]>
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: | |
19 | * page->first_page: points to the first component (0-order) page | |
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 | * | |
29 | * page->private (union with page->first_page): refers to the | |
30 | * component page after the first page | |
7b60a685 MK |
31 | * If the page is first_page for huge object, it stores handle. |
32 | * Look at size_class->huge. | |
2db51dae NG |
33 | * page->freelist: points to the first free object in zspage. |
34 | * Free objects are linked together using in-place | |
35 | * metadata. | |
36 | * page->objects: maximum number of objects we can store in this | |
37 | * zspage (class->zspage_order * PAGE_SIZE / class->size) | |
38 | * page->lru: links together first pages of various zspages. | |
39 | * Basically forming list of zspages in a fullness group. | |
40 | * page->mapping: class index and fullness group of the zspage | |
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> |
c60369f0 | 61 | #include <linux/hardirq.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 | NR_ZS_STAT_TYPE, |
170 | }; | |
171 | ||
0f050d99 GM |
172 | struct zs_size_stat { |
173 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
174 | }; | |
175 | ||
57244594 SS |
176 | #ifdef CONFIG_ZSMALLOC_STAT |
177 | static struct dentry *zs_stat_root; | |
0f050d99 GM |
178 | #endif |
179 | ||
40f9fb8c MG |
180 | /* |
181 | * number of size_classes | |
182 | */ | |
183 | static int zs_size_classes; | |
184 | ||
0959c63f SJ |
185 | /* |
186 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
187 | * n <= N / f, where | |
188 | * n = number of allocated objects | |
189 | * N = total number of objects zspage can store | |
6dd9737e | 190 | * f = fullness_threshold_frac |
0959c63f SJ |
191 | * |
192 | * Similarly, we assign zspage to: | |
193 | * ZS_ALMOST_FULL when n > N / f | |
194 | * ZS_EMPTY when n == 0 | |
195 | * ZS_FULL when n == N | |
196 | * | |
197 | * (see: fix_fullness_group()) | |
198 | */ | |
199 | static const int fullness_threshold_frac = 4; | |
200 | ||
201 | struct size_class { | |
57244594 SS |
202 | spinlock_t lock; |
203 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; | |
0959c63f SJ |
204 | /* |
205 | * Size of objects stored in this class. Must be multiple | |
206 | * of ZS_ALIGN. | |
207 | */ | |
208 | int size; | |
209 | unsigned int index; | |
210 | ||
211 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ | |
212 | int pages_per_zspage; | |
0f050d99 | 213 | struct zs_size_stat stats; |
0959c63f | 214 | |
57244594 SS |
215 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
216 | bool huge; | |
0959c63f SJ |
217 | }; |
218 | ||
219 | /* | |
220 | * Placed within free objects to form a singly linked list. | |
221 | * For every zspage, first_page->freelist gives head of this list. | |
222 | * | |
223 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
224 | */ | |
225 | struct link_free { | |
2e40e163 MK |
226 | union { |
227 | /* | |
228 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
229 | * It's valid for non-allocated object | |
230 | */ | |
231 | void *next; | |
232 | /* | |
233 | * Handle of allocated object. | |
234 | */ | |
235 | unsigned long handle; | |
236 | }; | |
0959c63f SJ |
237 | }; |
238 | ||
239 | struct zs_pool { | |
0f050d99 GM |
240 | char *name; |
241 | ||
40f9fb8c | 242 | struct size_class **size_class; |
2e40e163 | 243 | struct kmem_cache *handle_cachep; |
0959c63f SJ |
244 | |
245 | gfp_t flags; /* allocation flags used when growing pool */ | |
13de8933 | 246 | atomic_long_t pages_allocated; |
0f050d99 | 247 | |
7d3f3938 | 248 | struct zs_pool_stats stats; |
0f050d99 GM |
249 | #ifdef CONFIG_ZSMALLOC_STAT |
250 | struct dentry *stat_dentry; | |
251 | #endif | |
0959c63f | 252 | }; |
61989a80 NG |
253 | |
254 | /* | |
255 | * A zspage's class index and fullness group | |
256 | * are encoded in its (first)page->mapping | |
257 | */ | |
258 | #define CLASS_IDX_BITS 28 | |
259 | #define FULLNESS_BITS 4 | |
260 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
261 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
262 | ||
f553646a | 263 | struct mapping_area { |
1b945aee | 264 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
265 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
266 | #else | |
267 | char *vm_buf; /* copy buffer for objects that span pages */ | |
268 | #endif | |
269 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
270 | enum zs_mapmode vm_mm; /* mapping mode */ | |
7b60a685 | 271 | bool huge; |
f553646a SJ |
272 | }; |
273 | ||
2e40e163 MK |
274 | static int create_handle_cache(struct zs_pool *pool) |
275 | { | |
276 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
277 | 0, 0, NULL); | |
278 | return pool->handle_cachep ? 0 : 1; | |
279 | } | |
280 | ||
281 | static void destroy_handle_cache(struct zs_pool *pool) | |
282 | { | |
02f7b414 SS |
283 | if (pool->handle_cachep) |
284 | kmem_cache_destroy(pool->handle_cachep); | |
2e40e163 MK |
285 | } |
286 | ||
287 | static unsigned long alloc_handle(struct zs_pool *pool) | |
288 | { | |
289 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
290 | pool->flags & ~__GFP_HIGHMEM); | |
291 | } | |
292 | ||
293 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
294 | { | |
295 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
296 | } | |
297 | ||
298 | static void record_obj(unsigned long handle, unsigned long obj) | |
299 | { | |
300 | *(unsigned long *)handle = obj; | |
301 | } | |
302 | ||
c795779d DS |
303 | /* zpool driver */ |
304 | ||
305 | #ifdef CONFIG_ZPOOL | |
306 | ||
479305fd DS |
307 | static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops, |
308 | struct zpool *zpool) | |
c795779d | 309 | { |
3eba0c6a | 310 | return zs_create_pool(name, gfp); |
c795779d DS |
311 | } |
312 | ||
313 | static void zs_zpool_destroy(void *pool) | |
314 | { | |
315 | zs_destroy_pool(pool); | |
316 | } | |
317 | ||
318 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
319 | unsigned long *handle) | |
320 | { | |
321 | *handle = zs_malloc(pool, size); | |
322 | return *handle ? 0 : -1; | |
323 | } | |
324 | static void zs_zpool_free(void *pool, unsigned long handle) | |
325 | { | |
326 | zs_free(pool, handle); | |
327 | } | |
328 | ||
329 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
330 | unsigned int *reclaimed) | |
331 | { | |
332 | return -EINVAL; | |
333 | } | |
334 | ||
335 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
336 | enum zpool_mapmode mm) | |
337 | { | |
338 | enum zs_mapmode zs_mm; | |
339 | ||
340 | switch (mm) { | |
341 | case ZPOOL_MM_RO: | |
342 | zs_mm = ZS_MM_RO; | |
343 | break; | |
344 | case ZPOOL_MM_WO: | |
345 | zs_mm = ZS_MM_WO; | |
346 | break; | |
347 | case ZPOOL_MM_RW: /* fallthru */ | |
348 | default: | |
349 | zs_mm = ZS_MM_RW; | |
350 | break; | |
351 | } | |
352 | ||
353 | return zs_map_object(pool, handle, zs_mm); | |
354 | } | |
355 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
356 | { | |
357 | zs_unmap_object(pool, handle); | |
358 | } | |
359 | ||
360 | static u64 zs_zpool_total_size(void *pool) | |
361 | { | |
722cdc17 | 362 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
363 | } |
364 | ||
365 | static struct zpool_driver zs_zpool_driver = { | |
366 | .type = "zsmalloc", | |
367 | .owner = THIS_MODULE, | |
368 | .create = zs_zpool_create, | |
369 | .destroy = zs_zpool_destroy, | |
370 | .malloc = zs_zpool_malloc, | |
371 | .free = zs_zpool_free, | |
372 | .shrink = zs_zpool_shrink, | |
373 | .map = zs_zpool_map, | |
374 | .unmap = zs_zpool_unmap, | |
375 | .total_size = zs_zpool_total_size, | |
376 | }; | |
377 | ||
137f8cff | 378 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
379 | #endif /* CONFIG_ZPOOL */ |
380 | ||
248ca1b0 MK |
381 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
382 | { | |
383 | return pages_per_zspage * PAGE_SIZE / size; | |
384 | } | |
385 | ||
61989a80 NG |
386 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
387 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
388 | ||
389 | static int is_first_page(struct page *page) | |
390 | { | |
a27545bf | 391 | return PagePrivate(page); |
61989a80 NG |
392 | } |
393 | ||
394 | static int is_last_page(struct page *page) | |
395 | { | |
a27545bf | 396 | return PagePrivate2(page); |
61989a80 NG |
397 | } |
398 | ||
399 | static void get_zspage_mapping(struct page *page, unsigned int *class_idx, | |
400 | enum fullness_group *fullness) | |
401 | { | |
402 | unsigned long m; | |
403 | BUG_ON(!is_first_page(page)); | |
404 | ||
405 | m = (unsigned long)page->mapping; | |
406 | *fullness = m & FULLNESS_MASK; | |
407 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
408 | } | |
409 | ||
410 | static void set_zspage_mapping(struct page *page, unsigned int class_idx, | |
411 | enum fullness_group fullness) | |
412 | { | |
413 | unsigned long m; | |
414 | BUG_ON(!is_first_page(page)); | |
415 | ||
416 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
417 | (fullness & FULLNESS_MASK); | |
418 | page->mapping = (struct address_space *)m; | |
419 | } | |
420 | ||
c3e3e88a NC |
421 | /* |
422 | * zsmalloc divides the pool into various size classes where each | |
423 | * class maintains a list of zspages where each zspage is divided | |
424 | * into equal sized chunks. Each allocation falls into one of these | |
425 | * classes depending on its size. This function returns index of the | |
426 | * size class which has chunk size big enough to hold the give size. | |
427 | */ | |
61989a80 NG |
428 | static int get_size_class_index(int size) |
429 | { | |
430 | int idx = 0; | |
431 | ||
432 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
433 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
434 | ZS_SIZE_CLASS_DELTA); | |
435 | ||
7b60a685 | 436 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
437 | } |
438 | ||
248ca1b0 MK |
439 | static inline void zs_stat_inc(struct size_class *class, |
440 | enum zs_stat_type type, unsigned long cnt) | |
441 | { | |
442 | class->stats.objs[type] += cnt; | |
443 | } | |
444 | ||
445 | static inline void zs_stat_dec(struct size_class *class, | |
446 | enum zs_stat_type type, unsigned long cnt) | |
447 | { | |
448 | class->stats.objs[type] -= cnt; | |
449 | } | |
450 | ||
451 | static inline unsigned long zs_stat_get(struct size_class *class, | |
452 | enum zs_stat_type type) | |
453 | { | |
454 | return class->stats.objs[type]; | |
455 | } | |
456 | ||
57244594 SS |
457 | #ifdef CONFIG_ZSMALLOC_STAT |
458 | ||
248ca1b0 MK |
459 | static int __init zs_stat_init(void) |
460 | { | |
461 | if (!debugfs_initialized()) | |
462 | return -ENODEV; | |
463 | ||
464 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
465 | if (!zs_stat_root) | |
466 | return -ENOMEM; | |
467 | ||
468 | return 0; | |
469 | } | |
470 | ||
471 | static void __exit zs_stat_exit(void) | |
472 | { | |
473 | debugfs_remove_recursive(zs_stat_root); | |
474 | } | |
475 | ||
476 | static int zs_stats_size_show(struct seq_file *s, void *v) | |
477 | { | |
478 | int i; | |
479 | struct zs_pool *pool = s->private; | |
480 | struct size_class *class; | |
481 | int objs_per_zspage; | |
482 | unsigned long class_almost_full, class_almost_empty; | |
483 | unsigned long obj_allocated, obj_used, pages_used; | |
484 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; | |
485 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
486 | ||
487 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n", | |
488 | "class", "size", "almost_full", "almost_empty", | |
489 | "obj_allocated", "obj_used", "pages_used", | |
490 | "pages_per_zspage"); | |
491 | ||
492 | for (i = 0; i < zs_size_classes; i++) { | |
493 | class = pool->size_class[i]; | |
494 | ||
495 | if (class->index != i) | |
496 | continue; | |
497 | ||
498 | spin_lock(&class->lock); | |
499 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
500 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
501 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
502 | obj_used = zs_stat_get(class, OBJ_USED); | |
503 | spin_unlock(&class->lock); | |
504 | ||
505 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
506 | class->pages_per_zspage); | |
507 | pages_used = obj_allocated / objs_per_zspage * | |
508 | class->pages_per_zspage; | |
509 | ||
510 | seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n", | |
511 | i, class->size, class_almost_full, class_almost_empty, | |
512 | obj_allocated, obj_used, pages_used, | |
513 | class->pages_per_zspage); | |
514 | ||
515 | total_class_almost_full += class_almost_full; | |
516 | total_class_almost_empty += class_almost_empty; | |
517 | total_objs += obj_allocated; | |
518 | total_used_objs += obj_used; | |
519 | total_pages += pages_used; | |
520 | } | |
521 | ||
522 | seq_puts(s, "\n"); | |
523 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n", | |
524 | "Total", "", total_class_almost_full, | |
525 | total_class_almost_empty, total_objs, | |
526 | total_used_objs, total_pages); | |
527 | ||
528 | return 0; | |
529 | } | |
530 | ||
531 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
532 | { | |
533 | return single_open(file, zs_stats_size_show, inode->i_private); | |
534 | } | |
535 | ||
536 | static const struct file_operations zs_stat_size_ops = { | |
537 | .open = zs_stats_size_open, | |
538 | .read = seq_read, | |
539 | .llseek = seq_lseek, | |
540 | .release = single_release, | |
541 | }; | |
542 | ||
543 | static int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
544 | { | |
545 | struct dentry *entry; | |
546 | ||
547 | if (!zs_stat_root) | |
548 | return -ENODEV; | |
549 | ||
550 | entry = debugfs_create_dir(name, zs_stat_root); | |
551 | if (!entry) { | |
552 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
553 | return -ENOMEM; | |
554 | } | |
555 | pool->stat_dentry = entry; | |
556 | ||
557 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
558 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
559 | if (!entry) { | |
560 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
561 | name, "classes"); | |
562 | return -ENOMEM; | |
563 | } | |
564 | ||
565 | return 0; | |
566 | } | |
567 | ||
568 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
569 | { | |
570 | debugfs_remove_recursive(pool->stat_dentry); | |
571 | } | |
572 | ||
573 | #else /* CONFIG_ZSMALLOC_STAT */ | |
248ca1b0 MK |
574 | static int __init zs_stat_init(void) |
575 | { | |
576 | return 0; | |
577 | } | |
578 | ||
579 | static void __exit zs_stat_exit(void) | |
580 | { | |
581 | } | |
582 | ||
583 | static inline int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
584 | { | |
585 | return 0; | |
586 | } | |
587 | ||
588 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
589 | { | |
590 | } | |
248ca1b0 MK |
591 | #endif |
592 | ||
593 | ||
c3e3e88a NC |
594 | /* |
595 | * For each size class, zspages are divided into different groups | |
596 | * depending on how "full" they are. This was done so that we could | |
597 | * easily find empty or nearly empty zspages when we try to shrink | |
598 | * the pool (not yet implemented). This function returns fullness | |
599 | * status of the given page. | |
600 | */ | |
61989a80 NG |
601 | static enum fullness_group get_fullness_group(struct page *page) |
602 | { | |
603 | int inuse, max_objects; | |
604 | enum fullness_group fg; | |
605 | BUG_ON(!is_first_page(page)); | |
606 | ||
607 | inuse = page->inuse; | |
608 | max_objects = page->objects; | |
609 | ||
610 | if (inuse == 0) | |
611 | fg = ZS_EMPTY; | |
612 | else if (inuse == max_objects) | |
613 | fg = ZS_FULL; | |
d3d07c92 | 614 | else if (inuse <= 3 * max_objects / fullness_threshold_frac) |
61989a80 NG |
615 | fg = ZS_ALMOST_EMPTY; |
616 | else | |
617 | fg = ZS_ALMOST_FULL; | |
618 | ||
619 | return fg; | |
620 | } | |
621 | ||
c3e3e88a NC |
622 | /* |
623 | * Each size class maintains various freelists and zspages are assigned | |
624 | * to one of these freelists based on the number of live objects they | |
625 | * have. This functions inserts the given zspage into the freelist | |
626 | * identified by <class, fullness_group>. | |
627 | */ | |
61989a80 NG |
628 | static void insert_zspage(struct page *page, struct size_class *class, |
629 | enum fullness_group fullness) | |
630 | { | |
631 | struct page **head; | |
632 | ||
633 | BUG_ON(!is_first_page(page)); | |
634 | ||
635 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
636 | return; | |
637 | ||
638 | head = &class->fullness_list[fullness]; | |
639 | if (*head) | |
640 | list_add_tail(&page->lru, &(*head)->lru); | |
641 | ||
642 | *head = page; | |
248ca1b0 MK |
643 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
644 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
645 | } |
646 | ||
c3e3e88a NC |
647 | /* |
648 | * This function removes the given zspage from the freelist identified | |
649 | * by <class, fullness_group>. | |
650 | */ | |
61989a80 NG |
651 | static void remove_zspage(struct page *page, struct size_class *class, |
652 | enum fullness_group fullness) | |
653 | { | |
654 | struct page **head; | |
655 | ||
656 | BUG_ON(!is_first_page(page)); | |
657 | ||
658 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
659 | return; | |
660 | ||
661 | head = &class->fullness_list[fullness]; | |
662 | BUG_ON(!*head); | |
663 | if (list_empty(&(*head)->lru)) | |
664 | *head = NULL; | |
665 | else if (*head == page) | |
666 | *head = (struct page *)list_entry((*head)->lru.next, | |
667 | struct page, lru); | |
668 | ||
669 | list_del_init(&page->lru); | |
248ca1b0 MK |
670 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
671 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
672 | } |
673 | ||
c3e3e88a NC |
674 | /* |
675 | * Each size class maintains zspages in different fullness groups depending | |
676 | * on the number of live objects they contain. When allocating or freeing | |
677 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
678 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
679 | * a status change has occurred for the given page and accordingly moves the | |
680 | * page from the freelist of the old fullness group to that of the new | |
681 | * fullness group. | |
682 | */ | |
c7806261 | 683 | static enum fullness_group fix_fullness_group(struct size_class *class, |
61989a80 NG |
684 | struct page *page) |
685 | { | |
686 | int class_idx; | |
61989a80 NG |
687 | enum fullness_group currfg, newfg; |
688 | ||
689 | BUG_ON(!is_first_page(page)); | |
690 | ||
691 | get_zspage_mapping(page, &class_idx, &currfg); | |
692 | newfg = get_fullness_group(page); | |
693 | if (newfg == currfg) | |
694 | goto out; | |
695 | ||
61989a80 NG |
696 | remove_zspage(page, class, currfg); |
697 | insert_zspage(page, class, newfg); | |
698 | set_zspage_mapping(page, class_idx, newfg); | |
699 | ||
700 | out: | |
701 | return newfg; | |
702 | } | |
703 | ||
704 | /* | |
705 | * We have to decide on how many pages to link together | |
706 | * to form a zspage for each size class. This is important | |
707 | * to reduce wastage due to unusable space left at end of | |
708 | * each zspage which is given as: | |
888fa374 YX |
709 | * wastage = Zp % class_size |
710 | * usage = Zp - wastage | |
61989a80 NG |
711 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
712 | * | |
713 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
714 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
715 | * since then we can perfectly fit in 8 such objects. | |
716 | */ | |
2e3b6154 | 717 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
718 | { |
719 | int i, max_usedpc = 0; | |
720 | /* zspage order which gives maximum used size per KB */ | |
721 | int max_usedpc_order = 1; | |
722 | ||
84d4faab | 723 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
724 | int zspage_size; |
725 | int waste, usedpc; | |
726 | ||
727 | zspage_size = i * PAGE_SIZE; | |
728 | waste = zspage_size % class_size; | |
729 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
730 | ||
731 | if (usedpc > max_usedpc) { | |
732 | max_usedpc = usedpc; | |
733 | max_usedpc_order = i; | |
734 | } | |
735 | } | |
736 | ||
737 | return max_usedpc_order; | |
738 | } | |
739 | ||
740 | /* | |
741 | * A single 'zspage' is composed of many system pages which are | |
742 | * linked together using fields in struct page. This function finds | |
743 | * the first/head page, given any component page of a zspage. | |
744 | */ | |
745 | static struct page *get_first_page(struct page *page) | |
746 | { | |
747 | if (is_first_page(page)) | |
748 | return page; | |
749 | else | |
750 | return page->first_page; | |
751 | } | |
752 | ||
753 | static struct page *get_next_page(struct page *page) | |
754 | { | |
755 | struct page *next; | |
756 | ||
757 | if (is_last_page(page)) | |
758 | next = NULL; | |
759 | else if (is_first_page(page)) | |
e842b976 | 760 | next = (struct page *)page_private(page); |
61989a80 NG |
761 | else |
762 | next = list_entry(page->lru.next, struct page, lru); | |
763 | ||
764 | return next; | |
765 | } | |
766 | ||
67296874 OH |
767 | /* |
768 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 769 | * We use the least bit of handle for tagging. |
67296874 | 770 | */ |
312fcae2 | 771 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 772 | { |
312fcae2 | 773 | unsigned long obj; |
61989a80 NG |
774 | |
775 | if (!page) { | |
776 | BUG_ON(obj_idx); | |
777 | return NULL; | |
778 | } | |
779 | ||
312fcae2 MK |
780 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
781 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
782 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 783 | |
312fcae2 | 784 | return (void *)obj; |
61989a80 NG |
785 | } |
786 | ||
67296874 OH |
787 | /* |
788 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
789 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 790 | * location_to_obj(). |
67296874 | 791 | */ |
312fcae2 | 792 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
793 | unsigned long *obj_idx) |
794 | { | |
312fcae2 MK |
795 | obj >>= OBJ_TAG_BITS; |
796 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
797 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
798 | } |
799 | ||
2e40e163 MK |
800 | static unsigned long handle_to_obj(unsigned long handle) |
801 | { | |
802 | return *(unsigned long *)handle; | |
803 | } | |
804 | ||
7b60a685 MK |
805 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
806 | void *obj) | |
312fcae2 | 807 | { |
7b60a685 MK |
808 | if (class->huge) { |
809 | VM_BUG_ON(!is_first_page(page)); | |
810 | return *(unsigned long *)page_private(page); | |
811 | } else | |
812 | return *(unsigned long *)obj; | |
312fcae2 MK |
813 | } |
814 | ||
61989a80 NG |
815 | static unsigned long obj_idx_to_offset(struct page *page, |
816 | unsigned long obj_idx, int class_size) | |
817 | { | |
818 | unsigned long off = 0; | |
819 | ||
820 | if (!is_first_page(page)) | |
821 | off = page->index; | |
822 | ||
823 | return off + obj_idx * class_size; | |
824 | } | |
825 | ||
312fcae2 MK |
826 | static inline int trypin_tag(unsigned long handle) |
827 | { | |
828 | unsigned long *ptr = (unsigned long *)handle; | |
829 | ||
830 | return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr); | |
831 | } | |
832 | ||
833 | static void pin_tag(unsigned long handle) | |
834 | { | |
835 | while (!trypin_tag(handle)); | |
836 | } | |
837 | ||
838 | static void unpin_tag(unsigned long handle) | |
839 | { | |
840 | unsigned long *ptr = (unsigned long *)handle; | |
841 | ||
842 | clear_bit_unlock(HANDLE_PIN_BIT, ptr); | |
843 | } | |
844 | ||
f4477e90 NG |
845 | static void reset_page(struct page *page) |
846 | { | |
847 | clear_bit(PG_private, &page->flags); | |
848 | clear_bit(PG_private_2, &page->flags); | |
849 | set_page_private(page, 0); | |
850 | page->mapping = NULL; | |
851 | page->freelist = NULL; | |
22b751c3 | 852 | page_mapcount_reset(page); |
f4477e90 NG |
853 | } |
854 | ||
61989a80 NG |
855 | static void free_zspage(struct page *first_page) |
856 | { | |
f4477e90 | 857 | struct page *nextp, *tmp, *head_extra; |
61989a80 NG |
858 | |
859 | BUG_ON(!is_first_page(first_page)); | |
860 | BUG_ON(first_page->inuse); | |
861 | ||
f4477e90 | 862 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 863 | |
f4477e90 | 864 | reset_page(first_page); |
61989a80 NG |
865 | __free_page(first_page); |
866 | ||
867 | /* zspage with only 1 system page */ | |
f4477e90 | 868 | if (!head_extra) |
61989a80 NG |
869 | return; |
870 | ||
f4477e90 | 871 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 872 | list_del(&nextp->lru); |
f4477e90 | 873 | reset_page(nextp); |
61989a80 NG |
874 | __free_page(nextp); |
875 | } | |
f4477e90 NG |
876 | reset_page(head_extra); |
877 | __free_page(head_extra); | |
61989a80 NG |
878 | } |
879 | ||
880 | /* Initialize a newly allocated zspage */ | |
881 | static void init_zspage(struct page *first_page, struct size_class *class) | |
882 | { | |
883 | unsigned long off = 0; | |
884 | struct page *page = first_page; | |
885 | ||
886 | BUG_ON(!is_first_page(first_page)); | |
887 | while (page) { | |
888 | struct page *next_page; | |
889 | struct link_free *link; | |
5538c562 | 890 | unsigned int i = 1; |
af4ee5e9 | 891 | void *vaddr; |
61989a80 NG |
892 | |
893 | /* | |
894 | * page->index stores offset of first object starting | |
895 | * in the page. For the first page, this is always 0, | |
896 | * so we use first_page->index (aka ->freelist) to store | |
897 | * head of corresponding zspage's freelist. | |
898 | */ | |
899 | if (page != first_page) | |
900 | page->index = off; | |
901 | ||
af4ee5e9 MK |
902 | vaddr = kmap_atomic(page); |
903 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
904 | |
905 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 906 | link->next = location_to_obj(page, i++); |
5538c562 | 907 | link += class->size / sizeof(*link); |
61989a80 NG |
908 | } |
909 | ||
910 | /* | |
911 | * We now come to the last (full or partial) object on this | |
912 | * page, which must point to the first object on the next | |
913 | * page (if present) | |
914 | */ | |
915 | next_page = get_next_page(page); | |
312fcae2 | 916 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 917 | kunmap_atomic(vaddr); |
61989a80 | 918 | page = next_page; |
5538c562 | 919 | off %= PAGE_SIZE; |
61989a80 NG |
920 | } |
921 | } | |
922 | ||
923 | /* | |
924 | * Allocate a zspage for the given size class | |
925 | */ | |
926 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
927 | { | |
928 | int i, error; | |
b4b700c5 | 929 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
930 | |
931 | /* | |
932 | * Allocate individual pages and link them together as: | |
933 | * 1. first page->private = first sub-page | |
934 | * 2. all sub-pages are linked together using page->lru | |
935 | * 3. each sub-page is linked to the first page using page->first_page | |
936 | * | |
937 | * For each size class, First/Head pages are linked together using | |
938 | * page->lru. Also, we set PG_private to identify the first page | |
939 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
940 | * identify the last page. | |
941 | */ | |
942 | error = -ENOMEM; | |
2e3b6154 | 943 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 944 | struct page *page; |
61989a80 NG |
945 | |
946 | page = alloc_page(flags); | |
947 | if (!page) | |
948 | goto cleanup; | |
949 | ||
950 | INIT_LIST_HEAD(&page->lru); | |
951 | if (i == 0) { /* first page */ | |
a27545bf | 952 | SetPagePrivate(page); |
61989a80 NG |
953 | set_page_private(page, 0); |
954 | first_page = page; | |
955 | first_page->inuse = 0; | |
956 | } | |
957 | if (i == 1) | |
e842b976 | 958 | set_page_private(first_page, (unsigned long)page); |
61989a80 NG |
959 | if (i >= 1) |
960 | page->first_page = first_page; | |
961 | if (i >= 2) | |
962 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 963 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 964 | SetPagePrivate2(page); |
61989a80 NG |
965 | prev_page = page; |
966 | } | |
967 | ||
968 | init_zspage(first_page, class); | |
969 | ||
312fcae2 | 970 | first_page->freelist = location_to_obj(first_page, 0); |
61989a80 | 971 | /* Maximum number of objects we can store in this zspage */ |
2e3b6154 | 972 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
973 | |
974 | error = 0; /* Success */ | |
975 | ||
976 | cleanup: | |
977 | if (unlikely(error) && first_page) { | |
978 | free_zspage(first_page); | |
979 | first_page = NULL; | |
980 | } | |
981 | ||
982 | return first_page; | |
983 | } | |
984 | ||
985 | static struct page *find_get_zspage(struct size_class *class) | |
986 | { | |
987 | int i; | |
988 | struct page *page; | |
989 | ||
990 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
991 | page = class->fullness_list[i]; | |
992 | if (page) | |
993 | break; | |
994 | } | |
995 | ||
996 | return page; | |
997 | } | |
998 | ||
1b945aee | 999 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1000 | static inline int __zs_cpu_up(struct mapping_area *area) |
1001 | { | |
1002 | /* | |
1003 | * Make sure we don't leak memory if a cpu UP notification | |
1004 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1005 | */ | |
1006 | if (area->vm) | |
1007 | return 0; | |
1008 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1009 | if (!area->vm) | |
1010 | return -ENOMEM; | |
1011 | return 0; | |
1012 | } | |
1013 | ||
1014 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1015 | { | |
1016 | if (area->vm) | |
1017 | free_vm_area(area->vm); | |
1018 | area->vm = NULL; | |
1019 | } | |
1020 | ||
1021 | static inline void *__zs_map_object(struct mapping_area *area, | |
1022 | struct page *pages[2], int off, int size) | |
1023 | { | |
f6f8ed47 | 1024 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1025 | area->vm_addr = area->vm->addr; |
1026 | return area->vm_addr + off; | |
1027 | } | |
1028 | ||
1029 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1030 | struct page *pages[2], int off, int size) | |
1031 | { | |
1032 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1033 | |
d95abbbb | 1034 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1035 | } |
1036 | ||
1b945aee | 1037 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1038 | |
1039 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1040 | { | |
1041 | /* | |
1042 | * Make sure we don't leak memory if a cpu UP notification | |
1043 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1044 | */ | |
1045 | if (area->vm_buf) | |
1046 | return 0; | |
40f9fb8c | 1047 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1048 | if (!area->vm_buf) |
1049 | return -ENOMEM; | |
1050 | return 0; | |
1051 | } | |
1052 | ||
1053 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1054 | { | |
40f9fb8c | 1055 | kfree(area->vm_buf); |
f553646a SJ |
1056 | area->vm_buf = NULL; |
1057 | } | |
1058 | ||
1059 | static void *__zs_map_object(struct mapping_area *area, | |
1060 | struct page *pages[2], int off, int size) | |
5f601902 | 1061 | { |
5f601902 SJ |
1062 | int sizes[2]; |
1063 | void *addr; | |
f553646a | 1064 | char *buf = area->vm_buf; |
5f601902 | 1065 | |
f553646a SJ |
1066 | /* disable page faults to match kmap_atomic() return conditions */ |
1067 | pagefault_disable(); | |
1068 | ||
1069 | /* no read fastpath */ | |
1070 | if (area->vm_mm == ZS_MM_WO) | |
1071 | goto out; | |
5f601902 SJ |
1072 | |
1073 | sizes[0] = PAGE_SIZE - off; | |
1074 | sizes[1] = size - sizes[0]; | |
1075 | ||
5f601902 SJ |
1076 | /* copy object to per-cpu buffer */ |
1077 | addr = kmap_atomic(pages[0]); | |
1078 | memcpy(buf, addr + off, sizes[0]); | |
1079 | kunmap_atomic(addr); | |
1080 | addr = kmap_atomic(pages[1]); | |
1081 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1082 | kunmap_atomic(addr); | |
f553646a SJ |
1083 | out: |
1084 | return area->vm_buf; | |
5f601902 SJ |
1085 | } |
1086 | ||
f553646a SJ |
1087 | static void __zs_unmap_object(struct mapping_area *area, |
1088 | struct page *pages[2], int off, int size) | |
5f601902 | 1089 | { |
5f601902 SJ |
1090 | int sizes[2]; |
1091 | void *addr; | |
2e40e163 | 1092 | char *buf; |
5f601902 | 1093 | |
f553646a SJ |
1094 | /* no write fastpath */ |
1095 | if (area->vm_mm == ZS_MM_RO) | |
1096 | goto out; | |
5f601902 | 1097 | |
7b60a685 MK |
1098 | buf = area->vm_buf; |
1099 | if (!area->huge) { | |
1100 | buf = buf + ZS_HANDLE_SIZE; | |
1101 | size -= ZS_HANDLE_SIZE; | |
1102 | off += ZS_HANDLE_SIZE; | |
1103 | } | |
2e40e163 | 1104 | |
5f601902 SJ |
1105 | sizes[0] = PAGE_SIZE - off; |
1106 | sizes[1] = size - sizes[0]; | |
1107 | ||
1108 | /* copy per-cpu buffer to object */ | |
1109 | addr = kmap_atomic(pages[0]); | |
1110 | memcpy(addr + off, buf, sizes[0]); | |
1111 | kunmap_atomic(addr); | |
1112 | addr = kmap_atomic(pages[1]); | |
1113 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1114 | kunmap_atomic(addr); | |
f553646a SJ |
1115 | |
1116 | out: | |
1117 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1118 | pagefault_enable(); | |
5f601902 | 1119 | } |
61989a80 | 1120 | |
1b945aee | 1121 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1122 | |
61989a80 NG |
1123 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1124 | void *pcpu) | |
1125 | { | |
f553646a | 1126 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1127 | struct mapping_area *area; |
1128 | ||
1129 | switch (action) { | |
1130 | case CPU_UP_PREPARE: | |
1131 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1132 | ret = __zs_cpu_up(area); |
1133 | if (ret) | |
1134 | return notifier_from_errno(ret); | |
61989a80 NG |
1135 | break; |
1136 | case CPU_DEAD: | |
1137 | case CPU_UP_CANCELED: | |
1138 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1139 | __zs_cpu_down(area); |
61989a80 NG |
1140 | break; |
1141 | } | |
1142 | ||
1143 | return NOTIFY_OK; | |
1144 | } | |
1145 | ||
1146 | static struct notifier_block zs_cpu_nb = { | |
1147 | .notifier_call = zs_cpu_notifier | |
1148 | }; | |
1149 | ||
b1b00a5b | 1150 | static int zs_register_cpu_notifier(void) |
61989a80 | 1151 | { |
b1b00a5b | 1152 | int cpu, uninitialized_var(ret); |
61989a80 | 1153 | |
f0e71fcd SB |
1154 | cpu_notifier_register_begin(); |
1155 | ||
1156 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1157 | for_each_online_cpu(cpu) { |
1158 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1159 | if (notifier_to_errno(ret)) |
1160 | break; | |
61989a80 | 1161 | } |
f0e71fcd SB |
1162 | |
1163 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1164 | return notifier_to_errno(ret); |
1165 | } | |
f0e71fcd | 1166 | |
66cdef66 | 1167 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1168 | { |
66cdef66 | 1169 | int cpu; |
40f9fb8c | 1170 | |
66cdef66 | 1171 | cpu_notifier_register_begin(); |
40f9fb8c | 1172 | |
66cdef66 GM |
1173 | for_each_online_cpu(cpu) |
1174 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1175 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1176 | |
66cdef66 | 1177 | cpu_notifier_register_done(); |
b1b00a5b SS |
1178 | } |
1179 | ||
66cdef66 | 1180 | static void init_zs_size_classes(void) |
b1b00a5b | 1181 | { |
66cdef66 | 1182 | int nr; |
c795779d | 1183 | |
66cdef66 GM |
1184 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1185 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1186 | nr += 1; | |
40f9fb8c | 1187 | |
66cdef66 | 1188 | zs_size_classes = nr; |
61989a80 NG |
1189 | } |
1190 | ||
9eec4cd5 JK |
1191 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1192 | { | |
1193 | if (prev->pages_per_zspage != pages_per_zspage) | |
1194 | return false; | |
1195 | ||
1196 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1197 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1198 | return false; | |
1199 | ||
1200 | return true; | |
1201 | } | |
1202 | ||
312fcae2 MK |
1203 | static bool zspage_full(struct page *page) |
1204 | { | |
1205 | BUG_ON(!is_first_page(page)); | |
1206 | ||
1207 | return page->inuse == page->objects; | |
1208 | } | |
1209 | ||
66cdef66 GM |
1210 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1211 | { | |
1212 | return atomic_long_read(&pool->pages_allocated); | |
1213 | } | |
1214 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1215 | ||
4bbc0bc0 | 1216 | /** |
66cdef66 GM |
1217 | * zs_map_object - get address of allocated object from handle. |
1218 | * @pool: pool from which the object was allocated | |
1219 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1220 | * |
66cdef66 GM |
1221 | * Before using an object allocated from zs_malloc, it must be mapped using |
1222 | * this function. When done with the object, it must be unmapped using | |
1223 | * zs_unmap_object. | |
4bbc0bc0 | 1224 | * |
66cdef66 GM |
1225 | * Only one object can be mapped per cpu at a time. There is no protection |
1226 | * against nested mappings. | |
1227 | * | |
1228 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1229 | */ |
66cdef66 GM |
1230 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1231 | enum zs_mapmode mm) | |
61989a80 | 1232 | { |
66cdef66 | 1233 | struct page *page; |
2e40e163 | 1234 | unsigned long obj, obj_idx, off; |
61989a80 | 1235 | |
66cdef66 GM |
1236 | unsigned int class_idx; |
1237 | enum fullness_group fg; | |
1238 | struct size_class *class; | |
1239 | struct mapping_area *area; | |
1240 | struct page *pages[2]; | |
2e40e163 | 1241 | void *ret; |
61989a80 | 1242 | |
66cdef66 | 1243 | BUG_ON(!handle); |
40f9fb8c | 1244 | |
9eec4cd5 | 1245 | /* |
66cdef66 GM |
1246 | * Because we use per-cpu mapping areas shared among the |
1247 | * pools/users, we can't allow mapping in interrupt context | |
1248 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1249 | */ |
66cdef66 | 1250 | BUG_ON(in_interrupt()); |
61989a80 | 1251 | |
312fcae2 MK |
1252 | /* From now on, migration cannot move the object */ |
1253 | pin_tag(handle); | |
1254 | ||
2e40e163 MK |
1255 | obj = handle_to_obj(handle); |
1256 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1257 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1258 | class = pool->size_class[class_idx]; | |
1259 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1260 | |
66cdef66 GM |
1261 | area = &get_cpu_var(zs_map_area); |
1262 | area->vm_mm = mm; | |
1263 | if (off + class->size <= PAGE_SIZE) { | |
1264 | /* this object is contained entirely within a page */ | |
1265 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1266 | ret = area->vm_addr + off; |
1267 | goto out; | |
61989a80 NG |
1268 | } |
1269 | ||
66cdef66 GM |
1270 | /* this object spans two pages */ |
1271 | pages[0] = page; | |
1272 | pages[1] = get_next_page(page); | |
1273 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1274 | |
2e40e163 MK |
1275 | ret = __zs_map_object(area, pages, off, class->size); |
1276 | out: | |
7b60a685 MK |
1277 | if (!class->huge) |
1278 | ret += ZS_HANDLE_SIZE; | |
1279 | ||
1280 | return ret; | |
61989a80 | 1281 | } |
66cdef66 | 1282 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1283 | |
66cdef66 | 1284 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1285 | { |
66cdef66 | 1286 | struct page *page; |
2e40e163 | 1287 | unsigned long obj, obj_idx, off; |
61989a80 | 1288 | |
66cdef66 GM |
1289 | unsigned int class_idx; |
1290 | enum fullness_group fg; | |
1291 | struct size_class *class; | |
1292 | struct mapping_area *area; | |
9eec4cd5 | 1293 | |
66cdef66 | 1294 | BUG_ON(!handle); |
9eec4cd5 | 1295 | |
2e40e163 MK |
1296 | obj = handle_to_obj(handle); |
1297 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1298 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1299 | class = pool->size_class[class_idx]; | |
1300 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1301 | |
66cdef66 GM |
1302 | area = this_cpu_ptr(&zs_map_area); |
1303 | if (off + class->size <= PAGE_SIZE) | |
1304 | kunmap_atomic(area->vm_addr); | |
1305 | else { | |
1306 | struct page *pages[2]; | |
40f9fb8c | 1307 | |
66cdef66 GM |
1308 | pages[0] = page; |
1309 | pages[1] = get_next_page(page); | |
1310 | BUG_ON(!pages[1]); | |
1311 | ||
1312 | __zs_unmap_object(area, pages, off, class->size); | |
1313 | } | |
1314 | put_cpu_var(zs_map_area); | |
312fcae2 | 1315 | unpin_tag(handle); |
61989a80 | 1316 | } |
66cdef66 | 1317 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1318 | |
c7806261 MK |
1319 | static unsigned long obj_malloc(struct page *first_page, |
1320 | struct size_class *class, unsigned long handle) | |
1321 | { | |
1322 | unsigned long obj; | |
1323 | struct link_free *link; | |
1324 | ||
1325 | struct page *m_page; | |
1326 | unsigned long m_objidx, m_offset; | |
1327 | void *vaddr; | |
1328 | ||
312fcae2 | 1329 | handle |= OBJ_ALLOCATED_TAG; |
c7806261 MK |
1330 | obj = (unsigned long)first_page->freelist; |
1331 | obj_to_location(obj, &m_page, &m_objidx); | |
1332 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1333 | ||
1334 | vaddr = kmap_atomic(m_page); | |
1335 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1336 | first_page->freelist = link->next; | |
7b60a685 MK |
1337 | if (!class->huge) |
1338 | /* record handle in the header of allocated chunk */ | |
1339 | link->handle = handle; | |
1340 | else | |
1341 | /* record handle in first_page->private */ | |
1342 | set_page_private(first_page, handle); | |
c7806261 MK |
1343 | kunmap_atomic(vaddr); |
1344 | first_page->inuse++; | |
1345 | zs_stat_inc(class, OBJ_USED, 1); | |
1346 | ||
1347 | return obj; | |
1348 | } | |
1349 | ||
1350 | ||
61989a80 NG |
1351 | /** |
1352 | * zs_malloc - Allocate block of given size from pool. | |
1353 | * @pool: pool to allocate from | |
1354 | * @size: size of block to allocate | |
61989a80 | 1355 | * |
00a61d86 | 1356 | * On success, handle to the allocated object is returned, |
c2344348 | 1357 | * otherwise 0. |
61989a80 NG |
1358 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1359 | */ | |
c2344348 | 1360 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
61989a80 | 1361 | { |
2e40e163 | 1362 | unsigned long handle, obj; |
61989a80 | 1363 | struct size_class *class; |
c7806261 | 1364 | struct page *first_page; |
61989a80 | 1365 | |
7b60a685 | 1366 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1367 | return 0; |
1368 | ||
1369 | handle = alloc_handle(pool); | |
1370 | if (!handle) | |
c2344348 | 1371 | return 0; |
61989a80 | 1372 | |
2e40e163 MK |
1373 | /* extra space in chunk to keep the handle */ |
1374 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1375 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1376 | |
1377 | spin_lock(&class->lock); | |
1378 | first_page = find_get_zspage(class); | |
1379 | ||
1380 | if (!first_page) { | |
1381 | spin_unlock(&class->lock); | |
1382 | first_page = alloc_zspage(class, pool->flags); | |
2e40e163 MK |
1383 | if (unlikely(!first_page)) { |
1384 | free_handle(pool, handle); | |
c2344348 | 1385 | return 0; |
2e40e163 | 1386 | } |
61989a80 NG |
1387 | |
1388 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1389 | atomic_long_add(class->pages_per_zspage, |
1390 | &pool->pages_allocated); | |
0f050d99 | 1391 | |
61989a80 | 1392 | spin_lock(&class->lock); |
0f050d99 GM |
1393 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1394 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1395 | } |
1396 | ||
c7806261 | 1397 | obj = obj_malloc(first_page, class, handle); |
61989a80 | 1398 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1399 | fix_fullness_group(class, first_page); |
2e40e163 | 1400 | record_obj(handle, obj); |
61989a80 NG |
1401 | spin_unlock(&class->lock); |
1402 | ||
2e40e163 | 1403 | return handle; |
61989a80 NG |
1404 | } |
1405 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1406 | ||
c7806261 MK |
1407 | static void obj_free(struct zs_pool *pool, struct size_class *class, |
1408 | unsigned long obj) | |
61989a80 NG |
1409 | { |
1410 | struct link_free *link; | |
1411 | struct page *first_page, *f_page; | |
c7806261 | 1412 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1413 | void *vaddr; |
61989a80 | 1414 | int class_idx; |
61989a80 NG |
1415 | enum fullness_group fullness; |
1416 | ||
c7806261 | 1417 | BUG_ON(!obj); |
61989a80 | 1418 | |
312fcae2 | 1419 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1420 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1421 | first_page = get_first_page(f_page); |
1422 | ||
1423 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
61989a80 NG |
1424 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1425 | ||
c7806261 | 1426 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1427 | |
1428 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1429 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1430 | link->next = first_page->freelist; |
7b60a685 MK |
1431 | if (class->huge) |
1432 | set_page_private(first_page, 0); | |
af4ee5e9 | 1433 | kunmap_atomic(vaddr); |
c2344348 | 1434 | first_page->freelist = (void *)obj; |
61989a80 | 1435 | first_page->inuse--; |
0f050d99 | 1436 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1437 | } |
1438 | ||
1439 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1440 | { | |
1441 | struct page *first_page, *f_page; | |
1442 | unsigned long obj, f_objidx; | |
1443 | int class_idx; | |
1444 | struct size_class *class; | |
1445 | enum fullness_group fullness; | |
1446 | ||
1447 | if (unlikely(!handle)) | |
1448 | return; | |
1449 | ||
312fcae2 | 1450 | pin_tag(handle); |
c7806261 | 1451 | obj = handle_to_obj(handle); |
c7806261 MK |
1452 | obj_to_location(obj, &f_page, &f_objidx); |
1453 | first_page = get_first_page(f_page); | |
1454 | ||
1455 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1456 | class = pool->size_class[class_idx]; | |
1457 | ||
1458 | spin_lock(&class->lock); | |
1459 | obj_free(pool, class, obj); | |
1460 | fullness = fix_fullness_group(class, first_page); | |
312fcae2 | 1461 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1462 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1463 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1464 | atomic_long_sub(class->pages_per_zspage, |
1465 | &pool->pages_allocated); | |
1466 | free_zspage(first_page); | |
1467 | } | |
61989a80 | 1468 | spin_unlock(&class->lock); |
312fcae2 | 1469 | unpin_tag(handle); |
61989a80 | 1470 | |
312fcae2 MK |
1471 | free_handle(pool, handle); |
1472 | } | |
1473 | EXPORT_SYMBOL_GPL(zs_free); | |
1474 | ||
0dc63d48 | 1475 | static void zs_object_copy(unsigned long dst, unsigned long src, |
312fcae2 MK |
1476 | struct size_class *class) |
1477 | { | |
1478 | struct page *s_page, *d_page; | |
1479 | unsigned long s_objidx, d_objidx; | |
1480 | unsigned long s_off, d_off; | |
1481 | void *s_addr, *d_addr; | |
1482 | int s_size, d_size, size; | |
1483 | int written = 0; | |
1484 | ||
1485 | s_size = d_size = class->size; | |
1486 | ||
1487 | obj_to_location(src, &s_page, &s_objidx); | |
1488 | obj_to_location(dst, &d_page, &d_objidx); | |
1489 | ||
1490 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1491 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1492 | ||
1493 | if (s_off + class->size > PAGE_SIZE) | |
1494 | s_size = PAGE_SIZE - s_off; | |
1495 | ||
1496 | if (d_off + class->size > PAGE_SIZE) | |
1497 | d_size = PAGE_SIZE - d_off; | |
1498 | ||
1499 | s_addr = kmap_atomic(s_page); | |
1500 | d_addr = kmap_atomic(d_page); | |
1501 | ||
1502 | while (1) { | |
1503 | size = min(s_size, d_size); | |
1504 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1505 | written += size; | |
1506 | ||
1507 | if (written == class->size) | |
1508 | break; | |
1509 | ||
495819ea SS |
1510 | s_off += size; |
1511 | s_size -= size; | |
1512 | d_off += size; | |
1513 | d_size -= size; | |
1514 | ||
1515 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1516 | kunmap_atomic(d_addr); |
1517 | kunmap_atomic(s_addr); | |
1518 | s_page = get_next_page(s_page); | |
1519 | BUG_ON(!s_page); | |
1520 | s_addr = kmap_atomic(s_page); | |
1521 | d_addr = kmap_atomic(d_page); | |
1522 | s_size = class->size - written; | |
1523 | s_off = 0; | |
312fcae2 MK |
1524 | } |
1525 | ||
495819ea | 1526 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1527 | kunmap_atomic(d_addr); |
1528 | d_page = get_next_page(d_page); | |
1529 | BUG_ON(!d_page); | |
1530 | d_addr = kmap_atomic(d_page); | |
1531 | d_size = class->size - written; | |
1532 | d_off = 0; | |
312fcae2 MK |
1533 | } |
1534 | } | |
1535 | ||
1536 | kunmap_atomic(d_addr); | |
1537 | kunmap_atomic(s_addr); | |
1538 | } | |
1539 | ||
1540 | /* | |
1541 | * Find alloced object in zspage from index object and | |
1542 | * return handle. | |
1543 | */ | |
1544 | static unsigned long find_alloced_obj(struct page *page, int index, | |
1545 | struct size_class *class) | |
1546 | { | |
1547 | unsigned long head; | |
1548 | int offset = 0; | |
1549 | unsigned long handle = 0; | |
1550 | void *addr = kmap_atomic(page); | |
1551 | ||
1552 | if (!is_first_page(page)) | |
1553 | offset = page->index; | |
1554 | offset += class->size * index; | |
1555 | ||
1556 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1557 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1558 | if (head & OBJ_ALLOCATED_TAG) { |
1559 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1560 | if (trypin_tag(handle)) | |
1561 | break; | |
1562 | handle = 0; | |
1563 | } | |
1564 | ||
1565 | offset += class->size; | |
1566 | index++; | |
1567 | } | |
1568 | ||
1569 | kunmap_atomic(addr); | |
1570 | return handle; | |
1571 | } | |
1572 | ||
1573 | struct zs_compact_control { | |
1574 | /* Source page for migration which could be a subpage of zspage. */ | |
1575 | struct page *s_page; | |
1576 | /* Destination page for migration which should be a first page | |
1577 | * of zspage. */ | |
1578 | struct page *d_page; | |
1579 | /* Starting object index within @s_page which used for live object | |
1580 | * in the subpage. */ | |
1581 | int index; | |
7d3f3938 | 1582 | /* How many of objects were migrated */ |
312fcae2 MK |
1583 | int nr_migrated; |
1584 | }; | |
1585 | ||
1586 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1587 | struct zs_compact_control *cc) | |
1588 | { | |
1589 | unsigned long used_obj, free_obj; | |
1590 | unsigned long handle; | |
1591 | struct page *s_page = cc->s_page; | |
1592 | struct page *d_page = cc->d_page; | |
1593 | unsigned long index = cc->index; | |
312fcae2 MK |
1594 | int ret = 0; |
1595 | ||
1596 | while (1) { | |
1597 | handle = find_alloced_obj(s_page, index, class); | |
1598 | if (!handle) { | |
1599 | s_page = get_next_page(s_page); | |
1600 | if (!s_page) | |
1601 | break; | |
1602 | index = 0; | |
1603 | continue; | |
1604 | } | |
1605 | ||
1606 | /* Stop if there is no more space */ | |
1607 | if (zspage_full(d_page)) { | |
1608 | unpin_tag(handle); | |
1609 | ret = -ENOMEM; | |
1610 | break; | |
1611 | } | |
1612 | ||
1613 | used_obj = handle_to_obj(handle); | |
1614 | free_obj = obj_malloc(d_page, class, handle); | |
0dc63d48 | 1615 | zs_object_copy(free_obj, used_obj, class); |
312fcae2 MK |
1616 | index++; |
1617 | record_obj(handle, free_obj); | |
1618 | unpin_tag(handle); | |
1619 | obj_free(pool, class, used_obj); | |
7d3f3938 | 1620 | cc->nr_migrated++; |
312fcae2 MK |
1621 | } |
1622 | ||
1623 | /* Remember last position in this iteration */ | |
1624 | cc->s_page = s_page; | |
1625 | cc->index = index; | |
312fcae2 MK |
1626 | |
1627 | return ret; | |
1628 | } | |
1629 | ||
0dc63d48 | 1630 | static struct page *isolate_target_page(struct size_class *class) |
312fcae2 MK |
1631 | { |
1632 | int i; | |
1633 | struct page *page; | |
1634 | ||
1635 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1636 | page = class->fullness_list[i]; | |
1637 | if (page) { | |
1638 | remove_zspage(page, class, i); | |
1639 | break; | |
1640 | } | |
1641 | } | |
1642 | ||
1643 | return page; | |
1644 | } | |
1645 | ||
1646 | static void putback_zspage(struct zs_pool *pool, struct size_class *class, | |
1647 | struct page *first_page) | |
1648 | { | |
312fcae2 MK |
1649 | enum fullness_group fullness; |
1650 | ||
1651 | BUG_ON(!is_first_page(first_page)); | |
1652 | ||
839373e6 | 1653 | fullness = get_fullness_group(first_page); |
312fcae2 | 1654 | insert_zspage(first_page, class, fullness); |
839373e6 MK |
1655 | set_zspage_mapping(first_page, class->index, fullness); |
1656 | ||
13de8933 | 1657 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1658 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1659 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1660 | atomic_long_sub(class->pages_per_zspage, |
1661 | &pool->pages_allocated); | |
312fcae2 | 1662 | |
61989a80 | 1663 | free_zspage(first_page); |
13de8933 | 1664 | } |
61989a80 | 1665 | } |
312fcae2 MK |
1666 | |
1667 | static struct page *isolate_source_page(struct size_class *class) | |
1668 | { | |
1669 | struct page *page; | |
1670 | ||
1671 | page = class->fullness_list[ZS_ALMOST_EMPTY]; | |
1672 | if (page) | |
1673 | remove_zspage(page, class, ZS_ALMOST_EMPTY); | |
1674 | ||
1675 | return page; | |
1676 | } | |
1677 | ||
04f05909 SS |
1678 | /* |
1679 | * | |
1680 | * Based on the number of unused allocated objects calculate | |
1681 | * and return the number of pages that we can free. | |
1682 | * | |
1683 | * Should be called under class->lock. | |
1684 | */ | |
1685 | static unsigned long zs_can_compact(struct size_class *class) | |
1686 | { | |
1687 | unsigned long obj_wasted; | |
1688 | ||
1689 | if (!zs_stat_get(class, CLASS_ALMOST_EMPTY)) | |
1690 | return 0; | |
1691 | ||
1692 | obj_wasted = zs_stat_get(class, OBJ_ALLOCATED) - | |
1693 | zs_stat_get(class, OBJ_USED); | |
1694 | ||
1695 | obj_wasted /= get_maxobj_per_zspage(class->size, | |
1696 | class->pages_per_zspage); | |
1697 | ||
1698 | return obj_wasted * get_pages_per_zspage(class->size); | |
1699 | } | |
1700 | ||
7d3f3938 | 1701 | static void __zs_compact(struct zs_pool *pool, struct size_class *class) |
312fcae2 | 1702 | { |
312fcae2 MK |
1703 | struct zs_compact_control cc; |
1704 | struct page *src_page; | |
1705 | struct page *dst_page = NULL; | |
312fcae2 | 1706 | |
7d3f3938 | 1707 | cc.nr_migrated = 0; |
312fcae2 MK |
1708 | spin_lock(&class->lock); |
1709 | while ((src_page = isolate_source_page(class))) { | |
1710 | ||
1711 | BUG_ON(!is_first_page(src_page)); | |
1712 | ||
04f05909 SS |
1713 | if (!zs_can_compact(class)) |
1714 | break; | |
1715 | ||
312fcae2 MK |
1716 | cc.index = 0; |
1717 | cc.s_page = src_page; | |
1718 | ||
0dc63d48 | 1719 | while ((dst_page = isolate_target_page(class))) { |
312fcae2 MK |
1720 | cc.d_page = dst_page; |
1721 | /* | |
0dc63d48 SS |
1722 | * If there is no more space in dst_page, resched |
1723 | * and see if anyone had allocated another zspage. | |
312fcae2 MK |
1724 | */ |
1725 | if (!migrate_zspage(pool, class, &cc)) | |
1726 | break; | |
1727 | ||
1728 | putback_zspage(pool, class, dst_page); | |
312fcae2 MK |
1729 | } |
1730 | ||
1731 | /* Stop if we couldn't find slot */ | |
1732 | if (dst_page == NULL) | |
1733 | break; | |
1734 | ||
1735 | putback_zspage(pool, class, dst_page); | |
1736 | putback_zspage(pool, class, src_page); | |
1737 | spin_unlock(&class->lock); | |
312fcae2 MK |
1738 | cond_resched(); |
1739 | spin_lock(&class->lock); | |
1740 | } | |
1741 | ||
1742 | if (src_page) | |
1743 | putback_zspage(pool, class, src_page); | |
1744 | ||
7d3f3938 | 1745 | pool->stats.num_migrated += cc.nr_migrated; |
312fcae2 | 1746 | |
7d3f3938 | 1747 | spin_unlock(&class->lock); |
312fcae2 MK |
1748 | } |
1749 | ||
1750 | unsigned long zs_compact(struct zs_pool *pool) | |
1751 | { | |
1752 | int i; | |
312fcae2 MK |
1753 | struct size_class *class; |
1754 | ||
1755 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1756 | class = pool->size_class[i]; | |
1757 | if (!class) | |
1758 | continue; | |
1759 | if (class->index != i) | |
1760 | continue; | |
7d3f3938 | 1761 | __zs_compact(pool, class); |
312fcae2 MK |
1762 | } |
1763 | ||
7d3f3938 | 1764 | return pool->stats.num_migrated; |
312fcae2 MK |
1765 | } |
1766 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1767 | |
7d3f3938 SS |
1768 | void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats) |
1769 | { | |
1770 | memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats)); | |
1771 | } | |
1772 | EXPORT_SYMBOL_GPL(zs_pool_stats); | |
1773 | ||
00a61d86 | 1774 | /** |
66cdef66 GM |
1775 | * zs_create_pool - Creates an allocation pool to work from. |
1776 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1777 | * |
66cdef66 GM |
1778 | * This function must be called before anything when using |
1779 | * the zsmalloc allocator. | |
166cfda7 | 1780 | * |
66cdef66 GM |
1781 | * On success, a pointer to the newly created pool is returned, |
1782 | * otherwise NULL. | |
396b7fd6 | 1783 | */ |
3eba0c6a | 1784 | struct zs_pool *zs_create_pool(char *name, gfp_t flags) |
61989a80 | 1785 | { |
66cdef66 GM |
1786 | int i; |
1787 | struct zs_pool *pool; | |
1788 | struct size_class *prev_class = NULL; | |
61989a80 | 1789 | |
66cdef66 GM |
1790 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1791 | if (!pool) | |
1792 | return NULL; | |
61989a80 | 1793 | |
66cdef66 GM |
1794 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1795 | GFP_KERNEL); | |
1796 | if (!pool->size_class) { | |
1797 | kfree(pool); | |
1798 | return NULL; | |
1799 | } | |
61989a80 | 1800 | |
2e40e163 MK |
1801 | pool->name = kstrdup(name, GFP_KERNEL); |
1802 | if (!pool->name) | |
1803 | goto err; | |
1804 | ||
1805 | if (create_handle_cache(pool)) | |
1806 | goto err; | |
1807 | ||
c60369f0 | 1808 | /* |
66cdef66 GM |
1809 | * Iterate reversly, because, size of size_class that we want to use |
1810 | * for merging should be larger or equal to current size. | |
c60369f0 | 1811 | */ |
66cdef66 GM |
1812 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1813 | int size; | |
1814 | int pages_per_zspage; | |
1815 | struct size_class *class; | |
c60369f0 | 1816 | |
66cdef66 GM |
1817 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1818 | if (size > ZS_MAX_ALLOC_SIZE) | |
1819 | size = ZS_MAX_ALLOC_SIZE; | |
1820 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1821 | |
66cdef66 GM |
1822 | /* |
1823 | * size_class is used for normal zsmalloc operation such | |
1824 | * as alloc/free for that size. Although it is natural that we | |
1825 | * have one size_class for each size, there is a chance that we | |
1826 | * can get more memory utilization if we use one size_class for | |
1827 | * many different sizes whose size_class have same | |
1828 | * characteristics. So, we makes size_class point to | |
1829 | * previous size_class if possible. | |
1830 | */ | |
1831 | if (prev_class) { | |
1832 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1833 | pool->size_class[i] = prev_class; | |
1834 | continue; | |
1835 | } | |
1836 | } | |
1837 | ||
1838 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1839 | if (!class) | |
1840 | goto err; | |
1841 | ||
1842 | class->size = size; | |
1843 | class->index = i; | |
1844 | class->pages_per_zspage = pages_per_zspage; | |
7b60a685 MK |
1845 | if (pages_per_zspage == 1 && |
1846 | get_maxobj_per_zspage(size, pages_per_zspage) == 1) | |
1847 | class->huge = true; | |
66cdef66 GM |
1848 | spin_lock_init(&class->lock); |
1849 | pool->size_class[i] = class; | |
1850 | ||
1851 | prev_class = class; | |
61989a80 NG |
1852 | } |
1853 | ||
66cdef66 | 1854 | pool->flags = flags; |
b7418510 | 1855 | |
0f050d99 GM |
1856 | if (zs_pool_stat_create(name, pool)) |
1857 | goto err; | |
1858 | ||
66cdef66 GM |
1859 | return pool; |
1860 | ||
1861 | err: | |
1862 | zs_destroy_pool(pool); | |
1863 | return NULL; | |
61989a80 | 1864 | } |
66cdef66 | 1865 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1866 | |
66cdef66 | 1867 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1868 | { |
66cdef66 | 1869 | int i; |
61989a80 | 1870 | |
0f050d99 GM |
1871 | zs_pool_stat_destroy(pool); |
1872 | ||
66cdef66 GM |
1873 | for (i = 0; i < zs_size_classes; i++) { |
1874 | int fg; | |
1875 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1876 | |
66cdef66 GM |
1877 | if (!class) |
1878 | continue; | |
61989a80 | 1879 | |
66cdef66 GM |
1880 | if (class->index != i) |
1881 | continue; | |
61989a80 | 1882 | |
66cdef66 GM |
1883 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1884 | if (class->fullness_list[fg]) { | |
1885 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1886 | class->size, fg); | |
1887 | } | |
1888 | } | |
1889 | kfree(class); | |
1890 | } | |
f553646a | 1891 | |
2e40e163 | 1892 | destroy_handle_cache(pool); |
66cdef66 | 1893 | kfree(pool->size_class); |
0f050d99 | 1894 | kfree(pool->name); |
66cdef66 GM |
1895 | kfree(pool); |
1896 | } | |
1897 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 1898 | |
66cdef66 GM |
1899 | static int __init zs_init(void) |
1900 | { | |
1901 | int ret = zs_register_cpu_notifier(); | |
1902 | ||
0f050d99 GM |
1903 | if (ret) |
1904 | goto notifier_fail; | |
66cdef66 GM |
1905 | |
1906 | init_zs_size_classes(); | |
1907 | ||
1908 | #ifdef CONFIG_ZPOOL | |
1909 | zpool_register_driver(&zs_zpool_driver); | |
1910 | #endif | |
0f050d99 GM |
1911 | |
1912 | ret = zs_stat_init(); | |
1913 | if (ret) { | |
1914 | pr_err("zs stat initialization failed\n"); | |
1915 | goto stat_fail; | |
1916 | } | |
66cdef66 | 1917 | return 0; |
0f050d99 GM |
1918 | |
1919 | stat_fail: | |
1920 | #ifdef CONFIG_ZPOOL | |
1921 | zpool_unregister_driver(&zs_zpool_driver); | |
1922 | #endif | |
1923 | notifier_fail: | |
1924 | zs_unregister_cpu_notifier(); | |
1925 | ||
1926 | return ret; | |
61989a80 | 1927 | } |
61989a80 | 1928 | |
66cdef66 | 1929 | static void __exit zs_exit(void) |
61989a80 | 1930 | { |
66cdef66 GM |
1931 | #ifdef CONFIG_ZPOOL |
1932 | zpool_unregister_driver(&zs_zpool_driver); | |
1933 | #endif | |
1934 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
1935 | |
1936 | zs_stat_exit(); | |
61989a80 | 1937 | } |
069f101f BH |
1938 | |
1939 | module_init(zs_init); | |
1940 | module_exit(zs_exit); | |
1941 | ||
1942 | MODULE_LICENSE("Dual BSD/GPL"); | |
1943 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |