]>
Commit | Line | Data |
---|---|---|
61989a80 NG |
1 | /* |
2 | * zsmalloc memory allocator | |
3 | * | |
4 | * Copyright (C) 2011 Nitin Gupta | |
31fc00bb | 5 | * Copyright (C) 2012, 2013 Minchan Kim |
61989a80 NG |
6 | * |
7 | * This code is released using a dual license strategy: BSD/GPL | |
8 | * You can choose the license that better fits your requirements. | |
9 | * | |
10 | * Released under the terms of 3-clause BSD License | |
11 | * Released under the terms of GNU General Public License Version 2.0 | |
12 | */ | |
13 | ||
2db51dae | 14 | /* |
2db51dae NG |
15 | * Following is how we use various fields and flags of underlying |
16 | * struct page(s) to form a zspage. | |
17 | * | |
18 | * Usage of struct page fields: | |
3783689a | 19 | * page->private: points to zspage |
48b4800a MK |
20 | * page->freelist(index): links together all component pages of a zspage |
21 | * For the huge page, this is always 0, so we use this field | |
22 | * to store handle. | |
fd854463 | 23 | * page->units: first object offset in a subpage of zspage |
2db51dae NG |
24 | * |
25 | * Usage of struct page flags: | |
26 | * PG_private: identifies the first component page | |
399d8eeb | 27 | * PG_owner_priv_1: identifies the huge component page |
2db51dae NG |
28 | * |
29 | */ | |
30 | ||
4abaac9b DS |
31 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
32 | ||
61989a80 NG |
33 | #include <linux/module.h> |
34 | #include <linux/kernel.h> | |
312fcae2 | 35 | #include <linux/sched.h> |
50d34394 | 36 | #include <linux/magic.h> |
61989a80 NG |
37 | #include <linux/bitops.h> |
38 | #include <linux/errno.h> | |
39 | #include <linux/highmem.h> | |
61989a80 NG |
40 | #include <linux/string.h> |
41 | #include <linux/slab.h> | |
42 | #include <asm/tlbflush.h> | |
43 | #include <asm/pgtable.h> | |
44 | #include <linux/cpumask.h> | |
45 | #include <linux/cpu.h> | |
0cbb613f | 46 | #include <linux/vmalloc.h> |
759b26b2 | 47 | #include <linux/preempt.h> |
0959c63f | 48 | #include <linux/spinlock.h> |
93144ca3 | 49 | #include <linux/shrinker.h> |
0959c63f | 50 | #include <linux/types.h> |
0f050d99 | 51 | #include <linux/debugfs.h> |
bcf1647d | 52 | #include <linux/zsmalloc.h> |
c795779d | 53 | #include <linux/zpool.h> |
48b4800a | 54 | #include <linux/mount.h> |
dd4123f3 | 55 | #include <linux/migrate.h> |
48b4800a | 56 | #include <linux/pagemap.h> |
cdc346b3 | 57 | #include <linux/fs.h> |
48b4800a MK |
58 | |
59 | #define ZSPAGE_MAGIC 0x58 | |
0959c63f SJ |
60 | |
61 | /* | |
62 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
63 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
64 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
65 | * to restore link_free pointer values. | |
66 | */ | |
67 | #define ZS_ALIGN 8 | |
68 | ||
69 | /* | |
70 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
71 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
72 | */ | |
73 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
74 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
75 | ||
2e40e163 MK |
76 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
77 | ||
0959c63f SJ |
78 | /* |
79 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 80 | * as single (unsigned long) handle value. |
0959c63f | 81 | * |
bfd093f5 | 82 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
83 | * |
84 | * This is made more complicated by various memory models and PAE. | |
85 | */ | |
86 | ||
02390b87 KS |
87 | #ifndef MAX_POSSIBLE_PHYSMEM_BITS |
88 | #ifdef MAX_PHYSMEM_BITS | |
89 | #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS | |
90 | #else | |
0959c63f SJ |
91 | /* |
92 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
93 | * be PAGE_SHIFT | |
94 | */ | |
02390b87 | 95 | #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG |
0959c63f SJ |
96 | #endif |
97 | #endif | |
02390b87 KS |
98 | |
99 | #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
100 | |
101 | /* | |
102 | * Memory for allocating for handle keeps object position by | |
103 | * encoding <page, obj_idx> and the encoded value has a room | |
104 | * in least bit(ie, look at obj_to_location). | |
105 | * We use the bit to synchronize between object access by | |
106 | * user and migration. | |
107 | */ | |
108 | #define HANDLE_PIN_BIT 0 | |
109 | ||
110 | /* | |
111 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
112 | * to identify the object was allocated or not. | |
113 | * It's okay to add the status bit in the least bit because | |
114 | * header keeps handle which is 4byte-aligned address so we | |
115 | * have room for two bit at least. | |
116 | */ | |
117 | #define OBJ_ALLOCATED_TAG 1 | |
118 | #define OBJ_TAG_BITS 1 | |
119 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
120 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
121 | ||
cf8e0fed JM |
122 | #define FULLNESS_BITS 2 |
123 | #define CLASS_BITS 8 | |
124 | #define ISOLATED_BITS 3 | |
125 | #define MAGIC_VAL_BITS 8 | |
126 | ||
0959c63f SJ |
127 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) |
128 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
129 | #define ZS_MIN_ALLOC_SIZE \ | |
130 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 131 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 132 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
133 | |
134 | /* | |
7eb52512 | 135 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
136 | * trader-off here: |
137 | * - Large number of size classes is potentially wasteful as free page are | |
138 | * spread across these classes | |
139 | * - Small number of size classes causes large internal fragmentation | |
140 | * - Probably its better to use specific size classes (empirically | |
141 | * determined). NOTE: all those class sizes must be set as multiple of | |
142 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
143 | * | |
144 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
145 | * (reason above) | |
146 | */ | |
3783689a | 147 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
cf8e0fed JM |
148 | #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \ |
149 | ZS_SIZE_CLASS_DELTA) + 1) | |
0959c63f | 150 | |
0959c63f | 151 | enum fullness_group { |
0959c63f | 152 | ZS_EMPTY, |
48b4800a MK |
153 | ZS_ALMOST_EMPTY, |
154 | ZS_ALMOST_FULL, | |
155 | ZS_FULL, | |
156 | NR_ZS_FULLNESS, | |
0959c63f SJ |
157 | }; |
158 | ||
0f050d99 | 159 | enum zs_stat_type { |
48b4800a MK |
160 | CLASS_EMPTY, |
161 | CLASS_ALMOST_EMPTY, | |
162 | CLASS_ALMOST_FULL, | |
163 | CLASS_FULL, | |
0f050d99 GM |
164 | OBJ_ALLOCATED, |
165 | OBJ_USED, | |
48b4800a | 166 | NR_ZS_STAT_TYPE, |
0f050d99 GM |
167 | }; |
168 | ||
0f050d99 GM |
169 | struct zs_size_stat { |
170 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
171 | }; | |
172 | ||
57244594 SS |
173 | #ifdef CONFIG_ZSMALLOC_STAT |
174 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
175 | #endif |
176 | ||
48b4800a MK |
177 | #ifdef CONFIG_COMPACTION |
178 | static struct vfsmount *zsmalloc_mnt; | |
179 | #endif | |
180 | ||
0959c63f SJ |
181 | /* |
182 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
183 | * n <= N / f, where | |
184 | * n = number of allocated objects | |
185 | * N = total number of objects zspage can store | |
6dd9737e | 186 | * f = fullness_threshold_frac |
0959c63f SJ |
187 | * |
188 | * Similarly, we assign zspage to: | |
189 | * ZS_ALMOST_FULL when n > N / f | |
190 | * ZS_EMPTY when n == 0 | |
191 | * ZS_FULL when n == N | |
192 | * | |
193 | * (see: fix_fullness_group()) | |
194 | */ | |
195 | static const int fullness_threshold_frac = 4; | |
196 | ||
197 | struct size_class { | |
57244594 | 198 | spinlock_t lock; |
48b4800a | 199 | struct list_head fullness_list[NR_ZS_FULLNESS]; |
0959c63f SJ |
200 | /* |
201 | * Size of objects stored in this class. Must be multiple | |
202 | * of ZS_ALIGN. | |
203 | */ | |
204 | int size; | |
1fc6e27d | 205 | int objs_per_zspage; |
7dfa4612 WY |
206 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
207 | int pages_per_zspage; | |
48b4800a MK |
208 | |
209 | unsigned int index; | |
210 | struct zs_size_stat stats; | |
0959c63f SJ |
211 | }; |
212 | ||
48b4800a MK |
213 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
214 | static void SetPageHugeObject(struct page *page) | |
215 | { | |
216 | SetPageOwnerPriv1(page); | |
217 | } | |
218 | ||
219 | static void ClearPageHugeObject(struct page *page) | |
220 | { | |
221 | ClearPageOwnerPriv1(page); | |
222 | } | |
223 | ||
224 | static int PageHugeObject(struct page *page) | |
225 | { | |
226 | return PageOwnerPriv1(page); | |
227 | } | |
228 | ||
0959c63f SJ |
229 | /* |
230 | * Placed within free objects to form a singly linked list. | |
3783689a | 231 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
232 | * |
233 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
234 | */ | |
235 | struct link_free { | |
2e40e163 MK |
236 | union { |
237 | /* | |
bfd093f5 | 238 | * Free object index; |
2e40e163 MK |
239 | * It's valid for non-allocated object |
240 | */ | |
bfd093f5 | 241 | unsigned long next; |
2e40e163 MK |
242 | /* |
243 | * Handle of allocated object. | |
244 | */ | |
245 | unsigned long handle; | |
246 | }; | |
0959c63f SJ |
247 | }; |
248 | ||
249 | struct zs_pool { | |
6f3526d6 | 250 | const char *name; |
0f050d99 | 251 | |
cf8e0fed | 252 | struct size_class *size_class[ZS_SIZE_CLASSES]; |
2e40e163 | 253 | struct kmem_cache *handle_cachep; |
3783689a | 254 | struct kmem_cache *zspage_cachep; |
0959c63f | 255 | |
13de8933 | 256 | atomic_long_t pages_allocated; |
0f050d99 | 257 | |
7d3f3938 | 258 | struct zs_pool_stats stats; |
ab9d306d SS |
259 | |
260 | /* Compact classes */ | |
261 | struct shrinker shrinker; | |
93144ca3 | 262 | |
0f050d99 GM |
263 | #ifdef CONFIG_ZSMALLOC_STAT |
264 | struct dentry *stat_dentry; | |
265 | #endif | |
48b4800a MK |
266 | #ifdef CONFIG_COMPACTION |
267 | struct inode *inode; | |
268 | struct work_struct free_work; | |
269 | #endif | |
0959c63f | 270 | }; |
61989a80 | 271 | |
3783689a MK |
272 | struct zspage { |
273 | struct { | |
274 | unsigned int fullness:FULLNESS_BITS; | |
85d492f2 | 275 | unsigned int class:CLASS_BITS + 1; |
48b4800a MK |
276 | unsigned int isolated:ISOLATED_BITS; |
277 | unsigned int magic:MAGIC_VAL_BITS; | |
3783689a MK |
278 | }; |
279 | unsigned int inuse; | |
bfd093f5 | 280 | unsigned int freeobj; |
3783689a MK |
281 | struct page *first_page; |
282 | struct list_head list; /* fullness list */ | |
48b4800a MK |
283 | #ifdef CONFIG_COMPACTION |
284 | rwlock_t lock; | |
285 | #endif | |
3783689a | 286 | }; |
61989a80 | 287 | |
f553646a | 288 | struct mapping_area { |
1b945aee | 289 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
290 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
291 | #else | |
292 | char *vm_buf; /* copy buffer for objects that span pages */ | |
293 | #endif | |
294 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
295 | enum zs_mapmode vm_mm; /* mapping mode */ | |
296 | }; | |
297 | ||
48b4800a MK |
298 | #ifdef CONFIG_COMPACTION |
299 | static int zs_register_migration(struct zs_pool *pool); | |
300 | static void zs_unregister_migration(struct zs_pool *pool); | |
301 | static void migrate_lock_init(struct zspage *zspage); | |
302 | static void migrate_read_lock(struct zspage *zspage); | |
303 | static void migrate_read_unlock(struct zspage *zspage); | |
304 | static void kick_deferred_free(struct zs_pool *pool); | |
305 | static void init_deferred_free(struct zs_pool *pool); | |
306 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage); | |
307 | #else | |
308 | static int zsmalloc_mount(void) { return 0; } | |
309 | static void zsmalloc_unmount(void) {} | |
310 | static int zs_register_migration(struct zs_pool *pool) { return 0; } | |
311 | static void zs_unregister_migration(struct zs_pool *pool) {} | |
312 | static void migrate_lock_init(struct zspage *zspage) {} | |
313 | static void migrate_read_lock(struct zspage *zspage) {} | |
314 | static void migrate_read_unlock(struct zspage *zspage) {} | |
315 | static void kick_deferred_free(struct zs_pool *pool) {} | |
316 | static void init_deferred_free(struct zs_pool *pool) {} | |
317 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {} | |
318 | #endif | |
319 | ||
3783689a | 320 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
321 | { |
322 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
323 | 0, 0, NULL); | |
3783689a MK |
324 | if (!pool->handle_cachep) |
325 | return 1; | |
326 | ||
327 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
328 | 0, 0, NULL); | |
329 | if (!pool->zspage_cachep) { | |
330 | kmem_cache_destroy(pool->handle_cachep); | |
331 | pool->handle_cachep = NULL; | |
332 | return 1; | |
333 | } | |
334 | ||
335 | return 0; | |
2e40e163 MK |
336 | } |
337 | ||
3783689a | 338 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 339 | { |
cd10add0 | 340 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 341 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
342 | } |
343 | ||
3783689a | 344 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
345 | { |
346 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
48b4800a | 347 | gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
2e40e163 MK |
348 | } |
349 | ||
3783689a | 350 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
351 | { |
352 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
353 | } | |
354 | ||
3783689a MK |
355 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
356 | { | |
48b4800a MK |
357 | return kmem_cache_alloc(pool->zspage_cachep, |
358 | flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); | |
399d8eeb | 359 | } |
3783689a MK |
360 | |
361 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
362 | { | |
363 | kmem_cache_free(pool->zspage_cachep, zspage); | |
364 | } | |
365 | ||
2e40e163 MK |
366 | static void record_obj(unsigned long handle, unsigned long obj) |
367 | { | |
c102f07c JL |
368 | /* |
369 | * lsb of @obj represents handle lock while other bits | |
370 | * represent object value the handle is pointing so | |
371 | * updating shouldn't do store tearing. | |
372 | */ | |
373 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
374 | } |
375 | ||
c795779d DS |
376 | /* zpool driver */ |
377 | ||
378 | #ifdef CONFIG_ZPOOL | |
379 | ||
6f3526d6 | 380 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 381 | const struct zpool_ops *zpool_ops, |
479305fd | 382 | struct zpool *zpool) |
c795779d | 383 | { |
d0d8da2d SS |
384 | /* |
385 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
386 | * different contexts and its caller must provide a valid | |
387 | * gfp mask. | |
388 | */ | |
389 | return zs_create_pool(name); | |
c795779d DS |
390 | } |
391 | ||
392 | static void zs_zpool_destroy(void *pool) | |
393 | { | |
394 | zs_destroy_pool(pool); | |
395 | } | |
396 | ||
397 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
398 | unsigned long *handle) | |
399 | { | |
d0d8da2d | 400 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
401 | return *handle ? 0 : -1; |
402 | } | |
403 | static void zs_zpool_free(void *pool, unsigned long handle) | |
404 | { | |
405 | zs_free(pool, handle); | |
406 | } | |
407 | ||
c795779d DS |
408 | static void *zs_zpool_map(void *pool, unsigned long handle, |
409 | enum zpool_mapmode mm) | |
410 | { | |
411 | enum zs_mapmode zs_mm; | |
412 | ||
413 | switch (mm) { | |
414 | case ZPOOL_MM_RO: | |
415 | zs_mm = ZS_MM_RO; | |
416 | break; | |
417 | case ZPOOL_MM_WO: | |
418 | zs_mm = ZS_MM_WO; | |
419 | break; | |
420 | case ZPOOL_MM_RW: /* fallthru */ | |
421 | default: | |
422 | zs_mm = ZS_MM_RW; | |
423 | break; | |
424 | } | |
425 | ||
426 | return zs_map_object(pool, handle, zs_mm); | |
427 | } | |
428 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
429 | { | |
430 | zs_unmap_object(pool, handle); | |
431 | } | |
432 | ||
433 | static u64 zs_zpool_total_size(void *pool) | |
434 | { | |
722cdc17 | 435 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
436 | } |
437 | ||
438 | static struct zpool_driver zs_zpool_driver = { | |
439 | .type = "zsmalloc", | |
440 | .owner = THIS_MODULE, | |
441 | .create = zs_zpool_create, | |
442 | .destroy = zs_zpool_destroy, | |
443 | .malloc = zs_zpool_malloc, | |
444 | .free = zs_zpool_free, | |
c795779d DS |
445 | .map = zs_zpool_map, |
446 | .unmap = zs_zpool_unmap, | |
447 | .total_size = zs_zpool_total_size, | |
448 | }; | |
449 | ||
137f8cff | 450 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
451 | #endif /* CONFIG_ZPOOL */ |
452 | ||
61989a80 NG |
453 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
454 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
455 | ||
48b4800a MK |
456 | static bool is_zspage_isolated(struct zspage *zspage) |
457 | { | |
458 | return zspage->isolated; | |
459 | } | |
460 | ||
3457f414 | 461 | static __maybe_unused int is_first_page(struct page *page) |
61989a80 | 462 | { |
a27545bf | 463 | return PagePrivate(page); |
61989a80 NG |
464 | } |
465 | ||
48b4800a | 466 | /* Protected by class->lock */ |
3783689a | 467 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 468 | { |
3783689a | 469 | return zspage->inuse; |
4f42047b MK |
470 | } |
471 | ||
3783689a | 472 | static inline void set_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 473 | { |
3783689a | 474 | zspage->inuse = val; |
4f42047b MK |
475 | } |
476 | ||
3783689a | 477 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 478 | { |
3783689a | 479 | zspage->inuse += val; |
4f42047b MK |
480 | } |
481 | ||
48b4800a | 482 | static inline struct page *get_first_page(struct zspage *zspage) |
4f42047b | 483 | { |
48b4800a | 484 | struct page *first_page = zspage->first_page; |
3783689a | 485 | |
48b4800a MK |
486 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
487 | return first_page; | |
4f42047b MK |
488 | } |
489 | ||
48b4800a | 490 | static inline int get_first_obj_offset(struct page *page) |
4f42047b | 491 | { |
48b4800a MK |
492 | return page->units; |
493 | } | |
3783689a | 494 | |
48b4800a MK |
495 | static inline void set_first_obj_offset(struct page *page, int offset) |
496 | { | |
497 | page->units = offset; | |
4f42047b MK |
498 | } |
499 | ||
bfd093f5 | 500 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 501 | { |
bfd093f5 | 502 | return zspage->freeobj; |
4f42047b MK |
503 | } |
504 | ||
bfd093f5 | 505 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 506 | { |
bfd093f5 | 507 | zspage->freeobj = obj; |
4f42047b MK |
508 | } |
509 | ||
3783689a | 510 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 511 | unsigned int *class_idx, |
61989a80 NG |
512 | enum fullness_group *fullness) |
513 | { | |
48b4800a MK |
514 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); |
515 | ||
3783689a MK |
516 | *fullness = zspage->fullness; |
517 | *class_idx = zspage->class; | |
61989a80 NG |
518 | } |
519 | ||
3783689a | 520 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 521 | unsigned int class_idx, |
61989a80 NG |
522 | enum fullness_group fullness) |
523 | { | |
3783689a MK |
524 | zspage->class = class_idx; |
525 | zspage->fullness = fullness; | |
61989a80 NG |
526 | } |
527 | ||
c3e3e88a NC |
528 | /* |
529 | * zsmalloc divides the pool into various size classes where each | |
530 | * class maintains a list of zspages where each zspage is divided | |
531 | * into equal sized chunks. Each allocation falls into one of these | |
532 | * classes depending on its size. This function returns index of the | |
533 | * size class which has chunk size big enough to hold the give size. | |
534 | */ | |
61989a80 NG |
535 | static int get_size_class_index(int size) |
536 | { | |
537 | int idx = 0; | |
538 | ||
539 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
540 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
541 | ZS_SIZE_CLASS_DELTA); | |
542 | ||
cf8e0fed | 543 | return min_t(int, ZS_SIZE_CLASSES - 1, idx); |
61989a80 NG |
544 | } |
545 | ||
3eb95fea | 546 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 547 | static inline void zs_stat_inc(struct size_class *class, |
3eb95fea | 548 | int type, unsigned long cnt) |
248ca1b0 | 549 | { |
48b4800a | 550 | class->stats.objs[type] += cnt; |
248ca1b0 MK |
551 | } |
552 | ||
3eb95fea | 553 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 554 | static inline void zs_stat_dec(struct size_class *class, |
3eb95fea | 555 | int type, unsigned long cnt) |
248ca1b0 | 556 | { |
48b4800a | 557 | class->stats.objs[type] -= cnt; |
248ca1b0 MK |
558 | } |
559 | ||
3eb95fea | 560 | /* type can be of enum type zs_stat_type or fullness_group */ |
248ca1b0 | 561 | static inline unsigned long zs_stat_get(struct size_class *class, |
3eb95fea | 562 | int type) |
248ca1b0 | 563 | { |
48b4800a | 564 | return class->stats.objs[type]; |
248ca1b0 MK |
565 | } |
566 | ||
57244594 SS |
567 | #ifdef CONFIG_ZSMALLOC_STAT |
568 | ||
4abaac9b | 569 | static void __init zs_stat_init(void) |
248ca1b0 | 570 | { |
4abaac9b DS |
571 | if (!debugfs_initialized()) { |
572 | pr_warn("debugfs not available, stat dir not created\n"); | |
573 | return; | |
574 | } | |
248ca1b0 MK |
575 | |
576 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
577 | if (!zs_stat_root) | |
4abaac9b | 578 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
579 | } |
580 | ||
581 | static void __exit zs_stat_exit(void) | |
582 | { | |
583 | debugfs_remove_recursive(zs_stat_root); | |
584 | } | |
585 | ||
1120ed54 SS |
586 | static unsigned long zs_can_compact(struct size_class *class); |
587 | ||
248ca1b0 MK |
588 | static int zs_stats_size_show(struct seq_file *s, void *v) |
589 | { | |
590 | int i; | |
591 | struct zs_pool *pool = s->private; | |
592 | struct size_class *class; | |
593 | int objs_per_zspage; | |
594 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 595 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
596 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
597 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 598 | unsigned long total_freeable = 0; |
248ca1b0 | 599 | |
1120ed54 | 600 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
601 | "class", "size", "almost_full", "almost_empty", |
602 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 603 | "pages_per_zspage", "freeable"); |
248ca1b0 | 604 | |
cf8e0fed | 605 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
248ca1b0 MK |
606 | class = pool->size_class[i]; |
607 | ||
608 | if (class->index != i) | |
609 | continue; | |
610 | ||
611 | spin_lock(&class->lock); | |
612 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
613 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
614 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
615 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 616 | freeable = zs_can_compact(class); |
248ca1b0 MK |
617 | spin_unlock(&class->lock); |
618 | ||
b4fd07a0 | 619 | objs_per_zspage = class->objs_per_zspage; |
248ca1b0 MK |
620 | pages_used = obj_allocated / objs_per_zspage * |
621 | class->pages_per_zspage; | |
622 | ||
1120ed54 SS |
623 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
624 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
625 | i, class->size, class_almost_full, class_almost_empty, |
626 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 627 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
628 | |
629 | total_class_almost_full += class_almost_full; | |
630 | total_class_almost_empty += class_almost_empty; | |
631 | total_objs += obj_allocated; | |
632 | total_used_objs += obj_used; | |
633 | total_pages += pages_used; | |
1120ed54 | 634 | total_freeable += freeable; |
248ca1b0 MK |
635 | } |
636 | ||
637 | seq_puts(s, "\n"); | |
1120ed54 | 638 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
639 | "Total", "", total_class_almost_full, |
640 | total_class_almost_empty, total_objs, | |
1120ed54 | 641 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
642 | |
643 | return 0; | |
644 | } | |
5ad35093 | 645 | DEFINE_SHOW_ATTRIBUTE(zs_stats_size); |
248ca1b0 | 646 | |
d34f6157 | 647 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
648 | { |
649 | struct dentry *entry; | |
650 | ||
4abaac9b DS |
651 | if (!zs_stat_root) { |
652 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 653 | return; |
4abaac9b | 654 | } |
248ca1b0 MK |
655 | |
656 | entry = debugfs_create_dir(name, zs_stat_root); | |
657 | if (!entry) { | |
658 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 659 | return; |
248ca1b0 MK |
660 | } |
661 | pool->stat_dentry = entry; | |
662 | ||
663 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
5ad35093 | 664 | pool->stat_dentry, pool, &zs_stats_size_fops); |
248ca1b0 MK |
665 | if (!entry) { |
666 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
667 | name, "classes"); | |
4abaac9b DS |
668 | debugfs_remove_recursive(pool->stat_dentry); |
669 | pool->stat_dentry = NULL; | |
248ca1b0 | 670 | } |
248ca1b0 MK |
671 | } |
672 | ||
673 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
674 | { | |
675 | debugfs_remove_recursive(pool->stat_dentry); | |
676 | } | |
677 | ||
678 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 679 | static void __init zs_stat_init(void) |
248ca1b0 | 680 | { |
248ca1b0 MK |
681 | } |
682 | ||
683 | static void __exit zs_stat_exit(void) | |
684 | { | |
685 | } | |
686 | ||
d34f6157 | 687 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 688 | { |
248ca1b0 MK |
689 | } |
690 | ||
691 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
692 | { | |
693 | } | |
248ca1b0 MK |
694 | #endif |
695 | ||
48b4800a | 696 | |
c3e3e88a NC |
697 | /* |
698 | * For each size class, zspages are divided into different groups | |
699 | * depending on how "full" they are. This was done so that we could | |
700 | * easily find empty or nearly empty zspages when we try to shrink | |
701 | * the pool (not yet implemented). This function returns fullness | |
702 | * status of the given page. | |
703 | */ | |
1fc6e27d | 704 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 705 | struct zspage *zspage) |
61989a80 | 706 | { |
1fc6e27d | 707 | int inuse, objs_per_zspage; |
61989a80 | 708 | enum fullness_group fg; |
830e4bc5 | 709 | |
3783689a | 710 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 711 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
712 | |
713 | if (inuse == 0) | |
714 | fg = ZS_EMPTY; | |
1fc6e27d | 715 | else if (inuse == objs_per_zspage) |
61989a80 | 716 | fg = ZS_FULL; |
1fc6e27d | 717 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
718 | fg = ZS_ALMOST_EMPTY; |
719 | else | |
720 | fg = ZS_ALMOST_FULL; | |
721 | ||
722 | return fg; | |
723 | } | |
724 | ||
c3e3e88a NC |
725 | /* |
726 | * Each size class maintains various freelists and zspages are assigned | |
727 | * to one of these freelists based on the number of live objects they | |
728 | * have. This functions inserts the given zspage into the freelist | |
729 | * identified by <class, fullness_group>. | |
730 | */ | |
251cbb95 | 731 | static void insert_zspage(struct size_class *class, |
3783689a MK |
732 | struct zspage *zspage, |
733 | enum fullness_group fullness) | |
61989a80 | 734 | { |
3783689a | 735 | struct zspage *head; |
61989a80 | 736 | |
48b4800a | 737 | zs_stat_inc(class, fullness, 1); |
3783689a MK |
738 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
739 | struct zspage, list); | |
58f17117 | 740 | /* |
3783689a MK |
741 | * We want to see more ZS_FULL pages and less almost empty/full. |
742 | * Put pages with higher ->inuse first. | |
58f17117 | 743 | */ |
3783689a MK |
744 | if (head) { |
745 | if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) { | |
746 | list_add(&zspage->list, &head->list); | |
747 | return; | |
748 | } | |
749 | } | |
750 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
61989a80 NG |
751 | } |
752 | ||
c3e3e88a NC |
753 | /* |
754 | * This function removes the given zspage from the freelist identified | |
755 | * by <class, fullness_group>. | |
756 | */ | |
251cbb95 | 757 | static void remove_zspage(struct size_class *class, |
3783689a MK |
758 | struct zspage *zspage, |
759 | enum fullness_group fullness) | |
61989a80 | 760 | { |
3783689a | 761 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
48b4800a | 762 | VM_BUG_ON(is_zspage_isolated(zspage)); |
61989a80 | 763 | |
3783689a | 764 | list_del_init(&zspage->list); |
48b4800a | 765 | zs_stat_dec(class, fullness, 1); |
61989a80 NG |
766 | } |
767 | ||
c3e3e88a NC |
768 | /* |
769 | * Each size class maintains zspages in different fullness groups depending | |
770 | * on the number of live objects they contain. When allocating or freeing | |
771 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
772 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
773 | * a status change has occurred for the given page and accordingly moves the | |
774 | * page from the freelist of the old fullness group to that of the new | |
775 | * fullness group. | |
776 | */ | |
c7806261 | 777 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 778 | struct zspage *zspage) |
61989a80 NG |
779 | { |
780 | int class_idx; | |
61989a80 NG |
781 | enum fullness_group currfg, newfg; |
782 | ||
3783689a MK |
783 | get_zspage_mapping(zspage, &class_idx, &currfg); |
784 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
785 | if (newfg == currfg) |
786 | goto out; | |
787 | ||
48b4800a MK |
788 | if (!is_zspage_isolated(zspage)) { |
789 | remove_zspage(class, zspage, currfg); | |
790 | insert_zspage(class, zspage, newfg); | |
791 | } | |
792 | ||
3783689a | 793 | set_zspage_mapping(zspage, class_idx, newfg); |
61989a80 NG |
794 | |
795 | out: | |
796 | return newfg; | |
797 | } | |
798 | ||
799 | /* | |
800 | * We have to decide on how many pages to link together | |
801 | * to form a zspage for each size class. This is important | |
802 | * to reduce wastage due to unusable space left at end of | |
803 | * each zspage which is given as: | |
888fa374 YX |
804 | * wastage = Zp % class_size |
805 | * usage = Zp - wastage | |
61989a80 NG |
806 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
807 | * | |
808 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
809 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
810 | * since then we can perfectly fit in 8 such objects. | |
811 | */ | |
2e3b6154 | 812 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
813 | { |
814 | int i, max_usedpc = 0; | |
815 | /* zspage order which gives maximum used size per KB */ | |
816 | int max_usedpc_order = 1; | |
817 | ||
84d4faab | 818 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
819 | int zspage_size; |
820 | int waste, usedpc; | |
821 | ||
822 | zspage_size = i * PAGE_SIZE; | |
823 | waste = zspage_size % class_size; | |
824 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
825 | ||
826 | if (usedpc > max_usedpc) { | |
827 | max_usedpc = usedpc; | |
828 | max_usedpc_order = i; | |
829 | } | |
830 | } | |
831 | ||
832 | return max_usedpc_order; | |
833 | } | |
834 | ||
3783689a | 835 | static struct zspage *get_zspage(struct page *page) |
61989a80 | 836 | { |
48b4800a MK |
837 | struct zspage *zspage = (struct zspage *)page->private; |
838 | ||
839 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); | |
840 | return zspage; | |
61989a80 NG |
841 | } |
842 | ||
843 | static struct page *get_next_page(struct page *page) | |
844 | { | |
48b4800a MK |
845 | if (unlikely(PageHugeObject(page))) |
846 | return NULL; | |
847 | ||
848 | return page->freelist; | |
61989a80 NG |
849 | } |
850 | ||
bfd093f5 MK |
851 | /** |
852 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
853 | * @page: page object resides in zspage | |
854 | * @obj_idx: object index | |
67296874 | 855 | */ |
bfd093f5 MK |
856 | static void obj_to_location(unsigned long obj, struct page **page, |
857 | unsigned int *obj_idx) | |
61989a80 | 858 | { |
bfd093f5 MK |
859 | obj >>= OBJ_TAG_BITS; |
860 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
861 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
862 | } | |
61989a80 | 863 | |
bfd093f5 MK |
864 | /** |
865 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
866 | * @page: page object resides in zspage | |
867 | * @obj_idx: object index | |
868 | */ | |
869 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
870 | { | |
871 | unsigned long obj; | |
61989a80 | 872 | |
312fcae2 | 873 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 874 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 875 | obj <<= OBJ_TAG_BITS; |
61989a80 | 876 | |
bfd093f5 | 877 | return obj; |
61989a80 NG |
878 | } |
879 | ||
2e40e163 MK |
880 | static unsigned long handle_to_obj(unsigned long handle) |
881 | { | |
882 | return *(unsigned long *)handle; | |
883 | } | |
884 | ||
48b4800a | 885 | static unsigned long obj_to_head(struct page *page, void *obj) |
312fcae2 | 886 | { |
48b4800a | 887 | if (unlikely(PageHugeObject(page))) { |
830e4bc5 | 888 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3783689a | 889 | return page->index; |
7b60a685 MK |
890 | } else |
891 | return *(unsigned long *)obj; | |
312fcae2 MK |
892 | } |
893 | ||
48b4800a MK |
894 | static inline int testpin_tag(unsigned long handle) |
895 | { | |
896 | return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle); | |
897 | } | |
898 | ||
312fcae2 MK |
899 | static inline int trypin_tag(unsigned long handle) |
900 | { | |
1b8320b6 | 901 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
902 | } |
903 | ||
904 | static void pin_tag(unsigned long handle) | |
905 | { | |
1b8320b6 | 906 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
907 | } |
908 | ||
909 | static void unpin_tag(unsigned long handle) | |
910 | { | |
1b8320b6 | 911 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
912 | } |
913 | ||
f4477e90 NG |
914 | static void reset_page(struct page *page) |
915 | { | |
48b4800a | 916 | __ClearPageMovable(page); |
18fd06bf | 917 | ClearPagePrivate(page); |
f4477e90 | 918 | set_page_private(page, 0); |
48b4800a MK |
919 | page_mapcount_reset(page); |
920 | ClearPageHugeObject(page); | |
921 | page->freelist = NULL; | |
922 | } | |
923 | ||
924 | /* | |
925 | * To prevent zspage destroy during migration, zspage freeing should | |
926 | * hold locks of all pages in the zspage. | |
927 | */ | |
928 | void lock_zspage(struct zspage *zspage) | |
929 | { | |
930 | struct page *page = get_first_page(zspage); | |
931 | ||
932 | do { | |
933 | lock_page(page); | |
934 | } while ((page = get_next_page(page)) != NULL); | |
935 | } | |
936 | ||
937 | int trylock_zspage(struct zspage *zspage) | |
938 | { | |
939 | struct page *cursor, *fail; | |
940 | ||
941 | for (cursor = get_first_page(zspage); cursor != NULL; cursor = | |
942 | get_next_page(cursor)) { | |
943 | if (!trylock_page(cursor)) { | |
944 | fail = cursor; | |
945 | goto unlock; | |
946 | } | |
947 | } | |
948 | ||
949 | return 1; | |
950 | unlock: | |
951 | for (cursor = get_first_page(zspage); cursor != fail; cursor = | |
952 | get_next_page(cursor)) | |
953 | unlock_page(cursor); | |
954 | ||
955 | return 0; | |
f4477e90 NG |
956 | } |
957 | ||
48b4800a MK |
958 | static void __free_zspage(struct zs_pool *pool, struct size_class *class, |
959 | struct zspage *zspage) | |
61989a80 | 960 | { |
3783689a | 961 | struct page *page, *next; |
48b4800a MK |
962 | enum fullness_group fg; |
963 | unsigned int class_idx; | |
964 | ||
965 | get_zspage_mapping(zspage, &class_idx, &fg); | |
966 | ||
967 | assert_spin_locked(&class->lock); | |
61989a80 | 968 | |
3783689a | 969 | VM_BUG_ON(get_zspage_inuse(zspage)); |
48b4800a | 970 | VM_BUG_ON(fg != ZS_EMPTY); |
61989a80 | 971 | |
48b4800a | 972 | next = page = get_first_page(zspage); |
3783689a | 973 | do { |
48b4800a MK |
974 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
975 | next = get_next_page(page); | |
3783689a | 976 | reset_page(page); |
48b4800a | 977 | unlock_page(page); |
91537fee | 978 | dec_zone_page_state(page, NR_ZSPAGES); |
3783689a MK |
979 | put_page(page); |
980 | page = next; | |
981 | } while (page != NULL); | |
61989a80 | 982 | |
3783689a | 983 | cache_free_zspage(pool, zspage); |
48b4800a | 984 | |
b4fd07a0 | 985 | zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
986 | atomic_long_sub(class->pages_per_zspage, |
987 | &pool->pages_allocated); | |
988 | } | |
989 | ||
990 | static void free_zspage(struct zs_pool *pool, struct size_class *class, | |
991 | struct zspage *zspage) | |
992 | { | |
993 | VM_BUG_ON(get_zspage_inuse(zspage)); | |
994 | VM_BUG_ON(list_empty(&zspage->list)); | |
995 | ||
996 | if (!trylock_zspage(zspage)) { | |
997 | kick_deferred_free(pool); | |
998 | return; | |
999 | } | |
1000 | ||
1001 | remove_zspage(class, zspage, ZS_EMPTY); | |
1002 | __free_zspage(pool, class, zspage); | |
61989a80 NG |
1003 | } |
1004 | ||
1005 | /* Initialize a newly allocated zspage */ | |
3783689a | 1006 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 1007 | { |
bfd093f5 | 1008 | unsigned int freeobj = 1; |
61989a80 | 1009 | unsigned long off = 0; |
48b4800a | 1010 | struct page *page = get_first_page(zspage); |
830e4bc5 | 1011 | |
61989a80 NG |
1012 | while (page) { |
1013 | struct page *next_page; | |
1014 | struct link_free *link; | |
af4ee5e9 | 1015 | void *vaddr; |
61989a80 | 1016 | |
3783689a | 1017 | set_first_obj_offset(page, off); |
61989a80 | 1018 | |
af4ee5e9 MK |
1019 | vaddr = kmap_atomic(page); |
1020 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
1021 | |
1022 | while ((off += class->size) < PAGE_SIZE) { | |
3b1d9ca6 | 1023 | link->next = freeobj++ << OBJ_TAG_BITS; |
5538c562 | 1024 | link += class->size / sizeof(*link); |
61989a80 NG |
1025 | } |
1026 | ||
1027 | /* | |
1028 | * We now come to the last (full or partial) object on this | |
1029 | * page, which must point to the first object on the next | |
1030 | * page (if present) | |
1031 | */ | |
1032 | next_page = get_next_page(page); | |
bfd093f5 | 1033 | if (next_page) { |
3b1d9ca6 | 1034 | link->next = freeobj++ << OBJ_TAG_BITS; |
bfd093f5 MK |
1035 | } else { |
1036 | /* | |
3b1d9ca6 | 1037 | * Reset OBJ_TAG_BITS bit to last link to tell |
bfd093f5 MK |
1038 | * whether it's allocated object or not. |
1039 | */ | |
01a6ad9a | 1040 | link->next = -1UL << OBJ_TAG_BITS; |
bfd093f5 | 1041 | } |
af4ee5e9 | 1042 | kunmap_atomic(vaddr); |
61989a80 | 1043 | page = next_page; |
5538c562 | 1044 | off %= PAGE_SIZE; |
61989a80 | 1045 | } |
bdb0af7c | 1046 | |
bfd093f5 | 1047 | set_freeobj(zspage, 0); |
61989a80 NG |
1048 | } |
1049 | ||
48b4800a MK |
1050 | static void create_page_chain(struct size_class *class, struct zspage *zspage, |
1051 | struct page *pages[]) | |
61989a80 | 1052 | { |
bdb0af7c MK |
1053 | int i; |
1054 | struct page *page; | |
1055 | struct page *prev_page = NULL; | |
48b4800a | 1056 | int nr_pages = class->pages_per_zspage; |
61989a80 NG |
1057 | |
1058 | /* | |
1059 | * Allocate individual pages and link them together as: | |
48b4800a | 1060 | * 1. all pages are linked together using page->freelist |
3783689a | 1061 | * 2. each sub-page point to zspage using page->private |
61989a80 | 1062 | * |
3783689a | 1063 | * we set PG_private to identify the first page (i.e. no other sub-page |
22c5cef1 | 1064 | * has this flag set). |
61989a80 | 1065 | */ |
bdb0af7c MK |
1066 | for (i = 0; i < nr_pages; i++) { |
1067 | page = pages[i]; | |
3783689a | 1068 | set_page_private(page, (unsigned long)zspage); |
48b4800a | 1069 | page->freelist = NULL; |
bdb0af7c | 1070 | if (i == 0) { |
3783689a | 1071 | zspage->first_page = page; |
a27545bf | 1072 | SetPagePrivate(page); |
48b4800a MK |
1073 | if (unlikely(class->objs_per_zspage == 1 && |
1074 | class->pages_per_zspage == 1)) | |
1075 | SetPageHugeObject(page); | |
3783689a | 1076 | } else { |
48b4800a | 1077 | prev_page->freelist = page; |
61989a80 | 1078 | } |
61989a80 NG |
1079 | prev_page = page; |
1080 | } | |
bdb0af7c | 1081 | } |
61989a80 | 1082 | |
bdb0af7c MK |
1083 | /* |
1084 | * Allocate a zspage for the given size class | |
1085 | */ | |
3783689a MK |
1086 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
1087 | struct size_class *class, | |
1088 | gfp_t gfp) | |
bdb0af7c MK |
1089 | { |
1090 | int i; | |
bdb0af7c | 1091 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1092 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1093 | ||
1094 | if (!zspage) | |
1095 | return NULL; | |
1096 | ||
1097 | memset(zspage, 0, sizeof(struct zspage)); | |
48b4800a MK |
1098 | zspage->magic = ZSPAGE_MAGIC; |
1099 | migrate_lock_init(zspage); | |
61989a80 | 1100 | |
bdb0af7c MK |
1101 | for (i = 0; i < class->pages_per_zspage; i++) { |
1102 | struct page *page; | |
61989a80 | 1103 | |
3783689a | 1104 | page = alloc_page(gfp); |
bdb0af7c | 1105 | if (!page) { |
91537fee MK |
1106 | while (--i >= 0) { |
1107 | dec_zone_page_state(pages[i], NR_ZSPAGES); | |
bdb0af7c | 1108 | __free_page(pages[i]); |
91537fee | 1109 | } |
3783689a | 1110 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1111 | return NULL; |
1112 | } | |
91537fee MK |
1113 | |
1114 | inc_zone_page_state(page, NR_ZSPAGES); | |
bdb0af7c | 1115 | pages[i] = page; |
61989a80 NG |
1116 | } |
1117 | ||
48b4800a | 1118 | create_page_chain(class, zspage, pages); |
3783689a | 1119 | init_zspage(class, zspage); |
bdb0af7c | 1120 | |
3783689a | 1121 | return zspage; |
61989a80 NG |
1122 | } |
1123 | ||
3783689a | 1124 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1125 | { |
1126 | int i; | |
3783689a | 1127 | struct zspage *zspage; |
61989a80 | 1128 | |
48b4800a | 1129 | for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) { |
3783689a MK |
1130 | zspage = list_first_entry_or_null(&class->fullness_list[i], |
1131 | struct zspage, list); | |
1132 | if (zspage) | |
61989a80 NG |
1133 | break; |
1134 | } | |
1135 | ||
3783689a | 1136 | return zspage; |
61989a80 NG |
1137 | } |
1138 | ||
1b945aee | 1139 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1140 | static inline int __zs_cpu_up(struct mapping_area *area) |
1141 | { | |
1142 | /* | |
1143 | * Make sure we don't leak memory if a cpu UP notification | |
1144 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1145 | */ | |
1146 | if (area->vm) | |
1147 | return 0; | |
1148 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1149 | if (!area->vm) | |
1150 | return -ENOMEM; | |
1151 | return 0; | |
1152 | } | |
1153 | ||
1154 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1155 | { | |
1156 | if (area->vm) | |
1157 | free_vm_area(area->vm); | |
1158 | area->vm = NULL; | |
1159 | } | |
1160 | ||
1161 | static inline void *__zs_map_object(struct mapping_area *area, | |
1162 | struct page *pages[2], int off, int size) | |
1163 | { | |
f6f8ed47 | 1164 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1165 | area->vm_addr = area->vm->addr; |
1166 | return area->vm_addr + off; | |
1167 | } | |
1168 | ||
1169 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1170 | struct page *pages[2], int off, int size) | |
1171 | { | |
1172 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1173 | |
d95abbbb | 1174 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1175 | } |
1176 | ||
1b945aee | 1177 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1178 | |
1179 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1180 | { | |
1181 | /* | |
1182 | * Make sure we don't leak memory if a cpu UP notification | |
1183 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1184 | */ | |
1185 | if (area->vm_buf) | |
1186 | return 0; | |
40f9fb8c | 1187 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1188 | if (!area->vm_buf) |
1189 | return -ENOMEM; | |
1190 | return 0; | |
1191 | } | |
1192 | ||
1193 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1194 | { | |
40f9fb8c | 1195 | kfree(area->vm_buf); |
f553646a SJ |
1196 | area->vm_buf = NULL; |
1197 | } | |
1198 | ||
1199 | static void *__zs_map_object(struct mapping_area *area, | |
1200 | struct page *pages[2], int off, int size) | |
5f601902 | 1201 | { |
5f601902 SJ |
1202 | int sizes[2]; |
1203 | void *addr; | |
f553646a | 1204 | char *buf = area->vm_buf; |
5f601902 | 1205 | |
f553646a SJ |
1206 | /* disable page faults to match kmap_atomic() return conditions */ |
1207 | pagefault_disable(); | |
1208 | ||
1209 | /* no read fastpath */ | |
1210 | if (area->vm_mm == ZS_MM_WO) | |
1211 | goto out; | |
5f601902 SJ |
1212 | |
1213 | sizes[0] = PAGE_SIZE - off; | |
1214 | sizes[1] = size - sizes[0]; | |
1215 | ||
5f601902 SJ |
1216 | /* copy object to per-cpu buffer */ |
1217 | addr = kmap_atomic(pages[0]); | |
1218 | memcpy(buf, addr + off, sizes[0]); | |
1219 | kunmap_atomic(addr); | |
1220 | addr = kmap_atomic(pages[1]); | |
1221 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1222 | kunmap_atomic(addr); | |
f553646a SJ |
1223 | out: |
1224 | return area->vm_buf; | |
5f601902 SJ |
1225 | } |
1226 | ||
f553646a SJ |
1227 | static void __zs_unmap_object(struct mapping_area *area, |
1228 | struct page *pages[2], int off, int size) | |
5f601902 | 1229 | { |
5f601902 SJ |
1230 | int sizes[2]; |
1231 | void *addr; | |
2e40e163 | 1232 | char *buf; |
5f601902 | 1233 | |
f553646a SJ |
1234 | /* no write fastpath */ |
1235 | if (area->vm_mm == ZS_MM_RO) | |
1236 | goto out; | |
5f601902 | 1237 | |
7b60a685 | 1238 | buf = area->vm_buf; |
a82cbf07 YX |
1239 | buf = buf + ZS_HANDLE_SIZE; |
1240 | size -= ZS_HANDLE_SIZE; | |
1241 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1242 | |
5f601902 SJ |
1243 | sizes[0] = PAGE_SIZE - off; |
1244 | sizes[1] = size - sizes[0]; | |
1245 | ||
1246 | /* copy per-cpu buffer to object */ | |
1247 | addr = kmap_atomic(pages[0]); | |
1248 | memcpy(addr + off, buf, sizes[0]); | |
1249 | kunmap_atomic(addr); | |
1250 | addr = kmap_atomic(pages[1]); | |
1251 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1252 | kunmap_atomic(addr); | |
f553646a SJ |
1253 | |
1254 | out: | |
1255 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1256 | pagefault_enable(); | |
5f601902 | 1257 | } |
61989a80 | 1258 | |
1b945aee | 1259 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1260 | |
215c89d0 | 1261 | static int zs_cpu_prepare(unsigned int cpu) |
61989a80 | 1262 | { |
61989a80 NG |
1263 | struct mapping_area *area; |
1264 | ||
215c89d0 SAS |
1265 | area = &per_cpu(zs_map_area, cpu); |
1266 | return __zs_cpu_up(area); | |
61989a80 NG |
1267 | } |
1268 | ||
215c89d0 | 1269 | static int zs_cpu_dead(unsigned int cpu) |
61989a80 | 1270 | { |
215c89d0 | 1271 | struct mapping_area *area; |
40f9fb8c | 1272 | |
215c89d0 SAS |
1273 | area = &per_cpu(zs_map_area, cpu); |
1274 | __zs_cpu_down(area); | |
1275 | return 0; | |
b1b00a5b SS |
1276 | } |
1277 | ||
64d90465 GM |
1278 | static bool can_merge(struct size_class *prev, int pages_per_zspage, |
1279 | int objs_per_zspage) | |
9eec4cd5 | 1280 | { |
64d90465 GM |
1281 | if (prev->pages_per_zspage == pages_per_zspage && |
1282 | prev->objs_per_zspage == objs_per_zspage) | |
1283 | return true; | |
9eec4cd5 | 1284 | |
64d90465 | 1285 | return false; |
9eec4cd5 JK |
1286 | } |
1287 | ||
3783689a | 1288 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1289 | { |
3783689a | 1290 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1291 | } |
1292 | ||
66cdef66 GM |
1293 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1294 | { | |
1295 | return atomic_long_read(&pool->pages_allocated); | |
1296 | } | |
1297 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1298 | ||
4bbc0bc0 | 1299 | /** |
66cdef66 GM |
1300 | * zs_map_object - get address of allocated object from handle. |
1301 | * @pool: pool from which the object was allocated | |
1302 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1303 | * |
66cdef66 GM |
1304 | * Before using an object allocated from zs_malloc, it must be mapped using |
1305 | * this function. When done with the object, it must be unmapped using | |
1306 | * zs_unmap_object. | |
4bbc0bc0 | 1307 | * |
66cdef66 GM |
1308 | * Only one object can be mapped per cpu at a time. There is no protection |
1309 | * against nested mappings. | |
1310 | * | |
1311 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1312 | */ |
66cdef66 GM |
1313 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1314 | enum zs_mapmode mm) | |
61989a80 | 1315 | { |
3783689a | 1316 | struct zspage *zspage; |
66cdef66 | 1317 | struct page *page; |
bfd093f5 MK |
1318 | unsigned long obj, off; |
1319 | unsigned int obj_idx; | |
61989a80 | 1320 | |
66cdef66 GM |
1321 | unsigned int class_idx; |
1322 | enum fullness_group fg; | |
1323 | struct size_class *class; | |
1324 | struct mapping_area *area; | |
1325 | struct page *pages[2]; | |
2e40e163 | 1326 | void *ret; |
61989a80 | 1327 | |
9eec4cd5 | 1328 | /* |
66cdef66 GM |
1329 | * Because we use per-cpu mapping areas shared among the |
1330 | * pools/users, we can't allow mapping in interrupt context | |
1331 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1332 | */ |
1aedcafb | 1333 | BUG_ON(in_interrupt()); |
61989a80 | 1334 | |
312fcae2 MK |
1335 | /* From now on, migration cannot move the object */ |
1336 | pin_tag(handle); | |
1337 | ||
2e40e163 MK |
1338 | obj = handle_to_obj(handle); |
1339 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1340 | zspage = get_zspage(page); |
48b4800a MK |
1341 | |
1342 | /* migration cannot move any subpage in this zspage */ | |
1343 | migrate_read_lock(zspage); | |
1344 | ||
3783689a | 1345 | get_zspage_mapping(zspage, &class_idx, &fg); |
66cdef66 | 1346 | class = pool->size_class[class_idx]; |
bfd093f5 | 1347 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1348 | |
66cdef66 GM |
1349 | area = &get_cpu_var(zs_map_area); |
1350 | area->vm_mm = mm; | |
1351 | if (off + class->size <= PAGE_SIZE) { | |
1352 | /* this object is contained entirely within a page */ | |
1353 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1354 | ret = area->vm_addr + off; |
1355 | goto out; | |
61989a80 NG |
1356 | } |
1357 | ||
66cdef66 GM |
1358 | /* this object spans two pages */ |
1359 | pages[0] = page; | |
1360 | pages[1] = get_next_page(page); | |
1361 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1362 | |
2e40e163 MK |
1363 | ret = __zs_map_object(area, pages, off, class->size); |
1364 | out: | |
48b4800a | 1365 | if (likely(!PageHugeObject(page))) |
7b60a685 MK |
1366 | ret += ZS_HANDLE_SIZE; |
1367 | ||
1368 | return ret; | |
61989a80 | 1369 | } |
66cdef66 | 1370 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1371 | |
66cdef66 | 1372 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1373 | { |
3783689a | 1374 | struct zspage *zspage; |
66cdef66 | 1375 | struct page *page; |
bfd093f5 MK |
1376 | unsigned long obj, off; |
1377 | unsigned int obj_idx; | |
61989a80 | 1378 | |
66cdef66 GM |
1379 | unsigned int class_idx; |
1380 | enum fullness_group fg; | |
1381 | struct size_class *class; | |
1382 | struct mapping_area *area; | |
9eec4cd5 | 1383 | |
2e40e163 MK |
1384 | obj = handle_to_obj(handle); |
1385 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1386 | zspage = get_zspage(page); |
1387 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 | 1388 | class = pool->size_class[class_idx]; |
bfd093f5 | 1389 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1390 | |
66cdef66 GM |
1391 | area = this_cpu_ptr(&zs_map_area); |
1392 | if (off + class->size <= PAGE_SIZE) | |
1393 | kunmap_atomic(area->vm_addr); | |
1394 | else { | |
1395 | struct page *pages[2]; | |
40f9fb8c | 1396 | |
66cdef66 GM |
1397 | pages[0] = page; |
1398 | pages[1] = get_next_page(page); | |
1399 | BUG_ON(!pages[1]); | |
1400 | ||
1401 | __zs_unmap_object(area, pages, off, class->size); | |
1402 | } | |
1403 | put_cpu_var(zs_map_area); | |
48b4800a MK |
1404 | |
1405 | migrate_read_unlock(zspage); | |
312fcae2 | 1406 | unpin_tag(handle); |
61989a80 | 1407 | } |
66cdef66 | 1408 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1409 | |
251cbb95 | 1410 | static unsigned long obj_malloc(struct size_class *class, |
3783689a | 1411 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1412 | { |
bfd093f5 | 1413 | int i, nr_page, offset; |
c7806261 MK |
1414 | unsigned long obj; |
1415 | struct link_free *link; | |
1416 | ||
1417 | struct page *m_page; | |
bfd093f5 | 1418 | unsigned long m_offset; |
c7806261 MK |
1419 | void *vaddr; |
1420 | ||
312fcae2 | 1421 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1422 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1423 | |
1424 | offset = obj * class->size; | |
1425 | nr_page = offset >> PAGE_SHIFT; | |
1426 | m_offset = offset & ~PAGE_MASK; | |
1427 | m_page = get_first_page(zspage); | |
1428 | ||
1429 | for (i = 0; i < nr_page; i++) | |
1430 | m_page = get_next_page(m_page); | |
c7806261 MK |
1431 | |
1432 | vaddr = kmap_atomic(m_page); | |
1433 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3b1d9ca6 | 1434 | set_freeobj(zspage, link->next >> OBJ_TAG_BITS); |
48b4800a | 1435 | if (likely(!PageHugeObject(m_page))) |
7b60a685 MK |
1436 | /* record handle in the header of allocated chunk */ |
1437 | link->handle = handle; | |
1438 | else | |
3783689a MK |
1439 | /* record handle to page->index */ |
1440 | zspage->first_page->index = handle; | |
1441 | ||
c7806261 | 1442 | kunmap_atomic(vaddr); |
3783689a | 1443 | mod_zspage_inuse(zspage, 1); |
c7806261 MK |
1444 | zs_stat_inc(class, OBJ_USED, 1); |
1445 | ||
bfd093f5 MK |
1446 | obj = location_to_obj(m_page, obj); |
1447 | ||
c7806261 MK |
1448 | return obj; |
1449 | } | |
1450 | ||
1451 | ||
61989a80 NG |
1452 | /** |
1453 | * zs_malloc - Allocate block of given size from pool. | |
1454 | * @pool: pool to allocate from | |
1455 | * @size: size of block to allocate | |
fd854463 | 1456 | * @gfp: gfp flags when allocating object |
61989a80 | 1457 | * |
00a61d86 | 1458 | * On success, handle to the allocated object is returned, |
c2344348 | 1459 | * otherwise 0. |
61989a80 NG |
1460 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1461 | */ | |
d0d8da2d | 1462 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1463 | { |
2e40e163 | 1464 | unsigned long handle, obj; |
61989a80 | 1465 | struct size_class *class; |
48b4800a | 1466 | enum fullness_group newfg; |
3783689a | 1467 | struct zspage *zspage; |
61989a80 | 1468 | |
7b60a685 | 1469 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1470 | return 0; |
1471 | ||
3783689a | 1472 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1473 | if (!handle) |
c2344348 | 1474 | return 0; |
61989a80 | 1475 | |
2e40e163 MK |
1476 | /* extra space in chunk to keep the handle */ |
1477 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1478 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1479 | |
1480 | spin_lock(&class->lock); | |
3783689a | 1481 | zspage = find_get_zspage(class); |
48b4800a MK |
1482 | if (likely(zspage)) { |
1483 | obj = obj_malloc(class, zspage, handle); | |
1484 | /* Now move the zspage to another fullness group, if required */ | |
1485 | fix_fullness_group(class, zspage); | |
1486 | record_obj(handle, obj); | |
61989a80 | 1487 | spin_unlock(&class->lock); |
61989a80 | 1488 | |
48b4800a MK |
1489 | return handle; |
1490 | } | |
0f050d99 | 1491 | |
48b4800a MK |
1492 | spin_unlock(&class->lock); |
1493 | ||
1494 | zspage = alloc_zspage(pool, class, gfp); | |
1495 | if (!zspage) { | |
1496 | cache_free_handle(pool, handle); | |
1497 | return 0; | |
61989a80 NG |
1498 | } |
1499 | ||
48b4800a | 1500 | spin_lock(&class->lock); |
3783689a | 1501 | obj = obj_malloc(class, zspage, handle); |
48b4800a MK |
1502 | newfg = get_fullness_group(class, zspage); |
1503 | insert_zspage(class, zspage, newfg); | |
1504 | set_zspage_mapping(zspage, class->index, newfg); | |
2e40e163 | 1505 | record_obj(handle, obj); |
48b4800a MK |
1506 | atomic_long_add(class->pages_per_zspage, |
1507 | &pool->pages_allocated); | |
b4fd07a0 | 1508 | zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
1509 | |
1510 | /* We completely set up zspage so mark them as movable */ | |
1511 | SetZsPageMovable(pool, zspage); | |
61989a80 NG |
1512 | spin_unlock(&class->lock); |
1513 | ||
2e40e163 | 1514 | return handle; |
61989a80 NG |
1515 | } |
1516 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1517 | ||
1ee47165 | 1518 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1519 | { |
1520 | struct link_free *link; | |
3783689a MK |
1521 | struct zspage *zspage; |
1522 | struct page *f_page; | |
bfd093f5 MK |
1523 | unsigned long f_offset; |
1524 | unsigned int f_objidx; | |
af4ee5e9 | 1525 | void *vaddr; |
61989a80 | 1526 | |
312fcae2 | 1527 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1528 | obj_to_location(obj, &f_page, &f_objidx); |
bfd093f5 | 1529 | f_offset = (class->size * f_objidx) & ~PAGE_MASK; |
3783689a | 1530 | zspage = get_zspage(f_page); |
61989a80 | 1531 | |
c7806261 | 1532 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1533 | |
1534 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1535 | link = (struct link_free *)(vaddr + f_offset); |
3b1d9ca6 | 1536 | link->next = get_freeobj(zspage) << OBJ_TAG_BITS; |
af4ee5e9 | 1537 | kunmap_atomic(vaddr); |
bfd093f5 | 1538 | set_freeobj(zspage, f_objidx); |
3783689a | 1539 | mod_zspage_inuse(zspage, -1); |
0f050d99 | 1540 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1541 | } |
1542 | ||
1543 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1544 | { | |
3783689a MK |
1545 | struct zspage *zspage; |
1546 | struct page *f_page; | |
bfd093f5 MK |
1547 | unsigned long obj; |
1548 | unsigned int f_objidx; | |
c7806261 MK |
1549 | int class_idx; |
1550 | struct size_class *class; | |
1551 | enum fullness_group fullness; | |
48b4800a | 1552 | bool isolated; |
c7806261 MK |
1553 | |
1554 | if (unlikely(!handle)) | |
1555 | return; | |
1556 | ||
312fcae2 | 1557 | pin_tag(handle); |
c7806261 | 1558 | obj = handle_to_obj(handle); |
c7806261 | 1559 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1560 | zspage = get_zspage(f_page); |
c7806261 | 1561 | |
48b4800a MK |
1562 | migrate_read_lock(zspage); |
1563 | ||
3783689a | 1564 | get_zspage_mapping(zspage, &class_idx, &fullness); |
c7806261 MK |
1565 | class = pool->size_class[class_idx]; |
1566 | ||
1567 | spin_lock(&class->lock); | |
1ee47165 | 1568 | obj_free(class, obj); |
3783689a | 1569 | fullness = fix_fullness_group(class, zspage); |
48b4800a MK |
1570 | if (fullness != ZS_EMPTY) { |
1571 | migrate_read_unlock(zspage); | |
1572 | goto out; | |
312fcae2 | 1573 | } |
48b4800a MK |
1574 | |
1575 | isolated = is_zspage_isolated(zspage); | |
1576 | migrate_read_unlock(zspage); | |
1577 | /* If zspage is isolated, zs_page_putback will free the zspage */ | |
1578 | if (likely(!isolated)) | |
1579 | free_zspage(pool, class, zspage); | |
1580 | out: | |
1581 | ||
61989a80 | 1582 | spin_unlock(&class->lock); |
312fcae2 | 1583 | unpin_tag(handle); |
3783689a | 1584 | cache_free_handle(pool, handle); |
312fcae2 MK |
1585 | } |
1586 | EXPORT_SYMBOL_GPL(zs_free); | |
1587 | ||
251cbb95 MK |
1588 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1589 | unsigned long src) | |
312fcae2 MK |
1590 | { |
1591 | struct page *s_page, *d_page; | |
bfd093f5 | 1592 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1593 | unsigned long s_off, d_off; |
1594 | void *s_addr, *d_addr; | |
1595 | int s_size, d_size, size; | |
1596 | int written = 0; | |
1597 | ||
1598 | s_size = d_size = class->size; | |
1599 | ||
1600 | obj_to_location(src, &s_page, &s_objidx); | |
1601 | obj_to_location(dst, &d_page, &d_objidx); | |
1602 | ||
bfd093f5 MK |
1603 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1604 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1605 | |
1606 | if (s_off + class->size > PAGE_SIZE) | |
1607 | s_size = PAGE_SIZE - s_off; | |
1608 | ||
1609 | if (d_off + class->size > PAGE_SIZE) | |
1610 | d_size = PAGE_SIZE - d_off; | |
1611 | ||
1612 | s_addr = kmap_atomic(s_page); | |
1613 | d_addr = kmap_atomic(d_page); | |
1614 | ||
1615 | while (1) { | |
1616 | size = min(s_size, d_size); | |
1617 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1618 | written += size; | |
1619 | ||
1620 | if (written == class->size) | |
1621 | break; | |
1622 | ||
495819ea SS |
1623 | s_off += size; |
1624 | s_size -= size; | |
1625 | d_off += size; | |
1626 | d_size -= size; | |
1627 | ||
1628 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1629 | kunmap_atomic(d_addr); |
1630 | kunmap_atomic(s_addr); | |
1631 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1632 | s_addr = kmap_atomic(s_page); |
1633 | d_addr = kmap_atomic(d_page); | |
1634 | s_size = class->size - written; | |
1635 | s_off = 0; | |
312fcae2 MK |
1636 | } |
1637 | ||
495819ea | 1638 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1639 | kunmap_atomic(d_addr); |
1640 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1641 | d_addr = kmap_atomic(d_page); |
1642 | d_size = class->size - written; | |
1643 | d_off = 0; | |
312fcae2 MK |
1644 | } |
1645 | } | |
1646 | ||
1647 | kunmap_atomic(d_addr); | |
1648 | kunmap_atomic(s_addr); | |
1649 | } | |
1650 | ||
1651 | /* | |
1652 | * Find alloced object in zspage from index object and | |
1653 | * return handle. | |
1654 | */ | |
251cbb95 | 1655 | static unsigned long find_alloced_obj(struct size_class *class, |
cf675acb | 1656 | struct page *page, int *obj_idx) |
312fcae2 MK |
1657 | { |
1658 | unsigned long head; | |
1659 | int offset = 0; | |
cf675acb | 1660 | int index = *obj_idx; |
312fcae2 MK |
1661 | unsigned long handle = 0; |
1662 | void *addr = kmap_atomic(page); | |
1663 | ||
3783689a | 1664 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1665 | offset += class->size * index; |
1666 | ||
1667 | while (offset < PAGE_SIZE) { | |
48b4800a | 1668 | head = obj_to_head(page, addr + offset); |
312fcae2 MK |
1669 | if (head & OBJ_ALLOCATED_TAG) { |
1670 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1671 | if (trypin_tag(handle)) | |
1672 | break; | |
1673 | handle = 0; | |
1674 | } | |
1675 | ||
1676 | offset += class->size; | |
1677 | index++; | |
1678 | } | |
1679 | ||
1680 | kunmap_atomic(addr); | |
cf675acb GM |
1681 | |
1682 | *obj_idx = index; | |
1683 | ||
312fcae2 MK |
1684 | return handle; |
1685 | } | |
1686 | ||
1687 | struct zs_compact_control { | |
3783689a | 1688 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1689 | struct page *s_page; |
1690 | /* Destination page for migration which should be a first page | |
1691 | * of zspage. */ | |
1692 | struct page *d_page; | |
1693 | /* Starting object index within @s_page which used for live object | |
1694 | * in the subpage. */ | |
41b88e14 | 1695 | int obj_idx; |
312fcae2 MK |
1696 | }; |
1697 | ||
1698 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1699 | struct zs_compact_control *cc) | |
1700 | { | |
1701 | unsigned long used_obj, free_obj; | |
1702 | unsigned long handle; | |
1703 | struct page *s_page = cc->s_page; | |
1704 | struct page *d_page = cc->d_page; | |
41b88e14 | 1705 | int obj_idx = cc->obj_idx; |
312fcae2 MK |
1706 | int ret = 0; |
1707 | ||
1708 | while (1) { | |
cf675acb | 1709 | handle = find_alloced_obj(class, s_page, &obj_idx); |
312fcae2 MK |
1710 | if (!handle) { |
1711 | s_page = get_next_page(s_page); | |
1712 | if (!s_page) | |
1713 | break; | |
41b88e14 | 1714 | obj_idx = 0; |
312fcae2 MK |
1715 | continue; |
1716 | } | |
1717 | ||
1718 | /* Stop if there is no more space */ | |
3783689a | 1719 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1720 | unpin_tag(handle); |
1721 | ret = -ENOMEM; | |
1722 | break; | |
1723 | } | |
1724 | ||
1725 | used_obj = handle_to_obj(handle); | |
3783689a | 1726 | free_obj = obj_malloc(class, get_zspage(d_page), handle); |
251cbb95 | 1727 | zs_object_copy(class, free_obj, used_obj); |
41b88e14 | 1728 | obj_idx++; |
c102f07c JL |
1729 | /* |
1730 | * record_obj updates handle's value to free_obj and it will | |
1731 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1732 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1733 | * let's keep the lock bit. | |
1734 | */ | |
1735 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1736 | record_obj(handle, free_obj); |
1737 | unpin_tag(handle); | |
1ee47165 | 1738 | obj_free(class, used_obj); |
312fcae2 MK |
1739 | } |
1740 | ||
1741 | /* Remember last position in this iteration */ | |
1742 | cc->s_page = s_page; | |
41b88e14 | 1743 | cc->obj_idx = obj_idx; |
312fcae2 MK |
1744 | |
1745 | return ret; | |
1746 | } | |
1747 | ||
3783689a | 1748 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1749 | { |
1750 | int i; | |
3783689a MK |
1751 | struct zspage *zspage; |
1752 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1753 | |
3783689a MK |
1754 | if (!source) { |
1755 | fg[0] = ZS_ALMOST_FULL; | |
1756 | fg[1] = ZS_ALMOST_EMPTY; | |
1757 | } | |
1758 | ||
1759 | for (i = 0; i < 2; i++) { | |
1760 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1761 | struct zspage, list); | |
1762 | if (zspage) { | |
48b4800a | 1763 | VM_BUG_ON(is_zspage_isolated(zspage)); |
3783689a MK |
1764 | remove_zspage(class, zspage, fg[i]); |
1765 | return zspage; | |
312fcae2 MK |
1766 | } |
1767 | } | |
1768 | ||
3783689a | 1769 | return zspage; |
312fcae2 MK |
1770 | } |
1771 | ||
860c707d | 1772 | /* |
3783689a | 1773 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1774 | * @class: destination class |
3783689a | 1775 | * @zspage: target page |
860c707d | 1776 | * |
3783689a | 1777 | * Return @zspage's fullness_group |
860c707d | 1778 | */ |
4aa409ca | 1779 | static enum fullness_group putback_zspage(struct size_class *class, |
3783689a | 1780 | struct zspage *zspage) |
312fcae2 | 1781 | { |
312fcae2 MK |
1782 | enum fullness_group fullness; |
1783 | ||
48b4800a MK |
1784 | VM_BUG_ON(is_zspage_isolated(zspage)); |
1785 | ||
3783689a MK |
1786 | fullness = get_fullness_group(class, zspage); |
1787 | insert_zspage(class, zspage, fullness); | |
1788 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1789 | |
860c707d | 1790 | return fullness; |
61989a80 | 1791 | } |
312fcae2 | 1792 | |
48b4800a MK |
1793 | #ifdef CONFIG_COMPACTION |
1794 | static struct dentry *zs_mount(struct file_system_type *fs_type, | |
1795 | int flags, const char *dev_name, void *data) | |
1796 | { | |
1797 | static const struct dentry_operations ops = { | |
1798 | .d_dname = simple_dname, | |
1799 | }; | |
1800 | ||
1801 | return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC); | |
1802 | } | |
1803 | ||
1804 | static struct file_system_type zsmalloc_fs = { | |
1805 | .name = "zsmalloc", | |
1806 | .mount = zs_mount, | |
1807 | .kill_sb = kill_anon_super, | |
1808 | }; | |
1809 | ||
1810 | static int zsmalloc_mount(void) | |
1811 | { | |
1812 | int ret = 0; | |
1813 | ||
1814 | zsmalloc_mnt = kern_mount(&zsmalloc_fs); | |
1815 | if (IS_ERR(zsmalloc_mnt)) | |
1816 | ret = PTR_ERR(zsmalloc_mnt); | |
1817 | ||
1818 | return ret; | |
1819 | } | |
1820 | ||
1821 | static void zsmalloc_unmount(void) | |
1822 | { | |
1823 | kern_unmount(zsmalloc_mnt); | |
1824 | } | |
1825 | ||
1826 | static void migrate_lock_init(struct zspage *zspage) | |
1827 | { | |
1828 | rwlock_init(&zspage->lock); | |
1829 | } | |
1830 | ||
1831 | static void migrate_read_lock(struct zspage *zspage) | |
1832 | { | |
1833 | read_lock(&zspage->lock); | |
1834 | } | |
1835 | ||
1836 | static void migrate_read_unlock(struct zspage *zspage) | |
1837 | { | |
1838 | read_unlock(&zspage->lock); | |
1839 | } | |
1840 | ||
1841 | static void migrate_write_lock(struct zspage *zspage) | |
1842 | { | |
1843 | write_lock(&zspage->lock); | |
1844 | } | |
1845 | ||
1846 | static void migrate_write_unlock(struct zspage *zspage) | |
1847 | { | |
1848 | write_unlock(&zspage->lock); | |
1849 | } | |
1850 | ||
1851 | /* Number of isolated subpage for *page migration* in this zspage */ | |
1852 | static void inc_zspage_isolation(struct zspage *zspage) | |
1853 | { | |
1854 | zspage->isolated++; | |
1855 | } | |
1856 | ||
1857 | static void dec_zspage_isolation(struct zspage *zspage) | |
1858 | { | |
1859 | zspage->isolated--; | |
1860 | } | |
1861 | ||
1862 | static void replace_sub_page(struct size_class *class, struct zspage *zspage, | |
1863 | struct page *newpage, struct page *oldpage) | |
1864 | { | |
1865 | struct page *page; | |
1866 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, }; | |
1867 | int idx = 0; | |
1868 | ||
1869 | page = get_first_page(zspage); | |
1870 | do { | |
1871 | if (page == oldpage) | |
1872 | pages[idx] = newpage; | |
1873 | else | |
1874 | pages[idx] = page; | |
1875 | idx++; | |
1876 | } while ((page = get_next_page(page)) != NULL); | |
1877 | ||
1878 | create_page_chain(class, zspage, pages); | |
1879 | set_first_obj_offset(newpage, get_first_obj_offset(oldpage)); | |
1880 | if (unlikely(PageHugeObject(oldpage))) | |
1881 | newpage->index = oldpage->index; | |
1882 | __SetPageMovable(newpage, page_mapping(oldpage)); | |
1883 | } | |
1884 | ||
1885 | bool zs_page_isolate(struct page *page, isolate_mode_t mode) | |
1886 | { | |
1887 | struct zs_pool *pool; | |
1888 | struct size_class *class; | |
1889 | int class_idx; | |
1890 | enum fullness_group fullness; | |
1891 | struct zspage *zspage; | |
1892 | struct address_space *mapping; | |
1893 | ||
1894 | /* | |
1895 | * Page is locked so zspage couldn't be destroyed. For detail, look at | |
1896 | * lock_zspage in free_zspage. | |
1897 | */ | |
1898 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1899 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1900 | ||
1901 | zspage = get_zspage(page); | |
1902 | ||
1903 | /* | |
1904 | * Without class lock, fullness could be stale while class_idx is okay | |
1905 | * because class_idx is constant unless page is freed so we should get | |
1906 | * fullness again under class lock. | |
1907 | */ | |
1908 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1909 | mapping = page_mapping(page); | |
1910 | pool = mapping->private_data; | |
1911 | class = pool->size_class[class_idx]; | |
1912 | ||
1913 | spin_lock(&class->lock); | |
1914 | if (get_zspage_inuse(zspage) == 0) { | |
1915 | spin_unlock(&class->lock); | |
1916 | return false; | |
1917 | } | |
1918 | ||
1919 | /* zspage is isolated for object migration */ | |
1920 | if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
1921 | spin_unlock(&class->lock); | |
1922 | return false; | |
1923 | } | |
1924 | ||
1925 | /* | |
1926 | * If this is first time isolation for the zspage, isolate zspage from | |
1927 | * size_class to prevent further object allocation from the zspage. | |
1928 | */ | |
1929 | if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
1930 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1931 | remove_zspage(class, zspage, fullness); | |
1932 | } | |
1933 | ||
1934 | inc_zspage_isolation(zspage); | |
1935 | spin_unlock(&class->lock); | |
1936 | ||
1937 | return true; | |
1938 | } | |
1939 | ||
1940 | int zs_page_migrate(struct address_space *mapping, struct page *newpage, | |
1941 | struct page *page, enum migrate_mode mode) | |
1942 | { | |
1943 | struct zs_pool *pool; | |
1944 | struct size_class *class; | |
1945 | int class_idx; | |
1946 | enum fullness_group fullness; | |
1947 | struct zspage *zspage; | |
1948 | struct page *dummy; | |
1949 | void *s_addr, *d_addr, *addr; | |
1950 | int offset, pos; | |
1951 | unsigned long handle, head; | |
1952 | unsigned long old_obj, new_obj; | |
1953 | unsigned int obj_idx; | |
1954 | int ret = -EAGAIN; | |
1955 | ||
2916ecc0 JG |
1956 | /* |
1957 | * We cannot support the _NO_COPY case here, because copy needs to | |
1958 | * happen under the zs lock, which does not work with | |
1959 | * MIGRATE_SYNC_NO_COPY workflow. | |
1960 | */ | |
1961 | if (mode == MIGRATE_SYNC_NO_COPY) | |
1962 | return -EINVAL; | |
1963 | ||
48b4800a MK |
1964 | VM_BUG_ON_PAGE(!PageMovable(page), page); |
1965 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
1966 | ||
1967 | zspage = get_zspage(page); | |
1968 | ||
1969 | /* Concurrent compactor cannot migrate any subpage in zspage */ | |
1970 | migrate_write_lock(zspage); | |
1971 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1972 | pool = mapping->private_data; | |
1973 | class = pool->size_class[class_idx]; | |
1974 | offset = get_first_obj_offset(page); | |
1975 | ||
1976 | spin_lock(&class->lock); | |
1977 | if (!get_zspage_inuse(zspage)) { | |
77ff4657 HZ |
1978 | /* |
1979 | * Set "offset" to end of the page so that every loops | |
1980 | * skips unnecessary object scanning. | |
1981 | */ | |
1982 | offset = PAGE_SIZE; | |
48b4800a MK |
1983 | } |
1984 | ||
1985 | pos = offset; | |
1986 | s_addr = kmap_atomic(page); | |
1987 | while (pos < PAGE_SIZE) { | |
1988 | head = obj_to_head(page, s_addr + pos); | |
1989 | if (head & OBJ_ALLOCATED_TAG) { | |
1990 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1991 | if (!trypin_tag(handle)) | |
1992 | goto unpin_objects; | |
1993 | } | |
1994 | pos += class->size; | |
1995 | } | |
1996 | ||
1997 | /* | |
1998 | * Here, any user cannot access all objects in the zspage so let's move. | |
1999 | */ | |
2000 | d_addr = kmap_atomic(newpage); | |
2001 | memcpy(d_addr, s_addr, PAGE_SIZE); | |
2002 | kunmap_atomic(d_addr); | |
2003 | ||
2004 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2005 | addr += class->size) { | |
2006 | head = obj_to_head(page, addr); | |
2007 | if (head & OBJ_ALLOCATED_TAG) { | |
2008 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2009 | if (!testpin_tag(handle)) | |
2010 | BUG(); | |
2011 | ||
2012 | old_obj = handle_to_obj(handle); | |
2013 | obj_to_location(old_obj, &dummy, &obj_idx); | |
2014 | new_obj = (unsigned long)location_to_obj(newpage, | |
2015 | obj_idx); | |
2016 | new_obj |= BIT(HANDLE_PIN_BIT); | |
2017 | record_obj(handle, new_obj); | |
2018 | } | |
2019 | } | |
2020 | ||
2021 | replace_sub_page(class, zspage, newpage, page); | |
2022 | get_page(newpage); | |
2023 | ||
2024 | dec_zspage_isolation(zspage); | |
2025 | ||
2026 | /* | |
2027 | * Page migration is done so let's putback isolated zspage to | |
2028 | * the list if @page is final isolated subpage in the zspage. | |
2029 | */ | |
2030 | if (!is_zspage_isolated(zspage)) | |
2031 | putback_zspage(class, zspage); | |
2032 | ||
2033 | reset_page(page); | |
2034 | put_page(page); | |
2035 | page = newpage; | |
2036 | ||
dd4123f3 | 2037 | ret = MIGRATEPAGE_SUCCESS; |
48b4800a MK |
2038 | unpin_objects: |
2039 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2040 | addr += class->size) { | |
2041 | head = obj_to_head(page, addr); | |
2042 | if (head & OBJ_ALLOCATED_TAG) { | |
2043 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2044 | if (!testpin_tag(handle)) | |
2045 | BUG(); | |
2046 | unpin_tag(handle); | |
2047 | } | |
2048 | } | |
2049 | kunmap_atomic(s_addr); | |
48b4800a MK |
2050 | spin_unlock(&class->lock); |
2051 | migrate_write_unlock(zspage); | |
2052 | ||
2053 | return ret; | |
2054 | } | |
2055 | ||
2056 | void zs_page_putback(struct page *page) | |
2057 | { | |
2058 | struct zs_pool *pool; | |
2059 | struct size_class *class; | |
2060 | int class_idx; | |
2061 | enum fullness_group fg; | |
2062 | struct address_space *mapping; | |
2063 | struct zspage *zspage; | |
2064 | ||
2065 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
2066 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
2067 | ||
2068 | zspage = get_zspage(page); | |
2069 | get_zspage_mapping(zspage, &class_idx, &fg); | |
2070 | mapping = page_mapping(page); | |
2071 | pool = mapping->private_data; | |
2072 | class = pool->size_class[class_idx]; | |
2073 | ||
2074 | spin_lock(&class->lock); | |
2075 | dec_zspage_isolation(zspage); | |
2076 | if (!is_zspage_isolated(zspage)) { | |
2077 | fg = putback_zspage(class, zspage); | |
2078 | /* | |
2079 | * Due to page_lock, we cannot free zspage immediately | |
2080 | * so let's defer. | |
2081 | */ | |
2082 | if (fg == ZS_EMPTY) | |
2083 | schedule_work(&pool->free_work); | |
2084 | } | |
2085 | spin_unlock(&class->lock); | |
2086 | } | |
2087 | ||
2088 | const struct address_space_operations zsmalloc_aops = { | |
2089 | .isolate_page = zs_page_isolate, | |
2090 | .migratepage = zs_page_migrate, | |
2091 | .putback_page = zs_page_putback, | |
2092 | }; | |
2093 | ||
2094 | static int zs_register_migration(struct zs_pool *pool) | |
2095 | { | |
2096 | pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb); | |
2097 | if (IS_ERR(pool->inode)) { | |
2098 | pool->inode = NULL; | |
2099 | return 1; | |
2100 | } | |
2101 | ||
2102 | pool->inode->i_mapping->private_data = pool; | |
2103 | pool->inode->i_mapping->a_ops = &zsmalloc_aops; | |
2104 | return 0; | |
2105 | } | |
2106 | ||
2107 | static void zs_unregister_migration(struct zs_pool *pool) | |
2108 | { | |
2109 | flush_work(&pool->free_work); | |
c3491eca | 2110 | iput(pool->inode); |
48b4800a MK |
2111 | } |
2112 | ||
2113 | /* | |
2114 | * Caller should hold page_lock of all pages in the zspage | |
2115 | * In here, we cannot use zspage meta data. | |
2116 | */ | |
2117 | static void async_free_zspage(struct work_struct *work) | |
2118 | { | |
2119 | int i; | |
2120 | struct size_class *class; | |
2121 | unsigned int class_idx; | |
2122 | enum fullness_group fullness; | |
2123 | struct zspage *zspage, *tmp; | |
2124 | LIST_HEAD(free_pages); | |
2125 | struct zs_pool *pool = container_of(work, struct zs_pool, | |
2126 | free_work); | |
2127 | ||
cf8e0fed | 2128 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
48b4800a MK |
2129 | class = pool->size_class[i]; |
2130 | if (class->index != i) | |
2131 | continue; | |
2132 | ||
2133 | spin_lock(&class->lock); | |
2134 | list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages); | |
2135 | spin_unlock(&class->lock); | |
2136 | } | |
2137 | ||
2138 | ||
2139 | list_for_each_entry_safe(zspage, tmp, &free_pages, list) { | |
2140 | list_del(&zspage->list); | |
2141 | lock_zspage(zspage); | |
2142 | ||
2143 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2144 | VM_BUG_ON(fullness != ZS_EMPTY); | |
2145 | class = pool->size_class[class_idx]; | |
2146 | spin_lock(&class->lock); | |
2147 | __free_zspage(pool, pool->size_class[class_idx], zspage); | |
2148 | spin_unlock(&class->lock); | |
2149 | } | |
2150 | }; | |
2151 | ||
2152 | static void kick_deferred_free(struct zs_pool *pool) | |
2153 | { | |
2154 | schedule_work(&pool->free_work); | |
2155 | } | |
2156 | ||
2157 | static void init_deferred_free(struct zs_pool *pool) | |
2158 | { | |
2159 | INIT_WORK(&pool->free_work, async_free_zspage); | |
2160 | } | |
2161 | ||
2162 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) | |
2163 | { | |
2164 | struct page *page = get_first_page(zspage); | |
2165 | ||
2166 | do { | |
2167 | WARN_ON(!trylock_page(page)); | |
2168 | __SetPageMovable(page, pool->inode->i_mapping); | |
2169 | unlock_page(page); | |
2170 | } while ((page = get_next_page(page)) != NULL); | |
2171 | } | |
2172 | #endif | |
2173 | ||
04f05909 SS |
2174 | /* |
2175 | * | |
2176 | * Based on the number of unused allocated objects calculate | |
2177 | * and return the number of pages that we can free. | |
04f05909 SS |
2178 | */ |
2179 | static unsigned long zs_can_compact(struct size_class *class) | |
2180 | { | |
2181 | unsigned long obj_wasted; | |
44f43e99 SS |
2182 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
2183 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 2184 | |
44f43e99 SS |
2185 | if (obj_allocated <= obj_used) |
2186 | return 0; | |
04f05909 | 2187 | |
44f43e99 | 2188 | obj_wasted = obj_allocated - obj_used; |
b4fd07a0 | 2189 | obj_wasted /= class->objs_per_zspage; |
04f05909 | 2190 | |
6cbf16b3 | 2191 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
2192 | } |
2193 | ||
7d3f3938 | 2194 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 2195 | { |
312fcae2 | 2196 | struct zs_compact_control cc; |
3783689a MK |
2197 | struct zspage *src_zspage; |
2198 | struct zspage *dst_zspage = NULL; | |
312fcae2 | 2199 | |
312fcae2 | 2200 | spin_lock(&class->lock); |
3783689a | 2201 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 2202 | |
04f05909 SS |
2203 | if (!zs_can_compact(class)) |
2204 | break; | |
2205 | ||
41b88e14 | 2206 | cc.obj_idx = 0; |
48b4800a | 2207 | cc.s_page = get_first_page(src_zspage); |
312fcae2 | 2208 | |
3783689a | 2209 | while ((dst_zspage = isolate_zspage(class, false))) { |
48b4800a | 2210 | cc.d_page = get_first_page(dst_zspage); |
312fcae2 | 2211 | /* |
0dc63d48 SS |
2212 | * If there is no more space in dst_page, resched |
2213 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
2214 | */ |
2215 | if (!migrate_zspage(pool, class, &cc)) | |
2216 | break; | |
2217 | ||
4aa409ca | 2218 | putback_zspage(class, dst_zspage); |
312fcae2 MK |
2219 | } |
2220 | ||
2221 | /* Stop if we couldn't find slot */ | |
3783689a | 2222 | if (dst_zspage == NULL) |
312fcae2 MK |
2223 | break; |
2224 | ||
4aa409ca MK |
2225 | putback_zspage(class, dst_zspage); |
2226 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { | |
48b4800a | 2227 | free_zspage(pool, class, src_zspage); |
6cbf16b3 | 2228 | pool->stats.pages_compacted += class->pages_per_zspage; |
4aa409ca | 2229 | } |
312fcae2 | 2230 | spin_unlock(&class->lock); |
312fcae2 MK |
2231 | cond_resched(); |
2232 | spin_lock(&class->lock); | |
2233 | } | |
2234 | ||
3783689a | 2235 | if (src_zspage) |
4aa409ca | 2236 | putback_zspage(class, src_zspage); |
312fcae2 | 2237 | |
7d3f3938 | 2238 | spin_unlock(&class->lock); |
312fcae2 MK |
2239 | } |
2240 | ||
2241 | unsigned long zs_compact(struct zs_pool *pool) | |
2242 | { | |
2243 | int i; | |
312fcae2 MK |
2244 | struct size_class *class; |
2245 | ||
cf8e0fed | 2246 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
312fcae2 MK |
2247 | class = pool->size_class[i]; |
2248 | if (!class) | |
2249 | continue; | |
2250 | if (class->index != i) | |
2251 | continue; | |
7d3f3938 | 2252 | __zs_compact(pool, class); |
312fcae2 MK |
2253 | } |
2254 | ||
860c707d | 2255 | return pool->stats.pages_compacted; |
312fcae2 MK |
2256 | } |
2257 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 2258 | |
7d3f3938 SS |
2259 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
2260 | { | |
2261 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
2262 | } | |
2263 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
2264 | ||
ab9d306d SS |
2265 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
2266 | struct shrink_control *sc) | |
2267 | { | |
2268 | unsigned long pages_freed; | |
2269 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2270 | shrinker); | |
2271 | ||
2272 | pages_freed = pool->stats.pages_compacted; | |
2273 | /* | |
2274 | * Compact classes and calculate compaction delta. | |
2275 | * Can run concurrently with a manually triggered | |
2276 | * (by user) compaction. | |
2277 | */ | |
2278 | pages_freed = zs_compact(pool) - pages_freed; | |
2279 | ||
2280 | return pages_freed ? pages_freed : SHRINK_STOP; | |
2281 | } | |
2282 | ||
2283 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
2284 | struct shrink_control *sc) | |
2285 | { | |
2286 | int i; | |
2287 | struct size_class *class; | |
2288 | unsigned long pages_to_free = 0; | |
2289 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2290 | shrinker); | |
2291 | ||
cf8e0fed | 2292 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
ab9d306d SS |
2293 | class = pool->size_class[i]; |
2294 | if (!class) | |
2295 | continue; | |
2296 | if (class->index != i) | |
2297 | continue; | |
2298 | ||
ab9d306d | 2299 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
2300 | } |
2301 | ||
2302 | return pages_to_free; | |
2303 | } | |
2304 | ||
2305 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
2306 | { | |
93144ca3 | 2307 | unregister_shrinker(&pool->shrinker); |
ab9d306d SS |
2308 | } |
2309 | ||
2310 | static int zs_register_shrinker(struct zs_pool *pool) | |
2311 | { | |
2312 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
2313 | pool->shrinker.count_objects = zs_shrinker_count; | |
2314 | pool->shrinker.batch = 0; | |
2315 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
2316 | ||
2317 | return register_shrinker(&pool->shrinker); | |
2318 | } | |
2319 | ||
00a61d86 | 2320 | /** |
66cdef66 | 2321 | * zs_create_pool - Creates an allocation pool to work from. |
fd854463 | 2322 | * @name: pool name to be created |
166cfda7 | 2323 | * |
66cdef66 GM |
2324 | * This function must be called before anything when using |
2325 | * the zsmalloc allocator. | |
166cfda7 | 2326 | * |
66cdef66 GM |
2327 | * On success, a pointer to the newly created pool is returned, |
2328 | * otherwise NULL. | |
396b7fd6 | 2329 | */ |
d0d8da2d | 2330 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 2331 | { |
66cdef66 GM |
2332 | int i; |
2333 | struct zs_pool *pool; | |
2334 | struct size_class *prev_class = NULL; | |
61989a80 | 2335 | |
66cdef66 GM |
2336 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
2337 | if (!pool) | |
2338 | return NULL; | |
61989a80 | 2339 | |
48b4800a | 2340 | init_deferred_free(pool); |
61989a80 | 2341 | |
2e40e163 MK |
2342 | pool->name = kstrdup(name, GFP_KERNEL); |
2343 | if (!pool->name) | |
2344 | goto err; | |
2345 | ||
3783689a | 2346 | if (create_cache(pool)) |
2e40e163 MK |
2347 | goto err; |
2348 | ||
c60369f0 | 2349 | /* |
399d8eeb | 2350 | * Iterate reversely, because, size of size_class that we want to use |
66cdef66 | 2351 | * for merging should be larger or equal to current size. |
c60369f0 | 2352 | */ |
cf8e0fed | 2353 | for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) { |
66cdef66 GM |
2354 | int size; |
2355 | int pages_per_zspage; | |
64d90465 | 2356 | int objs_per_zspage; |
66cdef66 | 2357 | struct size_class *class; |
3783689a | 2358 | int fullness = 0; |
c60369f0 | 2359 | |
66cdef66 GM |
2360 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
2361 | if (size > ZS_MAX_ALLOC_SIZE) | |
2362 | size = ZS_MAX_ALLOC_SIZE; | |
2363 | pages_per_zspage = get_pages_per_zspage(size); | |
64d90465 | 2364 | objs_per_zspage = pages_per_zspage * PAGE_SIZE / size; |
61989a80 | 2365 | |
66cdef66 GM |
2366 | /* |
2367 | * size_class is used for normal zsmalloc operation such | |
2368 | * as alloc/free for that size. Although it is natural that we | |
2369 | * have one size_class for each size, there is a chance that we | |
2370 | * can get more memory utilization if we use one size_class for | |
2371 | * many different sizes whose size_class have same | |
2372 | * characteristics. So, we makes size_class point to | |
2373 | * previous size_class if possible. | |
2374 | */ | |
2375 | if (prev_class) { | |
64d90465 | 2376 | if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) { |
66cdef66 GM |
2377 | pool->size_class[i] = prev_class; |
2378 | continue; | |
2379 | } | |
2380 | } | |
2381 | ||
2382 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
2383 | if (!class) | |
2384 | goto err; | |
2385 | ||
2386 | class->size = size; | |
2387 | class->index = i; | |
2388 | class->pages_per_zspage = pages_per_zspage; | |
64d90465 | 2389 | class->objs_per_zspage = objs_per_zspage; |
66cdef66 GM |
2390 | spin_lock_init(&class->lock); |
2391 | pool->size_class[i] = class; | |
48b4800a MK |
2392 | for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS; |
2393 | fullness++) | |
3783689a | 2394 | INIT_LIST_HEAD(&class->fullness_list[fullness]); |
66cdef66 GM |
2395 | |
2396 | prev_class = class; | |
61989a80 NG |
2397 | } |
2398 | ||
d34f6157 DS |
2399 | /* debug only, don't abort if it fails */ |
2400 | zs_pool_stat_create(pool, name); | |
0f050d99 | 2401 | |
48b4800a MK |
2402 | if (zs_register_migration(pool)) |
2403 | goto err; | |
2404 | ||
ab9d306d | 2405 | /* |
93144ca3 AK |
2406 | * Not critical since shrinker is only used to trigger internal |
2407 | * defragmentation of the pool which is pretty optional thing. If | |
2408 | * registration fails we still can use the pool normally and user can | |
2409 | * trigger compaction manually. Thus, ignore return code. | |
ab9d306d | 2410 | */ |
93144ca3 AK |
2411 | zs_register_shrinker(pool); |
2412 | ||
66cdef66 GM |
2413 | return pool; |
2414 | ||
2415 | err: | |
2416 | zs_destroy_pool(pool); | |
2417 | return NULL; | |
61989a80 | 2418 | } |
66cdef66 | 2419 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2420 | |
66cdef66 | 2421 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2422 | { |
66cdef66 | 2423 | int i; |
61989a80 | 2424 | |
ab9d306d | 2425 | zs_unregister_shrinker(pool); |
48b4800a | 2426 | zs_unregister_migration(pool); |
0f050d99 GM |
2427 | zs_pool_stat_destroy(pool); |
2428 | ||
cf8e0fed | 2429 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
66cdef66 GM |
2430 | int fg; |
2431 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2432 | |
66cdef66 GM |
2433 | if (!class) |
2434 | continue; | |
61989a80 | 2435 | |
66cdef66 GM |
2436 | if (class->index != i) |
2437 | continue; | |
61989a80 | 2438 | |
48b4800a | 2439 | for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) { |
3783689a | 2440 | if (!list_empty(&class->fullness_list[fg])) { |
66cdef66 GM |
2441 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
2442 | class->size, fg); | |
2443 | } | |
2444 | } | |
2445 | kfree(class); | |
2446 | } | |
f553646a | 2447 | |
3783689a | 2448 | destroy_cache(pool); |
0f050d99 | 2449 | kfree(pool->name); |
66cdef66 GM |
2450 | kfree(pool); |
2451 | } | |
2452 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2453 | |
66cdef66 GM |
2454 | static int __init zs_init(void) |
2455 | { | |
48b4800a MK |
2456 | int ret; |
2457 | ||
2458 | ret = zsmalloc_mount(); | |
2459 | if (ret) | |
2460 | goto out; | |
2461 | ||
215c89d0 SAS |
2462 | ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare", |
2463 | zs_cpu_prepare, zs_cpu_dead); | |
0f050d99 | 2464 | if (ret) |
215c89d0 | 2465 | goto hp_setup_fail; |
66cdef66 | 2466 | |
66cdef66 GM |
2467 | #ifdef CONFIG_ZPOOL |
2468 | zpool_register_driver(&zs_zpool_driver); | |
2469 | #endif | |
0f050d99 | 2470 | |
4abaac9b DS |
2471 | zs_stat_init(); |
2472 | ||
66cdef66 | 2473 | return 0; |
0f050d99 | 2474 | |
215c89d0 | 2475 | hp_setup_fail: |
48b4800a MK |
2476 | zsmalloc_unmount(); |
2477 | out: | |
0f050d99 | 2478 | return ret; |
61989a80 | 2479 | } |
61989a80 | 2480 | |
66cdef66 | 2481 | static void __exit zs_exit(void) |
61989a80 | 2482 | { |
66cdef66 GM |
2483 | #ifdef CONFIG_ZPOOL |
2484 | zpool_unregister_driver(&zs_zpool_driver); | |
2485 | #endif | |
48b4800a | 2486 | zsmalloc_unmount(); |
215c89d0 | 2487 | cpuhp_remove_state(CPUHP_MM_ZS_PREPARE); |
0f050d99 GM |
2488 | |
2489 | zs_stat_exit(); | |
61989a80 | 2490 | } |
069f101f BH |
2491 | |
2492 | module_init(zs_init); | |
2493 | module_exit(zs_exit); | |
2494 | ||
2495 | MODULE_LICENSE("Dual BSD/GPL"); | |
2496 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |