]>
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. | |
2db51dae NG |
23 | * |
24 | * Usage of struct page flags: | |
25 | * PG_private: identifies the first component page | |
26 | * PG_private2: identifies the last component page | |
48b4800a | 27 | * PG_owner_priv_1: indentifies 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> |
61989a80 NG |
36 | #include <linux/bitops.h> |
37 | #include <linux/errno.h> | |
38 | #include <linux/highmem.h> | |
61989a80 NG |
39 | #include <linux/string.h> |
40 | #include <linux/slab.h> | |
41 | #include <asm/tlbflush.h> | |
42 | #include <asm/pgtable.h> | |
43 | #include <linux/cpumask.h> | |
44 | #include <linux/cpu.h> | |
0cbb613f | 45 | #include <linux/vmalloc.h> |
759b26b2 | 46 | #include <linux/preempt.h> |
0959c63f SJ |
47 | #include <linux/spinlock.h> |
48 | #include <linux/types.h> | |
0f050d99 | 49 | #include <linux/debugfs.h> |
bcf1647d | 50 | #include <linux/zsmalloc.h> |
c795779d | 51 | #include <linux/zpool.h> |
48b4800a | 52 | #include <linux/mount.h> |
dd4123f3 | 53 | #include <linux/migrate.h> |
48b4800a MK |
54 | #include <linux/pagemap.h> |
55 | ||
56 | #define ZSPAGE_MAGIC 0x58 | |
0959c63f SJ |
57 | |
58 | /* | |
59 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
60 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
61 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
62 | * to restore link_free pointer values. | |
63 | */ | |
64 | #define ZS_ALIGN 8 | |
65 | ||
66 | /* | |
67 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
68 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
69 | */ | |
70 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
71 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
72 | ||
2e40e163 MK |
73 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
74 | ||
0959c63f SJ |
75 | /* |
76 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 77 | * as single (unsigned long) handle value. |
0959c63f | 78 | * |
bfd093f5 | 79 | * Note that object index <obj_idx> starts from 0. |
0959c63f SJ |
80 | * |
81 | * This is made more complicated by various memory models and PAE. | |
82 | */ | |
83 | ||
84 | #ifndef MAX_PHYSMEM_BITS | |
85 | #ifdef CONFIG_HIGHMEM64G | |
86 | #define MAX_PHYSMEM_BITS 36 | |
87 | #else /* !CONFIG_HIGHMEM64G */ | |
88 | /* | |
89 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
90 | * be PAGE_SHIFT | |
91 | */ | |
92 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
93 | #endif | |
94 | #endif | |
95 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
96 | |
97 | /* | |
98 | * Memory for allocating for handle keeps object position by | |
99 | * encoding <page, obj_idx> and the encoded value has a room | |
100 | * in least bit(ie, look at obj_to_location). | |
101 | * We use the bit to synchronize between object access by | |
102 | * user and migration. | |
103 | */ | |
104 | #define HANDLE_PIN_BIT 0 | |
105 | ||
106 | /* | |
107 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
108 | * to identify the object was allocated or not. | |
109 | * It's okay to add the status bit in the least bit because | |
110 | * header keeps handle which is 4byte-aligned address so we | |
111 | * have room for two bit at least. | |
112 | */ | |
113 | #define OBJ_ALLOCATED_TAG 1 | |
114 | #define OBJ_TAG_BITS 1 | |
115 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
116 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
117 | ||
118 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
119 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
120 | #define ZS_MIN_ALLOC_SIZE \ | |
121 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 122 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 123 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
124 | |
125 | /* | |
7eb52512 | 126 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
127 | * trader-off here: |
128 | * - Large number of size classes is potentially wasteful as free page are | |
129 | * spread across these classes | |
130 | * - Small number of size classes causes large internal fragmentation | |
131 | * - Probably its better to use specific size classes (empirically | |
132 | * determined). NOTE: all those class sizes must be set as multiple of | |
133 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
134 | * | |
135 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
136 | * (reason above) | |
137 | */ | |
3783689a | 138 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS) |
0959c63f SJ |
139 | |
140 | /* | |
141 | * We do not maintain any list for completely empty or full pages | |
142 | */ | |
143 | enum fullness_group { | |
0959c63f | 144 | ZS_EMPTY, |
48b4800a MK |
145 | ZS_ALMOST_EMPTY, |
146 | ZS_ALMOST_FULL, | |
147 | ZS_FULL, | |
148 | NR_ZS_FULLNESS, | |
0959c63f SJ |
149 | }; |
150 | ||
0f050d99 | 151 | enum zs_stat_type { |
48b4800a MK |
152 | CLASS_EMPTY, |
153 | CLASS_ALMOST_EMPTY, | |
154 | CLASS_ALMOST_FULL, | |
155 | CLASS_FULL, | |
0f050d99 GM |
156 | OBJ_ALLOCATED, |
157 | OBJ_USED, | |
48b4800a | 158 | NR_ZS_STAT_TYPE, |
0f050d99 GM |
159 | }; |
160 | ||
0f050d99 GM |
161 | struct zs_size_stat { |
162 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
163 | }; | |
164 | ||
57244594 SS |
165 | #ifdef CONFIG_ZSMALLOC_STAT |
166 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
167 | #endif |
168 | ||
48b4800a MK |
169 | #ifdef CONFIG_COMPACTION |
170 | static struct vfsmount *zsmalloc_mnt; | |
171 | #endif | |
172 | ||
40f9fb8c MG |
173 | /* |
174 | * number of size_classes | |
175 | */ | |
176 | static int zs_size_classes; | |
177 | ||
0959c63f SJ |
178 | /* |
179 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
180 | * n <= N / f, where | |
181 | * n = number of allocated objects | |
182 | * N = total number of objects zspage can store | |
6dd9737e | 183 | * f = fullness_threshold_frac |
0959c63f SJ |
184 | * |
185 | * Similarly, we assign zspage to: | |
186 | * ZS_ALMOST_FULL when n > N / f | |
187 | * ZS_EMPTY when n == 0 | |
188 | * ZS_FULL when n == N | |
189 | * | |
190 | * (see: fix_fullness_group()) | |
191 | */ | |
192 | static const int fullness_threshold_frac = 4; | |
193 | ||
194 | struct size_class { | |
57244594 | 195 | spinlock_t lock; |
48b4800a | 196 | struct list_head fullness_list[NR_ZS_FULLNESS]; |
0959c63f SJ |
197 | /* |
198 | * Size of objects stored in this class. Must be multiple | |
199 | * of ZS_ALIGN. | |
200 | */ | |
201 | int size; | |
1fc6e27d | 202 | int objs_per_zspage; |
7dfa4612 WY |
203 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
204 | int pages_per_zspage; | |
48b4800a MK |
205 | |
206 | unsigned int index; | |
207 | struct zs_size_stat stats; | |
0959c63f SJ |
208 | }; |
209 | ||
48b4800a MK |
210 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
211 | static void SetPageHugeObject(struct page *page) | |
212 | { | |
213 | SetPageOwnerPriv1(page); | |
214 | } | |
215 | ||
216 | static void ClearPageHugeObject(struct page *page) | |
217 | { | |
218 | ClearPageOwnerPriv1(page); | |
219 | } | |
220 | ||
221 | static int PageHugeObject(struct page *page) | |
222 | { | |
223 | return PageOwnerPriv1(page); | |
224 | } | |
225 | ||
0959c63f SJ |
226 | /* |
227 | * Placed within free objects to form a singly linked list. | |
3783689a | 228 | * For every zspage, zspage->freeobj gives head of this list. |
0959c63f SJ |
229 | * |
230 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
231 | */ | |
232 | struct link_free { | |
2e40e163 MK |
233 | union { |
234 | /* | |
bfd093f5 | 235 | * Free object index; |
2e40e163 MK |
236 | * It's valid for non-allocated object |
237 | */ | |
bfd093f5 | 238 | unsigned long next; |
2e40e163 MK |
239 | /* |
240 | * Handle of allocated object. | |
241 | */ | |
242 | unsigned long handle; | |
243 | }; | |
0959c63f SJ |
244 | }; |
245 | ||
246 | struct zs_pool { | |
6f3526d6 | 247 | const char *name; |
0f050d99 | 248 | |
40f9fb8c | 249 | struct size_class **size_class; |
2e40e163 | 250 | struct kmem_cache *handle_cachep; |
3783689a | 251 | struct kmem_cache *zspage_cachep; |
0959c63f | 252 | |
13de8933 | 253 | atomic_long_t pages_allocated; |
0f050d99 | 254 | |
7d3f3938 | 255 | struct zs_pool_stats stats; |
ab9d306d SS |
256 | |
257 | /* Compact classes */ | |
258 | struct shrinker shrinker; | |
259 | /* | |
260 | * To signify that register_shrinker() was successful | |
261 | * and unregister_shrinker() will not Oops. | |
262 | */ | |
263 | bool shrinker_enabled; | |
0f050d99 GM |
264 | #ifdef CONFIG_ZSMALLOC_STAT |
265 | struct dentry *stat_dentry; | |
266 | #endif | |
48b4800a MK |
267 | #ifdef CONFIG_COMPACTION |
268 | struct inode *inode; | |
269 | struct work_struct free_work; | |
270 | #endif | |
0959c63f | 271 | }; |
61989a80 NG |
272 | |
273 | /* | |
274 | * A zspage's class index and fullness group | |
275 | * are encoded in its (first)page->mapping | |
276 | */ | |
3783689a MK |
277 | #define FULLNESS_BITS 2 |
278 | #define CLASS_BITS 8 | |
48b4800a MK |
279 | #define ISOLATED_BITS 3 |
280 | #define MAGIC_VAL_BITS 8 | |
4f42047b | 281 | |
3783689a MK |
282 | struct zspage { |
283 | struct { | |
284 | unsigned int fullness:FULLNESS_BITS; | |
285 | unsigned int class:CLASS_BITS; | |
48b4800a MK |
286 | unsigned int isolated:ISOLATED_BITS; |
287 | unsigned int magic:MAGIC_VAL_BITS; | |
3783689a MK |
288 | }; |
289 | unsigned int inuse; | |
bfd093f5 | 290 | unsigned int freeobj; |
3783689a MK |
291 | struct page *first_page; |
292 | struct list_head list; /* fullness list */ | |
48b4800a MK |
293 | #ifdef CONFIG_COMPACTION |
294 | rwlock_t lock; | |
295 | #endif | |
3783689a | 296 | }; |
61989a80 | 297 | |
f553646a | 298 | struct mapping_area { |
1b945aee | 299 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
300 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
301 | #else | |
302 | char *vm_buf; /* copy buffer for objects that span pages */ | |
303 | #endif | |
304 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
305 | enum zs_mapmode vm_mm; /* mapping mode */ | |
306 | }; | |
307 | ||
48b4800a MK |
308 | #ifdef CONFIG_COMPACTION |
309 | static int zs_register_migration(struct zs_pool *pool); | |
310 | static void zs_unregister_migration(struct zs_pool *pool); | |
311 | static void migrate_lock_init(struct zspage *zspage); | |
312 | static void migrate_read_lock(struct zspage *zspage); | |
313 | static void migrate_read_unlock(struct zspage *zspage); | |
314 | static void kick_deferred_free(struct zs_pool *pool); | |
315 | static void init_deferred_free(struct zs_pool *pool); | |
316 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage); | |
317 | #else | |
318 | static int zsmalloc_mount(void) { return 0; } | |
319 | static void zsmalloc_unmount(void) {} | |
320 | static int zs_register_migration(struct zs_pool *pool) { return 0; } | |
321 | static void zs_unregister_migration(struct zs_pool *pool) {} | |
322 | static void migrate_lock_init(struct zspage *zspage) {} | |
323 | static void migrate_read_lock(struct zspage *zspage) {} | |
324 | static void migrate_read_unlock(struct zspage *zspage) {} | |
325 | static void kick_deferred_free(struct zs_pool *pool) {} | |
326 | static void init_deferred_free(struct zs_pool *pool) {} | |
327 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {} | |
328 | #endif | |
329 | ||
3783689a | 330 | static int create_cache(struct zs_pool *pool) |
2e40e163 MK |
331 | { |
332 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
333 | 0, 0, NULL); | |
3783689a MK |
334 | if (!pool->handle_cachep) |
335 | return 1; | |
336 | ||
337 | pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage), | |
338 | 0, 0, NULL); | |
339 | if (!pool->zspage_cachep) { | |
340 | kmem_cache_destroy(pool->handle_cachep); | |
341 | pool->handle_cachep = NULL; | |
342 | return 1; | |
343 | } | |
344 | ||
345 | return 0; | |
2e40e163 MK |
346 | } |
347 | ||
3783689a | 348 | static void destroy_cache(struct zs_pool *pool) |
2e40e163 | 349 | { |
cd10add0 | 350 | kmem_cache_destroy(pool->handle_cachep); |
3783689a | 351 | kmem_cache_destroy(pool->zspage_cachep); |
2e40e163 MK |
352 | } |
353 | ||
3783689a | 354 | static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
355 | { |
356 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
48b4800a | 357 | gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); |
2e40e163 MK |
358 | } |
359 | ||
3783689a | 360 | static void cache_free_handle(struct zs_pool *pool, unsigned long handle) |
2e40e163 MK |
361 | { |
362 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
363 | } | |
364 | ||
3783689a MK |
365 | static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags) |
366 | { | |
48b4800a MK |
367 | return kmem_cache_alloc(pool->zspage_cachep, |
368 | flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE)); | |
3783689a MK |
369 | }; |
370 | ||
371 | static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage) | |
372 | { | |
373 | kmem_cache_free(pool->zspage_cachep, zspage); | |
374 | } | |
375 | ||
2e40e163 MK |
376 | static void record_obj(unsigned long handle, unsigned long obj) |
377 | { | |
c102f07c JL |
378 | /* |
379 | * lsb of @obj represents handle lock while other bits | |
380 | * represent object value the handle is pointing so | |
381 | * updating shouldn't do store tearing. | |
382 | */ | |
383 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
384 | } |
385 | ||
c795779d DS |
386 | /* zpool driver */ |
387 | ||
388 | #ifdef CONFIG_ZPOOL | |
389 | ||
6f3526d6 | 390 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 391 | const struct zpool_ops *zpool_ops, |
479305fd | 392 | struct zpool *zpool) |
c795779d | 393 | { |
d0d8da2d SS |
394 | /* |
395 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
396 | * different contexts and its caller must provide a valid | |
397 | * gfp mask. | |
398 | */ | |
399 | return zs_create_pool(name); | |
c795779d DS |
400 | } |
401 | ||
402 | static void zs_zpool_destroy(void *pool) | |
403 | { | |
404 | zs_destroy_pool(pool); | |
405 | } | |
406 | ||
407 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
408 | unsigned long *handle) | |
409 | { | |
d0d8da2d | 410 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
411 | return *handle ? 0 : -1; |
412 | } | |
413 | static void zs_zpool_free(void *pool, unsigned long handle) | |
414 | { | |
415 | zs_free(pool, handle); | |
416 | } | |
417 | ||
418 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
419 | unsigned int *reclaimed) | |
420 | { | |
421 | return -EINVAL; | |
422 | } | |
423 | ||
424 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
425 | enum zpool_mapmode mm) | |
426 | { | |
427 | enum zs_mapmode zs_mm; | |
428 | ||
429 | switch (mm) { | |
430 | case ZPOOL_MM_RO: | |
431 | zs_mm = ZS_MM_RO; | |
432 | break; | |
433 | case ZPOOL_MM_WO: | |
434 | zs_mm = ZS_MM_WO; | |
435 | break; | |
436 | case ZPOOL_MM_RW: /* fallthru */ | |
437 | default: | |
438 | zs_mm = ZS_MM_RW; | |
439 | break; | |
440 | } | |
441 | ||
442 | return zs_map_object(pool, handle, zs_mm); | |
443 | } | |
444 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
445 | { | |
446 | zs_unmap_object(pool, handle); | |
447 | } | |
448 | ||
449 | static u64 zs_zpool_total_size(void *pool) | |
450 | { | |
722cdc17 | 451 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
452 | } |
453 | ||
454 | static struct zpool_driver zs_zpool_driver = { | |
455 | .type = "zsmalloc", | |
456 | .owner = THIS_MODULE, | |
457 | .create = zs_zpool_create, | |
458 | .destroy = zs_zpool_destroy, | |
459 | .malloc = zs_zpool_malloc, | |
460 | .free = zs_zpool_free, | |
461 | .shrink = zs_zpool_shrink, | |
462 | .map = zs_zpool_map, | |
463 | .unmap = zs_zpool_unmap, | |
464 | .total_size = zs_zpool_total_size, | |
465 | }; | |
466 | ||
137f8cff | 467 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
468 | #endif /* CONFIG_ZPOOL */ |
469 | ||
248ca1b0 MK |
470 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
471 | { | |
472 | return pages_per_zspage * PAGE_SIZE / size; | |
473 | } | |
474 | ||
61989a80 NG |
475 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
476 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
477 | ||
48b4800a MK |
478 | static bool is_zspage_isolated(struct zspage *zspage) |
479 | { | |
480 | return zspage->isolated; | |
481 | } | |
482 | ||
61989a80 NG |
483 | static int is_first_page(struct page *page) |
484 | { | |
a27545bf | 485 | return PagePrivate(page); |
61989a80 NG |
486 | } |
487 | ||
48b4800a | 488 | /* Protected by class->lock */ |
3783689a | 489 | static inline int get_zspage_inuse(struct zspage *zspage) |
4f42047b | 490 | { |
3783689a | 491 | return zspage->inuse; |
4f42047b MK |
492 | } |
493 | ||
3783689a | 494 | static inline void set_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 495 | { |
3783689a | 496 | zspage->inuse = val; |
4f42047b MK |
497 | } |
498 | ||
3783689a | 499 | static inline void mod_zspage_inuse(struct zspage *zspage, int val) |
4f42047b | 500 | { |
3783689a | 501 | zspage->inuse += val; |
4f42047b MK |
502 | } |
503 | ||
48b4800a | 504 | static inline struct page *get_first_page(struct zspage *zspage) |
4f42047b | 505 | { |
48b4800a | 506 | struct page *first_page = zspage->first_page; |
3783689a | 507 | |
48b4800a MK |
508 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
509 | return first_page; | |
4f42047b MK |
510 | } |
511 | ||
48b4800a | 512 | static inline int get_first_obj_offset(struct page *page) |
4f42047b | 513 | { |
48b4800a MK |
514 | return page->units; |
515 | } | |
3783689a | 516 | |
48b4800a MK |
517 | static inline void set_first_obj_offset(struct page *page, int offset) |
518 | { | |
519 | page->units = offset; | |
4f42047b MK |
520 | } |
521 | ||
bfd093f5 | 522 | static inline unsigned int get_freeobj(struct zspage *zspage) |
4f42047b | 523 | { |
bfd093f5 | 524 | return zspage->freeobj; |
4f42047b MK |
525 | } |
526 | ||
bfd093f5 | 527 | static inline void set_freeobj(struct zspage *zspage, unsigned int obj) |
4f42047b | 528 | { |
bfd093f5 | 529 | zspage->freeobj = obj; |
4f42047b MK |
530 | } |
531 | ||
3783689a | 532 | static void get_zspage_mapping(struct zspage *zspage, |
a4209467 | 533 | unsigned int *class_idx, |
61989a80 NG |
534 | enum fullness_group *fullness) |
535 | { | |
48b4800a MK |
536 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); |
537 | ||
3783689a MK |
538 | *fullness = zspage->fullness; |
539 | *class_idx = zspage->class; | |
61989a80 NG |
540 | } |
541 | ||
3783689a | 542 | static void set_zspage_mapping(struct zspage *zspage, |
a4209467 | 543 | unsigned int class_idx, |
61989a80 NG |
544 | enum fullness_group fullness) |
545 | { | |
3783689a MK |
546 | zspage->class = class_idx; |
547 | zspage->fullness = fullness; | |
61989a80 NG |
548 | } |
549 | ||
c3e3e88a NC |
550 | /* |
551 | * zsmalloc divides the pool into various size classes where each | |
552 | * class maintains a list of zspages where each zspage is divided | |
553 | * into equal sized chunks. Each allocation falls into one of these | |
554 | * classes depending on its size. This function returns index of the | |
555 | * size class which has chunk size big enough to hold the give size. | |
556 | */ | |
61989a80 NG |
557 | static int get_size_class_index(int size) |
558 | { | |
559 | int idx = 0; | |
560 | ||
561 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
562 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
563 | ZS_SIZE_CLASS_DELTA); | |
564 | ||
7b60a685 | 565 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
566 | } |
567 | ||
248ca1b0 MK |
568 | static inline void zs_stat_inc(struct size_class *class, |
569 | enum zs_stat_type type, unsigned long cnt) | |
570 | { | |
48b4800a | 571 | class->stats.objs[type] += cnt; |
248ca1b0 MK |
572 | } |
573 | ||
574 | static inline void zs_stat_dec(struct size_class *class, | |
575 | enum zs_stat_type type, unsigned long cnt) | |
576 | { | |
48b4800a | 577 | class->stats.objs[type] -= cnt; |
248ca1b0 MK |
578 | } |
579 | ||
580 | static inline unsigned long zs_stat_get(struct size_class *class, | |
581 | enum zs_stat_type type) | |
582 | { | |
48b4800a | 583 | return class->stats.objs[type]; |
248ca1b0 MK |
584 | } |
585 | ||
57244594 SS |
586 | #ifdef CONFIG_ZSMALLOC_STAT |
587 | ||
4abaac9b | 588 | static void __init zs_stat_init(void) |
248ca1b0 | 589 | { |
4abaac9b DS |
590 | if (!debugfs_initialized()) { |
591 | pr_warn("debugfs not available, stat dir not created\n"); | |
592 | return; | |
593 | } | |
248ca1b0 MK |
594 | |
595 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
596 | if (!zs_stat_root) | |
4abaac9b | 597 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
598 | } |
599 | ||
600 | static void __exit zs_stat_exit(void) | |
601 | { | |
602 | debugfs_remove_recursive(zs_stat_root); | |
603 | } | |
604 | ||
1120ed54 SS |
605 | static unsigned long zs_can_compact(struct size_class *class); |
606 | ||
248ca1b0 MK |
607 | static int zs_stats_size_show(struct seq_file *s, void *v) |
608 | { | |
609 | int i; | |
610 | struct zs_pool *pool = s->private; | |
611 | struct size_class *class; | |
612 | int objs_per_zspage; | |
613 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 614 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
615 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
616 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 617 | unsigned long total_freeable = 0; |
248ca1b0 | 618 | |
1120ed54 | 619 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
620 | "class", "size", "almost_full", "almost_empty", |
621 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 622 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
623 | |
624 | for (i = 0; i < zs_size_classes; i++) { | |
625 | class = pool->size_class[i]; | |
626 | ||
627 | if (class->index != i) | |
628 | continue; | |
629 | ||
630 | spin_lock(&class->lock); | |
631 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
632 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
633 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
634 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 635 | freeable = zs_can_compact(class); |
248ca1b0 MK |
636 | spin_unlock(&class->lock); |
637 | ||
b4fd07a0 | 638 | objs_per_zspage = class->objs_per_zspage; |
248ca1b0 MK |
639 | pages_used = obj_allocated / objs_per_zspage * |
640 | class->pages_per_zspage; | |
641 | ||
1120ed54 SS |
642 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
643 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
644 | i, class->size, class_almost_full, class_almost_empty, |
645 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 646 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
647 | |
648 | total_class_almost_full += class_almost_full; | |
649 | total_class_almost_empty += class_almost_empty; | |
650 | total_objs += obj_allocated; | |
651 | total_used_objs += obj_used; | |
652 | total_pages += pages_used; | |
1120ed54 | 653 | total_freeable += freeable; |
248ca1b0 MK |
654 | } |
655 | ||
656 | seq_puts(s, "\n"); | |
1120ed54 | 657 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
658 | "Total", "", total_class_almost_full, |
659 | total_class_almost_empty, total_objs, | |
1120ed54 | 660 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
661 | |
662 | return 0; | |
663 | } | |
664 | ||
665 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
666 | { | |
667 | return single_open(file, zs_stats_size_show, inode->i_private); | |
668 | } | |
669 | ||
670 | static const struct file_operations zs_stat_size_ops = { | |
671 | .open = zs_stats_size_open, | |
672 | .read = seq_read, | |
673 | .llseek = seq_lseek, | |
674 | .release = single_release, | |
675 | }; | |
676 | ||
d34f6157 | 677 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
678 | { |
679 | struct dentry *entry; | |
680 | ||
4abaac9b DS |
681 | if (!zs_stat_root) { |
682 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 683 | return; |
4abaac9b | 684 | } |
248ca1b0 MK |
685 | |
686 | entry = debugfs_create_dir(name, zs_stat_root); | |
687 | if (!entry) { | |
688 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 689 | return; |
248ca1b0 MK |
690 | } |
691 | pool->stat_dentry = entry; | |
692 | ||
693 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
694 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
695 | if (!entry) { | |
696 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
697 | name, "classes"); | |
4abaac9b DS |
698 | debugfs_remove_recursive(pool->stat_dentry); |
699 | pool->stat_dentry = NULL; | |
248ca1b0 | 700 | } |
248ca1b0 MK |
701 | } |
702 | ||
703 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
704 | { | |
705 | debugfs_remove_recursive(pool->stat_dentry); | |
706 | } | |
707 | ||
708 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 709 | static void __init zs_stat_init(void) |
248ca1b0 | 710 | { |
248ca1b0 MK |
711 | } |
712 | ||
713 | static void __exit zs_stat_exit(void) | |
714 | { | |
715 | } | |
716 | ||
d34f6157 | 717 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 718 | { |
248ca1b0 MK |
719 | } |
720 | ||
721 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
722 | { | |
723 | } | |
248ca1b0 MK |
724 | #endif |
725 | ||
48b4800a | 726 | |
c3e3e88a NC |
727 | /* |
728 | * For each size class, zspages are divided into different groups | |
729 | * depending on how "full" they are. This was done so that we could | |
730 | * easily find empty or nearly empty zspages when we try to shrink | |
731 | * the pool (not yet implemented). This function returns fullness | |
732 | * status of the given page. | |
733 | */ | |
1fc6e27d | 734 | static enum fullness_group get_fullness_group(struct size_class *class, |
3783689a | 735 | struct zspage *zspage) |
61989a80 | 736 | { |
1fc6e27d | 737 | int inuse, objs_per_zspage; |
61989a80 | 738 | enum fullness_group fg; |
830e4bc5 | 739 | |
3783689a | 740 | inuse = get_zspage_inuse(zspage); |
1fc6e27d | 741 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
742 | |
743 | if (inuse == 0) | |
744 | fg = ZS_EMPTY; | |
1fc6e27d | 745 | else if (inuse == objs_per_zspage) |
61989a80 | 746 | fg = ZS_FULL; |
1fc6e27d | 747 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
748 | fg = ZS_ALMOST_EMPTY; |
749 | else | |
750 | fg = ZS_ALMOST_FULL; | |
751 | ||
752 | return fg; | |
753 | } | |
754 | ||
c3e3e88a NC |
755 | /* |
756 | * Each size class maintains various freelists and zspages are assigned | |
757 | * to one of these freelists based on the number of live objects they | |
758 | * have. This functions inserts the given zspage into the freelist | |
759 | * identified by <class, fullness_group>. | |
760 | */ | |
251cbb95 | 761 | static void insert_zspage(struct size_class *class, |
3783689a MK |
762 | struct zspage *zspage, |
763 | enum fullness_group fullness) | |
61989a80 | 764 | { |
3783689a | 765 | struct zspage *head; |
61989a80 | 766 | |
48b4800a | 767 | zs_stat_inc(class, fullness, 1); |
3783689a MK |
768 | head = list_first_entry_or_null(&class->fullness_list[fullness], |
769 | struct zspage, list); | |
58f17117 | 770 | /* |
3783689a MK |
771 | * We want to see more ZS_FULL pages and less almost empty/full. |
772 | * Put pages with higher ->inuse first. | |
58f17117 | 773 | */ |
3783689a MK |
774 | if (head) { |
775 | if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) { | |
776 | list_add(&zspage->list, &head->list); | |
777 | return; | |
778 | } | |
779 | } | |
780 | list_add(&zspage->list, &class->fullness_list[fullness]); | |
61989a80 NG |
781 | } |
782 | ||
c3e3e88a NC |
783 | /* |
784 | * This function removes the given zspage from the freelist identified | |
785 | * by <class, fullness_group>. | |
786 | */ | |
251cbb95 | 787 | static void remove_zspage(struct size_class *class, |
3783689a MK |
788 | struct zspage *zspage, |
789 | enum fullness_group fullness) | |
61989a80 | 790 | { |
3783689a | 791 | VM_BUG_ON(list_empty(&class->fullness_list[fullness])); |
48b4800a | 792 | VM_BUG_ON(is_zspage_isolated(zspage)); |
61989a80 | 793 | |
3783689a | 794 | list_del_init(&zspage->list); |
48b4800a | 795 | zs_stat_dec(class, fullness, 1); |
61989a80 NG |
796 | } |
797 | ||
c3e3e88a NC |
798 | /* |
799 | * Each size class maintains zspages in different fullness groups depending | |
800 | * on the number of live objects they contain. When allocating or freeing | |
801 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
802 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
803 | * a status change has occurred for the given page and accordingly moves the | |
804 | * page from the freelist of the old fullness group to that of the new | |
805 | * fullness group. | |
806 | */ | |
c7806261 | 807 | static enum fullness_group fix_fullness_group(struct size_class *class, |
3783689a | 808 | struct zspage *zspage) |
61989a80 NG |
809 | { |
810 | int class_idx; | |
61989a80 NG |
811 | enum fullness_group currfg, newfg; |
812 | ||
3783689a MK |
813 | get_zspage_mapping(zspage, &class_idx, &currfg); |
814 | newfg = get_fullness_group(class, zspage); | |
61989a80 NG |
815 | if (newfg == currfg) |
816 | goto out; | |
817 | ||
48b4800a MK |
818 | if (!is_zspage_isolated(zspage)) { |
819 | remove_zspage(class, zspage, currfg); | |
820 | insert_zspage(class, zspage, newfg); | |
821 | } | |
822 | ||
3783689a | 823 | set_zspage_mapping(zspage, class_idx, newfg); |
61989a80 NG |
824 | |
825 | out: | |
826 | return newfg; | |
827 | } | |
828 | ||
829 | /* | |
830 | * We have to decide on how many pages to link together | |
831 | * to form a zspage for each size class. This is important | |
832 | * to reduce wastage due to unusable space left at end of | |
833 | * each zspage which is given as: | |
888fa374 YX |
834 | * wastage = Zp % class_size |
835 | * usage = Zp - wastage | |
61989a80 NG |
836 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
837 | * | |
838 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
839 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
840 | * since then we can perfectly fit in 8 such objects. | |
841 | */ | |
2e3b6154 | 842 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
843 | { |
844 | int i, max_usedpc = 0; | |
845 | /* zspage order which gives maximum used size per KB */ | |
846 | int max_usedpc_order = 1; | |
847 | ||
84d4faab | 848 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
849 | int zspage_size; |
850 | int waste, usedpc; | |
851 | ||
852 | zspage_size = i * PAGE_SIZE; | |
853 | waste = zspage_size % class_size; | |
854 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
855 | ||
856 | if (usedpc > max_usedpc) { | |
857 | max_usedpc = usedpc; | |
858 | max_usedpc_order = i; | |
859 | } | |
860 | } | |
861 | ||
862 | return max_usedpc_order; | |
863 | } | |
864 | ||
3783689a | 865 | static struct zspage *get_zspage(struct page *page) |
61989a80 | 866 | { |
48b4800a MK |
867 | struct zspage *zspage = (struct zspage *)page->private; |
868 | ||
869 | BUG_ON(zspage->magic != ZSPAGE_MAGIC); | |
870 | return zspage; | |
61989a80 NG |
871 | } |
872 | ||
873 | static struct page *get_next_page(struct page *page) | |
874 | { | |
48b4800a MK |
875 | if (unlikely(PageHugeObject(page))) |
876 | return NULL; | |
877 | ||
878 | return page->freelist; | |
61989a80 NG |
879 | } |
880 | ||
bfd093f5 MK |
881 | /** |
882 | * obj_to_location - get (<page>, <obj_idx>) from encoded object value | |
883 | * @page: page object resides in zspage | |
884 | * @obj_idx: object index | |
67296874 | 885 | */ |
bfd093f5 MK |
886 | static void obj_to_location(unsigned long obj, struct page **page, |
887 | unsigned int *obj_idx) | |
61989a80 | 888 | { |
bfd093f5 MK |
889 | obj >>= OBJ_TAG_BITS; |
890 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
891 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
892 | } | |
61989a80 | 893 | |
bfd093f5 MK |
894 | /** |
895 | * location_to_obj - get obj value encoded from (<page>, <obj_idx>) | |
896 | * @page: page object resides in zspage | |
897 | * @obj_idx: object index | |
898 | */ | |
899 | static unsigned long location_to_obj(struct page *page, unsigned int obj_idx) | |
900 | { | |
901 | unsigned long obj; | |
61989a80 | 902 | |
312fcae2 | 903 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
bfd093f5 | 904 | obj |= obj_idx & OBJ_INDEX_MASK; |
312fcae2 | 905 | obj <<= OBJ_TAG_BITS; |
61989a80 | 906 | |
bfd093f5 | 907 | return obj; |
61989a80 NG |
908 | } |
909 | ||
2e40e163 MK |
910 | static unsigned long handle_to_obj(unsigned long handle) |
911 | { | |
912 | return *(unsigned long *)handle; | |
913 | } | |
914 | ||
48b4800a | 915 | static unsigned long obj_to_head(struct page *page, void *obj) |
312fcae2 | 916 | { |
48b4800a | 917 | if (unlikely(PageHugeObject(page))) { |
830e4bc5 | 918 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
3783689a | 919 | return page->index; |
7b60a685 MK |
920 | } else |
921 | return *(unsigned long *)obj; | |
312fcae2 MK |
922 | } |
923 | ||
48b4800a MK |
924 | static inline int testpin_tag(unsigned long handle) |
925 | { | |
926 | return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle); | |
927 | } | |
928 | ||
312fcae2 MK |
929 | static inline int trypin_tag(unsigned long handle) |
930 | { | |
1b8320b6 | 931 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
932 | } |
933 | ||
934 | static void pin_tag(unsigned long handle) | |
935 | { | |
1b8320b6 | 936 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
937 | } |
938 | ||
939 | static void unpin_tag(unsigned long handle) | |
940 | { | |
1b8320b6 | 941 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
942 | } |
943 | ||
f4477e90 NG |
944 | static void reset_page(struct page *page) |
945 | { | |
48b4800a | 946 | __ClearPageMovable(page); |
f4477e90 NG |
947 | clear_bit(PG_private, &page->flags); |
948 | clear_bit(PG_private_2, &page->flags); | |
949 | set_page_private(page, 0); | |
48b4800a MK |
950 | page_mapcount_reset(page); |
951 | ClearPageHugeObject(page); | |
952 | page->freelist = NULL; | |
953 | } | |
954 | ||
955 | /* | |
956 | * To prevent zspage destroy during migration, zspage freeing should | |
957 | * hold locks of all pages in the zspage. | |
958 | */ | |
959 | void lock_zspage(struct zspage *zspage) | |
960 | { | |
961 | struct page *page = get_first_page(zspage); | |
962 | ||
963 | do { | |
964 | lock_page(page); | |
965 | } while ((page = get_next_page(page)) != NULL); | |
966 | } | |
967 | ||
968 | int trylock_zspage(struct zspage *zspage) | |
969 | { | |
970 | struct page *cursor, *fail; | |
971 | ||
972 | for (cursor = get_first_page(zspage); cursor != NULL; cursor = | |
973 | get_next_page(cursor)) { | |
974 | if (!trylock_page(cursor)) { | |
975 | fail = cursor; | |
976 | goto unlock; | |
977 | } | |
978 | } | |
979 | ||
980 | return 1; | |
981 | unlock: | |
982 | for (cursor = get_first_page(zspage); cursor != fail; cursor = | |
983 | get_next_page(cursor)) | |
984 | unlock_page(cursor); | |
985 | ||
986 | return 0; | |
f4477e90 NG |
987 | } |
988 | ||
48b4800a MK |
989 | static void __free_zspage(struct zs_pool *pool, struct size_class *class, |
990 | struct zspage *zspage) | |
61989a80 | 991 | { |
3783689a | 992 | struct page *page, *next; |
48b4800a MK |
993 | enum fullness_group fg; |
994 | unsigned int class_idx; | |
995 | ||
996 | get_zspage_mapping(zspage, &class_idx, &fg); | |
997 | ||
998 | assert_spin_locked(&class->lock); | |
61989a80 | 999 | |
3783689a | 1000 | VM_BUG_ON(get_zspage_inuse(zspage)); |
48b4800a | 1001 | VM_BUG_ON(fg != ZS_EMPTY); |
61989a80 | 1002 | |
48b4800a | 1003 | next = page = get_first_page(zspage); |
3783689a | 1004 | do { |
48b4800a MK |
1005 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
1006 | next = get_next_page(page); | |
3783689a | 1007 | reset_page(page); |
48b4800a | 1008 | unlock_page(page); |
91537fee | 1009 | dec_zone_page_state(page, NR_ZSPAGES); |
3783689a MK |
1010 | put_page(page); |
1011 | page = next; | |
1012 | } while (page != NULL); | |
61989a80 | 1013 | |
3783689a | 1014 | cache_free_zspage(pool, zspage); |
48b4800a | 1015 | |
b4fd07a0 | 1016 | zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
1017 | atomic_long_sub(class->pages_per_zspage, |
1018 | &pool->pages_allocated); | |
1019 | } | |
1020 | ||
1021 | static void free_zspage(struct zs_pool *pool, struct size_class *class, | |
1022 | struct zspage *zspage) | |
1023 | { | |
1024 | VM_BUG_ON(get_zspage_inuse(zspage)); | |
1025 | VM_BUG_ON(list_empty(&zspage->list)); | |
1026 | ||
1027 | if (!trylock_zspage(zspage)) { | |
1028 | kick_deferred_free(pool); | |
1029 | return; | |
1030 | } | |
1031 | ||
1032 | remove_zspage(class, zspage, ZS_EMPTY); | |
1033 | __free_zspage(pool, class, zspage); | |
61989a80 NG |
1034 | } |
1035 | ||
1036 | /* Initialize a newly allocated zspage */ | |
3783689a | 1037 | static void init_zspage(struct size_class *class, struct zspage *zspage) |
61989a80 | 1038 | { |
bfd093f5 | 1039 | unsigned int freeobj = 1; |
61989a80 | 1040 | unsigned long off = 0; |
48b4800a | 1041 | struct page *page = get_first_page(zspage); |
830e4bc5 | 1042 | |
61989a80 NG |
1043 | while (page) { |
1044 | struct page *next_page; | |
1045 | struct link_free *link; | |
af4ee5e9 | 1046 | void *vaddr; |
61989a80 | 1047 | |
3783689a | 1048 | set_first_obj_offset(page, off); |
61989a80 | 1049 | |
af4ee5e9 MK |
1050 | vaddr = kmap_atomic(page); |
1051 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
1052 | |
1053 | while ((off += class->size) < PAGE_SIZE) { | |
3b1d9ca6 | 1054 | link->next = freeobj++ << OBJ_TAG_BITS; |
5538c562 | 1055 | link += class->size / sizeof(*link); |
61989a80 NG |
1056 | } |
1057 | ||
1058 | /* | |
1059 | * We now come to the last (full or partial) object on this | |
1060 | * page, which must point to the first object on the next | |
1061 | * page (if present) | |
1062 | */ | |
1063 | next_page = get_next_page(page); | |
bfd093f5 | 1064 | if (next_page) { |
3b1d9ca6 | 1065 | link->next = freeobj++ << OBJ_TAG_BITS; |
bfd093f5 MK |
1066 | } else { |
1067 | /* | |
3b1d9ca6 | 1068 | * Reset OBJ_TAG_BITS bit to last link to tell |
bfd093f5 MK |
1069 | * whether it's allocated object or not. |
1070 | */ | |
3b1d9ca6 | 1071 | link->next = -1 << OBJ_TAG_BITS; |
bfd093f5 | 1072 | } |
af4ee5e9 | 1073 | kunmap_atomic(vaddr); |
61989a80 | 1074 | page = next_page; |
5538c562 | 1075 | off %= PAGE_SIZE; |
61989a80 | 1076 | } |
bdb0af7c | 1077 | |
bfd093f5 | 1078 | set_freeobj(zspage, 0); |
61989a80 NG |
1079 | } |
1080 | ||
48b4800a MK |
1081 | static void create_page_chain(struct size_class *class, struct zspage *zspage, |
1082 | struct page *pages[]) | |
61989a80 | 1083 | { |
bdb0af7c MK |
1084 | int i; |
1085 | struct page *page; | |
1086 | struct page *prev_page = NULL; | |
48b4800a | 1087 | int nr_pages = class->pages_per_zspage; |
61989a80 NG |
1088 | |
1089 | /* | |
1090 | * Allocate individual pages and link them together as: | |
48b4800a | 1091 | * 1. all pages are linked together using page->freelist |
3783689a | 1092 | * 2. each sub-page point to zspage using page->private |
61989a80 | 1093 | * |
3783689a MK |
1094 | * we set PG_private to identify the first page (i.e. no other sub-page |
1095 | * has this flag set) and PG_private_2 to identify the last page. | |
61989a80 | 1096 | */ |
bdb0af7c MK |
1097 | for (i = 0; i < nr_pages; i++) { |
1098 | page = pages[i]; | |
3783689a | 1099 | set_page_private(page, (unsigned long)zspage); |
48b4800a | 1100 | page->freelist = NULL; |
bdb0af7c | 1101 | if (i == 0) { |
3783689a | 1102 | zspage->first_page = page; |
a27545bf | 1103 | SetPagePrivate(page); |
48b4800a MK |
1104 | if (unlikely(class->objs_per_zspage == 1 && |
1105 | class->pages_per_zspage == 1)) | |
1106 | SetPageHugeObject(page); | |
3783689a | 1107 | } else { |
48b4800a | 1108 | prev_page->freelist = page; |
61989a80 | 1109 | } |
48b4800a | 1110 | if (i == nr_pages - 1) |
a27545bf | 1111 | SetPagePrivate2(page); |
61989a80 NG |
1112 | prev_page = page; |
1113 | } | |
bdb0af7c | 1114 | } |
61989a80 | 1115 | |
bdb0af7c MK |
1116 | /* |
1117 | * Allocate a zspage for the given size class | |
1118 | */ | |
3783689a MK |
1119 | static struct zspage *alloc_zspage(struct zs_pool *pool, |
1120 | struct size_class *class, | |
1121 | gfp_t gfp) | |
bdb0af7c MK |
1122 | { |
1123 | int i; | |
bdb0af7c | 1124 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE]; |
3783689a MK |
1125 | struct zspage *zspage = cache_alloc_zspage(pool, gfp); |
1126 | ||
1127 | if (!zspage) | |
1128 | return NULL; | |
1129 | ||
1130 | memset(zspage, 0, sizeof(struct zspage)); | |
48b4800a MK |
1131 | zspage->magic = ZSPAGE_MAGIC; |
1132 | migrate_lock_init(zspage); | |
61989a80 | 1133 | |
bdb0af7c MK |
1134 | for (i = 0; i < class->pages_per_zspage; i++) { |
1135 | struct page *page; | |
61989a80 | 1136 | |
3783689a | 1137 | page = alloc_page(gfp); |
bdb0af7c | 1138 | if (!page) { |
91537fee MK |
1139 | while (--i >= 0) { |
1140 | dec_zone_page_state(pages[i], NR_ZSPAGES); | |
bdb0af7c | 1141 | __free_page(pages[i]); |
91537fee | 1142 | } |
3783689a | 1143 | cache_free_zspage(pool, zspage); |
bdb0af7c MK |
1144 | return NULL; |
1145 | } | |
91537fee MK |
1146 | |
1147 | inc_zone_page_state(page, NR_ZSPAGES); | |
bdb0af7c | 1148 | pages[i] = page; |
61989a80 NG |
1149 | } |
1150 | ||
48b4800a | 1151 | create_page_chain(class, zspage, pages); |
3783689a | 1152 | init_zspage(class, zspage); |
bdb0af7c | 1153 | |
3783689a | 1154 | return zspage; |
61989a80 NG |
1155 | } |
1156 | ||
3783689a | 1157 | static struct zspage *find_get_zspage(struct size_class *class) |
61989a80 NG |
1158 | { |
1159 | int i; | |
3783689a | 1160 | struct zspage *zspage; |
61989a80 | 1161 | |
48b4800a | 1162 | for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) { |
3783689a MK |
1163 | zspage = list_first_entry_or_null(&class->fullness_list[i], |
1164 | struct zspage, list); | |
1165 | if (zspage) | |
61989a80 NG |
1166 | break; |
1167 | } | |
1168 | ||
3783689a | 1169 | return zspage; |
61989a80 NG |
1170 | } |
1171 | ||
1b945aee | 1172 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1173 | static inline int __zs_cpu_up(struct mapping_area *area) |
1174 | { | |
1175 | /* | |
1176 | * Make sure we don't leak memory if a cpu UP notification | |
1177 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1178 | */ | |
1179 | if (area->vm) | |
1180 | return 0; | |
1181 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1182 | if (!area->vm) | |
1183 | return -ENOMEM; | |
1184 | return 0; | |
1185 | } | |
1186 | ||
1187 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1188 | { | |
1189 | if (area->vm) | |
1190 | free_vm_area(area->vm); | |
1191 | area->vm = NULL; | |
1192 | } | |
1193 | ||
1194 | static inline void *__zs_map_object(struct mapping_area *area, | |
1195 | struct page *pages[2], int off, int size) | |
1196 | { | |
f6f8ed47 | 1197 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1198 | area->vm_addr = area->vm->addr; |
1199 | return area->vm_addr + off; | |
1200 | } | |
1201 | ||
1202 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1203 | struct page *pages[2], int off, int size) | |
1204 | { | |
1205 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1206 | |
d95abbbb | 1207 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1208 | } |
1209 | ||
1b945aee | 1210 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1211 | |
1212 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1213 | { | |
1214 | /* | |
1215 | * Make sure we don't leak memory if a cpu UP notification | |
1216 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1217 | */ | |
1218 | if (area->vm_buf) | |
1219 | return 0; | |
40f9fb8c | 1220 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1221 | if (!area->vm_buf) |
1222 | return -ENOMEM; | |
1223 | return 0; | |
1224 | } | |
1225 | ||
1226 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1227 | { | |
40f9fb8c | 1228 | kfree(area->vm_buf); |
f553646a SJ |
1229 | area->vm_buf = NULL; |
1230 | } | |
1231 | ||
1232 | static void *__zs_map_object(struct mapping_area *area, | |
1233 | struct page *pages[2], int off, int size) | |
5f601902 | 1234 | { |
5f601902 SJ |
1235 | int sizes[2]; |
1236 | void *addr; | |
f553646a | 1237 | char *buf = area->vm_buf; |
5f601902 | 1238 | |
f553646a SJ |
1239 | /* disable page faults to match kmap_atomic() return conditions */ |
1240 | pagefault_disable(); | |
1241 | ||
1242 | /* no read fastpath */ | |
1243 | if (area->vm_mm == ZS_MM_WO) | |
1244 | goto out; | |
5f601902 SJ |
1245 | |
1246 | sizes[0] = PAGE_SIZE - off; | |
1247 | sizes[1] = size - sizes[0]; | |
1248 | ||
5f601902 SJ |
1249 | /* copy object to per-cpu buffer */ |
1250 | addr = kmap_atomic(pages[0]); | |
1251 | memcpy(buf, addr + off, sizes[0]); | |
1252 | kunmap_atomic(addr); | |
1253 | addr = kmap_atomic(pages[1]); | |
1254 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1255 | kunmap_atomic(addr); | |
f553646a SJ |
1256 | out: |
1257 | return area->vm_buf; | |
5f601902 SJ |
1258 | } |
1259 | ||
f553646a SJ |
1260 | static void __zs_unmap_object(struct mapping_area *area, |
1261 | struct page *pages[2], int off, int size) | |
5f601902 | 1262 | { |
5f601902 SJ |
1263 | int sizes[2]; |
1264 | void *addr; | |
2e40e163 | 1265 | char *buf; |
5f601902 | 1266 | |
f553646a SJ |
1267 | /* no write fastpath */ |
1268 | if (area->vm_mm == ZS_MM_RO) | |
1269 | goto out; | |
5f601902 | 1270 | |
7b60a685 | 1271 | buf = area->vm_buf; |
a82cbf07 YX |
1272 | buf = buf + ZS_HANDLE_SIZE; |
1273 | size -= ZS_HANDLE_SIZE; | |
1274 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1275 | |
5f601902 SJ |
1276 | sizes[0] = PAGE_SIZE - off; |
1277 | sizes[1] = size - sizes[0]; | |
1278 | ||
1279 | /* copy per-cpu buffer to object */ | |
1280 | addr = kmap_atomic(pages[0]); | |
1281 | memcpy(addr + off, buf, sizes[0]); | |
1282 | kunmap_atomic(addr); | |
1283 | addr = kmap_atomic(pages[1]); | |
1284 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1285 | kunmap_atomic(addr); | |
f553646a SJ |
1286 | |
1287 | out: | |
1288 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1289 | pagefault_enable(); | |
5f601902 | 1290 | } |
61989a80 | 1291 | |
1b945aee | 1292 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1293 | |
61989a80 NG |
1294 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1295 | void *pcpu) | |
1296 | { | |
f553646a | 1297 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1298 | struct mapping_area *area; |
1299 | ||
1300 | switch (action) { | |
1301 | case CPU_UP_PREPARE: | |
1302 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1303 | ret = __zs_cpu_up(area); |
1304 | if (ret) | |
1305 | return notifier_from_errno(ret); | |
61989a80 NG |
1306 | break; |
1307 | case CPU_DEAD: | |
1308 | case CPU_UP_CANCELED: | |
1309 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1310 | __zs_cpu_down(area); |
61989a80 NG |
1311 | break; |
1312 | } | |
1313 | ||
1314 | return NOTIFY_OK; | |
1315 | } | |
1316 | ||
1317 | static struct notifier_block zs_cpu_nb = { | |
1318 | .notifier_call = zs_cpu_notifier | |
1319 | }; | |
1320 | ||
b1b00a5b | 1321 | static int zs_register_cpu_notifier(void) |
61989a80 | 1322 | { |
b1b00a5b | 1323 | int cpu, uninitialized_var(ret); |
61989a80 | 1324 | |
f0e71fcd SB |
1325 | cpu_notifier_register_begin(); |
1326 | ||
1327 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1328 | for_each_online_cpu(cpu) { |
1329 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1330 | if (notifier_to_errno(ret)) |
1331 | break; | |
61989a80 | 1332 | } |
f0e71fcd SB |
1333 | |
1334 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1335 | return notifier_to_errno(ret); |
1336 | } | |
f0e71fcd | 1337 | |
66cdef66 | 1338 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1339 | { |
66cdef66 | 1340 | int cpu; |
40f9fb8c | 1341 | |
66cdef66 | 1342 | cpu_notifier_register_begin(); |
40f9fb8c | 1343 | |
66cdef66 GM |
1344 | for_each_online_cpu(cpu) |
1345 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1346 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1347 | |
66cdef66 | 1348 | cpu_notifier_register_done(); |
b1b00a5b SS |
1349 | } |
1350 | ||
66cdef66 | 1351 | static void init_zs_size_classes(void) |
b1b00a5b | 1352 | { |
66cdef66 | 1353 | int nr; |
c795779d | 1354 | |
66cdef66 GM |
1355 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1356 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1357 | nr += 1; | |
40f9fb8c | 1358 | |
66cdef66 | 1359 | zs_size_classes = nr; |
61989a80 NG |
1360 | } |
1361 | ||
9eec4cd5 JK |
1362 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1363 | { | |
1364 | if (prev->pages_per_zspage != pages_per_zspage) | |
1365 | return false; | |
1366 | ||
b4fd07a0 | 1367 | if (prev->objs_per_zspage |
9eec4cd5 JK |
1368 | != get_maxobj_per_zspage(size, pages_per_zspage)) |
1369 | return false; | |
1370 | ||
1371 | return true; | |
1372 | } | |
1373 | ||
3783689a | 1374 | static bool zspage_full(struct size_class *class, struct zspage *zspage) |
312fcae2 | 1375 | { |
3783689a | 1376 | return get_zspage_inuse(zspage) == class->objs_per_zspage; |
312fcae2 MK |
1377 | } |
1378 | ||
66cdef66 GM |
1379 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1380 | { | |
1381 | return atomic_long_read(&pool->pages_allocated); | |
1382 | } | |
1383 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1384 | ||
4bbc0bc0 | 1385 | /** |
66cdef66 GM |
1386 | * zs_map_object - get address of allocated object from handle. |
1387 | * @pool: pool from which the object was allocated | |
1388 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1389 | * |
66cdef66 GM |
1390 | * Before using an object allocated from zs_malloc, it must be mapped using |
1391 | * this function. When done with the object, it must be unmapped using | |
1392 | * zs_unmap_object. | |
4bbc0bc0 | 1393 | * |
66cdef66 GM |
1394 | * Only one object can be mapped per cpu at a time. There is no protection |
1395 | * against nested mappings. | |
1396 | * | |
1397 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1398 | */ |
66cdef66 GM |
1399 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1400 | enum zs_mapmode mm) | |
61989a80 | 1401 | { |
3783689a | 1402 | struct zspage *zspage; |
66cdef66 | 1403 | struct page *page; |
bfd093f5 MK |
1404 | unsigned long obj, off; |
1405 | unsigned int obj_idx; | |
61989a80 | 1406 | |
66cdef66 GM |
1407 | unsigned int class_idx; |
1408 | enum fullness_group fg; | |
1409 | struct size_class *class; | |
1410 | struct mapping_area *area; | |
1411 | struct page *pages[2]; | |
2e40e163 | 1412 | void *ret; |
61989a80 | 1413 | |
9eec4cd5 | 1414 | /* |
66cdef66 GM |
1415 | * Because we use per-cpu mapping areas shared among the |
1416 | * pools/users, we can't allow mapping in interrupt context | |
1417 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1418 | */ |
830e4bc5 | 1419 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1420 | |
312fcae2 MK |
1421 | /* From now on, migration cannot move the object */ |
1422 | pin_tag(handle); | |
1423 | ||
2e40e163 MK |
1424 | obj = handle_to_obj(handle); |
1425 | obj_to_location(obj, &page, &obj_idx); | |
3783689a | 1426 | zspage = get_zspage(page); |
48b4800a MK |
1427 | |
1428 | /* migration cannot move any subpage in this zspage */ | |
1429 | migrate_read_lock(zspage); | |
1430 | ||
3783689a | 1431 | get_zspage_mapping(zspage, &class_idx, &fg); |
66cdef66 | 1432 | class = pool->size_class[class_idx]; |
bfd093f5 | 1433 | off = (class->size * obj_idx) & ~PAGE_MASK; |
df8b5bb9 | 1434 | |
66cdef66 GM |
1435 | area = &get_cpu_var(zs_map_area); |
1436 | area->vm_mm = mm; | |
1437 | if (off + class->size <= PAGE_SIZE) { | |
1438 | /* this object is contained entirely within a page */ | |
1439 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1440 | ret = area->vm_addr + off; |
1441 | goto out; | |
61989a80 NG |
1442 | } |
1443 | ||
66cdef66 GM |
1444 | /* this object spans two pages */ |
1445 | pages[0] = page; | |
1446 | pages[1] = get_next_page(page); | |
1447 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1448 | |
2e40e163 MK |
1449 | ret = __zs_map_object(area, pages, off, class->size); |
1450 | out: | |
48b4800a | 1451 | if (likely(!PageHugeObject(page))) |
7b60a685 MK |
1452 | ret += ZS_HANDLE_SIZE; |
1453 | ||
1454 | return ret; | |
61989a80 | 1455 | } |
66cdef66 | 1456 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1457 | |
66cdef66 | 1458 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1459 | { |
3783689a | 1460 | struct zspage *zspage; |
66cdef66 | 1461 | struct page *page; |
bfd093f5 MK |
1462 | unsigned long obj, off; |
1463 | unsigned int obj_idx; | |
61989a80 | 1464 | |
66cdef66 GM |
1465 | unsigned int class_idx; |
1466 | enum fullness_group fg; | |
1467 | struct size_class *class; | |
1468 | struct mapping_area *area; | |
9eec4cd5 | 1469 | |
2e40e163 MK |
1470 | obj = handle_to_obj(handle); |
1471 | obj_to_location(obj, &page, &obj_idx); | |
3783689a MK |
1472 | zspage = get_zspage(page); |
1473 | get_zspage_mapping(zspage, &class_idx, &fg); | |
66cdef66 | 1474 | class = pool->size_class[class_idx]; |
bfd093f5 | 1475 | off = (class->size * obj_idx) & ~PAGE_MASK; |
61989a80 | 1476 | |
66cdef66 GM |
1477 | area = this_cpu_ptr(&zs_map_area); |
1478 | if (off + class->size <= PAGE_SIZE) | |
1479 | kunmap_atomic(area->vm_addr); | |
1480 | else { | |
1481 | struct page *pages[2]; | |
40f9fb8c | 1482 | |
66cdef66 GM |
1483 | pages[0] = page; |
1484 | pages[1] = get_next_page(page); | |
1485 | BUG_ON(!pages[1]); | |
1486 | ||
1487 | __zs_unmap_object(area, pages, off, class->size); | |
1488 | } | |
1489 | put_cpu_var(zs_map_area); | |
48b4800a MK |
1490 | |
1491 | migrate_read_unlock(zspage); | |
312fcae2 | 1492 | unpin_tag(handle); |
61989a80 | 1493 | } |
66cdef66 | 1494 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1495 | |
251cbb95 | 1496 | static unsigned long obj_malloc(struct size_class *class, |
3783689a | 1497 | struct zspage *zspage, unsigned long handle) |
c7806261 | 1498 | { |
bfd093f5 | 1499 | int i, nr_page, offset; |
c7806261 MK |
1500 | unsigned long obj; |
1501 | struct link_free *link; | |
1502 | ||
1503 | struct page *m_page; | |
bfd093f5 | 1504 | unsigned long m_offset; |
c7806261 MK |
1505 | void *vaddr; |
1506 | ||
312fcae2 | 1507 | handle |= OBJ_ALLOCATED_TAG; |
3783689a | 1508 | obj = get_freeobj(zspage); |
bfd093f5 MK |
1509 | |
1510 | offset = obj * class->size; | |
1511 | nr_page = offset >> PAGE_SHIFT; | |
1512 | m_offset = offset & ~PAGE_MASK; | |
1513 | m_page = get_first_page(zspage); | |
1514 | ||
1515 | for (i = 0; i < nr_page; i++) | |
1516 | m_page = get_next_page(m_page); | |
c7806261 MK |
1517 | |
1518 | vaddr = kmap_atomic(m_page); | |
1519 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
3b1d9ca6 | 1520 | set_freeobj(zspage, link->next >> OBJ_TAG_BITS); |
48b4800a | 1521 | if (likely(!PageHugeObject(m_page))) |
7b60a685 MK |
1522 | /* record handle in the header of allocated chunk */ |
1523 | link->handle = handle; | |
1524 | else | |
3783689a MK |
1525 | /* record handle to page->index */ |
1526 | zspage->first_page->index = handle; | |
1527 | ||
c7806261 | 1528 | kunmap_atomic(vaddr); |
3783689a | 1529 | mod_zspage_inuse(zspage, 1); |
c7806261 MK |
1530 | zs_stat_inc(class, OBJ_USED, 1); |
1531 | ||
bfd093f5 MK |
1532 | obj = location_to_obj(m_page, obj); |
1533 | ||
c7806261 MK |
1534 | return obj; |
1535 | } | |
1536 | ||
1537 | ||
61989a80 NG |
1538 | /** |
1539 | * zs_malloc - Allocate block of given size from pool. | |
1540 | * @pool: pool to allocate from | |
1541 | * @size: size of block to allocate | |
61989a80 | 1542 | * |
00a61d86 | 1543 | * On success, handle to the allocated object is returned, |
c2344348 | 1544 | * otherwise 0. |
61989a80 NG |
1545 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1546 | */ | |
d0d8da2d | 1547 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1548 | { |
2e40e163 | 1549 | unsigned long handle, obj; |
61989a80 | 1550 | struct size_class *class; |
48b4800a | 1551 | enum fullness_group newfg; |
3783689a | 1552 | struct zspage *zspage; |
61989a80 | 1553 | |
7b60a685 | 1554 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1555 | return 0; |
1556 | ||
3783689a | 1557 | handle = cache_alloc_handle(pool, gfp); |
2e40e163 | 1558 | if (!handle) |
c2344348 | 1559 | return 0; |
61989a80 | 1560 | |
2e40e163 MK |
1561 | /* extra space in chunk to keep the handle */ |
1562 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1563 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1564 | |
1565 | spin_lock(&class->lock); | |
3783689a | 1566 | zspage = find_get_zspage(class); |
48b4800a MK |
1567 | if (likely(zspage)) { |
1568 | obj = obj_malloc(class, zspage, handle); | |
1569 | /* Now move the zspage to another fullness group, if required */ | |
1570 | fix_fullness_group(class, zspage); | |
1571 | record_obj(handle, obj); | |
61989a80 | 1572 | spin_unlock(&class->lock); |
61989a80 | 1573 | |
48b4800a MK |
1574 | return handle; |
1575 | } | |
0f050d99 | 1576 | |
48b4800a MK |
1577 | spin_unlock(&class->lock); |
1578 | ||
1579 | zspage = alloc_zspage(pool, class, gfp); | |
1580 | if (!zspage) { | |
1581 | cache_free_handle(pool, handle); | |
1582 | return 0; | |
61989a80 NG |
1583 | } |
1584 | ||
48b4800a | 1585 | spin_lock(&class->lock); |
3783689a | 1586 | obj = obj_malloc(class, zspage, handle); |
48b4800a MK |
1587 | newfg = get_fullness_group(class, zspage); |
1588 | insert_zspage(class, zspage, newfg); | |
1589 | set_zspage_mapping(zspage, class->index, newfg); | |
2e40e163 | 1590 | record_obj(handle, obj); |
48b4800a MK |
1591 | atomic_long_add(class->pages_per_zspage, |
1592 | &pool->pages_allocated); | |
b4fd07a0 | 1593 | zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage); |
48b4800a MK |
1594 | |
1595 | /* We completely set up zspage so mark them as movable */ | |
1596 | SetZsPageMovable(pool, zspage); | |
61989a80 NG |
1597 | spin_unlock(&class->lock); |
1598 | ||
2e40e163 | 1599 | return handle; |
61989a80 NG |
1600 | } |
1601 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1602 | ||
1ee47165 | 1603 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1604 | { |
1605 | struct link_free *link; | |
3783689a MK |
1606 | struct zspage *zspage; |
1607 | struct page *f_page; | |
bfd093f5 MK |
1608 | unsigned long f_offset; |
1609 | unsigned int f_objidx; | |
af4ee5e9 | 1610 | void *vaddr; |
61989a80 | 1611 | |
312fcae2 | 1612 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1613 | obj_to_location(obj, &f_page, &f_objidx); |
bfd093f5 | 1614 | f_offset = (class->size * f_objidx) & ~PAGE_MASK; |
3783689a | 1615 | zspage = get_zspage(f_page); |
61989a80 | 1616 | |
c7806261 | 1617 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1618 | |
1619 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1620 | link = (struct link_free *)(vaddr + f_offset); |
3b1d9ca6 | 1621 | link->next = get_freeobj(zspage) << OBJ_TAG_BITS; |
af4ee5e9 | 1622 | kunmap_atomic(vaddr); |
bfd093f5 | 1623 | set_freeobj(zspage, f_objidx); |
3783689a | 1624 | mod_zspage_inuse(zspage, -1); |
0f050d99 | 1625 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1626 | } |
1627 | ||
1628 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1629 | { | |
3783689a MK |
1630 | struct zspage *zspage; |
1631 | struct page *f_page; | |
bfd093f5 MK |
1632 | unsigned long obj; |
1633 | unsigned int f_objidx; | |
c7806261 MK |
1634 | int class_idx; |
1635 | struct size_class *class; | |
1636 | enum fullness_group fullness; | |
48b4800a | 1637 | bool isolated; |
c7806261 MK |
1638 | |
1639 | if (unlikely(!handle)) | |
1640 | return; | |
1641 | ||
312fcae2 | 1642 | pin_tag(handle); |
c7806261 | 1643 | obj = handle_to_obj(handle); |
c7806261 | 1644 | obj_to_location(obj, &f_page, &f_objidx); |
3783689a | 1645 | zspage = get_zspage(f_page); |
c7806261 | 1646 | |
48b4800a MK |
1647 | migrate_read_lock(zspage); |
1648 | ||
3783689a | 1649 | get_zspage_mapping(zspage, &class_idx, &fullness); |
c7806261 MK |
1650 | class = pool->size_class[class_idx]; |
1651 | ||
1652 | spin_lock(&class->lock); | |
1ee47165 | 1653 | obj_free(class, obj); |
3783689a | 1654 | fullness = fix_fullness_group(class, zspage); |
48b4800a MK |
1655 | if (fullness != ZS_EMPTY) { |
1656 | migrate_read_unlock(zspage); | |
1657 | goto out; | |
312fcae2 | 1658 | } |
48b4800a MK |
1659 | |
1660 | isolated = is_zspage_isolated(zspage); | |
1661 | migrate_read_unlock(zspage); | |
1662 | /* If zspage is isolated, zs_page_putback will free the zspage */ | |
1663 | if (likely(!isolated)) | |
1664 | free_zspage(pool, class, zspage); | |
1665 | out: | |
1666 | ||
61989a80 | 1667 | spin_unlock(&class->lock); |
312fcae2 | 1668 | unpin_tag(handle); |
3783689a | 1669 | cache_free_handle(pool, handle); |
312fcae2 MK |
1670 | } |
1671 | EXPORT_SYMBOL_GPL(zs_free); | |
1672 | ||
251cbb95 MK |
1673 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1674 | unsigned long src) | |
312fcae2 MK |
1675 | { |
1676 | struct page *s_page, *d_page; | |
bfd093f5 | 1677 | unsigned int s_objidx, d_objidx; |
312fcae2 MK |
1678 | unsigned long s_off, d_off; |
1679 | void *s_addr, *d_addr; | |
1680 | int s_size, d_size, size; | |
1681 | int written = 0; | |
1682 | ||
1683 | s_size = d_size = class->size; | |
1684 | ||
1685 | obj_to_location(src, &s_page, &s_objidx); | |
1686 | obj_to_location(dst, &d_page, &d_objidx); | |
1687 | ||
bfd093f5 MK |
1688 | s_off = (class->size * s_objidx) & ~PAGE_MASK; |
1689 | d_off = (class->size * d_objidx) & ~PAGE_MASK; | |
312fcae2 MK |
1690 | |
1691 | if (s_off + class->size > PAGE_SIZE) | |
1692 | s_size = PAGE_SIZE - s_off; | |
1693 | ||
1694 | if (d_off + class->size > PAGE_SIZE) | |
1695 | d_size = PAGE_SIZE - d_off; | |
1696 | ||
1697 | s_addr = kmap_atomic(s_page); | |
1698 | d_addr = kmap_atomic(d_page); | |
1699 | ||
1700 | while (1) { | |
1701 | size = min(s_size, d_size); | |
1702 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1703 | written += size; | |
1704 | ||
1705 | if (written == class->size) | |
1706 | break; | |
1707 | ||
495819ea SS |
1708 | s_off += size; |
1709 | s_size -= size; | |
1710 | d_off += size; | |
1711 | d_size -= size; | |
1712 | ||
1713 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1714 | kunmap_atomic(d_addr); |
1715 | kunmap_atomic(s_addr); | |
1716 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1717 | s_addr = kmap_atomic(s_page); |
1718 | d_addr = kmap_atomic(d_page); | |
1719 | s_size = class->size - written; | |
1720 | s_off = 0; | |
312fcae2 MK |
1721 | } |
1722 | ||
495819ea | 1723 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1724 | kunmap_atomic(d_addr); |
1725 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1726 | d_addr = kmap_atomic(d_page); |
1727 | d_size = class->size - written; | |
1728 | d_off = 0; | |
312fcae2 MK |
1729 | } |
1730 | } | |
1731 | ||
1732 | kunmap_atomic(d_addr); | |
1733 | kunmap_atomic(s_addr); | |
1734 | } | |
1735 | ||
1736 | /* | |
1737 | * Find alloced object in zspage from index object and | |
1738 | * return handle. | |
1739 | */ | |
251cbb95 | 1740 | static unsigned long find_alloced_obj(struct size_class *class, |
cf675acb | 1741 | struct page *page, int *obj_idx) |
312fcae2 MK |
1742 | { |
1743 | unsigned long head; | |
1744 | int offset = 0; | |
cf675acb | 1745 | int index = *obj_idx; |
312fcae2 MK |
1746 | unsigned long handle = 0; |
1747 | void *addr = kmap_atomic(page); | |
1748 | ||
3783689a | 1749 | offset = get_first_obj_offset(page); |
312fcae2 MK |
1750 | offset += class->size * index; |
1751 | ||
1752 | while (offset < PAGE_SIZE) { | |
48b4800a | 1753 | head = obj_to_head(page, addr + offset); |
312fcae2 MK |
1754 | if (head & OBJ_ALLOCATED_TAG) { |
1755 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1756 | if (trypin_tag(handle)) | |
1757 | break; | |
1758 | handle = 0; | |
1759 | } | |
1760 | ||
1761 | offset += class->size; | |
1762 | index++; | |
1763 | } | |
1764 | ||
1765 | kunmap_atomic(addr); | |
cf675acb GM |
1766 | |
1767 | *obj_idx = index; | |
1768 | ||
312fcae2 MK |
1769 | return handle; |
1770 | } | |
1771 | ||
1772 | struct zs_compact_control { | |
3783689a | 1773 | /* Source spage for migration which could be a subpage of zspage */ |
312fcae2 MK |
1774 | struct page *s_page; |
1775 | /* Destination page for migration which should be a first page | |
1776 | * of zspage. */ | |
1777 | struct page *d_page; | |
1778 | /* Starting object index within @s_page which used for live object | |
1779 | * in the subpage. */ | |
41b88e14 | 1780 | int obj_idx; |
312fcae2 MK |
1781 | }; |
1782 | ||
1783 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1784 | struct zs_compact_control *cc) | |
1785 | { | |
1786 | unsigned long used_obj, free_obj; | |
1787 | unsigned long handle; | |
1788 | struct page *s_page = cc->s_page; | |
1789 | struct page *d_page = cc->d_page; | |
41b88e14 | 1790 | int obj_idx = cc->obj_idx; |
312fcae2 MK |
1791 | int ret = 0; |
1792 | ||
1793 | while (1) { | |
cf675acb | 1794 | handle = find_alloced_obj(class, s_page, &obj_idx); |
312fcae2 MK |
1795 | if (!handle) { |
1796 | s_page = get_next_page(s_page); | |
1797 | if (!s_page) | |
1798 | break; | |
41b88e14 | 1799 | obj_idx = 0; |
312fcae2 MK |
1800 | continue; |
1801 | } | |
1802 | ||
1803 | /* Stop if there is no more space */ | |
3783689a | 1804 | if (zspage_full(class, get_zspage(d_page))) { |
312fcae2 MK |
1805 | unpin_tag(handle); |
1806 | ret = -ENOMEM; | |
1807 | break; | |
1808 | } | |
1809 | ||
1810 | used_obj = handle_to_obj(handle); | |
3783689a | 1811 | free_obj = obj_malloc(class, get_zspage(d_page), handle); |
251cbb95 | 1812 | zs_object_copy(class, free_obj, used_obj); |
41b88e14 | 1813 | obj_idx++; |
c102f07c JL |
1814 | /* |
1815 | * record_obj updates handle's value to free_obj and it will | |
1816 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1817 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1818 | * let's keep the lock bit. | |
1819 | */ | |
1820 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1821 | record_obj(handle, free_obj); |
1822 | unpin_tag(handle); | |
1ee47165 | 1823 | obj_free(class, used_obj); |
312fcae2 MK |
1824 | } |
1825 | ||
1826 | /* Remember last position in this iteration */ | |
1827 | cc->s_page = s_page; | |
41b88e14 | 1828 | cc->obj_idx = obj_idx; |
312fcae2 MK |
1829 | |
1830 | return ret; | |
1831 | } | |
1832 | ||
3783689a | 1833 | static struct zspage *isolate_zspage(struct size_class *class, bool source) |
312fcae2 MK |
1834 | { |
1835 | int i; | |
3783689a MK |
1836 | struct zspage *zspage; |
1837 | enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL}; | |
312fcae2 | 1838 | |
3783689a MK |
1839 | if (!source) { |
1840 | fg[0] = ZS_ALMOST_FULL; | |
1841 | fg[1] = ZS_ALMOST_EMPTY; | |
1842 | } | |
1843 | ||
1844 | for (i = 0; i < 2; i++) { | |
1845 | zspage = list_first_entry_or_null(&class->fullness_list[fg[i]], | |
1846 | struct zspage, list); | |
1847 | if (zspage) { | |
48b4800a | 1848 | VM_BUG_ON(is_zspage_isolated(zspage)); |
3783689a MK |
1849 | remove_zspage(class, zspage, fg[i]); |
1850 | return zspage; | |
312fcae2 MK |
1851 | } |
1852 | } | |
1853 | ||
3783689a | 1854 | return zspage; |
312fcae2 MK |
1855 | } |
1856 | ||
860c707d | 1857 | /* |
3783689a | 1858 | * putback_zspage - add @zspage into right class's fullness list |
860c707d | 1859 | * @class: destination class |
3783689a | 1860 | * @zspage: target page |
860c707d | 1861 | * |
3783689a | 1862 | * Return @zspage's fullness_group |
860c707d | 1863 | */ |
4aa409ca | 1864 | static enum fullness_group putback_zspage(struct size_class *class, |
3783689a | 1865 | struct zspage *zspage) |
312fcae2 | 1866 | { |
312fcae2 MK |
1867 | enum fullness_group fullness; |
1868 | ||
48b4800a MK |
1869 | VM_BUG_ON(is_zspage_isolated(zspage)); |
1870 | ||
3783689a MK |
1871 | fullness = get_fullness_group(class, zspage); |
1872 | insert_zspage(class, zspage, fullness); | |
1873 | set_zspage_mapping(zspage, class->index, fullness); | |
839373e6 | 1874 | |
860c707d | 1875 | return fullness; |
61989a80 | 1876 | } |
312fcae2 | 1877 | |
48b4800a MK |
1878 | #ifdef CONFIG_COMPACTION |
1879 | static struct dentry *zs_mount(struct file_system_type *fs_type, | |
1880 | int flags, const char *dev_name, void *data) | |
1881 | { | |
1882 | static const struct dentry_operations ops = { | |
1883 | .d_dname = simple_dname, | |
1884 | }; | |
1885 | ||
1886 | return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC); | |
1887 | } | |
1888 | ||
1889 | static struct file_system_type zsmalloc_fs = { | |
1890 | .name = "zsmalloc", | |
1891 | .mount = zs_mount, | |
1892 | .kill_sb = kill_anon_super, | |
1893 | }; | |
1894 | ||
1895 | static int zsmalloc_mount(void) | |
1896 | { | |
1897 | int ret = 0; | |
1898 | ||
1899 | zsmalloc_mnt = kern_mount(&zsmalloc_fs); | |
1900 | if (IS_ERR(zsmalloc_mnt)) | |
1901 | ret = PTR_ERR(zsmalloc_mnt); | |
1902 | ||
1903 | return ret; | |
1904 | } | |
1905 | ||
1906 | static void zsmalloc_unmount(void) | |
1907 | { | |
1908 | kern_unmount(zsmalloc_mnt); | |
1909 | } | |
1910 | ||
1911 | static void migrate_lock_init(struct zspage *zspage) | |
1912 | { | |
1913 | rwlock_init(&zspage->lock); | |
1914 | } | |
1915 | ||
1916 | static void migrate_read_lock(struct zspage *zspage) | |
1917 | { | |
1918 | read_lock(&zspage->lock); | |
1919 | } | |
1920 | ||
1921 | static void migrate_read_unlock(struct zspage *zspage) | |
1922 | { | |
1923 | read_unlock(&zspage->lock); | |
1924 | } | |
1925 | ||
1926 | static void migrate_write_lock(struct zspage *zspage) | |
1927 | { | |
1928 | write_lock(&zspage->lock); | |
1929 | } | |
1930 | ||
1931 | static void migrate_write_unlock(struct zspage *zspage) | |
1932 | { | |
1933 | write_unlock(&zspage->lock); | |
1934 | } | |
1935 | ||
1936 | /* Number of isolated subpage for *page migration* in this zspage */ | |
1937 | static void inc_zspage_isolation(struct zspage *zspage) | |
1938 | { | |
1939 | zspage->isolated++; | |
1940 | } | |
1941 | ||
1942 | static void dec_zspage_isolation(struct zspage *zspage) | |
1943 | { | |
1944 | zspage->isolated--; | |
1945 | } | |
1946 | ||
1947 | static void replace_sub_page(struct size_class *class, struct zspage *zspage, | |
1948 | struct page *newpage, struct page *oldpage) | |
1949 | { | |
1950 | struct page *page; | |
1951 | struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, }; | |
1952 | int idx = 0; | |
1953 | ||
1954 | page = get_first_page(zspage); | |
1955 | do { | |
1956 | if (page == oldpage) | |
1957 | pages[idx] = newpage; | |
1958 | else | |
1959 | pages[idx] = page; | |
1960 | idx++; | |
1961 | } while ((page = get_next_page(page)) != NULL); | |
1962 | ||
1963 | create_page_chain(class, zspage, pages); | |
1964 | set_first_obj_offset(newpage, get_first_obj_offset(oldpage)); | |
1965 | if (unlikely(PageHugeObject(oldpage))) | |
1966 | newpage->index = oldpage->index; | |
1967 | __SetPageMovable(newpage, page_mapping(oldpage)); | |
1968 | } | |
1969 | ||
1970 | bool zs_page_isolate(struct page *page, isolate_mode_t mode) | |
1971 | { | |
1972 | struct zs_pool *pool; | |
1973 | struct size_class *class; | |
1974 | int class_idx; | |
1975 | enum fullness_group fullness; | |
1976 | struct zspage *zspage; | |
1977 | struct address_space *mapping; | |
1978 | ||
1979 | /* | |
1980 | * Page is locked so zspage couldn't be destroyed. For detail, look at | |
1981 | * lock_zspage in free_zspage. | |
1982 | */ | |
1983 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
1984 | VM_BUG_ON_PAGE(PageIsolated(page), page); | |
1985 | ||
1986 | zspage = get_zspage(page); | |
1987 | ||
1988 | /* | |
1989 | * Without class lock, fullness could be stale while class_idx is okay | |
1990 | * because class_idx is constant unless page is freed so we should get | |
1991 | * fullness again under class lock. | |
1992 | */ | |
1993 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
1994 | mapping = page_mapping(page); | |
1995 | pool = mapping->private_data; | |
1996 | class = pool->size_class[class_idx]; | |
1997 | ||
1998 | spin_lock(&class->lock); | |
1999 | if (get_zspage_inuse(zspage) == 0) { | |
2000 | spin_unlock(&class->lock); | |
2001 | return false; | |
2002 | } | |
2003 | ||
2004 | /* zspage is isolated for object migration */ | |
2005 | if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
2006 | spin_unlock(&class->lock); | |
2007 | return false; | |
2008 | } | |
2009 | ||
2010 | /* | |
2011 | * If this is first time isolation for the zspage, isolate zspage from | |
2012 | * size_class to prevent further object allocation from the zspage. | |
2013 | */ | |
2014 | if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) { | |
2015 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2016 | remove_zspage(class, zspage, fullness); | |
2017 | } | |
2018 | ||
2019 | inc_zspage_isolation(zspage); | |
2020 | spin_unlock(&class->lock); | |
2021 | ||
2022 | return true; | |
2023 | } | |
2024 | ||
2025 | int zs_page_migrate(struct address_space *mapping, struct page *newpage, | |
2026 | struct page *page, enum migrate_mode mode) | |
2027 | { | |
2028 | struct zs_pool *pool; | |
2029 | struct size_class *class; | |
2030 | int class_idx; | |
2031 | enum fullness_group fullness; | |
2032 | struct zspage *zspage; | |
2033 | struct page *dummy; | |
2034 | void *s_addr, *d_addr, *addr; | |
2035 | int offset, pos; | |
2036 | unsigned long handle, head; | |
2037 | unsigned long old_obj, new_obj; | |
2038 | unsigned int obj_idx; | |
2039 | int ret = -EAGAIN; | |
2040 | ||
2041 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
2042 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
2043 | ||
2044 | zspage = get_zspage(page); | |
2045 | ||
2046 | /* Concurrent compactor cannot migrate any subpage in zspage */ | |
2047 | migrate_write_lock(zspage); | |
2048 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2049 | pool = mapping->private_data; | |
2050 | class = pool->size_class[class_idx]; | |
2051 | offset = get_first_obj_offset(page); | |
2052 | ||
2053 | spin_lock(&class->lock); | |
2054 | if (!get_zspage_inuse(zspage)) { | |
2055 | ret = -EBUSY; | |
2056 | goto unlock_class; | |
2057 | } | |
2058 | ||
2059 | pos = offset; | |
2060 | s_addr = kmap_atomic(page); | |
2061 | while (pos < PAGE_SIZE) { | |
2062 | head = obj_to_head(page, s_addr + pos); | |
2063 | if (head & OBJ_ALLOCATED_TAG) { | |
2064 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2065 | if (!trypin_tag(handle)) | |
2066 | goto unpin_objects; | |
2067 | } | |
2068 | pos += class->size; | |
2069 | } | |
2070 | ||
2071 | /* | |
2072 | * Here, any user cannot access all objects in the zspage so let's move. | |
2073 | */ | |
2074 | d_addr = kmap_atomic(newpage); | |
2075 | memcpy(d_addr, s_addr, PAGE_SIZE); | |
2076 | kunmap_atomic(d_addr); | |
2077 | ||
2078 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2079 | addr += class->size) { | |
2080 | head = obj_to_head(page, addr); | |
2081 | if (head & OBJ_ALLOCATED_TAG) { | |
2082 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2083 | if (!testpin_tag(handle)) | |
2084 | BUG(); | |
2085 | ||
2086 | old_obj = handle_to_obj(handle); | |
2087 | obj_to_location(old_obj, &dummy, &obj_idx); | |
2088 | new_obj = (unsigned long)location_to_obj(newpage, | |
2089 | obj_idx); | |
2090 | new_obj |= BIT(HANDLE_PIN_BIT); | |
2091 | record_obj(handle, new_obj); | |
2092 | } | |
2093 | } | |
2094 | ||
2095 | replace_sub_page(class, zspage, newpage, page); | |
2096 | get_page(newpage); | |
2097 | ||
2098 | dec_zspage_isolation(zspage); | |
2099 | ||
2100 | /* | |
2101 | * Page migration is done so let's putback isolated zspage to | |
2102 | * the list if @page is final isolated subpage in the zspage. | |
2103 | */ | |
2104 | if (!is_zspage_isolated(zspage)) | |
2105 | putback_zspage(class, zspage); | |
2106 | ||
2107 | reset_page(page); | |
2108 | put_page(page); | |
2109 | page = newpage; | |
2110 | ||
dd4123f3 | 2111 | ret = MIGRATEPAGE_SUCCESS; |
48b4800a MK |
2112 | unpin_objects: |
2113 | for (addr = s_addr + offset; addr < s_addr + pos; | |
2114 | addr += class->size) { | |
2115 | head = obj_to_head(page, addr); | |
2116 | if (head & OBJ_ALLOCATED_TAG) { | |
2117 | handle = head & ~OBJ_ALLOCATED_TAG; | |
2118 | if (!testpin_tag(handle)) | |
2119 | BUG(); | |
2120 | unpin_tag(handle); | |
2121 | } | |
2122 | } | |
2123 | kunmap_atomic(s_addr); | |
2124 | unlock_class: | |
2125 | spin_unlock(&class->lock); | |
2126 | migrate_write_unlock(zspage); | |
2127 | ||
2128 | return ret; | |
2129 | } | |
2130 | ||
2131 | void zs_page_putback(struct page *page) | |
2132 | { | |
2133 | struct zs_pool *pool; | |
2134 | struct size_class *class; | |
2135 | int class_idx; | |
2136 | enum fullness_group fg; | |
2137 | struct address_space *mapping; | |
2138 | struct zspage *zspage; | |
2139 | ||
2140 | VM_BUG_ON_PAGE(!PageMovable(page), page); | |
2141 | VM_BUG_ON_PAGE(!PageIsolated(page), page); | |
2142 | ||
2143 | zspage = get_zspage(page); | |
2144 | get_zspage_mapping(zspage, &class_idx, &fg); | |
2145 | mapping = page_mapping(page); | |
2146 | pool = mapping->private_data; | |
2147 | class = pool->size_class[class_idx]; | |
2148 | ||
2149 | spin_lock(&class->lock); | |
2150 | dec_zspage_isolation(zspage); | |
2151 | if (!is_zspage_isolated(zspage)) { | |
2152 | fg = putback_zspage(class, zspage); | |
2153 | /* | |
2154 | * Due to page_lock, we cannot free zspage immediately | |
2155 | * so let's defer. | |
2156 | */ | |
2157 | if (fg == ZS_EMPTY) | |
2158 | schedule_work(&pool->free_work); | |
2159 | } | |
2160 | spin_unlock(&class->lock); | |
2161 | } | |
2162 | ||
2163 | const struct address_space_operations zsmalloc_aops = { | |
2164 | .isolate_page = zs_page_isolate, | |
2165 | .migratepage = zs_page_migrate, | |
2166 | .putback_page = zs_page_putback, | |
2167 | }; | |
2168 | ||
2169 | static int zs_register_migration(struct zs_pool *pool) | |
2170 | { | |
2171 | pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb); | |
2172 | if (IS_ERR(pool->inode)) { | |
2173 | pool->inode = NULL; | |
2174 | return 1; | |
2175 | } | |
2176 | ||
2177 | pool->inode->i_mapping->private_data = pool; | |
2178 | pool->inode->i_mapping->a_ops = &zsmalloc_aops; | |
2179 | return 0; | |
2180 | } | |
2181 | ||
2182 | static void zs_unregister_migration(struct zs_pool *pool) | |
2183 | { | |
2184 | flush_work(&pool->free_work); | |
2185 | if (pool->inode) | |
2186 | iput(pool->inode); | |
2187 | } | |
2188 | ||
2189 | /* | |
2190 | * Caller should hold page_lock of all pages in the zspage | |
2191 | * In here, we cannot use zspage meta data. | |
2192 | */ | |
2193 | static void async_free_zspage(struct work_struct *work) | |
2194 | { | |
2195 | int i; | |
2196 | struct size_class *class; | |
2197 | unsigned int class_idx; | |
2198 | enum fullness_group fullness; | |
2199 | struct zspage *zspage, *tmp; | |
2200 | LIST_HEAD(free_pages); | |
2201 | struct zs_pool *pool = container_of(work, struct zs_pool, | |
2202 | free_work); | |
2203 | ||
2204 | for (i = 0; i < zs_size_classes; i++) { | |
2205 | class = pool->size_class[i]; | |
2206 | if (class->index != i) | |
2207 | continue; | |
2208 | ||
2209 | spin_lock(&class->lock); | |
2210 | list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages); | |
2211 | spin_unlock(&class->lock); | |
2212 | } | |
2213 | ||
2214 | ||
2215 | list_for_each_entry_safe(zspage, tmp, &free_pages, list) { | |
2216 | list_del(&zspage->list); | |
2217 | lock_zspage(zspage); | |
2218 | ||
2219 | get_zspage_mapping(zspage, &class_idx, &fullness); | |
2220 | VM_BUG_ON(fullness != ZS_EMPTY); | |
2221 | class = pool->size_class[class_idx]; | |
2222 | spin_lock(&class->lock); | |
2223 | __free_zspage(pool, pool->size_class[class_idx], zspage); | |
2224 | spin_unlock(&class->lock); | |
2225 | } | |
2226 | }; | |
2227 | ||
2228 | static void kick_deferred_free(struct zs_pool *pool) | |
2229 | { | |
2230 | schedule_work(&pool->free_work); | |
2231 | } | |
2232 | ||
2233 | static void init_deferred_free(struct zs_pool *pool) | |
2234 | { | |
2235 | INIT_WORK(&pool->free_work, async_free_zspage); | |
2236 | } | |
2237 | ||
2238 | static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) | |
2239 | { | |
2240 | struct page *page = get_first_page(zspage); | |
2241 | ||
2242 | do { | |
2243 | WARN_ON(!trylock_page(page)); | |
2244 | __SetPageMovable(page, pool->inode->i_mapping); | |
2245 | unlock_page(page); | |
2246 | } while ((page = get_next_page(page)) != NULL); | |
2247 | } | |
2248 | #endif | |
2249 | ||
04f05909 SS |
2250 | /* |
2251 | * | |
2252 | * Based on the number of unused allocated objects calculate | |
2253 | * and return the number of pages that we can free. | |
04f05909 SS |
2254 | */ |
2255 | static unsigned long zs_can_compact(struct size_class *class) | |
2256 | { | |
2257 | unsigned long obj_wasted; | |
44f43e99 SS |
2258 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
2259 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 2260 | |
44f43e99 SS |
2261 | if (obj_allocated <= obj_used) |
2262 | return 0; | |
04f05909 | 2263 | |
44f43e99 | 2264 | obj_wasted = obj_allocated - obj_used; |
b4fd07a0 | 2265 | obj_wasted /= class->objs_per_zspage; |
04f05909 | 2266 | |
6cbf16b3 | 2267 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
2268 | } |
2269 | ||
7d3f3938 | 2270 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 2271 | { |
312fcae2 | 2272 | struct zs_compact_control cc; |
3783689a MK |
2273 | struct zspage *src_zspage; |
2274 | struct zspage *dst_zspage = NULL; | |
312fcae2 | 2275 | |
312fcae2 | 2276 | spin_lock(&class->lock); |
3783689a | 2277 | while ((src_zspage = isolate_zspage(class, true))) { |
312fcae2 | 2278 | |
04f05909 SS |
2279 | if (!zs_can_compact(class)) |
2280 | break; | |
2281 | ||
41b88e14 | 2282 | cc.obj_idx = 0; |
48b4800a | 2283 | cc.s_page = get_first_page(src_zspage); |
312fcae2 | 2284 | |
3783689a | 2285 | while ((dst_zspage = isolate_zspage(class, false))) { |
48b4800a | 2286 | cc.d_page = get_first_page(dst_zspage); |
312fcae2 | 2287 | /* |
0dc63d48 SS |
2288 | * If there is no more space in dst_page, resched |
2289 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
2290 | */ |
2291 | if (!migrate_zspage(pool, class, &cc)) | |
2292 | break; | |
2293 | ||
4aa409ca | 2294 | putback_zspage(class, dst_zspage); |
312fcae2 MK |
2295 | } |
2296 | ||
2297 | /* Stop if we couldn't find slot */ | |
3783689a | 2298 | if (dst_zspage == NULL) |
312fcae2 MK |
2299 | break; |
2300 | ||
4aa409ca MK |
2301 | putback_zspage(class, dst_zspage); |
2302 | if (putback_zspage(class, src_zspage) == ZS_EMPTY) { | |
48b4800a | 2303 | free_zspage(pool, class, src_zspage); |
6cbf16b3 | 2304 | pool->stats.pages_compacted += class->pages_per_zspage; |
4aa409ca | 2305 | } |
312fcae2 | 2306 | spin_unlock(&class->lock); |
312fcae2 MK |
2307 | cond_resched(); |
2308 | spin_lock(&class->lock); | |
2309 | } | |
2310 | ||
3783689a | 2311 | if (src_zspage) |
4aa409ca | 2312 | putback_zspage(class, src_zspage); |
312fcae2 | 2313 | |
7d3f3938 | 2314 | spin_unlock(&class->lock); |
312fcae2 MK |
2315 | } |
2316 | ||
2317 | unsigned long zs_compact(struct zs_pool *pool) | |
2318 | { | |
2319 | int i; | |
312fcae2 MK |
2320 | struct size_class *class; |
2321 | ||
2322 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
2323 | class = pool->size_class[i]; | |
2324 | if (!class) | |
2325 | continue; | |
2326 | if (class->index != i) | |
2327 | continue; | |
7d3f3938 | 2328 | __zs_compact(pool, class); |
312fcae2 MK |
2329 | } |
2330 | ||
860c707d | 2331 | return pool->stats.pages_compacted; |
312fcae2 MK |
2332 | } |
2333 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 2334 | |
7d3f3938 SS |
2335 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
2336 | { | |
2337 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
2338 | } | |
2339 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
2340 | ||
ab9d306d SS |
2341 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
2342 | struct shrink_control *sc) | |
2343 | { | |
2344 | unsigned long pages_freed; | |
2345 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2346 | shrinker); | |
2347 | ||
2348 | pages_freed = pool->stats.pages_compacted; | |
2349 | /* | |
2350 | * Compact classes and calculate compaction delta. | |
2351 | * Can run concurrently with a manually triggered | |
2352 | * (by user) compaction. | |
2353 | */ | |
2354 | pages_freed = zs_compact(pool) - pages_freed; | |
2355 | ||
2356 | return pages_freed ? pages_freed : SHRINK_STOP; | |
2357 | } | |
2358 | ||
2359 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
2360 | struct shrink_control *sc) | |
2361 | { | |
2362 | int i; | |
2363 | struct size_class *class; | |
2364 | unsigned long pages_to_free = 0; | |
2365 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
2366 | shrinker); | |
2367 | ||
ab9d306d SS |
2368 | for (i = zs_size_classes - 1; i >= 0; i--) { |
2369 | class = pool->size_class[i]; | |
2370 | if (!class) | |
2371 | continue; | |
2372 | if (class->index != i) | |
2373 | continue; | |
2374 | ||
ab9d306d | 2375 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
2376 | } |
2377 | ||
2378 | return pages_to_free; | |
2379 | } | |
2380 | ||
2381 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
2382 | { | |
2383 | if (pool->shrinker_enabled) { | |
2384 | unregister_shrinker(&pool->shrinker); | |
2385 | pool->shrinker_enabled = false; | |
2386 | } | |
2387 | } | |
2388 | ||
2389 | static int zs_register_shrinker(struct zs_pool *pool) | |
2390 | { | |
2391 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
2392 | pool->shrinker.count_objects = zs_shrinker_count; | |
2393 | pool->shrinker.batch = 0; | |
2394 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
2395 | ||
2396 | return register_shrinker(&pool->shrinker); | |
2397 | } | |
2398 | ||
00a61d86 | 2399 | /** |
66cdef66 GM |
2400 | * zs_create_pool - Creates an allocation pool to work from. |
2401 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 2402 | * |
66cdef66 GM |
2403 | * This function must be called before anything when using |
2404 | * the zsmalloc allocator. | |
166cfda7 | 2405 | * |
66cdef66 GM |
2406 | * On success, a pointer to the newly created pool is returned, |
2407 | * otherwise NULL. | |
396b7fd6 | 2408 | */ |
d0d8da2d | 2409 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 2410 | { |
66cdef66 GM |
2411 | int i; |
2412 | struct zs_pool *pool; | |
2413 | struct size_class *prev_class = NULL; | |
61989a80 | 2414 | |
66cdef66 GM |
2415 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
2416 | if (!pool) | |
2417 | return NULL; | |
61989a80 | 2418 | |
48b4800a | 2419 | init_deferred_free(pool); |
66cdef66 GM |
2420 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
2421 | GFP_KERNEL); | |
2422 | if (!pool->size_class) { | |
2423 | kfree(pool); | |
2424 | return NULL; | |
2425 | } | |
61989a80 | 2426 | |
2e40e163 MK |
2427 | pool->name = kstrdup(name, GFP_KERNEL); |
2428 | if (!pool->name) | |
2429 | goto err; | |
2430 | ||
3783689a | 2431 | if (create_cache(pool)) |
2e40e163 MK |
2432 | goto err; |
2433 | ||
c60369f0 | 2434 | /* |
66cdef66 GM |
2435 | * Iterate reversly, because, size of size_class that we want to use |
2436 | * for merging should be larger or equal to current size. | |
c60369f0 | 2437 | */ |
66cdef66 GM |
2438 | for (i = zs_size_classes - 1; i >= 0; i--) { |
2439 | int size; | |
2440 | int pages_per_zspage; | |
2441 | struct size_class *class; | |
3783689a | 2442 | int fullness = 0; |
c60369f0 | 2443 | |
66cdef66 GM |
2444 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
2445 | if (size > ZS_MAX_ALLOC_SIZE) | |
2446 | size = ZS_MAX_ALLOC_SIZE; | |
2447 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 2448 | |
66cdef66 GM |
2449 | /* |
2450 | * size_class is used for normal zsmalloc operation such | |
2451 | * as alloc/free for that size. Although it is natural that we | |
2452 | * have one size_class for each size, there is a chance that we | |
2453 | * can get more memory utilization if we use one size_class for | |
2454 | * many different sizes whose size_class have same | |
2455 | * characteristics. So, we makes size_class point to | |
2456 | * previous size_class if possible. | |
2457 | */ | |
2458 | if (prev_class) { | |
2459 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
2460 | pool->size_class[i] = prev_class; | |
2461 | continue; | |
2462 | } | |
2463 | } | |
2464 | ||
2465 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
2466 | if (!class) | |
2467 | goto err; | |
2468 | ||
2469 | class->size = size; | |
2470 | class->index = i; | |
2471 | class->pages_per_zspage = pages_per_zspage; | |
b4fd07a0 GM |
2472 | class->objs_per_zspage = get_maxobj_per_zspage(class->size, |
2473 | class->pages_per_zspage); | |
66cdef66 GM |
2474 | spin_lock_init(&class->lock); |
2475 | pool->size_class[i] = class; | |
48b4800a MK |
2476 | for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS; |
2477 | fullness++) | |
3783689a | 2478 | INIT_LIST_HEAD(&class->fullness_list[fullness]); |
66cdef66 GM |
2479 | |
2480 | prev_class = class; | |
61989a80 NG |
2481 | } |
2482 | ||
d34f6157 DS |
2483 | /* debug only, don't abort if it fails */ |
2484 | zs_pool_stat_create(pool, name); | |
0f050d99 | 2485 | |
48b4800a MK |
2486 | if (zs_register_migration(pool)) |
2487 | goto err; | |
2488 | ||
ab9d306d SS |
2489 | /* |
2490 | * Not critical, we still can use the pool | |
2491 | * and user can trigger compaction manually. | |
2492 | */ | |
2493 | if (zs_register_shrinker(pool) == 0) | |
2494 | pool->shrinker_enabled = true; | |
66cdef66 GM |
2495 | return pool; |
2496 | ||
2497 | err: | |
2498 | zs_destroy_pool(pool); | |
2499 | return NULL; | |
61989a80 | 2500 | } |
66cdef66 | 2501 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 2502 | |
66cdef66 | 2503 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 2504 | { |
66cdef66 | 2505 | int i; |
61989a80 | 2506 | |
ab9d306d | 2507 | zs_unregister_shrinker(pool); |
48b4800a | 2508 | zs_unregister_migration(pool); |
0f050d99 GM |
2509 | zs_pool_stat_destroy(pool); |
2510 | ||
66cdef66 GM |
2511 | for (i = 0; i < zs_size_classes; i++) { |
2512 | int fg; | |
2513 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 2514 | |
66cdef66 GM |
2515 | if (!class) |
2516 | continue; | |
61989a80 | 2517 | |
66cdef66 GM |
2518 | if (class->index != i) |
2519 | continue; | |
61989a80 | 2520 | |
48b4800a | 2521 | for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) { |
3783689a | 2522 | if (!list_empty(&class->fullness_list[fg])) { |
66cdef66 GM |
2523 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", |
2524 | class->size, fg); | |
2525 | } | |
2526 | } | |
2527 | kfree(class); | |
2528 | } | |
f553646a | 2529 | |
3783689a | 2530 | destroy_cache(pool); |
66cdef66 | 2531 | kfree(pool->size_class); |
0f050d99 | 2532 | kfree(pool->name); |
66cdef66 GM |
2533 | kfree(pool); |
2534 | } | |
2535 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2536 | |
66cdef66 GM |
2537 | static int __init zs_init(void) |
2538 | { | |
48b4800a MK |
2539 | int ret; |
2540 | ||
2541 | ret = zsmalloc_mount(); | |
2542 | if (ret) | |
2543 | goto out; | |
2544 | ||
2545 | ret = zs_register_cpu_notifier(); | |
66cdef66 | 2546 | |
0f050d99 GM |
2547 | if (ret) |
2548 | goto notifier_fail; | |
66cdef66 GM |
2549 | |
2550 | init_zs_size_classes(); | |
2551 | ||
2552 | #ifdef CONFIG_ZPOOL | |
2553 | zpool_register_driver(&zs_zpool_driver); | |
2554 | #endif | |
0f050d99 | 2555 | |
4abaac9b DS |
2556 | zs_stat_init(); |
2557 | ||
66cdef66 | 2558 | return 0; |
0f050d99 | 2559 | |
0f050d99 GM |
2560 | notifier_fail: |
2561 | zs_unregister_cpu_notifier(); | |
48b4800a MK |
2562 | zsmalloc_unmount(); |
2563 | out: | |
0f050d99 | 2564 | return ret; |
61989a80 | 2565 | } |
61989a80 | 2566 | |
66cdef66 | 2567 | static void __exit zs_exit(void) |
61989a80 | 2568 | { |
66cdef66 GM |
2569 | #ifdef CONFIG_ZPOOL |
2570 | zpool_unregister_driver(&zs_zpool_driver); | |
2571 | #endif | |
48b4800a | 2572 | zsmalloc_unmount(); |
66cdef66 | 2573 | zs_unregister_cpu_notifier(); |
0f050d99 GM |
2574 | |
2575 | zs_stat_exit(); | |
61989a80 | 2576 | } |
069f101f BH |
2577 | |
2578 | module_init(zs_init); | |
2579 | module_exit(zs_exit); | |
2580 | ||
2581 | MODULE_LICENSE("Dual BSD/GPL"); | |
2582 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |