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