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 | ||
172 | #ifdef CONFIG_ZSMALLOC_STAT | |
173 | ||
174 | static struct dentry *zs_stat_root; | |
175 | ||
176 | struct zs_size_stat { | |
177 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
178 | }; | |
179 | ||
180 | #endif | |
181 | ||
40f9fb8c MG |
182 | /* |
183 | * number of size_classes | |
184 | */ | |
185 | static int zs_size_classes; | |
186 | ||
0959c63f SJ |
187 | /* |
188 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
189 | * n <= N / f, where | |
190 | * n = number of allocated objects | |
191 | * N = total number of objects zspage can store | |
6dd9737e | 192 | * f = fullness_threshold_frac |
0959c63f SJ |
193 | * |
194 | * Similarly, we assign zspage to: | |
195 | * ZS_ALMOST_FULL when n > N / f | |
196 | * ZS_EMPTY when n == 0 | |
197 | * ZS_FULL when n == N | |
198 | * | |
199 | * (see: fix_fullness_group()) | |
200 | */ | |
201 | static const int fullness_threshold_frac = 4; | |
202 | ||
203 | struct size_class { | |
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; | |
7b60a685 MK |
213 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
214 | bool huge; | |
0959c63f | 215 | |
0f050d99 GM |
216 | #ifdef CONFIG_ZSMALLOC_STAT |
217 | struct zs_size_stat stats; | |
218 | #endif | |
219 | ||
0959c63f SJ |
220 | spinlock_t lock; |
221 | ||
0959c63f SJ |
222 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; |
223 | }; | |
224 | ||
225 | /* | |
226 | * Placed within free objects to form a singly linked list. | |
227 | * For every zspage, first_page->freelist gives head of this list. | |
228 | * | |
229 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
230 | */ | |
231 | struct link_free { | |
2e40e163 MK |
232 | union { |
233 | /* | |
234 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
235 | * It's valid for non-allocated object | |
236 | */ | |
237 | void *next; | |
238 | /* | |
239 | * Handle of allocated object. | |
240 | */ | |
241 | unsigned long handle; | |
242 | }; | |
0959c63f SJ |
243 | }; |
244 | ||
245 | struct zs_pool { | |
0f050d99 GM |
246 | char *name; |
247 | ||
40f9fb8c | 248 | struct size_class **size_class; |
2e40e163 | 249 | struct kmem_cache *handle_cachep; |
0959c63f SJ |
250 | |
251 | gfp_t flags; /* allocation flags used when growing pool */ | |
13de8933 | 252 | atomic_long_t pages_allocated; |
0f050d99 GM |
253 | |
254 | #ifdef CONFIG_ZSMALLOC_STAT | |
255 | struct dentry *stat_dentry; | |
256 | #endif | |
0959c63f | 257 | }; |
61989a80 NG |
258 | |
259 | /* | |
260 | * A zspage's class index and fullness group | |
261 | * are encoded in its (first)page->mapping | |
262 | */ | |
263 | #define CLASS_IDX_BITS 28 | |
264 | #define FULLNESS_BITS 4 | |
265 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
266 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
267 | ||
f553646a | 268 | struct mapping_area { |
1b945aee | 269 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
270 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
271 | #else | |
272 | char *vm_buf; /* copy buffer for objects that span pages */ | |
273 | #endif | |
274 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
275 | enum zs_mapmode vm_mm; /* mapping mode */ | |
7b60a685 | 276 | bool huge; |
f553646a SJ |
277 | }; |
278 | ||
2e40e163 MK |
279 | static int create_handle_cache(struct zs_pool *pool) |
280 | { | |
281 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
282 | 0, 0, NULL); | |
283 | return pool->handle_cachep ? 0 : 1; | |
284 | } | |
285 | ||
286 | static void destroy_handle_cache(struct zs_pool *pool) | |
287 | { | |
02f7b414 SS |
288 | if (pool->handle_cachep) |
289 | kmem_cache_destroy(pool->handle_cachep); | |
2e40e163 MK |
290 | } |
291 | ||
292 | static unsigned long alloc_handle(struct zs_pool *pool) | |
293 | { | |
294 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
295 | pool->flags & ~__GFP_HIGHMEM); | |
296 | } | |
297 | ||
298 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
299 | { | |
300 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
301 | } | |
302 | ||
303 | static void record_obj(unsigned long handle, unsigned long obj) | |
304 | { | |
305 | *(unsigned long *)handle = obj; | |
306 | } | |
307 | ||
c795779d DS |
308 | /* zpool driver */ |
309 | ||
310 | #ifdef CONFIG_ZPOOL | |
311 | ||
479305fd DS |
312 | static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops, |
313 | struct zpool *zpool) | |
c795779d | 314 | { |
3eba0c6a | 315 | return zs_create_pool(name, gfp); |
c795779d DS |
316 | } |
317 | ||
318 | static void zs_zpool_destroy(void *pool) | |
319 | { | |
320 | zs_destroy_pool(pool); | |
321 | } | |
322 | ||
323 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
324 | unsigned long *handle) | |
325 | { | |
326 | *handle = zs_malloc(pool, size); | |
327 | return *handle ? 0 : -1; | |
328 | } | |
329 | static void zs_zpool_free(void *pool, unsigned long handle) | |
330 | { | |
331 | zs_free(pool, handle); | |
332 | } | |
333 | ||
334 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
335 | unsigned int *reclaimed) | |
336 | { | |
337 | return -EINVAL; | |
338 | } | |
339 | ||
340 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
341 | enum zpool_mapmode mm) | |
342 | { | |
343 | enum zs_mapmode zs_mm; | |
344 | ||
345 | switch (mm) { | |
346 | case ZPOOL_MM_RO: | |
347 | zs_mm = ZS_MM_RO; | |
348 | break; | |
349 | case ZPOOL_MM_WO: | |
350 | zs_mm = ZS_MM_WO; | |
351 | break; | |
352 | case ZPOOL_MM_RW: /* fallthru */ | |
353 | default: | |
354 | zs_mm = ZS_MM_RW; | |
355 | break; | |
356 | } | |
357 | ||
358 | return zs_map_object(pool, handle, zs_mm); | |
359 | } | |
360 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
361 | { | |
362 | zs_unmap_object(pool, handle); | |
363 | } | |
364 | ||
365 | static u64 zs_zpool_total_size(void *pool) | |
366 | { | |
722cdc17 | 367 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
368 | } |
369 | ||
370 | static struct zpool_driver zs_zpool_driver = { | |
371 | .type = "zsmalloc", | |
372 | .owner = THIS_MODULE, | |
373 | .create = zs_zpool_create, | |
374 | .destroy = zs_zpool_destroy, | |
375 | .malloc = zs_zpool_malloc, | |
376 | .free = zs_zpool_free, | |
377 | .shrink = zs_zpool_shrink, | |
378 | .map = zs_zpool_map, | |
379 | .unmap = zs_zpool_unmap, | |
380 | .total_size = zs_zpool_total_size, | |
381 | }; | |
382 | ||
137f8cff | 383 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
384 | #endif /* CONFIG_ZPOOL */ |
385 | ||
248ca1b0 MK |
386 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
387 | { | |
388 | return pages_per_zspage * PAGE_SIZE / size; | |
389 | } | |
390 | ||
61989a80 NG |
391 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
392 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
393 | ||
394 | static int is_first_page(struct page *page) | |
395 | { | |
a27545bf | 396 | return PagePrivate(page); |
61989a80 NG |
397 | } |
398 | ||
399 | static int is_last_page(struct page *page) | |
400 | { | |
a27545bf | 401 | return PagePrivate2(page); |
61989a80 NG |
402 | } |
403 | ||
404 | static void get_zspage_mapping(struct page *page, unsigned int *class_idx, | |
405 | enum fullness_group *fullness) | |
406 | { | |
407 | unsigned long m; | |
408 | BUG_ON(!is_first_page(page)); | |
409 | ||
410 | m = (unsigned long)page->mapping; | |
411 | *fullness = m & FULLNESS_MASK; | |
412 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
413 | } | |
414 | ||
415 | static void set_zspage_mapping(struct page *page, unsigned int class_idx, | |
416 | enum fullness_group fullness) | |
417 | { | |
418 | unsigned long m; | |
419 | BUG_ON(!is_first_page(page)); | |
420 | ||
421 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
422 | (fullness & FULLNESS_MASK); | |
423 | page->mapping = (struct address_space *)m; | |
424 | } | |
425 | ||
c3e3e88a NC |
426 | /* |
427 | * zsmalloc divides the pool into various size classes where each | |
428 | * class maintains a list of zspages where each zspage is divided | |
429 | * into equal sized chunks. Each allocation falls into one of these | |
430 | * classes depending on its size. This function returns index of the | |
431 | * size class which has chunk size big enough to hold the give size. | |
432 | */ | |
61989a80 NG |
433 | static int get_size_class_index(int size) |
434 | { | |
435 | int idx = 0; | |
436 | ||
437 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
438 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
439 | ZS_SIZE_CLASS_DELTA); | |
440 | ||
7b60a685 | 441 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
442 | } |
443 | ||
248ca1b0 MK |
444 | #ifdef CONFIG_ZSMALLOC_STAT |
445 | ||
446 | static inline void zs_stat_inc(struct size_class *class, | |
447 | enum zs_stat_type type, unsigned long cnt) | |
448 | { | |
449 | class->stats.objs[type] += cnt; | |
450 | } | |
451 | ||
452 | static inline void zs_stat_dec(struct size_class *class, | |
453 | enum zs_stat_type type, unsigned long cnt) | |
454 | { | |
455 | class->stats.objs[type] -= cnt; | |
456 | } | |
457 | ||
458 | static inline unsigned long zs_stat_get(struct size_class *class, | |
459 | enum zs_stat_type type) | |
460 | { | |
461 | return class->stats.objs[type]; | |
462 | } | |
463 | ||
464 | static int __init zs_stat_init(void) | |
465 | { | |
466 | if (!debugfs_initialized()) | |
467 | return -ENODEV; | |
468 | ||
469 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
470 | if (!zs_stat_root) | |
471 | return -ENOMEM; | |
472 | ||
473 | return 0; | |
474 | } | |
475 | ||
476 | static void __exit zs_stat_exit(void) | |
477 | { | |
478 | debugfs_remove_recursive(zs_stat_root); | |
479 | } | |
480 | ||
481 | static int zs_stats_size_show(struct seq_file *s, void *v) | |
482 | { | |
483 | int i; | |
484 | struct zs_pool *pool = s->private; | |
485 | struct size_class *class; | |
486 | int objs_per_zspage; | |
487 | unsigned long class_almost_full, class_almost_empty; | |
488 | unsigned long obj_allocated, obj_used, pages_used; | |
489 | unsigned long total_class_almost_full = 0, total_class_almost_empty = 0; | |
490 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
491 | ||
492 | seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n", | |
493 | "class", "size", "almost_full", "almost_empty", | |
494 | "obj_allocated", "obj_used", "pages_used", | |
495 | "pages_per_zspage"); | |
496 | ||
497 | for (i = 0; i < zs_size_classes; i++) { | |
498 | class = pool->size_class[i]; | |
499 | ||
500 | if (class->index != i) | |
501 | continue; | |
502 | ||
503 | spin_lock(&class->lock); | |
504 | class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL); | |
505 | class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY); | |
506 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
507 | obj_used = zs_stat_get(class, OBJ_USED); | |
508 | spin_unlock(&class->lock); | |
509 | ||
510 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
511 | class->pages_per_zspage); | |
512 | pages_used = obj_allocated / objs_per_zspage * | |
513 | class->pages_per_zspage; | |
514 | ||
515 | seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n", | |
516 | i, class->size, class_almost_full, class_almost_empty, | |
517 | obj_allocated, obj_used, pages_used, | |
518 | class->pages_per_zspage); | |
519 | ||
520 | total_class_almost_full += class_almost_full; | |
521 | total_class_almost_empty += class_almost_empty; | |
522 | total_objs += obj_allocated; | |
523 | total_used_objs += obj_used; | |
524 | total_pages += pages_used; | |
525 | } | |
526 | ||
527 | seq_puts(s, "\n"); | |
528 | seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n", | |
529 | "Total", "", total_class_almost_full, | |
530 | total_class_almost_empty, total_objs, | |
531 | total_used_objs, total_pages); | |
532 | ||
533 | return 0; | |
534 | } | |
535 | ||
536 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
537 | { | |
538 | return single_open(file, zs_stats_size_show, inode->i_private); | |
539 | } | |
540 | ||
541 | static const struct file_operations zs_stat_size_ops = { | |
542 | .open = zs_stats_size_open, | |
543 | .read = seq_read, | |
544 | .llseek = seq_lseek, | |
545 | .release = single_release, | |
546 | }; | |
547 | ||
548 | static int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
549 | { | |
550 | struct dentry *entry; | |
551 | ||
552 | if (!zs_stat_root) | |
553 | return -ENODEV; | |
554 | ||
555 | entry = debugfs_create_dir(name, zs_stat_root); | |
556 | if (!entry) { | |
557 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
558 | return -ENOMEM; | |
559 | } | |
560 | pool->stat_dentry = entry; | |
561 | ||
562 | entry = debugfs_create_file("classes", S_IFREG | S_IRUGO, | |
563 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
564 | if (!entry) { | |
565 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
566 | name, "classes"); | |
567 | return -ENOMEM; | |
568 | } | |
569 | ||
570 | return 0; | |
571 | } | |
572 | ||
573 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
574 | { | |
575 | debugfs_remove_recursive(pool->stat_dentry); | |
576 | } | |
577 | ||
578 | #else /* CONFIG_ZSMALLOC_STAT */ | |
579 | ||
580 | static inline void zs_stat_inc(struct size_class *class, | |
581 | enum zs_stat_type type, unsigned long cnt) | |
582 | { | |
583 | } | |
584 | ||
585 | static inline void zs_stat_dec(struct size_class *class, | |
586 | enum zs_stat_type type, unsigned long cnt) | |
587 | { | |
588 | } | |
589 | ||
590 | static inline unsigned long zs_stat_get(struct size_class *class, | |
591 | enum zs_stat_type type) | |
592 | { | |
593 | return 0; | |
594 | } | |
595 | ||
596 | static int __init zs_stat_init(void) | |
597 | { | |
598 | return 0; | |
599 | } | |
600 | ||
601 | static void __exit zs_stat_exit(void) | |
602 | { | |
603 | } | |
604 | ||
605 | static inline int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
606 | { | |
607 | return 0; | |
608 | } | |
609 | ||
610 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
611 | { | |
612 | } | |
613 | ||
614 | #endif | |
615 | ||
616 | ||
c3e3e88a NC |
617 | /* |
618 | * For each size class, zspages are divided into different groups | |
619 | * depending on how "full" they are. This was done so that we could | |
620 | * easily find empty or nearly empty zspages when we try to shrink | |
621 | * the pool (not yet implemented). This function returns fullness | |
622 | * status of the given page. | |
623 | */ | |
61989a80 NG |
624 | static enum fullness_group get_fullness_group(struct page *page) |
625 | { | |
626 | int inuse, max_objects; | |
627 | enum fullness_group fg; | |
628 | BUG_ON(!is_first_page(page)); | |
629 | ||
630 | inuse = page->inuse; | |
631 | max_objects = page->objects; | |
632 | ||
633 | if (inuse == 0) | |
634 | fg = ZS_EMPTY; | |
635 | else if (inuse == max_objects) | |
636 | fg = ZS_FULL; | |
d3d07c92 | 637 | else if (inuse <= 3 * max_objects / fullness_threshold_frac) |
61989a80 NG |
638 | fg = ZS_ALMOST_EMPTY; |
639 | else | |
640 | fg = ZS_ALMOST_FULL; | |
641 | ||
642 | return fg; | |
643 | } | |
644 | ||
c3e3e88a NC |
645 | /* |
646 | * Each size class maintains various freelists and zspages are assigned | |
647 | * to one of these freelists based on the number of live objects they | |
648 | * have. This functions inserts the given zspage into the freelist | |
649 | * identified by <class, fullness_group>. | |
650 | */ | |
61989a80 NG |
651 | static void insert_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 | if (*head) | |
663 | list_add_tail(&page->lru, &(*head)->lru); | |
664 | ||
665 | *head = page; | |
248ca1b0 MK |
666 | zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ? |
667 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
668 | } |
669 | ||
c3e3e88a NC |
670 | /* |
671 | * This function removes the given zspage from the freelist identified | |
672 | * by <class, fullness_group>. | |
673 | */ | |
61989a80 NG |
674 | static void remove_zspage(struct page *page, struct size_class *class, |
675 | enum fullness_group fullness) | |
676 | { | |
677 | struct page **head; | |
678 | ||
679 | BUG_ON(!is_first_page(page)); | |
680 | ||
681 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
682 | return; | |
683 | ||
684 | head = &class->fullness_list[fullness]; | |
685 | BUG_ON(!*head); | |
686 | if (list_empty(&(*head)->lru)) | |
687 | *head = NULL; | |
688 | else if (*head == page) | |
689 | *head = (struct page *)list_entry((*head)->lru.next, | |
690 | struct page, lru); | |
691 | ||
692 | list_del_init(&page->lru); | |
248ca1b0 MK |
693 | zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ? |
694 | CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1); | |
61989a80 NG |
695 | } |
696 | ||
c3e3e88a NC |
697 | /* |
698 | * Each size class maintains zspages in different fullness groups depending | |
699 | * on the number of live objects they contain. When allocating or freeing | |
700 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
701 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
702 | * a status change has occurred for the given page and accordingly moves the | |
703 | * page from the freelist of the old fullness group to that of the new | |
704 | * fullness group. | |
705 | */ | |
c7806261 | 706 | static enum fullness_group fix_fullness_group(struct size_class *class, |
61989a80 NG |
707 | struct page *page) |
708 | { | |
709 | int class_idx; | |
61989a80 NG |
710 | enum fullness_group currfg, newfg; |
711 | ||
712 | BUG_ON(!is_first_page(page)); | |
713 | ||
714 | get_zspage_mapping(page, &class_idx, &currfg); | |
715 | newfg = get_fullness_group(page); | |
716 | if (newfg == currfg) | |
717 | goto out; | |
718 | ||
61989a80 NG |
719 | remove_zspage(page, class, currfg); |
720 | insert_zspage(page, class, newfg); | |
721 | set_zspage_mapping(page, class_idx, newfg); | |
722 | ||
723 | out: | |
724 | return newfg; | |
725 | } | |
726 | ||
727 | /* | |
728 | * We have to decide on how many pages to link together | |
729 | * to form a zspage for each size class. This is important | |
730 | * to reduce wastage due to unusable space left at end of | |
731 | * each zspage which is given as: | |
888fa374 YX |
732 | * wastage = Zp % class_size |
733 | * usage = Zp - wastage | |
61989a80 NG |
734 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
735 | * | |
736 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
737 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
738 | * since then we can perfectly fit in 8 such objects. | |
739 | */ | |
2e3b6154 | 740 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
741 | { |
742 | int i, max_usedpc = 0; | |
743 | /* zspage order which gives maximum used size per KB */ | |
744 | int max_usedpc_order = 1; | |
745 | ||
84d4faab | 746 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
747 | int zspage_size; |
748 | int waste, usedpc; | |
749 | ||
750 | zspage_size = i * PAGE_SIZE; | |
751 | waste = zspage_size % class_size; | |
752 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
753 | ||
754 | if (usedpc > max_usedpc) { | |
755 | max_usedpc = usedpc; | |
756 | max_usedpc_order = i; | |
757 | } | |
758 | } | |
759 | ||
760 | return max_usedpc_order; | |
761 | } | |
762 | ||
763 | /* | |
764 | * A single 'zspage' is composed of many system pages which are | |
765 | * linked together using fields in struct page. This function finds | |
766 | * the first/head page, given any component page of a zspage. | |
767 | */ | |
768 | static struct page *get_first_page(struct page *page) | |
769 | { | |
770 | if (is_first_page(page)) | |
771 | return page; | |
772 | else | |
773 | return page->first_page; | |
774 | } | |
775 | ||
776 | static struct page *get_next_page(struct page *page) | |
777 | { | |
778 | struct page *next; | |
779 | ||
780 | if (is_last_page(page)) | |
781 | next = NULL; | |
782 | else if (is_first_page(page)) | |
e842b976 | 783 | next = (struct page *)page_private(page); |
61989a80 NG |
784 | else |
785 | next = list_entry(page->lru.next, struct page, lru); | |
786 | ||
787 | return next; | |
788 | } | |
789 | ||
67296874 OH |
790 | /* |
791 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 792 | * We use the least bit of handle for tagging. |
67296874 | 793 | */ |
312fcae2 | 794 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 795 | { |
312fcae2 | 796 | unsigned long obj; |
61989a80 NG |
797 | |
798 | if (!page) { | |
799 | BUG_ON(obj_idx); | |
800 | return NULL; | |
801 | } | |
802 | ||
312fcae2 MK |
803 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
804 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
805 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 806 | |
312fcae2 | 807 | return (void *)obj; |
61989a80 NG |
808 | } |
809 | ||
67296874 OH |
810 | /* |
811 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
812 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 813 | * location_to_obj(). |
67296874 | 814 | */ |
312fcae2 | 815 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
816 | unsigned long *obj_idx) |
817 | { | |
312fcae2 MK |
818 | obj >>= OBJ_TAG_BITS; |
819 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
820 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
821 | } |
822 | ||
2e40e163 MK |
823 | static unsigned long handle_to_obj(unsigned long handle) |
824 | { | |
825 | return *(unsigned long *)handle; | |
826 | } | |
827 | ||
7b60a685 MK |
828 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
829 | void *obj) | |
312fcae2 | 830 | { |
7b60a685 MK |
831 | if (class->huge) { |
832 | VM_BUG_ON(!is_first_page(page)); | |
833 | return *(unsigned long *)page_private(page); | |
834 | } else | |
835 | return *(unsigned long *)obj; | |
312fcae2 MK |
836 | } |
837 | ||
61989a80 NG |
838 | static unsigned long obj_idx_to_offset(struct page *page, |
839 | unsigned long obj_idx, int class_size) | |
840 | { | |
841 | unsigned long off = 0; | |
842 | ||
843 | if (!is_first_page(page)) | |
844 | off = page->index; | |
845 | ||
846 | return off + obj_idx * class_size; | |
847 | } | |
848 | ||
312fcae2 MK |
849 | static inline int trypin_tag(unsigned long handle) |
850 | { | |
851 | unsigned long *ptr = (unsigned long *)handle; | |
852 | ||
853 | return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr); | |
854 | } | |
855 | ||
856 | static void pin_tag(unsigned long handle) | |
857 | { | |
858 | while (!trypin_tag(handle)); | |
859 | } | |
860 | ||
861 | static void unpin_tag(unsigned long handle) | |
862 | { | |
863 | unsigned long *ptr = (unsigned long *)handle; | |
864 | ||
865 | clear_bit_unlock(HANDLE_PIN_BIT, ptr); | |
866 | } | |
867 | ||
f4477e90 NG |
868 | static void reset_page(struct page *page) |
869 | { | |
870 | clear_bit(PG_private, &page->flags); | |
871 | clear_bit(PG_private_2, &page->flags); | |
872 | set_page_private(page, 0); | |
873 | page->mapping = NULL; | |
874 | page->freelist = NULL; | |
22b751c3 | 875 | page_mapcount_reset(page); |
f4477e90 NG |
876 | } |
877 | ||
61989a80 NG |
878 | static void free_zspage(struct page *first_page) |
879 | { | |
f4477e90 | 880 | struct page *nextp, *tmp, *head_extra; |
61989a80 NG |
881 | |
882 | BUG_ON(!is_first_page(first_page)); | |
883 | BUG_ON(first_page->inuse); | |
884 | ||
f4477e90 | 885 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 886 | |
f4477e90 | 887 | reset_page(first_page); |
61989a80 NG |
888 | __free_page(first_page); |
889 | ||
890 | /* zspage with only 1 system page */ | |
f4477e90 | 891 | if (!head_extra) |
61989a80 NG |
892 | return; |
893 | ||
f4477e90 | 894 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 895 | list_del(&nextp->lru); |
f4477e90 | 896 | reset_page(nextp); |
61989a80 NG |
897 | __free_page(nextp); |
898 | } | |
f4477e90 NG |
899 | reset_page(head_extra); |
900 | __free_page(head_extra); | |
61989a80 NG |
901 | } |
902 | ||
903 | /* Initialize a newly allocated zspage */ | |
904 | static void init_zspage(struct page *first_page, struct size_class *class) | |
905 | { | |
906 | unsigned long off = 0; | |
907 | struct page *page = first_page; | |
908 | ||
909 | BUG_ON(!is_first_page(first_page)); | |
910 | while (page) { | |
911 | struct page *next_page; | |
912 | struct link_free *link; | |
5538c562 | 913 | unsigned int i = 1; |
af4ee5e9 | 914 | void *vaddr; |
61989a80 NG |
915 | |
916 | /* | |
917 | * page->index stores offset of first object starting | |
918 | * in the page. For the first page, this is always 0, | |
919 | * so we use first_page->index (aka ->freelist) to store | |
920 | * head of corresponding zspage's freelist. | |
921 | */ | |
922 | if (page != first_page) | |
923 | page->index = off; | |
924 | ||
af4ee5e9 MK |
925 | vaddr = kmap_atomic(page); |
926 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
927 | |
928 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 929 | link->next = location_to_obj(page, i++); |
5538c562 | 930 | link += class->size / sizeof(*link); |
61989a80 NG |
931 | } |
932 | ||
933 | /* | |
934 | * We now come to the last (full or partial) object on this | |
935 | * page, which must point to the first object on the next | |
936 | * page (if present) | |
937 | */ | |
938 | next_page = get_next_page(page); | |
312fcae2 | 939 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 940 | kunmap_atomic(vaddr); |
61989a80 | 941 | page = next_page; |
5538c562 | 942 | off %= PAGE_SIZE; |
61989a80 NG |
943 | } |
944 | } | |
945 | ||
946 | /* | |
947 | * Allocate a zspage for the given size class | |
948 | */ | |
949 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
950 | { | |
951 | int i, error; | |
b4b700c5 | 952 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
953 | |
954 | /* | |
955 | * Allocate individual pages and link them together as: | |
956 | * 1. first page->private = first sub-page | |
957 | * 2. all sub-pages are linked together using page->lru | |
958 | * 3. each sub-page is linked to the first page using page->first_page | |
959 | * | |
960 | * For each size class, First/Head pages are linked together using | |
961 | * page->lru. Also, we set PG_private to identify the first page | |
962 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
963 | * identify the last page. | |
964 | */ | |
965 | error = -ENOMEM; | |
2e3b6154 | 966 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 967 | struct page *page; |
61989a80 NG |
968 | |
969 | page = alloc_page(flags); | |
970 | if (!page) | |
971 | goto cleanup; | |
972 | ||
973 | INIT_LIST_HEAD(&page->lru); | |
974 | if (i == 0) { /* first page */ | |
a27545bf | 975 | SetPagePrivate(page); |
61989a80 NG |
976 | set_page_private(page, 0); |
977 | first_page = page; | |
978 | first_page->inuse = 0; | |
979 | } | |
980 | if (i == 1) | |
e842b976 | 981 | set_page_private(first_page, (unsigned long)page); |
61989a80 NG |
982 | if (i >= 1) |
983 | page->first_page = first_page; | |
984 | if (i >= 2) | |
985 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 986 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 987 | SetPagePrivate2(page); |
61989a80 NG |
988 | prev_page = page; |
989 | } | |
990 | ||
991 | init_zspage(first_page, class); | |
992 | ||
312fcae2 | 993 | first_page->freelist = location_to_obj(first_page, 0); |
61989a80 | 994 | /* Maximum number of objects we can store in this zspage */ |
2e3b6154 | 995 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
996 | |
997 | error = 0; /* Success */ | |
998 | ||
999 | cleanup: | |
1000 | if (unlikely(error) && first_page) { | |
1001 | free_zspage(first_page); | |
1002 | first_page = NULL; | |
1003 | } | |
1004 | ||
1005 | return first_page; | |
1006 | } | |
1007 | ||
1008 | static struct page *find_get_zspage(struct size_class *class) | |
1009 | { | |
1010 | int i; | |
1011 | struct page *page; | |
1012 | ||
1013 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1014 | page = class->fullness_list[i]; | |
1015 | if (page) | |
1016 | break; | |
1017 | } | |
1018 | ||
1019 | return page; | |
1020 | } | |
1021 | ||
1b945aee | 1022 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
1023 | static inline int __zs_cpu_up(struct mapping_area *area) |
1024 | { | |
1025 | /* | |
1026 | * Make sure we don't leak memory if a cpu UP notification | |
1027 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1028 | */ | |
1029 | if (area->vm) | |
1030 | return 0; | |
1031 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
1032 | if (!area->vm) | |
1033 | return -ENOMEM; | |
1034 | return 0; | |
1035 | } | |
1036 | ||
1037 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1038 | { | |
1039 | if (area->vm) | |
1040 | free_vm_area(area->vm); | |
1041 | area->vm = NULL; | |
1042 | } | |
1043 | ||
1044 | static inline void *__zs_map_object(struct mapping_area *area, | |
1045 | struct page *pages[2], int off, int size) | |
1046 | { | |
f6f8ed47 | 1047 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
1048 | area->vm_addr = area->vm->addr; |
1049 | return area->vm_addr + off; | |
1050 | } | |
1051 | ||
1052 | static inline void __zs_unmap_object(struct mapping_area *area, | |
1053 | struct page *pages[2], int off, int size) | |
1054 | { | |
1055 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 1056 | |
d95abbbb | 1057 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
1058 | } |
1059 | ||
1b945aee | 1060 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
1061 | |
1062 | static inline int __zs_cpu_up(struct mapping_area *area) | |
1063 | { | |
1064 | /* | |
1065 | * Make sure we don't leak memory if a cpu UP notification | |
1066 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
1067 | */ | |
1068 | if (area->vm_buf) | |
1069 | return 0; | |
40f9fb8c | 1070 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
1071 | if (!area->vm_buf) |
1072 | return -ENOMEM; | |
1073 | return 0; | |
1074 | } | |
1075 | ||
1076 | static inline void __zs_cpu_down(struct mapping_area *area) | |
1077 | { | |
40f9fb8c | 1078 | kfree(area->vm_buf); |
f553646a SJ |
1079 | area->vm_buf = NULL; |
1080 | } | |
1081 | ||
1082 | static void *__zs_map_object(struct mapping_area *area, | |
1083 | struct page *pages[2], int off, int size) | |
5f601902 | 1084 | { |
5f601902 SJ |
1085 | int sizes[2]; |
1086 | void *addr; | |
f553646a | 1087 | char *buf = area->vm_buf; |
5f601902 | 1088 | |
f553646a SJ |
1089 | /* disable page faults to match kmap_atomic() return conditions */ |
1090 | pagefault_disable(); | |
1091 | ||
1092 | /* no read fastpath */ | |
1093 | if (area->vm_mm == ZS_MM_WO) | |
1094 | goto out; | |
5f601902 SJ |
1095 | |
1096 | sizes[0] = PAGE_SIZE - off; | |
1097 | sizes[1] = size - sizes[0]; | |
1098 | ||
5f601902 SJ |
1099 | /* copy object to per-cpu buffer */ |
1100 | addr = kmap_atomic(pages[0]); | |
1101 | memcpy(buf, addr + off, sizes[0]); | |
1102 | kunmap_atomic(addr); | |
1103 | addr = kmap_atomic(pages[1]); | |
1104 | memcpy(buf + sizes[0], addr, sizes[1]); | |
1105 | kunmap_atomic(addr); | |
f553646a SJ |
1106 | out: |
1107 | return area->vm_buf; | |
5f601902 SJ |
1108 | } |
1109 | ||
f553646a SJ |
1110 | static void __zs_unmap_object(struct mapping_area *area, |
1111 | struct page *pages[2], int off, int size) | |
5f601902 | 1112 | { |
5f601902 SJ |
1113 | int sizes[2]; |
1114 | void *addr; | |
2e40e163 | 1115 | char *buf; |
5f601902 | 1116 | |
f553646a SJ |
1117 | /* no write fastpath */ |
1118 | if (area->vm_mm == ZS_MM_RO) | |
1119 | goto out; | |
5f601902 | 1120 | |
7b60a685 MK |
1121 | buf = area->vm_buf; |
1122 | if (!area->huge) { | |
1123 | buf = buf + ZS_HANDLE_SIZE; | |
1124 | size -= ZS_HANDLE_SIZE; | |
1125 | off += ZS_HANDLE_SIZE; | |
1126 | } | |
2e40e163 | 1127 | |
5f601902 SJ |
1128 | sizes[0] = PAGE_SIZE - off; |
1129 | sizes[1] = size - sizes[0]; | |
1130 | ||
1131 | /* copy per-cpu buffer to object */ | |
1132 | addr = kmap_atomic(pages[0]); | |
1133 | memcpy(addr + off, buf, sizes[0]); | |
1134 | kunmap_atomic(addr); | |
1135 | addr = kmap_atomic(pages[1]); | |
1136 | memcpy(addr, buf + sizes[0], sizes[1]); | |
1137 | kunmap_atomic(addr); | |
f553646a SJ |
1138 | |
1139 | out: | |
1140 | /* enable page faults to match kunmap_atomic() return conditions */ | |
1141 | pagefault_enable(); | |
5f601902 | 1142 | } |
61989a80 | 1143 | |
1b945aee | 1144 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 1145 | |
61989a80 NG |
1146 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
1147 | void *pcpu) | |
1148 | { | |
f553646a | 1149 | int ret, cpu = (long)pcpu; |
61989a80 NG |
1150 | struct mapping_area *area; |
1151 | ||
1152 | switch (action) { | |
1153 | case CPU_UP_PREPARE: | |
1154 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1155 | ret = __zs_cpu_up(area); |
1156 | if (ret) | |
1157 | return notifier_from_errno(ret); | |
61989a80 NG |
1158 | break; |
1159 | case CPU_DEAD: | |
1160 | case CPU_UP_CANCELED: | |
1161 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1162 | __zs_cpu_down(area); |
61989a80 NG |
1163 | break; |
1164 | } | |
1165 | ||
1166 | return NOTIFY_OK; | |
1167 | } | |
1168 | ||
1169 | static struct notifier_block zs_cpu_nb = { | |
1170 | .notifier_call = zs_cpu_notifier | |
1171 | }; | |
1172 | ||
b1b00a5b | 1173 | static int zs_register_cpu_notifier(void) |
61989a80 | 1174 | { |
b1b00a5b | 1175 | int cpu, uninitialized_var(ret); |
61989a80 | 1176 | |
f0e71fcd SB |
1177 | cpu_notifier_register_begin(); |
1178 | ||
1179 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1180 | for_each_online_cpu(cpu) { |
1181 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1182 | if (notifier_to_errno(ret)) |
1183 | break; | |
61989a80 | 1184 | } |
f0e71fcd SB |
1185 | |
1186 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1187 | return notifier_to_errno(ret); |
1188 | } | |
f0e71fcd | 1189 | |
66cdef66 | 1190 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1191 | { |
66cdef66 | 1192 | int cpu; |
40f9fb8c | 1193 | |
66cdef66 | 1194 | cpu_notifier_register_begin(); |
40f9fb8c | 1195 | |
66cdef66 GM |
1196 | for_each_online_cpu(cpu) |
1197 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1198 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1199 | |
66cdef66 | 1200 | cpu_notifier_register_done(); |
b1b00a5b SS |
1201 | } |
1202 | ||
66cdef66 | 1203 | static void init_zs_size_classes(void) |
b1b00a5b | 1204 | { |
66cdef66 | 1205 | int nr; |
c795779d | 1206 | |
66cdef66 GM |
1207 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1208 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1209 | nr += 1; | |
40f9fb8c | 1210 | |
66cdef66 | 1211 | zs_size_classes = nr; |
61989a80 NG |
1212 | } |
1213 | ||
9eec4cd5 JK |
1214 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) |
1215 | { | |
1216 | if (prev->pages_per_zspage != pages_per_zspage) | |
1217 | return false; | |
1218 | ||
1219 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1220 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1221 | return false; | |
1222 | ||
1223 | return true; | |
1224 | } | |
1225 | ||
312fcae2 MK |
1226 | static bool zspage_full(struct page *page) |
1227 | { | |
1228 | BUG_ON(!is_first_page(page)); | |
1229 | ||
1230 | return page->inuse == page->objects; | |
1231 | } | |
1232 | ||
66cdef66 GM |
1233 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1234 | { | |
1235 | return atomic_long_read(&pool->pages_allocated); | |
1236 | } | |
1237 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1238 | ||
4bbc0bc0 | 1239 | /** |
66cdef66 GM |
1240 | * zs_map_object - get address of allocated object from handle. |
1241 | * @pool: pool from which the object was allocated | |
1242 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1243 | * |
66cdef66 GM |
1244 | * Before using an object allocated from zs_malloc, it must be mapped using |
1245 | * this function. When done with the object, it must be unmapped using | |
1246 | * zs_unmap_object. | |
4bbc0bc0 | 1247 | * |
66cdef66 GM |
1248 | * Only one object can be mapped per cpu at a time. There is no protection |
1249 | * against nested mappings. | |
1250 | * | |
1251 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1252 | */ |
66cdef66 GM |
1253 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1254 | enum zs_mapmode mm) | |
61989a80 | 1255 | { |
66cdef66 | 1256 | struct page *page; |
2e40e163 | 1257 | unsigned long obj, obj_idx, off; |
61989a80 | 1258 | |
66cdef66 GM |
1259 | unsigned int class_idx; |
1260 | enum fullness_group fg; | |
1261 | struct size_class *class; | |
1262 | struct mapping_area *area; | |
1263 | struct page *pages[2]; | |
2e40e163 | 1264 | void *ret; |
61989a80 | 1265 | |
66cdef66 | 1266 | BUG_ON(!handle); |
40f9fb8c | 1267 | |
9eec4cd5 | 1268 | /* |
66cdef66 GM |
1269 | * Because we use per-cpu mapping areas shared among the |
1270 | * pools/users, we can't allow mapping in interrupt context | |
1271 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1272 | */ |
66cdef66 | 1273 | BUG_ON(in_interrupt()); |
61989a80 | 1274 | |
312fcae2 MK |
1275 | /* From now on, migration cannot move the object */ |
1276 | pin_tag(handle); | |
1277 | ||
2e40e163 MK |
1278 | obj = handle_to_obj(handle); |
1279 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1280 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1281 | class = pool->size_class[class_idx]; | |
1282 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1283 | |
66cdef66 GM |
1284 | area = &get_cpu_var(zs_map_area); |
1285 | area->vm_mm = mm; | |
1286 | if (off + class->size <= PAGE_SIZE) { | |
1287 | /* this object is contained entirely within a page */ | |
1288 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1289 | ret = area->vm_addr + off; |
1290 | goto out; | |
61989a80 NG |
1291 | } |
1292 | ||
66cdef66 GM |
1293 | /* this object spans two pages */ |
1294 | pages[0] = page; | |
1295 | pages[1] = get_next_page(page); | |
1296 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1297 | |
2e40e163 MK |
1298 | ret = __zs_map_object(area, pages, off, class->size); |
1299 | out: | |
7b60a685 MK |
1300 | if (!class->huge) |
1301 | ret += ZS_HANDLE_SIZE; | |
1302 | ||
1303 | return ret; | |
61989a80 | 1304 | } |
66cdef66 | 1305 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1306 | |
66cdef66 | 1307 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1308 | { |
66cdef66 | 1309 | struct page *page; |
2e40e163 | 1310 | unsigned long obj, obj_idx, off; |
61989a80 | 1311 | |
66cdef66 GM |
1312 | unsigned int class_idx; |
1313 | enum fullness_group fg; | |
1314 | struct size_class *class; | |
1315 | struct mapping_area *area; | |
9eec4cd5 | 1316 | |
66cdef66 | 1317 | BUG_ON(!handle); |
9eec4cd5 | 1318 | |
2e40e163 MK |
1319 | obj = handle_to_obj(handle); |
1320 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1321 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1322 | class = pool->size_class[class_idx]; | |
1323 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1324 | |
66cdef66 GM |
1325 | area = this_cpu_ptr(&zs_map_area); |
1326 | if (off + class->size <= PAGE_SIZE) | |
1327 | kunmap_atomic(area->vm_addr); | |
1328 | else { | |
1329 | struct page *pages[2]; | |
40f9fb8c | 1330 | |
66cdef66 GM |
1331 | pages[0] = page; |
1332 | pages[1] = get_next_page(page); | |
1333 | BUG_ON(!pages[1]); | |
1334 | ||
1335 | __zs_unmap_object(area, pages, off, class->size); | |
1336 | } | |
1337 | put_cpu_var(zs_map_area); | |
312fcae2 | 1338 | unpin_tag(handle); |
61989a80 | 1339 | } |
66cdef66 | 1340 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1341 | |
c7806261 MK |
1342 | static unsigned long obj_malloc(struct page *first_page, |
1343 | struct size_class *class, unsigned long handle) | |
1344 | { | |
1345 | unsigned long obj; | |
1346 | struct link_free *link; | |
1347 | ||
1348 | struct page *m_page; | |
1349 | unsigned long m_objidx, m_offset; | |
1350 | void *vaddr; | |
1351 | ||
312fcae2 | 1352 | handle |= OBJ_ALLOCATED_TAG; |
c7806261 MK |
1353 | obj = (unsigned long)first_page->freelist; |
1354 | obj_to_location(obj, &m_page, &m_objidx); | |
1355 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1356 | ||
1357 | vaddr = kmap_atomic(m_page); | |
1358 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1359 | first_page->freelist = link->next; | |
7b60a685 MK |
1360 | if (!class->huge) |
1361 | /* record handle in the header of allocated chunk */ | |
1362 | link->handle = handle; | |
1363 | else | |
1364 | /* record handle in first_page->private */ | |
1365 | set_page_private(first_page, handle); | |
c7806261 MK |
1366 | kunmap_atomic(vaddr); |
1367 | first_page->inuse++; | |
1368 | zs_stat_inc(class, OBJ_USED, 1); | |
1369 | ||
1370 | return obj; | |
1371 | } | |
1372 | ||
1373 | ||
61989a80 NG |
1374 | /** |
1375 | * zs_malloc - Allocate block of given size from pool. | |
1376 | * @pool: pool to allocate from | |
1377 | * @size: size of block to allocate | |
61989a80 | 1378 | * |
00a61d86 | 1379 | * On success, handle to the allocated object is returned, |
c2344348 | 1380 | * otherwise 0. |
61989a80 NG |
1381 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1382 | */ | |
c2344348 | 1383 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
61989a80 | 1384 | { |
2e40e163 | 1385 | unsigned long handle, obj; |
61989a80 | 1386 | struct size_class *class; |
c7806261 | 1387 | struct page *first_page; |
61989a80 | 1388 | |
7b60a685 | 1389 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1390 | return 0; |
1391 | ||
1392 | handle = alloc_handle(pool); | |
1393 | if (!handle) | |
c2344348 | 1394 | return 0; |
61989a80 | 1395 | |
2e40e163 MK |
1396 | /* extra space in chunk to keep the handle */ |
1397 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1398 | class = pool->size_class[get_size_class_index(size)]; |
61989a80 NG |
1399 | |
1400 | spin_lock(&class->lock); | |
1401 | first_page = find_get_zspage(class); | |
1402 | ||
1403 | if (!first_page) { | |
1404 | spin_unlock(&class->lock); | |
1405 | first_page = alloc_zspage(class, pool->flags); | |
2e40e163 MK |
1406 | if (unlikely(!first_page)) { |
1407 | free_handle(pool, handle); | |
c2344348 | 1408 | return 0; |
2e40e163 | 1409 | } |
61989a80 NG |
1410 | |
1411 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1412 | atomic_long_add(class->pages_per_zspage, |
1413 | &pool->pages_allocated); | |
0f050d99 | 1414 | |
61989a80 | 1415 | spin_lock(&class->lock); |
0f050d99 GM |
1416 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1417 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1418 | } |
1419 | ||
c7806261 | 1420 | obj = obj_malloc(first_page, class, handle); |
61989a80 | 1421 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1422 | fix_fullness_group(class, first_page); |
2e40e163 | 1423 | record_obj(handle, obj); |
61989a80 NG |
1424 | spin_unlock(&class->lock); |
1425 | ||
2e40e163 | 1426 | return handle; |
61989a80 NG |
1427 | } |
1428 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1429 | ||
c7806261 MK |
1430 | static void obj_free(struct zs_pool *pool, struct size_class *class, |
1431 | unsigned long obj) | |
61989a80 NG |
1432 | { |
1433 | struct link_free *link; | |
1434 | struct page *first_page, *f_page; | |
c7806261 | 1435 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1436 | void *vaddr; |
61989a80 | 1437 | int class_idx; |
61989a80 NG |
1438 | enum fullness_group fullness; |
1439 | ||
c7806261 | 1440 | BUG_ON(!obj); |
61989a80 | 1441 | |
312fcae2 | 1442 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1443 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1444 | first_page = get_first_page(f_page); |
1445 | ||
1446 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
61989a80 NG |
1447 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1448 | ||
c7806261 | 1449 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1450 | |
1451 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1452 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1453 | link->next = first_page->freelist; |
7b60a685 MK |
1454 | if (class->huge) |
1455 | set_page_private(first_page, 0); | |
af4ee5e9 | 1456 | kunmap_atomic(vaddr); |
c2344348 | 1457 | first_page->freelist = (void *)obj; |
61989a80 | 1458 | first_page->inuse--; |
0f050d99 | 1459 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1460 | } |
1461 | ||
1462 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1463 | { | |
1464 | struct page *first_page, *f_page; | |
1465 | unsigned long obj, f_objidx; | |
1466 | int class_idx; | |
1467 | struct size_class *class; | |
1468 | enum fullness_group fullness; | |
1469 | ||
1470 | if (unlikely(!handle)) | |
1471 | return; | |
1472 | ||
312fcae2 | 1473 | pin_tag(handle); |
c7806261 | 1474 | obj = handle_to_obj(handle); |
c7806261 MK |
1475 | obj_to_location(obj, &f_page, &f_objidx); |
1476 | first_page = get_first_page(f_page); | |
1477 | ||
1478 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1479 | class = pool->size_class[class_idx]; | |
1480 | ||
1481 | spin_lock(&class->lock); | |
1482 | obj_free(pool, class, obj); | |
1483 | fullness = fix_fullness_group(class, first_page); | |
312fcae2 | 1484 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1485 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1486 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1487 | atomic_long_sub(class->pages_per_zspage, |
1488 | &pool->pages_allocated); | |
1489 | free_zspage(first_page); | |
1490 | } | |
61989a80 | 1491 | spin_unlock(&class->lock); |
312fcae2 | 1492 | unpin_tag(handle); |
61989a80 | 1493 | |
312fcae2 MK |
1494 | free_handle(pool, handle); |
1495 | } | |
1496 | EXPORT_SYMBOL_GPL(zs_free); | |
1497 | ||
1498 | static void zs_object_copy(unsigned long src, unsigned long dst, | |
1499 | struct size_class *class) | |
1500 | { | |
1501 | struct page *s_page, *d_page; | |
1502 | unsigned long s_objidx, d_objidx; | |
1503 | unsigned long s_off, d_off; | |
1504 | void *s_addr, *d_addr; | |
1505 | int s_size, d_size, size; | |
1506 | int written = 0; | |
1507 | ||
1508 | s_size = d_size = class->size; | |
1509 | ||
1510 | obj_to_location(src, &s_page, &s_objidx); | |
1511 | obj_to_location(dst, &d_page, &d_objidx); | |
1512 | ||
1513 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1514 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1515 | ||
1516 | if (s_off + class->size > PAGE_SIZE) | |
1517 | s_size = PAGE_SIZE - s_off; | |
1518 | ||
1519 | if (d_off + class->size > PAGE_SIZE) | |
1520 | d_size = PAGE_SIZE - d_off; | |
1521 | ||
1522 | s_addr = kmap_atomic(s_page); | |
1523 | d_addr = kmap_atomic(d_page); | |
1524 | ||
1525 | while (1) { | |
1526 | size = min(s_size, d_size); | |
1527 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1528 | written += size; | |
1529 | ||
1530 | if (written == class->size) | |
1531 | break; | |
1532 | ||
495819ea SS |
1533 | s_off += size; |
1534 | s_size -= size; | |
1535 | d_off += size; | |
1536 | d_size -= size; | |
1537 | ||
1538 | if (s_off >= PAGE_SIZE) { | |
312fcae2 MK |
1539 | kunmap_atomic(d_addr); |
1540 | kunmap_atomic(s_addr); | |
1541 | s_page = get_next_page(s_page); | |
1542 | BUG_ON(!s_page); | |
1543 | s_addr = kmap_atomic(s_page); | |
1544 | d_addr = kmap_atomic(d_page); | |
1545 | s_size = class->size - written; | |
1546 | s_off = 0; | |
312fcae2 MK |
1547 | } |
1548 | ||
495819ea | 1549 | if (d_off >= PAGE_SIZE) { |
312fcae2 MK |
1550 | kunmap_atomic(d_addr); |
1551 | d_page = get_next_page(d_page); | |
1552 | BUG_ON(!d_page); | |
1553 | d_addr = kmap_atomic(d_page); | |
1554 | d_size = class->size - written; | |
1555 | d_off = 0; | |
312fcae2 MK |
1556 | } |
1557 | } | |
1558 | ||
1559 | kunmap_atomic(d_addr); | |
1560 | kunmap_atomic(s_addr); | |
1561 | } | |
1562 | ||
1563 | /* | |
1564 | * Find alloced object in zspage from index object and | |
1565 | * return handle. | |
1566 | */ | |
1567 | static unsigned long find_alloced_obj(struct page *page, int index, | |
1568 | struct size_class *class) | |
1569 | { | |
1570 | unsigned long head; | |
1571 | int offset = 0; | |
1572 | unsigned long handle = 0; | |
1573 | void *addr = kmap_atomic(page); | |
1574 | ||
1575 | if (!is_first_page(page)) | |
1576 | offset = page->index; | |
1577 | offset += class->size * index; | |
1578 | ||
1579 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1580 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1581 | if (head & OBJ_ALLOCATED_TAG) { |
1582 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1583 | if (trypin_tag(handle)) | |
1584 | break; | |
1585 | handle = 0; | |
1586 | } | |
1587 | ||
1588 | offset += class->size; | |
1589 | index++; | |
1590 | } | |
1591 | ||
1592 | kunmap_atomic(addr); | |
1593 | return handle; | |
1594 | } | |
1595 | ||
1596 | struct zs_compact_control { | |
1597 | /* Source page for migration which could be a subpage of zspage. */ | |
1598 | struct page *s_page; | |
1599 | /* Destination page for migration which should be a first page | |
1600 | * of zspage. */ | |
1601 | struct page *d_page; | |
1602 | /* Starting object index within @s_page which used for live object | |
1603 | * in the subpage. */ | |
1604 | int index; | |
1605 | /* how many of objects are migrated */ | |
1606 | int nr_migrated; | |
1607 | }; | |
1608 | ||
1609 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1610 | struct zs_compact_control *cc) | |
1611 | { | |
1612 | unsigned long used_obj, free_obj; | |
1613 | unsigned long handle; | |
1614 | struct page *s_page = cc->s_page; | |
1615 | struct page *d_page = cc->d_page; | |
1616 | unsigned long index = cc->index; | |
1617 | int nr_migrated = 0; | |
1618 | int ret = 0; | |
1619 | ||
1620 | while (1) { | |
1621 | handle = find_alloced_obj(s_page, index, class); | |
1622 | if (!handle) { | |
1623 | s_page = get_next_page(s_page); | |
1624 | if (!s_page) | |
1625 | break; | |
1626 | index = 0; | |
1627 | continue; | |
1628 | } | |
1629 | ||
1630 | /* Stop if there is no more space */ | |
1631 | if (zspage_full(d_page)) { | |
1632 | unpin_tag(handle); | |
1633 | ret = -ENOMEM; | |
1634 | break; | |
1635 | } | |
1636 | ||
1637 | used_obj = handle_to_obj(handle); | |
1638 | free_obj = obj_malloc(d_page, class, handle); | |
1639 | zs_object_copy(used_obj, free_obj, class); | |
1640 | index++; | |
1641 | record_obj(handle, free_obj); | |
1642 | unpin_tag(handle); | |
1643 | obj_free(pool, class, used_obj); | |
1644 | nr_migrated++; | |
1645 | } | |
1646 | ||
1647 | /* Remember last position in this iteration */ | |
1648 | cc->s_page = s_page; | |
1649 | cc->index = index; | |
1650 | cc->nr_migrated = nr_migrated; | |
1651 | ||
1652 | return ret; | |
1653 | } | |
1654 | ||
1655 | static struct page *alloc_target_page(struct size_class *class) | |
1656 | { | |
1657 | int i; | |
1658 | struct page *page; | |
1659 | ||
1660 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1661 | page = class->fullness_list[i]; | |
1662 | if (page) { | |
1663 | remove_zspage(page, class, i); | |
1664 | break; | |
1665 | } | |
1666 | } | |
1667 | ||
1668 | return page; | |
1669 | } | |
1670 | ||
1671 | static void putback_zspage(struct zs_pool *pool, struct size_class *class, | |
1672 | struct page *first_page) | |
1673 | { | |
312fcae2 MK |
1674 | enum fullness_group fullness; |
1675 | ||
1676 | BUG_ON(!is_first_page(first_page)); | |
1677 | ||
839373e6 | 1678 | fullness = get_fullness_group(first_page); |
312fcae2 | 1679 | insert_zspage(first_page, class, fullness); |
839373e6 MK |
1680 | set_zspage_mapping(first_page, class->index, fullness); |
1681 | ||
13de8933 | 1682 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1683 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1684 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1685 | atomic_long_sub(class->pages_per_zspage, |
1686 | &pool->pages_allocated); | |
312fcae2 | 1687 | |
61989a80 | 1688 | free_zspage(first_page); |
13de8933 | 1689 | } |
61989a80 | 1690 | } |
312fcae2 MK |
1691 | |
1692 | static struct page *isolate_source_page(struct size_class *class) | |
1693 | { | |
1694 | struct page *page; | |
1695 | ||
1696 | page = class->fullness_list[ZS_ALMOST_EMPTY]; | |
1697 | if (page) | |
1698 | remove_zspage(page, class, ZS_ALMOST_EMPTY); | |
1699 | ||
1700 | return page; | |
1701 | } | |
1702 | ||
1703 | static unsigned long __zs_compact(struct zs_pool *pool, | |
1704 | struct size_class *class) | |
1705 | { | |
1706 | int nr_to_migrate; | |
1707 | struct zs_compact_control cc; | |
1708 | struct page *src_page; | |
1709 | struct page *dst_page = NULL; | |
1710 | unsigned long nr_total_migrated = 0; | |
1711 | ||
312fcae2 MK |
1712 | spin_lock(&class->lock); |
1713 | while ((src_page = isolate_source_page(class))) { | |
1714 | ||
1715 | BUG_ON(!is_first_page(src_page)); | |
1716 | ||
1717 | /* The goal is to migrate all live objects in source page */ | |
1718 | nr_to_migrate = src_page->inuse; | |
1719 | cc.index = 0; | |
1720 | cc.s_page = src_page; | |
1721 | ||
1722 | while ((dst_page = alloc_target_page(class))) { | |
1723 | cc.d_page = dst_page; | |
1724 | /* | |
1725 | * If there is no more space in dst_page, try to | |
1726 | * allocate another zspage. | |
1727 | */ | |
1728 | if (!migrate_zspage(pool, class, &cc)) | |
1729 | break; | |
1730 | ||
1731 | putback_zspage(pool, class, dst_page); | |
1732 | nr_total_migrated += cc.nr_migrated; | |
1733 | nr_to_migrate -= cc.nr_migrated; | |
1734 | } | |
1735 | ||
1736 | /* Stop if we couldn't find slot */ | |
1737 | if (dst_page == NULL) | |
1738 | break; | |
1739 | ||
1740 | putback_zspage(pool, class, dst_page); | |
1741 | putback_zspage(pool, class, src_page); | |
1742 | spin_unlock(&class->lock); | |
1743 | nr_total_migrated += cc.nr_migrated; | |
1744 | cond_resched(); | |
1745 | spin_lock(&class->lock); | |
1746 | } | |
1747 | ||
1748 | if (src_page) | |
1749 | putback_zspage(pool, class, src_page); | |
1750 | ||
1751 | spin_unlock(&class->lock); | |
1752 | ||
1753 | return nr_total_migrated; | |
1754 | } | |
1755 | ||
1756 | unsigned long zs_compact(struct zs_pool *pool) | |
1757 | { | |
1758 | int i; | |
1759 | unsigned long nr_migrated = 0; | |
1760 | struct size_class *class; | |
1761 | ||
1762 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1763 | class = pool->size_class[i]; | |
1764 | if (!class) | |
1765 | continue; | |
1766 | if (class->index != i) | |
1767 | continue; | |
1768 | nr_migrated += __zs_compact(pool, class); | |
1769 | } | |
1770 | ||
312fcae2 MK |
1771 | return nr_migrated; |
1772 | } | |
1773 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1774 | |
00a61d86 | 1775 | /** |
66cdef66 GM |
1776 | * zs_create_pool - Creates an allocation pool to work from. |
1777 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1778 | * |
66cdef66 GM |
1779 | * This function must be called before anything when using |
1780 | * the zsmalloc allocator. | |
166cfda7 | 1781 | * |
66cdef66 GM |
1782 | * On success, a pointer to the newly created pool is returned, |
1783 | * otherwise NULL. | |
396b7fd6 | 1784 | */ |
3eba0c6a | 1785 | struct zs_pool *zs_create_pool(char *name, gfp_t flags) |
61989a80 | 1786 | { |
66cdef66 GM |
1787 | int i; |
1788 | struct zs_pool *pool; | |
1789 | struct size_class *prev_class = NULL; | |
61989a80 | 1790 | |
66cdef66 GM |
1791 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1792 | if (!pool) | |
1793 | return NULL; | |
61989a80 | 1794 | |
66cdef66 GM |
1795 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1796 | GFP_KERNEL); | |
1797 | if (!pool->size_class) { | |
1798 | kfree(pool); | |
1799 | return NULL; | |
1800 | } | |
61989a80 | 1801 | |
2e40e163 MK |
1802 | pool->name = kstrdup(name, GFP_KERNEL); |
1803 | if (!pool->name) | |
1804 | goto err; | |
1805 | ||
1806 | if (create_handle_cache(pool)) | |
1807 | goto err; | |
1808 | ||
c60369f0 | 1809 | /* |
66cdef66 GM |
1810 | * Iterate reversly, because, size of size_class that we want to use |
1811 | * for merging should be larger or equal to current size. | |
c60369f0 | 1812 | */ |
66cdef66 GM |
1813 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1814 | int size; | |
1815 | int pages_per_zspage; | |
1816 | struct size_class *class; | |
c60369f0 | 1817 | |
66cdef66 GM |
1818 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1819 | if (size > ZS_MAX_ALLOC_SIZE) | |
1820 | size = ZS_MAX_ALLOC_SIZE; | |
1821 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1822 | |
66cdef66 GM |
1823 | /* |
1824 | * size_class is used for normal zsmalloc operation such | |
1825 | * as alloc/free for that size. Although it is natural that we | |
1826 | * have one size_class for each size, there is a chance that we | |
1827 | * can get more memory utilization if we use one size_class for | |
1828 | * many different sizes whose size_class have same | |
1829 | * characteristics. So, we makes size_class point to | |
1830 | * previous size_class if possible. | |
1831 | */ | |
1832 | if (prev_class) { | |
1833 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1834 | pool->size_class[i] = prev_class; | |
1835 | continue; | |
1836 | } | |
1837 | } | |
1838 | ||
1839 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1840 | if (!class) | |
1841 | goto err; | |
1842 | ||
1843 | class->size = size; | |
1844 | class->index = i; | |
1845 | class->pages_per_zspage = pages_per_zspage; | |
7b60a685 MK |
1846 | if (pages_per_zspage == 1 && |
1847 | get_maxobj_per_zspage(size, pages_per_zspage) == 1) | |
1848 | class->huge = true; | |
66cdef66 GM |
1849 | spin_lock_init(&class->lock); |
1850 | pool->size_class[i] = class; | |
1851 | ||
1852 | prev_class = class; | |
61989a80 NG |
1853 | } |
1854 | ||
66cdef66 | 1855 | pool->flags = flags; |
b7418510 | 1856 | |
0f050d99 GM |
1857 | if (zs_pool_stat_create(name, pool)) |
1858 | goto err; | |
1859 | ||
66cdef66 GM |
1860 | return pool; |
1861 | ||
1862 | err: | |
1863 | zs_destroy_pool(pool); | |
1864 | return NULL; | |
61989a80 | 1865 | } |
66cdef66 | 1866 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1867 | |
66cdef66 | 1868 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1869 | { |
66cdef66 | 1870 | int i; |
61989a80 | 1871 | |
0f050d99 GM |
1872 | zs_pool_stat_destroy(pool); |
1873 | ||
66cdef66 GM |
1874 | for (i = 0; i < zs_size_classes; i++) { |
1875 | int fg; | |
1876 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1877 | |
66cdef66 GM |
1878 | if (!class) |
1879 | continue; | |
61989a80 | 1880 | |
66cdef66 GM |
1881 | if (class->index != i) |
1882 | continue; | |
61989a80 | 1883 | |
66cdef66 GM |
1884 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1885 | if (class->fullness_list[fg]) { | |
1886 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1887 | class->size, fg); | |
1888 | } | |
1889 | } | |
1890 | kfree(class); | |
1891 | } | |
f553646a | 1892 | |
2e40e163 | 1893 | destroy_handle_cache(pool); |
66cdef66 | 1894 | kfree(pool->size_class); |
0f050d99 | 1895 | kfree(pool->name); |
66cdef66 GM |
1896 | kfree(pool); |
1897 | } | |
1898 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 1899 | |
66cdef66 GM |
1900 | static int __init zs_init(void) |
1901 | { | |
1902 | int ret = zs_register_cpu_notifier(); | |
1903 | ||
0f050d99 GM |
1904 | if (ret) |
1905 | goto notifier_fail; | |
66cdef66 GM |
1906 | |
1907 | init_zs_size_classes(); | |
1908 | ||
1909 | #ifdef CONFIG_ZPOOL | |
1910 | zpool_register_driver(&zs_zpool_driver); | |
1911 | #endif | |
0f050d99 GM |
1912 | |
1913 | ret = zs_stat_init(); | |
1914 | if (ret) { | |
1915 | pr_err("zs stat initialization failed\n"); | |
1916 | goto stat_fail; | |
1917 | } | |
66cdef66 | 1918 | return 0; |
0f050d99 GM |
1919 | |
1920 | stat_fail: | |
1921 | #ifdef CONFIG_ZPOOL | |
1922 | zpool_unregister_driver(&zs_zpool_driver); | |
1923 | #endif | |
1924 | notifier_fail: | |
1925 | zs_unregister_cpu_notifier(); | |
1926 | ||
1927 | return ret; | |
61989a80 | 1928 | } |
61989a80 | 1929 | |
66cdef66 | 1930 | static void __exit zs_exit(void) |
61989a80 | 1931 | { |
66cdef66 GM |
1932 | #ifdef CONFIG_ZPOOL |
1933 | zpool_unregister_driver(&zs_zpool_driver); | |
1934 | #endif | |
1935 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
1936 | |
1937 | zs_stat_exit(); | |
61989a80 | 1938 | } |
069f101f BH |
1939 | |
1940 | module_init(zs_init); | |
1941 | module_exit(zs_exit); | |
1942 | ||
1943 | MODULE_LICENSE("Dual BSD/GPL"); | |
1944 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |