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 |
48b4800a MK |
20 | * page->freelist(index): links together all component pages of a zspage |
21 | * For the huge page, this is always 0, so we use this field | |
22 | * to store handle. | |
fd854463 | 23 | * page->units: 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> | |
42 | #include <asm/tlbflush.h> | |
43 | #include <asm/pgtable.h> | |
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> |
dd4123f3 | 55 | #include <linux/migrate.h> |
48b4800a | 56 | #include <linux/pagemap.h> |
cdc346b3 | 57 | #include <linux/fs.h> |
48b4800a MK |
58 | |
59 | #define ZSPAGE_MAGIC 0x58 | |
0959c63f SJ |
60 | |
61 | /* | |
62 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
63 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
64 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
65 | * to restore link_free pointer values. | |
66 | */ | |
67 | #define ZS_ALIGN 8 | |
68 | ||
69 | /* | |
70 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
71 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
72 | */ | |
73 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
74 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
75 | ||
2e40e163 MK |
76 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
77 | ||
0959c63f SJ |
78 | /* |
79 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 80 | * as single (unsigned long) handle value. |
0959c63f | 81 | * |
bfd093f5 | 82 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
83 | * |
84 | * This is made more complicated by various memory models and PAE. | |
85 | */ | |
86 | ||
02390b87 KS |
87 | #ifndef MAX_POSSIBLE_PHYSMEM_BITS |
88 | #ifdef MAX_PHYSMEM_BITS | |
89 | #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS | |
90 | #else | |
0959c63f SJ |
91 | /* |
92 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
93 | * be PAGE_SHIFT | |
94 | */ | |
02390b87 | 95 | #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG |
0959c63f SJ |
96 | #endif |
97 | #endif | |
02390b87 KS |
98 | |
99 | #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
100 | |
101 | /* | |
102 | * Memory for allocating for handle keeps object position by | |
103 | * encoding <page, obj_idx> and the encoded value has a room | |
104 | * in least bit(ie, look at obj_to_location). | |
105 | * We use the bit to synchronize between object access by | |
106 | * user and migration. | |
107 | */ | |
108 | #define HANDLE_PIN_BIT 0 | |
109 | ||
110 | /* | |
111 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
112 | * to identify the object was allocated or not. | |
113 | * It's okay to add the status bit in the least bit because | |
114 | * header keeps handle which is 4byte-aligned address so we | |
115 | * have room for two bit at least. | |
116 | */ | |
117 | #define OBJ_ALLOCATED_TAG 1 | |
118 | #define OBJ_TAG_BITS 1 | |
119 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
120 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
121 | ||
cf8e0fed JM |
122 | #define FULLNESS_BITS 2 |
123 | #define CLASS_BITS 8 | |
124 | #define ISOLATED_BITS 3 | |
125 | #define MAGIC_VAL_BITS 8 | |
126 | ||
0959c63f SJ |
127 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) |
128 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
129 | #define ZS_MIN_ALLOC_SIZE \ | |
130 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 131 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 132 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
133 | |
134 | /* | |
7eb52512 | 135 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
136 | * trader-off here: |
137 | * - Large number of size classes is potentially wasteful as free page are | |
138 | * spread across these classes | |
139 | * - Small number of size classes causes large internal fragmentation | |
140 | * - Probably its better to use specific size classes (empirically | |
141 | * determined). NOTE: all those class sizes must be set as multiple of | |
142 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
143 | * | |
144 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
145 | * (reason above) | |
146 | */ | |
3783689a | 147 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
cf8e0fed JM |
148 | #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \ |
149 | ZS_SIZE_CLASS_DELTA) + 1) | |
0959c63f | 150 | |
0959c63f | 151 | enum fullness_group { |
0959c63f | 152 | ZS_EMPTY, |
48b4800a MK |
153 | ZS_ALMOST_EMPTY, |
154 | ZS_ALMOST_FULL, | |
155 | ZS_FULL, | |
156 | NR_ZS_FULLNESS, | |
0959c63f SJ |
157 | }; |
158 | ||
0f050d99 | 159 | enum zs_stat_type { |
48b4800a MK |
160 | CLASS_EMPTY, |
161 | CLASS_ALMOST_EMPTY, | |
162 | CLASS_ALMOST_FULL, | |
163 | CLASS_FULL, | |
0f050d99 GM |
164 | OBJ_ALLOCATED, |
165 | OBJ_USED, | |
48b4800a | 166 | NR_ZS_STAT_TYPE, |
0f050d99 GM |
167 | }; |
168 | ||
0f050d99 GM |
169 | struct zs_size_stat { |
170 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
171 | }; | |
172 | ||
57244594 SS |
173 | #ifdef CONFIG_ZSMALLOC_STAT |
174 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
175 | #endif |
176 | ||
48b4800a MK |
177 | #ifdef CONFIG_COMPACTION |
178 | static struct vfsmount *zsmalloc_mnt; | |
179 | #endif | |
180 | ||
0959c63f SJ |
181 | /* |
182 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
183 | * n <= N / f, where | |
184 | * n = number of allocated objects | |
185 | * N = total number of objects zspage can store | |
6dd9737e | 186 | * f = fullness_threshold_frac |
0959c63f SJ |
187 | * |
188 | * Similarly, we assign zspage to: | |
189 | * ZS_ALMOST_FULL when n > N / f | |
190 | * ZS_EMPTY when n == 0 | |
191 | * ZS_FULL when n == N | |
192 | * | |
193 | * (see: fix_fullness_group()) | |
194 | */ | |
195 | static const int fullness_threshold_frac = 4; | |
196 | ||
197 | struct size_class { | |
57244594 | 198 | spinlock_t lock; |
48b4800a | 199 | struct list_head fullness_list[NR_ZS_FULLNESS]; |
0959c63f SJ |
200 | /* |
201 | * Size of objects stored in this class. Must be multiple | |
202 | * of ZS_ALIGN. | |
203 | */ | |
204 | int size; | |
1fc6e27d | 205 | int objs_per_zspage; |
7dfa4612 WY |
206 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
207 | int pages_per_zspage; | |
48b4800a MK |
208 | |
209 | unsigned int index; | |
210 | struct zs_size_stat stats; | |
0959c63f SJ |
211 | }; |
212 | ||
48b4800a MK |
213 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
214 | static void SetPageHugeObject(struct page *page) | |
215 | { | |
216 | SetPageOwnerPriv1(page); | |
217 | } | |
218 | ||
219 | static void ClearPageHugeObject(struct page *page) | |
220 | { | |
221 | ClearPageOwnerPriv1(page); | |
222 | } | |
223 | ||
224 | static int PageHugeObject(struct page *page) | |
225 | { | |
226 | return PageOwnerPriv1(page); | |
227 | } | |
228 | ||
0959c63f SJ |
229 | /* |
230 | * Placed within free objects to form a singly linked list. | |
3783689a | 231 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
232 | * |
233 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
234 | */ | |
235 | struct link_free { | |
2e40e163 MK |
236 | union { |
237 | /* | |
bfd093f5 | 238 | * Free object index; |
2e40e163 MK |
239 | * It's valid for non-allocated object |
240 | */ | |
bfd093f5 | 241 | unsigned long next; |
2e40e163 MK |
242 | /* |
243 | * Handle of allocated object. | |
244 | */ | |
245 | unsigned long handle; | |
246 | }; | |
0959c63f SJ |
247 | }; |
248 | ||
249 | struct zs_pool { | |
6f3526d6 | 250 | const char *name; |
0f050d99 | 251 | |
cf8e0fed | 252 | struct size_class *size_class[ZS_SIZE_CLASSES]; |
2e40e163 | 253 | struct kmem_cache *handle_cachep; |
3783689a | 254 | struct kmem_cache *zspage_cachep; |
0959c63f | 255 | |
13de8933 | 256 | atomic_long_t pages_allocated; |
0f050d99 | 257 | |
7d3f3938 | 258 | struct zs_pool_stats stats; |
ab9d306d SS |
259 | |
260 | /* Compact classes */ | |
261 | struct shrinker shrinker; | |
93144ca3 | 262 | |
0f050d99 GM |
263 | #ifdef CONFIG_ZSMALLOC_STAT |
264 | struct dentry *stat_dentry; | |
265 | #endif | |
48b4800a MK |
266 | #ifdef CONFIG_COMPACTION |
267 | struct inode *inode; | |
268 | struct work_struct free_work; | |
269 | #endif | |
0959c63f | 270 | }; |
61989a80 | 271 | |
3783689a MK |
272 | struct zspage { |
273 | struct { | |
274 | unsigned int fullness:FULLNESS_BITS; | |
85d492f2 | 275 | unsigned int class:CLASS_BITS + 1; |
48b4800a MK |
276 | unsigned int isolated:ISOLATED_BITS; |
277 | unsigned int magic:MAGIC_VAL_BITS; | |
3783689a MK |
278 | }; |
279 | unsigned int inuse; | |
bfd093f5 | 280 | unsigned int freeobj; |
3783689a MK |
281 | struct page *first_page; |
282 | struct list_head list; /* fullness list */ | |
48b4800a MK |
283 | #ifdef CONFIG_COMPACTION |
284 | rwlock_t lock; | |
285 | #endif | |
3783689a | 286 | }; |
61989a80 | 287 | |
f553646a | 288 | struct mapping_area { |
1b945aee | 289 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
290 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
291 | #else | |
292 | char *vm_buf; /* copy buffer for objects that span pages */ | |
293 | #endif | |
294 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
295 | enum zs_mapmode vm_mm; /* mapping mode */ | |
296 | }; | |
297 | ||
48b4800a MK |
298 | #ifdef CONFIG_COMPACTION |
299 | static int zs_register_migration(struct zs_pool *pool); | |
300 | static void zs_unregister_migration(struct zs_pool *pool); | |
301 | static void migrate_lock_init(struct zspage *zspage); | |
302 | static void migrate_read_lock(struct zspage *zspage); | |
303 | static void migrate_read_unlock(struct zspage *zspage); | |
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) {} | |
315 | static void kick_deferred_free(struct zs_pool *pool) {} | |
316 | static void init_deferred_free(struct zs_pool *pool) {} | |
317 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {} | |
318 | #endif | |
319 | ||
3783689a | 320 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
321 | { |
322 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
323 | 0, 0, NULL); | |
3783689a MK |
324 | if (!pool->handle_cachep) |
325 | return 1; | |
326 | ||
327 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
328 | 0, 0, NULL); | |
329 | if (!pool->zspage_cachep) { | |
330 | kmem_cache_destroy(pool->handle_cachep); | |
331 | pool->handle_cachep = NULL; | |
332 | return 1; | |
333 | } | |
334 | ||
335 | return 0; | |
2e40e163 MK |
336 | } |
337 | ||
3783689a | 338 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 339 | { |
cd10add0 | 340 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 341 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
342 | } |
343 | ||
3783689a | 344 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
345 | { |
346 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
48b4800a | 347 | gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
2e40e163 MK |
348 | } |
349 | ||
3783689a | 350 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
351 | { |
352 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
353 | } | |
354 | ||
3783689a MK |
355 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
356 | { | |
48b4800a MK |
357 | return kmem_cache_alloc(pool->zspage_cachep, |
358 | flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); | |
399d8eeb | 359 | } |
3783689a MK |
360 | |
361 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
362 | { | |
363 | kmem_cache_free(pool->zspage_cachep, zspage); | |
364 | } | |
365 | ||
2e40e163 MK |
366 | static void record_obj(unsigned long handle, unsigned long obj) |
367 | { | |
c102f07c JL |
368 | /* |
369 | * lsb of @obj represents handle lock while other bits | |
370 | * represent object value the handle is pointing so | |
371 | * updating shouldn't do store tearing. | |
372 | */ | |
373 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
374 | } |
375 | ||
c795779d DS |
376 | /* zpool driver */ |
377 | ||
378 | #ifdef CONFIG_ZPOOL | |
379 | ||
6f3526d6 | 380 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 381 | const struct zpool_ops *zpool_ops, |
479305fd | 382 | struct zpool *zpool) |
c795779d | 383 | { |
d0d8da2d SS |
384 | /* |
385 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
386 | * different contexts and its caller must provide a valid | |
387 | * gfp mask. | |
388 | */ | |
389 | return zs_create_pool(name); | |
c795779d DS |
390 | } |
391 | ||
392 | static void zs_zpool_destroy(void *pool) | |
393 | { | |
394 | zs_destroy_pool(pool); | |
395 | } | |
396 | ||
397 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
398 | unsigned long *handle) | |
399 | { | |
d0d8da2d | 400 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
401 | return *handle ? 0 : -1; |
402 | } | |
403 | static void zs_zpool_free(void *pool, unsigned long handle) | |
404 | { | |
405 | zs_free(pool, handle); | |
406 | } | |
407 | ||
c795779d DS |
408 | static void *zs_zpool_map(void *pool, unsigned long handle, |
409 | enum zpool_mapmode mm) | |
410 | { | |
411 | enum zs_mapmode zs_mm; | |
412 | ||
413 | switch (mm) { | |
414 | case ZPOOL_MM_RO: | |
415 | zs_mm = ZS_MM_RO; | |
416 | break; | |
417 | case ZPOOL_MM_WO: | |
418 | zs_mm = ZS_MM_WO; | |
419 | break; | |
420 | case ZPOOL_MM_RW: /* fallthru */ | |
421 | default: | |
422 | zs_mm = ZS_MM_RW; | |
423 | break; | |
424 | } | |
425 | ||
426 | return zs_map_object(pool, handle, zs_mm); | |
427 | } | |
428 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
429 | { | |
430 | zs_unmap_object(pool, handle); | |
431 | } | |
432 | ||
433 | static u64 zs_zpool_total_size(void *pool) | |
434 | { | |
722cdc17 | 435 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
436 | } |
437 | ||
438 | static struct zpool_driver zs_zpool_driver = { | |
439 | .type = "zsmalloc", | |
440 | .owner = THIS_MODULE, | |
441 | .create = zs_zpool_create, | |
442 | .destroy = zs_zpool_destroy, | |
443 | .malloc = zs_zpool_malloc, | |
444 | .free = zs_zpool_free, | |
c795779d DS |
445 | .map = zs_zpool_map, |
446 | .unmap = zs_zpool_unmap, | |
447 | .total_size = zs_zpool_total_size, | |
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 | ||
48b4800a MK |
456 | static bool is_zspage_isolated(struct zspage *zspage) |
457 | { | |
458 | return zspage->isolated; | |
459 | } | |
460 | ||
3457f414 | 461 | static __maybe_unused int is_first_page(struct page *page) |
61989a80 | 462 | { |
a27545bf | 463 | return PagePrivate(page); |
61989a80 NG |
464 | } |
465 | ||
48b4800a | 466 | /* Protected by class->lock */ |
3783689a | 467 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 468 | { |
3783689a | 469 | return zspage->inuse; |
4f42047b MK |
470 | } |
471 | ||
3783689a | 472 | static inline void set_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 473 | { |
3783689a | 474 | zspage->inuse = val; |
4f42047b MK |
475 | } |
476 | ||
3783689a | 477 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 478 | { |
3783689a | 479 | zspage->inuse += val; |
4f42047b MK |
480 | } |
481 | ||
48b4800a | 482 | static inline struct page *get_first_page(struct zspage *zspage) |
4f42047b | 483 | { |
48b4800a | 484 | struct page *first_page = zspage->first_page; |
3783689a | 485 | |
48b4800a MK |
486 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
487 | return first_page; | |
4f42047b MK |
488 | } |
489 | ||
48b4800a | 490 | static inline int get_first_obj_offset(struct page *page) |
4f42047b | 491 | { |
48b4800a MK |
492 | return page->units; |
493 | } | |
3783689a | 494 | |
48b4800a MK |
495 | static inline void set_first_obj_offset(struct page *page, int offset) |
496 | { | |
497 | page->units = offset; | |
4f42047b MK |
498 | } |
499 | ||
bfd093f5 | 500 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 501 | { |
bfd093f5 | 502 | return zspage->freeobj; |
4f42047b MK |
503 | } |
504 | ||
bfd093f5 | 505 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 506 | { |
bfd093f5 | 507 | zspage->freeobj = obj; |
4f42047b MK |
508 | } |
509 | ||
3783689a | 510 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 511 | unsigned int *class_idx, |
61989a80 NG |
512 | enum fullness_group *fullness) |
513 | { | |
48b4800a MK |
514 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); |
515 | ||
3783689a MK |
516 | *fullness = zspage->fullness; |
517 | *class_idx = zspage->class; | |
61989a80 NG |
518 | } |
519 | ||
3783689a | 520 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 521 | unsigned int class_idx, |
61989a80 NG |
522 | enum fullness_group fullness) |
523 | { | |
3783689a MK |
524 | zspage->class = class_idx; |
525 | zspage->fullness = fullness; | |
61989a80 NG |
526 | } |
527 | ||
c3e3e88a NC |
528 | /* |
529 | * zsmalloc divides the pool into various size classes where each | |
530 | * class maintains a list of zspages where each zspage is divided | |
531 | * into equal sized chunks. Each allocation falls into one of these | |
532 | * classes depending on its size. This function returns index of the | |
533 | * size class which has chunk size big enough to hold the give size. | |
534 | */ | |
61989a80 NG |
535 | static int get_size_class_index(int size) |
536 | { | |
537 | int idx = 0; | |
538 | ||
539 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
540 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
541 | ZS_SIZE_CLASS_DELTA); | |
542 | ||
cf8e0fed | 543 | return min_t(int, ZS_SIZE_CLASSES - 1, idx); |
61989a80 NG |
544 | } |
545 | ||
3eb95fea | 546 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 547 | static inline void zs_stat_inc(struct size_class *class, |
3eb95fea | 548 | int type, unsigned long cnt) |
248ca1b0 | 549 | { |
48b4800a | 550 | class->stats.objs[type] += cnt; |
248ca1b0 MK |
551 | } |
552 | ||
3eb95fea | 553 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 554 | static inline void zs_stat_dec(struct size_class *class, |
3eb95fea | 555 | int type, unsigned long cnt) |
248ca1b0 | 556 | { |
48b4800a | 557 | class->stats.objs[type] -= cnt; |
248ca1b0 MK |
558 | } |
559 | ||
3eb95fea | 560 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 561 | static inline unsigned long zs_stat_get(struct size_class *class, |
3eb95fea | 562 | int type) |
248ca1b0 | 563 | { |
48b4800a | 564 | return class->stats.objs[type]; |
248ca1b0 MK |
565 | } |
566 | ||
57244594 SS |
567 | #ifdef CONFIG_ZSMALLOC_STAT |
568 | ||
4abaac9b | 569 | static void __init zs_stat_init(void) |
248ca1b0 | 570 | { |
4abaac9b DS |
571 | if (!debugfs_initialized()) { |
572 | pr_warn("debugfs not available, stat dir not created\n"); | |
573 | return; | |
574 | } | |
248ca1b0 MK |
575 | |
576 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
577 | if (!zs_stat_root) | |
4abaac9b | 578 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
579 | } |
580 | ||
581 | static void __exit zs_stat_exit(void) | |
582 | { | |
583 | debugfs_remove_recursive(zs_stat_root); | |
584 | } | |
585 | ||
1120ed54 SS |
586 | static unsigned long zs_can_compact(struct size_class *class); |
587 | ||
248ca1b0 MK |
588 | static int zs_stats_size_show(struct seq_file *s, void *v) |
589 | { | |
590 | int i; | |
591 | struct zs_pool *pool = s->private; | |
592 | struct size_class *class; | |
593 | int objs_per_zspage; | |
594 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 595 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
596 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
597 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 598 | unsigned long total_freeable = 0; |
248ca1b0 | 599 | |
1120ed54 | 600 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
601 | "class", "size", "almost_full", "almost_empty", |
602 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 603 | "pages_per_zspage", "freeable"); |
248ca1b0 | 604 | |
cf8e0fed | 605 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
248ca1b0 MK |
606 | class = pool->size_class[i]; |
607 | ||
608 | if (class->index != i) | |
609 | continue; | |
610 | ||
611 | spin_lock(&class->lock); | |
612 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
613 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
614 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
615 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 616 | freeable = zs_can_compact(class); |
248ca1b0 MK |
617 | spin_unlock(&class->lock); |
618 | ||
b4fd07a0 | 619 | objs_per_zspage = class->objs_per_zspage; |
248ca1b0 MK |
620 | pages_used = obj_allocated / objs_per_zspage * |
621 | class->pages_per_zspage; | |
622 | ||
1120ed54 SS |
623 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
624 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
625 | i, class->size, class_almost_full, class_almost_empty, |
626 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 627 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
628 | |
629 | total_class_almost_full += class_almost_full; | |
630 | total_class_almost_empty += class_almost_empty; | |
631 | total_objs += obj_allocated; | |
632 | total_used_objs += obj_used; | |
633 | total_pages += pages_used; | |
1120ed54 | 634 | total_freeable += freeable; |
248ca1b0 MK |
635 | } |
636 | ||
637 | seq_puts(s, "\n"); | |
1120ed54 | 638 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
639 | "Total", "", total_class_almost_full, |
640 | total_class_almost_empty, total_objs, | |
1120ed54 | 641 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
642 | |
643 | return 0; | |
644 | } | |
645 | ||
646 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
647 | { | |
648 | return single_open(file, zs_stats_size_show, inode->i_private); | |
649 | } | |
650 | ||
651 | static const struct file_operations zs_stat_size_ops = { | |
652 | .open = zs_stats_size_open, | |
653 | .read = seq_read, | |
654 | .llseek = seq_lseek, | |
655 | .release = single_release, | |
656 | }; | |
657 | ||
d34f6157 | 658 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
659 | { |
660 | struct dentry *entry; | |
661 | ||
4abaac9b DS |
662 | if (!zs_stat_root) { |
663 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 664 | return; |
4abaac9b | 665 | } |
248ca1b0 MK |
666 | |
667 | entry = debugfs_create_dir(name, zs_stat_root); | |
668 | if (!entry) { | |
669 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 670 | return; |
248ca1b0 MK |
671 | } |
672 | pool->stat_dentry = entry; | |
673 | ||
674 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
675 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
676 | if (!entry) { | |
677 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
678 | name, "classes"); | |
4abaac9b DS |
679 | debugfs_remove_recursive(pool->stat_dentry); |
680 | pool->stat_dentry = NULL; | |
248ca1b0 | 681 | } |
248ca1b0 MK |
682 | } |
683 | ||
684 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
685 | { | |
686 | debugfs_remove_recursive(pool->stat_dentry); | |
687 | } | |
688 | ||
689 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 690 | static void __init zs_stat_init(void) |
248ca1b0 | 691 | { |
248ca1b0 MK |
692 | } |
693 | ||
694 | static void __exit zs_stat_exit(void) | |
695 | { | |
696 | } | |
697 | ||
d34f6157 | 698 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 699 | { |
248ca1b0 MK |
700 | } |
701 | ||
702 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
703 | { | |
704 | } | |
248ca1b0 MK |
705 | #endif |
706 | ||
48b4800a | 707 | |
c3e3e88a NC |
708 | /* |
709 | * For each size class, zspages are divided into different groups | |
710 | * depending on how "full" they are. This was done so that we could | |
711 | * easily find empty or nearly empty zspages when we try to shrink | |
712 | * the pool (not yet implemented). This function returns fullness | |
713 | * status of the given page. | |
714 | */ | |
1fc6e27d | 715 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 716 | struct zspage *zspage) |
61989a80 | 717 | { |
1fc6e27d | 718 | int inuse, objs_per_zspage; |
61989a80 | 719 | enum fullness_group fg; |
830e4bc5 | 720 | |
3783689a | 721 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 722 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
723 | |
724 | if (inuse == 0) | |
725 | fg = ZS_EMPTY; | |
1fc6e27d | 726 | else if (inuse == objs_per_zspage) |
61989a80 | 727 | fg = ZS_FULL; |
1fc6e27d | 728 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
729 | fg = ZS_ALMOST_EMPTY; |
730 | else | |
731 | fg = ZS_ALMOST_FULL; | |
732 | ||
733 | return fg; | |
734 | } | |
735 | ||
c3e3e88a NC |
736 | /* |
737 | * Each size class maintains various freelists and zspages are assigned | |
738 | * to one of these freelists based on the number of live objects they | |
739 | * have. This functions inserts the given zspage into the freelist | |
740 | * identified by <class, fullness_group>. | |
741 | */ | |
251cbb95 | 742 | static void insert_zspage(struct size_class *class, |
3783689a MK |
743 | struct zspage *zspage, |
744 | enum fullness_group fullness) | |
61989a80 | 745 | { |
3783689a | 746 | struct zspage *head; |
61989a80 | 747 | |
48b4800a | 748 | zs_stat_inc(class, fullness, 1); |
3783689a MK |
749 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
750 | struct zspage, list); | |
58f17117 | 751 | /* |
3783689a MK |
752 | * We want to see more ZS_FULL pages and less almost empty/full. |
753 | * Put pages with higher ->inuse first. | |
58f17117 | 754 | */ |
3783689a MK |
755 | if (head) { |
756 | if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) { | |
757 | list_add(&zspage->list, &head->list); | |
758 | return; | |
759 | } | |
760 | } | |
761 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
61989a80 NG |
762 | } |
763 | ||
c3e3e88a NC |
764 | /* |
765 | * This function removes the given zspage from the freelist identified | |
766 | * by <class, fullness_group>. | |
767 | */ | |
251cbb95 | 768 | static void remove_zspage(struct size_class *class, |
3783689a MK |
769 | struct zspage *zspage, |
770 | enum fullness_group fullness) | |
61989a80 | 771 | { |
3783689a | 772 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
48b4800a | 773 | VM_BUG_ON(is_zspage_isolated(zspage)); |
61989a80 | 774 | |
3783689a | 775 | list_del_init(&zspage->list); |
48b4800a | 776 | zs_stat_dec(class, fullness, 1); |
61989a80 NG |
777 | } |
778 | ||
c3e3e88a NC |
779 | /* |
780 | * Each size class maintains zspages in different fullness groups depending | |
781 | * on the number of live objects they contain. When allocating or freeing | |
782 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
783 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
784 | * a status change has occurred for the given page and accordingly moves the | |
785 | * page from the freelist of the old fullness group to that of the new | |
786 | * fullness group. | |
787 | */ | |
c7806261 | 788 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 789 | struct zspage *zspage) |
61989a80 NG |
790 | { |
791 | int class_idx; | |
61989a80 NG |
792 | enum fullness_group currfg, newfg; |
793 | ||
3783689a MK |
794 | get_zspage_mapping(zspage, &class_idx, &currfg); |
795 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
796 | if (newfg == currfg) |
797 | goto out; | |
798 | ||
48b4800a MK |
799 | if (!is_zspage_isolated(zspage)) { |
800 | remove_zspage(class, zspage, currfg); | |
801 | insert_zspage(class, zspage, newfg); | |
802 | } | |
803 | ||
3783689a | 804 | set_zspage_mapping(zspage, class_idx, newfg); |
61989a80 NG |
805 | |
806 | out: | |
807 | return newfg; | |
808 | } | |
809 | ||
810 | /* | |
811 | * We have to decide on how many pages to link together | |
812 | * to form a zspage for each size class. This is important | |
813 | * to reduce wastage due to unusable space left at end of | |
814 | * each zspage which is given as: | |
888fa374 YX |
815 | * wastage = Zp % class_size |
816 | * usage = Zp - wastage | |
61989a80 NG |
817 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
818 | * | |
819 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
820 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
821 | * since then we can perfectly fit in 8 such objects. | |
822 | */ | |
2e3b6154 | 823 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
824 | { |
825 | int i, max_usedpc = 0; | |
826 | /* zspage order which gives maximum used size per KB */ | |
827 | int max_usedpc_order = 1; | |
828 | ||
84d4faab | 829 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
830 | int zspage_size; |
831 | int waste, usedpc; | |
832 | ||
833 | zspage_size = i * PAGE_SIZE; | |
834 | waste = zspage_size % class_size; | |
835 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
836 | ||
837 | if (usedpc > max_usedpc) { | |
838 | max_usedpc = usedpc; | |
839 | max_usedpc_order = i; | |
840 | } | |
841 | } | |
842 | ||
843 | return max_usedpc_order; | |
844 | } | |
845 | ||
3783689a | 846 | static struct zspage *get_zspage(struct page *page) |
61989a80 | 847 | { |
48b4800a MK |
848 | struct zspage *zspage = (struct zspage *)page->private; |
849 | ||
850 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); | |
851 | return zspage; | |
61989a80 NG |
852 | } |
853 | ||
854 | static struct page *get_next_page(struct page *page) | |
855 | { | |
48b4800a MK |
856 | if (unlikely(PageHugeObject(page))) |
857 | return NULL; | |
858 | ||
859 | return page->freelist; | |
61989a80 NG |
860 | } |
861 | ||
bfd093f5 MK |
862 | /** |
863 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
864 | * @page: page object resides in zspage | |
865 | * @obj_idx: object index | |
67296874 | 866 | */ |
bfd093f5 MK |
867 | static void obj_to_location(unsigned long obj, struct page **page, |
868 | unsigned int *obj_idx) | |
61989a80 | 869 | { |
bfd093f5 MK |
870 | obj >>= OBJ_TAG_BITS; |
871 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
872 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
873 | } | |
61989a80 | 874 | |
bfd093f5 MK |
875 | /** |
876 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
877 | * @page: page object resides in zspage | |
878 | * @obj_idx: object index | |
879 | */ | |
880 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
881 | { | |
882 | unsigned long obj; | |
61989a80 | 883 | |
312fcae2 | 884 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 885 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 886 | obj <<= OBJ_TAG_BITS; |
61989a80 | 887 | |
bfd093f5 | 888 | return obj; |
61989a80 NG |
889 | } |
890 | ||
2e40e163 MK |
891 | static unsigned long handle_to_obj(unsigned long handle) |
892 | { | |
893 | return *(unsigned long *)handle; | |
894 | } | |
895 | ||
48b4800a | 896 | static unsigned long obj_to_head(struct page *page, void *obj) |
312fcae2 | 897 | { |
48b4800a | 898 | if (unlikely(PageHugeObject(page))) { |
830e4bc5 | 899 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3783689a | 900 | return page->index; |
7b60a685 MK |
901 | } else |
902 | return *(unsigned long *)obj; | |
312fcae2 MK |
903 | } |
904 | ||
48b4800a MK |
905 | static inline int testpin_tag(unsigned long handle) |
906 | { | |
907 | return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle); | |
908 | } | |
909 | ||
312fcae2 MK |
910 | static inline int trypin_tag(unsigned long handle) |
911 | { | |
1b8320b6 | 912 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
913 | } |
914 | ||
915 | static void pin_tag(unsigned long handle) | |
916 | { | |
1b8320b6 | 917 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
918 | } |
919 | ||
920 | static void unpin_tag(unsigned long handle) | |
921 | { | |
1b8320b6 | 922 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
923 | } |
924 | ||
f4477e90 NG |
925 | static void reset_page(struct page *page) |
926 | { | |
48b4800a | 927 | __ClearPageMovable(page); |
18fd06bf | 928 | ClearPagePrivate(page); |
f4477e90 | 929 | set_page_private(page, 0); |
48b4800a MK |
930 | page_mapcount_reset(page); |
931 | ClearPageHugeObject(page); | |
932 | page->freelist = NULL; | |
933 | } | |
934 | ||
935 | /* | |
936 | * To prevent zspage destroy during migration, zspage freeing should | |
937 | * hold locks of all pages in the zspage. | |
938 | */ | |
939 | void lock_zspage(struct zspage *zspage) | |
940 | { | |
941 | struct page *page = get_first_page(zspage); | |
942 | ||
943 | do { | |
944 | lock_page(page); | |
945 | } while ((page = get_next_page(page)) != NULL); | |
946 | } | |
947 | ||
948 | int trylock_zspage(struct zspage *zspage) | |
949 | { | |
950 | struct page *cursor, *fail; | |
951 | ||
952 | for (cursor = get_first_page(zspage); cursor != NULL; cursor = | |
953 | get_next_page(cursor)) { | |
954 | if (!trylock_page(cursor)) { | |
955 | fail = cursor; | |
956 | goto unlock; | |
957 | } | |
958 | } | |
959 | ||
960 | return 1; | |
961 | unlock: | |
962 | for (cursor = get_first_page(zspage); cursor != fail; cursor = | |
963 | get_next_page(cursor)) | |
964 | unlock_page(cursor); | |
965 | ||
966 | return 0; | |
f4477e90 NG |
967 | } |
968 | ||
48b4800a MK |
969 | static void __free_zspage(struct zs_pool *pool, struct size_class *class, |
970 | struct zspage *zspage) | |
61989a80 | 971 | { |
3783689a | 972 | struct page *page, *next; |
48b4800a MK |
973 | enum fullness_group fg; |
974 | unsigned int class_idx; | |
975 | ||
976 | get_zspage_mapping(zspage, &class_idx, &fg); | |
977 | ||
978 | assert_spin_locked(&class->lock); | |
61989a80 | 979 | |
3783689a | 980 | VM_BUG_ON(get_zspage_inuse(zspage)); |
48b4800a | 981 | VM_BUG_ON(fg != ZS_EMPTY); |
61989a80 | 982 | |
48b4800a | 983 | next = page = get_first_page(zspage); |
3783689a | 984 | do { |
48b4800a MK |
985 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
986 | next = get_next_page(page); | |
3783689a | 987 | reset_page(page); |
48b4800a | 988 | unlock_page(page); |
91537fee | 989 | dec_zone_page_state(page, NR_ZSPAGES); |
3783689a MK |
990 | put_page(page); |
991 | page = next; | |
992 | } while (page != NULL); | |
61989a80 | 993 | |
3783689a | 994 | cache_free_zspage(pool, zspage); |
48b4800a | 995 | |
b4fd07a0 | 996 | zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
997 | atomic_long_sub(class->pages_per_zspage, |
998 | &pool->pages_allocated); | |
999 | } | |
1000 | ||
1001 | static void free_zspage(struct zs_pool *pool, struct size_class *class, | |
1002 | struct zspage *zspage) | |
1003 | { | |
1004 | VM_BUG_ON(get_zspage_inuse(zspage)); | |
1005 | VM_BUG_ON(list_empty(&zspage->list)); | |
1006 | ||
1007 | if (!trylock_zspage(zspage)) { | |
1008 | kick_deferred_free(pool); | |
1009 | return; | |
1010 | } | |
1011 | ||
1012 | remove_zspage(class, zspage, ZS_EMPTY); | |
1013 | __free_zspage(pool, class, zspage); | |
61989a80 NG |
1014 | } |
1015 | ||
1016 | /* Initialize a newly allocated zspage */ | |
3783689a | 1017 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 1018 | { |
bfd093f5 | 1019 | unsigned int freeobj = 1; |
61989a80 | 1020 | unsigned long off = 0; |
48b4800a | 1021 | struct page *page = get_first_page(zspage); |
830e4bc5 | 1022 | |
61989a80 NG |
1023 | while (page) { |
1024 | struct page *next_page; | |
1025 | struct link_free *link; | |
af4ee5e9 | 1026 | void *vaddr; |
61989a80 | 1027 | |
3783689a | 1028 | set_first_obj_offset(page, off); |
61989a80 | 1029 | |
af4ee5e9 MK |
1030 | vaddr = kmap_atomic(page); |
1031 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
1032 | |
1033 | while ((off += class->size) < PAGE_SIZE) { | |
3b1d9ca6 | 1034 | link->next = freeobj++ << OBJ_TAG_BITS; |
5538c562 | 1035 | link += class->size / sizeof(*link); |
61989a80 NG |
1036 | } |
1037 | ||
1038 | /* | |
1039 | * We now come to the last (full or partial) object on this | |
1040 | * page, which must point to the first object on the next | |
1041 | * page (if present) | |
1042 | */ | |
1043 | next_page = get_next_page(page); | |
bfd093f5 | 1044 | if (next_page) { |
3b1d9ca6 | 1045 | link->next = freeobj++ << OBJ_TAG_BITS; |
bfd093f5 MK |
1046 | } else { |
1047 | /* | |
3b1d9ca6 | 1048 | * Reset OBJ_TAG_BITS bit to last link to tell |
bfd093f5 MK |
1049 | * whether it's allocated object or not. |
1050 | */ | |
01a6ad9a | 1051 | link->next = -1UL << OBJ_TAG_BITS; |
bfd093f5 | 1052 | } |
af4ee5e9 | 1053 | kunmap_atomic(vaddr); |
61989a80 | 1054 | page = next_page; |
5538c562 | 1055 | off %= PAGE_SIZE; |
61989a80 | 1056 | } |
bdb0af7c | 1057 | |
bfd093f5 | 1058 | set_freeobj(zspage, 0); |
61989a80 NG |
1059 | } |
1060 | ||
48b4800a MK |
1061 | static void create_page_chain(struct size_class *class, struct zspage *zspage, |
1062 | struct page *pages[]) | |
61989a80 | 1063 | { |
bdb0af7c MK |
1064 | int i; |
1065 | struct page *page; | |
1066 | struct page *prev_page = NULL; | |
48b4800a | 1067 | int nr_pages = class->pages_per_zspage; |
61989a80 NG |
1068 | |
1069 | /* | |
1070 | * Allocate individual pages and link them together as: | |
48b4800a | 1071 | * 1. all pages are linked together using page->freelist |
3783689a | 1072 | * 2. each sub-page point to zspage using page->private |
61989a80 | 1073 | * |
3783689a | 1074 | * we set PG_private to identify the first page (i.e. no other sub-page |
22c5cef1 | 1075 | * has this flag set). |
61989a80 | 1076 | */ |
bdb0af7c MK |
1077 | for (i = 0; i < nr_pages; i++) { |
1078 | page = pages[i]; | |
3783689a | 1079 | set_page_private(page, (unsigned long)zspage); |
48b4800a | 1080 | page->freelist = NULL; |
bdb0af7c | 1081 | if (i == 0) { |
3783689a | 1082 | zspage->first_page = page; |
a27545bf | 1083 | SetPagePrivate(page); |
48b4800a MK |
1084 | if (unlikely(class->objs_per_zspage == 1 && |
1085 | class->pages_per_zspage == 1)) | |
1086 | SetPageHugeObject(page); | |
3783689a | 1087 | } else { |
48b4800a | 1088 | prev_page->freelist = page; |
61989a80 | 1089 | } |
61989a80 NG |
1090 | prev_page = page; |
1091 | } | |
bdb0af7c | 1092 | } |
61989a80 | 1093 | |
bdb0af7c MK |
1094 | /* |
1095 | * Allocate a zspage for the given size class | |
1096 | */ | |
3783689a MK |
1097 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
1098 | struct size_class *class, | |
1099 | gfp_t gfp) | |
bdb0af7c MK |
1100 | { |
1101 | int i; | |
bdb0af7c | 1102 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1103 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1104 | ||
1105 | if (!zspage) | |
1106 | return NULL; | |
1107 | ||
1108 | memset(zspage, 0, sizeof(struct zspage)); | |
48b4800a MK |
1109 | zspage->magic = ZSPAGE_MAGIC; |
1110 | migrate_lock_init(zspage); | |
61989a80 | 1111 | |
bdb0af7c MK |
1112 | for (i = 0; i < class->pages_per_zspage; i++) { |
1113 | struct page *page; | |
61989a80 | 1114 | |
3783689a | 1115 | page = alloc_page(gfp); |
bdb0af7c | 1116 | if (!page) { |
91537fee MK |
1117 | while (--i >= 0) { |
1118 | dec_zone_page_state(pages[i], NR_ZSPAGES); | |
bdb0af7c | 1119 | __free_page(pages[i]); |
91537fee | 1120 | } |
3783689a | 1121 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1122 | return NULL; |
1123 | } | |
91537fee MK |
1124 | |
1125 | inc_zone_page_state(page, NR_ZSPAGES); | |
bdb0af7c | 1126 | pages[i] = page; |
61989a80 NG |
1127 | } |
1128 | ||
48b4800a | 1129 | create_page_chain(class, zspage, pages); |
3783689a | 1130 | init_zspage(class, zspage); |
bdb0af7c | 1131 | |
3783689a | 1132 | return zspage; |
61989a80 NG |
1133 | } |
1134 | ||
3783689a | 1135 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1136 | { |
1137 | int i; | |
3783689a | 1138 | struct zspage *zspage; |
61989a80 | 1139 | |
48b4800a | 1140 | for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) { |
3783689a MK |
1141 | zspage = list_first_entry_or_null(&class->fullness_list[i], |
1142 | struct zspage, list); | |
1143 | if (zspage) | |
61989a80 NG |
1144 | break; |
1145 | } | |
1146 | ||
3783689a | 1147 | return zspage; |
61989a80 NG |
1148 | } |
1149 | ||
1b945aee | 1150 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1151 | static inline int __zs_cpu_up(struct mapping_area *area) |
1152 | { | |
1153 | /* | |
1154 | * Make sure we don't leak memory if a cpu UP notification | |
1155 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1156 | */ | |
1157 | if (area->vm) | |
1158 | return 0; | |
1159 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1160 | if (!area->vm) | |
1161 | return -ENOMEM; | |
1162 | return 0; | |
1163 | } | |
1164 | ||
1165 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1166 | { | |
1167 | if (area->vm) | |
1168 | free_vm_area(area->vm); | |
1169 | area->vm = NULL; | |
1170 | } | |
1171 | ||
1172 | static inline void *__zs_map_object(struct mapping_area *area, | |
1173 | struct page *pages[2], int off, int size) | |
1174 | { | |
f6f8ed47 | 1175 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1176 | area->vm_addr = area->vm->addr; |
1177 | return area->vm_addr + off; | |
1178 | } | |
1179 | ||
1180 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1181 | struct page *pages[2], int off, int size) | |
1182 | { | |
1183 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1184 | |
d95abbbb | 1185 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1186 | } |
1187 | ||
1b945aee | 1188 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1189 | |
1190 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1191 | { | |
1192 | /* | |
1193 | * Make sure we don't leak memory if a cpu UP notification | |
1194 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1195 | */ | |
1196 | if (area->vm_buf) | |
1197 | return 0; | |
40f9fb8c | 1198 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1199 | if (!area->vm_buf) |
1200 | return -ENOMEM; | |
1201 | return 0; | |
1202 | } | |
1203 | ||
1204 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1205 | { | |
40f9fb8c | 1206 | kfree(area->vm_buf); |
f553646a SJ |
1207 | area->vm_buf = NULL; |
1208 | } | |
1209 | ||
1210 | static void *__zs_map_object(struct mapping_area *area, | |
1211 | struct page *pages[2], int off, int size) | |
5f601902 | 1212 | { |
5f601902 SJ |
1213 | int sizes[2]; |
1214 | void *addr; | |
f553646a | 1215 | char *buf = area->vm_buf; |
5f601902 | 1216 | |
f553646a SJ |
1217 | /* disable page faults to match kmap_atomic() return conditions */ |
1218 | pagefault_disable(); | |
1219 | ||
1220 | /* no read fastpath */ | |
1221 | if (area->vm_mm == ZS_MM_WO) | |
1222 | goto out; | |
5f601902 SJ |
1223 | |
1224 | sizes[0] = PAGE_SIZE - off; | |
1225 | sizes[1] = size - sizes[0]; | |
1226 | ||
5f601902 SJ |
1227 | /* copy object to per-cpu buffer */ |
1228 | addr = kmap_atomic(pages[0]); | |
1229 | memcpy(buf, addr + off, sizes[0]); | |
1230 | kunmap_atomic(addr); | |
1231 | addr = kmap_atomic(pages[1]); | |
1232 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1233 | kunmap_atomic(addr); | |
f553646a SJ |
1234 | out: |
1235 | return area->vm_buf; | |
5f601902 SJ |
1236 | } |
1237 | ||
f553646a SJ |
1238 | static void __zs_unmap_object(struct mapping_area *area, |
1239 | struct page *pages[2], int off, int size) | |
5f601902 | 1240 | { |
5f601902 SJ |
1241 | int sizes[2]; |
1242 | void *addr; | |
2e40e163 | 1243 | char *buf; |
5f601902 | 1244 | |
f553646a SJ |
1245 | /* no write fastpath */ |
1246 | if (area->vm_mm == ZS_MM_RO) | |
1247 | goto out; | |
5f601902 | 1248 | |
7b60a685 | 1249 | buf = area->vm_buf; |
a82cbf07 YX |
1250 | buf = buf + ZS_HANDLE_SIZE; |
1251 | size -= ZS_HANDLE_SIZE; | |
1252 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1253 | |
5f601902 SJ |
1254 | sizes[0] = PAGE_SIZE - off; |
1255 | sizes[1] = size - sizes[0]; | |
1256 | ||
1257 | /* copy per-cpu buffer to object */ | |
1258 | addr = kmap_atomic(pages[0]); | |
1259 | memcpy(addr + off, buf, sizes[0]); | |
1260 | kunmap_atomic(addr); | |
1261 | addr = kmap_atomic(pages[1]); | |
1262 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1263 | kunmap_atomic(addr); | |
f553646a SJ |
1264 | |
1265 | out: | |
1266 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1267 | pagefault_enable(); | |
5f601902 | 1268 | } |
61989a80 | 1269 | |
1b945aee | 1270 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1271 | |
215c89d0 | 1272 | static int zs_cpu_prepare(unsigned int cpu) |
61989a80 | 1273 | { |
61989a80 NG |
1274 | struct mapping_area *area; |
1275 | ||
215c89d0 SAS |
1276 | area = &per_cpu(zs_map_area, cpu); |
1277 | return __zs_cpu_up(area); | |
61989a80 NG |
1278 | } |
1279 | ||
215c89d0 | 1280 | static int zs_cpu_dead(unsigned int cpu) |
61989a80 | 1281 | { |
215c89d0 | 1282 | struct mapping_area *area; |
40f9fb8c | 1283 | |
215c89d0 SAS |
1284 | area = &per_cpu(zs_map_area, cpu); |
1285 | __zs_cpu_down(area); | |
1286 | return 0; | |
b1b00a5b SS |
1287 | } |
1288 | ||
64d90465 GM |
1289 | static bool can_merge(struct size_class *prev, int pages_per_zspage, |
1290 | int objs_per_zspage) | |
9eec4cd5 | 1291 | { |
64d90465 GM |
1292 | if (prev->pages_per_zspage == pages_per_zspage && |
1293 | prev->objs_per_zspage == objs_per_zspage) | |
1294 | return true; | |
9eec4cd5 | 1295 | |
64d90465 | 1296 | return false; |
9eec4cd5 JK |
1297 | } |
1298 | ||
3783689a | 1299 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1300 | { |
3783689a | 1301 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1302 | } |
1303 | ||
66cdef66 GM |
1304 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1305 | { | |
1306 | return atomic_long_read(&pool->pages_allocated); | |
1307 | } | |
1308 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1309 | ||
4bbc0bc0 | 1310 | /** |
66cdef66 GM |
1311 | * zs_map_object - get address of allocated object from handle. |
1312 | * @pool: pool from which the object was allocated | |
1313 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1314 | * |
66cdef66 GM |
1315 | * Before using an object allocated from zs_malloc, it must be mapped using |
1316 | * this function. When done with the object, it must be unmapped using | |
1317 | * zs_unmap_object. | |
4bbc0bc0 | 1318 | * |
66cdef66 GM |
1319 | * Only one object can be mapped per cpu at a time. There is no protection |
1320 | * against nested mappings. | |
1321 | * | |
1322 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1323 | */ |
66cdef66 GM |
1324 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1325 | enum zs_mapmode mm) | |
61989a80 | 1326 | { |
3783689a | 1327 | struct zspage *zspage; |
66cdef66 | 1328 | struct page *page; |
bfd093f5 MK |
1329 | unsigned long obj, off; |
1330 | unsigned int obj_idx; | |
61989a80 | 1331 | |
66cdef66 GM |
1332 | unsigned int class_idx; |
1333 | enum fullness_group fg; | |
1334 | struct size_class *class; | |
1335 | struct mapping_area *area; | |
1336 | struct page *pages[2]; | |
2e40e163 | 1337 | void *ret; |
61989a80 | 1338 | |
9eec4cd5 | 1339 | /* |
66cdef66 GM |
1340 | * Because we use per-cpu mapping areas shared among the |
1341 | * pools/users, we can't allow mapping in interrupt context | |
1342 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1343 | */ |
1aedcafb | 1344 | BUG_ON(in_interrupt()); |
61989a80 | 1345 | |
312fcae2 MK |
1346 | /* From now on, migration cannot move the object */ |
1347 | pin_tag(handle); | |
1348 | ||
2e40e163 MK |
1349 | obj = handle_to_obj(handle); |
1350 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1351 | zspage = get_zspage(page); |
48b4800a MK |
1352 | |
1353 | /* migration cannot move any subpage in this zspage */ | |
1354 | migrate_read_lock(zspage); | |
1355 | ||
3783689a | 1356 | get_zspage_mapping(zspage, &class_idx, &fg); |
66cdef66 | 1357 | class = pool->size_class[class_idx]; |
bfd093f5 | 1358 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1359 | |
66cdef66 GM |
1360 | area = &get_cpu_var(zs_map_area); |
1361 | area->vm_mm = mm; | |
1362 | if (off + class->size <= PAGE_SIZE) { | |
1363 | /* this object is contained entirely within a page */ | |
1364 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1365 | ret = area->vm_addr + off; |
1366 | goto out; | |
61989a80 NG |
1367 | } |
1368 | ||
66cdef66 GM |
1369 | /* this object spans two pages */ |
1370 | pages[0] = page; | |
1371 | pages[1] = get_next_page(page); | |
1372 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1373 | |
2e40e163 MK |
1374 | ret = __zs_map_object(area, pages, off, class->size); |
1375 | out: | |
48b4800a | 1376 | if (likely(!PageHugeObject(page))) |
7b60a685 MK |
1377 | ret += ZS_HANDLE_SIZE; |
1378 | ||
1379 | return ret; | |
61989a80 | 1380 | } |
66cdef66 | 1381 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1382 | |
66cdef66 | 1383 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1384 | { |
3783689a | 1385 | struct zspage *zspage; |
66cdef66 | 1386 | struct page *page; |
bfd093f5 MK |
1387 | unsigned long obj, off; |
1388 | unsigned int obj_idx; | |
61989a80 | 1389 | |
66cdef66 GM |
1390 | unsigned int class_idx; |
1391 | enum fullness_group fg; | |
1392 | struct size_class *class; | |
1393 | struct mapping_area *area; | |
9eec4cd5 | 1394 | |
2e40e163 MK |
1395 | obj = handle_to_obj(handle); |
1396 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1397 | zspage = get_zspage(page); |
1398 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 | 1399 | class = pool->size_class[class_idx]; |
bfd093f5 | 1400 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1401 | |
66cdef66 GM |
1402 | area = this_cpu_ptr(&zs_map_area); |
1403 | if (off + class->size <= PAGE_SIZE) | |
1404 | kunmap_atomic(area->vm_addr); | |
1405 | else { | |
1406 | struct page *pages[2]; | |
40f9fb8c | 1407 | |
66cdef66 GM |
1408 | pages[0] = page; |
1409 | pages[1] = get_next_page(page); | |
1410 | BUG_ON(!pages[1]); | |
1411 | ||
1412 | __zs_unmap_object(area, pages, off, class->size); | |
1413 | } | |
1414 | put_cpu_var(zs_map_area); | |
48b4800a MK |
1415 | |
1416 | migrate_read_unlock(zspage); | |
312fcae2 | 1417 | unpin_tag(handle); |
61989a80 | 1418 | } |
66cdef66 | 1419 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1420 | |
251cbb95 | 1421 | static unsigned long obj_malloc(struct size_class *class, |
3783689a | 1422 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1423 | { |
bfd093f5 | 1424 | int i, nr_page, offset; |
c7806261 MK |
1425 | unsigned long obj; |
1426 | struct link_free *link; | |
1427 | ||
1428 | struct page *m_page; | |
bfd093f5 | 1429 | unsigned long m_offset; |
c7806261 MK |
1430 | void *vaddr; |
1431 | ||
312fcae2 | 1432 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1433 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1434 | |
1435 | offset = obj * class->size; | |
1436 | nr_page = offset >> PAGE_SHIFT; | |
1437 | m_offset = offset & ~PAGE_MASK; | |
1438 | m_page = get_first_page(zspage); | |
1439 | ||
1440 | for (i = 0; i < nr_page; i++) | |
1441 | m_page = get_next_page(m_page); | |
c7806261 MK |
1442 | |
1443 | vaddr = kmap_atomic(m_page); | |
1444 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3b1d9ca6 | 1445 | set_freeobj(zspage, link->next >> OBJ_TAG_BITS); |
48b4800a | 1446 | if (likely(!PageHugeObject(m_page))) |
7b60a685 MK |
1447 | /* record handle in the header of allocated chunk */ |
1448 | link->handle = handle; | |
1449 | else | |
3783689a MK |
1450 | /* record handle to page->index */ |
1451 | zspage->first_page->index = handle; | |
1452 | ||
c7806261 | 1453 | kunmap_atomic(vaddr); |
3783689a | 1454 | mod_zspage_inuse(zspage, 1); |
c7806261 MK |
1455 | zs_stat_inc(class, OBJ_USED, 1); |
1456 | ||
bfd093f5 MK |
1457 | obj = location_to_obj(m_page, obj); |
1458 | ||
c7806261 MK |
1459 | return obj; |
1460 | } | |
1461 | ||
1462 | ||
61989a80 NG |
1463 | /** |
1464 | * zs_malloc - Allocate block of given size from pool. | |
1465 | * @pool: pool to allocate from | |
1466 | * @size: size of block to allocate | |
fd854463 | 1467 | * @gfp: gfp flags when allocating object |
61989a80 | 1468 | * |
00a61d86 | 1469 | * On success, handle to the allocated object is returned, |
c2344348 | 1470 | * otherwise 0. |
61989a80 NG |
1471 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1472 | */ | |
d0d8da2d | 1473 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1474 | { |
2e40e163 | 1475 | unsigned long handle, obj; |
61989a80 | 1476 | struct size_class *class; |
48b4800a | 1477 | enum fullness_group newfg; |
3783689a | 1478 | struct zspage *zspage; |
61989a80 | 1479 | |
7b60a685 | 1480 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1481 | return 0; |
1482 | ||
3783689a | 1483 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1484 | if (!handle) |
c2344348 | 1485 | return 0; |
61989a80 | 1486 | |
2e40e163 MK |
1487 | /* extra space in chunk to keep the handle */ |
1488 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1489 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1490 | |
1491 | spin_lock(&class->lock); | |
3783689a | 1492 | zspage = find_get_zspage(class); |
48b4800a MK |
1493 | if (likely(zspage)) { |
1494 | obj = obj_malloc(class, zspage, handle); | |
1495 | /* Now move the zspage to another fullness group, if required */ | |
1496 | fix_fullness_group(class, zspage); | |
1497 | record_obj(handle, obj); | |
61989a80 | 1498 | spin_unlock(&class->lock); |
61989a80 | 1499 | |
48b4800a MK |
1500 | return handle; |
1501 | } | |
0f050d99 | 1502 | |
48b4800a MK |
1503 | spin_unlock(&class->lock); |
1504 | ||
1505 | zspage = alloc_zspage(pool, class, gfp); | |
1506 | if (!zspage) { | |
1507 | cache_free_handle(pool, handle); | |
1508 | return 0; | |
61989a80 NG |
1509 | } |
1510 | ||
48b4800a | 1511 | spin_lock(&class->lock); |
3783689a | 1512 | obj = obj_malloc(class, zspage, handle); |
48b4800a MK |
1513 | newfg = get_fullness_group(class, zspage); |
1514 | insert_zspage(class, zspage, newfg); | |
1515 | set_zspage_mapping(zspage, class->index, newfg); | |
2e40e163 | 1516 | record_obj(handle, obj); |
48b4800a MK |
1517 | atomic_long_add(class->pages_per_zspage, |
1518 | &pool->pages_allocated); | |
b4fd07a0 | 1519 | zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
1520 | |
1521 | /* We completely set up zspage so mark them as movable */ | |
1522 | SetZsPageMovable(pool, zspage); | |
61989a80 NG |
1523 | spin_unlock(&class->lock); |
1524 | ||
2e40e163 | 1525 | return handle; |
61989a80 NG |
1526 | } |
1527 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1528 | ||
1ee47165 | 1529 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1530 | { |
1531 | struct link_free *link; | |
3783689a MK |
1532 | struct zspage *zspage; |
1533 | struct page *f_page; | |
bfd093f5 MK |
1534 | unsigned long f_offset; |
1535 | unsigned int f_objidx; | |
af4ee5e9 | 1536 | void *vaddr; |
61989a80 | 1537 | |
312fcae2 | 1538 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1539 | obj_to_location(obj, &f_page, &f_objidx); |
bfd093f5 | 1540 | f_offset = (class->size * f_objidx) & ~PAGE_MASK; |
3783689a | 1541 | zspage = get_zspage(f_page); |
61989a80 | 1542 | |
c7806261 | 1543 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1544 | |
1545 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1546 | link = (struct link_free *)(vaddr + f_offset); |
3b1d9ca6 | 1547 | link->next = get_freeobj(zspage) << OBJ_TAG_BITS; |
af4ee5e9 | 1548 | kunmap_atomic(vaddr); |
bfd093f5 | 1549 | set_freeobj(zspage, f_objidx); |
3783689a | 1550 | mod_zspage_inuse(zspage, -1); |
0f050d99 | 1551 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1552 | } |
1553 | ||
1554 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1555 | { | |
3783689a MK |
1556 | struct zspage *zspage; |
1557 | struct page *f_page; | |
bfd093f5 MK |
1558 | unsigned long obj; |
1559 | unsigned int f_objidx; | |
c7806261 MK |
1560 | int class_idx; |
1561 | struct size_class *class; | |
1562 | enum fullness_group fullness; | |
48b4800a | 1563 | bool isolated; |
c7806261 MK |
1564 | |
1565 | if (unlikely(!handle)) | |
1566 | return; | |
1567 | ||
312fcae2 | 1568 | pin_tag(handle); |
c7806261 | 1569 | obj = handle_to_obj(handle); |
c7806261 | 1570 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1571 | zspage = get_zspage(f_page); |
c7806261 | 1572 | |
48b4800a MK |
1573 | migrate_read_lock(zspage); |
1574 | ||
3783689a | 1575 | get_zspage_mapping(zspage, &class_idx, &fullness); |
c7806261 MK |
1576 | class = pool->size_class[class_idx]; |
1577 | ||
1578 | spin_lock(&class->lock); | |
1ee47165 | 1579 | obj_free(class, obj); |
3783689a | 1580 | fullness = fix_fullness_group(class, zspage); |
48b4800a MK |
1581 | if (fullness != ZS_EMPTY) { |
1582 | migrate_read_unlock(zspage); | |
1583 | goto out; | |
312fcae2 | 1584 | } |
48b4800a MK |
1585 | |
1586 | isolated = is_zspage_isolated(zspage); | |
1587 | migrate_read_unlock(zspage); | |
1588 | /* If zspage is isolated, zs_page_putback will free the zspage */ | |
1589 | if (likely(!isolated)) | |
1590 | free_zspage(pool, class, zspage); | |
1591 | out: | |
1592 | ||
61989a80 | 1593 | spin_unlock(&class->lock); |
312fcae2 | 1594 | unpin_tag(handle); |
3783689a | 1595 | cache_free_handle(pool, handle); |
312fcae2 MK |
1596 | } |
1597 | EXPORT_SYMBOL_GPL(zs_free); | |
1598 | ||
251cbb95 MK |
1599 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1600 | unsigned long src) | |
312fcae2 MK |
1601 | { |
1602 | struct page *s_page, *d_page; | |
bfd093f5 | 1603 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1604 | unsigned long s_off, d_off; |
1605 | void *s_addr, *d_addr; | |
1606 | int s_size, d_size, size; | |
1607 | int written = 0; | |
1608 | ||
1609 | s_size = d_size = class->size; | |
1610 | ||
1611 | obj_to_location(src, &s_page, &s_objidx); | |
1612 | obj_to_location(dst, &d_page, &d_objidx); | |
1613 | ||
bfd093f5 MK |
1614 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1615 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1616 | |
1617 | if (s_off + class->size > PAGE_SIZE) | |
1618 | s_size = PAGE_SIZE - s_off; | |
1619 | ||
1620 | if (d_off + class->size > PAGE_SIZE) | |
1621 | d_size = PAGE_SIZE - d_off; | |
1622 | ||
1623 | s_addr = kmap_atomic(s_page); | |
1624 | d_addr = kmap_atomic(d_page); | |
1625 | ||
1626 | while (1) { | |
1627 | size = min(s_size, d_size); | |
1628 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1629 | written += size; | |
1630 | ||
1631 | if (written == class->size) | |
1632 | break; | |
1633 | ||
495819ea SS |
1634 | s_off += size; |
1635 | s_size -= size; | |
1636 | d_off += size; | |
1637 | d_size -= size; | |
1638 | ||
1639 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1640 | kunmap_atomic(d_addr); |
1641 | kunmap_atomic(s_addr); | |
1642 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1643 | s_addr = kmap_atomic(s_page); |
1644 | d_addr = kmap_atomic(d_page); | |
1645 | s_size = class->size - written; | |
1646 | s_off = 0; | |
312fcae2 MK |
1647 | } |
1648 | ||
495819ea | 1649 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1650 | kunmap_atomic(d_addr); |
1651 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1652 | d_addr = kmap_atomic(d_page); |
1653 | d_size = class->size - written; | |
1654 | d_off = 0; | |
312fcae2 MK |
1655 | } |
1656 | } | |
1657 | ||
1658 | kunmap_atomic(d_addr); | |
1659 | kunmap_atomic(s_addr); | |
1660 | } | |
1661 | ||
1662 | /* | |
1663 | * Find alloced object in zspage from index object and | |
1664 | * return handle. | |
1665 | */ | |
251cbb95 | 1666 | static unsigned long find_alloced_obj(struct size_class *class, |
cf675acb | 1667 | struct page *page, int *obj_idx) |
312fcae2 MK |
1668 | { |
1669 | unsigned long head; | |
1670 | int offset = 0; | |
cf675acb | 1671 | int index = *obj_idx; |
312fcae2 MK |
1672 | unsigned long handle = 0; |
1673 | void *addr = kmap_atomic(page); | |
1674 | ||
3783689a | 1675 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1676 | offset += class->size * index; |
1677 | ||
1678 | while (offset < PAGE_SIZE) { | |
48b4800a | 1679 | head = obj_to_head(page, addr + offset); |
312fcae2 MK |
1680 | if (head & OBJ_ALLOCATED_TAG) { |
1681 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1682 | if (trypin_tag(handle)) | |
1683 | break; | |
1684 | handle = 0; | |
1685 | } | |
1686 | ||
1687 | offset += class->size; | |
1688 | index++; | |
1689 | } | |
1690 | ||
1691 | kunmap_atomic(addr); | |
cf675acb GM |
1692 | |
1693 | *obj_idx = index; | |
1694 | ||
312fcae2 MK |
1695 | return handle; |
1696 | } | |
1697 | ||
1698 | struct zs_compact_control { | |
3783689a | 1699 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1700 | struct page *s_page; |
1701 | /* Destination page for migration which should be a first page | |
1702 | * of zspage. */ | |
1703 | struct page *d_page; | |
1704 | /* Starting object index within @s_page which used for live object | |
1705 | * in the subpage. */ | |
41b88e14 | 1706 | int obj_idx; |
312fcae2 MK |
1707 | }; |
1708 | ||
1709 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1710 | struct zs_compact_control *cc) | |
1711 | { | |
1712 | unsigned long used_obj, free_obj; | |
1713 | unsigned long handle; | |
1714 | struct page *s_page = cc->s_page; | |
1715 | struct page *d_page = cc->d_page; | |
41b88e14 | 1716 | int obj_idx = cc->obj_idx; |
312fcae2 MK |
1717 | int ret = 0; |
1718 | ||
1719 | while (1) { | |
cf675acb | 1720 | handle = find_alloced_obj(class, s_page, &obj_idx); |
312fcae2 MK |
1721 | if (!handle) { |
1722 | s_page = get_next_page(s_page); | |
1723 | if (!s_page) | |
1724 | break; | |
41b88e14 | 1725 | obj_idx = 0; |
312fcae2 MK |
1726 | continue; |
1727 | } | |
1728 | ||
1729 | /* Stop if there is no more space */ | |
3783689a | 1730 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1731 | unpin_tag(handle); |
1732 | ret = -ENOMEM; | |
1733 | break; | |
1734 | } | |
1735 | ||
1736 | used_obj = handle_to_obj(handle); | |
3783689a | 1737 | free_obj = obj_malloc(class, get_zspage(d_page), handle); |
251cbb95 | 1738 | zs_object_copy(class, free_obj, used_obj); |
41b88e14 | 1739 | obj_idx++; |
c102f07c JL |
1740 | /* |
1741 | * record_obj updates handle's value to free_obj and it will | |
1742 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1743 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1744 | * let's keep the lock bit. | |
1745 | */ | |
1746 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1747 | record_obj(handle, free_obj); |
1748 | unpin_tag(handle); | |
1ee47165 | 1749 | obj_free(class, used_obj); |
312fcae2 MK |
1750 | } |
1751 | ||
1752 | /* Remember last position in this iteration */ | |
1753 | cc->s_page = s_page; | |
41b88e14 | 1754 | cc->obj_idx = obj_idx; |
312fcae2 MK |
1755 | |
1756 | return ret; | |
1757 | } | |
1758 | ||
3783689a | 1759 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1760 | { |
1761 | int i; | |
3783689a MK |
1762 | struct zspage *zspage; |
1763 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1764 | |
3783689a MK |
1765 | if (!source) { |
1766 | fg[0] = ZS_ALMOST_FULL; | |
1767 | fg[1] = ZS_ALMOST_EMPTY; | |
1768 | } | |
1769 | ||
1770 | for (i = 0; i < 2; i++) { | |
1771 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1772 | struct zspage, list); | |
1773 | if (zspage) { | |
48b4800a | 1774 | VM_BUG_ON(is_zspage_isolated(zspage)); |
3783689a MK |
1775 | remove_zspage(class, zspage, fg[i]); |
1776 | return zspage; | |
312fcae2 MK |
1777 | } |
1778 | } | |
1779 | ||
3783689a | 1780 | return zspage; |
312fcae2 MK |
1781 | } |
1782 | ||
860c707d | 1783 | /* |
3783689a | 1784 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1785 | * @class: destination class |
3783689a | 1786 | * @zspage: target page |
860c707d | 1787 | * |
3783689a | 1788 | * Return @zspage's fullness_group |
860c707d | 1789 | */ |
4aa409ca | 1790 | static enum fullness_group putback_zspage(struct size_class *class, |
3783689a | 1791 | struct zspage *zspage) |
312fcae2 | 1792 | { |
312fcae2 MK |
1793 | enum fullness_group fullness; |
1794 | ||
48b4800a MK |
1795 | VM_BUG_ON(is_zspage_isolated(zspage)); |
1796 | ||
3783689a MK |
1797 | fullness = get_fullness_group(class, zspage); |
1798 | insert_zspage(class, zspage, fullness); | |
1799 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1800 | |
860c707d | 1801 | return fullness; |
61989a80 | 1802 | } |
312fcae2 | 1803 | |
48b4800a MK |
1804 | #ifdef CONFIG_COMPACTION |
1805 | static struct dentry *zs_mount(struct file_system_type *fs_type, | |
1806 | int flags, const char *dev_name, void *data) | |
1807 | { | |
1808 | static const struct dentry_operations ops = { | |
1809 | .d_dname = simple_dname, | |
1810 | }; | |
1811 | ||
1812 | return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC); | |
1813 | } | |
1814 | ||
1815 | static struct file_system_type zsmalloc_fs = { | |
1816 | .name = "zsmalloc", | |
1817 | .mount = zs_mount, | |
1818 | .kill_sb = kill_anon_super, | |
1819 | }; | |
1820 | ||
1821 | static int zsmalloc_mount(void) | |
1822 | { | |
1823 | int ret = 0; | |
1824 | ||
1825 | zsmalloc_mnt = kern_mount(&zsmalloc_fs); | |
1826 | if (IS_ERR(zsmalloc_mnt)) | |
1827 | ret = PTR_ERR(zsmalloc_mnt); | |
1828 | ||
1829 | return ret; | |
1830 | } | |
1831 | ||
1832 | static void zsmalloc_unmount(void) | |
1833 | { | |
1834 | kern_unmount(zsmalloc_mnt); | |
1835 | } | |
1836 | ||
1837 | static void migrate_lock_init(struct zspage *zspage) | |
1838 | { | |
1839 | rwlock_init(&zspage->lock); | |
1840 | } | |
1841 | ||
1842 | static void migrate_read_lock(struct zspage *zspage) | |
1843 | { | |
1844 | read_lock(&zspage->lock); | |
1845 | } | |
1846 | ||
1847 | static void migrate_read_unlock(struct zspage *zspage) | |
1848 | { | |
1849 | read_unlock(&zspage->lock); | |
1850 | } | |
1851 | ||
1852 | static void migrate_write_lock(struct zspage *zspage) | |
1853 | { | |
1854 | write_lock(&zspage->lock); | |
1855 | } | |
1856 | ||
1857 | static void migrate_write_unlock(struct zspage *zspage) | |
1858 | { | |
1859 | write_unlock(&zspage->lock); | |
1860 | } | |
1861 | ||
1862 | /* Number of isolated subpage for *page migration* in this zspage */ | |
1863 | static void inc_zspage_isolation(struct zspage *zspage) | |
1864 | { | |
1865 | zspage->isolated++; | |
1866 | } | |
1867 | ||
1868 | static void dec_zspage_isolation(struct zspage *zspage) | |
1869 | { | |
1870 | zspage->isolated--; | |
1871 | } | |
1872 | ||
1873 | static void replace_sub_page(struct size_class *class, struct zspage *zspage, | |
1874 | struct page *newpage, struct page *oldpage) | |
1875 | { | |
1876 | struct page *page; | |
1877 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, }; | |
1878 | int idx = 0; | |
1879 | ||
1880 | page = get_first_page(zspage); | |
1881 | do { | |
1882 | if (page == oldpage) | |
1883 | pages[idx] = newpage; | |
1884 | else | |
1885 | pages[idx] = page; | |
1886 | idx++; | |
1887 | } while ((page = get_next_page(page)) != NULL); | |
1888 | ||
1889 | create_page_chain(class, zspage, pages); | |
1890 | set_first_obj_offset(newpage, get_first_obj_offset(oldpage)); | |
1891 | if (unlikely(PageHugeObject(oldpage))) | |
1892 | newpage->index = oldpage->index; | |
1893 | __SetPageMovable(newpage, page_mapping(oldpage)); | |
1894 | } | |
1895 | ||
1896 | bool zs_page_isolate(struct page *page, isolate_mode_t mode) | |
1897 | { | |
1898 | struct zs_pool *pool; | |
1899 | struct size_class *class; | |
1900 | int class_idx; | |
1901 | enum fullness_group fullness; | |
1902 | struct zspage *zspage; | |
1903 | struct address_space *mapping; | |
1904 | ||
1905 | /* | |
1906 | * Page is locked so zspage couldn't be destroyed. For detail, look at | |
1907 | * lock_zspage in free_zspage. | |
1908 | */ | |
1909 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1910 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1911 | ||
1912 | zspage = get_zspage(page); | |
1913 | ||
1914 | /* | |
1915 | * Without class lock, fullness could be stale while class_idx is okay | |
1916 | * because class_idx is constant unless page is freed so we should get | |
1917 | * fullness again under class lock. | |
1918 | */ | |
1919 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1920 | mapping = page_mapping(page); | |
1921 | pool = mapping->private_data; | |
1922 | class = pool->size_class[class_idx]; | |
1923 | ||
1924 | spin_lock(&class->lock); | |
1925 | if (get_zspage_inuse(zspage) == 0) { | |
1926 | spin_unlock(&class->lock); | |
1927 | return false; | |
1928 | } | |
1929 | ||
1930 | /* zspage is isolated for object migration */ | |
1931 | if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
1932 | spin_unlock(&class->lock); | |
1933 | return false; | |
1934 | } | |
1935 | ||
1936 | /* | |
1937 | * If this is first time isolation for the zspage, isolate zspage from | |
1938 | * size_class to prevent further object allocation from the zspage. | |
1939 | */ | |
1940 | if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
1941 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1942 | remove_zspage(class, zspage, fullness); | |
1943 | } | |
1944 | ||
1945 | inc_zspage_isolation(zspage); | |
1946 | spin_unlock(&class->lock); | |
1947 | ||
1948 | return true; | |
1949 | } | |
1950 | ||
1951 | int zs_page_migrate(struct address_space *mapping, struct page *newpage, | |
1952 | struct page *page, enum migrate_mode mode) | |
1953 | { | |
1954 | struct zs_pool *pool; | |
1955 | struct size_class *class; | |
1956 | int class_idx; | |
1957 | enum fullness_group fullness; | |
1958 | struct zspage *zspage; | |
1959 | struct page *dummy; | |
1960 | void *s_addr, *d_addr, *addr; | |
1961 | int offset, pos; | |
1962 | unsigned long handle, head; | |
1963 | unsigned long old_obj, new_obj; | |
1964 | unsigned int obj_idx; | |
1965 | int ret = -EAGAIN; | |
1966 | ||
2916ecc0 JG |
1967 | /* |
1968 | * We cannot support the _NO_COPY case here, because copy needs to | |
1969 | * happen under the zs lock, which does not work with | |
1970 | * MIGRATE_SYNC_NO_COPY workflow. | |
1971 | */ | |
1972 | if (mode == MIGRATE_SYNC_NO_COPY) | |
1973 | return -EINVAL; | |
1974 | ||
48b4800a MK |
1975 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
1976 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
1977 | ||
1978 | zspage = get_zspage(page); | |
1979 | ||
1980 | /* Concurrent compactor cannot migrate any subpage in zspage */ | |
1981 | migrate_write_lock(zspage); | |
1982 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1983 | pool = mapping->private_data; | |
1984 | class = pool->size_class[class_idx]; | |
1985 | offset = get_first_obj_offset(page); | |
1986 | ||
1987 | spin_lock(&class->lock); | |
1988 | if (!get_zspage_inuse(zspage)) { | |
77ff4657 HZ |
1989 | /* |
1990 | * Set "offset" to end of the page so that every loops | |
1991 | * skips unnecessary object scanning. | |
1992 | */ | |
1993 | offset = PAGE_SIZE; | |
48b4800a MK |
1994 | } |
1995 | ||
1996 | pos = offset; | |
1997 | s_addr = kmap_atomic(page); | |
1998 | while (pos < PAGE_SIZE) { | |
1999 | head = obj_to_head(page, s_addr + pos); | |
2000 | if (head & OBJ_ALLOCATED_TAG) { | |
2001 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2002 | if (!trypin_tag(handle)) | |
2003 | goto unpin_objects; | |
2004 | } | |
2005 | pos += class->size; | |
2006 | } | |
2007 | ||
2008 | /* | |
2009 | * Here, any user cannot access all objects in the zspage so let's move. | |
2010 | */ | |
2011 | d_addr = kmap_atomic(newpage); | |
2012 | memcpy(d_addr, s_addr, PAGE_SIZE); | |
2013 | kunmap_atomic(d_addr); | |
2014 | ||
2015 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2016 | addr += class->size) { | |
2017 | head = obj_to_head(page, addr); | |
2018 | if (head & OBJ_ALLOCATED_TAG) { | |
2019 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2020 | if (!testpin_tag(handle)) | |
2021 | BUG(); | |
2022 | ||
2023 | old_obj = handle_to_obj(handle); | |
2024 | obj_to_location(old_obj, &dummy, &obj_idx); | |
2025 | new_obj = (unsigned long)location_to_obj(newpage, | |
2026 | obj_idx); | |
2027 | new_obj |= BIT(HANDLE_PIN_BIT); | |
2028 | record_obj(handle, new_obj); | |
2029 | } | |
2030 | } | |
2031 | ||
2032 | replace_sub_page(class, zspage, newpage, page); | |
2033 | get_page(newpage); | |
2034 | ||
2035 | dec_zspage_isolation(zspage); | |
2036 | ||
2037 | /* | |
2038 | * Page migration is done so let's putback isolated zspage to | |
2039 | * the list if @page is final isolated subpage in the zspage. | |
2040 | */ | |
2041 | if (!is_zspage_isolated(zspage)) | |
2042 | putback_zspage(class, zspage); | |
2043 | ||
2044 | reset_page(page); | |
2045 | put_page(page); | |
2046 | page = newpage; | |
2047 | ||
dd4123f3 | 2048 | ret = MIGRATEPAGE_SUCCESS; |
48b4800a MK |
2049 | unpin_objects: |
2050 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2051 | addr += class->size) { | |
2052 | head = obj_to_head(page, addr); | |
2053 | if (head & OBJ_ALLOCATED_TAG) { | |
2054 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2055 | if (!testpin_tag(handle)) | |
2056 | BUG(); | |
2057 | unpin_tag(handle); | |
2058 | } | |
2059 | } | |
2060 | kunmap_atomic(s_addr); | |
48b4800a MK |
2061 | spin_unlock(&class->lock); |
2062 | migrate_write_unlock(zspage); | |
2063 | ||
2064 | return ret; | |
2065 | } | |
2066 | ||
2067 | void zs_page_putback(struct page *page) | |
2068 | { | |
2069 | struct zs_pool *pool; | |
2070 | struct size_class *class; | |
2071 | int class_idx; | |
2072 | enum fullness_group fg; | |
2073 | struct address_space *mapping; | |
2074 | struct zspage *zspage; | |
2075 | ||
2076 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
2077 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
2078 | ||
2079 | zspage = get_zspage(page); | |
2080 | get_zspage_mapping(zspage, &class_idx, &fg); | |
2081 | mapping = page_mapping(page); | |
2082 | pool = mapping->private_data; | |
2083 | class = pool->size_class[class_idx]; | |
2084 | ||
2085 | spin_lock(&class->lock); | |
2086 | dec_zspage_isolation(zspage); | |
2087 | if (!is_zspage_isolated(zspage)) { | |
2088 | fg = putback_zspage(class, zspage); | |
2089 | /* | |
2090 | * Due to page_lock, we cannot free zspage immediately | |
2091 | * so let's defer. | |
2092 | */ | |
2093 | if (fg == ZS_EMPTY) | |
2094 | schedule_work(&pool->free_work); | |
2095 | } | |
2096 | spin_unlock(&class->lock); | |
2097 | } | |
2098 | ||
2099 | const struct address_space_operations zsmalloc_aops = { | |
2100 | .isolate_page = zs_page_isolate, | |
2101 | .migratepage = zs_page_migrate, | |
2102 | .putback_page = zs_page_putback, | |
2103 | }; | |
2104 | ||
2105 | static int zs_register_migration(struct zs_pool *pool) | |
2106 | { | |
2107 | pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb); | |
2108 | if (IS_ERR(pool->inode)) { | |
2109 | pool->inode = NULL; | |
2110 | return 1; | |
2111 | } | |
2112 | ||
2113 | pool->inode->i_mapping->private_data = pool; | |
2114 | pool->inode->i_mapping->a_ops = &zsmalloc_aops; | |
2115 | return 0; | |
2116 | } | |
2117 | ||
2118 | static void zs_unregister_migration(struct zs_pool *pool) | |
2119 | { | |
2120 | flush_work(&pool->free_work); | |
c3491eca | 2121 | iput(pool->inode); |
48b4800a MK |
2122 | } |
2123 | ||
2124 | /* | |
2125 | * Caller should hold page_lock of all pages in the zspage | |
2126 | * In here, we cannot use zspage meta data. | |
2127 | */ | |
2128 | static void async_free_zspage(struct work_struct *work) | |
2129 | { | |
2130 | int i; | |
2131 | struct size_class *class; | |
2132 | unsigned int class_idx; | |
2133 | enum fullness_group fullness; | |
2134 | struct zspage *zspage, *tmp; | |
2135 | LIST_HEAD(free_pages); | |
2136 | struct zs_pool *pool = container_of(work, struct zs_pool, | |
2137 | free_work); | |
2138 | ||
cf8e0fed | 2139 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
48b4800a MK |
2140 | class = pool->size_class[i]; |
2141 | if (class->index != i) | |
2142 | continue; | |
2143 | ||
2144 | spin_lock(&class->lock); | |
2145 | list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages); | |
2146 | spin_unlock(&class->lock); | |
2147 | } | |
2148 | ||
2149 | ||
2150 | list_for_each_entry_safe(zspage, tmp, &free_pages, list) { | |
2151 | list_del(&zspage->list); | |
2152 | lock_zspage(zspage); | |
2153 | ||
2154 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2155 | VM_BUG_ON(fullness != ZS_EMPTY); | |
2156 | class = pool->size_class[class_idx]; | |
2157 | spin_lock(&class->lock); | |
2158 | __free_zspage(pool, pool->size_class[class_idx], zspage); | |
2159 | spin_unlock(&class->lock); | |
2160 | } | |
2161 | }; | |
2162 | ||
2163 | static void kick_deferred_free(struct zs_pool *pool) | |
2164 | { | |
2165 | schedule_work(&pool->free_work); | |
2166 | } | |
2167 | ||
2168 | static void init_deferred_free(struct zs_pool *pool) | |
2169 | { | |
2170 | INIT_WORK(&pool->free_work, async_free_zspage); | |
2171 | } | |
2172 | ||
2173 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) | |
2174 | { | |
2175 | struct page *page = get_first_page(zspage); | |
2176 | ||
2177 | do { | |
2178 | WARN_ON(!trylock_page(page)); | |
2179 | __SetPageMovable(page, pool->inode->i_mapping); | |
2180 | unlock_page(page); | |
2181 | } while ((page = get_next_page(page)) != NULL); | |
2182 | } | |
2183 | #endif | |
2184 | ||
04f05909 SS |
2185 | /* |
2186 | * | |
2187 | * Based on the number of unused allocated objects calculate | |
2188 | * and return the number of pages that we can free. | |
04f05909 SS |
2189 | */ |
2190 | static unsigned long zs_can_compact(struct size_class *class) | |
2191 | { | |
2192 | unsigned long obj_wasted; | |
44f43e99 SS |
2193 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
2194 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 2195 | |
44f43e99 SS |
2196 | if (obj_allocated <= obj_used) |
2197 | return 0; | |
04f05909 | 2198 | |
44f43e99 | 2199 | obj_wasted = obj_allocated - obj_used; |
b4fd07a0 | 2200 | obj_wasted /= class->objs_per_zspage; |
04f05909 | 2201 | |
6cbf16b3 | 2202 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
2203 | } |
2204 | ||
7d3f3938 | 2205 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 2206 | { |
312fcae2 | 2207 | struct zs_compact_control cc; |
3783689a MK |
2208 | struct zspage *src_zspage; |
2209 | struct zspage *dst_zspage = NULL; | |
312fcae2 | 2210 | |
312fcae2 | 2211 | spin_lock(&class->lock); |
3783689a | 2212 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 2213 | |
04f05909 SS |
2214 | if (!zs_can_compact(class)) |
2215 | break; | |
2216 | ||
41b88e14 | 2217 | cc.obj_idx = 0; |
48b4800a | 2218 | cc.s_page = get_first_page(src_zspage); |
312fcae2 | 2219 | |
3783689a | 2220 | while ((dst_zspage = isolate_zspage(class, false))) { |
48b4800a | 2221 | cc.d_page = get_first_page(dst_zspage); |
312fcae2 | 2222 | /* |
0dc63d48 SS |
2223 | * If there is no more space in dst_page, resched |
2224 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
2225 | */ |
2226 | if (!migrate_zspage(pool, class, &cc)) | |
2227 | break; | |
2228 | ||
4aa409ca | 2229 | putback_zspage(class, dst_zspage); |
312fcae2 MK |
2230 | } |
2231 | ||
2232 | /* Stop if we couldn't find slot */ | |
3783689a | 2233 | if (dst_zspage == NULL) |
312fcae2 MK |
2234 | break; |
2235 | ||
4aa409ca MK |
2236 | putback_zspage(class, dst_zspage); |
2237 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { | |
48b4800a | 2238 | free_zspage(pool, class, src_zspage); |
6cbf16b3 | 2239 | pool->stats.pages_compacted += class->pages_per_zspage; |
4aa409ca | 2240 | } |
312fcae2 | 2241 | spin_unlock(&class->lock); |
312fcae2 MK |
2242 | cond_resched(); |
2243 | spin_lock(&class->lock); | |
2244 | } | |
2245 | ||
3783689a | 2246 | if (src_zspage) |
4aa409ca | 2247 | putback_zspage(class, src_zspage); |
312fcae2 | 2248 | |
7d3f3938 | 2249 | spin_unlock(&class->lock); |
312fcae2 MK |
2250 | } |
2251 | ||
2252 | unsigned long zs_compact(struct zs_pool *pool) | |
2253 | { | |
2254 | int i; | |
312fcae2 MK |
2255 | struct size_class *class; |
2256 | ||
cf8e0fed | 2257 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
312fcae2 MK |
2258 | class = pool->size_class[i]; |
2259 | if (!class) | |
2260 | continue; | |
2261 | if (class->index != i) | |
2262 | continue; | |
7d3f3938 | 2263 | __zs_compact(pool, class); |
312fcae2 MK |
2264 | } |
2265 | ||
860c707d | 2266 | return pool->stats.pages_compacted; |
312fcae2 MK |
2267 | } |
2268 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 2269 | |
7d3f3938 SS |
2270 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
2271 | { | |
2272 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
2273 | } | |
2274 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
2275 | ||
ab9d306d SS |
2276 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
2277 | struct shrink_control *sc) | |
2278 | { | |
2279 | unsigned long pages_freed; | |
2280 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2281 | shrinker); | |
2282 | ||
2283 | pages_freed = pool->stats.pages_compacted; | |
2284 | /* | |
2285 | * Compact classes and calculate compaction delta. | |
2286 | * Can run concurrently with a manually triggered | |
2287 | * (by user) compaction. | |
2288 | */ | |
2289 | pages_freed = zs_compact(pool) - pages_freed; | |
2290 | ||
2291 | return pages_freed ? pages_freed : SHRINK_STOP; | |
2292 | } | |
2293 | ||
2294 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
2295 | struct shrink_control *sc) | |
2296 | { | |
2297 | int i; | |
2298 | struct size_class *class; | |
2299 | unsigned long pages_to_free = 0; | |
2300 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2301 | shrinker); | |
2302 | ||
cf8e0fed | 2303 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
ab9d306d SS |
2304 | class = pool->size_class[i]; |
2305 | if (!class) | |
2306 | continue; | |
2307 | if (class->index != i) | |
2308 | continue; | |
2309 | ||
ab9d306d | 2310 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
2311 | } |
2312 | ||
2313 | return pages_to_free; | |
2314 | } | |
2315 | ||
2316 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
2317 | { | |
93144ca3 | 2318 | unregister_shrinker(&pool->shrinker); |
ab9d306d SS |
2319 | } |
2320 | ||
2321 | static int zs_register_shrinker(struct zs_pool *pool) | |
2322 | { | |
2323 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
2324 | pool->shrinker.count_objects = zs_shrinker_count; | |
2325 | pool->shrinker.batch = 0; | |
2326 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
2327 | ||
2328 | return register_shrinker(&pool->shrinker); | |
2329 | } | |
2330 | ||
00a61d86 | 2331 | /** |
66cdef66 | 2332 | * zs_create_pool - Creates an allocation pool to work from. |
fd854463 | 2333 | * @name: pool name to be created |
166cfda7 | 2334 | * |
66cdef66 GM |
2335 | * This function must be called before anything when using |
2336 | * the zsmalloc allocator. | |
166cfda7 | 2337 | * |
66cdef66 GM |
2338 | * On success, a pointer to the newly created pool is returned, |
2339 | * otherwise NULL. | |
396b7fd6 | 2340 | */ |
d0d8da2d | 2341 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 2342 | { |
66cdef66 GM |
2343 | int i; |
2344 | struct zs_pool *pool; | |
2345 | struct size_class *prev_class = NULL; | |
61989a80 | 2346 | |
66cdef66 GM |
2347 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
2348 | if (!pool) | |
2349 | return NULL; | |
61989a80 | 2350 | |
48b4800a | 2351 | init_deferred_free(pool); |
61989a80 | 2352 | |
2e40e163 MK |
2353 | pool->name = kstrdup(name, GFP_KERNEL); |
2354 | if (!pool->name) | |
2355 | goto err; | |
2356 | ||
3783689a | 2357 | if (create_cache(pool)) |
2e40e163 MK |
2358 | goto err; |
2359 | ||
c60369f0 | 2360 | /* |
399d8eeb | 2361 | * Iterate reversely, because, size of size_class that we want to use |
66cdef66 | 2362 | * for merging should be larger or equal to current size. |
c60369f0 | 2363 | */ |
cf8e0fed | 2364 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
66cdef66 GM |
2365 | int size; |
2366 | int pages_per_zspage; | |
64d90465 | 2367 | int objs_per_zspage; |
66cdef66 | 2368 | struct size_class *class; |
3783689a | 2369 | int fullness = 0; |
c60369f0 | 2370 | |
66cdef66 GM |
2371 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
2372 | if (size > ZS_MAX_ALLOC_SIZE) | |
2373 | size = ZS_MAX_ALLOC_SIZE; | |
2374 | pages_per_zspage = get_pages_per_zspage(size); | |
64d90465 | 2375 | objs_per_zspage = pages_per_zspage * PAGE_SIZE / size; |
61989a80 | 2376 | |
66cdef66 GM |
2377 | /* |
2378 | * size_class is used for normal zsmalloc operation such | |
2379 | * as alloc/free for that size. Although it is natural that we | |
2380 | * have one size_class for each size, there is a chance that we | |
2381 | * can get more memory utilization if we use one size_class for | |
2382 | * many different sizes whose size_class have same | |
2383 | * characteristics. So, we makes size_class point to | |
2384 | * previous size_class if possible. | |
2385 | */ | |
2386 | if (prev_class) { | |
64d90465 | 2387 | if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) { |
66cdef66 GM |
2388 | pool->size_class[i] = prev_class; |
2389 | continue; | |
2390 | } | |
2391 | } | |
2392 | ||
2393 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
2394 | if (!class) | |
2395 | goto err; | |
2396 | ||
2397 | class->size = size; | |
2398 | class->index = i; | |
2399 | class->pages_per_zspage = pages_per_zspage; | |
64d90465 | 2400 | class->objs_per_zspage = objs_per_zspage; |
66cdef66 GM |
2401 | spin_lock_init(&class->lock); |
2402 | pool->size_class[i] = class; | |
48b4800a MK |
2403 | for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS; |
2404 | fullness++) | |
3783689a | 2405 | INIT_LIST_HEAD(&class->fullness_list[fullness]); |
66cdef66 GM |
2406 | |
2407 | prev_class = class; | |
61989a80 NG |
2408 | } |
2409 | ||
d34f6157 DS |
2410 | /* debug only, don't abort if it fails */ |
2411 | zs_pool_stat_create(pool, name); | |
0f050d99 | 2412 | |
48b4800a MK |
2413 | if (zs_register_migration(pool)) |
2414 | goto err; | |
2415 | ||
ab9d306d | 2416 | /* |
93144ca3 AK |
2417 | * Not critical since shrinker is only used to trigger internal |
2418 | * defragmentation of the pool which is pretty optional thing. If | |
2419 | * registration fails we still can use the pool normally and user can | |
2420 | * trigger compaction manually. Thus, ignore return code. | |
ab9d306d | 2421 | */ |
93144ca3 AK |
2422 | zs_register_shrinker(pool); |
2423 | ||
66cdef66 GM |
2424 | return pool; |
2425 | ||
2426 | err: | |
2427 | zs_destroy_pool(pool); | |
2428 | return NULL; | |
61989a80 | 2429 | } |
66cdef66 | 2430 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2431 | |
66cdef66 | 2432 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2433 | { |
66cdef66 | 2434 | int i; |
61989a80 | 2435 | |
ab9d306d | 2436 | zs_unregister_shrinker(pool); |
48b4800a | 2437 | zs_unregister_migration(pool); |
0f050d99 GM |
2438 | zs_pool_stat_destroy(pool); |
2439 | ||
cf8e0fed | 2440 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
66cdef66 GM |
2441 | int fg; |
2442 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2443 | |
66cdef66 GM |
2444 | if (!class) |
2445 | continue; | |
61989a80 | 2446 | |
66cdef66 GM |
2447 | if (class->index != i) |
2448 | continue; | |
61989a80 | 2449 | |
48b4800a | 2450 | for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) { |
3783689a | 2451 | if (!list_empty(&class->fullness_list[fg])) { |
66cdef66 GM |
2452 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
2453 | class->size, fg); | |
2454 | } | |
2455 | } | |
2456 | kfree(class); | |
2457 | } | |
f553646a | 2458 | |
3783689a | 2459 | destroy_cache(pool); |
0f050d99 | 2460 | kfree(pool->name); |
66cdef66 GM |
2461 | kfree(pool); |
2462 | } | |
2463 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2464 | |
66cdef66 GM |
2465 | static int __init zs_init(void) |
2466 | { | |
48b4800a MK |
2467 | int ret; |
2468 | ||
2469 | ret = zsmalloc_mount(); | |
2470 | if (ret) | |
2471 | goto out; | |
2472 | ||
215c89d0 SAS |
2473 | ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare", |
2474 | zs_cpu_prepare, zs_cpu_dead); | |
0f050d99 | 2475 | if (ret) |
215c89d0 | 2476 | goto hp_setup_fail; |
66cdef66 | 2477 | |
66cdef66 GM |
2478 | #ifdef CONFIG_ZPOOL |
2479 | zpool_register_driver(&zs_zpool_driver); | |
2480 | #endif | |
0f050d99 | 2481 | |
4abaac9b DS |
2482 | zs_stat_init(); |
2483 | ||
66cdef66 | 2484 | return 0; |
0f050d99 | 2485 | |
215c89d0 | 2486 | hp_setup_fail: |
48b4800a MK |
2487 | zsmalloc_unmount(); |
2488 | out: | |
0f050d99 | 2489 | return ret; |
61989a80 | 2490 | } |
61989a80 | 2491 | |
66cdef66 | 2492 | static void __exit zs_exit(void) |
61989a80 | 2493 | { |
66cdef66 GM |
2494 | #ifdef CONFIG_ZPOOL |
2495 | zpool_unregister_driver(&zs_zpool_driver); | |
2496 | #endif | |
48b4800a | 2497 | zsmalloc_unmount(); |
215c89d0 | 2498 | cpuhp_remove_state(CPUHP_MM_ZS_PREPARE); |
0f050d99 GM |
2499 | |
2500 | zs_stat_exit(); | |
61989a80 | 2501 | } |
069f101f BH |
2502 | |
2503 | module_init(zs_init); | |
2504 | module_exit(zs_exit); | |
2505 | ||
2506 | MODULE_LICENSE("Dual BSD/GPL"); | |
2507 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |