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 | ||
61989a80 NG |
33 | #include <linux/module.h> |
34 | #include <linux/kernel.h> | |
312fcae2 | 35 | #include <linux/sched.h> |
50d34394 | 36 | #include <linux/magic.h> |
61989a80 NG |
37 | #include <linux/bitops.h> |
38 | #include <linux/errno.h> | |
39 | #include <linux/highmem.h> | |
61989a80 NG |
40 | #include <linux/string.h> |
41 | #include <linux/slab.h> | |
ca5999fd | 42 | #include <linux/pgtable.h> |
65fddcfc | 43 | #include <asm/tlbflush.h> |
61989a80 NG |
44 | #include <linux/cpumask.h> |
45 | #include <linux/cpu.h> | |
0cbb613f | 46 | #include <linux/vmalloc.h> |
759b26b2 | 47 | #include <linux/preempt.h> |
0959c63f | 48 | #include <linux/spinlock.h> |
93144ca3 | 49 | #include <linux/shrinker.h> |
0959c63f | 50 | #include <linux/types.h> |
0f050d99 | 51 | #include <linux/debugfs.h> |
bcf1647d | 52 | #include <linux/zsmalloc.h> |
c795779d | 53 | #include <linux/zpool.h> |
48b4800a | 54 | #include <linux/mount.h> |
8e9231f8 | 55 | #include <linux/pseudo_fs.h> |
dd4123f3 | 56 | #include <linux/migrate.h> |
701d6785 | 57 | #include <linux/wait.h> |
48b4800a | 58 | #include <linux/pagemap.h> |
cdc346b3 | 59 | #include <linux/fs.h> |
48b4800a MK |
60 | |
61 | #define ZSPAGE_MAGIC 0x58 | |
0959c63f SJ |
62 | |
63 | /* | |
cb152a1a | 64 | * This must be power of 2 and greater than or equal to sizeof(link_free). |
0959c63f SJ |
65 | * These two conditions ensure that any 'struct link_free' itself doesn't |
66 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
67 | * to restore link_free pointer values. | |
68 | */ | |
69 | #define ZS_ALIGN 8 | |
70 | ||
71 | /* | |
72 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
73 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
74 | */ | |
75 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
76 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
77 | ||
2e40e163 MK |
78 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
79 | ||
0959c63f SJ |
80 | /* |
81 | * Object location (<PFN>, <obj_idx>) is encoded as | |
b956b5ac | 82 | * a single (unsigned long) handle value. |
0959c63f | 83 | * |
bfd093f5 | 84 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
85 | * |
86 | * This is made more complicated by various memory models and PAE. | |
87 | */ | |
88 | ||
02390b87 KS |
89 | #ifndef MAX_POSSIBLE_PHYSMEM_BITS |
90 | #ifdef MAX_PHYSMEM_BITS | |
91 | #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS | |
92 | #else | |
0959c63f SJ |
93 | /* |
94 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
95 | * be PAGE_SHIFT | |
96 | */ | |
02390b87 | 97 | #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG |
0959c63f SJ |
98 | #endif |
99 | #endif | |
02390b87 KS |
100 | |
101 | #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
102 | |
103 | /* | |
104 | * Memory for allocating for handle keeps object position by | |
105 | * encoding <page, obj_idx> and the encoded value has a room | |
106 | * in least bit(ie, look at obj_to_location). | |
107 | * We use the bit to synchronize between object access by | |
108 | * user and migration. | |
109 | */ | |
110 | #define HANDLE_PIN_BIT 0 | |
111 | ||
112 | /* | |
113 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
114 | * to identify the object was allocated or not. | |
115 | * It's okay to add the status bit in the least bit because | |
116 | * header keeps handle which is 4byte-aligned address so we | |
117 | * have room for two bit at least. | |
118 | */ | |
119 | #define OBJ_ALLOCATED_TAG 1 | |
120 | #define OBJ_TAG_BITS 1 | |
121 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
122 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
123 | ||
a41ec880 | 124 | #define HUGE_BITS 1 |
cf8e0fed JM |
125 | #define FULLNESS_BITS 2 |
126 | #define CLASS_BITS 8 | |
127 | #define ISOLATED_BITS 3 | |
128 | #define MAGIC_VAL_BITS 8 | |
129 | ||
0959c63f SJ |
130 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) |
131 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
132 | #define ZS_MIN_ALLOC_SIZE \ | |
133 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 134 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 135 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
136 | |
137 | /* | |
7eb52512 | 138 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
139 | * trader-off here: |
140 | * - Large number of size classes is potentially wasteful as free page are | |
141 | * spread across these classes | |
142 | * - Small number of size classes causes large internal fragmentation | |
143 | * - Probably its better to use specific size classes (empirically | |
144 | * determined). NOTE: all those class sizes must be set as multiple of | |
145 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
146 | * | |
147 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
148 | * (reason above) | |
149 | */ | |
3783689a | 150 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
cf8e0fed JM |
151 | #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \ |
152 | ZS_SIZE_CLASS_DELTA) + 1) | |
0959c63f | 153 | |
0959c63f | 154 | enum fullness_group { |
0959c63f | 155 | ZS_EMPTY, |
48b4800a MK |
156 | ZS_ALMOST_EMPTY, |
157 | ZS_ALMOST_FULL, | |
158 | ZS_FULL, | |
159 | NR_ZS_FULLNESS, | |
0959c63f SJ |
160 | }; |
161 | ||
3828a764 | 162 | enum class_stat_type { |
48b4800a MK |
163 | CLASS_EMPTY, |
164 | CLASS_ALMOST_EMPTY, | |
165 | CLASS_ALMOST_FULL, | |
166 | CLASS_FULL, | |
0f050d99 GM |
167 | OBJ_ALLOCATED, |
168 | OBJ_USED, | |
48b4800a | 169 | NR_ZS_STAT_TYPE, |
0f050d99 GM |
170 | }; |
171 | ||
0f050d99 GM |
172 | struct zs_size_stat { |
173 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
174 | }; | |
175 | ||
57244594 SS |
176 | #ifdef CONFIG_ZSMALLOC_STAT |
177 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
178 | #endif |
179 | ||
48b4800a MK |
180 | #ifdef CONFIG_COMPACTION |
181 | static struct vfsmount *zsmalloc_mnt; | |
182 | #endif | |
183 | ||
0959c63f SJ |
184 | /* |
185 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
186 | * n <= N / f, where | |
187 | * n = number of allocated objects | |
188 | * N = total number of objects zspage can store | |
6dd9737e | 189 | * f = fullness_threshold_frac |
0959c63f SJ |
190 | * |
191 | * Similarly, we assign zspage to: | |
192 | * ZS_ALMOST_FULL when n > N / f | |
193 | * ZS_EMPTY when n == 0 | |
194 | * ZS_FULL when n == N | |
195 | * | |
196 | * (see: fix_fullness_group()) | |
197 | */ | |
198 | static const int fullness_threshold_frac = 4; | |
010b495e | 199 | static size_t huge_class_size; |
0959c63f SJ |
200 | |
201 | struct size_class { | |
57244594 | 202 | spinlock_t lock; |
48b4800a | 203 | struct list_head fullness_list[NR_ZS_FULLNESS]; |
0959c63f SJ |
204 | /* |
205 | * Size of objects stored in this class. Must be multiple | |
206 | * of ZS_ALIGN. | |
207 | */ | |
208 | int size; | |
1fc6e27d | 209 | int objs_per_zspage; |
7dfa4612 WY |
210 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
211 | int pages_per_zspage; | |
48b4800a MK |
212 | |
213 | unsigned int index; | |
214 | struct zs_size_stat stats; | |
0959c63f SJ |
215 | }; |
216 | ||
217 | /* | |
218 | * Placed within free objects to form a singly linked list. | |
3783689a | 219 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
220 | * |
221 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
222 | */ | |
223 | struct link_free { | |
2e40e163 MK |
224 | union { |
225 | /* | |
bfd093f5 | 226 | * Free object index; |
2e40e163 MK |
227 | * It's valid for non-allocated object |
228 | */ | |
bfd093f5 | 229 | unsigned long next; |
2e40e163 MK |
230 | /* |
231 | * Handle of allocated object. | |
232 | */ | |
233 | unsigned long handle; | |
234 | }; | |
0959c63f SJ |
235 | }; |
236 | ||
237 | struct zs_pool { | |
6f3526d6 | 238 | const char *name; |
0f050d99 | 239 | |
cf8e0fed | 240 | struct size_class *size_class[ZS_SIZE_CLASSES]; |
2e40e163 | 241 | struct kmem_cache *handle_cachep; |
3783689a | 242 | struct kmem_cache *zspage_cachep; |
0959c63f | 243 | |
13de8933 | 244 | atomic_long_t pages_allocated; |
0f050d99 | 245 | |
7d3f3938 | 246 | struct zs_pool_stats stats; |
ab9d306d SS |
247 | |
248 | /* Compact classes */ | |
249 | struct shrinker shrinker; | |
93144ca3 | 250 | |
0f050d99 GM |
251 | #ifdef CONFIG_ZSMALLOC_STAT |
252 | struct dentry *stat_dentry; | |
253 | #endif | |
48b4800a MK |
254 | #ifdef CONFIG_COMPACTION |
255 | struct inode *inode; | |
256 | struct work_struct free_work; | |
257 | #endif | |
0959c63f | 258 | }; |
61989a80 | 259 | |
3783689a MK |
260 | struct zspage { |
261 | struct { | |
a41ec880 | 262 | unsigned int huge:HUGE_BITS; |
3783689a | 263 | unsigned int fullness:FULLNESS_BITS; |
85d492f2 | 264 | unsigned int class:CLASS_BITS + 1; |
48b4800a MK |
265 | unsigned int isolated:ISOLATED_BITS; |
266 | unsigned int magic:MAGIC_VAL_BITS; | |
3783689a MK |
267 | }; |
268 | unsigned int inuse; | |
bfd093f5 | 269 | unsigned int freeobj; |
3783689a MK |
270 | struct page *first_page; |
271 | struct list_head list; /* fullness list */ | |
48b4800a MK |
272 | #ifdef CONFIG_COMPACTION |
273 | rwlock_t lock; | |
274 | #endif | |
3783689a | 275 | }; |
61989a80 | 276 | |
f553646a | 277 | struct mapping_area { |
f553646a | 278 | char *vm_buf; /* copy buffer for objects that span pages */ |
f553646a SJ |
279 | char *vm_addr; /* address of kmap_atomic()'ed pages */ |
280 | enum zs_mapmode vm_mm; /* mapping mode */ | |
281 | }; | |
282 | ||
a41ec880 MK |
283 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
284 | static void SetZsHugePage(struct zspage *zspage) | |
285 | { | |
286 | zspage->huge = 1; | |
287 | } | |
288 | ||
289 | static bool ZsHugePage(struct zspage *zspage) | |
290 | { | |
291 | return zspage->huge; | |
292 | } | |
293 | ||
48b4800a MK |
294 | #ifdef CONFIG_COMPACTION |
295 | static int zs_register_migration(struct zs_pool *pool); | |
296 | static void zs_unregister_migration(struct zs_pool *pool); | |
297 | static void migrate_lock_init(struct zspage *zspage); | |
298 | static void migrate_read_lock(struct zspage *zspage); | |
299 | static void migrate_read_unlock(struct zspage *zspage); | |
300 | static void kick_deferred_free(struct zs_pool *pool); | |
301 | static void init_deferred_free(struct zs_pool *pool); | |
302 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage); | |
303 | #else | |
304 | static int zsmalloc_mount(void) { return 0; } | |
305 | static void zsmalloc_unmount(void) {} | |
306 | static int zs_register_migration(struct zs_pool *pool) { return 0; } | |
307 | static void zs_unregister_migration(struct zs_pool *pool) {} | |
308 | static void migrate_lock_init(struct zspage *zspage) {} | |
309 | static void migrate_read_lock(struct zspage *zspage) {} | |
310 | static void migrate_read_unlock(struct zspage *zspage) {} | |
311 | static void kick_deferred_free(struct zs_pool *pool) {} | |
312 | static void init_deferred_free(struct zs_pool *pool) {} | |
313 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {} | |
314 | #endif | |
315 | ||
3783689a | 316 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
317 | { |
318 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
319 | 0, 0, NULL); | |
3783689a MK |
320 | if (!pool->handle_cachep) |
321 | return 1; | |
322 | ||
323 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
324 | 0, 0, NULL); | |
325 | if (!pool->zspage_cachep) { | |
326 | kmem_cache_destroy(pool->handle_cachep); | |
327 | pool->handle_cachep = NULL; | |
328 | return 1; | |
329 | } | |
330 | ||
331 | return 0; | |
2e40e163 MK |
332 | } |
333 | ||
3783689a | 334 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 335 | { |
cd10add0 | 336 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 337 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
338 | } |
339 | ||
3783689a | 340 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
341 | { |
342 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
48b4800a | 343 | gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
2e40e163 MK |
344 | } |
345 | ||
3783689a | 346 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
347 | { |
348 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
349 | } | |
350 | ||
3783689a MK |
351 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
352 | { | |
f0231305 | 353 | return kmem_cache_zalloc(pool->zspage_cachep, |
48b4800a | 354 | flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
399d8eeb | 355 | } |
3783689a MK |
356 | |
357 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
358 | { | |
359 | kmem_cache_free(pool->zspage_cachep, zspage); | |
360 | } | |
361 | ||
2e40e163 MK |
362 | static void record_obj(unsigned long handle, unsigned long obj) |
363 | { | |
c102f07c JL |
364 | /* |
365 | * lsb of @obj represents handle lock while other bits | |
366 | * represent object value the handle is pointing so | |
367 | * updating shouldn't do store tearing. | |
368 | */ | |
369 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
370 | } |
371 | ||
c795779d DS |
372 | /* zpool driver */ |
373 | ||
374 | #ifdef CONFIG_ZPOOL | |
375 | ||
6f3526d6 | 376 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 377 | const struct zpool_ops *zpool_ops, |
479305fd | 378 | struct zpool *zpool) |
c795779d | 379 | { |
d0d8da2d SS |
380 | /* |
381 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
382 | * different contexts and its caller must provide a valid | |
383 | * gfp mask. | |
384 | */ | |
385 | return zs_create_pool(name); | |
c795779d DS |
386 | } |
387 | ||
388 | static void zs_zpool_destroy(void *pool) | |
389 | { | |
390 | zs_destroy_pool(pool); | |
391 | } | |
392 | ||
393 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
394 | unsigned long *handle) | |
395 | { | |
d0d8da2d | 396 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
397 | return *handle ? 0 : -1; |
398 | } | |
399 | static void zs_zpool_free(void *pool, unsigned long handle) | |
400 | { | |
401 | zs_free(pool, handle); | |
402 | } | |
403 | ||
c795779d DS |
404 | static void *zs_zpool_map(void *pool, unsigned long handle, |
405 | enum zpool_mapmode mm) | |
406 | { | |
407 | enum zs_mapmode zs_mm; | |
408 | ||
409 | switch (mm) { | |
410 | case ZPOOL_MM_RO: | |
411 | zs_mm = ZS_MM_RO; | |
412 | break; | |
413 | case ZPOOL_MM_WO: | |
414 | zs_mm = ZS_MM_WO; | |
415 | break; | |
e4a9bc58 | 416 | case ZPOOL_MM_RW: |
c795779d DS |
417 | default: |
418 | zs_mm = ZS_MM_RW; | |
419 | break; | |
420 | } | |
421 | ||
422 | return zs_map_object(pool, handle, zs_mm); | |
423 | } | |
424 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
425 | { | |
426 | zs_unmap_object(pool, handle); | |
427 | } | |
428 | ||
429 | static u64 zs_zpool_total_size(void *pool) | |
430 | { | |
722cdc17 | 431 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
432 | } |
433 | ||
434 | static struct zpool_driver zs_zpool_driver = { | |
c165f25d HZ |
435 | .type = "zsmalloc", |
436 | .owner = THIS_MODULE, | |
437 | .create = zs_zpool_create, | |
438 | .destroy = zs_zpool_destroy, | |
439 | .malloc_support_movable = true, | |
440 | .malloc = zs_zpool_malloc, | |
441 | .free = zs_zpool_free, | |
442 | .map = zs_zpool_map, | |
443 | .unmap = zs_zpool_unmap, | |
444 | .total_size = zs_zpool_total_size, | |
c795779d DS |
445 | }; |
446 | ||
137f8cff | 447 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
448 | #endif /* CONFIG_ZPOOL */ |
449 | ||
61989a80 NG |
450 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
451 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
452 | ||
3457f414 | 453 | static __maybe_unused int is_first_page(struct page *page) |
61989a80 | 454 | { |
a27545bf | 455 | return PagePrivate(page); |
61989a80 NG |
456 | } |
457 | ||
48b4800a | 458 | /* Protected by class->lock */ |
3783689a | 459 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 460 | { |
3783689a | 461 | return zspage->inuse; |
4f42047b MK |
462 | } |
463 | ||
4f42047b | 464 | |
3783689a | 465 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 466 | { |
3783689a | 467 | zspage->inuse += val; |
4f42047b MK |
468 | } |
469 | ||
48b4800a | 470 | static inline struct page *get_first_page(struct zspage *zspage) |
4f42047b | 471 | { |
48b4800a | 472 | struct page *first_page = zspage->first_page; |
3783689a | 473 | |
48b4800a MK |
474 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
475 | return first_page; | |
4f42047b MK |
476 | } |
477 | ||
48b4800a | 478 | static inline int get_first_obj_offset(struct page *page) |
4f42047b | 479 | { |
ffedd09f | 480 | return page->page_type; |
48b4800a | 481 | } |
3783689a | 482 | |
48b4800a MK |
483 | static inline void set_first_obj_offset(struct page *page, int offset) |
484 | { | |
ffedd09f | 485 | page->page_type = offset; |
4f42047b MK |
486 | } |
487 | ||
bfd093f5 | 488 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 489 | { |
bfd093f5 | 490 | return zspage->freeobj; |
4f42047b MK |
491 | } |
492 | ||
bfd093f5 | 493 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 494 | { |
bfd093f5 | 495 | zspage->freeobj = obj; |
4f42047b MK |
496 | } |
497 | ||
3783689a | 498 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 499 | unsigned int *class_idx, |
61989a80 NG |
500 | enum fullness_group *fullness) |
501 | { | |
48b4800a MK |
502 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); |
503 | ||
3783689a MK |
504 | *fullness = zspage->fullness; |
505 | *class_idx = zspage->class; | |
61989a80 NG |
506 | } |
507 | ||
67f1c9cd MK |
508 | static struct size_class *zspage_class(struct zs_pool *pool, |
509 | struct zspage *zspage) | |
510 | { | |
511 | return pool->size_class[zspage->class]; | |
512 | } | |
513 | ||
3783689a | 514 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 515 | unsigned int class_idx, |
61989a80 NG |
516 | enum fullness_group fullness) |
517 | { | |
3783689a MK |
518 | zspage->class = class_idx; |
519 | zspage->fullness = fullness; | |
61989a80 NG |
520 | } |
521 | ||
c3e3e88a NC |
522 | /* |
523 | * zsmalloc divides the pool into various size classes where each | |
524 | * class maintains a list of zspages where each zspage is divided | |
525 | * into equal sized chunks. Each allocation falls into one of these | |
526 | * classes depending on its size. This function returns index of the | |
cb152a1a | 527 | * size class which has chunk size big enough to hold the given size. |
c3e3e88a | 528 | */ |
61989a80 NG |
529 | static int get_size_class_index(int size) |
530 | { | |
531 | int idx = 0; | |
532 | ||
533 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
534 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
535 | ZS_SIZE_CLASS_DELTA); | |
536 | ||
cf8e0fed | 537 | return min_t(int, ZS_SIZE_CLASSES - 1, idx); |
61989a80 NG |
538 | } |
539 | ||
3828a764 MK |
540 | /* type can be of enum type class_stat_type or fullness_group */ |
541 | static inline void class_stat_inc(struct size_class *class, | |
3eb95fea | 542 | int type, unsigned long cnt) |
248ca1b0 | 543 | { |
48b4800a | 544 | class->stats.objs[type] += cnt; |
248ca1b0 MK |
545 | } |
546 | ||
3828a764 MK |
547 | /* type can be of enum type class_stat_type or fullness_group */ |
548 | static inline void class_stat_dec(struct size_class *class, | |
3eb95fea | 549 | int type, unsigned long cnt) |
248ca1b0 | 550 | { |
48b4800a | 551 | class->stats.objs[type] -= cnt; |
248ca1b0 MK |
552 | } |
553 | ||
3828a764 | 554 | /* type can be of enum type class_stat_type or fullness_group */ |
248ca1b0 | 555 | static inline unsigned long zs_stat_get(struct size_class *class, |
3eb95fea | 556 | int type) |
248ca1b0 | 557 | { |
48b4800a | 558 | return class->stats.objs[type]; |
248ca1b0 MK |
559 | } |
560 | ||
57244594 SS |
561 | #ifdef CONFIG_ZSMALLOC_STAT |
562 | ||
4abaac9b | 563 | static void __init zs_stat_init(void) |
248ca1b0 | 564 | { |
4abaac9b DS |
565 | if (!debugfs_initialized()) { |
566 | pr_warn("debugfs not available, stat dir not created\n"); | |
567 | return; | |
568 | } | |
248ca1b0 MK |
569 | |
570 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
248ca1b0 MK |
571 | } |
572 | ||
573 | static void __exit zs_stat_exit(void) | |
574 | { | |
575 | debugfs_remove_recursive(zs_stat_root); | |
576 | } | |
577 | ||
1120ed54 SS |
578 | static unsigned long zs_can_compact(struct size_class *class); |
579 | ||
248ca1b0 MK |
580 | static int zs_stats_size_show(struct seq_file *s, void *v) |
581 | { | |
582 | int i; | |
583 | struct zs_pool *pool = s->private; | |
584 | struct size_class *class; | |
585 | int objs_per_zspage; | |
586 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 587 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
588 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
589 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 590 | unsigned long total_freeable = 0; |
248ca1b0 | 591 | |
1120ed54 | 592 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
593 | "class", "size", "almost_full", "almost_empty", |
594 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 595 | "pages_per_zspage", "freeable"); |
248ca1b0 | 596 | |
cf8e0fed | 597 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
248ca1b0 MK |
598 | class = pool->size_class[i]; |
599 | ||
600 | if (class->index != i) | |
601 | continue; | |
602 | ||
603 | spin_lock(&class->lock); | |
604 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
605 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
606 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
607 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 608 | freeable = zs_can_compact(class); |
248ca1b0 MK |
609 | spin_unlock(&class->lock); |
610 | ||
b4fd07a0 | 611 | objs_per_zspage = class->objs_per_zspage; |
248ca1b0 MK |
612 | pages_used = obj_allocated / objs_per_zspage * |
613 | class->pages_per_zspage; | |
614 | ||
1120ed54 SS |
615 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
616 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
617 | i, class->size, class_almost_full, class_almost_empty, |
618 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 619 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
620 | |
621 | total_class_almost_full += class_almost_full; | |
622 | total_class_almost_empty += class_almost_empty; | |
623 | total_objs += obj_allocated; | |
624 | total_used_objs += obj_used; | |
625 | total_pages += pages_used; | |
1120ed54 | 626 | total_freeable += freeable; |
248ca1b0 MK |
627 | } |
628 | ||
629 | seq_puts(s, "\n"); | |
1120ed54 | 630 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
631 | "Total", "", total_class_almost_full, |
632 | total_class_almost_empty, total_objs, | |
1120ed54 | 633 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
634 | |
635 | return 0; | |
636 | } | |
5ad35093 | 637 | DEFINE_SHOW_ATTRIBUTE(zs_stats_size); |
248ca1b0 | 638 | |
d34f6157 | 639 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 640 | { |
4abaac9b DS |
641 | if (!zs_stat_root) { |
642 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 643 | return; |
4abaac9b | 644 | } |
248ca1b0 | 645 | |
4268509a GKH |
646 | pool->stat_dentry = debugfs_create_dir(name, zs_stat_root); |
647 | ||
648 | debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool, | |
649 | &zs_stats_size_fops); | |
248ca1b0 MK |
650 | } |
651 | ||
652 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
653 | { | |
654 | debugfs_remove_recursive(pool->stat_dentry); | |
655 | } | |
656 | ||
657 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 658 | static void __init zs_stat_init(void) |
248ca1b0 | 659 | { |
248ca1b0 MK |
660 | } |
661 | ||
662 | static void __exit zs_stat_exit(void) | |
663 | { | |
664 | } | |
665 | ||
d34f6157 | 666 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 667 | { |
248ca1b0 MK |
668 | } |
669 | ||
670 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
671 | { | |
672 | } | |
248ca1b0 MK |
673 | #endif |
674 | ||
48b4800a | 675 | |
c3e3e88a NC |
676 | /* |
677 | * For each size class, zspages are divided into different groups | |
678 | * depending on how "full" they are. This was done so that we could | |
679 | * easily find empty or nearly empty zspages when we try to shrink | |
680 | * the pool (not yet implemented). This function returns fullness | |
681 | * status of the given page. | |
682 | */ | |
1fc6e27d | 683 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 684 | struct zspage *zspage) |
61989a80 | 685 | { |
1fc6e27d | 686 | int inuse, objs_per_zspage; |
61989a80 | 687 | enum fullness_group fg; |
830e4bc5 | 688 | |
3783689a | 689 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 690 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
691 | |
692 | if (inuse == 0) | |
693 | fg = ZS_EMPTY; | |
1fc6e27d | 694 | else if (inuse == objs_per_zspage) |
61989a80 | 695 | fg = ZS_FULL; |
1fc6e27d | 696 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
697 | fg = ZS_ALMOST_EMPTY; |
698 | else | |
699 | fg = ZS_ALMOST_FULL; | |
700 | ||
701 | return fg; | |
702 | } | |
703 | ||
c3e3e88a NC |
704 | /* |
705 | * Each size class maintains various freelists and zspages are assigned | |
706 | * to one of these freelists based on the number of live objects they | |
707 | * have. This functions inserts the given zspage into the freelist | |
708 | * identified by <class, fullness_group>. | |
709 | */ | |
251cbb95 | 710 | static void insert_zspage(struct size_class *class, |
3783689a MK |
711 | struct zspage *zspage, |
712 | enum fullness_group fullness) | |
61989a80 | 713 | { |
3783689a | 714 | struct zspage *head; |
61989a80 | 715 | |
3828a764 | 716 | class_stat_inc(class, fullness, 1); |
3783689a MK |
717 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
718 | struct zspage, list); | |
58f17117 | 719 | /* |
3783689a MK |
720 | * We want to see more ZS_FULL pages and less almost empty/full. |
721 | * Put pages with higher ->inuse first. | |
58f17117 | 722 | */ |
110ceb82 ML |
723 | if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head)) |
724 | list_add(&zspage->list, &head->list); | |
725 | else | |
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 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
61989a80 | 738 | |
3783689a | 739 | list_del_init(&zspage->list); |
3828a764 | 740 | class_stat_dec(class, fullness, 1); |
61989a80 NG |
741 | } |
742 | ||
c3e3e88a NC |
743 | /* |
744 | * Each size class maintains zspages in different fullness groups depending | |
745 | * on the number of live objects they contain. When allocating or freeing | |
746 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
747 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
748 | * a status change has occurred for the given page and accordingly moves the | |
749 | * page from the freelist of the old fullness group to that of the new | |
750 | * fullness group. | |
751 | */ | |
c7806261 | 752 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 753 | struct zspage *zspage) |
61989a80 NG |
754 | { |
755 | int class_idx; | |
61989a80 NG |
756 | enum fullness_group currfg, newfg; |
757 | ||
3783689a MK |
758 | get_zspage_mapping(zspage, &class_idx, &currfg); |
759 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
760 | if (newfg == currfg) |
761 | goto out; | |
762 | ||
c4549b87 MK |
763 | remove_zspage(class, zspage, currfg); |
764 | insert_zspage(class, zspage, newfg); | |
3783689a | 765 | set_zspage_mapping(zspage, class_idx, newfg); |
61989a80 NG |
766 | out: |
767 | return newfg; | |
768 | } | |
769 | ||
770 | /* | |
771 | * We have to decide on how many pages to link together | |
772 | * to form a zspage for each size class. This is important | |
773 | * to reduce wastage due to unusable space left at end of | |
774 | * each zspage which is given as: | |
888fa374 YX |
775 | * wastage = Zp % class_size |
776 | * usage = Zp - wastage | |
61989a80 NG |
777 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
778 | * | |
779 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
780 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
781 | * since then we can perfectly fit in 8 such objects. | |
782 | */ | |
2e3b6154 | 783 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
784 | { |
785 | int i, max_usedpc = 0; | |
786 | /* zspage order which gives maximum used size per KB */ | |
787 | int max_usedpc_order = 1; | |
788 | ||
84d4faab | 789 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
790 | int zspage_size; |
791 | int waste, usedpc; | |
792 | ||
793 | zspage_size = i * PAGE_SIZE; | |
794 | waste = zspage_size % class_size; | |
795 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
796 | ||
797 | if (usedpc > max_usedpc) { | |
798 | max_usedpc = usedpc; | |
799 | max_usedpc_order = i; | |
800 | } | |
801 | } | |
802 | ||
803 | return max_usedpc_order; | |
804 | } | |
805 | ||
3783689a | 806 | static struct zspage *get_zspage(struct page *page) |
61989a80 | 807 | { |
a6c5e0f7 | 808 | struct zspage *zspage = (struct zspage *)page_private(page); |
48b4800a MK |
809 | |
810 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); | |
811 | return zspage; | |
61989a80 NG |
812 | } |
813 | ||
814 | static struct page *get_next_page(struct page *page) | |
815 | { | |
a41ec880 MK |
816 | struct zspage *zspage = get_zspage(page); |
817 | ||
818 | if (unlikely(ZsHugePage(zspage))) | |
48b4800a MK |
819 | return NULL; |
820 | ||
ffedd09f | 821 | return (struct page *)page->index; |
61989a80 NG |
822 | } |
823 | ||
bfd093f5 MK |
824 | /** |
825 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
e8b098fc | 826 | * @obj: the encoded object value |
bfd093f5 MK |
827 | * @page: page object resides in zspage |
828 | * @obj_idx: object index | |
67296874 | 829 | */ |
bfd093f5 MK |
830 | static void obj_to_location(unsigned long obj, struct page **page, |
831 | unsigned int *obj_idx) | |
61989a80 | 832 | { |
bfd093f5 MK |
833 | obj >>= OBJ_TAG_BITS; |
834 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
835 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
836 | } | |
61989a80 | 837 | |
67f1c9cd MK |
838 | static void obj_to_page(unsigned long obj, struct page **page) |
839 | { | |
840 | obj >>= OBJ_TAG_BITS; | |
841 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
842 | } | |
843 | ||
bfd093f5 MK |
844 | /** |
845 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
846 | * @page: page object resides in zspage | |
847 | * @obj_idx: object index | |
848 | */ | |
849 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
850 | { | |
851 | unsigned long obj; | |
61989a80 | 852 | |
312fcae2 | 853 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 854 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 855 | obj <<= OBJ_TAG_BITS; |
61989a80 | 856 | |
bfd093f5 | 857 | return obj; |
61989a80 NG |
858 | } |
859 | ||
2e40e163 MK |
860 | static unsigned long handle_to_obj(unsigned long handle) |
861 | { | |
862 | return *(unsigned long *)handle; | |
863 | } | |
864 | ||
3ae92ac2 | 865 | static bool obj_allocated(struct page *page, void *obj, unsigned long *phandle) |
312fcae2 | 866 | { |
3ae92ac2 | 867 | unsigned long handle; |
a41ec880 | 868 | struct zspage *zspage = get_zspage(page); |
3ae92ac2 | 869 | |
a41ec880 | 870 | if (unlikely(ZsHugePage(zspage))) { |
830e4bc5 | 871 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3ae92ac2 | 872 | handle = page->index; |
7b60a685 | 873 | } else |
3ae92ac2 MK |
874 | handle = *(unsigned long *)obj; |
875 | ||
876 | if (!(handle & OBJ_ALLOCATED_TAG)) | |
877 | return false; | |
878 | ||
879 | *phandle = handle & ~OBJ_ALLOCATED_TAG; | |
880 | return true; | |
312fcae2 MK |
881 | } |
882 | ||
48b4800a MK |
883 | static inline int testpin_tag(unsigned long handle) |
884 | { | |
885 | return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle); | |
886 | } | |
887 | ||
312fcae2 MK |
888 | static inline int trypin_tag(unsigned long handle) |
889 | { | |
1b8320b6 | 890 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
891 | } |
892 | ||
70c7ec95 | 893 | static void pin_tag(unsigned long handle) __acquires(bitlock) |
312fcae2 | 894 | { |
1b8320b6 | 895 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
896 | } |
897 | ||
bc22b18b | 898 | static void unpin_tag(unsigned long handle) __releases(bitlock) |
312fcae2 | 899 | { |
1b8320b6 | 900 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
901 | } |
902 | ||
f4477e90 NG |
903 | static void reset_page(struct page *page) |
904 | { | |
48b4800a | 905 | __ClearPageMovable(page); |
18fd06bf | 906 | ClearPagePrivate(page); |
f4477e90 | 907 | set_page_private(page, 0); |
48b4800a | 908 | page_mapcount_reset(page); |
ffedd09f | 909 | page->index = 0; |
48b4800a MK |
910 | } |
911 | ||
4d0a5402 | 912 | static int trylock_zspage(struct zspage *zspage) |
48b4800a MK |
913 | { |
914 | struct page *cursor, *fail; | |
915 | ||
916 | for (cursor = get_first_page(zspage); cursor != NULL; cursor = | |
917 | get_next_page(cursor)) { | |
918 | if (!trylock_page(cursor)) { | |
919 | fail = cursor; | |
920 | goto unlock; | |
921 | } | |
922 | } | |
923 | ||
924 | return 1; | |
925 | unlock: | |
926 | for (cursor = get_first_page(zspage); cursor != fail; cursor = | |
927 | get_next_page(cursor)) | |
928 | unlock_page(cursor); | |
929 | ||
930 | return 0; | |
f4477e90 NG |
931 | } |
932 | ||
48b4800a MK |
933 | static void __free_zspage(struct zs_pool *pool, struct size_class *class, |
934 | struct zspage *zspage) | |
61989a80 | 935 | { |
3783689a | 936 | struct page *page, *next; |
48b4800a MK |
937 | enum fullness_group fg; |
938 | unsigned int class_idx; | |
939 | ||
940 | get_zspage_mapping(zspage, &class_idx, &fg); | |
941 | ||
942 | assert_spin_locked(&class->lock); | |
61989a80 | 943 | |
3783689a | 944 | VM_BUG_ON(get_zspage_inuse(zspage)); |
48b4800a | 945 | VM_BUG_ON(fg != ZS_EMPTY); |
61989a80 | 946 | |
48b4800a | 947 | next = page = get_first_page(zspage); |
3783689a | 948 | do { |
48b4800a MK |
949 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
950 | next = get_next_page(page); | |
3783689a | 951 | reset_page(page); |
48b4800a | 952 | unlock_page(page); |
91537fee | 953 | dec_zone_page_state(page, NR_ZSPAGES); |
3783689a MK |
954 | put_page(page); |
955 | page = next; | |
956 | } while (page != NULL); | |
61989a80 | 957 | |
3783689a | 958 | cache_free_zspage(pool, zspage); |
48b4800a | 959 | |
3828a764 | 960 | class_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
961 | atomic_long_sub(class->pages_per_zspage, |
962 | &pool->pages_allocated); | |
963 | } | |
964 | ||
965 | static void free_zspage(struct zs_pool *pool, struct size_class *class, | |
966 | struct zspage *zspage) | |
967 | { | |
968 | VM_BUG_ON(get_zspage_inuse(zspage)); | |
969 | VM_BUG_ON(list_empty(&zspage->list)); | |
970 | ||
971 | if (!trylock_zspage(zspage)) { | |
972 | kick_deferred_free(pool); | |
973 | return; | |
974 | } | |
975 | ||
976 | remove_zspage(class, zspage, ZS_EMPTY); | |
977 | __free_zspage(pool, class, zspage); | |
61989a80 NG |
978 | } |
979 | ||
980 | /* Initialize a newly allocated zspage */ | |
3783689a | 981 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 982 | { |
bfd093f5 | 983 | unsigned int freeobj = 1; |
61989a80 | 984 | unsigned long off = 0; |
48b4800a | 985 | struct page *page = get_first_page(zspage); |
830e4bc5 | 986 | |
61989a80 NG |
987 | while (page) { |
988 | struct page *next_page; | |
989 | struct link_free *link; | |
af4ee5e9 | 990 | void *vaddr; |
61989a80 | 991 | |
3783689a | 992 | set_first_obj_offset(page, off); |
61989a80 | 993 | |
af4ee5e9 MK |
994 | vaddr = kmap_atomic(page); |
995 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
996 | |
997 | while ((off += class->size) < PAGE_SIZE) { | |
3b1d9ca6 | 998 | link->next = freeobj++ << OBJ_TAG_BITS; |
5538c562 | 999 | link += class->size / sizeof(*link); |
61989a80 NG |
1000 | } |
1001 | ||
1002 | /* | |
1003 | * We now come to the last (full or partial) object on this | |
1004 | * page, which must point to the first object on the next | |
1005 | * page (if present) | |
1006 | */ | |
1007 | next_page = get_next_page(page); | |
bfd093f5 | 1008 | if (next_page) { |
3b1d9ca6 | 1009 | link->next = freeobj++ << OBJ_TAG_BITS; |
bfd093f5 MK |
1010 | } else { |
1011 | /* | |
3b1d9ca6 | 1012 | * Reset OBJ_TAG_BITS bit to last link to tell |
bfd093f5 MK |
1013 | * whether it's allocated object or not. |
1014 | */ | |
01a6ad9a | 1015 | link->next = -1UL << OBJ_TAG_BITS; |
bfd093f5 | 1016 | } |
af4ee5e9 | 1017 | kunmap_atomic(vaddr); |
61989a80 | 1018 | page = next_page; |
5538c562 | 1019 | off %= PAGE_SIZE; |
61989a80 | 1020 | } |
bdb0af7c | 1021 | |
bfd093f5 | 1022 | set_freeobj(zspage, 0); |
61989a80 NG |
1023 | } |
1024 | ||
48b4800a MK |
1025 | static void create_page_chain(struct size_class *class, struct zspage *zspage, |
1026 | struct page *pages[]) | |
61989a80 | 1027 | { |
bdb0af7c MK |
1028 | int i; |
1029 | struct page *page; | |
1030 | struct page *prev_page = NULL; | |
48b4800a | 1031 | int nr_pages = class->pages_per_zspage; |
61989a80 NG |
1032 | |
1033 | /* | |
1034 | * Allocate individual pages and link them together as: | |
ffedd09f | 1035 | * 1. all pages are linked together using page->index |
3783689a | 1036 | * 2. each sub-page point to zspage using page->private |
61989a80 | 1037 | * |
3783689a | 1038 | * we set PG_private to identify the first page (i.e. no other sub-page |
22c5cef1 | 1039 | * has this flag set). |
61989a80 | 1040 | */ |
bdb0af7c MK |
1041 | for (i = 0; i < nr_pages; i++) { |
1042 | page = pages[i]; | |
3783689a | 1043 | set_page_private(page, (unsigned long)zspage); |
ffedd09f | 1044 | page->index = 0; |
bdb0af7c | 1045 | if (i == 0) { |
3783689a | 1046 | zspage->first_page = page; |
a27545bf | 1047 | SetPagePrivate(page); |
48b4800a MK |
1048 | if (unlikely(class->objs_per_zspage == 1 && |
1049 | class->pages_per_zspage == 1)) | |
a41ec880 | 1050 | SetZsHugePage(zspage); |
3783689a | 1051 | } else { |
ffedd09f | 1052 | prev_page->index = (unsigned long)page; |
61989a80 | 1053 | } |
61989a80 NG |
1054 | prev_page = page; |
1055 | } | |
bdb0af7c | 1056 | } |
61989a80 | 1057 | |
bdb0af7c MK |
1058 | /* |
1059 | * Allocate a zspage for the given size class | |
1060 | */ | |
3783689a MK |
1061 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
1062 | struct size_class *class, | |
1063 | gfp_t gfp) | |
bdb0af7c MK |
1064 | { |
1065 | int i; | |
bdb0af7c | 1066 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1067 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1068 | ||
1069 | if (!zspage) | |
1070 | return NULL; | |
1071 | ||
48b4800a MK |
1072 | zspage->magic = ZSPAGE_MAGIC; |
1073 | migrate_lock_init(zspage); | |
61989a80 | 1074 | |
bdb0af7c MK |
1075 | for (i = 0; i < class->pages_per_zspage; i++) { |
1076 | struct page *page; | |
61989a80 | 1077 | |
3783689a | 1078 | page = alloc_page(gfp); |
bdb0af7c | 1079 | if (!page) { |
91537fee MK |
1080 | while (--i >= 0) { |
1081 | dec_zone_page_state(pages[i], NR_ZSPAGES); | |
bdb0af7c | 1082 | __free_page(pages[i]); |
91537fee | 1083 | } |
3783689a | 1084 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1085 | return NULL; |
1086 | } | |
91537fee MK |
1087 | |
1088 | inc_zone_page_state(page, NR_ZSPAGES); | |
bdb0af7c | 1089 | pages[i] = page; |
61989a80 NG |
1090 | } |
1091 | ||
48b4800a | 1092 | create_page_chain(class, zspage, pages); |
3783689a | 1093 | init_zspage(class, zspage); |
bdb0af7c | 1094 | |
3783689a | 1095 | return zspage; |
61989a80 NG |
1096 | } |
1097 | ||
3783689a | 1098 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1099 | { |
1100 | int i; | |
3783689a | 1101 | struct zspage *zspage; |
61989a80 | 1102 | |
48b4800a | 1103 | for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) { |
3783689a MK |
1104 | zspage = list_first_entry_or_null(&class->fullness_list[i], |
1105 | struct zspage, list); | |
1106 | if (zspage) | |
61989a80 NG |
1107 | break; |
1108 | } | |
1109 | ||
3783689a | 1110 | return zspage; |
61989a80 NG |
1111 | } |
1112 | ||
f553646a SJ |
1113 | static inline int __zs_cpu_up(struct mapping_area *area) |
1114 | { | |
1115 | /* | |
1116 | * Make sure we don't leak memory if a cpu UP notification | |
1117 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1118 | */ | |
1119 | if (area->vm_buf) | |
1120 | return 0; | |
40f9fb8c | 1121 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1122 | if (!area->vm_buf) |
1123 | return -ENOMEM; | |
1124 | return 0; | |
1125 | } | |
1126 | ||
1127 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1128 | { | |
40f9fb8c | 1129 | kfree(area->vm_buf); |
f553646a SJ |
1130 | area->vm_buf = NULL; |
1131 | } | |
1132 | ||
1133 | static void *__zs_map_object(struct mapping_area *area, | |
1134 | struct page *pages[2], int off, int size) | |
5f601902 | 1135 | { |
5f601902 SJ |
1136 | int sizes[2]; |
1137 | void *addr; | |
f553646a | 1138 | char *buf = area->vm_buf; |
5f601902 | 1139 | |
f553646a SJ |
1140 | /* disable page faults to match kmap_atomic() return conditions */ |
1141 | pagefault_disable(); | |
1142 | ||
1143 | /* no read fastpath */ | |
1144 | if (area->vm_mm == ZS_MM_WO) | |
1145 | goto out; | |
5f601902 SJ |
1146 | |
1147 | sizes[0] = PAGE_SIZE - off; | |
1148 | sizes[1] = size - sizes[0]; | |
1149 | ||
5f601902 SJ |
1150 | /* copy object to per-cpu buffer */ |
1151 | addr = kmap_atomic(pages[0]); | |
1152 | memcpy(buf, addr + off, sizes[0]); | |
1153 | kunmap_atomic(addr); | |
1154 | addr = kmap_atomic(pages[1]); | |
1155 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1156 | kunmap_atomic(addr); | |
f553646a SJ |
1157 | out: |
1158 | return area->vm_buf; | |
5f601902 SJ |
1159 | } |
1160 | ||
f553646a SJ |
1161 | static void __zs_unmap_object(struct mapping_area *area, |
1162 | struct page *pages[2], int off, int size) | |
5f601902 | 1163 | { |
5f601902 SJ |
1164 | int sizes[2]; |
1165 | void *addr; | |
2e40e163 | 1166 | char *buf; |
5f601902 | 1167 | |
f553646a SJ |
1168 | /* no write fastpath */ |
1169 | if (area->vm_mm == ZS_MM_RO) | |
1170 | goto out; | |
5f601902 | 1171 | |
7b60a685 | 1172 | buf = area->vm_buf; |
a82cbf07 YX |
1173 | buf = buf + ZS_HANDLE_SIZE; |
1174 | size -= ZS_HANDLE_SIZE; | |
1175 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1176 | |
5f601902 SJ |
1177 | sizes[0] = PAGE_SIZE - off; |
1178 | sizes[1] = size - sizes[0]; | |
1179 | ||
1180 | /* copy per-cpu buffer to object */ | |
1181 | addr = kmap_atomic(pages[0]); | |
1182 | memcpy(addr + off, buf, sizes[0]); | |
1183 | kunmap_atomic(addr); | |
1184 | addr = kmap_atomic(pages[1]); | |
1185 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1186 | kunmap_atomic(addr); | |
f553646a SJ |
1187 | |
1188 | out: | |
1189 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1190 | pagefault_enable(); | |
5f601902 | 1191 | } |
61989a80 | 1192 | |
215c89d0 | 1193 | static int zs_cpu_prepare(unsigned int cpu) |
61989a80 | 1194 | { |
61989a80 NG |
1195 | struct mapping_area *area; |
1196 | ||
215c89d0 SAS |
1197 | area = &per_cpu(zs_map_area, cpu); |
1198 | return __zs_cpu_up(area); | |
61989a80 NG |
1199 | } |
1200 | ||
215c89d0 | 1201 | static int zs_cpu_dead(unsigned int cpu) |
61989a80 | 1202 | { |
215c89d0 | 1203 | struct mapping_area *area; |
40f9fb8c | 1204 | |
215c89d0 SAS |
1205 | area = &per_cpu(zs_map_area, cpu); |
1206 | __zs_cpu_down(area); | |
1207 | return 0; | |
b1b00a5b SS |
1208 | } |
1209 | ||
64d90465 GM |
1210 | static bool can_merge(struct size_class *prev, int pages_per_zspage, |
1211 | int objs_per_zspage) | |
9eec4cd5 | 1212 | { |
64d90465 GM |
1213 | if (prev->pages_per_zspage == pages_per_zspage && |
1214 | prev->objs_per_zspage == objs_per_zspage) | |
1215 | return true; | |
9eec4cd5 | 1216 | |
64d90465 | 1217 | return false; |
9eec4cd5 JK |
1218 | } |
1219 | ||
3783689a | 1220 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1221 | { |
3783689a | 1222 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1223 | } |
1224 | ||
66cdef66 GM |
1225 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1226 | { | |
1227 | return atomic_long_read(&pool->pages_allocated); | |
1228 | } | |
1229 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1230 | ||
4bbc0bc0 | 1231 | /** |
66cdef66 GM |
1232 | * zs_map_object - get address of allocated object from handle. |
1233 | * @pool: pool from which the object was allocated | |
1234 | * @handle: handle returned from zs_malloc | |
f0953a1b | 1235 | * @mm: mapping mode to use |
4bbc0bc0 | 1236 | * |
66cdef66 GM |
1237 | * Before using an object allocated from zs_malloc, it must be mapped using |
1238 | * this function. When done with the object, it must be unmapped using | |
1239 | * zs_unmap_object. | |
4bbc0bc0 | 1240 | * |
66cdef66 GM |
1241 | * Only one object can be mapped per cpu at a time. There is no protection |
1242 | * against nested mappings. | |
1243 | * | |
1244 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1245 | */ |
66cdef66 GM |
1246 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1247 | enum zs_mapmode mm) | |
61989a80 | 1248 | { |
3783689a | 1249 | struct zspage *zspage; |
66cdef66 | 1250 | struct page *page; |
bfd093f5 MK |
1251 | unsigned long obj, off; |
1252 | unsigned int obj_idx; | |
61989a80 | 1253 | |
66cdef66 GM |
1254 | struct size_class *class; |
1255 | struct mapping_area *area; | |
1256 | struct page *pages[2]; | |
2e40e163 | 1257 | void *ret; |
61989a80 | 1258 | |
9eec4cd5 | 1259 | /* |
66cdef66 GM |
1260 | * Because we use per-cpu mapping areas shared among the |
1261 | * pools/users, we can't allow mapping in interrupt context | |
1262 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1263 | */ |
1aedcafb | 1264 | BUG_ON(in_interrupt()); |
61989a80 | 1265 | |
312fcae2 MK |
1266 | /* From now on, migration cannot move the object */ |
1267 | pin_tag(handle); | |
1268 | ||
2e40e163 MK |
1269 | obj = handle_to_obj(handle); |
1270 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1271 | zspage = get_zspage(page); |
48b4800a MK |
1272 | |
1273 | /* migration cannot move any subpage in this zspage */ | |
1274 | migrate_read_lock(zspage); | |
1275 | ||
67f1c9cd | 1276 | class = zspage_class(pool, zspage); |
bfd093f5 | 1277 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1278 | |
66cdef66 GM |
1279 | area = &get_cpu_var(zs_map_area); |
1280 | area->vm_mm = mm; | |
1281 | if (off + class->size <= PAGE_SIZE) { | |
1282 | /* this object is contained entirely within a page */ | |
1283 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1284 | ret = area->vm_addr + off; |
1285 | goto out; | |
61989a80 NG |
1286 | } |
1287 | ||
66cdef66 GM |
1288 | /* this object spans two pages */ |
1289 | pages[0] = page; | |
1290 | pages[1] = get_next_page(page); | |
1291 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1292 | |
2e40e163 MK |
1293 | ret = __zs_map_object(area, pages, off, class->size); |
1294 | out: | |
a41ec880 | 1295 | if (likely(!ZsHugePage(zspage))) |
7b60a685 MK |
1296 | ret += ZS_HANDLE_SIZE; |
1297 | ||
1298 | return ret; | |
61989a80 | 1299 | } |
66cdef66 | 1300 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1301 | |
66cdef66 | 1302 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1303 | { |
3783689a | 1304 | struct zspage *zspage; |
66cdef66 | 1305 | struct page *page; |
bfd093f5 MK |
1306 | unsigned long obj, off; |
1307 | unsigned int obj_idx; | |
61989a80 | 1308 | |
66cdef66 GM |
1309 | struct size_class *class; |
1310 | struct mapping_area *area; | |
9eec4cd5 | 1311 | |
2e40e163 MK |
1312 | obj = handle_to_obj(handle); |
1313 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1314 | zspage = get_zspage(page); |
67f1c9cd | 1315 | class = zspage_class(pool, zspage); |
bfd093f5 | 1316 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1317 | |
66cdef66 GM |
1318 | area = this_cpu_ptr(&zs_map_area); |
1319 | if (off + class->size <= PAGE_SIZE) | |
1320 | kunmap_atomic(area->vm_addr); | |
1321 | else { | |
1322 | struct page *pages[2]; | |
40f9fb8c | 1323 | |
66cdef66 GM |
1324 | pages[0] = page; |
1325 | pages[1] = get_next_page(page); | |
1326 | BUG_ON(!pages[1]); | |
1327 | ||
1328 | __zs_unmap_object(area, pages, off, class->size); | |
1329 | } | |
1330 | put_cpu_var(zs_map_area); | |
48b4800a MK |
1331 | |
1332 | migrate_read_unlock(zspage); | |
312fcae2 | 1333 | unpin_tag(handle); |
61989a80 | 1334 | } |
66cdef66 | 1335 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1336 | |
010b495e SS |
1337 | /** |
1338 | * zs_huge_class_size() - Returns the size (in bytes) of the first huge | |
1339 | * zsmalloc &size_class. | |
1340 | * @pool: zsmalloc pool to use | |
1341 | * | |
1342 | * The function returns the size of the first huge class - any object of equal | |
1343 | * or bigger size will be stored in zspage consisting of a single physical | |
1344 | * page. | |
1345 | * | |
1346 | * Context: Any context. | |
1347 | * | |
1348 | * Return: the size (in bytes) of the first huge zsmalloc &size_class. | |
1349 | */ | |
1350 | size_t zs_huge_class_size(struct zs_pool *pool) | |
1351 | { | |
1352 | return huge_class_size; | |
1353 | } | |
1354 | EXPORT_SYMBOL_GPL(zs_huge_class_size); | |
1355 | ||
0a5f079b | 1356 | static unsigned long obj_malloc(struct zs_pool *pool, |
3783689a | 1357 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1358 | { |
bfd093f5 | 1359 | int i, nr_page, offset; |
c7806261 MK |
1360 | unsigned long obj; |
1361 | struct link_free *link; | |
0a5f079b | 1362 | struct size_class *class; |
c7806261 MK |
1363 | |
1364 | struct page *m_page; | |
bfd093f5 | 1365 | unsigned long m_offset; |
c7806261 MK |
1366 | void *vaddr; |
1367 | ||
0a5f079b | 1368 | class = pool->size_class[zspage->class]; |
312fcae2 | 1369 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1370 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1371 | |
1372 | offset = obj * class->size; | |
1373 | nr_page = offset >> PAGE_SHIFT; | |
1374 | m_offset = offset & ~PAGE_MASK; | |
1375 | m_page = get_first_page(zspage); | |
1376 | ||
1377 | for (i = 0; i < nr_page; i++) | |
1378 | m_page = get_next_page(m_page); | |
c7806261 MK |
1379 | |
1380 | vaddr = kmap_atomic(m_page); | |
1381 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3b1d9ca6 | 1382 | set_freeobj(zspage, link->next >> OBJ_TAG_BITS); |
a41ec880 | 1383 | if (likely(!ZsHugePage(zspage))) |
7b60a685 MK |
1384 | /* record handle in the header of allocated chunk */ |
1385 | link->handle = handle; | |
1386 | else | |
3783689a MK |
1387 | /* record handle to page->index */ |
1388 | zspage->first_page->index = handle; | |
1389 | ||
c7806261 | 1390 | kunmap_atomic(vaddr); |
3783689a | 1391 | mod_zspage_inuse(zspage, 1); |
c7806261 | 1392 | |
bfd093f5 MK |
1393 | obj = location_to_obj(m_page, obj); |
1394 | ||
c7806261 MK |
1395 | return obj; |
1396 | } | |
1397 | ||
1398 | ||
61989a80 NG |
1399 | /** |
1400 | * zs_malloc - Allocate block of given size from pool. | |
1401 | * @pool: pool to allocate from | |
1402 | * @size: size of block to allocate | |
fd854463 | 1403 | * @gfp: gfp flags when allocating object |
61989a80 | 1404 | * |
00a61d86 | 1405 | * On success, handle to the allocated object is returned, |
c2344348 | 1406 | * otherwise 0. |
61989a80 NG |
1407 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1408 | */ | |
d0d8da2d | 1409 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1410 | { |
2e40e163 | 1411 | unsigned long handle, obj; |
61989a80 | 1412 | struct size_class *class; |
48b4800a | 1413 | enum fullness_group newfg; |
3783689a | 1414 | struct zspage *zspage; |
61989a80 | 1415 | |
7b60a685 | 1416 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1417 | return 0; |
1418 | ||
3783689a | 1419 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1420 | if (!handle) |
c2344348 | 1421 | return 0; |
61989a80 | 1422 | |
2e40e163 MK |
1423 | /* extra space in chunk to keep the handle */ |
1424 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1425 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1426 | |
1427 | spin_lock(&class->lock); | |
3783689a | 1428 | zspage = find_get_zspage(class); |
48b4800a | 1429 | if (likely(zspage)) { |
0a5f079b | 1430 | obj = obj_malloc(pool, zspage, handle); |
48b4800a MK |
1431 | /* Now move the zspage to another fullness group, if required */ |
1432 | fix_fullness_group(class, zspage); | |
1433 | record_obj(handle, obj); | |
0a5f079b | 1434 | class_stat_inc(class, OBJ_USED, 1); |
61989a80 | 1435 | spin_unlock(&class->lock); |
61989a80 | 1436 | |
48b4800a MK |
1437 | return handle; |
1438 | } | |
0f050d99 | 1439 | |
48b4800a MK |
1440 | spin_unlock(&class->lock); |
1441 | ||
1442 | zspage = alloc_zspage(pool, class, gfp); | |
1443 | if (!zspage) { | |
1444 | cache_free_handle(pool, handle); | |
1445 | return 0; | |
61989a80 NG |
1446 | } |
1447 | ||
48b4800a | 1448 | spin_lock(&class->lock); |
0a5f079b | 1449 | obj = obj_malloc(pool, zspage, handle); |
48b4800a MK |
1450 | newfg = get_fullness_group(class, zspage); |
1451 | insert_zspage(class, zspage, newfg); | |
1452 | set_zspage_mapping(zspage, class->index, newfg); | |
2e40e163 | 1453 | record_obj(handle, obj); |
48b4800a MK |
1454 | atomic_long_add(class->pages_per_zspage, |
1455 | &pool->pages_allocated); | |
3828a764 | 1456 | class_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage); |
0a5f079b | 1457 | class_stat_inc(class, OBJ_USED, 1); |
48b4800a MK |
1458 | |
1459 | /* We completely set up zspage so mark them as movable */ | |
1460 | SetZsPageMovable(pool, zspage); | |
61989a80 NG |
1461 | spin_unlock(&class->lock); |
1462 | ||
2e40e163 | 1463 | return handle; |
61989a80 NG |
1464 | } |
1465 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1466 | ||
0a5f079b | 1467 | static void obj_free(int class_size, unsigned long obj) |
61989a80 NG |
1468 | { |
1469 | struct link_free *link; | |
3783689a MK |
1470 | struct zspage *zspage; |
1471 | struct page *f_page; | |
bfd093f5 MK |
1472 | unsigned long f_offset; |
1473 | unsigned int f_objidx; | |
af4ee5e9 | 1474 | void *vaddr; |
61989a80 | 1475 | |
2e40e163 | 1476 | obj_to_location(obj, &f_page, &f_objidx); |
0a5f079b | 1477 | f_offset = (class_size * f_objidx) & ~PAGE_MASK; |
3783689a | 1478 | zspage = get_zspage(f_page); |
61989a80 | 1479 | |
c7806261 | 1480 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1481 | |
1482 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1483 | link = (struct link_free *)(vaddr + f_offset); |
a41ec880 MK |
1484 | if (likely(!ZsHugePage(zspage))) |
1485 | link->next = get_freeobj(zspage) << OBJ_TAG_BITS; | |
1486 | else | |
1487 | f_page->index = 0; | |
af4ee5e9 | 1488 | kunmap_atomic(vaddr); |
bfd093f5 | 1489 | set_freeobj(zspage, f_objidx); |
3783689a | 1490 | mod_zspage_inuse(zspage, -1); |
c7806261 MK |
1491 | } |
1492 | ||
1493 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1494 | { | |
3783689a MK |
1495 | struct zspage *zspage; |
1496 | struct page *f_page; | |
bfd093f5 | 1497 | unsigned long obj; |
c7806261 MK |
1498 | struct size_class *class; |
1499 | enum fullness_group fullness; | |
1500 | ||
1501 | if (unlikely(!handle)) | |
1502 | return; | |
1503 | ||
312fcae2 | 1504 | pin_tag(handle); |
c7806261 | 1505 | obj = handle_to_obj(handle); |
67f1c9cd | 1506 | obj_to_page(obj, &f_page); |
3783689a | 1507 | zspage = get_zspage(f_page); |
c7806261 | 1508 | |
48b4800a | 1509 | migrate_read_lock(zspage); |
67f1c9cd | 1510 | class = zspage_class(pool, zspage); |
c7806261 MK |
1511 | |
1512 | spin_lock(&class->lock); | |
0a5f079b MK |
1513 | obj_free(class->size, obj); |
1514 | class_stat_dec(class, OBJ_USED, 1); | |
3783689a | 1515 | fullness = fix_fullness_group(class, zspage); |
48b4800a MK |
1516 | if (fullness != ZS_EMPTY) { |
1517 | migrate_read_unlock(zspage); | |
1518 | goto out; | |
312fcae2 | 1519 | } |
48b4800a | 1520 | |
48b4800a MK |
1521 | migrate_read_unlock(zspage); |
1522 | /* If zspage is isolated, zs_page_putback will free the zspage */ | |
c4549b87 | 1523 | free_zspage(pool, class, zspage); |
48b4800a MK |
1524 | out: |
1525 | ||
61989a80 | 1526 | spin_unlock(&class->lock); |
312fcae2 | 1527 | unpin_tag(handle); |
3783689a | 1528 | cache_free_handle(pool, handle); |
312fcae2 MK |
1529 | } |
1530 | EXPORT_SYMBOL_GPL(zs_free); | |
1531 | ||
251cbb95 MK |
1532 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1533 | unsigned long src) | |
312fcae2 MK |
1534 | { |
1535 | struct page *s_page, *d_page; | |
bfd093f5 | 1536 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1537 | unsigned long s_off, d_off; |
1538 | void *s_addr, *d_addr; | |
1539 | int s_size, d_size, size; | |
1540 | int written = 0; | |
1541 | ||
1542 | s_size = d_size = class->size; | |
1543 | ||
1544 | obj_to_location(src, &s_page, &s_objidx); | |
1545 | obj_to_location(dst, &d_page, &d_objidx); | |
1546 | ||
bfd093f5 MK |
1547 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1548 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1549 | |
1550 | if (s_off + class->size > PAGE_SIZE) | |
1551 | s_size = PAGE_SIZE - s_off; | |
1552 | ||
1553 | if (d_off + class->size > PAGE_SIZE) | |
1554 | d_size = PAGE_SIZE - d_off; | |
1555 | ||
1556 | s_addr = kmap_atomic(s_page); | |
1557 | d_addr = kmap_atomic(d_page); | |
1558 | ||
1559 | while (1) { | |
1560 | size = min(s_size, d_size); | |
1561 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1562 | written += size; | |
1563 | ||
1564 | if (written == class->size) | |
1565 | break; | |
1566 | ||
495819ea SS |
1567 | s_off += size; |
1568 | s_size -= size; | |
1569 | d_off += size; | |
1570 | d_size -= size; | |
1571 | ||
1572 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1573 | kunmap_atomic(d_addr); |
1574 | kunmap_atomic(s_addr); | |
1575 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1576 | s_addr = kmap_atomic(s_page); |
1577 | d_addr = kmap_atomic(d_page); | |
1578 | s_size = class->size - written; | |
1579 | s_off = 0; | |
312fcae2 MK |
1580 | } |
1581 | ||
495819ea | 1582 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1583 | kunmap_atomic(d_addr); |
1584 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1585 | d_addr = kmap_atomic(d_page); |
1586 | d_size = class->size - written; | |
1587 | d_off = 0; | |
312fcae2 MK |
1588 | } |
1589 | } | |
1590 | ||
1591 | kunmap_atomic(d_addr); | |
1592 | kunmap_atomic(s_addr); | |
1593 | } | |
1594 | ||
1595 | /* | |
1596 | * Find alloced object in zspage from index object and | |
1597 | * return handle. | |
1598 | */ | |
251cbb95 | 1599 | static unsigned long find_alloced_obj(struct size_class *class, |
cf675acb | 1600 | struct page *page, int *obj_idx) |
312fcae2 | 1601 | { |
312fcae2 | 1602 | int offset = 0; |
cf675acb | 1603 | int index = *obj_idx; |
312fcae2 MK |
1604 | unsigned long handle = 0; |
1605 | void *addr = kmap_atomic(page); | |
1606 | ||
3783689a | 1607 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1608 | offset += class->size * index; |
1609 | ||
1610 | while (offset < PAGE_SIZE) { | |
3ae92ac2 | 1611 | if (obj_allocated(page, addr + offset, &handle)) { |
312fcae2 MK |
1612 | if (trypin_tag(handle)) |
1613 | break; | |
1614 | handle = 0; | |
1615 | } | |
1616 | ||
1617 | offset += class->size; | |
1618 | index++; | |
1619 | } | |
1620 | ||
1621 | kunmap_atomic(addr); | |
cf675acb GM |
1622 | |
1623 | *obj_idx = index; | |
1624 | ||
312fcae2 MK |
1625 | return handle; |
1626 | } | |
1627 | ||
1628 | struct zs_compact_control { | |
3783689a | 1629 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1630 | struct page *s_page; |
1631 | /* Destination page for migration which should be a first page | |
1632 | * of zspage. */ | |
1633 | struct page *d_page; | |
1634 | /* Starting object index within @s_page which used for live object | |
1635 | * in the subpage. */ | |
41b88e14 | 1636 | int obj_idx; |
312fcae2 MK |
1637 | }; |
1638 | ||
1639 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1640 | struct zs_compact_control *cc) | |
1641 | { | |
1642 | unsigned long used_obj, free_obj; | |
1643 | unsigned long handle; | |
1644 | struct page *s_page = cc->s_page; | |
1645 | struct page *d_page = cc->d_page; | |
41b88e14 | 1646 | int obj_idx = cc->obj_idx; |
312fcae2 MK |
1647 | int ret = 0; |
1648 | ||
1649 | while (1) { | |
cf675acb | 1650 | handle = find_alloced_obj(class, s_page, &obj_idx); |
312fcae2 MK |
1651 | if (!handle) { |
1652 | s_page = get_next_page(s_page); | |
1653 | if (!s_page) | |
1654 | break; | |
41b88e14 | 1655 | obj_idx = 0; |
312fcae2 MK |
1656 | continue; |
1657 | } | |
1658 | ||
1659 | /* Stop if there is no more space */ | |
3783689a | 1660 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1661 | unpin_tag(handle); |
1662 | ret = -ENOMEM; | |
1663 | break; | |
1664 | } | |
1665 | ||
1666 | used_obj = handle_to_obj(handle); | |
0a5f079b | 1667 | free_obj = obj_malloc(pool, get_zspage(d_page), handle); |
251cbb95 | 1668 | zs_object_copy(class, free_obj, used_obj); |
41b88e14 | 1669 | obj_idx++; |
c102f07c JL |
1670 | /* |
1671 | * record_obj updates handle's value to free_obj and it will | |
1672 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1673 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1674 | * let's keep the lock bit. | |
1675 | */ | |
1676 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1677 | record_obj(handle, free_obj); |
1678 | unpin_tag(handle); | |
0a5f079b | 1679 | obj_free(class->size, used_obj); |
312fcae2 MK |
1680 | } |
1681 | ||
1682 | /* Remember last position in this iteration */ | |
1683 | cc->s_page = s_page; | |
41b88e14 | 1684 | cc->obj_idx = obj_idx; |
312fcae2 MK |
1685 | |
1686 | return ret; | |
1687 | } | |
1688 | ||
3783689a | 1689 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1690 | { |
1691 | int i; | |
3783689a MK |
1692 | struct zspage *zspage; |
1693 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1694 | |
3783689a MK |
1695 | if (!source) { |
1696 | fg[0] = ZS_ALMOST_FULL; | |
1697 | fg[1] = ZS_ALMOST_EMPTY; | |
1698 | } | |
1699 | ||
1700 | for (i = 0; i < 2; i++) { | |
1701 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1702 | struct zspage, list); | |
1703 | if (zspage) { | |
1704 | remove_zspage(class, zspage, fg[i]); | |
1705 | return zspage; | |
312fcae2 MK |
1706 | } |
1707 | } | |
1708 | ||
3783689a | 1709 | return zspage; |
312fcae2 MK |
1710 | } |
1711 | ||
860c707d | 1712 | /* |
3783689a | 1713 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1714 | * @class: destination class |
3783689a | 1715 | * @zspage: target page |
860c707d | 1716 | * |
3783689a | 1717 | * Return @zspage's fullness_group |
860c707d | 1718 | */ |
4aa409ca | 1719 | static enum fullness_group putback_zspage(struct size_class *class, |
3783689a | 1720 | struct zspage *zspage) |
312fcae2 | 1721 | { |
312fcae2 MK |
1722 | enum fullness_group fullness; |
1723 | ||
3783689a MK |
1724 | fullness = get_fullness_group(class, zspage); |
1725 | insert_zspage(class, zspage, fullness); | |
1726 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1727 | |
860c707d | 1728 | return fullness; |
61989a80 | 1729 | } |
312fcae2 | 1730 | |
48b4800a | 1731 | #ifdef CONFIG_COMPACTION |
4d0a5402 CIK |
1732 | /* |
1733 | * To prevent zspage destroy during migration, zspage freeing should | |
1734 | * hold locks of all pages in the zspage. | |
1735 | */ | |
1736 | static void lock_zspage(struct zspage *zspage) | |
1737 | { | |
1738 | struct page *page = get_first_page(zspage); | |
1739 | ||
1740 | do { | |
1741 | lock_page(page); | |
1742 | } while ((page = get_next_page(page)) != NULL); | |
1743 | } | |
1744 | ||
8e9231f8 | 1745 | static int zs_init_fs_context(struct fs_context *fc) |
48b4800a | 1746 | { |
8e9231f8 | 1747 | return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM; |
48b4800a MK |
1748 | } |
1749 | ||
1750 | static struct file_system_type zsmalloc_fs = { | |
1751 | .name = "zsmalloc", | |
8e9231f8 | 1752 | .init_fs_context = zs_init_fs_context, |
48b4800a MK |
1753 | .kill_sb = kill_anon_super, |
1754 | }; | |
1755 | ||
1756 | static int zsmalloc_mount(void) | |
1757 | { | |
1758 | int ret = 0; | |
1759 | ||
1760 | zsmalloc_mnt = kern_mount(&zsmalloc_fs); | |
1761 | if (IS_ERR(zsmalloc_mnt)) | |
1762 | ret = PTR_ERR(zsmalloc_mnt); | |
1763 | ||
1764 | return ret; | |
1765 | } | |
1766 | ||
1767 | static void zsmalloc_unmount(void) | |
1768 | { | |
1769 | kern_unmount(zsmalloc_mnt); | |
1770 | } | |
1771 | ||
1772 | static void migrate_lock_init(struct zspage *zspage) | |
1773 | { | |
1774 | rwlock_init(&zspage->lock); | |
1775 | } | |
1776 | ||
cfc451cf | 1777 | static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock) |
48b4800a MK |
1778 | { |
1779 | read_lock(&zspage->lock); | |
1780 | } | |
1781 | ||
8a374ccc | 1782 | static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock) |
48b4800a MK |
1783 | { |
1784 | read_unlock(&zspage->lock); | |
1785 | } | |
1786 | ||
1787 | static void migrate_write_lock(struct zspage *zspage) | |
1788 | { | |
1789 | write_lock(&zspage->lock); | |
1790 | } | |
1791 | ||
1792 | static void migrate_write_unlock(struct zspage *zspage) | |
1793 | { | |
1794 | write_unlock(&zspage->lock); | |
1795 | } | |
1796 | ||
1797 | /* Number of isolated subpage for *page migration* in this zspage */ | |
1798 | static void inc_zspage_isolation(struct zspage *zspage) | |
1799 | { | |
1800 | zspage->isolated++; | |
1801 | } | |
1802 | ||
1803 | static void dec_zspage_isolation(struct zspage *zspage) | |
1804 | { | |
c4549b87 | 1805 | VM_BUG_ON(zspage->isolated == 0); |
48b4800a MK |
1806 | zspage->isolated--; |
1807 | } | |
1808 | ||
1809 | static void replace_sub_page(struct size_class *class, struct zspage *zspage, | |
1810 | struct page *newpage, struct page *oldpage) | |
1811 | { | |
1812 | struct page *page; | |
1813 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, }; | |
1814 | int idx = 0; | |
1815 | ||
1816 | page = get_first_page(zspage); | |
1817 | do { | |
1818 | if (page == oldpage) | |
1819 | pages[idx] = newpage; | |
1820 | else | |
1821 | pages[idx] = page; | |
1822 | idx++; | |
1823 | } while ((page = get_next_page(page)) != NULL); | |
1824 | ||
1825 | create_page_chain(class, zspage, pages); | |
1826 | set_first_obj_offset(newpage, get_first_obj_offset(oldpage)); | |
a41ec880 | 1827 | if (unlikely(ZsHugePage(zspage))) |
48b4800a MK |
1828 | newpage->index = oldpage->index; |
1829 | __SetPageMovable(newpage, page_mapping(oldpage)); | |
1830 | } | |
1831 | ||
4d0a5402 | 1832 | static bool zs_page_isolate(struct page *page, isolate_mode_t mode) |
48b4800a | 1833 | { |
48b4800a | 1834 | struct zspage *zspage; |
48b4800a MK |
1835 | |
1836 | /* | |
1837 | * Page is locked so zspage couldn't be destroyed. For detail, look at | |
1838 | * lock_zspage in free_zspage. | |
1839 | */ | |
1840 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1841 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1842 | ||
1843 | zspage = get_zspage(page); | |
c4549b87 | 1844 | migrate_write_lock(zspage); |
48b4800a | 1845 | inc_zspage_isolation(zspage); |
c4549b87 | 1846 | migrate_write_unlock(zspage); |
48b4800a MK |
1847 | |
1848 | return true; | |
1849 | } | |
1850 | ||
4d0a5402 | 1851 | static int zs_page_migrate(struct address_space *mapping, struct page *newpage, |
48b4800a MK |
1852 | struct page *page, enum migrate_mode mode) |
1853 | { | |
1854 | struct zs_pool *pool; | |
1855 | struct size_class *class; | |
48b4800a MK |
1856 | struct zspage *zspage; |
1857 | struct page *dummy; | |
1858 | void *s_addr, *d_addr, *addr; | |
1859 | int offset, pos; | |
3ae92ac2 | 1860 | unsigned long handle; |
48b4800a MK |
1861 | unsigned long old_obj, new_obj; |
1862 | unsigned int obj_idx; | |
1863 | int ret = -EAGAIN; | |
1864 | ||
2916ecc0 JG |
1865 | /* |
1866 | * We cannot support the _NO_COPY case here, because copy needs to | |
1867 | * happen under the zs lock, which does not work with | |
1868 | * MIGRATE_SYNC_NO_COPY workflow. | |
1869 | */ | |
1870 | if (mode == MIGRATE_SYNC_NO_COPY) | |
1871 | return -EINVAL; | |
1872 | ||
48b4800a MK |
1873 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
1874 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
1875 | ||
1876 | zspage = get_zspage(page); | |
1877 | ||
1878 | /* Concurrent compactor cannot migrate any subpage in zspage */ | |
1879 | migrate_write_lock(zspage); | |
48b4800a | 1880 | pool = mapping->private_data; |
67f1c9cd | 1881 | class = zspage_class(pool, zspage); |
48b4800a MK |
1882 | offset = get_first_obj_offset(page); |
1883 | ||
1884 | spin_lock(&class->lock); | |
1885 | if (!get_zspage_inuse(zspage)) { | |
77ff4657 HZ |
1886 | /* |
1887 | * Set "offset" to end of the page so that every loops | |
1888 | * skips unnecessary object scanning. | |
1889 | */ | |
1890 | offset = PAGE_SIZE; | |
48b4800a MK |
1891 | } |
1892 | ||
1893 | pos = offset; | |
1894 | s_addr = kmap_atomic(page); | |
1895 | while (pos < PAGE_SIZE) { | |
3ae92ac2 | 1896 | if (obj_allocated(page, s_addr + pos, &handle)) { |
48b4800a MK |
1897 | if (!trypin_tag(handle)) |
1898 | goto unpin_objects; | |
1899 | } | |
1900 | pos += class->size; | |
1901 | } | |
1902 | ||
1903 | /* | |
1904 | * Here, any user cannot access all objects in the zspage so let's move. | |
1905 | */ | |
1906 | d_addr = kmap_atomic(newpage); | |
1907 | memcpy(d_addr, s_addr, PAGE_SIZE); | |
1908 | kunmap_atomic(d_addr); | |
1909 | ||
1910 | for (addr = s_addr + offset; addr < s_addr + pos; | |
1911 | addr += class->size) { | |
3ae92ac2 | 1912 | if (obj_allocated(page, addr, &handle)) { |
ecfc2bda | 1913 | BUG_ON(!testpin_tag(handle)); |
48b4800a MK |
1914 | |
1915 | old_obj = handle_to_obj(handle); | |
1916 | obj_to_location(old_obj, &dummy, &obj_idx); | |
1917 | new_obj = (unsigned long)location_to_obj(newpage, | |
1918 | obj_idx); | |
1919 | new_obj |= BIT(HANDLE_PIN_BIT); | |
1920 | record_obj(handle, new_obj); | |
1921 | } | |
1922 | } | |
1923 | ||
1924 | replace_sub_page(class, zspage, newpage, page); | |
1925 | get_page(newpage); | |
1926 | ||
1927 | dec_zspage_isolation(zspage); | |
1928 | ||
ac8f05da CM |
1929 | if (page_zone(newpage) != page_zone(page)) { |
1930 | dec_zone_page_state(page, NR_ZSPAGES); | |
1931 | inc_zone_page_state(newpage, NR_ZSPAGES); | |
1932 | } | |
1933 | ||
48b4800a MK |
1934 | reset_page(page); |
1935 | put_page(page); | |
1936 | page = newpage; | |
1937 | ||
dd4123f3 | 1938 | ret = MIGRATEPAGE_SUCCESS; |
48b4800a MK |
1939 | unpin_objects: |
1940 | for (addr = s_addr + offset; addr < s_addr + pos; | |
1941 | addr += class->size) { | |
3ae92ac2 | 1942 | if (obj_allocated(page, addr, &handle)) { |
ecfc2bda | 1943 | BUG_ON(!testpin_tag(handle)); |
48b4800a MK |
1944 | unpin_tag(handle); |
1945 | } | |
1946 | } | |
1947 | kunmap_atomic(s_addr); | |
48b4800a MK |
1948 | spin_unlock(&class->lock); |
1949 | migrate_write_unlock(zspage); | |
1950 | ||
1951 | return ret; | |
1952 | } | |
1953 | ||
4d0a5402 | 1954 | static void zs_page_putback(struct page *page) |
48b4800a | 1955 | { |
48b4800a MK |
1956 | struct zspage *zspage; |
1957 | ||
1958 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1959 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
1960 | ||
1961 | zspage = get_zspage(page); | |
c4549b87 | 1962 | migrate_write_lock(zspage); |
48b4800a | 1963 | dec_zspage_isolation(zspage); |
c4549b87 | 1964 | migrate_write_unlock(zspage); |
48b4800a MK |
1965 | } |
1966 | ||
4d0a5402 | 1967 | static const struct address_space_operations zsmalloc_aops = { |
48b4800a MK |
1968 | .isolate_page = zs_page_isolate, |
1969 | .migratepage = zs_page_migrate, | |
1970 | .putback_page = zs_page_putback, | |
1971 | }; | |
1972 | ||
1973 | static int zs_register_migration(struct zs_pool *pool) | |
1974 | { | |
1975 | pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb); | |
1976 | if (IS_ERR(pool->inode)) { | |
1977 | pool->inode = NULL; | |
1978 | return 1; | |
1979 | } | |
1980 | ||
1981 | pool->inode->i_mapping->private_data = pool; | |
1982 | pool->inode->i_mapping->a_ops = &zsmalloc_aops; | |
1983 | return 0; | |
1984 | } | |
1985 | ||
1986 | static void zs_unregister_migration(struct zs_pool *pool) | |
1987 | { | |
1988 | flush_work(&pool->free_work); | |
c3491eca | 1989 | iput(pool->inode); |
48b4800a MK |
1990 | } |
1991 | ||
1992 | /* | |
1993 | * Caller should hold page_lock of all pages in the zspage | |
1994 | * In here, we cannot use zspage meta data. | |
1995 | */ | |
1996 | static void async_free_zspage(struct work_struct *work) | |
1997 | { | |
1998 | int i; | |
1999 | struct size_class *class; | |
2000 | unsigned int class_idx; | |
2001 | enum fullness_group fullness; | |
2002 | struct zspage *zspage, *tmp; | |
2003 | LIST_HEAD(free_pages); | |
2004 | struct zs_pool *pool = container_of(work, struct zs_pool, | |
2005 | free_work); | |
2006 | ||
cf8e0fed | 2007 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
48b4800a MK |
2008 | class = pool->size_class[i]; |
2009 | if (class->index != i) | |
2010 | continue; | |
2011 | ||
2012 | spin_lock(&class->lock); | |
2013 | list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages); | |
2014 | spin_unlock(&class->lock); | |
2015 | } | |
2016 | ||
48b4800a MK |
2017 | list_for_each_entry_safe(zspage, tmp, &free_pages, list) { |
2018 | list_del(&zspage->list); | |
2019 | lock_zspage(zspage); | |
2020 | ||
2021 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2022 | VM_BUG_ON(fullness != ZS_EMPTY); | |
2023 | class = pool->size_class[class_idx]; | |
2024 | spin_lock(&class->lock); | |
33848337 | 2025 | __free_zspage(pool, class, zspage); |
48b4800a MK |
2026 | spin_unlock(&class->lock); |
2027 | } | |
2028 | }; | |
2029 | ||
2030 | static void kick_deferred_free(struct zs_pool *pool) | |
2031 | { | |
2032 | schedule_work(&pool->free_work); | |
2033 | } | |
2034 | ||
2035 | static void init_deferred_free(struct zs_pool *pool) | |
2036 | { | |
2037 | INIT_WORK(&pool->free_work, async_free_zspage); | |
2038 | } | |
2039 | ||
2040 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) | |
2041 | { | |
2042 | struct page *page = get_first_page(zspage); | |
2043 | ||
2044 | do { | |
2045 | WARN_ON(!trylock_page(page)); | |
2046 | __SetPageMovable(page, pool->inode->i_mapping); | |
2047 | unlock_page(page); | |
2048 | } while ((page = get_next_page(page)) != NULL); | |
2049 | } | |
2050 | #endif | |
2051 | ||
04f05909 SS |
2052 | /* |
2053 | * | |
2054 | * Based on the number of unused allocated objects calculate | |
2055 | * and return the number of pages that we can free. | |
04f05909 SS |
2056 | */ |
2057 | static unsigned long zs_can_compact(struct size_class *class) | |
2058 | { | |
2059 | unsigned long obj_wasted; | |
44f43e99 SS |
2060 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
2061 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 2062 | |
44f43e99 SS |
2063 | if (obj_allocated <= obj_used) |
2064 | return 0; | |
04f05909 | 2065 | |
44f43e99 | 2066 | obj_wasted = obj_allocated - obj_used; |
b4fd07a0 | 2067 | obj_wasted /= class->objs_per_zspage; |
04f05909 | 2068 | |
6cbf16b3 | 2069 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
2070 | } |
2071 | ||
23959281 RY |
2072 | static unsigned long __zs_compact(struct zs_pool *pool, |
2073 | struct size_class *class) | |
312fcae2 | 2074 | { |
312fcae2 | 2075 | struct zs_compact_control cc; |
3783689a MK |
2076 | struct zspage *src_zspage; |
2077 | struct zspage *dst_zspage = NULL; | |
23959281 | 2078 | unsigned long pages_freed = 0; |
312fcae2 | 2079 | |
312fcae2 | 2080 | spin_lock(&class->lock); |
3783689a | 2081 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 2082 | |
04f05909 SS |
2083 | if (!zs_can_compact(class)) |
2084 | break; | |
2085 | ||
41b88e14 | 2086 | cc.obj_idx = 0; |
48b4800a | 2087 | cc.s_page = get_first_page(src_zspage); |
312fcae2 | 2088 | |
3783689a | 2089 | while ((dst_zspage = isolate_zspage(class, false))) { |
48b4800a | 2090 | cc.d_page = get_first_page(dst_zspage); |
312fcae2 | 2091 | /* |
0dc63d48 SS |
2092 | * If there is no more space in dst_page, resched |
2093 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
2094 | */ |
2095 | if (!migrate_zspage(pool, class, &cc)) | |
2096 | break; | |
2097 | ||
4aa409ca | 2098 | putback_zspage(class, dst_zspage); |
312fcae2 MK |
2099 | } |
2100 | ||
2101 | /* Stop if we couldn't find slot */ | |
3783689a | 2102 | if (dst_zspage == NULL) |
312fcae2 MK |
2103 | break; |
2104 | ||
4aa409ca MK |
2105 | putback_zspage(class, dst_zspage); |
2106 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { | |
48b4800a | 2107 | free_zspage(pool, class, src_zspage); |
23959281 | 2108 | pages_freed += class->pages_per_zspage; |
4aa409ca | 2109 | } |
312fcae2 | 2110 | spin_unlock(&class->lock); |
312fcae2 MK |
2111 | cond_resched(); |
2112 | spin_lock(&class->lock); | |
2113 | } | |
2114 | ||
3783689a | 2115 | if (src_zspage) |
4aa409ca | 2116 | putback_zspage(class, src_zspage); |
312fcae2 | 2117 | |
7d3f3938 | 2118 | spin_unlock(&class->lock); |
23959281 RY |
2119 | |
2120 | return pages_freed; | |
312fcae2 MK |
2121 | } |
2122 | ||
2123 | unsigned long zs_compact(struct zs_pool *pool) | |
2124 | { | |
2125 | int i; | |
312fcae2 | 2126 | struct size_class *class; |
23959281 | 2127 | unsigned long pages_freed = 0; |
312fcae2 | 2128 | |
cf8e0fed | 2129 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
312fcae2 MK |
2130 | class = pool->size_class[i]; |
2131 | if (!class) | |
2132 | continue; | |
2133 | if (class->index != i) | |
2134 | continue; | |
23959281 | 2135 | pages_freed += __zs_compact(pool, class); |
312fcae2 | 2136 | } |
23959281 | 2137 | atomic_long_add(pages_freed, &pool->stats.pages_compacted); |
312fcae2 | 2138 | |
23959281 | 2139 | return pages_freed; |
312fcae2 MK |
2140 | } |
2141 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 2142 | |
7d3f3938 SS |
2143 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
2144 | { | |
2145 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
2146 | } | |
2147 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
2148 | ||
ab9d306d SS |
2149 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
2150 | struct shrink_control *sc) | |
2151 | { | |
2152 | unsigned long pages_freed; | |
2153 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2154 | shrinker); | |
2155 | ||
ab9d306d SS |
2156 | /* |
2157 | * Compact classes and calculate compaction delta. | |
2158 | * Can run concurrently with a manually triggered | |
2159 | * (by user) compaction. | |
2160 | */ | |
23959281 | 2161 | pages_freed = zs_compact(pool); |
ab9d306d SS |
2162 | |
2163 | return pages_freed ? pages_freed : SHRINK_STOP; | |
2164 | } | |
2165 | ||
2166 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
2167 | struct shrink_control *sc) | |
2168 | { | |
2169 | int i; | |
2170 | struct size_class *class; | |
2171 | unsigned long pages_to_free = 0; | |
2172 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2173 | shrinker); | |
2174 | ||
cf8e0fed | 2175 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
ab9d306d SS |
2176 | class = pool->size_class[i]; |
2177 | if (!class) | |
2178 | continue; | |
2179 | if (class->index != i) | |
2180 | continue; | |
2181 | ||
ab9d306d | 2182 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
2183 | } |
2184 | ||
2185 | return pages_to_free; | |
2186 | } | |
2187 | ||
2188 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
2189 | { | |
93144ca3 | 2190 | unregister_shrinker(&pool->shrinker); |
ab9d306d SS |
2191 | } |
2192 | ||
2193 | static int zs_register_shrinker(struct zs_pool *pool) | |
2194 | { | |
2195 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
2196 | pool->shrinker.count_objects = zs_shrinker_count; | |
2197 | pool->shrinker.batch = 0; | |
2198 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
2199 | ||
2200 | return register_shrinker(&pool->shrinker); | |
2201 | } | |
2202 | ||
00a61d86 | 2203 | /** |
66cdef66 | 2204 | * zs_create_pool - Creates an allocation pool to work from. |
fd854463 | 2205 | * @name: pool name to be created |
166cfda7 | 2206 | * |
66cdef66 GM |
2207 | * This function must be called before anything when using |
2208 | * the zsmalloc allocator. | |
166cfda7 | 2209 | * |
66cdef66 GM |
2210 | * On success, a pointer to the newly created pool is returned, |
2211 | * otherwise NULL. | |
396b7fd6 | 2212 | */ |
d0d8da2d | 2213 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 2214 | { |
66cdef66 GM |
2215 | int i; |
2216 | struct zs_pool *pool; | |
2217 | struct size_class *prev_class = NULL; | |
61989a80 | 2218 | |
66cdef66 GM |
2219 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
2220 | if (!pool) | |
2221 | return NULL; | |
61989a80 | 2222 | |
48b4800a | 2223 | init_deferred_free(pool); |
61989a80 | 2224 | |
2e40e163 MK |
2225 | pool->name = kstrdup(name, GFP_KERNEL); |
2226 | if (!pool->name) | |
2227 | goto err; | |
2228 | ||
3783689a | 2229 | if (create_cache(pool)) |
2e40e163 MK |
2230 | goto err; |
2231 | ||
c60369f0 | 2232 | /* |
399d8eeb | 2233 | * Iterate reversely, because, size of size_class that we want to use |
66cdef66 | 2234 | * for merging should be larger or equal to current size. |
c60369f0 | 2235 | */ |
cf8e0fed | 2236 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
66cdef66 GM |
2237 | int size; |
2238 | int pages_per_zspage; | |
64d90465 | 2239 | int objs_per_zspage; |
66cdef66 | 2240 | struct size_class *class; |
3783689a | 2241 | int fullness = 0; |
c60369f0 | 2242 | |
66cdef66 GM |
2243 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
2244 | if (size > ZS_MAX_ALLOC_SIZE) | |
2245 | size = ZS_MAX_ALLOC_SIZE; | |
2246 | pages_per_zspage = get_pages_per_zspage(size); | |
64d90465 | 2247 | objs_per_zspage = pages_per_zspage * PAGE_SIZE / size; |
61989a80 | 2248 | |
010b495e SS |
2249 | /* |
2250 | * We iterate from biggest down to smallest classes, | |
2251 | * so huge_class_size holds the size of the first huge | |
2252 | * class. Any object bigger than or equal to that will | |
2253 | * endup in the huge class. | |
2254 | */ | |
2255 | if (pages_per_zspage != 1 && objs_per_zspage != 1 && | |
2256 | !huge_class_size) { | |
2257 | huge_class_size = size; | |
2258 | /* | |
2259 | * The object uses ZS_HANDLE_SIZE bytes to store the | |
2260 | * handle. We need to subtract it, because zs_malloc() | |
2261 | * unconditionally adds handle size before it performs | |
2262 | * size class search - so object may be smaller than | |
2263 | * huge class size, yet it still can end up in the huge | |
2264 | * class because it grows by ZS_HANDLE_SIZE extra bytes | |
2265 | * right before class lookup. | |
2266 | */ | |
2267 | huge_class_size -= (ZS_HANDLE_SIZE - 1); | |
2268 | } | |
2269 | ||
66cdef66 GM |
2270 | /* |
2271 | * size_class is used for normal zsmalloc operation such | |
2272 | * as alloc/free for that size. Although it is natural that we | |
2273 | * have one size_class for each size, there is a chance that we | |
2274 | * can get more memory utilization if we use one size_class for | |
2275 | * many different sizes whose size_class have same | |
2276 | * characteristics. So, we makes size_class point to | |
2277 | * previous size_class if possible. | |
2278 | */ | |
2279 | if (prev_class) { | |
64d90465 | 2280 | if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) { |
66cdef66 GM |
2281 | pool->size_class[i] = prev_class; |
2282 | continue; | |
2283 | } | |
2284 | } | |
2285 | ||
2286 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
2287 | if (!class) | |
2288 | goto err; | |
2289 | ||
2290 | class->size = size; | |
2291 | class->index = i; | |
2292 | class->pages_per_zspage = pages_per_zspage; | |
64d90465 | 2293 | class->objs_per_zspage = objs_per_zspage; |
66cdef66 GM |
2294 | spin_lock_init(&class->lock); |
2295 | pool->size_class[i] = class; | |
48b4800a MK |
2296 | for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS; |
2297 | fullness++) | |
3783689a | 2298 | INIT_LIST_HEAD(&class->fullness_list[fullness]); |
66cdef66 GM |
2299 | |
2300 | prev_class = class; | |
61989a80 NG |
2301 | } |
2302 | ||
d34f6157 DS |
2303 | /* debug only, don't abort if it fails */ |
2304 | zs_pool_stat_create(pool, name); | |
0f050d99 | 2305 | |
48b4800a MK |
2306 | if (zs_register_migration(pool)) |
2307 | goto err; | |
2308 | ||
ab9d306d | 2309 | /* |
93144ca3 AK |
2310 | * Not critical since shrinker is only used to trigger internal |
2311 | * defragmentation of the pool which is pretty optional thing. If | |
2312 | * registration fails we still can use the pool normally and user can | |
2313 | * trigger compaction manually. Thus, ignore return code. | |
ab9d306d | 2314 | */ |
93144ca3 AK |
2315 | zs_register_shrinker(pool); |
2316 | ||
66cdef66 GM |
2317 | return pool; |
2318 | ||
2319 | err: | |
2320 | zs_destroy_pool(pool); | |
2321 | return NULL; | |
61989a80 | 2322 | } |
66cdef66 | 2323 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2324 | |
66cdef66 | 2325 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2326 | { |
66cdef66 | 2327 | int i; |
61989a80 | 2328 | |
ab9d306d | 2329 | zs_unregister_shrinker(pool); |
48b4800a | 2330 | zs_unregister_migration(pool); |
0f050d99 GM |
2331 | zs_pool_stat_destroy(pool); |
2332 | ||
cf8e0fed | 2333 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
66cdef66 GM |
2334 | int fg; |
2335 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2336 | |
66cdef66 GM |
2337 | if (!class) |
2338 | continue; | |
61989a80 | 2339 | |
66cdef66 GM |
2340 | if (class->index != i) |
2341 | continue; | |
61989a80 | 2342 | |
48b4800a | 2343 | for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) { |
3783689a | 2344 | if (!list_empty(&class->fullness_list[fg])) { |
66cdef66 GM |
2345 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
2346 | class->size, fg); | |
2347 | } | |
2348 | } | |
2349 | kfree(class); | |
2350 | } | |
f553646a | 2351 | |
3783689a | 2352 | destroy_cache(pool); |
0f050d99 | 2353 | kfree(pool->name); |
66cdef66 GM |
2354 | kfree(pool); |
2355 | } | |
2356 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2357 | |
66cdef66 GM |
2358 | static int __init zs_init(void) |
2359 | { | |
48b4800a MK |
2360 | int ret; |
2361 | ||
2362 | ret = zsmalloc_mount(); | |
2363 | if (ret) | |
2364 | goto out; | |
2365 | ||
215c89d0 SAS |
2366 | ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare", |
2367 | zs_cpu_prepare, zs_cpu_dead); | |
0f050d99 | 2368 | if (ret) |
215c89d0 | 2369 | goto hp_setup_fail; |
66cdef66 | 2370 | |
66cdef66 GM |
2371 | #ifdef CONFIG_ZPOOL |
2372 | zpool_register_driver(&zs_zpool_driver); | |
2373 | #endif | |
0f050d99 | 2374 | |
4abaac9b DS |
2375 | zs_stat_init(); |
2376 | ||
66cdef66 | 2377 | return 0; |
0f050d99 | 2378 | |
215c89d0 | 2379 | hp_setup_fail: |
48b4800a MK |
2380 | zsmalloc_unmount(); |
2381 | out: | |
0f050d99 | 2382 | return ret; |
61989a80 | 2383 | } |
61989a80 | 2384 | |
66cdef66 | 2385 | static void __exit zs_exit(void) |
61989a80 | 2386 | { |
66cdef66 GM |
2387 | #ifdef CONFIG_ZPOOL |
2388 | zpool_unregister_driver(&zs_zpool_driver); | |
2389 | #endif | |
48b4800a | 2390 | zsmalloc_unmount(); |
215c89d0 | 2391 | cpuhp_remove_state(CPUHP_MM_ZS_PREPARE); |
0f050d99 GM |
2392 | |
2393 | zs_stat_exit(); | |
61989a80 | 2394 | } |
069f101f BH |
2395 | |
2396 | module_init(zs_init); | |
2397 | module_exit(zs_exit); | |
2398 | ||
2399 | MODULE_LICENSE("Dual BSD/GPL"); | |
2400 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |