]>
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 | /* |
c3e3e88a NC |
15 | * This allocator is designed for use with zram. Thus, the allocator is |
16 | * supposed to work well under low memory conditions. In particular, it | |
17 | * never attempts higher order page allocation which is very likely to | |
18 | * fail under memory pressure. On the other hand, if we just use single | |
19 | * (0-order) pages, it would suffer from very high fragmentation -- | |
20 | * any object of size PAGE_SIZE/2 or larger would occupy an entire page. | |
21 | * This was one of the major issues with its predecessor (xvmalloc). | |
2db51dae NG |
22 | * |
23 | * To overcome these issues, zsmalloc allocates a bunch of 0-order pages | |
24 | * and links them together using various 'struct page' fields. These linked | |
25 | * pages act as a single higher-order page i.e. an object can span 0-order | |
26 | * page boundaries. The code refers to these linked pages as a single entity | |
27 | * called zspage. | |
28 | * | |
c3e3e88a NC |
29 | * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE |
30 | * since this satisfies the requirements of all its current users (in the | |
31 | * worst case, page is incompressible and is thus stored "as-is" i.e. in | |
32 | * uncompressed form). For allocation requests larger than this size, failure | |
33 | * is returned (see zs_malloc). | |
34 | * | |
35 | * Additionally, zs_malloc() does not return a dereferenceable pointer. | |
36 | * Instead, it returns an opaque handle (unsigned long) which encodes actual | |
37 | * location of the allocated object. The reason for this indirection is that | |
38 | * zsmalloc does not keep zspages permanently mapped since that would cause | |
39 | * issues on 32-bit systems where the VA region for kernel space mappings | |
40 | * is very small. So, before using the allocating memory, the object has to | |
41 | * be mapped using zs_map_object() to get a usable pointer and subsequently | |
42 | * unmapped using zs_unmap_object(). | |
43 | * | |
2db51dae NG |
44 | * Following is how we use various fields and flags of underlying |
45 | * struct page(s) to form a zspage. | |
46 | * | |
47 | * Usage of struct page fields: | |
48 | * page->first_page: points to the first component (0-order) page | |
49 | * page->index (union with page->freelist): offset of the first object | |
50 | * starting in this page. For the first page, this is | |
51 | * always 0, so we use this field (aka freelist) to point | |
52 | * to the first free object in zspage. | |
53 | * page->lru: links together all component pages (except the first page) | |
54 | * of a zspage | |
55 | * | |
56 | * For _first_ page only: | |
57 | * | |
58 | * page->private (union with page->first_page): refers to the | |
59 | * component page after the first page | |
7b60a685 MK |
60 | * If the page is first_page for huge object, it stores handle. |
61 | * Look at size_class->huge. | |
2db51dae NG |
62 | * page->freelist: points to the first free object in zspage. |
63 | * Free objects are linked together using in-place | |
64 | * metadata. | |
65 | * page->objects: maximum number of objects we can store in this | |
66 | * zspage (class->zspage_order * PAGE_SIZE / class->size) | |
67 | * page->lru: links together first pages of various zspages. | |
68 | * Basically forming list of zspages in a fullness group. | |
69 | * page->mapping: class index and fullness group of the zspage | |
70 | * | |
71 | * Usage of struct page flags: | |
72 | * PG_private: identifies the first component page | |
73 | * PG_private2: identifies the last component page | |
74 | * | |
75 | */ | |
76 | ||
61989a80 NG |
77 | #ifdef CONFIG_ZSMALLOC_DEBUG |
78 | #define DEBUG | |
79 | #endif | |
80 | ||
81 | #include <linux/module.h> | |
82 | #include <linux/kernel.h> | |
312fcae2 | 83 | #include <linux/sched.h> |
61989a80 NG |
84 | #include <linux/bitops.h> |
85 | #include <linux/errno.h> | |
86 | #include <linux/highmem.h> | |
61989a80 NG |
87 | #include <linux/string.h> |
88 | #include <linux/slab.h> | |
89 | #include <asm/tlbflush.h> | |
90 | #include <asm/pgtable.h> | |
91 | #include <linux/cpumask.h> | |
92 | #include <linux/cpu.h> | |
0cbb613f | 93 | #include <linux/vmalloc.h> |
c60369f0 | 94 | #include <linux/hardirq.h> |
0959c63f SJ |
95 | #include <linux/spinlock.h> |
96 | #include <linux/types.h> | |
0f050d99 | 97 | #include <linux/debugfs.h> |
bcf1647d | 98 | #include <linux/zsmalloc.h> |
c795779d | 99 | #include <linux/zpool.h> |
0959c63f SJ |
100 | |
101 | /* | |
102 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
103 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
104 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
105 | * to restore link_free pointer values. | |
106 | */ | |
107 | #define ZS_ALIGN 8 | |
108 | ||
109 | /* | |
110 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
111 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
112 | */ | |
113 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
114 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
115 | ||
2e40e163 MK |
116 | #define ZS_HANDLE_SIZE (sizeof(unsigned long)) |
117 | ||
0959c63f SJ |
118 | /* |
119 | * Object location (<PFN>, <obj_idx>) is encoded as | |
c3e3e88a | 120 | * as single (unsigned long) handle value. |
0959c63f SJ |
121 | * |
122 | * Note that object index <obj_idx> is relative to system | |
123 | * page <PFN> it is stored in, so for each sub-page belonging | |
124 | * to a zspage, obj_idx starts with 0. | |
125 | * | |
126 | * This is made more complicated by various memory models and PAE. | |
127 | */ | |
128 | ||
129 | #ifndef MAX_PHYSMEM_BITS | |
130 | #ifdef CONFIG_HIGHMEM64G | |
131 | #define MAX_PHYSMEM_BITS 36 | |
132 | #else /* !CONFIG_HIGHMEM64G */ | |
133 | /* | |
134 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
135 | * be PAGE_SHIFT | |
136 | */ | |
137 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
138 | #endif | |
139 | #endif | |
140 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
312fcae2 MK |
141 | |
142 | /* | |
143 | * Memory for allocating for handle keeps object position by | |
144 | * encoding <page, obj_idx> and the encoded value has a room | |
145 | * in least bit(ie, look at obj_to_location). | |
146 | * We use the bit to synchronize between object access by | |
147 | * user and migration. | |
148 | */ | |
149 | #define HANDLE_PIN_BIT 0 | |
150 | ||
151 | /* | |
152 | * Head in allocated object should have OBJ_ALLOCATED_TAG | |
153 | * to identify the object was allocated or not. | |
154 | * It's okay to add the status bit in the least bit because | |
155 | * header keeps handle which is 4byte-aligned address so we | |
156 | * have room for two bit at least. | |
157 | */ | |
158 | #define OBJ_ALLOCATED_TAG 1 | |
159 | #define OBJ_TAG_BITS 1 | |
160 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS) | |
0959c63f SJ |
161 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) |
162 | ||
163 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
164 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
165 | #define ZS_MIN_ALLOC_SIZE \ | |
166 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
2e40e163 | 167 | /* each chunk includes extra space to keep handle */ |
7b60a685 | 168 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE |
0959c63f SJ |
169 | |
170 | /* | |
7eb52512 | 171 | * On systems with 4K page size, this gives 255 size classes! There is a |
0959c63f SJ |
172 | * trader-off here: |
173 | * - Large number of size classes is potentially wasteful as free page are | |
174 | * spread across these classes | |
175 | * - Small number of size classes causes large internal fragmentation | |
176 | * - Probably its better to use specific size classes (empirically | |
177 | * determined). NOTE: all those class sizes must be set as multiple of | |
178 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
179 | * | |
180 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
181 | * (reason above) | |
182 | */ | |
d662b8eb | 183 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8) |
0959c63f SJ |
184 | |
185 | /* | |
186 | * We do not maintain any list for completely empty or full pages | |
187 | */ | |
188 | enum fullness_group { | |
189 | ZS_ALMOST_FULL, | |
190 | ZS_ALMOST_EMPTY, | |
191 | _ZS_NR_FULLNESS_GROUPS, | |
192 | ||
193 | ZS_EMPTY, | |
194 | ZS_FULL | |
195 | }; | |
196 | ||
0f050d99 GM |
197 | enum zs_stat_type { |
198 | OBJ_ALLOCATED, | |
199 | OBJ_USED, | |
200 | NR_ZS_STAT_TYPE, | |
201 | }; | |
202 | ||
203 | #ifdef CONFIG_ZSMALLOC_STAT | |
204 | ||
205 | static struct dentry *zs_stat_root; | |
206 | ||
207 | struct zs_size_stat { | |
208 | unsigned long objs[NR_ZS_STAT_TYPE]; | |
209 | }; | |
210 | ||
211 | #endif | |
212 | ||
40f9fb8c MG |
213 | /* |
214 | * number of size_classes | |
215 | */ | |
216 | static int zs_size_classes; | |
217 | ||
0959c63f SJ |
218 | /* |
219 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
220 | * n <= N / f, where | |
221 | * n = number of allocated objects | |
222 | * N = total number of objects zspage can store | |
6dd9737e | 223 | * f = fullness_threshold_frac |
0959c63f SJ |
224 | * |
225 | * Similarly, we assign zspage to: | |
226 | * ZS_ALMOST_FULL when n > N / f | |
227 | * ZS_EMPTY when n == 0 | |
228 | * ZS_FULL when n == N | |
229 | * | |
230 | * (see: fix_fullness_group()) | |
231 | */ | |
232 | static const int fullness_threshold_frac = 4; | |
233 | ||
234 | struct size_class { | |
235 | /* | |
236 | * Size of objects stored in this class. Must be multiple | |
237 | * of ZS_ALIGN. | |
238 | */ | |
239 | int size; | |
240 | unsigned int index; | |
241 | ||
242 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ | |
243 | int pages_per_zspage; | |
7b60a685 MK |
244 | /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */ |
245 | bool huge; | |
0959c63f | 246 | |
0f050d99 GM |
247 | #ifdef CONFIG_ZSMALLOC_STAT |
248 | struct zs_size_stat stats; | |
249 | #endif | |
250 | ||
0959c63f SJ |
251 | spinlock_t lock; |
252 | ||
0959c63f SJ |
253 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; |
254 | }; | |
255 | ||
256 | /* | |
257 | * Placed within free objects to form a singly linked list. | |
258 | * For every zspage, first_page->freelist gives head of this list. | |
259 | * | |
260 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
261 | */ | |
262 | struct link_free { | |
2e40e163 MK |
263 | union { |
264 | /* | |
265 | * Position of next free chunk (encodes <PFN, obj_idx>) | |
266 | * It's valid for non-allocated object | |
267 | */ | |
268 | void *next; | |
269 | /* | |
270 | * Handle of allocated object. | |
271 | */ | |
272 | unsigned long handle; | |
273 | }; | |
0959c63f SJ |
274 | }; |
275 | ||
276 | struct zs_pool { | |
0f050d99 GM |
277 | char *name; |
278 | ||
40f9fb8c | 279 | struct size_class **size_class; |
2e40e163 | 280 | struct kmem_cache *handle_cachep; |
0959c63f SJ |
281 | |
282 | gfp_t flags; /* allocation flags used when growing pool */ | |
13de8933 | 283 | atomic_long_t pages_allocated; |
0f050d99 GM |
284 | |
285 | #ifdef CONFIG_ZSMALLOC_STAT | |
286 | struct dentry *stat_dentry; | |
287 | #endif | |
0959c63f | 288 | }; |
61989a80 NG |
289 | |
290 | /* | |
291 | * A zspage's class index and fullness group | |
292 | * are encoded in its (first)page->mapping | |
293 | */ | |
294 | #define CLASS_IDX_BITS 28 | |
295 | #define FULLNESS_BITS 4 | |
296 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
297 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
298 | ||
f553646a | 299 | struct mapping_area { |
1b945aee | 300 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
301 | struct vm_struct *vm; /* vm area for mapping object that span pages */ |
302 | #else | |
303 | char *vm_buf; /* copy buffer for objects that span pages */ | |
304 | #endif | |
305 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
306 | enum zs_mapmode vm_mm; /* mapping mode */ | |
7b60a685 | 307 | bool huge; |
f553646a SJ |
308 | }; |
309 | ||
2e40e163 MK |
310 | static int create_handle_cache(struct zs_pool *pool) |
311 | { | |
312 | pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE, | |
313 | 0, 0, NULL); | |
314 | return pool->handle_cachep ? 0 : 1; | |
315 | } | |
316 | ||
317 | static void destroy_handle_cache(struct zs_pool *pool) | |
318 | { | |
319 | kmem_cache_destroy(pool->handle_cachep); | |
320 | } | |
321 | ||
322 | static unsigned long alloc_handle(struct zs_pool *pool) | |
323 | { | |
324 | return (unsigned long)kmem_cache_alloc(pool->handle_cachep, | |
325 | pool->flags & ~__GFP_HIGHMEM); | |
326 | } | |
327 | ||
328 | static void free_handle(struct zs_pool *pool, unsigned long handle) | |
329 | { | |
330 | kmem_cache_free(pool->handle_cachep, (void *)handle); | |
331 | } | |
332 | ||
333 | static void record_obj(unsigned long handle, unsigned long obj) | |
334 | { | |
335 | *(unsigned long *)handle = obj; | |
336 | } | |
337 | ||
c795779d DS |
338 | /* zpool driver */ |
339 | ||
340 | #ifdef CONFIG_ZPOOL | |
341 | ||
3eba0c6a | 342 | static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops) |
c795779d | 343 | { |
3eba0c6a | 344 | return zs_create_pool(name, gfp); |
c795779d DS |
345 | } |
346 | ||
347 | static void zs_zpool_destroy(void *pool) | |
348 | { | |
349 | zs_destroy_pool(pool); | |
350 | } | |
351 | ||
352 | static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp, | |
353 | unsigned long *handle) | |
354 | { | |
355 | *handle = zs_malloc(pool, size); | |
356 | return *handle ? 0 : -1; | |
357 | } | |
358 | static void zs_zpool_free(void *pool, unsigned long handle) | |
359 | { | |
360 | zs_free(pool, handle); | |
361 | } | |
362 | ||
363 | static int zs_zpool_shrink(void *pool, unsigned int pages, | |
364 | unsigned int *reclaimed) | |
365 | { | |
366 | return -EINVAL; | |
367 | } | |
368 | ||
369 | static void *zs_zpool_map(void *pool, unsigned long handle, | |
370 | enum zpool_mapmode mm) | |
371 | { | |
372 | enum zs_mapmode zs_mm; | |
373 | ||
374 | switch (mm) { | |
375 | case ZPOOL_MM_RO: | |
376 | zs_mm = ZS_MM_RO; | |
377 | break; | |
378 | case ZPOOL_MM_WO: | |
379 | zs_mm = ZS_MM_WO; | |
380 | break; | |
381 | case ZPOOL_MM_RW: /* fallthru */ | |
382 | default: | |
383 | zs_mm = ZS_MM_RW; | |
384 | break; | |
385 | } | |
386 | ||
387 | return zs_map_object(pool, handle, zs_mm); | |
388 | } | |
389 | static void zs_zpool_unmap(void *pool, unsigned long handle) | |
390 | { | |
391 | zs_unmap_object(pool, handle); | |
392 | } | |
393 | ||
394 | static u64 zs_zpool_total_size(void *pool) | |
395 | { | |
722cdc17 | 396 | return zs_get_total_pages(pool) << PAGE_SHIFT; |
c795779d DS |
397 | } |
398 | ||
399 | static struct zpool_driver zs_zpool_driver = { | |
400 | .type = "zsmalloc", | |
401 | .owner = THIS_MODULE, | |
402 | .create = zs_zpool_create, | |
403 | .destroy = zs_zpool_destroy, | |
404 | .malloc = zs_zpool_malloc, | |
405 | .free = zs_zpool_free, | |
406 | .shrink = zs_zpool_shrink, | |
407 | .map = zs_zpool_map, | |
408 | .unmap = zs_zpool_unmap, | |
409 | .total_size = zs_zpool_total_size, | |
410 | }; | |
411 | ||
137f8cff | 412 | MODULE_ALIAS("zpool-zsmalloc"); |
c795779d DS |
413 | #endif /* CONFIG_ZPOOL */ |
414 | ||
61989a80 NG |
415 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
416 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
417 | ||
418 | static int is_first_page(struct page *page) | |
419 | { | |
a27545bf | 420 | return PagePrivate(page); |
61989a80 NG |
421 | } |
422 | ||
423 | static int is_last_page(struct page *page) | |
424 | { | |
a27545bf | 425 | return PagePrivate2(page); |
61989a80 NG |
426 | } |
427 | ||
428 | static void get_zspage_mapping(struct page *page, unsigned int *class_idx, | |
429 | enum fullness_group *fullness) | |
430 | { | |
431 | unsigned long m; | |
432 | BUG_ON(!is_first_page(page)); | |
433 | ||
434 | m = (unsigned long)page->mapping; | |
435 | *fullness = m & FULLNESS_MASK; | |
436 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
437 | } | |
438 | ||
439 | static void set_zspage_mapping(struct page *page, unsigned int class_idx, | |
440 | enum fullness_group fullness) | |
441 | { | |
442 | unsigned long m; | |
443 | BUG_ON(!is_first_page(page)); | |
444 | ||
445 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
446 | (fullness & FULLNESS_MASK); | |
447 | page->mapping = (struct address_space *)m; | |
448 | } | |
449 | ||
c3e3e88a NC |
450 | /* |
451 | * zsmalloc divides the pool into various size classes where each | |
452 | * class maintains a list of zspages where each zspage is divided | |
453 | * into equal sized chunks. Each allocation falls into one of these | |
454 | * classes depending on its size. This function returns index of the | |
455 | * size class which has chunk size big enough to hold the give size. | |
456 | */ | |
61989a80 NG |
457 | static int get_size_class_index(int size) |
458 | { | |
459 | int idx = 0; | |
460 | ||
461 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
462 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
463 | ZS_SIZE_CLASS_DELTA); | |
464 | ||
7b60a685 | 465 | return min(zs_size_classes - 1, idx); |
61989a80 NG |
466 | } |
467 | ||
c3e3e88a NC |
468 | /* |
469 | * For each size class, zspages are divided into different groups | |
470 | * depending on how "full" they are. This was done so that we could | |
471 | * easily find empty or nearly empty zspages when we try to shrink | |
472 | * the pool (not yet implemented). This function returns fullness | |
473 | * status of the given page. | |
474 | */ | |
61989a80 NG |
475 | static enum fullness_group get_fullness_group(struct page *page) |
476 | { | |
477 | int inuse, max_objects; | |
478 | enum fullness_group fg; | |
479 | BUG_ON(!is_first_page(page)); | |
480 | ||
481 | inuse = page->inuse; | |
482 | max_objects = page->objects; | |
483 | ||
484 | if (inuse == 0) | |
485 | fg = ZS_EMPTY; | |
486 | else if (inuse == max_objects) | |
487 | fg = ZS_FULL; | |
d3d07c92 | 488 | else if (inuse <= 3 * max_objects / fullness_threshold_frac) |
61989a80 NG |
489 | fg = ZS_ALMOST_EMPTY; |
490 | else | |
491 | fg = ZS_ALMOST_FULL; | |
492 | ||
493 | return fg; | |
494 | } | |
495 | ||
c3e3e88a NC |
496 | /* |
497 | * Each size class maintains various freelists and zspages are assigned | |
498 | * to one of these freelists based on the number of live objects they | |
499 | * have. This functions inserts the given zspage into the freelist | |
500 | * identified by <class, fullness_group>. | |
501 | */ | |
61989a80 NG |
502 | static void insert_zspage(struct page *page, struct size_class *class, |
503 | enum fullness_group fullness) | |
504 | { | |
505 | struct page **head; | |
506 | ||
507 | BUG_ON(!is_first_page(page)); | |
508 | ||
509 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
510 | return; | |
511 | ||
512 | head = &class->fullness_list[fullness]; | |
513 | if (*head) | |
514 | list_add_tail(&page->lru, &(*head)->lru); | |
515 | ||
516 | *head = page; | |
517 | } | |
518 | ||
c3e3e88a NC |
519 | /* |
520 | * This function removes the given zspage from the freelist identified | |
521 | * by <class, fullness_group>. | |
522 | */ | |
61989a80 NG |
523 | static void remove_zspage(struct page *page, struct size_class *class, |
524 | enum fullness_group fullness) | |
525 | { | |
526 | struct page **head; | |
527 | ||
528 | BUG_ON(!is_first_page(page)); | |
529 | ||
530 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
531 | return; | |
532 | ||
533 | head = &class->fullness_list[fullness]; | |
534 | BUG_ON(!*head); | |
535 | if (list_empty(&(*head)->lru)) | |
536 | *head = NULL; | |
537 | else if (*head == page) | |
538 | *head = (struct page *)list_entry((*head)->lru.next, | |
539 | struct page, lru); | |
540 | ||
541 | list_del_init(&page->lru); | |
542 | } | |
543 | ||
c3e3e88a NC |
544 | /* |
545 | * Each size class maintains zspages in different fullness groups depending | |
546 | * on the number of live objects they contain. When allocating or freeing | |
547 | * objects, the fullness status of the page can change, say, from ALMOST_FULL | |
548 | * to ALMOST_EMPTY when freeing an object. This function checks if such | |
549 | * a status change has occurred for the given page and accordingly moves the | |
550 | * page from the freelist of the old fullness group to that of the new | |
551 | * fullness group. | |
552 | */ | |
c7806261 | 553 | static enum fullness_group fix_fullness_group(struct size_class *class, |
61989a80 NG |
554 | struct page *page) |
555 | { | |
556 | int class_idx; | |
61989a80 NG |
557 | enum fullness_group currfg, newfg; |
558 | ||
559 | BUG_ON(!is_first_page(page)); | |
560 | ||
561 | get_zspage_mapping(page, &class_idx, &currfg); | |
562 | newfg = get_fullness_group(page); | |
563 | if (newfg == currfg) | |
564 | goto out; | |
565 | ||
61989a80 NG |
566 | remove_zspage(page, class, currfg); |
567 | insert_zspage(page, class, newfg); | |
568 | set_zspage_mapping(page, class_idx, newfg); | |
569 | ||
570 | out: | |
571 | return newfg; | |
572 | } | |
573 | ||
574 | /* | |
575 | * We have to decide on how many pages to link together | |
576 | * to form a zspage for each size class. This is important | |
577 | * to reduce wastage due to unusable space left at end of | |
578 | * each zspage which is given as: | |
579 | * wastage = Zp - Zp % size_class | |
580 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... | |
581 | * | |
582 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
583 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
584 | * since then we can perfectly fit in 8 such objects. | |
585 | */ | |
2e3b6154 | 586 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
587 | { |
588 | int i, max_usedpc = 0; | |
589 | /* zspage order which gives maximum used size per KB */ | |
590 | int max_usedpc_order = 1; | |
591 | ||
84d4faab | 592 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
593 | int zspage_size; |
594 | int waste, usedpc; | |
595 | ||
596 | zspage_size = i * PAGE_SIZE; | |
597 | waste = zspage_size % class_size; | |
598 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
599 | ||
600 | if (usedpc > max_usedpc) { | |
601 | max_usedpc = usedpc; | |
602 | max_usedpc_order = i; | |
603 | } | |
604 | } | |
605 | ||
606 | return max_usedpc_order; | |
607 | } | |
608 | ||
609 | /* | |
610 | * A single 'zspage' is composed of many system pages which are | |
611 | * linked together using fields in struct page. This function finds | |
612 | * the first/head page, given any component page of a zspage. | |
613 | */ | |
614 | static struct page *get_first_page(struct page *page) | |
615 | { | |
616 | if (is_first_page(page)) | |
617 | return page; | |
618 | else | |
619 | return page->first_page; | |
620 | } | |
621 | ||
622 | static struct page *get_next_page(struct page *page) | |
623 | { | |
624 | struct page *next; | |
625 | ||
626 | if (is_last_page(page)) | |
627 | next = NULL; | |
628 | else if (is_first_page(page)) | |
e842b976 | 629 | next = (struct page *)page_private(page); |
61989a80 NG |
630 | else |
631 | next = list_entry(page->lru.next, struct page, lru); | |
632 | ||
633 | return next; | |
634 | } | |
635 | ||
67296874 OH |
636 | /* |
637 | * Encode <page, obj_idx> as a single handle value. | |
312fcae2 | 638 | * We use the least bit of handle for tagging. |
67296874 | 639 | */ |
312fcae2 | 640 | static void *location_to_obj(struct page *page, unsigned long obj_idx) |
61989a80 | 641 | { |
312fcae2 | 642 | unsigned long obj; |
61989a80 NG |
643 | |
644 | if (!page) { | |
645 | BUG_ON(obj_idx); | |
646 | return NULL; | |
647 | } | |
648 | ||
312fcae2 MK |
649 | obj = page_to_pfn(page) << OBJ_INDEX_BITS; |
650 | obj |= ((obj_idx) & OBJ_INDEX_MASK); | |
651 | obj <<= OBJ_TAG_BITS; | |
61989a80 | 652 | |
312fcae2 | 653 | return (void *)obj; |
61989a80 NG |
654 | } |
655 | ||
67296874 OH |
656 | /* |
657 | * Decode <page, obj_idx> pair from the given object handle. We adjust the | |
658 | * decoded obj_idx back to its original value since it was adjusted in | |
312fcae2 | 659 | * location_to_obj(). |
67296874 | 660 | */ |
312fcae2 | 661 | static void obj_to_location(unsigned long obj, struct page **page, |
61989a80 NG |
662 | unsigned long *obj_idx) |
663 | { | |
312fcae2 MK |
664 | obj >>= OBJ_TAG_BITS; |
665 | *page = pfn_to_page(obj >> OBJ_INDEX_BITS); | |
666 | *obj_idx = (obj & OBJ_INDEX_MASK); | |
61989a80 NG |
667 | } |
668 | ||
2e40e163 MK |
669 | static unsigned long handle_to_obj(unsigned long handle) |
670 | { | |
671 | return *(unsigned long *)handle; | |
672 | } | |
673 | ||
7b60a685 MK |
674 | static unsigned long obj_to_head(struct size_class *class, struct page *page, |
675 | void *obj) | |
312fcae2 | 676 | { |
7b60a685 MK |
677 | if (class->huge) { |
678 | VM_BUG_ON(!is_first_page(page)); | |
679 | return *(unsigned long *)page_private(page); | |
680 | } else | |
681 | return *(unsigned long *)obj; | |
312fcae2 MK |
682 | } |
683 | ||
61989a80 NG |
684 | static unsigned long obj_idx_to_offset(struct page *page, |
685 | unsigned long obj_idx, int class_size) | |
686 | { | |
687 | unsigned long off = 0; | |
688 | ||
689 | if (!is_first_page(page)) | |
690 | off = page->index; | |
691 | ||
692 | return off + obj_idx * class_size; | |
693 | } | |
694 | ||
312fcae2 MK |
695 | static inline int trypin_tag(unsigned long handle) |
696 | { | |
697 | unsigned long *ptr = (unsigned long *)handle; | |
698 | ||
699 | return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr); | |
700 | } | |
701 | ||
702 | static void pin_tag(unsigned long handle) | |
703 | { | |
704 | while (!trypin_tag(handle)); | |
705 | } | |
706 | ||
707 | static void unpin_tag(unsigned long handle) | |
708 | { | |
709 | unsigned long *ptr = (unsigned long *)handle; | |
710 | ||
711 | clear_bit_unlock(HANDLE_PIN_BIT, ptr); | |
712 | } | |
713 | ||
f4477e90 NG |
714 | static void reset_page(struct page *page) |
715 | { | |
716 | clear_bit(PG_private, &page->flags); | |
717 | clear_bit(PG_private_2, &page->flags); | |
718 | set_page_private(page, 0); | |
719 | page->mapping = NULL; | |
720 | page->freelist = NULL; | |
22b751c3 | 721 | page_mapcount_reset(page); |
f4477e90 NG |
722 | } |
723 | ||
61989a80 NG |
724 | static void free_zspage(struct page *first_page) |
725 | { | |
f4477e90 | 726 | struct page *nextp, *tmp, *head_extra; |
61989a80 NG |
727 | |
728 | BUG_ON(!is_first_page(first_page)); | |
729 | BUG_ON(first_page->inuse); | |
730 | ||
f4477e90 | 731 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 732 | |
f4477e90 | 733 | reset_page(first_page); |
61989a80 NG |
734 | __free_page(first_page); |
735 | ||
736 | /* zspage with only 1 system page */ | |
f4477e90 | 737 | if (!head_extra) |
61989a80 NG |
738 | return; |
739 | ||
f4477e90 | 740 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 741 | list_del(&nextp->lru); |
f4477e90 | 742 | reset_page(nextp); |
61989a80 NG |
743 | __free_page(nextp); |
744 | } | |
f4477e90 NG |
745 | reset_page(head_extra); |
746 | __free_page(head_extra); | |
61989a80 NG |
747 | } |
748 | ||
749 | /* Initialize a newly allocated zspage */ | |
750 | static void init_zspage(struct page *first_page, struct size_class *class) | |
751 | { | |
752 | unsigned long off = 0; | |
753 | struct page *page = first_page; | |
754 | ||
755 | BUG_ON(!is_first_page(first_page)); | |
756 | while (page) { | |
757 | struct page *next_page; | |
758 | struct link_free *link; | |
5538c562 | 759 | unsigned int i = 1; |
af4ee5e9 | 760 | void *vaddr; |
61989a80 NG |
761 | |
762 | /* | |
763 | * page->index stores offset of first object starting | |
764 | * in the page. For the first page, this is always 0, | |
765 | * so we use first_page->index (aka ->freelist) to store | |
766 | * head of corresponding zspage's freelist. | |
767 | */ | |
768 | if (page != first_page) | |
769 | page->index = off; | |
770 | ||
af4ee5e9 MK |
771 | vaddr = kmap_atomic(page); |
772 | link = (struct link_free *)vaddr + off / sizeof(*link); | |
5538c562 DS |
773 | |
774 | while ((off += class->size) < PAGE_SIZE) { | |
312fcae2 | 775 | link->next = location_to_obj(page, i++); |
5538c562 | 776 | link += class->size / sizeof(*link); |
61989a80 NG |
777 | } |
778 | ||
779 | /* | |
780 | * We now come to the last (full or partial) object on this | |
781 | * page, which must point to the first object on the next | |
782 | * page (if present) | |
783 | */ | |
784 | next_page = get_next_page(page); | |
312fcae2 | 785 | link->next = location_to_obj(next_page, 0); |
af4ee5e9 | 786 | kunmap_atomic(vaddr); |
61989a80 | 787 | page = next_page; |
5538c562 | 788 | off %= PAGE_SIZE; |
61989a80 NG |
789 | } |
790 | } | |
791 | ||
792 | /* | |
793 | * Allocate a zspage for the given size class | |
794 | */ | |
795 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
796 | { | |
797 | int i, error; | |
b4b700c5 | 798 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
799 | |
800 | /* | |
801 | * Allocate individual pages and link them together as: | |
802 | * 1. first page->private = first sub-page | |
803 | * 2. all sub-pages are linked together using page->lru | |
804 | * 3. each sub-page is linked to the first page using page->first_page | |
805 | * | |
806 | * For each size class, First/Head pages are linked together using | |
807 | * page->lru. Also, we set PG_private to identify the first page | |
808 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
809 | * identify the last page. | |
810 | */ | |
811 | error = -ENOMEM; | |
2e3b6154 | 812 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 813 | struct page *page; |
61989a80 NG |
814 | |
815 | page = alloc_page(flags); | |
816 | if (!page) | |
817 | goto cleanup; | |
818 | ||
819 | INIT_LIST_HEAD(&page->lru); | |
820 | if (i == 0) { /* first page */ | |
a27545bf | 821 | SetPagePrivate(page); |
61989a80 NG |
822 | set_page_private(page, 0); |
823 | first_page = page; | |
824 | first_page->inuse = 0; | |
825 | } | |
826 | if (i == 1) | |
e842b976 | 827 | set_page_private(first_page, (unsigned long)page); |
61989a80 NG |
828 | if (i >= 1) |
829 | page->first_page = first_page; | |
830 | if (i >= 2) | |
831 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 832 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 833 | SetPagePrivate2(page); |
61989a80 NG |
834 | prev_page = page; |
835 | } | |
836 | ||
837 | init_zspage(first_page, class); | |
838 | ||
312fcae2 | 839 | first_page->freelist = location_to_obj(first_page, 0); |
61989a80 | 840 | /* Maximum number of objects we can store in this zspage */ |
2e3b6154 | 841 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
842 | |
843 | error = 0; /* Success */ | |
844 | ||
845 | cleanup: | |
846 | if (unlikely(error) && first_page) { | |
847 | free_zspage(first_page); | |
848 | first_page = NULL; | |
849 | } | |
850 | ||
851 | return first_page; | |
852 | } | |
853 | ||
854 | static struct page *find_get_zspage(struct size_class *class) | |
855 | { | |
856 | int i; | |
857 | struct page *page; | |
858 | ||
859 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
860 | page = class->fullness_list[i]; | |
861 | if (page) | |
862 | break; | |
863 | } | |
864 | ||
865 | return page; | |
866 | } | |
867 | ||
1b945aee | 868 | #ifdef CONFIG_PGTABLE_MAPPING |
f553646a SJ |
869 | static inline int __zs_cpu_up(struct mapping_area *area) |
870 | { | |
871 | /* | |
872 | * Make sure we don't leak memory if a cpu UP notification | |
873 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
874 | */ | |
875 | if (area->vm) | |
876 | return 0; | |
877 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
878 | if (!area->vm) | |
879 | return -ENOMEM; | |
880 | return 0; | |
881 | } | |
882 | ||
883 | static inline void __zs_cpu_down(struct mapping_area *area) | |
884 | { | |
885 | if (area->vm) | |
886 | free_vm_area(area->vm); | |
887 | area->vm = NULL; | |
888 | } | |
889 | ||
890 | static inline void *__zs_map_object(struct mapping_area *area, | |
891 | struct page *pages[2], int off, int size) | |
892 | { | |
f6f8ed47 | 893 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages)); |
f553646a SJ |
894 | area->vm_addr = area->vm->addr; |
895 | return area->vm_addr + off; | |
896 | } | |
897 | ||
898 | static inline void __zs_unmap_object(struct mapping_area *area, | |
899 | struct page *pages[2], int off, int size) | |
900 | { | |
901 | unsigned long addr = (unsigned long)area->vm_addr; | |
f553646a | 902 | |
d95abbbb | 903 | unmap_kernel_range(addr, PAGE_SIZE * 2); |
f553646a SJ |
904 | } |
905 | ||
1b945aee | 906 | #else /* CONFIG_PGTABLE_MAPPING */ |
f553646a SJ |
907 | |
908 | static inline int __zs_cpu_up(struct mapping_area *area) | |
909 | { | |
910 | /* | |
911 | * Make sure we don't leak memory if a cpu UP notification | |
912 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
913 | */ | |
914 | if (area->vm_buf) | |
915 | return 0; | |
40f9fb8c | 916 | area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL); |
f553646a SJ |
917 | if (!area->vm_buf) |
918 | return -ENOMEM; | |
919 | return 0; | |
920 | } | |
921 | ||
922 | static inline void __zs_cpu_down(struct mapping_area *area) | |
923 | { | |
40f9fb8c | 924 | kfree(area->vm_buf); |
f553646a SJ |
925 | area->vm_buf = NULL; |
926 | } | |
927 | ||
928 | static void *__zs_map_object(struct mapping_area *area, | |
929 | struct page *pages[2], int off, int size) | |
5f601902 | 930 | { |
5f601902 SJ |
931 | int sizes[2]; |
932 | void *addr; | |
f553646a | 933 | char *buf = area->vm_buf; |
5f601902 | 934 | |
f553646a SJ |
935 | /* disable page faults to match kmap_atomic() return conditions */ |
936 | pagefault_disable(); | |
937 | ||
938 | /* no read fastpath */ | |
939 | if (area->vm_mm == ZS_MM_WO) | |
940 | goto out; | |
5f601902 SJ |
941 | |
942 | sizes[0] = PAGE_SIZE - off; | |
943 | sizes[1] = size - sizes[0]; | |
944 | ||
5f601902 SJ |
945 | /* copy object to per-cpu buffer */ |
946 | addr = kmap_atomic(pages[0]); | |
947 | memcpy(buf, addr + off, sizes[0]); | |
948 | kunmap_atomic(addr); | |
949 | addr = kmap_atomic(pages[1]); | |
950 | memcpy(buf + sizes[0], addr, sizes[1]); | |
951 | kunmap_atomic(addr); | |
f553646a SJ |
952 | out: |
953 | return area->vm_buf; | |
5f601902 SJ |
954 | } |
955 | ||
f553646a SJ |
956 | static void __zs_unmap_object(struct mapping_area *area, |
957 | struct page *pages[2], int off, int size) | |
5f601902 | 958 | { |
5f601902 SJ |
959 | int sizes[2]; |
960 | void *addr; | |
2e40e163 | 961 | char *buf; |
5f601902 | 962 | |
f553646a SJ |
963 | /* no write fastpath */ |
964 | if (area->vm_mm == ZS_MM_RO) | |
965 | goto out; | |
5f601902 | 966 | |
7b60a685 MK |
967 | buf = area->vm_buf; |
968 | if (!area->huge) { | |
969 | buf = buf + ZS_HANDLE_SIZE; | |
970 | size -= ZS_HANDLE_SIZE; | |
971 | off += ZS_HANDLE_SIZE; | |
972 | } | |
2e40e163 | 973 | |
5f601902 SJ |
974 | sizes[0] = PAGE_SIZE - off; |
975 | sizes[1] = size - sizes[0]; | |
976 | ||
977 | /* copy per-cpu buffer to object */ | |
978 | addr = kmap_atomic(pages[0]); | |
979 | memcpy(addr + off, buf, sizes[0]); | |
980 | kunmap_atomic(addr); | |
981 | addr = kmap_atomic(pages[1]); | |
982 | memcpy(addr, buf + sizes[0], sizes[1]); | |
983 | kunmap_atomic(addr); | |
f553646a SJ |
984 | |
985 | out: | |
986 | /* enable page faults to match kunmap_atomic() return conditions */ | |
987 | pagefault_enable(); | |
5f601902 | 988 | } |
61989a80 | 989 | |
1b945aee | 990 | #endif /* CONFIG_PGTABLE_MAPPING */ |
f553646a | 991 | |
61989a80 NG |
992 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
993 | void *pcpu) | |
994 | { | |
f553646a | 995 | int ret, cpu = (long)pcpu; |
61989a80 NG |
996 | struct mapping_area *area; |
997 | ||
998 | switch (action) { | |
999 | case CPU_UP_PREPARE: | |
1000 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
1001 | ret = __zs_cpu_up(area); |
1002 | if (ret) | |
1003 | return notifier_from_errno(ret); | |
61989a80 NG |
1004 | break; |
1005 | case CPU_DEAD: | |
1006 | case CPU_UP_CANCELED: | |
1007 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 1008 | __zs_cpu_down(area); |
61989a80 NG |
1009 | break; |
1010 | } | |
1011 | ||
1012 | return NOTIFY_OK; | |
1013 | } | |
1014 | ||
1015 | static struct notifier_block zs_cpu_nb = { | |
1016 | .notifier_call = zs_cpu_notifier | |
1017 | }; | |
1018 | ||
b1b00a5b | 1019 | static int zs_register_cpu_notifier(void) |
61989a80 | 1020 | { |
b1b00a5b | 1021 | int cpu, uninitialized_var(ret); |
61989a80 | 1022 | |
f0e71fcd SB |
1023 | cpu_notifier_register_begin(); |
1024 | ||
1025 | __register_cpu_notifier(&zs_cpu_nb); | |
61989a80 NG |
1026 | for_each_online_cpu(cpu) { |
1027 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
b1b00a5b SS |
1028 | if (notifier_to_errno(ret)) |
1029 | break; | |
61989a80 | 1030 | } |
f0e71fcd SB |
1031 | |
1032 | cpu_notifier_register_done(); | |
b1b00a5b SS |
1033 | return notifier_to_errno(ret); |
1034 | } | |
f0e71fcd | 1035 | |
66cdef66 | 1036 | static void zs_unregister_cpu_notifier(void) |
40f9fb8c | 1037 | { |
66cdef66 | 1038 | int cpu; |
40f9fb8c | 1039 | |
66cdef66 | 1040 | cpu_notifier_register_begin(); |
40f9fb8c | 1041 | |
66cdef66 GM |
1042 | for_each_online_cpu(cpu) |
1043 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
1044 | __unregister_cpu_notifier(&zs_cpu_nb); | |
40f9fb8c | 1045 | |
66cdef66 | 1046 | cpu_notifier_register_done(); |
b1b00a5b SS |
1047 | } |
1048 | ||
66cdef66 | 1049 | static void init_zs_size_classes(void) |
b1b00a5b | 1050 | { |
66cdef66 | 1051 | int nr; |
c795779d | 1052 | |
66cdef66 GM |
1053 | nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1; |
1054 | if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA) | |
1055 | nr += 1; | |
40f9fb8c | 1056 | |
66cdef66 | 1057 | zs_size_classes = nr; |
61989a80 NG |
1058 | } |
1059 | ||
9eec4cd5 JK |
1060 | static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage) |
1061 | { | |
1062 | return pages_per_zspage * PAGE_SIZE / size; | |
1063 | } | |
1064 | ||
1065 | static bool can_merge(struct size_class *prev, int size, int pages_per_zspage) | |
1066 | { | |
1067 | if (prev->pages_per_zspage != pages_per_zspage) | |
1068 | return false; | |
1069 | ||
1070 | if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage) | |
1071 | != get_maxobj_per_zspage(size, pages_per_zspage)) | |
1072 | return false; | |
1073 | ||
1074 | return true; | |
1075 | } | |
1076 | ||
312fcae2 MK |
1077 | static bool zspage_full(struct page *page) |
1078 | { | |
1079 | BUG_ON(!is_first_page(page)); | |
1080 | ||
1081 | return page->inuse == page->objects; | |
1082 | } | |
1083 | ||
0f050d99 GM |
1084 | #ifdef CONFIG_ZSMALLOC_STAT |
1085 | ||
1086 | static inline void zs_stat_inc(struct size_class *class, | |
1087 | enum zs_stat_type type, unsigned long cnt) | |
1088 | { | |
1089 | class->stats.objs[type] += cnt; | |
1090 | } | |
1091 | ||
1092 | static inline void zs_stat_dec(struct size_class *class, | |
1093 | enum zs_stat_type type, unsigned long cnt) | |
1094 | { | |
1095 | class->stats.objs[type] -= cnt; | |
1096 | } | |
1097 | ||
1098 | static inline unsigned long zs_stat_get(struct size_class *class, | |
1099 | enum zs_stat_type type) | |
1100 | { | |
1101 | return class->stats.objs[type]; | |
1102 | } | |
1103 | ||
1104 | static int __init zs_stat_init(void) | |
1105 | { | |
1106 | if (!debugfs_initialized()) | |
1107 | return -ENODEV; | |
1108 | ||
1109 | zs_stat_root = debugfs_create_dir("zsmalloc", NULL); | |
1110 | if (!zs_stat_root) | |
1111 | return -ENOMEM; | |
1112 | ||
1113 | return 0; | |
1114 | } | |
1115 | ||
1116 | static void __exit zs_stat_exit(void) | |
1117 | { | |
1118 | debugfs_remove_recursive(zs_stat_root); | |
1119 | } | |
1120 | ||
1121 | static int zs_stats_size_show(struct seq_file *s, void *v) | |
1122 | { | |
1123 | int i; | |
1124 | struct zs_pool *pool = s->private; | |
1125 | struct size_class *class; | |
1126 | int objs_per_zspage; | |
1127 | unsigned long obj_allocated, obj_used, pages_used; | |
1128 | unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0; | |
1129 | ||
1130 | seq_printf(s, " %5s %5s %13s %10s %10s\n", "class", "size", | |
1131 | "obj_allocated", "obj_used", "pages_used"); | |
1132 | ||
1133 | for (i = 0; i < zs_size_classes; i++) { | |
1134 | class = pool->size_class[i]; | |
1135 | ||
1136 | if (class->index != i) | |
1137 | continue; | |
1138 | ||
1139 | spin_lock(&class->lock); | |
1140 | obj_allocated = zs_stat_get(class, OBJ_ALLOCATED); | |
1141 | obj_used = zs_stat_get(class, OBJ_USED); | |
1142 | spin_unlock(&class->lock); | |
1143 | ||
1144 | objs_per_zspage = get_maxobj_per_zspage(class->size, | |
1145 | class->pages_per_zspage); | |
1146 | pages_used = obj_allocated / objs_per_zspage * | |
1147 | class->pages_per_zspage; | |
1148 | ||
1149 | seq_printf(s, " %5u %5u %10lu %10lu %10lu\n", i, | |
1150 | class->size, obj_allocated, obj_used, pages_used); | |
1151 | ||
1152 | total_objs += obj_allocated; | |
1153 | total_used_objs += obj_used; | |
1154 | total_pages += pages_used; | |
1155 | } | |
1156 | ||
1157 | seq_puts(s, "\n"); | |
1158 | seq_printf(s, " %5s %5s %10lu %10lu %10lu\n", "Total", "", | |
1159 | total_objs, total_used_objs, total_pages); | |
1160 | ||
1161 | return 0; | |
1162 | } | |
1163 | ||
1164 | static int zs_stats_size_open(struct inode *inode, struct file *file) | |
1165 | { | |
1166 | return single_open(file, zs_stats_size_show, inode->i_private); | |
1167 | } | |
1168 | ||
1169 | static const struct file_operations zs_stat_size_ops = { | |
1170 | .open = zs_stats_size_open, | |
1171 | .read = seq_read, | |
1172 | .llseek = seq_lseek, | |
1173 | .release = single_release, | |
1174 | }; | |
1175 | ||
1176 | static int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
1177 | { | |
1178 | struct dentry *entry; | |
1179 | ||
1180 | if (!zs_stat_root) | |
1181 | return -ENODEV; | |
1182 | ||
1183 | entry = debugfs_create_dir(name, zs_stat_root); | |
1184 | if (!entry) { | |
1185 | pr_warn("debugfs dir <%s> creation failed\n", name); | |
1186 | return -ENOMEM; | |
1187 | } | |
1188 | pool->stat_dentry = entry; | |
1189 | ||
1190 | entry = debugfs_create_file("obj_in_classes", S_IFREG | S_IRUGO, | |
1191 | pool->stat_dentry, pool, &zs_stat_size_ops); | |
1192 | if (!entry) { | |
1193 | pr_warn("%s: debugfs file entry <%s> creation failed\n", | |
1194 | name, "obj_in_classes"); | |
1195 | return -ENOMEM; | |
1196 | } | |
1197 | ||
1198 | return 0; | |
1199 | } | |
1200 | ||
1201 | static void zs_pool_stat_destroy(struct zs_pool *pool) | |
1202 | { | |
1203 | debugfs_remove_recursive(pool->stat_dentry); | |
1204 | } | |
1205 | ||
1206 | #else /* CONFIG_ZSMALLOC_STAT */ | |
1207 | ||
1208 | static inline void zs_stat_inc(struct size_class *class, | |
1209 | enum zs_stat_type type, unsigned long cnt) | |
1210 | { | |
1211 | } | |
1212 | ||
1213 | static inline void zs_stat_dec(struct size_class *class, | |
1214 | enum zs_stat_type type, unsigned long cnt) | |
1215 | { | |
1216 | } | |
1217 | ||
1218 | static inline unsigned long zs_stat_get(struct size_class *class, | |
1219 | enum zs_stat_type type) | |
1220 | { | |
1221 | return 0; | |
1222 | } | |
1223 | ||
1224 | static int __init zs_stat_init(void) | |
1225 | { | |
1226 | return 0; | |
1227 | } | |
1228 | ||
1229 | static void __exit zs_stat_exit(void) | |
1230 | { | |
1231 | } | |
1232 | ||
1233 | static inline int zs_pool_stat_create(char *name, struct zs_pool *pool) | |
1234 | { | |
1235 | return 0; | |
1236 | } | |
1237 | ||
1238 | static inline void zs_pool_stat_destroy(struct zs_pool *pool) | |
1239 | { | |
1240 | } | |
1241 | ||
1242 | #endif | |
1243 | ||
66cdef66 GM |
1244 | unsigned long zs_get_total_pages(struct zs_pool *pool) |
1245 | { | |
1246 | return atomic_long_read(&pool->pages_allocated); | |
1247 | } | |
1248 | EXPORT_SYMBOL_GPL(zs_get_total_pages); | |
1249 | ||
4bbc0bc0 | 1250 | /** |
66cdef66 GM |
1251 | * zs_map_object - get address of allocated object from handle. |
1252 | * @pool: pool from which the object was allocated | |
1253 | * @handle: handle returned from zs_malloc | |
4bbc0bc0 | 1254 | * |
66cdef66 GM |
1255 | * Before using an object allocated from zs_malloc, it must be mapped using |
1256 | * this function. When done with the object, it must be unmapped using | |
1257 | * zs_unmap_object. | |
4bbc0bc0 | 1258 | * |
66cdef66 GM |
1259 | * Only one object can be mapped per cpu at a time. There is no protection |
1260 | * against nested mappings. | |
1261 | * | |
1262 | * This function returns with preemption and page faults disabled. | |
4bbc0bc0 | 1263 | */ |
66cdef66 GM |
1264 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
1265 | enum zs_mapmode mm) | |
61989a80 | 1266 | { |
66cdef66 | 1267 | struct page *page; |
2e40e163 | 1268 | unsigned long obj, obj_idx, off; |
61989a80 | 1269 | |
66cdef66 GM |
1270 | unsigned int class_idx; |
1271 | enum fullness_group fg; | |
1272 | struct size_class *class; | |
1273 | struct mapping_area *area; | |
1274 | struct page *pages[2]; | |
2e40e163 | 1275 | void *ret; |
61989a80 | 1276 | |
66cdef66 | 1277 | BUG_ON(!handle); |
40f9fb8c | 1278 | |
9eec4cd5 | 1279 | /* |
66cdef66 GM |
1280 | * Because we use per-cpu mapping areas shared among the |
1281 | * pools/users, we can't allow mapping in interrupt context | |
1282 | * because it can corrupt another users mappings. | |
9eec4cd5 | 1283 | */ |
66cdef66 | 1284 | BUG_ON(in_interrupt()); |
61989a80 | 1285 | |
312fcae2 MK |
1286 | /* From now on, migration cannot move the object */ |
1287 | pin_tag(handle); | |
1288 | ||
2e40e163 MK |
1289 | obj = handle_to_obj(handle); |
1290 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1291 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1292 | class = pool->size_class[class_idx]; | |
1293 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
df8b5bb9 | 1294 | |
66cdef66 GM |
1295 | area = &get_cpu_var(zs_map_area); |
1296 | area->vm_mm = mm; | |
1297 | if (off + class->size <= PAGE_SIZE) { | |
1298 | /* this object is contained entirely within a page */ | |
1299 | area->vm_addr = kmap_atomic(page); | |
2e40e163 MK |
1300 | ret = area->vm_addr + off; |
1301 | goto out; | |
61989a80 NG |
1302 | } |
1303 | ||
66cdef66 GM |
1304 | /* this object spans two pages */ |
1305 | pages[0] = page; | |
1306 | pages[1] = get_next_page(page); | |
1307 | BUG_ON(!pages[1]); | |
9eec4cd5 | 1308 | |
2e40e163 MK |
1309 | ret = __zs_map_object(area, pages, off, class->size); |
1310 | out: | |
7b60a685 MK |
1311 | if (!class->huge) |
1312 | ret += ZS_HANDLE_SIZE; | |
1313 | ||
1314 | return ret; | |
61989a80 | 1315 | } |
66cdef66 | 1316 | EXPORT_SYMBOL_GPL(zs_map_object); |
61989a80 | 1317 | |
66cdef66 | 1318 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 | 1319 | { |
66cdef66 | 1320 | struct page *page; |
2e40e163 | 1321 | unsigned long obj, obj_idx, off; |
61989a80 | 1322 | |
66cdef66 GM |
1323 | unsigned int class_idx; |
1324 | enum fullness_group fg; | |
1325 | struct size_class *class; | |
1326 | struct mapping_area *area; | |
9eec4cd5 | 1327 | |
66cdef66 | 1328 | BUG_ON(!handle); |
9eec4cd5 | 1329 | |
2e40e163 MK |
1330 | obj = handle_to_obj(handle); |
1331 | obj_to_location(obj, &page, &obj_idx); | |
66cdef66 GM |
1332 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
1333 | class = pool->size_class[class_idx]; | |
1334 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
61989a80 | 1335 | |
66cdef66 GM |
1336 | area = this_cpu_ptr(&zs_map_area); |
1337 | if (off + class->size <= PAGE_SIZE) | |
1338 | kunmap_atomic(area->vm_addr); | |
1339 | else { | |
1340 | struct page *pages[2]; | |
40f9fb8c | 1341 | |
66cdef66 GM |
1342 | pages[0] = page; |
1343 | pages[1] = get_next_page(page); | |
1344 | BUG_ON(!pages[1]); | |
1345 | ||
1346 | __zs_unmap_object(area, pages, off, class->size); | |
1347 | } | |
1348 | put_cpu_var(zs_map_area); | |
312fcae2 | 1349 | unpin_tag(handle); |
61989a80 | 1350 | } |
66cdef66 | 1351 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
61989a80 | 1352 | |
c7806261 MK |
1353 | static unsigned long obj_malloc(struct page *first_page, |
1354 | struct size_class *class, unsigned long handle) | |
1355 | { | |
1356 | unsigned long obj; | |
1357 | struct link_free *link; | |
1358 | ||
1359 | struct page *m_page; | |
1360 | unsigned long m_objidx, m_offset; | |
1361 | void *vaddr; | |
1362 | ||
312fcae2 | 1363 | handle |= OBJ_ALLOCATED_TAG; |
c7806261 MK |
1364 | obj = (unsigned long)first_page->freelist; |
1365 | obj_to_location(obj, &m_page, &m_objidx); | |
1366 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
1367 | ||
1368 | vaddr = kmap_atomic(m_page); | |
1369 | link = (struct link_free *)vaddr + m_offset / sizeof(*link); | |
1370 | first_page->freelist = link->next; | |
7b60a685 MK |
1371 | if (!class->huge) |
1372 | /* record handle in the header of allocated chunk */ | |
1373 | link->handle = handle; | |
1374 | else | |
1375 | /* record handle in first_page->private */ | |
1376 | set_page_private(first_page, handle); | |
c7806261 MK |
1377 | kunmap_atomic(vaddr); |
1378 | first_page->inuse++; | |
1379 | zs_stat_inc(class, OBJ_USED, 1); | |
1380 | ||
1381 | return obj; | |
1382 | } | |
1383 | ||
1384 | ||
61989a80 NG |
1385 | /** |
1386 | * zs_malloc - Allocate block of given size from pool. | |
1387 | * @pool: pool to allocate from | |
1388 | * @size: size of block to allocate | |
61989a80 | 1389 | * |
00a61d86 | 1390 | * On success, handle to the allocated object is returned, |
c2344348 | 1391 | * otherwise 0. |
61989a80 NG |
1392 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
1393 | */ | |
c2344348 | 1394 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
61989a80 | 1395 | { |
2e40e163 | 1396 | unsigned long handle, obj; |
61989a80 | 1397 | struct size_class *class; |
c7806261 | 1398 | struct page *first_page; |
61989a80 | 1399 | |
7b60a685 | 1400 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
2e40e163 MK |
1401 | return 0; |
1402 | ||
1403 | handle = alloc_handle(pool); | |
1404 | if (!handle) | |
c2344348 | 1405 | return 0; |
61989a80 | 1406 | |
2e40e163 MK |
1407 | /* extra space in chunk to keep the handle */ |
1408 | size += ZS_HANDLE_SIZE; | |
9eec4cd5 | 1409 | class = pool->size_class[get_size_class_index(size)]; |
7b60a685 MK |
1410 | /* In huge class size, we store the handle into first_page->private */ |
1411 | if (class->huge) { | |
1412 | size -= ZS_HANDLE_SIZE; | |
1413 | class = pool->size_class[get_size_class_index(size)]; | |
1414 | } | |
61989a80 NG |
1415 | |
1416 | spin_lock(&class->lock); | |
1417 | first_page = find_get_zspage(class); | |
1418 | ||
1419 | if (!first_page) { | |
1420 | spin_unlock(&class->lock); | |
1421 | first_page = alloc_zspage(class, pool->flags); | |
2e40e163 MK |
1422 | if (unlikely(!first_page)) { |
1423 | free_handle(pool, handle); | |
c2344348 | 1424 | return 0; |
2e40e163 | 1425 | } |
61989a80 NG |
1426 | |
1427 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
13de8933 MK |
1428 | atomic_long_add(class->pages_per_zspage, |
1429 | &pool->pages_allocated); | |
0f050d99 | 1430 | |
61989a80 | 1431 | spin_lock(&class->lock); |
0f050d99 GM |
1432 | zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1433 | class->size, class->pages_per_zspage)); | |
61989a80 NG |
1434 | } |
1435 | ||
c7806261 | 1436 | obj = obj_malloc(first_page, class, handle); |
61989a80 | 1437 | /* Now move the zspage to another fullness group, if required */ |
c7806261 | 1438 | fix_fullness_group(class, first_page); |
2e40e163 | 1439 | record_obj(handle, obj); |
61989a80 NG |
1440 | spin_unlock(&class->lock); |
1441 | ||
2e40e163 | 1442 | return handle; |
61989a80 NG |
1443 | } |
1444 | EXPORT_SYMBOL_GPL(zs_malloc); | |
1445 | ||
c7806261 MK |
1446 | static void obj_free(struct zs_pool *pool, struct size_class *class, |
1447 | unsigned long obj) | |
61989a80 NG |
1448 | { |
1449 | struct link_free *link; | |
1450 | struct page *first_page, *f_page; | |
c7806261 | 1451 | unsigned long f_objidx, f_offset; |
af4ee5e9 | 1452 | void *vaddr; |
61989a80 | 1453 | int class_idx; |
61989a80 NG |
1454 | enum fullness_group fullness; |
1455 | ||
c7806261 | 1456 | BUG_ON(!obj); |
61989a80 | 1457 | |
312fcae2 | 1458 | obj &= ~OBJ_ALLOCATED_TAG; |
2e40e163 | 1459 | obj_to_location(obj, &f_page, &f_objidx); |
61989a80 NG |
1460 | first_page = get_first_page(f_page); |
1461 | ||
1462 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
61989a80 NG |
1463 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
1464 | ||
c7806261 | 1465 | vaddr = kmap_atomic(f_page); |
61989a80 NG |
1466 | |
1467 | /* Insert this object in containing zspage's freelist */ | |
af4ee5e9 | 1468 | link = (struct link_free *)(vaddr + f_offset); |
61989a80 | 1469 | link->next = first_page->freelist; |
7b60a685 MK |
1470 | if (class->huge) |
1471 | set_page_private(first_page, 0); | |
af4ee5e9 | 1472 | kunmap_atomic(vaddr); |
c2344348 | 1473 | first_page->freelist = (void *)obj; |
61989a80 | 1474 | first_page->inuse--; |
0f050d99 | 1475 | zs_stat_dec(class, OBJ_USED, 1); |
c7806261 MK |
1476 | } |
1477 | ||
1478 | void zs_free(struct zs_pool *pool, unsigned long handle) | |
1479 | { | |
1480 | struct page *first_page, *f_page; | |
1481 | unsigned long obj, f_objidx; | |
1482 | int class_idx; | |
1483 | struct size_class *class; | |
1484 | enum fullness_group fullness; | |
1485 | ||
1486 | if (unlikely(!handle)) | |
1487 | return; | |
1488 | ||
312fcae2 | 1489 | pin_tag(handle); |
c7806261 | 1490 | obj = handle_to_obj(handle); |
c7806261 MK |
1491 | obj_to_location(obj, &f_page, &f_objidx); |
1492 | first_page = get_first_page(f_page); | |
1493 | ||
1494 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1495 | class = pool->size_class[class_idx]; | |
1496 | ||
1497 | spin_lock(&class->lock); | |
1498 | obj_free(pool, class, obj); | |
1499 | fullness = fix_fullness_group(class, first_page); | |
312fcae2 | 1500 | if (fullness == ZS_EMPTY) { |
0f050d99 GM |
1501 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1502 | class->size, class->pages_per_zspage)); | |
312fcae2 MK |
1503 | atomic_long_sub(class->pages_per_zspage, |
1504 | &pool->pages_allocated); | |
1505 | free_zspage(first_page); | |
1506 | } | |
61989a80 | 1507 | spin_unlock(&class->lock); |
312fcae2 | 1508 | unpin_tag(handle); |
61989a80 | 1509 | |
312fcae2 MK |
1510 | free_handle(pool, handle); |
1511 | } | |
1512 | EXPORT_SYMBOL_GPL(zs_free); | |
1513 | ||
1514 | static void zs_object_copy(unsigned long src, unsigned long dst, | |
1515 | struct size_class *class) | |
1516 | { | |
1517 | struct page *s_page, *d_page; | |
1518 | unsigned long s_objidx, d_objidx; | |
1519 | unsigned long s_off, d_off; | |
1520 | void *s_addr, *d_addr; | |
1521 | int s_size, d_size, size; | |
1522 | int written = 0; | |
1523 | ||
1524 | s_size = d_size = class->size; | |
1525 | ||
1526 | obj_to_location(src, &s_page, &s_objidx); | |
1527 | obj_to_location(dst, &d_page, &d_objidx); | |
1528 | ||
1529 | s_off = obj_idx_to_offset(s_page, s_objidx, class->size); | |
1530 | d_off = obj_idx_to_offset(d_page, d_objidx, class->size); | |
1531 | ||
1532 | if (s_off + class->size > PAGE_SIZE) | |
1533 | s_size = PAGE_SIZE - s_off; | |
1534 | ||
1535 | if (d_off + class->size > PAGE_SIZE) | |
1536 | d_size = PAGE_SIZE - d_off; | |
1537 | ||
1538 | s_addr = kmap_atomic(s_page); | |
1539 | d_addr = kmap_atomic(d_page); | |
1540 | ||
1541 | while (1) { | |
1542 | size = min(s_size, d_size); | |
1543 | memcpy(d_addr + d_off, s_addr + s_off, size); | |
1544 | written += size; | |
1545 | ||
1546 | if (written == class->size) | |
1547 | break; | |
1548 | ||
1549 | if (s_off + size >= PAGE_SIZE) { | |
1550 | kunmap_atomic(d_addr); | |
1551 | kunmap_atomic(s_addr); | |
1552 | s_page = get_next_page(s_page); | |
1553 | BUG_ON(!s_page); | |
1554 | s_addr = kmap_atomic(s_page); | |
1555 | d_addr = kmap_atomic(d_page); | |
1556 | s_size = class->size - written; | |
1557 | s_off = 0; | |
1558 | } else { | |
1559 | s_off += size; | |
1560 | s_size -= size; | |
1561 | } | |
1562 | ||
1563 | if (d_off + size >= PAGE_SIZE) { | |
1564 | kunmap_atomic(d_addr); | |
1565 | d_page = get_next_page(d_page); | |
1566 | BUG_ON(!d_page); | |
1567 | d_addr = kmap_atomic(d_page); | |
1568 | d_size = class->size - written; | |
1569 | d_off = 0; | |
1570 | } else { | |
1571 | d_off += size; | |
1572 | d_size -= size; | |
1573 | } | |
1574 | } | |
1575 | ||
1576 | kunmap_atomic(d_addr); | |
1577 | kunmap_atomic(s_addr); | |
1578 | } | |
1579 | ||
1580 | /* | |
1581 | * Find alloced object in zspage from index object and | |
1582 | * return handle. | |
1583 | */ | |
1584 | static unsigned long find_alloced_obj(struct page *page, int index, | |
1585 | struct size_class *class) | |
1586 | { | |
1587 | unsigned long head; | |
1588 | int offset = 0; | |
1589 | unsigned long handle = 0; | |
1590 | void *addr = kmap_atomic(page); | |
1591 | ||
1592 | if (!is_first_page(page)) | |
1593 | offset = page->index; | |
1594 | offset += class->size * index; | |
1595 | ||
1596 | while (offset < PAGE_SIZE) { | |
7b60a685 | 1597 | head = obj_to_head(class, page, addr + offset); |
312fcae2 MK |
1598 | if (head & OBJ_ALLOCATED_TAG) { |
1599 | handle = head & ~OBJ_ALLOCATED_TAG; | |
1600 | if (trypin_tag(handle)) | |
1601 | break; | |
1602 | handle = 0; | |
1603 | } | |
1604 | ||
1605 | offset += class->size; | |
1606 | index++; | |
1607 | } | |
1608 | ||
1609 | kunmap_atomic(addr); | |
1610 | return handle; | |
1611 | } | |
1612 | ||
1613 | struct zs_compact_control { | |
1614 | /* Source page for migration which could be a subpage of zspage. */ | |
1615 | struct page *s_page; | |
1616 | /* Destination page for migration which should be a first page | |
1617 | * of zspage. */ | |
1618 | struct page *d_page; | |
1619 | /* Starting object index within @s_page which used for live object | |
1620 | * in the subpage. */ | |
1621 | int index; | |
1622 | /* how many of objects are migrated */ | |
1623 | int nr_migrated; | |
1624 | }; | |
1625 | ||
1626 | static int migrate_zspage(struct zs_pool *pool, struct size_class *class, | |
1627 | struct zs_compact_control *cc) | |
1628 | { | |
1629 | unsigned long used_obj, free_obj; | |
1630 | unsigned long handle; | |
1631 | struct page *s_page = cc->s_page; | |
1632 | struct page *d_page = cc->d_page; | |
1633 | unsigned long index = cc->index; | |
1634 | int nr_migrated = 0; | |
1635 | int ret = 0; | |
1636 | ||
1637 | while (1) { | |
1638 | handle = find_alloced_obj(s_page, index, class); | |
1639 | if (!handle) { | |
1640 | s_page = get_next_page(s_page); | |
1641 | if (!s_page) | |
1642 | break; | |
1643 | index = 0; | |
1644 | continue; | |
1645 | } | |
1646 | ||
1647 | /* Stop if there is no more space */ | |
1648 | if (zspage_full(d_page)) { | |
1649 | unpin_tag(handle); | |
1650 | ret = -ENOMEM; | |
1651 | break; | |
1652 | } | |
1653 | ||
1654 | used_obj = handle_to_obj(handle); | |
1655 | free_obj = obj_malloc(d_page, class, handle); | |
1656 | zs_object_copy(used_obj, free_obj, class); | |
1657 | index++; | |
1658 | record_obj(handle, free_obj); | |
1659 | unpin_tag(handle); | |
1660 | obj_free(pool, class, used_obj); | |
1661 | nr_migrated++; | |
1662 | } | |
1663 | ||
1664 | /* Remember last position in this iteration */ | |
1665 | cc->s_page = s_page; | |
1666 | cc->index = index; | |
1667 | cc->nr_migrated = nr_migrated; | |
1668 | ||
1669 | return ret; | |
1670 | } | |
1671 | ||
1672 | static struct page *alloc_target_page(struct size_class *class) | |
1673 | { | |
1674 | int i; | |
1675 | struct page *page; | |
1676 | ||
1677 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
1678 | page = class->fullness_list[i]; | |
1679 | if (page) { | |
1680 | remove_zspage(page, class, i); | |
1681 | break; | |
1682 | } | |
1683 | } | |
1684 | ||
1685 | return page; | |
1686 | } | |
1687 | ||
1688 | static void putback_zspage(struct zs_pool *pool, struct size_class *class, | |
1689 | struct page *first_page) | |
1690 | { | |
1691 | int class_idx; | |
1692 | enum fullness_group fullness; | |
1693 | ||
1694 | BUG_ON(!is_first_page(first_page)); | |
1695 | ||
1696 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
1697 | insert_zspage(first_page, class, fullness); | |
1698 | fullness = fix_fullness_group(class, first_page); | |
13de8933 | 1699 | if (fullness == ZS_EMPTY) { |
312fcae2 MK |
1700 | zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage( |
1701 | class->size, class->pages_per_zspage)); | |
13de8933 MK |
1702 | atomic_long_sub(class->pages_per_zspage, |
1703 | &pool->pages_allocated); | |
312fcae2 | 1704 | |
61989a80 | 1705 | free_zspage(first_page); |
13de8933 | 1706 | } |
61989a80 | 1707 | } |
312fcae2 MK |
1708 | |
1709 | static struct page *isolate_source_page(struct size_class *class) | |
1710 | { | |
1711 | struct page *page; | |
1712 | ||
1713 | page = class->fullness_list[ZS_ALMOST_EMPTY]; | |
1714 | if (page) | |
1715 | remove_zspage(page, class, ZS_ALMOST_EMPTY); | |
1716 | ||
1717 | return page; | |
1718 | } | |
1719 | ||
1720 | static unsigned long __zs_compact(struct zs_pool *pool, | |
1721 | struct size_class *class) | |
1722 | { | |
1723 | int nr_to_migrate; | |
1724 | struct zs_compact_control cc; | |
1725 | struct page *src_page; | |
1726 | struct page *dst_page = NULL; | |
1727 | unsigned long nr_total_migrated = 0; | |
1728 | ||
1729 | cond_resched(); | |
1730 | ||
1731 | spin_lock(&class->lock); | |
1732 | while ((src_page = isolate_source_page(class))) { | |
1733 | ||
1734 | BUG_ON(!is_first_page(src_page)); | |
1735 | ||
1736 | /* The goal is to migrate all live objects in source page */ | |
1737 | nr_to_migrate = src_page->inuse; | |
1738 | cc.index = 0; | |
1739 | cc.s_page = src_page; | |
1740 | ||
1741 | while ((dst_page = alloc_target_page(class))) { | |
1742 | cc.d_page = dst_page; | |
1743 | /* | |
1744 | * If there is no more space in dst_page, try to | |
1745 | * allocate another zspage. | |
1746 | */ | |
1747 | if (!migrate_zspage(pool, class, &cc)) | |
1748 | break; | |
1749 | ||
1750 | putback_zspage(pool, class, dst_page); | |
1751 | nr_total_migrated += cc.nr_migrated; | |
1752 | nr_to_migrate -= cc.nr_migrated; | |
1753 | } | |
1754 | ||
1755 | /* Stop if we couldn't find slot */ | |
1756 | if (dst_page == NULL) | |
1757 | break; | |
1758 | ||
1759 | putback_zspage(pool, class, dst_page); | |
1760 | putback_zspage(pool, class, src_page); | |
1761 | spin_unlock(&class->lock); | |
1762 | nr_total_migrated += cc.nr_migrated; | |
1763 | cond_resched(); | |
1764 | spin_lock(&class->lock); | |
1765 | } | |
1766 | ||
1767 | if (src_page) | |
1768 | putback_zspage(pool, class, src_page); | |
1769 | ||
1770 | spin_unlock(&class->lock); | |
1771 | ||
1772 | return nr_total_migrated; | |
1773 | } | |
1774 | ||
1775 | unsigned long zs_compact(struct zs_pool *pool) | |
1776 | { | |
1777 | int i; | |
1778 | unsigned long nr_migrated = 0; | |
1779 | struct size_class *class; | |
1780 | ||
1781 | for (i = zs_size_classes - 1; i >= 0; i--) { | |
1782 | class = pool->size_class[i]; | |
1783 | if (!class) | |
1784 | continue; | |
1785 | if (class->index != i) | |
1786 | continue; | |
1787 | nr_migrated += __zs_compact(pool, class); | |
1788 | } | |
1789 | ||
1790 | synchronize_rcu(); | |
1791 | ||
1792 | return nr_migrated; | |
1793 | } | |
1794 | EXPORT_SYMBOL_GPL(zs_compact); | |
61989a80 | 1795 | |
00a61d86 | 1796 | /** |
66cdef66 GM |
1797 | * zs_create_pool - Creates an allocation pool to work from. |
1798 | * @flags: allocation flags used to allocate pool metadata | |
166cfda7 | 1799 | * |
66cdef66 GM |
1800 | * This function must be called before anything when using |
1801 | * the zsmalloc allocator. | |
166cfda7 | 1802 | * |
66cdef66 GM |
1803 | * On success, a pointer to the newly created pool is returned, |
1804 | * otherwise NULL. | |
396b7fd6 | 1805 | */ |
3eba0c6a | 1806 | struct zs_pool *zs_create_pool(char *name, gfp_t flags) |
61989a80 | 1807 | { |
66cdef66 GM |
1808 | int i; |
1809 | struct zs_pool *pool; | |
1810 | struct size_class *prev_class = NULL; | |
61989a80 | 1811 | |
66cdef66 GM |
1812 | pool = kzalloc(sizeof(*pool), GFP_KERNEL); |
1813 | if (!pool) | |
1814 | return NULL; | |
61989a80 | 1815 | |
66cdef66 GM |
1816 | pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *), |
1817 | GFP_KERNEL); | |
1818 | if (!pool->size_class) { | |
1819 | kfree(pool); | |
1820 | return NULL; | |
1821 | } | |
61989a80 | 1822 | |
2e40e163 MK |
1823 | pool->name = kstrdup(name, GFP_KERNEL); |
1824 | if (!pool->name) | |
1825 | goto err; | |
1826 | ||
1827 | if (create_handle_cache(pool)) | |
1828 | goto err; | |
1829 | ||
c60369f0 | 1830 | /* |
66cdef66 GM |
1831 | * Iterate reversly, because, size of size_class that we want to use |
1832 | * for merging should be larger or equal to current size. | |
c60369f0 | 1833 | */ |
66cdef66 GM |
1834 | for (i = zs_size_classes - 1; i >= 0; i--) { |
1835 | int size; | |
1836 | int pages_per_zspage; | |
1837 | struct size_class *class; | |
c60369f0 | 1838 | |
66cdef66 GM |
1839 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
1840 | if (size > ZS_MAX_ALLOC_SIZE) | |
1841 | size = ZS_MAX_ALLOC_SIZE; | |
1842 | pages_per_zspage = get_pages_per_zspage(size); | |
61989a80 | 1843 | |
66cdef66 GM |
1844 | /* |
1845 | * size_class is used for normal zsmalloc operation such | |
1846 | * as alloc/free for that size. Although it is natural that we | |
1847 | * have one size_class for each size, there is a chance that we | |
1848 | * can get more memory utilization if we use one size_class for | |
1849 | * many different sizes whose size_class have same | |
1850 | * characteristics. So, we makes size_class point to | |
1851 | * previous size_class if possible. | |
1852 | */ | |
1853 | if (prev_class) { | |
1854 | if (can_merge(prev_class, size, pages_per_zspage)) { | |
1855 | pool->size_class[i] = prev_class; | |
1856 | continue; | |
1857 | } | |
1858 | } | |
1859 | ||
1860 | class = kzalloc(sizeof(struct size_class), GFP_KERNEL); | |
1861 | if (!class) | |
1862 | goto err; | |
1863 | ||
1864 | class->size = size; | |
1865 | class->index = i; | |
1866 | class->pages_per_zspage = pages_per_zspage; | |
7b60a685 MK |
1867 | if (pages_per_zspage == 1 && |
1868 | get_maxobj_per_zspage(size, pages_per_zspage) == 1) | |
1869 | class->huge = true; | |
66cdef66 GM |
1870 | spin_lock_init(&class->lock); |
1871 | pool->size_class[i] = class; | |
1872 | ||
1873 | prev_class = class; | |
61989a80 NG |
1874 | } |
1875 | ||
66cdef66 | 1876 | pool->flags = flags; |
b7418510 | 1877 | |
0f050d99 GM |
1878 | if (zs_pool_stat_create(name, pool)) |
1879 | goto err; | |
1880 | ||
66cdef66 GM |
1881 | return pool; |
1882 | ||
1883 | err: | |
1884 | zs_destroy_pool(pool); | |
1885 | return NULL; | |
61989a80 | 1886 | } |
66cdef66 | 1887 | EXPORT_SYMBOL_GPL(zs_create_pool); |
61989a80 | 1888 | |
66cdef66 | 1889 | void zs_destroy_pool(struct zs_pool *pool) |
61989a80 | 1890 | { |
66cdef66 | 1891 | int i; |
61989a80 | 1892 | |
0f050d99 GM |
1893 | zs_pool_stat_destroy(pool); |
1894 | ||
66cdef66 GM |
1895 | for (i = 0; i < zs_size_classes; i++) { |
1896 | int fg; | |
1897 | struct size_class *class = pool->size_class[i]; | |
61989a80 | 1898 | |
66cdef66 GM |
1899 | if (!class) |
1900 | continue; | |
61989a80 | 1901 | |
66cdef66 GM |
1902 | if (class->index != i) |
1903 | continue; | |
61989a80 | 1904 | |
66cdef66 GM |
1905 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
1906 | if (class->fullness_list[fg]) { | |
1907 | pr_info("Freeing non-empty class with size %db, fullness group %d\n", | |
1908 | class->size, fg); | |
1909 | } | |
1910 | } | |
1911 | kfree(class); | |
1912 | } | |
f553646a | 1913 | |
2e40e163 | 1914 | destroy_handle_cache(pool); |
66cdef66 | 1915 | kfree(pool->size_class); |
0f050d99 | 1916 | kfree(pool->name); |
66cdef66 GM |
1917 | kfree(pool); |
1918 | } | |
1919 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
b7418510 | 1920 | |
66cdef66 GM |
1921 | static int __init zs_init(void) |
1922 | { | |
1923 | int ret = zs_register_cpu_notifier(); | |
1924 | ||
0f050d99 GM |
1925 | if (ret) |
1926 | goto notifier_fail; | |
66cdef66 GM |
1927 | |
1928 | init_zs_size_classes(); | |
1929 | ||
1930 | #ifdef CONFIG_ZPOOL | |
1931 | zpool_register_driver(&zs_zpool_driver); | |
1932 | #endif | |
0f050d99 GM |
1933 | |
1934 | ret = zs_stat_init(); | |
1935 | if (ret) { | |
1936 | pr_err("zs stat initialization failed\n"); | |
1937 | goto stat_fail; | |
1938 | } | |
66cdef66 | 1939 | return 0; |
0f050d99 GM |
1940 | |
1941 | stat_fail: | |
1942 | #ifdef CONFIG_ZPOOL | |
1943 | zpool_unregister_driver(&zs_zpool_driver); | |
1944 | #endif | |
1945 | notifier_fail: | |
1946 | zs_unregister_cpu_notifier(); | |
1947 | ||
1948 | return ret; | |
61989a80 | 1949 | } |
61989a80 | 1950 | |
66cdef66 | 1951 | static void __exit zs_exit(void) |
61989a80 | 1952 | { |
66cdef66 GM |
1953 | #ifdef CONFIG_ZPOOL |
1954 | zpool_unregister_driver(&zs_zpool_driver); | |
1955 | #endif | |
1956 | zs_unregister_cpu_notifier(); | |
0f050d99 GM |
1957 | |
1958 | zs_stat_exit(); | |
61989a80 | 1959 | } |
069f101f BH |
1960 | |
1961 | module_init(zs_init); | |
1962 | module_exit(zs_exit); | |
1963 | ||
1964 | MODULE_LICENSE("Dual BSD/GPL"); | |
1965 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |