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: | |
32e7ba1e | 19 | * page->private: points to the first component (0-order) page |
2db51dae NG |
20 | * page->index (union with page->freelist): offset of the first object |
21 | * starting in this page. For the first page, this is | |
22 | * always 0, so we use this field (aka freelist) to point | |
23 | * to the first free object in zspage. | |
24 | * page->lru: links together all component pages (except the first page) | |
25 | * of a zspage | |
26 | * | |
27 | * For _first_ page only: | |
28 | * | |
32e7ba1e | 29 | * page->private: refers to the component page after the first page |
7b60a685 MK |
30 | * If the page is first_page for huge object, it stores handle. |
31 | * Look at size_class->huge. | |
2db51dae NG |
32 | * page->freelist: points to the first free object in zspage. |
33 | * Free objects are linked together using in-place | |
34 | * metadata. | |
35 | * page->objects: maximum number of objects we can store in this | |
36 | * zspage (class->zspage_order * PAGE_SIZE / class->size) | |
37 | * page->lru: links together first pages of various zspages. | |
38 | * Basically forming list of zspages in a fullness group. | |
39 | * page->mapping: class index and fullness group of the zspage | |
8f958c98 | 40 | * page->inuse: the number of objects that are used in this zspage |
2db51dae NG |
41 | * |
42 | * Usage of struct page flags: | |
43 | * PG_private: identifies the first component page | |
44 | * PG_private2: identifies the last component page | |
45 | * | |
46 | */ | |
47 | ||
61989a80 NG |
48 | #include <linux/module.h> |
49 | #include <linux/kernel.h> | |
312fcae2 | 50 | #include <linux/sched.h> |
61989a80 NG |
51 | #include <linux/bitops.h> |
52 | #include <linux/errno.h> | |
53 | #include <linux/highmem.h> | |
61989a80 NG |
54 | #include <linux/string.h> |
55 | #include <linux/slab.h> | |
56 | #include <asm/tlbflush.h> | |
57 | #include <asm/pgtable.h> | |
58 | #include <linux/cpumask.h> | |
59 | #include <linux/cpu.h> | |
0cbb613f | 60 | #include <linux/vmalloc.h> |
759b26b2 | 61 | #include <linux/preempt.h> |
0959c63f SJ |
62 | #include <linux/spinlock.h> |
63 | #include <linux/types.h> | |
0f050d99 | 64 | #include <linux/debugfs.h> |
bcf1647d | 65 | #include <linux/zsmalloc.h> |
c795779d | 66 | #include <linux/zpool.h> |
0959c63f SJ |
67 | |
68 | /* | |
69 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
70 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
71 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
72 | * to restore link_free pointer values. | |
73 | */ | |
74 | #define ZS_ALIGN 8 | |
75 | ||
76 | /* | |
77 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
78 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
79 | */ | |
80 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
81 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
82 | ||
2e40e163 MK |
83 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
84 | ||
0959c63f SJ |
85 | /* |
86 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 87 | * as single (unsigned long) handle value. |
0959c63f SJ |
88 | * |
89 | * Note that object index <obj_idx> is relative to system | |
90 | * page <PFN> it is stored in, so for each sub-page belonging | |
91 | * to a zspage, obj_idx starts with 0. | |
92 | * | |
93 | * This is made more complicated by various memory models and PAE. | |
94 | */ | |
95 | ||
96 | #ifndef MAX_PHYSMEM_BITS | |
97 | #ifdef CONFIG_HIGHMEM64G | |
98 | #define MAX_PHYSMEM_BITS 36 | |
99 | #else /* !CONFIG_HIGHMEM64G */ | |
100 | /* | |
101 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
102 | * be PAGE_SHIFT | |
103 | */ | |
104 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
105 | #endif | |
106 | #endif | |
107 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
108 | |
109 | /* | |
110 | * Memory for allocating for handle keeps object position by | |
111 | * encoding <page, obj_idx> and the encoded value has a room | |
112 | * in least bit(ie, look at obj_to_location). | |
113 | * We use the bit to synchronize between object access by | |
114 | * user and migration. | |
115 | */ | |
116 | #define HANDLE_PIN_BIT 0 | |
117 | ||
118 | /* | |
119 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
120 | * to identify the object was allocated or not. | |
121 | * It's okay to add the status bit in the least bit because | |
122 | * header keeps handle which is 4byte-aligned address so we | |
123 | * have room for two bit at least. | |
124 | */ | |
125 | #define OBJ_ALLOCATED_TAG 1 | |
126 | #define OBJ_TAG_BITS 1 | |
127 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
128 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
129 | ||
130 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
131 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
132 | #define ZS_MIN_ALLOC_SIZE \ | |
133 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 134 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 135 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
136 | |
137 | /* | |
7eb52512 | 138 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
139 | * trader-off here: |
140 | * - Large number of size classes is potentially wasteful as free page are | |
141 | * spread across these classes | |
142 | * - Small number of size classes causes large internal fragmentation | |
143 | * - Probably its better to use specific size classes (empirically | |
144 | * determined). NOTE: all those class sizes must be set as multiple of | |
145 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
146 | * | |
147 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
148 | * (reason above) | |
149 | */ | |
d662b8eb | 150 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8) |
0959c63f SJ |
151 | |
152 | /* | |
153 | * We do not maintain any list for completely empty or full pages | |
154 | */ | |
155 | enum fullness_group { | |
156 | ZS_ALMOST_FULL, | |
157 | ZS_ALMOST_EMPTY, | |
158 | _ZS_NR_FULLNESS_GROUPS, | |
159 | ||
160 | ZS_EMPTY, | |
161 | ZS_FULL | |
162 | }; | |
163 | ||
0f050d99 GM |
164 | enum zs_stat_type { |
165 | OBJ_ALLOCATED, | |
166 | OBJ_USED, | |
248ca1b0 MK |
167 | CLASS_ALMOST_FULL, |
168 | CLASS_ALMOST_EMPTY, | |
0f050d99 GM |
169 | }; |
170 | ||
6fe5186f SS |
171 | #ifdef CONFIG_ZSMALLOC_STAT |
172 | #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1) | |
173 | #else | |
174 | #define NR_ZS_STAT_TYPE (OBJ_USED + 1) | |
175 | #endif | |
176 | ||
0f050d99 GM |
177 | struct zs_size_stat { |
178 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
179 | }; | |
180 | ||
57244594 SS |
181 | #ifdef CONFIG_ZSMALLOC_STAT |
182 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
183 | #endif |
184 | ||
40f9fb8c MG |
185 | /* |
186 | * number of size_classes | |
187 | */ | |
188 | static int zs_size_classes; | |
189 | ||
0959c63f SJ |
190 | /* |
191 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
192 | * n <= N / f, where | |
193 | * n = number of allocated objects | |
194 | * N = total number of objects zspage can store | |
6dd9737e | 195 | * f = fullness_threshold_frac |
0959c63f SJ |
196 | * |
197 | * Similarly, we assign zspage to: | |
198 | * ZS_ALMOST_FULL when n > N / f | |
199 | * ZS_EMPTY when n == 0 | |
200 | * ZS_FULL when n == N | |
201 | * | |
202 | * (see: fix_fullness_group()) | |
203 | */ | |
204 | static const int fullness_threshold_frac = 4; | |
205 | ||
206 | struct size_class { | |
57244594 SS |
207 | spinlock_t lock; |
208 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; | |
0959c63f SJ |
209 | /* |
210 | * Size of objects stored in this class. Must be multiple | |
211 | * of ZS_ALIGN. | |
212 | */ | |
213 | int size; | |
214 | unsigned int index; | |
215 | ||
0f050d99 | 216 | struct zs_size_stat stats; |
0959c63f | 217 | |
7dfa4612 WY |
218 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
219 | int pages_per_zspage; | |
57244594 SS |
220 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
221 | bool huge; | |
0959c63f SJ |
222 | }; |
223 | ||
224 | /* | |
225 | * Placed within free objects to form a singly linked list. | |
226 | * For every zspage, first_page->freelist gives head of this list. | |
227 | * | |
228 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
229 | */ | |
230 | struct link_free { | |
2e40e163 MK |
231 | union { |
232 | /* | |
233 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
234 | * It's valid for non-allocated object | |
235 | */ | |
236 | void *next; | |
237 | /* | |
238 | * Handle of allocated object. | |
239 | */ | |
240 | unsigned long handle; | |
241 | }; | |
0959c63f SJ |
242 | }; |
243 | ||
244 | struct zs_pool { | |
6f3526d6 | 245 | const char *name; |
0f050d99 | 246 | |
40f9fb8c | 247 | struct size_class **size_class; |
2e40e163 | 248 | struct kmem_cache *handle_cachep; |
0959c63f | 249 | |
13de8933 | 250 | atomic_long_t pages_allocated; |
0f050d99 | 251 | |
7d3f3938 | 252 | struct zs_pool_stats stats; |
ab9d306d SS |
253 | |
254 | /* Compact classes */ | |
255 | struct shrinker shrinker; | |
256 | /* | |
257 | * To signify that register_shrinker() was successful | |
258 | * and unregister_shrinker() will not Oops. | |
259 | */ | |
260 | bool shrinker_enabled; | |
0f050d99 GM |
261 | #ifdef CONFIG_ZSMALLOC_STAT |
262 | struct dentry *stat_dentry; | |
263 | #endif | |
0959c63f | 264 | }; |
61989a80 NG |
265 | |
266 | /* | |
267 | * A zspage's class index and fullness group | |
268 | * are encoded in its (first)page->mapping | |
269 | */ | |
270 | #define CLASS_IDX_BITS 28 | |
271 | #define FULLNESS_BITS 4 | |
272 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
273 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
274 | ||
f553646a | 275 | struct mapping_area { |
1b945aee | 276 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
277 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
278 | #else | |
279 | char *vm_buf; /* copy buffer for objects that span pages */ | |
280 | #endif | |
281 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
282 | enum zs_mapmode vm_mm; /* mapping mode */ | |
283 | }; | |
284 | ||
2e40e163 MK |
285 | static int create_handle_cache(struct zs_pool *pool) |
286 | { | |
287 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
288 | 0, 0, NULL); | |
289 | return pool->handle_cachep ? 0 : 1; | |
290 | } | |
291 | ||
292 | static void destroy_handle_cache(struct zs_pool *pool) | |
293 | { | |
cd10add0 | 294 | kmem_cache_destroy(pool->handle_cachep); |
2e40e163 MK |
295 | } |
296 | ||
d0d8da2d | 297 | static unsigned long alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
298 | { |
299 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
d0d8da2d | 300 | gfp & ~__GFP_HIGHMEM); |
2e40e163 MK |
301 | } |
302 | ||
303 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
304 | { | |
305 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
306 | } | |
307 | ||
308 | static void record_obj(unsigned long handle, unsigned long obj) | |
309 | { | |
c102f07c JL |
310 | /* |
311 | * lsb of @obj represents handle lock while other bits | |
312 | * represent object value the handle is pointing so | |
313 | * updating shouldn't do store tearing. | |
314 | */ | |
315 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
316 | } |
317 | ||
c795779d DS |
318 | /* zpool driver */ |
319 | ||
320 | #ifdef CONFIG_ZPOOL | |
321 | ||
6f3526d6 | 322 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 323 | const struct zpool_ops *zpool_ops, |
479305fd | 324 | struct zpool *zpool) |
c795779d | 325 | { |
d0d8da2d SS |
326 | /* |
327 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
328 | * different contexts and its caller must provide a valid | |
329 | * gfp mask. | |
330 | */ | |
331 | return zs_create_pool(name); | |
c795779d DS |
332 | } |
333 | ||
334 | static void zs_zpool_destroy(void *pool) | |
335 | { | |
336 | zs_destroy_pool(pool); | |
337 | } | |
338 | ||
339 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
340 | unsigned long *handle) | |
341 | { | |
d0d8da2d | 342 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
343 | return *handle ? 0 : -1; |
344 | } | |
345 | static void zs_zpool_free(void *pool, unsigned long handle) | |
346 | { | |
347 | zs_free(pool, handle); | |
348 | } | |
349 | ||
350 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
351 | unsigned int *reclaimed) | |
352 | { | |
353 | return -EINVAL; | |
354 | } | |
355 | ||
356 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
357 | enum zpool_mapmode mm) | |
358 | { | |
359 | enum zs_mapmode zs_mm; | |
360 | ||
361 | switch (mm) { | |
362 | case ZPOOL_MM_RO: | |
363 | zs_mm = ZS_MM_RO; | |
364 | break; | |
365 | case ZPOOL_MM_WO: | |
366 | zs_mm = ZS_MM_WO; | |
367 | break; | |
368 | case ZPOOL_MM_RW: /* fallthru */ | |
369 | default: | |
370 | zs_mm = ZS_MM_RW; | |
371 | break; | |
372 | } | |
373 | ||
374 | return zs_map_object(pool, handle, zs_mm); | |
375 | } | |
376 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
377 | { | |
378 | zs_unmap_object(pool, handle); | |
379 | } | |
380 | ||
381 | static u64 zs_zpool_total_size(void *pool) | |
382 | { | |
722cdc17 | 383 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
384 | } |
385 | ||
386 | static struct zpool_driver zs_zpool_driver = { | |
387 | .type = "zsmalloc", | |
388 | .owner = THIS_MODULE, | |
389 | .create = zs_zpool_create, | |
390 | .destroy = zs_zpool_destroy, | |
391 | .malloc = zs_zpool_malloc, | |
392 | .free = zs_zpool_free, | |
393 | .shrink = zs_zpool_shrink, | |
394 | .map = zs_zpool_map, | |
395 | .unmap = zs_zpool_unmap, | |
396 | .total_size = zs_zpool_total_size, | |
397 | }; | |
398 | ||
137f8cff | 399 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
400 | #endif /* CONFIG_ZPOOL */ |
401 | ||
248ca1b0 MK |
402 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
403 | { | |
404 | return pages_per_zspage * PAGE_SIZE / size; | |
405 | } | |
406 | ||
61989a80 NG |
407 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
408 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
409 | ||
410 | static int is_first_page(struct page *page) | |
411 | { | |
a27545bf | 412 | return PagePrivate(page); |
61989a80 NG |
413 | } |
414 | ||
415 | static int is_last_page(struct page *page) | |
416 | { | |
a27545bf | 417 | return PagePrivate2(page); |
61989a80 NG |
418 | } |
419 | ||
a4209467 MK |
420 | static void get_zspage_mapping(struct page *first_page, |
421 | unsigned int *class_idx, | |
61989a80 NG |
422 | enum fullness_group *fullness) |
423 | { | |
424 | unsigned long m; | |
830e4bc5 | 425 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 | 426 | |
a4209467 | 427 | m = (unsigned long)first_page->mapping; |
61989a80 NG |
428 | *fullness = m & FULLNESS_MASK; |
429 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
430 | } | |
431 | ||
a4209467 MK |
432 | static void set_zspage_mapping(struct page *first_page, |
433 | unsigned int class_idx, | |
61989a80 NG |
434 | enum fullness_group fullness) |
435 | { | |
436 | unsigned long m; | |
830e4bc5 | 437 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
438 | |
439 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
440 | (fullness & FULLNESS_MASK); | |
a4209467 | 441 | first_page->mapping = (struct address_space *)m; |
61989a80 NG |
442 | } |
443 | ||
c3e3e88a NC |
444 | /* |
445 | * zsmalloc divides the pool into various size classes where each | |
446 | * class maintains a list of zspages where each zspage is divided | |
447 | * into equal sized chunks. Each allocation falls into one of these | |
448 | * classes depending on its size. This function returns index of the | |
449 | * size class which has chunk size big enough to hold the give size. | |
450 | */ | |
61989a80 NG |
451 | static int get_size_class_index(int size) |
452 | { | |
453 | int idx = 0; | |
454 | ||
455 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
456 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
457 | ZS_SIZE_CLASS_DELTA); | |
458 | ||
7b60a685 | 459 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
460 | } |
461 | ||
248ca1b0 MK |
462 | static inline void zs_stat_inc(struct size_class *class, |
463 | enum zs_stat_type type, unsigned long cnt) | |
464 | { | |
6fe5186f SS |
465 | if (type < NR_ZS_STAT_TYPE) |
466 | class->stats.objs[type] += cnt; | |
248ca1b0 MK |
467 | } |
468 | ||
469 | static inline void zs_stat_dec(struct size_class *class, | |
470 | enum zs_stat_type type, unsigned long cnt) | |
471 | { | |
6fe5186f SS |
472 | if (type < NR_ZS_STAT_TYPE) |
473 | class->stats.objs[type] -= cnt; | |
248ca1b0 MK |
474 | } |
475 | ||
476 | static inline unsigned long zs_stat_get(struct size_class *class, | |
477 | enum zs_stat_type type) | |
478 | { | |
6fe5186f SS |
479 | if (type < NR_ZS_STAT_TYPE) |
480 | return class->stats.objs[type]; | |
481 | return 0; | |
248ca1b0 MK |
482 | } |
483 | ||
57244594 SS |
484 | #ifdef CONFIG_ZSMALLOC_STAT |
485 | ||
248ca1b0 MK |
486 | static int __init zs_stat_init(void) |
487 | { | |
488 | if (!debugfs_initialized()) | |
489 | return -ENODEV; | |
490 | ||
491 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
492 | if (!zs_stat_root) | |
493 | return -ENOMEM; | |
494 | ||
495 | return 0; | |
496 | } | |
497 | ||
498 | static void __exit zs_stat_exit(void) | |
499 | { | |
500 | debugfs_remove_recursive(zs_stat_root); | |
501 | } | |
502 | ||
1120ed54 SS |
503 | static unsigned long zs_can_compact(struct size_class *class); |
504 | ||
248ca1b0 MK |
505 | static int zs_stats_size_show(struct seq_file *s, void *v) |
506 | { | |
507 | int i; | |
508 | struct zs_pool *pool = s->private; | |
509 | struct size_class *class; | |
510 | int objs_per_zspage; | |
511 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 512 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
513 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
514 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 515 | unsigned long total_freeable = 0; |
248ca1b0 | 516 | |
1120ed54 | 517 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
518 | "class", "size", "almost_full", "almost_empty", |
519 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 520 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
521 | |
522 | for (i = 0; i < zs_size_classes; i++) { | |
523 | class = pool->size_class[i]; | |
524 | ||
525 | if (class->index != i) | |
526 | continue; | |
527 | ||
528 | spin_lock(&class->lock); | |
529 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
530 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
531 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
532 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 533 | freeable = zs_can_compact(class); |
248ca1b0 MK |
534 | spin_unlock(&class->lock); |
535 | ||
536 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
537 | class->pages_per_zspage); | |
538 | pages_used = obj_allocated / objs_per_zspage * | |
539 | class->pages_per_zspage; | |
540 | ||
1120ed54 SS |
541 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
542 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
543 | i, class->size, class_almost_full, class_almost_empty, |
544 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 545 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
546 | |
547 | total_class_almost_full += class_almost_full; | |
548 | total_class_almost_empty += class_almost_empty; | |
549 | total_objs += obj_allocated; | |
550 | total_used_objs += obj_used; | |
551 | total_pages += pages_used; | |
1120ed54 | 552 | total_freeable += freeable; |
248ca1b0 MK |
553 | } |
554 | ||
555 | seq_puts(s, "\n"); | |
1120ed54 | 556 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
557 | "Total", "", total_class_almost_full, |
558 | total_class_almost_empty, total_objs, | |
1120ed54 | 559 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
560 | |
561 | return 0; | |
562 | } | |
563 | ||
564 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
565 | { | |
566 | return single_open(file, zs_stats_size_show, inode->i_private); | |
567 | } | |
568 | ||
569 | static const struct file_operations zs_stat_size_ops = { | |
570 | .open = zs_stats_size_open, | |
571 | .read = seq_read, | |
572 | .llseek = seq_lseek, | |
573 | .release = single_release, | |
574 | }; | |
575 | ||
d34f6157 | 576 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
577 | { |
578 | struct dentry *entry; | |
579 | ||
580 | if (!zs_stat_root) | |
d34f6157 | 581 | return; |
248ca1b0 MK |
582 | |
583 | entry = debugfs_create_dir(name, zs_stat_root); | |
584 | if (!entry) { | |
585 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 586 | return; |
248ca1b0 MK |
587 | } |
588 | pool->stat_dentry = entry; | |
589 | ||
590 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
591 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
592 | if (!entry) { | |
593 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
594 | name, "classes"); | |
d34f6157 | 595 | return; |
248ca1b0 | 596 | } |
248ca1b0 MK |
597 | } |
598 | ||
599 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
600 | { | |
601 | debugfs_remove_recursive(pool->stat_dentry); | |
602 | } | |
603 | ||
604 | #else /* CONFIG_ZSMALLOC_STAT */ | |
248ca1b0 MK |
605 | static int __init zs_stat_init(void) |
606 | { | |
607 | return 0; | |
608 | } | |
609 | ||
610 | static void __exit zs_stat_exit(void) | |
611 | { | |
612 | } | |
613 | ||
d34f6157 | 614 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 615 | { |
248ca1b0 MK |
616 | } |
617 | ||
618 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
619 | { | |
620 | } | |
248ca1b0 MK |
621 | #endif |
622 | ||
c3e3e88a NC |
623 | /* |
624 | * For each size class, zspages are divided into different groups | |
625 | * depending on how "full" they are. This was done so that we could | |
626 | * easily find empty or nearly empty zspages when we try to shrink | |
627 | * the pool (not yet implemented). This function returns fullness | |
628 | * status of the given page. | |
629 | */ | |
a4209467 | 630 | static enum fullness_group get_fullness_group(struct page *first_page) |
61989a80 NG |
631 | { |
632 | int inuse, max_objects; | |
633 | enum fullness_group fg; | |
830e4bc5 MK |
634 | |
635 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); | |
61989a80 | 636 | |
a4209467 MK |
637 | inuse = first_page->inuse; |
638 | max_objects = first_page->objects; | |
61989a80 NG |
639 | |
640 | if (inuse == 0) | |
641 | fg = ZS_EMPTY; | |
642 | else if (inuse == max_objects) | |
643 | fg = ZS_FULL; | |
d3d07c92 | 644 | else if (inuse <= 3 * max_objects / fullness_threshold_frac) |
61989a80 NG |
645 | fg = ZS_ALMOST_EMPTY; |
646 | else | |
647 | fg = ZS_ALMOST_FULL; | |
648 | ||
649 | return fg; | |
650 | } | |
651 | ||
c3e3e88a NC |
652 | /* |
653 | * Each size class maintains various freelists and zspages are assigned | |
654 | * to one of these freelists based on the number of live objects they | |
655 | * have. This functions inserts the given zspage into the freelist | |
656 | * identified by <class, fullness_group>. | |
657 | */ | |
251cbb95 MK |
658 | static void insert_zspage(struct size_class *class, |
659 | enum fullness_group fullness, | |
660 | struct page *first_page) | |
61989a80 NG |
661 | { |
662 | struct page **head; | |
663 | ||
830e4bc5 | 664 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
665 | |
666 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
667 | return; | |
668 | ||
248ca1b0 MK |
669 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
670 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
58f17117 SS |
671 | |
672 | head = &class->fullness_list[fullness]; | |
673 | if (!*head) { | |
a4209467 | 674 | *head = first_page; |
58f17117 SS |
675 | return; |
676 | } | |
677 | ||
678 | /* | |
679 | * We want to see more ZS_FULL pages and less almost | |
680 | * empty/full. Put pages with higher ->inuse first. | |
681 | */ | |
a4209467 MK |
682 | list_add_tail(&first_page->lru, &(*head)->lru); |
683 | if (first_page->inuse >= (*head)->inuse) | |
684 | *head = first_page; | |
61989a80 NG |
685 | } |
686 | ||
c3e3e88a NC |
687 | /* |
688 | * This function removes the given zspage from the freelist identified | |
689 | * by <class, fullness_group>. | |
690 | */ | |
251cbb95 MK |
691 | static void remove_zspage(struct size_class *class, |
692 | enum fullness_group fullness, | |
693 | struct page *first_page) | |
61989a80 NG |
694 | { |
695 | struct page **head; | |
696 | ||
830e4bc5 | 697 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
698 | |
699 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
700 | return; | |
701 | ||
702 | head = &class->fullness_list[fullness]; | |
830e4bc5 | 703 | VM_BUG_ON_PAGE(!*head, first_page); |
61989a80 NG |
704 | if (list_empty(&(*head)->lru)) |
705 | *head = NULL; | |
a4209467 | 706 | else if (*head == first_page) |
61989a80 NG |
707 | *head = (struct page *)list_entry((*head)->lru.next, |
708 | struct page, lru); | |
709 | ||
a4209467 | 710 | list_del_init(&first_page->lru); |
248ca1b0 MK |
711 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
712 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
713 | } |
714 | ||
c3e3e88a NC |
715 | /* |
716 | * Each size class maintains zspages in different fullness groups depending | |
717 | * on the number of live objects they contain. When allocating or freeing | |
718 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
719 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
720 | * a status change has occurred for the given page and accordingly moves the | |
721 | * page from the freelist of the old fullness group to that of the new | |
722 | * fullness group. | |
723 | */ | |
c7806261 | 724 | static enum fullness_group fix_fullness_group(struct size_class *class, |
a4209467 | 725 | struct page *first_page) |
61989a80 NG |
726 | { |
727 | int class_idx; | |
61989a80 NG |
728 | enum fullness_group currfg, newfg; |
729 | ||
a4209467 MK |
730 | get_zspage_mapping(first_page, &class_idx, &currfg); |
731 | newfg = get_fullness_group(first_page); | |
61989a80 NG |
732 | if (newfg == currfg) |
733 | goto out; | |
734 | ||
251cbb95 MK |
735 | remove_zspage(class, currfg, first_page); |
736 | insert_zspage(class, newfg, first_page); | |
a4209467 | 737 | set_zspage_mapping(first_page, class_idx, newfg); |
61989a80 NG |
738 | |
739 | out: | |
740 | return newfg; | |
741 | } | |
742 | ||
743 | /* | |
744 | * We have to decide on how many pages to link together | |
745 | * to form a zspage for each size class. This is important | |
746 | * to reduce wastage due to unusable space left at end of | |
747 | * each zspage which is given as: | |
888fa374 YX |
748 | * wastage = Zp % class_size |
749 | * usage = Zp - wastage | |
61989a80 NG |
750 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
751 | * | |
752 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
753 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
754 | * since then we can perfectly fit in 8 such objects. | |
755 | */ | |
2e3b6154 | 756 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
757 | { |
758 | int i, max_usedpc = 0; | |
759 | /* zspage order which gives maximum used size per KB */ | |
760 | int max_usedpc_order = 1; | |
761 | ||
84d4faab | 762 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
763 | int zspage_size; |
764 | int waste, usedpc; | |
765 | ||
766 | zspage_size = i * PAGE_SIZE; | |
767 | waste = zspage_size % class_size; | |
768 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
769 | ||
770 | if (usedpc > max_usedpc) { | |
771 | max_usedpc = usedpc; | |
772 | max_usedpc_order = i; | |
773 | } | |
774 | } | |
775 | ||
776 | return max_usedpc_order; | |
777 | } | |
778 | ||
779 | /* | |
780 | * A single 'zspage' is composed of many system pages which are | |
781 | * linked together using fields in struct page. This function finds | |
782 | * the first/head page, given any component page of a zspage. | |
783 | */ | |
784 | static struct page *get_first_page(struct page *page) | |
785 | { | |
786 | if (is_first_page(page)) | |
787 | return page; | |
788 | else | |
32e7ba1e | 789 | return (struct page *)page_private(page); |
61989a80 NG |
790 | } |
791 | ||
792 | static struct page *get_next_page(struct page *page) | |
793 | { | |
794 | struct page *next; | |
795 | ||
796 | if (is_last_page(page)) | |
797 | next = NULL; | |
798 | else if (is_first_page(page)) | |
e842b976 | 799 | next = (struct page *)page_private(page); |
61989a80 NG |
800 | else |
801 | next = list_entry(page->lru.next, struct page, lru); | |
802 | ||
803 | return next; | |
804 | } | |
805 | ||
67296874 OH |
806 | /* |
807 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 808 | * We use the least bit of handle for tagging. |
67296874 | 809 | */ |
312fcae2 | 810 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 811 | { |
312fcae2 | 812 | unsigned long obj; |
61989a80 NG |
813 | |
814 | if (!page) { | |
830e4bc5 | 815 | VM_BUG_ON(obj_idx); |
61989a80 NG |
816 | return NULL; |
817 | } | |
818 | ||
312fcae2 MK |
819 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
820 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
821 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 822 | |
312fcae2 | 823 | return (void *)obj; |
61989a80 NG |
824 | } |
825 | ||
67296874 OH |
826 | /* |
827 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
828 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 829 | * location_to_obj(). |
67296874 | 830 | */ |
312fcae2 | 831 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
832 | unsigned long *obj_idx) |
833 | { | |
312fcae2 MK |
834 | obj >>= OBJ_TAG_BITS; |
835 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
836 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
837 | } |
838 | ||
2e40e163 MK |
839 | static unsigned long handle_to_obj(unsigned long handle) |
840 | { | |
841 | return *(unsigned long *)handle; | |
842 | } | |
843 | ||
7b60a685 MK |
844 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
845 | void *obj) | |
312fcae2 | 846 | { |
7b60a685 | 847 | if (class->huge) { |
830e4bc5 | 848 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
12a7bfad | 849 | return page_private(page); |
7b60a685 MK |
850 | } else |
851 | return *(unsigned long *)obj; | |
312fcae2 MK |
852 | } |
853 | ||
61989a80 NG |
854 | static unsigned long obj_idx_to_offset(struct page *page, |
855 | unsigned long obj_idx, int class_size) | |
856 | { | |
857 | unsigned long off = 0; | |
858 | ||
859 | if (!is_first_page(page)) | |
860 | off = page->index; | |
861 | ||
862 | return off + obj_idx * class_size; | |
863 | } | |
864 | ||
312fcae2 MK |
865 | static inline int trypin_tag(unsigned long handle) |
866 | { | |
867 | unsigned long *ptr = (unsigned long *)handle; | |
868 | ||
869 | return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr); | |
870 | } | |
871 | ||
872 | static void pin_tag(unsigned long handle) | |
873 | { | |
874 | while (!trypin_tag(handle)); | |
875 | } | |
876 | ||
877 | static void unpin_tag(unsigned long handle) | |
878 | { | |
879 | unsigned long *ptr = (unsigned long *)handle; | |
880 | ||
881 | clear_bit_unlock(HANDLE_PIN_BIT, ptr); | |
882 | } | |
883 | ||
f4477e90 NG |
884 | static void reset_page(struct page *page) |
885 | { | |
886 | clear_bit(PG_private, &page->flags); | |
887 | clear_bit(PG_private_2, &page->flags); | |
888 | set_page_private(page, 0); | |
889 | page->mapping = NULL; | |
890 | page->freelist = NULL; | |
22b751c3 | 891 | page_mapcount_reset(page); |
f4477e90 NG |
892 | } |
893 | ||
61989a80 NG |
894 | static void free_zspage(struct page *first_page) |
895 | { | |
f4477e90 | 896 | struct page *nextp, *tmp, *head_extra; |
61989a80 | 897 | |
830e4bc5 MK |
898 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
899 | VM_BUG_ON_PAGE(first_page->inuse, first_page); | |
61989a80 | 900 | |
f4477e90 | 901 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 902 | |
f4477e90 | 903 | reset_page(first_page); |
61989a80 NG |
904 | __free_page(first_page); |
905 | ||
906 | /* zspage with only 1 system page */ | |
f4477e90 | 907 | if (!head_extra) |
61989a80 NG |
908 | return; |
909 | ||
f4477e90 | 910 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 911 | list_del(&nextp->lru); |
f4477e90 | 912 | reset_page(nextp); |
61989a80 NG |
913 | __free_page(nextp); |
914 | } | |
f4477e90 NG |
915 | reset_page(head_extra); |
916 | __free_page(head_extra); | |
61989a80 NG |
917 | } |
918 | ||
919 | /* Initialize a newly allocated zspage */ | |
251cbb95 | 920 | static void init_zspage(struct size_class *class, struct page *first_page) |
61989a80 NG |
921 | { |
922 | unsigned long off = 0; | |
923 | struct page *page = first_page; | |
924 | ||
830e4bc5 MK |
925 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
926 | ||
61989a80 NG |
927 | while (page) { |
928 | struct page *next_page; | |
929 | struct link_free *link; | |
5538c562 | 930 | unsigned int i = 1; |
af4ee5e9 | 931 | void *vaddr; |
61989a80 NG |
932 | |
933 | /* | |
934 | * page->index stores offset of first object starting | |
935 | * in the page. For the first page, this is always 0, | |
936 | * so we use first_page->index (aka ->freelist) to store | |
937 | * head of corresponding zspage's freelist. | |
938 | */ | |
939 | if (page != first_page) | |
940 | page->index = off; | |
941 | ||
af4ee5e9 MK |
942 | vaddr = kmap_atomic(page); |
943 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
944 | |
945 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 946 | link->next = location_to_obj(page, i++); |
5538c562 | 947 | link += class->size / sizeof(*link); |
61989a80 NG |
948 | } |
949 | ||
950 | /* | |
951 | * We now come to the last (full or partial) object on this | |
952 | * page, which must point to the first object on the next | |
953 | * page (if present) | |
954 | */ | |
955 | next_page = get_next_page(page); | |
312fcae2 | 956 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 957 | kunmap_atomic(vaddr); |
61989a80 | 958 | page = next_page; |
5538c562 | 959 | off %= PAGE_SIZE; |
61989a80 NG |
960 | } |
961 | } | |
962 | ||
963 | /* | |
964 | * Allocate a zspage for the given size class | |
965 | */ | |
966 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
967 | { | |
968 | int i, error; | |
b4b700c5 | 969 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
970 | |
971 | /* | |
972 | * Allocate individual pages and link them together as: | |
973 | * 1. first page->private = first sub-page | |
974 | * 2. all sub-pages are linked together using page->lru | |
32e7ba1e | 975 | * 3. each sub-page is linked to the first page using page->private |
61989a80 NG |
976 | * |
977 | * For each size class, First/Head pages are linked together using | |
978 | * page->lru. Also, we set PG_private to identify the first page | |
979 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
980 | * identify the last page. | |
981 | */ | |
982 | error = -ENOMEM; | |
2e3b6154 | 983 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 984 | struct page *page; |
61989a80 NG |
985 | |
986 | page = alloc_page(flags); | |
987 | if (!page) | |
988 | goto cleanup; | |
989 | ||
990 | INIT_LIST_HEAD(&page->lru); | |
991 | if (i == 0) { /* first page */ | |
a27545bf | 992 | SetPagePrivate(page); |
61989a80 NG |
993 | set_page_private(page, 0); |
994 | first_page = page; | |
995 | first_page->inuse = 0; | |
996 | } | |
997 | if (i == 1) | |
e842b976 | 998 | set_page_private(first_page, (unsigned long)page); |
61989a80 | 999 | if (i >= 1) |
32e7ba1e | 1000 | set_page_private(page, (unsigned long)first_page); |
61989a80 NG |
1001 | if (i >= 2) |
1002 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 1003 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 1004 | SetPagePrivate2(page); |
61989a80 NG |
1005 | prev_page = page; |
1006 | } | |
1007 | ||
251cbb95 | 1008 | init_zspage(class, first_page); |
61989a80 | 1009 | |
312fcae2 | 1010 | first_page->freelist = location_to_obj(first_page, 0); |
61989a80 | 1011 | /* Maximum number of objects we can store in this zspage */ |
2e3b6154 | 1012 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
1013 | |
1014 | error = 0; /* Success */ | |
1015 | ||
1016 | cleanup: | |
1017 | if (unlikely(error) && first_page) { | |
1018 | free_zspage(first_page); | |
1019 | first_page = NULL; | |
1020 | } | |
1021 | ||
1022 | return first_page; | |
1023 | } | |
1024 | ||
1025 | static struct page *find_get_zspage(struct size_class *class) | |
1026 | { | |
1027 | int i; | |
1028 | struct page *page; | |
1029 | ||
1030 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1031 | page = class->fullness_list[i]; | |
1032 | if (page) | |
1033 | break; | |
1034 | } | |
1035 | ||
1036 | return page; | |
1037 | } | |
1038 | ||
1b945aee | 1039 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1040 | static inline int __zs_cpu_up(struct mapping_area *area) |
1041 | { | |
1042 | /* | |
1043 | * Make sure we don't leak memory if a cpu UP notification | |
1044 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1045 | */ | |
1046 | if (area->vm) | |
1047 | return 0; | |
1048 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1049 | if (!area->vm) | |
1050 | return -ENOMEM; | |
1051 | return 0; | |
1052 | } | |
1053 | ||
1054 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1055 | { | |
1056 | if (area->vm) | |
1057 | free_vm_area(area->vm); | |
1058 | area->vm = NULL; | |
1059 | } | |
1060 | ||
1061 | static inline void *__zs_map_object(struct mapping_area *area, | |
1062 | struct page *pages[2], int off, int size) | |
1063 | { | |
f6f8ed47 | 1064 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1065 | area->vm_addr = area->vm->addr; |
1066 | return area->vm_addr + off; | |
1067 | } | |
1068 | ||
1069 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1070 | struct page *pages[2], int off, int size) | |
1071 | { | |
1072 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1073 | |
d95abbbb | 1074 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1075 | } |
1076 | ||
1b945aee | 1077 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1078 | |
1079 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1080 | { | |
1081 | /* | |
1082 | * Make sure we don't leak memory if a cpu UP notification | |
1083 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1084 | */ | |
1085 | if (area->vm_buf) | |
1086 | return 0; | |
40f9fb8c | 1087 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1088 | if (!area->vm_buf) |
1089 | return -ENOMEM; | |
1090 | return 0; | |
1091 | } | |
1092 | ||
1093 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1094 | { | |
40f9fb8c | 1095 | kfree(area->vm_buf); |
f553646a SJ |
1096 | area->vm_buf = NULL; |
1097 | } | |
1098 | ||
1099 | static void *__zs_map_object(struct mapping_area *area, | |
1100 | struct page *pages[2], int off, int size) | |
5f601902 | 1101 | { |
5f601902 SJ |
1102 | int sizes[2]; |
1103 | void *addr; | |
f553646a | 1104 | char *buf = area->vm_buf; |
5f601902 | 1105 | |
f553646a SJ |
1106 | /* disable page faults to match kmap_atomic() return conditions */ |
1107 | pagefault_disable(); | |
1108 | ||
1109 | /* no read fastpath */ | |
1110 | if (area->vm_mm == ZS_MM_WO) | |
1111 | goto out; | |
5f601902 SJ |
1112 | |
1113 | sizes[0] = PAGE_SIZE - off; | |
1114 | sizes[1] = size - sizes[0]; | |
1115 | ||
5f601902 SJ |
1116 | /* copy object to per-cpu buffer */ |
1117 | addr = kmap_atomic(pages[0]); | |
1118 | memcpy(buf, addr + off, sizes[0]); | |
1119 | kunmap_atomic(addr); | |
1120 | addr = kmap_atomic(pages[1]); | |
1121 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1122 | kunmap_atomic(addr); | |
f553646a SJ |
1123 | out: |
1124 | return area->vm_buf; | |
5f601902 SJ |
1125 | } |
1126 | ||
f553646a SJ |
1127 | static void __zs_unmap_object(struct mapping_area *area, |
1128 | struct page *pages[2], int off, int size) | |
5f601902 | 1129 | { |
5f601902 SJ |
1130 | int sizes[2]; |
1131 | void *addr; | |
2e40e163 | 1132 | char *buf; |
5f601902 | 1133 | |
f553646a SJ |
1134 | /* no write fastpath */ |
1135 | if (area->vm_mm == ZS_MM_RO) | |
1136 | goto out; | |
5f601902 | 1137 | |
7b60a685 | 1138 | buf = area->vm_buf; |
a82cbf07 YX |
1139 | buf = buf + ZS_HANDLE_SIZE; |
1140 | size -= ZS_HANDLE_SIZE; | |
1141 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1142 | |
5f601902 SJ |
1143 | sizes[0] = PAGE_SIZE - off; |
1144 | sizes[1] = size - sizes[0]; | |
1145 | ||
1146 | /* copy per-cpu buffer to object */ | |
1147 | addr = kmap_atomic(pages[0]); | |
1148 | memcpy(addr + off, buf, sizes[0]); | |
1149 | kunmap_atomic(addr); | |
1150 | addr = kmap_atomic(pages[1]); | |
1151 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1152 | kunmap_atomic(addr); | |
f553646a SJ |
1153 | |
1154 | out: | |
1155 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1156 | pagefault_enable(); | |
5f601902 | 1157 | } |
61989a80 | 1158 | |
1b945aee | 1159 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1160 | |
61989a80 NG |
1161 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1162 | void *pcpu) | |
1163 | { | |
f553646a | 1164 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1165 | struct mapping_area *area; |
1166 | ||
1167 | switch (action) { | |
1168 | case CPU_UP_PREPARE: | |
1169 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1170 | ret = __zs_cpu_up(area); |
1171 | if (ret) | |
1172 | return notifier_from_errno(ret); | |
61989a80 NG |
1173 | break; |
1174 | case CPU_DEAD: | |
1175 | case CPU_UP_CANCELED: | |
1176 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1177 | __zs_cpu_down(area); |
61989a80 NG |
1178 | break; |
1179 | } | |
1180 | ||
1181 | return NOTIFY_OK; | |
1182 | } | |
1183 | ||
1184 | static struct notifier_block zs_cpu_nb = { | |
1185 | .notifier_call = zs_cpu_notifier | |
1186 | }; | |
1187 | ||
b1b00a5b | 1188 | static int zs_register_cpu_notifier(void) |
61989a80 | 1189 | { |
b1b00a5b | 1190 | int cpu, uninitialized_var(ret); |
61989a80 | 1191 | |
f0e71fcd SB |
1192 | cpu_notifier_register_begin(); |
1193 | ||
1194 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1195 | for_each_online_cpu(cpu) { |
1196 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1197 | if (notifier_to_errno(ret)) |
1198 | break; | |
61989a80 | 1199 | } |
f0e71fcd SB |
1200 | |
1201 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1202 | return notifier_to_errno(ret); |
1203 | } | |
f0e71fcd | 1204 | |
66cdef66 | 1205 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1206 | { |
66cdef66 | 1207 | int cpu; |
40f9fb8c | 1208 | |
66cdef66 | 1209 | cpu_notifier_register_begin(); |
40f9fb8c | 1210 | |
66cdef66 GM |
1211 | for_each_online_cpu(cpu) |
1212 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1213 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1214 | |
66cdef66 | 1215 | cpu_notifier_register_done(); |
b1b00a5b SS |
1216 | } |
1217 | ||
66cdef66 | 1218 | static void init_zs_size_classes(void) |
b1b00a5b | 1219 | { |
66cdef66 | 1220 | int nr; |
c795779d | 1221 | |
66cdef66 GM |
1222 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1223 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1224 | nr += 1; | |
40f9fb8c | 1225 | |
66cdef66 | 1226 | zs_size_classes = nr; |
61989a80 NG |
1227 | } |
1228 | ||
9eec4cd5 JK |
1229 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1230 | { | |
1231 | if (prev->pages_per_zspage != pages_per_zspage) | |
1232 | return false; | |
1233 | ||
1234 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1235 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1236 | return false; | |
1237 | ||
1238 | return true; | |
1239 | } | |
1240 | ||
a4209467 | 1241 | static bool zspage_full(struct page *first_page) |
312fcae2 | 1242 | { |
830e4bc5 | 1243 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
312fcae2 | 1244 | |
a4209467 | 1245 | return first_page->inuse == first_page->objects; |
312fcae2 MK |
1246 | } |
1247 | ||
66cdef66 GM |
1248 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1249 | { | |
1250 | return atomic_long_read(&pool->pages_allocated); | |
1251 | } | |
1252 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1253 | ||
4bbc0bc0 | 1254 | /** |
66cdef66 GM |
1255 | * zs_map_object - get address of allocated object from handle. |
1256 | * @pool: pool from which the object was allocated | |
1257 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1258 | * |
66cdef66 GM |
1259 | * Before using an object allocated from zs_malloc, it must be mapped using |
1260 | * this function. When done with the object, it must be unmapped using | |
1261 | * zs_unmap_object. | |
4bbc0bc0 | 1262 | * |
66cdef66 GM |
1263 | * Only one object can be mapped per cpu at a time. There is no protection |
1264 | * against nested mappings. | |
1265 | * | |
1266 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1267 | */ |
66cdef66 GM |
1268 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1269 | enum zs_mapmode mm) | |
61989a80 | 1270 | { |
66cdef66 | 1271 | struct page *page; |
2e40e163 | 1272 | unsigned long obj, obj_idx, off; |
61989a80 | 1273 | |
66cdef66 GM |
1274 | unsigned int class_idx; |
1275 | enum fullness_group fg; | |
1276 | struct size_class *class; | |
1277 | struct mapping_area *area; | |
1278 | struct page *pages[2]; | |
2e40e163 | 1279 | void *ret; |
61989a80 | 1280 | |
9eec4cd5 | 1281 | /* |
66cdef66 GM |
1282 | * Because we use per-cpu mapping areas shared among the |
1283 | * pools/users, we can't allow mapping in interrupt context | |
1284 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1285 | */ |
830e4bc5 | 1286 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1287 | |
312fcae2 MK |
1288 | /* From now on, migration cannot move the object */ |
1289 | pin_tag(handle); | |
1290 | ||
2e40e163 MK |
1291 | obj = handle_to_obj(handle); |
1292 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1293 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1294 | class = pool->size_class[class_idx]; | |
1295 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1296 | |
66cdef66 GM |
1297 | area = &get_cpu_var(zs_map_area); |
1298 | area->vm_mm = mm; | |
1299 | if (off + class->size <= PAGE_SIZE) { | |
1300 | /* this object is contained entirely within a page */ | |
1301 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1302 | ret = area->vm_addr + off; |
1303 | goto out; | |
61989a80 NG |
1304 | } |
1305 | ||
66cdef66 GM |
1306 | /* this object spans two pages */ |
1307 | pages[0] = page; | |
1308 | pages[1] = get_next_page(page); | |
1309 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1310 | |
2e40e163 MK |
1311 | ret = __zs_map_object(area, pages, off, class->size); |
1312 | out: | |
7b60a685 MK |
1313 | if (!class->huge) |
1314 | ret += ZS_HANDLE_SIZE; | |
1315 | ||
1316 | return ret; | |
61989a80 | 1317 | } |
66cdef66 | 1318 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1319 | |
66cdef66 | 1320 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1321 | { |
66cdef66 | 1322 | struct page *page; |
2e40e163 | 1323 | unsigned long obj, obj_idx, off; |
61989a80 | 1324 | |
66cdef66 GM |
1325 | unsigned int class_idx; |
1326 | enum fullness_group fg; | |
1327 | struct size_class *class; | |
1328 | struct mapping_area *area; | |
9eec4cd5 | 1329 | |
2e40e163 MK |
1330 | obj = handle_to_obj(handle); |
1331 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1332 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1333 | class = pool->size_class[class_idx]; | |
1334 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1335 | |
66cdef66 GM |
1336 | area = this_cpu_ptr(&zs_map_area); |
1337 | if (off + class->size <= PAGE_SIZE) | |
1338 | kunmap_atomic(area->vm_addr); | |
1339 | else { | |
1340 | struct page *pages[2]; | |
40f9fb8c | 1341 | |
66cdef66 GM |
1342 | pages[0] = page; |
1343 | pages[1] = get_next_page(page); | |
1344 | BUG_ON(!pages[1]); | |
1345 | ||
1346 | __zs_unmap_object(area, pages, off, class->size); | |
1347 | } | |
1348 | put_cpu_var(zs_map_area); | |
312fcae2 | 1349 | unpin_tag(handle); |
61989a80 | 1350 | } |
66cdef66 | 1351 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1352 | |
251cbb95 MK |
1353 | static unsigned long obj_malloc(struct size_class *class, |
1354 | struct page *first_page, unsigned long handle) | |
c7806261 MK |
1355 | { |
1356 | unsigned long obj; | |
1357 | struct link_free *link; | |
1358 | ||
1359 | struct page *m_page; | |
1360 | unsigned long m_objidx, m_offset; | |
1361 | void *vaddr; | |
1362 | ||
312fcae2 | 1363 | handle |= OBJ_ALLOCATED_TAG; |
c7806261 MK |
1364 | obj = (unsigned long)first_page->freelist; |
1365 | obj_to_location(obj, &m_page, &m_objidx); | |
1366 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1367 | ||
1368 | vaddr = kmap_atomic(m_page); | |
1369 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1370 | first_page->freelist = link->next; | |
7b60a685 MK |
1371 | if (!class->huge) |
1372 | /* record handle in the header of allocated chunk */ | |
1373 | link->handle = handle; | |
1374 | else | |
1375 | /* record handle in first_page->private */ | |
1376 | set_page_private(first_page, handle); | |
c7806261 MK |
1377 | kunmap_atomic(vaddr); |
1378 | first_page->inuse++; | |
1379 | zs_stat_inc(class, OBJ_USED, 1); | |
1380 | ||
1381 | return obj; | |
1382 | } | |
1383 | ||
1384 | ||
61989a80 NG |
1385 | /** |
1386 | * zs_malloc - Allocate block of given size from pool. | |
1387 | * @pool: pool to allocate from | |
1388 | * @size: size of block to allocate | |
61989a80 | 1389 | * |
00a61d86 | 1390 | * On success, handle to the allocated object is returned, |
c2344348 | 1391 | * otherwise 0. |
61989a80 NG |
1392 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1393 | */ | |
d0d8da2d | 1394 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1395 | { |
2e40e163 | 1396 | unsigned long handle, obj; |
61989a80 | 1397 | struct size_class *class; |
c7806261 | 1398 | struct page *first_page; |
61989a80 | 1399 | |
7b60a685 | 1400 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1401 | return 0; |
1402 | ||
d0d8da2d | 1403 | handle = alloc_handle(pool, gfp); |
2e40e163 | 1404 | if (!handle) |
c2344348 | 1405 | return 0; |
61989a80 | 1406 | |
2e40e163 MK |
1407 | /* extra space in chunk to keep the handle */ |
1408 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1409 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1410 | |
1411 | spin_lock(&class->lock); | |
1412 | first_page = find_get_zspage(class); | |
1413 | ||
1414 | if (!first_page) { | |
1415 | spin_unlock(&class->lock); | |
d0d8da2d | 1416 | first_page = alloc_zspage(class, gfp); |
2e40e163 MK |
1417 | if (unlikely(!first_page)) { |
1418 | free_handle(pool, handle); | |
c2344348 | 1419 | return 0; |
2e40e163 | 1420 | } |
61989a80 NG |
1421 | |
1422 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1423 | atomic_long_add(class->pages_per_zspage, |
1424 | &pool->pages_allocated); | |
0f050d99 | 1425 | |
61989a80 | 1426 | spin_lock(&class->lock); |
0f050d99 GM |
1427 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1428 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1429 | } |
1430 | ||
251cbb95 | 1431 | obj = obj_malloc(class, first_page, handle); |
61989a80 | 1432 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1433 | fix_fullness_group(class, first_page); |
2e40e163 | 1434 | record_obj(handle, obj); |
61989a80 NG |
1435 | spin_unlock(&class->lock); |
1436 | ||
2e40e163 | 1437 | return handle; |
61989a80 NG |
1438 | } |
1439 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1440 | ||
1ee47165 | 1441 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1442 | { |
1443 | struct link_free *link; | |
1444 | struct page *first_page, *f_page; | |
c7806261 | 1445 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1446 | void *vaddr; |
61989a80 | 1447 | |
312fcae2 | 1448 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1449 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1450 | first_page = get_first_page(f_page); |
1451 | ||
61989a80 NG |
1452 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1453 | ||
c7806261 | 1454 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1455 | |
1456 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1457 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1458 | link->next = first_page->freelist; |
7b60a685 MK |
1459 | if (class->huge) |
1460 | set_page_private(first_page, 0); | |
af4ee5e9 | 1461 | kunmap_atomic(vaddr); |
c2344348 | 1462 | first_page->freelist = (void *)obj; |
61989a80 | 1463 | first_page->inuse--; |
0f050d99 | 1464 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1465 | } |
1466 | ||
1467 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1468 | { | |
1469 | struct page *first_page, *f_page; | |
1470 | unsigned long obj, f_objidx; | |
1471 | int class_idx; | |
1472 | struct size_class *class; | |
1473 | enum fullness_group fullness; | |
1474 | ||
1475 | if (unlikely(!handle)) | |
1476 | return; | |
1477 | ||
312fcae2 | 1478 | pin_tag(handle); |
c7806261 | 1479 | obj = handle_to_obj(handle); |
c7806261 MK |
1480 | obj_to_location(obj, &f_page, &f_objidx); |
1481 | first_page = get_first_page(f_page); | |
1482 | ||
1483 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1484 | class = pool->size_class[class_idx]; | |
1485 | ||
1486 | spin_lock(&class->lock); | |
1ee47165 | 1487 | obj_free(class, obj); |
c7806261 | 1488 | fullness = fix_fullness_group(class, first_page); |
312fcae2 | 1489 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1490 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1491 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1492 | atomic_long_sub(class->pages_per_zspage, |
1493 | &pool->pages_allocated); | |
1494 | free_zspage(first_page); | |
1495 | } | |
61989a80 | 1496 | spin_unlock(&class->lock); |
312fcae2 | 1497 | unpin_tag(handle); |
61989a80 | 1498 | |
312fcae2 MK |
1499 | free_handle(pool, handle); |
1500 | } | |
1501 | EXPORT_SYMBOL_GPL(zs_free); | |
1502 | ||
251cbb95 MK |
1503 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1504 | unsigned long src) | |
312fcae2 MK |
1505 | { |
1506 | struct page *s_page, *d_page; | |
1507 | unsigned long s_objidx, d_objidx; | |
1508 | unsigned long s_off, d_off; | |
1509 | void *s_addr, *d_addr; | |
1510 | int s_size, d_size, size; | |
1511 | int written = 0; | |
1512 | ||
1513 | s_size = d_size = class->size; | |
1514 | ||
1515 | obj_to_location(src, &s_page, &s_objidx); | |
1516 | obj_to_location(dst, &d_page, &d_objidx); | |
1517 | ||
1518 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1519 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1520 | ||
1521 | if (s_off + class->size > PAGE_SIZE) | |
1522 | s_size = PAGE_SIZE - s_off; | |
1523 | ||
1524 | if (d_off + class->size > PAGE_SIZE) | |
1525 | d_size = PAGE_SIZE - d_off; | |
1526 | ||
1527 | s_addr = kmap_atomic(s_page); | |
1528 | d_addr = kmap_atomic(d_page); | |
1529 | ||
1530 | while (1) { | |
1531 | size = min(s_size, d_size); | |
1532 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1533 | written += size; | |
1534 | ||
1535 | if (written == class->size) | |
1536 | break; | |
1537 | ||
495819ea SS |
1538 | s_off += size; |
1539 | s_size -= size; | |
1540 | d_off += size; | |
1541 | d_size -= size; | |
1542 | ||
1543 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1544 | kunmap_atomic(d_addr); |
1545 | kunmap_atomic(s_addr); | |
1546 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1547 | s_addr = kmap_atomic(s_page); |
1548 | d_addr = kmap_atomic(d_page); | |
1549 | s_size = class->size - written; | |
1550 | s_off = 0; | |
312fcae2 MK |
1551 | } |
1552 | ||
495819ea | 1553 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1554 | kunmap_atomic(d_addr); |
1555 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1556 | d_addr = kmap_atomic(d_page); |
1557 | d_size = class->size - written; | |
1558 | d_off = 0; | |
312fcae2 MK |
1559 | } |
1560 | } | |
1561 | ||
1562 | kunmap_atomic(d_addr); | |
1563 | kunmap_atomic(s_addr); | |
1564 | } | |
1565 | ||
1566 | /* | |
1567 | * Find alloced object in zspage from index object and | |
1568 | * return handle. | |
1569 | */ | |
251cbb95 MK |
1570 | static unsigned long find_alloced_obj(struct size_class *class, |
1571 | struct page *page, int index) | |
312fcae2 MK |
1572 | { |
1573 | unsigned long head; | |
1574 | int offset = 0; | |
1575 | unsigned long handle = 0; | |
1576 | void *addr = kmap_atomic(page); | |
1577 | ||
1578 | if (!is_first_page(page)) | |
1579 | offset = page->index; | |
1580 | offset += class->size * index; | |
1581 | ||
1582 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1583 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1584 | if (head & OBJ_ALLOCATED_TAG) { |
1585 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1586 | if (trypin_tag(handle)) | |
1587 | break; | |
1588 | handle = 0; | |
1589 | } | |
1590 | ||
1591 | offset += class->size; | |
1592 | index++; | |
1593 | } | |
1594 | ||
1595 | kunmap_atomic(addr); | |
1596 | return handle; | |
1597 | } | |
1598 | ||
1599 | struct zs_compact_control { | |
1600 | /* Source page for migration which could be a subpage of zspage. */ | |
1601 | struct page *s_page; | |
1602 | /* Destination page for migration which should be a first page | |
1603 | * of zspage. */ | |
1604 | struct page *d_page; | |
1605 | /* Starting object index within @s_page which used for live object | |
1606 | * in the subpage. */ | |
1607 | int index; | |
312fcae2 MK |
1608 | }; |
1609 | ||
1610 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1611 | struct zs_compact_control *cc) | |
1612 | { | |
1613 | unsigned long used_obj, free_obj; | |
1614 | unsigned long handle; | |
1615 | struct page *s_page = cc->s_page; | |
1616 | struct page *d_page = cc->d_page; | |
1617 | unsigned long index = cc->index; | |
312fcae2 MK |
1618 | int ret = 0; |
1619 | ||
1620 | while (1) { | |
251cbb95 | 1621 | handle = find_alloced_obj(class, s_page, index); |
312fcae2 MK |
1622 | if (!handle) { |
1623 | s_page = get_next_page(s_page); | |
1624 | if (!s_page) | |
1625 | break; | |
1626 | index = 0; | |
1627 | continue; | |
1628 | } | |
1629 | ||
1630 | /* Stop if there is no more space */ | |
1631 | if (zspage_full(d_page)) { | |
1632 | unpin_tag(handle); | |
1633 | ret = -ENOMEM; | |
1634 | break; | |
1635 | } | |
1636 | ||
1637 | used_obj = handle_to_obj(handle); | |
251cbb95 MK |
1638 | free_obj = obj_malloc(class, d_page, handle); |
1639 | zs_object_copy(class, free_obj, used_obj); | |
312fcae2 | 1640 | index++; |
c102f07c JL |
1641 | /* |
1642 | * record_obj updates handle's value to free_obj and it will | |
1643 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1644 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1645 | * let's keep the lock bit. | |
1646 | */ | |
1647 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1648 | record_obj(handle, free_obj); |
1649 | unpin_tag(handle); | |
1ee47165 | 1650 | obj_free(class, used_obj); |
312fcae2 MK |
1651 | } |
1652 | ||
1653 | /* Remember last position in this iteration */ | |
1654 | cc->s_page = s_page; | |
1655 | cc->index = index; | |
312fcae2 MK |
1656 | |
1657 | return ret; | |
1658 | } | |
1659 | ||
0dc63d48 | 1660 | static struct page *isolate_target_page(struct size_class *class) |
312fcae2 MK |
1661 | { |
1662 | int i; | |
1663 | struct page *page; | |
1664 | ||
1665 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1666 | page = class->fullness_list[i]; | |
1667 | if (page) { | |
251cbb95 | 1668 | remove_zspage(class, i, page); |
312fcae2 MK |
1669 | break; |
1670 | } | |
1671 | } | |
1672 | ||
1673 | return page; | |
1674 | } | |
1675 | ||
860c707d SS |
1676 | /* |
1677 | * putback_zspage - add @first_page into right class's fullness list | |
1678 | * @pool: target pool | |
1679 | * @class: destination class | |
1680 | * @first_page: target page | |
1681 | * | |
1682 | * Return @fist_page's fullness_group | |
1683 | */ | |
1684 | static enum fullness_group putback_zspage(struct zs_pool *pool, | |
1685 | struct size_class *class, | |
1686 | struct page *first_page) | |
312fcae2 | 1687 | { |
312fcae2 MK |
1688 | enum fullness_group fullness; |
1689 | ||
839373e6 | 1690 | fullness = get_fullness_group(first_page); |
251cbb95 | 1691 | insert_zspage(class, fullness, first_page); |
839373e6 MK |
1692 | set_zspage_mapping(first_page, class->index, fullness); |
1693 | ||
13de8933 | 1694 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1695 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1696 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1697 | atomic_long_sub(class->pages_per_zspage, |
1698 | &pool->pages_allocated); | |
312fcae2 | 1699 | |
61989a80 | 1700 | free_zspage(first_page); |
13de8933 | 1701 | } |
860c707d SS |
1702 | |
1703 | return fullness; | |
61989a80 | 1704 | } |
312fcae2 MK |
1705 | |
1706 | static struct page *isolate_source_page(struct size_class *class) | |
1707 | { | |
ad9d5e17 MK |
1708 | int i; |
1709 | struct page *page = NULL; | |
1710 | ||
1711 | for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) { | |
1712 | page = class->fullness_list[i]; | |
1713 | if (!page) | |
1714 | continue; | |
312fcae2 | 1715 | |
251cbb95 | 1716 | remove_zspage(class, i, page); |
ad9d5e17 MK |
1717 | break; |
1718 | } | |
312fcae2 MK |
1719 | |
1720 | return page; | |
1721 | } | |
1722 | ||
04f05909 SS |
1723 | /* |
1724 | * | |
1725 | * Based on the number of unused allocated objects calculate | |
1726 | * and return the number of pages that we can free. | |
04f05909 SS |
1727 | */ |
1728 | static unsigned long zs_can_compact(struct size_class *class) | |
1729 | { | |
1730 | unsigned long obj_wasted; | |
44f43e99 SS |
1731 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
1732 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 1733 | |
44f43e99 SS |
1734 | if (obj_allocated <= obj_used) |
1735 | return 0; | |
04f05909 | 1736 | |
44f43e99 | 1737 | obj_wasted = obj_allocated - obj_used; |
04f05909 SS |
1738 | obj_wasted /= get_maxobj_per_zspage(class->size, |
1739 | class->pages_per_zspage); | |
1740 | ||
6cbf16b3 | 1741 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
1742 | } |
1743 | ||
7d3f3938 | 1744 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1745 | { |
312fcae2 MK |
1746 | struct zs_compact_control cc; |
1747 | struct page *src_page; | |
1748 | struct page *dst_page = NULL; | |
312fcae2 | 1749 | |
312fcae2 MK |
1750 | spin_lock(&class->lock); |
1751 | while ((src_page = isolate_source_page(class))) { | |
1752 | ||
04f05909 SS |
1753 | if (!zs_can_compact(class)) |
1754 | break; | |
1755 | ||
312fcae2 MK |
1756 | cc.index = 0; |
1757 | cc.s_page = src_page; | |
1758 | ||
0dc63d48 | 1759 | while ((dst_page = isolate_target_page(class))) { |
312fcae2 MK |
1760 | cc.d_page = dst_page; |
1761 | /* | |
0dc63d48 SS |
1762 | * If there is no more space in dst_page, resched |
1763 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1764 | */ |
1765 | if (!migrate_zspage(pool, class, &cc)) | |
1766 | break; | |
1767 | ||
1768 | putback_zspage(pool, class, dst_page); | |
312fcae2 MK |
1769 | } |
1770 | ||
1771 | /* Stop if we couldn't find slot */ | |
1772 | if (dst_page == NULL) | |
1773 | break; | |
1774 | ||
1775 | putback_zspage(pool, class, dst_page); | |
860c707d | 1776 | if (putback_zspage(pool, class, src_page) == ZS_EMPTY) |
6cbf16b3 | 1777 | pool->stats.pages_compacted += class->pages_per_zspage; |
312fcae2 | 1778 | spin_unlock(&class->lock); |
312fcae2 MK |
1779 | cond_resched(); |
1780 | spin_lock(&class->lock); | |
1781 | } | |
1782 | ||
1783 | if (src_page) | |
1784 | putback_zspage(pool, class, src_page); | |
1785 | ||
7d3f3938 | 1786 | spin_unlock(&class->lock); |
312fcae2 MK |
1787 | } |
1788 | ||
1789 | unsigned long zs_compact(struct zs_pool *pool) | |
1790 | { | |
1791 | int i; | |
312fcae2 MK |
1792 | struct size_class *class; |
1793 | ||
1794 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1795 | class = pool->size_class[i]; | |
1796 | if (!class) | |
1797 | continue; | |
1798 | if (class->index != i) | |
1799 | continue; | |
7d3f3938 | 1800 | __zs_compact(pool, class); |
312fcae2 MK |
1801 | } |
1802 | ||
860c707d | 1803 | return pool->stats.pages_compacted; |
312fcae2 MK |
1804 | } |
1805 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1806 | |
7d3f3938 SS |
1807 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1808 | { | |
1809 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1810 | } | |
1811 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1812 | ||
ab9d306d SS |
1813 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
1814 | struct shrink_control *sc) | |
1815 | { | |
1816 | unsigned long pages_freed; | |
1817 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1818 | shrinker); | |
1819 | ||
1820 | pages_freed = pool->stats.pages_compacted; | |
1821 | /* | |
1822 | * Compact classes and calculate compaction delta. | |
1823 | * Can run concurrently with a manually triggered | |
1824 | * (by user) compaction. | |
1825 | */ | |
1826 | pages_freed = zs_compact(pool) - pages_freed; | |
1827 | ||
1828 | return pages_freed ? pages_freed : SHRINK_STOP; | |
1829 | } | |
1830 | ||
1831 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
1832 | struct shrink_control *sc) | |
1833 | { | |
1834 | int i; | |
1835 | struct size_class *class; | |
1836 | unsigned long pages_to_free = 0; | |
1837 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1838 | shrinker); | |
1839 | ||
ab9d306d SS |
1840 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1841 | class = pool->size_class[i]; | |
1842 | if (!class) | |
1843 | continue; | |
1844 | if (class->index != i) | |
1845 | continue; | |
1846 | ||
ab9d306d | 1847 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
1848 | } |
1849 | ||
1850 | return pages_to_free; | |
1851 | } | |
1852 | ||
1853 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
1854 | { | |
1855 | if (pool->shrinker_enabled) { | |
1856 | unregister_shrinker(&pool->shrinker); | |
1857 | pool->shrinker_enabled = false; | |
1858 | } | |
1859 | } | |
1860 | ||
1861 | static int zs_register_shrinker(struct zs_pool *pool) | |
1862 | { | |
1863 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
1864 | pool->shrinker.count_objects = zs_shrinker_count; | |
1865 | pool->shrinker.batch = 0; | |
1866 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
1867 | ||
1868 | return register_shrinker(&pool->shrinker); | |
1869 | } | |
1870 | ||
00a61d86 | 1871 | /** |
66cdef66 GM |
1872 | * zs_create_pool - Creates an allocation pool to work from. |
1873 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1874 | * |
66cdef66 GM |
1875 | * This function must be called before anything when using |
1876 | * the zsmalloc allocator. | |
166cfda7 | 1877 | * |
66cdef66 GM |
1878 | * On success, a pointer to the newly created pool is returned, |
1879 | * otherwise NULL. | |
396b7fd6 | 1880 | */ |
d0d8da2d | 1881 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 1882 | { |
66cdef66 GM |
1883 | int i; |
1884 | struct zs_pool *pool; | |
1885 | struct size_class *prev_class = NULL; | |
61989a80 | 1886 | |
66cdef66 GM |
1887 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1888 | if (!pool) | |
1889 | return NULL; | |
61989a80 | 1890 | |
66cdef66 GM |
1891 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1892 | GFP_KERNEL); | |
1893 | if (!pool->size_class) { | |
1894 | kfree(pool); | |
1895 | return NULL; | |
1896 | } | |
61989a80 | 1897 | |
2e40e163 MK |
1898 | pool->name = kstrdup(name, GFP_KERNEL); |
1899 | if (!pool->name) | |
1900 | goto err; | |
1901 | ||
1902 | if (create_handle_cache(pool)) | |
1903 | goto err; | |
1904 | ||
c60369f0 | 1905 | /* |
66cdef66 GM |
1906 | * Iterate reversly, because, size of size_class that we want to use |
1907 | * for merging should be larger or equal to current size. | |
c60369f0 | 1908 | */ |
66cdef66 GM |
1909 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1910 | int size; | |
1911 | int pages_per_zspage; | |
1912 | struct size_class *class; | |
c60369f0 | 1913 | |
66cdef66 GM |
1914 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1915 | if (size > ZS_MAX_ALLOC_SIZE) | |
1916 | size = ZS_MAX_ALLOC_SIZE; | |
1917 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1918 | |
66cdef66 GM |
1919 | /* |
1920 | * size_class is used for normal zsmalloc operation such | |
1921 | * as alloc/free for that size. Although it is natural that we | |
1922 | * have one size_class for each size, there is a chance that we | |
1923 | * can get more memory utilization if we use one size_class for | |
1924 | * many different sizes whose size_class have same | |
1925 | * characteristics. So, we makes size_class point to | |
1926 | * previous size_class if possible. | |
1927 | */ | |
1928 | if (prev_class) { | |
1929 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1930 | pool->size_class[i] = prev_class; | |
1931 | continue; | |
1932 | } | |
1933 | } | |
1934 | ||
1935 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1936 | if (!class) | |
1937 | goto err; | |
1938 | ||
1939 | class->size = size; | |
1940 | class->index = i; | |
1941 | class->pages_per_zspage = pages_per_zspage; | |
7b60a685 MK |
1942 | if (pages_per_zspage == 1 && |
1943 | get_maxobj_per_zspage(size, pages_per_zspage) == 1) | |
1944 | class->huge = true; | |
66cdef66 GM |
1945 | spin_lock_init(&class->lock); |
1946 | pool->size_class[i] = class; | |
1947 | ||
1948 | prev_class = class; | |
61989a80 NG |
1949 | } |
1950 | ||
d34f6157 DS |
1951 | /* debug only, don't abort if it fails */ |
1952 | zs_pool_stat_create(pool, name); | |
0f050d99 | 1953 | |
ab9d306d SS |
1954 | /* |
1955 | * Not critical, we still can use the pool | |
1956 | * and user can trigger compaction manually. | |
1957 | */ | |
1958 | if (zs_register_shrinker(pool) == 0) | |
1959 | pool->shrinker_enabled = true; | |
66cdef66 GM |
1960 | return pool; |
1961 | ||
1962 | err: | |
1963 | zs_destroy_pool(pool); | |
1964 | return NULL; | |
61989a80 | 1965 | } |
66cdef66 | 1966 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1967 | |
66cdef66 | 1968 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1969 | { |
66cdef66 | 1970 | int i; |
61989a80 | 1971 | |
ab9d306d | 1972 | zs_unregister_shrinker(pool); |
0f050d99 GM |
1973 | zs_pool_stat_destroy(pool); |
1974 | ||
66cdef66 GM |
1975 | for (i = 0; i < zs_size_classes; i++) { |
1976 | int fg; | |
1977 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1978 | |
66cdef66 GM |
1979 | if (!class) |
1980 | continue; | |
61989a80 | 1981 | |
66cdef66 GM |
1982 | if (class->index != i) |
1983 | continue; | |
61989a80 | 1984 | |
66cdef66 GM |
1985 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1986 | if (class->fullness_list[fg]) { | |
1987 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1988 | class->size, fg); | |
1989 | } | |
1990 | } | |
1991 | kfree(class); | |
1992 | } | |
f553646a | 1993 | |
2e40e163 | 1994 | destroy_handle_cache(pool); |
66cdef66 | 1995 | kfree(pool->size_class); |
0f050d99 | 1996 | kfree(pool->name); |
66cdef66 GM |
1997 | kfree(pool); |
1998 | } | |
1999 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 2000 | |
66cdef66 GM |
2001 | static int __init zs_init(void) |
2002 | { | |
2003 | int ret = zs_register_cpu_notifier(); | |
2004 | ||
0f050d99 GM |
2005 | if (ret) |
2006 | goto notifier_fail; | |
66cdef66 GM |
2007 | |
2008 | init_zs_size_classes(); | |
2009 | ||
2010 | #ifdef CONFIG_ZPOOL | |
2011 | zpool_register_driver(&zs_zpool_driver); | |
2012 | #endif | |
0f050d99 GM |
2013 | |
2014 | ret = zs_stat_init(); | |
2015 | if (ret) { | |
2016 | pr_err("zs stat initialization failed\n"); | |
2017 | goto stat_fail; | |
2018 | } | |
66cdef66 | 2019 | return 0; |
0f050d99 GM |
2020 | |
2021 | stat_fail: | |
2022 | #ifdef CONFIG_ZPOOL | |
2023 | zpool_unregister_driver(&zs_zpool_driver); | |
2024 | #endif | |
2025 | notifier_fail: | |
2026 | zs_unregister_cpu_notifier(); | |
2027 | ||
2028 | return ret; | |
61989a80 | 2029 | } |
61989a80 | 2030 | |
66cdef66 | 2031 | static void __exit zs_exit(void) |
61989a80 | 2032 | { |
66cdef66 GM |
2033 | #ifdef CONFIG_ZPOOL |
2034 | zpool_unregister_driver(&zs_zpool_driver); | |
2035 | #endif | |
2036 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
2037 | |
2038 | zs_stat_exit(); | |
61989a80 | 2039 | } |
069f101f BH |
2040 | |
2041 | module_init(zs_init); | |
2042 | module_exit(zs_exit); | |
2043 | ||
2044 | MODULE_LICENSE("Dual BSD/GPL"); | |
2045 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |