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 | 19 | * page->private: points to zspage |
ffedd09f | 20 | * page->index: links together all component pages of a zspage |
48b4800a MK |
21 | * For the huge page, this is always 0, so we use this field |
22 | * to store handle. | |
ffedd09f | 23 | * page->page_type: first object offset in a subpage of zspage |
2db51dae NG |
24 | * |
25 | * Usage of struct page flags: | |
26 | * PG_private: identifies the first component page | |
399d8eeb | 27 | * PG_owner_priv_1: identifies the huge component page |
2db51dae NG |
28 | * |
29 | */ | |
30 | ||
4abaac9b DS |
31 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
32 | ||
b475d42d MK |
33 | /* |
34 | * lock ordering: | |
35 | * page_lock | |
36 | * pool->migrate_lock | |
37 | * class->lock | |
38 | * zspage->lock | |
39 | */ | |
40 | ||
61989a80 NG |
41 | #include <linux/module.h> |
42 | #include <linux/kernel.h> | |
312fcae2 | 43 | #include <linux/sched.h> |
50d34394 | 44 | #include <linux/magic.h> |
61989a80 NG |
45 | #include <linux/bitops.h> |
46 | #include <linux/errno.h> | |
47 | #include <linux/highmem.h> | |
61989a80 NG |
48 | #include <linux/string.h> |
49 | #include <linux/slab.h> | |
ca5999fd | 50 | #include <linux/pgtable.h> |
65fddcfc | 51 | #include <asm/tlbflush.h> |
61989a80 NG |
52 | #include <linux/cpumask.h> |
53 | #include <linux/cpu.h> | |
0cbb613f | 54 | #include <linux/vmalloc.h> |
759b26b2 | 55 | #include <linux/preempt.h> |
0959c63f | 56 | #include <linux/spinlock.h> |
93144ca3 | 57 | #include <linux/shrinker.h> |
0959c63f | 58 | #include <linux/types.h> |
0f050d99 | 59 | #include <linux/debugfs.h> |
bcf1647d | 60 | #include <linux/zsmalloc.h> |
c795779d | 61 | #include <linux/zpool.h> |
48b4800a | 62 | #include <linux/mount.h> |
8e9231f8 | 63 | #include <linux/pseudo_fs.h> |
dd4123f3 | 64 | #include <linux/migrate.h> |
701d6785 | 65 | #include <linux/wait.h> |
48b4800a | 66 | #include <linux/pagemap.h> |
cdc346b3 | 67 | #include <linux/fs.h> |
48b4800a MK |
68 | |
69 | #define ZSPAGE_MAGIC 0x58 | |
0959c63f SJ |
70 | |
71 | /* | |
cb152a1a | 72 | * This must be power of 2 and greater than or equal to sizeof(link_free). |
0959c63f SJ |
73 | * These two conditions ensure that any 'struct link_free' itself doesn't |
74 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
75 | * to restore link_free pointer values. | |
76 | */ | |
77 | #define ZS_ALIGN 8 | |
78 | ||
79 | /* | |
80 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
81 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
82 | */ | |
83 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
84 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
85 | ||
2e40e163 MK |
86 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
87 | ||
0959c63f SJ |
88 | /* |
89 | * Object location (<PFN>, <obj_idx>) is encoded as | |
b956b5ac | 90 | * a single (unsigned long) handle value. |
0959c63f | 91 | * |
bfd093f5 | 92 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
93 | * |
94 | * This is made more complicated by various memory models and PAE. | |
95 | */ | |
96 | ||
02390b87 KS |
97 | #ifndef MAX_POSSIBLE_PHYSMEM_BITS |
98 | #ifdef MAX_PHYSMEM_BITS | |
99 | #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS | |
100 | #else | |
0959c63f SJ |
101 | /* |
102 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
103 | * be PAGE_SHIFT | |
104 | */ | |
02390b87 | 105 | #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG |
0959c63f SJ |
106 | #endif |
107 | #endif | |
02390b87 KS |
108 | |
109 | #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 | 110 | |
312fcae2 MK |
111 | /* |
112 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
113 | * to identify the object was allocated or not. | |
114 | * It's okay to add the status bit in the least bit because | |
115 | * header keeps handle which is 4byte-aligned address so we | |
116 | * have room for two bit at least. | |
117 | */ | |
118 | #define OBJ_ALLOCATED_TAG 1 | |
119 | #define OBJ_TAG_BITS 1 | |
120 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
121 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
122 | ||
a41ec880 | 123 | #define HUGE_BITS 1 |
cf8e0fed JM |
124 | #define FULLNESS_BITS 2 |
125 | #define CLASS_BITS 8 | |
126 | #define ISOLATED_BITS 3 | |
127 | #define MAGIC_VAL_BITS 8 | |
128 | ||
0959c63f SJ |
129 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) |
130 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
131 | #define ZS_MIN_ALLOC_SIZE \ | |
132 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 133 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 134 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
135 | |
136 | /* | |
7eb52512 | 137 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
138 | * trader-off here: |
139 | * - Large number of size classes is potentially wasteful as free page are | |
140 | * spread across these classes | |
141 | * - Small number of size classes causes large internal fragmentation | |
142 | * - Probably its better to use specific size classes (empirically | |
143 | * determined). NOTE: all those class sizes must be set as multiple of | |
144 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
145 | * | |
146 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
147 | * (reason above) | |
148 | */ | |
3783689a | 149 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
cf8e0fed JM |
150 | #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \ |
151 | ZS_SIZE_CLASS_DELTA) + 1) | |
0959c63f | 152 | |
0959c63f | 153 | enum fullness_group { |
0959c63f | 154 | ZS_EMPTY, |
48b4800a MK |
155 | ZS_ALMOST_EMPTY, |
156 | ZS_ALMOST_FULL, | |
157 | ZS_FULL, | |
158 | NR_ZS_FULLNESS, | |
0959c63f SJ |
159 | }; |
160 | ||
3828a764 | 161 | enum class_stat_type { |
48b4800a MK |
162 | CLASS_EMPTY, |
163 | CLASS_ALMOST_EMPTY, | |
164 | CLASS_ALMOST_FULL, | |
165 | CLASS_FULL, | |
0f050d99 GM |
166 | OBJ_ALLOCATED, |
167 | OBJ_USED, | |
48b4800a | 168 | NR_ZS_STAT_TYPE, |
0f050d99 GM |
169 | }; |
170 | ||
0f050d99 GM |
171 | struct zs_size_stat { |
172 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
173 | }; | |
174 | ||
57244594 SS |
175 | #ifdef CONFIG_ZSMALLOC_STAT |
176 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
177 | #endif |
178 | ||
48b4800a MK |
179 | #ifdef CONFIG_COMPACTION |
180 | static struct vfsmount *zsmalloc_mnt; | |
181 | #endif | |
182 | ||
0959c63f SJ |
183 | /* |
184 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
185 | * n <= N / f, where | |
186 | * n = number of allocated objects | |
187 | * N = total number of objects zspage can store | |
6dd9737e | 188 | * f = fullness_threshold_frac |
0959c63f SJ |
189 | * |
190 | * Similarly, we assign zspage to: | |
191 | * ZS_ALMOST_FULL when n > N / f | |
192 | * ZS_EMPTY when n == 0 | |
193 | * ZS_FULL when n == N | |
194 | * | |
195 | * (see: fix_fullness_group()) | |
196 | */ | |
197 | static const int fullness_threshold_frac = 4; | |
010b495e | 198 | static size_t huge_class_size; |
0959c63f SJ |
199 | |
200 | struct size_class { | |
57244594 | 201 | spinlock_t lock; |
48b4800a | 202 | struct list_head fullness_list[NR_ZS_FULLNESS]; |
0959c63f SJ |
203 | /* |
204 | * Size of objects stored in this class. Must be multiple | |
205 | * of ZS_ALIGN. | |
206 | */ | |
207 | int size; | |
1fc6e27d | 208 | int objs_per_zspage; |
7dfa4612 WY |
209 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
210 | int pages_per_zspage; | |
48b4800a MK |
211 | |
212 | unsigned int index; | |
213 | struct zs_size_stat stats; | |
0959c63f SJ |
214 | }; |
215 | ||
216 | /* | |
217 | * Placed within free objects to form a singly linked list. | |
3783689a | 218 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
219 | * |
220 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
221 | */ | |
222 | struct link_free { | |
2e40e163 MK |
223 | union { |
224 | /* | |
bfd093f5 | 225 | * Free object index; |
2e40e163 MK |
226 | * It's valid for non-allocated object |
227 | */ | |
bfd093f5 | 228 | unsigned long next; |
2e40e163 MK |
229 | /* |
230 | * Handle of allocated object. | |
231 | */ | |
232 | unsigned long handle; | |
233 | }; | |
0959c63f SJ |
234 | }; |
235 | ||
236 | struct zs_pool { | |
6f3526d6 | 237 | const char *name; |
0f050d99 | 238 | |
cf8e0fed | 239 | struct size_class *size_class[ZS_SIZE_CLASSES]; |
2e40e163 | 240 | struct kmem_cache *handle_cachep; |
3783689a | 241 | struct kmem_cache *zspage_cachep; |
0959c63f | 242 | |
13de8933 | 243 | atomic_long_t pages_allocated; |
0f050d99 | 244 | |
7d3f3938 | 245 | struct zs_pool_stats stats; |
ab9d306d SS |
246 | |
247 | /* Compact classes */ | |
248 | struct shrinker shrinker; | |
93144ca3 | 249 | |
0f050d99 GM |
250 | #ifdef CONFIG_ZSMALLOC_STAT |
251 | struct dentry *stat_dentry; | |
252 | #endif | |
48b4800a MK |
253 | #ifdef CONFIG_COMPACTION |
254 | struct inode *inode; | |
255 | struct work_struct free_work; | |
256 | #endif | |
b475d42d MK |
257 | /* protect page/zspage migration */ |
258 | rwlock_t migrate_lock; | |
0959c63f | 259 | }; |
61989a80 | 260 | |
3783689a MK |
261 | struct zspage { |
262 | struct { | |
a41ec880 | 263 | unsigned int huge:HUGE_BITS; |
3783689a | 264 | unsigned int fullness:FULLNESS_BITS; |
85d492f2 | 265 | unsigned int class:CLASS_BITS + 1; |
48b4800a MK |
266 | unsigned int isolated:ISOLATED_BITS; |
267 | unsigned int magic:MAGIC_VAL_BITS; | |
3783689a MK |
268 | }; |
269 | unsigned int inuse; | |
bfd093f5 | 270 | unsigned int freeobj; |
3783689a MK |
271 | struct page *first_page; |
272 | struct list_head list; /* fullness list */ | |
48b4800a MK |
273 | #ifdef CONFIG_COMPACTION |
274 | rwlock_t lock; | |
275 | #endif | |
3783689a | 276 | }; |
61989a80 | 277 | |
f553646a | 278 | struct mapping_area { |
f553646a | 279 | char *vm_buf; /* copy buffer for objects that span pages */ |
f553646a SJ |
280 | char *vm_addr; /* address of kmap_atomic()'ed pages */ |
281 | enum zs_mapmode vm_mm; /* mapping mode */ | |
282 | }; | |
283 | ||
a41ec880 MK |
284 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
285 | static void SetZsHugePage(struct zspage *zspage) | |
286 | { | |
287 | zspage->huge = 1; | |
288 | } | |
289 | ||
290 | static bool ZsHugePage(struct zspage *zspage) | |
291 | { | |
292 | return zspage->huge; | |
293 | } | |
294 | ||
48b4800a MK |
295 | #ifdef CONFIG_COMPACTION |
296 | static int zs_register_migration(struct zs_pool *pool); | |
297 | static void zs_unregister_migration(struct zs_pool *pool); | |
298 | static void migrate_lock_init(struct zspage *zspage); | |
299 | static void migrate_read_lock(struct zspage *zspage); | |
300 | static void migrate_read_unlock(struct zspage *zspage); | |
b475d42d MK |
301 | static void migrate_write_lock(struct zspage *zspage); |
302 | static void migrate_write_lock_nested(struct zspage *zspage); | |
303 | static void migrate_write_unlock(struct zspage *zspage); | |
48b4800a MK |
304 | static void kick_deferred_free(struct zs_pool *pool); |
305 | static void init_deferred_free(struct zs_pool *pool); | |
306 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage); | |
307 | #else | |
308 | static int zsmalloc_mount(void) { return 0; } | |
309 | static void zsmalloc_unmount(void) {} | |
310 | static int zs_register_migration(struct zs_pool *pool) { return 0; } | |
311 | static void zs_unregister_migration(struct zs_pool *pool) {} | |
312 | static void migrate_lock_init(struct zspage *zspage) {} | |
313 | static void migrate_read_lock(struct zspage *zspage) {} | |
314 | static void migrate_read_unlock(struct zspage *zspage) {} | |
b475d42d MK |
315 | static void migrate_write_lock(struct zspage *zspage) {} |
316 | static void migrate_write_lock_nested(struct zspage *zspage) {} | |
317 | static void migrate_write_unlock(struct zspage *zspage) {} | |
48b4800a MK |
318 | static void kick_deferred_free(struct zs_pool *pool) {} |
319 | static void init_deferred_free(struct zs_pool *pool) {} | |
320 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {} | |
321 | #endif | |
322 | ||
3783689a | 323 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
324 | { |
325 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
326 | 0, 0, NULL); | |
3783689a MK |
327 | if (!pool->handle_cachep) |
328 | return 1; | |
329 | ||
330 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
331 | 0, 0, NULL); | |
332 | if (!pool->zspage_cachep) { | |
333 | kmem_cache_destroy(pool->handle_cachep); | |
334 | pool->handle_cachep = NULL; | |
335 | return 1; | |
336 | } | |
337 | ||
338 | return 0; | |
2e40e163 MK |
339 | } |
340 | ||
3783689a | 341 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 342 | { |
cd10add0 | 343 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 344 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
345 | } |
346 | ||
3783689a | 347 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
348 | { |
349 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
48b4800a | 350 | gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
2e40e163 MK |
351 | } |
352 | ||
3783689a | 353 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
354 | { |
355 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
356 | } | |
357 | ||
3783689a MK |
358 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
359 | { | |
f0231305 | 360 | return kmem_cache_zalloc(pool->zspage_cachep, |
48b4800a | 361 | flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
399d8eeb | 362 | } |
3783689a MK |
363 | |
364 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
365 | { | |
366 | kmem_cache_free(pool->zspage_cachep, zspage); | |
367 | } | |
368 | ||
b475d42d | 369 | /* class->lock(which owns the handle) synchronizes races */ |
2e40e163 MK |
370 | static void record_obj(unsigned long handle, unsigned long obj) |
371 | { | |
b475d42d | 372 | *(unsigned long *)handle = obj; |
2e40e163 MK |
373 | } |
374 | ||
c795779d DS |
375 | /* zpool driver */ |
376 | ||
377 | #ifdef CONFIG_ZPOOL | |
378 | ||
6f3526d6 | 379 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 380 | const struct zpool_ops *zpool_ops, |
479305fd | 381 | struct zpool *zpool) |
c795779d | 382 | { |
d0d8da2d SS |
383 | /* |
384 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
385 | * different contexts and its caller must provide a valid | |
386 | * gfp mask. | |
387 | */ | |
388 | return zs_create_pool(name); | |
c795779d DS |
389 | } |
390 | ||
391 | static void zs_zpool_destroy(void *pool) | |
392 | { | |
393 | zs_destroy_pool(pool); | |
394 | } | |
395 | ||
396 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
397 | unsigned long *handle) | |
398 | { | |
d0d8da2d | 399 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
400 | return *handle ? 0 : -1; |
401 | } | |
402 | static void zs_zpool_free(void *pool, unsigned long handle) | |
403 | { | |
404 | zs_free(pool, handle); | |
405 | } | |
406 | ||
c795779d DS |
407 | static void *zs_zpool_map(void *pool, unsigned long handle, |
408 | enum zpool_mapmode mm) | |
409 | { | |
410 | enum zs_mapmode zs_mm; | |
411 | ||
412 | switch (mm) { | |
413 | case ZPOOL_MM_RO: | |
414 | zs_mm = ZS_MM_RO; | |
415 | break; | |
416 | case ZPOOL_MM_WO: | |
417 | zs_mm = ZS_MM_WO; | |
418 | break; | |
e4a9bc58 | 419 | case ZPOOL_MM_RW: |
c795779d DS |
420 | default: |
421 | zs_mm = ZS_MM_RW; | |
422 | break; | |
423 | } | |
424 | ||
425 | return zs_map_object(pool, handle, zs_mm); | |
426 | } | |
427 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
428 | { | |
429 | zs_unmap_object(pool, handle); | |
430 | } | |
431 | ||
432 | static u64 zs_zpool_total_size(void *pool) | |
433 | { | |
722cdc17 | 434 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
435 | } |
436 | ||
437 | static struct zpool_driver zs_zpool_driver = { | |
c165f25d HZ |
438 | .type = "zsmalloc", |
439 | .owner = THIS_MODULE, | |
440 | .create = zs_zpool_create, | |
441 | .destroy = zs_zpool_destroy, | |
442 | .malloc_support_movable = true, | |
443 | .malloc = zs_zpool_malloc, | |
444 | .free = zs_zpool_free, | |
445 | .map = zs_zpool_map, | |
446 | .unmap = zs_zpool_unmap, | |
447 | .total_size = zs_zpool_total_size, | |
c795779d DS |
448 | }; |
449 | ||
137f8cff | 450 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
451 | #endif /* CONFIG_ZPOOL */ |
452 | ||
61989a80 NG |
453 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
454 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
455 | ||
3457f414 | 456 | static __maybe_unused int is_first_page(struct page *page) |
61989a80 | 457 | { |
a27545bf | 458 | return PagePrivate(page); |
61989a80 NG |
459 | } |
460 | ||
48b4800a | 461 | /* Protected by class->lock */ |
3783689a | 462 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 463 | { |
3783689a | 464 | return zspage->inuse; |
4f42047b MK |
465 | } |
466 | ||
4f42047b | 467 | |
3783689a | 468 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 469 | { |
3783689a | 470 | zspage->inuse += val; |
4f42047b MK |
471 | } |
472 | ||
48b4800a | 473 | static inline struct page *get_first_page(struct zspage *zspage) |
4f42047b | 474 | { |
48b4800a | 475 | struct page *first_page = zspage->first_page; |
3783689a | 476 | |
48b4800a MK |
477 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
478 | return first_page; | |
4f42047b MK |
479 | } |
480 | ||
48b4800a | 481 | static inline int get_first_obj_offset(struct page *page) |
4f42047b | 482 | { |
ffedd09f | 483 | return page->page_type; |
48b4800a | 484 | } |
3783689a | 485 | |
48b4800a MK |
486 | static inline void set_first_obj_offset(struct page *page, int offset) |
487 | { | |
ffedd09f | 488 | page->page_type = offset; |
4f42047b MK |
489 | } |
490 | ||
bfd093f5 | 491 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 492 | { |
bfd093f5 | 493 | return zspage->freeobj; |
4f42047b MK |
494 | } |
495 | ||
bfd093f5 | 496 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 497 | { |
bfd093f5 | 498 | zspage->freeobj = obj; |
4f42047b MK |
499 | } |
500 | ||
3783689a | 501 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 502 | unsigned int *class_idx, |
61989a80 NG |
503 | enum fullness_group *fullness) |
504 | { | |
48b4800a MK |
505 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); |
506 | ||
3783689a MK |
507 | *fullness = zspage->fullness; |
508 | *class_idx = zspage->class; | |
61989a80 NG |
509 | } |
510 | ||
67f1c9cd MK |
511 | static struct size_class *zspage_class(struct zs_pool *pool, |
512 | struct zspage *zspage) | |
513 | { | |
514 | return pool->size_class[zspage->class]; | |
515 | } | |
516 | ||
3783689a | 517 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 518 | unsigned int class_idx, |
61989a80 NG |
519 | enum fullness_group fullness) |
520 | { | |
3783689a MK |
521 | zspage->class = class_idx; |
522 | zspage->fullness = fullness; | |
61989a80 NG |
523 | } |
524 | ||
c3e3e88a NC |
525 | /* |
526 | * zsmalloc divides the pool into various size classes where each | |
527 | * class maintains a list of zspages where each zspage is divided | |
528 | * into equal sized chunks. Each allocation falls into one of these | |
529 | * classes depending on its size. This function returns index of the | |
cb152a1a | 530 | * size class which has chunk size big enough to hold the given size. |
c3e3e88a | 531 | */ |
61989a80 NG |
532 | static int get_size_class_index(int size) |
533 | { | |
534 | int idx = 0; | |
535 | ||
536 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
537 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
538 | ZS_SIZE_CLASS_DELTA); | |
539 | ||
cf8e0fed | 540 | return min_t(int, ZS_SIZE_CLASSES - 1, idx); |
61989a80 NG |
541 | } |
542 | ||
3828a764 MK |
543 | /* type can be of enum type class_stat_type or fullness_group */ |
544 | static inline void class_stat_inc(struct size_class *class, | |
3eb95fea | 545 | int type, unsigned long cnt) |
248ca1b0 | 546 | { |
48b4800a | 547 | class->stats.objs[type] += cnt; |
248ca1b0 MK |
548 | } |
549 | ||
3828a764 MK |
550 | /* type can be of enum type class_stat_type or fullness_group */ |
551 | static inline void class_stat_dec(struct size_class *class, | |
3eb95fea | 552 | int type, unsigned long cnt) |
248ca1b0 | 553 | { |
48b4800a | 554 | class->stats.objs[type] -= cnt; |
248ca1b0 MK |
555 | } |
556 | ||
3828a764 | 557 | /* type can be of enum type class_stat_type or fullness_group */ |
248ca1b0 | 558 | static inline unsigned long zs_stat_get(struct size_class *class, |
3eb95fea | 559 | int type) |
248ca1b0 | 560 | { |
48b4800a | 561 | return class->stats.objs[type]; |
248ca1b0 MK |
562 | } |
563 | ||
57244594 SS |
564 | #ifdef CONFIG_ZSMALLOC_STAT |
565 | ||
4abaac9b | 566 | static void __init zs_stat_init(void) |
248ca1b0 | 567 | { |
4abaac9b DS |
568 | if (!debugfs_initialized()) { |
569 | pr_warn("debugfs not available, stat dir not created\n"); | |
570 | return; | |
571 | } | |
248ca1b0 MK |
572 | |
573 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
248ca1b0 MK |
574 | } |
575 | ||
576 | static void __exit zs_stat_exit(void) | |
577 | { | |
578 | debugfs_remove_recursive(zs_stat_root); | |
579 | } | |
580 | ||
1120ed54 SS |
581 | static unsigned long zs_can_compact(struct size_class *class); |
582 | ||
248ca1b0 MK |
583 | static int zs_stats_size_show(struct seq_file *s, void *v) |
584 | { | |
585 | int i; | |
586 | struct zs_pool *pool = s->private; | |
587 | struct size_class *class; | |
588 | int objs_per_zspage; | |
589 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 590 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
591 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
592 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 593 | unsigned long total_freeable = 0; |
248ca1b0 | 594 | |
1120ed54 | 595 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
596 | "class", "size", "almost_full", "almost_empty", |
597 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 598 | "pages_per_zspage", "freeable"); |
248ca1b0 | 599 | |
cf8e0fed | 600 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
248ca1b0 MK |
601 | class = pool->size_class[i]; |
602 | ||
603 | if (class->index != i) | |
604 | continue; | |
605 | ||
606 | spin_lock(&class->lock); | |
607 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
608 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
609 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
610 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 611 | freeable = zs_can_compact(class); |
248ca1b0 MK |
612 | spin_unlock(&class->lock); |
613 | ||
b4fd07a0 | 614 | objs_per_zspage = class->objs_per_zspage; |
248ca1b0 MK |
615 | pages_used = obj_allocated / objs_per_zspage * |
616 | class->pages_per_zspage; | |
617 | ||
1120ed54 SS |
618 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
619 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
620 | i, class->size, class_almost_full, class_almost_empty, |
621 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 622 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
623 | |
624 | total_class_almost_full += class_almost_full; | |
625 | total_class_almost_empty += class_almost_empty; | |
626 | total_objs += obj_allocated; | |
627 | total_used_objs += obj_used; | |
628 | total_pages += pages_used; | |
1120ed54 | 629 | total_freeable += freeable; |
248ca1b0 MK |
630 | } |
631 | ||
632 | seq_puts(s, "\n"); | |
1120ed54 | 633 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
634 | "Total", "", total_class_almost_full, |
635 | total_class_almost_empty, total_objs, | |
1120ed54 | 636 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
637 | |
638 | return 0; | |
639 | } | |
5ad35093 | 640 | DEFINE_SHOW_ATTRIBUTE(zs_stats_size); |
248ca1b0 | 641 | |
d34f6157 | 642 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 643 | { |
4abaac9b DS |
644 | if (!zs_stat_root) { |
645 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 646 | return; |
4abaac9b | 647 | } |
248ca1b0 | 648 | |
4268509a GKH |
649 | pool->stat_dentry = debugfs_create_dir(name, zs_stat_root); |
650 | ||
651 | debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool, | |
652 | &zs_stats_size_fops); | |
248ca1b0 MK |
653 | } |
654 | ||
655 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
656 | { | |
657 | debugfs_remove_recursive(pool->stat_dentry); | |
658 | } | |
659 | ||
660 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 661 | static void __init zs_stat_init(void) |
248ca1b0 | 662 | { |
248ca1b0 MK |
663 | } |
664 | ||
665 | static void __exit zs_stat_exit(void) | |
666 | { | |
667 | } | |
668 | ||
d34f6157 | 669 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 670 | { |
248ca1b0 MK |
671 | } |
672 | ||
673 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
674 | { | |
675 | } | |
248ca1b0 MK |
676 | #endif |
677 | ||
48b4800a | 678 | |
c3e3e88a NC |
679 | /* |
680 | * For each size class, zspages are divided into different groups | |
681 | * depending on how "full" they are. This was done so that we could | |
682 | * easily find empty or nearly empty zspages when we try to shrink | |
683 | * the pool (not yet implemented). This function returns fullness | |
684 | * status of the given page. | |
685 | */ | |
1fc6e27d | 686 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 687 | struct zspage *zspage) |
61989a80 | 688 | { |
1fc6e27d | 689 | int inuse, objs_per_zspage; |
61989a80 | 690 | enum fullness_group fg; |
830e4bc5 | 691 | |
3783689a | 692 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 693 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
694 | |
695 | if (inuse == 0) | |
696 | fg = ZS_EMPTY; | |
1fc6e27d | 697 | else if (inuse == objs_per_zspage) |
61989a80 | 698 | fg = ZS_FULL; |
1fc6e27d | 699 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
700 | fg = ZS_ALMOST_EMPTY; |
701 | else | |
702 | fg = ZS_ALMOST_FULL; | |
703 | ||
704 | return fg; | |
705 | } | |
706 | ||
c3e3e88a NC |
707 | /* |
708 | * Each size class maintains various freelists and zspages are assigned | |
709 | * to one of these freelists based on the number of live objects they | |
710 | * have. This functions inserts the given zspage into the freelist | |
711 | * identified by <class, fullness_group>. | |
712 | */ | |
251cbb95 | 713 | static void insert_zspage(struct size_class *class, |
3783689a MK |
714 | struct zspage *zspage, |
715 | enum fullness_group fullness) | |
61989a80 | 716 | { |
3783689a | 717 | struct zspage *head; |
61989a80 | 718 | |
3828a764 | 719 | class_stat_inc(class, fullness, 1); |
3783689a MK |
720 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
721 | struct zspage, list); | |
58f17117 | 722 | /* |
3783689a MK |
723 | * We want to see more ZS_FULL pages and less almost empty/full. |
724 | * Put pages with higher ->inuse first. | |
58f17117 | 725 | */ |
110ceb82 ML |
726 | if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head)) |
727 | list_add(&zspage->list, &head->list); | |
728 | else | |
729 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
61989a80 NG |
730 | } |
731 | ||
c3e3e88a NC |
732 | /* |
733 | * This function removes the given zspage from the freelist identified | |
734 | * by <class, fullness_group>. | |
735 | */ | |
251cbb95 | 736 | static void remove_zspage(struct size_class *class, |
3783689a MK |
737 | struct zspage *zspage, |
738 | enum fullness_group fullness) | |
61989a80 | 739 | { |
3783689a | 740 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
61989a80 | 741 | |
3783689a | 742 | list_del_init(&zspage->list); |
3828a764 | 743 | class_stat_dec(class, fullness, 1); |
61989a80 NG |
744 | } |
745 | ||
c3e3e88a NC |
746 | /* |
747 | * Each size class maintains zspages in different fullness groups depending | |
748 | * on the number of live objects they contain. When allocating or freeing | |
749 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
750 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
751 | * a status change has occurred for the given page and accordingly moves the | |
752 | * page from the freelist of the old fullness group to that of the new | |
753 | * fullness group. | |
754 | */ | |
c7806261 | 755 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 756 | struct zspage *zspage) |
61989a80 NG |
757 | { |
758 | int class_idx; | |
61989a80 NG |
759 | enum fullness_group currfg, newfg; |
760 | ||
3783689a MK |
761 | get_zspage_mapping(zspage, &class_idx, &currfg); |
762 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
763 | if (newfg == currfg) |
764 | goto out; | |
765 | ||
c4549b87 MK |
766 | remove_zspage(class, zspage, currfg); |
767 | insert_zspage(class, zspage, newfg); | |
3783689a | 768 | set_zspage_mapping(zspage, class_idx, newfg); |
61989a80 NG |
769 | out: |
770 | return newfg; | |
771 | } | |
772 | ||
773 | /* | |
774 | * We have to decide on how many pages to link together | |
775 | * to form a zspage for each size class. This is important | |
776 | * to reduce wastage due to unusable space left at end of | |
777 | * each zspage which is given as: | |
888fa374 YX |
778 | * wastage = Zp % class_size |
779 | * usage = Zp - wastage | |
61989a80 NG |
780 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
781 | * | |
782 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
783 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
784 | * since then we can perfectly fit in 8 such objects. | |
785 | */ | |
2e3b6154 | 786 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
787 | { |
788 | int i, max_usedpc = 0; | |
789 | /* zspage order which gives maximum used size per KB */ | |
790 | int max_usedpc_order = 1; | |
791 | ||
84d4faab | 792 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
793 | int zspage_size; |
794 | int waste, usedpc; | |
795 | ||
796 | zspage_size = i * PAGE_SIZE; | |
797 | waste = zspage_size % class_size; | |
798 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
799 | ||
800 | if (usedpc > max_usedpc) { | |
801 | max_usedpc = usedpc; | |
802 | max_usedpc_order = i; | |
803 | } | |
804 | } | |
805 | ||
806 | return max_usedpc_order; | |
807 | } | |
808 | ||
3783689a | 809 | static struct zspage *get_zspage(struct page *page) |
61989a80 | 810 | { |
a6c5e0f7 | 811 | struct zspage *zspage = (struct zspage *)page_private(page); |
48b4800a MK |
812 | |
813 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); | |
814 | return zspage; | |
61989a80 NG |
815 | } |
816 | ||
817 | static struct page *get_next_page(struct page *page) | |
818 | { | |
a41ec880 MK |
819 | struct zspage *zspage = get_zspage(page); |
820 | ||
821 | if (unlikely(ZsHugePage(zspage))) | |
48b4800a MK |
822 | return NULL; |
823 | ||
ffedd09f | 824 | return (struct page *)page->index; |
61989a80 NG |
825 | } |
826 | ||
bfd093f5 MK |
827 | /** |
828 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
e8b098fc | 829 | * @obj: the encoded object value |
bfd093f5 MK |
830 | * @page: page object resides in zspage |
831 | * @obj_idx: object index | |
67296874 | 832 | */ |
bfd093f5 MK |
833 | static void obj_to_location(unsigned long obj, struct page **page, |
834 | unsigned int *obj_idx) | |
61989a80 | 835 | { |
bfd093f5 MK |
836 | obj >>= OBJ_TAG_BITS; |
837 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
838 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
839 | } | |
61989a80 | 840 | |
67f1c9cd MK |
841 | static void obj_to_page(unsigned long obj, struct page **page) |
842 | { | |
843 | obj >>= OBJ_TAG_BITS; | |
844 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
845 | } | |
846 | ||
bfd093f5 MK |
847 | /** |
848 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
849 | * @page: page object resides in zspage | |
850 | * @obj_idx: object index | |
851 | */ | |
852 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
853 | { | |
854 | unsigned long obj; | |
61989a80 | 855 | |
312fcae2 | 856 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 857 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 858 | obj <<= OBJ_TAG_BITS; |
61989a80 | 859 | |
bfd093f5 | 860 | return obj; |
61989a80 NG |
861 | } |
862 | ||
2e40e163 MK |
863 | static unsigned long handle_to_obj(unsigned long handle) |
864 | { | |
865 | return *(unsigned long *)handle; | |
866 | } | |
867 | ||
3ae92ac2 | 868 | static bool obj_allocated(struct page *page, void *obj, unsigned long *phandle) |
312fcae2 | 869 | { |
3ae92ac2 | 870 | unsigned long handle; |
a41ec880 | 871 | struct zspage *zspage = get_zspage(page); |
3ae92ac2 | 872 | |
a41ec880 | 873 | if (unlikely(ZsHugePage(zspage))) { |
830e4bc5 | 874 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3ae92ac2 | 875 | handle = page->index; |
7b60a685 | 876 | } else |
3ae92ac2 MK |
877 | handle = *(unsigned long *)obj; |
878 | ||
879 | if (!(handle & OBJ_ALLOCATED_TAG)) | |
880 | return false; | |
881 | ||
882 | *phandle = handle & ~OBJ_ALLOCATED_TAG; | |
883 | return true; | |
312fcae2 MK |
884 | } |
885 | ||
f4477e90 NG |
886 | static void reset_page(struct page *page) |
887 | { | |
48b4800a | 888 | __ClearPageMovable(page); |
18fd06bf | 889 | ClearPagePrivate(page); |
f4477e90 | 890 | set_page_private(page, 0); |
48b4800a | 891 | page_mapcount_reset(page); |
ffedd09f | 892 | page->index = 0; |
48b4800a MK |
893 | } |
894 | ||
4d0a5402 | 895 | static int trylock_zspage(struct zspage *zspage) |
48b4800a MK |
896 | { |
897 | struct page *cursor, *fail; | |
898 | ||
899 | for (cursor = get_first_page(zspage); cursor != NULL; cursor = | |
900 | get_next_page(cursor)) { | |
901 | if (!trylock_page(cursor)) { | |
902 | fail = cursor; | |
903 | goto unlock; | |
904 | } | |
905 | } | |
906 | ||
907 | return 1; | |
908 | unlock: | |
909 | for (cursor = get_first_page(zspage); cursor != fail; cursor = | |
910 | get_next_page(cursor)) | |
911 | unlock_page(cursor); | |
912 | ||
913 | return 0; | |
f4477e90 NG |
914 | } |
915 | ||
48b4800a MK |
916 | static void __free_zspage(struct zs_pool *pool, struct size_class *class, |
917 | struct zspage *zspage) | |
61989a80 | 918 | { |
3783689a | 919 | struct page *page, *next; |
48b4800a MK |
920 | enum fullness_group fg; |
921 | unsigned int class_idx; | |
922 | ||
923 | get_zspage_mapping(zspage, &class_idx, &fg); | |
924 | ||
925 | assert_spin_locked(&class->lock); | |
61989a80 | 926 | |
3783689a | 927 | VM_BUG_ON(get_zspage_inuse(zspage)); |
48b4800a | 928 | VM_BUG_ON(fg != ZS_EMPTY); |
61989a80 | 929 | |
48b4800a | 930 | next = page = get_first_page(zspage); |
3783689a | 931 | do { |
48b4800a MK |
932 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
933 | next = get_next_page(page); | |
3783689a | 934 | reset_page(page); |
48b4800a | 935 | unlock_page(page); |
91537fee | 936 | dec_zone_page_state(page, NR_ZSPAGES); |
3783689a MK |
937 | put_page(page); |
938 | page = next; | |
939 | } while (page != NULL); | |
61989a80 | 940 | |
3783689a | 941 | cache_free_zspage(pool, zspage); |
48b4800a | 942 | |
3828a764 | 943 | class_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
944 | atomic_long_sub(class->pages_per_zspage, |
945 | &pool->pages_allocated); | |
946 | } | |
947 | ||
948 | static void free_zspage(struct zs_pool *pool, struct size_class *class, | |
949 | struct zspage *zspage) | |
950 | { | |
951 | VM_BUG_ON(get_zspage_inuse(zspage)); | |
952 | VM_BUG_ON(list_empty(&zspage->list)); | |
953 | ||
b475d42d MK |
954 | /* |
955 | * Since zs_free couldn't be sleepable, this function cannot call | |
956 | * lock_page. The page locks trylock_zspage got will be released | |
957 | * by __free_zspage. | |
958 | */ | |
48b4800a MK |
959 | if (!trylock_zspage(zspage)) { |
960 | kick_deferred_free(pool); | |
961 | return; | |
962 | } | |
963 | ||
964 | remove_zspage(class, zspage, ZS_EMPTY); | |
965 | __free_zspage(pool, class, zspage); | |
61989a80 NG |
966 | } |
967 | ||
968 | /* Initialize a newly allocated zspage */ | |
3783689a | 969 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 970 | { |
bfd093f5 | 971 | unsigned int freeobj = 1; |
61989a80 | 972 | unsigned long off = 0; |
48b4800a | 973 | struct page *page = get_first_page(zspage); |
830e4bc5 | 974 | |
61989a80 NG |
975 | while (page) { |
976 | struct page *next_page; | |
977 | struct link_free *link; | |
af4ee5e9 | 978 | void *vaddr; |
61989a80 | 979 | |
3783689a | 980 | set_first_obj_offset(page, off); |
61989a80 | 981 | |
af4ee5e9 MK |
982 | vaddr = kmap_atomic(page); |
983 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
984 | |
985 | while ((off += class->size) < PAGE_SIZE) { | |
3b1d9ca6 | 986 | link->next = freeobj++ << OBJ_TAG_BITS; |
5538c562 | 987 | link += class->size / sizeof(*link); |
61989a80 NG |
988 | } |
989 | ||
990 | /* | |
991 | * We now come to the last (full or partial) object on this | |
992 | * page, which must point to the first object on the next | |
993 | * page (if present) | |
994 | */ | |
995 | next_page = get_next_page(page); | |
bfd093f5 | 996 | if (next_page) { |
3b1d9ca6 | 997 | link->next = freeobj++ << OBJ_TAG_BITS; |
bfd093f5 MK |
998 | } else { |
999 | /* | |
3b1d9ca6 | 1000 | * Reset OBJ_TAG_BITS bit to last link to tell |
bfd093f5 MK |
1001 | * whether it's allocated object or not. |
1002 | */ | |
01a6ad9a | 1003 | link->next = -1UL << OBJ_TAG_BITS; |
bfd093f5 | 1004 | } |
af4ee5e9 | 1005 | kunmap_atomic(vaddr); |
61989a80 | 1006 | page = next_page; |
5538c562 | 1007 | off %= PAGE_SIZE; |
61989a80 | 1008 | } |
bdb0af7c | 1009 | |
bfd093f5 | 1010 | set_freeobj(zspage, 0); |
61989a80 NG |
1011 | } |
1012 | ||
48b4800a MK |
1013 | static void create_page_chain(struct size_class *class, struct zspage *zspage, |
1014 | struct page *pages[]) | |
61989a80 | 1015 | { |
bdb0af7c MK |
1016 | int i; |
1017 | struct page *page; | |
1018 | struct page *prev_page = NULL; | |
48b4800a | 1019 | int nr_pages = class->pages_per_zspage; |
61989a80 NG |
1020 | |
1021 | /* | |
1022 | * Allocate individual pages and link them together as: | |
ffedd09f | 1023 | * 1. all pages are linked together using page->index |
3783689a | 1024 | * 2. each sub-page point to zspage using page->private |
61989a80 | 1025 | * |
3783689a | 1026 | * we set PG_private to identify the first page (i.e. no other sub-page |
22c5cef1 | 1027 | * has this flag set). |
61989a80 | 1028 | */ |
bdb0af7c MK |
1029 | for (i = 0; i < nr_pages; i++) { |
1030 | page = pages[i]; | |
3783689a | 1031 | set_page_private(page, (unsigned long)zspage); |
ffedd09f | 1032 | page->index = 0; |
bdb0af7c | 1033 | if (i == 0) { |
3783689a | 1034 | zspage->first_page = page; |
a27545bf | 1035 | SetPagePrivate(page); |
48b4800a MK |
1036 | if (unlikely(class->objs_per_zspage == 1 && |
1037 | class->pages_per_zspage == 1)) | |
a41ec880 | 1038 | SetZsHugePage(zspage); |
3783689a | 1039 | } else { |
ffedd09f | 1040 | prev_page->index = (unsigned long)page; |
61989a80 | 1041 | } |
61989a80 NG |
1042 | prev_page = page; |
1043 | } | |
bdb0af7c | 1044 | } |
61989a80 | 1045 | |
bdb0af7c MK |
1046 | /* |
1047 | * Allocate a zspage for the given size class | |
1048 | */ | |
3783689a MK |
1049 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
1050 | struct size_class *class, | |
1051 | gfp_t gfp) | |
bdb0af7c MK |
1052 | { |
1053 | int i; | |
bdb0af7c | 1054 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1055 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1056 | ||
1057 | if (!zspage) | |
1058 | return NULL; | |
1059 | ||
48b4800a MK |
1060 | zspage->magic = ZSPAGE_MAGIC; |
1061 | migrate_lock_init(zspage); | |
61989a80 | 1062 | |
bdb0af7c MK |
1063 | for (i = 0; i < class->pages_per_zspage; i++) { |
1064 | struct page *page; | |
61989a80 | 1065 | |
3783689a | 1066 | page = alloc_page(gfp); |
bdb0af7c | 1067 | if (!page) { |
91537fee MK |
1068 | while (--i >= 0) { |
1069 | dec_zone_page_state(pages[i], NR_ZSPAGES); | |
bdb0af7c | 1070 | __free_page(pages[i]); |
91537fee | 1071 | } |
3783689a | 1072 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1073 | return NULL; |
1074 | } | |
91537fee MK |
1075 | |
1076 | inc_zone_page_state(page, NR_ZSPAGES); | |
bdb0af7c | 1077 | pages[i] = page; |
61989a80 NG |
1078 | } |
1079 | ||
48b4800a | 1080 | create_page_chain(class, zspage, pages); |
3783689a | 1081 | init_zspage(class, zspage); |
bdb0af7c | 1082 | |
3783689a | 1083 | return zspage; |
61989a80 NG |
1084 | } |
1085 | ||
3783689a | 1086 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1087 | { |
1088 | int i; | |
3783689a | 1089 | struct zspage *zspage; |
61989a80 | 1090 | |
48b4800a | 1091 | for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) { |
3783689a MK |
1092 | zspage = list_first_entry_or_null(&class->fullness_list[i], |
1093 | struct zspage, list); | |
1094 | if (zspage) | |
61989a80 NG |
1095 | break; |
1096 | } | |
1097 | ||
3783689a | 1098 | return zspage; |
61989a80 NG |
1099 | } |
1100 | ||
f553646a SJ |
1101 | static inline int __zs_cpu_up(struct mapping_area *area) |
1102 | { | |
1103 | /* | |
1104 | * Make sure we don't leak memory if a cpu UP notification | |
1105 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1106 | */ | |
1107 | if (area->vm_buf) | |
1108 | return 0; | |
40f9fb8c | 1109 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1110 | if (!area->vm_buf) |
1111 | return -ENOMEM; | |
1112 | return 0; | |
1113 | } | |
1114 | ||
1115 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1116 | { | |
40f9fb8c | 1117 | kfree(area->vm_buf); |
f553646a SJ |
1118 | area->vm_buf = NULL; |
1119 | } | |
1120 | ||
1121 | static void *__zs_map_object(struct mapping_area *area, | |
1122 | struct page *pages[2], int off, int size) | |
5f601902 | 1123 | { |
5f601902 SJ |
1124 | int sizes[2]; |
1125 | void *addr; | |
f553646a | 1126 | char *buf = area->vm_buf; |
5f601902 | 1127 | |
f553646a SJ |
1128 | /* disable page faults to match kmap_atomic() return conditions */ |
1129 | pagefault_disable(); | |
1130 | ||
1131 | /* no read fastpath */ | |
1132 | if (area->vm_mm == ZS_MM_WO) | |
1133 | goto out; | |
5f601902 SJ |
1134 | |
1135 | sizes[0] = PAGE_SIZE - off; | |
1136 | sizes[1] = size - sizes[0]; | |
1137 | ||
5f601902 SJ |
1138 | /* copy object to per-cpu buffer */ |
1139 | addr = kmap_atomic(pages[0]); | |
1140 | memcpy(buf, addr + off, sizes[0]); | |
1141 | kunmap_atomic(addr); | |
1142 | addr = kmap_atomic(pages[1]); | |
1143 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1144 | kunmap_atomic(addr); | |
f553646a SJ |
1145 | out: |
1146 | return area->vm_buf; | |
5f601902 SJ |
1147 | } |
1148 | ||
f553646a SJ |
1149 | static void __zs_unmap_object(struct mapping_area *area, |
1150 | struct page *pages[2], int off, int size) | |
5f601902 | 1151 | { |
5f601902 SJ |
1152 | int sizes[2]; |
1153 | void *addr; | |
2e40e163 | 1154 | char *buf; |
5f601902 | 1155 | |
f553646a SJ |
1156 | /* no write fastpath */ |
1157 | if (area->vm_mm == ZS_MM_RO) | |
1158 | goto out; | |
5f601902 | 1159 | |
7b60a685 | 1160 | buf = area->vm_buf; |
a82cbf07 YX |
1161 | buf = buf + ZS_HANDLE_SIZE; |
1162 | size -= ZS_HANDLE_SIZE; | |
1163 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1164 | |
5f601902 SJ |
1165 | sizes[0] = PAGE_SIZE - off; |
1166 | sizes[1] = size - sizes[0]; | |
1167 | ||
1168 | /* copy per-cpu buffer to object */ | |
1169 | addr = kmap_atomic(pages[0]); | |
1170 | memcpy(addr + off, buf, sizes[0]); | |
1171 | kunmap_atomic(addr); | |
1172 | addr = kmap_atomic(pages[1]); | |
1173 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1174 | kunmap_atomic(addr); | |
f553646a SJ |
1175 | |
1176 | out: | |
1177 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1178 | pagefault_enable(); | |
5f601902 | 1179 | } |
61989a80 | 1180 | |
215c89d0 | 1181 | static int zs_cpu_prepare(unsigned int cpu) |
61989a80 | 1182 | { |
61989a80 NG |
1183 | struct mapping_area *area; |
1184 | ||
215c89d0 SAS |
1185 | area = &per_cpu(zs_map_area, cpu); |
1186 | return __zs_cpu_up(area); | |
61989a80 NG |
1187 | } |
1188 | ||
215c89d0 | 1189 | static int zs_cpu_dead(unsigned int cpu) |
61989a80 | 1190 | { |
215c89d0 | 1191 | struct mapping_area *area; |
40f9fb8c | 1192 | |
215c89d0 SAS |
1193 | area = &per_cpu(zs_map_area, cpu); |
1194 | __zs_cpu_down(area); | |
1195 | return 0; | |
b1b00a5b SS |
1196 | } |
1197 | ||
64d90465 GM |
1198 | static bool can_merge(struct size_class *prev, int pages_per_zspage, |
1199 | int objs_per_zspage) | |
9eec4cd5 | 1200 | { |
64d90465 GM |
1201 | if (prev->pages_per_zspage == pages_per_zspage && |
1202 | prev->objs_per_zspage == objs_per_zspage) | |
1203 | return true; | |
9eec4cd5 | 1204 | |
64d90465 | 1205 | return false; |
9eec4cd5 JK |
1206 | } |
1207 | ||
3783689a | 1208 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1209 | { |
3783689a | 1210 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1211 | } |
1212 | ||
66cdef66 GM |
1213 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1214 | { | |
1215 | return atomic_long_read(&pool->pages_allocated); | |
1216 | } | |
1217 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1218 | ||
4bbc0bc0 | 1219 | /** |
66cdef66 GM |
1220 | * zs_map_object - get address of allocated object from handle. |
1221 | * @pool: pool from which the object was allocated | |
1222 | * @handle: handle returned from zs_malloc | |
f0953a1b | 1223 | * @mm: mapping mode to use |
4bbc0bc0 | 1224 | * |
66cdef66 GM |
1225 | * Before using an object allocated from zs_malloc, it must be mapped using |
1226 | * this function. When done with the object, it must be unmapped using | |
1227 | * zs_unmap_object. | |
4bbc0bc0 | 1228 | * |
66cdef66 GM |
1229 | * Only one object can be mapped per cpu at a time. There is no protection |
1230 | * against nested mappings. | |
1231 | * | |
1232 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1233 | */ |
66cdef66 GM |
1234 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1235 | enum zs_mapmode mm) | |
61989a80 | 1236 | { |
3783689a | 1237 | struct zspage *zspage; |
66cdef66 | 1238 | struct page *page; |
bfd093f5 MK |
1239 | unsigned long obj, off; |
1240 | unsigned int obj_idx; | |
61989a80 | 1241 | |
66cdef66 GM |
1242 | struct size_class *class; |
1243 | struct mapping_area *area; | |
1244 | struct page *pages[2]; | |
2e40e163 | 1245 | void *ret; |
61989a80 | 1246 | |
9eec4cd5 | 1247 | /* |
66cdef66 GM |
1248 | * Because we use per-cpu mapping areas shared among the |
1249 | * pools/users, we can't allow mapping in interrupt context | |
1250 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1251 | */ |
1aedcafb | 1252 | BUG_ON(in_interrupt()); |
61989a80 | 1253 | |
b475d42d MK |
1254 | /* It guarantees it can get zspage from handle safely */ |
1255 | read_lock(&pool->migrate_lock); | |
2e40e163 MK |
1256 | obj = handle_to_obj(handle); |
1257 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1258 | zspage = get_zspage(page); |
48b4800a | 1259 | |
b475d42d MK |
1260 | /* |
1261 | * migration cannot move any zpages in this zspage. Here, class->lock | |
1262 | * is too heavy since callers would take some time until they calls | |
1263 | * zs_unmap_object API so delegate the locking from class to zspage | |
1264 | * which is smaller granularity. | |
1265 | */ | |
48b4800a | 1266 | migrate_read_lock(zspage); |
b475d42d | 1267 | read_unlock(&pool->migrate_lock); |
48b4800a | 1268 | |
67f1c9cd | 1269 | class = zspage_class(pool, zspage); |
bfd093f5 | 1270 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1271 | |
66cdef66 GM |
1272 | area = &get_cpu_var(zs_map_area); |
1273 | area->vm_mm = mm; | |
1274 | if (off + class->size <= PAGE_SIZE) { | |
1275 | /* this object is contained entirely within a page */ | |
1276 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1277 | ret = area->vm_addr + off; |
1278 | goto out; | |
61989a80 NG |
1279 | } |
1280 | ||
66cdef66 GM |
1281 | /* this object spans two pages */ |
1282 | pages[0] = page; | |
1283 | pages[1] = get_next_page(page); | |
1284 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1285 | |
2e40e163 MK |
1286 | ret = __zs_map_object(area, pages, off, class->size); |
1287 | out: | |
a41ec880 | 1288 | if (likely(!ZsHugePage(zspage))) |
7b60a685 MK |
1289 | ret += ZS_HANDLE_SIZE; |
1290 | ||
1291 | return ret; | |
61989a80 | 1292 | } |
66cdef66 | 1293 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1294 | |
66cdef66 | 1295 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1296 | { |
3783689a | 1297 | struct zspage *zspage; |
66cdef66 | 1298 | struct page *page; |
bfd093f5 MK |
1299 | unsigned long obj, off; |
1300 | unsigned int obj_idx; | |
61989a80 | 1301 | |
66cdef66 GM |
1302 | struct size_class *class; |
1303 | struct mapping_area *area; | |
9eec4cd5 | 1304 | |
2e40e163 MK |
1305 | obj = handle_to_obj(handle); |
1306 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1307 | zspage = get_zspage(page); |
67f1c9cd | 1308 | class = zspage_class(pool, zspage); |
bfd093f5 | 1309 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1310 | |
66cdef66 GM |
1311 | area = this_cpu_ptr(&zs_map_area); |
1312 | if (off + class->size <= PAGE_SIZE) | |
1313 | kunmap_atomic(area->vm_addr); | |
1314 | else { | |
1315 | struct page *pages[2]; | |
40f9fb8c | 1316 | |
66cdef66 GM |
1317 | pages[0] = page; |
1318 | pages[1] = get_next_page(page); | |
1319 | BUG_ON(!pages[1]); | |
1320 | ||
1321 | __zs_unmap_object(area, pages, off, class->size); | |
1322 | } | |
1323 | put_cpu_var(zs_map_area); | |
48b4800a MK |
1324 | |
1325 | migrate_read_unlock(zspage); | |
61989a80 | 1326 | } |
66cdef66 | 1327 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1328 | |
010b495e SS |
1329 | /** |
1330 | * zs_huge_class_size() - Returns the size (in bytes) of the first huge | |
1331 | * zsmalloc &size_class. | |
1332 | * @pool: zsmalloc pool to use | |
1333 | * | |
1334 | * The function returns the size of the first huge class - any object of equal | |
1335 | * or bigger size will be stored in zspage consisting of a single physical | |
1336 | * page. | |
1337 | * | |
1338 | * Context: Any context. | |
1339 | * | |
1340 | * Return: the size (in bytes) of the first huge zsmalloc &size_class. | |
1341 | */ | |
1342 | size_t zs_huge_class_size(struct zs_pool *pool) | |
1343 | { | |
1344 | return huge_class_size; | |
1345 | } | |
1346 | EXPORT_SYMBOL_GPL(zs_huge_class_size); | |
1347 | ||
0a5f079b | 1348 | static unsigned long obj_malloc(struct zs_pool *pool, |
3783689a | 1349 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1350 | { |
bfd093f5 | 1351 | int i, nr_page, offset; |
c7806261 MK |
1352 | unsigned long obj; |
1353 | struct link_free *link; | |
0a5f079b | 1354 | struct size_class *class; |
c7806261 MK |
1355 | |
1356 | struct page *m_page; | |
bfd093f5 | 1357 | unsigned long m_offset; |
c7806261 MK |
1358 | void *vaddr; |
1359 | ||
0a5f079b | 1360 | class = pool->size_class[zspage->class]; |
312fcae2 | 1361 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1362 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1363 | |
1364 | offset = obj * class->size; | |
1365 | nr_page = offset >> PAGE_SHIFT; | |
1366 | m_offset = offset & ~PAGE_MASK; | |
1367 | m_page = get_first_page(zspage); | |
1368 | ||
1369 | for (i = 0; i < nr_page; i++) | |
1370 | m_page = get_next_page(m_page); | |
c7806261 MK |
1371 | |
1372 | vaddr = kmap_atomic(m_page); | |
1373 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3b1d9ca6 | 1374 | set_freeobj(zspage, link->next >> OBJ_TAG_BITS); |
a41ec880 | 1375 | if (likely(!ZsHugePage(zspage))) |
7b60a685 MK |
1376 | /* record handle in the header of allocated chunk */ |
1377 | link->handle = handle; | |
1378 | else | |
3783689a MK |
1379 | /* record handle to page->index */ |
1380 | zspage->first_page->index = handle; | |
1381 | ||
c7806261 | 1382 | kunmap_atomic(vaddr); |
3783689a | 1383 | mod_zspage_inuse(zspage, 1); |
c7806261 | 1384 | |
bfd093f5 MK |
1385 | obj = location_to_obj(m_page, obj); |
1386 | ||
c7806261 MK |
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 | |
fd854463 | 1395 | * @gfp: gfp flags when allocating object |
61989a80 | 1396 | * |
00a61d86 | 1397 | * On success, handle to the allocated object is returned, |
c2344348 | 1398 | * otherwise 0. |
61989a80 NG |
1399 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1400 | */ | |
d0d8da2d | 1401 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1402 | { |
2e40e163 | 1403 | unsigned long handle, obj; |
61989a80 | 1404 | struct size_class *class; |
48b4800a | 1405 | enum fullness_group newfg; |
3783689a | 1406 | struct zspage *zspage; |
61989a80 | 1407 | |
7b60a685 | 1408 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1409 | return 0; |
1410 | ||
3783689a | 1411 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1412 | if (!handle) |
c2344348 | 1413 | return 0; |
61989a80 | 1414 | |
2e40e163 MK |
1415 | /* extra space in chunk to keep the handle */ |
1416 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1417 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 | 1418 | |
b475d42d | 1419 | /* class->lock effectively protects the zpage migration */ |
61989a80 | 1420 | spin_lock(&class->lock); |
3783689a | 1421 | zspage = find_get_zspage(class); |
48b4800a | 1422 | if (likely(zspage)) { |
0a5f079b | 1423 | obj = obj_malloc(pool, zspage, handle); |
48b4800a MK |
1424 | /* Now move the zspage to another fullness group, if required */ |
1425 | fix_fullness_group(class, zspage); | |
1426 | record_obj(handle, obj); | |
0a5f079b | 1427 | class_stat_inc(class, OBJ_USED, 1); |
61989a80 | 1428 | spin_unlock(&class->lock); |
61989a80 | 1429 | |
48b4800a MK |
1430 | return handle; |
1431 | } | |
0f050d99 | 1432 | |
48b4800a MK |
1433 | spin_unlock(&class->lock); |
1434 | ||
1435 | zspage = alloc_zspage(pool, class, gfp); | |
1436 | if (!zspage) { | |
1437 | cache_free_handle(pool, handle); | |
1438 | return 0; | |
61989a80 NG |
1439 | } |
1440 | ||
48b4800a | 1441 | spin_lock(&class->lock); |
0a5f079b | 1442 | obj = obj_malloc(pool, zspage, handle); |
48b4800a MK |
1443 | newfg = get_fullness_group(class, zspage); |
1444 | insert_zspage(class, zspage, newfg); | |
1445 | set_zspage_mapping(zspage, class->index, newfg); | |
2e40e163 | 1446 | record_obj(handle, obj); |
48b4800a MK |
1447 | atomic_long_add(class->pages_per_zspage, |
1448 | &pool->pages_allocated); | |
3828a764 | 1449 | class_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage); |
0a5f079b | 1450 | class_stat_inc(class, OBJ_USED, 1); |
48b4800a MK |
1451 | |
1452 | /* We completely set up zspage so mark them as movable */ | |
1453 | SetZsPageMovable(pool, zspage); | |
61989a80 NG |
1454 | spin_unlock(&class->lock); |
1455 | ||
2e40e163 | 1456 | return handle; |
61989a80 NG |
1457 | } |
1458 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1459 | ||
0a5f079b | 1460 | static void obj_free(int class_size, unsigned long obj) |
61989a80 NG |
1461 | { |
1462 | struct link_free *link; | |
3783689a MK |
1463 | struct zspage *zspage; |
1464 | struct page *f_page; | |
bfd093f5 MK |
1465 | unsigned long f_offset; |
1466 | unsigned int f_objidx; | |
af4ee5e9 | 1467 | void *vaddr; |
61989a80 | 1468 | |
2e40e163 | 1469 | obj_to_location(obj, &f_page, &f_objidx); |
0a5f079b | 1470 | f_offset = (class_size * f_objidx) & ~PAGE_MASK; |
3783689a | 1471 | zspage = get_zspage(f_page); |
61989a80 | 1472 | |
c7806261 | 1473 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1474 | |
1475 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1476 | link = (struct link_free *)(vaddr + f_offset); |
a41ec880 MK |
1477 | if (likely(!ZsHugePage(zspage))) |
1478 | link->next = get_freeobj(zspage) << OBJ_TAG_BITS; | |
1479 | else | |
1480 | f_page->index = 0; | |
af4ee5e9 | 1481 | kunmap_atomic(vaddr); |
bfd093f5 | 1482 | set_freeobj(zspage, f_objidx); |
3783689a | 1483 | mod_zspage_inuse(zspage, -1); |
c7806261 MK |
1484 | } |
1485 | ||
1486 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1487 | { | |
3783689a MK |
1488 | struct zspage *zspage; |
1489 | struct page *f_page; | |
bfd093f5 | 1490 | unsigned long obj; |
c7806261 MK |
1491 | struct size_class *class; |
1492 | enum fullness_group fullness; | |
1493 | ||
1494 | if (unlikely(!handle)) | |
1495 | return; | |
1496 | ||
b475d42d MK |
1497 | /* |
1498 | * The pool->migrate_lock protects the race with zpage's migration | |
1499 | * so it's safe to get the page from handle. | |
1500 | */ | |
1501 | read_lock(&pool->migrate_lock); | |
c7806261 | 1502 | obj = handle_to_obj(handle); |
67f1c9cd | 1503 | obj_to_page(obj, &f_page); |
3783689a | 1504 | zspage = get_zspage(f_page); |
67f1c9cd | 1505 | class = zspage_class(pool, zspage); |
c7806261 | 1506 | spin_lock(&class->lock); |
b475d42d MK |
1507 | read_unlock(&pool->migrate_lock); |
1508 | ||
0a5f079b MK |
1509 | obj_free(class->size, obj); |
1510 | class_stat_dec(class, OBJ_USED, 1); | |
3783689a | 1511 | fullness = fix_fullness_group(class, zspage); |
b475d42d | 1512 | if (fullness != ZS_EMPTY) |
48b4800a | 1513 | goto out; |
48b4800a | 1514 | |
c4549b87 | 1515 | free_zspage(pool, class, zspage); |
48b4800a | 1516 | out: |
61989a80 | 1517 | spin_unlock(&class->lock); |
3783689a | 1518 | cache_free_handle(pool, handle); |
312fcae2 MK |
1519 | } |
1520 | EXPORT_SYMBOL_GPL(zs_free); | |
1521 | ||
251cbb95 MK |
1522 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1523 | unsigned long src) | |
312fcae2 MK |
1524 | { |
1525 | struct page *s_page, *d_page; | |
bfd093f5 | 1526 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1527 | unsigned long s_off, d_off; |
1528 | void *s_addr, *d_addr; | |
1529 | int s_size, d_size, size; | |
1530 | int written = 0; | |
1531 | ||
1532 | s_size = d_size = class->size; | |
1533 | ||
1534 | obj_to_location(src, &s_page, &s_objidx); | |
1535 | obj_to_location(dst, &d_page, &d_objidx); | |
1536 | ||
bfd093f5 MK |
1537 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1538 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1539 | |
1540 | if (s_off + class->size > PAGE_SIZE) | |
1541 | s_size = PAGE_SIZE - s_off; | |
1542 | ||
1543 | if (d_off + class->size > PAGE_SIZE) | |
1544 | d_size = PAGE_SIZE - d_off; | |
1545 | ||
1546 | s_addr = kmap_atomic(s_page); | |
1547 | d_addr = kmap_atomic(d_page); | |
1548 | ||
1549 | while (1) { | |
1550 | size = min(s_size, d_size); | |
1551 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1552 | written += size; | |
1553 | ||
1554 | if (written == class->size) | |
1555 | break; | |
1556 | ||
495819ea SS |
1557 | s_off += size; |
1558 | s_size -= size; | |
1559 | d_off += size; | |
1560 | d_size -= size; | |
1561 | ||
1562 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1563 | kunmap_atomic(d_addr); |
1564 | kunmap_atomic(s_addr); | |
1565 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1566 | s_addr = kmap_atomic(s_page); |
1567 | d_addr = kmap_atomic(d_page); | |
1568 | s_size = class->size - written; | |
1569 | s_off = 0; | |
312fcae2 MK |
1570 | } |
1571 | ||
495819ea | 1572 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1573 | kunmap_atomic(d_addr); |
1574 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1575 | d_addr = kmap_atomic(d_page); |
1576 | d_size = class->size - written; | |
1577 | d_off = 0; | |
312fcae2 MK |
1578 | } |
1579 | } | |
1580 | ||
1581 | kunmap_atomic(d_addr); | |
1582 | kunmap_atomic(s_addr); | |
1583 | } | |
1584 | ||
1585 | /* | |
1586 | * Find alloced object in zspage from index object and | |
1587 | * return handle. | |
1588 | */ | |
251cbb95 | 1589 | static unsigned long find_alloced_obj(struct size_class *class, |
cf675acb | 1590 | struct page *page, int *obj_idx) |
312fcae2 | 1591 | { |
312fcae2 | 1592 | int offset = 0; |
cf675acb | 1593 | int index = *obj_idx; |
312fcae2 MK |
1594 | unsigned long handle = 0; |
1595 | void *addr = kmap_atomic(page); | |
1596 | ||
3783689a | 1597 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1598 | offset += class->size * index; |
1599 | ||
1600 | while (offset < PAGE_SIZE) { | |
b475d42d MK |
1601 | if (obj_allocated(page, addr + offset, &handle)) |
1602 | break; | |
312fcae2 MK |
1603 | |
1604 | offset += class->size; | |
1605 | index++; | |
1606 | } | |
1607 | ||
1608 | kunmap_atomic(addr); | |
cf675acb GM |
1609 | |
1610 | *obj_idx = index; | |
1611 | ||
312fcae2 MK |
1612 | return handle; |
1613 | } | |
1614 | ||
1615 | struct zs_compact_control { | |
3783689a | 1616 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1617 | struct page *s_page; |
1618 | /* Destination page for migration which should be a first page | |
1619 | * of zspage. */ | |
1620 | struct page *d_page; | |
1621 | /* Starting object index within @s_page which used for live object | |
1622 | * in the subpage. */ | |
41b88e14 | 1623 | int obj_idx; |
312fcae2 MK |
1624 | }; |
1625 | ||
1626 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1627 | struct zs_compact_control *cc) | |
1628 | { | |
1629 | unsigned long used_obj, free_obj; | |
1630 | unsigned long handle; | |
1631 | struct page *s_page = cc->s_page; | |
1632 | struct page *d_page = cc->d_page; | |
41b88e14 | 1633 | int obj_idx = cc->obj_idx; |
312fcae2 MK |
1634 | int ret = 0; |
1635 | ||
1636 | while (1) { | |
cf675acb | 1637 | handle = find_alloced_obj(class, s_page, &obj_idx); |
312fcae2 MK |
1638 | if (!handle) { |
1639 | s_page = get_next_page(s_page); | |
1640 | if (!s_page) | |
1641 | break; | |
41b88e14 | 1642 | obj_idx = 0; |
312fcae2 MK |
1643 | continue; |
1644 | } | |
1645 | ||
1646 | /* Stop if there is no more space */ | |
3783689a | 1647 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1648 | ret = -ENOMEM; |
1649 | break; | |
1650 | } | |
1651 | ||
1652 | used_obj = handle_to_obj(handle); | |
0a5f079b | 1653 | free_obj = obj_malloc(pool, get_zspage(d_page), handle); |
251cbb95 | 1654 | zs_object_copy(class, free_obj, used_obj); |
41b88e14 | 1655 | obj_idx++; |
312fcae2 | 1656 | record_obj(handle, free_obj); |
0a5f079b | 1657 | obj_free(class->size, used_obj); |
312fcae2 MK |
1658 | } |
1659 | ||
1660 | /* Remember last position in this iteration */ | |
1661 | cc->s_page = s_page; | |
41b88e14 | 1662 | cc->obj_idx = obj_idx; |
312fcae2 MK |
1663 | |
1664 | return ret; | |
1665 | } | |
1666 | ||
3783689a | 1667 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1668 | { |
1669 | int i; | |
3783689a MK |
1670 | struct zspage *zspage; |
1671 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1672 | |
3783689a MK |
1673 | if (!source) { |
1674 | fg[0] = ZS_ALMOST_FULL; | |
1675 | fg[1] = ZS_ALMOST_EMPTY; | |
1676 | } | |
1677 | ||
1678 | for (i = 0; i < 2; i++) { | |
1679 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1680 | struct zspage, list); | |
1681 | if (zspage) { | |
1682 | remove_zspage(class, zspage, fg[i]); | |
1683 | return zspage; | |
312fcae2 MK |
1684 | } |
1685 | } | |
1686 | ||
3783689a | 1687 | return zspage; |
312fcae2 MK |
1688 | } |
1689 | ||
860c707d | 1690 | /* |
3783689a | 1691 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1692 | * @class: destination class |
3783689a | 1693 | * @zspage: target page |
860c707d | 1694 | * |
3783689a | 1695 | * Return @zspage's fullness_group |
860c707d | 1696 | */ |
4aa409ca | 1697 | static enum fullness_group putback_zspage(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 | |
860c707d | 1706 | return fullness; |
61989a80 | 1707 | } |
312fcae2 | 1708 | |
48b4800a | 1709 | #ifdef CONFIG_COMPACTION |
4d0a5402 CIK |
1710 | /* |
1711 | * To prevent zspage destroy during migration, zspage freeing should | |
1712 | * hold locks of all pages in the zspage. | |
1713 | */ | |
1714 | static void lock_zspage(struct zspage *zspage) | |
1715 | { | |
1716 | struct page *page = get_first_page(zspage); | |
1717 | ||
1718 | do { | |
1719 | lock_page(page); | |
1720 | } while ((page = get_next_page(page)) != NULL); | |
1721 | } | |
1722 | ||
8e9231f8 | 1723 | static int zs_init_fs_context(struct fs_context *fc) |
48b4800a | 1724 | { |
8e9231f8 | 1725 | return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM; |
48b4800a MK |
1726 | } |
1727 | ||
1728 | static struct file_system_type zsmalloc_fs = { | |
1729 | .name = "zsmalloc", | |
8e9231f8 | 1730 | .init_fs_context = zs_init_fs_context, |
48b4800a MK |
1731 | .kill_sb = kill_anon_super, |
1732 | }; | |
1733 | ||
1734 | static int zsmalloc_mount(void) | |
1735 | { | |
1736 | int ret = 0; | |
1737 | ||
1738 | zsmalloc_mnt = kern_mount(&zsmalloc_fs); | |
1739 | if (IS_ERR(zsmalloc_mnt)) | |
1740 | ret = PTR_ERR(zsmalloc_mnt); | |
1741 | ||
1742 | return ret; | |
1743 | } | |
1744 | ||
1745 | static void zsmalloc_unmount(void) | |
1746 | { | |
1747 | kern_unmount(zsmalloc_mnt); | |
1748 | } | |
1749 | ||
1750 | static void migrate_lock_init(struct zspage *zspage) | |
1751 | { | |
1752 | rwlock_init(&zspage->lock); | |
1753 | } | |
1754 | ||
cfc451cf | 1755 | static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock) |
48b4800a MK |
1756 | { |
1757 | read_lock(&zspage->lock); | |
1758 | } | |
1759 | ||
8a374ccc | 1760 | static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock) |
48b4800a MK |
1761 | { |
1762 | read_unlock(&zspage->lock); | |
1763 | } | |
1764 | ||
1765 | static void migrate_write_lock(struct zspage *zspage) | |
1766 | { | |
1767 | write_lock(&zspage->lock); | |
1768 | } | |
1769 | ||
b475d42d MK |
1770 | static void migrate_write_lock_nested(struct zspage *zspage) |
1771 | { | |
1772 | write_lock_nested(&zspage->lock, SINGLE_DEPTH_NESTING); | |
1773 | } | |
1774 | ||
48b4800a MK |
1775 | static void migrate_write_unlock(struct zspage *zspage) |
1776 | { | |
1777 | write_unlock(&zspage->lock); | |
1778 | } | |
1779 | ||
1780 | /* Number of isolated subpage for *page migration* in this zspage */ | |
1781 | static void inc_zspage_isolation(struct zspage *zspage) | |
1782 | { | |
1783 | zspage->isolated++; | |
1784 | } | |
1785 | ||
1786 | static void dec_zspage_isolation(struct zspage *zspage) | |
1787 | { | |
c4549b87 | 1788 | VM_BUG_ON(zspage->isolated == 0); |
48b4800a MK |
1789 | zspage->isolated--; |
1790 | } | |
1791 | ||
1792 | static void replace_sub_page(struct size_class *class, struct zspage *zspage, | |
1793 | struct page *newpage, struct page *oldpage) | |
1794 | { | |
1795 | struct page *page; | |
1796 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, }; | |
1797 | int idx = 0; | |
1798 | ||
1799 | page = get_first_page(zspage); | |
1800 | do { | |
1801 | if (page == oldpage) | |
1802 | pages[idx] = newpage; | |
1803 | else | |
1804 | pages[idx] = page; | |
1805 | idx++; | |
1806 | } while ((page = get_next_page(page)) != NULL); | |
1807 | ||
1808 | create_page_chain(class, zspage, pages); | |
1809 | set_first_obj_offset(newpage, get_first_obj_offset(oldpage)); | |
a41ec880 | 1810 | if (unlikely(ZsHugePage(zspage))) |
48b4800a MK |
1811 | newpage->index = oldpage->index; |
1812 | __SetPageMovable(newpage, page_mapping(oldpage)); | |
1813 | } | |
1814 | ||
4d0a5402 | 1815 | static bool zs_page_isolate(struct page *page, isolate_mode_t mode) |
48b4800a | 1816 | { |
48b4800a | 1817 | struct zspage *zspage; |
48b4800a MK |
1818 | |
1819 | /* | |
1820 | * Page is locked so zspage couldn't be destroyed. For detail, look at | |
1821 | * lock_zspage in free_zspage. | |
1822 | */ | |
1823 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1824 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1825 | ||
1826 | zspage = get_zspage(page); | |
c4549b87 | 1827 | migrate_write_lock(zspage); |
48b4800a | 1828 | inc_zspage_isolation(zspage); |
c4549b87 | 1829 | migrate_write_unlock(zspage); |
48b4800a MK |
1830 | |
1831 | return true; | |
1832 | } | |
1833 | ||
4d0a5402 | 1834 | static int zs_page_migrate(struct address_space *mapping, struct page *newpage, |
48b4800a MK |
1835 | struct page *page, enum migrate_mode mode) |
1836 | { | |
1837 | struct zs_pool *pool; | |
1838 | struct size_class *class; | |
48b4800a MK |
1839 | struct zspage *zspage; |
1840 | struct page *dummy; | |
1841 | void *s_addr, *d_addr, *addr; | |
b475d42d | 1842 | int offset; |
3ae92ac2 | 1843 | unsigned long handle; |
48b4800a MK |
1844 | unsigned long old_obj, new_obj; |
1845 | unsigned int obj_idx; | |
48b4800a | 1846 | |
2916ecc0 JG |
1847 | /* |
1848 | * We cannot support the _NO_COPY case here, because copy needs to | |
1849 | * happen under the zs lock, which does not work with | |
1850 | * MIGRATE_SYNC_NO_COPY workflow. | |
1851 | */ | |
1852 | if (mode == MIGRATE_SYNC_NO_COPY) | |
1853 | return -EINVAL; | |
1854 | ||
48b4800a MK |
1855 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
1856 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
1857 | ||
48b4800a | 1858 | pool = mapping->private_data; |
b475d42d MK |
1859 | |
1860 | /* | |
1861 | * The pool migrate_lock protects the race between zpage migration | |
1862 | * and zs_free. | |
1863 | */ | |
1864 | write_lock(&pool->migrate_lock); | |
1865 | zspage = get_zspage(page); | |
67f1c9cd | 1866 | class = zspage_class(pool, zspage); |
48b4800a | 1867 | |
b475d42d MK |
1868 | /* |
1869 | * the class lock protects zpage alloc/free in the zspage. | |
1870 | */ | |
48b4800a | 1871 | spin_lock(&class->lock); |
b475d42d MK |
1872 | /* the migrate_write_lock protects zpage access via zs_map_object */ |
1873 | migrate_write_lock(zspage); | |
48b4800a | 1874 | |
b475d42d | 1875 | offset = get_first_obj_offset(page); |
48b4800a | 1876 | s_addr = kmap_atomic(page); |
48b4800a MK |
1877 | |
1878 | /* | |
1879 | * Here, any user cannot access all objects in the zspage so let's move. | |
1880 | */ | |
1881 | d_addr = kmap_atomic(newpage); | |
1882 | memcpy(d_addr, s_addr, PAGE_SIZE); | |
1883 | kunmap_atomic(d_addr); | |
1884 | ||
b475d42d | 1885 | for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE; |
48b4800a | 1886 | addr += class->size) { |
3ae92ac2 | 1887 | if (obj_allocated(page, addr, &handle)) { |
48b4800a MK |
1888 | |
1889 | old_obj = handle_to_obj(handle); | |
1890 | obj_to_location(old_obj, &dummy, &obj_idx); | |
1891 | new_obj = (unsigned long)location_to_obj(newpage, | |
1892 | obj_idx); | |
48b4800a MK |
1893 | record_obj(handle, new_obj); |
1894 | } | |
1895 | } | |
b475d42d | 1896 | kunmap_atomic(s_addr); |
48b4800a MK |
1897 | |
1898 | replace_sub_page(class, zspage, newpage, page); | |
b475d42d MK |
1899 | /* |
1900 | * Since we complete the data copy and set up new zspage structure, | |
1901 | * it's okay to release migration_lock. | |
1902 | */ | |
1903 | write_unlock(&pool->migrate_lock); | |
1904 | spin_unlock(&class->lock); | |
48b4800a | 1905 | dec_zspage_isolation(zspage); |
b475d42d | 1906 | migrate_write_unlock(zspage); |
48b4800a | 1907 | |
b475d42d | 1908 | get_page(newpage); |
ac8f05da CM |
1909 | if (page_zone(newpage) != page_zone(page)) { |
1910 | dec_zone_page_state(page, NR_ZSPAGES); | |
1911 | inc_zone_page_state(newpage, NR_ZSPAGES); | |
1912 | } | |
1913 | ||
48b4800a MK |
1914 | reset_page(page); |
1915 | put_page(page); | |
48b4800a | 1916 | |
b475d42d | 1917 | return MIGRATEPAGE_SUCCESS; |
48b4800a MK |
1918 | } |
1919 | ||
4d0a5402 | 1920 | static void zs_page_putback(struct page *page) |
48b4800a | 1921 | { |
48b4800a MK |
1922 | struct zspage *zspage; |
1923 | ||
1924 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1925 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
1926 | ||
1927 | zspage = get_zspage(page); | |
c4549b87 | 1928 | migrate_write_lock(zspage); |
48b4800a | 1929 | dec_zspage_isolation(zspage); |
c4549b87 | 1930 | migrate_write_unlock(zspage); |
48b4800a MK |
1931 | } |
1932 | ||
4d0a5402 | 1933 | static const struct address_space_operations zsmalloc_aops = { |
48b4800a MK |
1934 | .isolate_page = zs_page_isolate, |
1935 | .migratepage = zs_page_migrate, | |
1936 | .putback_page = zs_page_putback, | |
1937 | }; | |
1938 | ||
1939 | static int zs_register_migration(struct zs_pool *pool) | |
1940 | { | |
1941 | pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb); | |
1942 | if (IS_ERR(pool->inode)) { | |
1943 | pool->inode = NULL; | |
1944 | return 1; | |
1945 | } | |
1946 | ||
1947 | pool->inode->i_mapping->private_data = pool; | |
1948 | pool->inode->i_mapping->a_ops = &zsmalloc_aops; | |
1949 | return 0; | |
1950 | } | |
1951 | ||
1952 | static void zs_unregister_migration(struct zs_pool *pool) | |
1953 | { | |
1954 | flush_work(&pool->free_work); | |
c3491eca | 1955 | iput(pool->inode); |
48b4800a MK |
1956 | } |
1957 | ||
1958 | /* | |
1959 | * Caller should hold page_lock of all pages in the zspage | |
1960 | * In here, we cannot use zspage meta data. | |
1961 | */ | |
1962 | static void async_free_zspage(struct work_struct *work) | |
1963 | { | |
1964 | int i; | |
1965 | struct size_class *class; | |
1966 | unsigned int class_idx; | |
1967 | enum fullness_group fullness; | |
1968 | struct zspage *zspage, *tmp; | |
1969 | LIST_HEAD(free_pages); | |
1970 | struct zs_pool *pool = container_of(work, struct zs_pool, | |
1971 | free_work); | |
1972 | ||
cf8e0fed | 1973 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
48b4800a MK |
1974 | class = pool->size_class[i]; |
1975 | if (class->index != i) | |
1976 | continue; | |
1977 | ||
1978 | spin_lock(&class->lock); | |
1979 | list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages); | |
1980 | spin_unlock(&class->lock); | |
1981 | } | |
1982 | ||
48b4800a MK |
1983 | list_for_each_entry_safe(zspage, tmp, &free_pages, list) { |
1984 | list_del(&zspage->list); | |
1985 | lock_zspage(zspage); | |
1986 | ||
1987 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1988 | VM_BUG_ON(fullness != ZS_EMPTY); | |
1989 | class = pool->size_class[class_idx]; | |
1990 | spin_lock(&class->lock); | |
33848337 | 1991 | __free_zspage(pool, class, zspage); |
48b4800a MK |
1992 | spin_unlock(&class->lock); |
1993 | } | |
1994 | }; | |
1995 | ||
1996 | static void kick_deferred_free(struct zs_pool *pool) | |
1997 | { | |
1998 | schedule_work(&pool->free_work); | |
1999 | } | |
2000 | ||
2001 | static void init_deferred_free(struct zs_pool *pool) | |
2002 | { | |
2003 | INIT_WORK(&pool->free_work, async_free_zspage); | |
2004 | } | |
2005 | ||
2006 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) | |
2007 | { | |
2008 | struct page *page = get_first_page(zspage); | |
2009 | ||
2010 | do { | |
2011 | WARN_ON(!trylock_page(page)); | |
2012 | __SetPageMovable(page, pool->inode->i_mapping); | |
2013 | unlock_page(page); | |
2014 | } while ((page = get_next_page(page)) != NULL); | |
2015 | } | |
2016 | #endif | |
2017 | ||
04f05909 SS |
2018 | /* |
2019 | * | |
2020 | * Based on the number of unused allocated objects calculate | |
2021 | * and return the number of pages that we can free. | |
04f05909 SS |
2022 | */ |
2023 | static unsigned long zs_can_compact(struct size_class *class) | |
2024 | { | |
2025 | unsigned long obj_wasted; | |
44f43e99 SS |
2026 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
2027 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 2028 | |
44f43e99 SS |
2029 | if (obj_allocated <= obj_used) |
2030 | return 0; | |
04f05909 | 2031 | |
44f43e99 | 2032 | obj_wasted = obj_allocated - obj_used; |
b4fd07a0 | 2033 | obj_wasted /= class->objs_per_zspage; |
04f05909 | 2034 | |
6cbf16b3 | 2035 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
2036 | } |
2037 | ||
23959281 RY |
2038 | static unsigned long __zs_compact(struct zs_pool *pool, |
2039 | struct size_class *class) | |
312fcae2 | 2040 | { |
312fcae2 | 2041 | struct zs_compact_control cc; |
3783689a MK |
2042 | struct zspage *src_zspage; |
2043 | struct zspage *dst_zspage = NULL; | |
23959281 | 2044 | unsigned long pages_freed = 0; |
312fcae2 | 2045 | |
b475d42d MK |
2046 | /* protect the race between zpage migration and zs_free */ |
2047 | write_lock(&pool->migrate_lock); | |
2048 | /* protect zpage allocation/free */ | |
312fcae2 | 2049 | spin_lock(&class->lock); |
3783689a | 2050 | while ((src_zspage = isolate_zspage(class, true))) { |
b475d42d MK |
2051 | /* protect someone accessing the zspage(i.e., zs_map_object) */ |
2052 | migrate_write_lock(src_zspage); | |
312fcae2 | 2053 | |
04f05909 SS |
2054 | if (!zs_can_compact(class)) |
2055 | break; | |
2056 | ||
41b88e14 | 2057 | cc.obj_idx = 0; |
48b4800a | 2058 | cc.s_page = get_first_page(src_zspage); |
312fcae2 | 2059 | |
3783689a | 2060 | while ((dst_zspage = isolate_zspage(class, false))) { |
b475d42d MK |
2061 | migrate_write_lock_nested(dst_zspage); |
2062 | ||
48b4800a | 2063 | cc.d_page = get_first_page(dst_zspage); |
312fcae2 | 2064 | /* |
0dc63d48 SS |
2065 | * If there is no more space in dst_page, resched |
2066 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
2067 | */ |
2068 | if (!migrate_zspage(pool, class, &cc)) | |
2069 | break; | |
2070 | ||
4aa409ca | 2071 | putback_zspage(class, dst_zspage); |
b475d42d MK |
2072 | migrate_write_unlock(dst_zspage); |
2073 | dst_zspage = NULL; | |
2074 | if (rwlock_is_contended(&pool->migrate_lock)) | |
2075 | break; | |
312fcae2 MK |
2076 | } |
2077 | ||
2078 | /* Stop if we couldn't find slot */ | |
3783689a | 2079 | if (dst_zspage == NULL) |
312fcae2 MK |
2080 | break; |
2081 | ||
4aa409ca | 2082 | putback_zspage(class, dst_zspage); |
b475d42d MK |
2083 | migrate_write_unlock(dst_zspage); |
2084 | ||
4aa409ca | 2085 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { |
b475d42d | 2086 | migrate_write_unlock(src_zspage); |
48b4800a | 2087 | free_zspage(pool, class, src_zspage); |
23959281 | 2088 | pages_freed += class->pages_per_zspage; |
b475d42d MK |
2089 | } else |
2090 | migrate_write_unlock(src_zspage); | |
312fcae2 | 2091 | spin_unlock(&class->lock); |
b475d42d | 2092 | write_unlock(&pool->migrate_lock); |
312fcae2 | 2093 | cond_resched(); |
b475d42d | 2094 | write_lock(&pool->migrate_lock); |
312fcae2 MK |
2095 | spin_lock(&class->lock); |
2096 | } | |
2097 | ||
b475d42d | 2098 | if (src_zspage) { |
4aa409ca | 2099 | putback_zspage(class, src_zspage); |
b475d42d MK |
2100 | migrate_write_unlock(src_zspage); |
2101 | } | |
312fcae2 | 2102 | |
7d3f3938 | 2103 | spin_unlock(&class->lock); |
b475d42d | 2104 | write_unlock(&pool->migrate_lock); |
23959281 RY |
2105 | |
2106 | return pages_freed; | |
312fcae2 MK |
2107 | } |
2108 | ||
2109 | unsigned long zs_compact(struct zs_pool *pool) | |
2110 | { | |
2111 | int i; | |
312fcae2 | 2112 | struct size_class *class; |
23959281 | 2113 | unsigned long pages_freed = 0; |
312fcae2 | 2114 | |
cf8e0fed | 2115 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
312fcae2 MK |
2116 | class = pool->size_class[i]; |
2117 | if (!class) | |
2118 | continue; | |
2119 | if (class->index != i) | |
2120 | continue; | |
23959281 | 2121 | pages_freed += __zs_compact(pool, class); |
312fcae2 | 2122 | } |
23959281 | 2123 | atomic_long_add(pages_freed, &pool->stats.pages_compacted); |
312fcae2 | 2124 | |
23959281 | 2125 | return pages_freed; |
312fcae2 MK |
2126 | } |
2127 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 2128 | |
7d3f3938 SS |
2129 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
2130 | { | |
2131 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
2132 | } | |
2133 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
2134 | ||
ab9d306d SS |
2135 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
2136 | struct shrink_control *sc) | |
2137 | { | |
2138 | unsigned long pages_freed; | |
2139 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2140 | shrinker); | |
2141 | ||
ab9d306d SS |
2142 | /* |
2143 | * Compact classes and calculate compaction delta. | |
2144 | * Can run concurrently with a manually triggered | |
2145 | * (by user) compaction. | |
2146 | */ | |
23959281 | 2147 | pages_freed = zs_compact(pool); |
ab9d306d SS |
2148 | |
2149 | return pages_freed ? pages_freed : SHRINK_STOP; | |
2150 | } | |
2151 | ||
2152 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
2153 | struct shrink_control *sc) | |
2154 | { | |
2155 | int i; | |
2156 | struct size_class *class; | |
2157 | unsigned long pages_to_free = 0; | |
2158 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2159 | shrinker); | |
2160 | ||
cf8e0fed | 2161 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
ab9d306d SS |
2162 | class = pool->size_class[i]; |
2163 | if (!class) | |
2164 | continue; | |
2165 | if (class->index != i) | |
2166 | continue; | |
2167 | ||
ab9d306d | 2168 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
2169 | } |
2170 | ||
2171 | return pages_to_free; | |
2172 | } | |
2173 | ||
2174 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
2175 | { | |
93144ca3 | 2176 | unregister_shrinker(&pool->shrinker); |
ab9d306d SS |
2177 | } |
2178 | ||
2179 | static int zs_register_shrinker(struct zs_pool *pool) | |
2180 | { | |
2181 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
2182 | pool->shrinker.count_objects = zs_shrinker_count; | |
2183 | pool->shrinker.batch = 0; | |
2184 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
2185 | ||
2186 | return register_shrinker(&pool->shrinker); | |
2187 | } | |
2188 | ||
00a61d86 | 2189 | /** |
66cdef66 | 2190 | * zs_create_pool - Creates an allocation pool to work from. |
fd854463 | 2191 | * @name: pool name to be created |
166cfda7 | 2192 | * |
66cdef66 GM |
2193 | * This function must be called before anything when using |
2194 | * the zsmalloc allocator. | |
166cfda7 | 2195 | * |
66cdef66 GM |
2196 | * On success, a pointer to the newly created pool is returned, |
2197 | * otherwise NULL. | |
396b7fd6 | 2198 | */ |
d0d8da2d | 2199 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 2200 | { |
66cdef66 GM |
2201 | int i; |
2202 | struct zs_pool *pool; | |
2203 | struct size_class *prev_class = NULL; | |
61989a80 | 2204 | |
66cdef66 GM |
2205 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
2206 | if (!pool) | |
2207 | return NULL; | |
61989a80 | 2208 | |
48b4800a | 2209 | init_deferred_free(pool); |
b475d42d | 2210 | rwlock_init(&pool->migrate_lock); |
61989a80 | 2211 | |
2e40e163 MK |
2212 | pool->name = kstrdup(name, GFP_KERNEL); |
2213 | if (!pool->name) | |
2214 | goto err; | |
2215 | ||
3783689a | 2216 | if (create_cache(pool)) |
2e40e163 MK |
2217 | goto err; |
2218 | ||
c60369f0 | 2219 | /* |
399d8eeb | 2220 | * Iterate reversely, because, size of size_class that we want to use |
66cdef66 | 2221 | * for merging should be larger or equal to current size. |
c60369f0 | 2222 | */ |
cf8e0fed | 2223 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
66cdef66 GM |
2224 | int size; |
2225 | int pages_per_zspage; | |
64d90465 | 2226 | int objs_per_zspage; |
66cdef66 | 2227 | struct size_class *class; |
3783689a | 2228 | int fullness = 0; |
c60369f0 | 2229 | |
66cdef66 GM |
2230 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
2231 | if (size > ZS_MAX_ALLOC_SIZE) | |
2232 | size = ZS_MAX_ALLOC_SIZE; | |
2233 | pages_per_zspage = get_pages_per_zspage(size); | |
64d90465 | 2234 | objs_per_zspage = pages_per_zspage * PAGE_SIZE / size; |
61989a80 | 2235 | |
010b495e SS |
2236 | /* |
2237 | * We iterate from biggest down to smallest classes, | |
2238 | * so huge_class_size holds the size of the first huge | |
2239 | * class. Any object bigger than or equal to that will | |
2240 | * endup in the huge class. | |
2241 | */ | |
2242 | if (pages_per_zspage != 1 && objs_per_zspage != 1 && | |
2243 | !huge_class_size) { | |
2244 | huge_class_size = size; | |
2245 | /* | |
2246 | * The object uses ZS_HANDLE_SIZE bytes to store the | |
2247 | * handle. We need to subtract it, because zs_malloc() | |
2248 | * unconditionally adds handle size before it performs | |
2249 | * size class search - so object may be smaller than | |
2250 | * huge class size, yet it still can end up in the huge | |
2251 | * class because it grows by ZS_HANDLE_SIZE extra bytes | |
2252 | * right before class lookup. | |
2253 | */ | |
2254 | huge_class_size -= (ZS_HANDLE_SIZE - 1); | |
2255 | } | |
2256 | ||
66cdef66 GM |
2257 | /* |
2258 | * size_class is used for normal zsmalloc operation such | |
2259 | * as alloc/free for that size. Although it is natural that we | |
2260 | * have one size_class for each size, there is a chance that we | |
2261 | * can get more memory utilization if we use one size_class for | |
2262 | * many different sizes whose size_class have same | |
2263 | * characteristics. So, we makes size_class point to | |
2264 | * previous size_class if possible. | |
2265 | */ | |
2266 | if (prev_class) { | |
64d90465 | 2267 | if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) { |
66cdef66 GM |
2268 | pool->size_class[i] = prev_class; |
2269 | continue; | |
2270 | } | |
2271 | } | |
2272 | ||
2273 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
2274 | if (!class) | |
2275 | goto err; | |
2276 | ||
2277 | class->size = size; | |
2278 | class->index = i; | |
2279 | class->pages_per_zspage = pages_per_zspage; | |
64d90465 | 2280 | class->objs_per_zspage = objs_per_zspage; |
66cdef66 GM |
2281 | spin_lock_init(&class->lock); |
2282 | pool->size_class[i] = class; | |
48b4800a MK |
2283 | for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS; |
2284 | fullness++) | |
3783689a | 2285 | INIT_LIST_HEAD(&class->fullness_list[fullness]); |
66cdef66 GM |
2286 | |
2287 | prev_class = class; | |
61989a80 NG |
2288 | } |
2289 | ||
d34f6157 DS |
2290 | /* debug only, don't abort if it fails */ |
2291 | zs_pool_stat_create(pool, name); | |
0f050d99 | 2292 | |
48b4800a MK |
2293 | if (zs_register_migration(pool)) |
2294 | goto err; | |
2295 | ||
ab9d306d | 2296 | /* |
93144ca3 AK |
2297 | * Not critical since shrinker is only used to trigger internal |
2298 | * defragmentation of the pool which is pretty optional thing. If | |
2299 | * registration fails we still can use the pool normally and user can | |
2300 | * trigger compaction manually. Thus, ignore return code. | |
ab9d306d | 2301 | */ |
93144ca3 AK |
2302 | zs_register_shrinker(pool); |
2303 | ||
66cdef66 GM |
2304 | return pool; |
2305 | ||
2306 | err: | |
2307 | zs_destroy_pool(pool); | |
2308 | return NULL; | |
61989a80 | 2309 | } |
66cdef66 | 2310 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2311 | |
66cdef66 | 2312 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2313 | { |
66cdef66 | 2314 | int i; |
61989a80 | 2315 | |
ab9d306d | 2316 | zs_unregister_shrinker(pool); |
48b4800a | 2317 | zs_unregister_migration(pool); |
0f050d99 GM |
2318 | zs_pool_stat_destroy(pool); |
2319 | ||
cf8e0fed | 2320 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
66cdef66 GM |
2321 | int fg; |
2322 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2323 | |
66cdef66 GM |
2324 | if (!class) |
2325 | continue; | |
61989a80 | 2326 | |
66cdef66 GM |
2327 | if (class->index != i) |
2328 | continue; | |
61989a80 | 2329 | |
48b4800a | 2330 | for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) { |
3783689a | 2331 | if (!list_empty(&class->fullness_list[fg])) { |
66cdef66 GM |
2332 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
2333 | class->size, fg); | |
2334 | } | |
2335 | } | |
2336 | kfree(class); | |
2337 | } | |
f553646a | 2338 | |
3783689a | 2339 | destroy_cache(pool); |
0f050d99 | 2340 | kfree(pool->name); |
66cdef66 GM |
2341 | kfree(pool); |
2342 | } | |
2343 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2344 | |
66cdef66 GM |
2345 | static int __init zs_init(void) |
2346 | { | |
48b4800a MK |
2347 | int ret; |
2348 | ||
2349 | ret = zsmalloc_mount(); | |
2350 | if (ret) | |
2351 | goto out; | |
2352 | ||
215c89d0 SAS |
2353 | ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare", |
2354 | zs_cpu_prepare, zs_cpu_dead); | |
0f050d99 | 2355 | if (ret) |
215c89d0 | 2356 | goto hp_setup_fail; |
66cdef66 | 2357 | |
66cdef66 GM |
2358 | #ifdef CONFIG_ZPOOL |
2359 | zpool_register_driver(&zs_zpool_driver); | |
2360 | #endif | |
0f050d99 | 2361 | |
4abaac9b DS |
2362 | zs_stat_init(); |
2363 | ||
66cdef66 | 2364 | return 0; |
0f050d99 | 2365 | |
215c89d0 | 2366 | hp_setup_fail: |
48b4800a MK |
2367 | zsmalloc_unmount(); |
2368 | out: | |
0f050d99 | 2369 | return ret; |
61989a80 | 2370 | } |
61989a80 | 2371 | |
66cdef66 | 2372 | static void __exit zs_exit(void) |
61989a80 | 2373 | { |
66cdef66 GM |
2374 | #ifdef CONFIG_ZPOOL |
2375 | zpool_unregister_driver(&zs_zpool_driver); | |
2376 | #endif | |
48b4800a | 2377 | zsmalloc_unmount(); |
215c89d0 | 2378 | cpuhp_remove_state(CPUHP_MM_ZS_PREPARE); |
0f050d99 GM |
2379 | |
2380 | zs_stat_exit(); | |
61989a80 | 2381 | } |
069f101f BH |
2382 | |
2383 | module_init(zs_init); | |
2384 | module_exit(zs_exit); | |
2385 | ||
2386 | MODULE_LICENSE("Dual BSD/GPL"); | |
2387 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |