blob: a8b5e749e84e7dbd50d325eecf84a47316145598 [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 Cuptac3e3e88a2013-12-11 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{
Sergey Senozhatsky02f7b412015-06-10 11:14:57 -0700292 if (pool->handle_cachep)
293 kmem_cache_destroy(pool->handle_cachep);
Minchan Kim2e40e162015-04-15 16:15:23 -0700294}
295
296static unsigned long alloc_handle(struct zs_pool *pool)
297{
298 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
299 pool->flags & ~__GFP_HIGHMEM);
300}
301
302static void free_handle(struct zs_pool *pool, unsigned long handle)
303{
304 kmem_cache_free(pool->handle_cachep, (void *)handle);
305}
306
307static void record_obj(unsigned long handle, unsigned long obj)
308{
309 *(unsigned long *)handle = obj;
310}
311
Dan Streetmanc7957792014-08-06 16:08:38 -0700312/* zpool driver */
313
314#ifdef CONFIG_ZPOOL
315
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -0800316static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops)
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 Cuptac3e3e88a2013-12-11 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 -0700447#ifdef CONFIG_ZSMALLOC_STAT
448
449static inline void zs_stat_inc(struct size_class *class,
450 enum zs_stat_type type, unsigned long cnt)
451{
452 class->stats.objs[type] += cnt;
453}
454
455static inline void zs_stat_dec(struct size_class *class,
456 enum zs_stat_type type, unsigned long cnt)
457{
458 class->stats.objs[type] -= cnt;
459}
460
461static inline unsigned long zs_stat_get(struct size_class *class,
462 enum zs_stat_type type)
463{
464 return class->stats.objs[type];
465}
466
467static 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 */
582
583static inline void zs_stat_inc(struct size_class *class,
584 enum zs_stat_type type, unsigned long cnt)
585{
586}
587
588static inline void zs_stat_dec(struct size_class *class,
589 enum zs_stat_type type, unsigned long cnt)
590{
591}
592
593static inline unsigned long zs_stat_get(struct size_class *class,
594 enum zs_stat_type type)
595{
596 return 0;
597}
598
599static int __init zs_stat_init(void)
600{
601 return 0;
602}
603
604static void __exit zs_stat_exit(void)
605{
606}
607
608static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
609{
610 return 0;
611}
612
613static inline void zs_pool_stat_destroy(struct zs_pool *pool)
614{
615}
616
617#endif
618
619
Nitin Cuptac3e3e88a2013-12-11 11:04:37 +0900620/*
621 * For each size class, zspages are divided into different groups
622 * depending on how "full" they are. This was done so that we could
623 * easily find empty or nearly empty zspages when we try to shrink
624 * the pool (not yet implemented). This function returns fullness
625 * status of the given page.
626 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600627static enum fullness_group get_fullness_group(struct page *page)
628{
629 int inuse, max_objects;
630 enum fullness_group fg;
631 BUG_ON(!is_first_page(page));
632
633 inuse = page->inuse;
634 max_objects = page->objects;
635
636 if (inuse == 0)
637 fg = ZS_EMPTY;
638 else if (inuse == max_objects)
639 fg = ZS_FULL;
Minchan Kimd3d07c92015-04-15 16:15:33 -0700640 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
Nitin Gupta61989a82012-01-09 16:51:56 -0600641 fg = ZS_ALMOST_EMPTY;
642 else
643 fg = ZS_ALMOST_FULL;
644
645 return fg;
646}
647
Nitin Cuptac3e3e88a2013-12-11 11:04:37 +0900648/*
649 * Each size class maintains various freelists and zspages are assigned
650 * to one of these freelists based on the number of live objects they
651 * have. This functions inserts the given zspage into the freelist
652 * identified by <class, fullness_group>.
653 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600654static void insert_zspage(struct page *page, struct size_class *class,
655 enum fullness_group fullness)
656{
657 struct page **head;
658
659 BUG_ON(!is_first_page(page));
660
661 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
662 return;
663
664 head = &class->fullness_list[fullness];
665 if (*head)
666 list_add_tail(&page->lru, &(*head)->lru);
667
668 *head = page;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700669 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
670 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600671}
672
Nitin Cuptac3e3e88a2013-12-11 11:04:37 +0900673/*
674 * This function removes the given zspage from the freelist identified
675 * by <class, fullness_group>.
676 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600677static void remove_zspage(struct page *page, struct size_class *class,
678 enum fullness_group fullness)
679{
680 struct page **head;
681
682 BUG_ON(!is_first_page(page));
683
684 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
685 return;
686
687 head = &class->fullness_list[fullness];
688 BUG_ON(!*head);
689 if (list_empty(&(*head)->lru))
690 *head = NULL;
691 else if (*head == page)
692 *head = (struct page *)list_entry((*head)->lru.next,
693 struct page, lru);
694
695 list_del_init(&page->lru);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700696 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
697 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600698}
699
Nitin Cuptac3e3e88a2013-12-11 11:04:37 +0900700/*
701 * Each size class maintains zspages in different fullness groups depending
702 * on the number of live objects they contain. When allocating or freeing
703 * objects, the fullness status of the page can change, say, from ALMOST_FULL
704 * to ALMOST_EMPTY when freeing an object. This function checks if such
705 * a status change has occurred for the given page and accordingly moves the
706 * page from the freelist of the old fullness group to that of the new
707 * fullness group.
708 */
Minchan Kimc7806262015-04-15 16:15:26 -0700709static enum fullness_group fix_fullness_group(struct size_class *class,
Nitin Gupta61989a82012-01-09 16:51:56 -0600710 struct page *page)
711{
712 int class_idx;
Nitin Gupta61989a82012-01-09 16:51:56 -0600713 enum fullness_group currfg, newfg;
714
715 BUG_ON(!is_first_page(page));
716
717 get_zspage_mapping(page, &class_idx, &currfg);
718 newfg = get_fullness_group(page);
719 if (newfg == currfg)
720 goto out;
721
Nitin Gupta61989a82012-01-09 16:51:56 -0600722 remove_zspage(page, class, currfg);
723 insert_zspage(page, class, newfg);
724 set_zspage_mapping(page, class_idx, newfg);
725
726out:
727 return newfg;
728}
729
730/*
731 * We have to decide on how many pages to link together
732 * to form a zspage for each size class. This is important
733 * to reduce wastage due to unusable space left at end of
734 * each zspage which is given as:
Yinghao Xie888fa3742015-04-15 16:15:49 -0700735 * wastage = Zp % class_size
736 * usage = Zp - wastage
Nitin Gupta61989a82012-01-09 16:51:56 -0600737 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
738 *
739 * For example, for size class of 3/8 * PAGE_SIZE, we should
740 * link together 3 PAGE_SIZE sized pages to form a zspage
741 * since then we can perfectly fit in 8 such objects.
742 */
Minchan Kim2e3b61542012-05-03 15:40:39 +0900743static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600744{
745 int i, max_usedpc = 0;
746 /* zspage order which gives maximum used size per KB */
747 int max_usedpc_order = 1;
748
Seth Jennings84d4faa2012-03-05 11:33:21 -0600749 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600750 int zspage_size;
751 int waste, usedpc;
752
753 zspage_size = i * PAGE_SIZE;
754 waste = zspage_size % class_size;
755 usedpc = (zspage_size - waste) * 100 / zspage_size;
756
757 if (usedpc > max_usedpc) {
758 max_usedpc = usedpc;
759 max_usedpc_order = i;
760 }
761 }
762
763 return max_usedpc_order;
764}
765
766/*
767 * A single 'zspage' is composed of many system pages which are
768 * linked together using fields in struct page. This function finds
769 * the first/head page, given any component page of a zspage.
770 */
771static struct page *get_first_page(struct page *page)
772{
773 if (is_first_page(page))
774 return page;
775 else
776 return page->first_page;
777}
778
779static struct page *get_next_page(struct page *page)
780{
781 struct page *next;
782
783 if (is_last_page(page))
784 next = NULL;
785 else if (is_first_page(page))
Sunghan Suhe842b972013-07-12 16:08:13 +0900786 next = (struct page *)page_private(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600787 else
788 next = list_entry(page->lru.next, struct page, lru);
789
790 return next;
791}
792
Olav Haugan67296872013-11-22 09:30:41 -0800793/*
794 * Encode <page, obj_idx> as a single handle value.
Minchan Kim312fcae2015-04-15 16:15:30 -0700795 * We use the least bit of handle for tagging.
Olav Haugan67296872013-11-22 09:30:41 -0800796 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700797static void *location_to_obj(struct page *page, unsigned long obj_idx)
Nitin Gupta61989a82012-01-09 16:51:56 -0600798{
Minchan Kim312fcae2015-04-15 16:15:30 -0700799 unsigned long obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600800
801 if (!page) {
802 BUG_ON(obj_idx);
803 return NULL;
804 }
805
Minchan Kim312fcae2015-04-15 16:15:30 -0700806 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
807 obj |= ((obj_idx) & OBJ_INDEX_MASK);
808 obj <<= OBJ_TAG_BITS;
Nitin Gupta61989a82012-01-09 16:51:56 -0600809
Minchan Kim312fcae2015-04-15 16:15:30 -0700810 return (void *)obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600811}
812
Olav Haugan67296872013-11-22 09:30:41 -0800813/*
814 * Decode <page, obj_idx> pair from the given object handle. We adjust the
815 * decoded obj_idx back to its original value since it was adjusted in
Minchan Kim312fcae2015-04-15 16:15:30 -0700816 * location_to_obj().
Olav Haugan67296872013-11-22 09:30:41 -0800817 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700818static void obj_to_location(unsigned long obj, struct page **page,
Nitin Gupta61989a82012-01-09 16:51:56 -0600819 unsigned long *obj_idx)
820{
Minchan Kim312fcae2015-04-15 16:15:30 -0700821 obj >>= OBJ_TAG_BITS;
822 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
823 *obj_idx = (obj & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600824}
825
Minchan Kim2e40e162015-04-15 16:15:23 -0700826static unsigned long handle_to_obj(unsigned long handle)
827{
828 return *(unsigned long *)handle;
829}
830
Minchan Kim7b60a682015-04-15 16:15:39 -0700831static unsigned long obj_to_head(struct size_class *class, struct page *page,
832 void *obj)
Minchan Kim312fcae2015-04-15 16:15:30 -0700833{
Minchan Kim7b60a682015-04-15 16:15:39 -0700834 if (class->huge) {
835 VM_BUG_ON(!is_first_page(page));
836 return *(unsigned long *)page_private(page);
837 } else
838 return *(unsigned long *)obj;
Minchan Kim312fcae2015-04-15 16:15:30 -0700839}
840
Nitin Gupta61989a82012-01-09 16:51:56 -0600841static unsigned long obj_idx_to_offset(struct page *page,
842 unsigned long obj_idx, int class_size)
843{
844 unsigned long off = 0;
845
846 if (!is_first_page(page))
847 off = page->index;
848
849 return off + obj_idx * class_size;
850}
851
Minchan Kim312fcae2015-04-15 16:15:30 -0700852static inline int trypin_tag(unsigned long handle)
853{
854 unsigned long *ptr = (unsigned long *)handle;
855
856 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
857}
858
859static void pin_tag(unsigned long handle)
860{
861 while (!trypin_tag(handle));
862}
863
864static void unpin_tag(unsigned long handle)
865{
866 unsigned long *ptr = (unsigned long *)handle;
867
868 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
869}
870
Nitin Guptaf4477e92012-04-02 09:13:56 -0500871static void reset_page(struct page *page)
872{
873 clear_bit(PG_private, &page->flags);
874 clear_bit(PG_private_2, &page->flags);
875 set_page_private(page, 0);
876 page->mapping = NULL;
877 page->freelist = NULL;
Mel Gorman22b751c2013-02-22 16:34:59 -0800878 page_mapcount_reset(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500879}
880
Nitin Gupta61989a82012-01-09 16:51:56 -0600881static void free_zspage(struct page *first_page)
882{
Nitin Guptaf4477e92012-04-02 09:13:56 -0500883 struct page *nextp, *tmp, *head_extra;
Nitin Gupta61989a82012-01-09 16:51:56 -0600884
885 BUG_ON(!is_first_page(first_page));
886 BUG_ON(first_page->inuse);
887
Nitin Guptaf4477e92012-04-02 09:13:56 -0500888 head_extra = (struct page *)page_private(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600889
Nitin Guptaf4477e92012-04-02 09:13:56 -0500890 reset_page(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600891 __free_page(first_page);
892
893 /* zspage with only 1 system page */
Nitin Guptaf4477e92012-04-02 09:13:56 -0500894 if (!head_extra)
Nitin Gupta61989a82012-01-09 16:51:56 -0600895 return;
896
Nitin Guptaf4477e92012-04-02 09:13:56 -0500897 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600898 list_del(&nextp->lru);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500899 reset_page(nextp);
Nitin Gupta61989a82012-01-09 16:51:56 -0600900 __free_page(nextp);
901 }
Nitin Guptaf4477e92012-04-02 09:13:56 -0500902 reset_page(head_extra);
903 __free_page(head_extra);
Nitin Gupta61989a82012-01-09 16:51:56 -0600904}
905
906/* Initialize a newly allocated zspage */
907static void init_zspage(struct page *first_page, struct size_class *class)
908{
909 unsigned long off = 0;
910 struct page *page = first_page;
911
912 BUG_ON(!is_first_page(first_page));
913 while (page) {
914 struct page *next_page;
915 struct link_free *link;
Dan Streetman5538c562014-10-09 15:30:01 -0700916 unsigned int i = 1;
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800917 void *vaddr;
Nitin Gupta61989a82012-01-09 16:51:56 -0600918
919 /*
920 * page->index stores offset of first object starting
921 * in the page. For the first page, this is always 0,
922 * so we use first_page->index (aka ->freelist) to store
923 * head of corresponding zspage's freelist.
924 */
925 if (page != first_page)
926 page->index = off;
927
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800928 vaddr = kmap_atomic(page);
929 link = (struct link_free *)vaddr + off / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600930
Dan Streetman5538c562014-10-09 15:30:01 -0700931 while ((off += class->size) < PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -0700932 link->next = location_to_obj(page, i++);
Dan Streetman5538c562014-10-09 15:30:01 -0700933 link += class->size / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600934 }
935
936 /*
937 * We now come to the last (full or partial) object on this
938 * page, which must point to the first object on the next
939 * page (if present)
940 */
941 next_page = get_next_page(page);
Minchan Kim312fcae2015-04-15 16:15:30 -0700942 link->next = location_to_obj(next_page, 0);
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800943 kunmap_atomic(vaddr);
Nitin Gupta61989a82012-01-09 16:51:56 -0600944 page = next_page;
Dan Streetman5538c562014-10-09 15:30:01 -0700945 off %= PAGE_SIZE;
Nitin Gupta61989a82012-01-09 16:51:56 -0600946 }
947}
948
949/*
950 * Allocate a zspage for the given size class
951 */
952static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
953{
954 int i, error;
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500955 struct page *first_page = NULL, *uninitialized_var(prev_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600956
957 /*
958 * Allocate individual pages and link them together as:
959 * 1. first page->private = first sub-page
960 * 2. all sub-pages are linked together using page->lru
961 * 3. each sub-page is linked to the first page using page->first_page
962 *
963 * For each size class, First/Head pages are linked together using
964 * page->lru. Also, we set PG_private to identify the first page
965 * (i.e. no other sub-page has this flag set) and PG_private_2 to
966 * identify the last page.
967 */
968 error = -ENOMEM;
Minchan Kim2e3b61542012-05-03 15:40:39 +0900969 for (i = 0; i < class->pages_per_zspage; i++) {
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500970 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600971
972 page = alloc_page(flags);
973 if (!page)
974 goto cleanup;
975
976 INIT_LIST_HEAD(&page->lru);
977 if (i == 0) { /* first page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900978 SetPagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600979 set_page_private(page, 0);
980 first_page = page;
981 first_page->inuse = 0;
982 }
983 if (i == 1)
Sunghan Suhe842b972013-07-12 16:08:13 +0900984 set_page_private(first_page, (unsigned long)page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600985 if (i >= 1)
986 page->first_page = first_page;
987 if (i >= 2)
988 list_add(&page->lru, &prev_page->lru);
Minchan Kim2e3b61542012-05-03 15:40:39 +0900989 if (i == class->pages_per_zspage - 1) /* last page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900990 SetPagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600991 prev_page = page;
992 }
993
994 init_zspage(first_page, class);
995
Minchan Kim312fcae2015-04-15 16:15:30 -0700996 first_page->freelist = location_to_obj(first_page, 0);
Nitin Gupta61989a82012-01-09 16:51:56 -0600997 /* Maximum number of objects we can store in this zspage */
Minchan Kim2e3b61542012-05-03 15:40:39 +0900998 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
Nitin Gupta61989a82012-01-09 16:51:56 -0600999
1000 error = 0; /* Success */
1001
1002cleanup:
1003 if (unlikely(error) && first_page) {
1004 free_zspage(first_page);
1005 first_page = NULL;
1006 }
1007
1008 return first_page;
1009}
1010
1011static struct page *find_get_zspage(struct size_class *class)
1012{
1013 int i;
1014 struct page *page;
1015
1016 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1017 page = class->fullness_list[i];
1018 if (page)
1019 break;
1020 }
1021
1022 return page;
1023}
1024
Minchan Kim1b945ae2013-12-11 11:04:36 +09001025#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -05001026static inline int __zs_cpu_up(struct mapping_area *area)
Seth Jennings5f6019022012-07-02 16:15:49 -05001027{
Seth Jenningsf5536462012-07-18 11:55:56 -05001028 /*
1029 * Make sure we don't leak memory if a cpu UP notification
1030 * and zs_init() race and both call zs_cpu_up() on the same cpu
1031 */
1032 if (area->vm)
1033 return 0;
1034 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1035 if (!area->vm)
1036 return -ENOMEM;
1037 return 0;
1038}
1039
1040static inline void __zs_cpu_down(struct mapping_area *area)
1041{
1042 if (area->vm)
1043 free_vm_area(area->vm);
1044 area->vm = NULL;
1045}
1046
1047static inline void *__zs_map_object(struct mapping_area *area,
1048 struct page *pages[2], int off, int size)
1049{
WANG Chaof6f8ed42014-08-06 16:06:58 -07001050 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
Seth Jenningsf5536462012-07-18 11:55:56 -05001051 area->vm_addr = area->vm->addr;
1052 return area->vm_addr + off;
1053}
1054
1055static inline void __zs_unmap_object(struct mapping_area *area,
1056 struct page *pages[2], int off, int size)
1057{
1058 unsigned long addr = (unsigned long)area->vm_addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001059
Joerg Roedeld95abbb2013-03-27 01:43:14 +01001060 unmap_kernel_range(addr, PAGE_SIZE * 2);
Seth Jenningsf5536462012-07-18 11:55:56 -05001061}
1062
Minchan Kim1b945ae2013-12-11 11:04:36 +09001063#else /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001064
1065static inline int __zs_cpu_up(struct mapping_area *area)
1066{
1067 /*
1068 * Make sure we don't leak memory if a cpu UP notification
1069 * and zs_init() race and both call zs_cpu_up() on the same cpu
1070 */
1071 if (area->vm_buf)
1072 return 0;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001073 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
Seth Jenningsf5536462012-07-18 11:55:56 -05001074 if (!area->vm_buf)
1075 return -ENOMEM;
1076 return 0;
1077}
1078
1079static inline void __zs_cpu_down(struct mapping_area *area)
1080{
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001081 kfree(area->vm_buf);
Seth Jenningsf5536462012-07-18 11:55:56 -05001082 area->vm_buf = NULL;
1083}
1084
1085static void *__zs_map_object(struct mapping_area *area,
1086 struct page *pages[2], int off, int size)
1087{
Seth Jennings5f6019022012-07-02 16:15:49 -05001088 int sizes[2];
1089 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001090 char *buf = area->vm_buf;
Seth Jennings5f6019022012-07-02 16:15:49 -05001091
Seth Jenningsf5536462012-07-18 11:55:56 -05001092 /* disable page faults to match kmap_atomic() return conditions */
1093 pagefault_disable();
1094
1095 /* no read fastpath */
1096 if (area->vm_mm == ZS_MM_WO)
1097 goto out;
Seth Jennings5f6019022012-07-02 16:15:49 -05001098
1099 sizes[0] = PAGE_SIZE - off;
1100 sizes[1] = size - sizes[0];
1101
Seth Jennings5f6019022012-07-02 16:15:49 -05001102 /* copy object to per-cpu buffer */
1103 addr = kmap_atomic(pages[0]);
1104 memcpy(buf, addr + off, sizes[0]);
1105 kunmap_atomic(addr);
1106 addr = kmap_atomic(pages[1]);
1107 memcpy(buf + sizes[0], addr, sizes[1]);
1108 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001109out:
1110 return area->vm_buf;
Seth Jennings5f6019022012-07-02 16:15:49 -05001111}
1112
Seth Jenningsf5536462012-07-18 11:55:56 -05001113static void __zs_unmap_object(struct mapping_area *area,
1114 struct page *pages[2], int off, int size)
Seth Jennings5f6019022012-07-02 16:15:49 -05001115{
Seth Jennings5f6019022012-07-02 16:15:49 -05001116 int sizes[2];
1117 void *addr;
Minchan Kim2e40e162015-04-15 16:15:23 -07001118 char *buf;
Seth Jennings5f6019022012-07-02 16:15:49 -05001119
Seth Jenningsf5536462012-07-18 11:55:56 -05001120 /* no write fastpath */
1121 if (area->vm_mm == ZS_MM_RO)
1122 goto out;
Seth Jennings5f6019022012-07-02 16:15:49 -05001123
Minchan Kim7b60a682015-04-15 16:15:39 -07001124 buf = area->vm_buf;
1125 if (!area->huge) {
1126 buf = buf + ZS_HANDLE_SIZE;
1127 size -= ZS_HANDLE_SIZE;
1128 off += ZS_HANDLE_SIZE;
1129 }
Minchan Kim2e40e162015-04-15 16:15:23 -07001130
Seth Jennings5f6019022012-07-02 16:15:49 -05001131 sizes[0] = PAGE_SIZE - off;
1132 sizes[1] = size - sizes[0];
1133
1134 /* copy per-cpu buffer to object */
1135 addr = kmap_atomic(pages[0]);
1136 memcpy(addr + off, buf, sizes[0]);
1137 kunmap_atomic(addr);
1138 addr = kmap_atomic(pages[1]);
1139 memcpy(addr, buf + sizes[0], sizes[1]);
1140 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001141
1142out:
1143 /* enable page faults to match kunmap_atomic() return conditions */
1144 pagefault_enable();
Seth Jennings5f6019022012-07-02 16:15:49 -05001145}
Nitin Gupta61989a82012-01-09 16:51:56 -06001146
Minchan Kim1b945ae2013-12-11 11:04:36 +09001147#endif /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001148
Nitin Gupta61989a82012-01-09 16:51:56 -06001149static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1150 void *pcpu)
1151{
Seth Jenningsf5536462012-07-18 11:55:56 -05001152 int ret, cpu = (long)pcpu;
Nitin Gupta61989a82012-01-09 16:51:56 -06001153 struct mapping_area *area;
1154
1155 switch (action) {
1156 case CPU_UP_PREPARE:
1157 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001158 ret = __zs_cpu_up(area);
1159 if (ret)
1160 return notifier_from_errno(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001161 break;
1162 case CPU_DEAD:
1163 case CPU_UP_CANCELED:
1164 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001165 __zs_cpu_down(area);
Nitin Gupta61989a82012-01-09 16:51:56 -06001166 break;
1167 }
1168
1169 return NOTIFY_OK;
1170}
1171
1172static struct notifier_block zs_cpu_nb = {
1173 .notifier_call = zs_cpu_notifier
1174};
1175
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001176static int zs_register_cpu_notifier(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001177{
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001178 int cpu, uninitialized_var(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001179
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301180 cpu_notifier_register_begin();
1181
1182 __register_cpu_notifier(&zs_cpu_nb);
Nitin Gupta61989a82012-01-09 16:51:56 -06001183 for_each_online_cpu(cpu) {
1184 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001185 if (notifier_to_errno(ret))
1186 break;
Nitin Gupta61989a82012-01-09 16:51:56 -06001187 }
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301188
1189 cpu_notifier_register_done();
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001190 return notifier_to_errno(ret);
1191}
1192
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001193static void zs_unregister_cpu_notifier(void)
1194{
1195 int cpu;
1196
1197 cpu_notifier_register_begin();
1198
1199 for_each_online_cpu(cpu)
1200 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1201 __unregister_cpu_notifier(&zs_cpu_nb);
1202
1203 cpu_notifier_register_done();
1204}
1205
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001206static void init_zs_size_classes(void)
1207{
1208 int nr;
1209
1210 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1211 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1212 nr += 1;
1213
1214 zs_size_classes = nr;
1215}
1216
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001217static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1218{
1219 if (prev->pages_per_zspage != pages_per_zspage)
1220 return false;
1221
1222 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1223 != get_maxobj_per_zspage(size, pages_per_zspage))
1224 return false;
1225
1226 return true;
1227}
1228
Minchan Kim312fcae2015-04-15 16:15:30 -07001229static bool zspage_full(struct page *page)
1230{
1231 BUG_ON(!is_first_page(page));
1232
1233 return page->inuse == page->objects;
1234}
1235
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001236unsigned long zs_get_total_pages(struct zs_pool *pool)
1237{
1238 return atomic_long_read(&pool->pages_allocated);
1239}
1240EXPORT_SYMBOL_GPL(zs_get_total_pages);
1241
1242/**
1243 * zs_map_object - get address of allocated object from handle.
1244 * @pool: pool from which the object was allocated
1245 * @handle: handle returned from zs_malloc
1246 *
1247 * Before using an object allocated from zs_malloc, it must be mapped using
1248 * this function. When done with the object, it must be unmapped using
1249 * zs_unmap_object.
1250 *
1251 * Only one object can be mapped per cpu at a time. There is no protection
1252 * against nested mappings.
1253 *
1254 * This function returns with preemption and page faults disabled.
1255 */
1256void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1257 enum zs_mapmode mm)
1258{
1259 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001260 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001261
1262 unsigned int class_idx;
1263 enum fullness_group fg;
1264 struct size_class *class;
1265 struct mapping_area *area;
1266 struct page *pages[2];
Minchan Kim2e40e162015-04-15 16:15:23 -07001267 void *ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001268
1269 BUG_ON(!handle);
1270
1271 /*
1272 * Because we use per-cpu mapping areas shared among the
1273 * pools/users, we can't allow mapping in interrupt context
1274 * because it can corrupt another users mappings.
1275 */
1276 BUG_ON(in_interrupt());
1277
Minchan Kim312fcae2015-04-15 16:15:30 -07001278 /* From now on, migration cannot move the object */
1279 pin_tag(handle);
1280
Minchan Kim2e40e162015-04-15 16:15:23 -07001281 obj = handle_to_obj(handle);
1282 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001283 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1284 class = pool->size_class[class_idx];
1285 off = obj_idx_to_offset(page, obj_idx, class->size);
1286
1287 area = &get_cpu_var(zs_map_area);
1288 area->vm_mm = mm;
1289 if (off + class->size <= PAGE_SIZE) {
1290 /* this object is contained entirely within a page */
1291 area->vm_addr = kmap_atomic(page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001292 ret = area->vm_addr + off;
1293 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001294 }
1295
1296 /* this object spans two pages */
1297 pages[0] = page;
1298 pages[1] = get_next_page(page);
1299 BUG_ON(!pages[1]);
1300
Minchan Kim2e40e162015-04-15 16:15:23 -07001301 ret = __zs_map_object(area, pages, off, class->size);
1302out:
Minchan Kim7b60a682015-04-15 16:15:39 -07001303 if (!class->huge)
1304 ret += ZS_HANDLE_SIZE;
1305
1306 return ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001307}
1308EXPORT_SYMBOL_GPL(zs_map_object);
1309
1310void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1311{
1312 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001313 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001314
1315 unsigned int class_idx;
1316 enum fullness_group fg;
1317 struct size_class *class;
1318 struct mapping_area *area;
1319
1320 BUG_ON(!handle);
1321
Minchan Kim2e40e162015-04-15 16:15:23 -07001322 obj = handle_to_obj(handle);
1323 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001324 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1325 class = pool->size_class[class_idx];
1326 off = obj_idx_to_offset(page, obj_idx, class->size);
1327
1328 area = this_cpu_ptr(&zs_map_area);
1329 if (off + class->size <= PAGE_SIZE)
1330 kunmap_atomic(area->vm_addr);
1331 else {
1332 struct page *pages[2];
1333
1334 pages[0] = page;
1335 pages[1] = get_next_page(page);
1336 BUG_ON(!pages[1]);
1337
1338 __zs_unmap_object(area, pages, off, class->size);
1339 }
1340 put_cpu_var(zs_map_area);
Minchan Kim312fcae2015-04-15 16:15:30 -07001341 unpin_tag(handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001342}
1343EXPORT_SYMBOL_GPL(zs_unmap_object);
1344
Minchan Kimc7806262015-04-15 16:15:26 -07001345static unsigned long obj_malloc(struct page *first_page,
1346 struct size_class *class, unsigned long handle)
1347{
1348 unsigned long obj;
1349 struct link_free *link;
1350
1351 struct page *m_page;
1352 unsigned long m_objidx, m_offset;
1353 void *vaddr;
1354
Minchan Kim312fcae2015-04-15 16:15:30 -07001355 handle |= OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001356 obj = (unsigned long)first_page->freelist;
1357 obj_to_location(obj, &m_page, &m_objidx);
1358 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1359
1360 vaddr = kmap_atomic(m_page);
1361 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1362 first_page->freelist = link->next;
Minchan Kim7b60a682015-04-15 16:15:39 -07001363 if (!class->huge)
1364 /* record handle in the header of allocated chunk */
1365 link->handle = handle;
1366 else
1367 /* record handle in first_page->private */
1368 set_page_private(first_page, handle);
Minchan Kimc7806262015-04-15 16:15:26 -07001369 kunmap_atomic(vaddr);
1370 first_page->inuse++;
1371 zs_stat_inc(class, OBJ_USED, 1);
1372
1373 return obj;
1374}
1375
1376
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001377/**
1378 * zs_malloc - Allocate block of given size from pool.
1379 * @pool: pool to allocate from
1380 * @size: size of block to allocate
1381 *
1382 * On success, handle to the allocated object is returned,
1383 * otherwise 0.
1384 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1385 */
1386unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1387{
Minchan Kim2e40e162015-04-15 16:15:23 -07001388 unsigned long handle, obj;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001389 struct size_class *class;
Minchan Kimc7806262015-04-15 16:15:26 -07001390 struct page *first_page;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001391
Minchan Kim7b60a682015-04-15 16:15:39 -07001392 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001393 return 0;
1394
Minchan Kim2e40e162015-04-15 16:15:23 -07001395 handle = alloc_handle(pool);
1396 if (!handle)
1397 return 0;
1398
1399 /* extra space in chunk to keep the handle */
1400 size += ZS_HANDLE_SIZE;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001401 class = pool->size_class[get_size_class_index(size)];
1402
1403 spin_lock(&class->lock);
1404 first_page = find_get_zspage(class);
1405
1406 if (!first_page) {
1407 spin_unlock(&class->lock);
1408 first_page = alloc_zspage(class, pool->flags);
Minchan Kim2e40e162015-04-15 16:15:23 -07001409 if (unlikely(!first_page)) {
1410 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001411 return 0;
Minchan Kim2e40e162015-04-15 16:15:23 -07001412 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001413
1414 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1415 atomic_long_add(class->pages_per_zspage,
1416 &pool->pages_allocated);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001417
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001418 spin_lock(&class->lock);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001419 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1420 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001421 }
1422
Minchan Kimc7806262015-04-15 16:15:26 -07001423 obj = obj_malloc(first_page, class, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001424 /* Now move the zspage to another fullness group, if required */
Minchan Kimc7806262015-04-15 16:15:26 -07001425 fix_fullness_group(class, first_page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001426 record_obj(handle, obj);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001427 spin_unlock(&class->lock);
1428
Minchan Kim2e40e162015-04-15 16:15:23 -07001429 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001430}
1431EXPORT_SYMBOL_GPL(zs_malloc);
1432
Minchan Kimc7806262015-04-15 16:15:26 -07001433static void obj_free(struct zs_pool *pool, struct size_class *class,
1434 unsigned long obj)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001435{
1436 struct link_free *link;
1437 struct page *first_page, *f_page;
Minchan Kimc7806262015-04-15 16:15:26 -07001438 unsigned long f_objidx, f_offset;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001439 void *vaddr;
Minchan Kimc7806262015-04-15 16:15:26 -07001440 int class_idx;
1441 enum fullness_group fullness;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001442
Minchan Kimc7806262015-04-15 16:15:26 -07001443 BUG_ON(!obj);
1444
Minchan Kim312fcae2015-04-15 16:15:30 -07001445 obj &= ~OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001446 obj_to_location(obj, &f_page, &f_objidx);
1447 first_page = get_first_page(f_page);
1448
1449 get_zspage_mapping(first_page, &class_idx, &fullness);
1450 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1451
1452 vaddr = kmap_atomic(f_page);
1453
1454 /* Insert this object in containing zspage's freelist */
1455 link = (struct link_free *)(vaddr + f_offset);
1456 link->next = first_page->freelist;
Minchan Kim7b60a682015-04-15 16:15:39 -07001457 if (class->huge)
1458 set_page_private(first_page, 0);
Minchan Kimc7806262015-04-15 16:15:26 -07001459 kunmap_atomic(vaddr);
1460 first_page->freelist = (void *)obj;
1461 first_page->inuse--;
1462 zs_stat_dec(class, OBJ_USED, 1);
1463}
1464
1465void zs_free(struct zs_pool *pool, unsigned long handle)
1466{
1467 struct page *first_page, *f_page;
1468 unsigned long obj, f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001469 int class_idx;
1470 struct size_class *class;
1471 enum fullness_group fullness;
1472
Minchan Kim2e40e162015-04-15 16:15:23 -07001473 if (unlikely(!handle))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001474 return;
1475
Minchan Kim312fcae2015-04-15 16:15:30 -07001476 pin_tag(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001477 obj = handle_to_obj(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001478 obj_to_location(obj, &f_page, &f_objidx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001479 first_page = get_first_page(f_page);
1480
1481 get_zspage_mapping(first_page, &class_idx, &fullness);
1482 class = pool->size_class[class_idx];
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001483
1484 spin_lock(&class->lock);
Minchan Kimc7806262015-04-15 16:15:26 -07001485 obj_free(pool, class, obj);
1486 fullness = fix_fullness_group(class, first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001487 if (fullness == ZS_EMPTY) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001488 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1489 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001490 atomic_long_sub(class->pages_per_zspage,
1491 &pool->pages_allocated);
1492 free_zspage(first_page);
1493 }
Minchan Kim312fcae2015-04-15 16:15:30 -07001494 spin_unlock(&class->lock);
1495 unpin_tag(handle);
1496
1497 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001498}
1499EXPORT_SYMBOL_GPL(zs_free);
1500
Minchan Kim312fcae2015-04-15 16:15:30 -07001501static void zs_object_copy(unsigned long src, unsigned long dst,
1502 struct size_class *class)
1503{
1504 struct page *s_page, *d_page;
1505 unsigned long s_objidx, d_objidx;
1506 unsigned long s_off, d_off;
1507 void *s_addr, *d_addr;
1508 int s_size, d_size, size;
1509 int written = 0;
1510
1511 s_size = d_size = class->size;
1512
1513 obj_to_location(src, &s_page, &s_objidx);
1514 obj_to_location(dst, &d_page, &d_objidx);
1515
1516 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1517 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1518
1519 if (s_off + class->size > PAGE_SIZE)
1520 s_size = PAGE_SIZE - s_off;
1521
1522 if (d_off + class->size > PAGE_SIZE)
1523 d_size = PAGE_SIZE - d_off;
1524
1525 s_addr = kmap_atomic(s_page);
1526 d_addr = kmap_atomic(d_page);
1527
1528 while (1) {
1529 size = min(s_size, d_size);
1530 memcpy(d_addr + d_off, s_addr + s_off, size);
1531 written += size;
1532
1533 if (written == class->size)
1534 break;
1535
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001536 s_off += size;
1537 s_size -= size;
1538 d_off += size;
1539 d_size -= size;
1540
1541 if (s_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001542 kunmap_atomic(d_addr);
1543 kunmap_atomic(s_addr);
1544 s_page = get_next_page(s_page);
1545 BUG_ON(!s_page);
1546 s_addr = kmap_atomic(s_page);
1547 d_addr = kmap_atomic(d_page);
1548 s_size = class->size - written;
1549 s_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001550 }
1551
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001552 if (d_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001553 kunmap_atomic(d_addr);
1554 d_page = get_next_page(d_page);
1555 BUG_ON(!d_page);
1556 d_addr = kmap_atomic(d_page);
1557 d_size = class->size - written;
1558 d_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001559 }
1560 }
1561
1562 kunmap_atomic(d_addr);
1563 kunmap_atomic(s_addr);
1564}
1565
1566/*
1567 * Find alloced object in zspage from index object and
1568 * return handle.
1569 */
1570static unsigned long find_alloced_obj(struct page *page, int index,
1571 struct size_class *class)
1572{
1573 unsigned long head;
1574 int offset = 0;
1575 unsigned long handle = 0;
1576 void *addr = kmap_atomic(page);
1577
1578 if (!is_first_page(page))
1579 offset = page->index;
1580 offset += class->size * index;
1581
1582 while (offset < PAGE_SIZE) {
Minchan Kim7b60a682015-04-15 16:15:39 -07001583 head = obj_to_head(class, page, addr + offset);
Minchan Kim312fcae2015-04-15 16:15:30 -07001584 if (head & OBJ_ALLOCATED_TAG) {
1585 handle = head & ~OBJ_ALLOCATED_TAG;
1586 if (trypin_tag(handle))
1587 break;
1588 handle = 0;
1589 }
1590
1591 offset += class->size;
1592 index++;
1593 }
1594
1595 kunmap_atomic(addr);
1596 return handle;
1597}
1598
1599struct zs_compact_control {
1600 /* Source page for migration which could be a subpage of zspage. */
1601 struct page *s_page;
1602 /* Destination page for migration which should be a first page
1603 * of zspage. */
1604 struct page *d_page;
1605 /* Starting object index within @s_page which used for live object
1606 * in the subpage. */
1607 int index;
1608 /* how many of objects are migrated */
1609 int nr_migrated;
1610};
1611
1612static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1613 struct zs_compact_control *cc)
1614{
1615 unsigned long used_obj, free_obj;
1616 unsigned long handle;
1617 struct page *s_page = cc->s_page;
1618 struct page *d_page = cc->d_page;
1619 unsigned long index = cc->index;
1620 int nr_migrated = 0;
1621 int ret = 0;
1622
1623 while (1) {
1624 handle = find_alloced_obj(s_page, index, class);
1625 if (!handle) {
1626 s_page = get_next_page(s_page);
1627 if (!s_page)
1628 break;
1629 index = 0;
1630 continue;
1631 }
1632
1633 /* Stop if there is no more space */
1634 if (zspage_full(d_page)) {
1635 unpin_tag(handle);
1636 ret = -ENOMEM;
1637 break;
1638 }
1639
1640 used_obj = handle_to_obj(handle);
1641 free_obj = obj_malloc(d_page, class, handle);
1642 zs_object_copy(used_obj, free_obj, class);
1643 index++;
1644 record_obj(handle, free_obj);
1645 unpin_tag(handle);
1646 obj_free(pool, class, used_obj);
1647 nr_migrated++;
1648 }
1649
1650 /* Remember last position in this iteration */
1651 cc->s_page = s_page;
1652 cc->index = index;
1653 cc->nr_migrated = nr_migrated;
1654
1655 return ret;
1656}
1657
1658static struct page *alloc_target_page(struct size_class *class)
1659{
1660 int i;
1661 struct page *page;
1662
1663 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1664 page = class->fullness_list[i];
1665 if (page) {
1666 remove_zspage(page, class, i);
1667 break;
1668 }
1669 }
1670
1671 return page;
1672}
1673
1674static void putback_zspage(struct zs_pool *pool, struct size_class *class,
1675 struct page *first_page)
1676{
Minchan Kim312fcae2015-04-15 16:15:30 -07001677 enum fullness_group fullness;
1678
1679 BUG_ON(!is_first_page(first_page));
1680
Minchan Kim839373e2015-04-15 16:16:18 -07001681 fullness = get_fullness_group(first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001682 insert_zspage(first_page, class, fullness);
Minchan Kim839373e2015-04-15 16:16:18 -07001683 set_zspage_mapping(first_page, class->index, fullness);
1684
Minchan Kim312fcae2015-04-15 16:15:30 -07001685 if (fullness == ZS_EMPTY) {
1686 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1687 class->size, class->pages_per_zspage));
1688 atomic_long_sub(class->pages_per_zspage,
1689 &pool->pages_allocated);
1690
1691 free_zspage(first_page);
1692 }
1693}
1694
1695static struct page *isolate_source_page(struct size_class *class)
1696{
1697 struct page *page;
1698
1699 page = class->fullness_list[ZS_ALMOST_EMPTY];
1700 if (page)
1701 remove_zspage(page, class, ZS_ALMOST_EMPTY);
1702
1703 return page;
1704}
1705
1706static unsigned long __zs_compact(struct zs_pool *pool,
1707 struct size_class *class)
1708{
1709 int nr_to_migrate;
1710 struct zs_compact_control cc;
1711 struct page *src_page;
1712 struct page *dst_page = NULL;
1713 unsigned long nr_total_migrated = 0;
1714
Minchan Kim312fcae2015-04-15 16:15:30 -07001715 spin_lock(&class->lock);
1716 while ((src_page = isolate_source_page(class))) {
1717
1718 BUG_ON(!is_first_page(src_page));
1719
1720 /* The goal is to migrate all live objects in source page */
1721 nr_to_migrate = src_page->inuse;
1722 cc.index = 0;
1723 cc.s_page = src_page;
1724
1725 while ((dst_page = alloc_target_page(class))) {
1726 cc.d_page = dst_page;
1727 /*
1728 * If there is no more space in dst_page, try to
1729 * allocate another zspage.
1730 */
1731 if (!migrate_zspage(pool, class, &cc))
1732 break;
1733
1734 putback_zspage(pool, class, dst_page);
1735 nr_total_migrated += cc.nr_migrated;
1736 nr_to_migrate -= cc.nr_migrated;
1737 }
1738
1739 /* Stop if we couldn't find slot */
1740 if (dst_page == NULL)
1741 break;
1742
1743 putback_zspage(pool, class, dst_page);
1744 putback_zspage(pool, class, src_page);
1745 spin_unlock(&class->lock);
1746 nr_total_migrated += cc.nr_migrated;
1747 cond_resched();
1748 spin_lock(&class->lock);
1749 }
1750
1751 if (src_page)
1752 putback_zspage(pool, class, src_page);
1753
1754 spin_unlock(&class->lock);
1755
1756 return nr_total_migrated;
1757}
1758
1759unsigned long zs_compact(struct zs_pool *pool)
1760{
1761 int i;
1762 unsigned long nr_migrated = 0;
1763 struct size_class *class;
1764
1765 for (i = zs_size_classes - 1; i >= 0; i--) {
1766 class = pool->size_class[i];
1767 if (!class)
1768 continue;
1769 if (class->index != i)
1770 continue;
1771 nr_migrated += __zs_compact(pool, class);
1772 }
1773
Minchan Kim312fcae2015-04-15 16:15:30 -07001774 return nr_migrated;
1775}
1776EXPORT_SYMBOL_GPL(zs_compact);
1777
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001778/**
1779 * zs_create_pool - Creates an allocation pool to work from.
Seth Jennings0d145a52013-01-30 09:36:52 -06001780 * @flags: allocation flags used to allocate pool metadata
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001781 *
1782 * This function must be called before anything when using
1783 * the zsmalloc allocator.
1784 *
1785 * On success, a pointer to the newly created pool is returned,
1786 * otherwise NULL.
1787 */
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -08001788struct zs_pool *zs_create_pool(char *name, gfp_t flags)
Nitin Gupta61989a82012-01-09 16:51:56 -06001789{
Ganesh Mahendran18136652014-12-12 16:57:10 -08001790 int i;
Nitin Gupta61989a82012-01-09 16:51:56 -06001791 struct zs_pool *pool;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001792 struct size_class *prev_class = NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001793
Ganesh Mahendran18136652014-12-12 16:57:10 -08001794 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
Nitin Gupta61989a82012-01-09 16:51:56 -06001795 if (!pool)
1796 return NULL;
1797
Minchan Kim2e40e162015-04-15 16:15:23 -07001798 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1799 GFP_KERNEL);
1800 if (!pool->size_class) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001801 kfree(pool);
1802 return NULL;
1803 }
1804
Minchan Kim2e40e162015-04-15 16:15:23 -07001805 pool->name = kstrdup(name, GFP_KERNEL);
1806 if (!pool->name)
1807 goto err;
1808
1809 if (create_handle_cache(pool))
1810 goto err;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001811
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001812 /*
1813 * Iterate reversly, because, size of size_class that we want to use
1814 * for merging should be larger or equal to current size.
1815 */
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001816 for (i = zs_size_classes - 1; i >= 0; i--) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001817 int size;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001818 int pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001819 struct size_class *class;
1820
1821 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1822 if (size > ZS_MAX_ALLOC_SIZE)
1823 size = ZS_MAX_ALLOC_SIZE;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001824 pages_per_zspage = get_pages_per_zspage(size);
Nitin Gupta61989a82012-01-09 16:51:56 -06001825
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001826 /*
1827 * size_class is used for normal zsmalloc operation such
1828 * as alloc/free for that size. Although it is natural that we
1829 * have one size_class for each size, there is a chance that we
1830 * can get more memory utilization if we use one size_class for
1831 * many different sizes whose size_class have same
1832 * characteristics. So, we makes size_class point to
1833 * previous size_class if possible.
1834 */
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001835 if (prev_class) {
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001836 if (can_merge(prev_class, size, pages_per_zspage)) {
1837 pool->size_class[i] = prev_class;
1838 continue;
1839 }
1840 }
1841
1842 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1843 if (!class)
1844 goto err;
1845
Nitin Gupta61989a82012-01-09 16:51:56 -06001846 class->size = size;
1847 class->index = i;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001848 class->pages_per_zspage = pages_per_zspage;
Minchan Kim7b60a682015-04-15 16:15:39 -07001849 if (pages_per_zspage == 1 &&
1850 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1851 class->huge = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06001852 spin_lock_init(&class->lock);
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001853 pool->size_class[i] = class;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001854
1855 prev_class = class;
Nitin Gupta61989a82012-01-09 16:51:56 -06001856 }
1857
Nitin Gupta61989a82012-01-09 16:51:56 -06001858 pool->flags = flags;
Nitin Gupta61989a82012-01-09 16:51:56 -06001859
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001860 if (zs_pool_stat_create(name, pool))
1861 goto err;
1862
Nitin Gupta61989a82012-01-09 16:51:56 -06001863 return pool;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001864
1865err:
1866 zs_destroy_pool(pool);
1867 return NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001868}
1869EXPORT_SYMBOL_GPL(zs_create_pool);
1870
1871void zs_destroy_pool(struct zs_pool *pool)
1872{
1873 int i;
1874
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001875 zs_pool_stat_destroy(pool);
1876
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001877 for (i = 0; i < zs_size_classes; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001878 int fg;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001879 struct size_class *class = pool->size_class[i];
1880
1881 if (!class)
1882 continue;
1883
1884 if (class->index != i)
1885 continue;
Nitin Gupta61989a82012-01-09 16:51:56 -06001886
1887 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1888 if (class->fullness_list[fg]) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -04001889 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -06001890 class->size, fg);
1891 }
1892 }
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001893 kfree(class);
Nitin Gupta61989a82012-01-09 16:51:56 -06001894 }
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001895
Minchan Kim2e40e162015-04-15 16:15:23 -07001896 destroy_handle_cache(pool);
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001897 kfree(pool->size_class);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001898 kfree(pool->name);
Nitin Gupta61989a82012-01-09 16:51:56 -06001899 kfree(pool);
1900}
1901EXPORT_SYMBOL_GPL(zs_destroy_pool);
1902
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001903static int __init zs_init(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001904{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001905 int ret = zs_register_cpu_notifier();
Nitin Gupta61989a82012-01-09 16:51:56 -06001906
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001907 if (ret)
1908 goto notifier_fail;
Nitin Gupta61989a82012-01-09 16:51:56 -06001909
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001910 init_zs_size_classes();
Nitin Gupta61989a82012-01-09 16:51:56 -06001911
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001912#ifdef CONFIG_ZPOOL
1913 zpool_register_driver(&zs_zpool_driver);
1914#endif
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001915
1916 ret = zs_stat_init();
1917 if (ret) {
1918 pr_err("zs stat initialization failed\n");
1919 goto stat_fail;
1920 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001921 return 0;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001922
1923stat_fail:
1924#ifdef CONFIG_ZPOOL
1925 zpool_unregister_driver(&zs_zpool_driver);
1926#endif
1927notifier_fail:
1928 zs_unregister_cpu_notifier();
1929
1930 return ret;
Nitin Gupta61989a82012-01-09 16:51:56 -06001931}
Nitin Gupta61989a82012-01-09 16:51:56 -06001932
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001933static void __exit zs_exit(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001934{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001935#ifdef CONFIG_ZPOOL
1936 zpool_unregister_driver(&zs_zpool_driver);
1937#endif
1938 zs_unregister_cpu_notifier();
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001939
1940 zs_stat_exit();
Nitin Gupta61989a82012-01-09 16:51:56 -06001941}
Ben Hutchings069f1012012-06-20 02:31:11 +01001942
1943module_init(zs_init);
1944module_exit(zs_exit);
1945
1946MODULE_LICENSE("Dual BSD/GPL");
1947MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");