blob: 08bd7a3d464a9c6959a39e269d2284600e750a50 [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#ifdef CONFIG_ZSMALLOC_DEBUG
49#define DEBUG
50#endif
51
52#include <linux/module.h>
53#include <linux/kernel.h>
Minchan Kim312fcae2015-04-15 16:15:30 -070054#include <linux/sched.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060055#include <linux/bitops.h>
56#include <linux/errno.h>
57#include <linux/highmem.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060058#include <linux/string.h>
59#include <linux/slab.h>
60#include <asm/tlbflush.h>
61#include <asm/pgtable.h>
62#include <linux/cpumask.h>
63#include <linux/cpu.h>
Seth Jennings0cbb6132012-02-13 08:47:49 -060064#include <linux/vmalloc.h>
Seth Jenningsc60369f2012-07-18 11:55:55 -050065#include <linux/hardirq.h>
Seth Jennings0959c632012-08-08 15:12:17 +090066#include <linux/spinlock.h>
67#include <linux/types.h>
Ganesh Mahendran0f050d92015-02-12 15:00:54 -080068#include <linux/debugfs.h>
Minchan Kimbcf16472014-01-30 15:45:50 -080069#include <linux/zsmalloc.h>
Dan Streetmanc7957792014-08-06 16:08:38 -070070#include <linux/zpool.h>
Seth Jennings0959c632012-08-08 15:12:17 +090071
72/*
73 * This must be power of 2 and greater than of equal to sizeof(link_free).
74 * These two conditions ensure that any 'struct link_free' itself doesn't
75 * span more than 1 page which avoids complex case of mapping 2 pages simply
76 * to restore link_free pointer values.
77 */
78#define ZS_ALIGN 8
79
80/*
81 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
82 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
83 */
84#define ZS_MAX_ZSPAGE_ORDER 2
85#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
86
Minchan Kim2e40e162015-04-15 16:15:23 -070087#define ZS_HANDLE_SIZE (sizeof(unsigned long))
88
Seth Jennings0959c632012-08-08 15:12:17 +090089/*
90 * Object location (<PFN>, <obj_idx>) is encoded as
Nitin Cuptac3e3e882013-12-11 11:04:37 +090091 * as single (unsigned long) handle value.
Seth Jennings0959c632012-08-08 15:12:17 +090092 *
93 * Note that object index <obj_idx> is relative to system
94 * page <PFN> it is stored in, so for each sub-page belonging
95 * to a zspage, obj_idx starts with 0.
96 *
97 * This is made more complicated by various memory models and PAE.
98 */
99
100#ifndef MAX_PHYSMEM_BITS
101#ifdef CONFIG_HIGHMEM64G
102#define MAX_PHYSMEM_BITS 36
103#else /* !CONFIG_HIGHMEM64G */
104/*
105 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
106 * be PAGE_SHIFT
107 */
108#define MAX_PHYSMEM_BITS BITS_PER_LONG
109#endif
110#endif
111#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
Minchan Kim312fcae2015-04-15 16:15:30 -0700112
113/*
114 * Memory for allocating for handle keeps object position by
115 * encoding <page, obj_idx> and the encoded value has a room
116 * in least bit(ie, look at obj_to_location).
117 * We use the bit to synchronize between object access by
118 * user and migration.
119 */
120#define HANDLE_PIN_BIT 0
121
122/*
123 * Head in allocated object should have OBJ_ALLOCATED_TAG
124 * to identify the object was allocated or not.
125 * It's okay to add the status bit in the least bit because
126 * header keeps handle which is 4byte-aligned address so we
127 * have room for two bit at least.
128 */
129#define OBJ_ALLOCATED_TAG 1
130#define OBJ_TAG_BITS 1
131#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
Seth Jennings0959c632012-08-08 15:12:17 +0900132#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
133
134#define MAX(a, b) ((a) >= (b) ? (a) : (b))
135/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
136#define ZS_MIN_ALLOC_SIZE \
137 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
Minchan Kim2e40e162015-04-15 16:15:23 -0700138/* each chunk includes extra space to keep handle */
Minchan Kim7b60a682015-04-15 16:15:39 -0700139#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
Seth Jennings0959c632012-08-08 15:12:17 +0900140
141/*
Weijie Yang7eb52512014-06-04 16:11:08 -0700142 * On systems with 4K page size, this gives 255 size classes! There is a
Seth Jennings0959c632012-08-08 15:12:17 +0900143 * trader-off here:
144 * - Large number of size classes is potentially wasteful as free page are
145 * spread across these classes
146 * - Small number of size classes causes large internal fragmentation
147 * - Probably its better to use specific size classes (empirically
148 * determined). NOTE: all those class sizes must be set as multiple of
149 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
150 *
151 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
152 * (reason above)
153 */
Seth Jenningsd662b8e2013-01-25 11:46:18 -0600154#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
Seth Jennings0959c632012-08-08 15:12:17 +0900155
156/*
157 * We do not maintain any list for completely empty or full pages
158 */
159enum fullness_group {
160 ZS_ALMOST_FULL,
161 ZS_ALMOST_EMPTY,
162 _ZS_NR_FULLNESS_GROUPS,
163
164 ZS_EMPTY,
165 ZS_FULL
166};
167
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800168enum zs_stat_type {
169 OBJ_ALLOCATED,
170 OBJ_USED,
Minchan Kim248ca1b2015-04-15 16:15:42 -0700171 CLASS_ALMOST_FULL,
172 CLASS_ALMOST_EMPTY,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800173 NR_ZS_STAT_TYPE,
174};
175
176#ifdef CONFIG_ZSMALLOC_STAT
177
178static struct dentry *zs_stat_root;
179
180struct zs_size_stat {
181 unsigned long objs[NR_ZS_STAT_TYPE];
182};
183
184#endif
185
Seth Jennings0959c632012-08-08 15:12:17 +0900186/*
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800187 * number of size_classes
188 */
189static int zs_size_classes;
190
191/*
Seth Jennings0959c632012-08-08 15:12:17 +0900192 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
193 * n <= N / f, where
194 * n = number of allocated objects
195 * N = total number of objects zspage can store
Wang Sheng-Hui6dd97372014-10-09 15:29:59 -0700196 * f = fullness_threshold_frac
Seth Jennings0959c632012-08-08 15:12:17 +0900197 *
198 * Similarly, we assign zspage to:
199 * ZS_ALMOST_FULL when n > N / f
200 * ZS_EMPTY when n == 0
201 * ZS_FULL when n == N
202 *
203 * (see: fix_fullness_group())
204 */
205static const int fullness_threshold_frac = 4;
206
207struct size_class {
208 /*
209 * Size of objects stored in this class. Must be multiple
210 * of ZS_ALIGN.
211 */
212 int size;
213 unsigned int index;
214
215 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
216 int pages_per_zspage;
Minchan Kim7b60a682015-04-15 16:15:39 -0700217 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
218 bool huge;
Seth Jennings0959c632012-08-08 15:12:17 +0900219
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800220#ifdef CONFIG_ZSMALLOC_STAT
221 struct zs_size_stat stats;
222#endif
223
Seth Jennings0959c632012-08-08 15:12:17 +0900224 spinlock_t lock;
225
Seth Jennings0959c632012-08-08 15:12:17 +0900226 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
227};
228
229/*
230 * Placed within free objects to form a singly linked list.
231 * For every zspage, first_page->freelist gives head of this list.
232 *
233 * This must be power of 2 and less than or equal to ZS_ALIGN
234 */
235struct link_free {
Minchan Kim2e40e162015-04-15 16:15:23 -0700236 union {
237 /*
238 * Position of next free chunk (encodes <PFN, obj_idx>)
239 * It's valid for non-allocated object
240 */
241 void *next;
242 /*
243 * Handle of allocated object.
244 */
245 unsigned long handle;
246 };
Seth Jennings0959c632012-08-08 15:12:17 +0900247};
248
249struct zs_pool {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800250 char *name;
251
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800252 struct size_class **size_class;
Minchan Kim2e40e162015-04-15 16:15:23 -0700253 struct kmem_cache *handle_cachep;
Seth Jennings0959c632012-08-08 15:12:17 +0900254
255 gfp_t flags; /* allocation flags used when growing pool */
Minchan Kim13de8932014-10-09 15:29:48 -0700256 atomic_long_t pages_allocated;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800257
258#ifdef CONFIG_ZSMALLOC_STAT
259 struct dentry *stat_dentry;
260#endif
Seth Jennings0959c632012-08-08 15:12:17 +0900261};
Nitin Gupta61989a82012-01-09 16:51:56 -0600262
263/*
264 * A zspage's class index and fullness group
265 * are encoded in its (first)page->mapping
266 */
267#define CLASS_IDX_BITS 28
268#define FULLNESS_BITS 4
269#define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
270#define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
271
Seth Jenningsf5536462012-07-18 11:55:56 -0500272struct mapping_area {
Minchan Kim1b945ae2013-12-11 11:04:36 +0900273#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -0500274 struct vm_struct *vm; /* vm area for mapping object that span pages */
275#else
276 char *vm_buf; /* copy buffer for objects that span pages */
277#endif
278 char *vm_addr; /* address of kmap_atomic()'ed pages */
279 enum zs_mapmode vm_mm; /* mapping mode */
Minchan Kim7b60a682015-04-15 16:15:39 -0700280 bool huge;
Seth Jenningsf5536462012-07-18 11:55:56 -0500281};
282
Minchan Kim2e40e162015-04-15 16:15:23 -0700283static int create_handle_cache(struct zs_pool *pool)
284{
285 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
286 0, 0, NULL);
287 return pool->handle_cachep ? 0 : 1;
288}
289
290static void destroy_handle_cache(struct zs_pool *pool)
291{
292 kmem_cache_destroy(pool->handle_cachep);
293}
294
295static unsigned long alloc_handle(struct zs_pool *pool)
296{
297 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
298 pool->flags & ~__GFP_HIGHMEM);
299}
300
301static void free_handle(struct zs_pool *pool, unsigned long handle)
302{
303 kmem_cache_free(pool->handle_cachep, (void *)handle);
304}
305
306static void record_obj(unsigned long handle, unsigned long obj)
307{
308 *(unsigned long *)handle = obj;
309}
310
Dan Streetmanc7957792014-08-06 16:08:38 -0700311/* zpool driver */
312
313#ifdef CONFIG_ZPOOL
314
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -0800315static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops)
Dan Streetmanc7957792014-08-06 16:08:38 -0700316{
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -0800317 return zs_create_pool(name, gfp);
Dan Streetmanc7957792014-08-06 16:08:38 -0700318}
319
320static void zs_zpool_destroy(void *pool)
321{
322 zs_destroy_pool(pool);
323}
324
325static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
326 unsigned long *handle)
327{
328 *handle = zs_malloc(pool, size);
329 return *handle ? 0 : -1;
330}
331static void zs_zpool_free(void *pool, unsigned long handle)
332{
333 zs_free(pool, handle);
334}
335
336static int zs_zpool_shrink(void *pool, unsigned int pages,
337 unsigned int *reclaimed)
338{
339 return -EINVAL;
340}
341
342static void *zs_zpool_map(void *pool, unsigned long handle,
343 enum zpool_mapmode mm)
344{
345 enum zs_mapmode zs_mm;
346
347 switch (mm) {
348 case ZPOOL_MM_RO:
349 zs_mm = ZS_MM_RO;
350 break;
351 case ZPOOL_MM_WO:
352 zs_mm = ZS_MM_WO;
353 break;
354 case ZPOOL_MM_RW: /* fallthru */
355 default:
356 zs_mm = ZS_MM_RW;
357 break;
358 }
359
360 return zs_map_object(pool, handle, zs_mm);
361}
362static void zs_zpool_unmap(void *pool, unsigned long handle)
363{
364 zs_unmap_object(pool, handle);
365}
366
367static u64 zs_zpool_total_size(void *pool)
368{
Minchan Kim722cdc12014-10-09 15:29:50 -0700369 return zs_get_total_pages(pool) << PAGE_SHIFT;
Dan Streetmanc7957792014-08-06 16:08:38 -0700370}
371
372static struct zpool_driver zs_zpool_driver = {
373 .type = "zsmalloc",
374 .owner = THIS_MODULE,
375 .create = zs_zpool_create,
376 .destroy = zs_zpool_destroy,
377 .malloc = zs_zpool_malloc,
378 .free = zs_zpool_free,
379 .shrink = zs_zpool_shrink,
380 .map = zs_zpool_map,
381 .unmap = zs_zpool_unmap,
382 .total_size = zs_zpool_total_size,
383};
384
Kees Cook137f8cf2014-08-29 15:18:40 -0700385MODULE_ALIAS("zpool-zsmalloc");
Dan Streetmanc7957792014-08-06 16:08:38 -0700386#endif /* CONFIG_ZPOOL */
387
Minchan Kim248ca1b2015-04-15 16:15:42 -0700388static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
389{
390 return pages_per_zspage * PAGE_SIZE / size;
391}
392
Nitin Gupta61989a82012-01-09 16:51:56 -0600393/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
394static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
395
396static int is_first_page(struct page *page)
397{
Minchan Kima27545bf2012-04-25 15:23:09 +0900398 return PagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600399}
400
401static int is_last_page(struct page *page)
402{
Minchan Kima27545bf2012-04-25 15:23:09 +0900403 return PagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600404}
405
406static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
407 enum fullness_group *fullness)
408{
409 unsigned long m;
410 BUG_ON(!is_first_page(page));
411
412 m = (unsigned long)page->mapping;
413 *fullness = m & FULLNESS_MASK;
414 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
415}
416
417static void set_zspage_mapping(struct page *page, unsigned int class_idx,
418 enum fullness_group fullness)
419{
420 unsigned long m;
421 BUG_ON(!is_first_page(page));
422
423 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
424 (fullness & FULLNESS_MASK);
425 page->mapping = (struct address_space *)m;
426}
427
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900428/*
429 * zsmalloc divides the pool into various size classes where each
430 * class maintains a list of zspages where each zspage is divided
431 * into equal sized chunks. Each allocation falls into one of these
432 * classes depending on its size. This function returns index of the
433 * size class which has chunk size big enough to hold the give size.
434 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600435static int get_size_class_index(int size)
436{
437 int idx = 0;
438
439 if (likely(size > ZS_MIN_ALLOC_SIZE))
440 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
441 ZS_SIZE_CLASS_DELTA);
442
Minchan Kim7b60a682015-04-15 16:15:39 -0700443 return min(zs_size_classes - 1, idx);
Nitin Gupta61989a82012-01-09 16:51:56 -0600444}
445
Minchan Kim248ca1b2015-04-15 16:15:42 -0700446#ifdef CONFIG_ZSMALLOC_STAT
447
448static inline void zs_stat_inc(struct size_class *class,
449 enum zs_stat_type type, unsigned long cnt)
450{
451 class->stats.objs[type] += cnt;
452}
453
454static inline void zs_stat_dec(struct size_class *class,
455 enum zs_stat_type type, unsigned long cnt)
456{
457 class->stats.objs[type] -= cnt;
458}
459
460static inline unsigned long zs_stat_get(struct size_class *class,
461 enum zs_stat_type type)
462{
463 return class->stats.objs[type];
464}
465
466static int __init zs_stat_init(void)
467{
468 if (!debugfs_initialized())
469 return -ENODEV;
470
471 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
472 if (!zs_stat_root)
473 return -ENOMEM;
474
475 return 0;
476}
477
478static void __exit zs_stat_exit(void)
479{
480 debugfs_remove_recursive(zs_stat_root);
481}
482
483static int zs_stats_size_show(struct seq_file *s, void *v)
484{
485 int i;
486 struct zs_pool *pool = s->private;
487 struct size_class *class;
488 int objs_per_zspage;
489 unsigned long class_almost_full, class_almost_empty;
490 unsigned long obj_allocated, obj_used, pages_used;
491 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
492 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
493
494 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
495 "class", "size", "almost_full", "almost_empty",
496 "obj_allocated", "obj_used", "pages_used",
497 "pages_per_zspage");
498
499 for (i = 0; i < zs_size_classes; i++) {
500 class = pool->size_class[i];
501
502 if (class->index != i)
503 continue;
504
505 spin_lock(&class->lock);
506 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
507 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
508 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
509 obj_used = zs_stat_get(class, OBJ_USED);
510 spin_unlock(&class->lock);
511
512 objs_per_zspage = get_maxobj_per_zspage(class->size,
513 class->pages_per_zspage);
514 pages_used = obj_allocated / objs_per_zspage *
515 class->pages_per_zspage;
516
517 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
518 i, class->size, class_almost_full, class_almost_empty,
519 obj_allocated, obj_used, pages_used,
520 class->pages_per_zspage);
521
522 total_class_almost_full += class_almost_full;
523 total_class_almost_empty += class_almost_empty;
524 total_objs += obj_allocated;
525 total_used_objs += obj_used;
526 total_pages += pages_used;
527 }
528
529 seq_puts(s, "\n");
530 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
531 "Total", "", total_class_almost_full,
532 total_class_almost_empty, total_objs,
533 total_used_objs, total_pages);
534
535 return 0;
536}
537
538static int zs_stats_size_open(struct inode *inode, struct file *file)
539{
540 return single_open(file, zs_stats_size_show, inode->i_private);
541}
542
543static const struct file_operations zs_stat_size_ops = {
544 .open = zs_stats_size_open,
545 .read = seq_read,
546 .llseek = seq_lseek,
547 .release = single_release,
548};
549
550static int zs_pool_stat_create(char *name, struct zs_pool *pool)
551{
552 struct dentry *entry;
553
554 if (!zs_stat_root)
555 return -ENODEV;
556
557 entry = debugfs_create_dir(name, zs_stat_root);
558 if (!entry) {
559 pr_warn("debugfs dir <%s> creation failed\n", name);
560 return -ENOMEM;
561 }
562 pool->stat_dentry = entry;
563
564 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
565 pool->stat_dentry, pool, &zs_stat_size_ops);
566 if (!entry) {
567 pr_warn("%s: debugfs file entry <%s> creation failed\n",
568 name, "classes");
569 return -ENOMEM;
570 }
571
572 return 0;
573}
574
575static void zs_pool_stat_destroy(struct zs_pool *pool)
576{
577 debugfs_remove_recursive(pool->stat_dentry);
578}
579
580#else /* CONFIG_ZSMALLOC_STAT */
581
582static inline void zs_stat_inc(struct size_class *class,
583 enum zs_stat_type type, unsigned long cnt)
584{
585}
586
587static inline void zs_stat_dec(struct size_class *class,
588 enum zs_stat_type type, unsigned long cnt)
589{
590}
591
592static inline unsigned long zs_stat_get(struct size_class *class,
593 enum zs_stat_type type)
594{
595 return 0;
596}
597
598static int __init zs_stat_init(void)
599{
600 return 0;
601}
602
603static void __exit zs_stat_exit(void)
604{
605}
606
607static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
608{
609 return 0;
610}
611
612static inline void zs_pool_stat_destroy(struct zs_pool *pool)
613{
614}
615
616#endif
617
618
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900619/*
620 * For each size class, zspages are divided into different groups
621 * depending on how "full" they are. This was done so that we could
622 * easily find empty or nearly empty zspages when we try to shrink
623 * the pool (not yet implemented). This function returns fullness
624 * status of the given page.
625 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600626static enum fullness_group get_fullness_group(struct page *page)
627{
628 int inuse, max_objects;
629 enum fullness_group fg;
630 BUG_ON(!is_first_page(page));
631
632 inuse = page->inuse;
633 max_objects = page->objects;
634
635 if (inuse == 0)
636 fg = ZS_EMPTY;
637 else if (inuse == max_objects)
638 fg = ZS_FULL;
Minchan Kimd3d07c92015-04-15 16:15:33 -0700639 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
Nitin Gupta61989a82012-01-09 16:51:56 -0600640 fg = ZS_ALMOST_EMPTY;
641 else
642 fg = ZS_ALMOST_FULL;
643
644 return fg;
645}
646
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900647/*
648 * Each size class maintains various freelists and zspages are assigned
649 * to one of these freelists based on the number of live objects they
650 * have. This functions inserts the given zspage into the freelist
651 * identified by <class, fullness_group>.
652 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600653static void insert_zspage(struct page *page, struct size_class *class,
654 enum fullness_group fullness)
655{
656 struct page **head;
657
658 BUG_ON(!is_first_page(page));
659
660 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
661 return;
662
663 head = &class->fullness_list[fullness];
664 if (*head)
665 list_add_tail(&page->lru, &(*head)->lru);
666
667 *head = page;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700668 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
669 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600670}
671
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900672/*
673 * This function removes the given zspage from the freelist identified
674 * by <class, fullness_group>.
675 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600676static void remove_zspage(struct page *page, struct size_class *class,
677 enum fullness_group fullness)
678{
679 struct page **head;
680
681 BUG_ON(!is_first_page(page));
682
683 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
684 return;
685
686 head = &class->fullness_list[fullness];
687 BUG_ON(!*head);
688 if (list_empty(&(*head)->lru))
689 *head = NULL;
690 else if (*head == page)
691 *head = (struct page *)list_entry((*head)->lru.next,
692 struct page, lru);
693
694 list_del_init(&page->lru);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700695 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
696 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600697}
698
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900699/*
700 * Each size class maintains zspages in different fullness groups depending
701 * on the number of live objects they contain. When allocating or freeing
702 * objects, the fullness status of the page can change, say, from ALMOST_FULL
703 * to ALMOST_EMPTY when freeing an object. This function checks if such
704 * a status change has occurred for the given page and accordingly moves the
705 * page from the freelist of the old fullness group to that of the new
706 * fullness group.
707 */
Minchan Kimc7806262015-04-15 16:15:26 -0700708static enum fullness_group fix_fullness_group(struct size_class *class,
Nitin Gupta61989a82012-01-09 16:51:56 -0600709 struct page *page)
710{
711 int class_idx;
Nitin Gupta61989a82012-01-09 16:51:56 -0600712 enum fullness_group currfg, newfg;
713
714 BUG_ON(!is_first_page(page));
715
716 get_zspage_mapping(page, &class_idx, &currfg);
717 newfg = get_fullness_group(page);
718 if (newfg == currfg)
719 goto out;
720
Nitin Gupta61989a82012-01-09 16:51:56 -0600721 remove_zspage(page, class, currfg);
722 insert_zspage(page, class, newfg);
723 set_zspage_mapping(page, class_idx, newfg);
724
725out:
726 return newfg;
727}
728
729/*
730 * We have to decide on how many pages to link together
731 * to form a zspage for each size class. This is important
732 * to reduce wastage due to unusable space left at end of
733 * each zspage which is given as:
Yinghao Xie888fa3742015-04-15 16:15:49 -0700734 * wastage = Zp % class_size
735 * usage = Zp - wastage
Nitin Gupta61989a82012-01-09 16:51:56 -0600736 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
737 *
738 * For example, for size class of 3/8 * PAGE_SIZE, we should
739 * link together 3 PAGE_SIZE sized pages to form a zspage
740 * since then we can perfectly fit in 8 such objects.
741 */
Minchan Kim2e3b6152012-05-03 15:40:39 +0900742static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600743{
744 int i, max_usedpc = 0;
745 /* zspage order which gives maximum used size per KB */
746 int max_usedpc_order = 1;
747
Seth Jennings84d4faa2012-03-05 11:33:21 -0600748 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600749 int zspage_size;
750 int waste, usedpc;
751
752 zspage_size = i * PAGE_SIZE;
753 waste = zspage_size % class_size;
754 usedpc = (zspage_size - waste) * 100 / zspage_size;
755
756 if (usedpc > max_usedpc) {
757 max_usedpc = usedpc;
758 max_usedpc_order = i;
759 }
760 }
761
762 return max_usedpc_order;
763}
764
765/*
766 * A single 'zspage' is composed of many system pages which are
767 * linked together using fields in struct page. This function finds
768 * the first/head page, given any component page of a zspage.
769 */
770static struct page *get_first_page(struct page *page)
771{
772 if (is_first_page(page))
773 return page;
774 else
775 return page->first_page;
776}
777
778static struct page *get_next_page(struct page *page)
779{
780 struct page *next;
781
782 if (is_last_page(page))
783 next = NULL;
784 else if (is_first_page(page))
Sunghan Suhe842b972013-07-12 16:08:13 +0900785 next = (struct page *)page_private(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600786 else
787 next = list_entry(page->lru.next, struct page, lru);
788
789 return next;
790}
791
Olav Haugan67296872013-11-22 09:30:41 -0800792/*
793 * Encode <page, obj_idx> as a single handle value.
Minchan Kim312fcae2015-04-15 16:15:30 -0700794 * We use the least bit of handle for tagging.
Olav Haugan67296872013-11-22 09:30:41 -0800795 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700796static void *location_to_obj(struct page *page, unsigned long obj_idx)
Nitin Gupta61989a82012-01-09 16:51:56 -0600797{
Minchan Kim312fcae2015-04-15 16:15:30 -0700798 unsigned long obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600799
800 if (!page) {
801 BUG_ON(obj_idx);
802 return NULL;
803 }
804
Minchan Kim312fcae2015-04-15 16:15:30 -0700805 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
806 obj |= ((obj_idx) & OBJ_INDEX_MASK);
807 obj <<= OBJ_TAG_BITS;
Nitin Gupta61989a82012-01-09 16:51:56 -0600808
Minchan Kim312fcae2015-04-15 16:15:30 -0700809 return (void *)obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600810}
811
Olav Haugan67296872013-11-22 09:30:41 -0800812/*
813 * Decode <page, obj_idx> pair from the given object handle. We adjust the
814 * decoded obj_idx back to its original value since it was adjusted in
Minchan Kim312fcae2015-04-15 16:15:30 -0700815 * location_to_obj().
Olav Haugan67296872013-11-22 09:30:41 -0800816 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700817static void obj_to_location(unsigned long obj, struct page **page,
Nitin Gupta61989a82012-01-09 16:51:56 -0600818 unsigned long *obj_idx)
819{
Minchan Kim312fcae2015-04-15 16:15:30 -0700820 obj >>= OBJ_TAG_BITS;
821 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
822 *obj_idx = (obj & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600823}
824
Minchan Kim2e40e162015-04-15 16:15:23 -0700825static unsigned long handle_to_obj(unsigned long handle)
826{
827 return *(unsigned long *)handle;
828}
829
Minchan Kim7b60a682015-04-15 16:15:39 -0700830static unsigned long obj_to_head(struct size_class *class, struct page *page,
831 void *obj)
Minchan Kim312fcae2015-04-15 16:15:30 -0700832{
Minchan Kim7b60a682015-04-15 16:15:39 -0700833 if (class->huge) {
834 VM_BUG_ON(!is_first_page(page));
835 return *(unsigned long *)page_private(page);
836 } else
837 return *(unsigned long *)obj;
Minchan Kim312fcae2015-04-15 16:15:30 -0700838}
839
Nitin Gupta61989a82012-01-09 16:51:56 -0600840static unsigned long obj_idx_to_offset(struct page *page,
841 unsigned long obj_idx, int class_size)
842{
843 unsigned long off = 0;
844
845 if (!is_first_page(page))
846 off = page->index;
847
848 return off + obj_idx * class_size;
849}
850
Minchan Kim312fcae2015-04-15 16:15:30 -0700851static inline int trypin_tag(unsigned long handle)
852{
853 unsigned long *ptr = (unsigned long *)handle;
854
855 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
856}
857
858static void pin_tag(unsigned long handle)
859{
860 while (!trypin_tag(handle));
861}
862
863static void unpin_tag(unsigned long handle)
864{
865 unsigned long *ptr = (unsigned long *)handle;
866
867 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
868}
869
Nitin Guptaf4477e92012-04-02 09:13:56 -0500870static void reset_page(struct page *page)
871{
872 clear_bit(PG_private, &page->flags);
873 clear_bit(PG_private_2, &page->flags);
874 set_page_private(page, 0);
875 page->mapping = NULL;
876 page->freelist = NULL;
Mel Gorman22b751c2013-02-22 16:34:59 -0800877 page_mapcount_reset(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500878}
879
Nitin Gupta61989a82012-01-09 16:51:56 -0600880static void free_zspage(struct page *first_page)
881{
Nitin Guptaf4477e92012-04-02 09:13:56 -0500882 struct page *nextp, *tmp, *head_extra;
Nitin Gupta61989a82012-01-09 16:51:56 -0600883
884 BUG_ON(!is_first_page(first_page));
885 BUG_ON(first_page->inuse);
886
Nitin Guptaf4477e92012-04-02 09:13:56 -0500887 head_extra = (struct page *)page_private(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600888
Nitin Guptaf4477e92012-04-02 09:13:56 -0500889 reset_page(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600890 __free_page(first_page);
891
892 /* zspage with only 1 system page */
Nitin Guptaf4477e92012-04-02 09:13:56 -0500893 if (!head_extra)
Nitin Gupta61989a82012-01-09 16:51:56 -0600894 return;
895
Nitin Guptaf4477e92012-04-02 09:13:56 -0500896 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600897 list_del(&nextp->lru);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500898 reset_page(nextp);
Nitin Gupta61989a82012-01-09 16:51:56 -0600899 __free_page(nextp);
900 }
Nitin Guptaf4477e92012-04-02 09:13:56 -0500901 reset_page(head_extra);
902 __free_page(head_extra);
Nitin Gupta61989a82012-01-09 16:51:56 -0600903}
904
905/* Initialize a newly allocated zspage */
906static void init_zspage(struct page *first_page, struct size_class *class)
907{
908 unsigned long off = 0;
909 struct page *page = first_page;
910
911 BUG_ON(!is_first_page(first_page));
912 while (page) {
913 struct page *next_page;
914 struct link_free *link;
Dan Streetman5538c562014-10-09 15:30:01 -0700915 unsigned int i = 1;
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800916 void *vaddr;
Nitin Gupta61989a82012-01-09 16:51:56 -0600917
918 /*
919 * page->index stores offset of first object starting
920 * in the page. For the first page, this is always 0,
921 * so we use first_page->index (aka ->freelist) to store
922 * head of corresponding zspage's freelist.
923 */
924 if (page != first_page)
925 page->index = off;
926
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800927 vaddr = kmap_atomic(page);
928 link = (struct link_free *)vaddr + off / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600929
Dan Streetman5538c562014-10-09 15:30:01 -0700930 while ((off += class->size) < PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -0700931 link->next = location_to_obj(page, i++);
Dan Streetman5538c562014-10-09 15:30:01 -0700932 link += class->size / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600933 }
934
935 /*
936 * We now come to the last (full or partial) object on this
937 * page, which must point to the first object on the next
938 * page (if present)
939 */
940 next_page = get_next_page(page);
Minchan Kim312fcae2015-04-15 16:15:30 -0700941 link->next = location_to_obj(next_page, 0);
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800942 kunmap_atomic(vaddr);
Nitin Gupta61989a82012-01-09 16:51:56 -0600943 page = next_page;
Dan Streetman5538c562014-10-09 15:30:01 -0700944 off %= PAGE_SIZE;
Nitin Gupta61989a82012-01-09 16:51:56 -0600945 }
946}
947
948/*
949 * Allocate a zspage for the given size class
950 */
951static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
952{
953 int i, error;
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500954 struct page *first_page = NULL, *uninitialized_var(prev_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600955
956 /*
957 * Allocate individual pages and link them together as:
958 * 1. first page->private = first sub-page
959 * 2. all sub-pages are linked together using page->lru
960 * 3. each sub-page is linked to the first page using page->first_page
961 *
962 * For each size class, First/Head pages are linked together using
963 * page->lru. Also, we set PG_private to identify the first page
964 * (i.e. no other sub-page has this flag set) and PG_private_2 to
965 * identify the last page.
966 */
967 error = -ENOMEM;
Minchan Kim2e3b6152012-05-03 15:40:39 +0900968 for (i = 0; i < class->pages_per_zspage; i++) {
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500969 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600970
971 page = alloc_page(flags);
972 if (!page)
973 goto cleanup;
974
975 INIT_LIST_HEAD(&page->lru);
976 if (i == 0) { /* first page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900977 SetPagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600978 set_page_private(page, 0);
979 first_page = page;
980 first_page->inuse = 0;
981 }
982 if (i == 1)
Sunghan Suhe842b972013-07-12 16:08:13 +0900983 set_page_private(first_page, (unsigned long)page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600984 if (i >= 1)
985 page->first_page = first_page;
986 if (i >= 2)
987 list_add(&page->lru, &prev_page->lru);
Minchan Kim2e3b6152012-05-03 15:40:39 +0900988 if (i == class->pages_per_zspage - 1) /* last page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900989 SetPagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600990 prev_page = page;
991 }
992
993 init_zspage(first_page, class);
994
Minchan Kim312fcae2015-04-15 16:15:30 -0700995 first_page->freelist = location_to_obj(first_page, 0);
Nitin Gupta61989a82012-01-09 16:51:56 -0600996 /* Maximum number of objects we can store in this zspage */
Minchan Kim2e3b6152012-05-03 15:40:39 +0900997 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
Nitin Gupta61989a82012-01-09 16:51:56 -0600998
999 error = 0; /* Success */
1000
1001cleanup:
1002 if (unlikely(error) && first_page) {
1003 free_zspage(first_page);
1004 first_page = NULL;
1005 }
1006
1007 return first_page;
1008}
1009
1010static struct page *find_get_zspage(struct size_class *class)
1011{
1012 int i;
1013 struct page *page;
1014
1015 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1016 page = class->fullness_list[i];
1017 if (page)
1018 break;
1019 }
1020
1021 return page;
1022}
1023
Minchan Kim1b945ae2013-12-11 11:04:36 +09001024#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -05001025static inline int __zs_cpu_up(struct mapping_area *area)
Seth Jennings5f601902012-07-02 16:15:49 -05001026{
Seth Jenningsf5536462012-07-18 11:55:56 -05001027 /*
1028 * Make sure we don't leak memory if a cpu UP notification
1029 * and zs_init() race and both call zs_cpu_up() on the same cpu
1030 */
1031 if (area->vm)
1032 return 0;
1033 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1034 if (!area->vm)
1035 return -ENOMEM;
1036 return 0;
1037}
1038
1039static inline void __zs_cpu_down(struct mapping_area *area)
1040{
1041 if (area->vm)
1042 free_vm_area(area->vm);
1043 area->vm = NULL;
1044}
1045
1046static inline void *__zs_map_object(struct mapping_area *area,
1047 struct page *pages[2], int off, int size)
1048{
WANG Chaof6f8ed42014-08-06 16:06:58 -07001049 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
Seth Jenningsf5536462012-07-18 11:55:56 -05001050 area->vm_addr = area->vm->addr;
1051 return area->vm_addr + off;
1052}
1053
1054static inline void __zs_unmap_object(struct mapping_area *area,
1055 struct page *pages[2], int off, int size)
1056{
1057 unsigned long addr = (unsigned long)area->vm_addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001058
Joerg Roedeld95abbb2013-03-27 01:43:14 +01001059 unmap_kernel_range(addr, PAGE_SIZE * 2);
Seth Jenningsf5536462012-07-18 11:55:56 -05001060}
1061
Minchan Kim1b945ae2013-12-11 11:04:36 +09001062#else /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001063
1064static inline int __zs_cpu_up(struct mapping_area *area)
1065{
1066 /*
1067 * Make sure we don't leak memory if a cpu UP notification
1068 * and zs_init() race and both call zs_cpu_up() on the same cpu
1069 */
1070 if (area->vm_buf)
1071 return 0;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001072 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
Seth Jenningsf5536462012-07-18 11:55:56 -05001073 if (!area->vm_buf)
1074 return -ENOMEM;
1075 return 0;
1076}
1077
1078static inline void __zs_cpu_down(struct mapping_area *area)
1079{
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001080 kfree(area->vm_buf);
Seth Jenningsf5536462012-07-18 11:55:56 -05001081 area->vm_buf = NULL;
1082}
1083
1084static void *__zs_map_object(struct mapping_area *area,
1085 struct page *pages[2], int off, int size)
1086{
Seth Jennings5f601902012-07-02 16:15:49 -05001087 int sizes[2];
1088 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001089 char *buf = area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001090
Seth Jenningsf5536462012-07-18 11:55:56 -05001091 /* disable page faults to match kmap_atomic() return conditions */
1092 pagefault_disable();
1093
1094 /* no read fastpath */
1095 if (area->vm_mm == ZS_MM_WO)
1096 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001097
1098 sizes[0] = PAGE_SIZE - off;
1099 sizes[1] = size - sizes[0];
1100
Seth Jennings5f601902012-07-02 16:15:49 -05001101 /* copy object to per-cpu buffer */
1102 addr = kmap_atomic(pages[0]);
1103 memcpy(buf, addr + off, sizes[0]);
1104 kunmap_atomic(addr);
1105 addr = kmap_atomic(pages[1]);
1106 memcpy(buf + sizes[0], addr, sizes[1]);
1107 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001108out:
1109 return area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001110}
1111
Seth Jenningsf5536462012-07-18 11:55:56 -05001112static void __zs_unmap_object(struct mapping_area *area,
1113 struct page *pages[2], int off, int size)
Seth Jennings5f601902012-07-02 16:15:49 -05001114{
Seth Jennings5f601902012-07-02 16:15:49 -05001115 int sizes[2];
1116 void *addr;
Minchan Kim2e40e162015-04-15 16:15:23 -07001117 char *buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001118
Seth Jenningsf5536462012-07-18 11:55:56 -05001119 /* no write fastpath */
1120 if (area->vm_mm == ZS_MM_RO)
1121 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001122
Minchan Kim7b60a682015-04-15 16:15:39 -07001123 buf = area->vm_buf;
1124 if (!area->huge) {
1125 buf = buf + ZS_HANDLE_SIZE;
1126 size -= ZS_HANDLE_SIZE;
1127 off += ZS_HANDLE_SIZE;
1128 }
Minchan Kim2e40e162015-04-15 16:15:23 -07001129
Seth Jennings5f601902012-07-02 16:15:49 -05001130 sizes[0] = PAGE_SIZE - off;
1131 sizes[1] = size - sizes[0];
1132
1133 /* copy per-cpu buffer to object */
1134 addr = kmap_atomic(pages[0]);
1135 memcpy(addr + off, buf, sizes[0]);
1136 kunmap_atomic(addr);
1137 addr = kmap_atomic(pages[1]);
1138 memcpy(addr, buf + sizes[0], sizes[1]);
1139 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001140
1141out:
1142 /* enable page faults to match kunmap_atomic() return conditions */
1143 pagefault_enable();
Seth Jennings5f601902012-07-02 16:15:49 -05001144}
Nitin Gupta61989a82012-01-09 16:51:56 -06001145
Minchan Kim1b945ae2013-12-11 11:04:36 +09001146#endif /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001147
Nitin Gupta61989a82012-01-09 16:51:56 -06001148static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1149 void *pcpu)
1150{
Seth Jenningsf5536462012-07-18 11:55:56 -05001151 int ret, cpu = (long)pcpu;
Nitin Gupta61989a82012-01-09 16:51:56 -06001152 struct mapping_area *area;
1153
1154 switch (action) {
1155 case CPU_UP_PREPARE:
1156 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001157 ret = __zs_cpu_up(area);
1158 if (ret)
1159 return notifier_from_errno(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001160 break;
1161 case CPU_DEAD:
1162 case CPU_UP_CANCELED:
1163 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001164 __zs_cpu_down(area);
Nitin Gupta61989a82012-01-09 16:51:56 -06001165 break;
1166 }
1167
1168 return NOTIFY_OK;
1169}
1170
1171static struct notifier_block zs_cpu_nb = {
1172 .notifier_call = zs_cpu_notifier
1173};
1174
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001175static int zs_register_cpu_notifier(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001176{
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001177 int cpu, uninitialized_var(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001178
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301179 cpu_notifier_register_begin();
1180
1181 __register_cpu_notifier(&zs_cpu_nb);
Nitin Gupta61989a82012-01-09 16:51:56 -06001182 for_each_online_cpu(cpu) {
1183 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001184 if (notifier_to_errno(ret))
1185 break;
Nitin Gupta61989a82012-01-09 16:51:56 -06001186 }
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301187
1188 cpu_notifier_register_done();
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001189 return notifier_to_errno(ret);
1190}
1191
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001192static void zs_unregister_cpu_notifier(void)
1193{
1194 int cpu;
1195
1196 cpu_notifier_register_begin();
1197
1198 for_each_online_cpu(cpu)
1199 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1200 __unregister_cpu_notifier(&zs_cpu_nb);
1201
1202 cpu_notifier_register_done();
1203}
1204
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001205static void init_zs_size_classes(void)
1206{
1207 int nr;
1208
1209 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1210 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1211 nr += 1;
1212
1213 zs_size_classes = nr;
1214}
1215
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001216static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1217{
1218 if (prev->pages_per_zspage != pages_per_zspage)
1219 return false;
1220
1221 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1222 != get_maxobj_per_zspage(size, pages_per_zspage))
1223 return false;
1224
1225 return true;
1226}
1227
Minchan Kim312fcae2015-04-15 16:15:30 -07001228static bool zspage_full(struct page *page)
1229{
1230 BUG_ON(!is_first_page(page));
1231
1232 return page->inuse == page->objects;
1233}
1234
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001235unsigned long zs_get_total_pages(struct zs_pool *pool)
1236{
1237 return atomic_long_read(&pool->pages_allocated);
1238}
1239EXPORT_SYMBOL_GPL(zs_get_total_pages);
1240
1241/**
1242 * zs_map_object - get address of allocated object from handle.
1243 * @pool: pool from which the object was allocated
1244 * @handle: handle returned from zs_malloc
1245 *
1246 * Before using an object allocated from zs_malloc, it must be mapped using
1247 * this function. When done with the object, it must be unmapped using
1248 * zs_unmap_object.
1249 *
1250 * Only one object can be mapped per cpu at a time. There is no protection
1251 * against nested mappings.
1252 *
1253 * This function returns with preemption and page faults disabled.
1254 */
1255void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1256 enum zs_mapmode mm)
1257{
1258 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001259 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001260
1261 unsigned int class_idx;
1262 enum fullness_group fg;
1263 struct size_class *class;
1264 struct mapping_area *area;
1265 struct page *pages[2];
Minchan Kim2e40e162015-04-15 16:15:23 -07001266 void *ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001267
1268 BUG_ON(!handle);
1269
1270 /*
1271 * Because we use per-cpu mapping areas shared among the
1272 * pools/users, we can't allow mapping in interrupt context
1273 * because it can corrupt another users mappings.
1274 */
1275 BUG_ON(in_interrupt());
1276
Minchan Kim312fcae2015-04-15 16:15:30 -07001277 /* From now on, migration cannot move the object */
1278 pin_tag(handle);
1279
Minchan Kim2e40e162015-04-15 16:15:23 -07001280 obj = handle_to_obj(handle);
1281 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001282 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1283 class = pool->size_class[class_idx];
1284 off = obj_idx_to_offset(page, obj_idx, class->size);
1285
1286 area = &get_cpu_var(zs_map_area);
1287 area->vm_mm = mm;
1288 if (off + class->size <= PAGE_SIZE) {
1289 /* this object is contained entirely within a page */
1290 area->vm_addr = kmap_atomic(page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001291 ret = area->vm_addr + off;
1292 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001293 }
1294
1295 /* this object spans two pages */
1296 pages[0] = page;
1297 pages[1] = get_next_page(page);
1298 BUG_ON(!pages[1]);
1299
Minchan Kim2e40e162015-04-15 16:15:23 -07001300 ret = __zs_map_object(area, pages, off, class->size);
1301out:
Minchan Kim7b60a682015-04-15 16:15:39 -07001302 if (!class->huge)
1303 ret += ZS_HANDLE_SIZE;
1304
1305 return ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001306}
1307EXPORT_SYMBOL_GPL(zs_map_object);
1308
1309void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1310{
1311 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001312 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001313
1314 unsigned int class_idx;
1315 enum fullness_group fg;
1316 struct size_class *class;
1317 struct mapping_area *area;
1318
1319 BUG_ON(!handle);
1320
Minchan Kim2e40e162015-04-15 16:15:23 -07001321 obj = handle_to_obj(handle);
1322 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001323 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1324 class = pool->size_class[class_idx];
1325 off = obj_idx_to_offset(page, obj_idx, class->size);
1326
1327 area = this_cpu_ptr(&zs_map_area);
1328 if (off + class->size <= PAGE_SIZE)
1329 kunmap_atomic(area->vm_addr);
1330 else {
1331 struct page *pages[2];
1332
1333 pages[0] = page;
1334 pages[1] = get_next_page(page);
1335 BUG_ON(!pages[1]);
1336
1337 __zs_unmap_object(area, pages, off, class->size);
1338 }
1339 put_cpu_var(zs_map_area);
Minchan Kim312fcae2015-04-15 16:15:30 -07001340 unpin_tag(handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001341}
1342EXPORT_SYMBOL_GPL(zs_unmap_object);
1343
Minchan Kimc7806262015-04-15 16:15:26 -07001344static unsigned long obj_malloc(struct page *first_page,
1345 struct size_class *class, unsigned long handle)
1346{
1347 unsigned long obj;
1348 struct link_free *link;
1349
1350 struct page *m_page;
1351 unsigned long m_objidx, m_offset;
1352 void *vaddr;
1353
Minchan Kim312fcae2015-04-15 16:15:30 -07001354 handle |= OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001355 obj = (unsigned long)first_page->freelist;
1356 obj_to_location(obj, &m_page, &m_objidx);
1357 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1358
1359 vaddr = kmap_atomic(m_page);
1360 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1361 first_page->freelist = link->next;
Minchan Kim7b60a682015-04-15 16:15:39 -07001362 if (!class->huge)
1363 /* record handle in the header of allocated chunk */
1364 link->handle = handle;
1365 else
1366 /* record handle in first_page->private */
1367 set_page_private(first_page, handle);
Minchan Kimc7806262015-04-15 16:15:26 -07001368 kunmap_atomic(vaddr);
1369 first_page->inuse++;
1370 zs_stat_inc(class, OBJ_USED, 1);
1371
1372 return obj;
1373}
1374
1375
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001376/**
1377 * zs_malloc - Allocate block of given size from pool.
1378 * @pool: pool to allocate from
1379 * @size: size of block to allocate
1380 *
1381 * On success, handle to the allocated object is returned,
1382 * otherwise 0.
1383 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1384 */
1385unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1386{
Minchan Kim2e40e162015-04-15 16:15:23 -07001387 unsigned long handle, obj;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001388 struct size_class *class;
Minchan Kimc7806262015-04-15 16:15:26 -07001389 struct page *first_page;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001390
Minchan Kim7b60a682015-04-15 16:15:39 -07001391 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001392 return 0;
1393
Minchan Kim2e40e162015-04-15 16:15:23 -07001394 handle = alloc_handle(pool);
1395 if (!handle)
1396 return 0;
1397
1398 /* extra space in chunk to keep the handle */
1399 size += ZS_HANDLE_SIZE;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001400 class = pool->size_class[get_size_class_index(size)];
1401
1402 spin_lock(&class->lock);
1403 first_page = find_get_zspage(class);
1404
1405 if (!first_page) {
1406 spin_unlock(&class->lock);
1407 first_page = alloc_zspage(class, pool->flags);
Minchan Kim2e40e162015-04-15 16:15:23 -07001408 if (unlikely(!first_page)) {
1409 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001410 return 0;
Minchan Kim2e40e162015-04-15 16:15:23 -07001411 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001412
1413 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1414 atomic_long_add(class->pages_per_zspage,
1415 &pool->pages_allocated);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001416
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001417 spin_lock(&class->lock);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001418 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1419 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001420 }
1421
Minchan Kimc7806262015-04-15 16:15:26 -07001422 obj = obj_malloc(first_page, class, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001423 /* Now move the zspage to another fullness group, if required */
Minchan Kimc7806262015-04-15 16:15:26 -07001424 fix_fullness_group(class, first_page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001425 record_obj(handle, obj);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001426 spin_unlock(&class->lock);
1427
Minchan Kim2e40e162015-04-15 16:15:23 -07001428 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001429}
1430EXPORT_SYMBOL_GPL(zs_malloc);
1431
Minchan Kimc7806262015-04-15 16:15:26 -07001432static void obj_free(struct zs_pool *pool, struct size_class *class,
1433 unsigned long obj)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001434{
1435 struct link_free *link;
1436 struct page *first_page, *f_page;
Minchan Kimc7806262015-04-15 16:15:26 -07001437 unsigned long f_objidx, f_offset;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001438 void *vaddr;
Minchan Kimc7806262015-04-15 16:15:26 -07001439 int class_idx;
1440 enum fullness_group fullness;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001441
Minchan Kimc7806262015-04-15 16:15:26 -07001442 BUG_ON(!obj);
1443
Minchan Kim312fcae2015-04-15 16:15:30 -07001444 obj &= ~OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001445 obj_to_location(obj, &f_page, &f_objidx);
1446 first_page = get_first_page(f_page);
1447
1448 get_zspage_mapping(first_page, &class_idx, &fullness);
1449 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1450
1451 vaddr = kmap_atomic(f_page);
1452
1453 /* Insert this object in containing zspage's freelist */
1454 link = (struct link_free *)(vaddr + f_offset);
1455 link->next = first_page->freelist;
Minchan Kim7b60a682015-04-15 16:15:39 -07001456 if (class->huge)
1457 set_page_private(first_page, 0);
Minchan Kimc7806262015-04-15 16:15:26 -07001458 kunmap_atomic(vaddr);
1459 first_page->freelist = (void *)obj;
1460 first_page->inuse--;
1461 zs_stat_dec(class, OBJ_USED, 1);
1462}
1463
1464void zs_free(struct zs_pool *pool, unsigned long handle)
1465{
1466 struct page *first_page, *f_page;
1467 unsigned long obj, f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001468 int class_idx;
1469 struct size_class *class;
1470 enum fullness_group fullness;
1471
Minchan Kim2e40e162015-04-15 16:15:23 -07001472 if (unlikely(!handle))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001473 return;
1474
Minchan Kim312fcae2015-04-15 16:15:30 -07001475 pin_tag(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001476 obj = handle_to_obj(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001477 obj_to_location(obj, &f_page, &f_objidx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001478 first_page = get_first_page(f_page);
1479
1480 get_zspage_mapping(first_page, &class_idx, &fullness);
1481 class = pool->size_class[class_idx];
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001482
1483 spin_lock(&class->lock);
Minchan Kimc7806262015-04-15 16:15:26 -07001484 obj_free(pool, class, obj);
1485 fullness = fix_fullness_group(class, first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001486 if (fullness == ZS_EMPTY) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001487 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1488 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001489 atomic_long_sub(class->pages_per_zspage,
1490 &pool->pages_allocated);
1491 free_zspage(first_page);
1492 }
Minchan Kim312fcae2015-04-15 16:15:30 -07001493 spin_unlock(&class->lock);
1494 unpin_tag(handle);
1495
1496 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001497}
1498EXPORT_SYMBOL_GPL(zs_free);
1499
Minchan Kim312fcae2015-04-15 16:15:30 -07001500static void zs_object_copy(unsigned long src, unsigned long dst,
1501 struct size_class *class)
1502{
1503 struct page *s_page, *d_page;
1504 unsigned long s_objidx, d_objidx;
1505 unsigned long s_off, d_off;
1506 void *s_addr, *d_addr;
1507 int s_size, d_size, size;
1508 int written = 0;
1509
1510 s_size = d_size = class->size;
1511
1512 obj_to_location(src, &s_page, &s_objidx);
1513 obj_to_location(dst, &d_page, &d_objidx);
1514
1515 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1516 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1517
1518 if (s_off + class->size > PAGE_SIZE)
1519 s_size = PAGE_SIZE - s_off;
1520
1521 if (d_off + class->size > PAGE_SIZE)
1522 d_size = PAGE_SIZE - d_off;
1523
1524 s_addr = kmap_atomic(s_page);
1525 d_addr = kmap_atomic(d_page);
1526
1527 while (1) {
1528 size = min(s_size, d_size);
1529 memcpy(d_addr + d_off, s_addr + s_off, size);
1530 written += size;
1531
1532 if (written == class->size)
1533 break;
1534
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001535 s_off += size;
1536 s_size -= size;
1537 d_off += size;
1538 d_size -= size;
1539
1540 if (s_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001541 kunmap_atomic(d_addr);
1542 kunmap_atomic(s_addr);
1543 s_page = get_next_page(s_page);
1544 BUG_ON(!s_page);
1545 s_addr = kmap_atomic(s_page);
1546 d_addr = kmap_atomic(d_page);
1547 s_size = class->size - written;
1548 s_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001549 }
1550
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001551 if (d_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001552 kunmap_atomic(d_addr);
1553 d_page = get_next_page(d_page);
1554 BUG_ON(!d_page);
1555 d_addr = kmap_atomic(d_page);
1556 d_size = class->size - written;
1557 d_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001558 }
1559 }
1560
1561 kunmap_atomic(d_addr);
1562 kunmap_atomic(s_addr);
1563}
1564
1565/*
1566 * Find alloced object in zspage from index object and
1567 * return handle.
1568 */
1569static unsigned long find_alloced_obj(struct page *page, int index,
1570 struct size_class *class)
1571{
1572 unsigned long head;
1573 int offset = 0;
1574 unsigned long handle = 0;
1575 void *addr = kmap_atomic(page);
1576
1577 if (!is_first_page(page))
1578 offset = page->index;
1579 offset += class->size * index;
1580
1581 while (offset < PAGE_SIZE) {
Minchan Kim7b60a682015-04-15 16:15:39 -07001582 head = obj_to_head(class, page, addr + offset);
Minchan Kim312fcae2015-04-15 16:15:30 -07001583 if (head & OBJ_ALLOCATED_TAG) {
1584 handle = head & ~OBJ_ALLOCATED_TAG;
1585 if (trypin_tag(handle))
1586 break;
1587 handle = 0;
1588 }
1589
1590 offset += class->size;
1591 index++;
1592 }
1593
1594 kunmap_atomic(addr);
1595 return handle;
1596}
1597
1598struct zs_compact_control {
1599 /* Source page for migration which could be a subpage of zspage. */
1600 struct page *s_page;
1601 /* Destination page for migration which should be a first page
1602 * of zspage. */
1603 struct page *d_page;
1604 /* Starting object index within @s_page which used for live object
1605 * in the subpage. */
1606 int index;
1607 /* how many of objects are migrated */
1608 int nr_migrated;
1609};
1610
1611static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1612 struct zs_compact_control *cc)
1613{
1614 unsigned long used_obj, free_obj;
1615 unsigned long handle;
1616 struct page *s_page = cc->s_page;
1617 struct page *d_page = cc->d_page;
1618 unsigned long index = cc->index;
1619 int nr_migrated = 0;
1620 int ret = 0;
1621
1622 while (1) {
1623 handle = find_alloced_obj(s_page, index, class);
1624 if (!handle) {
1625 s_page = get_next_page(s_page);
1626 if (!s_page)
1627 break;
1628 index = 0;
1629 continue;
1630 }
1631
1632 /* Stop if there is no more space */
1633 if (zspage_full(d_page)) {
1634 unpin_tag(handle);
1635 ret = -ENOMEM;
1636 break;
1637 }
1638
1639 used_obj = handle_to_obj(handle);
1640 free_obj = obj_malloc(d_page, class, handle);
1641 zs_object_copy(used_obj, free_obj, class);
1642 index++;
1643 record_obj(handle, free_obj);
1644 unpin_tag(handle);
1645 obj_free(pool, class, used_obj);
1646 nr_migrated++;
1647 }
1648
1649 /* Remember last position in this iteration */
1650 cc->s_page = s_page;
1651 cc->index = index;
1652 cc->nr_migrated = nr_migrated;
1653
1654 return ret;
1655}
1656
1657static struct page *alloc_target_page(struct size_class *class)
1658{
1659 int i;
1660 struct page *page;
1661
1662 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1663 page = class->fullness_list[i];
1664 if (page) {
1665 remove_zspage(page, class, i);
1666 break;
1667 }
1668 }
1669
1670 return page;
1671}
1672
1673static void putback_zspage(struct zs_pool *pool, struct size_class *class,
1674 struct page *first_page)
1675{
Minchan Kim312fcae2015-04-15 16:15:30 -07001676 enum fullness_group fullness;
1677
1678 BUG_ON(!is_first_page(first_page));
1679
Minchan Kim839373e2015-04-15 16:16:18 -07001680 fullness = get_fullness_group(first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001681 insert_zspage(first_page, class, fullness);
Minchan Kim839373e2015-04-15 16:16:18 -07001682 set_zspage_mapping(first_page, class->index, fullness);
1683
Minchan Kim312fcae2015-04-15 16:15:30 -07001684 if (fullness == ZS_EMPTY) {
1685 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1686 class->size, class->pages_per_zspage));
1687 atomic_long_sub(class->pages_per_zspage,
1688 &pool->pages_allocated);
1689
1690 free_zspage(first_page);
1691 }
1692}
1693
1694static struct page *isolate_source_page(struct size_class *class)
1695{
1696 struct page *page;
1697
1698 page = class->fullness_list[ZS_ALMOST_EMPTY];
1699 if (page)
1700 remove_zspage(page, class, ZS_ALMOST_EMPTY);
1701
1702 return page;
1703}
1704
1705static unsigned long __zs_compact(struct zs_pool *pool,
1706 struct size_class *class)
1707{
1708 int nr_to_migrate;
1709 struct zs_compact_control cc;
1710 struct page *src_page;
1711 struct page *dst_page = NULL;
1712 unsigned long nr_total_migrated = 0;
1713
Minchan Kim312fcae2015-04-15 16:15:30 -07001714 spin_lock(&class->lock);
1715 while ((src_page = isolate_source_page(class))) {
1716
1717 BUG_ON(!is_first_page(src_page));
1718
1719 /* The goal is to migrate all live objects in source page */
1720 nr_to_migrate = src_page->inuse;
1721 cc.index = 0;
1722 cc.s_page = src_page;
1723
1724 while ((dst_page = alloc_target_page(class))) {
1725 cc.d_page = dst_page;
1726 /*
1727 * If there is no more space in dst_page, try to
1728 * allocate another zspage.
1729 */
1730 if (!migrate_zspage(pool, class, &cc))
1731 break;
1732
1733 putback_zspage(pool, class, dst_page);
1734 nr_total_migrated += cc.nr_migrated;
1735 nr_to_migrate -= cc.nr_migrated;
1736 }
1737
1738 /* Stop if we couldn't find slot */
1739 if (dst_page == NULL)
1740 break;
1741
1742 putback_zspage(pool, class, dst_page);
1743 putback_zspage(pool, class, src_page);
1744 spin_unlock(&class->lock);
1745 nr_total_migrated += cc.nr_migrated;
1746 cond_resched();
1747 spin_lock(&class->lock);
1748 }
1749
1750 if (src_page)
1751 putback_zspage(pool, class, src_page);
1752
1753 spin_unlock(&class->lock);
1754
1755 return nr_total_migrated;
1756}
1757
1758unsigned long zs_compact(struct zs_pool *pool)
1759{
1760 int i;
1761 unsigned long nr_migrated = 0;
1762 struct size_class *class;
1763
1764 for (i = zs_size_classes - 1; i >= 0; i--) {
1765 class = pool->size_class[i];
1766 if (!class)
1767 continue;
1768 if (class->index != i)
1769 continue;
1770 nr_migrated += __zs_compact(pool, class);
1771 }
1772
Minchan Kim312fcae2015-04-15 16:15:30 -07001773 return nr_migrated;
1774}
1775EXPORT_SYMBOL_GPL(zs_compact);
1776
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001777/**
1778 * zs_create_pool - Creates an allocation pool to work from.
Seth Jennings0d145a52013-01-30 09:36:52 -06001779 * @flags: allocation flags used to allocate pool metadata
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001780 *
1781 * This function must be called before anything when using
1782 * the zsmalloc allocator.
1783 *
1784 * On success, a pointer to the newly created pool is returned,
1785 * otherwise NULL.
1786 */
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -08001787struct zs_pool *zs_create_pool(char *name, gfp_t flags)
Nitin Gupta61989a82012-01-09 16:51:56 -06001788{
Ganesh Mahendran18136652014-12-12 16:57:10 -08001789 int i;
Nitin Gupta61989a82012-01-09 16:51:56 -06001790 struct zs_pool *pool;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001791 struct size_class *prev_class = NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001792
Ganesh Mahendran18136652014-12-12 16:57:10 -08001793 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
Nitin Gupta61989a82012-01-09 16:51:56 -06001794 if (!pool)
1795 return NULL;
1796
Minchan Kim2e40e162015-04-15 16:15:23 -07001797 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1798 GFP_KERNEL);
1799 if (!pool->size_class) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001800 kfree(pool);
1801 return NULL;
1802 }
1803
Minchan Kim2e40e162015-04-15 16:15:23 -07001804 pool->name = kstrdup(name, GFP_KERNEL);
1805 if (!pool->name)
1806 goto err;
1807
1808 if (create_handle_cache(pool))
1809 goto err;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001810
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001811 /*
1812 * Iterate reversly, because, size of size_class that we want to use
1813 * for merging should be larger or equal to current size.
1814 */
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001815 for (i = zs_size_classes - 1; i >= 0; i--) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001816 int size;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001817 int pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001818 struct size_class *class;
1819
1820 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1821 if (size > ZS_MAX_ALLOC_SIZE)
1822 size = ZS_MAX_ALLOC_SIZE;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001823 pages_per_zspage = get_pages_per_zspage(size);
Nitin Gupta61989a82012-01-09 16:51:56 -06001824
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001825 /*
1826 * size_class is used for normal zsmalloc operation such
1827 * as alloc/free for that size. Although it is natural that we
1828 * have one size_class for each size, there is a chance that we
1829 * can get more memory utilization if we use one size_class for
1830 * many different sizes whose size_class have same
1831 * characteristics. So, we makes size_class point to
1832 * previous size_class if possible.
1833 */
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001834 if (prev_class) {
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001835 if (can_merge(prev_class, size, pages_per_zspage)) {
1836 pool->size_class[i] = prev_class;
1837 continue;
1838 }
1839 }
1840
1841 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1842 if (!class)
1843 goto err;
1844
Nitin Gupta61989a82012-01-09 16:51:56 -06001845 class->size = size;
1846 class->index = i;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001847 class->pages_per_zspage = pages_per_zspage;
Minchan Kim7b60a682015-04-15 16:15:39 -07001848 if (pages_per_zspage == 1 &&
1849 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1850 class->huge = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06001851 spin_lock_init(&class->lock);
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001852 pool->size_class[i] = class;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001853
1854 prev_class = class;
Nitin Gupta61989a82012-01-09 16:51:56 -06001855 }
1856
Nitin Gupta61989a82012-01-09 16:51:56 -06001857 pool->flags = flags;
Nitin Gupta61989a82012-01-09 16:51:56 -06001858
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001859 if (zs_pool_stat_create(name, pool))
1860 goto err;
1861
Nitin Gupta61989a82012-01-09 16:51:56 -06001862 return pool;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001863
1864err:
1865 zs_destroy_pool(pool);
1866 return NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001867}
1868EXPORT_SYMBOL_GPL(zs_create_pool);
1869
1870void zs_destroy_pool(struct zs_pool *pool)
1871{
1872 int i;
1873
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001874 zs_pool_stat_destroy(pool);
1875
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001876 for (i = 0; i < zs_size_classes; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001877 int fg;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001878 struct size_class *class = pool->size_class[i];
1879
1880 if (!class)
1881 continue;
1882
1883 if (class->index != i)
1884 continue;
Nitin Gupta61989a82012-01-09 16:51:56 -06001885
1886 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1887 if (class->fullness_list[fg]) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -04001888 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -06001889 class->size, fg);
1890 }
1891 }
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001892 kfree(class);
Nitin Gupta61989a82012-01-09 16:51:56 -06001893 }
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001894
Minchan Kim2e40e162015-04-15 16:15:23 -07001895 destroy_handle_cache(pool);
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001896 kfree(pool->size_class);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001897 kfree(pool->name);
Nitin Gupta61989a82012-01-09 16:51:56 -06001898 kfree(pool);
1899}
1900EXPORT_SYMBOL_GPL(zs_destroy_pool);
1901
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001902static int __init zs_init(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001903{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001904 int ret = zs_register_cpu_notifier();
Nitin Gupta61989a82012-01-09 16:51:56 -06001905
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001906 if (ret)
1907 goto notifier_fail;
Nitin Gupta61989a82012-01-09 16:51:56 -06001908
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001909 init_zs_size_classes();
Nitin Gupta61989a82012-01-09 16:51:56 -06001910
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001911#ifdef CONFIG_ZPOOL
1912 zpool_register_driver(&zs_zpool_driver);
1913#endif
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001914
1915 ret = zs_stat_init();
1916 if (ret) {
1917 pr_err("zs stat initialization failed\n");
1918 goto stat_fail;
1919 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001920 return 0;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001921
1922stat_fail:
1923#ifdef CONFIG_ZPOOL
1924 zpool_unregister_driver(&zs_zpool_driver);
1925#endif
1926notifier_fail:
1927 zs_unregister_cpu_notifier();
1928
1929 return ret;
Nitin Gupta61989a82012-01-09 16:51:56 -06001930}
Nitin Gupta61989a82012-01-09 16:51:56 -06001931
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001932static void __exit zs_exit(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001933{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001934#ifdef CONFIG_ZPOOL
1935 zpool_unregister_driver(&zs_zpool_driver);
1936#endif
1937 zs_unregister_cpu_notifier();
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001938
1939 zs_stat_exit();
Nitin Gupta61989a82012-01-09 16:51:56 -06001940}
Ben Hutchings069f1012012-06-20 02:31:11 +01001941
1942module_init(zs_init);
1943module_exit(zs_exit);
1944
1945MODULE_LICENSE("Dual BSD/GPL");
1946MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");