blob: 7a0b79b0a689967102663cdd9759b970e22bc24e [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:
Minchan Kim37836892016-07-26 15:23:23 -070019 * page->private: points to zspage
Minchan Kim48b48002016-07-26 15:23:31 -070020 * page->freelist(index): links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
22 * to store handle.
Ganesh Mahendranfd854462016-07-28 15:47:54 -070023 * page->units: first object offset in a subpage of zspage
Nitin Gupta2db51da2012-06-09 17:41:14 -070024 *
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
Xishi Qiu399d8ee2017-02-22 15:45:01 -080027 * PG_owner_priv_1: identifies the huge component page
Nitin Gupta2db51da2012-06-09 17:41:14 -070028 *
29 */
30
Dan Streetman4abaac92016-05-26 15:16:27 -070031#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32
Nitin Gupta61989a82012-01-09 16:51:56 -060033#include <linux/module.h>
34#include <linux/kernel.h>
Minchan Kim312fcae2015-04-15 16:15:30 -070035#include <linux/sched.h>
Ingo Molnar50d34392017-02-05 16:03:58 +010036#include <linux/magic.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060037#include <linux/bitops.h>
38#include <linux/errno.h>
39#include <linux/highmem.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060040#include <linux/string.h>
41#include <linux/slab.h>
Mike Rapoportca5999f2020-06-08 21:32:38 -070042#include <linux/pgtable.h>
Mike Rapoport65fddcf2020-06-08 21:32:42 -070043#include <asm/tlbflush.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060044#include <linux/cpumask.h>
45#include <linux/cpu.h>
Seth Jennings0cbb6132012-02-13 08:47:49 -060046#include <linux/vmalloc.h>
Sergey Senozhatsky759b26b2015-11-06 16:29:29 -080047#include <linux/preempt.h>
Seth Jennings0959c632012-08-08 15:12:17 +090048#include <linux/spinlock.h>
Aliaksei Karaliou93144ca2018-01-31 16:18:40 -080049#include <linux/shrinker.h>
Seth Jennings0959c632012-08-08 15:12:17 +090050#include <linux/types.h>
Ganesh Mahendran0f050d92015-02-12 15:00:54 -080051#include <linux/debugfs.h>
Minchan Kimbcf16472014-01-30 15:45:50 -080052#include <linux/zsmalloc.h>
Dan Streetmanc7957792014-08-06 16:08:38 -070053#include <linux/zpool.h>
Minchan Kim48b48002016-07-26 15:23:31 -070054#include <linux/mount.h>
David Howells8e9231f2019-03-25 16:38:23 +000055#include <linux/pseudo_fs.h>
Minchan Kimdd4123f2016-07-26 15:26:50 -070056#include <linux/migrate.h>
Henry Burns701d6782019-08-24 17:55:06 -070057#include <linux/wait.h>
Minchan Kim48b48002016-07-26 15:23:31 -070058#include <linux/pagemap.h>
Sergey Senozhatskycdc346b2018-01-04 16:18:02 -080059#include <linux/fs.h>
Minchan Kim48b48002016-07-26 15:23:31 -070060
61#define ZSPAGE_MAGIC 0x58
Seth Jennings0959c632012-08-08 15:12:17 +090062
63/*
64 * This must be power of 2 and greater than of equal to sizeof(link_free).
65 * These two conditions ensure that any 'struct link_free' itself doesn't
66 * span more than 1 page which avoids complex case of mapping 2 pages simply
67 * to restore link_free pointer values.
68 */
69#define ZS_ALIGN 8
70
71/*
72 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
73 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
74 */
75#define ZS_MAX_ZSPAGE_ORDER 2
76#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
77
Minchan Kim2e40e162015-04-15 16:15:23 -070078#define ZS_HANDLE_SIZE (sizeof(unsigned long))
79
Seth Jennings0959c632012-08-08 15:12:17 +090080/*
81 * Object location (<PFN>, <obj_idx>) is encoded as
Randy Dunlapb956b5a2020-08-11 18:33:31 -070082 * a single (unsigned long) handle value.
Seth Jennings0959c632012-08-08 15:12:17 +090083 *
Minchan Kimbfd093f2016-07-26 15:23:28 -070084 * Note that object index <obj_idx> starts from 0.
Seth Jennings0959c632012-08-08 15:12:17 +090085 *
86 * This is made more complicated by various memory models and PAE.
87 */
88
Kirill A. Shutemov02390b82018-02-14 14:16:49 +030089#ifndef MAX_POSSIBLE_PHYSMEM_BITS
90#ifdef MAX_PHYSMEM_BITS
91#define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
92#else
Seth Jennings0959c632012-08-08 15:12:17 +090093/*
94 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
95 * be PAGE_SHIFT
96 */
Kirill A. Shutemov02390b82018-02-14 14:16:49 +030097#define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
Seth Jennings0959c632012-08-08 15:12:17 +090098#endif
99#endif
Kirill A. Shutemov02390b82018-02-14 14:16:49 +0300100
101#define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
Minchan Kim312fcae2015-04-15 16:15:30 -0700102
103/*
104 * Memory for allocating for handle keeps object position by
105 * encoding <page, obj_idx> and the encoded value has a room
106 * in least bit(ie, look at obj_to_location).
107 * We use the bit to synchronize between object access by
108 * user and migration.
109 */
110#define HANDLE_PIN_BIT 0
111
112/*
113 * Head in allocated object should have OBJ_ALLOCATED_TAG
114 * to identify the object was allocated or not.
115 * It's okay to add the status bit in the least bit because
116 * header keeps handle which is 4byte-aligned address so we
117 * have room for two bit at least.
118 */
119#define OBJ_ALLOCATED_TAG 1
120#define OBJ_TAG_BITS 1
121#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
Seth Jennings0959c632012-08-08 15:12:17 +0900122#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
123
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -0700124#define FULLNESS_BITS 2
125#define CLASS_BITS 8
126#define ISOLATED_BITS 3
127#define MAGIC_VAL_BITS 8
128
Seth Jennings0959c632012-08-08 15:12:17 +0900129#define MAX(a, b) ((a) >= (b) ? (a) : (b))
130/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
131#define ZS_MIN_ALLOC_SIZE \
132 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
Minchan Kim2e40e162015-04-15 16:15:23 -0700133/* each chunk includes extra space to keep handle */
Minchan Kim7b60a682015-04-15 16:15:39 -0700134#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
Seth Jennings0959c632012-08-08 15:12:17 +0900135
136/*
Weijie Yang7eb52512014-06-04 16:11:08 -0700137 * On systems with 4K page size, this gives 255 size classes! There is a
Seth Jennings0959c632012-08-08 15:12:17 +0900138 * trader-off here:
139 * - Large number of size classes is potentially wasteful as free page are
140 * spread across these classes
141 * - Small number of size classes causes large internal fragmentation
142 * - Probably its better to use specific size classes (empirically
143 * determined). NOTE: all those class sizes must be set as multiple of
144 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
145 *
146 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
147 * (reason above)
148 */
Minchan Kim37836892016-07-26 15:23:23 -0700149#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -0700150#define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
151 ZS_SIZE_CLASS_DELTA) + 1)
Seth Jennings0959c632012-08-08 15:12:17 +0900152
Seth Jennings0959c632012-08-08 15:12:17 +0900153enum fullness_group {
Seth Jennings0959c632012-08-08 15:12:17 +0900154 ZS_EMPTY,
Minchan Kim48b48002016-07-26 15:23:31 -0700155 ZS_ALMOST_EMPTY,
156 ZS_ALMOST_FULL,
157 ZS_FULL,
158 NR_ZS_FULLNESS,
Seth Jennings0959c632012-08-08 15:12:17 +0900159};
160
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800161enum zs_stat_type {
Minchan Kim48b48002016-07-26 15:23:31 -0700162 CLASS_EMPTY,
163 CLASS_ALMOST_EMPTY,
164 CLASS_ALMOST_FULL,
165 CLASS_FULL,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800166 OBJ_ALLOCATED,
167 OBJ_USED,
Minchan Kim48b48002016-07-26 15:23:31 -0700168 NR_ZS_STAT_TYPE,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800169};
170
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800171struct zs_size_stat {
172 unsigned long objs[NR_ZS_STAT_TYPE];
173};
174
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700175#ifdef CONFIG_ZSMALLOC_STAT
176static struct dentry *zs_stat_root;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800177#endif
178
Minchan Kim48b48002016-07-26 15:23:31 -0700179#ifdef CONFIG_COMPACTION
180static struct vfsmount *zsmalloc_mnt;
181#endif
182
Seth Jennings0959c632012-08-08 15:12:17 +0900183/*
184 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
185 * n <= N / f, where
186 * n = number of allocated objects
187 * N = total number of objects zspage can store
Wang Sheng-Hui6dd97372014-10-09 15:29:59 -0700188 * f = fullness_threshold_frac
Seth Jennings0959c632012-08-08 15:12:17 +0900189 *
190 * Similarly, we assign zspage to:
191 * ZS_ALMOST_FULL when n > N / f
192 * ZS_EMPTY when n == 0
193 * ZS_FULL when n == N
194 *
195 * (see: fix_fullness_group())
196 */
197static const int fullness_threshold_frac = 4;
Sergey Senozhatsky010b4952018-04-05 16:24:43 -0700198static size_t huge_class_size;
Seth Jennings0959c632012-08-08 15:12:17 +0900199
200struct size_class {
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700201 spinlock_t lock;
Minchan Kim48b48002016-07-26 15:23:31 -0700202 struct list_head fullness_list[NR_ZS_FULLNESS];
Seth Jennings0959c632012-08-08 15:12:17 +0900203 /*
204 * Size of objects stored in this class. Must be multiple
205 * of ZS_ALIGN.
206 */
207 int size;
Minchan Kim1fc6e272016-07-26 15:23:11 -0700208 int objs_per_zspage;
Weijie Yang7dfa4612016-01-14 15:22:40 -0800209 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
210 int pages_per_zspage;
Minchan Kim48b48002016-07-26 15:23:31 -0700211
212 unsigned int index;
213 struct zs_size_stat stats;
Seth Jennings0959c632012-08-08 15:12:17 +0900214};
215
Minchan Kim48b48002016-07-26 15:23:31 -0700216/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
217static void SetPageHugeObject(struct page *page)
218{
219 SetPageOwnerPriv1(page);
220}
221
222static void ClearPageHugeObject(struct page *page)
223{
224 ClearPageOwnerPriv1(page);
225}
226
227static int PageHugeObject(struct page *page)
228{
229 return PageOwnerPriv1(page);
230}
231
Seth Jennings0959c632012-08-08 15:12:17 +0900232/*
233 * Placed within free objects to form a singly linked list.
Minchan Kim37836892016-07-26 15:23:23 -0700234 * For every zspage, zspage->freeobj gives head of this list.
Seth Jennings0959c632012-08-08 15:12:17 +0900235 *
236 * This must be power of 2 and less than or equal to ZS_ALIGN
237 */
238struct link_free {
Minchan Kim2e40e162015-04-15 16:15:23 -0700239 union {
240 /*
Minchan Kimbfd093f2016-07-26 15:23:28 -0700241 * Free object index;
Minchan Kim2e40e162015-04-15 16:15:23 -0700242 * It's valid for non-allocated object
243 */
Minchan Kimbfd093f2016-07-26 15:23:28 -0700244 unsigned long next;
Minchan Kim2e40e162015-04-15 16:15:23 -0700245 /*
246 * Handle of allocated object.
247 */
248 unsigned long handle;
249 };
Seth Jennings0959c632012-08-08 15:12:17 +0900250};
251
252struct zs_pool {
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800253 const char *name;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800254
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -0700255 struct size_class *size_class[ZS_SIZE_CLASSES];
Minchan Kim2e40e162015-04-15 16:15:23 -0700256 struct kmem_cache *handle_cachep;
Minchan Kim37836892016-07-26 15:23:23 -0700257 struct kmem_cache *zspage_cachep;
Seth Jennings0959c632012-08-08 15:12:17 +0900258
Minchan Kim13de8932014-10-09 15:29:48 -0700259 atomic_long_t pages_allocated;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800260
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -0700261 struct zs_pool_stats stats;
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -0700262
263 /* Compact classes */
264 struct shrinker shrinker;
Aliaksei Karaliou93144ca2018-01-31 16:18:40 -0800265
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800266#ifdef CONFIG_ZSMALLOC_STAT
267 struct dentry *stat_dentry;
268#endif
Minchan Kim48b48002016-07-26 15:23:31 -0700269#ifdef CONFIG_COMPACTION
270 struct inode *inode;
271 struct work_struct free_work;
Henry Burns701d6782019-08-24 17:55:06 -0700272 /* A wait queue for when migration races with async_free_zspage() */
273 struct wait_queue_head migration_wait;
274 atomic_long_t isolated_pages;
275 bool destroying;
Minchan Kim48b48002016-07-26 15:23:31 -0700276#endif
Seth Jennings0959c632012-08-08 15:12:17 +0900277};
Nitin Gupta61989a82012-01-09 16:51:56 -0600278
Minchan Kim37836892016-07-26 15:23:23 -0700279struct zspage {
280 struct {
281 unsigned int fullness:FULLNESS_BITS;
Minchan Kim85d492f2017-04-13 14:56:40 -0700282 unsigned int class:CLASS_BITS + 1;
Minchan Kim48b48002016-07-26 15:23:31 -0700283 unsigned int isolated:ISOLATED_BITS;
284 unsigned int magic:MAGIC_VAL_BITS;
Minchan Kim37836892016-07-26 15:23:23 -0700285 };
286 unsigned int inuse;
Minchan Kimbfd093f2016-07-26 15:23:28 -0700287 unsigned int freeobj;
Minchan Kim37836892016-07-26 15:23:23 -0700288 struct page *first_page;
289 struct list_head list; /* fullness list */
Minchan Kim48b48002016-07-26 15:23:31 -0700290#ifdef CONFIG_COMPACTION
291 rwlock_t lock;
292#endif
Minchan Kim37836892016-07-26 15:23:23 -0700293};
Nitin Gupta61989a82012-01-09 16:51:56 -0600294
Seth Jenningsf5536462012-07-18 11:55:56 -0500295struct mapping_area {
Seth Jenningsf5536462012-07-18 11:55:56 -0500296 char *vm_buf; /* copy buffer for objects that span pages */
Seth Jenningsf5536462012-07-18 11:55:56 -0500297 char *vm_addr; /* address of kmap_atomic()'ed pages */
298 enum zs_mapmode vm_mm; /* mapping mode */
299};
300
Minchan Kim48b48002016-07-26 15:23:31 -0700301#ifdef CONFIG_COMPACTION
302static int zs_register_migration(struct zs_pool *pool);
303static void zs_unregister_migration(struct zs_pool *pool);
304static void migrate_lock_init(struct zspage *zspage);
305static void migrate_read_lock(struct zspage *zspage);
306static void migrate_read_unlock(struct zspage *zspage);
307static void kick_deferred_free(struct zs_pool *pool);
308static void init_deferred_free(struct zs_pool *pool);
309static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
310#else
311static int zsmalloc_mount(void) { return 0; }
312static void zsmalloc_unmount(void) {}
313static int zs_register_migration(struct zs_pool *pool) { return 0; }
314static void zs_unregister_migration(struct zs_pool *pool) {}
315static void migrate_lock_init(struct zspage *zspage) {}
316static void migrate_read_lock(struct zspage *zspage) {}
317static void migrate_read_unlock(struct zspage *zspage) {}
318static void kick_deferred_free(struct zs_pool *pool) {}
319static void init_deferred_free(struct zs_pool *pool) {}
320static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
321#endif
322
Minchan Kim37836892016-07-26 15:23:23 -0700323static int create_cache(struct zs_pool *pool)
Minchan Kim2e40e162015-04-15 16:15:23 -0700324{
325 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
326 0, 0, NULL);
Minchan Kim37836892016-07-26 15:23:23 -0700327 if (!pool->handle_cachep)
328 return 1;
329
330 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
331 0, 0, NULL);
332 if (!pool->zspage_cachep) {
333 kmem_cache_destroy(pool->handle_cachep);
334 pool->handle_cachep = NULL;
335 return 1;
336 }
337
338 return 0;
Minchan Kim2e40e162015-04-15 16:15:23 -0700339}
340
Minchan Kim37836892016-07-26 15:23:23 -0700341static void destroy_cache(struct zs_pool *pool)
Minchan Kim2e40e162015-04-15 16:15:23 -0700342{
Sergey Senozhatskycd10add2015-09-08 15:04:55 -0700343 kmem_cache_destroy(pool->handle_cachep);
Minchan Kim37836892016-07-26 15:23:23 -0700344 kmem_cache_destroy(pool->zspage_cachep);
Minchan Kim2e40e162015-04-15 16:15:23 -0700345}
346
Minchan Kim37836892016-07-26 15:23:23 -0700347static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
Minchan Kim2e40e162015-04-15 16:15:23 -0700348{
349 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
Minchan Kim48b48002016-07-26 15:23:31 -0700350 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
Minchan Kim2e40e162015-04-15 16:15:23 -0700351}
352
Minchan Kim37836892016-07-26 15:23:23 -0700353static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
Minchan Kim2e40e162015-04-15 16:15:23 -0700354{
355 kmem_cache_free(pool->handle_cachep, (void *)handle);
356}
357
Minchan Kim37836892016-07-26 15:23:23 -0700358static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
359{
Minchan Kim48b48002016-07-26 15:23:31 -0700360 return kmem_cache_alloc(pool->zspage_cachep,
361 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
Xishi Qiu399d8ee2017-02-22 15:45:01 -0800362}
Minchan Kim37836892016-07-26 15:23:23 -0700363
364static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
365{
366 kmem_cache_free(pool->zspage_cachep, zspage);
367}
368
Minchan Kim2e40e162015-04-15 16:15:23 -0700369static void record_obj(unsigned long handle, unsigned long obj)
370{
Junil Leec102f07c2016-01-20 14:58:18 -0800371 /*
372 * lsb of @obj represents handle lock while other bits
373 * represent object value the handle is pointing so
374 * updating shouldn't do store tearing.
375 */
376 WRITE_ONCE(*(unsigned long *)handle, obj);
Minchan Kim2e40e162015-04-15 16:15:23 -0700377}
378
Dan Streetmanc7957792014-08-06 16:08:38 -0700379/* zpool driver */
380
381#ifdef CONFIG_ZPOOL
382
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800383static void *zs_zpool_create(const char *name, gfp_t gfp,
Krzysztof Kozlowski78672772015-09-08 15:05:03 -0700384 const struct zpool_ops *zpool_ops,
Dan Streetman479305f2015-06-25 15:00:40 -0700385 struct zpool *zpool)
Dan Streetmanc7957792014-08-06 16:08:38 -0700386{
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -0700387 /*
388 * Ignore global gfp flags: zs_malloc() may be invoked from
389 * different contexts and its caller must provide a valid
390 * gfp mask.
391 */
392 return zs_create_pool(name);
Dan Streetmanc7957792014-08-06 16:08:38 -0700393}
394
395static void zs_zpool_destroy(void *pool)
396{
397 zs_destroy_pool(pool);
398}
399
400static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
401 unsigned long *handle)
402{
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -0700403 *handle = zs_malloc(pool, size, gfp);
Dan Streetmanc7957792014-08-06 16:08:38 -0700404 return *handle ? 0 : -1;
405}
406static void zs_zpool_free(void *pool, unsigned long handle)
407{
408 zs_free(pool, handle);
409}
410
Dan Streetmanc7957792014-08-06 16:08:38 -0700411static void *zs_zpool_map(void *pool, unsigned long handle,
412 enum zpool_mapmode mm)
413{
414 enum zs_mapmode zs_mm;
415
416 switch (mm) {
417 case ZPOOL_MM_RO:
418 zs_mm = ZS_MM_RO;
419 break;
420 case ZPOOL_MM_WO:
421 zs_mm = ZS_MM_WO;
422 break;
Joe Perchese4a9bc52020-04-06 20:08:39 -0700423 case ZPOOL_MM_RW:
Dan Streetmanc7957792014-08-06 16:08:38 -0700424 default:
425 zs_mm = ZS_MM_RW;
426 break;
427 }
428
429 return zs_map_object(pool, handle, zs_mm);
430}
431static void zs_zpool_unmap(void *pool, unsigned long handle)
432{
433 zs_unmap_object(pool, handle);
434}
435
436static u64 zs_zpool_total_size(void *pool)
437{
Minchan Kim722cdc12014-10-09 15:29:50 -0700438 return zs_get_total_pages(pool) << PAGE_SHIFT;
Dan Streetmanc7957792014-08-06 16:08:38 -0700439}
440
441static struct zpool_driver zs_zpool_driver = {
Hui Zhuc165f252019-09-23 15:39:37 -0700442 .type = "zsmalloc",
443 .owner = THIS_MODULE,
444 .create = zs_zpool_create,
445 .destroy = zs_zpool_destroy,
446 .malloc_support_movable = true,
447 .malloc = zs_zpool_malloc,
448 .free = zs_zpool_free,
449 .map = zs_zpool_map,
450 .unmap = zs_zpool_unmap,
451 .total_size = zs_zpool_total_size,
Dan Streetmanc7957792014-08-06 16:08:38 -0700452};
453
Kees Cook137f8cf2014-08-29 15:18:40 -0700454MODULE_ALIAS("zpool-zsmalloc");
Dan Streetmanc7957792014-08-06 16:08:38 -0700455#endif /* CONFIG_ZPOOL */
456
Nitin Gupta61989a82012-01-09 16:51:56 -0600457/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
458static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
459
Minchan Kim48b48002016-07-26 15:23:31 -0700460static bool is_zspage_isolated(struct zspage *zspage)
461{
462 return zspage->isolated;
463}
464
Nick Desaulniers3457f412017-07-10 15:47:26 -0700465static __maybe_unused int is_first_page(struct page *page)
Nitin Gupta61989a82012-01-09 16:51:56 -0600466{
Minchan Kima27545bf2012-04-25 15:23:09 +0900467 return PagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600468}
469
Minchan Kim48b48002016-07-26 15:23:31 -0700470/* Protected by class->lock */
Minchan Kim37836892016-07-26 15:23:23 -0700471static inline int get_zspage_inuse(struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600472{
Minchan Kim37836892016-07-26 15:23:23 -0700473 return zspage->inuse;
Nitin Gupta61989a82012-01-09 16:51:56 -0600474}
475
Minchan Kim4f420472016-07-26 15:23:17 -0700476
Minchan Kim37836892016-07-26 15:23:23 -0700477static inline void mod_zspage_inuse(struct zspage *zspage, int val)
Minchan Kim4f420472016-07-26 15:23:17 -0700478{
Minchan Kim37836892016-07-26 15:23:23 -0700479 zspage->inuse += val;
Minchan Kim4f420472016-07-26 15:23:17 -0700480}
481
Minchan Kim48b48002016-07-26 15:23:31 -0700482static inline struct page *get_first_page(struct zspage *zspage)
483{
484 struct page *first_page = zspage->first_page;
485
486 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
487 return first_page;
488}
489
Minchan Kim4f420472016-07-26 15:23:17 -0700490static inline int get_first_obj_offset(struct page *page)
491{
Minchan Kim48b48002016-07-26 15:23:31 -0700492 return page->units;
Minchan Kim4f420472016-07-26 15:23:17 -0700493}
494
495static inline void set_first_obj_offset(struct page *page, int offset)
496{
Minchan Kim48b48002016-07-26 15:23:31 -0700497 page->units = offset;
Minchan Kim4f420472016-07-26 15:23:17 -0700498}
499
Minchan Kimbfd093f2016-07-26 15:23:28 -0700500static inline unsigned int get_freeobj(struct zspage *zspage)
Minchan Kim4f420472016-07-26 15:23:17 -0700501{
Minchan Kimbfd093f2016-07-26 15:23:28 -0700502 return zspage->freeobj;
Minchan Kim4f420472016-07-26 15:23:17 -0700503}
504
Minchan Kimbfd093f2016-07-26 15:23:28 -0700505static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
Minchan Kim4f420472016-07-26 15:23:17 -0700506{
Minchan Kimbfd093f2016-07-26 15:23:28 -0700507 zspage->freeobj = obj;
Minchan Kim4f420472016-07-26 15:23:17 -0700508}
509
Minchan Kim37836892016-07-26 15:23:23 -0700510static void get_zspage_mapping(struct zspage *zspage,
Minchan Kima4209462016-05-20 16:59:36 -0700511 unsigned int *class_idx,
Nitin Gupta61989a82012-01-09 16:51:56 -0600512 enum fullness_group *fullness)
513{
Minchan Kim48b48002016-07-26 15:23:31 -0700514 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
515
Minchan Kim37836892016-07-26 15:23:23 -0700516 *fullness = zspage->fullness;
517 *class_idx = zspage->class;
Nitin Gupta61989a82012-01-09 16:51:56 -0600518}
519
Minchan Kim37836892016-07-26 15:23:23 -0700520static void set_zspage_mapping(struct zspage *zspage,
Minchan Kima4209462016-05-20 16:59:36 -0700521 unsigned int class_idx,
Nitin Gupta61989a82012-01-09 16:51:56 -0600522 enum fullness_group fullness)
523{
Minchan Kim37836892016-07-26 15:23:23 -0700524 zspage->class = class_idx;
525 zspage->fullness = fullness;
Nitin Gupta61989a82012-01-09 16:51:56 -0600526}
527
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900528/*
529 * zsmalloc divides the pool into various size classes where each
530 * class maintains a list of zspages where each zspage is divided
531 * into equal sized chunks. Each allocation falls into one of these
532 * classes depending on its size. This function returns index of the
533 * size class which has chunk size big enough to hold the give size.
534 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600535static int get_size_class_index(int size)
536{
537 int idx = 0;
538
539 if (likely(size > ZS_MIN_ALLOC_SIZE))
540 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
541 ZS_SIZE_CLASS_DELTA);
542
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -0700543 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
Nitin Gupta61989a82012-01-09 16:51:56 -0600544}
545
Matthias Kaehlcke3eb95fe2017-09-08 16:13:02 -0700546/* type can be of enum type zs_stat_type or fullness_group */
Minchan Kim248ca1b2015-04-15 16:15:42 -0700547static inline void zs_stat_inc(struct size_class *class,
Matthias Kaehlcke3eb95fe2017-09-08 16:13:02 -0700548 int type, unsigned long cnt)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700549{
Minchan Kim48b48002016-07-26 15:23:31 -0700550 class->stats.objs[type] += cnt;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700551}
552
Matthias Kaehlcke3eb95fe2017-09-08 16:13:02 -0700553/* type can be of enum type zs_stat_type or fullness_group */
Minchan Kim248ca1b2015-04-15 16:15:42 -0700554static inline void zs_stat_dec(struct size_class *class,
Matthias Kaehlcke3eb95fe2017-09-08 16:13:02 -0700555 int type, unsigned long cnt)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700556{
Minchan Kim48b48002016-07-26 15:23:31 -0700557 class->stats.objs[type] -= cnt;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700558}
559
Matthias Kaehlcke3eb95fe2017-09-08 16:13:02 -0700560/* type can be of enum type zs_stat_type or fullness_group */
Minchan Kim248ca1b2015-04-15 16:15:42 -0700561static inline unsigned long zs_stat_get(struct size_class *class,
Matthias Kaehlcke3eb95fe2017-09-08 16:13:02 -0700562 int type)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700563{
Minchan Kim48b48002016-07-26 15:23:31 -0700564 return class->stats.objs[type];
Minchan Kim248ca1b2015-04-15 16:15:42 -0700565}
566
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700567#ifdef CONFIG_ZSMALLOC_STAT
568
Dan Streetman4abaac92016-05-26 15:16:27 -0700569static void __init zs_stat_init(void)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700570{
Dan Streetman4abaac92016-05-26 15:16:27 -0700571 if (!debugfs_initialized()) {
572 pr_warn("debugfs not available, stat dir not created\n");
573 return;
574 }
Minchan Kim248ca1b2015-04-15 16:15:42 -0700575
576 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700577}
578
579static void __exit zs_stat_exit(void)
580{
581 debugfs_remove_recursive(zs_stat_root);
582}
583
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700584static unsigned long zs_can_compact(struct size_class *class);
585
Minchan Kim248ca1b2015-04-15 16:15:42 -0700586static int zs_stats_size_show(struct seq_file *s, void *v)
587{
588 int i;
589 struct zs_pool *pool = s->private;
590 struct size_class *class;
591 int objs_per_zspage;
592 unsigned long class_almost_full, class_almost_empty;
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700593 unsigned long obj_allocated, obj_used, pages_used, freeable;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700594 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
595 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700596 unsigned long total_freeable = 0;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700597
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700598 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
Minchan Kim248ca1b2015-04-15 16:15:42 -0700599 "class", "size", "almost_full", "almost_empty",
600 "obj_allocated", "obj_used", "pages_used",
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700601 "pages_per_zspage", "freeable");
Minchan Kim248ca1b2015-04-15 16:15:42 -0700602
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -0700603 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
Minchan Kim248ca1b2015-04-15 16:15:42 -0700604 class = pool->size_class[i];
605
606 if (class->index != i)
607 continue;
608
609 spin_lock(&class->lock);
610 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
611 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
612 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
613 obj_used = zs_stat_get(class, OBJ_USED);
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700614 freeable = zs_can_compact(class);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700615 spin_unlock(&class->lock);
616
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -0700617 objs_per_zspage = class->objs_per_zspage;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700618 pages_used = obj_allocated / objs_per_zspage *
619 class->pages_per_zspage;
620
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700621 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
622 " %10lu %10lu %16d %8lu\n",
Minchan Kim248ca1b2015-04-15 16:15:42 -0700623 i, class->size, class_almost_full, class_almost_empty,
624 obj_allocated, obj_used, pages_used,
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700625 class->pages_per_zspage, freeable);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700626
627 total_class_almost_full += class_almost_full;
628 total_class_almost_empty += class_almost_empty;
629 total_objs += obj_allocated;
630 total_used_objs += obj_used;
631 total_pages += pages_used;
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700632 total_freeable += freeable;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700633 }
634
635 seq_puts(s, "\n");
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700636 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
Minchan Kim248ca1b2015-04-15 16:15:42 -0700637 "Total", "", total_class_almost_full,
638 total_class_almost_empty, total_objs,
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700639 total_used_objs, total_pages, "", total_freeable);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700640
641 return 0;
642}
Andy Shevchenko5ad35092018-04-05 16:23:16 -0700643DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700644
Dan Streetmand34f6152016-05-20 16:59:56 -0700645static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700646{
Dan Streetman4abaac92016-05-26 15:16:27 -0700647 if (!zs_stat_root) {
648 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
Dan Streetmand34f6152016-05-20 16:59:56 -0700649 return;
Dan Streetman4abaac92016-05-26 15:16:27 -0700650 }
Minchan Kim248ca1b2015-04-15 16:15:42 -0700651
Greg Kroah-Hartman42685092019-01-22 16:21:09 +0100652 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700653
Greg Kroah-Hartman42685092019-01-22 16:21:09 +0100654 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
655 &zs_stats_size_fops);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700656}
657
658static void zs_pool_stat_destroy(struct zs_pool *pool)
659{
660 debugfs_remove_recursive(pool->stat_dentry);
661}
662
663#else /* CONFIG_ZSMALLOC_STAT */
Dan Streetman4abaac92016-05-26 15:16:27 -0700664static void __init zs_stat_init(void)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700665{
Minchan Kim248ca1b2015-04-15 16:15:42 -0700666}
667
668static void __exit zs_stat_exit(void)
669{
670}
671
Dan Streetmand34f6152016-05-20 16:59:56 -0700672static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700673{
Minchan Kim248ca1b2015-04-15 16:15:42 -0700674}
675
676static inline void zs_pool_stat_destroy(struct zs_pool *pool)
677{
678}
Minchan Kim248ca1b2015-04-15 16:15:42 -0700679#endif
680
Minchan Kim48b48002016-07-26 15:23:31 -0700681
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900682/*
683 * For each size class, zspages are divided into different groups
684 * depending on how "full" they are. This was done so that we could
685 * easily find empty or nearly empty zspages when we try to shrink
686 * the pool (not yet implemented). This function returns fullness
687 * status of the given page.
688 */
Minchan Kim1fc6e272016-07-26 15:23:11 -0700689static enum fullness_group get_fullness_group(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700690 struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600691{
Minchan Kim1fc6e272016-07-26 15:23:11 -0700692 int inuse, objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600693 enum fullness_group fg;
Minchan Kim830e4bc2016-05-20 16:59:39 -0700694
Minchan Kim37836892016-07-26 15:23:23 -0700695 inuse = get_zspage_inuse(zspage);
Minchan Kim1fc6e272016-07-26 15:23:11 -0700696 objs_per_zspage = class->objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600697
698 if (inuse == 0)
699 fg = ZS_EMPTY;
Minchan Kim1fc6e272016-07-26 15:23:11 -0700700 else if (inuse == objs_per_zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600701 fg = ZS_FULL;
Minchan Kim1fc6e272016-07-26 15:23:11 -0700702 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
Nitin Gupta61989a82012-01-09 16:51:56 -0600703 fg = ZS_ALMOST_EMPTY;
704 else
705 fg = ZS_ALMOST_FULL;
706
707 return fg;
708}
709
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900710/*
711 * Each size class maintains various freelists and zspages are assigned
712 * to one of these freelists based on the number of live objects they
713 * have. This functions inserts the given zspage into the freelist
714 * identified by <class, fullness_group>.
715 */
Minchan Kim251cbb92016-05-20 16:59:42 -0700716static void insert_zspage(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700717 struct zspage *zspage,
718 enum fullness_group fullness)
Nitin Gupta61989a82012-01-09 16:51:56 -0600719{
Minchan Kim37836892016-07-26 15:23:23 -0700720 struct zspage *head;
Nitin Gupta61989a82012-01-09 16:51:56 -0600721
Minchan Kim48b48002016-07-26 15:23:31 -0700722 zs_stat_inc(class, fullness, 1);
Minchan Kim37836892016-07-26 15:23:23 -0700723 head = list_first_entry_or_null(&class->fullness_list[fullness],
724 struct zspage, list);
Sergey Senozhatsky58f17112015-09-08 15:04:44 -0700725 /*
Minchan Kim37836892016-07-26 15:23:23 -0700726 * We want to see more ZS_FULL pages and less almost empty/full.
727 * Put pages with higher ->inuse first.
Sergey Senozhatsky58f17112015-09-08 15:04:44 -0700728 */
Minchan Kim37836892016-07-26 15:23:23 -0700729 if (head) {
730 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
731 list_add(&zspage->list, &head->list);
732 return;
733 }
734 }
735 list_add(&zspage->list, &class->fullness_list[fullness]);
Nitin Gupta61989a82012-01-09 16:51:56 -0600736}
737
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900738/*
739 * This function removes the given zspage from the freelist identified
740 * by <class, fullness_group>.
741 */
Minchan Kim251cbb92016-05-20 16:59:42 -0700742static void remove_zspage(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700743 struct zspage *zspage,
744 enum fullness_group fullness)
Nitin Gupta61989a82012-01-09 16:51:56 -0600745{
Minchan Kim37836892016-07-26 15:23:23 -0700746 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
Minchan Kim48b48002016-07-26 15:23:31 -0700747 VM_BUG_ON(is_zspage_isolated(zspage));
Nitin Gupta61989a82012-01-09 16:51:56 -0600748
Minchan Kim37836892016-07-26 15:23:23 -0700749 list_del_init(&zspage->list);
Minchan Kim48b48002016-07-26 15:23:31 -0700750 zs_stat_dec(class, fullness, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600751}
752
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900753/*
754 * Each size class maintains zspages in different fullness groups depending
755 * on the number of live objects they contain. When allocating or freeing
756 * objects, the fullness status of the page can change, say, from ALMOST_FULL
757 * to ALMOST_EMPTY when freeing an object. This function checks if such
758 * a status change has occurred for the given page and accordingly moves the
759 * page from the freelist of the old fullness group to that of the new
760 * fullness group.
761 */
Minchan Kimc7806262015-04-15 16:15:26 -0700762static enum fullness_group fix_fullness_group(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700763 struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600764{
765 int class_idx;
Nitin Gupta61989a82012-01-09 16:51:56 -0600766 enum fullness_group currfg, newfg;
767
Minchan Kim37836892016-07-26 15:23:23 -0700768 get_zspage_mapping(zspage, &class_idx, &currfg);
769 newfg = get_fullness_group(class, zspage);
Nitin Gupta61989a82012-01-09 16:51:56 -0600770 if (newfg == currfg)
771 goto out;
772
Minchan Kim48b48002016-07-26 15:23:31 -0700773 if (!is_zspage_isolated(zspage)) {
774 remove_zspage(class, zspage, currfg);
775 insert_zspage(class, zspage, newfg);
776 }
777
Minchan Kim37836892016-07-26 15:23:23 -0700778 set_zspage_mapping(zspage, class_idx, newfg);
Nitin Gupta61989a82012-01-09 16:51:56 -0600779
780out:
781 return newfg;
782}
783
784/*
785 * We have to decide on how many pages to link together
786 * to form a zspage for each size class. This is important
787 * to reduce wastage due to unusable space left at end of
788 * each zspage which is given as:
Yinghao Xie888fa3742015-04-15 16:15:49 -0700789 * wastage = Zp % class_size
790 * usage = Zp - wastage
Nitin Gupta61989a82012-01-09 16:51:56 -0600791 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
792 *
793 * For example, for size class of 3/8 * PAGE_SIZE, we should
794 * link together 3 PAGE_SIZE sized pages to form a zspage
795 * since then we can perfectly fit in 8 such objects.
796 */
Minchan Kim2e3b61542012-05-03 15:40:39 +0900797static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600798{
799 int i, max_usedpc = 0;
800 /* zspage order which gives maximum used size per KB */
801 int max_usedpc_order = 1;
802
Seth Jennings84d4faa2012-03-05 11:33:21 -0600803 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600804 int zspage_size;
805 int waste, usedpc;
806
807 zspage_size = i * PAGE_SIZE;
808 waste = zspage_size % class_size;
809 usedpc = (zspage_size - waste) * 100 / zspage_size;
810
811 if (usedpc > max_usedpc) {
812 max_usedpc = usedpc;
813 max_usedpc_order = i;
814 }
815 }
816
817 return max_usedpc_order;
818}
819
Minchan Kim37836892016-07-26 15:23:23 -0700820static struct zspage *get_zspage(struct page *page)
Nitin Gupta61989a82012-01-09 16:51:56 -0600821{
Minchan Kim48b48002016-07-26 15:23:31 -0700822 struct zspage *zspage = (struct zspage *)page->private;
823
824 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
825 return zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600826}
827
828static struct page *get_next_page(struct page *page)
829{
Minchan Kim48b48002016-07-26 15:23:31 -0700830 if (unlikely(PageHugeObject(page)))
831 return NULL;
832
833 return page->freelist;
Nitin Gupta61989a82012-01-09 16:51:56 -0600834}
835
Minchan Kimbfd093f2016-07-26 15:23:28 -0700836/**
837 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
Mike Rapoporte8b098f2018-04-05 16:24:57 -0700838 * @obj: the encoded object value
Minchan Kimbfd093f2016-07-26 15:23:28 -0700839 * @page: page object resides in zspage
840 * @obj_idx: object index
Olav Haugan67296872013-11-22 09:30:41 -0800841 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700842static void obj_to_location(unsigned long obj, struct page **page,
Minchan Kimbfd093f2016-07-26 15:23:28 -0700843 unsigned int *obj_idx)
Nitin Gupta61989a82012-01-09 16:51:56 -0600844{
Minchan Kim312fcae2015-04-15 16:15:30 -0700845 obj >>= OBJ_TAG_BITS;
846 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
847 *obj_idx = (obj & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600848}
849
Minchan Kimbfd093f2016-07-26 15:23:28 -0700850/**
851 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
852 * @page: page object resides in zspage
853 * @obj_idx: object index
854 */
855static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
856{
857 unsigned long obj;
858
859 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
860 obj |= obj_idx & OBJ_INDEX_MASK;
861 obj <<= OBJ_TAG_BITS;
862
863 return obj;
864}
865
Minchan Kim2e40e162015-04-15 16:15:23 -0700866static unsigned long handle_to_obj(unsigned long handle)
867{
868 return *(unsigned long *)handle;
869}
870
Minchan Kim48b48002016-07-26 15:23:31 -0700871static unsigned long obj_to_head(struct page *page, void *obj)
Minchan Kim312fcae2015-04-15 16:15:30 -0700872{
Minchan Kim48b48002016-07-26 15:23:31 -0700873 if (unlikely(PageHugeObject(page))) {
Minchan Kim830e4bc2016-05-20 16:59:39 -0700874 VM_BUG_ON_PAGE(!is_first_page(page), page);
Minchan Kim37836892016-07-26 15:23:23 -0700875 return page->index;
Minchan Kim7b60a682015-04-15 16:15:39 -0700876 } else
877 return *(unsigned long *)obj;
Minchan Kim312fcae2015-04-15 16:15:30 -0700878}
879
Minchan Kim48b48002016-07-26 15:23:31 -0700880static inline int testpin_tag(unsigned long handle)
881{
882 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
883}
884
Minchan Kim312fcae2015-04-15 16:15:30 -0700885static inline int trypin_tag(unsigned long handle)
886{
Minchan Kim1b8320b2016-07-26 15:23:14 -0700887 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
Minchan Kim312fcae2015-04-15 16:15:30 -0700888}
889
Jules Irenge70c7ec92020-04-06 20:08:27 -0700890static void pin_tag(unsigned long handle) __acquires(bitlock)
Minchan Kim312fcae2015-04-15 16:15:30 -0700891{
Minchan Kim1b8320b2016-07-26 15:23:14 -0700892 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
Minchan Kim312fcae2015-04-15 16:15:30 -0700893}
894
Jules Irengebc22b182020-04-06 20:08:30 -0700895static void unpin_tag(unsigned long handle) __releases(bitlock)
Minchan Kim312fcae2015-04-15 16:15:30 -0700896{
Minchan Kim1b8320b2016-07-26 15:23:14 -0700897 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
Minchan Kim312fcae2015-04-15 16:15:30 -0700898}
899
Nitin Guptaf4477e92012-04-02 09:13:56 -0500900static void reset_page(struct page *page)
901{
Minchan Kim48b48002016-07-26 15:23:31 -0700902 __ClearPageMovable(page);
Ganesh Mahendran18fd06b2016-07-28 15:48:00 -0700903 ClearPagePrivate(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500904 set_page_private(page, 0);
Minchan Kim48b48002016-07-26 15:23:31 -0700905 page_mapcount_reset(page);
906 ClearPageHugeObject(page);
907 page->freelist = NULL;
Nitin Guptaf4477e92012-04-02 09:13:56 -0500908}
909
Colin Ian King4d0a5402018-08-17 15:46:50 -0700910static int trylock_zspage(struct zspage *zspage)
Minchan Kim48b48002016-07-26 15:23:31 -0700911{
912 struct page *cursor, *fail;
913
914 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
915 get_next_page(cursor)) {
916 if (!trylock_page(cursor)) {
917 fail = cursor;
918 goto unlock;
919 }
920 }
921
922 return 1;
923unlock:
924 for (cursor = get_first_page(zspage); cursor != fail; cursor =
925 get_next_page(cursor))
926 unlock_page(cursor);
927
928 return 0;
929}
930
931static void __free_zspage(struct zs_pool *pool, struct size_class *class,
932 struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600933{
Minchan Kim37836892016-07-26 15:23:23 -0700934 struct page *page, *next;
Minchan Kim48b48002016-07-26 15:23:31 -0700935 enum fullness_group fg;
936 unsigned int class_idx;
937
938 get_zspage_mapping(zspage, &class_idx, &fg);
939
940 assert_spin_locked(&class->lock);
Nitin Gupta61989a82012-01-09 16:51:56 -0600941
Minchan Kim37836892016-07-26 15:23:23 -0700942 VM_BUG_ON(get_zspage_inuse(zspage));
Minchan Kim48b48002016-07-26 15:23:31 -0700943 VM_BUG_ON(fg != ZS_EMPTY);
Nitin Gupta61989a82012-01-09 16:51:56 -0600944
Minchan Kim48b48002016-07-26 15:23:31 -0700945 next = page = get_first_page(zspage);
Minchan Kim37836892016-07-26 15:23:23 -0700946 do {
Minchan Kim48b48002016-07-26 15:23:31 -0700947 VM_BUG_ON_PAGE(!PageLocked(page), page);
948 next = get_next_page(page);
Minchan Kim37836892016-07-26 15:23:23 -0700949 reset_page(page);
Minchan Kim48b48002016-07-26 15:23:31 -0700950 unlock_page(page);
Minchan Kim91537fe2016-07-26 15:24:45 -0700951 dec_zone_page_state(page, NR_ZSPAGES);
Minchan Kim37836892016-07-26 15:23:23 -0700952 put_page(page);
953 page = next;
954 } while (page != NULL);
Nitin Gupta61989a82012-01-09 16:51:56 -0600955
Minchan Kim37836892016-07-26 15:23:23 -0700956 cache_free_zspage(pool, zspage);
Minchan Kim48b48002016-07-26 15:23:31 -0700957
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -0700958 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
Minchan Kim48b48002016-07-26 15:23:31 -0700959 atomic_long_sub(class->pages_per_zspage,
960 &pool->pages_allocated);
961}
962
963static void free_zspage(struct zs_pool *pool, struct size_class *class,
964 struct zspage *zspage)
965{
966 VM_BUG_ON(get_zspage_inuse(zspage));
967 VM_BUG_ON(list_empty(&zspage->list));
968
969 if (!trylock_zspage(zspage)) {
970 kick_deferred_free(pool);
971 return;
972 }
973
974 remove_zspage(class, zspage, ZS_EMPTY);
975 __free_zspage(pool, class, zspage);
Nitin Gupta61989a82012-01-09 16:51:56 -0600976}
977
978/* Initialize a newly allocated zspage */
Minchan Kim37836892016-07-26 15:23:23 -0700979static void init_zspage(struct size_class *class, struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600980{
Minchan Kimbfd093f2016-07-26 15:23:28 -0700981 unsigned int freeobj = 1;
Nitin Gupta61989a82012-01-09 16:51:56 -0600982 unsigned long off = 0;
Minchan Kim48b48002016-07-26 15:23:31 -0700983 struct page *page = get_first_page(zspage);
Minchan Kim830e4bc2016-05-20 16:59:39 -0700984
Nitin Gupta61989a82012-01-09 16:51:56 -0600985 while (page) {
986 struct page *next_page;
987 struct link_free *link;
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800988 void *vaddr;
Nitin Gupta61989a82012-01-09 16:51:56 -0600989
Minchan Kim37836892016-07-26 15:23:23 -0700990 set_first_obj_offset(page, off);
Nitin Gupta61989a82012-01-09 16:51:56 -0600991
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800992 vaddr = kmap_atomic(page);
993 link = (struct link_free *)vaddr + off / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600994
Dan Streetman5538c562014-10-09 15:30:01 -0700995 while ((off += class->size) < PAGE_SIZE) {
Minchan Kim3b1d9ca2016-07-26 15:23:37 -0700996 link->next = freeobj++ << OBJ_TAG_BITS;
Dan Streetman5538c562014-10-09 15:30:01 -0700997 link += class->size / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600998 }
999
1000 /*
1001 * We now come to the last (full or partial) object on this
1002 * page, which must point to the first object on the next
1003 * page (if present)
1004 */
1005 next_page = get_next_page(page);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001006 if (next_page) {
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001007 link->next = freeobj++ << OBJ_TAG_BITS;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001008 } else {
1009 /*
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001010 * Reset OBJ_TAG_BITS bit to last link to tell
Minchan Kimbfd093f2016-07-26 15:23:28 -07001011 * whether it's allocated object or not.
1012 */
Nick Desaulniers01a6ad92018-01-31 16:20:15 -08001013 link->next = -1UL << OBJ_TAG_BITS;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001014 }
Minchan Kimaf4ee5e2014-12-12 16:56:58 -08001015 kunmap_atomic(vaddr);
Nitin Gupta61989a82012-01-09 16:51:56 -06001016 page = next_page;
Dan Streetman5538c562014-10-09 15:30:01 -07001017 off %= PAGE_SIZE;
Nitin Gupta61989a82012-01-09 16:51:56 -06001018 }
Minchan Kimbdb0af72016-07-26 15:23:20 -07001019
Minchan Kimbfd093f2016-07-26 15:23:28 -07001020 set_freeobj(zspage, 0);
Nitin Gupta61989a82012-01-09 16:51:56 -06001021}
1022
Minchan Kim48b48002016-07-26 15:23:31 -07001023static void create_page_chain(struct size_class *class, struct zspage *zspage,
1024 struct page *pages[])
Nitin Gupta61989a82012-01-09 16:51:56 -06001025{
Minchan Kimbdb0af72016-07-26 15:23:20 -07001026 int i;
1027 struct page *page;
1028 struct page *prev_page = NULL;
Minchan Kim48b48002016-07-26 15:23:31 -07001029 int nr_pages = class->pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001030
1031 /*
1032 * Allocate individual pages and link them together as:
Minchan Kim48b48002016-07-26 15:23:31 -07001033 * 1. all pages are linked together using page->freelist
Minchan Kim37836892016-07-26 15:23:23 -07001034 * 2. each sub-page point to zspage using page->private
Nitin Gupta61989a82012-01-09 16:51:56 -06001035 *
Minchan Kim37836892016-07-26 15:23:23 -07001036 * we set PG_private to identify the first page (i.e. no other sub-page
Yisheng Xie22c5cef2017-02-24 14:59:42 -08001037 * has this flag set).
Nitin Gupta61989a82012-01-09 16:51:56 -06001038 */
Minchan Kimbdb0af72016-07-26 15:23:20 -07001039 for (i = 0; i < nr_pages; i++) {
1040 page = pages[i];
Minchan Kim37836892016-07-26 15:23:23 -07001041 set_page_private(page, (unsigned long)zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001042 page->freelist = NULL;
Minchan Kimbdb0af72016-07-26 15:23:20 -07001043 if (i == 0) {
Minchan Kim37836892016-07-26 15:23:23 -07001044 zspage->first_page = page;
Minchan Kima27545bf2012-04-25 15:23:09 +09001045 SetPagePrivate(page);
Minchan Kim48b48002016-07-26 15:23:31 -07001046 if (unlikely(class->objs_per_zspage == 1 &&
1047 class->pages_per_zspage == 1))
1048 SetPageHugeObject(page);
Minchan Kim37836892016-07-26 15:23:23 -07001049 } else {
Minchan Kim48b48002016-07-26 15:23:31 -07001050 prev_page->freelist = page;
Nitin Gupta61989a82012-01-09 16:51:56 -06001051 }
Nitin Gupta61989a82012-01-09 16:51:56 -06001052 prev_page = page;
1053 }
Minchan Kimbdb0af72016-07-26 15:23:20 -07001054}
Nitin Gupta61989a82012-01-09 16:51:56 -06001055
Minchan Kimbdb0af72016-07-26 15:23:20 -07001056/*
1057 * Allocate a zspage for the given size class
1058 */
Minchan Kim37836892016-07-26 15:23:23 -07001059static struct zspage *alloc_zspage(struct zs_pool *pool,
1060 struct size_class *class,
1061 gfp_t gfp)
Minchan Kimbdb0af72016-07-26 15:23:20 -07001062{
1063 int i;
Minchan Kimbdb0af72016-07-26 15:23:20 -07001064 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
Minchan Kim37836892016-07-26 15:23:23 -07001065 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1066
1067 if (!zspage)
1068 return NULL;
1069
1070 memset(zspage, 0, sizeof(struct zspage));
Minchan Kim48b48002016-07-26 15:23:31 -07001071 zspage->magic = ZSPAGE_MAGIC;
1072 migrate_lock_init(zspage);
Nitin Gupta61989a82012-01-09 16:51:56 -06001073
Minchan Kimbdb0af72016-07-26 15:23:20 -07001074 for (i = 0; i < class->pages_per_zspage; i++) {
1075 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -06001076
Minchan Kim37836892016-07-26 15:23:23 -07001077 page = alloc_page(gfp);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001078 if (!page) {
Minchan Kim91537fe2016-07-26 15:24:45 -07001079 while (--i >= 0) {
1080 dec_zone_page_state(pages[i], NR_ZSPAGES);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001081 __free_page(pages[i]);
Minchan Kim91537fe2016-07-26 15:24:45 -07001082 }
Minchan Kim37836892016-07-26 15:23:23 -07001083 cache_free_zspage(pool, zspage);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001084 return NULL;
1085 }
Minchan Kim91537fe2016-07-26 15:24:45 -07001086
1087 inc_zone_page_state(page, NR_ZSPAGES);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001088 pages[i] = page;
Nitin Gupta61989a82012-01-09 16:51:56 -06001089 }
1090
Minchan Kim48b48002016-07-26 15:23:31 -07001091 create_page_chain(class, zspage, pages);
Minchan Kim37836892016-07-26 15:23:23 -07001092 init_zspage(class, zspage);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001093
Minchan Kim37836892016-07-26 15:23:23 -07001094 return zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001095}
1096
Minchan Kim37836892016-07-26 15:23:23 -07001097static struct zspage *find_get_zspage(struct size_class *class)
Nitin Gupta61989a82012-01-09 16:51:56 -06001098{
1099 int i;
Minchan Kim37836892016-07-26 15:23:23 -07001100 struct zspage *zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001101
Minchan Kim48b48002016-07-26 15:23:31 -07001102 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
Minchan Kim37836892016-07-26 15:23:23 -07001103 zspage = list_first_entry_or_null(&class->fullness_list[i],
1104 struct zspage, list);
1105 if (zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -06001106 break;
1107 }
1108
Minchan Kim37836892016-07-26 15:23:23 -07001109 return zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001110}
1111
Seth Jenningsf5536462012-07-18 11:55:56 -05001112static inline int __zs_cpu_up(struct mapping_area *area)
1113{
1114 /*
1115 * Make sure we don't leak memory if a cpu UP notification
1116 * and zs_init() race and both call zs_cpu_up() on the same cpu
1117 */
1118 if (area->vm_buf)
1119 return 0;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001120 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
Seth Jenningsf5536462012-07-18 11:55:56 -05001121 if (!area->vm_buf)
1122 return -ENOMEM;
1123 return 0;
1124}
1125
1126static inline void __zs_cpu_down(struct mapping_area *area)
1127{
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001128 kfree(area->vm_buf);
Seth Jenningsf5536462012-07-18 11:55:56 -05001129 area->vm_buf = NULL;
1130}
1131
1132static void *__zs_map_object(struct mapping_area *area,
1133 struct page *pages[2], int off, int size)
1134{
Seth Jennings5f6019022012-07-02 16:15:49 -05001135 int sizes[2];
1136 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001137 char *buf = area->vm_buf;
Seth Jennings5f6019022012-07-02 16:15:49 -05001138
Seth Jenningsf5536462012-07-18 11:55:56 -05001139 /* disable page faults to match kmap_atomic() return conditions */
1140 pagefault_disable();
1141
1142 /* no read fastpath */
1143 if (area->vm_mm == ZS_MM_WO)
1144 goto out;
Seth Jennings5f6019022012-07-02 16:15:49 -05001145
1146 sizes[0] = PAGE_SIZE - off;
1147 sizes[1] = size - sizes[0];
1148
Seth Jennings5f6019022012-07-02 16:15:49 -05001149 /* copy object to per-cpu buffer */
1150 addr = kmap_atomic(pages[0]);
1151 memcpy(buf, addr + off, sizes[0]);
1152 kunmap_atomic(addr);
1153 addr = kmap_atomic(pages[1]);
1154 memcpy(buf + sizes[0], addr, sizes[1]);
1155 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001156out:
1157 return area->vm_buf;
Seth Jennings5f6019022012-07-02 16:15:49 -05001158}
1159
Seth Jenningsf5536462012-07-18 11:55:56 -05001160static void __zs_unmap_object(struct mapping_area *area,
1161 struct page *pages[2], int off, int size)
Seth Jennings5f6019022012-07-02 16:15:49 -05001162{
Seth Jennings5f6019022012-07-02 16:15:49 -05001163 int sizes[2];
1164 void *addr;
Minchan Kim2e40e162015-04-15 16:15:23 -07001165 char *buf;
Seth Jennings5f6019022012-07-02 16:15:49 -05001166
Seth Jenningsf5536462012-07-18 11:55:56 -05001167 /* no write fastpath */
1168 if (area->vm_mm == ZS_MM_RO)
1169 goto out;
Seth Jennings5f6019022012-07-02 16:15:49 -05001170
Minchan Kim7b60a682015-04-15 16:15:39 -07001171 buf = area->vm_buf;
YiPing Xua82cbf02016-03-17 14:20:39 -07001172 buf = buf + ZS_HANDLE_SIZE;
1173 size -= ZS_HANDLE_SIZE;
1174 off += ZS_HANDLE_SIZE;
Minchan Kim2e40e162015-04-15 16:15:23 -07001175
Seth Jennings5f6019022012-07-02 16:15:49 -05001176 sizes[0] = PAGE_SIZE - off;
1177 sizes[1] = size - sizes[0];
1178
1179 /* copy per-cpu buffer to object */
1180 addr = kmap_atomic(pages[0]);
1181 memcpy(addr + off, buf, sizes[0]);
1182 kunmap_atomic(addr);
1183 addr = kmap_atomic(pages[1]);
1184 memcpy(addr, buf + sizes[0], sizes[1]);
1185 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001186
1187out:
1188 /* enable page faults to match kunmap_atomic() return conditions */
1189 pagefault_enable();
Seth Jennings5f6019022012-07-02 16:15:49 -05001190}
Nitin Gupta61989a82012-01-09 16:51:56 -06001191
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01001192static int zs_cpu_prepare(unsigned int cpu)
Nitin Gupta61989a82012-01-09 16:51:56 -06001193{
Nitin Gupta61989a82012-01-09 16:51:56 -06001194 struct mapping_area *area;
1195
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01001196 area = &per_cpu(zs_map_area, cpu);
1197 return __zs_cpu_up(area);
Nitin Gupta61989a82012-01-09 16:51:56 -06001198}
1199
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01001200static int zs_cpu_dead(unsigned int cpu)
Nitin Gupta61989a82012-01-09 16:51:56 -06001201{
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01001202 struct mapping_area *area;
Nitin Gupta61989a82012-01-09 16:51:56 -06001203
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01001204 area = &per_cpu(zs_map_area, cpu);
1205 __zs_cpu_down(area);
1206 return 0;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001207}
1208
Ganesh Mahendran64d90462016-07-28 15:47:51 -07001209static bool can_merge(struct size_class *prev, int pages_per_zspage,
1210 int objs_per_zspage)
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001211{
Ganesh Mahendran64d90462016-07-28 15:47:51 -07001212 if (prev->pages_per_zspage == pages_per_zspage &&
1213 prev->objs_per_zspage == objs_per_zspage)
1214 return true;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001215
Ganesh Mahendran64d90462016-07-28 15:47:51 -07001216 return false;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001217}
1218
Minchan Kim37836892016-07-26 15:23:23 -07001219static bool zspage_full(struct size_class *class, struct zspage *zspage)
Minchan Kim312fcae2015-04-15 16:15:30 -07001220{
Minchan Kim37836892016-07-26 15:23:23 -07001221 return get_zspage_inuse(zspage) == class->objs_per_zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001222}
1223
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001224unsigned long zs_get_total_pages(struct zs_pool *pool)
1225{
1226 return atomic_long_read(&pool->pages_allocated);
1227}
1228EXPORT_SYMBOL_GPL(zs_get_total_pages);
1229
1230/**
1231 * zs_map_object - get address of allocated object from handle.
1232 * @pool: pool from which the object was allocated
1233 * @handle: handle returned from zs_malloc
Mike Rapoporte8b098f2018-04-05 16:24:57 -07001234 * @mm: maping mode to use
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001235 *
1236 * Before using an object allocated from zs_malloc, it must be mapped using
1237 * this function. When done with the object, it must be unmapped using
1238 * zs_unmap_object.
1239 *
1240 * Only one object can be mapped per cpu at a time. There is no protection
1241 * against nested mappings.
1242 *
1243 * This function returns with preemption and page faults disabled.
1244 */
1245void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1246 enum zs_mapmode mm)
1247{
Minchan Kim37836892016-07-26 15:23:23 -07001248 struct zspage *zspage;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001249 struct page *page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001250 unsigned long obj, off;
1251 unsigned int obj_idx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001252
1253 unsigned int class_idx;
1254 enum fullness_group fg;
1255 struct size_class *class;
1256 struct mapping_area *area;
1257 struct page *pages[2];
Minchan Kim2e40e162015-04-15 16:15:23 -07001258 void *ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001259
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001260 /*
1261 * Because we use per-cpu mapping areas shared among the
1262 * pools/users, we can't allow mapping in interrupt context
1263 * because it can corrupt another users mappings.
1264 */
Sergey Senozhatsky1aedcaf2017-11-15 17:34:03 -08001265 BUG_ON(in_interrupt());
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001266
Minchan Kim312fcae2015-04-15 16:15:30 -07001267 /* From now on, migration cannot move the object */
1268 pin_tag(handle);
1269
Minchan Kim2e40e162015-04-15 16:15:23 -07001270 obj = handle_to_obj(handle);
1271 obj_to_location(obj, &page, &obj_idx);
Minchan Kim37836892016-07-26 15:23:23 -07001272 zspage = get_zspage(page);
Minchan Kim48b48002016-07-26 15:23:31 -07001273
1274 /* migration cannot move any subpage in this zspage */
1275 migrate_read_lock(zspage);
1276
Minchan Kim37836892016-07-26 15:23:23 -07001277 get_zspage_mapping(zspage, &class_idx, &fg);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001278 class = pool->size_class[class_idx];
Minchan Kimbfd093f2016-07-26 15:23:28 -07001279 off = (class->size * obj_idx) & ~PAGE_MASK;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001280
1281 area = &get_cpu_var(zs_map_area);
1282 area->vm_mm = mm;
1283 if (off + class->size <= PAGE_SIZE) {
1284 /* this object is contained entirely within a page */
1285 area->vm_addr = kmap_atomic(page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001286 ret = area->vm_addr + off;
1287 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001288 }
1289
1290 /* this object spans two pages */
1291 pages[0] = page;
1292 pages[1] = get_next_page(page);
1293 BUG_ON(!pages[1]);
1294
Minchan Kim2e40e162015-04-15 16:15:23 -07001295 ret = __zs_map_object(area, pages, off, class->size);
1296out:
Minchan Kim48b48002016-07-26 15:23:31 -07001297 if (likely(!PageHugeObject(page)))
Minchan Kim7b60a682015-04-15 16:15:39 -07001298 ret += ZS_HANDLE_SIZE;
1299
1300 return ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001301}
1302EXPORT_SYMBOL_GPL(zs_map_object);
1303
1304void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1305{
Minchan Kim37836892016-07-26 15:23:23 -07001306 struct zspage *zspage;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001307 struct page *page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001308 unsigned long obj, off;
1309 unsigned int obj_idx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001310
1311 unsigned int class_idx;
1312 enum fullness_group fg;
1313 struct size_class *class;
1314 struct mapping_area *area;
1315
Minchan Kim2e40e162015-04-15 16:15:23 -07001316 obj = handle_to_obj(handle);
1317 obj_to_location(obj, &page, &obj_idx);
Minchan Kim37836892016-07-26 15:23:23 -07001318 zspage = get_zspage(page);
1319 get_zspage_mapping(zspage, &class_idx, &fg);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001320 class = pool->size_class[class_idx];
Minchan Kimbfd093f2016-07-26 15:23:28 -07001321 off = (class->size * obj_idx) & ~PAGE_MASK;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001322
1323 area = this_cpu_ptr(&zs_map_area);
1324 if (off + class->size <= PAGE_SIZE)
1325 kunmap_atomic(area->vm_addr);
1326 else {
1327 struct page *pages[2];
1328
1329 pages[0] = page;
1330 pages[1] = get_next_page(page);
1331 BUG_ON(!pages[1]);
1332
1333 __zs_unmap_object(area, pages, off, class->size);
1334 }
1335 put_cpu_var(zs_map_area);
Minchan Kim48b48002016-07-26 15:23:31 -07001336
1337 migrate_read_unlock(zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07001338 unpin_tag(handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001339}
1340EXPORT_SYMBOL_GPL(zs_unmap_object);
1341
Sergey Senozhatsky010b4952018-04-05 16:24:43 -07001342/**
1343 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1344 * zsmalloc &size_class.
1345 * @pool: zsmalloc pool to use
1346 *
1347 * The function returns the size of the first huge class - any object of equal
1348 * or bigger size will be stored in zspage consisting of a single physical
1349 * page.
1350 *
1351 * Context: Any context.
1352 *
1353 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1354 */
1355size_t zs_huge_class_size(struct zs_pool *pool)
1356{
1357 return huge_class_size;
1358}
1359EXPORT_SYMBOL_GPL(zs_huge_class_size);
1360
Minchan Kim251cbb92016-05-20 16:59:42 -07001361static unsigned long obj_malloc(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -07001362 struct zspage *zspage, unsigned long handle)
Minchan Kimc7806262015-04-15 16:15:26 -07001363{
Minchan Kimbfd093f2016-07-26 15:23:28 -07001364 int i, nr_page, offset;
Minchan Kimc7806262015-04-15 16:15:26 -07001365 unsigned long obj;
1366 struct link_free *link;
1367
1368 struct page *m_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001369 unsigned long m_offset;
Minchan Kimc7806262015-04-15 16:15:26 -07001370 void *vaddr;
1371
Minchan Kim312fcae2015-04-15 16:15:30 -07001372 handle |= OBJ_ALLOCATED_TAG;
Minchan Kim37836892016-07-26 15:23:23 -07001373 obj = get_freeobj(zspage);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001374
1375 offset = obj * class->size;
1376 nr_page = offset >> PAGE_SHIFT;
1377 m_offset = offset & ~PAGE_MASK;
1378 m_page = get_first_page(zspage);
1379
1380 for (i = 0; i < nr_page; i++)
1381 m_page = get_next_page(m_page);
Minchan Kimc7806262015-04-15 16:15:26 -07001382
1383 vaddr = kmap_atomic(m_page);
1384 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001385 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
Minchan Kim48b48002016-07-26 15:23:31 -07001386 if (likely(!PageHugeObject(m_page)))
Minchan Kim7b60a682015-04-15 16:15:39 -07001387 /* record handle in the header of allocated chunk */
1388 link->handle = handle;
1389 else
Minchan Kim37836892016-07-26 15:23:23 -07001390 /* record handle to page->index */
1391 zspage->first_page->index = handle;
1392
Minchan Kimc7806262015-04-15 16:15:26 -07001393 kunmap_atomic(vaddr);
Minchan Kim37836892016-07-26 15:23:23 -07001394 mod_zspage_inuse(zspage, 1);
Minchan Kimc7806262015-04-15 16:15:26 -07001395 zs_stat_inc(class, OBJ_USED, 1);
1396
Minchan Kimbfd093f2016-07-26 15:23:28 -07001397 obj = location_to_obj(m_page, obj);
1398
Minchan Kimc7806262015-04-15 16:15:26 -07001399 return obj;
1400}
1401
1402
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001403/**
1404 * zs_malloc - Allocate block of given size from pool.
1405 * @pool: pool to allocate from
1406 * @size: size of block to allocate
Ganesh Mahendranfd854462016-07-28 15:47:54 -07001407 * @gfp: gfp flags when allocating object
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001408 *
1409 * On success, handle to the allocated object is returned,
1410 * otherwise 0.
1411 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1412 */
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -07001413unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001414{
Minchan Kim2e40e162015-04-15 16:15:23 -07001415 unsigned long handle, obj;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001416 struct size_class *class;
Minchan Kim48b48002016-07-26 15:23:31 -07001417 enum fullness_group newfg;
Minchan Kim37836892016-07-26 15:23:23 -07001418 struct zspage *zspage;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001419
Minchan Kim7b60a682015-04-15 16:15:39 -07001420 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001421 return 0;
1422
Minchan Kim37836892016-07-26 15:23:23 -07001423 handle = cache_alloc_handle(pool, gfp);
Minchan Kim2e40e162015-04-15 16:15:23 -07001424 if (!handle)
1425 return 0;
1426
1427 /* extra space in chunk to keep the handle */
1428 size += ZS_HANDLE_SIZE;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001429 class = pool->size_class[get_size_class_index(size)];
1430
1431 spin_lock(&class->lock);
Minchan Kim37836892016-07-26 15:23:23 -07001432 zspage = find_get_zspage(class);
Minchan Kim48b48002016-07-26 15:23:31 -07001433 if (likely(zspage)) {
1434 obj = obj_malloc(class, zspage, handle);
1435 /* Now move the zspage to another fullness group, if required */
1436 fix_fullness_group(class, zspage);
1437 record_obj(handle, obj);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001438 spin_unlock(&class->lock);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001439
Minchan Kim48b48002016-07-26 15:23:31 -07001440 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001441 }
1442
Minchan Kim48b48002016-07-26 15:23:31 -07001443 spin_unlock(&class->lock);
1444
1445 zspage = alloc_zspage(pool, class, gfp);
1446 if (!zspage) {
1447 cache_free_handle(pool, handle);
1448 return 0;
1449 }
1450
1451 spin_lock(&class->lock);
Minchan Kim37836892016-07-26 15:23:23 -07001452 obj = obj_malloc(class, zspage, handle);
Minchan Kim48b48002016-07-26 15:23:31 -07001453 newfg = get_fullness_group(class, zspage);
1454 insert_zspage(class, zspage, newfg);
1455 set_zspage_mapping(zspage, class->index, newfg);
Minchan Kim2e40e162015-04-15 16:15:23 -07001456 record_obj(handle, obj);
Minchan Kim48b48002016-07-26 15:23:31 -07001457 atomic_long_add(class->pages_per_zspage,
1458 &pool->pages_allocated);
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -07001459 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001460
1461 /* We completely set up zspage so mark them as movable */
1462 SetZsPageMovable(pool, zspage);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001463 spin_unlock(&class->lock);
1464
Minchan Kim2e40e162015-04-15 16:15:23 -07001465 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001466}
1467EXPORT_SYMBOL_GPL(zs_malloc);
1468
Minchan Kim1ee47162016-05-20 16:59:45 -07001469static void obj_free(struct size_class *class, unsigned long obj)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001470{
1471 struct link_free *link;
Minchan Kim37836892016-07-26 15:23:23 -07001472 struct zspage *zspage;
1473 struct page *f_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001474 unsigned long f_offset;
1475 unsigned int f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001476 void *vaddr;
1477
Minchan Kim312fcae2015-04-15 16:15:30 -07001478 obj &= ~OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001479 obj_to_location(obj, &f_page, &f_objidx);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001480 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
Minchan Kim37836892016-07-26 15:23:23 -07001481 zspage = get_zspage(f_page);
Minchan Kimc7806262015-04-15 16:15:26 -07001482
Minchan Kimc7806262015-04-15 16:15:26 -07001483 vaddr = kmap_atomic(f_page);
1484
1485 /* Insert this object in containing zspage's freelist */
1486 link = (struct link_free *)(vaddr + f_offset);
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001487 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
Minchan Kimc7806262015-04-15 16:15:26 -07001488 kunmap_atomic(vaddr);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001489 set_freeobj(zspage, f_objidx);
Minchan Kim37836892016-07-26 15:23:23 -07001490 mod_zspage_inuse(zspage, -1);
Minchan Kimc7806262015-04-15 16:15:26 -07001491 zs_stat_dec(class, OBJ_USED, 1);
1492}
1493
1494void zs_free(struct zs_pool *pool, unsigned long handle)
1495{
Minchan Kim37836892016-07-26 15:23:23 -07001496 struct zspage *zspage;
1497 struct page *f_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001498 unsigned long obj;
1499 unsigned int f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001500 int class_idx;
1501 struct size_class *class;
1502 enum fullness_group fullness;
Minchan Kim48b48002016-07-26 15:23:31 -07001503 bool isolated;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001504
Minchan Kim2e40e162015-04-15 16:15:23 -07001505 if (unlikely(!handle))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001506 return;
1507
Minchan Kim312fcae2015-04-15 16:15:30 -07001508 pin_tag(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001509 obj = handle_to_obj(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001510 obj_to_location(obj, &f_page, &f_objidx);
Minchan Kim37836892016-07-26 15:23:23 -07001511 zspage = get_zspage(f_page);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001512
Minchan Kim48b48002016-07-26 15:23:31 -07001513 migrate_read_lock(zspage);
1514
Minchan Kim37836892016-07-26 15:23:23 -07001515 get_zspage_mapping(zspage, &class_idx, &fullness);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001516 class = pool->size_class[class_idx];
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001517
1518 spin_lock(&class->lock);
Minchan Kim1ee47162016-05-20 16:59:45 -07001519 obj_free(class, obj);
Minchan Kim37836892016-07-26 15:23:23 -07001520 fullness = fix_fullness_group(class, zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001521 if (fullness != ZS_EMPTY) {
1522 migrate_read_unlock(zspage);
1523 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001524 }
Minchan Kim48b48002016-07-26 15:23:31 -07001525
1526 isolated = is_zspage_isolated(zspage);
1527 migrate_read_unlock(zspage);
1528 /* If zspage is isolated, zs_page_putback will free the zspage */
1529 if (likely(!isolated))
1530 free_zspage(pool, class, zspage);
1531out:
1532
Minchan Kim312fcae2015-04-15 16:15:30 -07001533 spin_unlock(&class->lock);
1534 unpin_tag(handle);
Minchan Kim37836892016-07-26 15:23:23 -07001535 cache_free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001536}
1537EXPORT_SYMBOL_GPL(zs_free);
1538
Minchan Kim251cbb92016-05-20 16:59:42 -07001539static void zs_object_copy(struct size_class *class, unsigned long dst,
1540 unsigned long src)
Minchan Kim312fcae2015-04-15 16:15:30 -07001541{
1542 struct page *s_page, *d_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001543 unsigned int s_objidx, d_objidx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001544 unsigned long s_off, d_off;
1545 void *s_addr, *d_addr;
1546 int s_size, d_size, size;
1547 int written = 0;
1548
1549 s_size = d_size = class->size;
1550
1551 obj_to_location(src, &s_page, &s_objidx);
1552 obj_to_location(dst, &d_page, &d_objidx);
1553
Minchan Kimbfd093f2016-07-26 15:23:28 -07001554 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1555 d_off = (class->size * d_objidx) & ~PAGE_MASK;
Minchan Kim312fcae2015-04-15 16:15:30 -07001556
1557 if (s_off + class->size > PAGE_SIZE)
1558 s_size = PAGE_SIZE - s_off;
1559
1560 if (d_off + class->size > PAGE_SIZE)
1561 d_size = PAGE_SIZE - d_off;
1562
1563 s_addr = kmap_atomic(s_page);
1564 d_addr = kmap_atomic(d_page);
1565
1566 while (1) {
1567 size = min(s_size, d_size);
1568 memcpy(d_addr + d_off, s_addr + s_off, size);
1569 written += size;
1570
1571 if (written == class->size)
1572 break;
1573
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001574 s_off += size;
1575 s_size -= size;
1576 d_off += size;
1577 d_size -= size;
1578
1579 if (s_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001580 kunmap_atomic(d_addr);
1581 kunmap_atomic(s_addr);
1582 s_page = get_next_page(s_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001583 s_addr = kmap_atomic(s_page);
1584 d_addr = kmap_atomic(d_page);
1585 s_size = class->size - written;
1586 s_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001587 }
1588
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001589 if (d_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001590 kunmap_atomic(d_addr);
1591 d_page = get_next_page(d_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001592 d_addr = kmap_atomic(d_page);
1593 d_size = class->size - written;
1594 d_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001595 }
1596 }
1597
1598 kunmap_atomic(d_addr);
1599 kunmap_atomic(s_addr);
1600}
1601
1602/*
1603 * Find alloced object in zspage from index object and
1604 * return handle.
1605 */
Minchan Kim251cbb92016-05-20 16:59:42 -07001606static unsigned long find_alloced_obj(struct size_class *class,
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001607 struct page *page, int *obj_idx)
Minchan Kim312fcae2015-04-15 16:15:30 -07001608{
1609 unsigned long head;
1610 int offset = 0;
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001611 int index = *obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001612 unsigned long handle = 0;
1613 void *addr = kmap_atomic(page);
1614
Minchan Kim37836892016-07-26 15:23:23 -07001615 offset = get_first_obj_offset(page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001616 offset += class->size * index;
1617
1618 while (offset < PAGE_SIZE) {
Minchan Kim48b48002016-07-26 15:23:31 -07001619 head = obj_to_head(page, addr + offset);
Minchan Kim312fcae2015-04-15 16:15:30 -07001620 if (head & OBJ_ALLOCATED_TAG) {
1621 handle = head & ~OBJ_ALLOCATED_TAG;
1622 if (trypin_tag(handle))
1623 break;
1624 handle = 0;
1625 }
1626
1627 offset += class->size;
1628 index++;
1629 }
1630
1631 kunmap_atomic(addr);
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001632
1633 *obj_idx = index;
1634
Minchan Kim312fcae2015-04-15 16:15:30 -07001635 return handle;
1636}
1637
1638struct zs_compact_control {
Minchan Kim37836892016-07-26 15:23:23 -07001639 /* Source spage for migration which could be a subpage of zspage */
Minchan Kim312fcae2015-04-15 16:15:30 -07001640 struct page *s_page;
1641 /* Destination page for migration which should be a first page
1642 * of zspage. */
1643 struct page *d_page;
1644 /* Starting object index within @s_page which used for live object
1645 * in the subpage. */
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001646 int obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001647};
1648
1649static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1650 struct zs_compact_control *cc)
1651{
1652 unsigned long used_obj, free_obj;
1653 unsigned long handle;
1654 struct page *s_page = cc->s_page;
1655 struct page *d_page = cc->d_page;
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001656 int obj_idx = cc->obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001657 int ret = 0;
1658
1659 while (1) {
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001660 handle = find_alloced_obj(class, s_page, &obj_idx);
Minchan Kim312fcae2015-04-15 16:15:30 -07001661 if (!handle) {
1662 s_page = get_next_page(s_page);
1663 if (!s_page)
1664 break;
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001665 obj_idx = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001666 continue;
1667 }
1668
1669 /* Stop if there is no more space */
Minchan Kim37836892016-07-26 15:23:23 -07001670 if (zspage_full(class, get_zspage(d_page))) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001671 unpin_tag(handle);
1672 ret = -ENOMEM;
1673 break;
1674 }
1675
1676 used_obj = handle_to_obj(handle);
Minchan Kim37836892016-07-26 15:23:23 -07001677 free_obj = obj_malloc(class, get_zspage(d_page), handle);
Minchan Kim251cbb92016-05-20 16:59:42 -07001678 zs_object_copy(class, free_obj, used_obj);
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001679 obj_idx++;
Junil Leec102f07c2016-01-20 14:58:18 -08001680 /*
1681 * record_obj updates handle's value to free_obj and it will
1682 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1683 * breaks synchronization using pin_tag(e,g, zs_free) so
1684 * let's keep the lock bit.
1685 */
1686 free_obj |= BIT(HANDLE_PIN_BIT);
Minchan Kim312fcae2015-04-15 16:15:30 -07001687 record_obj(handle, free_obj);
1688 unpin_tag(handle);
Minchan Kim1ee47162016-05-20 16:59:45 -07001689 obj_free(class, used_obj);
Minchan Kim312fcae2015-04-15 16:15:30 -07001690 }
1691
1692 /* Remember last position in this iteration */
1693 cc->s_page = s_page;
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001694 cc->obj_idx = obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001695
1696 return ret;
1697}
1698
Minchan Kim37836892016-07-26 15:23:23 -07001699static struct zspage *isolate_zspage(struct size_class *class, bool source)
Minchan Kim312fcae2015-04-15 16:15:30 -07001700{
1701 int i;
Minchan Kim37836892016-07-26 15:23:23 -07001702 struct zspage *zspage;
1703 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
Minchan Kim312fcae2015-04-15 16:15:30 -07001704
Minchan Kim37836892016-07-26 15:23:23 -07001705 if (!source) {
1706 fg[0] = ZS_ALMOST_FULL;
1707 fg[1] = ZS_ALMOST_EMPTY;
1708 }
1709
1710 for (i = 0; i < 2; i++) {
1711 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1712 struct zspage, list);
1713 if (zspage) {
Minchan Kim48b48002016-07-26 15:23:31 -07001714 VM_BUG_ON(is_zspage_isolated(zspage));
Minchan Kim37836892016-07-26 15:23:23 -07001715 remove_zspage(class, zspage, fg[i]);
1716 return zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001717 }
1718 }
1719
Minchan Kim37836892016-07-26 15:23:23 -07001720 return zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001721}
1722
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001723/*
Minchan Kim37836892016-07-26 15:23:23 -07001724 * putback_zspage - add @zspage into right class's fullness list
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001725 * @class: destination class
Minchan Kim37836892016-07-26 15:23:23 -07001726 * @zspage: target page
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001727 *
Minchan Kim37836892016-07-26 15:23:23 -07001728 * Return @zspage's fullness_group
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001729 */
Minchan Kim4aa409c2016-07-26 15:23:26 -07001730static enum fullness_group putback_zspage(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -07001731 struct zspage *zspage)
Minchan Kim312fcae2015-04-15 16:15:30 -07001732{
Minchan Kim312fcae2015-04-15 16:15:30 -07001733 enum fullness_group fullness;
1734
Minchan Kim48b48002016-07-26 15:23:31 -07001735 VM_BUG_ON(is_zspage_isolated(zspage));
1736
Minchan Kim37836892016-07-26 15:23:23 -07001737 fullness = get_fullness_group(class, zspage);
1738 insert_zspage(class, zspage, fullness);
1739 set_zspage_mapping(zspage, class->index, fullness);
Minchan Kim839373e2015-04-15 16:16:18 -07001740
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001741 return fullness;
Minchan Kim312fcae2015-04-15 16:15:30 -07001742}
1743
Minchan Kim48b48002016-07-26 15:23:31 -07001744#ifdef CONFIG_COMPACTION
Colin Ian King4d0a5402018-08-17 15:46:50 -07001745/*
1746 * To prevent zspage destroy during migration, zspage freeing should
1747 * hold locks of all pages in the zspage.
1748 */
1749static void lock_zspage(struct zspage *zspage)
1750{
1751 struct page *page = get_first_page(zspage);
1752
1753 do {
1754 lock_page(page);
1755 } while ((page = get_next_page(page)) != NULL);
1756}
1757
David Howells8e9231f2019-03-25 16:38:23 +00001758static int zs_init_fs_context(struct fs_context *fc)
Minchan Kim48b48002016-07-26 15:23:31 -07001759{
David Howells8e9231f2019-03-25 16:38:23 +00001760 return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
Minchan Kim48b48002016-07-26 15:23:31 -07001761}
1762
1763static struct file_system_type zsmalloc_fs = {
1764 .name = "zsmalloc",
David Howells8e9231f2019-03-25 16:38:23 +00001765 .init_fs_context = zs_init_fs_context,
Minchan Kim48b48002016-07-26 15:23:31 -07001766 .kill_sb = kill_anon_super,
1767};
1768
1769static int zsmalloc_mount(void)
1770{
1771 int ret = 0;
1772
1773 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1774 if (IS_ERR(zsmalloc_mnt))
1775 ret = PTR_ERR(zsmalloc_mnt);
1776
1777 return ret;
1778}
1779
1780static void zsmalloc_unmount(void)
1781{
1782 kern_unmount(zsmalloc_mnt);
1783}
1784
1785static void migrate_lock_init(struct zspage *zspage)
1786{
1787 rwlock_init(&zspage->lock);
1788}
1789
Jules Irengecfc451cf2020-04-06 20:08:21 -07001790static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
Minchan Kim48b48002016-07-26 15:23:31 -07001791{
1792 read_lock(&zspage->lock);
1793}
1794
Jules Irenge8a374cc2020-04-06 20:08:24 -07001795static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
Minchan Kim48b48002016-07-26 15:23:31 -07001796{
1797 read_unlock(&zspage->lock);
1798}
1799
1800static void migrate_write_lock(struct zspage *zspage)
1801{
1802 write_lock(&zspage->lock);
1803}
1804
1805static void migrate_write_unlock(struct zspage *zspage)
1806{
1807 write_unlock(&zspage->lock);
1808}
1809
1810/* Number of isolated subpage for *page migration* in this zspage */
1811static void inc_zspage_isolation(struct zspage *zspage)
1812{
1813 zspage->isolated++;
1814}
1815
1816static void dec_zspage_isolation(struct zspage *zspage)
1817{
1818 zspage->isolated--;
1819}
1820
Henry Burns1a87aa02019-08-24 17:55:03 -07001821static void putback_zspage_deferred(struct zs_pool *pool,
1822 struct size_class *class,
1823 struct zspage *zspage)
1824{
1825 enum fullness_group fg;
1826
1827 fg = putback_zspage(class, zspage);
1828 if (fg == ZS_EMPTY)
1829 schedule_work(&pool->free_work);
1830
1831}
1832
Henry Burns701d6782019-08-24 17:55:06 -07001833static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1834{
1835 VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1836 atomic_long_dec(&pool->isolated_pages);
1837 /*
1838 * There's no possibility of racing, since wait_for_isolated_drain()
1839 * checks the isolated count under &class->lock after enqueuing
1840 * on migration_wait.
1841 */
1842 if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1843 wake_up_all(&pool->migration_wait);
1844}
1845
Minchan Kim48b48002016-07-26 15:23:31 -07001846static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1847 struct page *newpage, struct page *oldpage)
1848{
1849 struct page *page;
1850 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1851 int idx = 0;
1852
1853 page = get_first_page(zspage);
1854 do {
1855 if (page == oldpage)
1856 pages[idx] = newpage;
1857 else
1858 pages[idx] = page;
1859 idx++;
1860 } while ((page = get_next_page(page)) != NULL);
1861
1862 create_page_chain(class, zspage, pages);
1863 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1864 if (unlikely(PageHugeObject(oldpage)))
1865 newpage->index = oldpage->index;
1866 __SetPageMovable(newpage, page_mapping(oldpage));
1867}
1868
Colin Ian King4d0a5402018-08-17 15:46:50 -07001869static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
Minchan Kim48b48002016-07-26 15:23:31 -07001870{
1871 struct zs_pool *pool;
1872 struct size_class *class;
1873 int class_idx;
1874 enum fullness_group fullness;
1875 struct zspage *zspage;
1876 struct address_space *mapping;
1877
1878 /*
1879 * Page is locked so zspage couldn't be destroyed. For detail, look at
1880 * lock_zspage in free_zspage.
1881 */
1882 VM_BUG_ON_PAGE(!PageMovable(page), page);
1883 VM_BUG_ON_PAGE(PageIsolated(page), page);
1884
1885 zspage = get_zspage(page);
1886
1887 /*
1888 * Without class lock, fullness could be stale while class_idx is okay
1889 * because class_idx is constant unless page is freed so we should get
1890 * fullness again under class lock.
1891 */
1892 get_zspage_mapping(zspage, &class_idx, &fullness);
1893 mapping = page_mapping(page);
1894 pool = mapping->private_data;
1895 class = pool->size_class[class_idx];
1896
1897 spin_lock(&class->lock);
1898 if (get_zspage_inuse(zspage) == 0) {
1899 spin_unlock(&class->lock);
1900 return false;
1901 }
1902
1903 /* zspage is isolated for object migration */
1904 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1905 spin_unlock(&class->lock);
1906 return false;
1907 }
1908
1909 /*
1910 * If this is first time isolation for the zspage, isolate zspage from
1911 * size_class to prevent further object allocation from the zspage.
1912 */
1913 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1914 get_zspage_mapping(zspage, &class_idx, &fullness);
Henry Burns701d6782019-08-24 17:55:06 -07001915 atomic_long_inc(&pool->isolated_pages);
Minchan Kim48b48002016-07-26 15:23:31 -07001916 remove_zspage(class, zspage, fullness);
1917 }
1918
1919 inc_zspage_isolation(zspage);
1920 spin_unlock(&class->lock);
1921
1922 return true;
1923}
1924
Colin Ian King4d0a5402018-08-17 15:46:50 -07001925static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
Minchan Kim48b48002016-07-26 15:23:31 -07001926 struct page *page, enum migrate_mode mode)
1927{
1928 struct zs_pool *pool;
1929 struct size_class *class;
1930 int class_idx;
1931 enum fullness_group fullness;
1932 struct zspage *zspage;
1933 struct page *dummy;
1934 void *s_addr, *d_addr, *addr;
1935 int offset, pos;
1936 unsigned long handle, head;
1937 unsigned long old_obj, new_obj;
1938 unsigned int obj_idx;
1939 int ret = -EAGAIN;
1940
Jérôme Glisse2916ecc2017-09-08 16:12:06 -07001941 /*
1942 * We cannot support the _NO_COPY case here, because copy needs to
1943 * happen under the zs lock, which does not work with
1944 * MIGRATE_SYNC_NO_COPY workflow.
1945 */
1946 if (mode == MIGRATE_SYNC_NO_COPY)
1947 return -EINVAL;
1948
Minchan Kim48b48002016-07-26 15:23:31 -07001949 VM_BUG_ON_PAGE(!PageMovable(page), page);
1950 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1951
1952 zspage = get_zspage(page);
1953
1954 /* Concurrent compactor cannot migrate any subpage in zspage */
1955 migrate_write_lock(zspage);
1956 get_zspage_mapping(zspage, &class_idx, &fullness);
1957 pool = mapping->private_data;
1958 class = pool->size_class[class_idx];
1959 offset = get_first_obj_offset(page);
1960
1961 spin_lock(&class->lock);
1962 if (!get_zspage_inuse(zspage)) {
Hui Zhu77ff4652017-09-06 16:21:08 -07001963 /*
1964 * Set "offset" to end of the page so that every loops
1965 * skips unnecessary object scanning.
1966 */
1967 offset = PAGE_SIZE;
Minchan Kim48b48002016-07-26 15:23:31 -07001968 }
1969
1970 pos = offset;
1971 s_addr = kmap_atomic(page);
1972 while (pos < PAGE_SIZE) {
1973 head = obj_to_head(page, s_addr + pos);
1974 if (head & OBJ_ALLOCATED_TAG) {
1975 handle = head & ~OBJ_ALLOCATED_TAG;
1976 if (!trypin_tag(handle))
1977 goto unpin_objects;
1978 }
1979 pos += class->size;
1980 }
1981
1982 /*
1983 * Here, any user cannot access all objects in the zspage so let's move.
1984 */
1985 d_addr = kmap_atomic(newpage);
1986 memcpy(d_addr, s_addr, PAGE_SIZE);
1987 kunmap_atomic(d_addr);
1988
1989 for (addr = s_addr + offset; addr < s_addr + pos;
1990 addr += class->size) {
1991 head = obj_to_head(page, addr);
1992 if (head & OBJ_ALLOCATED_TAG) {
1993 handle = head & ~OBJ_ALLOCATED_TAG;
1994 if (!testpin_tag(handle))
1995 BUG();
1996
1997 old_obj = handle_to_obj(handle);
1998 obj_to_location(old_obj, &dummy, &obj_idx);
1999 new_obj = (unsigned long)location_to_obj(newpage,
2000 obj_idx);
2001 new_obj |= BIT(HANDLE_PIN_BIT);
2002 record_obj(handle, new_obj);
2003 }
2004 }
2005
2006 replace_sub_page(class, zspage, newpage, page);
2007 get_page(newpage);
2008
2009 dec_zspage_isolation(zspage);
2010
2011 /*
2012 * Page migration is done so let's putback isolated zspage to
2013 * the list if @page is final isolated subpage in the zspage.
2014 */
Henry Burns701d6782019-08-24 17:55:06 -07002015 if (!is_zspage_isolated(zspage)) {
2016 /*
2017 * We cannot race with zs_destroy_pool() here because we wait
2018 * for isolation to hit zero before we start destroying.
2019 * Also, we ensure that everyone can see pool->destroying before
2020 * we start waiting.
2021 */
Henry Burns1a87aa02019-08-24 17:55:03 -07002022 putback_zspage_deferred(pool, class, zspage);
Henry Burns701d6782019-08-24 17:55:06 -07002023 zs_pool_dec_isolated(pool);
2024 }
Minchan Kim48b48002016-07-26 15:23:31 -07002025
Chanho Minac8f05d2020-01-04 12:59:36 -08002026 if (page_zone(newpage) != page_zone(page)) {
2027 dec_zone_page_state(page, NR_ZSPAGES);
2028 inc_zone_page_state(newpage, NR_ZSPAGES);
2029 }
2030
Minchan Kim48b48002016-07-26 15:23:31 -07002031 reset_page(page);
2032 put_page(page);
2033 page = newpage;
2034
Minchan Kimdd4123f2016-07-26 15:26:50 -07002035 ret = MIGRATEPAGE_SUCCESS;
Minchan Kim48b48002016-07-26 15:23:31 -07002036unpin_objects:
2037 for (addr = s_addr + offset; addr < s_addr + pos;
2038 addr += class->size) {
2039 head = obj_to_head(page, addr);
2040 if (head & OBJ_ALLOCATED_TAG) {
2041 handle = head & ~OBJ_ALLOCATED_TAG;
2042 if (!testpin_tag(handle))
2043 BUG();
2044 unpin_tag(handle);
2045 }
2046 }
2047 kunmap_atomic(s_addr);
Minchan Kim48b48002016-07-26 15:23:31 -07002048 spin_unlock(&class->lock);
2049 migrate_write_unlock(zspage);
2050
2051 return ret;
2052}
2053
Colin Ian King4d0a5402018-08-17 15:46:50 -07002054static void zs_page_putback(struct page *page)
Minchan Kim48b48002016-07-26 15:23:31 -07002055{
2056 struct zs_pool *pool;
2057 struct size_class *class;
2058 int class_idx;
2059 enum fullness_group fg;
2060 struct address_space *mapping;
2061 struct zspage *zspage;
2062
2063 VM_BUG_ON_PAGE(!PageMovable(page), page);
2064 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2065
2066 zspage = get_zspage(page);
2067 get_zspage_mapping(zspage, &class_idx, &fg);
2068 mapping = page_mapping(page);
2069 pool = mapping->private_data;
2070 class = pool->size_class[class_idx];
2071
2072 spin_lock(&class->lock);
2073 dec_zspage_isolation(zspage);
2074 if (!is_zspage_isolated(zspage)) {
Minchan Kim48b48002016-07-26 15:23:31 -07002075 /*
2076 * Due to page_lock, we cannot free zspage immediately
2077 * so let's defer.
2078 */
Henry Burns1a87aa02019-08-24 17:55:03 -07002079 putback_zspage_deferred(pool, class, zspage);
Henry Burns701d6782019-08-24 17:55:06 -07002080 zs_pool_dec_isolated(pool);
Minchan Kim48b48002016-07-26 15:23:31 -07002081 }
2082 spin_unlock(&class->lock);
2083}
2084
Colin Ian King4d0a5402018-08-17 15:46:50 -07002085static const struct address_space_operations zsmalloc_aops = {
Minchan Kim48b48002016-07-26 15:23:31 -07002086 .isolate_page = zs_page_isolate,
2087 .migratepage = zs_page_migrate,
2088 .putback_page = zs_page_putback,
2089};
2090
2091static int zs_register_migration(struct zs_pool *pool)
2092{
2093 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2094 if (IS_ERR(pool->inode)) {
2095 pool->inode = NULL;
2096 return 1;
2097 }
2098
2099 pool->inode->i_mapping->private_data = pool;
2100 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2101 return 0;
2102}
2103
Henry Burns701d6782019-08-24 17:55:06 -07002104static bool pool_isolated_are_drained(struct zs_pool *pool)
2105{
2106 return atomic_long_read(&pool->isolated_pages) == 0;
2107}
2108
2109/* Function for resolving migration */
2110static void wait_for_isolated_drain(struct zs_pool *pool)
2111{
2112
2113 /*
2114 * We're in the process of destroying the pool, so there are no
2115 * active allocations. zs_page_isolate() fails for completely free
2116 * zspages, so we need only wait for the zs_pool's isolated
2117 * count to hit zero.
2118 */
2119 wait_event(pool->migration_wait,
2120 pool_isolated_are_drained(pool));
2121}
2122
Minchan Kim48b48002016-07-26 15:23:31 -07002123static void zs_unregister_migration(struct zs_pool *pool)
2124{
Henry Burns701d6782019-08-24 17:55:06 -07002125 pool->destroying = true;
2126 /*
2127 * We need a memory barrier here to ensure global visibility of
2128 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2129 * case we don't care, or it will be > 0 and pool->destroying will
2130 * ensure that we wake up once isolation hits 0.
2131 */
2132 smp_mb();
2133 wait_for_isolated_drain(pool); /* This can block */
Minchan Kim48b48002016-07-26 15:23:31 -07002134 flush_work(&pool->free_work);
Markus Elfringc3491ec2016-07-28 15:48:59 -07002135 iput(pool->inode);
Minchan Kim48b48002016-07-26 15:23:31 -07002136}
2137
2138/*
2139 * Caller should hold page_lock of all pages in the zspage
2140 * In here, we cannot use zspage meta data.
2141 */
2142static void async_free_zspage(struct work_struct *work)
2143{
2144 int i;
2145 struct size_class *class;
2146 unsigned int class_idx;
2147 enum fullness_group fullness;
2148 struct zspage *zspage, *tmp;
2149 LIST_HEAD(free_pages);
2150 struct zs_pool *pool = container_of(work, struct zs_pool,
2151 free_work);
2152
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -07002153 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
Minchan Kim48b48002016-07-26 15:23:31 -07002154 class = pool->size_class[i];
2155 if (class->index != i)
2156 continue;
2157
2158 spin_lock(&class->lock);
2159 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2160 spin_unlock(&class->lock);
2161 }
2162
2163
2164 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2165 list_del(&zspage->list);
2166 lock_zspage(zspage);
2167
2168 get_zspage_mapping(zspage, &class_idx, &fullness);
2169 VM_BUG_ON(fullness != ZS_EMPTY);
2170 class = pool->size_class[class_idx];
2171 spin_lock(&class->lock);
2172 __free_zspage(pool, pool->size_class[class_idx], zspage);
2173 spin_unlock(&class->lock);
2174 }
2175};
2176
2177static void kick_deferred_free(struct zs_pool *pool)
2178{
2179 schedule_work(&pool->free_work);
2180}
2181
2182static void init_deferred_free(struct zs_pool *pool)
2183{
2184 INIT_WORK(&pool->free_work, async_free_zspage);
2185}
2186
2187static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2188{
2189 struct page *page = get_first_page(zspage);
2190
2191 do {
2192 WARN_ON(!trylock_page(page));
2193 __SetPageMovable(page, pool->inode->i_mapping);
2194 unlock_page(page);
2195 } while ((page = get_next_page(page)) != NULL);
2196}
2197#endif
2198
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002199/*
2200 *
2201 * Based on the number of unused allocated objects calculate
2202 * and return the number of pages that we can free.
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002203 */
2204static unsigned long zs_can_compact(struct size_class *class)
2205{
2206 unsigned long obj_wasted;
Sergey Senozhatsky44f43e92016-05-09 16:28:49 -07002207 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2208 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002209
Sergey Senozhatsky44f43e92016-05-09 16:28:49 -07002210 if (obj_allocated <= obj_used)
2211 return 0;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002212
Sergey Senozhatsky44f43e92016-05-09 16:28:49 -07002213 obj_wasted = obj_allocated - obj_used;
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -07002214 obj_wasted /= class->objs_per_zspage;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002215
Minchan Kim6cbf16b2015-09-08 15:04:49 -07002216 return obj_wasted * class->pages_per_zspage;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002217}
2218
Rokudo Yan02f768e2021-02-25 17:18:31 -08002219static unsigned long __zs_compact(struct zs_pool *pool,
2220 struct size_class *class)
Minchan Kim312fcae2015-04-15 16:15:30 -07002221{
Minchan Kim312fcae2015-04-15 16:15:30 -07002222 struct zs_compact_control cc;
Minchan Kim37836892016-07-26 15:23:23 -07002223 struct zspage *src_zspage;
2224 struct zspage *dst_zspage = NULL;
Rokudo Yan02f768e2021-02-25 17:18:31 -08002225 unsigned long pages_freed = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07002226
Minchan Kim312fcae2015-04-15 16:15:30 -07002227 spin_lock(&class->lock);
Minchan Kim37836892016-07-26 15:23:23 -07002228 while ((src_zspage = isolate_zspage(class, true))) {
Minchan Kim312fcae2015-04-15 16:15:30 -07002229
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002230 if (!zs_can_compact(class))
2231 break;
2232
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07002233 cc.obj_idx = 0;
Minchan Kim48b48002016-07-26 15:23:31 -07002234 cc.s_page = get_first_page(src_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002235
Minchan Kim37836892016-07-26 15:23:23 -07002236 while ((dst_zspage = isolate_zspage(class, false))) {
Minchan Kim48b48002016-07-26 15:23:31 -07002237 cc.d_page = get_first_page(dst_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002238 /*
Sergey Senozhatsky0dc63d482015-09-08 15:04:33 -07002239 * If there is no more space in dst_page, resched
2240 * and see if anyone had allocated another zspage.
Minchan Kim312fcae2015-04-15 16:15:30 -07002241 */
2242 if (!migrate_zspage(pool, class, &cc))
2243 break;
2244
Minchan Kim4aa409c2016-07-26 15:23:26 -07002245 putback_zspage(class, dst_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002246 }
2247
2248 /* Stop if we couldn't find slot */
Minchan Kim37836892016-07-26 15:23:23 -07002249 if (dst_zspage == NULL)
Minchan Kim312fcae2015-04-15 16:15:30 -07002250 break;
2251
Minchan Kim4aa409c2016-07-26 15:23:26 -07002252 putback_zspage(class, dst_zspage);
2253 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
Minchan Kim48b48002016-07-26 15:23:31 -07002254 free_zspage(pool, class, src_zspage);
Rokudo Yan02f768e2021-02-25 17:18:31 -08002255 pages_freed += class->pages_per_zspage;
Minchan Kim4aa409c2016-07-26 15:23:26 -07002256 }
Minchan Kim312fcae2015-04-15 16:15:30 -07002257 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07002258 cond_resched();
2259 spin_lock(&class->lock);
2260 }
2261
Minchan Kim37836892016-07-26 15:23:23 -07002262 if (src_zspage)
Minchan Kim4aa409c2016-07-26 15:23:26 -07002263 putback_zspage(class, src_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002264
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07002265 spin_unlock(&class->lock);
Rokudo Yan02f768e2021-02-25 17:18:31 -08002266
2267 return pages_freed;
Minchan Kim312fcae2015-04-15 16:15:30 -07002268}
2269
2270unsigned long zs_compact(struct zs_pool *pool)
2271{
2272 int i;
Minchan Kim312fcae2015-04-15 16:15:30 -07002273 struct size_class *class;
Rokudo Yan02f768e2021-02-25 17:18:31 -08002274 unsigned long pages_freed = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07002275
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -07002276 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
Minchan Kim312fcae2015-04-15 16:15:30 -07002277 class = pool->size_class[i];
2278 if (!class)
2279 continue;
2280 if (class->index != i)
2281 continue;
Rokudo Yan02f768e2021-02-25 17:18:31 -08002282 pages_freed += __zs_compact(pool, class);
Minchan Kim312fcae2015-04-15 16:15:30 -07002283 }
Rokudo Yan02f768e2021-02-25 17:18:31 -08002284 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
Minchan Kim312fcae2015-04-15 16:15:30 -07002285
Rokudo Yan02f768e2021-02-25 17:18:31 -08002286 return pages_freed;
Minchan Kim312fcae2015-04-15 16:15:30 -07002287}
2288EXPORT_SYMBOL_GPL(zs_compact);
2289
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07002290void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2291{
2292 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2293}
2294EXPORT_SYMBOL_GPL(zs_pool_stats);
2295
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002296static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2297 struct shrink_control *sc)
2298{
2299 unsigned long pages_freed;
2300 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2301 shrinker);
2302
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002303 /*
2304 * Compact classes and calculate compaction delta.
2305 * Can run concurrently with a manually triggered
2306 * (by user) compaction.
2307 */
Rokudo Yan02f768e2021-02-25 17:18:31 -08002308 pages_freed = zs_compact(pool);
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002309
2310 return pages_freed ? pages_freed : SHRINK_STOP;
2311}
2312
2313static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2314 struct shrink_control *sc)
2315{
2316 int i;
2317 struct size_class *class;
2318 unsigned long pages_to_free = 0;
2319 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2320 shrinker);
2321
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -07002322 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002323 class = pool->size_class[i];
2324 if (!class)
2325 continue;
2326 if (class->index != i)
2327 continue;
2328
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002329 pages_to_free += zs_can_compact(class);
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002330 }
2331
2332 return pages_to_free;
2333}
2334
2335static void zs_unregister_shrinker(struct zs_pool *pool)
2336{
Aliaksei Karaliou93144ca2018-01-31 16:18:40 -08002337 unregister_shrinker(&pool->shrinker);
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002338}
2339
2340static int zs_register_shrinker(struct zs_pool *pool)
2341{
2342 pool->shrinker.scan_objects = zs_shrinker_scan;
2343 pool->shrinker.count_objects = zs_shrinker_count;
2344 pool->shrinker.batch = 0;
2345 pool->shrinker.seeks = DEFAULT_SEEKS;
2346
2347 return register_shrinker(&pool->shrinker);
2348}
2349
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08002350/**
2351 * zs_create_pool - Creates an allocation pool to work from.
Ganesh Mahendranfd854462016-07-28 15:47:54 -07002352 * @name: pool name to be created
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08002353 *
2354 * This function must be called before anything when using
2355 * the zsmalloc allocator.
2356 *
2357 * On success, a pointer to the newly created pool is returned,
2358 * otherwise NULL.
2359 */
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -07002360struct zs_pool *zs_create_pool(const char *name)
Nitin Gupta61989a82012-01-09 16:51:56 -06002361{
Ganesh Mahendran18136652014-12-12 16:57:10 -08002362 int i;
Nitin Gupta61989a82012-01-09 16:51:56 -06002363 struct zs_pool *pool;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08002364 struct size_class *prev_class = NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06002365
Ganesh Mahendran18136652014-12-12 16:57:10 -08002366 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
Nitin Gupta61989a82012-01-09 16:51:56 -06002367 if (!pool)
2368 return NULL;
2369
Minchan Kim48b48002016-07-26 15:23:31 -07002370 init_deferred_free(pool);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002371
Minchan Kim2e40e162015-04-15 16:15:23 -07002372 pool->name = kstrdup(name, GFP_KERNEL);
2373 if (!pool->name)
2374 goto err;
2375
Andrew Morton441e2542019-08-30 16:04:35 -07002376#ifdef CONFIG_COMPACTION
Henry Burns701d6782019-08-24 17:55:06 -07002377 init_waitqueue_head(&pool->migration_wait);
Andrew Morton441e2542019-08-30 16:04:35 -07002378#endif
Henry Burns701d6782019-08-24 17:55:06 -07002379
Minchan Kim37836892016-07-26 15:23:23 -07002380 if (create_cache(pool))
Minchan Kim2e40e162015-04-15 16:15:23 -07002381 goto err;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002382
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002383 /*
Xishi Qiu399d8ee2017-02-22 15:45:01 -08002384 * Iterate reversely, because, size of size_class that we want to use
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002385 * for merging should be larger or equal to current size.
2386 */
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -07002387 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
Nitin Gupta61989a82012-01-09 16:51:56 -06002388 int size;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002389 int pages_per_zspage;
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002390 int objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06002391 struct size_class *class;
Minchan Kim37836892016-07-26 15:23:23 -07002392 int fullness = 0;
Nitin Gupta61989a82012-01-09 16:51:56 -06002393
2394 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2395 if (size > ZS_MAX_ALLOC_SIZE)
2396 size = ZS_MAX_ALLOC_SIZE;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002397 pages_per_zspage = get_pages_per_zspage(size);
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002398 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
Nitin Gupta61989a82012-01-09 16:51:56 -06002399
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002400 /*
Sergey Senozhatsky010b4952018-04-05 16:24:43 -07002401 * We iterate from biggest down to smallest classes,
2402 * so huge_class_size holds the size of the first huge
2403 * class. Any object bigger than or equal to that will
2404 * endup in the huge class.
2405 */
2406 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2407 !huge_class_size) {
2408 huge_class_size = size;
2409 /*
2410 * The object uses ZS_HANDLE_SIZE bytes to store the
2411 * handle. We need to subtract it, because zs_malloc()
2412 * unconditionally adds handle size before it performs
2413 * size class search - so object may be smaller than
2414 * huge class size, yet it still can end up in the huge
2415 * class because it grows by ZS_HANDLE_SIZE extra bytes
2416 * right before class lookup.
2417 */
2418 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2419 }
2420
2421 /*
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002422 * size_class is used for normal zsmalloc operation such
2423 * as alloc/free for that size. Although it is natural that we
2424 * have one size_class for each size, there is a chance that we
2425 * can get more memory utilization if we use one size_class for
2426 * many different sizes whose size_class have same
2427 * characteristics. So, we makes size_class point to
2428 * previous size_class if possible.
2429 */
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08002430 if (prev_class) {
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002431 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002432 pool->size_class[i] = prev_class;
2433 continue;
2434 }
2435 }
2436
2437 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2438 if (!class)
2439 goto err;
2440
Nitin Gupta61989a82012-01-09 16:51:56 -06002441 class->size = size;
2442 class->index = i;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002443 class->pages_per_zspage = pages_per_zspage;
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002444 class->objs_per_zspage = objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06002445 spin_lock_init(&class->lock);
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002446 pool->size_class[i] = class;
Minchan Kim48b48002016-07-26 15:23:31 -07002447 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2448 fullness++)
Minchan Kim37836892016-07-26 15:23:23 -07002449 INIT_LIST_HEAD(&class->fullness_list[fullness]);
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08002450
2451 prev_class = class;
Nitin Gupta61989a82012-01-09 16:51:56 -06002452 }
2453
Dan Streetmand34f6152016-05-20 16:59:56 -07002454 /* debug only, don't abort if it fails */
2455 zs_pool_stat_create(pool, name);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002456
Minchan Kim48b48002016-07-26 15:23:31 -07002457 if (zs_register_migration(pool))
2458 goto err;
2459
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002460 /*
Aliaksei Karaliou93144ca2018-01-31 16:18:40 -08002461 * Not critical since shrinker is only used to trigger internal
2462 * defragmentation of the pool which is pretty optional thing. If
2463 * registration fails we still can use the pool normally and user can
2464 * trigger compaction manually. Thus, ignore return code.
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002465 */
Aliaksei Karaliou93144ca2018-01-31 16:18:40 -08002466 zs_register_shrinker(pool);
2467
Nitin Gupta61989a82012-01-09 16:51:56 -06002468 return pool;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002469
2470err:
2471 zs_destroy_pool(pool);
2472 return NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06002473}
2474EXPORT_SYMBOL_GPL(zs_create_pool);
2475
2476void zs_destroy_pool(struct zs_pool *pool)
2477{
2478 int i;
2479
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002480 zs_unregister_shrinker(pool);
Minchan Kim48b48002016-07-26 15:23:31 -07002481 zs_unregister_migration(pool);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002482 zs_pool_stat_destroy(pool);
2483
Jerome Marchandcf8e0fe2017-07-10 15:50:18 -07002484 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -06002485 int fg;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002486 struct size_class *class = pool->size_class[i];
2487
2488 if (!class)
2489 continue;
2490
2491 if (class->index != i)
2492 continue;
Nitin Gupta61989a82012-01-09 16:51:56 -06002493
Minchan Kim48b48002016-07-26 15:23:31 -07002494 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
Minchan Kim37836892016-07-26 15:23:23 -07002495 if (!list_empty(&class->fullness_list[fg])) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -04002496 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -06002497 class->size, fg);
2498 }
2499 }
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002500 kfree(class);
Nitin Gupta61989a82012-01-09 16:51:56 -06002501 }
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002502
Minchan Kim37836892016-07-26 15:23:23 -07002503 destroy_cache(pool);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002504 kfree(pool->name);
Nitin Gupta61989a82012-01-09 16:51:56 -06002505 kfree(pool);
2506}
2507EXPORT_SYMBOL_GPL(zs_destroy_pool);
2508
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002509static int __init zs_init(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002510{
Minchan Kim48b48002016-07-26 15:23:31 -07002511 int ret;
2512
2513 ret = zsmalloc_mount();
2514 if (ret)
2515 goto out;
2516
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01002517 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2518 zs_cpu_prepare, zs_cpu_dead);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002519 if (ret)
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01002520 goto hp_setup_fail;
Nitin Gupta61989a82012-01-09 16:51:56 -06002521
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002522#ifdef CONFIG_ZPOOL
2523 zpool_register_driver(&zs_zpool_driver);
2524#endif
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002525
Dan Streetman4abaac92016-05-26 15:16:27 -07002526 zs_stat_init();
2527
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002528 return 0;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002529
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01002530hp_setup_fail:
Minchan Kim48b48002016-07-26 15:23:31 -07002531 zsmalloc_unmount();
2532out:
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002533 return ret;
Nitin Gupta61989a82012-01-09 16:51:56 -06002534}
Nitin Gupta61989a82012-01-09 16:51:56 -06002535
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002536static void __exit zs_exit(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002537{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002538#ifdef CONFIG_ZPOOL
2539 zpool_unregister_driver(&zs_zpool_driver);
2540#endif
Minchan Kim48b48002016-07-26 15:23:31 -07002541 zsmalloc_unmount();
Sebastian Andrzej Siewior215c89d2016-11-27 00:13:38 +01002542 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002543
2544 zs_stat_exit();
Nitin Gupta61989a82012-01-09 16:51:56 -06002545}
Ben Hutchings069f1012012-06-20 02:31:11 +01002546
2547module_init(zs_init);
2548module_exit(zs_exit);
2549
2550MODULE_LICENSE("Dual BSD/GPL");
2551MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");