Commit | Line | Data |
---|---|---|
61989a80 NG |
1 | /* |
2 | * zsmalloc memory allocator | |
3 | * | |
4 | * Copyright (C) 2011 Nitin Gupta | |
5 | * | |
6 | * This code is released using a dual license strategy: BSD/GPL | |
7 | * You can choose the license that better fits your requirements. | |
8 | * | |
9 | * Released under the terms of 3-clause BSD License | |
10 | * Released under the terms of GNU General Public License Version 2.0 | |
11 | */ | |
12 | ||
2db51dae NG |
13 | |
14 | /* | |
15 | * This allocator is designed for use with zcache and zram. Thus, the | |
16 | * allocator is supposed to work well under low memory conditions. In | |
17 | * particular, it never attempts higher order page allocation which is | |
18 | * very likely to fail under memory pressure. On the other hand, if we | |
19 | * just use single (0-order) pages, it would suffer from very high | |
20 | * fragmentation -- any object of size PAGE_SIZE/2 or larger would occupy | |
21 | * an entire page. This was one of the major issues with its predecessor | |
22 | * (xvmalloc). | |
23 | * | |
24 | * To overcome these issues, zsmalloc allocates a bunch of 0-order pages | |
25 | * and links them together using various 'struct page' fields. These linked | |
26 | * pages act as a single higher-order page i.e. an object can span 0-order | |
27 | * page boundaries. The code refers to these linked pages as a single entity | |
28 | * called zspage. | |
29 | * | |
30 | * Following is how we use various fields and flags of underlying | |
31 | * struct page(s) to form a zspage. | |
32 | * | |
33 | * Usage of struct page fields: | |
34 | * page->first_page: points to the first component (0-order) page | |
35 | * page->index (union with page->freelist): offset of the first object | |
36 | * starting in this page. For the first page, this is | |
37 | * always 0, so we use this field (aka freelist) to point | |
38 | * to the first free object in zspage. | |
39 | * page->lru: links together all component pages (except the first page) | |
40 | * of a zspage | |
41 | * | |
42 | * For _first_ page only: | |
43 | * | |
44 | * page->private (union with page->first_page): refers to the | |
45 | * component page after the first page | |
46 | * page->freelist: points to the first free object in zspage. | |
47 | * Free objects are linked together using in-place | |
48 | * metadata. | |
49 | * page->objects: maximum number of objects we can store in this | |
50 | * zspage (class->zspage_order * PAGE_SIZE / class->size) | |
51 | * page->lru: links together first pages of various zspages. | |
52 | * Basically forming list of zspages in a fullness group. | |
53 | * page->mapping: class index and fullness group of the zspage | |
54 | * | |
55 | * Usage of struct page flags: | |
56 | * PG_private: identifies the first component page | |
57 | * PG_private2: identifies the last component page | |
58 | * | |
59 | */ | |
60 | ||
61989a80 NG |
61 | #ifdef CONFIG_ZSMALLOC_DEBUG |
62 | #define DEBUG | |
63 | #endif | |
64 | ||
65 | #include <linux/module.h> | |
66 | #include <linux/kernel.h> | |
67 | #include <linux/bitops.h> | |
68 | #include <linux/errno.h> | |
69 | #include <linux/highmem.h> | |
70 | #include <linux/init.h> | |
71 | #include <linux/string.h> | |
72 | #include <linux/slab.h> | |
73 | #include <asm/tlbflush.h> | |
74 | #include <asm/pgtable.h> | |
75 | #include <linux/cpumask.h> | |
76 | #include <linux/cpu.h> | |
0cbb613f | 77 | #include <linux/vmalloc.h> |
c60369f0 | 78 | #include <linux/hardirq.h> |
0959c63f SJ |
79 | #include <linux/spinlock.h> |
80 | #include <linux/types.h> | |
61989a80 NG |
81 | |
82 | #include "zsmalloc.h" | |
0959c63f SJ |
83 | |
84 | /* | |
85 | * This must be power of 2 and greater than of equal to sizeof(link_free). | |
86 | * These two conditions ensure that any 'struct link_free' itself doesn't | |
87 | * span more than 1 page which avoids complex case of mapping 2 pages simply | |
88 | * to restore link_free pointer values. | |
89 | */ | |
90 | #define ZS_ALIGN 8 | |
91 | ||
92 | /* | |
93 | * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single) | |
94 | * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N. | |
95 | */ | |
96 | #define ZS_MAX_ZSPAGE_ORDER 2 | |
97 | #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER) | |
98 | ||
99 | /* | |
100 | * Object location (<PFN>, <obj_idx>) is encoded as | |
101 | * as single (void *) handle value. | |
102 | * | |
103 | * Note that object index <obj_idx> is relative to system | |
104 | * page <PFN> it is stored in, so for each sub-page belonging | |
105 | * to a zspage, obj_idx starts with 0. | |
106 | * | |
107 | * This is made more complicated by various memory models and PAE. | |
108 | */ | |
109 | ||
110 | #ifndef MAX_PHYSMEM_BITS | |
111 | #ifdef CONFIG_HIGHMEM64G | |
112 | #define MAX_PHYSMEM_BITS 36 | |
113 | #else /* !CONFIG_HIGHMEM64G */ | |
114 | /* | |
115 | * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just | |
116 | * be PAGE_SHIFT | |
117 | */ | |
118 | #define MAX_PHYSMEM_BITS BITS_PER_LONG | |
119 | #endif | |
120 | #endif | |
121 | #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT) | |
122 | #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS) | |
123 | #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1) | |
124 | ||
125 | #define MAX(a, b) ((a) >= (b) ? (a) : (b)) | |
126 | /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */ | |
127 | #define ZS_MIN_ALLOC_SIZE \ | |
128 | MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS)) | |
129 | #define ZS_MAX_ALLOC_SIZE PAGE_SIZE | |
130 | ||
131 | /* | |
132 | * On systems with 4K page size, this gives 254 size classes! There is a | |
133 | * trader-off here: | |
134 | * - Large number of size classes is potentially wasteful as free page are | |
135 | * spread across these classes | |
136 | * - Small number of size classes causes large internal fragmentation | |
137 | * - Probably its better to use specific size classes (empirically | |
138 | * determined). NOTE: all those class sizes must be set as multiple of | |
139 | * ZS_ALIGN to make sure link_free itself never has to span 2 pages. | |
140 | * | |
141 | * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN | |
142 | * (reason above) | |
143 | */ | |
d662b8eb | 144 | #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8) |
0959c63f SJ |
145 | #define ZS_SIZE_CLASSES ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \ |
146 | ZS_SIZE_CLASS_DELTA + 1) | |
147 | ||
148 | /* | |
149 | * We do not maintain any list for completely empty or full pages | |
150 | */ | |
151 | enum fullness_group { | |
152 | ZS_ALMOST_FULL, | |
153 | ZS_ALMOST_EMPTY, | |
154 | _ZS_NR_FULLNESS_GROUPS, | |
155 | ||
156 | ZS_EMPTY, | |
157 | ZS_FULL | |
158 | }; | |
159 | ||
160 | /* | |
161 | * We assign a page to ZS_ALMOST_EMPTY fullness group when: | |
162 | * n <= N / f, where | |
163 | * n = number of allocated objects | |
164 | * N = total number of objects zspage can store | |
165 | * f = 1/fullness_threshold_frac | |
166 | * | |
167 | * Similarly, we assign zspage to: | |
168 | * ZS_ALMOST_FULL when n > N / f | |
169 | * ZS_EMPTY when n == 0 | |
170 | * ZS_FULL when n == N | |
171 | * | |
172 | * (see: fix_fullness_group()) | |
173 | */ | |
174 | static const int fullness_threshold_frac = 4; | |
175 | ||
176 | struct size_class { | |
177 | /* | |
178 | * Size of objects stored in this class. Must be multiple | |
179 | * of ZS_ALIGN. | |
180 | */ | |
181 | int size; | |
182 | unsigned int index; | |
183 | ||
184 | /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */ | |
185 | int pages_per_zspage; | |
186 | ||
187 | spinlock_t lock; | |
188 | ||
189 | /* stats */ | |
190 | u64 pages_allocated; | |
191 | ||
192 | struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS]; | |
193 | }; | |
194 | ||
195 | /* | |
196 | * Placed within free objects to form a singly linked list. | |
197 | * For every zspage, first_page->freelist gives head of this list. | |
198 | * | |
199 | * This must be power of 2 and less than or equal to ZS_ALIGN | |
200 | */ | |
201 | struct link_free { | |
202 | /* Handle of next free chunk (encodes <PFN, obj_idx>) */ | |
203 | void *next; | |
204 | }; | |
205 | ||
206 | struct zs_pool { | |
207 | struct size_class size_class[ZS_SIZE_CLASSES]; | |
208 | ||
209 | gfp_t flags; /* allocation flags used when growing pool */ | |
210 | const char *name; | |
211 | }; | |
61989a80 NG |
212 | |
213 | /* | |
214 | * A zspage's class index and fullness group | |
215 | * are encoded in its (first)page->mapping | |
216 | */ | |
217 | #define CLASS_IDX_BITS 28 | |
218 | #define FULLNESS_BITS 4 | |
219 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) | |
220 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) | |
221 | ||
f553646a SJ |
222 | /* |
223 | * By default, zsmalloc uses a copy-based object mapping method to access | |
224 | * allocations that span two pages. However, if a particular architecture | |
99155188 MK |
225 | * performs VM mapping faster than copying, then it should be added here |
226 | * so that USE_PGTABLE_MAPPING is defined. This causes zsmalloc to use | |
227 | * page table mapping rather than copying for object mapping. | |
f553646a SJ |
228 | */ |
229 | #if defined(CONFIG_ARM) | |
230 | #define USE_PGTABLE_MAPPING | |
231 | #endif | |
232 | ||
233 | struct mapping_area { | |
234 | #ifdef USE_PGTABLE_MAPPING | |
235 | struct vm_struct *vm; /* vm area for mapping object that span pages */ | |
236 | #else | |
237 | char *vm_buf; /* copy buffer for objects that span pages */ | |
238 | #endif | |
239 | char *vm_addr; /* address of kmap_atomic()'ed pages */ | |
240 | enum zs_mapmode vm_mm; /* mapping mode */ | |
241 | }; | |
242 | ||
243 | ||
61989a80 NG |
244 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
245 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); | |
246 | ||
247 | static int is_first_page(struct page *page) | |
248 | { | |
a27545bf | 249 | return PagePrivate(page); |
61989a80 NG |
250 | } |
251 | ||
252 | static int is_last_page(struct page *page) | |
253 | { | |
a27545bf | 254 | return PagePrivate2(page); |
61989a80 NG |
255 | } |
256 | ||
257 | static void get_zspage_mapping(struct page *page, unsigned int *class_idx, | |
258 | enum fullness_group *fullness) | |
259 | { | |
260 | unsigned long m; | |
261 | BUG_ON(!is_first_page(page)); | |
262 | ||
263 | m = (unsigned long)page->mapping; | |
264 | *fullness = m & FULLNESS_MASK; | |
265 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; | |
266 | } | |
267 | ||
268 | static void set_zspage_mapping(struct page *page, unsigned int class_idx, | |
269 | enum fullness_group fullness) | |
270 | { | |
271 | unsigned long m; | |
272 | BUG_ON(!is_first_page(page)); | |
273 | ||
274 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | | |
275 | (fullness & FULLNESS_MASK); | |
276 | page->mapping = (struct address_space *)m; | |
277 | } | |
278 | ||
279 | static int get_size_class_index(int size) | |
280 | { | |
281 | int idx = 0; | |
282 | ||
283 | if (likely(size > ZS_MIN_ALLOC_SIZE)) | |
284 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, | |
285 | ZS_SIZE_CLASS_DELTA); | |
286 | ||
287 | return idx; | |
288 | } | |
289 | ||
290 | static enum fullness_group get_fullness_group(struct page *page) | |
291 | { | |
292 | int inuse, max_objects; | |
293 | enum fullness_group fg; | |
294 | BUG_ON(!is_first_page(page)); | |
295 | ||
296 | inuse = page->inuse; | |
297 | max_objects = page->objects; | |
298 | ||
299 | if (inuse == 0) | |
300 | fg = ZS_EMPTY; | |
301 | else if (inuse == max_objects) | |
302 | fg = ZS_FULL; | |
303 | else if (inuse <= max_objects / fullness_threshold_frac) | |
304 | fg = ZS_ALMOST_EMPTY; | |
305 | else | |
306 | fg = ZS_ALMOST_FULL; | |
307 | ||
308 | return fg; | |
309 | } | |
310 | ||
311 | static void insert_zspage(struct page *page, struct size_class *class, | |
312 | enum fullness_group fullness) | |
313 | { | |
314 | struct page **head; | |
315 | ||
316 | BUG_ON(!is_first_page(page)); | |
317 | ||
318 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
319 | return; | |
320 | ||
321 | head = &class->fullness_list[fullness]; | |
322 | if (*head) | |
323 | list_add_tail(&page->lru, &(*head)->lru); | |
324 | ||
325 | *head = page; | |
326 | } | |
327 | ||
328 | static void remove_zspage(struct page *page, struct size_class *class, | |
329 | enum fullness_group fullness) | |
330 | { | |
331 | struct page **head; | |
332 | ||
333 | BUG_ON(!is_first_page(page)); | |
334 | ||
335 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) | |
336 | return; | |
337 | ||
338 | head = &class->fullness_list[fullness]; | |
339 | BUG_ON(!*head); | |
340 | if (list_empty(&(*head)->lru)) | |
341 | *head = NULL; | |
342 | else if (*head == page) | |
343 | *head = (struct page *)list_entry((*head)->lru.next, | |
344 | struct page, lru); | |
345 | ||
346 | list_del_init(&page->lru); | |
347 | } | |
348 | ||
349 | static enum fullness_group fix_fullness_group(struct zs_pool *pool, | |
350 | struct page *page) | |
351 | { | |
352 | int class_idx; | |
353 | struct size_class *class; | |
354 | enum fullness_group currfg, newfg; | |
355 | ||
356 | BUG_ON(!is_first_page(page)); | |
357 | ||
358 | get_zspage_mapping(page, &class_idx, &currfg); | |
359 | newfg = get_fullness_group(page); | |
360 | if (newfg == currfg) | |
361 | goto out; | |
362 | ||
363 | class = &pool->size_class[class_idx]; | |
364 | remove_zspage(page, class, currfg); | |
365 | insert_zspage(page, class, newfg); | |
366 | set_zspage_mapping(page, class_idx, newfg); | |
367 | ||
368 | out: | |
369 | return newfg; | |
370 | } | |
371 | ||
372 | /* | |
373 | * We have to decide on how many pages to link together | |
374 | * to form a zspage for each size class. This is important | |
375 | * to reduce wastage due to unusable space left at end of | |
376 | * each zspage which is given as: | |
377 | * wastage = Zp - Zp % size_class | |
378 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... | |
379 | * | |
380 | * For example, for size class of 3/8 * PAGE_SIZE, we should | |
381 | * link together 3 PAGE_SIZE sized pages to form a zspage | |
382 | * since then we can perfectly fit in 8 such objects. | |
383 | */ | |
2e3b6154 | 384 | static int get_pages_per_zspage(int class_size) |
61989a80 NG |
385 | { |
386 | int i, max_usedpc = 0; | |
387 | /* zspage order which gives maximum used size per KB */ | |
388 | int max_usedpc_order = 1; | |
389 | ||
84d4faab | 390 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
61989a80 NG |
391 | int zspage_size; |
392 | int waste, usedpc; | |
393 | ||
394 | zspage_size = i * PAGE_SIZE; | |
395 | waste = zspage_size % class_size; | |
396 | usedpc = (zspage_size - waste) * 100 / zspage_size; | |
397 | ||
398 | if (usedpc > max_usedpc) { | |
399 | max_usedpc = usedpc; | |
400 | max_usedpc_order = i; | |
401 | } | |
402 | } | |
403 | ||
404 | return max_usedpc_order; | |
405 | } | |
406 | ||
407 | /* | |
408 | * A single 'zspage' is composed of many system pages which are | |
409 | * linked together using fields in struct page. This function finds | |
410 | * the first/head page, given any component page of a zspage. | |
411 | */ | |
412 | static struct page *get_first_page(struct page *page) | |
413 | { | |
414 | if (is_first_page(page)) | |
415 | return page; | |
416 | else | |
417 | return page->first_page; | |
418 | } | |
419 | ||
420 | static struct page *get_next_page(struct page *page) | |
421 | { | |
422 | struct page *next; | |
423 | ||
424 | if (is_last_page(page)) | |
425 | next = NULL; | |
426 | else if (is_first_page(page)) | |
427 | next = (struct page *)page->private; | |
428 | else | |
429 | next = list_entry(page->lru.next, struct page, lru); | |
430 | ||
431 | return next; | |
432 | } | |
433 | ||
434 | /* Encode <page, obj_idx> as a single handle value */ | |
435 | static void *obj_location_to_handle(struct page *page, unsigned long obj_idx) | |
436 | { | |
437 | unsigned long handle; | |
438 | ||
439 | if (!page) { | |
440 | BUG_ON(obj_idx); | |
441 | return NULL; | |
442 | } | |
443 | ||
444 | handle = page_to_pfn(page) << OBJ_INDEX_BITS; | |
445 | handle |= (obj_idx & OBJ_INDEX_MASK); | |
446 | ||
447 | return (void *)handle; | |
448 | } | |
449 | ||
450 | /* Decode <page, obj_idx> pair from the given object handle */ | |
c2344348 | 451 | static void obj_handle_to_location(unsigned long handle, struct page **page, |
61989a80 NG |
452 | unsigned long *obj_idx) |
453 | { | |
c2344348 MK |
454 | *page = pfn_to_page(handle >> OBJ_INDEX_BITS); |
455 | *obj_idx = handle & OBJ_INDEX_MASK; | |
61989a80 NG |
456 | } |
457 | ||
458 | static unsigned long obj_idx_to_offset(struct page *page, | |
459 | unsigned long obj_idx, int class_size) | |
460 | { | |
461 | unsigned long off = 0; | |
462 | ||
463 | if (!is_first_page(page)) | |
464 | off = page->index; | |
465 | ||
466 | return off + obj_idx * class_size; | |
467 | } | |
468 | ||
f4477e90 NG |
469 | static void reset_page(struct page *page) |
470 | { | |
471 | clear_bit(PG_private, &page->flags); | |
472 | clear_bit(PG_private_2, &page->flags); | |
473 | set_page_private(page, 0); | |
474 | page->mapping = NULL; | |
475 | page->freelist = NULL; | |
476 | reset_page_mapcount(page); | |
477 | } | |
478 | ||
61989a80 NG |
479 | static void free_zspage(struct page *first_page) |
480 | { | |
f4477e90 | 481 | struct page *nextp, *tmp, *head_extra; |
61989a80 NG |
482 | |
483 | BUG_ON(!is_first_page(first_page)); | |
484 | BUG_ON(first_page->inuse); | |
485 | ||
f4477e90 | 486 | head_extra = (struct page *)page_private(first_page); |
61989a80 | 487 | |
f4477e90 | 488 | reset_page(first_page); |
61989a80 NG |
489 | __free_page(first_page); |
490 | ||
491 | /* zspage with only 1 system page */ | |
f4477e90 | 492 | if (!head_extra) |
61989a80 NG |
493 | return; |
494 | ||
f4477e90 | 495 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
61989a80 | 496 | list_del(&nextp->lru); |
f4477e90 | 497 | reset_page(nextp); |
61989a80 NG |
498 | __free_page(nextp); |
499 | } | |
f4477e90 NG |
500 | reset_page(head_extra); |
501 | __free_page(head_extra); | |
61989a80 NG |
502 | } |
503 | ||
504 | /* Initialize a newly allocated zspage */ | |
505 | static void init_zspage(struct page *first_page, struct size_class *class) | |
506 | { | |
507 | unsigned long off = 0; | |
508 | struct page *page = first_page; | |
509 | ||
510 | BUG_ON(!is_first_page(first_page)); | |
511 | while (page) { | |
512 | struct page *next_page; | |
513 | struct link_free *link; | |
514 | unsigned int i, objs_on_page; | |
515 | ||
516 | /* | |
517 | * page->index stores offset of first object starting | |
518 | * in the page. For the first page, this is always 0, | |
519 | * so we use first_page->index (aka ->freelist) to store | |
520 | * head of corresponding zspage's freelist. | |
521 | */ | |
522 | if (page != first_page) | |
523 | page->index = off; | |
524 | ||
525 | link = (struct link_free *)kmap_atomic(page) + | |
526 | off / sizeof(*link); | |
527 | objs_on_page = (PAGE_SIZE - off) / class->size; | |
528 | ||
529 | for (i = 1; i <= objs_on_page; i++) { | |
530 | off += class->size; | |
531 | if (off < PAGE_SIZE) { | |
532 | link->next = obj_location_to_handle(page, i); | |
533 | link += class->size / sizeof(*link); | |
534 | } | |
535 | } | |
536 | ||
537 | /* | |
538 | * We now come to the last (full or partial) object on this | |
539 | * page, which must point to the first object on the next | |
540 | * page (if present) | |
541 | */ | |
542 | next_page = get_next_page(page); | |
543 | link->next = obj_location_to_handle(next_page, 0); | |
544 | kunmap_atomic(link); | |
545 | page = next_page; | |
546 | off = (off + class->size) % PAGE_SIZE; | |
547 | } | |
548 | } | |
549 | ||
550 | /* | |
551 | * Allocate a zspage for the given size class | |
552 | */ | |
553 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) | |
554 | { | |
555 | int i, error; | |
b4b700c5 | 556 | struct page *first_page = NULL, *uninitialized_var(prev_page); |
61989a80 NG |
557 | |
558 | /* | |
559 | * Allocate individual pages and link them together as: | |
560 | * 1. first page->private = first sub-page | |
561 | * 2. all sub-pages are linked together using page->lru | |
562 | * 3. each sub-page is linked to the first page using page->first_page | |
563 | * | |
564 | * For each size class, First/Head pages are linked together using | |
565 | * page->lru. Also, we set PG_private to identify the first page | |
566 | * (i.e. no other sub-page has this flag set) and PG_private_2 to | |
567 | * identify the last page. | |
568 | */ | |
569 | error = -ENOMEM; | |
2e3b6154 | 570 | for (i = 0; i < class->pages_per_zspage; i++) { |
b4b700c5 | 571 | struct page *page; |
61989a80 NG |
572 | |
573 | page = alloc_page(flags); | |
574 | if (!page) | |
575 | goto cleanup; | |
576 | ||
577 | INIT_LIST_HEAD(&page->lru); | |
578 | if (i == 0) { /* first page */ | |
a27545bf | 579 | SetPagePrivate(page); |
61989a80 NG |
580 | set_page_private(page, 0); |
581 | first_page = page; | |
582 | first_page->inuse = 0; | |
583 | } | |
584 | if (i == 1) | |
585 | first_page->private = (unsigned long)page; | |
586 | if (i >= 1) | |
587 | page->first_page = first_page; | |
588 | if (i >= 2) | |
589 | list_add(&page->lru, &prev_page->lru); | |
2e3b6154 | 590 | if (i == class->pages_per_zspage - 1) /* last page */ |
a27545bf | 591 | SetPagePrivate2(page); |
61989a80 NG |
592 | prev_page = page; |
593 | } | |
594 | ||
595 | init_zspage(first_page, class); | |
596 | ||
597 | first_page->freelist = obj_location_to_handle(first_page, 0); | |
598 | /* Maximum number of objects we can store in this zspage */ | |
2e3b6154 | 599 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
61989a80 NG |
600 | |
601 | error = 0; /* Success */ | |
602 | ||
603 | cleanup: | |
604 | if (unlikely(error) && first_page) { | |
605 | free_zspage(first_page); | |
606 | first_page = NULL; | |
607 | } | |
608 | ||
609 | return first_page; | |
610 | } | |
611 | ||
612 | static struct page *find_get_zspage(struct size_class *class) | |
613 | { | |
614 | int i; | |
615 | struct page *page; | |
616 | ||
617 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { | |
618 | page = class->fullness_list[i]; | |
619 | if (page) | |
620 | break; | |
621 | } | |
622 | ||
623 | return page; | |
624 | } | |
625 | ||
f553646a SJ |
626 | #ifdef USE_PGTABLE_MAPPING |
627 | static inline int __zs_cpu_up(struct mapping_area *area) | |
628 | { | |
629 | /* | |
630 | * Make sure we don't leak memory if a cpu UP notification | |
631 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
632 | */ | |
633 | if (area->vm) | |
634 | return 0; | |
635 | area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL); | |
636 | if (!area->vm) | |
637 | return -ENOMEM; | |
638 | return 0; | |
639 | } | |
640 | ||
641 | static inline void __zs_cpu_down(struct mapping_area *area) | |
642 | { | |
643 | if (area->vm) | |
644 | free_vm_area(area->vm); | |
645 | area->vm = NULL; | |
646 | } | |
647 | ||
648 | static inline void *__zs_map_object(struct mapping_area *area, | |
649 | struct page *pages[2], int off, int size) | |
650 | { | |
651 | BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, &pages)); | |
652 | area->vm_addr = area->vm->addr; | |
653 | return area->vm_addr + off; | |
654 | } | |
655 | ||
656 | static inline void __zs_unmap_object(struct mapping_area *area, | |
657 | struct page *pages[2], int off, int size) | |
658 | { | |
659 | unsigned long addr = (unsigned long)area->vm_addr; | |
660 | unsigned long end = addr + (PAGE_SIZE * 2); | |
661 | ||
662 | flush_cache_vunmap(addr, end); | |
663 | unmap_kernel_range_noflush(addr, PAGE_SIZE * 2); | |
99155188 | 664 | flush_tlb_kernel_range(addr, end); |
f553646a SJ |
665 | } |
666 | ||
667 | #else /* USE_PGTABLE_MAPPING */ | |
668 | ||
669 | static inline int __zs_cpu_up(struct mapping_area *area) | |
670 | { | |
671 | /* | |
672 | * Make sure we don't leak memory if a cpu UP notification | |
673 | * and zs_init() race and both call zs_cpu_up() on the same cpu | |
674 | */ | |
675 | if (area->vm_buf) | |
676 | return 0; | |
677 | area->vm_buf = (char *)__get_free_page(GFP_KERNEL); | |
678 | if (!area->vm_buf) | |
679 | return -ENOMEM; | |
680 | return 0; | |
681 | } | |
682 | ||
683 | static inline void __zs_cpu_down(struct mapping_area *area) | |
684 | { | |
685 | if (area->vm_buf) | |
686 | free_page((unsigned long)area->vm_buf); | |
687 | area->vm_buf = NULL; | |
688 | } | |
689 | ||
690 | static void *__zs_map_object(struct mapping_area *area, | |
691 | struct page *pages[2], int off, int size) | |
5f601902 | 692 | { |
5f601902 SJ |
693 | int sizes[2]; |
694 | void *addr; | |
f553646a | 695 | char *buf = area->vm_buf; |
5f601902 | 696 | |
f553646a SJ |
697 | /* disable page faults to match kmap_atomic() return conditions */ |
698 | pagefault_disable(); | |
699 | ||
700 | /* no read fastpath */ | |
701 | if (area->vm_mm == ZS_MM_WO) | |
702 | goto out; | |
5f601902 SJ |
703 | |
704 | sizes[0] = PAGE_SIZE - off; | |
705 | sizes[1] = size - sizes[0]; | |
706 | ||
5f601902 SJ |
707 | /* copy object to per-cpu buffer */ |
708 | addr = kmap_atomic(pages[0]); | |
709 | memcpy(buf, addr + off, sizes[0]); | |
710 | kunmap_atomic(addr); | |
711 | addr = kmap_atomic(pages[1]); | |
712 | memcpy(buf + sizes[0], addr, sizes[1]); | |
713 | kunmap_atomic(addr); | |
f553646a SJ |
714 | out: |
715 | return area->vm_buf; | |
5f601902 SJ |
716 | } |
717 | ||
f553646a SJ |
718 | static void __zs_unmap_object(struct mapping_area *area, |
719 | struct page *pages[2], int off, int size) | |
5f601902 | 720 | { |
5f601902 SJ |
721 | int sizes[2]; |
722 | void *addr; | |
f553646a | 723 | char *buf = area->vm_buf; |
5f601902 | 724 | |
f553646a SJ |
725 | /* no write fastpath */ |
726 | if (area->vm_mm == ZS_MM_RO) | |
727 | goto out; | |
5f601902 SJ |
728 | |
729 | sizes[0] = PAGE_SIZE - off; | |
730 | sizes[1] = size - sizes[0]; | |
731 | ||
732 | /* copy per-cpu buffer to object */ | |
733 | addr = kmap_atomic(pages[0]); | |
734 | memcpy(addr + off, buf, sizes[0]); | |
735 | kunmap_atomic(addr); | |
736 | addr = kmap_atomic(pages[1]); | |
737 | memcpy(addr, buf + sizes[0], sizes[1]); | |
738 | kunmap_atomic(addr); | |
f553646a SJ |
739 | |
740 | out: | |
741 | /* enable page faults to match kunmap_atomic() return conditions */ | |
742 | pagefault_enable(); | |
5f601902 | 743 | } |
61989a80 | 744 | |
f553646a SJ |
745 | #endif /* USE_PGTABLE_MAPPING */ |
746 | ||
61989a80 NG |
747 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
748 | void *pcpu) | |
749 | { | |
f553646a | 750 | int ret, cpu = (long)pcpu; |
61989a80 NG |
751 | struct mapping_area *area; |
752 | ||
753 | switch (action) { | |
754 | case CPU_UP_PREPARE: | |
755 | area = &per_cpu(zs_map_area, cpu); | |
f553646a SJ |
756 | ret = __zs_cpu_up(area); |
757 | if (ret) | |
758 | return notifier_from_errno(ret); | |
61989a80 NG |
759 | break; |
760 | case CPU_DEAD: | |
761 | case CPU_UP_CANCELED: | |
762 | area = &per_cpu(zs_map_area, cpu); | |
f553646a | 763 | __zs_cpu_down(area); |
61989a80 NG |
764 | break; |
765 | } | |
766 | ||
767 | return NOTIFY_OK; | |
768 | } | |
769 | ||
770 | static struct notifier_block zs_cpu_nb = { | |
771 | .notifier_call = zs_cpu_notifier | |
772 | }; | |
773 | ||
774 | static void zs_exit(void) | |
775 | { | |
776 | int cpu; | |
777 | ||
778 | for_each_online_cpu(cpu) | |
779 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); | |
780 | unregister_cpu_notifier(&zs_cpu_nb); | |
781 | } | |
782 | ||
783 | static int zs_init(void) | |
784 | { | |
785 | int cpu, ret; | |
786 | ||
787 | register_cpu_notifier(&zs_cpu_nb); | |
788 | for_each_online_cpu(cpu) { | |
789 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); | |
790 | if (notifier_to_errno(ret)) | |
791 | goto fail; | |
792 | } | |
793 | return 0; | |
794 | fail: | |
795 | zs_exit(); | |
796 | return notifier_to_errno(ret); | |
797 | } | |
798 | ||
4bbc0bc0 DB |
799 | /** |
800 | * zs_create_pool - Creates an allocation pool to work from. | |
801 | * @name: name of the pool to be created | |
802 | * @flags: allocation flags used when growing pool | |
803 | * | |
804 | * This function must be called before anything when using | |
805 | * the zsmalloc allocator. | |
806 | * | |
807 | * On success, a pointer to the newly created pool is returned, | |
808 | * otherwise NULL. | |
809 | */ | |
61989a80 NG |
810 | struct zs_pool *zs_create_pool(const char *name, gfp_t flags) |
811 | { | |
069f101f | 812 | int i, ovhd_size; |
61989a80 NG |
813 | struct zs_pool *pool; |
814 | ||
815 | if (!name) | |
816 | return NULL; | |
817 | ||
818 | ovhd_size = roundup(sizeof(*pool), PAGE_SIZE); | |
819 | pool = kzalloc(ovhd_size, GFP_KERNEL); | |
820 | if (!pool) | |
821 | return NULL; | |
822 | ||
823 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { | |
824 | int size; | |
825 | struct size_class *class; | |
826 | ||
827 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; | |
828 | if (size > ZS_MAX_ALLOC_SIZE) | |
829 | size = ZS_MAX_ALLOC_SIZE; | |
830 | ||
831 | class = &pool->size_class[i]; | |
832 | class->size = size; | |
833 | class->index = i; | |
834 | spin_lock_init(&class->lock); | |
2e3b6154 | 835 | class->pages_per_zspage = get_pages_per_zspage(size); |
61989a80 NG |
836 | |
837 | } | |
838 | ||
61989a80 NG |
839 | pool->flags = flags; |
840 | pool->name = name; | |
841 | ||
61989a80 NG |
842 | return pool; |
843 | } | |
844 | EXPORT_SYMBOL_GPL(zs_create_pool); | |
845 | ||
846 | void zs_destroy_pool(struct zs_pool *pool) | |
847 | { | |
848 | int i; | |
849 | ||
850 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { | |
851 | int fg; | |
852 | struct size_class *class = &pool->size_class[i]; | |
853 | ||
854 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { | |
855 | if (class->fullness_list[fg]) { | |
856 | pr_info("Freeing non-empty class with size " | |
857 | "%db, fullness group %d\n", | |
858 | class->size, fg); | |
859 | } | |
860 | } | |
861 | } | |
862 | kfree(pool); | |
863 | } | |
864 | EXPORT_SYMBOL_GPL(zs_destroy_pool); | |
865 | ||
866 | /** | |
867 | * zs_malloc - Allocate block of given size from pool. | |
868 | * @pool: pool to allocate from | |
869 | * @size: size of block to allocate | |
61989a80 | 870 | * |
00a61d86 | 871 | * On success, handle to the allocated object is returned, |
c2344348 | 872 | * otherwise 0. |
61989a80 NG |
873 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
874 | */ | |
c2344348 | 875 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
61989a80 | 876 | { |
c2344348 | 877 | unsigned long obj; |
61989a80 NG |
878 | struct link_free *link; |
879 | int class_idx; | |
880 | struct size_class *class; | |
881 | ||
882 | struct page *first_page, *m_page; | |
883 | unsigned long m_objidx, m_offset; | |
884 | ||
885 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) | |
c2344348 | 886 | return 0; |
61989a80 NG |
887 | |
888 | class_idx = get_size_class_index(size); | |
889 | class = &pool->size_class[class_idx]; | |
890 | BUG_ON(class_idx != class->index); | |
891 | ||
892 | spin_lock(&class->lock); | |
893 | first_page = find_get_zspage(class); | |
894 | ||
895 | if (!first_page) { | |
896 | spin_unlock(&class->lock); | |
897 | first_page = alloc_zspage(class, pool->flags); | |
898 | if (unlikely(!first_page)) | |
c2344348 | 899 | return 0; |
61989a80 NG |
900 | |
901 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); | |
902 | spin_lock(&class->lock); | |
2e3b6154 | 903 | class->pages_allocated += class->pages_per_zspage; |
61989a80 NG |
904 | } |
905 | ||
c2344348 | 906 | obj = (unsigned long)first_page->freelist; |
61989a80 NG |
907 | obj_handle_to_location(obj, &m_page, &m_objidx); |
908 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); | |
909 | ||
910 | link = (struct link_free *)kmap_atomic(m_page) + | |
911 | m_offset / sizeof(*link); | |
912 | first_page->freelist = link->next; | |
913 | memset(link, POISON_INUSE, sizeof(*link)); | |
914 | kunmap_atomic(link); | |
915 | ||
916 | first_page->inuse++; | |
917 | /* Now move the zspage to another fullness group, if required */ | |
918 | fix_fullness_group(pool, first_page); | |
919 | spin_unlock(&class->lock); | |
920 | ||
921 | return obj; | |
922 | } | |
923 | EXPORT_SYMBOL_GPL(zs_malloc); | |
924 | ||
c2344348 | 925 | void zs_free(struct zs_pool *pool, unsigned long obj) |
61989a80 NG |
926 | { |
927 | struct link_free *link; | |
928 | struct page *first_page, *f_page; | |
929 | unsigned long f_objidx, f_offset; | |
930 | ||
931 | int class_idx; | |
932 | struct size_class *class; | |
933 | enum fullness_group fullness; | |
934 | ||
935 | if (unlikely(!obj)) | |
936 | return; | |
937 | ||
938 | obj_handle_to_location(obj, &f_page, &f_objidx); | |
939 | first_page = get_first_page(f_page); | |
940 | ||
941 | get_zspage_mapping(first_page, &class_idx, &fullness); | |
942 | class = &pool->size_class[class_idx]; | |
943 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); | |
944 | ||
945 | spin_lock(&class->lock); | |
946 | ||
947 | /* Insert this object in containing zspage's freelist */ | |
948 | link = (struct link_free *)((unsigned char *)kmap_atomic(f_page) | |
949 | + f_offset); | |
950 | link->next = first_page->freelist; | |
951 | kunmap_atomic(link); | |
c2344348 | 952 | first_page->freelist = (void *)obj; |
61989a80 NG |
953 | |
954 | first_page->inuse--; | |
955 | fullness = fix_fullness_group(pool, first_page); | |
956 | ||
957 | if (fullness == ZS_EMPTY) | |
2e3b6154 | 958 | class->pages_allocated -= class->pages_per_zspage; |
61989a80 NG |
959 | |
960 | spin_unlock(&class->lock); | |
961 | ||
962 | if (fullness == ZS_EMPTY) | |
963 | free_zspage(first_page); | |
964 | } | |
965 | EXPORT_SYMBOL_GPL(zs_free); | |
966 | ||
00a61d86 MK |
967 | /** |
968 | * zs_map_object - get address of allocated object from handle. | |
969 | * @pool: pool from which the object was allocated | |
970 | * @handle: handle returned from zs_malloc | |
971 | * | |
972 | * Before using an object allocated from zs_malloc, it must be mapped using | |
973 | * this function. When done with the object, it must be unmapped using | |
166cfda7 SJ |
974 | * zs_unmap_object. |
975 | * | |
976 | * Only one object can be mapped per cpu at a time. There is no protection | |
977 | * against nested mappings. | |
978 | * | |
979 | * This function returns with preemption and page faults disabled. | |
00a61d86 | 980 | */ |
b7418510 SJ |
981 | void *zs_map_object(struct zs_pool *pool, unsigned long handle, |
982 | enum zs_mapmode mm) | |
61989a80 NG |
983 | { |
984 | struct page *page; | |
985 | unsigned long obj_idx, off; | |
986 | ||
987 | unsigned int class_idx; | |
988 | enum fullness_group fg; | |
989 | struct size_class *class; | |
990 | struct mapping_area *area; | |
f553646a | 991 | struct page *pages[2]; |
61989a80 NG |
992 | |
993 | BUG_ON(!handle); | |
994 | ||
c60369f0 SJ |
995 | /* |
996 | * Because we use per-cpu mapping areas shared among the | |
997 | * pools/users, we can't allow mapping in interrupt context | |
998 | * because it can corrupt another users mappings. | |
999 | */ | |
1000 | BUG_ON(in_interrupt()); | |
1001 | ||
61989a80 NG |
1002 | obj_handle_to_location(handle, &page, &obj_idx); |
1003 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); | |
1004 | class = &pool->size_class[class_idx]; | |
1005 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
1006 | ||
1007 | area = &get_cpu_var(zs_map_area); | |
f553646a | 1008 | area->vm_mm = mm; |
61989a80 NG |
1009 | if (off + class->size <= PAGE_SIZE) { |
1010 | /* this object is contained entirely within a page */ | |
1011 | area->vm_addr = kmap_atomic(page); | |
5f601902 | 1012 | return area->vm_addr + off; |
61989a80 NG |
1013 | } |
1014 | ||
f553646a SJ |
1015 | /* this object spans two pages */ |
1016 | pages[0] = page; | |
1017 | pages[1] = get_next_page(page); | |
1018 | BUG_ON(!pages[1]); | |
b7418510 | 1019 | |
f553646a | 1020 | return __zs_map_object(area, pages, off, class->size); |
61989a80 NG |
1021 | } |
1022 | EXPORT_SYMBOL_GPL(zs_map_object); | |
1023 | ||
c2344348 | 1024 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
61989a80 NG |
1025 | { |
1026 | struct page *page; | |
1027 | unsigned long obj_idx, off; | |
1028 | ||
1029 | unsigned int class_idx; | |
1030 | enum fullness_group fg; | |
1031 | struct size_class *class; | |
1032 | struct mapping_area *area; | |
1033 | ||
1034 | BUG_ON(!handle); | |
1035 | ||
1036 | obj_handle_to_location(handle, &page, &obj_idx); | |
1037 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); | |
1038 | class = &pool->size_class[class_idx]; | |
1039 | off = obj_idx_to_offset(page, obj_idx, class->size); | |
1040 | ||
f553646a SJ |
1041 | area = &__get_cpu_var(zs_map_area); |
1042 | if (off + class->size <= PAGE_SIZE) | |
1043 | kunmap_atomic(area->vm_addr); | |
1044 | else { | |
1045 | struct page *pages[2]; | |
1046 | ||
1047 | pages[0] = page; | |
1048 | pages[1] = get_next_page(page); | |
1049 | BUG_ON(!pages[1]); | |
b7418510 | 1050 | |
f553646a SJ |
1051 | __zs_unmap_object(area, pages, off, class->size); |
1052 | } | |
61989a80 NG |
1053 | put_cpu_var(zs_map_area); |
1054 | } | |
1055 | EXPORT_SYMBOL_GPL(zs_unmap_object); | |
1056 | ||
1057 | u64 zs_get_total_size_bytes(struct zs_pool *pool) | |
1058 | { | |
1059 | int i; | |
1060 | u64 npages = 0; | |
1061 | ||
1062 | for (i = 0; i < ZS_SIZE_CLASSES; i++) | |
1063 | npages += pool->size_class[i].pages_allocated; | |
1064 | ||
1065 | return npages << PAGE_SHIFT; | |
1066 | } | |
1067 | EXPORT_SYMBOL_GPL(zs_get_total_size_bytes); | |
069f101f BH |
1068 | |
1069 | module_init(zs_init); | |
1070 | module_exit(zs_exit); | |
1071 | ||
1072 | MODULE_LICENSE("Dual BSD/GPL"); | |
1073 | MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>"); |