 | kmem_cache_destroy(pool->handle_cachep); | |
288 | pool->handle_cachep = NULL; | |
289 | return 1; | |
290 | } | |
291 | ||
292 | return 0; | |
2e40e163 MK |
293 | } |
294 | ||
3783689a | 295 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 296 | { |
cd10add0 | 297 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 298 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
299 | } |
300 | ||
3783689a | 301 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
302 | { |
303 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
d0d8da2d | 304 | gfp & ~__GFP_HIGHMEM); |
2e40e163 MK |
305 | } |
306 | ||
3783689a | 307 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
308 | { |
309 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
310 | } | |
311 | ||
3783689a MK |
312 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
313 | { | |
314 | return kmem_cache_alloc(pool->zspage_cachep, flags & ~__GFP_HIGHMEM); | |
315 | }; | |
316 | ||
317 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
318 | { | |
319 | kmem_cache_free(pool->zspage_cachep, zspage); | |
320 | } | |
321 | ||
2e40e163 MK |
322 | static void record_obj(unsigned long handle, unsigned long obj) |
323 | { | |
c102f07c JL |
324 | /* |
325 | * lsb of @obj represents handle lock while other bits | |
326 | * represent object value the handle is pointing so | |
327 | * updating shouldn't do store tearing. | |
328 | */ | |
329 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
330 | } |
331 | ||
c795779d DS |
332 | /* zpool driver */ |
333 | ||
334 | #ifdef CONFIG_ZPOOL | |
335 | ||
6f3526d6 | 336 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 337 | const struct zpool_ops *zpool_ops, |
479305fd | 338 | struct zpool *zpool) |
c795779d | 339 | { |
d0d8da2d SS |
340 | /* |
341 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
342 | * different contexts and its caller must provide a valid | |
343 | * gfp mask. | |
344 | */ | |
345 | return zs_create_pool(name); | |
c795779d DS |
346 | } |
347 | ||
348 | static void zs_zpool_destroy(void *pool) | |
349 | { | |
350 | zs_destroy_pool(pool); | |
351 | } | |
352 | ||
353 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
354 | unsigned long *handle) | |
355 | { | |
d0d8da2d | 356 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
357 | return *handle ? 0 : -1; |
358 | } | |
359 | static void zs_zpool_free(void *pool, unsigned long handle) | |
360 | { | |
361 | zs_free(pool, handle); | |
362 | } | |
363 | ||
364 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
365 | unsigned int *reclaimed) | |
366 | { | |
367 | return -EINVAL; | |
368 | } | |
369 | ||
370 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
371 | enum zpool_mapmode mm) | |
372 | { | |
373 | enum zs_mapmode zs_mm; | |
374 | ||
375 | switch (mm) { | |
376 | case ZPOOL_MM_RO: | |
377 | zs_mm = ZS_MM_RO; | |
378 | break; | |
379 | case ZPOOL_MM_WO: | |
380 | zs_mm = ZS_MM_WO; | |
381 | break; | |
382 | case ZPOOL_MM_RW: /* fallthru */ | |
383 | default: | |
384 | zs_mm = ZS_MM_RW; | |
385 | break; | |
386 | } | |
387 | ||
388 | return zs_map_object(pool, handle, zs_mm); | |
389 | } | |
390 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
391 | { | |
392 | zs_unmap_object(pool, handle); | |
393 | } | |
394 | ||
395 | static u64 zs_zpool_total_size(void *pool) | |
396 | { | |
722cdc17 | 397 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
398 | } |
399 | ||
400 | static struct zpool_driver zs_zpool_driver = { | |
401 | .type = "zsmalloc", | |
402 | .owner = THIS_MODULE, | |
403 | .create = zs_zpool_create, | |
404 | .destroy = zs_zpool_destroy, | |
405 | .malloc = zs_zpool_malloc, | |
406 | .free = zs_zpool_free, | |
407 | .shrink = zs_zpool_shrink, | |
408 | .map = zs_zpool_map, | |
409 | .unmap = zs_zpool_unmap, | |
410 | .total_size = zs_zpool_total_size, | |
411 | }; | |
412 | ||
137f8cff | 413 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
414 | #endif /* CONFIG_ZPOOL */ |
415 | ||
248ca1b0 MK |
416 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
417 | { | |
418 | return pages_per_zspage * PAGE_SIZE / size; | |
419 | } | |
420 | ||
61989a80 NG |
421 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
422 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
423 | ||
424 | static int is_first_page(struct page *page) | |
425 | { | |
a27545bf | 426 | return PagePrivate(page); |
61989a80 NG |
427 | } |
428 | ||
3783689a | 429 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 430 | { |
3783689a | 431 | return zspage->inuse; |
4f42047b MK |
432 | } |
433 | ||
3783689a | 434 | static inline void set_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 435 | { |
3783689a | 436 | zspage->inuse = val; |
4f42047b MK |
437 | } |
438 | ||
3783689a | 439 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 440 | { |
3783689a | 441 | zspage->inuse += val; |
4f42047b MK |
442 | } |
443 | ||
444 | static inline int get_first_obj_offset(struct page *page) | |
445 | { | |
3783689a MK |
446 | if (is_first_page(page)) |
447 | return 0; | |
448 | ||
4f42047b MK |
449 | return page->index; |
450 | } | |
451 | ||
452 | static inline void set_first_obj_offset(struct page *page, int offset) | |
453 | { | |
3783689a MK |
454 | if (is_first_page(page)) |
455 | return; | |
456 | ||
4f42047b MK |
457 | page->index = offset; |
458 | } | |
459 | ||
bfd093f5 | 460 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 461 | { |
bfd093f5 | 462 | return zspage->freeobj; |
4f42047b MK |
463 | } |
464 | ||
bfd093f5 | 465 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 466 | { |
bfd093f5 | 467 | zspage->freeobj = obj; |
4f42047b MK |
468 | } |
469 | ||
3783689a | 470 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 471 | unsigned int *class_idx, |
61989a80 NG |
472 | enum fullness_group *fullness) |
473 | { | |
3783689a MK |
474 | *fullness = zspage->fullness; |
475 | *class_idx = zspage->class; | |
61989a80 NG |
476 | } |
477 | ||
3783689a | 478 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 479 | unsigned int class_idx, |
61989a80 NG |
480 | enum fullness_group fullness) |
481 | { | |
3783689a MK |
482 | zspage->class = class_idx; |
483 | zspage->fullness = fullness; | |
61989a80 NG |
484 | } |
485 | ||
c3e3e88a NC |
486 | /* |
487 | * zsmalloc divides the pool into various size classes where each | |
488 | * class maintains a list of zspages where each zspage is divided | |
489 | * into equal sized chunks. Each allocation falls into one of these | |
490 | * classes depending on its size. This function returns index of the | |
491 | * size class which has chunk size big enough to hold the give size. | |
492 | */ | |
61989a80 NG |
493 | static int get_size_class_index(int size) |
494 | { | |
495 | int idx = 0; | |
496 | ||
497 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
498 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
499 | ZS_SIZE_CLASS_DELTA); | |
500 | ||
7b60a685 | 501 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
502 | } |
503 | ||
248ca1b0 MK |
504 | static inline void zs_stat_inc(struct size_class *class, |
505 | enum zs_stat_type type, unsigned long cnt) | |
506 | { | |
6fe5186f SS |
507 | if (type < NR_ZS_STAT_TYPE) |
508 | class->stats.objs[type] += cnt; | |
248ca1b0 MK |
509 | } |
510 | ||
511 | static inline void zs_stat_dec(struct size_class *class, | |
512 | enum zs_stat_type type, unsigned long cnt) | |
513 | { | |
6fe5186f SS |
514 | if (type < NR_ZS_STAT_TYPE) |
515 | class->stats.objs[type] -= cnt; | |
248ca1b0 MK |
516 | } |
517 | ||
518 | static inline unsigned long zs_stat_get(struct size_class *class, | |
519 | enum zs_stat_type type) | |
520 | { | |
6fe5186f SS |
521 | if (type < NR_ZS_STAT_TYPE) |
522 | return class->stats.objs[type]; | |
523 | return 0; | |
248ca1b0 MK |
524 | } |
525 | ||
57244594 SS |
526 | #ifdef CONFIG_ZSMALLOC_STAT |
527 | ||
4abaac9b | 528 | static void __init zs_stat_init(void) |
248ca1b0 | 529 | { |
4abaac9b DS |
530 | if (!debugfs_initialized()) { |
531 | pr_warn("debugfs not available, stat dir not created\n"); | |
532 | return; | |
533 | } | |
248ca1b0 MK |
534 | |
535 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
536 | if (!zs_stat_root) | |
4abaac9b | 537 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
538 | } |
539 | ||
540 | static void __exit zs_stat_exit(void) | |
541 | { | |
542 | debugfs_remove_recursive(zs_stat_root); | |
543 | } | |
544 | ||
1120ed54 SS |
545 | static unsigned long zs_can_compact(struct size_class *class); |
546 | ||
248ca1b0 MK |
547 | static int zs_stats_size_show(struct seq_file *s, void *v) |
548 | { | |
549 | int i; | |
550 | struct zs_pool *pool = s->private; | |
551 | struct size_class *class; | |
552 | int objs_per_zspage; | |
553 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 554 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
555 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
556 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 557 | unsigned long total_freeable = 0; |
248ca1b0 | 558 | |
1120ed54 | 559 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
560 | "class", "size", "almost_full", "almost_empty", |
561 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 562 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
563 | |
564 | for (i = 0; i < zs_size_classes; i++) { | |
565 | class = pool->size_class[i]; | |
566 | ||
567 | if (class->index != i) | |
568 | continue; | |
569 | ||
570 | spin_lock(&class->lock); | |
571 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
572 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
573 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
574 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 575 | freeable = zs_can_compact(class); |
248ca1b0 MK |
576 | spin_unlock(&class->lock); |
577 | ||
578 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
579 | class->pages_per_zspage); | |
580 | pages_used = obj_allocated / objs_per_zspage * | |
581 | class->pages_per_zspage; | |
582 | ||
1120ed54 SS |
583 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
584 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
585 | i, class->size, class_almost_full, class_almost_empty, |
586 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 587 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
588 | |
589 | total_class_almost_full += class_almost_full; | |
590 | total_class_almost_empty += class_almost_empty; | |
591 | total_objs += obj_allocated; | |
592 | total_used_objs += obj_used; | |
593 | total_pages += pages_used; | |
1120ed54 | 594 | total_freeable += freeable; |
248ca1b0 MK |
595 | } |
596 | ||
597 | seq_puts(s, "\n"); | |
1120ed54 | 598 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
599 | "Total", "", total_class_almost_full, |
600 | total_class_almost_empty, total_objs, | |
1120ed54 | 601 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
602 | |
603 | return 0; | |
604 | } | |
605 | ||
606 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
607 | { | |
608 | return single_open(file, zs_stats_size_show, inode->i_private); | |
609 | } | |
610 | ||
611 | static const struct file_operations zs_stat_size_ops = { | |
612 | .open = zs_stats_size_open, | |
613 | .read = seq_read, | |
614 | .llseek = seq_lseek, | |
615 | .release = single_release, | |
616 | }; | |
617 | ||
d34f6157 | 618 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
619 | { |
620 | struct dentry *entry; | |
621 | ||
4abaac9b DS |
622 | if (!zs_stat_root) { |
623 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 624 | return; |
4abaac9b | 625 | } |
248ca1b0 MK |
626 | |
627 | entry = debugfs_create_dir(name, zs_stat_root); | |
628 | if (!entry) { | |
629 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 630 | return; |
248ca1b0 MK |
631 | } |
632 | pool->stat_dentry = entry; | |
633 | ||
634 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
635 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
636 | if (!entry) { | |
637 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
638 | name, "classes"); | |
4abaac9b DS |
639 | debugfs_remove_recursive(pool->stat_dentry); |
640 | pool->stat_dentry = NULL; | |
248ca1b0 | 641 | } |
248ca1b0 MK |
642 | } |
643 | ||
644 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
645 | { | |
646 | debugfs_remove_recursive(pool->stat_dentry); | |
647 | } | |
648 | ||
649 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 650 | static void __init zs_stat_init(void) |
248ca1b0 | 651 | { |
248ca1b0 MK |
652 | } |
653 | ||
654 | static void __exit zs_stat_exit(void) | |
655 | { | |
656 | } | |
657 | ||
d34f6157 | 658 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 659 | { |
248ca1b0 MK |
660 | } |
661 | ||
662 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
663 | { | |
664 | } | |
248ca1b0 MK |
665 | #endif |
666 | ||
c3e3e88a NC |
667 | /* |
668 | * For each size class, zspages are divided into different groups | |
669 | * depending on how "full" they are. This was done so that we could | |
670 | * easily find empty or nearly empty zspages when we try to shrink | |
671 | * the pool (not yet implemented). This function returns fullness | |
672 | * status of the given page. | |
673 | */ | |
1fc6e27d | 674 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 675 | struct zspage *zspage) |
61989a80 | 676 | { |
1fc6e27d | 677 | int inuse, objs_per_zspage; |
61989a80 | 678 | enum fullness_group fg; |
830e4bc5 | 679 | |
3783689a | 680 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 681 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
682 | |
683 | if (inuse == 0) | |
684 | fg = ZS_EMPTY; | |
1fc6e27d | 685 | else if (inuse == objs_per_zspage) |
61989a80 | 686 | fg = ZS_FULL; |
1fc6e27d | 687 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
688 | fg = ZS_ALMOST_EMPTY; |
689 | else | |
690 | fg = ZS_ALMOST_FULL; | |
691 | ||
692 | return fg; | |
693 | } | |
694 | ||
c3e3e88a NC |
695 | /* |
696 | * Each size class maintains various freelists and zspages are assigned | |
697 | * to one of these freelists based on the number of live objects they | |
698 | * have. This functions inserts the given zspage into the freelist | |
699 | * identified by <class, fullness_group>. | |
700 | */ | |
251cbb95 | 701 | static void insert_zspage(struct size_class *class, |
3783689a MK |
702 | struct zspage *zspage, |
703 | enum fullness_group fullness) | |
61989a80 | 704 | { |
3783689a | 705 | struct zspage *head; |
61989a80 | 706 | |
3783689a | 707 | if (fullness >= ZS_EMPTY) |
61989a80 NG |
708 | return; |
709 | ||
3783689a MK |
710 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
711 | struct zspage, list); | |
712 | ||
248ca1b0 MK |
713 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
714 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
58f17117 | 715 | |
58f17117 | 716 | /* |
3783689a MK |
717 | * We want to see more ZS_FULL pages and less almost empty/full. |
718 | * Put pages with higher ->inuse first. | |
58f17117 | 719 | */ |
3783689a MK |
720 | if (head) { |
721 | if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) { | |
722 | list_add(&zspage->list, &head->list); | |
723 | return; | |
724 | } | |
725 | } | |
726 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
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 | 733 | static void remove_zspage(struct size_class *class, |
3783689a MK |
734 | struct zspage *zspage, |
735 | enum fullness_group fullness) | |
61989a80 | 736 | { |
3783689a | 737 | if (fullness >= ZS_EMPTY) |
61989a80 NG |
738 | return; |
739 | ||
3783689a | 740 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
61989a80 | 741 | |
3783689a | 742 | list_del_init(&zspage->list); |
248ca1b0 MK |
743 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
744 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
745 | } |
746 | ||
c3e3e88a NC |
747 | /* |
748 | * Each size class maintains zspages in different fullness groups depending | |
749 | * on the number of live objects they contain. When allocating or freeing | |
750 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
751 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
752 | * a status change has occurred for the given page and accordingly moves the | |
753 | * page from the freelist of the old fullness group to that of the new | |
754 | * fullness group. | |
755 | */ | |
c7806261 | 756 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 757 | struct zspage *zspage) |
61989a80 NG |
758 | { |
759 | int class_idx; | |
61989a80 NG |
760 | enum fullness_group currfg, newfg; |
761 | ||
3783689a MK |
762 | get_zspage_mapping(zspage, &class_idx, &currfg); |
763 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
764 | if (newfg == currfg) |
765 | goto out; | |
766 | ||
3783689a MK |
767 | remove_zspage(class, zspage, currfg); |
768 | insert_zspage(class, zspage, newfg); | |
769 | set_zspage_mapping(zspage, class_idx, newfg); | |
61989a80 NG |
770 | |
771 | out: | |
772 | return newfg; | |
773 | } | |
774 | ||
775 | /* | |
776 | * We have to decide on how many pages to link together | |
777 | * to form a zspage for each size class. This is important | |
778 | * to reduce wastage due to unusable space left at end of | |
779 | * each zspage which is given as: | |
888fa374 YX |
780 | * wastage = Zp % class_size |
781 | * usage = Zp - wastage | |
61989a80 NG |
782 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
783 | * | |
784 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
785 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
786 | * since then we can perfectly fit in 8 such objects. | |
787 | */ | |
2e3b6154 | 788 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
789 | { |
790 | int i, max_usedpc = 0; | |
791 | /* zspage order which gives maximum used size per KB */ | |
792 | int max_usedpc_order = 1; | |
793 | ||
84d4faab | 794 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
795 | int zspage_size; |
796 | int waste, usedpc; | |
797 | ||
798 | zspage_size = i * PAGE_SIZE; | |
799 | waste = zspage_size % class_size; | |
800 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
801 | ||
802 | if (usedpc > max_usedpc) { | |
803 | max_usedpc = usedpc; | |
804 | max_usedpc_order = i; | |
805 | } | |
806 | } | |
807 | ||
808 | return max_usedpc_order; | |
809 | } | |
810 | ||
bfd093f5 MK |
811 | static struct page *get_first_page(struct zspage *zspage) |
812 | { | |
813 | return zspage->first_page; | |
814 | } | |
3783689a MK |
815 | |
816 | static struct zspage *get_zspage(struct page *page) | |
61989a80 | 817 | { |
3783689a | 818 | return (struct zspage *)page->private; |
61989a80 NG |
819 | } |
820 | ||
821 | static struct page *get_next_page(struct page *page) | |
822 | { | |
3783689a | 823 | return page->next; |
61989a80 NG |
824 | } |
825 | ||
bfd093f5 MK |
826 | /** |
827 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
828 | * @page: page object resides in zspage | |
829 | * @obj_idx: object index | |
67296874 | 830 | */ |
bfd093f5 MK |
831 | static void obj_to_location(unsigned long obj, struct page **page, |
832 | unsigned int *obj_idx) | |
61989a80 | 833 | { |
bfd093f5 MK |
834 | obj >>= OBJ_TAG_BITS; |
835 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
836 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
837 | } | |
61989a80 | 838 | |
bfd093f5 MK |
839 | /** |
840 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
841 | * @page: page object resides in zspage | |
842 | * @obj_idx: object index | |
843 | */ | |
844 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
845 | { | |
846 | unsigned long obj; | |
61989a80 | 847 | |
312fcae2 | 848 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 849 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 850 | obj <<= OBJ_TAG_BITS; |
61989a80 | 851 | |
bfd093f5 | 852 | return obj; |
61989a80 NG |
853 | } |
854 | ||
2e40e163 MK |
855 | static unsigned long handle_to_obj(unsigned long handle) |
856 | { | |
857 | return *(unsigned long *)handle; | |
858 | } | |
859 | ||
7b60a685 MK |
860 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
861 | void *obj) | |
312fcae2 | 862 | { |
7b60a685 | 863 | if (class->huge) { |
830e4bc5 | 864 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3783689a | 865 | return page->index; |
7b60a685 MK |
866 | } else |
867 | return *(unsigned long *)obj; | |
312fcae2 MK |
868 | } |
869 | ||
312fcae2 MK |
870 | static inline int trypin_tag(unsigned long handle) |
871 | { | |
1b8320b6 | 872 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
873 | } |
874 | ||
875 | static void pin_tag(unsigned long handle) | |
876 | { | |
1b8320b6 | 877 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
878 | } |
879 | ||
880 | static void unpin_tag(unsigned long handle) | |
881 | { | |
1b8320b6 | 882 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
883 | } |
884 | ||
f4477e90 NG |
885 | static void reset_page(struct page *page) |
886 | { | |
887 | clear_bit(PG_private, &page->flags); | |
888 | clear_bit(PG_private_2, &page->flags); | |
889 | set_page_private(page, 0); | |
3783689a | 890 | page->index = 0; |
f4477e90 NG |
891 | } |
892 | ||
3783689a | 893 | static void free_zspage(struct zs_pool *pool, struct zspage *zspage) |
61989a80 | 894 | { |
3783689a | 895 | struct page *page, *next; |
61989a80 | 896 | |
3783689a | 897 | VM_BUG_ON(get_zspage_inuse(zspage)); |
61989a80 | 898 | |
3783689a MK |
899 | next = page = zspage->first_page; |
900 | do { | |
901 | next = page->next; | |
902 | reset_page(page); | |
903 | put_page(page); | |
904 | page = next; | |
905 | } while (page != NULL); | |
61989a80 | 906 | |
3783689a | 907 | cache_free_zspage(pool, zspage); |
61989a80 NG |
908 | } |
909 | ||
910 | /* Initialize a newly allocated zspage */ | |
3783689a | 911 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 912 | { |
bfd093f5 | 913 | unsigned int freeobj = 1; |
61989a80 | 914 | unsigned long off = 0; |
3783689a | 915 | struct page *page = zspage->first_page; |
830e4bc5 | 916 | |
61989a80 NG |
917 | while (page) { |
918 | struct page *next_page; | |
919 | struct link_free *link; | |
af4ee5e9 | 920 | void *vaddr; |
61989a80 | 921 | |
3783689a | 922 | set_first_obj_offset(page, off); |
61989a80 | 923 | |
af4ee5e9 MK |
924 | vaddr = kmap_atomic(page); |
925 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
926 | |
927 | while ((off += class->size) < PAGE_SIZE) { | |
bfd093f5 | 928 | link->next = freeobj++ << OBJ_ALLOCATED_TAG; |
5538c562 | 929 | link += class->size / sizeof(*link); |
61989a80 NG |
930 | } |
931 | ||
932 | /* | |
933 | * We now come to the last (full or partial) object on this | |
934 | * page, which must point to the first object on the next | |
935 | * page (if present) | |
936 | */ | |
937 | next_page = get_next_page(page); | |
bfd093f5 MK |
938 | if (next_page) { |
939 | link->next = freeobj++ << OBJ_ALLOCATED_TAG; | |
940 | } else { | |
941 | /* | |
942 | * Reset OBJ_ALLOCATED_TAG bit to last link to tell | |
943 | * whether it's allocated object or not. | |
944 | */ | |
945 | link->next = -1 << OBJ_ALLOCATED_TAG; | |
946 | } | |
af4ee5e9 | 947 | kunmap_atomic(vaddr); |
61989a80 | 948 | page = next_page; |
5538c562 | 949 | off %= PAGE_SIZE; |
61989a80 | 950 | } |
bdb0af7c | 951 | |
bfd093f5 | 952 | set_freeobj(zspage, 0); |
61989a80 NG |
953 | } |
954 | ||
3783689a MK |
955 | static void create_page_chain(struct zspage *zspage, struct page *pages[], |
956 | int nr_pages) | |
61989a80 | 957 | { |
bdb0af7c MK |
958 | int i; |
959 | struct page *page; | |
960 | struct page *prev_page = NULL; | |
61989a80 NG |
961 | |
962 | /* | |
963 | * Allocate individual pages and link them together as: | |
3783689a MK |
964 | * 1. all pages are linked together using page->next |
965 | * 2. each sub-page point to zspage using page->private | |
61989a80 | 966 | * |
3783689a MK |
967 | * we set PG_private to identify the first page (i.e. no other sub-page |
968 | * has this flag set) and PG_private_2 to identify the last page. | |
61989a80 | 969 | */ |
bdb0af7c MK |
970 | for (i = 0; i < nr_pages; i++) { |
971 | page = pages[i]; | |
3783689a | 972 | set_page_private(page, (unsigned long)zspage); |
bdb0af7c | 973 | if (i == 0) { |
3783689a | 974 | zspage->first_page = page; |
a27545bf | 975 | SetPagePrivate(page); |
3783689a MK |
976 | } else { |
977 | prev_page->next = page; | |
61989a80 | 978 | } |
3783689a | 979 | if (i == nr_pages - 1) { |
a27545bf | 980 | SetPagePrivate2(page); |
3783689a MK |
981 | page->next = NULL; |
982 | } | |
61989a80 NG |
983 | prev_page = page; |
984 | } | |
bdb0af7c | 985 | } |
61989a80 | 986 | |
bdb0af7c MK |
987 | /* |
988 | * Allocate a zspage for the given size class | |
989 | */ | |
3783689a MK |
990 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
991 | struct size_class *class, | |
992 | gfp_t gfp) | |
bdb0af7c MK |
993 | { |
994 | int i; | |
bdb0af7c | 995 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
996 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
997 | ||
998 | if (!zspage) | |
999 | return NULL; | |
1000 | ||
1001 | memset(zspage, 0, sizeof(struct zspage)); | |
61989a80 | 1002 | |
bdb0af7c MK |
1003 | for (i = 0; i < class->pages_per_zspage; i++) { |
1004 | struct page *page; | |
61989a80 | 1005 | |
3783689a | 1006 | page = alloc_page(gfp); |
bdb0af7c MK |
1007 | if (!page) { |
1008 | while (--i >= 0) | |
1009 | __free_page(pages[i]); | |
3783689a | 1010 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1011 | return NULL; |
1012 | } | |
1013 | pages[i] = page; | |
61989a80 NG |
1014 | } |
1015 | ||
3783689a MK |
1016 | create_page_chain(zspage, pages, class->pages_per_zspage); |
1017 | init_zspage(class, zspage); | |
bdb0af7c | 1018 | |
3783689a | 1019 | return zspage; |
61989a80 NG |
1020 | } |
1021 | ||
3783689a | 1022 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1023 | { |
1024 | int i; | |
3783689a | 1025 | struct zspage *zspage; |
61989a80 | 1026 | |
3783689a MK |
1027 | for (i = ZS_ALMOST_FULL; i <= ZS_ALMOST_EMPTY; i++) { |
1028 | zspage = list_first_entry_or_null(&class->fullness_list[i], | |
1029 | struct zspage, list); | |
1030 | if (zspage) | |
61989a80 NG |
1031 | break; |
1032 | } | |
1033 | ||
3783689a | 1034 | return zspage; |
61989a80 NG |
1035 | } |
1036 | ||
1b945aee | 1037 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1038 | static inline int __zs_cpu_up(struct mapping_area *area) |
1039 | { | |
1040 | /* | |
1041 | * Make sure we don't leak memory if a cpu UP notification | |
1042 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1043 | */ | |
1044 | if (area->vm) | |
1045 | return 0; | |
1046 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1047 | if (!area->vm) | |
1048 | return -ENOMEM; | |
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1053 | { | |
1054 | if (area->vm) | |
1055 | free_vm_area(area->vm); | |
1056 | area->vm = NULL; | |
1057 | } | |
1058 | ||
1059 | static inline void *__zs_map_object(struct mapping_area *area, | |
1060 | struct page *pages[2], int off, int size) | |
1061 | { | |
f6f8ed47 | 1062 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1063 | area->vm_addr = area->vm->addr; |
1064 | return area->vm_addr + off; | |
1065 | } | |
1066 | ||
1067 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1068 | struct page *pages[2], int off, int size) | |
1069 | { | |
1070 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1071 | |
d95abbbb | 1072 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1073 | } |
1074 | ||
1b945aee | 1075 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1076 | |
1077 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1078 | { | |
1079 | /* | |
1080 | * Make sure we don't leak memory if a cpu UP notification | |
1081 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1082 | */ | |
1083 | if (area->vm_buf) | |
1084 | return 0; | |
40f9fb8c | 1085 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1086 | if (!area->vm_buf) |
1087 | return -ENOMEM; | |
1088 | return 0; | |
1089 | } | |
1090 | ||
1091 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1092 | { | |
40f9fb8c | 1093 | kfree(area->vm_buf); |
f553646a SJ |
1094 | area->vm_buf = NULL; |
1095 | } | |
1096 | ||
1097 | static void *__zs_map_object(struct mapping_area *area, | |
1098 | struct page *pages[2], int off, int size) | |
5f601902 | 1099 | { |
5f601902 SJ |
1100 | int sizes[2]; |
1101 | void *addr; | |
f553646a | 1102 | char *buf = area->vm_buf; |
5f601902 | 1103 | |
f553646a SJ |
1104 | /* disable page faults to match kmap_atomic() return conditions */ |
1105 | pagefault_disable(); | |
1106 | ||
1107 | /* no read fastpath */ | |
1108 | if (area->vm_mm == ZS_MM_WO) | |
1109 | goto out; | |
5f601902 SJ |
1110 | |
1111 | sizes[0] = PAGE_SIZE - off; | |
1112 | sizes[1] = size - sizes[0]; | |
1113 | ||
5f601902 SJ |
1114 | /* copy object to per-cpu buffer */ |
1115 | addr = kmap_atomic(pages[0]); | |
1116 | memcpy(buf, addr + off, sizes[0]); | |
1117 | kunmap_atomic(addr); | |
1118 | addr = kmap_atomic(pages[1]); | |
1119 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1120 | kunmap_atomic(addr); | |
f553646a SJ |
1121 | out: |
1122 | return area->vm_buf; | |
5f601902 SJ |
1123 | } |
1124 | ||
f553646a SJ |
1125 | static void __zs_unmap_object(struct mapping_area *area, |
1126 | struct page *pages[2], int off, int size) | |
5f601902 | 1127 | { |
5f601902 SJ |
1128 | int sizes[2]; |
1129 | void *addr; | |
2e40e163 | 1130 | char *buf; |
5f601902 | 1131 | |
f553646a SJ |
1132 | /* no write fastpath */ |
1133 | if (area->vm_mm == ZS_MM_RO) | |
1134 | goto out; | |
5f601902 | 1135 | |
7b60a685 | 1136 | buf = area->vm_buf; |
a82cbf07 YX |
1137 | buf = buf + ZS_HANDLE_SIZE; |
1138 | size -= ZS_HANDLE_SIZE; | |
1139 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1140 | |
5f601902 SJ |
1141 | sizes[0] = PAGE_SIZE - off; |
1142 | sizes[1] = size - sizes[0]; | |
1143 | ||
1144 | /* copy per-cpu buffer to object */ | |
1145 | addr = kmap_atomic(pages[0]); | |
1146 | memcpy(addr + off, buf, sizes[0]); | |
1147 | kunmap_atomic(addr); | |
1148 | addr = kmap_atomic(pages[1]); | |
1149 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1150 | kunmap_atomic(addr); | |
f553646a SJ |
1151 | |
1152 | out: | |
1153 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1154 | pagefault_enable(); | |
5f601902 | 1155 | } |
61989a80 | 1156 | |
1b945aee | 1157 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1158 | |
61989a80 NG |
1159 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1160 | void *pcpu) | |
1161 | { | |
f553646a | 1162 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1163 | struct mapping_area *area; |
1164 | ||
1165 | switch (action) { | |
1166 | case CPU_UP_PREPARE: | |
1167 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1168 | ret = __zs_cpu_up(area); |
1169 | if (ret) | |
1170 | return notifier_from_errno(ret); | |
61989a80 NG |
1171 | break; |
1172 | case CPU_DEAD: | |
1173 | case CPU_UP_CANCELED: | |
1174 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1175 | __zs_cpu_down(area); |
61989a80 NG |
1176 | break; |
1177 | } | |
1178 | ||
1179 | return NOTIFY_OK; | |
1180 | } | |
1181 | ||
1182 | static struct notifier_block zs_cpu_nb = { | |
1183 | .notifier_call = zs_cpu_notifier | |
1184 | }; | |
1185 | ||
b1b00a5b | 1186 | static int zs_register_cpu_notifier(void) |
61989a80 | 1187 | { |
b1b00a5b | 1188 | int cpu, uninitialized_var(ret); |
61989a80 | 1189 | |
f0e71fcd SB |
1190 | cpu_notifier_register_begin(); |
1191 | ||
1192 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1193 | for_each_online_cpu(cpu) { |
1194 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1195 | if (notifier_to_errno(ret)) |
1196 | break; | |
61989a80 | 1197 | } |
f0e71fcd SB |
1198 | |
1199 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1200 | return notifier_to_errno(ret); |
1201 | } | |
f0e71fcd | 1202 | |
66cdef66 | 1203 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1204 | { |
66cdef66 | 1205 | int cpu; |
40f9fb8c | 1206 | |
66cdef66 | 1207 | cpu_notifier_register_begin(); |
40f9fb8c | 1208 | |
66cdef66 GM |
1209 | for_each_online_cpu(cpu) |
1210 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1211 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1212 | |
66cdef66 | 1213 | cpu_notifier_register_done(); |
b1b00a5b SS |
1214 | } |
1215 | ||
66cdef66 | 1216 | static void init_zs_size_classes(void) |
b1b00a5b | 1217 | { |
66cdef66 | 1218 | int nr; |
c795779d | 1219 | |
66cdef66 GM |
1220 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1221 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1222 | nr += 1; | |
40f9fb8c | 1223 | |
66cdef66 | 1224 | zs_size_classes = nr; |
61989a80 NG |
1225 | } |
1226 | ||
9eec4cd5 JK |
1227 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1228 | { | |
1229 | if (prev->pages_per_zspage != pages_per_zspage) | |
1230 | return false; | |
1231 | ||
1232 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1233 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1234 | return false; | |
1235 | ||
1236 | return true; | |
1237 | } | |
1238 | ||
3783689a | 1239 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1240 | { |
3783689a | 1241 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1242 | } |
1243 | ||
66cdef66 GM |
1244 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1245 | { | |
1246 | return atomic_long_read(&pool->pages_allocated); | |
1247 | } | |
1248 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1249 | ||
4bbc0bc0 | 1250 | /** |
66cdef66 GM |
1251 | * zs_map_object - get address of allocated object from handle. |
1252 | * @pool: pool from which the object was allocated | |
1253 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1254 | * |
66cdef66 GM |
1255 | * Before using an object allocated from zs_malloc, it must be mapped using |
1256 | * this function. When done with the object, it must be unmapped using | |
1257 | * zs_unmap_object. | |
4bbc0bc0 | 1258 | * |
66cdef66 GM |
1259 | * Only one object can be mapped per cpu at a time. There is no protection |
1260 | * against nested mappings. | |
1261 | * | |
1262 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1263 | */ |
66cdef66 GM |
1264 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1265 | enum zs_mapmode mm) | |
61989a80 | 1266 | { |
3783689a | 1267 | struct zspage *zspage; |
66cdef66 | 1268 | struct page *page; |
bfd093f5 MK |
1269 | unsigned long obj, off; |
1270 | unsigned int obj_idx; | |
61989a80 | 1271 | |
66cdef66 GM |
1272 | unsigned int class_idx; |
1273 | enum fullness_group fg; | |
1274 | struct size_class *class; | |
1275 | struct mapping_area *area; | |
1276 | struct page *pages[2]; | |
2e40e163 | 1277 | void *ret; |
61989a80 | 1278 | |
9eec4cd5 | 1279 | /* |
66cdef66 GM |
1280 | * Because we use per-cpu mapping areas shared among the |
1281 | * pools/users, we can't allow mapping in interrupt context | |
1282 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1283 | */ |
830e4bc5 | 1284 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1285 | |
312fcae2 MK |
1286 | /* From now on, migration cannot move the object */ |
1287 | pin_tag(handle); | |
1288 | ||
2e40e163 MK |
1289 | obj = handle_to_obj(handle); |
1290 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1291 | zspage = get_zspage(page); |
1292 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 | 1293 | class = pool->size_class[class_idx]; |
bfd093f5 | 1294 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1295 | |
66cdef66 GM |
1296 | area = &get_cpu_var(zs_map_area); |
1297 | area->vm_mm = mm; | |
1298 | if (off + class->size <= PAGE_SIZE) { | |
1299 | /* this object is contained entirely within a page */ | |
1300 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1301 | ret = area->vm_addr + off; |
1302 | goto out; | |
61989a80 NG |
1303 | } |
1304 | ||
66cdef66 GM |
1305 | /* this object spans two pages */ |
1306 | pages[0] = page; | |
1307 | pages[1] = get_next_page(page); | |
1308 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1309 | |
2e40e163 MK |
1310 | ret = __zs_map_object(area, pages, off, class->size); |
1311 | out: | |
7b60a685 MK |
1312 | if (!class->huge) |
1313 | ret += ZS_HANDLE_SIZE; | |
1314 | ||
1315 | return ret; | |
61989a80 | 1316 | } |
66cdef66 | 1317 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1318 | |
66cdef66 | 1319 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1320 | { |
3783689a | 1321 | struct zspage *zspage; |
66cdef66 | 1322 | struct page *page; |
bfd093f5 MK |
1323 | unsigned long obj, off; |
1324 | unsigned int obj_idx; | |
61989a80 | 1325 | |
66cdef66 GM |
1326 | unsigned int class_idx; |
1327 | enum fullness_group fg; | |
1328 | struct size_class *class; | |
1329 | struct mapping_area *area; | |
9eec4cd5 | 1330 | |
2e40e163 MK |
1331 | obj = handle_to_obj(handle); |
1332 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1333 | zspage = get_zspage(page); |
1334 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 | 1335 | class = pool->size_class[class_idx]; |
bfd093f5 | 1336 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1337 | |
66cdef66 GM |
1338 | area = this_cpu_ptr(&zs_map_area); |
1339 | if (off + class->size <= PAGE_SIZE) | |
1340 | kunmap_atomic(area->vm_addr); | |
1341 | else { | |
1342 | struct page *pages[2]; | |
40f9fb8c | 1343 | |
66cdef66 GM |
1344 | pages[0] = page; |
1345 | pages[1] = get_next_page(page); | |
1346 | BUG_ON(!pages[1]); | |
1347 | ||
1348 | __zs_unmap_object(area, pages, off, class->size); | |
1349 | } | |
1350 | put_cpu_var(zs_map_area); | |
312fcae2 | 1351 | unpin_tag(handle); |
61989a80 | 1352 | } |
66cdef66 | 1353 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1354 | |
251cbb95 | 1355 | static unsigned long obj_malloc(struct size_class *class, |
3783689a | 1356 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1357 | { |
bfd093f5 | 1358 | int i, nr_page, offset; |
c7806261 MK |
1359 | unsigned long obj; |
1360 | struct link_free *link; | |
1361 | ||
1362 | struct page *m_page; | |
bfd093f5 | 1363 | unsigned long m_offset; |
c7806261 MK |
1364 | void *vaddr; |
1365 | ||
312fcae2 | 1366 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1367 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1368 | |
1369 | offset = obj * class->size; | |
1370 | nr_page = offset >> PAGE_SHIFT; | |
1371 | m_offset = offset & ~PAGE_MASK; | |
1372 | m_page = get_first_page(zspage); | |
1373 | ||
1374 | for (i = 0; i < nr_page; i++) | |
1375 | m_page = get_next_page(m_page); | |
c7806261 MK |
1376 | |
1377 | vaddr = kmap_atomic(m_page); | |
1378 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
bfd093f5 | 1379 | set_freeobj(zspage, link->next >> OBJ_ALLOCATED_TAG); |
7b60a685 MK |
1380 | if (!class->huge) |
1381 | /* record handle in the header of allocated chunk */ | |
1382 | link->handle = handle; | |
1383 | else | |
3783689a MK |
1384 | /* record handle to page->index */ |
1385 | zspage->first_page->index = handle; | |
1386 | ||
c7806261 | 1387 | kunmap_atomic(vaddr); |
3783689a | 1388 | mod_zspage_inuse(zspage, 1); |
c7806261 MK |
1389 | zs_stat_inc(class, OBJ_USED, 1); |
1390 | ||
bfd093f5 MK |
1391 | obj = location_to_obj(m_page, obj); |
1392 | ||
c7806261 MK |
1393 | return obj; |
1394 | } | |
1395 | ||
1396 | ||
61989a80 NG |
1397 | /** |
1398 | * zs_malloc - Allocate block of given size from pool. | |
1399 | * @pool: pool to allocate from | |
1400 | * @size: size of block to allocate | |
61989a80 | 1401 | * |
00a61d86 | 1402 | * On success, handle to the allocated object is returned, |
c2344348 | 1403 | * otherwise 0. |
61989a80 NG |
1404 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1405 | */ | |
d0d8da2d | 1406 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1407 | { |
2e40e163 | 1408 | unsigned long handle, obj; |
61989a80 | 1409 | struct size_class *class; |
3783689a | 1410 | struct zspage *zspage; |
61989a80 | 1411 | |
7b60a685 | 1412 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1413 | return 0; |
1414 | ||
3783689a | 1415 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1416 | if (!handle) |
c2344348 | 1417 | return 0; |
61989a80 | 1418 | |
2e40e163 MK |
1419 | /* extra space in chunk to keep the handle */ |
1420 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1421 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1422 | |
1423 | spin_lock(&class->lock); | |
3783689a | 1424 | zspage = find_get_zspage(class); |
61989a80 | 1425 | |
3783689a | 1426 | if (!zspage) { |
61989a80 | 1427 | spin_unlock(&class->lock); |
3783689a MK |
1428 | zspage = alloc_zspage(pool, class, gfp); |
1429 | if (unlikely(!zspage)) { | |
1430 | cache_free_handle(pool, handle); | |
c2344348 | 1431 | return 0; |
2e40e163 | 1432 | } |
61989a80 | 1433 | |
3783689a | 1434 | set_zspage_mapping(zspage, class->index, ZS_EMPTY); |
13de8933 MK |
1435 | atomic_long_add(class->pages_per_zspage, |
1436 | &pool->pages_allocated); | |
0f050d99 | 1437 | |
61989a80 | 1438 | spin_lock(&class->lock); |
0f050d99 GM |
1439 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1440 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1441 | } |
1442 | ||
3783689a | 1443 | obj = obj_malloc(class, zspage, handle); |
61989a80 | 1444 | /* Now move the zspage to another fullness group, if required */ |
3783689a | 1445 | fix_fullness_group(class, zspage); |
2e40e163 | 1446 | record_obj(handle, obj); |
61989a80 NG |
1447 | spin_unlock(&class->lock); |
1448 | ||
2e40e163 | 1449 | return handle; |
61989a80 NG |
1450 | } |
1451 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1452 | ||
1ee47165 | 1453 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1454 | { |
1455 | struct link_free *link; | |
3783689a MK |
1456 | struct zspage *zspage; |
1457 | struct page *f_page; | |
bfd093f5 MK |
1458 | unsigned long f_offset; |
1459 | unsigned int f_objidx; | |
af4ee5e9 | 1460 | void *vaddr; |
61989a80 | 1461 | |
312fcae2 | 1462 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1463 | obj_to_location(obj, &f_page, &f_objidx); |
bfd093f5 | 1464 | f_offset = (class->size * f_objidx) & ~PAGE_MASK; |
3783689a | 1465 | zspage = get_zspage(f_page); |
61989a80 | 1466 | |
c7806261 | 1467 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1468 | |
1469 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1470 | link = (struct link_free *)(vaddr + f_offset); |
bfd093f5 | 1471 | link->next = get_freeobj(zspage) << OBJ_ALLOCATED_TAG; |
af4ee5e9 | 1472 | kunmap_atomic(vaddr); |
bfd093f5 | 1473 | set_freeobj(zspage, f_objidx); |
3783689a | 1474 | mod_zspage_inuse(zspage, -1); |
0f050d99 | 1475 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1476 | } |
1477 | ||
1478 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1479 | { | |
3783689a MK |
1480 | struct zspage *zspage; |
1481 | struct page *f_page; | |
bfd093f5 MK |
1482 | unsigned long obj; |
1483 | unsigned int f_objidx; | |
c7806261 MK |
1484 | int class_idx; |
1485 | struct size_class *class; | |
1486 | enum fullness_group fullness; | |
1487 | ||
1488 | if (unlikely(!handle)) | |
1489 | return; | |
1490 | ||
312fcae2 | 1491 | pin_tag(handle); |
c7806261 | 1492 | obj = handle_to_obj(handle); |
c7806261 | 1493 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1494 | zspage = get_zspage(f_page); |
c7806261 | 1495 | |
3783689a | 1496 | get_zspage_mapping(zspage, &class_idx, &fullness); |
c7806261 MK |
1497 | class = pool->size_class[class_idx]; |
1498 | ||
1499 | spin_lock(&class->lock); | |
1ee47165 | 1500 | obj_free(class, obj); |
3783689a | 1501 | fullness = fix_fullness_group(class, zspage); |
312fcae2 | 1502 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1503 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1504 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1505 | atomic_long_sub(class->pages_per_zspage, |
1506 | &pool->pages_allocated); | |
3783689a | 1507 | free_zspage(pool, zspage); |
312fcae2 | 1508 | } |
61989a80 | 1509 | spin_unlock(&class->lock); |
312fcae2 | 1510 | unpin_tag(handle); |
61989a80 | 1511 | |
3783689a | 1512 | cache_free_handle(pool, handle); |
312fcae2 MK |
1513 | } |
1514 | EXPORT_SYMBOL_GPL(zs_free); | |
1515 | ||
251cbb95 MK |
1516 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1517 | unsigned long src) | |
312fcae2 MK |
1518 | { |
1519 | struct page *s_page, *d_page; | |
bfd093f5 | 1520 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1521 | unsigned long s_off, d_off; |
1522 | void *s_addr, *d_addr; | |
1523 | int s_size, d_size, size; | |
1524 | int written = 0; | |
1525 | ||
1526 | s_size = d_size = class->size; | |
1527 | ||
1528 | obj_to_location(src, &s_page, &s_objidx); | |
1529 | obj_to_location(dst, &d_page, &d_objidx); | |
1530 | ||
bfd093f5 MK |
1531 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1532 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1533 | |
1534 | if (s_off + class->size > PAGE_SIZE) | |
1535 | s_size = PAGE_SIZE - s_off; | |
1536 | ||
1537 | if (d_off + class->size > PAGE_SIZE) | |
1538 | d_size = PAGE_SIZE - d_off; | |
1539 | ||
1540 | s_addr = kmap_atomic(s_page); | |
1541 | d_addr = kmap_atomic(d_page); | |
1542 | ||
1543 | while (1) { | |
1544 | size = min(s_size, d_size); | |
1545 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1546 | written += size; | |
1547 | ||
1548 | if (written == class->size) | |
1549 | break; | |
1550 | ||
495819ea SS |
1551 | s_off += size; |
1552 | s_size -= size; | |
1553 | d_off += size; | |
1554 | d_size -= size; | |
1555 | ||
1556 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1557 | kunmap_atomic(d_addr); |
1558 | kunmap_atomic(s_addr); | |
1559 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1560 | s_addr = kmap_atomic(s_page); |
1561 | d_addr = kmap_atomic(d_page); | |
1562 | s_size = class->size - written; | |
1563 | s_off = 0; | |
312fcae2 MK |
1564 | } |
1565 | ||
495819ea | 1566 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1567 | kunmap_atomic(d_addr); |
1568 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1569 | d_addr = kmap_atomic(d_page); |
1570 | d_size = class->size - written; | |
1571 | d_off = 0; | |
312fcae2 MK |
1572 | } |
1573 | } | |
1574 | ||
1575 | kunmap_atomic(d_addr); | |
1576 | kunmap_atomic(s_addr); | |
1577 | } | |
1578 | ||
1579 | /* | |
1580 | * Find alloced object in zspage from index object and | |
1581 | * return handle. | |
1582 | */ | |
251cbb95 MK |
1583 | static unsigned long find_alloced_obj(struct size_class *class, |
1584 | struct page *page, int index) | |
312fcae2 MK |
1585 | { |
1586 | unsigned long head; | |
1587 | int offset = 0; | |
1588 | unsigned long handle = 0; | |
1589 | void *addr = kmap_atomic(page); | |
1590 | ||
3783689a | 1591 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1592 | offset += class->size * index; |
1593 | ||
1594 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1595 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1596 | if (head & OBJ_ALLOCATED_TAG) { |
1597 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1598 | if (trypin_tag(handle)) | |
1599 | break; | |
1600 | handle = 0; | |
1601 | } | |
1602 | ||
1603 | offset += class->size; | |
1604 | index++; | |
1605 | } | |
1606 | ||
1607 | kunmap_atomic(addr); | |
1608 | return handle; | |
1609 | } | |
1610 | ||
1611 | struct zs_compact_control { | |
3783689a | 1612 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1613 | struct page *s_page; |
1614 | /* Destination page for migration which should be a first page | |
1615 | * of zspage. */ | |
1616 | struct page *d_page; | |
1617 | /* Starting object index within @s_page which used for live object | |
1618 | * in the subpage. */ | |
1619 | int index; | |
312fcae2 MK |
1620 | }; |
1621 | ||
1622 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1623 | struct zs_compact_control *cc) | |
1624 | { | |
1625 | unsigned long used_obj, free_obj; | |
1626 | unsigned long handle; | |
1627 | struct page *s_page = cc->s_page; | |
1628 | struct page *d_page = cc->d_page; | |
1629 | unsigned long index = cc->index; | |
312fcae2 MK |
1630 | int ret = 0; |
1631 | ||
1632 | while (1) { | |
251cbb95 | 1633 | handle = find_alloced_obj(class, s_page, index); |
312fcae2 MK |
1634 | if (!handle) { |
1635 | s_page = get_next_page(s_page); | |
1636 | if (!s_page) | |
1637 | break; | |
1638 | index = 0; | |
1639 | continue; | |
1640 | } | |
1641 | ||
1642 | /* Stop if there is no more space */ | |
3783689a | 1643 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1644 | unpin_tag(handle); |
1645 | ret = -ENOMEM; | |
1646 | break; | |
1647 | } | |
1648 | ||
1649 | used_obj = handle_to_obj(handle); | |
3783689a | 1650 | free_obj = obj_malloc(class, get_zspage(d_page), handle); |
251cbb95 | 1651 | zs_object_copy(class, free_obj, used_obj); |
312fcae2 | 1652 | index++; |
c102f07c JL |
1653 | /* |
1654 | * record_obj updates handle's value to free_obj and it will | |
1655 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1656 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1657 | * let's keep the lock bit. | |
1658 | */ | |
1659 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1660 | record_obj(handle, free_obj); |
1661 | unpin_tag(handle); | |
1ee47165 | 1662 | obj_free(class, used_obj); |
312fcae2 MK |
1663 | } |
1664 | ||
1665 | /* Remember last position in this iteration */ | |
1666 | cc->s_page = s_page; | |
1667 | cc->index = index; | |
312fcae2 MK |
1668 | |
1669 | return ret; | |
1670 | } | |
1671 | ||
3783689a | 1672 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1673 | { |
1674 | int i; | |
3783689a MK |
1675 | struct zspage *zspage; |
1676 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1677 | |
3783689a MK |
1678 | if (!source) { |
1679 | fg[0] = ZS_ALMOST_FULL; | |
1680 | fg[1] = ZS_ALMOST_EMPTY; | |
1681 | } | |
1682 | ||
1683 | for (i = 0; i < 2; i++) { | |
1684 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1685 | struct zspage, list); | |
1686 | if (zspage) { | |
1687 | remove_zspage(class, zspage, fg[i]); | |
1688 | return zspage; | |
312fcae2 MK |
1689 | } |
1690 | } | |
1691 | ||
3783689a | 1692 | return zspage; |
312fcae2 MK |
1693 | } |
1694 | ||
860c707d | 1695 | /* |
3783689a | 1696 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1697 | * @class: destination class |
3783689a | 1698 | * @zspage: target page |
860c707d | 1699 | * |
3783689a | 1700 | * Return @zspage's fullness_group |
860c707d | 1701 | */ |
4aa409ca | 1702 | static enum fullness_group putback_zspage(struct size_class *class, |
3783689a | 1703 | struct zspage *zspage) |
312fcae2 | 1704 | { |
312fcae2 MK |
1705 | enum fullness_group fullness; |
1706 | ||
3783689a MK |
1707 | fullness = get_fullness_group(class, zspage); |
1708 | insert_zspage(class, zspage, fullness); | |
1709 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1710 | |
860c707d | 1711 | return fullness; |
61989a80 | 1712 | } |
312fcae2 | 1713 | |
04f05909 SS |
1714 | /* |
1715 | * | |
1716 | * Based on the number of unused allocated objects calculate | |
1717 | * and return the number of pages that we can free. | |
04f05909 SS |
1718 | */ |
1719 | static unsigned long zs_can_compact(struct size_class *class) | |
1720 | { | |
1721 | unsigned long obj_wasted; | |
44f43e99 SS |
1722 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
1723 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 1724 | |
44f43e99 SS |
1725 | if (obj_allocated <= obj_used) |
1726 | return 0; | |
04f05909 | 1727 | |
44f43e99 | 1728 | obj_wasted = obj_allocated - obj_used; |
04f05909 SS |
1729 | obj_wasted /= get_maxobj_per_zspage(class->size, |
1730 | class->pages_per_zspage); | |
1731 | ||
6cbf16b3 | 1732 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
1733 | } |
1734 | ||
7d3f3938 | 1735 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1736 | { |
312fcae2 | 1737 | struct zs_compact_control cc; |
3783689a MK |
1738 | struct zspage *src_zspage; |
1739 | struct zspage *dst_zspage = NULL; | |
312fcae2 | 1740 | |
312fcae2 | 1741 | spin_lock(&class->lock); |
3783689a | 1742 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 1743 | |
04f05909 SS |
1744 | if (!zs_can_compact(class)) |
1745 | break; | |
1746 | ||
312fcae2 | 1747 | cc.index = 0; |
3783689a | 1748 | cc.s_page = src_zspage->first_page; |
312fcae2 | 1749 | |
3783689a MK |
1750 | while ((dst_zspage = isolate_zspage(class, false))) { |
1751 | cc.d_page = dst_zspage->first_page; | |
312fcae2 | 1752 | /* |
0dc63d48 SS |
1753 | * If there is no more space in dst_page, resched |
1754 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1755 | */ |
1756 | if (!migrate_zspage(pool, class, &cc)) | |
1757 | break; | |
1758 | ||
4aa409ca | 1759 | putback_zspage(class, dst_zspage); |
312fcae2 MK |
1760 | } |
1761 | ||
1762 | /* Stop if we couldn't find slot */ | |
3783689a | 1763 | if (dst_zspage == NULL) |
312fcae2 MK |
1764 | break; |
1765 | ||
4aa409ca MK |
1766 | putback_zspage(class, dst_zspage); |
1767 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { | |
1768 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( | |
1769 | class->size, class->pages_per_zspage)); | |
1770 | atomic_long_sub(class->pages_per_zspage, | |
1771 | &pool->pages_allocated); | |
1772 | free_zspage(pool, src_zspage); | |
6cbf16b3 | 1773 | pool->stats.pages_compacted += class->pages_per_zspage; |
4aa409ca | 1774 | } |
312fcae2 | 1775 | spin_unlock(&class->lock); |
312fcae2 MK |
1776 | cond_resched(); |
1777 | spin_lock(&class->lock); | |
1778 | } | |
1779 | ||
3783689a | 1780 | if (src_zspage) |
4aa409ca | 1781 | putback_zspage(class, src_zspage); |
312fcae2 | 1782 | |
7d3f3938 | 1783 | spin_unlock(&class->lock); |
312fcae2 MK |
1784 | } |
1785 | ||
1786 | unsigned long zs_compact(struct zs_pool *pool) | |
1787 | { | |
1788 | int i; | |
312fcae2 MK |
1789 | struct size_class *class; |
1790 | ||
1791 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1792 | class = pool->size_class[i]; | |
1793 | if (!class) | |
1794 | continue; | |
1795 | if (class->index != i) | |
1796 | continue; | |
7d3f3938 | 1797 | __zs_compact(pool, class); |
312fcae2 MK |
1798 | } |
1799 | ||
860c707d | 1800 | return pool->stats.pages_compacted; |
312fcae2 MK |
1801 | } |
1802 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1803 | |
7d3f3938 SS |
1804 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1805 | { | |
1806 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1807 | } | |
1808 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1809 | ||
ab9d306d SS |
1810 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
1811 | struct shrink_control *sc) | |
1812 | { | |
1813 | unsigned long pages_freed; | |
1814 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1815 | shrinker); | |
1816 | ||
1817 | pages_freed = pool->stats.pages_compacted; | |
1818 | /* | |
1819 | * Compact classes and calculate compaction delta. | |
1820 | * Can run concurrently with a manually triggered | |
1821 | * (by user) compaction. | |
1822 | */ | |
1823 | pages_freed = zs_compact(pool) - pages_freed; | |
1824 | ||
1825 | return pages_freed ? pages_freed : SHRINK_STOP; | |
1826 | } | |
1827 | ||
1828 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
1829 | struct shrink_control *sc) | |
1830 | { | |
1831 | int i; | |
1832 | struct size_class *class; | |
1833 | unsigned long pages_to_free = 0; | |
1834 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1835 | shrinker); | |
1836 | ||
ab9d306d SS |
1837 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1838 | class = pool->size_class[i]; | |
1839 | if (!class) | |
1840 | continue; | |
1841 | if (class->index != i) | |
1842 | continue; | |
1843 | ||
ab9d306d | 1844 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
1845 | } |
1846 | ||
1847 | return pages_to_free; | |
1848 | } | |
1849 | ||
1850 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
1851 | { | |
1852 | if (pool->shrinker_enabled) { | |
1853 | unregister_shrinker(&pool->shrinker); | |
1854 | pool->shrinker_enabled = false; | |
1855 | } | |
1856 | } | |
1857 | ||
1858 | static int zs_register_shrinker(struct zs_pool *pool) | |
1859 | { | |
1860 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
1861 | pool->shrinker.count_objects = zs_shrinker_count; | |
1862 | pool->shrinker.batch = 0; | |
1863 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
1864 | ||
1865 | return register_shrinker(&pool->shrinker); | |
1866 | } | |
1867 | ||
00a61d86 | 1868 | /** |
66cdef66 GM |
1869 | * zs_create_pool - Creates an allocation pool to work from. |
1870 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1871 | * |
66cdef66 GM |
1872 | * This function must be called before anything when using |
1873 | * the zsmalloc allocator. | |
166cfda7 | 1874 | * |
66cdef66 GM |
1875 | * On success, a pointer to the newly created pool is returned, |
1876 | * otherwise NULL. | |
396b7fd6 | 1877 | */ |
d0d8da2d | 1878 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 1879 | { |
66cdef66 GM |
1880 | int i; |
1881 | struct zs_pool *pool; | |
1882 | struct size_class *prev_class = NULL; | |
61989a80 | 1883 | |
66cdef66 GM |
1884 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1885 | if (!pool) | |
1886 | return NULL; | |
61989a80 | 1887 | |
66cdef66 GM |
1888 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1889 | GFP_KERNEL); | |
1890 | if (!pool->size_class) { | |
1891 | kfree(pool); | |
1892 | return NULL; | |
1893 | } | |
61989a80 | 1894 | |
2e40e163 MK |
1895 | pool->name = kstrdup(name, GFP_KERNEL); |
1896 | if (!pool->name) | |
1897 | goto err; | |
1898 | ||
3783689a | 1899 | if (create_cache(pool)) |
2e40e163 MK |
1900 | goto err; |
1901 | ||
c60369f0 | 1902 | /* |
66cdef66 GM |
1903 | * Iterate reversly, because, size of size_class that we want to use |
1904 | * for merging should be larger or equal to current size. | |
c60369f0 | 1905 | */ |
66cdef66 GM |
1906 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1907 | int size; | |
1908 | int pages_per_zspage; | |
1909 | struct size_class *class; | |
3783689a | 1910 | int fullness = 0; |
c60369f0 | 1911 | |
66cdef66 GM |
1912 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1913 | if (size > ZS_MAX_ALLOC_SIZE) | |
1914 | size = ZS_MAX_ALLOC_SIZE; | |
1915 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1916 | |
66cdef66 GM |
1917 | /* |
1918 | * size_class is used for normal zsmalloc operation such | |
1919 | * as alloc/free for that size. Although it is natural that we | |
1920 | * have one size_class for each size, there is a chance that we | |
1921 | * can get more memory utilization if we use one size_class for | |
1922 | * many different sizes whose size_class have same | |
1923 | * characteristics. So, we makes size_class point to | |
1924 | * previous size_class if possible. | |
1925 | */ | |
1926 | if (prev_class) { | |
1927 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1928 | pool->size_class[i] = prev_class; | |
1929 | continue; | |
1930 | } | |
1931 | } | |
1932 | ||
1933 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1934 | if (!class) | |
1935 | goto err; | |
1936 | ||
1937 | class->size = size; | |
1938 | class->index = i; | |
1939 | class->pages_per_zspage = pages_per_zspage; | |
1fc6e27d MK |
1940 | class->objs_per_zspage = class->pages_per_zspage * |
1941 | PAGE_SIZE / class->size; | |
1942 | if (pages_per_zspage == 1 && class->objs_per_zspage == 1) | |
7b60a685 | 1943 | class->huge = true; |
66cdef66 GM |
1944 | spin_lock_init(&class->lock); |
1945 | pool->size_class[i] = class; | |
3783689a MK |
1946 | for (fullness = ZS_ALMOST_FULL; fullness <= ZS_ALMOST_EMPTY; |
1947 | fullness++) | |
1948 | INIT_LIST_HEAD(&class->fullness_list[fullness]); | |
66cdef66 GM |
1949 | |
1950 | prev_class = class; | |
61989a80 NG |
1951 | } |
1952 | ||
d34f6157 DS |
1953 | /* debug only, don't abort if it fails */ |
1954 | zs_pool_stat_create(pool, name); | |
0f050d99 | 1955 | |
ab9d306d SS |
1956 | /* |
1957 | * Not critical, we still can use the pool | |
1958 | * and user can trigger compaction manually. | |
1959 | */ | |
1960 | if (zs_register_shrinker(pool) == 0) | |
1961 | pool->shrinker_enabled = true; | |
66cdef66 GM |
1962 | return pool; |
1963 | ||
1964 | err: | |
1965 | zs_destroy_pool(pool); | |
1966 | return NULL; | |
61989a80 | 1967 | } |
66cdef66 | 1968 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1969 | |
66cdef66 | 1970 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1971 | { |
66cdef66 | 1972 | int i; |
61989a80 | 1973 | |
ab9d306d | 1974 | zs_unregister_shrinker(pool); |
0f050d99 GM |
1975 | zs_pool_stat_destroy(pool); |
1976 | ||
66cdef66 GM |
1977 | for (i = 0; i < zs_size_classes; i++) { |
1978 | int fg; | |
1979 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1980 | |
66cdef66 GM |
1981 | if (!class) |
1982 | continue; | |
61989a80 | 1983 | |
66cdef66 GM |
1984 | if (class->index != i) |
1985 | continue; | |
61989a80 | 1986 | |
3783689a MK |
1987 | for (fg = ZS_ALMOST_FULL; fg <= ZS_ALMOST_EMPTY; fg++) { |
1988 | if (!list_empty(&class->fullness_list[fg])) { | |
66cdef66 GM |
1989 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
1990 | class->size, fg); | |
1991 | } | |
1992 | } | |
1993 | kfree(class); | |
1994 | } | |
f553646a | 1995 | |
3783689a | 1996 | destroy_cache(pool); |
66cdef66 | 1997 | kfree(pool->size_class); |
0f050d99 | 1998 | kfree(pool->name); |
66cdef66 GM |
1999 | kfree(pool); |
2000 | } | |
2001 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2002 | |
66cdef66 GM |
2003 | static int __init zs_init(void) |
2004 | { | |
2005 | int ret = zs_register_cpu_notifier(); | |
2006 | ||
0f050d99 GM |
2007 | if (ret) |
2008 | goto notifier_fail; | |
66cdef66 GM |
2009 | |
2010 | init_zs_size_classes(); | |
2011 | ||
2012 | #ifdef CONFIG_ZPOOL | |
2013 | zpool_register_driver(&zs_zpool_driver); | |
2014 | #endif | |
0f050d99 | 2015 | |
4abaac9b DS |
2016 | zs_stat_init(); |
2017 | ||
66cdef66 | 2018 | return 0; |
0f050d99 | 2019 | |
0f050d99 GM |
2020 | notifier_fail: |
2021 | zs_unregister_cpu_notifier(); | |
2022 | ||
2023 | return ret; | |
61989a80 | 2024 | } |
61989a80 | 2025 | |
66cdef66 | 2026 | static void __exit zs_exit(void) |
61989a80 | 2027 | { |
66cdef66 GM |
2028 | #ifdef CONFIG_ZPOOL |
2029 | zpool_unregister_driver(&zs_zpool_driver); | |
2030 | #endif | |
2031 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
2032 | |
2033 | zs_stat_exit(); | |
61989a80 | 2034 | } |
069f101f BH |
2035 | |
2036 | module_init(zs_init); | |
2037 | module_exit(zs_exit); | |
2038 | ||
2039 | MODULE_LICENSE("Dual BSD/GPL"); | |
2040 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |