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. | |
2db51dae NG |
35 | * page->lru: links together first pages of various zspages. |
36 | * Basically forming list of zspages in a fullness group. | |
37 | * page->mapping: class index and fullness group of the zspage | |
8f958c98 | 38 | * page->inuse: the number of objects that are used in this zspage |
2db51dae NG |
39 | * |
40 | * Usage of struct page flags: | |
41 | * PG_private: identifies the first component page | |
42 | * PG_private2: identifies the last component page | |
43 | * | |
44 | */ | |
45 | ||
4abaac9b DS |
46 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
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; | |
1fc6e27d | 214 | int objs_per_zspage; |
0959c63f SJ |
215 | unsigned int index; |
216 | ||
0f050d99 | 217 | struct zs_size_stat stats; |
0959c63f | 218 | |
7dfa4612 WY |
219 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ |
220 | int pages_per_zspage; | |
57244594 SS |
221 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
222 | bool huge; | |
0959c63f SJ |
223 | }; |
224 | ||
225 | /* | |
226 | * Placed within free objects to form a singly linked list. | |
227 | * For every zspage, first_page->freelist gives head of this list. | |
228 | * | |
229 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
230 | */ | |
231 | struct link_free { | |
2e40e163 MK |
232 | union { |
233 | /* | |
234 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
235 | * It's valid for non-allocated object | |
236 | */ | |
237 | void *next; | |
238 | /* | |
239 | * Handle of allocated object. | |
240 | */ | |
241 | unsigned long handle; | |
242 | }; | |
0959c63f SJ |
243 | }; |
244 | ||
245 | struct zs_pool { | |
6f3526d6 | 246 | const char *name; |
0f050d99 | 247 | |
40f9fb8c | 248 | struct size_class **size_class; |
2e40e163 | 249 | struct kmem_cache *handle_cachep; |
0959c63f | 250 | |
13de8933 | 251 | atomic_long_t pages_allocated; |
0f050d99 | 252 | |
7d3f3938 | 253 | struct zs_pool_stats stats; |
ab9d306d SS |
254 | |
255 | /* Compact classes */ | |
256 | struct shrinker shrinker; | |
257 | /* | |
258 | * To signify that register_shrinker() was successful | |
259 | * and unregister_shrinker() will not Oops. | |
260 | */ | |
261 | bool shrinker_enabled; | |
0f050d99 GM |
262 | #ifdef CONFIG_ZSMALLOC_STAT |
263 | struct dentry *stat_dentry; | |
264 | #endif | |
0959c63f | 265 | }; |
61989a80 NG |
266 | |
267 | /* | |
268 | * A zspage's class index and fullness group | |
269 | * are encoded in its (first)page->mapping | |
270 | */ | |
271 | #define CLASS_IDX_BITS 28 | |
272 | #define FULLNESS_BITS 4 | |
273 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
274 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
275 | ||
f553646a | 276 | struct mapping_area { |
1b945aee | 277 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
278 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
279 | #else | |
280 | char *vm_buf; /* copy buffer for objects that span pages */ | |
281 | #endif | |
282 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
283 | enum zs_mapmode vm_mm; /* mapping mode */ | |
284 | }; | |
285 | ||
2e40e163 MK |
286 | static int create_handle_cache(struct zs_pool *pool) |
287 | { | |
288 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
289 | 0, 0, NULL); | |
290 | return pool->handle_cachep ? 0 : 1; | |
291 | } | |
292 | ||
293 | static void destroy_handle_cache(struct zs_pool *pool) | |
294 | { | |
cd10add0 | 295 | kmem_cache_destroy(pool->handle_cachep); |
2e40e163 MK |
296 | } |
297 | ||
d0d8da2d | 298 | static unsigned long alloc_handle(struct zs_pool *pool, gfp_t gfp) |
2e40e163 MK |
299 | { |
300 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
d0d8da2d | 301 | gfp & ~__GFP_HIGHMEM); |
2e40e163 MK |
302 | } |
303 | ||
304 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
305 | { | |
306 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
307 | } | |
308 | ||
309 | static void record_obj(unsigned long handle, unsigned long obj) | |
310 | { | |
c102f07c JL |
311 | /* |
312 | * lsb of @obj represents handle lock while other bits | |
313 | * represent object value the handle is pointing so | |
314 | * updating shouldn't do store tearing. | |
315 | */ | |
316 | WRITE_ONCE(*(unsigned long *)handle, obj); | |
2e40e163 MK |
317 | } |
318 | ||
c795779d DS |
319 | /* zpool driver */ |
320 | ||
321 | #ifdef CONFIG_ZPOOL | |
322 | ||
6f3526d6 | 323 | static void *zs_zpool_create(const char *name, gfp_t gfp, |
78672779 | 324 | const struct zpool_ops *zpool_ops, |
479305fd | 325 | struct zpool *zpool) |
c795779d | 326 | { |
d0d8da2d SS |
327 | /* |
328 | * Ignore global gfp flags: zs_malloc() may be invoked from | |
329 | * different contexts and its caller must provide a valid | |
330 | * gfp mask. | |
331 | */ | |
332 | return zs_create_pool(name); | |
c795779d DS |
333 | } |
334 | ||
335 | static void zs_zpool_destroy(void *pool) | |
336 | { | |
337 | zs_destroy_pool(pool); | |
338 | } | |
339 | ||
340 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
341 | unsigned long *handle) | |
342 | { | |
d0d8da2d | 343 | *handle = zs_malloc(pool, size, gfp); |
c795779d DS |
344 | return *handle ? 0 : -1; |
345 | } | |
346 | static void zs_zpool_free(void *pool, unsigned long handle) | |
347 | { | |
348 | zs_free(pool, handle); | |
349 | } | |
350 | ||
351 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
352 | unsigned int *reclaimed) | |
353 | { | |
354 | return -EINVAL; | |
355 | } | |
356 | ||
357 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
358 | enum zpool_mapmode mm) | |
359 | { | |
360 | enum zs_mapmode zs_mm; | |
361 | ||
362 | switch (mm) { | |
363 | case ZPOOL_MM_RO: | |
364 | zs_mm = ZS_MM_RO; | |
365 | break; | |
366 | case ZPOOL_MM_WO: | |
367 | zs_mm = ZS_MM_WO; | |
368 | break; | |
369 | case ZPOOL_MM_RW: /* fallthru */ | |
370 | default: | |
371 | zs_mm = ZS_MM_RW; | |
372 | break; | |
373 | } | |
374 | ||
375 | return zs_map_object(pool, handle, zs_mm); | |
376 | } | |
377 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
378 | { | |
379 | zs_unmap_object(pool, handle); | |
380 | } | |
381 | ||
382 | static u64 zs_zpool_total_size(void *pool) | |
383 | { | |
722cdc17 | 384 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
385 | } |
386 | ||
387 | static struct zpool_driver zs_zpool_driver = { | |
388 | .type = "zsmalloc", | |
389 | .owner = THIS_MODULE, | |
390 | .create = zs_zpool_create, | |
391 | .destroy = zs_zpool_destroy, | |
392 | .malloc = zs_zpool_malloc, | |
393 | .free = zs_zpool_free, | |
394 | .shrink = zs_zpool_shrink, | |
395 | .map = zs_zpool_map, | |
396 | .unmap = zs_zpool_unmap, | |
397 | .total_size = zs_zpool_total_size, | |
398 | }; | |
399 | ||
137f8cff | 400 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
401 | #endif /* CONFIG_ZPOOL */ |
402 | ||
248ca1b0 MK |
403 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
404 | { | |
405 | return pages_per_zspage * PAGE_SIZE / size; | |
406 | } | |
407 | ||
61989a80 NG |
408 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
409 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
410 | ||
411 | static int is_first_page(struct page *page) | |
412 | { | |
a27545bf | 413 | return PagePrivate(page); |
61989a80 NG |
414 | } |
415 | ||
416 | static int is_last_page(struct page *page) | |
417 | { | |
a27545bf | 418 | return PagePrivate2(page); |
61989a80 NG |
419 | } |
420 | ||
a4209467 MK |
421 | static void get_zspage_mapping(struct page *first_page, |
422 | unsigned int *class_idx, | |
61989a80 NG |
423 | enum fullness_group *fullness) |
424 | { | |
425 | unsigned long m; | |
830e4bc5 | 426 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 | 427 | |
a4209467 | 428 | m = (unsigned long)first_page->mapping; |
61989a80 NG |
429 | *fullness = m & FULLNESS_MASK; |
430 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
431 | } | |
432 | ||
a4209467 MK |
433 | static void set_zspage_mapping(struct page *first_page, |
434 | unsigned int class_idx, | |
61989a80 NG |
435 | enum fullness_group fullness) |
436 | { | |
437 | unsigned long m; | |
830e4bc5 | 438 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
439 | |
440 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
441 | (fullness & FULLNESS_MASK); | |
a4209467 | 442 | first_page->mapping = (struct address_space *)m; |
61989a80 NG |
443 | } |
444 | ||
c3e3e88a NC |
445 | /* |
446 | * zsmalloc divides the pool into various size classes where each | |
447 | * class maintains a list of zspages where each zspage is divided | |
448 | * into equal sized chunks. Each allocation falls into one of these | |
449 | * classes depending on its size. This function returns index of the | |
450 | * size class which has chunk size big enough to hold the give size. | |
451 | */ | |
61989a80 NG |
452 | static int get_size_class_index(int size) |
453 | { | |
454 | int idx = 0; | |
455 | ||
456 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
457 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
458 | ZS_SIZE_CLASS_DELTA); | |
459 | ||
7b60a685 | 460 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
461 | } |
462 | ||
248ca1b0 MK |
463 | static inline void zs_stat_inc(struct size_class *class, |
464 | enum zs_stat_type type, unsigned long cnt) | |
465 | { | |
6fe5186f SS |
466 | if (type < NR_ZS_STAT_TYPE) |
467 | class->stats.objs[type] += cnt; | |
248ca1b0 MK |
468 | } |
469 | ||
470 | static inline void zs_stat_dec(struct size_class *class, | |
471 | enum zs_stat_type type, unsigned long cnt) | |
472 | { | |
6fe5186f SS |
473 | if (type < NR_ZS_STAT_TYPE) |
474 | class->stats.objs[type] -= cnt; | |
248ca1b0 MK |
475 | } |
476 | ||
477 | static inline unsigned long zs_stat_get(struct size_class *class, | |
478 | enum zs_stat_type type) | |
479 | { | |
6fe5186f SS |
480 | if (type < NR_ZS_STAT_TYPE) |
481 | return class->stats.objs[type]; | |
482 | return 0; | |
248ca1b0 MK |
483 | } |
484 | ||
57244594 SS |
485 | #ifdef CONFIG_ZSMALLOC_STAT |
486 | ||
4abaac9b | 487 | static void __init zs_stat_init(void) |
248ca1b0 | 488 | { |
4abaac9b DS |
489 | if (!debugfs_initialized()) { |
490 | pr_warn("debugfs not available, stat dir not created\n"); | |
491 | return; | |
492 | } | |
248ca1b0 MK |
493 | |
494 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
495 | if (!zs_stat_root) | |
4abaac9b | 496 | pr_warn("debugfs 'zsmalloc' stat dir creation failed\n"); |
248ca1b0 MK |
497 | } |
498 | ||
499 | static void __exit zs_stat_exit(void) | |
500 | { | |
501 | debugfs_remove_recursive(zs_stat_root); | |
502 | } | |
503 | ||
1120ed54 SS |
504 | static unsigned long zs_can_compact(struct size_class *class); |
505 | ||
248ca1b0 MK |
506 | static int zs_stats_size_show(struct seq_file *s, void *v) |
507 | { | |
508 | int i; | |
509 | struct zs_pool *pool = s->private; | |
510 | struct size_class *class; | |
511 | int objs_per_zspage; | |
512 | unsigned long class_almost_full, class_almost_empty; | |
1120ed54 | 513 | unsigned long obj_allocated, obj_used, pages_used, freeable; |
248ca1b0 MK |
514 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; |
515 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1120ed54 | 516 | unsigned long total_freeable = 0; |
248ca1b0 | 517 | |
1120ed54 | 518 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n", |
248ca1b0 MK |
519 | "class", "size", "almost_full", "almost_empty", |
520 | "obj_allocated", "obj_used", "pages_used", | |
1120ed54 | 521 | "pages_per_zspage", "freeable"); |
248ca1b0 MK |
522 | |
523 | for (i = 0; i < zs_size_classes; i++) { | |
524 | class = pool->size_class[i]; | |
525 | ||
526 | if (class->index != i) | |
527 | continue; | |
528 | ||
529 | spin_lock(&class->lock); | |
530 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
531 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
532 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
533 | obj_used = zs_stat_get(class, OBJ_USED); | |
1120ed54 | 534 | freeable = zs_can_compact(class); |
248ca1b0 MK |
535 | spin_unlock(&class->lock); |
536 | ||
537 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
538 | class->pages_per_zspage); | |
539 | pages_used = obj_allocated / objs_per_zspage * | |
540 | class->pages_per_zspage; | |
541 | ||
1120ed54 SS |
542 | seq_printf(s, " %5u %5u %11lu %12lu %13lu" |
543 | " %10lu %10lu %16d %8lu\n", | |
248ca1b0 MK |
544 | i, class->size, class_almost_full, class_almost_empty, |
545 | obj_allocated, obj_used, pages_used, | |
1120ed54 | 546 | class->pages_per_zspage, freeable); |
248ca1b0 MK |
547 | |
548 | total_class_almost_full += class_almost_full; | |
549 | total_class_almost_empty += class_almost_empty; | |
550 | total_objs += obj_allocated; | |
551 | total_used_objs += obj_used; | |
552 | total_pages += pages_used; | |
1120ed54 | 553 | total_freeable += freeable; |
248ca1b0 MK |
554 | } |
555 | ||
556 | seq_puts(s, "\n"); | |
1120ed54 | 557 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n", |
248ca1b0 MK |
558 | "Total", "", total_class_almost_full, |
559 | total_class_almost_empty, total_objs, | |
1120ed54 | 560 | total_used_objs, total_pages, "", total_freeable); |
248ca1b0 MK |
561 | |
562 | return 0; | |
563 | } | |
564 | ||
565 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
566 | { | |
567 | return single_open(file, zs_stats_size_show, inode->i_private); | |
568 | } | |
569 | ||
570 | static const struct file_operations zs_stat_size_ops = { | |
571 | .open = zs_stats_size_open, | |
572 | .read = seq_read, | |
573 | .llseek = seq_lseek, | |
574 | .release = single_release, | |
575 | }; | |
576 | ||
d34f6157 | 577 | static void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 MK |
578 | { |
579 | struct dentry *entry; | |
580 | ||
4abaac9b DS |
581 | if (!zs_stat_root) { |
582 | pr_warn("no root stat dir, not creating <%s> stat dir\n", name); | |
d34f6157 | 583 | return; |
4abaac9b | 584 | } |
248ca1b0 MK |
585 | |
586 | entry = debugfs_create_dir(name, zs_stat_root); | |
587 | if (!entry) { | |
588 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
d34f6157 | 589 | return; |
248ca1b0 MK |
590 | } |
591 | pool->stat_dentry = entry; | |
592 | ||
593 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
594 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
595 | if (!entry) { | |
596 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
597 | name, "classes"); | |
4abaac9b DS |
598 | debugfs_remove_recursive(pool->stat_dentry); |
599 | pool->stat_dentry = NULL; | |
248ca1b0 | 600 | } |
248ca1b0 MK |
601 | } |
602 | ||
603 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
604 | { | |
605 | debugfs_remove_recursive(pool->stat_dentry); | |
606 | } | |
607 | ||
608 | #else /* CONFIG_ZSMALLOC_STAT */ | |
4abaac9b | 609 | static void __init zs_stat_init(void) |
248ca1b0 | 610 | { |
248ca1b0 MK |
611 | } |
612 | ||
613 | static void __exit zs_stat_exit(void) | |
614 | { | |
615 | } | |
616 | ||
d34f6157 | 617 | static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name) |
248ca1b0 | 618 | { |
248ca1b0 MK |
619 | } |
620 | ||
621 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
622 | { | |
623 | } | |
248ca1b0 MK |
624 | #endif |
625 | ||
c3e3e88a NC |
626 | /* |
627 | * For each size class, zspages are divided into different groups | |
628 | * depending on how "full" they are. This was done so that we could | |
629 | * easily find empty or nearly empty zspages when we try to shrink | |
630 | * the pool (not yet implemented). This function returns fullness | |
631 | * status of the given page. | |
632 | */ | |
1fc6e27d MK |
633 | static enum fullness_group get_fullness_group(struct size_class *class, |
634 | struct page *first_page) | |
61989a80 | 635 | { |
1fc6e27d | 636 | int inuse, objs_per_zspage; |
61989a80 | 637 | enum fullness_group fg; |
830e4bc5 MK |
638 | |
639 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); | |
61989a80 | 640 | |
a4209467 | 641 | inuse = first_page->inuse; |
1fc6e27d | 642 | objs_per_zspage = class->objs_per_zspage; |
61989a80 NG |
643 | |
644 | if (inuse == 0) | |
645 | fg = ZS_EMPTY; | |
1fc6e27d | 646 | else if (inuse == objs_per_zspage) |
61989a80 | 647 | fg = ZS_FULL; |
1fc6e27d | 648 | else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac) |
61989a80 NG |
649 | fg = ZS_ALMOST_EMPTY; |
650 | else | |
651 | fg = ZS_ALMOST_FULL; | |
652 | ||
653 | return fg; | |
654 | } | |
655 | ||
c3e3e88a NC |
656 | /* |
657 | * Each size class maintains various freelists and zspages are assigned | |
658 | * to one of these freelists based on the number of live objects they | |
659 | * have. This functions inserts the given zspage into the freelist | |
660 | * identified by <class, fullness_group>. | |
661 | */ | |
251cbb95 MK |
662 | static void insert_zspage(struct size_class *class, |
663 | enum fullness_group fullness, | |
664 | struct page *first_page) | |
61989a80 NG |
665 | { |
666 | struct page **head; | |
667 | ||
830e4bc5 | 668 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
669 | |
670 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
671 | return; | |
672 | ||
248ca1b0 MK |
673 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
674 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
58f17117 SS |
675 | |
676 | head = &class->fullness_list[fullness]; | |
677 | if (!*head) { | |
a4209467 | 678 | *head = first_page; |
58f17117 SS |
679 | return; |
680 | } | |
681 | ||
682 | /* | |
683 | * We want to see more ZS_FULL pages and less almost | |
684 | * empty/full. Put pages with higher ->inuse first. | |
685 | */ | |
a4209467 MK |
686 | list_add_tail(&first_page->lru, &(*head)->lru); |
687 | if (first_page->inuse >= (*head)->inuse) | |
688 | *head = first_page; | |
61989a80 NG |
689 | } |
690 | ||
c3e3e88a NC |
691 | /* |
692 | * This function removes the given zspage from the freelist identified | |
693 | * by <class, fullness_group>. | |
694 | */ | |
251cbb95 MK |
695 | static void remove_zspage(struct size_class *class, |
696 | enum fullness_group fullness, | |
697 | struct page *first_page) | |
61989a80 NG |
698 | { |
699 | struct page **head; | |
700 | ||
830e4bc5 | 701 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
61989a80 NG |
702 | |
703 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
704 | return; | |
705 | ||
706 | head = &class->fullness_list[fullness]; | |
830e4bc5 | 707 | VM_BUG_ON_PAGE(!*head, first_page); |
61989a80 NG |
708 | if (list_empty(&(*head)->lru)) |
709 | *head = NULL; | |
a4209467 | 710 | else if (*head == first_page) |
61989a80 NG |
711 | *head = (struct page *)list_entry((*head)->lru.next, |
712 | struct page, lru); | |
713 | ||
a4209467 | 714 | list_del_init(&first_page->lru); |
248ca1b0 MK |
715 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
716 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
717 | } |
718 | ||
c3e3e88a NC |
719 | /* |
720 | * Each size class maintains zspages in different fullness groups depending | |
721 | * on the number of live objects they contain. When allocating or freeing | |
722 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
723 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
724 | * a status change has occurred for the given page and accordingly moves the | |
725 | * page from the freelist of the old fullness group to that of the new | |
726 | * fullness group. | |
727 | */ | |
c7806261 | 728 | static enum fullness_group fix_fullness_group(struct size_class *class, |
a4209467 | 729 | struct page *first_page) |
61989a80 NG |
730 | { |
731 | int class_idx; | |
61989a80 NG |
732 | enum fullness_group currfg, newfg; |
733 | ||
a4209467 | 734 | get_zspage_mapping(first_page, &class_idx, &currfg); |
1fc6e27d | 735 | newfg = get_fullness_group(class, first_page); |
61989a80 NG |
736 | if (newfg == currfg) |
737 | goto out; | |
738 | ||
251cbb95 MK |
739 | remove_zspage(class, currfg, first_page); |
740 | insert_zspage(class, newfg, first_page); | |
a4209467 | 741 | set_zspage_mapping(first_page, class_idx, newfg); |
61989a80 NG |
742 | |
743 | out: | |
744 | return newfg; | |
745 | } | |
746 | ||
747 | /* | |
748 | * We have to decide on how many pages to link together | |
749 | * to form a zspage for each size class. This is important | |
750 | * to reduce wastage due to unusable space left at end of | |
751 | * each zspage which is given as: | |
888fa374 YX |
752 | * wastage = Zp % class_size |
753 | * usage = Zp - wastage | |
61989a80 NG |
754 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
755 | * | |
756 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
757 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
758 | * since then we can perfectly fit in 8 such objects. | |
759 | */ | |
2e3b6154 | 760 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
761 | { |
762 | int i, max_usedpc = 0; | |
763 | /* zspage order which gives maximum used size per KB */ | |
764 | int max_usedpc_order = 1; | |
765 | ||
84d4faab | 766 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
767 | int zspage_size; |
768 | int waste, usedpc; | |
769 | ||
770 | zspage_size = i * PAGE_SIZE; | |
771 | waste = zspage_size % class_size; | |
772 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
773 | ||
774 | if (usedpc > max_usedpc) { | |
775 | max_usedpc = usedpc; | |
776 | max_usedpc_order = i; | |
777 | } | |
778 | } | |
779 | ||
780 | return max_usedpc_order; | |
781 | } | |
782 | ||
783 | /* | |
784 | * A single 'zspage' is composed of many system pages which are | |
785 | * linked together using fields in struct page. This function finds | |
786 | * the first/head page, given any component page of a zspage. | |
787 | */ | |
788 | static struct page *get_first_page(struct page *page) | |
789 | { | |
790 | if (is_first_page(page)) | |
791 | return page; | |
792 | else | |
32e7ba1e | 793 | return (struct page *)page_private(page); |
61989a80 NG |
794 | } |
795 | ||
796 | static struct page *get_next_page(struct page *page) | |
797 | { | |
798 | struct page *next; | |
799 | ||
800 | if (is_last_page(page)) | |
801 | next = NULL; | |
802 | else if (is_first_page(page)) | |
e842b976 | 803 | next = (struct page *)page_private(page); |
61989a80 NG |
804 | else |
805 | next = list_entry(page->lru.next, struct page, lru); | |
806 | ||
807 | return next; | |
808 | } | |
809 | ||
67296874 OH |
810 | /* |
811 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 812 | * We use the least bit of handle for tagging. |
67296874 | 813 | */ |
312fcae2 | 814 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 815 | { |
312fcae2 | 816 | unsigned long obj; |
61989a80 NG |
817 | |
818 | if (!page) { | |
830e4bc5 | 819 | VM_BUG_ON(obj_idx); |
61989a80 NG |
820 | return NULL; |
821 | } | |
822 | ||
312fcae2 MK |
823 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
824 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
825 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 826 | |
312fcae2 | 827 | return (void *)obj; |
61989a80 NG |
828 | } |
829 | ||
67296874 OH |
830 | /* |
831 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
832 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 833 | * location_to_obj(). |
67296874 | 834 | */ |
312fcae2 | 835 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
836 | unsigned long *obj_idx) |
837 | { | |
312fcae2 MK |
838 | obj >>= OBJ_TAG_BITS; |
839 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
840 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
841 | } |
842 | ||
2e40e163 MK |
843 | static unsigned long handle_to_obj(unsigned long handle) |
844 | { | |
845 | return *(unsigned long *)handle; | |
846 | } | |
847 | ||
7b60a685 MK |
848 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
849 | void *obj) | |
312fcae2 | 850 | { |
7b60a685 | 851 | if (class->huge) { |
830e4bc5 | 852 | VM_BUG_ON_PAGE(!is_first_page(page), page); |
12a7bfad | 853 | return page_private(page); |
7b60a685 MK |
854 | } else |
855 | return *(unsigned long *)obj; | |
312fcae2 MK |
856 | } |
857 | ||
61989a80 NG |
858 | static unsigned long obj_idx_to_offset(struct page *page, |
859 | unsigned long obj_idx, int class_size) | |
860 | { | |
861 | unsigned long off = 0; | |
862 | ||
863 | if (!is_first_page(page)) | |
864 | off = page->index; | |
865 | ||
866 | return off + obj_idx * class_size; | |
867 | } | |
868 | ||
312fcae2 MK |
869 | static inline int trypin_tag(unsigned long handle) |
870 | { | |
1b8320b6 | 871 | return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
872 | } |
873 | ||
874 | static void pin_tag(unsigned long handle) | |
875 | { | |
1b8320b6 | 876 | bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
877 | } |
878 | ||
879 | static void unpin_tag(unsigned long handle) | |
880 | { | |
1b8320b6 | 881 | bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle); |
312fcae2 MK |
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 NG |
1011 | error = 0; /* Success */ |
1012 | ||
1013 | cleanup: | |
1014 | if (unlikely(error) && first_page) { | |
1015 | free_zspage(first_page); | |
1016 | first_page = NULL; | |
1017 | } | |
1018 | ||
1019 | return first_page; | |
1020 | } | |
1021 | ||
1022 | static struct page *find_get_zspage(struct size_class *class) | |
1023 | { | |
1024 | int i; | |
1025 | struct page *page; | |
1026 | ||
1027 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1028 | page = class->fullness_list[i]; | |
1029 | if (page) | |
1030 | break; | |
1031 | } | |
1032 | ||
1033 | return page; | |
1034 | } | |
1035 | ||
1b945aee | 1036 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1037 | static inline int __zs_cpu_up(struct mapping_area *area) |
1038 | { | |
1039 | /* | |
1040 | * Make sure we don't leak memory if a cpu UP notification | |
1041 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1042 | */ | |
1043 | if (area->vm) | |
1044 | return 0; | |
1045 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1046 | if (!area->vm) | |
1047 | return -ENOMEM; | |
1048 | return 0; | |
1049 | } | |
1050 | ||
1051 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1052 | { | |
1053 | if (area->vm) | |
1054 | free_vm_area(area->vm); | |
1055 | area->vm = NULL; | |
1056 | } | |
1057 | ||
1058 | static inline void *__zs_map_object(struct mapping_area *area, | |
1059 | struct page *pages[2], int off, int size) | |
1060 | { | |
f6f8ed47 | 1061 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1062 | area->vm_addr = area->vm->addr; |
1063 | return area->vm_addr + off; | |
1064 | } | |
1065 | ||
1066 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1067 | struct page *pages[2], int off, int size) | |
1068 | { | |
1069 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1070 | |
d95abbbb | 1071 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1072 | } |
1073 | ||
1b945aee | 1074 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1075 | |
1076 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1077 | { | |
1078 | /* | |
1079 | * Make sure we don't leak memory if a cpu UP notification | |
1080 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1081 | */ | |
1082 | if (area->vm_buf) | |
1083 | return 0; | |
40f9fb8c | 1084 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1085 | if (!area->vm_buf) |
1086 | return -ENOMEM; | |
1087 | return 0; | |
1088 | } | |
1089 | ||
1090 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1091 | { | |
40f9fb8c | 1092 | kfree(area->vm_buf); |
f553646a SJ |
1093 | area->vm_buf = NULL; |
1094 | } | |
1095 | ||
1096 | static void *__zs_map_object(struct mapping_area *area, | |
1097 | struct page *pages[2], int off, int size) | |
5f601902 | 1098 | { |
5f601902 SJ |
1099 | int sizes[2]; |
1100 | void *addr; | |
f553646a | 1101 | char *buf = area->vm_buf; |
5f601902 | 1102 | |
f553646a SJ |
1103 | /* disable page faults to match kmap_atomic() return conditions */ |
1104 | pagefault_disable(); | |
1105 | ||
1106 | /* no read fastpath */ | |
1107 | if (area->vm_mm == ZS_MM_WO) | |
1108 | goto out; | |
5f601902 SJ |
1109 | |
1110 | sizes[0] = PAGE_SIZE - off; | |
1111 | sizes[1] = size - sizes[0]; | |
1112 | ||
5f601902 SJ |
1113 | /* copy object to per-cpu buffer */ |
1114 | addr = kmap_atomic(pages[0]); | |
1115 | memcpy(buf, addr + off, sizes[0]); | |
1116 | kunmap_atomic(addr); | |
1117 | addr = kmap_atomic(pages[1]); | |
1118 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1119 | kunmap_atomic(addr); | |
f553646a SJ |
1120 | out: |
1121 | return area->vm_buf; | |
5f601902 SJ |
1122 | } |
1123 | ||
f553646a SJ |
1124 | static void __zs_unmap_object(struct mapping_area *area, |
1125 | struct page *pages[2], int off, int size) | |
5f601902 | 1126 | { |
5f601902 SJ |
1127 | int sizes[2]; |
1128 | void *addr; | |
2e40e163 | 1129 | char *buf; |
5f601902 | 1130 | |
f553646a SJ |
1131 | /* no write fastpath */ |
1132 | if (area->vm_mm == ZS_MM_RO) | |
1133 | goto out; | |
5f601902 | 1134 | |
7b60a685 | 1135 | buf = area->vm_buf; |
a82cbf07 YX |
1136 | buf = buf + ZS_HANDLE_SIZE; |
1137 | size -= ZS_HANDLE_SIZE; | |
1138 | off += ZS_HANDLE_SIZE; | |
2e40e163 | 1139 | |
5f601902 SJ |
1140 | sizes[0] = PAGE_SIZE - off; |
1141 | sizes[1] = size - sizes[0]; | |
1142 | ||
1143 | /* copy per-cpu buffer to object */ | |
1144 | addr = kmap_atomic(pages[0]); | |
1145 | memcpy(addr + off, buf, sizes[0]); | |
1146 | kunmap_atomic(addr); | |
1147 | addr = kmap_atomic(pages[1]); | |
1148 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1149 | kunmap_atomic(addr); | |
f553646a SJ |
1150 | |
1151 | out: | |
1152 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1153 | pagefault_enable(); | |
5f601902 | 1154 | } |
61989a80 | 1155 | |
1b945aee | 1156 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1157 | |
61989a80 NG |
1158 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1159 | void *pcpu) | |
1160 | { | |
f553646a | 1161 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1162 | struct mapping_area *area; |
1163 | ||
1164 | switch (action) { | |
1165 | case CPU_UP_PREPARE: | |
1166 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1167 | ret = __zs_cpu_up(area); |
1168 | if (ret) | |
1169 | return notifier_from_errno(ret); | |
61989a80 NG |
1170 | break; |
1171 | case CPU_DEAD: | |
1172 | case CPU_UP_CANCELED: | |
1173 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1174 | __zs_cpu_down(area); |
61989a80 NG |
1175 | break; |
1176 | } | |
1177 | ||
1178 | return NOTIFY_OK; | |
1179 | } | |
1180 | ||
1181 | static struct notifier_block zs_cpu_nb = { | |
1182 | .notifier_call = zs_cpu_notifier | |
1183 | }; | |
1184 | ||
b1b00a5b | 1185 | static int zs_register_cpu_notifier(void) |
61989a80 | 1186 | { |
b1b00a5b | 1187 | int cpu, uninitialized_var(ret); |
61989a80 | 1188 | |
f0e71fcd SB |
1189 | cpu_notifier_register_begin(); |
1190 | ||
1191 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1192 | for_each_online_cpu(cpu) { |
1193 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1194 | if (notifier_to_errno(ret)) |
1195 | break; | |
61989a80 | 1196 | } |
f0e71fcd SB |
1197 | |
1198 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1199 | return notifier_to_errno(ret); |
1200 | } | |
f0e71fcd | 1201 | |
66cdef66 | 1202 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1203 | { |
66cdef66 | 1204 | int cpu; |
40f9fb8c | 1205 | |
66cdef66 | 1206 | cpu_notifier_register_begin(); |
40f9fb8c | 1207 | |
66cdef66 GM |
1208 | for_each_online_cpu(cpu) |
1209 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1210 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1211 | |
66cdef66 | 1212 | cpu_notifier_register_done(); |
b1b00a5b SS |
1213 | } |
1214 | ||
66cdef66 | 1215 | static void init_zs_size_classes(void) |
b1b00a5b | 1216 | { |
66cdef66 | 1217 | int nr; |
c795779d | 1218 | |
66cdef66 GM |
1219 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1220 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1221 | nr += 1; | |
40f9fb8c | 1222 | |
66cdef66 | 1223 | zs_size_classes = nr; |
61989a80 NG |
1224 | } |
1225 | ||
9eec4cd5 JK |
1226 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1227 | { | |
1228 | if (prev->pages_per_zspage != pages_per_zspage) | |
1229 | return false; | |
1230 | ||
1231 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1232 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1233 | return false; | |
1234 | ||
1235 | return true; | |
1236 | } | |
1237 | ||
1fc6e27d | 1238 | static bool zspage_full(struct size_class *class, struct page *first_page) |
312fcae2 | 1239 | { |
830e4bc5 | 1240 | VM_BUG_ON_PAGE(!is_first_page(first_page), first_page); |
312fcae2 | 1241 | |
1fc6e27d | 1242 | return first_page->inuse == class->objs_per_zspage; |
312fcae2 MK |
1243 | } |
1244 | ||
66cdef66 GM |
1245 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1246 | { | |
1247 | return atomic_long_read(&pool->pages_allocated); | |
1248 | } | |
1249 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1250 | ||
4bbc0bc0 | 1251 | /** |
66cdef66 GM |
1252 | * zs_map_object - get address of allocated object from handle. |
1253 | * @pool: pool from which the object was allocated | |
1254 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1255 | * |
66cdef66 GM |
1256 | * Before using an object allocated from zs_malloc, it must be mapped using |
1257 | * this function. When done with the object, it must be unmapped using | |
1258 | * zs_unmap_object. | |
4bbc0bc0 | 1259 | * |
66cdef66 GM |
1260 | * Only one object can be mapped per cpu at a time. There is no protection |
1261 | * against nested mappings. | |
1262 | * | |
1263 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1264 | */ |
66cdef66 GM |
1265 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1266 | enum zs_mapmode mm) | |
61989a80 | 1267 | { |
66cdef66 | 1268 | struct page *page; |
2e40e163 | 1269 | unsigned long obj, obj_idx, off; |
61989a80 | 1270 | |
66cdef66 GM |
1271 | unsigned int class_idx; |
1272 | enum fullness_group fg; | |
1273 | struct size_class *class; | |
1274 | struct mapping_area *area; | |
1275 | struct page *pages[2]; | |
2e40e163 | 1276 | void *ret; |
61989a80 | 1277 | |
9eec4cd5 | 1278 | /* |
66cdef66 GM |
1279 | * Because we use per-cpu mapping areas shared among the |
1280 | * pools/users, we can't allow mapping in interrupt context | |
1281 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1282 | */ |
830e4bc5 | 1283 | WARN_ON_ONCE(in_interrupt()); |
61989a80 | 1284 | |
312fcae2 MK |
1285 | /* From now on, migration cannot move the object */ |
1286 | pin_tag(handle); | |
1287 | ||
2e40e163 MK |
1288 | obj = handle_to_obj(handle); |
1289 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1290 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1291 | class = pool->size_class[class_idx]; | |
1292 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1293 | |
66cdef66 GM |
1294 | area = &get_cpu_var(zs_map_area); |
1295 | area->vm_mm = mm; | |
1296 | if (off + class->size <= PAGE_SIZE) { | |
1297 | /* this object is contained entirely within a page */ | |
1298 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1299 | ret = area->vm_addr + off; |
1300 | goto out; | |
61989a80 NG |
1301 | } |
1302 | ||
66cdef66 GM |
1303 | /* this object spans two pages */ |
1304 | pages[0] = page; | |
1305 | pages[1] = get_next_page(page); | |
1306 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1307 | |
2e40e163 MK |
1308 | ret = __zs_map_object(area, pages, off, class->size); |
1309 | out: | |
7b60a685 MK |
1310 | if (!class->huge) |
1311 | ret += ZS_HANDLE_SIZE; | |
1312 | ||
1313 | return ret; | |
61989a80 | 1314 | } |
66cdef66 | 1315 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1316 | |
66cdef66 | 1317 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1318 | { |
66cdef66 | 1319 | struct page *page; |
2e40e163 | 1320 | unsigned long obj, obj_idx, off; |
61989a80 | 1321 | |
66cdef66 GM |
1322 | unsigned int class_idx; |
1323 | enum fullness_group fg; | |
1324 | struct size_class *class; | |
1325 | struct mapping_area *area; | |
9eec4cd5 | 1326 | |
2e40e163 MK |
1327 | obj = handle_to_obj(handle); |
1328 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1329 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1330 | class = pool->size_class[class_idx]; | |
1331 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1332 | |
66cdef66 GM |
1333 | area = this_cpu_ptr(&zs_map_area); |
1334 | if (off + class->size <= PAGE_SIZE) | |
1335 | kunmap_atomic(area->vm_addr); | |
1336 | else { | |
1337 | struct page *pages[2]; | |
40f9fb8c | 1338 | |
66cdef66 GM |
1339 | pages[0] = page; |
1340 | pages[1] = get_next_page(page); | |
1341 | BUG_ON(!pages[1]); | |
1342 | ||
1343 | __zs_unmap_object(area, pages, off, class->size); | |
1344 | } | |
1345 | put_cpu_var(zs_map_area); | |
312fcae2 | 1346 | unpin_tag(handle); |
61989a80 | 1347 | } |
66cdef66 | 1348 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1349 | |
251cbb95 MK |
1350 | static unsigned long obj_malloc(struct size_class *class, |
1351 | struct page *first_page, unsigned long handle) | |
c7806261 MK |
1352 | { |
1353 | unsigned long obj; | |
1354 | struct link_free *link; | |
1355 | ||
1356 | struct page *m_page; | |
1357 | unsigned long m_objidx, m_offset; | |
1358 | void *vaddr; | |
1359 | ||
312fcae2 | 1360 | handle |= OBJ_ALLOCATED_TAG; |
c7806261 MK |
1361 | obj = (unsigned long)first_page->freelist; |
1362 | obj_to_location(obj, &m_page, &m_objidx); | |
1363 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1364 | ||
1365 | vaddr = kmap_atomic(m_page); | |
1366 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1367 | first_page->freelist = link->next; | |
7b60a685 MK |
1368 | if (!class->huge) |
1369 | /* record handle in the header of allocated chunk */ | |
1370 | link->handle = handle; | |
1371 | else | |
1372 | /* record handle in first_page->private */ | |
1373 | set_page_private(first_page, handle); | |
c7806261 MK |
1374 | kunmap_atomic(vaddr); |
1375 | first_page->inuse++; | |
1376 | zs_stat_inc(class, OBJ_USED, 1); | |
1377 | ||
1378 | return obj; | |
1379 | } | |
1380 | ||
1381 | ||
61989a80 NG |
1382 | /** |
1383 | * zs_malloc - Allocate block of given size from pool. | |
1384 | * @pool: pool to allocate from | |
1385 | * @size: size of block to allocate | |
61989a80 | 1386 | * |
00a61d86 | 1387 | * On success, handle to the allocated object is returned, |
c2344348 | 1388 | * otherwise 0. |
61989a80 NG |
1389 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1390 | */ | |
d0d8da2d | 1391 | unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp) |
61989a80 | 1392 | { |
2e40e163 | 1393 | unsigned long handle, obj; |
61989a80 | 1394 | struct size_class *class; |
c7806261 | 1395 | struct page *first_page; |
61989a80 | 1396 | |
7b60a685 | 1397 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1398 | return 0; |
1399 | ||
d0d8da2d | 1400 | handle = alloc_handle(pool, gfp); |
2e40e163 | 1401 | if (!handle) |
c2344348 | 1402 | return 0; |
61989a80 | 1403 | |
2e40e163 MK |
1404 | /* extra space in chunk to keep the handle */ |
1405 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1406 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1407 | |
1408 | spin_lock(&class->lock); | |
1409 | first_page = find_get_zspage(class); | |
1410 | ||
1411 | if (!first_page) { | |
1412 | spin_unlock(&class->lock); | |
d0d8da2d | 1413 | first_page = alloc_zspage(class, gfp); |
2e40e163 MK |
1414 | if (unlikely(!first_page)) { |
1415 | free_handle(pool, handle); | |
c2344348 | 1416 | return 0; |
2e40e163 | 1417 | } |
61989a80 NG |
1418 | |
1419 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1420 | atomic_long_add(class->pages_per_zspage, |
1421 | &pool->pages_allocated); | |
0f050d99 | 1422 | |
61989a80 | 1423 | spin_lock(&class->lock); |
0f050d99 GM |
1424 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1425 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1426 | } |
1427 | ||
251cbb95 | 1428 | obj = obj_malloc(class, first_page, handle); |
61989a80 | 1429 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1430 | fix_fullness_group(class, first_page); |
2e40e163 | 1431 | record_obj(handle, obj); |
61989a80 NG |
1432 | spin_unlock(&class->lock); |
1433 | ||
2e40e163 | 1434 | return handle; |
61989a80 NG |
1435 | } |
1436 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1437 | ||
1ee47165 | 1438 | static void obj_free(struct size_class *class, unsigned long obj) |
61989a80 NG |
1439 | { |
1440 | struct link_free *link; | |
1441 | struct page *first_page, *f_page; | |
c7806261 | 1442 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1443 | void *vaddr; |
61989a80 | 1444 | |
312fcae2 | 1445 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1446 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1447 | first_page = get_first_page(f_page); |
1448 | ||
61989a80 NG |
1449 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1450 | ||
c7806261 | 1451 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1452 | |
1453 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1454 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1455 | link->next = first_page->freelist; |
7b60a685 MK |
1456 | if (class->huge) |
1457 | set_page_private(first_page, 0); | |
af4ee5e9 | 1458 | kunmap_atomic(vaddr); |
c2344348 | 1459 | first_page->freelist = (void *)obj; |
61989a80 | 1460 | first_page->inuse--; |
0f050d99 | 1461 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1462 | } |
1463 | ||
1464 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1465 | { | |
1466 | struct page *first_page, *f_page; | |
1467 | unsigned long obj, f_objidx; | |
1468 | int class_idx; | |
1469 | struct size_class *class; | |
1470 | enum fullness_group fullness; | |
1471 | ||
1472 | if (unlikely(!handle)) | |
1473 | return; | |
1474 | ||
312fcae2 | 1475 | pin_tag(handle); |
c7806261 | 1476 | obj = handle_to_obj(handle); |
c7806261 MK |
1477 | obj_to_location(obj, &f_page, &f_objidx); |
1478 | first_page = get_first_page(f_page); | |
1479 | ||
1480 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1481 | class = pool->size_class[class_idx]; | |
1482 | ||
1483 | spin_lock(&class->lock); | |
1ee47165 | 1484 | obj_free(class, obj); |
c7806261 | 1485 | fullness = fix_fullness_group(class, first_page); |
312fcae2 | 1486 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1487 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1488 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1489 | atomic_long_sub(class->pages_per_zspage, |
1490 | &pool->pages_allocated); | |
1491 | free_zspage(first_page); | |
1492 | } | |
61989a80 | 1493 | spin_unlock(&class->lock); |
312fcae2 | 1494 | unpin_tag(handle); |
61989a80 | 1495 | |
312fcae2 MK |
1496 | free_handle(pool, handle); |
1497 | } | |
1498 | EXPORT_SYMBOL_GPL(zs_free); | |
1499 | ||
251cbb95 MK |
1500 | static void zs_object_copy(struct size_class *class, unsigned long dst, |
1501 | unsigned long src) | |
312fcae2 MK |
1502 | { |
1503 | struct page *s_page, *d_page; | |
1504 | unsigned long s_objidx, d_objidx; | |
1505 | unsigned long s_off, d_off; | |
1506 | void *s_addr, *d_addr; | |
1507 | int s_size, d_size, size; | |
1508 | int written = 0; | |
1509 | ||
1510 | s_size = d_size = class->size; | |
1511 | ||
1512 | obj_to_location(src, &s_page, &s_objidx); | |
1513 | obj_to_location(dst, &d_page, &d_objidx); | |
1514 | ||
1515 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1516 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1517 | ||
1518 | if (s_off + class->size > PAGE_SIZE) | |
1519 | s_size = PAGE_SIZE - s_off; | |
1520 | ||
1521 | if (d_off + class->size > PAGE_SIZE) | |
1522 | d_size = PAGE_SIZE - d_off; | |
1523 | ||
1524 | s_addr = kmap_atomic(s_page); | |
1525 | d_addr = kmap_atomic(d_page); | |
1526 | ||
1527 | while (1) { | |
1528 | size = min(s_size, d_size); | |
1529 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1530 | written += size; | |
1531 | ||
1532 | if (written == class->size) | |
1533 | break; | |
1534 | ||
495819ea SS |
1535 | s_off += size; |
1536 | s_size -= size; | |
1537 | d_off += size; | |
1538 | d_size -= size; | |
1539 | ||
1540 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1541 | kunmap_atomic(d_addr); |
1542 | kunmap_atomic(s_addr); | |
1543 | s_page = get_next_page(s_page); | |
312fcae2 MK |
1544 | s_addr = kmap_atomic(s_page); |
1545 | d_addr = kmap_atomic(d_page); | |
1546 | s_size = class->size - written; | |
1547 | s_off = 0; | |
312fcae2 MK |
1548 | } |
1549 | ||
495819ea | 1550 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1551 | kunmap_atomic(d_addr); |
1552 | d_page = get_next_page(d_page); | |
312fcae2 MK |
1553 | d_addr = kmap_atomic(d_page); |
1554 | d_size = class->size - written; | |
1555 | d_off = 0; | |
312fcae2 MK |
1556 | } |
1557 | } | |
1558 | ||
1559 | kunmap_atomic(d_addr); | |
1560 | kunmap_atomic(s_addr); | |
1561 | } | |
1562 | ||
1563 | /* | |
1564 | * Find alloced object in zspage from index object and | |
1565 | * return handle. | |
1566 | */ | |
251cbb95 MK |
1567 | static unsigned long find_alloced_obj(struct size_class *class, |
1568 | struct page *page, int index) | |
312fcae2 MK |
1569 | { |
1570 | unsigned long head; | |
1571 | int offset = 0; | |
1572 | unsigned long handle = 0; | |
1573 | void *addr = kmap_atomic(page); | |
1574 | ||
1575 | if (!is_first_page(page)) | |
1576 | offset = page->index; | |
1577 | offset += class->size * index; | |
1578 | ||
1579 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1580 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1581 | if (head & OBJ_ALLOCATED_TAG) { |
1582 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1583 | if (trypin_tag(handle)) | |
1584 | break; | |
1585 | handle = 0; | |
1586 | } | |
1587 | ||
1588 | offset += class->size; | |
1589 | index++; | |
1590 | } | |
1591 | ||
1592 | kunmap_atomic(addr); | |
1593 | return handle; | |
1594 | } | |
1595 | ||
1596 | struct zs_compact_control { | |
1597 | /* Source page for migration which could be a subpage of zspage. */ | |
1598 | struct page *s_page; | |
1599 | /* Destination page for migration which should be a first page | |
1600 | * of zspage. */ | |
1601 | struct page *d_page; | |
1602 | /* Starting object index within @s_page which used for live object | |
1603 | * in the subpage. */ | |
1604 | int index; | |
312fcae2 MK |
1605 | }; |
1606 | ||
1607 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1608 | struct zs_compact_control *cc) | |
1609 | { | |
1610 | unsigned long used_obj, free_obj; | |
1611 | unsigned long handle; | |
1612 | struct page *s_page = cc->s_page; | |
1613 | struct page *d_page = cc->d_page; | |
1614 | unsigned long index = cc->index; | |
312fcae2 MK |
1615 | int ret = 0; |
1616 | ||
1617 | while (1) { | |
251cbb95 | 1618 | handle = find_alloced_obj(class, s_page, index); |
312fcae2 MK |
1619 | if (!handle) { |
1620 | s_page = get_next_page(s_page); | |
1621 | if (!s_page) | |
1622 | break; | |
1623 | index = 0; | |
1624 | continue; | |
1625 | } | |
1626 | ||
1627 | /* Stop if there is no more space */ | |
1fc6e27d | 1628 | if (zspage_full(class, d_page)) { |
312fcae2 MK |
1629 | unpin_tag(handle); |
1630 | ret = -ENOMEM; | |
1631 | break; | |
1632 | } | |
1633 | ||
1634 | used_obj = handle_to_obj(handle); | |
251cbb95 MK |
1635 | free_obj = obj_malloc(class, d_page, handle); |
1636 | zs_object_copy(class, free_obj, used_obj); | |
312fcae2 | 1637 | index++; |
c102f07c JL |
1638 | /* |
1639 | * record_obj updates handle's value to free_obj and it will | |
1640 | * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which | |
1641 | * breaks synchronization using pin_tag(e,g, zs_free) so | |
1642 | * let's keep the lock bit. | |
1643 | */ | |
1644 | free_obj |= BIT(HANDLE_PIN_BIT); | |
312fcae2 MK |
1645 | record_obj(handle, free_obj); |
1646 | unpin_tag(handle); | |
1ee47165 | 1647 | obj_free(class, used_obj); |
312fcae2 MK |
1648 | } |
1649 | ||
1650 | /* Remember last position in this iteration */ | |
1651 | cc->s_page = s_page; | |
1652 | cc->index = index; | |
312fcae2 MK |
1653 | |
1654 | return ret; | |
1655 | } | |
1656 | ||
0dc63d48 | 1657 | static struct page *isolate_target_page(struct size_class *class) |
312fcae2 MK |
1658 | { |
1659 | int i; | |
1660 | struct page *page; | |
1661 | ||
1662 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1663 | page = class->fullness_list[i]; | |
1664 | if (page) { | |
251cbb95 | 1665 | remove_zspage(class, i, page); |
312fcae2 MK |
1666 | break; |
1667 | } | |
1668 | } | |
1669 | ||
1670 | return page; | |
1671 | } | |
1672 | ||
860c707d SS |
1673 | /* |
1674 | * putback_zspage - add @first_page into right class's fullness list | |
1675 | * @pool: target pool | |
1676 | * @class: destination class | |
1677 | * @first_page: target page | |
1678 | * | |
1679 | * Return @fist_page's fullness_group | |
1680 | */ | |
1681 | static enum fullness_group putback_zspage(struct zs_pool *pool, | |
1682 | struct size_class *class, | |
1683 | struct page *first_page) | |
312fcae2 | 1684 | { |
312fcae2 MK |
1685 | enum fullness_group fullness; |
1686 | ||
1fc6e27d | 1687 | fullness = get_fullness_group(class, first_page); |
251cbb95 | 1688 | insert_zspage(class, fullness, first_page); |
839373e6 MK |
1689 | set_zspage_mapping(first_page, class->index, fullness); |
1690 | ||
13de8933 | 1691 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1692 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1693 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1694 | atomic_long_sub(class->pages_per_zspage, |
1695 | &pool->pages_allocated); | |
312fcae2 | 1696 | |
61989a80 | 1697 | free_zspage(first_page); |
13de8933 | 1698 | } |
860c707d SS |
1699 | |
1700 | return fullness; | |
61989a80 | 1701 | } |
312fcae2 MK |
1702 | |
1703 | static struct page *isolate_source_page(struct size_class *class) | |
1704 | { | |
ad9d5e17 MK |
1705 | int i; |
1706 | struct page *page = NULL; | |
1707 | ||
1708 | for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) { | |
1709 | page = class->fullness_list[i]; | |
1710 | if (!page) | |
1711 | continue; | |
312fcae2 | 1712 | |
251cbb95 | 1713 | remove_zspage(class, i, page); |
ad9d5e17 MK |
1714 | break; |
1715 | } | |
312fcae2 MK |
1716 | |
1717 | return page; | |
1718 | } | |
1719 | ||
04f05909 SS |
1720 | /* |
1721 | * | |
1722 | * Based on the number of unused allocated objects calculate | |
1723 | * and return the number of pages that we can free. | |
04f05909 SS |
1724 | */ |
1725 | static unsigned long zs_can_compact(struct size_class *class) | |
1726 | { | |
1727 | unsigned long obj_wasted; | |
44f43e99 SS |
1728 | unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); |
1729 | unsigned long obj_used = zs_stat_get(class, OBJ_USED); | |
04f05909 | 1730 | |
44f43e99 SS |
1731 | if (obj_allocated <= obj_used) |
1732 | return 0; | |
04f05909 | 1733 | |
44f43e99 | 1734 | obj_wasted = obj_allocated - obj_used; |
04f05909 SS |
1735 | obj_wasted /= get_maxobj_per_zspage(class->size, |
1736 | class->pages_per_zspage); | |
1737 | ||
6cbf16b3 | 1738 | return obj_wasted * class->pages_per_zspage; |
04f05909 SS |
1739 | } |
1740 | ||
7d3f3938 | 1741 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1742 | { |
312fcae2 MK |
1743 | struct zs_compact_control cc; |
1744 | struct page *src_page; | |
1745 | struct page *dst_page = NULL; | |
312fcae2 | 1746 | |
312fcae2 MK |
1747 | spin_lock(&class->lock); |
1748 | while ((src_page = isolate_source_page(class))) { | |
1749 | ||
04f05909 SS |
1750 | if (!zs_can_compact(class)) |
1751 | break; | |
1752 | ||
312fcae2 MK |
1753 | cc.index = 0; |
1754 | cc.s_page = src_page; | |
1755 | ||
0dc63d48 | 1756 | while ((dst_page = isolate_target_page(class))) { |
312fcae2 MK |
1757 | cc.d_page = dst_page; |
1758 | /* | |
0dc63d48 SS |
1759 | * If there is no more space in dst_page, resched |
1760 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1761 | */ |
1762 | if (!migrate_zspage(pool, class, &cc)) | |
1763 | break; | |
1764 | ||
1765 | putback_zspage(pool, class, dst_page); | |
312fcae2 MK |
1766 | } |
1767 | ||
1768 | /* Stop if we couldn't find slot */ | |
1769 | if (dst_page == NULL) | |
1770 | break; | |
1771 | ||
1772 | putback_zspage(pool, class, dst_page); | |
860c707d | 1773 | if (putback_zspage(pool, class, src_page) == ZS_EMPTY) |
6cbf16b3 | 1774 | pool->stats.pages_compacted += class->pages_per_zspage; |
312fcae2 | 1775 | spin_unlock(&class->lock); |
312fcae2 MK |
1776 | cond_resched(); |
1777 | spin_lock(&class->lock); | |
1778 | } | |
1779 | ||
1780 | if (src_page) | |
1781 | putback_zspage(pool, class, src_page); | |
1782 | ||
7d3f3938 | 1783 | spin_unlock(&class->lock); |
312fcae2 MK |
1784 | } |
1785 | ||
1786 | unsigned long zs_compact(struct zs_pool *pool) | |
1787 | { | |
1788 | int i; | |
312fcae2 MK |
1789 | struct size_class *class; |
1790 | ||
1791 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1792 | class = pool->size_class[i]; | |
1793 | if (!class) | |
1794 | continue; | |
1795 | if (class->index != i) | |
1796 | continue; | |
7d3f3938 | 1797 | __zs_compact(pool, class); |
312fcae2 MK |
1798 | } |
1799 | ||
860c707d | 1800 | return pool->stats.pages_compacted; |
312fcae2 MK |
1801 | } |
1802 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1803 | |
7d3f3938 SS |
1804 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1805 | { | |
1806 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1807 | } | |
1808 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1809 | ||
ab9d306d SS |
1810 | static unsigned long zs_shrinker_scan(struct shrinker *shrinker, |
1811 | struct shrink_control *sc) | |
1812 | { | |
1813 | unsigned long pages_freed; | |
1814 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1815 | shrinker); | |
1816 | ||
1817 | pages_freed = pool->stats.pages_compacted; | |
1818 | /* | |
1819 | * Compact classes and calculate compaction delta. | |
1820 | * Can run concurrently with a manually triggered | |
1821 | * (by user) compaction. | |
1822 | */ | |
1823 | pages_freed = zs_compact(pool) - pages_freed; | |
1824 | ||
1825 | return pages_freed ? pages_freed : SHRINK_STOP; | |
1826 | } | |
1827 | ||
1828 | static unsigned long zs_shrinker_count(struct shrinker *shrinker, | |
1829 | struct shrink_control *sc) | |
1830 | { | |
1831 | int i; | |
1832 | struct size_class *class; | |
1833 | unsigned long pages_to_free = 0; | |
1834 | struct zs_pool *pool = container_of(shrinker, struct zs_pool, | |
1835 | shrinker); | |
1836 | ||
ab9d306d SS |
1837 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1838 | class = pool->size_class[i]; | |
1839 | if (!class) | |
1840 | continue; | |
1841 | if (class->index != i) | |
1842 | continue; | |
1843 | ||
ab9d306d | 1844 | pages_to_free += zs_can_compact(class); |
ab9d306d SS |
1845 | } |
1846 | ||
1847 | return pages_to_free; | |
1848 | } | |
1849 | ||
1850 | static void zs_unregister_shrinker(struct zs_pool *pool) | |
1851 | { | |
1852 | if (pool->shrinker_enabled) { | |
1853 | unregister_shrinker(&pool->shrinker); | |
1854 | pool->shrinker_enabled = false; | |
1855 | } | |
1856 | } | |
1857 | ||
1858 | static int zs_register_shrinker(struct zs_pool *pool) | |
1859 | { | |
1860 | pool->shrinker.scan_objects = zs_shrinker_scan; | |
1861 | pool->shrinker.count_objects = zs_shrinker_count; | |
1862 | pool->shrinker.batch = 0; | |
1863 | pool->shrinker.seeks = DEFAULT_SEEKS; | |
1864 | ||
1865 | return register_shrinker(&pool->shrinker); | |
1866 | } | |
1867 | ||
00a61d86 | 1868 | /** |
66cdef66 GM |
1869 | * zs_create_pool - Creates an allocation pool to work from. |
1870 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1871 | * |
66cdef66 GM |
1872 | * This function must be called before anything when using |
1873 | * the zsmalloc allocator. | |
166cfda7 | 1874 | * |
66cdef66 GM |
1875 | * On success, a pointer to the newly created pool is returned, |
1876 | * otherwise NULL. | |
396b7fd6 | 1877 | */ |
d0d8da2d | 1878 | struct zs_pool *zs_create_pool(const char *name) |
61989a80 | 1879 | { |
66cdef66 GM |
1880 | int i; |
1881 | struct zs_pool *pool; | |
1882 | struct size_class *prev_class = NULL; | |
61989a80 | 1883 | |
66cdef66 GM |
1884 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1885 | if (!pool) | |
1886 | return NULL; | |
61989a80 | 1887 | |
66cdef66 GM |
1888 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1889 | GFP_KERNEL); | |
1890 | if (!pool->size_class) { | |
1891 | kfree(pool); | |
1892 | return NULL; | |
1893 | } | |
61989a80 | 1894 | |
2e40e163 MK |
1895 | pool->name = kstrdup(name, GFP_KERNEL); |
1896 | if (!pool->name) | |
1897 | goto err; | |
1898 | ||
1899 | if (create_handle_cache(pool)) | |
1900 | goto err; | |
1901 | ||
c60369f0 | 1902 | /* |
66cdef66 GM |
1903 | * Iterate reversly, because, size of size_class that we want to use |
1904 | * for merging should be larger or equal to current size. | |
c60369f0 | 1905 | */ |
66cdef66 GM |
1906 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1907 | int size; | |
1908 | int pages_per_zspage; | |
1909 | struct size_class *class; | |
c60369f0 | 1910 | |
66cdef66 GM |
1911 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1912 | if (size > ZS_MAX_ALLOC_SIZE) | |
1913 | size = ZS_MAX_ALLOC_SIZE; | |
1914 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1915 | |
66cdef66 GM |
1916 | /* |
1917 | * size_class is used for normal zsmalloc operation such | |
1918 | * as alloc/free for that size. Although it is natural that we | |
1919 | * have one size_class for each size, there is a chance that we | |
1920 | * can get more memory utilization if we use one size_class for | |
1921 | * many different sizes whose size_class have same | |
1922 | * characteristics. So, we makes size_class point to | |
1923 | * previous size_class if possible. | |
1924 | */ | |
1925 | if (prev_class) { | |
1926 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1927 | pool->size_class[i] = prev_class; | |
1928 | continue; | |
1929 | } | |
1930 | } | |
1931 | ||
1932 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1933 | if (!class) | |
1934 | goto err; | |
1935 | ||
1936 | class->size = size; | |
1937 | class->index = i; | |
1938 | class->pages_per_zspage = pages_per_zspage; | |
1fc6e27d MK |
1939 | class->objs_per_zspage = class->pages_per_zspage * |
1940 | PAGE_SIZE / class->size; | |
1941 | if (pages_per_zspage == 1 && class->objs_per_zspage == 1) | |
7b60a685 | 1942 | class->huge = true; |
66cdef66 GM |
1943 | spin_lock_init(&class->lock); |
1944 | pool->size_class[i] = class; | |
1945 | ||
1946 | prev_class = class; | |
61989a80 NG |
1947 | } |
1948 | ||
d34f6157 DS |
1949 | /* debug only, don't abort if it fails */ |
1950 | zs_pool_stat_create(pool, name); | |
0f050d99 | 1951 | |
ab9d306d SS |
1952 | /* |
1953 | * Not critical, we still can use the pool | |
1954 | * and user can trigger compaction manually. | |
1955 | */ | |
1956 | if (zs_register_shrinker(pool) == 0) | |
1957 | pool->shrinker_enabled = true; | |
66cdef66 GM |
1958 | return pool; |
1959 | ||
1960 | err: | |
1961 | zs_destroy_pool(pool); | |
1962 | return NULL; | |
61989a80 | 1963 | } |
66cdef66 | 1964 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1965 | |
66cdef66 | 1966 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1967 | { |
66cdef66 | 1968 | int i; |
61989a80 | 1969 | |
ab9d306d | 1970 | zs_unregister_shrinker(pool); |
0f050d99 GM |
1971 | zs_pool_stat_destroy(pool); |
1972 | ||
66cdef66 GM |
1973 | for (i = 0; i < zs_size_classes; i++) { |
1974 | int fg; | |
1975 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1976 | |
66cdef66 GM |
1977 | if (!class) |
1978 | continue; | |
61989a80 | 1979 | |
66cdef66 GM |
1980 | if (class->index != i) |
1981 | continue; | |
61989a80 | 1982 | |
66cdef66 GM |
1983 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1984 | if (class->fullness_list[fg]) { | |
1985 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1986 | class->size, fg); | |
1987 | } | |
1988 | } | |
1989 | kfree(class); | |
1990 | } | |
f553646a | 1991 | |
2e40e163 | 1992 | destroy_handle_cache(pool); |
66cdef66 | 1993 | kfree(pool->size_class); |
0f050d99 | 1994 | kfree(pool->name); |
66cdef66 GM |
1995 | kfree(pool); |
1996 | } | |
1997 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 1998 | |
66cdef66 GM |
1999 | static int __init zs_init(void) |
2000 | { | |
2001 | int ret = zs_register_cpu_notifier(); | |
2002 | ||
0f050d99 GM |
2003 | if (ret) |
2004 | goto notifier_fail; | |
66cdef66 GM |
2005 | |
2006 | init_zs_size_classes(); | |
2007 | ||
2008 | #ifdef CONFIG_ZPOOL | |
2009 | zpool_register_driver(&zs_zpool_driver); | |
2010 | #endif | |
0f050d99 | 2011 | |
4abaac9b DS |
2012 | zs_stat_init(); |
2013 | ||
66cdef66 | 2014 | return 0; |
0f050d99 | 2015 | |
0f050d99 GM |
2016 | notifier_fail: |
2017 | zs_unregister_cpu_notifier(); | |
2018 | ||
2019 | return ret; | |
61989a80 | 2020 | } |
61989a80 | 2021 | |
66cdef66 | 2022 | static void __exit zs_exit(void) |
61989a80 | 2023 | { |
66cdef66 GM |
2024 | #ifdef CONFIG_ZPOOL |
2025 | zpool_unregister_driver(&zs_zpool_driver); | |
2026 | #endif | |
2027 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
2028 | |
2029 | zs_stat_exit(); | |
61989a80 | 2030 | } |
069f101f BH |
2031 | |
2032 | module_init(zs_init); | |
2033 | module_exit(zs_exit); | |
2034 | ||
2035 | MODULE_LICENSE("Dual BSD/GPL"); | |
2036 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |