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 SJ |
48 | #include <linux/spinlock.h> |
49 | #include <linux/types.h> | |
0f050d99 | 50 | #include <linux/debugfs.h> |
bcf1647d | 51 | #include <linux/zsmalloc.h> |
c795779d | 52 | #include <linux/zpool.h> |
48b4800a | 53 | #include <linux/mount.h> |
dd4123f3 | 54 | #include <linux/migrate.h> |
48b4800a MK |
55 | #include <linux/pagemap.h> |
56 | ||
57 | #define ZSPAGE_MAGIC 0x58 | |
0959c63f SJ |
58 | |
59 | /* | |
60 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
61 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
62 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
63 | * to restore link_free pointer values. | |
64 | */ | |
65 | #define ZS_ALIGN 8 | |
66 | ||
67 | /* | |
68 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
69 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
70 | */ | |
71 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
72 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
73 | ||
2e40e163 MK |
74 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
75 | ||
0959c63f SJ |
76 | /* |
77 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 78 | * as single (unsigned long) handle value. |
0959c63f | 79 | * |
bfd093f5 | 80 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
81 | * |
82 | * This is made more complicated by various memory models and PAE. | |
83 | */ | |
84 | ||
85 | #ifndef MAX_PHYSMEM_BITS | |
86 | #ifdef CONFIG_HIGHMEM64G | |
87 | #define MAX_PHYSMEM_BITS 36 | |
88 | #else /* !CONFIG_HIGHMEM64G */ | |
89 | /* | |
90 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
91 | * be PAGE_SHIFT | |
92 | */ | |
93 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
94 | #endif | |
95 | #endif | |
96 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
97 | |
98 | /* | |
99 | * Memory for allocating for handle keeps object position by | |
100 | * encoding <page, obj_idx> and the encoded value has a room | |
101 | * in least bit(ie, look at obj_to_location). | |
102 | * We use the bit to synchronize between object access by | |
103 | * user and migration. | |
104 | */ | |
105 | #define HANDLE_PIN_BIT 0 | |
106 | ||
107 | /* | |
108 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
109 | * to identify the object was allocated or not. | |
110 | * It's okay to add the status bit in the least bit because | |
111 | * header keeps handle which is 4byte-aligned address so we | |
112 | * have room for two bit at least. | |
113 | */ | |
114 | #define OBJ_ALLOCATED_TAG 1 | |
115 | #define OBJ_TAG_BITS 1 | |
116 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
117 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
118 | ||
cf8e0fed JM |
119 | #define FULLNESS_BITS 2 |
120 | #define CLASS_BITS 8 | |
121 | #define ISOLATED_BITS 3 | |
122 | #define MAGIC_VAL_BITS 8 | |
123 | ||
0959c63f SJ |
124 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) |
125 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
126 | #define ZS_MIN_ALLOC_SIZE \ | |
127 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 128 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 129 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
130 | |
131 | /* | |
7eb52512 | 132 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
133 | * trader-off here: |
134 | * - Large number of size classes is potentially wasteful as free page are | |
135 | * spread across these classes | |
136 | * - Small number of size classes causes large internal fragmentation | |
137 | * - Probably its better to use specific size classes (empirically | |
138 | * determined). NOTE: all those class sizes must be set as multiple of | |
139 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
140 | * | |
141 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
142 | * (reason above) | |
143 | */ | |
3783689a | 144 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
cf8e0fed JM |
145 | #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \ |
146 | ZS_SIZE_CLASS_DELTA) + 1) | |
0959c63f | 147 | |
0959c63f | 148 | enum fullness_group { |
0959c63f | 149 | ZS_EMPTY, |
48b4800a MK |
150 | ZS_ALMOST_EMPTY, |
151 | ZS_ALMOST_FULL, | |
152 | ZS_FULL, | |
153 | NR_ZS_FULLNESS, | |
0959c63f SJ |
154 | }; |
155 | ||
0f050d99 | 156 | enum zs_stat_type { |
48b4800a MK |
157 | CLASS_EMPTY, |
158 | CLASS_ALMOST_EMPTY, | |
159 | CLASS_ALMOST_FULL, | |
160 | CLASS_FULL, | |
0f050d99 GM |
161 | OBJ_ALLOCATED, |
162 | OBJ_USED, | |
48b4800a | 163 | NR_ZS_STAT_TYPE, |
0f050d99 GM |
164 | }; |
165 | ||
0f050d99 GM |
166 | struct zs_size_stat { |
167 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
168 | }; | |
169 | ||
57244594 SS |
170 | #ifdef CONFIG_ZSMALLOC_STAT |
171 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
172 | #endif |
173 | ||
48b4800a MK |
174 | #ifdef CONFIG_COMPACTION |
175 | static struct vfsmount *zsmalloc_mnt; | |
176 | #endif | |
177 | ||
0959c63f SJ |
178 | /* |
179 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
180 | * n <= N / f, where | |
181 | * n = number of allocated objects | |
182 | * N = total number of objects zspage can store | |
6dd9737e | 183 | * f = fullness_threshold_frac |
0959c63f SJ |
184 | * |
185 | * Similarly, we assign zspage to: | |
186 | * ZS_ALMOST_FULL when n > N / f | |
187 | * ZS_EMPTY when n == 0 | |
188 | * ZS_FULL when n == N | |
189 | * | |
190 | * (see: fix_fullness_group()) | |
191 | */ | |
192 | static const int fullness_threshold_frac = 4; | |
7f47f23b | 193 | static size_t huge_class_size; |
0959c63f SJ |
194 | |
195 | struct size_class { | |
57244594 | 196 | spinlock_t lock; |
48b4800a | 197 | struct list_head fullness_list[NR_ZS_FULLNESS]; |
0959c63f SJ |
198 | /* |
199 | * Size of objects stored in this class. Must be multiple | |
200 | * of ZS_ALIGN. | |
201 | */ | |
202 | int size; | |
1fc6e27d | 203 | int objs_per_zspage; |
7dfa4612 WY |
204 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
205 | int pages_per_zspage; | |
48b4800a MK |
206 | |
207 | unsigned int index; | |
208 | struct zs_size_stat stats; | |
0959c63f SJ |
209 | }; |
210 | ||
48b4800a MK |
211 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
212 | static void SetPageHugeObject(struct page *page) | |
213 | { | |
214 | SetPageOwnerPriv1(page); | |
215 | } | |
216 | ||
217 | static void ClearPageHugeObject(struct page *page) | |
218 | { | |
219 | ClearPageOwnerPriv1(page); | |
220 | } | |
221 | ||
222 | static int PageHugeObject(struct page *page) | |
223 | { | |
224 | return PageOwnerPriv1(page); | |
225 | } | |
226 | ||
0959c63f SJ |
227 | /* |
228 | * Placed within free objects to form a singly linked list. | |
3783689a | 229 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
230 | * |
231 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
232 | */ | |
233 | struct link_free { | |
2e40e163 MK |
234 | union { |
235 | /* | |
bfd093f5 | 236 | * Free object index; |
2e40e163 MK |
237 | * It's valid for non-allocated object |
238 | */ | |
bfd093f5 | 239 | unsigned long next; |
2e40e163 MK |
240 | /* |
241 | * Handle of allocated object. | |
242 | */ | |
243 | unsigned long handle; | |
244 | }; | |
0959c63f SJ |
245 | }; |
246 | ||
247 | struct zs_pool { | |
6f3526d6 | 248 | const char *name; |
0f050d99 | 249 | |
cf8e0fed | 250 | struct size_class *size_class[ZS_SIZE_CLASSES]; |
2e40e163 | 251 | struct kmem_cache *handle_cachep; |
3783689a | 252 | struct kmem_cache *zspage_cachep; |
0959c63f | 253 | |
13de8933 | 254 | atomic_long_t pages_allocated; |
0f050d99 | 255 | |
7d3f3938 | 256 | struct zs_pool_stats stats; |
ab9d306d SS |
257 | |
258 | /* Compact classes */ | |
259 | struct shrinker shrinker; | |
260 | /* | |
261 | * To signify that register_shrinker() was successful | |
262 | * and unregister_shrinker() will not Oops. | |
263 | */ | |
264 | bool shrinker_enabled; | |
0f050d99 GM |
265 | #ifdef CONFIG_ZSMALLOC_STAT |
266 | struct dentry *stat_dentry; | |
267 | #endif | |
48b4800a MK |
268 | #ifdef CONFIG_COMPACTION |
269 | struct inode *inode; | |
270 | struct work_struct free_work; | |
271 | #endif | |
0959c63f | 272 | }; |
61989a80 | 273 | |
3783689a MK |
274 | struct zspage { |
275 | struct { | |
276 | unsigned int fullness:FULLNESS_BITS; | |
85d492f2 | 277 | unsigned int class:CLASS_BITS + 1; |
48b4800a MK |
278 | unsigned int isolated:ISOLATED_BITS; |
279 | unsigned int magic:MAGIC_VAL_BITS; | |
3783689a MK |
280 | }; |
281 | unsigned int inuse; | |
bfd093f5 | 282 | unsigned int freeobj; |
3783689a MK |
283 | struct page *first_page; |
284 | struct list_head list; /* fullness list */ | |
48b4800a MK |
285 | #ifdef CONFIG_COMPACTION |
286 | rwlock_t lock; | |
287 | #endif | |
3783689a | 288 | }; |
61989a80 | 289 | |
f553646a | 290 | struct mapping_area { |
1b945aee | 291 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
292 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
293 | #else | |
294 | char *vm_buf; /* copy buffer for objects that span pages */ | |
295 | #endif | |
296 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
297 | enum zs_mapmode vm_mm; /* mapping mode */ | |
298 | }; | |
299 | ||
48b4800a MK |
300 | #ifdef CONFIG_COMPACTION |
301 | static int zs_register_migration(struct zs_pool *pool); | |
302 | static void zs_unregister_migration(struct zs_pool *pool); | |
303 | static void migrate_lock_init(struct zspage *zspage); | |
304 | static void migrate_read_lock(struct zspage *zspage); | |
305 | static void migrate_read_unlock(struct zspage *zspage); | |
306 | static void kick_deferred_free(struct zs_pool *pool); | |
307 | static void init_deferred_free(struct zs_pool *pool); | |
308 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage); | |
309 | #else | |
310 | static int zsmalloc_mount(void) { return 0; } | |
311 | static void zsmalloc_unmount(void) {} | |
312 | static int zs_register_migration(struct zs_pool *pool) { return 0; } | |
313 | static void zs_unregister_migration(struct zs_pool *pool) {} | |
314 | static void migrate_lock_init(struct zspage *zspage) {} | |
315 | static void migrate_read_lock(struct zspage *zspage) {} | |
316 | static void migrate_read_unlock(struct zspage *zspage) {} | |
317 | static void kick_deferred_free(struct zs_pool *pool) {} | |
318 | static void init_deferred_free(struct zs_pool *pool) {} | |
319 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {} | |
320 | #endif | |
321 | ||
3783689a | 322 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
323 | { |
324 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
325 | 0, 0, NULL); | |
3783689a MK |
326 | if (!pool->handle_cachep) |
327 | return 1; | |
328 | ||
329 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
330 | 0, 0, NULL); | |
331 | if (!pool->zspage_cachep) { | |
332 | kmem_cache_destroy(pool->handle_cachep); | |
333 | pool->handle_cachep = NULL; | |
334 | return 1; | |
335 | } | |
336 | ||
337 | return 0; | |
2e40e163 MK |
338 | } |
339 | ||
3783689a | 340 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 341 | { |
cd10add0 | 342 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 343 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
344 | } |
345 | ||
3783689a | 346 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
347 | { |
348 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
48b4800a | 349 | gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
2e40e163 MK |
350 | } |
351 | ||
3783689a | 352 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
353 | { |
354 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
355 | } | |
356 | ||
3783689a MK |
357 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
358 | { | |
48b4800a MK |
359 | return kmem_cache_alloc(pool->zspage_cachep, |
360 | flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); | |
399d8eeb | 361 | } |
3783689a MK |
362 | |
363 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
364 | { | |
365 | kmem_cache_free(pool->zspage_cachep, zspage); | |
366 | } | |
367 | ||
2e40e163 MK |
368 | static void record_obj(unsigned long handle, unsigned long obj) |
369 | { | |
c102f07c JL |
370 | /* |
371 | * lsb of @obj represents handle lock while other bits | |
372 | * represent object value the handle is pointing so | |
373 | * updating shouldn't do store tearing. | |
374 | */ | |
375 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
376 | } |
377 | ||
c795779d DS |
378 | /* zpool driver */ |
379 | ||
380 | #ifdef CONFIG_ZPOOL | |
381 | ||
6f3526d6 | 382 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 383 | const struct zpool_ops *zpool_ops, |
479305fd | 384 | struct zpool *zpool) |
c795779d | 385 | { |
d0d8da2d SS |
386 | /* |
387 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
388 | * different contexts and its caller must provide a valid | |
389 | * gfp mask. | |
390 | */ | |
391 | return zs_create_pool(name); | |
c795779d DS |
392 | } |
393 | ||
394 | static void zs_zpool_destroy(void *pool) | |
395 | { | |
396 | zs_destroy_pool(pool); | |
397 | } | |
398 | ||
399 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
400 | unsigned long *handle) | |
401 | { | |
d0d8da2d | 402 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
403 | return *handle ? 0 : -1; |
404 | } | |
405 | static void zs_zpool_free(void *pool, unsigned long handle) | |
406 | { | |
407 | zs_free(pool, handle); | |
408 | } | |
409 | ||
410 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
411 | unsigned int *reclaimed) | |
412 | { | |
413 | return -EINVAL; | |
414 | } | |
415 | ||
416 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
417 | enum zpool_mapmode mm) | |
418 | { | |
419 | enum zs_mapmode zs_mm; | |
420 | ||
421 | switch (mm) { | |
422 | case ZPOOL_MM_RO: | |
423 | zs_mm = ZS_MM_RO; | |
424 | break; | |
425 | case ZPOOL_MM_WO: | |
426 | zs_mm = ZS_MM_WO; | |
427 | break; | |
428 | case ZPOOL_MM_RW: /* fallthru */ | |
429 | default: | |
430 | zs_mm = ZS_MM_RW; | |
431 | break; | |
432 | } | |
433 | ||
434 | return zs_map_object(pool, handle, zs_mm); | |
435 | } | |
436 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
437 | { | |
438 | zs_unmap_object(pool, handle); | |
439 | } | |
440 | ||
441 | static u64 zs_zpool_total_size(void *pool) | |
442 | { | |
722cdc17 | 443 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
444 | } |
445 | ||
446 | static struct zpool_driver zs_zpool_driver = { | |
447 | .type = "zsmalloc", | |
448 | .owner = THIS_MODULE, | |
449 | .create = zs_zpool_create, | |
450 | .destroy = zs_zpool_destroy, | |
451 | .malloc = zs_zpool_malloc, | |
452 | .free = zs_zpool_free, | |
453 | .shrink = zs_zpool_shrink, | |
454 | .map = zs_zpool_map, | |
455 | .unmap = zs_zpool_unmap, | |
456 | .total_size = zs_zpool_total_size, | |
457 | }; | |
458 | ||
137f8cff | 459 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
460 | #endif /* CONFIG_ZPOOL */ |
461 | ||
61989a80 NG |
462 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
463 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
464 | ||
48b4800a MK |
465 | static bool is_zspage_isolated(struct zspage *zspage) |
466 | { | |
467 | return zspage->isolated; | |
468 | } | |
469 | ||
3457f414 | 470 | static __maybe_unused int is_first_page(struct page *page) |
61989a80 | 471 | { |
a27545bf | 472 | return PagePrivate(page); |
61989a80 NG |
473 | } |
474 | ||
48b4800a | 475 | /* Protected by class->lock */ |
3783689a | 476 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 477 | { |
3783689a | 478 | return zspage->inuse; |
4f42047b MK |
479 | } |
480 | ||
3783689a | 481 | static inline void set_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 482 | { |
3783689a | 483 | zspage->inuse = val; |
4f42047b MK |
484 | } |
485 | ||
3783689a | 486 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 487 | { |
3783689a | 488 | zspage->inuse += val; |
4f42047b MK |
489 | } |
490 | ||
48b4800a | 491 | static inline struct page *get_first_page(struct zspage *zspage) |
4f42047b | 492 | { |
48b4800a | 493 | struct page *first_page = zspage->first_page; |
3783689a | 494 | |
48b4800a MK |
495 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
496 | return first_page; | |
4f42047b MK |
497 | } |
498 | ||
48b4800a | 499 | static inline int get_first_obj_offset(struct page *page) |
4f42047b | 500 | { |
48b4800a MK |
501 | return page->units; |
502 | } | |
3783689a | 503 | |
48b4800a MK |
504 | static inline void set_first_obj_offset(struct page *page, int offset) |
505 | { | |
506 | page->units = offset; | |
4f42047b MK |
507 | } |
508 | ||
bfd093f5 | 509 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 510 | { |
bfd093f5 | 511 | return zspage->freeobj; |
4f42047b MK |
512 | } |
513 | ||
bfd093f5 | 514 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 515 | { |
bfd093f5 | 516 | zspage->freeobj = obj; |
4f42047b MK |
517 | } |
518 | ||
3783689a | 519 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 520 | unsigned int *class_idx, |
61989a80 NG |
521 | enum fullness_group *fullness) |
522 | { | |
48b4800a MK |
523 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); |
524 | ||
3783689a MK |
525 | *fullness = zspage->fullness; |
526 | *class_idx = zspage->class; | |
61989a80 NG |
527 | } |
528 | ||
3783689a | 529 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 530 | unsigned int class_idx, |
61989a80 NG |
531 | enum fullness_group fullness) |
532 | { | |
3783689a MK |
533 | zspage->class = class_idx; |
534 | zspage->fullness = fullness; | |
61989a80 NG |
535 | } |
536 | ||
c3e3e88a NC |
537 | /* |
538 | * zsmalloc divides the pool into various size classes where each | |
539 | * class maintains a list of zspages where each zspage is divided | |
540 | * into equal sized chunks. Each allocation falls into one of these | |
541 | * classes depending on its size. This function returns index of the | |
542 | * size class which has chunk size big enough to hold the give size. | |
543 | */ | |
61989a80 NG |
544 | static int get_size_class_index(int size) |
545 | { | |
546 | int idx = 0; | |
547 | ||
548 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
549 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
550 | ZS_SIZE_CLASS_DELTA); | |
551 | ||
cf8e0fed | 552 | return min_t(int, ZS_SIZE_CLASSES - 1, idx); |
61989a80 NG |
553 | } |
554 | ||
3eb95fea | 555 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 556 | static inline void zs_stat_inc(struct size_class *class, |
3eb95fea | 557 | int type, unsigned long cnt) |
248ca1b0 | 558 | { |
48b4800a | 559 | class->stats.objs[type] += cnt; |
248ca1b0 MK |
560 | } |
561 | ||
3eb95fea | 562 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 563 | static inline void zs_stat_dec(struct size_class *class, |
3eb95fea | 564 | int type, unsigned long cnt) |
248ca1b0 | 565 | { |
48b4800a | 566 | class->stats.objs[type] -= cnt; |
248ca1b0 MK |
567 | } |
568 | ||
3eb95fea | 569 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 570 | static inline unsigned long zs_stat_get(struct size_class *class, |
3eb95fea | 571 | int type) |
248ca1b0 | 572 | { |
48b4800a | 573 | return class->stats.objs[type]; |
248ca1b0 MK |
574 | } |
575 | ||
57244594 SS |
576 | #ifdef CONFIG_ZSMALLOC_STAT |
577 | ||
4abaac9b | 578 | static void __init zs_stat_init(void) |
248ca1b0 | 579 | { |
4abaac9b DS |
580 | if (!debugfs_initialized()) { |
581 | pr_warn("debugfs not available, stat dir not created\n"); | |
582 | return; | |
583 | } | |
248ca1b0 MK |
584 | |
585 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
586 | if (!zs_stat_root) | |
4abaac9b | 587 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
588 | } |
589 | ||
590 | static void __exit zs_stat_exit(void) | |
591 | { | |
592 | debugfs_remove_recursive(zs_stat_root); | |
593 | } | |
594 | ||
1120ed54 SS |
595 | static unsigned long zs_can_compact(struct size_class *class); |
596 | ||
248ca1b0 MK |
597 | static int zs_stats_size_show(struct seq_file *s, void *v) |
598 | { | |
599 | int i; | |
600 | struct zs_pool *pool = s->private; | |
601 | struct size_class *class; | |
602 | int objs_per_zspage; | |
603 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 604 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
605 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
606 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 607 | unsigned long total_freeable = 0; |
248ca1b0 | 608 | |
1120ed54 | 609 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
610 | "class", "size", "almost_full", "almost_empty", |
611 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 612 | "pages_per_zspage", "freeable"); |
248ca1b0 | 613 | |
cf8e0fed | 614 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
248ca1b0 MK |
615 | class = pool->size_class[i]; |
616 | ||
617 | if (class->index != i) | |
618 | continue; | |
619 | ||
620 | spin_lock(&class->lock); | |
621 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
622 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
623 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
624 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 625 | freeable = zs_can_compact(class); |
248ca1b0 MK |
626 | spin_unlock(&class->lock); |
627 | ||
b4fd07a0 | 628 | objs_per_zspage = class->objs_per_zspage; |
248ca1b0 MK |
629 | pages_used = obj_allocated / objs_per_zspage * |
630 | class->pages_per_zspage; | |
631 | ||
1120ed54 SS |
632 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
633 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
634 | i, class->size, class_almost_full, class_almost_empty, |
635 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 636 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
637 | |
638 | total_class_almost_full += class_almost_full; | |
639 | total_class_almost_empty += class_almost_empty; | |
640 | total_objs += obj_allocated; | |
641 | total_used_objs += obj_used; | |
642 | total_pages += pages_used; | |
1120ed54 | 643 | total_freeable += freeable; |
248ca1b0 MK |
644 | } |
645 | ||
646 | seq_puts(s, "\n"); | |
1120ed54 | 647 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
648 | "Total", "", total_class_almost_full, |
649 | total_class_almost_empty, total_objs, | |
1120ed54 | 650 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
651 | |
652 | return 0; | |
653 | } | |
654 | ||
655 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
656 | { | |
657 | return single_open(file, zs_stats_size_show, inode->i_private); | |
658 | } | |
659 | ||
660 | static const struct file_operations zs_stat_size_ops = { | |
661 | .open = zs_stats_size_open, | |
662 | .read = seq_read, | |
663 | .llseek = seq_lseek, | |
664 | .release = single_release, | |
665 | }; | |
666 | ||
d34f6157 | 667 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
668 | { |
669 | struct dentry *entry; | |
670 | ||
4abaac9b DS |
671 | if (!zs_stat_root) { |
672 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 673 | return; |
4abaac9b | 674 | } |
248ca1b0 MK |
675 | |
676 | entry = debugfs_create_dir(name, zs_stat_root); | |
677 | if (!entry) { | |
678 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 679 | return; |
248ca1b0 MK |
680 | } |
681 | pool->stat_dentry = entry; | |
682 | ||
683 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
684 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
685 | if (!entry) { | |
686 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
687 | name, "classes"); | |
4abaac9b DS |
688 | debugfs_remove_recursive(pool->stat_dentry); |
689 | pool->stat_dentry = NULL; | |
248ca1b0 | 690 | } |
248ca1b0 MK |
691 | } |
692 | ||
693 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
694 | { | |
695 | debugfs_remove_recursive(pool->stat_dentry); | |
696 | } | |
697 | ||
698 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 699 | static void __init zs_stat_init(void) |
248ca1b0 | 700 | { |
248ca1b0 MK |
701 | } |
702 | ||
703 | static void __exit zs_stat_exit(void) | |
704 | { | |
705 | } | |
706 | ||
d34f6157 | 707 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 708 | { |
248ca1b0 MK |
709 | } |
710 | ||
711 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
712 | { | |
713 | } | |
248ca1b0 MK |
714 | #endif |
715 | ||
48b4800a | 716 | |
c3e3e88a NC |
717 | /* |
718 | * For each size class, zspages are divided into different groups | |
719 | * depending on how "full" they are. This was done so that we could | |
720 | * easily find empty or nearly empty zspages when we try to shrink | |
721 | * the pool (not yet implemented). This function returns fullness | |
722 | * status of the given page. | |
723 | */ | |
1fc6e27d | 724 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 725 | struct zspage *zspage) |
61989a80 | 726 | { |
1fc6e27d | 727 | int inuse, objs_per_zspage; |
61989a80 | 728 | enum fullness_group fg; |
830e4bc5 | 729 | |
3783689a | 730 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 731 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
732 | |
733 | if (inuse == 0) | |
734 | fg = ZS_EMPTY; | |
1fc6e27d | 735 | else if (inuse == objs_per_zspage) |
61989a80 | 736 | fg = ZS_FULL; |
1fc6e27d | 737 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
738 | fg = ZS_ALMOST_EMPTY; |
739 | else | |
740 | fg = ZS_ALMOST_FULL; | |
741 | ||
742 | return fg; | |
743 | } | |
744 | ||
c3e3e88a NC |
745 | /* |
746 | * Each size class maintains various freelists and zspages are assigned | |
747 | * to one of these freelists based on the number of live objects they | |
748 | * have. This functions inserts the given zspage into the freelist | |
749 | * identified by <class, fullness_group>. | |
750 | */ | |
251cbb95 | 751 | static void insert_zspage(struct size_class *class, |
3783689a MK |
752 | struct zspage *zspage, |
753 | enum fullness_group fullness) | |
61989a80 | 754 | { |
3783689a | 755 | struct zspage *head; |
61989a80 | 756 | |
48b4800a | 757 | zs_stat_inc(class, fullness, 1); |
3783689a MK |
758 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
759 | struct zspage, list); | |
58f17117 | 760 | /* |
3783689a MK |
761 | * We want to see more ZS_FULL pages and less almost empty/full. |
762 | * Put pages with higher ->inuse first. | |
58f17117 | 763 | */ |
3783689a MK |
764 | if (head) { |
765 | if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) { | |
766 | list_add(&zspage->list, &head->list); | |
767 | return; | |
768 | } | |
769 | } | |
770 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
61989a80 NG |
771 | } |
772 | ||
c3e3e88a NC |
773 | /* |
774 | * This function removes the given zspage from the freelist identified | |
775 | * by <class, fullness_group>. | |
776 | */ | |
251cbb95 | 777 | static void remove_zspage(struct size_class *class, |
3783689a MK |
778 | struct zspage *zspage, |
779 | enum fullness_group fullness) | |
61989a80 | 780 | { |
3783689a | 781 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
48b4800a | 782 | VM_BUG_ON(is_zspage_isolated(zspage)); |
61989a80 | 783 | |
3783689a | 784 | list_del_init(&zspage->list); |
48b4800a | 785 | zs_stat_dec(class, fullness, 1); |
61989a80 NG |
786 | } |
787 | ||
c3e3e88a NC |
788 | /* |
789 | * Each size class maintains zspages in different fullness groups depending | |
790 | * on the number of live objects they contain. When allocating or freeing | |
791 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
792 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
793 | * a status change has occurred for the given page and accordingly moves the | |
794 | * page from the freelist of the old fullness group to that of the new | |
795 | * fullness group. | |
796 | */ | |
c7806261 | 797 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 798 | struct zspage *zspage) |
61989a80 NG |
799 | { |
800 | int class_idx; | |
61989a80 NG |
801 | enum fullness_group currfg, newfg; |
802 | ||
3783689a MK |
803 | get_zspage_mapping(zspage, &class_idx, &currfg); |
804 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
805 | if (newfg == currfg) |
806 | goto out; | |
807 | ||
48b4800a MK |
808 | if (!is_zspage_isolated(zspage)) { |
809 | remove_zspage(class, zspage, currfg); | |
810 | insert_zspage(class, zspage, newfg); | |
811 | } | |
812 | ||
3783689a | 813 | set_zspage_mapping(zspage, class_idx, newfg); |
61989a80 NG |
814 | |
815 | out: | |
816 | return newfg; | |
817 | } | |
818 | ||
819 | /* | |
820 | * We have to decide on how many pages to link together | |
821 | * to form a zspage for each size class. This is important | |
822 | * to reduce wastage due to unusable space left at end of | |
823 | * each zspage which is given as: | |
888fa374 YX |
824 | * wastage = Zp % class_size |
825 | * usage = Zp - wastage | |
61989a80 NG |
826 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
827 | * | |
828 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
829 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
830 | * since then we can perfectly fit in 8 such objects. | |
831 | */ | |
2e3b6154 | 832 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
833 | { |
834 | int i, max_usedpc = 0; | |
835 | /* zspage order which gives maximum used size per KB */ | |
836 | int max_usedpc_order = 1; | |
837 | ||
84d4faab | 838 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
839 | int zspage_size; |
840 | int waste, usedpc; | |
841 | ||
842 | zspage_size = i * PAGE_SIZE; | |
843 | waste = zspage_size % class_size; | |
844 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
845 | ||
846 | if (usedpc > max_usedpc) { | |
847 | max_usedpc = usedpc; | |
848 | max_usedpc_order = i; | |
849 | } | |
850 | } | |
851 | ||
852 | return max_usedpc_order; | |
853 | } | |
854 | ||
3783689a | 855 | static struct zspage *get_zspage(struct page *page) |
61989a80 | 856 | { |
48b4800a MK |
857 | struct zspage *zspage = (struct zspage *)page->private; |
858 | ||
859 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); | |
860 | return zspage; | |
61989a80 NG |
861 | } |
862 | ||
863 | static struct page *get_next_page(struct page *page) | |
864 | { | |
48b4800a MK |
865 | if (unlikely(PageHugeObject(page))) |
866 | return NULL; | |
867 | ||
868 | return page->freelist; | |
61989a80 NG |
869 | } |
870 | ||
bfd093f5 MK |
871 | /** |
872 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
873 | * @page: page object resides in zspage | |
874 | * @obj_idx: object index | |
67296874 | 875 | */ |
bfd093f5 MK |
876 | static void obj_to_location(unsigned long obj, struct page **page, |
877 | unsigned int *obj_idx) | |
61989a80 | 878 | { |
bfd093f5 MK |
879 | obj >>= OBJ_TAG_BITS; |
880 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
881 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
882 | } | |
61989a80 | 883 | |
bfd093f5 MK |
884 | /** |
885 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
886 | * @page: page object resides in zspage | |
887 | * @obj_idx: object index | |
888 | */ | |
889 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
890 | { | |
891 | unsigned long obj; | |
61989a80 | 892 | |
312fcae2 | 893 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 894 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 895 | obj <<= OBJ_TAG_BITS; |
61989a80 | 896 | |
bfd093f5 | 897 | return obj; |
61989a80 NG |
898 | } |
899 | ||
2e40e163 MK |
900 | static unsigned long handle_to_obj(unsigned long handle) |
901 | { | |
902 | return *(unsigned long *)handle; | |
903 | } | |
904 | ||
48b4800a | 905 | static unsigned long obj_to_head(struct page *page, void *obj) |
312fcae2 | 906 | { |
48b4800a | 907 | if (unlikely(PageHugeObject(page))) { |
830e4bc5 | 908 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3783689a | 909 | return page->index; |
7b60a685 MK |
910 | } else |
911 | return *(unsigned long *)obj; | |
312fcae2 MK |
912 | } |
913 | ||
48b4800a MK |
914 | static inline int testpin_tag(unsigned long handle) |
915 | { | |
916 | return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle); | |
917 | } | |
918 | ||
312fcae2 MK |
919 | static inline int trypin_tag(unsigned long handle) |
920 | { | |
1b8320b6 | 921 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
922 | } |
923 | ||
924 | static void pin_tag(unsigned long handle) | |
925 | { | |
1b8320b6 | 926 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
927 | } |
928 | ||
929 | static void unpin_tag(unsigned long handle) | |
930 | { | |
1b8320b6 | 931 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
932 | } |
933 | ||
f4477e90 NG |
934 | static void reset_page(struct page *page) |
935 | { | |
48b4800a | 936 | __ClearPageMovable(page); |
18fd06bf | 937 | ClearPagePrivate(page); |
f4477e90 | 938 | set_page_private(page, 0); |
48b4800a MK |
939 | page_mapcount_reset(page); |
940 | ClearPageHugeObject(page); | |
941 | page->freelist = NULL; | |
942 | } | |
943 | ||
944 | /* | |
945 | * To prevent zspage destroy during migration, zspage freeing should | |
946 | * hold locks of all pages in the zspage. | |
947 | */ | |
948 | void lock_zspage(struct zspage *zspage) | |
949 | { | |
950 | struct page *page = get_first_page(zspage); | |
951 | ||
952 | do { | |
953 | lock_page(page); | |
954 | } while ((page = get_next_page(page)) != NULL); | |
955 | } | |
956 | ||
957 | int trylock_zspage(struct zspage *zspage) | |
958 | { | |
959 | struct page *cursor, *fail; | |
960 | ||
961 | for (cursor = get_first_page(zspage); cursor != NULL; cursor = | |
962 | get_next_page(cursor)) { | |
963 | if (!trylock_page(cursor)) { | |
964 | fail = cursor; | |
965 | goto unlock; | |
966 | } | |
967 | } | |
968 | ||
969 | return 1; | |
970 | unlock: | |
971 | for (cursor = get_first_page(zspage); cursor != fail; cursor = | |
972 | get_next_page(cursor)) | |
973 | unlock_page(cursor); | |
974 | ||
975 | return 0; | |
f4477e90 NG |
976 | } |
977 | ||
48b4800a MK |
978 | static void __free_zspage(struct zs_pool *pool, struct size_class *class, |
979 | struct zspage *zspage) | |
61989a80 | 980 | { |
3783689a | 981 | struct page *page, *next; |
48b4800a MK |
982 | enum fullness_group fg; |
983 | unsigned int class_idx; | |
984 | ||
985 | get_zspage_mapping(zspage, &class_idx, &fg); | |
986 | ||
987 | assert_spin_locked(&class->lock); | |
61989a80 | 988 | |
3783689a | 989 | VM_BUG_ON(get_zspage_inuse(zspage)); |
48b4800a | 990 | VM_BUG_ON(fg != ZS_EMPTY); |
61989a80 | 991 | |
48b4800a | 992 | next = page = get_first_page(zspage); |
3783689a | 993 | do { |
48b4800a MK |
994 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
995 | next = get_next_page(page); | |
3783689a | 996 | reset_page(page); |
48b4800a | 997 | unlock_page(page); |
91537fee | 998 | dec_zone_page_state(page, NR_ZSPAGES); |
3783689a MK |
999 | put_page(page); |
1000 | page = next; | |
1001 | } while (page != NULL); | |
61989a80 | 1002 | |
3783689a | 1003 | cache_free_zspage(pool, zspage); |
48b4800a | 1004 | |
b4fd07a0 | 1005 | zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
1006 | atomic_long_sub(class->pages_per_zspage, |
1007 | &pool->pages_allocated); | |
1008 | } | |
1009 | ||
1010 | static void free_zspage(struct zs_pool *pool, struct size_class *class, | |
1011 | struct zspage *zspage) | |
1012 | { | |
1013 | VM_BUG_ON(get_zspage_inuse(zspage)); | |
1014 | VM_BUG_ON(list_empty(&zspage->list)); | |
1015 | ||
1016 | if (!trylock_zspage(zspage)) { | |
1017 | kick_deferred_free(pool); | |
1018 | return; | |
1019 | } | |
1020 | ||
1021 | remove_zspage(class, zspage, ZS_EMPTY); | |
1022 | __free_zspage(pool, class, zspage); | |
61989a80 NG |
1023 | } |
1024 | ||
1025 | /* Initialize a newly allocated zspage */ | |
3783689a | 1026 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 1027 | { |
bfd093f5 | 1028 | unsigned int freeobj = 1; |
61989a80 | 1029 | unsigned long off = 0; |
48b4800a | 1030 | struct page *page = get_first_page(zspage); |
830e4bc5 | 1031 | |
61989a80 NG |
1032 | while (page) { |
1033 | struct page *next_page; | |
1034 | struct link_free *link; | |
af4ee5e9 | 1035 | void *vaddr; |
61989a80 | 1036 | |
3783689a | 1037 | set_first_obj_offset(page, off); |
61989a80 | 1038 | |
af4ee5e9 MK |
1039 | vaddr = kmap_atomic(page); |
1040 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
1041 | |
1042 | while ((off += class->size) < PAGE_SIZE) { | |
3b1d9ca6 | 1043 | link->next = freeobj++ << OBJ_TAG_BITS; |
5538c562 | 1044 | link += class->size / sizeof(*link); |
61989a80 NG |
1045 | } |
1046 | ||
1047 | /* | |
1048 | * We now come to the last (full or partial) object on this | |
1049 | * page, which must point to the first object on the next | |
1050 | * page (if present) | |
1051 | */ | |
1052 | next_page = get_next_page(page); | |
bfd093f5 | 1053 | if (next_page) { |
3b1d9ca6 | 1054 | link->next = freeobj++ << OBJ_TAG_BITS; |
bfd093f5 MK |
1055 | } else { |
1056 | /* | |
3b1d9ca6 | 1057 | * Reset OBJ_TAG_BITS bit to last link to tell |
bfd093f5 MK |
1058 | * whether it's allocated object or not. |
1059 | */ | |
3b1d9ca6 | 1060 | link->next = -1 << OBJ_TAG_BITS; |
bfd093f5 | 1061 | } |
af4ee5e9 | 1062 | kunmap_atomic(vaddr); |
61989a80 | 1063 | page = next_page; |
5538c562 | 1064 | off %= PAGE_SIZE; |
61989a80 | 1065 | } |
bdb0af7c | 1066 | |
bfd093f5 | 1067 | set_freeobj(zspage, 0); |
61989a80 NG |
1068 | } |
1069 | ||
48b4800a MK |
1070 | static void create_page_chain(struct size_class *class, struct zspage *zspage, |
1071 | struct page *pages[]) | |
61989a80 | 1072 | { |
bdb0af7c MK |
1073 | int i; |
1074 | struct page *page; | |
1075 | struct page *prev_page = NULL; | |
48b4800a | 1076 | int nr_pages = class->pages_per_zspage; |
61989a80 NG |
1077 | |
1078 | /* | |
1079 | * Allocate individual pages and link them together as: | |
48b4800a | 1080 | * 1. all pages are linked together using page->freelist |
3783689a | 1081 | * 2. each sub-page point to zspage using page->private |
61989a80 | 1082 | * |
3783689a | 1083 | * we set PG_private to identify the first page (i.e. no other sub-page |
22c5cef1 | 1084 | * has this flag set). |
61989a80 | 1085 | */ |
bdb0af7c MK |
1086 | for (i = 0; i < nr_pages; i++) { |
1087 | page = pages[i]; | |
3783689a | 1088 | set_page_private(page, (unsigned long)zspage); |
48b4800a | 1089 | page->freelist = NULL; |
bdb0af7c | 1090 | if (i == 0) { |
3783689a | 1091 | zspage->first_page = page; |
a27545bf | 1092 | SetPagePrivate(page); |
48b4800a MK |
1093 | if (unlikely(class->objs_per_zspage == 1 && |
1094 | class->pages_per_zspage == 1)) | |
1095 | SetPageHugeObject(page); | |
3783689a | 1096 | } else { |
48b4800a | 1097 | prev_page->freelist = page; |
61989a80 | 1098 | } |
61989a80 NG |
1099 | prev_page = page; |
1100 | } | |
bdb0af7c | 1101 | } |
61989a80 | 1102 | |
bdb0af7c MK |
1103 | /* |
1104 | * Allocate a zspage for the given size class | |
1105 | */ | |
3783689a MK |
1106 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
1107 | struct size_class *class, | |
1108 | gfp_t gfp) | |
bdb0af7c MK |
1109 | { |
1110 | int i; | |
bdb0af7c | 1111 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1112 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1113 | ||
1114 | if (!zspage) | |
1115 | return NULL; | |
1116 | ||
1117 | memset(zspage, 0, sizeof(struct zspage)); | |
48b4800a MK |
1118 | zspage->magic = ZSPAGE_MAGIC; |
1119 | migrate_lock_init(zspage); | |
61989a80 | 1120 | |
bdb0af7c MK |
1121 | for (i = 0; i < class->pages_per_zspage; i++) { |
1122 | struct page *page; | |
61989a80 | 1123 | |
3783689a | 1124 | page = alloc_page(gfp); |
bdb0af7c | 1125 | if (!page) { |
91537fee MK |
1126 | while (--i >= 0) { |
1127 | dec_zone_page_state(pages[i], NR_ZSPAGES); | |
bdb0af7c | 1128 | __free_page(pages[i]); |
91537fee | 1129 | } |
3783689a | 1130 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1131 | return NULL; |
1132 | } | |
91537fee MK |
1133 | |
1134 | inc_zone_page_state(page, NR_ZSPAGES); | |
bdb0af7c | 1135 | pages[i] = page; |
61989a80 NG |
1136 | } |
1137 | ||
48b4800a | 1138 | create_page_chain(class, zspage, pages); |
3783689a | 1139 | init_zspage(class, zspage); |
bdb0af7c | 1140 | |
3783689a | 1141 | return zspage; |
61989a80 NG |
1142 | } |
1143 | ||
3783689a | 1144 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1145 | { |
1146 | int i; | |
3783689a | 1147 | struct zspage *zspage; |
61989a80 | 1148 | |
48b4800a | 1149 | for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) { |
3783689a MK |
1150 | zspage = list_first_entry_or_null(&class->fullness_list[i], |
1151 | struct zspage, list); | |
1152 | if (zspage) | |
61989a80 NG |
1153 | break; |
1154 | } | |
1155 | ||
3783689a | 1156 | return zspage; |
61989a80 NG |
1157 | } |
1158 | ||
1b945aee | 1159 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1160 | static inline int __zs_cpu_up(struct mapping_area *area) |
1161 | { | |
1162 | /* | |
1163 | * Make sure we don't leak memory if a cpu UP notification | |
1164 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1165 | */ | |
1166 | if (area->vm) | |
1167 | return 0; | |
1168 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1169 | if (!area->vm) | |
1170 | return -ENOMEM; | |
1171 | return 0; | |
1172 | } | |
1173 | ||
1174 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1175 | { | |
1176 | if (area->vm) | |
1177 | free_vm_area(area->vm); | |
1178 | area->vm = NULL; | |
1179 | } | |
1180 | ||
1181 | static inline void *__zs_map_object(struct mapping_area *area, | |
1182 | struct page *pages[2], int off, int size) | |
1183 | { | |
f6f8ed47 | 1184 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1185 | area->vm_addr = area->vm->addr; |
1186 | return area->vm_addr + off; | |
1187 | } | |
1188 | ||
1189 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1190 | struct page *pages[2], int off, int size) | |
1191 | { | |
1192 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1193 | |
d95abbbb | 1194 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1195 | } |
1196 | ||
1b945aee | 1197 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1198 | |
1199 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1200 | { | |
1201 | /* | |
1202 | * Make sure we don't leak memory if a cpu UP notification | |
1203 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1204 | */ | |
1205 | if (area->vm_buf) | |
1206 | return 0; | |
40f9fb8c | 1207 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1208 | if (!area->vm_buf) |
1209 | return -ENOMEM; | |
1210 | return 0; | |
1211 | } | |
1212 | ||
1213 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1214 | { | |
40f9fb8c | 1215 | kfree(area->vm_buf); |
f553646a SJ |
1216 | area->vm_buf = NULL; |
1217 | } | |
1218 | ||
1219 | static void *__zs_map_object(struct mapping_area *area, | |
1220 | struct page *pages[2], int off, int size) | |
5f601902 | 1221 | { |
5f601902 SJ |
1222 | int sizes[2]; |
1223 | void *addr; | |
f553646a | 1224 | char *buf = area->vm_buf; |
5f601902 | 1225 | |
f553646a SJ |
1226 | /* disable page faults to match kmap_atomic() return conditions */ |
1227 | pagefault_disable(); | |
1228 | ||
1229 | /* no read fastpath */ | |
1230 | if (area->vm_mm == ZS_MM_WO) | |
1231 | goto out; | |
5f601902 SJ |
1232 | |
1233 | sizes[0] = PAGE_SIZE - off; | |
1234 | sizes[1] = size - sizes[0]; | |
1235 | ||
5f601902 SJ |
1236 | /* copy object to per-cpu buffer */ |
1237 | addr = kmap_atomic(pages[0]); | |
1238 | memcpy(buf, addr + off, sizes[0]); | |
1239 | kunmap_atomic(addr); | |
1240 | addr = kmap_atomic(pages[1]); | |
1241 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1242 | kunmap_atomic(addr); | |
f553646a SJ |
1243 | out: |
1244 | return area->vm_buf; | |
5f601902 SJ |
1245 | } |
1246 | ||
f553646a SJ |
1247 | static void __zs_unmap_object(struct mapping_area *area, |
1248 | struct page *pages[2], int off, int size) | |
5f601902 | 1249 | { |
5f601902 SJ |
1250 | int sizes[2]; |
1251 | void *addr; | |
2e40e163 | 1252 | char *buf; |
5f601902 | 1253 | |
f553646a SJ |
1254 | /* no write fastpath */ |
1255 | if (area->vm_mm == ZS_MM_RO) | |
1256 | goto out; | |
5f601902 | 1257 | |
7b60a685 | 1258 | buf = area->vm_buf; |
a82cbf07 YX |
1259 | buf = buf + ZS_HANDLE_SIZE; |
1260 | size -= ZS_HANDLE_SIZE; | |
1261 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1262 | |
5f601902 SJ |
1263 | sizes[0] = PAGE_SIZE - off; |
1264 | sizes[1] = size - sizes[0]; | |
1265 | ||
1266 | /* copy per-cpu buffer to object */ | |
1267 | addr = kmap_atomic(pages[0]); | |
1268 | memcpy(addr + off, buf, sizes[0]); | |
1269 | kunmap_atomic(addr); | |
1270 | addr = kmap_atomic(pages[1]); | |
1271 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1272 | kunmap_atomic(addr); | |
f553646a SJ |
1273 | |
1274 | out: | |
1275 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1276 | pagefault_enable(); | |
5f601902 | 1277 | } |
61989a80 | 1278 | |
1b945aee | 1279 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1280 | |
215c89d0 | 1281 | static int zs_cpu_prepare(unsigned int cpu) |
61989a80 | 1282 | { |
61989a80 NG |
1283 | struct mapping_area *area; |
1284 | ||
215c89d0 SAS |
1285 | area = &per_cpu(zs_map_area, cpu); |
1286 | return __zs_cpu_up(area); | |
61989a80 NG |
1287 | } |
1288 | ||
215c89d0 | 1289 | static int zs_cpu_dead(unsigned int cpu) |
61989a80 | 1290 | { |
215c89d0 | 1291 | struct mapping_area *area; |
40f9fb8c | 1292 | |
215c89d0 SAS |
1293 | area = &per_cpu(zs_map_area, cpu); |
1294 | __zs_cpu_down(area); | |
1295 | return 0; | |
b1b00a5b SS |
1296 | } |
1297 | ||
64d90465 GM |
1298 | static bool can_merge(struct size_class *prev, int pages_per_zspage, |
1299 | int objs_per_zspage) | |
9eec4cd5 | 1300 | { |
64d90465 GM |
1301 | if (prev->pages_per_zspage == pages_per_zspage && |
1302 | prev->objs_per_zspage == objs_per_zspage) | |
1303 | return true; | |
9eec4cd5 | 1304 | |
64d90465 | 1305 | return false; |
9eec4cd5 JK |
1306 | } |
1307 | ||
3783689a | 1308 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1309 | { |
3783689a | 1310 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1311 | } |
1312 | ||
66cdef66 GM |
1313 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1314 | { | |
1315 | return atomic_long_read(&pool->pages_allocated); | |
1316 | } | |
1317 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1318 | ||
4bbc0bc0 | 1319 | /** |
66cdef66 GM |
1320 | * zs_map_object - get address of allocated object from handle. |
1321 | * @pool: pool from which the object was allocated | |
1322 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1323 | * |
66cdef66 GM |
1324 | * Before using an object allocated from zs_malloc, it must be mapped using |
1325 | * this function. When done with the object, it must be unmapped using | |
1326 | * zs_unmap_object. | |
4bbc0bc0 | 1327 | * |
66cdef66 GM |
1328 | * Only one object can be mapped per cpu at a time. There is no protection |
1329 | * against nested mappings. | |
1330 | * | |
1331 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1332 | */ |
66cdef66 GM |
1333 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1334 | enum zs_mapmode mm) | |
61989a80 | 1335 | { |
3783689a | 1336 | struct zspage *zspage; |
66cdef66 | 1337 | struct page *page; |
bfd093f5 MK |
1338 | unsigned long obj, off; |
1339 | unsigned int obj_idx; | |
61989a80 | 1340 | |
66cdef66 GM |
1341 | unsigned int class_idx; |
1342 | enum fullness_group fg; | |
1343 | struct size_class *class; | |
1344 | struct mapping_area *area; | |
1345 | struct page *pages[2]; | |
2e40e163 | 1346 | void *ret; |
61989a80 | 1347 | |
9eec4cd5 | 1348 | /* |
66cdef66 GM |
1349 | * Because we use per-cpu mapping areas shared among the |
1350 | * pools/users, we can't allow mapping in interrupt context | |
1351 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1352 | */ |
12383340 | 1353 | BUG_ON(in_interrupt()); |
61989a80 | 1354 | |
312fcae2 MK |
1355 | /* From now on, migration cannot move the object */ |
1356 | pin_tag(handle); | |
1357 | ||
2e40e163 MK |
1358 | obj = handle_to_obj(handle); |
1359 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1360 | zspage = get_zspage(page); |
48b4800a MK |
1361 | |
1362 | /* migration cannot move any subpage in this zspage */ | |
1363 | migrate_read_lock(zspage); | |
1364 | ||
3783689a | 1365 | get_zspage_mapping(zspage, &class_idx, &fg); |
66cdef66 | 1366 | class = pool->size_class[class_idx]; |
bfd093f5 | 1367 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1368 | |
66cdef66 GM |
1369 | area = &get_cpu_var(zs_map_area); |
1370 | area->vm_mm = mm; | |
1371 | if (off + class->size <= PAGE_SIZE) { | |
1372 | /* this object is contained entirely within a page */ | |
1373 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1374 | ret = area->vm_addr + off; |
1375 | goto out; | |
61989a80 NG |
1376 | } |
1377 | ||
66cdef66 GM |
1378 | /* this object spans two pages */ |
1379 | pages[0] = page; | |
1380 | pages[1] = get_next_page(page); | |
1381 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1382 | |
2e40e163 MK |
1383 | ret = __zs_map_object(area, pages, off, class->size); |
1384 | out: | |
48b4800a | 1385 | if (likely(!PageHugeObject(page))) |
7b60a685 MK |
1386 | ret += ZS_HANDLE_SIZE; |
1387 | ||
1388 | return ret; | |
61989a80 | 1389 | } |
66cdef66 | 1390 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1391 | |
66cdef66 | 1392 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1393 | { |
3783689a | 1394 | struct zspage *zspage; |
66cdef66 | 1395 | struct page *page; |
bfd093f5 MK |
1396 | unsigned long obj, off; |
1397 | unsigned int obj_idx; | |
61989a80 | 1398 | |
66cdef66 GM |
1399 | unsigned int class_idx; |
1400 | enum fullness_group fg; | |
1401 | struct size_class *class; | |
1402 | struct mapping_area *area; | |
9eec4cd5 | 1403 | |
2e40e163 MK |
1404 | obj = handle_to_obj(handle); |
1405 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1406 | zspage = get_zspage(page); |
1407 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 | 1408 | class = pool->size_class[class_idx]; |
bfd093f5 | 1409 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1410 | |
66cdef66 GM |
1411 | area = this_cpu_ptr(&zs_map_area); |
1412 | if (off + class->size <= PAGE_SIZE) | |
1413 | kunmap_atomic(area->vm_addr); | |
1414 | else { | |
1415 | struct page *pages[2]; | |
40f9fb8c | 1416 | |
66cdef66 GM |
1417 | pages[0] = page; |
1418 | pages[1] = get_next_page(page); | |
1419 | BUG_ON(!pages[1]); | |
1420 | ||
1421 | __zs_unmap_object(area, pages, off, class->size); | |
1422 | } | |
1423 | put_cpu_var(zs_map_area); | |
48b4800a MK |
1424 | |
1425 | migrate_read_unlock(zspage); | |
312fcae2 | 1426 | unpin_tag(handle); |
61989a80 | 1427 | } |
66cdef66 | 1428 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1429 | |
7f47f23b SS |
1430 | /** |
1431 | * zs_huge_class_size() - Returns the size (in bytes) of the first huge | |
1432 | * zsmalloc &size_class. | |
1433 | * @pool: zsmalloc pool to use | |
1434 | * | |
1435 | * The function returns the size of the first huge class - any object of equal | |
1436 | * or bigger size will be stored in zspage consisting of a single physical | |
1437 | * page. | |
1438 | * | |
1439 | * Context: Any context. | |
1440 | * | |
1441 | * Return: the size (in bytes) of the first huge zsmalloc &size_class. | |
1442 | */ | |
1443 | size_t zs_huge_class_size(struct zs_pool *pool) | |
1444 | { | |
1445 | return huge_class_size; | |
1446 | } | |
1447 | EXPORT_SYMBOL_GPL(zs_huge_class_size); | |
1448 | ||
251cbb95 | 1449 | static unsigned long obj_malloc(struct size_class *class, |
3783689a | 1450 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1451 | { |
bfd093f5 | 1452 | int i, nr_page, offset; |
c7806261 MK |
1453 | unsigned long obj; |
1454 | struct link_free *link; | |
1455 | ||
1456 | struct page *m_page; | |
bfd093f5 | 1457 | unsigned long m_offset; |
c7806261 MK |
1458 | void *vaddr; |
1459 | ||
312fcae2 | 1460 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1461 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1462 | |
1463 | offset = obj * class->size; | |
1464 | nr_page = offset >> PAGE_SHIFT; | |
1465 | m_offset = offset & ~PAGE_MASK; | |
1466 | m_page = get_first_page(zspage); | |
1467 | ||
1468 | for (i = 0; i < nr_page; i++) | |
1469 | m_page = get_next_page(m_page); | |
c7806261 MK |
1470 | |
1471 | vaddr = kmap_atomic(m_page); | |
1472 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3b1d9ca6 | 1473 | set_freeobj(zspage, link->next >> OBJ_TAG_BITS); |
48b4800a | 1474 | if (likely(!PageHugeObject(m_page))) |
7b60a685 MK |
1475 | /* record handle in the header of allocated chunk */ |
1476 | link->handle = handle; | |
1477 | else | |
3783689a MK |
1478 | /* record handle to page->index */ |
1479 | zspage->first_page->index = handle; | |
1480 | ||
c7806261 | 1481 | kunmap_atomic(vaddr); |
3783689a | 1482 | mod_zspage_inuse(zspage, 1); |
c7806261 MK |
1483 | zs_stat_inc(class, OBJ_USED, 1); |
1484 | ||
bfd093f5 MK |
1485 | obj = location_to_obj(m_page, obj); |
1486 | ||
c7806261 MK |
1487 | return obj; |
1488 | } | |
1489 | ||
1490 | ||
61989a80 NG |
1491 | /** |
1492 | * zs_malloc - Allocate block of given size from pool. | |
1493 | * @pool: pool to allocate from | |
1494 | * @size: size of block to allocate | |
fd854463 | 1495 | * @gfp: gfp flags when allocating object |
61989a80 | 1496 | * |
00a61d86 | 1497 | * On success, handle to the allocated object is returned, |
c2344348 | 1498 | * otherwise 0. |
61989a80 NG |
1499 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1500 | */ | |
d0d8da2d | 1501 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1502 | { |
2e40e163 | 1503 | unsigned long handle, obj; |
61989a80 | 1504 | struct size_class *class; |
48b4800a | 1505 | enum fullness_group newfg; |
3783689a | 1506 | struct zspage *zspage; |
61989a80 | 1507 | |
7b60a685 | 1508 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1509 | return 0; |
1510 | ||
3783689a | 1511 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1512 | if (!handle) |
c2344348 | 1513 | return 0; |
61989a80 | 1514 | |
2e40e163 MK |
1515 | /* extra space in chunk to keep the handle */ |
1516 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1517 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1518 | |
1519 | spin_lock(&class->lock); | |
3783689a | 1520 | zspage = find_get_zspage(class); |
48b4800a MK |
1521 | if (likely(zspage)) { |
1522 | obj = obj_malloc(class, zspage, handle); | |
1523 | /* Now move the zspage to another fullness group, if required */ | |
1524 | fix_fullness_group(class, zspage); | |
1525 | record_obj(handle, obj); | |
61989a80 | 1526 | spin_unlock(&class->lock); |
61989a80 | 1527 | |
48b4800a MK |
1528 | return handle; |
1529 | } | |
0f050d99 | 1530 | |
48b4800a MK |
1531 | spin_unlock(&class->lock); |
1532 | ||
1533 | zspage = alloc_zspage(pool, class, gfp); | |
1534 | if (!zspage) { | |
1535 | cache_free_handle(pool, handle); | |
1536 | return 0; | |
61989a80 NG |
1537 | } |
1538 | ||
48b4800a | 1539 | spin_lock(&class->lock); |
3783689a | 1540 | obj = obj_malloc(class, zspage, handle); |
48b4800a MK |
1541 | newfg = get_fullness_group(class, zspage); |
1542 | insert_zspage(class, zspage, newfg); | |
1543 | set_zspage_mapping(zspage, class->index, newfg); | |
2e40e163 | 1544 | record_obj(handle, obj); |
48b4800a MK |
1545 | atomic_long_add(class->pages_per_zspage, |
1546 | &pool->pages_allocated); | |
b4fd07a0 | 1547 | zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
1548 | |
1549 | /* We completely set up zspage so mark them as movable */ | |
1550 | SetZsPageMovable(pool, zspage); | |
61989a80 NG |
1551 | spin_unlock(&class->lock); |
1552 | ||
2e40e163 | 1553 | return handle; |
61989a80 NG |
1554 | } |
1555 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1556 | ||
1ee47165 | 1557 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1558 | { |
1559 | struct link_free *link; | |
3783689a MK |
1560 | struct zspage *zspage; |
1561 | struct page *f_page; | |
bfd093f5 MK |
1562 | unsigned long f_offset; |
1563 | unsigned int f_objidx; | |
af4ee5e9 | 1564 | void *vaddr; |
61989a80 | 1565 | |
312fcae2 | 1566 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1567 | obj_to_location(obj, &f_page, &f_objidx); |
bfd093f5 | 1568 | f_offset = (class->size * f_objidx) & ~PAGE_MASK; |
3783689a | 1569 | zspage = get_zspage(f_page); |
61989a80 | 1570 | |
c7806261 | 1571 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1572 | |
1573 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1574 | link = (struct link_free *)(vaddr + f_offset); |
3b1d9ca6 | 1575 | link->next = get_freeobj(zspage) << OBJ_TAG_BITS; |
af4ee5e9 | 1576 | kunmap_atomic(vaddr); |
bfd093f5 | 1577 | set_freeobj(zspage, f_objidx); |
3783689a | 1578 | mod_zspage_inuse(zspage, -1); |
0f050d99 | 1579 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1580 | } |
1581 | ||
1582 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1583 | { | |
3783689a MK |
1584 | struct zspage *zspage; |
1585 | struct page *f_page; | |
bfd093f5 MK |
1586 | unsigned long obj; |
1587 | unsigned int f_objidx; | |
c7806261 MK |
1588 | int class_idx; |
1589 | struct size_class *class; | |
1590 | enum fullness_group fullness; | |
48b4800a | 1591 | bool isolated; |
c7806261 MK |
1592 | |
1593 | if (unlikely(!handle)) | |
1594 | return; | |
1595 | ||
312fcae2 | 1596 | pin_tag(handle); |
c7806261 | 1597 | obj = handle_to_obj(handle); |
c7806261 | 1598 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1599 | zspage = get_zspage(f_page); |
c7806261 | 1600 | |
48b4800a MK |
1601 | migrate_read_lock(zspage); |
1602 | ||
3783689a | 1603 | get_zspage_mapping(zspage, &class_idx, &fullness); |
c7806261 MK |
1604 | class = pool->size_class[class_idx]; |
1605 | ||
1606 | spin_lock(&class->lock); | |
1ee47165 | 1607 | obj_free(class, obj); |
3783689a | 1608 | fullness = fix_fullness_group(class, zspage); |
48b4800a MK |
1609 | if (fullness != ZS_EMPTY) { |
1610 | migrate_read_unlock(zspage); | |
1611 | goto out; | |
312fcae2 | 1612 | } |
48b4800a MK |
1613 | |
1614 | isolated = is_zspage_isolated(zspage); | |
1615 | migrate_read_unlock(zspage); | |
1616 | /* If zspage is isolated, zs_page_putback will free the zspage */ | |
1617 | if (likely(!isolated)) | |
1618 | free_zspage(pool, class, zspage); | |
1619 | out: | |
1620 | ||
61989a80 | 1621 | spin_unlock(&class->lock); |
312fcae2 | 1622 | unpin_tag(handle); |
3783689a | 1623 | cache_free_handle(pool, handle); |
312fcae2 MK |
1624 | } |
1625 | EXPORT_SYMBOL_GPL(zs_free); | |
1626 | ||
251cbb95 MK |
1627 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1628 | unsigned long src) | |
312fcae2 MK |
1629 | { |
1630 | struct page *s_page, *d_page; | |
bfd093f5 | 1631 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1632 | unsigned long s_off, d_off; |
1633 | void *s_addr, *d_addr; | |
1634 | int s_size, d_size, size; | |
1635 | int written = 0; | |
1636 | ||
1637 | s_size = d_size = class->size; | |
1638 | ||
1639 | obj_to_location(src, &s_page, &s_objidx); | |
1640 | obj_to_location(dst, &d_page, &d_objidx); | |
1641 | ||
bfd093f5 MK |
1642 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1643 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1644 | |
1645 | if (s_off + class->size > PAGE_SIZE) | |
1646 | s_size = PAGE_SIZE - s_off; | |
1647 | ||
1648 | if (d_off + class->size > PAGE_SIZE) | |
1649 | d_size = PAGE_SIZE - d_off; | |
1650 | ||
1651 | s_addr = kmap_atomic(s_page); | |
1652 | d_addr = kmap_atomic(d_page); | |
1653 | ||
1654 | while (1) { | |
1655 | size = min(s_size, d_size); | |
1656 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1657 | written += size; | |
1658 | ||
1659 | if (written == class->size) | |
1660 | break; | |
1661 | ||
495819ea SS |
1662 | s_off += size; |
1663 | s_size -= size; | |
1664 | d_off += size; | |
1665 | d_size -= size; | |
1666 | ||
1667 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1668 | kunmap_atomic(d_addr); |
1669 | kunmap_atomic(s_addr); | |
1670 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1671 | s_addr = kmap_atomic(s_page); |
1672 | d_addr = kmap_atomic(d_page); | |
1673 | s_size = class->size - written; | |
1674 | s_off = 0; | |
312fcae2 MK |
1675 | } |
1676 | ||
495819ea | 1677 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1678 | kunmap_atomic(d_addr); |
1679 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1680 | d_addr = kmap_atomic(d_page); |
1681 | d_size = class->size - written; | |
1682 | d_off = 0; | |
312fcae2 MK |
1683 | } |
1684 | } | |
1685 | ||
1686 | kunmap_atomic(d_addr); | |
1687 | kunmap_atomic(s_addr); | |
1688 | } | |
1689 | ||
1690 | /* | |
1691 | * Find alloced object in zspage from index object and | |
1692 | * return handle. | |
1693 | */ | |
251cbb95 | 1694 | static unsigned long find_alloced_obj(struct size_class *class, |
cf675acb | 1695 | struct page *page, int *obj_idx) |
312fcae2 MK |
1696 | { |
1697 | unsigned long head; | |
1698 | int offset = 0; | |
cf675acb | 1699 | int index = *obj_idx; |
312fcae2 MK |
1700 | unsigned long handle = 0; |
1701 | void *addr = kmap_atomic(page); | |
1702 | ||
3783689a | 1703 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1704 | offset += class->size * index; |
1705 | ||
1706 | while (offset < PAGE_SIZE) { | |
48b4800a | 1707 | head = obj_to_head(page, addr + offset); |
312fcae2 MK |
1708 | if (head & OBJ_ALLOCATED_TAG) { |
1709 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1710 | if (trypin_tag(handle)) | |
1711 | break; | |
1712 | handle = 0; | |
1713 | } | |
1714 | ||
1715 | offset += class->size; | |
1716 | index++; | |
1717 | } | |
1718 | ||
1719 | kunmap_atomic(addr); | |
cf675acb GM |
1720 | |
1721 | *obj_idx = index; | |
1722 | ||
312fcae2 MK |
1723 | return handle; |
1724 | } | |
1725 | ||
1726 | struct zs_compact_control { | |
3783689a | 1727 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1728 | struct page *s_page; |
1729 | /* Destination page for migration which should be a first page | |
1730 | * of zspage. */ | |
1731 | struct page *d_page; | |
1732 | /* Starting object index within @s_page which used for live object | |
1733 | * in the subpage. */ | |
41b88e14 | 1734 | int obj_idx; |
312fcae2 MK |
1735 | }; |
1736 | ||
1737 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1738 | struct zs_compact_control *cc) | |
1739 | { | |
1740 | unsigned long used_obj, free_obj; | |
1741 | unsigned long handle; | |
1742 | struct page *s_page = cc->s_page; | |
1743 | struct page *d_page = cc->d_page; | |
41b88e14 | 1744 | int obj_idx = cc->obj_idx; |
312fcae2 MK |
1745 | int ret = 0; |
1746 | ||
1747 | while (1) { | |
cf675acb | 1748 | handle = find_alloced_obj(class, s_page, &obj_idx); |
312fcae2 MK |
1749 | if (!handle) { |
1750 | s_page = get_next_page(s_page); | |
1751 | if (!s_page) | |
1752 | break; | |
41b88e14 | 1753 | obj_idx = 0; |
312fcae2 MK |
1754 | continue; |
1755 | } | |
1756 | ||
1757 | /* Stop if there is no more space */ | |
3783689a | 1758 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1759 | unpin_tag(handle); |
1760 | ret = -ENOMEM; | |
1761 | break; | |
1762 | } | |
1763 | ||
1764 | used_obj = handle_to_obj(handle); | |
3783689a | 1765 | free_obj = obj_malloc(class, get_zspage(d_page), handle); |
251cbb95 | 1766 | zs_object_copy(class, free_obj, used_obj); |
41b88e14 | 1767 | obj_idx++; |
c102f07c JL |
1768 | /* |
1769 | * record_obj updates handle's value to free_obj and it will | |
1770 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1771 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1772 | * let's keep the lock bit. | |
1773 | */ | |
1774 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1775 | record_obj(handle, free_obj); |
1776 | unpin_tag(handle); | |
1ee47165 | 1777 | obj_free(class, used_obj); |
312fcae2 MK |
1778 | } |
1779 | ||
1780 | /* Remember last position in this iteration */ | |
1781 | cc->s_page = s_page; | |
41b88e14 | 1782 | cc->obj_idx = obj_idx; |
312fcae2 MK |
1783 | |
1784 | return ret; | |
1785 | } | |
1786 | ||
3783689a | 1787 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1788 | { |
1789 | int i; | |
3783689a MK |
1790 | struct zspage *zspage; |
1791 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1792 | |
3783689a MK |
1793 | if (!source) { |
1794 | fg[0] = ZS_ALMOST_FULL; | |
1795 | fg[1] = ZS_ALMOST_EMPTY; | |
1796 | } | |
1797 | ||
1798 | for (i = 0; i < 2; i++) { | |
1799 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1800 | struct zspage, list); | |
1801 | if (zspage) { | |
48b4800a | 1802 | VM_BUG_ON(is_zspage_isolated(zspage)); |
3783689a MK |
1803 | remove_zspage(class, zspage, fg[i]); |
1804 | return zspage; | |
312fcae2 MK |
1805 | } |
1806 | } | |
1807 | ||
3783689a | 1808 | return zspage; |
312fcae2 MK |
1809 | } |
1810 | ||
860c707d | 1811 | /* |
3783689a | 1812 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1813 | * @class: destination class |
3783689a | 1814 | * @zspage: target page |
860c707d | 1815 | * |
3783689a | 1816 | * Return @zspage's fullness_group |
860c707d | 1817 | */ |
4aa409ca | 1818 | static enum fullness_group putback_zspage(struct size_class *class, |
3783689a | 1819 | struct zspage *zspage) |
312fcae2 | 1820 | { |
312fcae2 MK |
1821 | enum fullness_group fullness; |
1822 | ||
48b4800a MK |
1823 | VM_BUG_ON(is_zspage_isolated(zspage)); |
1824 | ||
3783689a MK |
1825 | fullness = get_fullness_group(class, zspage); |
1826 | insert_zspage(class, zspage, fullness); | |
1827 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1828 | |
860c707d | 1829 | return fullness; |
61989a80 | 1830 | } |
312fcae2 | 1831 | |
48b4800a MK |
1832 | #ifdef CONFIG_COMPACTION |
1833 | static struct dentry *zs_mount(struct file_system_type *fs_type, | |
1834 | int flags, const char *dev_name, void *data) | |
1835 | { | |
1836 | static const struct dentry_operations ops = { | |
1837 | .d_dname = simple_dname, | |
1838 | }; | |
1839 | ||
1840 | return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC); | |
1841 | } | |
1842 | ||
1843 | static struct file_system_type zsmalloc_fs = { | |
1844 | .name = "zsmalloc", | |
1845 | .mount = zs_mount, | |
1846 | .kill_sb = kill_anon_super, | |
1847 | }; | |
1848 | ||
1849 | static int zsmalloc_mount(void) | |
1850 | { | |
1851 | int ret = 0; | |
1852 | ||
1853 | zsmalloc_mnt = kern_mount(&zsmalloc_fs); | |
1854 | if (IS_ERR(zsmalloc_mnt)) | |
1855 | ret = PTR_ERR(zsmalloc_mnt); | |
1856 | ||
1857 | return ret; | |
1858 | } | |
1859 | ||
1860 | static void zsmalloc_unmount(void) | |
1861 | { | |
1862 | kern_unmount(zsmalloc_mnt); | |
1863 | } | |
1864 | ||
1865 | static void migrate_lock_init(struct zspage *zspage) | |
1866 | { | |
1867 | rwlock_init(&zspage->lock); | |
1868 | } | |
1869 | ||
1870 | static void migrate_read_lock(struct zspage *zspage) | |
1871 | { | |
1872 | read_lock(&zspage->lock); | |
1873 | } | |
1874 | ||
1875 | static void migrate_read_unlock(struct zspage *zspage) | |
1876 | { | |
1877 | read_unlock(&zspage->lock); | |
1878 | } | |
1879 | ||
1880 | static void migrate_write_lock(struct zspage *zspage) | |
1881 | { | |
1882 | write_lock(&zspage->lock); | |
1883 | } | |
1884 | ||
1885 | static void migrate_write_unlock(struct zspage *zspage) | |
1886 | { | |
1887 | write_unlock(&zspage->lock); | |
1888 | } | |
1889 | ||
1890 | /* Number of isolated subpage for *page migration* in this zspage */ | |
1891 | static void inc_zspage_isolation(struct zspage *zspage) | |
1892 | { | |
1893 | zspage->isolated++; | |
1894 | } | |
1895 | ||
1896 | static void dec_zspage_isolation(struct zspage *zspage) | |
1897 | { | |
1898 | zspage->isolated--; | |
1899 | } | |
1900 | ||
1901 | static void replace_sub_page(struct size_class *class, struct zspage *zspage, | |
1902 | struct page *newpage, struct page *oldpage) | |
1903 | { | |
1904 | struct page *page; | |
1905 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, }; | |
1906 | int idx = 0; | |
1907 | ||
1908 | page = get_first_page(zspage); | |
1909 | do { | |
1910 | if (page == oldpage) | |
1911 | pages[idx] = newpage; | |
1912 | else | |
1913 | pages[idx] = page; | |
1914 | idx++; | |
1915 | } while ((page = get_next_page(page)) != NULL); | |
1916 | ||
1917 | create_page_chain(class, zspage, pages); | |
1918 | set_first_obj_offset(newpage, get_first_obj_offset(oldpage)); | |
1919 | if (unlikely(PageHugeObject(oldpage))) | |
1920 | newpage->index = oldpage->index; | |
1921 | __SetPageMovable(newpage, page_mapping(oldpage)); | |
1922 | } | |
1923 | ||
1924 | bool zs_page_isolate(struct page *page, isolate_mode_t mode) | |
1925 | { | |
1926 | struct zs_pool *pool; | |
1927 | struct size_class *class; | |
1928 | int class_idx; | |
1929 | enum fullness_group fullness; | |
1930 | struct zspage *zspage; | |
1931 | struct address_space *mapping; | |
1932 | ||
1933 | /* | |
1934 | * Page is locked so zspage couldn't be destroyed. For detail, look at | |
1935 | * lock_zspage in free_zspage. | |
1936 | */ | |
1937 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1938 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1939 | ||
1940 | zspage = get_zspage(page); | |
1941 | ||
1942 | /* | |
1943 | * Without class lock, fullness could be stale while class_idx is okay | |
1944 | * because class_idx is constant unless page is freed so we should get | |
1945 | * fullness again under class lock. | |
1946 | */ | |
1947 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1948 | mapping = page_mapping(page); | |
1949 | pool = mapping->private_data; | |
1950 | class = pool->size_class[class_idx]; | |
1951 | ||
1952 | spin_lock(&class->lock); | |
1953 | if (get_zspage_inuse(zspage) == 0) { | |
1954 | spin_unlock(&class->lock); | |
1955 | return false; | |
1956 | } | |
1957 | ||
1958 | /* zspage is isolated for object migration */ | |
1959 | if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
1960 | spin_unlock(&class->lock); | |
1961 | return false; | |
1962 | } | |
1963 | ||
1964 | /* | |
1965 | * If this is first time isolation for the zspage, isolate zspage from | |
1966 | * size_class to prevent further object allocation from the zspage. | |
1967 | */ | |
1968 | if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
1969 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1970 | remove_zspage(class, zspage, fullness); | |
1971 | } | |
1972 | ||
1973 | inc_zspage_isolation(zspage); | |
1974 | spin_unlock(&class->lock); | |
1975 | ||
1976 | return true; | |
1977 | } | |
1978 | ||
1979 | int zs_page_migrate(struct address_space *mapping, struct page *newpage, | |
1980 | struct page *page, enum migrate_mode mode) | |
1981 | { | |
1982 | struct zs_pool *pool; | |
1983 | struct size_class *class; | |
1984 | int class_idx; | |
1985 | enum fullness_group fullness; | |
1986 | struct zspage *zspage; | |
1987 | struct page *dummy; | |
1988 | void *s_addr, *d_addr, *addr; | |
1989 | int offset, pos; | |
1990 | unsigned long handle, head; | |
1991 | unsigned long old_obj, new_obj; | |
1992 | unsigned int obj_idx; | |
1993 | int ret = -EAGAIN; | |
1994 | ||
2916ecc0 JG |
1995 | /* |
1996 | * We cannot support the _NO_COPY case here, because copy needs to | |
1997 | * happen under the zs lock, which does not work with | |
1998 | * MIGRATE_SYNC_NO_COPY workflow. | |
1999 | */ | |
2000 | if (mode == MIGRATE_SYNC_NO_COPY) | |
2001 | return -EINVAL; | |
2002 | ||
48b4800a MK |
2003 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
2004 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
2005 | ||
2006 | zspage = get_zspage(page); | |
2007 | ||
2008 | /* Concurrent compactor cannot migrate any subpage in zspage */ | |
2009 | migrate_write_lock(zspage); | |
2010 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2011 | pool = mapping->private_data; | |
2012 | class = pool->size_class[class_idx]; | |
2013 | offset = get_first_obj_offset(page); | |
2014 | ||
2015 | spin_lock(&class->lock); | |
2016 | if (!get_zspage_inuse(zspage)) { | |
77ff4657 HZ |
2017 | /* |
2018 | * Set "offset" to end of the page so that every loops | |
2019 | * skips unnecessary object scanning. | |
2020 | */ | |
2021 | offset = PAGE_SIZE; | |
48b4800a MK |
2022 | } |
2023 | ||
2024 | pos = offset; | |
2025 | s_addr = kmap_atomic(page); | |
2026 | while (pos < PAGE_SIZE) { | |
2027 | head = obj_to_head(page, s_addr + pos); | |
2028 | if (head & OBJ_ALLOCATED_TAG) { | |
2029 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2030 | if (!trypin_tag(handle)) | |
2031 | goto unpin_objects; | |
2032 | } | |
2033 | pos += class->size; | |
2034 | } | |
2035 | ||
2036 | /* | |
2037 | * Here, any user cannot access all objects in the zspage so let's move. | |
2038 | */ | |
2039 | d_addr = kmap_atomic(newpage); | |
2040 | memcpy(d_addr, s_addr, PAGE_SIZE); | |
2041 | kunmap_atomic(d_addr); | |
2042 | ||
2043 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2044 | addr += class->size) { | |
2045 | head = obj_to_head(page, addr); | |
2046 | if (head & OBJ_ALLOCATED_TAG) { | |
2047 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2048 | if (!testpin_tag(handle)) | |
2049 | BUG(); | |
2050 | ||
2051 | old_obj = handle_to_obj(handle); | |
2052 | obj_to_location(old_obj, &dummy, &obj_idx); | |
2053 | new_obj = (unsigned long)location_to_obj(newpage, | |
2054 | obj_idx); | |
2055 | new_obj |= BIT(HANDLE_PIN_BIT); | |
2056 | record_obj(handle, new_obj); | |
2057 | } | |
2058 | } | |
2059 | ||
2060 | replace_sub_page(class, zspage, newpage, page); | |
2061 | get_page(newpage); | |
2062 | ||
2063 | dec_zspage_isolation(zspage); | |
2064 | ||
2065 | /* | |
2066 | * Page migration is done so let's putback isolated zspage to | |
2067 | * the list if @page is final isolated subpage in the zspage. | |
2068 | */ | |
2069 | if (!is_zspage_isolated(zspage)) | |
2070 | putback_zspage(class, zspage); | |
2071 | ||
2072 | reset_page(page); | |
2073 | put_page(page); | |
2074 | page = newpage; | |
2075 | ||
dd4123f3 | 2076 | ret = MIGRATEPAGE_SUCCESS; |
48b4800a MK |
2077 | unpin_objects: |
2078 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2079 | addr += class->size) { | |
2080 | head = obj_to_head(page, addr); | |
2081 | if (head & OBJ_ALLOCATED_TAG) { | |
2082 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2083 | if (!testpin_tag(handle)) | |
2084 | BUG(); | |
2085 | unpin_tag(handle); | |
2086 | } | |
2087 | } | |
2088 | kunmap_atomic(s_addr); | |
48b4800a MK |
2089 | spin_unlock(&class->lock); |
2090 | migrate_write_unlock(zspage); | |
2091 | ||
2092 | return ret; | |
2093 | } | |
2094 | ||
2095 | void zs_page_putback(struct page *page) | |
2096 | { | |
2097 | struct zs_pool *pool; | |
2098 | struct size_class *class; | |
2099 | int class_idx; | |
2100 | enum fullness_group fg; | |
2101 | struct address_space *mapping; | |
2102 | struct zspage *zspage; | |
2103 | ||
2104 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
2105 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
2106 | ||
2107 | zspage = get_zspage(page); | |
2108 | get_zspage_mapping(zspage, &class_idx, &fg); | |
2109 | mapping = page_mapping(page); | |
2110 | pool = mapping->private_data; | |
2111 | class = pool->size_class[class_idx]; | |
2112 | ||
2113 | spin_lock(&class->lock); | |
2114 | dec_zspage_isolation(zspage); | |
2115 | if (!is_zspage_isolated(zspage)) { | |
2116 | fg = putback_zspage(class, zspage); | |
2117 | /* | |
2118 | * Due to page_lock, we cannot free zspage immediately | |
2119 | * so let's defer. | |
2120 | */ | |
2121 | if (fg == ZS_EMPTY) | |
2122 | schedule_work(&pool->free_work); | |
2123 | } | |
2124 | spin_unlock(&class->lock); | |
2125 | } | |
2126 | ||
2127 | const struct address_space_operations zsmalloc_aops = { | |
2128 | .isolate_page = zs_page_isolate, | |
2129 | .migratepage = zs_page_migrate, | |
2130 | .putback_page = zs_page_putback, | |
2131 | }; | |
2132 | ||
2133 | static int zs_register_migration(struct zs_pool *pool) | |
2134 | { | |
2135 | pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb); | |
2136 | if (IS_ERR(pool->inode)) { | |
2137 | pool->inode = NULL; | |
2138 | return 1; | |
2139 | } | |
2140 | ||
2141 | pool->inode->i_mapping->private_data = pool; | |
2142 | pool->inode->i_mapping->a_ops = &zsmalloc_aops; | |
2143 | return 0; | |
2144 | } | |
2145 | ||
2146 | static void zs_unregister_migration(struct zs_pool *pool) | |
2147 | { | |
2148 | flush_work(&pool->free_work); | |
c3491eca | 2149 | iput(pool->inode); |
48b4800a MK |
2150 | } |
2151 | ||
2152 | /* | |
2153 | * Caller should hold page_lock of all pages in the zspage | |
2154 | * In here, we cannot use zspage meta data. | |
2155 | */ | |
2156 | static void async_free_zspage(struct work_struct *work) | |
2157 | { | |
2158 | int i; | |
2159 | struct size_class *class; | |
2160 | unsigned int class_idx; | |
2161 | enum fullness_group fullness; | |
2162 | struct zspage *zspage, *tmp; | |
2163 | LIST_HEAD(free_pages); | |
2164 | struct zs_pool *pool = container_of(work, struct zs_pool, | |
2165 | free_work); | |
2166 | ||
cf8e0fed | 2167 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
48b4800a MK |
2168 | class = pool->size_class[i]; |
2169 | if (class->index != i) | |
2170 | continue; | |
2171 | ||
2172 | spin_lock(&class->lock); | |
2173 | list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages); | |
2174 | spin_unlock(&class->lock); | |
2175 | } | |
2176 | ||
2177 | ||
2178 | list_for_each_entry_safe(zspage, tmp, &free_pages, list) { | |
2179 | list_del(&zspage->list); | |
2180 | lock_zspage(zspage); | |
2181 | ||
2182 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2183 | VM_BUG_ON(fullness != ZS_EMPTY); | |
2184 | class = pool->size_class[class_idx]; | |
2185 | spin_lock(&class->lock); | |
2186 | __free_zspage(pool, pool->size_class[class_idx], zspage); | |
2187 | spin_unlock(&class->lock); | |
2188 | } | |
2189 | }; | |
2190 | ||
2191 | static void kick_deferred_free(struct zs_pool *pool) | |
2192 | { | |
2193 | schedule_work(&pool->free_work); | |
2194 | } | |
2195 | ||
2196 | static void init_deferred_free(struct zs_pool *pool) | |
2197 | { | |
2198 | INIT_WORK(&pool->free_work, async_free_zspage); | |
2199 | } | |
2200 | ||
2201 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) | |
2202 | { | |
2203 | struct page *page = get_first_page(zspage); | |
2204 | ||
2205 | do { | |
2206 | WARN_ON(!trylock_page(page)); | |
2207 | __SetPageMovable(page, pool->inode->i_mapping); | |
2208 | unlock_page(page); | |
2209 | } while ((page = get_next_page(page)) != NULL); | |
2210 | } | |
2211 | #endif | |
2212 | ||
04f05909 SS |
2213 | /* |
2214 | * | |
2215 | * Based on the number of unused allocated objects calculate | |
2216 | * and return the number of pages that we can free. | |
04f05909 SS |
2217 | */ |
2218 | static unsigned long zs_can_compact(struct size_class *class) | |
2219 | { | |
2220 | unsigned long obj_wasted; | |
44f43e99 SS |
2221 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
2222 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 2223 | |
44f43e99 SS |
2224 | if (obj_allocated <= obj_used) |
2225 | return 0; | |
04f05909 | 2226 | |
44f43e99 | 2227 | obj_wasted = obj_allocated - obj_used; |
b4fd07a0 | 2228 | obj_wasted /= class->objs_per_zspage; |
04f05909 | 2229 | |
6cbf16b3 | 2230 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
2231 | } |
2232 | ||
7d3f3938 | 2233 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 2234 | { |
312fcae2 | 2235 | struct zs_compact_control cc; |
3783689a MK |
2236 | struct zspage *src_zspage; |
2237 | struct zspage *dst_zspage = NULL; | |
312fcae2 | 2238 | |
312fcae2 | 2239 | spin_lock(&class->lock); |
3783689a | 2240 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 2241 | |
04f05909 SS |
2242 | if (!zs_can_compact(class)) |
2243 | break; | |
2244 | ||
41b88e14 | 2245 | cc.obj_idx = 0; |
48b4800a | 2246 | cc.s_page = get_first_page(src_zspage); |
312fcae2 | 2247 | |
3783689a | 2248 | while ((dst_zspage = isolate_zspage(class, false))) { |
48b4800a | 2249 | cc.d_page = get_first_page(dst_zspage); |
312fcae2 | 2250 | /* |
0dc63d48 SS |
2251 | * If there is no more space in dst_page, resched |
2252 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
2253 | */ |
2254 | if (!migrate_zspage(pool, class, &cc)) | |
2255 | break; | |
2256 | ||
4aa409ca | 2257 | putback_zspage(class, dst_zspage); |
312fcae2 MK |
2258 | } |
2259 | ||
2260 | /* Stop if we couldn't find slot */ | |
3783689a | 2261 | if (dst_zspage == NULL) |
312fcae2 MK |
2262 | break; |
2263 | ||
4aa409ca MK |
2264 | putback_zspage(class, dst_zspage); |
2265 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { | |
48b4800a | 2266 | free_zspage(pool, class, src_zspage); |
6cbf16b3 | 2267 | pool->stats.pages_compacted += class->pages_per_zspage; |
4aa409ca | 2268 | } |
312fcae2 | 2269 | spin_unlock(&class->lock); |
312fcae2 MK |
2270 | cond_resched(); |
2271 | spin_lock(&class->lock); | |
2272 | } | |
2273 | ||
3783689a | 2274 | if (src_zspage) |
4aa409ca | 2275 | putback_zspage(class, src_zspage); |
312fcae2 | 2276 | |
7d3f3938 | 2277 | spin_unlock(&class->lock); |
312fcae2 MK |
2278 | } |
2279 | ||
2280 | unsigned long zs_compact(struct zs_pool *pool) | |
2281 | { | |
2282 | int i; | |
312fcae2 MK |
2283 | struct size_class *class; |
2284 | ||
cf8e0fed | 2285 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
312fcae2 MK |
2286 | class = pool->size_class[i]; |
2287 | if (!class) | |
2288 | continue; | |
2289 | if (class->index != i) | |
2290 | continue; | |
7d3f3938 | 2291 | __zs_compact(pool, class); |
312fcae2 MK |
2292 | } |
2293 | ||
860c707d | 2294 | return pool->stats.pages_compacted; |
312fcae2 MK |
2295 | } |
2296 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 2297 | |
7d3f3938 SS |
2298 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
2299 | { | |
2300 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
2301 | } | |
2302 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
2303 | ||
ab9d306d SS |
2304 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
2305 | struct shrink_control *sc) | |
2306 | { | |
2307 | unsigned long pages_freed; | |
2308 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2309 | shrinker); | |
2310 | ||
2311 | pages_freed = pool->stats.pages_compacted; | |
2312 | /* | |
2313 | * Compact classes and calculate compaction delta. | |
2314 | * Can run concurrently with a manually triggered | |
2315 | * (by user) compaction. | |
2316 | */ | |
2317 | pages_freed = zs_compact(pool) - pages_freed; | |
2318 | ||
2319 | return pages_freed ? pages_freed : SHRINK_STOP; | |
2320 | } | |
2321 | ||
2322 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
2323 | struct shrink_control *sc) | |
2324 | { | |
2325 | int i; | |
2326 | struct size_class *class; | |
2327 | unsigned long pages_to_free = 0; | |
2328 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2329 | shrinker); | |
2330 | ||
cf8e0fed | 2331 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
ab9d306d SS |
2332 | class = pool->size_class[i]; |
2333 | if (!class) | |
2334 | continue; | |
2335 | if (class->index != i) | |
2336 | continue; | |
2337 | ||
ab9d306d | 2338 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
2339 | } |
2340 | ||
2341 | return pages_to_free; | |
2342 | } | |
2343 | ||
2344 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
2345 | { | |
2346 | if (pool->shrinker_enabled) { | |
2347 | unregister_shrinker(&pool->shrinker); | |
2348 | pool->shrinker_enabled = false; | |
2349 | } | |
2350 | } | |
2351 | ||
2352 | static int zs_register_shrinker(struct zs_pool *pool) | |
2353 | { | |
2354 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
2355 | pool->shrinker.count_objects = zs_shrinker_count; | |
2356 | pool->shrinker.batch = 0; | |
2357 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
2358 | ||
2359 | return register_shrinker(&pool->shrinker); | |
2360 | } | |
2361 | ||
00a61d86 | 2362 | /** |
66cdef66 | 2363 | * zs_create_pool - Creates an allocation pool to work from. |
fd854463 | 2364 | * @name: pool name to be created |
166cfda7 | 2365 | * |
66cdef66 GM |
2366 | * This function must be called before anything when using |
2367 | * the zsmalloc allocator. | |
166cfda7 | 2368 | * |
66cdef66 GM |
2369 | * On success, a pointer to the newly created pool is returned, |
2370 | * otherwise NULL. | |
396b7fd6 | 2371 | */ |
d0d8da2d | 2372 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 2373 | { |
66cdef66 GM |
2374 | int i; |
2375 | struct zs_pool *pool; | |
2376 | struct size_class *prev_class = NULL; | |
61989a80 | 2377 | |
66cdef66 GM |
2378 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
2379 | if (!pool) | |
2380 | return NULL; | |
61989a80 | 2381 | |
48b4800a | 2382 | init_deferred_free(pool); |
61989a80 | 2383 | |
2e40e163 MK |
2384 | pool->name = kstrdup(name, GFP_KERNEL); |
2385 | if (!pool->name) | |
2386 | goto err; | |
2387 | ||
3783689a | 2388 | if (create_cache(pool)) |
2e40e163 MK |
2389 | goto err; |
2390 | ||
c60369f0 | 2391 | /* |
399d8eeb | 2392 | * Iterate reversely, because, size of size_class that we want to use |
66cdef66 | 2393 | * for merging should be larger or equal to current size. |
c60369f0 | 2394 | */ |
cf8e0fed | 2395 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
66cdef66 GM |
2396 | int size; |
2397 | int pages_per_zspage; | |
64d90465 | 2398 | int objs_per_zspage; |
66cdef66 | 2399 | struct size_class *class; |
3783689a | 2400 | int fullness = 0; |
c60369f0 | 2401 | |
66cdef66 GM |
2402 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
2403 | if (size > ZS_MAX_ALLOC_SIZE) | |
2404 | size = ZS_MAX_ALLOC_SIZE; | |
2405 | pages_per_zspage = get_pages_per_zspage(size); | |
64d90465 | 2406 | objs_per_zspage = pages_per_zspage * PAGE_SIZE / size; |
61989a80 | 2407 | |
7f47f23b SS |
2408 | /* |
2409 | * We iterate from biggest down to smallest classes, | |
2410 | * so huge_class_size holds the size of the first huge | |
2411 | * class. Any object bigger than or equal to that will | |
2412 | * endup in the huge class. | |
2413 | */ | |
2414 | if (pages_per_zspage != 1 && objs_per_zspage != 1 && | |
2415 | !huge_class_size) { | |
2416 | huge_class_size = size; | |
2417 | /* | |
2418 | * The object uses ZS_HANDLE_SIZE bytes to store the | |
2419 | * handle. We need to subtract it, because zs_malloc() | |
2420 | * unconditionally adds handle size before it performs | |
2421 | * size class search - so object may be smaller than | |
2422 | * huge class size, yet it still can end up in the huge | |
2423 | * class because it grows by ZS_HANDLE_SIZE extra bytes | |
2424 | * right before class lookup. | |
2425 | */ | |
2426 | huge_class_size -= (ZS_HANDLE_SIZE - 1); | |
2427 | } | |
2428 | ||
66cdef66 GM |
2429 | /* |
2430 | * size_class is used for normal zsmalloc operation such | |
2431 | * as alloc/free for that size. Although it is natural that we | |
2432 | * have one size_class for each size, there is a chance that we | |
2433 | * can get more memory utilization if we use one size_class for | |
2434 | * many different sizes whose size_class have same | |
2435 | * characteristics. So, we makes size_class point to | |
2436 | * previous size_class if possible. | |
2437 | */ | |
2438 | if (prev_class) { | |
64d90465 | 2439 | if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) { |
66cdef66 GM |
2440 | pool->size_class[i] = prev_class; |
2441 | continue; | |
2442 | } | |
2443 | } | |
2444 | ||
2445 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
2446 | if (!class) | |
2447 | goto err; | |
2448 | ||
2449 | class->size = size; | |
2450 | class->index = i; | |
2451 | class->pages_per_zspage = pages_per_zspage; | |
64d90465 | 2452 | class->objs_per_zspage = objs_per_zspage; |
66cdef66 GM |
2453 | spin_lock_init(&class->lock); |
2454 | pool->size_class[i] = class; | |
48b4800a MK |
2455 | for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS; |
2456 | fullness++) | |
3783689a | 2457 | INIT_LIST_HEAD(&class->fullness_list[fullness]); |
66cdef66 GM |
2458 | |
2459 | prev_class = class; | |
61989a80 NG |
2460 | } |
2461 | ||
d34f6157 DS |
2462 | /* debug only, don't abort if it fails */ |
2463 | zs_pool_stat_create(pool, name); | |
0f050d99 | 2464 | |
48b4800a MK |
2465 | if (zs_register_migration(pool)) |
2466 | goto err; | |
2467 | ||
ab9d306d SS |
2468 | /* |
2469 | * Not critical, we still can use the pool | |
2470 | * and user can trigger compaction manually. | |
2471 | */ | |
2472 | if (zs_register_shrinker(pool) == 0) | |
2473 | pool->shrinker_enabled = true; | |
66cdef66 GM |
2474 | return pool; |
2475 | ||
2476 | err: | |
2477 | zs_destroy_pool(pool); | |
2478 | return NULL; | |
61989a80 | 2479 | } |
66cdef66 | 2480 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2481 | |
66cdef66 | 2482 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2483 | { |
66cdef66 | 2484 | int i; |
61989a80 | 2485 | |
ab9d306d | 2486 | zs_unregister_shrinker(pool); |
48b4800a | 2487 | zs_unregister_migration(pool); |
0f050d99 GM |
2488 | zs_pool_stat_destroy(pool); |
2489 | ||
cf8e0fed | 2490 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
66cdef66 GM |
2491 | int fg; |
2492 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2493 | |
66cdef66 GM |
2494 | if (!class) |
2495 | continue; | |
61989a80 | 2496 | |
66cdef66 GM |
2497 | if (class->index != i) |
2498 | continue; | |
61989a80 | 2499 | |
48b4800a | 2500 | for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) { |
3783689a | 2501 | if (!list_empty(&class->fullness_list[fg])) { |
66cdef66 GM |
2502 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
2503 | class->size, fg); | |
2504 | } | |
2505 | } | |
2506 | kfree(class); | |
2507 | } | |
f553646a | 2508 | |
3783689a | 2509 | destroy_cache(pool); |
0f050d99 | 2510 | kfree(pool->name); |
66cdef66 GM |
2511 | kfree(pool); |
2512 | } | |
2513 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2514 | |
66cdef66 GM |
2515 | static int __init zs_init(void) |
2516 | { | |
48b4800a MK |
2517 | int ret; |
2518 | ||
2519 | ret = zsmalloc_mount(); | |
2520 | if (ret) | |
2521 | goto out; | |
2522 | ||
215c89d0 SAS |
2523 | ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare", |
2524 | zs_cpu_prepare, zs_cpu_dead); | |
0f050d99 | 2525 | if (ret) |
215c89d0 | 2526 | goto hp_setup_fail; |
66cdef66 | 2527 | |
66cdef66 GM |
2528 | #ifdef CONFIG_ZPOOL |
2529 | zpool_register_driver(&zs_zpool_driver); | |
2530 | #endif | |
0f050d99 | 2531 | |
4abaac9b DS |
2532 | zs_stat_init(); |
2533 | ||
66cdef66 | 2534 | return 0; |
0f050d99 | 2535 | |
215c89d0 | 2536 | hp_setup_fail: |
48b4800a MK |
2537 | zsmalloc_unmount(); |
2538 | out: | |
0f050d99 | 2539 | return ret; |
61989a80 | 2540 | } |
61989a80 | 2541 | |
66cdef66 | 2542 | static void __exit zs_exit(void) |
61989a80 | 2543 | { |
66cdef66 GM |
2544 | #ifdef CONFIG_ZPOOL |
2545 | zpool_unregister_driver(&zs_zpool_driver); | |
2546 | #endif | |
48b4800a | 2547 | zsmalloc_unmount(); |
215c89d0 | 2548 | cpuhp_remove_state(CPUHP_MM_ZS_PREPARE); |
0f050d99 GM |
2549 | |
2550 | zs_stat_exit(); | |
61989a80 | 2551 | } |
069f101f BH |
2552 | |
2553 | module_init(zs_init); | |
2554 | module_exit(zs_exit); | |
2555 | ||
2556 | MODULE_LICENSE("Dual BSD/GPL"); | |
2557 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |