Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 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 | |
Nitin Gupta | 2db51da | 2012-06-09 17:41:14 -0700 | [diff] [blame^] | 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 | |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 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> |
Seth Jennings | 0cbb613 | 2012-02-13 08:47:49 -0600 | [diff] [blame] | 77 | #include <linux/vmalloc.h> |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 78 | |
| 79 | #include "zsmalloc.h" |
| 80 | #include "zsmalloc_int.h" |
| 81 | |
| 82 | /* |
| 83 | * A zspage's class index and fullness group |
| 84 | * are encoded in its (first)page->mapping |
| 85 | */ |
| 86 | #define CLASS_IDX_BITS 28 |
| 87 | #define FULLNESS_BITS 4 |
| 88 | #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1) |
| 89 | #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1) |
| 90 | |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 91 | /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */ |
| 92 | static DEFINE_PER_CPU(struct mapping_area, zs_map_area); |
| 93 | |
| 94 | static int is_first_page(struct page *page) |
| 95 | { |
Minchan Kim | a27545bf | 2012-04-25 15:23:09 +0900 | [diff] [blame] | 96 | return PagePrivate(page); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 97 | } |
| 98 | |
| 99 | static int is_last_page(struct page *page) |
| 100 | { |
Minchan Kim | a27545bf | 2012-04-25 15:23:09 +0900 | [diff] [blame] | 101 | return PagePrivate2(page); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 102 | } |
| 103 | |
| 104 | static void get_zspage_mapping(struct page *page, unsigned int *class_idx, |
| 105 | enum fullness_group *fullness) |
| 106 | { |
| 107 | unsigned long m; |
| 108 | BUG_ON(!is_first_page(page)); |
| 109 | |
| 110 | m = (unsigned long)page->mapping; |
| 111 | *fullness = m & FULLNESS_MASK; |
| 112 | *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK; |
| 113 | } |
| 114 | |
| 115 | static void set_zspage_mapping(struct page *page, unsigned int class_idx, |
| 116 | enum fullness_group fullness) |
| 117 | { |
| 118 | unsigned long m; |
| 119 | BUG_ON(!is_first_page(page)); |
| 120 | |
| 121 | m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) | |
| 122 | (fullness & FULLNESS_MASK); |
| 123 | page->mapping = (struct address_space *)m; |
| 124 | } |
| 125 | |
| 126 | static int get_size_class_index(int size) |
| 127 | { |
| 128 | int idx = 0; |
| 129 | |
| 130 | if (likely(size > ZS_MIN_ALLOC_SIZE)) |
| 131 | idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE, |
| 132 | ZS_SIZE_CLASS_DELTA); |
| 133 | |
| 134 | return idx; |
| 135 | } |
| 136 | |
| 137 | static enum fullness_group get_fullness_group(struct page *page) |
| 138 | { |
| 139 | int inuse, max_objects; |
| 140 | enum fullness_group fg; |
| 141 | BUG_ON(!is_first_page(page)); |
| 142 | |
| 143 | inuse = page->inuse; |
| 144 | max_objects = page->objects; |
| 145 | |
| 146 | if (inuse == 0) |
| 147 | fg = ZS_EMPTY; |
| 148 | else if (inuse == max_objects) |
| 149 | fg = ZS_FULL; |
| 150 | else if (inuse <= max_objects / fullness_threshold_frac) |
| 151 | fg = ZS_ALMOST_EMPTY; |
| 152 | else |
| 153 | fg = ZS_ALMOST_FULL; |
| 154 | |
| 155 | return fg; |
| 156 | } |
| 157 | |
| 158 | static void insert_zspage(struct page *page, struct size_class *class, |
| 159 | enum fullness_group fullness) |
| 160 | { |
| 161 | struct page **head; |
| 162 | |
| 163 | BUG_ON(!is_first_page(page)); |
| 164 | |
| 165 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) |
| 166 | return; |
| 167 | |
| 168 | head = &class->fullness_list[fullness]; |
| 169 | if (*head) |
| 170 | list_add_tail(&page->lru, &(*head)->lru); |
| 171 | |
| 172 | *head = page; |
| 173 | } |
| 174 | |
| 175 | static void remove_zspage(struct page *page, struct size_class *class, |
| 176 | enum fullness_group fullness) |
| 177 | { |
| 178 | struct page **head; |
| 179 | |
| 180 | BUG_ON(!is_first_page(page)); |
| 181 | |
| 182 | if (fullness >= _ZS_NR_FULLNESS_GROUPS) |
| 183 | return; |
| 184 | |
| 185 | head = &class->fullness_list[fullness]; |
| 186 | BUG_ON(!*head); |
| 187 | if (list_empty(&(*head)->lru)) |
| 188 | *head = NULL; |
| 189 | else if (*head == page) |
| 190 | *head = (struct page *)list_entry((*head)->lru.next, |
| 191 | struct page, lru); |
| 192 | |
| 193 | list_del_init(&page->lru); |
| 194 | } |
| 195 | |
| 196 | static enum fullness_group fix_fullness_group(struct zs_pool *pool, |
| 197 | struct page *page) |
| 198 | { |
| 199 | int class_idx; |
| 200 | struct size_class *class; |
| 201 | enum fullness_group currfg, newfg; |
| 202 | |
| 203 | BUG_ON(!is_first_page(page)); |
| 204 | |
| 205 | get_zspage_mapping(page, &class_idx, &currfg); |
| 206 | newfg = get_fullness_group(page); |
| 207 | if (newfg == currfg) |
| 208 | goto out; |
| 209 | |
| 210 | class = &pool->size_class[class_idx]; |
| 211 | remove_zspage(page, class, currfg); |
| 212 | insert_zspage(page, class, newfg); |
| 213 | set_zspage_mapping(page, class_idx, newfg); |
| 214 | |
| 215 | out: |
| 216 | return newfg; |
| 217 | } |
| 218 | |
| 219 | /* |
| 220 | * We have to decide on how many pages to link together |
| 221 | * to form a zspage for each size class. This is important |
| 222 | * to reduce wastage due to unusable space left at end of |
| 223 | * each zspage which is given as: |
| 224 | * wastage = Zp - Zp % size_class |
| 225 | * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ... |
| 226 | * |
| 227 | * For example, for size class of 3/8 * PAGE_SIZE, we should |
| 228 | * link together 3 PAGE_SIZE sized pages to form a zspage |
| 229 | * since then we can perfectly fit in 8 such objects. |
| 230 | */ |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 231 | static int get_pages_per_zspage(int class_size) |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 232 | { |
| 233 | int i, max_usedpc = 0; |
| 234 | /* zspage order which gives maximum used size per KB */ |
| 235 | int max_usedpc_order = 1; |
| 236 | |
Seth Jennings | 84d4faa | 2012-03-05 11:33:21 -0600 | [diff] [blame] | 237 | for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) { |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 238 | int zspage_size; |
| 239 | int waste, usedpc; |
| 240 | |
| 241 | zspage_size = i * PAGE_SIZE; |
| 242 | waste = zspage_size % class_size; |
| 243 | usedpc = (zspage_size - waste) * 100 / zspage_size; |
| 244 | |
| 245 | if (usedpc > max_usedpc) { |
| 246 | max_usedpc = usedpc; |
| 247 | max_usedpc_order = i; |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | return max_usedpc_order; |
| 252 | } |
| 253 | |
| 254 | /* |
| 255 | * A single 'zspage' is composed of many system pages which are |
| 256 | * linked together using fields in struct page. This function finds |
| 257 | * the first/head page, given any component page of a zspage. |
| 258 | */ |
| 259 | static struct page *get_first_page(struct page *page) |
| 260 | { |
| 261 | if (is_first_page(page)) |
| 262 | return page; |
| 263 | else |
| 264 | return page->first_page; |
| 265 | } |
| 266 | |
| 267 | static struct page *get_next_page(struct page *page) |
| 268 | { |
| 269 | struct page *next; |
| 270 | |
| 271 | if (is_last_page(page)) |
| 272 | next = NULL; |
| 273 | else if (is_first_page(page)) |
| 274 | next = (struct page *)page->private; |
| 275 | else |
| 276 | next = list_entry(page->lru.next, struct page, lru); |
| 277 | |
| 278 | return next; |
| 279 | } |
| 280 | |
| 281 | /* Encode <page, obj_idx> as a single handle value */ |
| 282 | static void *obj_location_to_handle(struct page *page, unsigned long obj_idx) |
| 283 | { |
| 284 | unsigned long handle; |
| 285 | |
| 286 | if (!page) { |
| 287 | BUG_ON(obj_idx); |
| 288 | return NULL; |
| 289 | } |
| 290 | |
| 291 | handle = page_to_pfn(page) << OBJ_INDEX_BITS; |
| 292 | handle |= (obj_idx & OBJ_INDEX_MASK); |
| 293 | |
| 294 | return (void *)handle; |
| 295 | } |
| 296 | |
| 297 | /* Decode <page, obj_idx> pair from the given object handle */ |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 298 | static void obj_handle_to_location(unsigned long handle, struct page **page, |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 299 | unsigned long *obj_idx) |
| 300 | { |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 301 | *page = pfn_to_page(handle >> OBJ_INDEX_BITS); |
| 302 | *obj_idx = handle & OBJ_INDEX_MASK; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 303 | } |
| 304 | |
| 305 | static unsigned long obj_idx_to_offset(struct page *page, |
| 306 | unsigned long obj_idx, int class_size) |
| 307 | { |
| 308 | unsigned long off = 0; |
| 309 | |
| 310 | if (!is_first_page(page)) |
| 311 | off = page->index; |
| 312 | |
| 313 | return off + obj_idx * class_size; |
| 314 | } |
| 315 | |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 316 | static void reset_page(struct page *page) |
| 317 | { |
| 318 | clear_bit(PG_private, &page->flags); |
| 319 | clear_bit(PG_private_2, &page->flags); |
| 320 | set_page_private(page, 0); |
| 321 | page->mapping = NULL; |
| 322 | page->freelist = NULL; |
| 323 | reset_page_mapcount(page); |
| 324 | } |
| 325 | |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 326 | static void free_zspage(struct page *first_page) |
| 327 | { |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 328 | struct page *nextp, *tmp, *head_extra; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 329 | |
| 330 | BUG_ON(!is_first_page(first_page)); |
| 331 | BUG_ON(first_page->inuse); |
| 332 | |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 333 | head_extra = (struct page *)page_private(first_page); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 334 | |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 335 | reset_page(first_page); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 336 | __free_page(first_page); |
| 337 | |
| 338 | /* zspage with only 1 system page */ |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 339 | if (!head_extra) |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 340 | return; |
| 341 | |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 342 | list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) { |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 343 | list_del(&nextp->lru); |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 344 | reset_page(nextp); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 345 | __free_page(nextp); |
| 346 | } |
Nitin Gupta | f4477e9 | 2012-04-02 09:13:56 -0500 | [diff] [blame] | 347 | reset_page(head_extra); |
| 348 | __free_page(head_extra); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 349 | } |
| 350 | |
| 351 | /* Initialize a newly allocated zspage */ |
| 352 | static void init_zspage(struct page *first_page, struct size_class *class) |
| 353 | { |
| 354 | unsigned long off = 0; |
| 355 | struct page *page = first_page; |
| 356 | |
| 357 | BUG_ON(!is_first_page(first_page)); |
| 358 | while (page) { |
| 359 | struct page *next_page; |
| 360 | struct link_free *link; |
| 361 | unsigned int i, objs_on_page; |
| 362 | |
| 363 | /* |
| 364 | * page->index stores offset of first object starting |
| 365 | * in the page. For the first page, this is always 0, |
| 366 | * so we use first_page->index (aka ->freelist) to store |
| 367 | * head of corresponding zspage's freelist. |
| 368 | */ |
| 369 | if (page != first_page) |
| 370 | page->index = off; |
| 371 | |
| 372 | link = (struct link_free *)kmap_atomic(page) + |
| 373 | off / sizeof(*link); |
| 374 | objs_on_page = (PAGE_SIZE - off) / class->size; |
| 375 | |
| 376 | for (i = 1; i <= objs_on_page; i++) { |
| 377 | off += class->size; |
| 378 | if (off < PAGE_SIZE) { |
| 379 | link->next = obj_location_to_handle(page, i); |
| 380 | link += class->size / sizeof(*link); |
| 381 | } |
| 382 | } |
| 383 | |
| 384 | /* |
| 385 | * We now come to the last (full or partial) object on this |
| 386 | * page, which must point to the first object on the next |
| 387 | * page (if present) |
| 388 | */ |
| 389 | next_page = get_next_page(page); |
| 390 | link->next = obj_location_to_handle(next_page, 0); |
| 391 | kunmap_atomic(link); |
| 392 | page = next_page; |
| 393 | off = (off + class->size) % PAGE_SIZE; |
| 394 | } |
| 395 | } |
| 396 | |
| 397 | /* |
| 398 | * Allocate a zspage for the given size class |
| 399 | */ |
| 400 | static struct page *alloc_zspage(struct size_class *class, gfp_t flags) |
| 401 | { |
| 402 | int i, error; |
| 403 | struct page *first_page = NULL; |
| 404 | |
| 405 | /* |
| 406 | * Allocate individual pages and link them together as: |
| 407 | * 1. first page->private = first sub-page |
| 408 | * 2. all sub-pages are linked together using page->lru |
| 409 | * 3. each sub-page is linked to the first page using page->first_page |
| 410 | * |
| 411 | * For each size class, First/Head pages are linked together using |
| 412 | * page->lru. Also, we set PG_private to identify the first page |
| 413 | * (i.e. no other sub-page has this flag set) and PG_private_2 to |
| 414 | * identify the last page. |
| 415 | */ |
| 416 | error = -ENOMEM; |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 417 | for (i = 0; i < class->pages_per_zspage; i++) { |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 418 | struct page *page, *prev_page; |
| 419 | |
| 420 | page = alloc_page(flags); |
| 421 | if (!page) |
| 422 | goto cleanup; |
| 423 | |
| 424 | INIT_LIST_HEAD(&page->lru); |
| 425 | if (i == 0) { /* first page */ |
Minchan Kim | a27545bf | 2012-04-25 15:23:09 +0900 | [diff] [blame] | 426 | SetPagePrivate(page); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 427 | set_page_private(page, 0); |
| 428 | first_page = page; |
| 429 | first_page->inuse = 0; |
| 430 | } |
| 431 | if (i == 1) |
| 432 | first_page->private = (unsigned long)page; |
| 433 | if (i >= 1) |
| 434 | page->first_page = first_page; |
| 435 | if (i >= 2) |
| 436 | list_add(&page->lru, &prev_page->lru); |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 437 | if (i == class->pages_per_zspage - 1) /* last page */ |
Minchan Kim | a27545bf | 2012-04-25 15:23:09 +0900 | [diff] [blame] | 438 | SetPagePrivate2(page); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 439 | prev_page = page; |
| 440 | } |
| 441 | |
| 442 | init_zspage(first_page, class); |
| 443 | |
| 444 | first_page->freelist = obj_location_to_handle(first_page, 0); |
| 445 | /* Maximum number of objects we can store in this zspage */ |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 446 | first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 447 | |
| 448 | error = 0; /* Success */ |
| 449 | |
| 450 | cleanup: |
| 451 | if (unlikely(error) && first_page) { |
| 452 | free_zspage(first_page); |
| 453 | first_page = NULL; |
| 454 | } |
| 455 | |
| 456 | return first_page; |
| 457 | } |
| 458 | |
| 459 | static struct page *find_get_zspage(struct size_class *class) |
| 460 | { |
| 461 | int i; |
| 462 | struct page *page; |
| 463 | |
| 464 | for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) { |
| 465 | page = class->fullness_list[i]; |
| 466 | if (page) |
| 467 | break; |
| 468 | } |
| 469 | |
| 470 | return page; |
| 471 | } |
| 472 | |
| 473 | |
| 474 | /* |
| 475 | * If this becomes a separate module, register zs_init() with |
| 476 | * module_init(), zs_exit with module_exit(), and remove zs_initialized |
| 477 | */ |
| 478 | static int zs_initialized; |
| 479 | |
| 480 | static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action, |
| 481 | void *pcpu) |
| 482 | { |
| 483 | int cpu = (long)pcpu; |
| 484 | struct mapping_area *area; |
| 485 | |
| 486 | switch (action) { |
| 487 | case CPU_UP_PREPARE: |
| 488 | area = &per_cpu(zs_map_area, cpu); |
| 489 | if (area->vm) |
| 490 | break; |
| 491 | area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes); |
| 492 | if (!area->vm) |
| 493 | return notifier_from_errno(-ENOMEM); |
| 494 | break; |
| 495 | case CPU_DEAD: |
| 496 | case CPU_UP_CANCELED: |
| 497 | area = &per_cpu(zs_map_area, cpu); |
| 498 | if (area->vm) |
| 499 | free_vm_area(area->vm); |
| 500 | area->vm = NULL; |
| 501 | break; |
| 502 | } |
| 503 | |
| 504 | return NOTIFY_OK; |
| 505 | } |
| 506 | |
| 507 | static struct notifier_block zs_cpu_nb = { |
| 508 | .notifier_call = zs_cpu_notifier |
| 509 | }; |
| 510 | |
| 511 | static void zs_exit(void) |
| 512 | { |
| 513 | int cpu; |
| 514 | |
| 515 | for_each_online_cpu(cpu) |
| 516 | zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu); |
| 517 | unregister_cpu_notifier(&zs_cpu_nb); |
| 518 | } |
| 519 | |
| 520 | static int zs_init(void) |
| 521 | { |
| 522 | int cpu, ret; |
| 523 | |
| 524 | register_cpu_notifier(&zs_cpu_nb); |
| 525 | for_each_online_cpu(cpu) { |
| 526 | ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu); |
| 527 | if (notifier_to_errno(ret)) |
| 528 | goto fail; |
| 529 | } |
| 530 | return 0; |
| 531 | fail: |
| 532 | zs_exit(); |
| 533 | return notifier_to_errno(ret); |
| 534 | } |
| 535 | |
| 536 | struct zs_pool *zs_create_pool(const char *name, gfp_t flags) |
| 537 | { |
| 538 | int i, error, ovhd_size; |
| 539 | struct zs_pool *pool; |
| 540 | |
| 541 | if (!name) |
| 542 | return NULL; |
| 543 | |
| 544 | ovhd_size = roundup(sizeof(*pool), PAGE_SIZE); |
| 545 | pool = kzalloc(ovhd_size, GFP_KERNEL); |
| 546 | if (!pool) |
| 547 | return NULL; |
| 548 | |
| 549 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
| 550 | int size; |
| 551 | struct size_class *class; |
| 552 | |
| 553 | size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA; |
| 554 | if (size > ZS_MAX_ALLOC_SIZE) |
| 555 | size = ZS_MAX_ALLOC_SIZE; |
| 556 | |
| 557 | class = &pool->size_class[i]; |
| 558 | class->size = size; |
| 559 | class->index = i; |
| 560 | spin_lock_init(&class->lock); |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 561 | class->pages_per_zspage = get_pages_per_zspage(size); |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 562 | |
| 563 | } |
| 564 | |
| 565 | /* |
| 566 | * If this becomes a separate module, register zs_init with |
| 567 | * module_init, and remove this block |
| 568 | */ |
| 569 | if (!zs_initialized) { |
| 570 | error = zs_init(); |
| 571 | if (error) |
| 572 | goto cleanup; |
| 573 | zs_initialized = 1; |
| 574 | } |
| 575 | |
| 576 | pool->flags = flags; |
| 577 | pool->name = name; |
| 578 | |
| 579 | error = 0; /* Success */ |
| 580 | |
| 581 | cleanup: |
| 582 | if (error) { |
| 583 | zs_destroy_pool(pool); |
| 584 | pool = NULL; |
| 585 | } |
| 586 | |
| 587 | return pool; |
| 588 | } |
| 589 | EXPORT_SYMBOL_GPL(zs_create_pool); |
| 590 | |
| 591 | void zs_destroy_pool(struct zs_pool *pool) |
| 592 | { |
| 593 | int i; |
| 594 | |
| 595 | for (i = 0; i < ZS_SIZE_CLASSES; i++) { |
| 596 | int fg; |
| 597 | struct size_class *class = &pool->size_class[i]; |
| 598 | |
| 599 | for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) { |
| 600 | if (class->fullness_list[fg]) { |
| 601 | pr_info("Freeing non-empty class with size " |
| 602 | "%db, fullness group %d\n", |
| 603 | class->size, fg); |
| 604 | } |
| 605 | } |
| 606 | } |
| 607 | kfree(pool); |
| 608 | } |
| 609 | EXPORT_SYMBOL_GPL(zs_destroy_pool); |
| 610 | |
| 611 | /** |
| 612 | * zs_malloc - Allocate block of given size from pool. |
| 613 | * @pool: pool to allocate from |
| 614 | * @size: size of block to allocate |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 615 | * |
Minchan Kim | 00a61d8 | 2012-05-03 15:40:40 +0900 | [diff] [blame] | 616 | * On success, handle to the allocated object is returned, |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 617 | * otherwise 0. |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 618 | * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail. |
| 619 | */ |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 620 | unsigned long zs_malloc(struct zs_pool *pool, size_t size) |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 621 | { |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 622 | unsigned long obj; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 623 | struct link_free *link; |
| 624 | int class_idx; |
| 625 | struct size_class *class; |
| 626 | |
| 627 | struct page *first_page, *m_page; |
| 628 | unsigned long m_objidx, m_offset; |
| 629 | |
| 630 | if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE)) |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 631 | return 0; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 632 | |
| 633 | class_idx = get_size_class_index(size); |
| 634 | class = &pool->size_class[class_idx]; |
| 635 | BUG_ON(class_idx != class->index); |
| 636 | |
| 637 | spin_lock(&class->lock); |
| 638 | first_page = find_get_zspage(class); |
| 639 | |
| 640 | if (!first_page) { |
| 641 | spin_unlock(&class->lock); |
| 642 | first_page = alloc_zspage(class, pool->flags); |
| 643 | if (unlikely(!first_page)) |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 644 | return 0; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 645 | |
| 646 | set_zspage_mapping(first_page, class->index, ZS_EMPTY); |
| 647 | spin_lock(&class->lock); |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 648 | class->pages_allocated += class->pages_per_zspage; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 649 | } |
| 650 | |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 651 | obj = (unsigned long)first_page->freelist; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 652 | obj_handle_to_location(obj, &m_page, &m_objidx); |
| 653 | m_offset = obj_idx_to_offset(m_page, m_objidx, class->size); |
| 654 | |
| 655 | link = (struct link_free *)kmap_atomic(m_page) + |
| 656 | m_offset / sizeof(*link); |
| 657 | first_page->freelist = link->next; |
| 658 | memset(link, POISON_INUSE, sizeof(*link)); |
| 659 | kunmap_atomic(link); |
| 660 | |
| 661 | first_page->inuse++; |
| 662 | /* Now move the zspage to another fullness group, if required */ |
| 663 | fix_fullness_group(pool, first_page); |
| 664 | spin_unlock(&class->lock); |
| 665 | |
| 666 | return obj; |
| 667 | } |
| 668 | EXPORT_SYMBOL_GPL(zs_malloc); |
| 669 | |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 670 | void zs_free(struct zs_pool *pool, unsigned long obj) |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 671 | { |
| 672 | struct link_free *link; |
| 673 | struct page *first_page, *f_page; |
| 674 | unsigned long f_objidx, f_offset; |
| 675 | |
| 676 | int class_idx; |
| 677 | struct size_class *class; |
| 678 | enum fullness_group fullness; |
| 679 | |
| 680 | if (unlikely(!obj)) |
| 681 | return; |
| 682 | |
| 683 | obj_handle_to_location(obj, &f_page, &f_objidx); |
| 684 | first_page = get_first_page(f_page); |
| 685 | |
| 686 | get_zspage_mapping(first_page, &class_idx, &fullness); |
| 687 | class = &pool->size_class[class_idx]; |
| 688 | f_offset = obj_idx_to_offset(f_page, f_objidx, class->size); |
| 689 | |
| 690 | spin_lock(&class->lock); |
| 691 | |
| 692 | /* Insert this object in containing zspage's freelist */ |
| 693 | link = (struct link_free *)((unsigned char *)kmap_atomic(f_page) |
| 694 | + f_offset); |
| 695 | link->next = first_page->freelist; |
| 696 | kunmap_atomic(link); |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 697 | first_page->freelist = (void *)obj; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 698 | |
| 699 | first_page->inuse--; |
| 700 | fullness = fix_fullness_group(pool, first_page); |
| 701 | |
| 702 | if (fullness == ZS_EMPTY) |
Minchan Kim | 2e3b615 | 2012-05-03 15:40:39 +0900 | [diff] [blame] | 703 | class->pages_allocated -= class->pages_per_zspage; |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 704 | |
| 705 | spin_unlock(&class->lock); |
| 706 | |
| 707 | if (fullness == ZS_EMPTY) |
| 708 | free_zspage(first_page); |
| 709 | } |
| 710 | EXPORT_SYMBOL_GPL(zs_free); |
| 711 | |
Minchan Kim | 00a61d8 | 2012-05-03 15:40:40 +0900 | [diff] [blame] | 712 | /** |
| 713 | * zs_map_object - get address of allocated object from handle. |
| 714 | * @pool: pool from which the object was allocated |
| 715 | * @handle: handle returned from zs_malloc |
| 716 | * |
| 717 | * Before using an object allocated from zs_malloc, it must be mapped using |
| 718 | * this function. When done with the object, it must be unmapped using |
| 719 | * zs_unmap_object |
| 720 | */ |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 721 | void *zs_map_object(struct zs_pool *pool, unsigned long handle) |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 722 | { |
| 723 | struct page *page; |
| 724 | unsigned long obj_idx, off; |
| 725 | |
| 726 | unsigned int class_idx; |
| 727 | enum fullness_group fg; |
| 728 | struct size_class *class; |
| 729 | struct mapping_area *area; |
| 730 | |
| 731 | BUG_ON(!handle); |
| 732 | |
| 733 | obj_handle_to_location(handle, &page, &obj_idx); |
| 734 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
| 735 | class = &pool->size_class[class_idx]; |
| 736 | off = obj_idx_to_offset(page, obj_idx, class->size); |
| 737 | |
| 738 | area = &get_cpu_var(zs_map_area); |
| 739 | if (off + class->size <= PAGE_SIZE) { |
| 740 | /* this object is contained entirely within a page */ |
| 741 | area->vm_addr = kmap_atomic(page); |
| 742 | } else { |
| 743 | /* this object spans two pages */ |
| 744 | struct page *nextp; |
| 745 | |
| 746 | nextp = get_next_page(page); |
| 747 | BUG_ON(!nextp); |
| 748 | |
| 749 | |
| 750 | set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL)); |
| 751 | set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL)); |
| 752 | |
| 753 | /* We pre-allocated VM area so mapping can never fail */ |
| 754 | area->vm_addr = area->vm->addr; |
| 755 | } |
| 756 | |
| 757 | return area->vm_addr + off; |
| 758 | } |
| 759 | EXPORT_SYMBOL_GPL(zs_map_object); |
| 760 | |
Minchan Kim | c234434 | 2012-06-08 15:39:25 +0900 | [diff] [blame] | 761 | void zs_unmap_object(struct zs_pool *pool, unsigned long handle) |
Nitin Gupta | 61989a8 | 2012-01-09 16:51:56 -0600 | [diff] [blame] | 762 | { |
| 763 | struct page *page; |
| 764 | unsigned long obj_idx, off; |
| 765 | |
| 766 | unsigned int class_idx; |
| 767 | enum fullness_group fg; |
| 768 | struct size_class *class; |
| 769 | struct mapping_area *area; |
| 770 | |
| 771 | BUG_ON(!handle); |
| 772 | |
| 773 | obj_handle_to_location(handle, &page, &obj_idx); |
| 774 | get_zspage_mapping(get_first_page(page), &class_idx, &fg); |
| 775 | class = &pool->size_class[class_idx]; |
| 776 | off = obj_idx_to_offset(page, obj_idx, class->size); |
| 777 | |
| 778 | area = &__get_cpu_var(zs_map_area); |
| 779 | if (off + class->size <= PAGE_SIZE) { |
| 780 | kunmap_atomic(area->vm_addr); |
| 781 | } else { |
| 782 | set_pte(area->vm_ptes[0], __pte(0)); |
| 783 | set_pte(area->vm_ptes[1], __pte(0)); |
| 784 | __flush_tlb_one((unsigned long)area->vm_addr); |
| 785 | __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE); |
| 786 | } |
| 787 | put_cpu_var(zs_map_area); |
| 788 | } |
| 789 | EXPORT_SYMBOL_GPL(zs_unmap_object); |
| 790 | |
| 791 | u64 zs_get_total_size_bytes(struct zs_pool *pool) |
| 792 | { |
| 793 | int i; |
| 794 | u64 npages = 0; |
| 795 | |
| 796 | for (i = 0; i < ZS_SIZE_CLASSES; i++) |
| 797 | npages += pool->size_class[i].pages_allocated; |
| 798 | |
| 799 | return npages << PAGE_SHIFT; |
| 800 | } |
| 801 | EXPORT_SYMBOL_GPL(zs_get_total_size_bytes); |