blob: f135b1b6fcdcab49aaf0845e078c6fc299b4b28b [file] [log] [blame]
Nitin Gupta61989a82012-01-09 16:51:56 -06001/*
2 * zsmalloc memory allocator
3 *
4 * Copyright (C) 2011 Nitin Gupta
Minchan Kim31fc00b2014-01-30 15:45:55 -08005 * Copyright (C) 2012, 2013 Minchan Kim
Nitin Gupta61989a82012-01-09 16:51:56 -06006 *
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
Nitin Gupta2db51da2012-06-09 17:41:14 -070014/*
Nitin Gupta2db51da2012-06-09 17:41:14 -070015 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
17 *
18 * Usage of struct page fields:
19 * page->first_page: points to the first component (0-order) page
20 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
25 * of a zspage
26 *
27 * For _first_ page only:
28 *
29 * page->private (union with page->first_page): refers to the
30 * component page after the first page
Minchan Kim7b60a682015-04-15 16:15:39 -070031 * If the page is first_page for huge object, it stores handle.
32 * Look at size_class->huge.
Nitin Gupta2db51da2012-06-09 17:41:14 -070033 * page->freelist: points to the first free object in zspage.
34 * Free objects are linked together using in-place
35 * metadata.
36 * page->objects: maximum number of objects we can store in this
37 * zspage (class->zspage_order * PAGE_SIZE / class->size)
38 * page->lru: links together first pages of various zspages.
39 * Basically forming list of zspages in a fullness group.
40 * page->mapping: class index and fullness group of the zspage
41 *
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
45 *
46 */
47
Nitin Gupta61989a82012-01-09 16:51:56 -060048#include <linux/module.h>
49#include <linux/kernel.h>
Minchan Kim312fcae2015-04-15 16:15:30 -070050#include <linux/sched.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060051#include <linux/bitops.h>
52#include <linux/errno.h>
53#include <linux/highmem.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060054#include <linux/string.h>
55#include <linux/slab.h>
56#include <asm/tlbflush.h>
57#include <asm/pgtable.h>
58#include <linux/cpumask.h>
59#include <linux/cpu.h>
Seth Jennings0cbb6132012-02-13 08:47:49 -060060#include <linux/vmalloc.h>
Seth Jenningsc60369f2012-07-18 11:55:55 -050061#include <linux/hardirq.h>
Seth Jennings0959c632012-08-08 15:12:17 +090062#include <linux/spinlock.h>
63#include <linux/types.h>
Ganesh Mahendran0f050d92015-02-12 15:00:54 -080064#include <linux/debugfs.h>
Minchan Kimbcf16472014-01-30 15:45:50 -080065#include <linux/zsmalloc.h>
Dan Streetmanc7957792014-08-06 16:08:38 -070066#include <linux/zpool.h>
Seth Jennings0959c632012-08-08 15:12:17 +090067
68/*
69 * This must be power of 2 and greater than of equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
73 */
74#define ZS_ALIGN 8
75
76/*
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
79 */
80#define ZS_MAX_ZSPAGE_ORDER 2
81#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
82
Minchan Kim2e40e162015-04-15 16:15:23 -070083#define ZS_HANDLE_SIZE (sizeof(unsigned long))
84
Seth Jennings0959c632012-08-08 15:12:17 +090085/*
86 * Object location (<PFN>, <obj_idx>) is encoded as
Nitin Cuptac3e3e882013-12-11 11:04:37 +090087 * as single (unsigned long) handle value.
Seth Jennings0959c632012-08-08 15:12:17 +090088 *
89 * Note that object index <obj_idx> is relative to system
90 * page <PFN> it is stored in, so for each sub-page belonging
91 * to a zspage, obj_idx starts with 0.
92 *
93 * This is made more complicated by various memory models and PAE.
94 */
95
96#ifndef MAX_PHYSMEM_BITS
97#ifdef CONFIG_HIGHMEM64G
98#define MAX_PHYSMEM_BITS 36
99#else /* !CONFIG_HIGHMEM64G */
100/*
101 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
102 * be PAGE_SHIFT
103 */
104#define MAX_PHYSMEM_BITS BITS_PER_LONG
105#endif
106#endif
107#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
Minchan Kim312fcae2015-04-15 16:15:30 -0700108
109/*
110 * Memory for allocating for handle keeps object position by
111 * encoding <page, obj_idx> and the encoded value has a room
112 * in least bit(ie, look at obj_to_location).
113 * We use the bit to synchronize between object access by
114 * user and migration.
115 */
116#define HANDLE_PIN_BIT 0
117
118/*
119 * Head in allocated object should have OBJ_ALLOCATED_TAG
120 * to identify the object was allocated or not.
121 * It's okay to add the status bit in the least bit because
122 * header keeps handle which is 4byte-aligned address so we
123 * have room for two bit at least.
124 */
125#define OBJ_ALLOCATED_TAG 1
126#define OBJ_TAG_BITS 1
127#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
Seth Jennings0959c632012-08-08 15:12:17 +0900128#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
129
130#define MAX(a, b) ((a) >= (b) ? (a) : (b))
131/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132#define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
Minchan Kim2e40e162015-04-15 16:15:23 -0700134/* each chunk includes extra space to keep handle */
Minchan Kim7b60a682015-04-15 16:15:39 -0700135#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
Seth Jennings0959c632012-08-08 15:12:17 +0900136
137/*
Weijie Yang7eb52512014-06-04 16:11:08 -0700138 * On systems with 4K page size, this gives 255 size classes! There is a
Seth Jennings0959c632012-08-08 15:12:17 +0900139 * trader-off here:
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
146 *
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
148 * (reason above)
149 */
Seth Jenningsd662b8e2013-01-25 11:46:18 -0600150#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
Seth Jennings0959c632012-08-08 15:12:17 +0900151
152/*
153 * We do not maintain any list for completely empty or full pages
154 */
155enum fullness_group {
156 ZS_ALMOST_FULL,
157 ZS_ALMOST_EMPTY,
158 _ZS_NR_FULLNESS_GROUPS,
159
160 ZS_EMPTY,
161 ZS_FULL
162};
163
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800164enum zs_stat_type {
165 OBJ_ALLOCATED,
166 OBJ_USED,
Minchan Kim248ca1b2015-04-15 16:15:42 -0700167 CLASS_ALMOST_FULL,
168 CLASS_ALMOST_EMPTY,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800169 NR_ZS_STAT_TYPE,
170};
171
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800172struct zs_size_stat {
173 unsigned long objs[NR_ZS_STAT_TYPE];
174};
175
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700176#ifdef CONFIG_ZSMALLOC_STAT
177static struct dentry *zs_stat_root;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800178#endif
179
Seth Jennings0959c632012-08-08 15:12:17 +0900180/*
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800181 * number of size_classes
182 */
183static int zs_size_classes;
184
185/*
Seth Jennings0959c632012-08-08 15:12:17 +0900186 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
187 * n <= N / f, where
188 * n = number of allocated objects
189 * N = total number of objects zspage can store
Wang Sheng-Hui6dd97372014-10-09 15:29:59 -0700190 * f = fullness_threshold_frac
Seth Jennings0959c632012-08-08 15:12:17 +0900191 *
192 * Similarly, we assign zspage to:
193 * ZS_ALMOST_FULL when n > N / f
194 * ZS_EMPTY when n == 0
195 * ZS_FULL when n == N
196 *
197 * (see: fix_fullness_group())
198 */
199static const int fullness_threshold_frac = 4;
200
201struct size_class {
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700202 spinlock_t lock;
203 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
Seth Jennings0959c632012-08-08 15:12:17 +0900204 /*
205 * Size of objects stored in this class. Must be multiple
206 * of ZS_ALIGN.
207 */
208 int size;
209 unsigned int index;
210
211 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
212 int pages_per_zspage;
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700213 struct zs_size_stat stats;
214
Minchan Kim7b60a682015-04-15 16:15:39 -0700215 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
216 bool huge;
Seth Jennings0959c632012-08-08 15:12:17 +0900217};
218
219/*
220 * Placed within free objects to form a singly linked list.
221 * For every zspage, first_page->freelist gives head of this list.
222 *
223 * This must be power of 2 and less than or equal to ZS_ALIGN
224 */
225struct link_free {
Minchan Kim2e40e162015-04-15 16:15:23 -0700226 union {
227 /*
228 * Position of next free chunk (encodes <PFN, obj_idx>)
229 * It's valid for non-allocated object
230 */
231 void *next;
232 /*
233 * Handle of allocated object.
234 */
235 unsigned long handle;
236 };
Seth Jennings0959c632012-08-08 15:12:17 +0900237};
238
239struct zs_pool {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800240 char *name;
241
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800242 struct size_class **size_class;
Minchan Kim2e40e162015-04-15 16:15:23 -0700243 struct kmem_cache *handle_cachep;
Seth Jennings0959c632012-08-08 15:12:17 +0900244
245 gfp_t flags; /* allocation flags used when growing pool */
Minchan Kim13de8932014-10-09 15:29:48 -0700246 atomic_long_t pages_allocated;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800247
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -0700248 struct zs_pool_stats stats;
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -0700249
250 /* Compact classes */
251 struct shrinker shrinker;
252 /*
253 * To signify that register_shrinker() was successful
254 * and unregister_shrinker() will not Oops.
255 */
256 bool shrinker_enabled;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800257#ifdef CONFIG_ZSMALLOC_STAT
258 struct dentry *stat_dentry;
259#endif
Seth Jennings0959c632012-08-08 15:12:17 +0900260};
Nitin Gupta61989a82012-01-09 16:51:56 -0600261
262/*
263 * A zspage's class index and fullness group
264 * are encoded in its (first)page->mapping
265 */
266#define CLASS_IDX_BITS 28
267#define FULLNESS_BITS 4
268#define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
269#define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
270
Seth Jenningsf5536462012-07-18 11:55:56 -0500271struct mapping_area {
Minchan Kim1b945ae2013-12-11 11:04:36 +0900272#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -0500273 struct vm_struct *vm; /* vm area for mapping object that span pages */
274#else
275 char *vm_buf; /* copy buffer for objects that span pages */
276#endif
277 char *vm_addr; /* address of kmap_atomic()'ed pages */
278 enum zs_mapmode vm_mm; /* mapping mode */
Minchan Kim7b60a682015-04-15 16:15:39 -0700279 bool huge;
Seth Jenningsf5536462012-07-18 11:55:56 -0500280};
281
Minchan Kim2e40e162015-04-15 16:15:23 -0700282static int create_handle_cache(struct zs_pool *pool)
283{
284 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
285 0, 0, NULL);
286 return pool->handle_cachep ? 0 : 1;
287}
288
289static void destroy_handle_cache(struct zs_pool *pool)
290{
Sergey Senozhatskycd10add2015-09-08 15:04:55 -0700291 kmem_cache_destroy(pool->handle_cachep);
Minchan Kim2e40e162015-04-15 16:15:23 -0700292}
293
294static unsigned long alloc_handle(struct zs_pool *pool)
295{
296 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
297 pool->flags & ~__GFP_HIGHMEM);
298}
299
300static void free_handle(struct zs_pool *pool, unsigned long handle)
301{
302 kmem_cache_free(pool->handle_cachep, (void *)handle);
303}
304
305static void record_obj(unsigned long handle, unsigned long obj)
306{
307 *(unsigned long *)handle = obj;
308}
309
Dan Streetmanc7957792014-08-06 16:08:38 -0700310/* zpool driver */
311
312#ifdef CONFIG_ZPOOL
313
Krzysztof Kozlowski78672772015-09-08 15:05:03 -0700314static void *zs_zpool_create(char *name, gfp_t gfp,
315 const struct zpool_ops *zpool_ops,
Dan Streetman479305f2015-06-25 15:00:40 -0700316 struct zpool *zpool)
Dan Streetmanc7957792014-08-06 16:08:38 -0700317{
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -0800318 return zs_create_pool(name, gfp);
Dan Streetmanc7957792014-08-06 16:08:38 -0700319}
320
321static void zs_zpool_destroy(void *pool)
322{
323 zs_destroy_pool(pool);
324}
325
326static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
327 unsigned long *handle)
328{
329 *handle = zs_malloc(pool, size);
330 return *handle ? 0 : -1;
331}
332static void zs_zpool_free(void *pool, unsigned long handle)
333{
334 zs_free(pool, handle);
335}
336
337static int zs_zpool_shrink(void *pool, unsigned int pages,
338 unsigned int *reclaimed)
339{
340 return -EINVAL;
341}
342
343static void *zs_zpool_map(void *pool, unsigned long handle,
344 enum zpool_mapmode mm)
345{
346 enum zs_mapmode zs_mm;
347
348 switch (mm) {
349 case ZPOOL_MM_RO:
350 zs_mm = ZS_MM_RO;
351 break;
352 case ZPOOL_MM_WO:
353 zs_mm = ZS_MM_WO;
354 break;
355 case ZPOOL_MM_RW: /* fallthru */
356 default:
357 zs_mm = ZS_MM_RW;
358 break;
359 }
360
361 return zs_map_object(pool, handle, zs_mm);
362}
363static void zs_zpool_unmap(void *pool, unsigned long handle)
364{
365 zs_unmap_object(pool, handle);
366}
367
368static u64 zs_zpool_total_size(void *pool)
369{
Minchan Kim722cdc12014-10-09 15:29:50 -0700370 return zs_get_total_pages(pool) << PAGE_SHIFT;
Dan Streetmanc7957792014-08-06 16:08:38 -0700371}
372
373static struct zpool_driver zs_zpool_driver = {
374 .type = "zsmalloc",
375 .owner = THIS_MODULE,
376 .create = zs_zpool_create,
377 .destroy = zs_zpool_destroy,
378 .malloc = zs_zpool_malloc,
379 .free = zs_zpool_free,
380 .shrink = zs_zpool_shrink,
381 .map = zs_zpool_map,
382 .unmap = zs_zpool_unmap,
383 .total_size = zs_zpool_total_size,
384};
385
Kees Cook137f8cf2014-08-29 15:18:40 -0700386MODULE_ALIAS("zpool-zsmalloc");
Dan Streetmanc7957792014-08-06 16:08:38 -0700387#endif /* CONFIG_ZPOOL */
388
Minchan Kim248ca1b2015-04-15 16:15:42 -0700389static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
390{
391 return pages_per_zspage * PAGE_SIZE / size;
392}
393
Nitin Gupta61989a82012-01-09 16:51:56 -0600394/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
395static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
396
397static int is_first_page(struct page *page)
398{
Minchan Kima27545bf2012-04-25 15:23:09 +0900399 return PagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600400}
401
402static int is_last_page(struct page *page)
403{
Minchan Kima27545bf2012-04-25 15:23:09 +0900404 return PagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600405}
406
407static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
408 enum fullness_group *fullness)
409{
410 unsigned long m;
411 BUG_ON(!is_first_page(page));
412
413 m = (unsigned long)page->mapping;
414 *fullness = m & FULLNESS_MASK;
415 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
416}
417
418static void set_zspage_mapping(struct page *page, unsigned int class_idx,
419 enum fullness_group fullness)
420{
421 unsigned long m;
422 BUG_ON(!is_first_page(page));
423
424 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
425 (fullness & FULLNESS_MASK);
426 page->mapping = (struct address_space *)m;
427}
428
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900429/*
430 * zsmalloc divides the pool into various size classes where each
431 * class maintains a list of zspages where each zspage is divided
432 * into equal sized chunks. Each allocation falls into one of these
433 * classes depending on its size. This function returns index of the
434 * size class which has chunk size big enough to hold the give size.
435 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600436static int get_size_class_index(int size)
437{
438 int idx = 0;
439
440 if (likely(size > ZS_MIN_ALLOC_SIZE))
441 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
442 ZS_SIZE_CLASS_DELTA);
443
Minchan Kim7b60a682015-04-15 16:15:39 -0700444 return min(zs_size_classes - 1, idx);
Nitin Gupta61989a82012-01-09 16:51:56 -0600445}
446
Minchan Kim248ca1b2015-04-15 16:15:42 -0700447static inline void zs_stat_inc(struct size_class *class,
448 enum zs_stat_type type, unsigned long cnt)
449{
450 class->stats.objs[type] += cnt;
451}
452
453static inline void zs_stat_dec(struct size_class *class,
454 enum zs_stat_type type, unsigned long cnt)
455{
456 class->stats.objs[type] -= cnt;
457}
458
459static inline unsigned long zs_stat_get(struct size_class *class,
460 enum zs_stat_type type)
461{
462 return class->stats.objs[type];
463}
464
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700465#ifdef CONFIG_ZSMALLOC_STAT
466
Minchan Kim248ca1b2015-04-15 16:15:42 -0700467static int __init zs_stat_init(void)
468{
469 if (!debugfs_initialized())
470 return -ENODEV;
471
472 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
473 if (!zs_stat_root)
474 return -ENOMEM;
475
476 return 0;
477}
478
479static void __exit zs_stat_exit(void)
480{
481 debugfs_remove_recursive(zs_stat_root);
482}
483
484static int zs_stats_size_show(struct seq_file *s, void *v)
485{
486 int i;
487 struct zs_pool *pool = s->private;
488 struct size_class *class;
489 int objs_per_zspage;
490 unsigned long class_almost_full, class_almost_empty;
491 unsigned long obj_allocated, obj_used, pages_used;
492 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
493 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
494
495 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
496 "class", "size", "almost_full", "almost_empty",
497 "obj_allocated", "obj_used", "pages_used",
498 "pages_per_zspage");
499
500 for (i = 0; i < zs_size_classes; i++) {
501 class = pool->size_class[i];
502
503 if (class->index != i)
504 continue;
505
506 spin_lock(&class->lock);
507 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
508 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
509 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
510 obj_used = zs_stat_get(class, OBJ_USED);
511 spin_unlock(&class->lock);
512
513 objs_per_zspage = get_maxobj_per_zspage(class->size,
514 class->pages_per_zspage);
515 pages_used = obj_allocated / objs_per_zspage *
516 class->pages_per_zspage;
517
518 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
519 i, class->size, class_almost_full, class_almost_empty,
520 obj_allocated, obj_used, pages_used,
521 class->pages_per_zspage);
522
523 total_class_almost_full += class_almost_full;
524 total_class_almost_empty += class_almost_empty;
525 total_objs += obj_allocated;
526 total_used_objs += obj_used;
527 total_pages += pages_used;
528 }
529
530 seq_puts(s, "\n");
531 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
532 "Total", "", total_class_almost_full,
533 total_class_almost_empty, total_objs,
534 total_used_objs, total_pages);
535
536 return 0;
537}
538
539static int zs_stats_size_open(struct inode *inode, struct file *file)
540{
541 return single_open(file, zs_stats_size_show, inode->i_private);
542}
543
544static const struct file_operations zs_stat_size_ops = {
545 .open = zs_stats_size_open,
546 .read = seq_read,
547 .llseek = seq_lseek,
548 .release = single_release,
549};
550
551static int zs_pool_stat_create(char *name, struct zs_pool *pool)
552{
553 struct dentry *entry;
554
555 if (!zs_stat_root)
556 return -ENODEV;
557
558 entry = debugfs_create_dir(name, zs_stat_root);
559 if (!entry) {
560 pr_warn("debugfs dir <%s> creation failed\n", name);
561 return -ENOMEM;
562 }
563 pool->stat_dentry = entry;
564
565 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
566 pool->stat_dentry, pool, &zs_stat_size_ops);
567 if (!entry) {
568 pr_warn("%s: debugfs file entry <%s> creation failed\n",
569 name, "classes");
570 return -ENOMEM;
571 }
572
573 return 0;
574}
575
576static void zs_pool_stat_destroy(struct zs_pool *pool)
577{
578 debugfs_remove_recursive(pool->stat_dentry);
579}
580
581#else /* CONFIG_ZSMALLOC_STAT */
Minchan Kim248ca1b2015-04-15 16:15:42 -0700582static int __init zs_stat_init(void)
583{
584 return 0;
585}
586
587static void __exit zs_stat_exit(void)
588{
589}
590
591static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
592{
593 return 0;
594}
595
596static inline void zs_pool_stat_destroy(struct zs_pool *pool)
597{
598}
Minchan Kim248ca1b2015-04-15 16:15:42 -0700599#endif
600
601
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900602/*
603 * For each size class, zspages are divided into different groups
604 * depending on how "full" they are. This was done so that we could
605 * easily find empty or nearly empty zspages when we try to shrink
606 * the pool (not yet implemented). This function returns fullness
607 * status of the given page.
608 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600609static enum fullness_group get_fullness_group(struct page *page)
610{
611 int inuse, max_objects;
612 enum fullness_group fg;
613 BUG_ON(!is_first_page(page));
614
615 inuse = page->inuse;
616 max_objects = page->objects;
617
618 if (inuse == 0)
619 fg = ZS_EMPTY;
620 else if (inuse == max_objects)
621 fg = ZS_FULL;
Minchan Kimd3d07c92015-04-15 16:15:33 -0700622 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
Nitin Gupta61989a82012-01-09 16:51:56 -0600623 fg = ZS_ALMOST_EMPTY;
624 else
625 fg = ZS_ALMOST_FULL;
626
627 return fg;
628}
629
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900630/*
631 * Each size class maintains various freelists and zspages are assigned
632 * to one of these freelists based on the number of live objects they
633 * have. This functions inserts the given zspage into the freelist
634 * identified by <class, fullness_group>.
635 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600636static void insert_zspage(struct page *page, struct size_class *class,
637 enum fullness_group fullness)
638{
639 struct page **head;
640
641 BUG_ON(!is_first_page(page));
642
643 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
644 return;
645
Minchan Kim248ca1b2015-04-15 16:15:42 -0700646 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
647 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Sergey Senozhatsky58f17112015-09-08 15:04:44 -0700648
649 head = &class->fullness_list[fullness];
650 if (!*head) {
651 *head = page;
652 return;
653 }
654
655 /*
656 * We want to see more ZS_FULL pages and less almost
657 * empty/full. Put pages with higher ->inuse first.
658 */
659 list_add_tail(&page->lru, &(*head)->lru);
660 if (page->inuse >= (*head)->inuse)
661 *head = page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600662}
663
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900664/*
665 * This function removes the given zspage from the freelist identified
666 * by <class, fullness_group>.
667 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600668static void remove_zspage(struct page *page, struct size_class *class,
669 enum fullness_group fullness)
670{
671 struct page **head;
672
673 BUG_ON(!is_first_page(page));
674
675 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
676 return;
677
678 head = &class->fullness_list[fullness];
679 BUG_ON(!*head);
680 if (list_empty(&(*head)->lru))
681 *head = NULL;
682 else if (*head == page)
683 *head = (struct page *)list_entry((*head)->lru.next,
684 struct page, lru);
685
686 list_del_init(&page->lru);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700687 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
688 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600689}
690
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900691/*
692 * Each size class maintains zspages in different fullness groups depending
693 * on the number of live objects they contain. When allocating or freeing
694 * objects, the fullness status of the page can change, say, from ALMOST_FULL
695 * to ALMOST_EMPTY when freeing an object. This function checks if such
696 * a status change has occurred for the given page and accordingly moves the
697 * page from the freelist of the old fullness group to that of the new
698 * fullness group.
699 */
Minchan Kimc7806262015-04-15 16:15:26 -0700700static enum fullness_group fix_fullness_group(struct size_class *class,
Nitin Gupta61989a82012-01-09 16:51:56 -0600701 struct page *page)
702{
703 int class_idx;
Nitin Gupta61989a82012-01-09 16:51:56 -0600704 enum fullness_group currfg, newfg;
705
706 BUG_ON(!is_first_page(page));
707
708 get_zspage_mapping(page, &class_idx, &currfg);
709 newfg = get_fullness_group(page);
710 if (newfg == currfg)
711 goto out;
712
Nitin Gupta61989a82012-01-09 16:51:56 -0600713 remove_zspage(page, class, currfg);
714 insert_zspage(page, class, newfg);
715 set_zspage_mapping(page, class_idx, newfg);
716
717out:
718 return newfg;
719}
720
721/*
722 * We have to decide on how many pages to link together
723 * to form a zspage for each size class. This is important
724 * to reduce wastage due to unusable space left at end of
725 * each zspage which is given as:
Yinghao Xie888fa3742015-04-15 16:15:49 -0700726 * wastage = Zp % class_size
727 * usage = Zp - wastage
Nitin Gupta61989a82012-01-09 16:51:56 -0600728 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
729 *
730 * For example, for size class of 3/8 * PAGE_SIZE, we should
731 * link together 3 PAGE_SIZE sized pages to form a zspage
732 * since then we can perfectly fit in 8 such objects.
733 */
Minchan Kim2e3b6152012-05-03 15:40:39 +0900734static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600735{
736 int i, max_usedpc = 0;
737 /* zspage order which gives maximum used size per KB */
738 int max_usedpc_order = 1;
739
Seth Jennings84d4faa2012-03-05 11:33:21 -0600740 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600741 int zspage_size;
742 int waste, usedpc;
743
744 zspage_size = i * PAGE_SIZE;
745 waste = zspage_size % class_size;
746 usedpc = (zspage_size - waste) * 100 / zspage_size;
747
748 if (usedpc > max_usedpc) {
749 max_usedpc = usedpc;
750 max_usedpc_order = i;
751 }
752 }
753
754 return max_usedpc_order;
755}
756
757/*
758 * A single 'zspage' is composed of many system pages which are
759 * linked together using fields in struct page. This function finds
760 * the first/head page, given any component page of a zspage.
761 */
762static struct page *get_first_page(struct page *page)
763{
764 if (is_first_page(page))
765 return page;
766 else
767 return page->first_page;
768}
769
770static struct page *get_next_page(struct page *page)
771{
772 struct page *next;
773
774 if (is_last_page(page))
775 next = NULL;
776 else if (is_first_page(page))
Sunghan Suhe842b972013-07-12 16:08:13 +0900777 next = (struct page *)page_private(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600778 else
779 next = list_entry(page->lru.next, struct page, lru);
780
781 return next;
782}
783
Olav Haugan67296872013-11-22 09:30:41 -0800784/*
785 * Encode <page, obj_idx> as a single handle value.
Minchan Kim312fcae2015-04-15 16:15:30 -0700786 * We use the least bit of handle for tagging.
Olav Haugan67296872013-11-22 09:30:41 -0800787 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700788static void *location_to_obj(struct page *page, unsigned long obj_idx)
Nitin Gupta61989a82012-01-09 16:51:56 -0600789{
Minchan Kim312fcae2015-04-15 16:15:30 -0700790 unsigned long obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600791
792 if (!page) {
793 BUG_ON(obj_idx);
794 return NULL;
795 }
796
Minchan Kim312fcae2015-04-15 16:15:30 -0700797 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
798 obj |= ((obj_idx) & OBJ_INDEX_MASK);
799 obj <<= OBJ_TAG_BITS;
Nitin Gupta61989a82012-01-09 16:51:56 -0600800
Minchan Kim312fcae2015-04-15 16:15:30 -0700801 return (void *)obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600802}
803
Olav Haugan67296872013-11-22 09:30:41 -0800804/*
805 * Decode <page, obj_idx> pair from the given object handle. We adjust the
806 * decoded obj_idx back to its original value since it was adjusted in
Minchan Kim312fcae2015-04-15 16:15:30 -0700807 * location_to_obj().
Olav Haugan67296872013-11-22 09:30:41 -0800808 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700809static void obj_to_location(unsigned long obj, struct page **page,
Nitin Gupta61989a82012-01-09 16:51:56 -0600810 unsigned long *obj_idx)
811{
Minchan Kim312fcae2015-04-15 16:15:30 -0700812 obj >>= OBJ_TAG_BITS;
813 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
814 *obj_idx = (obj & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600815}
816
Minchan Kim2e40e162015-04-15 16:15:23 -0700817static unsigned long handle_to_obj(unsigned long handle)
818{
819 return *(unsigned long *)handle;
820}
821
Minchan Kim7b60a682015-04-15 16:15:39 -0700822static unsigned long obj_to_head(struct size_class *class, struct page *page,
823 void *obj)
Minchan Kim312fcae2015-04-15 16:15:30 -0700824{
Minchan Kim7b60a682015-04-15 16:15:39 -0700825 if (class->huge) {
826 VM_BUG_ON(!is_first_page(page));
827 return *(unsigned long *)page_private(page);
828 } else
829 return *(unsigned long *)obj;
Minchan Kim312fcae2015-04-15 16:15:30 -0700830}
831
Nitin Gupta61989a82012-01-09 16:51:56 -0600832static unsigned long obj_idx_to_offset(struct page *page,
833 unsigned long obj_idx, int class_size)
834{
835 unsigned long off = 0;
836
837 if (!is_first_page(page))
838 off = page->index;
839
840 return off + obj_idx * class_size;
841}
842
Minchan Kim312fcae2015-04-15 16:15:30 -0700843static inline int trypin_tag(unsigned long handle)
844{
845 unsigned long *ptr = (unsigned long *)handle;
846
847 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
848}
849
850static void pin_tag(unsigned long handle)
851{
852 while (!trypin_tag(handle));
853}
854
855static void unpin_tag(unsigned long handle)
856{
857 unsigned long *ptr = (unsigned long *)handle;
858
859 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
860}
861
Nitin Guptaf4477e92012-04-02 09:13:56 -0500862static void reset_page(struct page *page)
863{
864 clear_bit(PG_private, &page->flags);
865 clear_bit(PG_private_2, &page->flags);
866 set_page_private(page, 0);
867 page->mapping = NULL;
868 page->freelist = NULL;
Mel Gorman22b751c2013-02-22 16:34:59 -0800869 page_mapcount_reset(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500870}
871
Nitin Gupta61989a82012-01-09 16:51:56 -0600872static void free_zspage(struct page *first_page)
873{
Nitin Guptaf4477e92012-04-02 09:13:56 -0500874 struct page *nextp, *tmp, *head_extra;
Nitin Gupta61989a82012-01-09 16:51:56 -0600875
876 BUG_ON(!is_first_page(first_page));
877 BUG_ON(first_page->inuse);
878
Nitin Guptaf4477e92012-04-02 09:13:56 -0500879 head_extra = (struct page *)page_private(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600880
Nitin Guptaf4477e92012-04-02 09:13:56 -0500881 reset_page(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600882 __free_page(first_page);
883
884 /* zspage with only 1 system page */
Nitin Guptaf4477e92012-04-02 09:13:56 -0500885 if (!head_extra)
Nitin Gupta61989a82012-01-09 16:51:56 -0600886 return;
887
Nitin Guptaf4477e92012-04-02 09:13:56 -0500888 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600889 list_del(&nextp->lru);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500890 reset_page(nextp);
Nitin Gupta61989a82012-01-09 16:51:56 -0600891 __free_page(nextp);
892 }
Nitin Guptaf4477e92012-04-02 09:13:56 -0500893 reset_page(head_extra);
894 __free_page(head_extra);
Nitin Gupta61989a82012-01-09 16:51:56 -0600895}
896
897/* Initialize a newly allocated zspage */
898static void init_zspage(struct page *first_page, struct size_class *class)
899{
900 unsigned long off = 0;
901 struct page *page = first_page;
902
903 BUG_ON(!is_first_page(first_page));
904 while (page) {
905 struct page *next_page;
906 struct link_free *link;
Dan Streetman5538c562014-10-09 15:30:01 -0700907 unsigned int i = 1;
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800908 void *vaddr;
Nitin Gupta61989a82012-01-09 16:51:56 -0600909
910 /*
911 * page->index stores offset of first object starting
912 * in the page. For the first page, this is always 0,
913 * so we use first_page->index (aka ->freelist) to store
914 * head of corresponding zspage's freelist.
915 */
916 if (page != first_page)
917 page->index = off;
918
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800919 vaddr = kmap_atomic(page);
920 link = (struct link_free *)vaddr + off / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600921
Dan Streetman5538c562014-10-09 15:30:01 -0700922 while ((off += class->size) < PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -0700923 link->next = location_to_obj(page, i++);
Dan Streetman5538c562014-10-09 15:30:01 -0700924 link += class->size / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600925 }
926
927 /*
928 * We now come to the last (full or partial) object on this
929 * page, which must point to the first object on the next
930 * page (if present)
931 */
932 next_page = get_next_page(page);
Minchan Kim312fcae2015-04-15 16:15:30 -0700933 link->next = location_to_obj(next_page, 0);
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800934 kunmap_atomic(vaddr);
Nitin Gupta61989a82012-01-09 16:51:56 -0600935 page = next_page;
Dan Streetman5538c562014-10-09 15:30:01 -0700936 off %= PAGE_SIZE;
Nitin Gupta61989a82012-01-09 16:51:56 -0600937 }
938}
939
940/*
941 * Allocate a zspage for the given size class
942 */
943static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
944{
945 int i, error;
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500946 struct page *first_page = NULL, *uninitialized_var(prev_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600947
948 /*
949 * Allocate individual pages and link them together as:
950 * 1. first page->private = first sub-page
951 * 2. all sub-pages are linked together using page->lru
952 * 3. each sub-page is linked to the first page using page->first_page
953 *
954 * For each size class, First/Head pages are linked together using
955 * page->lru. Also, we set PG_private to identify the first page
956 * (i.e. no other sub-page has this flag set) and PG_private_2 to
957 * identify the last page.
958 */
959 error = -ENOMEM;
Minchan Kim2e3b6152012-05-03 15:40:39 +0900960 for (i = 0; i < class->pages_per_zspage; i++) {
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500961 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600962
963 page = alloc_page(flags);
964 if (!page)
965 goto cleanup;
966
967 INIT_LIST_HEAD(&page->lru);
968 if (i == 0) { /* first page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900969 SetPagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600970 set_page_private(page, 0);
971 first_page = page;
972 first_page->inuse = 0;
973 }
974 if (i == 1)
Sunghan Suhe842b972013-07-12 16:08:13 +0900975 set_page_private(first_page, (unsigned long)page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600976 if (i >= 1)
977 page->first_page = first_page;
978 if (i >= 2)
979 list_add(&page->lru, &prev_page->lru);
Minchan Kim2e3b6152012-05-03 15:40:39 +0900980 if (i == class->pages_per_zspage - 1) /* last page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900981 SetPagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600982 prev_page = page;
983 }
984
985 init_zspage(first_page, class);
986
Minchan Kim312fcae2015-04-15 16:15:30 -0700987 first_page->freelist = location_to_obj(first_page, 0);
Nitin Gupta61989a82012-01-09 16:51:56 -0600988 /* Maximum number of objects we can store in this zspage */
Minchan Kim2e3b6152012-05-03 15:40:39 +0900989 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
Nitin Gupta61989a82012-01-09 16:51:56 -0600990
991 error = 0; /* Success */
992
993cleanup:
994 if (unlikely(error) && first_page) {
995 free_zspage(first_page);
996 first_page = NULL;
997 }
998
999 return first_page;
1000}
1001
1002static struct page *find_get_zspage(struct size_class *class)
1003{
1004 int i;
1005 struct page *page;
1006
1007 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1008 page = class->fullness_list[i];
1009 if (page)
1010 break;
1011 }
1012
1013 return page;
1014}
1015
Minchan Kim1b945ae2013-12-11 11:04:36 +09001016#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -05001017static inline int __zs_cpu_up(struct mapping_area *area)
Seth Jennings5f601902012-07-02 16:15:49 -05001018{
Seth Jenningsf5536462012-07-18 11:55:56 -05001019 /*
1020 * Make sure we don't leak memory if a cpu UP notification
1021 * and zs_init() race and both call zs_cpu_up() on the same cpu
1022 */
1023 if (area->vm)
1024 return 0;
1025 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1026 if (!area->vm)
1027 return -ENOMEM;
1028 return 0;
1029}
1030
1031static inline void __zs_cpu_down(struct mapping_area *area)
1032{
1033 if (area->vm)
1034 free_vm_area(area->vm);
1035 area->vm = NULL;
1036}
1037
1038static inline void *__zs_map_object(struct mapping_area *area,
1039 struct page *pages[2], int off, int size)
1040{
WANG Chaof6f8ed42014-08-06 16:06:58 -07001041 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
Seth Jenningsf5536462012-07-18 11:55:56 -05001042 area->vm_addr = area->vm->addr;
1043 return area->vm_addr + off;
1044}
1045
1046static inline void __zs_unmap_object(struct mapping_area *area,
1047 struct page *pages[2], int off, int size)
1048{
1049 unsigned long addr = (unsigned long)area->vm_addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001050
Joerg Roedeld95abbb2013-03-27 01:43:14 +01001051 unmap_kernel_range(addr, PAGE_SIZE * 2);
Seth Jenningsf5536462012-07-18 11:55:56 -05001052}
1053
Minchan Kim1b945ae2013-12-11 11:04:36 +09001054#else /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001055
1056static inline int __zs_cpu_up(struct mapping_area *area)
1057{
1058 /*
1059 * Make sure we don't leak memory if a cpu UP notification
1060 * and zs_init() race and both call zs_cpu_up() on the same cpu
1061 */
1062 if (area->vm_buf)
1063 return 0;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001064 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
Seth Jenningsf5536462012-07-18 11:55:56 -05001065 if (!area->vm_buf)
1066 return -ENOMEM;
1067 return 0;
1068}
1069
1070static inline void __zs_cpu_down(struct mapping_area *area)
1071{
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001072 kfree(area->vm_buf);
Seth Jenningsf5536462012-07-18 11:55:56 -05001073 area->vm_buf = NULL;
1074}
1075
1076static void *__zs_map_object(struct mapping_area *area,
1077 struct page *pages[2], int off, int size)
1078{
Seth Jennings5f601902012-07-02 16:15:49 -05001079 int sizes[2];
1080 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001081 char *buf = area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001082
Seth Jenningsf5536462012-07-18 11:55:56 -05001083 /* disable page faults to match kmap_atomic() return conditions */
1084 pagefault_disable();
1085
1086 /* no read fastpath */
1087 if (area->vm_mm == ZS_MM_WO)
1088 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001089
1090 sizes[0] = PAGE_SIZE - off;
1091 sizes[1] = size - sizes[0];
1092
Seth Jennings5f601902012-07-02 16:15:49 -05001093 /* copy object to per-cpu buffer */
1094 addr = kmap_atomic(pages[0]);
1095 memcpy(buf, addr + off, sizes[0]);
1096 kunmap_atomic(addr);
1097 addr = kmap_atomic(pages[1]);
1098 memcpy(buf + sizes[0], addr, sizes[1]);
1099 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001100out:
1101 return area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001102}
1103
Seth Jenningsf5536462012-07-18 11:55:56 -05001104static void __zs_unmap_object(struct mapping_area *area,
1105 struct page *pages[2], int off, int size)
Seth Jennings5f601902012-07-02 16:15:49 -05001106{
Seth Jennings5f601902012-07-02 16:15:49 -05001107 int sizes[2];
1108 void *addr;
Minchan Kim2e40e162015-04-15 16:15:23 -07001109 char *buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001110
Seth Jenningsf5536462012-07-18 11:55:56 -05001111 /* no write fastpath */
1112 if (area->vm_mm == ZS_MM_RO)
1113 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001114
Minchan Kim7b60a682015-04-15 16:15:39 -07001115 buf = area->vm_buf;
1116 if (!area->huge) {
1117 buf = buf + ZS_HANDLE_SIZE;
1118 size -= ZS_HANDLE_SIZE;
1119 off += ZS_HANDLE_SIZE;
1120 }
Minchan Kim2e40e162015-04-15 16:15:23 -07001121
Seth Jennings5f601902012-07-02 16:15:49 -05001122 sizes[0] = PAGE_SIZE - off;
1123 sizes[1] = size - sizes[0];
1124
1125 /* copy per-cpu buffer to object */
1126 addr = kmap_atomic(pages[0]);
1127 memcpy(addr + off, buf, sizes[0]);
1128 kunmap_atomic(addr);
1129 addr = kmap_atomic(pages[1]);
1130 memcpy(addr, buf + sizes[0], sizes[1]);
1131 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001132
1133out:
1134 /* enable page faults to match kunmap_atomic() return conditions */
1135 pagefault_enable();
Seth Jennings5f601902012-07-02 16:15:49 -05001136}
Nitin Gupta61989a82012-01-09 16:51:56 -06001137
Minchan Kim1b945ae2013-12-11 11:04:36 +09001138#endif /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001139
Nitin Gupta61989a82012-01-09 16:51:56 -06001140static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1141 void *pcpu)
1142{
Seth Jenningsf5536462012-07-18 11:55:56 -05001143 int ret, cpu = (long)pcpu;
Nitin Gupta61989a82012-01-09 16:51:56 -06001144 struct mapping_area *area;
1145
1146 switch (action) {
1147 case CPU_UP_PREPARE:
1148 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001149 ret = __zs_cpu_up(area);
1150 if (ret)
1151 return notifier_from_errno(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001152 break;
1153 case CPU_DEAD:
1154 case CPU_UP_CANCELED:
1155 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001156 __zs_cpu_down(area);
Nitin Gupta61989a82012-01-09 16:51:56 -06001157 break;
1158 }
1159
1160 return NOTIFY_OK;
1161}
1162
1163static struct notifier_block zs_cpu_nb = {
1164 .notifier_call = zs_cpu_notifier
1165};
1166
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001167static int zs_register_cpu_notifier(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001168{
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001169 int cpu, uninitialized_var(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001170
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301171 cpu_notifier_register_begin();
1172
1173 __register_cpu_notifier(&zs_cpu_nb);
Nitin Gupta61989a82012-01-09 16:51:56 -06001174 for_each_online_cpu(cpu) {
1175 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001176 if (notifier_to_errno(ret))
1177 break;
Nitin Gupta61989a82012-01-09 16:51:56 -06001178 }
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301179
1180 cpu_notifier_register_done();
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001181 return notifier_to_errno(ret);
1182}
1183
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001184static void zs_unregister_cpu_notifier(void)
1185{
1186 int cpu;
1187
1188 cpu_notifier_register_begin();
1189
1190 for_each_online_cpu(cpu)
1191 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1192 __unregister_cpu_notifier(&zs_cpu_nb);
1193
1194 cpu_notifier_register_done();
1195}
1196
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001197static void init_zs_size_classes(void)
1198{
1199 int nr;
1200
1201 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1202 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1203 nr += 1;
1204
1205 zs_size_classes = nr;
1206}
1207
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001208static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1209{
1210 if (prev->pages_per_zspage != pages_per_zspage)
1211 return false;
1212
1213 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1214 != get_maxobj_per_zspage(size, pages_per_zspage))
1215 return false;
1216
1217 return true;
1218}
1219
Minchan Kim312fcae2015-04-15 16:15:30 -07001220static bool zspage_full(struct page *page)
1221{
1222 BUG_ON(!is_first_page(page));
1223
1224 return page->inuse == page->objects;
1225}
1226
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001227unsigned long zs_get_total_pages(struct zs_pool *pool)
1228{
1229 return atomic_long_read(&pool->pages_allocated);
1230}
1231EXPORT_SYMBOL_GPL(zs_get_total_pages);
1232
1233/**
1234 * zs_map_object - get address of allocated object from handle.
1235 * @pool: pool from which the object was allocated
1236 * @handle: handle returned from zs_malloc
1237 *
1238 * Before using an object allocated from zs_malloc, it must be mapped using
1239 * this function. When done with the object, it must be unmapped using
1240 * zs_unmap_object.
1241 *
1242 * Only one object can be mapped per cpu at a time. There is no protection
1243 * against nested mappings.
1244 *
1245 * This function returns with preemption and page faults disabled.
1246 */
1247void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1248 enum zs_mapmode mm)
1249{
1250 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001251 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001252
1253 unsigned int class_idx;
1254 enum fullness_group fg;
1255 struct size_class *class;
1256 struct mapping_area *area;
1257 struct page *pages[2];
Minchan Kim2e40e162015-04-15 16:15:23 -07001258 void *ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001259
1260 BUG_ON(!handle);
1261
1262 /*
1263 * Because we use per-cpu mapping areas shared among the
1264 * pools/users, we can't allow mapping in interrupt context
1265 * because it can corrupt another users mappings.
1266 */
1267 BUG_ON(in_interrupt());
1268
Minchan Kim312fcae2015-04-15 16:15:30 -07001269 /* From now on, migration cannot move the object */
1270 pin_tag(handle);
1271
Minchan Kim2e40e162015-04-15 16:15:23 -07001272 obj = handle_to_obj(handle);
1273 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001274 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1275 class = pool->size_class[class_idx];
1276 off = obj_idx_to_offset(page, obj_idx, class->size);
1277
1278 area = &get_cpu_var(zs_map_area);
1279 area->vm_mm = mm;
1280 if (off + class->size <= PAGE_SIZE) {
1281 /* this object is contained entirely within a page */
1282 area->vm_addr = kmap_atomic(page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001283 ret = area->vm_addr + off;
1284 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001285 }
1286
1287 /* this object spans two pages */
1288 pages[0] = page;
1289 pages[1] = get_next_page(page);
1290 BUG_ON(!pages[1]);
1291
Minchan Kim2e40e162015-04-15 16:15:23 -07001292 ret = __zs_map_object(area, pages, off, class->size);
1293out:
Minchan Kim7b60a682015-04-15 16:15:39 -07001294 if (!class->huge)
1295 ret += ZS_HANDLE_SIZE;
1296
1297 return ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001298}
1299EXPORT_SYMBOL_GPL(zs_map_object);
1300
1301void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1302{
1303 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001304 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001305
1306 unsigned int class_idx;
1307 enum fullness_group fg;
1308 struct size_class *class;
1309 struct mapping_area *area;
1310
1311 BUG_ON(!handle);
1312
Minchan Kim2e40e162015-04-15 16:15:23 -07001313 obj = handle_to_obj(handle);
1314 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001315 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1316 class = pool->size_class[class_idx];
1317 off = obj_idx_to_offset(page, obj_idx, class->size);
1318
1319 area = this_cpu_ptr(&zs_map_area);
1320 if (off + class->size <= PAGE_SIZE)
1321 kunmap_atomic(area->vm_addr);
1322 else {
1323 struct page *pages[2];
1324
1325 pages[0] = page;
1326 pages[1] = get_next_page(page);
1327 BUG_ON(!pages[1]);
1328
1329 __zs_unmap_object(area, pages, off, class->size);
1330 }
1331 put_cpu_var(zs_map_area);
Minchan Kim312fcae2015-04-15 16:15:30 -07001332 unpin_tag(handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001333}
1334EXPORT_SYMBOL_GPL(zs_unmap_object);
1335
Minchan Kimc7806262015-04-15 16:15:26 -07001336static unsigned long obj_malloc(struct page *first_page,
1337 struct size_class *class, unsigned long handle)
1338{
1339 unsigned long obj;
1340 struct link_free *link;
1341
1342 struct page *m_page;
1343 unsigned long m_objidx, m_offset;
1344 void *vaddr;
1345
Minchan Kim312fcae2015-04-15 16:15:30 -07001346 handle |= OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001347 obj = (unsigned long)first_page->freelist;
1348 obj_to_location(obj, &m_page, &m_objidx);
1349 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1350
1351 vaddr = kmap_atomic(m_page);
1352 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1353 first_page->freelist = link->next;
Minchan Kim7b60a682015-04-15 16:15:39 -07001354 if (!class->huge)
1355 /* record handle in the header of allocated chunk */
1356 link->handle = handle;
1357 else
1358 /* record handle in first_page->private */
1359 set_page_private(first_page, handle);
Minchan Kimc7806262015-04-15 16:15:26 -07001360 kunmap_atomic(vaddr);
1361 first_page->inuse++;
1362 zs_stat_inc(class, OBJ_USED, 1);
1363
1364 return obj;
1365}
1366
1367
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001368/**
1369 * zs_malloc - Allocate block of given size from pool.
1370 * @pool: pool to allocate from
1371 * @size: size of block to allocate
1372 *
1373 * On success, handle to the allocated object is returned,
1374 * otherwise 0.
1375 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1376 */
1377unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1378{
Minchan Kim2e40e162015-04-15 16:15:23 -07001379 unsigned long handle, obj;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001380 struct size_class *class;
Minchan Kimc7806262015-04-15 16:15:26 -07001381 struct page *first_page;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001382
Minchan Kim7b60a682015-04-15 16:15:39 -07001383 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001384 return 0;
1385
Minchan Kim2e40e162015-04-15 16:15:23 -07001386 handle = alloc_handle(pool);
1387 if (!handle)
1388 return 0;
1389
1390 /* extra space in chunk to keep the handle */
1391 size += ZS_HANDLE_SIZE;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001392 class = pool->size_class[get_size_class_index(size)];
1393
1394 spin_lock(&class->lock);
1395 first_page = find_get_zspage(class);
1396
1397 if (!first_page) {
1398 spin_unlock(&class->lock);
1399 first_page = alloc_zspage(class, pool->flags);
Minchan Kim2e40e162015-04-15 16:15:23 -07001400 if (unlikely(!first_page)) {
1401 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001402 return 0;
Minchan Kim2e40e162015-04-15 16:15:23 -07001403 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001404
1405 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1406 atomic_long_add(class->pages_per_zspage,
1407 &pool->pages_allocated);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001408
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001409 spin_lock(&class->lock);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001410 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1411 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001412 }
1413
Minchan Kimc7806262015-04-15 16:15:26 -07001414 obj = obj_malloc(first_page, class, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001415 /* Now move the zspage to another fullness group, if required */
Minchan Kimc7806262015-04-15 16:15:26 -07001416 fix_fullness_group(class, first_page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001417 record_obj(handle, obj);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001418 spin_unlock(&class->lock);
1419
Minchan Kim2e40e162015-04-15 16:15:23 -07001420 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001421}
1422EXPORT_SYMBOL_GPL(zs_malloc);
1423
Minchan Kimc7806262015-04-15 16:15:26 -07001424static void obj_free(struct zs_pool *pool, struct size_class *class,
1425 unsigned long obj)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001426{
1427 struct link_free *link;
1428 struct page *first_page, *f_page;
Minchan Kimc7806262015-04-15 16:15:26 -07001429 unsigned long f_objidx, f_offset;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001430 void *vaddr;
Minchan Kimc7806262015-04-15 16:15:26 -07001431 int class_idx;
1432 enum fullness_group fullness;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001433
Minchan Kimc7806262015-04-15 16:15:26 -07001434 BUG_ON(!obj);
1435
Minchan Kim312fcae2015-04-15 16:15:30 -07001436 obj &= ~OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001437 obj_to_location(obj, &f_page, &f_objidx);
1438 first_page = get_first_page(f_page);
1439
1440 get_zspage_mapping(first_page, &class_idx, &fullness);
1441 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1442
1443 vaddr = kmap_atomic(f_page);
1444
1445 /* Insert this object in containing zspage's freelist */
1446 link = (struct link_free *)(vaddr + f_offset);
1447 link->next = first_page->freelist;
Minchan Kim7b60a682015-04-15 16:15:39 -07001448 if (class->huge)
1449 set_page_private(first_page, 0);
Minchan Kimc7806262015-04-15 16:15:26 -07001450 kunmap_atomic(vaddr);
1451 first_page->freelist = (void *)obj;
1452 first_page->inuse--;
1453 zs_stat_dec(class, OBJ_USED, 1);
1454}
1455
1456void zs_free(struct zs_pool *pool, unsigned long handle)
1457{
1458 struct page *first_page, *f_page;
1459 unsigned long obj, f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001460 int class_idx;
1461 struct size_class *class;
1462 enum fullness_group fullness;
1463
Minchan Kim2e40e162015-04-15 16:15:23 -07001464 if (unlikely(!handle))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001465 return;
1466
Minchan Kim312fcae2015-04-15 16:15:30 -07001467 pin_tag(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001468 obj = handle_to_obj(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001469 obj_to_location(obj, &f_page, &f_objidx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001470 first_page = get_first_page(f_page);
1471
1472 get_zspage_mapping(first_page, &class_idx, &fullness);
1473 class = pool->size_class[class_idx];
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001474
1475 spin_lock(&class->lock);
Minchan Kimc7806262015-04-15 16:15:26 -07001476 obj_free(pool, class, obj);
1477 fullness = fix_fullness_group(class, first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001478 if (fullness == ZS_EMPTY) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001479 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1480 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001481 atomic_long_sub(class->pages_per_zspage,
1482 &pool->pages_allocated);
1483 free_zspage(first_page);
1484 }
Minchan Kim312fcae2015-04-15 16:15:30 -07001485 spin_unlock(&class->lock);
1486 unpin_tag(handle);
1487
1488 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001489}
1490EXPORT_SYMBOL_GPL(zs_free);
1491
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001492static void zs_object_copy(unsigned long dst, unsigned long src,
Minchan Kim312fcae2015-04-15 16:15:30 -07001493 struct size_class *class)
1494{
1495 struct page *s_page, *d_page;
1496 unsigned long s_objidx, d_objidx;
1497 unsigned long s_off, d_off;
1498 void *s_addr, *d_addr;
1499 int s_size, d_size, size;
1500 int written = 0;
1501
1502 s_size = d_size = class->size;
1503
1504 obj_to_location(src, &s_page, &s_objidx);
1505 obj_to_location(dst, &d_page, &d_objidx);
1506
1507 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1508 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1509
1510 if (s_off + class->size > PAGE_SIZE)
1511 s_size = PAGE_SIZE - s_off;
1512
1513 if (d_off + class->size > PAGE_SIZE)
1514 d_size = PAGE_SIZE - d_off;
1515
1516 s_addr = kmap_atomic(s_page);
1517 d_addr = kmap_atomic(d_page);
1518
1519 while (1) {
1520 size = min(s_size, d_size);
1521 memcpy(d_addr + d_off, s_addr + s_off, size);
1522 written += size;
1523
1524 if (written == class->size)
1525 break;
1526
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001527 s_off += size;
1528 s_size -= size;
1529 d_off += size;
1530 d_size -= size;
1531
1532 if (s_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001533 kunmap_atomic(d_addr);
1534 kunmap_atomic(s_addr);
1535 s_page = get_next_page(s_page);
1536 BUG_ON(!s_page);
1537 s_addr = kmap_atomic(s_page);
1538 d_addr = kmap_atomic(d_page);
1539 s_size = class->size - written;
1540 s_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001541 }
1542
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001543 if (d_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001544 kunmap_atomic(d_addr);
1545 d_page = get_next_page(d_page);
1546 BUG_ON(!d_page);
1547 d_addr = kmap_atomic(d_page);
1548 d_size = class->size - written;
1549 d_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001550 }
1551 }
1552
1553 kunmap_atomic(d_addr);
1554 kunmap_atomic(s_addr);
1555}
1556
1557/*
1558 * Find alloced object in zspage from index object and
1559 * return handle.
1560 */
1561static unsigned long find_alloced_obj(struct page *page, int index,
1562 struct size_class *class)
1563{
1564 unsigned long head;
1565 int offset = 0;
1566 unsigned long handle = 0;
1567 void *addr = kmap_atomic(page);
1568
1569 if (!is_first_page(page))
1570 offset = page->index;
1571 offset += class->size * index;
1572
1573 while (offset < PAGE_SIZE) {
Minchan Kim7b60a682015-04-15 16:15:39 -07001574 head = obj_to_head(class, page, addr + offset);
Minchan Kim312fcae2015-04-15 16:15:30 -07001575 if (head & OBJ_ALLOCATED_TAG) {
1576 handle = head & ~OBJ_ALLOCATED_TAG;
1577 if (trypin_tag(handle))
1578 break;
1579 handle = 0;
1580 }
1581
1582 offset += class->size;
1583 index++;
1584 }
1585
1586 kunmap_atomic(addr);
1587 return handle;
1588}
1589
1590struct zs_compact_control {
1591 /* Source page for migration which could be a subpage of zspage. */
1592 struct page *s_page;
1593 /* Destination page for migration which should be a first page
1594 * of zspage. */
1595 struct page *d_page;
1596 /* Starting object index within @s_page which used for live object
1597 * in the subpage. */
1598 int index;
Minchan Kim312fcae2015-04-15 16:15:30 -07001599};
1600
1601static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1602 struct zs_compact_control *cc)
1603{
1604 unsigned long used_obj, free_obj;
1605 unsigned long handle;
1606 struct page *s_page = cc->s_page;
1607 struct page *d_page = cc->d_page;
1608 unsigned long index = cc->index;
Minchan Kim312fcae2015-04-15 16:15:30 -07001609 int ret = 0;
1610
1611 while (1) {
1612 handle = find_alloced_obj(s_page, index, class);
1613 if (!handle) {
1614 s_page = get_next_page(s_page);
1615 if (!s_page)
1616 break;
1617 index = 0;
1618 continue;
1619 }
1620
1621 /* Stop if there is no more space */
1622 if (zspage_full(d_page)) {
1623 unpin_tag(handle);
1624 ret = -ENOMEM;
1625 break;
1626 }
1627
1628 used_obj = handle_to_obj(handle);
1629 free_obj = obj_malloc(d_page, class, handle);
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001630 zs_object_copy(free_obj, used_obj, class);
Minchan Kim312fcae2015-04-15 16:15:30 -07001631 index++;
1632 record_obj(handle, free_obj);
1633 unpin_tag(handle);
1634 obj_free(pool, class, used_obj);
Minchan Kim312fcae2015-04-15 16:15:30 -07001635 }
1636
1637 /* Remember last position in this iteration */
1638 cc->s_page = s_page;
1639 cc->index = index;
Minchan Kim312fcae2015-04-15 16:15:30 -07001640
1641 return ret;
1642}
1643
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001644static struct page *isolate_target_page(struct size_class *class)
Minchan Kim312fcae2015-04-15 16:15:30 -07001645{
1646 int i;
1647 struct page *page;
1648
1649 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1650 page = class->fullness_list[i];
1651 if (page) {
1652 remove_zspage(page, class, i);
1653 break;
1654 }
1655 }
1656
1657 return page;
1658}
1659
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001660/*
1661 * putback_zspage - add @first_page into right class's fullness list
1662 * @pool: target pool
1663 * @class: destination class
1664 * @first_page: target page
1665 *
1666 * Return @fist_page's fullness_group
1667 */
1668static enum fullness_group putback_zspage(struct zs_pool *pool,
1669 struct size_class *class,
1670 struct page *first_page)
Minchan Kim312fcae2015-04-15 16:15:30 -07001671{
Minchan Kim312fcae2015-04-15 16:15:30 -07001672 enum fullness_group fullness;
1673
1674 BUG_ON(!is_first_page(first_page));
1675
Minchan Kim839373e2015-04-15 16:16:18 -07001676 fullness = get_fullness_group(first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001677 insert_zspage(first_page, class, fullness);
Minchan Kim839373e2015-04-15 16:16:18 -07001678 set_zspage_mapping(first_page, class->index, fullness);
1679
Minchan Kim312fcae2015-04-15 16:15:30 -07001680 if (fullness == ZS_EMPTY) {
1681 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1682 class->size, class->pages_per_zspage));
1683 atomic_long_sub(class->pages_per_zspage,
1684 &pool->pages_allocated);
1685
1686 free_zspage(first_page);
1687 }
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001688
1689 return fullness;
Minchan Kim312fcae2015-04-15 16:15:30 -07001690}
1691
1692static struct page *isolate_source_page(struct size_class *class)
1693{
Minchan Kimad9d5e12015-09-08 15:04:47 -07001694 int i;
1695 struct page *page = NULL;
Minchan Kim312fcae2015-04-15 16:15:30 -07001696
Minchan Kimad9d5e12015-09-08 15:04:47 -07001697 for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) {
1698 page = class->fullness_list[i];
1699 if (!page)
1700 continue;
1701
1702 remove_zspage(page, class, i);
1703 break;
1704 }
Minchan Kim312fcae2015-04-15 16:15:30 -07001705
1706 return page;
1707}
1708
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001709/*
1710 *
1711 * Based on the number of unused allocated objects calculate
1712 * and return the number of pages that we can free.
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001713 */
1714static unsigned long zs_can_compact(struct size_class *class)
1715{
1716 unsigned long obj_wasted;
1717
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001718 obj_wasted = zs_stat_get(class, OBJ_ALLOCATED) -
1719 zs_stat_get(class, OBJ_USED);
1720
1721 obj_wasted /= get_maxobj_per_zspage(class->size,
1722 class->pages_per_zspage);
1723
Minchan Kim6cbf16b2015-09-08 15:04:49 -07001724 return obj_wasted * class->pages_per_zspage;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001725}
1726
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001727static void __zs_compact(struct zs_pool *pool, struct size_class *class)
Minchan Kim312fcae2015-04-15 16:15:30 -07001728{
Minchan Kim312fcae2015-04-15 16:15:30 -07001729 struct zs_compact_control cc;
1730 struct page *src_page;
1731 struct page *dst_page = NULL;
Minchan Kim312fcae2015-04-15 16:15:30 -07001732
Minchan Kim312fcae2015-04-15 16:15:30 -07001733 spin_lock(&class->lock);
1734 while ((src_page = isolate_source_page(class))) {
1735
1736 BUG_ON(!is_first_page(src_page));
1737
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001738 if (!zs_can_compact(class))
1739 break;
1740
Minchan Kim312fcae2015-04-15 16:15:30 -07001741 cc.index = 0;
1742 cc.s_page = src_page;
1743
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001744 while ((dst_page = isolate_target_page(class))) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001745 cc.d_page = dst_page;
1746 /*
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001747 * If there is no more space in dst_page, resched
1748 * and see if anyone had allocated another zspage.
Minchan Kim312fcae2015-04-15 16:15:30 -07001749 */
1750 if (!migrate_zspage(pool, class, &cc))
1751 break;
1752
1753 putback_zspage(pool, class, dst_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001754 }
1755
1756 /* Stop if we couldn't find slot */
1757 if (dst_page == NULL)
1758 break;
1759
1760 putback_zspage(pool, class, dst_page);
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001761 if (putback_zspage(pool, class, src_page) == ZS_EMPTY)
Minchan Kim6cbf16b2015-09-08 15:04:49 -07001762 pool->stats.pages_compacted += class->pages_per_zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001763 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07001764 cond_resched();
1765 spin_lock(&class->lock);
1766 }
1767
1768 if (src_page)
1769 putback_zspage(pool, class, src_page);
1770
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001771 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07001772}
1773
1774unsigned long zs_compact(struct zs_pool *pool)
1775{
1776 int i;
Minchan Kim312fcae2015-04-15 16:15:30 -07001777 struct size_class *class;
1778
1779 for (i = zs_size_classes - 1; i >= 0; i--) {
1780 class = pool->size_class[i];
1781 if (!class)
1782 continue;
1783 if (class->index != i)
1784 continue;
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001785 __zs_compact(pool, class);
Minchan Kim312fcae2015-04-15 16:15:30 -07001786 }
1787
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001788 return pool->stats.pages_compacted;
Minchan Kim312fcae2015-04-15 16:15:30 -07001789}
1790EXPORT_SYMBOL_GPL(zs_compact);
1791
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001792void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
1793{
1794 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
1795}
1796EXPORT_SYMBOL_GPL(zs_pool_stats);
1797
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001798static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
1799 struct shrink_control *sc)
1800{
1801 unsigned long pages_freed;
1802 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1803 shrinker);
1804
1805 pages_freed = pool->stats.pages_compacted;
1806 /*
1807 * Compact classes and calculate compaction delta.
1808 * Can run concurrently with a manually triggered
1809 * (by user) compaction.
1810 */
1811 pages_freed = zs_compact(pool) - pages_freed;
1812
1813 return pages_freed ? pages_freed : SHRINK_STOP;
1814}
1815
1816static unsigned long zs_shrinker_count(struct shrinker *shrinker,
1817 struct shrink_control *sc)
1818{
1819 int i;
1820 struct size_class *class;
1821 unsigned long pages_to_free = 0;
1822 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1823 shrinker);
1824
1825 if (!pool->shrinker_enabled)
1826 return 0;
1827
1828 for (i = zs_size_classes - 1; i >= 0; i--) {
1829 class = pool->size_class[i];
1830 if (!class)
1831 continue;
1832 if (class->index != i)
1833 continue;
1834
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001835 pages_to_free += zs_can_compact(class);
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001836 }
1837
1838 return pages_to_free;
1839}
1840
1841static void zs_unregister_shrinker(struct zs_pool *pool)
1842{
1843 if (pool->shrinker_enabled) {
1844 unregister_shrinker(&pool->shrinker);
1845 pool->shrinker_enabled = false;
1846 }
1847}
1848
1849static int zs_register_shrinker(struct zs_pool *pool)
1850{
1851 pool->shrinker.scan_objects = zs_shrinker_scan;
1852 pool->shrinker.count_objects = zs_shrinker_count;
1853 pool->shrinker.batch = 0;
1854 pool->shrinker.seeks = DEFAULT_SEEKS;
1855
1856 return register_shrinker(&pool->shrinker);
1857}
1858
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001859/**
1860 * zs_create_pool - Creates an allocation pool to work from.
Seth Jennings0d145a52013-01-30 09:36:52 -06001861 * @flags: allocation flags used to allocate pool metadata
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001862 *
1863 * This function must be called before anything when using
1864 * the zsmalloc allocator.
1865 *
1866 * On success, a pointer to the newly created pool is returned,
1867 * otherwise NULL.
1868 */
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -08001869struct zs_pool *zs_create_pool(char *name, gfp_t flags)
Nitin Gupta61989a82012-01-09 16:51:56 -06001870{
Ganesh Mahendran18136652014-12-12 16:57:10 -08001871 int i;
Nitin Gupta61989a82012-01-09 16:51:56 -06001872 struct zs_pool *pool;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001873 struct size_class *prev_class = NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001874
Ganesh Mahendran18136652014-12-12 16:57:10 -08001875 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
Nitin Gupta61989a82012-01-09 16:51:56 -06001876 if (!pool)
1877 return NULL;
1878
Minchan Kim2e40e162015-04-15 16:15:23 -07001879 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1880 GFP_KERNEL);
1881 if (!pool->size_class) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001882 kfree(pool);
1883 return NULL;
1884 }
1885
Minchan Kim2e40e162015-04-15 16:15:23 -07001886 pool->name = kstrdup(name, GFP_KERNEL);
1887 if (!pool->name)
1888 goto err;
1889
1890 if (create_handle_cache(pool))
1891 goto err;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001892
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001893 /*
1894 * Iterate reversly, because, size of size_class that we want to use
1895 * for merging should be larger or equal to current size.
1896 */
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001897 for (i = zs_size_classes - 1; i >= 0; i--) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001898 int size;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001899 int pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001900 struct size_class *class;
1901
1902 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1903 if (size > ZS_MAX_ALLOC_SIZE)
1904 size = ZS_MAX_ALLOC_SIZE;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001905 pages_per_zspage = get_pages_per_zspage(size);
Nitin Gupta61989a82012-01-09 16:51:56 -06001906
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001907 /*
1908 * size_class is used for normal zsmalloc operation such
1909 * as alloc/free for that size. Although it is natural that we
1910 * have one size_class for each size, there is a chance that we
1911 * can get more memory utilization if we use one size_class for
1912 * many different sizes whose size_class have same
1913 * characteristics. So, we makes size_class point to
1914 * previous size_class if possible.
1915 */
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001916 if (prev_class) {
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001917 if (can_merge(prev_class, size, pages_per_zspage)) {
1918 pool->size_class[i] = prev_class;
1919 continue;
1920 }
1921 }
1922
1923 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1924 if (!class)
1925 goto err;
1926
Nitin Gupta61989a82012-01-09 16:51:56 -06001927 class->size = size;
1928 class->index = i;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001929 class->pages_per_zspage = pages_per_zspage;
Minchan Kim7b60a682015-04-15 16:15:39 -07001930 if (pages_per_zspage == 1 &&
1931 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1932 class->huge = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06001933 spin_lock_init(&class->lock);
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001934 pool->size_class[i] = class;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001935
1936 prev_class = class;
Nitin Gupta61989a82012-01-09 16:51:56 -06001937 }
1938
Nitin Gupta61989a82012-01-09 16:51:56 -06001939 pool->flags = flags;
Nitin Gupta61989a82012-01-09 16:51:56 -06001940
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001941 if (zs_pool_stat_create(name, pool))
1942 goto err;
1943
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001944 /*
1945 * Not critical, we still can use the pool
1946 * and user can trigger compaction manually.
1947 */
1948 if (zs_register_shrinker(pool) == 0)
1949 pool->shrinker_enabled = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06001950 return pool;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001951
1952err:
1953 zs_destroy_pool(pool);
1954 return NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001955}
1956EXPORT_SYMBOL_GPL(zs_create_pool);
1957
1958void zs_destroy_pool(struct zs_pool *pool)
1959{
1960 int i;
1961
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001962 zs_unregister_shrinker(pool);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001963 zs_pool_stat_destroy(pool);
1964
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001965 for (i = 0; i < zs_size_classes; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001966 int fg;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001967 struct size_class *class = pool->size_class[i];
1968
1969 if (!class)
1970 continue;
1971
1972 if (class->index != i)
1973 continue;
Nitin Gupta61989a82012-01-09 16:51:56 -06001974
1975 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1976 if (class->fullness_list[fg]) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -04001977 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -06001978 class->size, fg);
1979 }
1980 }
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001981 kfree(class);
Nitin Gupta61989a82012-01-09 16:51:56 -06001982 }
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001983
Minchan Kim2e40e162015-04-15 16:15:23 -07001984 destroy_handle_cache(pool);
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001985 kfree(pool->size_class);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001986 kfree(pool->name);
Nitin Gupta61989a82012-01-09 16:51:56 -06001987 kfree(pool);
1988}
1989EXPORT_SYMBOL_GPL(zs_destroy_pool);
1990
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001991static int __init zs_init(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001992{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001993 int ret = zs_register_cpu_notifier();
Nitin Gupta61989a82012-01-09 16:51:56 -06001994
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001995 if (ret)
1996 goto notifier_fail;
Nitin Gupta61989a82012-01-09 16:51:56 -06001997
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001998 init_zs_size_classes();
Nitin Gupta61989a82012-01-09 16:51:56 -06001999
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002000#ifdef CONFIG_ZPOOL
2001 zpool_register_driver(&zs_zpool_driver);
2002#endif
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002003
2004 ret = zs_stat_init();
2005 if (ret) {
2006 pr_err("zs stat initialization failed\n");
2007 goto stat_fail;
2008 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002009 return 0;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002010
2011stat_fail:
2012#ifdef CONFIG_ZPOOL
2013 zpool_unregister_driver(&zs_zpool_driver);
2014#endif
2015notifier_fail:
2016 zs_unregister_cpu_notifier();
2017
2018 return ret;
Nitin Gupta61989a82012-01-09 16:51:56 -06002019}
Nitin Gupta61989a82012-01-09 16:51:56 -06002020
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002021static void __exit zs_exit(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002022{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002023#ifdef CONFIG_ZPOOL
2024 zpool_unregister_driver(&zs_zpool_driver);
2025#endif
2026 zs_unregister_cpu_notifier();
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002027
2028 zs_stat_exit();
Nitin Gupta61989a82012-01-09 16:51:56 -06002029}
Ben Hutchings069f1012012-06-20 02:31:11 +01002030
2031module_init(zs_init);
2032module_exit(zs_exit);
2033
2034MODULE_LICENSE("Dual BSD/GPL");
2035MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");