blob: 1689bb58e0d170053cc00a8b41c24404cbf0c3f4 [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
27 * PG_private2: identifies the last component page
Minchan Kim48b48002016-07-26 15:23:31 -070028 * PG_owner_priv_1: indentifies the huge component page
Nitin Gupta2db51da2012-06-09 17:41:14 -070029 *
30 */
31
Dan Streetman4abaac92016-05-26 15:16:27 -070032#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33
Nitin Gupta61989a82012-01-09 16:51:56 -060034#include <linux/module.h>
35#include <linux/kernel.h>
Minchan Kim312fcae2015-04-15 16:15:30 -070036#include <linux/sched.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>
42#include <asm/tlbflush.h>
43#include <asm/pgtable.h>
44#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>
49#include <linux/types.h>
Ganesh Mahendran0f050d92015-02-12 15:00:54 -080050#include <linux/debugfs.h>
Minchan Kimbcf16472014-01-30 15:45:50 -080051#include <linux/zsmalloc.h>
Dan Streetmanc7957792014-08-06 16:08:38 -070052#include <linux/zpool.h>
Minchan Kim48b48002016-07-26 15:23:31 -070053#include <linux/mount.h>
Minchan Kimdd4123f2016-07-26 15:26:50 -070054#include <linux/migrate.h>
Minchan Kim48b48002016-07-26 15:23:31 -070055#include <linux/pagemap.h>
56
57#define ZSPAGE_MAGIC 0x58
Seth Jennings0959c632012-08-08 15:12:17 +090058
59/*
60 * This must be power of 2 and greater than of equal to sizeof(link_free).
61 * These two conditions ensure that any 'struct link_free' itself doesn't
62 * span more than 1 page which avoids complex case of mapping 2 pages simply
63 * to restore link_free pointer values.
64 */
65#define ZS_ALIGN 8
66
67/*
68 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
69 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
70 */
71#define ZS_MAX_ZSPAGE_ORDER 2
72#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
73
Minchan Kim2e40e162015-04-15 16:15:23 -070074#define ZS_HANDLE_SIZE (sizeof(unsigned long))
75
Seth Jennings0959c632012-08-08 15:12:17 +090076/*
77 * Object location (<PFN>, <obj_idx>) is encoded as
Nitin Cuptac3e3e882013-12-11 11:04:37 +090078 * as single (unsigned long) handle value.
Seth Jennings0959c632012-08-08 15:12:17 +090079 *
Minchan Kimbfd093f2016-07-26 15:23:28 -070080 * Note that object index <obj_idx> starts from 0.
Seth Jennings0959c632012-08-08 15:12:17 +090081 *
82 * This is made more complicated by various memory models and PAE.
83 */
84
85#ifndef MAX_PHYSMEM_BITS
86#ifdef CONFIG_HIGHMEM64G
87#define MAX_PHYSMEM_BITS 36
88#else /* !CONFIG_HIGHMEM64G */
89/*
90 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
91 * be PAGE_SHIFT
92 */
93#define MAX_PHYSMEM_BITS BITS_PER_LONG
94#endif
95#endif
96#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
Minchan Kim312fcae2015-04-15 16:15:30 -070097
98/*
99 * Memory for allocating for handle keeps object position by
100 * encoding <page, obj_idx> and the encoded value has a room
101 * in least bit(ie, look at obj_to_location).
102 * We use the bit to synchronize between object access by
103 * user and migration.
104 */
105#define HANDLE_PIN_BIT 0
106
107/*
108 * Head in allocated object should have OBJ_ALLOCATED_TAG
109 * to identify the object was allocated or not.
110 * It's okay to add the status bit in the least bit because
111 * header keeps handle which is 4byte-aligned address so we
112 * have room for two bit at least.
113 */
114#define OBJ_ALLOCATED_TAG 1
115#define OBJ_TAG_BITS 1
116#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
Seth Jennings0959c632012-08-08 15:12:17 +0900117#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
118
119#define MAX(a, b) ((a) >= (b) ? (a) : (b))
120/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
121#define ZS_MIN_ALLOC_SIZE \
122 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
Minchan Kim2e40e162015-04-15 16:15:23 -0700123/* each chunk includes extra space to keep handle */
Minchan Kim7b60a682015-04-15 16:15:39 -0700124#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
Seth Jennings0959c632012-08-08 15:12:17 +0900125
126/*
Weijie Yang7eb52512014-06-04 16:11:08 -0700127 * On systems with 4K page size, this gives 255 size classes! There is a
Seth Jennings0959c632012-08-08 15:12:17 +0900128 * trader-off here:
129 * - Large number of size classes is potentially wasteful as free page are
130 * spread across these classes
131 * - Small number of size classes causes large internal fragmentation
132 * - Probably its better to use specific size classes (empirically
133 * determined). NOTE: all those class sizes must be set as multiple of
134 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
135 *
136 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
137 * (reason above)
138 */
Minchan Kim37836892016-07-26 15:23:23 -0700139#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
Seth Jennings0959c632012-08-08 15:12:17 +0900140
Seth Jennings0959c632012-08-08 15:12:17 +0900141enum fullness_group {
Seth Jennings0959c632012-08-08 15:12:17 +0900142 ZS_EMPTY,
Minchan Kim48b48002016-07-26 15:23:31 -0700143 ZS_ALMOST_EMPTY,
144 ZS_ALMOST_FULL,
145 ZS_FULL,
146 NR_ZS_FULLNESS,
Seth Jennings0959c632012-08-08 15:12:17 +0900147};
148
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800149enum zs_stat_type {
Minchan Kim48b48002016-07-26 15:23:31 -0700150 CLASS_EMPTY,
151 CLASS_ALMOST_EMPTY,
152 CLASS_ALMOST_FULL,
153 CLASS_FULL,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800154 OBJ_ALLOCATED,
155 OBJ_USED,
Minchan Kim48b48002016-07-26 15:23:31 -0700156 NR_ZS_STAT_TYPE,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800157};
158
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800159struct zs_size_stat {
160 unsigned long objs[NR_ZS_STAT_TYPE];
161};
162
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700163#ifdef CONFIG_ZSMALLOC_STAT
164static struct dentry *zs_stat_root;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800165#endif
166
Minchan Kim48b48002016-07-26 15:23:31 -0700167#ifdef CONFIG_COMPACTION
168static struct vfsmount *zsmalloc_mnt;
169#endif
170
Seth Jennings0959c632012-08-08 15:12:17 +0900171/*
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800172 * number of size_classes
173 */
174static int zs_size_classes;
175
176/*
Seth Jennings0959c632012-08-08 15:12:17 +0900177 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
178 * n <= N / f, where
179 * n = number of allocated objects
180 * N = total number of objects zspage can store
Wang Sheng-Hui6dd97372014-10-09 15:29:59 -0700181 * f = fullness_threshold_frac
Seth Jennings0959c632012-08-08 15:12:17 +0900182 *
183 * Similarly, we assign zspage to:
184 * ZS_ALMOST_FULL when n > N / f
185 * ZS_EMPTY when n == 0
186 * ZS_FULL when n == N
187 *
188 * (see: fix_fullness_group())
189 */
190static const int fullness_threshold_frac = 4;
191
192struct size_class {
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700193 spinlock_t lock;
Minchan Kim48b48002016-07-26 15:23:31 -0700194 struct list_head fullness_list[NR_ZS_FULLNESS];
Seth Jennings0959c632012-08-08 15:12:17 +0900195 /*
196 * Size of objects stored in this class. Must be multiple
197 * of ZS_ALIGN.
198 */
199 int size;
Minchan Kim1fc6e272016-07-26 15:23:11 -0700200 int objs_per_zspage;
Weijie Yang7dfa4612016-01-14 15:22:40 -0800201 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
202 int pages_per_zspage;
Minchan Kim48b48002016-07-26 15:23:31 -0700203
204 unsigned int index;
205 struct zs_size_stat stats;
Seth Jennings0959c632012-08-08 15:12:17 +0900206};
207
Minchan Kim48b48002016-07-26 15:23:31 -0700208/* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
209static void SetPageHugeObject(struct page *page)
210{
211 SetPageOwnerPriv1(page);
212}
213
214static void ClearPageHugeObject(struct page *page)
215{
216 ClearPageOwnerPriv1(page);
217}
218
219static int PageHugeObject(struct page *page)
220{
221 return PageOwnerPriv1(page);
222}
223
Seth Jennings0959c632012-08-08 15:12:17 +0900224/*
225 * Placed within free objects to form a singly linked list.
Minchan Kim37836892016-07-26 15:23:23 -0700226 * For every zspage, zspage->freeobj gives head of this list.
Seth Jennings0959c632012-08-08 15:12:17 +0900227 *
228 * This must be power of 2 and less than or equal to ZS_ALIGN
229 */
230struct link_free {
Minchan Kim2e40e162015-04-15 16:15:23 -0700231 union {
232 /*
Minchan Kimbfd093f2016-07-26 15:23:28 -0700233 * Free object index;
Minchan Kim2e40e162015-04-15 16:15:23 -0700234 * It's valid for non-allocated object
235 */
Minchan Kimbfd093f2016-07-26 15:23:28 -0700236 unsigned long next;
Minchan Kim2e40e162015-04-15 16:15:23 -0700237 /*
238 * Handle of allocated object.
239 */
240 unsigned long handle;
241 };
Seth Jennings0959c632012-08-08 15:12:17 +0900242};
243
244struct zs_pool {
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800245 const char *name;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800246
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800247 struct size_class **size_class;
Minchan Kim2e40e162015-04-15 16:15:23 -0700248 struct kmem_cache *handle_cachep;
Minchan Kim37836892016-07-26 15:23:23 -0700249 struct kmem_cache *zspage_cachep;
Seth Jennings0959c632012-08-08 15:12:17 +0900250
Minchan Kim13de8932014-10-09 15:29:48 -0700251 atomic_long_t pages_allocated;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800252
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -0700253 struct zs_pool_stats stats;
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -0700254
255 /* Compact classes */
256 struct shrinker shrinker;
257 /*
258 * To signify that register_shrinker() was successful
259 * and unregister_shrinker() will not Oops.
260 */
261 bool shrinker_enabled;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800262#ifdef CONFIG_ZSMALLOC_STAT
263 struct dentry *stat_dentry;
264#endif
Minchan Kim48b48002016-07-26 15:23:31 -0700265#ifdef CONFIG_COMPACTION
266 struct inode *inode;
267 struct work_struct free_work;
268#endif
Seth Jennings0959c632012-08-08 15:12:17 +0900269};
Nitin Gupta61989a82012-01-09 16:51:56 -0600270
271/*
272 * A zspage's class index and fullness group
273 * are encoded in its (first)page->mapping
274 */
Minchan Kim37836892016-07-26 15:23:23 -0700275#define FULLNESS_BITS 2
276#define CLASS_BITS 8
Minchan Kim48b48002016-07-26 15:23:31 -0700277#define ISOLATED_BITS 3
278#define MAGIC_VAL_BITS 8
Minchan Kim4f420472016-07-26 15:23:17 -0700279
Minchan Kim37836892016-07-26 15:23:23 -0700280struct zspage {
281 struct {
282 unsigned int fullness:FULLNESS_BITS;
Minchan Kimd19f7452017-04-13 14:56:40 -0700283 unsigned int class:CLASS_BITS + 1;
Minchan Kim48b48002016-07-26 15:23:31 -0700284 unsigned int isolated:ISOLATED_BITS;
285 unsigned int magic:MAGIC_VAL_BITS;
Minchan Kim37836892016-07-26 15:23:23 -0700286 };
287 unsigned int inuse;
Minchan Kimbfd093f2016-07-26 15:23:28 -0700288 unsigned int freeobj;
Minchan Kim37836892016-07-26 15:23:23 -0700289 struct page *first_page;
290 struct list_head list; /* fullness list */
Minchan Kim48b48002016-07-26 15:23:31 -0700291#ifdef CONFIG_COMPACTION
292 rwlock_t lock;
293#endif
Minchan Kim37836892016-07-26 15:23:23 -0700294};
Nitin Gupta61989a82012-01-09 16:51:56 -0600295
Seth Jenningsf5536462012-07-18 11:55:56 -0500296struct mapping_area {
Minchan Kim1b945ae2013-12-11 11:04:36 +0900297#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -0500298 struct vm_struct *vm; /* vm area for mapping object that span pages */
299#else
300 char *vm_buf; /* copy buffer for objects that span pages */
301#endif
302 char *vm_addr; /* address of kmap_atomic()'ed pages */
303 enum zs_mapmode vm_mm; /* mapping mode */
304};
305
Minchan Kim48b48002016-07-26 15:23:31 -0700306#ifdef CONFIG_COMPACTION
307static int zs_register_migration(struct zs_pool *pool);
308static void zs_unregister_migration(struct zs_pool *pool);
309static void migrate_lock_init(struct zspage *zspage);
310static void migrate_read_lock(struct zspage *zspage);
311static void migrate_read_unlock(struct zspage *zspage);
312static void kick_deferred_free(struct zs_pool *pool);
313static void init_deferred_free(struct zs_pool *pool);
314static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
315#else
316static int zsmalloc_mount(void) { return 0; }
317static void zsmalloc_unmount(void) {}
318static int zs_register_migration(struct zs_pool *pool) { return 0; }
319static void zs_unregister_migration(struct zs_pool *pool) {}
320static void migrate_lock_init(struct zspage *zspage) {}
321static void migrate_read_lock(struct zspage *zspage) {}
322static void migrate_read_unlock(struct zspage *zspage) {}
323static void kick_deferred_free(struct zs_pool *pool) {}
324static void init_deferred_free(struct zs_pool *pool) {}
325static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
326#endif
327
Minchan Kim37836892016-07-26 15:23:23 -0700328static int create_cache(struct zs_pool *pool)
Minchan Kim2e40e162015-04-15 16:15:23 -0700329{
330 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
331 0, 0, NULL);
Minchan Kim37836892016-07-26 15:23:23 -0700332 if (!pool->handle_cachep)
333 return 1;
334
335 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
336 0, 0, NULL);
337 if (!pool->zspage_cachep) {
338 kmem_cache_destroy(pool->handle_cachep);
339 pool->handle_cachep = NULL;
340 return 1;
341 }
342
343 return 0;
Minchan Kim2e40e162015-04-15 16:15:23 -0700344}
345
Minchan Kim37836892016-07-26 15:23:23 -0700346static void destroy_cache(struct zs_pool *pool)
Minchan Kim2e40e162015-04-15 16:15:23 -0700347{
Sergey Senozhatskycd10add2015-09-08 15:04:55 -0700348 kmem_cache_destroy(pool->handle_cachep);
Minchan Kim37836892016-07-26 15:23:23 -0700349 kmem_cache_destroy(pool->zspage_cachep);
Minchan Kim2e40e162015-04-15 16:15:23 -0700350}
351
Minchan Kim37836892016-07-26 15:23:23 -0700352static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
Minchan Kim2e40e162015-04-15 16:15:23 -0700353{
354 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
Minchan Kim48b48002016-07-26 15:23:31 -0700355 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
Minchan Kim2e40e162015-04-15 16:15:23 -0700356}
357
Minchan Kim37836892016-07-26 15:23:23 -0700358static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
Minchan Kim2e40e162015-04-15 16:15:23 -0700359{
360 kmem_cache_free(pool->handle_cachep, (void *)handle);
361}
362
Minchan Kim37836892016-07-26 15:23:23 -0700363static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
364{
Minchan Kim48b48002016-07-26 15:23:31 -0700365 return kmem_cache_alloc(pool->zspage_cachep,
366 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
Minchan Kim37836892016-07-26 15:23:23 -0700367};
368
369static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
370{
371 kmem_cache_free(pool->zspage_cachep, zspage);
372}
373
Minchan Kim2e40e162015-04-15 16:15:23 -0700374static void record_obj(unsigned long handle, unsigned long obj)
375{
Junil Leec102f072016-01-20 14:58:18 -0800376 /*
377 * lsb of @obj represents handle lock while other bits
378 * represent object value the handle is pointing so
379 * updating shouldn't do store tearing.
380 */
381 WRITE_ONCE(*(unsigned long *)handle, obj);
Minchan Kim2e40e162015-04-15 16:15:23 -0700382}
383
Dan Streetmanc7957792014-08-06 16:08:38 -0700384/* zpool driver */
385
386#ifdef CONFIG_ZPOOL
387
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800388static void *zs_zpool_create(const char *name, gfp_t gfp,
Krzysztof Kozlowski78672772015-09-08 15:05:03 -0700389 const struct zpool_ops *zpool_ops,
Dan Streetman479305f2015-06-25 15:00:40 -0700390 struct zpool *zpool)
Dan Streetmanc7957792014-08-06 16:08:38 -0700391{
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -0700392 /*
393 * Ignore global gfp flags: zs_malloc() may be invoked from
394 * different contexts and its caller must provide a valid
395 * gfp mask.
396 */
397 return zs_create_pool(name);
Dan Streetmanc7957792014-08-06 16:08:38 -0700398}
399
400static void zs_zpool_destroy(void *pool)
401{
402 zs_destroy_pool(pool);
403}
404
405static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
406 unsigned long *handle)
407{
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -0700408 *handle = zs_malloc(pool, size, gfp);
Dan Streetmanc7957792014-08-06 16:08:38 -0700409 return *handle ? 0 : -1;
410}
411static void zs_zpool_free(void *pool, unsigned long handle)
412{
413 zs_free(pool, handle);
414}
415
416static int zs_zpool_shrink(void *pool, unsigned int pages,
417 unsigned int *reclaimed)
418{
419 return -EINVAL;
420}
421
422static void *zs_zpool_map(void *pool, unsigned long handle,
423 enum zpool_mapmode mm)
424{
425 enum zs_mapmode zs_mm;
426
427 switch (mm) {
428 case ZPOOL_MM_RO:
429 zs_mm = ZS_MM_RO;
430 break;
431 case ZPOOL_MM_WO:
432 zs_mm = ZS_MM_WO;
433 break;
434 case ZPOOL_MM_RW: /* fallthru */
435 default:
436 zs_mm = ZS_MM_RW;
437 break;
438 }
439
440 return zs_map_object(pool, handle, zs_mm);
441}
442static void zs_zpool_unmap(void *pool, unsigned long handle)
443{
444 zs_unmap_object(pool, handle);
445}
446
447static u64 zs_zpool_total_size(void *pool)
448{
Minchan Kim722cdc12014-10-09 15:29:50 -0700449 return zs_get_total_pages(pool) << PAGE_SHIFT;
Dan Streetmanc7957792014-08-06 16:08:38 -0700450}
451
452static struct zpool_driver zs_zpool_driver = {
453 .type = "zsmalloc",
454 .owner = THIS_MODULE,
455 .create = zs_zpool_create,
456 .destroy = zs_zpool_destroy,
457 .malloc = zs_zpool_malloc,
458 .free = zs_zpool_free,
459 .shrink = zs_zpool_shrink,
460 .map = zs_zpool_map,
461 .unmap = zs_zpool_unmap,
462 .total_size = zs_zpool_total_size,
463};
464
Kees Cook137f8cf2014-08-29 15:18:40 -0700465MODULE_ALIAS("zpool-zsmalloc");
Dan Streetmanc7957792014-08-06 16:08:38 -0700466#endif /* CONFIG_ZPOOL */
467
Nitin Gupta61989a82012-01-09 16:51:56 -0600468/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
469static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
470
Minchan Kim48b48002016-07-26 15:23:31 -0700471static bool is_zspage_isolated(struct zspage *zspage)
472{
473 return zspage->isolated;
474}
475
Nitin Gupta61989a82012-01-09 16:51:56 -0600476static int is_first_page(struct page *page)
477{
Minchan Kima27545bf2012-04-25 15:23:09 +0900478 return PagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600479}
480
Minchan Kim48b48002016-07-26 15:23:31 -0700481/* Protected by class->lock */
Minchan Kim37836892016-07-26 15:23:23 -0700482static inline int get_zspage_inuse(struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600483{
Minchan Kim37836892016-07-26 15:23:23 -0700484 return zspage->inuse;
Nitin Gupta61989a82012-01-09 16:51:56 -0600485}
486
Minchan Kim37836892016-07-26 15:23:23 -0700487static inline void set_zspage_inuse(struct zspage *zspage, int val)
Minchan Kim4f420472016-07-26 15:23:17 -0700488{
Minchan Kim37836892016-07-26 15:23:23 -0700489 zspage->inuse = val;
Minchan Kim4f420472016-07-26 15:23:17 -0700490}
491
Minchan Kim37836892016-07-26 15:23:23 -0700492static inline void mod_zspage_inuse(struct zspage *zspage, int val)
Minchan Kim4f420472016-07-26 15:23:17 -0700493{
Minchan Kim37836892016-07-26 15:23:23 -0700494 zspage->inuse += val;
Minchan Kim4f420472016-07-26 15:23:17 -0700495}
496
Minchan Kim48b48002016-07-26 15:23:31 -0700497static inline struct page *get_first_page(struct zspage *zspage)
498{
499 struct page *first_page = zspage->first_page;
500
501 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
502 return first_page;
503}
504
Minchan Kim4f420472016-07-26 15:23:17 -0700505static inline int get_first_obj_offset(struct page *page)
506{
Minchan Kim48b48002016-07-26 15:23:31 -0700507 return page->units;
Minchan Kim4f420472016-07-26 15:23:17 -0700508}
509
510static inline void set_first_obj_offset(struct page *page, int offset)
511{
Minchan Kim48b48002016-07-26 15:23:31 -0700512 page->units = offset;
Minchan Kim4f420472016-07-26 15:23:17 -0700513}
514
Minchan Kimbfd093f2016-07-26 15:23:28 -0700515static inline unsigned int get_freeobj(struct zspage *zspage)
Minchan Kim4f420472016-07-26 15:23:17 -0700516{
Minchan Kimbfd093f2016-07-26 15:23:28 -0700517 return zspage->freeobj;
Minchan Kim4f420472016-07-26 15:23:17 -0700518}
519
Minchan Kimbfd093f2016-07-26 15:23:28 -0700520static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
Minchan Kim4f420472016-07-26 15:23:17 -0700521{
Minchan Kimbfd093f2016-07-26 15:23:28 -0700522 zspage->freeobj = obj;
Minchan Kim4f420472016-07-26 15:23:17 -0700523}
524
Minchan Kim37836892016-07-26 15:23:23 -0700525static void get_zspage_mapping(struct zspage *zspage,
Minchan Kima4209462016-05-20 16:59:36 -0700526 unsigned int *class_idx,
Nitin Gupta61989a82012-01-09 16:51:56 -0600527 enum fullness_group *fullness)
528{
Minchan Kim48b48002016-07-26 15:23:31 -0700529 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
530
Minchan Kim37836892016-07-26 15:23:23 -0700531 *fullness = zspage->fullness;
532 *class_idx = zspage->class;
Nitin Gupta61989a82012-01-09 16:51:56 -0600533}
534
Minchan Kim37836892016-07-26 15:23:23 -0700535static void set_zspage_mapping(struct zspage *zspage,
Minchan Kima4209462016-05-20 16:59:36 -0700536 unsigned int class_idx,
Nitin Gupta61989a82012-01-09 16:51:56 -0600537 enum fullness_group fullness)
538{
Minchan Kim37836892016-07-26 15:23:23 -0700539 zspage->class = class_idx;
540 zspage->fullness = fullness;
Nitin Gupta61989a82012-01-09 16:51:56 -0600541}
542
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900543/*
544 * zsmalloc divides the pool into various size classes where each
545 * class maintains a list of zspages where each zspage is divided
546 * into equal sized chunks. Each allocation falls into one of these
547 * classes depending on its size. This function returns index of the
548 * size class which has chunk size big enough to hold the give size.
549 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600550static int get_size_class_index(int size)
551{
552 int idx = 0;
553
554 if (likely(size > ZS_MIN_ALLOC_SIZE))
555 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
556 ZS_SIZE_CLASS_DELTA);
557
Minchan Kim7b60a682015-04-15 16:15:39 -0700558 return min(zs_size_classes - 1, idx);
Nitin Gupta61989a82012-01-09 16:51:56 -0600559}
560
Minchan Kim248ca1b2015-04-15 16:15:42 -0700561static inline void zs_stat_inc(struct size_class *class,
562 enum zs_stat_type type, unsigned long cnt)
563{
Minchan Kim48b48002016-07-26 15:23:31 -0700564 class->stats.objs[type] += cnt;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700565}
566
567static inline void zs_stat_dec(struct size_class *class,
568 enum zs_stat_type type, unsigned long cnt)
569{
Minchan Kim48b48002016-07-26 15:23:31 -0700570 class->stats.objs[type] -= cnt;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700571}
572
573static inline unsigned long zs_stat_get(struct size_class *class,
574 enum zs_stat_type type)
575{
Minchan Kim48b48002016-07-26 15:23:31 -0700576 return class->stats.objs[type];
Minchan Kim248ca1b2015-04-15 16:15:42 -0700577}
578
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700579#ifdef CONFIG_ZSMALLOC_STAT
580
Dan Streetman4abaac92016-05-26 15:16:27 -0700581static void __init zs_stat_init(void)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700582{
Dan Streetman4abaac92016-05-26 15:16:27 -0700583 if (!debugfs_initialized()) {
584 pr_warn("debugfs not available, stat dir not created\n");
585 return;
586 }
Minchan Kim248ca1b2015-04-15 16:15:42 -0700587
588 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
589 if (!zs_stat_root)
Dan Streetman4abaac92016-05-26 15:16:27 -0700590 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
Minchan Kim248ca1b2015-04-15 16:15:42 -0700591}
592
593static void __exit zs_stat_exit(void)
594{
595 debugfs_remove_recursive(zs_stat_root);
596}
597
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700598static unsigned long zs_can_compact(struct size_class *class);
599
Minchan Kim248ca1b2015-04-15 16:15:42 -0700600static int zs_stats_size_show(struct seq_file *s, void *v)
601{
602 int i;
603 struct zs_pool *pool = s->private;
604 struct size_class *class;
605 int objs_per_zspage;
606 unsigned long class_almost_full, class_almost_empty;
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700607 unsigned long obj_allocated, obj_used, pages_used, freeable;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700608 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
609 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700610 unsigned long total_freeable = 0;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700611
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700612 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
Minchan Kim248ca1b2015-04-15 16:15:42 -0700613 "class", "size", "almost_full", "almost_empty",
614 "obj_allocated", "obj_used", "pages_used",
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700615 "pages_per_zspage", "freeable");
Minchan Kim248ca1b2015-04-15 16:15:42 -0700616
617 for (i = 0; i < zs_size_classes; i++) {
618 class = pool->size_class[i];
619
620 if (class->index != i)
621 continue;
622
623 spin_lock(&class->lock);
624 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
625 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
626 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
627 obj_used = zs_stat_get(class, OBJ_USED);
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700628 freeable = zs_can_compact(class);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700629 spin_unlock(&class->lock);
630
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -0700631 objs_per_zspage = class->objs_per_zspage;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700632 pages_used = obj_allocated / objs_per_zspage *
633 class->pages_per_zspage;
634
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700635 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
636 " %10lu %10lu %16d %8lu\n",
Minchan Kim248ca1b2015-04-15 16:15:42 -0700637 i, class->size, class_almost_full, class_almost_empty,
638 obj_allocated, obj_used, pages_used,
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700639 class->pages_per_zspage, freeable);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700640
641 total_class_almost_full += class_almost_full;
642 total_class_almost_empty += class_almost_empty;
643 total_objs += obj_allocated;
644 total_used_objs += obj_used;
645 total_pages += pages_used;
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700646 total_freeable += freeable;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700647 }
648
649 seq_puts(s, "\n");
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700650 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
Minchan Kim248ca1b2015-04-15 16:15:42 -0700651 "Total", "", total_class_almost_full,
652 total_class_almost_empty, total_objs,
Sergey Senozhatsky1120ed52016-03-17 14:20:42 -0700653 total_used_objs, total_pages, "", total_freeable);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700654
655 return 0;
656}
657
658static int zs_stats_size_open(struct inode *inode, struct file *file)
659{
660 return single_open(file, zs_stats_size_show, inode->i_private);
661}
662
663static const struct file_operations zs_stat_size_ops = {
664 .open = zs_stats_size_open,
665 .read = seq_read,
666 .llseek = seq_lseek,
667 .release = single_release,
668};
669
Dan Streetmand34f6152016-05-20 16:59:56 -0700670static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700671{
672 struct dentry *entry;
673
Dan Streetman4abaac92016-05-26 15:16:27 -0700674 if (!zs_stat_root) {
675 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
Dan Streetmand34f6152016-05-20 16:59:56 -0700676 return;
Dan Streetman4abaac92016-05-26 15:16:27 -0700677 }
Minchan Kim248ca1b2015-04-15 16:15:42 -0700678
679 entry = debugfs_create_dir(name, zs_stat_root);
680 if (!entry) {
681 pr_warn("debugfs dir <%s> creation failed\n", name);
Dan Streetmand34f6152016-05-20 16:59:56 -0700682 return;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700683 }
684 pool->stat_dentry = entry;
685
686 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
687 pool->stat_dentry, pool, &zs_stat_size_ops);
688 if (!entry) {
689 pr_warn("%s: debugfs file entry <%s> creation failed\n",
690 name, "classes");
Dan Streetman4abaac92016-05-26 15:16:27 -0700691 debugfs_remove_recursive(pool->stat_dentry);
692 pool->stat_dentry = NULL;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700693 }
Minchan Kim248ca1b2015-04-15 16:15:42 -0700694}
695
696static void zs_pool_stat_destroy(struct zs_pool *pool)
697{
698 debugfs_remove_recursive(pool->stat_dentry);
699}
700
701#else /* CONFIG_ZSMALLOC_STAT */
Dan Streetman4abaac92016-05-26 15:16:27 -0700702static void __init zs_stat_init(void)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700703{
Minchan Kim248ca1b2015-04-15 16:15:42 -0700704}
705
706static void __exit zs_stat_exit(void)
707{
708}
709
Dan Streetmand34f6152016-05-20 16:59:56 -0700710static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700711{
Minchan Kim248ca1b2015-04-15 16:15:42 -0700712}
713
714static inline void zs_pool_stat_destroy(struct zs_pool *pool)
715{
716}
Minchan Kim248ca1b2015-04-15 16:15:42 -0700717#endif
718
Minchan Kim48b48002016-07-26 15:23:31 -0700719
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900720/*
721 * For each size class, zspages are divided into different groups
722 * depending on how "full" they are. This was done so that we could
723 * easily find empty or nearly empty zspages when we try to shrink
724 * the pool (not yet implemented). This function returns fullness
725 * status of the given page.
726 */
Minchan Kim1fc6e272016-07-26 15:23:11 -0700727static enum fullness_group get_fullness_group(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700728 struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600729{
Minchan Kim1fc6e272016-07-26 15:23:11 -0700730 int inuse, objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600731 enum fullness_group fg;
Minchan Kim830e4bc2016-05-20 16:59:39 -0700732
Minchan Kim37836892016-07-26 15:23:23 -0700733 inuse = get_zspage_inuse(zspage);
Minchan Kim1fc6e272016-07-26 15:23:11 -0700734 objs_per_zspage = class->objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600735
736 if (inuse == 0)
737 fg = ZS_EMPTY;
Minchan Kim1fc6e272016-07-26 15:23:11 -0700738 else if (inuse == objs_per_zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600739 fg = ZS_FULL;
Minchan Kim1fc6e272016-07-26 15:23:11 -0700740 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
Nitin Gupta61989a82012-01-09 16:51:56 -0600741 fg = ZS_ALMOST_EMPTY;
742 else
743 fg = ZS_ALMOST_FULL;
744
745 return fg;
746}
747
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900748/*
749 * Each size class maintains various freelists and zspages are assigned
750 * to one of these freelists based on the number of live objects they
751 * have. This functions inserts the given zspage into the freelist
752 * identified by <class, fullness_group>.
753 */
Minchan Kim251cbb92016-05-20 16:59:42 -0700754static void insert_zspage(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700755 struct zspage *zspage,
756 enum fullness_group fullness)
Nitin Gupta61989a82012-01-09 16:51:56 -0600757{
Minchan Kim37836892016-07-26 15:23:23 -0700758 struct zspage *head;
Nitin Gupta61989a82012-01-09 16:51:56 -0600759
Minchan Kim48b48002016-07-26 15:23:31 -0700760 zs_stat_inc(class, fullness, 1);
Minchan Kim37836892016-07-26 15:23:23 -0700761 head = list_first_entry_or_null(&class->fullness_list[fullness],
762 struct zspage, list);
Sergey Senozhatsky58f17112015-09-08 15:04:44 -0700763 /*
Minchan Kim37836892016-07-26 15:23:23 -0700764 * We want to see more ZS_FULL pages and less almost empty/full.
765 * Put pages with higher ->inuse first.
Sergey Senozhatsky58f17112015-09-08 15:04:44 -0700766 */
Minchan Kim37836892016-07-26 15:23:23 -0700767 if (head) {
768 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
769 list_add(&zspage->list, &head->list);
770 return;
771 }
772 }
773 list_add(&zspage->list, &class->fullness_list[fullness]);
Nitin Gupta61989a82012-01-09 16:51:56 -0600774}
775
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900776/*
777 * This function removes the given zspage from the freelist identified
778 * by <class, fullness_group>.
779 */
Minchan Kim251cbb92016-05-20 16:59:42 -0700780static void remove_zspage(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700781 struct zspage *zspage,
782 enum fullness_group fullness)
Nitin Gupta61989a82012-01-09 16:51:56 -0600783{
Minchan Kim37836892016-07-26 15:23:23 -0700784 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
Minchan Kim48b48002016-07-26 15:23:31 -0700785 VM_BUG_ON(is_zspage_isolated(zspage));
Nitin Gupta61989a82012-01-09 16:51:56 -0600786
Minchan Kim37836892016-07-26 15:23:23 -0700787 list_del_init(&zspage->list);
Minchan Kim48b48002016-07-26 15:23:31 -0700788 zs_stat_dec(class, fullness, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600789}
790
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900791/*
792 * Each size class maintains zspages in different fullness groups depending
793 * on the number of live objects they contain. When allocating or freeing
794 * objects, the fullness status of the page can change, say, from ALMOST_FULL
795 * to ALMOST_EMPTY when freeing an object. This function checks if such
796 * a status change has occurred for the given page and accordingly moves the
797 * page from the freelist of the old fullness group to that of the new
798 * fullness group.
799 */
Minchan Kimc7806262015-04-15 16:15:26 -0700800static enum fullness_group fix_fullness_group(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -0700801 struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600802{
803 int class_idx;
Nitin Gupta61989a82012-01-09 16:51:56 -0600804 enum fullness_group currfg, newfg;
805
Minchan Kim37836892016-07-26 15:23:23 -0700806 get_zspage_mapping(zspage, &class_idx, &currfg);
807 newfg = get_fullness_group(class, zspage);
Nitin Gupta61989a82012-01-09 16:51:56 -0600808 if (newfg == currfg)
809 goto out;
810
Minchan Kim48b48002016-07-26 15:23:31 -0700811 if (!is_zspage_isolated(zspage)) {
812 remove_zspage(class, zspage, currfg);
813 insert_zspage(class, zspage, newfg);
814 }
815
Minchan Kim37836892016-07-26 15:23:23 -0700816 set_zspage_mapping(zspage, class_idx, newfg);
Nitin Gupta61989a82012-01-09 16:51:56 -0600817
818out:
819 return newfg;
820}
821
822/*
823 * We have to decide on how many pages to link together
824 * to form a zspage for each size class. This is important
825 * to reduce wastage due to unusable space left at end of
826 * each zspage which is given as:
Yinghao Xie888fa3742015-04-15 16:15:49 -0700827 * wastage = Zp % class_size
828 * usage = Zp - wastage
Nitin Gupta61989a82012-01-09 16:51:56 -0600829 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
830 *
831 * For example, for size class of 3/8 * PAGE_SIZE, we should
832 * link together 3 PAGE_SIZE sized pages to form a zspage
833 * since then we can perfectly fit in 8 such objects.
834 */
Minchan Kim2e3b6152012-05-03 15:40:39 +0900835static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600836{
837 int i, max_usedpc = 0;
838 /* zspage order which gives maximum used size per KB */
839 int max_usedpc_order = 1;
840
Seth Jennings84d4faa2012-03-05 11:33:21 -0600841 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600842 int zspage_size;
843 int waste, usedpc;
844
845 zspage_size = i * PAGE_SIZE;
846 waste = zspage_size % class_size;
847 usedpc = (zspage_size - waste) * 100 / zspage_size;
848
849 if (usedpc > max_usedpc) {
850 max_usedpc = usedpc;
851 max_usedpc_order = i;
852 }
853 }
854
855 return max_usedpc_order;
856}
857
Minchan Kim37836892016-07-26 15:23:23 -0700858static struct zspage *get_zspage(struct page *page)
Nitin Gupta61989a82012-01-09 16:51:56 -0600859{
Minchan Kim48b48002016-07-26 15:23:31 -0700860 struct zspage *zspage = (struct zspage *)page->private;
861
862 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
863 return zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -0600864}
865
866static struct page *get_next_page(struct page *page)
867{
Minchan Kim48b48002016-07-26 15:23:31 -0700868 if (unlikely(PageHugeObject(page)))
869 return NULL;
870
871 return page->freelist;
Nitin Gupta61989a82012-01-09 16:51:56 -0600872}
873
Minchan Kimbfd093f2016-07-26 15:23:28 -0700874/**
875 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
876 * @page: page object resides in zspage
877 * @obj_idx: object index
Olav Haugan67296872013-11-22 09:30:41 -0800878 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700879static void obj_to_location(unsigned long obj, struct page **page,
Minchan Kimbfd093f2016-07-26 15:23:28 -0700880 unsigned int *obj_idx)
Nitin Gupta61989a82012-01-09 16:51:56 -0600881{
Minchan Kim312fcae2015-04-15 16:15:30 -0700882 obj >>= OBJ_TAG_BITS;
883 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
884 *obj_idx = (obj & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600885}
886
Minchan Kimbfd093f2016-07-26 15:23:28 -0700887/**
888 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
889 * @page: page object resides in zspage
890 * @obj_idx: object index
891 */
892static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
893{
894 unsigned long obj;
895
896 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
897 obj |= obj_idx & OBJ_INDEX_MASK;
898 obj <<= OBJ_TAG_BITS;
899
900 return obj;
901}
902
Minchan Kim2e40e162015-04-15 16:15:23 -0700903static unsigned long handle_to_obj(unsigned long handle)
904{
905 return *(unsigned long *)handle;
906}
907
Minchan Kim48b48002016-07-26 15:23:31 -0700908static unsigned long obj_to_head(struct page *page, void *obj)
Minchan Kim312fcae2015-04-15 16:15:30 -0700909{
Minchan Kim48b48002016-07-26 15:23:31 -0700910 if (unlikely(PageHugeObject(page))) {
Minchan Kim830e4bc2016-05-20 16:59:39 -0700911 VM_BUG_ON_PAGE(!is_first_page(page), page);
Minchan Kim37836892016-07-26 15:23:23 -0700912 return page->index;
Minchan Kim7b60a682015-04-15 16:15:39 -0700913 } else
914 return *(unsigned long *)obj;
Minchan Kim312fcae2015-04-15 16:15:30 -0700915}
916
Minchan Kim48b48002016-07-26 15:23:31 -0700917static inline int testpin_tag(unsigned long handle)
918{
919 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
920}
921
Minchan Kim312fcae2015-04-15 16:15:30 -0700922static inline int trypin_tag(unsigned long handle)
923{
Minchan Kim1b8320b2016-07-26 15:23:14 -0700924 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
Minchan Kim312fcae2015-04-15 16:15:30 -0700925}
926
927static void pin_tag(unsigned long handle)
928{
Minchan Kim1b8320b2016-07-26 15:23:14 -0700929 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
Minchan Kim312fcae2015-04-15 16:15:30 -0700930}
931
932static void unpin_tag(unsigned long handle)
933{
Minchan Kim1b8320b2016-07-26 15:23:14 -0700934 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
Minchan Kim312fcae2015-04-15 16:15:30 -0700935}
936
Nitin Guptaf4477e92012-04-02 09:13:56 -0500937static void reset_page(struct page *page)
938{
Minchan Kim48b48002016-07-26 15:23:31 -0700939 __ClearPageMovable(page);
Ganesh Mahendran18fd06b2016-07-28 15:48:00 -0700940 ClearPagePrivate(page);
941 ClearPagePrivate2(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500942 set_page_private(page, 0);
Minchan Kim48b48002016-07-26 15:23:31 -0700943 page_mapcount_reset(page);
944 ClearPageHugeObject(page);
945 page->freelist = NULL;
Nitin Guptaf4477e92012-04-02 09:13:56 -0500946}
947
Minchan Kim48b48002016-07-26 15:23:31 -0700948/*
949 * To prevent zspage destroy during migration, zspage freeing should
950 * hold locks of all pages in the zspage.
951 */
952void lock_zspage(struct zspage *zspage)
953{
954 struct page *page = get_first_page(zspage);
955
956 do {
957 lock_page(page);
958 } while ((page = get_next_page(page)) != NULL);
959}
960
961int trylock_zspage(struct zspage *zspage)
962{
963 struct page *cursor, *fail;
964
965 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
966 get_next_page(cursor)) {
967 if (!trylock_page(cursor)) {
968 fail = cursor;
969 goto unlock;
970 }
971 }
972
973 return 1;
974unlock:
975 for (cursor = get_first_page(zspage); cursor != fail; cursor =
976 get_next_page(cursor))
977 unlock_page(cursor);
978
979 return 0;
980}
981
982static void __free_zspage(struct zs_pool *pool, struct size_class *class,
983 struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -0600984{
Minchan Kim37836892016-07-26 15:23:23 -0700985 struct page *page, *next;
Minchan Kim48b48002016-07-26 15:23:31 -0700986 enum fullness_group fg;
987 unsigned int class_idx;
988
989 get_zspage_mapping(zspage, &class_idx, &fg);
990
991 assert_spin_locked(&class->lock);
Nitin Gupta61989a82012-01-09 16:51:56 -0600992
Minchan Kim37836892016-07-26 15:23:23 -0700993 VM_BUG_ON(get_zspage_inuse(zspage));
Minchan Kim48b48002016-07-26 15:23:31 -0700994 VM_BUG_ON(fg != ZS_EMPTY);
Nitin Gupta61989a82012-01-09 16:51:56 -0600995
Minchan Kim48b48002016-07-26 15:23:31 -0700996 next = page = get_first_page(zspage);
Minchan Kim37836892016-07-26 15:23:23 -0700997 do {
Minchan Kim48b48002016-07-26 15:23:31 -0700998 VM_BUG_ON_PAGE(!PageLocked(page), page);
999 next = get_next_page(page);
Minchan Kim37836892016-07-26 15:23:23 -07001000 reset_page(page);
Minchan Kim48b48002016-07-26 15:23:31 -07001001 unlock_page(page);
Minchan Kim91537fe2016-07-26 15:24:45 -07001002 dec_zone_page_state(page, NR_ZSPAGES);
Minchan Kim37836892016-07-26 15:23:23 -07001003 put_page(page);
1004 page = next;
1005 } while (page != NULL);
Nitin Gupta61989a82012-01-09 16:51:56 -06001006
Minchan Kim37836892016-07-26 15:23:23 -07001007 cache_free_zspage(pool, zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001008
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -07001009 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001010 atomic_long_sub(class->pages_per_zspage,
1011 &pool->pages_allocated);
1012}
1013
1014static void free_zspage(struct zs_pool *pool, struct size_class *class,
1015 struct zspage *zspage)
1016{
1017 VM_BUG_ON(get_zspage_inuse(zspage));
1018 VM_BUG_ON(list_empty(&zspage->list));
1019
1020 if (!trylock_zspage(zspage)) {
1021 kick_deferred_free(pool);
1022 return;
1023 }
1024
1025 remove_zspage(class, zspage, ZS_EMPTY);
1026 __free_zspage(pool, class, zspage);
Nitin Gupta61989a82012-01-09 16:51:56 -06001027}
1028
1029/* Initialize a newly allocated zspage */
Minchan Kim37836892016-07-26 15:23:23 -07001030static void init_zspage(struct size_class *class, struct zspage *zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -06001031{
Minchan Kimbfd093f2016-07-26 15:23:28 -07001032 unsigned int freeobj = 1;
Nitin Gupta61989a82012-01-09 16:51:56 -06001033 unsigned long off = 0;
Minchan Kim48b48002016-07-26 15:23:31 -07001034 struct page *page = get_first_page(zspage);
Minchan Kim830e4bc2016-05-20 16:59:39 -07001035
Nitin Gupta61989a82012-01-09 16:51:56 -06001036 while (page) {
1037 struct page *next_page;
1038 struct link_free *link;
Minchan Kimaf4ee5e2014-12-12 16:56:58 -08001039 void *vaddr;
Nitin Gupta61989a82012-01-09 16:51:56 -06001040
Minchan Kim37836892016-07-26 15:23:23 -07001041 set_first_obj_offset(page, off);
Nitin Gupta61989a82012-01-09 16:51:56 -06001042
Minchan Kimaf4ee5e2014-12-12 16:56:58 -08001043 vaddr = kmap_atomic(page);
1044 link = (struct link_free *)vaddr + off / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -06001045
Dan Streetman5538c562014-10-09 15:30:01 -07001046 while ((off += class->size) < PAGE_SIZE) {
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001047 link->next = freeobj++ << OBJ_TAG_BITS;
Dan Streetman5538c562014-10-09 15:30:01 -07001048 link += class->size / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -06001049 }
1050
1051 /*
1052 * We now come to the last (full or partial) object on this
1053 * page, which must point to the first object on the next
1054 * page (if present)
1055 */
1056 next_page = get_next_page(page);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001057 if (next_page) {
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001058 link->next = freeobj++ << OBJ_TAG_BITS;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001059 } else {
1060 /*
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001061 * Reset OBJ_TAG_BITS bit to last link to tell
Minchan Kimbfd093f2016-07-26 15:23:28 -07001062 * whether it's allocated object or not.
1063 */
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001064 link->next = -1 << OBJ_TAG_BITS;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001065 }
Minchan Kimaf4ee5e2014-12-12 16:56:58 -08001066 kunmap_atomic(vaddr);
Nitin Gupta61989a82012-01-09 16:51:56 -06001067 page = next_page;
Dan Streetman5538c562014-10-09 15:30:01 -07001068 off %= PAGE_SIZE;
Nitin Gupta61989a82012-01-09 16:51:56 -06001069 }
Minchan Kimbdb0af72016-07-26 15:23:20 -07001070
Minchan Kimbfd093f2016-07-26 15:23:28 -07001071 set_freeobj(zspage, 0);
Nitin Gupta61989a82012-01-09 16:51:56 -06001072}
1073
Minchan Kim48b48002016-07-26 15:23:31 -07001074static void create_page_chain(struct size_class *class, struct zspage *zspage,
1075 struct page *pages[])
Nitin Gupta61989a82012-01-09 16:51:56 -06001076{
Minchan Kimbdb0af72016-07-26 15:23:20 -07001077 int i;
1078 struct page *page;
1079 struct page *prev_page = NULL;
Minchan Kim48b48002016-07-26 15:23:31 -07001080 int nr_pages = class->pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001081
1082 /*
1083 * Allocate individual pages and link them together as:
Minchan Kim48b48002016-07-26 15:23:31 -07001084 * 1. all pages are linked together using page->freelist
Minchan Kim37836892016-07-26 15:23:23 -07001085 * 2. each sub-page point to zspage using page->private
Nitin Gupta61989a82012-01-09 16:51:56 -06001086 *
Minchan Kim37836892016-07-26 15:23:23 -07001087 * we set PG_private to identify the first page (i.e. no other sub-page
1088 * has this flag set) and PG_private_2 to identify the last page.
Nitin Gupta61989a82012-01-09 16:51:56 -06001089 */
Minchan Kimbdb0af72016-07-26 15:23:20 -07001090 for (i = 0; i < nr_pages; i++) {
1091 page = pages[i];
Minchan Kim37836892016-07-26 15:23:23 -07001092 set_page_private(page, (unsigned long)zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001093 page->freelist = NULL;
Minchan Kimbdb0af72016-07-26 15:23:20 -07001094 if (i == 0) {
Minchan Kim37836892016-07-26 15:23:23 -07001095 zspage->first_page = page;
Minchan Kima27545bf2012-04-25 15:23:09 +09001096 SetPagePrivate(page);
Minchan Kim48b48002016-07-26 15:23:31 -07001097 if (unlikely(class->objs_per_zspage == 1 &&
1098 class->pages_per_zspage == 1))
1099 SetPageHugeObject(page);
Minchan Kim37836892016-07-26 15:23:23 -07001100 } else {
Minchan Kim48b48002016-07-26 15:23:31 -07001101 prev_page->freelist = page;
Nitin Gupta61989a82012-01-09 16:51:56 -06001102 }
Minchan Kim48b48002016-07-26 15:23:31 -07001103 if (i == nr_pages - 1)
Minchan Kima27545bf2012-04-25 15:23:09 +09001104 SetPagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -06001105 prev_page = page;
1106 }
Minchan Kimbdb0af72016-07-26 15:23:20 -07001107}
Nitin Gupta61989a82012-01-09 16:51:56 -06001108
Minchan Kimbdb0af72016-07-26 15:23:20 -07001109/*
1110 * Allocate a zspage for the given size class
1111 */
Minchan Kim37836892016-07-26 15:23:23 -07001112static struct zspage *alloc_zspage(struct zs_pool *pool,
1113 struct size_class *class,
1114 gfp_t gfp)
Minchan Kimbdb0af72016-07-26 15:23:20 -07001115{
1116 int i;
Minchan Kimbdb0af72016-07-26 15:23:20 -07001117 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
Minchan Kim37836892016-07-26 15:23:23 -07001118 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1119
1120 if (!zspage)
1121 return NULL;
1122
1123 memset(zspage, 0, sizeof(struct zspage));
Minchan Kim48b48002016-07-26 15:23:31 -07001124 zspage->magic = ZSPAGE_MAGIC;
1125 migrate_lock_init(zspage);
Nitin Gupta61989a82012-01-09 16:51:56 -06001126
Minchan Kimbdb0af72016-07-26 15:23:20 -07001127 for (i = 0; i < class->pages_per_zspage; i++) {
1128 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -06001129
Minchan Kim37836892016-07-26 15:23:23 -07001130 page = alloc_page(gfp);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001131 if (!page) {
Minchan Kim91537fe2016-07-26 15:24:45 -07001132 while (--i >= 0) {
1133 dec_zone_page_state(pages[i], NR_ZSPAGES);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001134 __free_page(pages[i]);
Minchan Kim91537fe2016-07-26 15:24:45 -07001135 }
Minchan Kim37836892016-07-26 15:23:23 -07001136 cache_free_zspage(pool, zspage);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001137 return NULL;
1138 }
Minchan Kim91537fe2016-07-26 15:24:45 -07001139
1140 inc_zone_page_state(page, NR_ZSPAGES);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001141 pages[i] = page;
Nitin Gupta61989a82012-01-09 16:51:56 -06001142 }
1143
Minchan Kim48b48002016-07-26 15:23:31 -07001144 create_page_chain(class, zspage, pages);
Minchan Kim37836892016-07-26 15:23:23 -07001145 init_zspage(class, zspage);
Minchan Kimbdb0af72016-07-26 15:23:20 -07001146
Minchan Kim37836892016-07-26 15:23:23 -07001147 return zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001148}
1149
Minchan Kim37836892016-07-26 15:23:23 -07001150static struct zspage *find_get_zspage(struct size_class *class)
Nitin Gupta61989a82012-01-09 16:51:56 -06001151{
1152 int i;
Minchan Kim37836892016-07-26 15:23:23 -07001153 struct zspage *zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001154
Minchan Kim48b48002016-07-26 15:23:31 -07001155 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
Minchan Kim37836892016-07-26 15:23:23 -07001156 zspage = list_first_entry_or_null(&class->fullness_list[i],
1157 struct zspage, list);
1158 if (zspage)
Nitin Gupta61989a82012-01-09 16:51:56 -06001159 break;
1160 }
1161
Minchan Kim37836892016-07-26 15:23:23 -07001162 return zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001163}
1164
Minchan Kim1b945ae2013-12-11 11:04:36 +09001165#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -05001166static inline int __zs_cpu_up(struct mapping_area *area)
Seth Jennings5f601902012-07-02 16:15:49 -05001167{
Seth Jenningsf5536462012-07-18 11:55:56 -05001168 /*
1169 * Make sure we don't leak memory if a cpu UP notification
1170 * and zs_init() race and both call zs_cpu_up() on the same cpu
1171 */
1172 if (area->vm)
1173 return 0;
1174 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1175 if (!area->vm)
1176 return -ENOMEM;
1177 return 0;
1178}
1179
1180static inline void __zs_cpu_down(struct mapping_area *area)
1181{
1182 if (area->vm)
1183 free_vm_area(area->vm);
1184 area->vm = NULL;
1185}
1186
1187static inline void *__zs_map_object(struct mapping_area *area,
1188 struct page *pages[2], int off, int size)
1189{
WANG Chaof6f8ed42014-08-06 16:06:58 -07001190 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
Seth Jenningsf5536462012-07-18 11:55:56 -05001191 area->vm_addr = area->vm->addr;
1192 return area->vm_addr + off;
1193}
1194
1195static inline void __zs_unmap_object(struct mapping_area *area,
1196 struct page *pages[2], int off, int size)
1197{
1198 unsigned long addr = (unsigned long)area->vm_addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001199
Joerg Roedeld95abbb2013-03-27 01:43:14 +01001200 unmap_kernel_range(addr, PAGE_SIZE * 2);
Seth Jenningsf5536462012-07-18 11:55:56 -05001201}
1202
Minchan Kim1b945ae2013-12-11 11:04:36 +09001203#else /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001204
1205static inline int __zs_cpu_up(struct mapping_area *area)
1206{
1207 /*
1208 * Make sure we don't leak memory if a cpu UP notification
1209 * and zs_init() race and both call zs_cpu_up() on the same cpu
1210 */
1211 if (area->vm_buf)
1212 return 0;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001213 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
Seth Jenningsf5536462012-07-18 11:55:56 -05001214 if (!area->vm_buf)
1215 return -ENOMEM;
1216 return 0;
1217}
1218
1219static inline void __zs_cpu_down(struct mapping_area *area)
1220{
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001221 kfree(area->vm_buf);
Seth Jenningsf5536462012-07-18 11:55:56 -05001222 area->vm_buf = NULL;
1223}
1224
1225static void *__zs_map_object(struct mapping_area *area,
1226 struct page *pages[2], int off, int size)
1227{
Seth Jennings5f601902012-07-02 16:15:49 -05001228 int sizes[2];
1229 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001230 char *buf = area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001231
Seth Jenningsf5536462012-07-18 11:55:56 -05001232 /* disable page faults to match kmap_atomic() return conditions */
1233 pagefault_disable();
1234
1235 /* no read fastpath */
1236 if (area->vm_mm == ZS_MM_WO)
1237 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001238
1239 sizes[0] = PAGE_SIZE - off;
1240 sizes[1] = size - sizes[0];
1241
Seth Jennings5f601902012-07-02 16:15:49 -05001242 /* copy object to per-cpu buffer */
1243 addr = kmap_atomic(pages[0]);
1244 memcpy(buf, addr + off, sizes[0]);
1245 kunmap_atomic(addr);
1246 addr = kmap_atomic(pages[1]);
1247 memcpy(buf + sizes[0], addr, sizes[1]);
1248 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001249out:
1250 return area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001251}
1252
Seth Jenningsf5536462012-07-18 11:55:56 -05001253static void __zs_unmap_object(struct mapping_area *area,
1254 struct page *pages[2], int off, int size)
Seth Jennings5f601902012-07-02 16:15:49 -05001255{
Seth Jennings5f601902012-07-02 16:15:49 -05001256 int sizes[2];
1257 void *addr;
Minchan Kim2e40e162015-04-15 16:15:23 -07001258 char *buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001259
Seth Jenningsf5536462012-07-18 11:55:56 -05001260 /* no write fastpath */
1261 if (area->vm_mm == ZS_MM_RO)
1262 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001263
Minchan Kim7b60a682015-04-15 16:15:39 -07001264 buf = area->vm_buf;
YiPing Xua82cbf02016-03-17 14:20:39 -07001265 buf = buf + ZS_HANDLE_SIZE;
1266 size -= ZS_HANDLE_SIZE;
1267 off += ZS_HANDLE_SIZE;
Minchan Kim2e40e162015-04-15 16:15:23 -07001268
Seth Jennings5f601902012-07-02 16:15:49 -05001269 sizes[0] = PAGE_SIZE - off;
1270 sizes[1] = size - sizes[0];
1271
1272 /* copy per-cpu buffer to object */
1273 addr = kmap_atomic(pages[0]);
1274 memcpy(addr + off, buf, sizes[0]);
1275 kunmap_atomic(addr);
1276 addr = kmap_atomic(pages[1]);
1277 memcpy(addr, buf + sizes[0], sizes[1]);
1278 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001279
1280out:
1281 /* enable page faults to match kunmap_atomic() return conditions */
1282 pagefault_enable();
Seth Jennings5f601902012-07-02 16:15:49 -05001283}
Nitin Gupta61989a82012-01-09 16:51:56 -06001284
Minchan Kim1b945ae2013-12-11 11:04:36 +09001285#endif /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001286
Nitin Gupta61989a82012-01-09 16:51:56 -06001287static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1288 void *pcpu)
1289{
Seth Jenningsf5536462012-07-18 11:55:56 -05001290 int ret, cpu = (long)pcpu;
Nitin Gupta61989a82012-01-09 16:51:56 -06001291 struct mapping_area *area;
1292
1293 switch (action) {
1294 case CPU_UP_PREPARE:
1295 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001296 ret = __zs_cpu_up(area);
1297 if (ret)
1298 return notifier_from_errno(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001299 break;
1300 case CPU_DEAD:
1301 case CPU_UP_CANCELED:
1302 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001303 __zs_cpu_down(area);
Nitin Gupta61989a82012-01-09 16:51:56 -06001304 break;
1305 }
1306
1307 return NOTIFY_OK;
1308}
1309
1310static struct notifier_block zs_cpu_nb = {
1311 .notifier_call = zs_cpu_notifier
1312};
1313
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001314static int zs_register_cpu_notifier(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001315{
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001316 int cpu, uninitialized_var(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001317
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301318 cpu_notifier_register_begin();
1319
1320 __register_cpu_notifier(&zs_cpu_nb);
Nitin Gupta61989a82012-01-09 16:51:56 -06001321 for_each_online_cpu(cpu) {
1322 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001323 if (notifier_to_errno(ret))
1324 break;
Nitin Gupta61989a82012-01-09 16:51:56 -06001325 }
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301326
1327 cpu_notifier_register_done();
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001328 return notifier_to_errno(ret);
1329}
1330
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001331static void zs_unregister_cpu_notifier(void)
1332{
1333 int cpu;
1334
1335 cpu_notifier_register_begin();
1336
1337 for_each_online_cpu(cpu)
1338 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1339 __unregister_cpu_notifier(&zs_cpu_nb);
1340
1341 cpu_notifier_register_done();
1342}
1343
Ganesh Mahendran35b34452016-07-28 15:47:57 -07001344static void __init init_zs_size_classes(void)
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001345{
1346 int nr;
1347
1348 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1349 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1350 nr += 1;
1351
1352 zs_size_classes = nr;
1353}
1354
Ganesh Mahendran64d90462016-07-28 15:47:51 -07001355static bool can_merge(struct size_class *prev, int pages_per_zspage,
1356 int objs_per_zspage)
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001357{
Ganesh Mahendran64d90462016-07-28 15:47:51 -07001358 if (prev->pages_per_zspage == pages_per_zspage &&
1359 prev->objs_per_zspage == objs_per_zspage)
1360 return true;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001361
Ganesh Mahendran64d90462016-07-28 15:47:51 -07001362 return false;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001363}
1364
Minchan Kim37836892016-07-26 15:23:23 -07001365static bool zspage_full(struct size_class *class, struct zspage *zspage)
Minchan Kim312fcae2015-04-15 16:15:30 -07001366{
Minchan Kim37836892016-07-26 15:23:23 -07001367 return get_zspage_inuse(zspage) == class->objs_per_zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001368}
1369
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001370unsigned long zs_get_total_pages(struct zs_pool *pool)
1371{
1372 return atomic_long_read(&pool->pages_allocated);
1373}
1374EXPORT_SYMBOL_GPL(zs_get_total_pages);
1375
1376/**
1377 * zs_map_object - get address of allocated object from handle.
1378 * @pool: pool from which the object was allocated
1379 * @handle: handle returned from zs_malloc
1380 *
1381 * Before using an object allocated from zs_malloc, it must be mapped using
1382 * this function. When done with the object, it must be unmapped using
1383 * zs_unmap_object.
1384 *
1385 * Only one object can be mapped per cpu at a time. There is no protection
1386 * against nested mappings.
1387 *
1388 * This function returns with preemption and page faults disabled.
1389 */
1390void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1391 enum zs_mapmode mm)
1392{
Minchan Kim37836892016-07-26 15:23:23 -07001393 struct zspage *zspage;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001394 struct page *page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001395 unsigned long obj, off;
1396 unsigned int obj_idx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001397
1398 unsigned int class_idx;
1399 enum fullness_group fg;
1400 struct size_class *class;
1401 struct mapping_area *area;
1402 struct page *pages[2];
Minchan Kim2e40e162015-04-15 16:15:23 -07001403 void *ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001404
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001405 /*
1406 * Because we use per-cpu mapping areas shared among the
1407 * pools/users, we can't allow mapping in interrupt context
1408 * because it can corrupt another users mappings.
1409 */
Minchan Kim830e4bc2016-05-20 16:59:39 -07001410 WARN_ON_ONCE(in_interrupt());
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001411
Minchan Kim312fcae2015-04-15 16:15:30 -07001412 /* From now on, migration cannot move the object */
1413 pin_tag(handle);
1414
Minchan Kim2e40e162015-04-15 16:15:23 -07001415 obj = handle_to_obj(handle);
1416 obj_to_location(obj, &page, &obj_idx);
Minchan Kim37836892016-07-26 15:23:23 -07001417 zspage = get_zspage(page);
Minchan Kim48b48002016-07-26 15:23:31 -07001418
1419 /* migration cannot move any subpage in this zspage */
1420 migrate_read_lock(zspage);
1421
Minchan Kim37836892016-07-26 15:23:23 -07001422 get_zspage_mapping(zspage, &class_idx, &fg);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001423 class = pool->size_class[class_idx];
Minchan Kimbfd093f2016-07-26 15:23:28 -07001424 off = (class->size * obj_idx) & ~PAGE_MASK;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001425
1426 area = &get_cpu_var(zs_map_area);
1427 area->vm_mm = mm;
1428 if (off + class->size <= PAGE_SIZE) {
1429 /* this object is contained entirely within a page */
1430 area->vm_addr = kmap_atomic(page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001431 ret = area->vm_addr + off;
1432 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001433 }
1434
1435 /* this object spans two pages */
1436 pages[0] = page;
1437 pages[1] = get_next_page(page);
1438 BUG_ON(!pages[1]);
1439
Minchan Kim2e40e162015-04-15 16:15:23 -07001440 ret = __zs_map_object(area, pages, off, class->size);
1441out:
Minchan Kim48b48002016-07-26 15:23:31 -07001442 if (likely(!PageHugeObject(page)))
Minchan Kim7b60a682015-04-15 16:15:39 -07001443 ret += ZS_HANDLE_SIZE;
1444
1445 return ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001446}
1447EXPORT_SYMBOL_GPL(zs_map_object);
1448
1449void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1450{
Minchan Kim37836892016-07-26 15:23:23 -07001451 struct zspage *zspage;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001452 struct page *page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001453 unsigned long obj, off;
1454 unsigned int obj_idx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001455
1456 unsigned int class_idx;
1457 enum fullness_group fg;
1458 struct size_class *class;
1459 struct mapping_area *area;
1460
Minchan Kim2e40e162015-04-15 16:15:23 -07001461 obj = handle_to_obj(handle);
1462 obj_to_location(obj, &page, &obj_idx);
Minchan Kim37836892016-07-26 15:23:23 -07001463 zspage = get_zspage(page);
1464 get_zspage_mapping(zspage, &class_idx, &fg);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001465 class = pool->size_class[class_idx];
Minchan Kimbfd093f2016-07-26 15:23:28 -07001466 off = (class->size * obj_idx) & ~PAGE_MASK;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001467
1468 area = this_cpu_ptr(&zs_map_area);
1469 if (off + class->size <= PAGE_SIZE)
1470 kunmap_atomic(area->vm_addr);
1471 else {
1472 struct page *pages[2];
1473
1474 pages[0] = page;
1475 pages[1] = get_next_page(page);
1476 BUG_ON(!pages[1]);
1477
1478 __zs_unmap_object(area, pages, off, class->size);
1479 }
1480 put_cpu_var(zs_map_area);
Minchan Kim48b48002016-07-26 15:23:31 -07001481
1482 migrate_read_unlock(zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07001483 unpin_tag(handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001484}
1485EXPORT_SYMBOL_GPL(zs_unmap_object);
1486
Minchan Kim251cbb92016-05-20 16:59:42 -07001487static unsigned long obj_malloc(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -07001488 struct zspage *zspage, unsigned long handle)
Minchan Kimc7806262015-04-15 16:15:26 -07001489{
Minchan Kimbfd093f2016-07-26 15:23:28 -07001490 int i, nr_page, offset;
Minchan Kimc7806262015-04-15 16:15:26 -07001491 unsigned long obj;
1492 struct link_free *link;
1493
1494 struct page *m_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001495 unsigned long m_offset;
Minchan Kimc7806262015-04-15 16:15:26 -07001496 void *vaddr;
1497
Minchan Kim312fcae2015-04-15 16:15:30 -07001498 handle |= OBJ_ALLOCATED_TAG;
Minchan Kim37836892016-07-26 15:23:23 -07001499 obj = get_freeobj(zspage);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001500
1501 offset = obj * class->size;
1502 nr_page = offset >> PAGE_SHIFT;
1503 m_offset = offset & ~PAGE_MASK;
1504 m_page = get_first_page(zspage);
1505
1506 for (i = 0; i < nr_page; i++)
1507 m_page = get_next_page(m_page);
Minchan Kimc7806262015-04-15 16:15:26 -07001508
1509 vaddr = kmap_atomic(m_page);
1510 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001511 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
Minchan Kim48b48002016-07-26 15:23:31 -07001512 if (likely(!PageHugeObject(m_page)))
Minchan Kim7b60a682015-04-15 16:15:39 -07001513 /* record handle in the header of allocated chunk */
1514 link->handle = handle;
1515 else
Minchan Kim37836892016-07-26 15:23:23 -07001516 /* record handle to page->index */
1517 zspage->first_page->index = handle;
1518
Minchan Kimc7806262015-04-15 16:15:26 -07001519 kunmap_atomic(vaddr);
Minchan Kim37836892016-07-26 15:23:23 -07001520 mod_zspage_inuse(zspage, 1);
Minchan Kimc7806262015-04-15 16:15:26 -07001521 zs_stat_inc(class, OBJ_USED, 1);
1522
Minchan Kimbfd093f2016-07-26 15:23:28 -07001523 obj = location_to_obj(m_page, obj);
1524
Minchan Kimc7806262015-04-15 16:15:26 -07001525 return obj;
1526}
1527
1528
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001529/**
1530 * zs_malloc - Allocate block of given size from pool.
1531 * @pool: pool to allocate from
1532 * @size: size of block to allocate
Ganesh Mahendranfd854462016-07-28 15:47:54 -07001533 * @gfp: gfp flags when allocating object
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001534 *
1535 * On success, handle to the allocated object is returned,
1536 * otherwise 0.
1537 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1538 */
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -07001539unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001540{
Minchan Kim2e40e162015-04-15 16:15:23 -07001541 unsigned long handle, obj;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001542 struct size_class *class;
Minchan Kim48b48002016-07-26 15:23:31 -07001543 enum fullness_group newfg;
Minchan Kim37836892016-07-26 15:23:23 -07001544 struct zspage *zspage;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001545
Minchan Kim7b60a682015-04-15 16:15:39 -07001546 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001547 return 0;
1548
Minchan Kim37836892016-07-26 15:23:23 -07001549 handle = cache_alloc_handle(pool, gfp);
Minchan Kim2e40e162015-04-15 16:15:23 -07001550 if (!handle)
1551 return 0;
1552
1553 /* extra space in chunk to keep the handle */
1554 size += ZS_HANDLE_SIZE;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001555 class = pool->size_class[get_size_class_index(size)];
1556
1557 spin_lock(&class->lock);
Minchan Kim37836892016-07-26 15:23:23 -07001558 zspage = find_get_zspage(class);
Minchan Kim48b48002016-07-26 15:23:31 -07001559 if (likely(zspage)) {
1560 obj = obj_malloc(class, zspage, handle);
1561 /* Now move the zspage to another fullness group, if required */
1562 fix_fullness_group(class, zspage);
1563 record_obj(handle, obj);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001564 spin_unlock(&class->lock);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001565
Minchan Kim48b48002016-07-26 15:23:31 -07001566 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001567 }
1568
Minchan Kim48b48002016-07-26 15:23:31 -07001569 spin_unlock(&class->lock);
1570
1571 zspage = alloc_zspage(pool, class, gfp);
1572 if (!zspage) {
1573 cache_free_handle(pool, handle);
1574 return 0;
1575 }
1576
1577 spin_lock(&class->lock);
Minchan Kim37836892016-07-26 15:23:23 -07001578 obj = obj_malloc(class, zspage, handle);
Minchan Kim48b48002016-07-26 15:23:31 -07001579 newfg = get_fullness_group(class, zspage);
1580 insert_zspage(class, zspage, newfg);
1581 set_zspage_mapping(zspage, class->index, newfg);
Minchan Kim2e40e162015-04-15 16:15:23 -07001582 record_obj(handle, obj);
Minchan Kim48b48002016-07-26 15:23:31 -07001583 atomic_long_add(class->pages_per_zspage,
1584 &pool->pages_allocated);
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -07001585 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001586
1587 /* We completely set up zspage so mark them as movable */
1588 SetZsPageMovable(pool, zspage);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001589 spin_unlock(&class->lock);
1590
Minchan Kim2e40e162015-04-15 16:15:23 -07001591 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001592}
1593EXPORT_SYMBOL_GPL(zs_malloc);
1594
Minchan Kim1ee47162016-05-20 16:59:45 -07001595static void obj_free(struct size_class *class, unsigned long obj)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001596{
1597 struct link_free *link;
Minchan Kim37836892016-07-26 15:23:23 -07001598 struct zspage *zspage;
1599 struct page *f_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001600 unsigned long f_offset;
1601 unsigned int f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001602 void *vaddr;
1603
Minchan Kim312fcae2015-04-15 16:15:30 -07001604 obj &= ~OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001605 obj_to_location(obj, &f_page, &f_objidx);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001606 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
Minchan Kim37836892016-07-26 15:23:23 -07001607 zspage = get_zspage(f_page);
Minchan Kimc7806262015-04-15 16:15:26 -07001608
Minchan Kimc7806262015-04-15 16:15:26 -07001609 vaddr = kmap_atomic(f_page);
1610
1611 /* Insert this object in containing zspage's freelist */
1612 link = (struct link_free *)(vaddr + f_offset);
Minchan Kim3b1d9ca2016-07-26 15:23:37 -07001613 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
Minchan Kimc7806262015-04-15 16:15:26 -07001614 kunmap_atomic(vaddr);
Minchan Kimbfd093f2016-07-26 15:23:28 -07001615 set_freeobj(zspage, f_objidx);
Minchan Kim37836892016-07-26 15:23:23 -07001616 mod_zspage_inuse(zspage, -1);
Minchan Kimc7806262015-04-15 16:15:26 -07001617 zs_stat_dec(class, OBJ_USED, 1);
1618}
1619
1620void zs_free(struct zs_pool *pool, unsigned long handle)
1621{
Minchan Kim37836892016-07-26 15:23:23 -07001622 struct zspage *zspage;
1623 struct page *f_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001624 unsigned long obj;
1625 unsigned int f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001626 int class_idx;
1627 struct size_class *class;
1628 enum fullness_group fullness;
Minchan Kim48b48002016-07-26 15:23:31 -07001629 bool isolated;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001630
Minchan Kim2e40e162015-04-15 16:15:23 -07001631 if (unlikely(!handle))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001632 return;
1633
Minchan Kim312fcae2015-04-15 16:15:30 -07001634 pin_tag(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001635 obj = handle_to_obj(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001636 obj_to_location(obj, &f_page, &f_objidx);
Minchan Kim37836892016-07-26 15:23:23 -07001637 zspage = get_zspage(f_page);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001638
Minchan Kim48b48002016-07-26 15:23:31 -07001639 migrate_read_lock(zspage);
1640
Minchan Kim37836892016-07-26 15:23:23 -07001641 get_zspage_mapping(zspage, &class_idx, &fullness);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001642 class = pool->size_class[class_idx];
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001643
1644 spin_lock(&class->lock);
Minchan Kim1ee47162016-05-20 16:59:45 -07001645 obj_free(class, obj);
Minchan Kim37836892016-07-26 15:23:23 -07001646 fullness = fix_fullness_group(class, zspage);
Minchan Kim48b48002016-07-26 15:23:31 -07001647 if (fullness != ZS_EMPTY) {
1648 migrate_read_unlock(zspage);
1649 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001650 }
Minchan Kim48b48002016-07-26 15:23:31 -07001651
1652 isolated = is_zspage_isolated(zspage);
1653 migrate_read_unlock(zspage);
1654 /* If zspage is isolated, zs_page_putback will free the zspage */
1655 if (likely(!isolated))
1656 free_zspage(pool, class, zspage);
1657out:
1658
Minchan Kim312fcae2015-04-15 16:15:30 -07001659 spin_unlock(&class->lock);
1660 unpin_tag(handle);
Minchan Kim37836892016-07-26 15:23:23 -07001661 cache_free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001662}
1663EXPORT_SYMBOL_GPL(zs_free);
1664
Minchan Kim251cbb92016-05-20 16:59:42 -07001665static void zs_object_copy(struct size_class *class, unsigned long dst,
1666 unsigned long src)
Minchan Kim312fcae2015-04-15 16:15:30 -07001667{
1668 struct page *s_page, *d_page;
Minchan Kimbfd093f2016-07-26 15:23:28 -07001669 unsigned int s_objidx, d_objidx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001670 unsigned long s_off, d_off;
1671 void *s_addr, *d_addr;
1672 int s_size, d_size, size;
1673 int written = 0;
1674
1675 s_size = d_size = class->size;
1676
1677 obj_to_location(src, &s_page, &s_objidx);
1678 obj_to_location(dst, &d_page, &d_objidx);
1679
Minchan Kimbfd093f2016-07-26 15:23:28 -07001680 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1681 d_off = (class->size * d_objidx) & ~PAGE_MASK;
Minchan Kim312fcae2015-04-15 16:15:30 -07001682
1683 if (s_off + class->size > PAGE_SIZE)
1684 s_size = PAGE_SIZE - s_off;
1685
1686 if (d_off + class->size > PAGE_SIZE)
1687 d_size = PAGE_SIZE - d_off;
1688
1689 s_addr = kmap_atomic(s_page);
1690 d_addr = kmap_atomic(d_page);
1691
1692 while (1) {
1693 size = min(s_size, d_size);
1694 memcpy(d_addr + d_off, s_addr + s_off, size);
1695 written += size;
1696
1697 if (written == class->size)
1698 break;
1699
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001700 s_off += size;
1701 s_size -= size;
1702 d_off += size;
1703 d_size -= size;
1704
1705 if (s_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001706 kunmap_atomic(d_addr);
1707 kunmap_atomic(s_addr);
1708 s_page = get_next_page(s_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001709 s_addr = kmap_atomic(s_page);
1710 d_addr = kmap_atomic(d_page);
1711 s_size = class->size - written;
1712 s_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001713 }
1714
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001715 if (d_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001716 kunmap_atomic(d_addr);
1717 d_page = get_next_page(d_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001718 d_addr = kmap_atomic(d_page);
1719 d_size = class->size - written;
1720 d_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001721 }
1722 }
1723
1724 kunmap_atomic(d_addr);
1725 kunmap_atomic(s_addr);
1726}
1727
1728/*
1729 * Find alloced object in zspage from index object and
1730 * return handle.
1731 */
Minchan Kim251cbb92016-05-20 16:59:42 -07001732static unsigned long find_alloced_obj(struct size_class *class,
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001733 struct page *page, int *obj_idx)
Minchan Kim312fcae2015-04-15 16:15:30 -07001734{
1735 unsigned long head;
1736 int offset = 0;
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001737 int index = *obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001738 unsigned long handle = 0;
1739 void *addr = kmap_atomic(page);
1740
Minchan Kim37836892016-07-26 15:23:23 -07001741 offset = get_first_obj_offset(page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001742 offset += class->size * index;
1743
1744 while (offset < PAGE_SIZE) {
Minchan Kim48b48002016-07-26 15:23:31 -07001745 head = obj_to_head(page, addr + offset);
Minchan Kim312fcae2015-04-15 16:15:30 -07001746 if (head & OBJ_ALLOCATED_TAG) {
1747 handle = head & ~OBJ_ALLOCATED_TAG;
1748 if (trypin_tag(handle))
1749 break;
1750 handle = 0;
1751 }
1752
1753 offset += class->size;
1754 index++;
1755 }
1756
1757 kunmap_atomic(addr);
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001758
1759 *obj_idx = index;
1760
Minchan Kim312fcae2015-04-15 16:15:30 -07001761 return handle;
1762}
1763
1764struct zs_compact_control {
Minchan Kim37836892016-07-26 15:23:23 -07001765 /* Source spage for migration which could be a subpage of zspage */
Minchan Kim312fcae2015-04-15 16:15:30 -07001766 struct page *s_page;
1767 /* Destination page for migration which should be a first page
1768 * of zspage. */
1769 struct page *d_page;
1770 /* Starting object index within @s_page which used for live object
1771 * in the subpage. */
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001772 int obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001773};
1774
1775static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1776 struct zs_compact_control *cc)
1777{
1778 unsigned long used_obj, free_obj;
1779 unsigned long handle;
1780 struct page *s_page = cc->s_page;
1781 struct page *d_page = cc->d_page;
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001782 int obj_idx = cc->obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001783 int ret = 0;
1784
1785 while (1) {
Ganesh Mahendrancf675ac2016-07-28 15:47:46 -07001786 handle = find_alloced_obj(class, s_page, &obj_idx);
Minchan Kim312fcae2015-04-15 16:15:30 -07001787 if (!handle) {
1788 s_page = get_next_page(s_page);
1789 if (!s_page)
1790 break;
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001791 obj_idx = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001792 continue;
1793 }
1794
1795 /* Stop if there is no more space */
Minchan Kim37836892016-07-26 15:23:23 -07001796 if (zspage_full(class, get_zspage(d_page))) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001797 unpin_tag(handle);
1798 ret = -ENOMEM;
1799 break;
1800 }
1801
1802 used_obj = handle_to_obj(handle);
Minchan Kim37836892016-07-26 15:23:23 -07001803 free_obj = obj_malloc(class, get_zspage(d_page), handle);
Minchan Kim251cbb92016-05-20 16:59:42 -07001804 zs_object_copy(class, free_obj, used_obj);
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001805 obj_idx++;
Junil Leec102f072016-01-20 14:58:18 -08001806 /*
1807 * record_obj updates handle's value to free_obj and it will
1808 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1809 * breaks synchronization using pin_tag(e,g, zs_free) so
1810 * let's keep the lock bit.
1811 */
1812 free_obj |= BIT(HANDLE_PIN_BIT);
Minchan Kim312fcae2015-04-15 16:15:30 -07001813 record_obj(handle, free_obj);
1814 unpin_tag(handle);
Minchan Kim1ee47162016-05-20 16:59:45 -07001815 obj_free(class, used_obj);
Minchan Kim312fcae2015-04-15 16:15:30 -07001816 }
1817
1818 /* Remember last position in this iteration */
1819 cc->s_page = s_page;
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07001820 cc->obj_idx = obj_idx;
Minchan Kim312fcae2015-04-15 16:15:30 -07001821
1822 return ret;
1823}
1824
Minchan Kim37836892016-07-26 15:23:23 -07001825static struct zspage *isolate_zspage(struct size_class *class, bool source)
Minchan Kim312fcae2015-04-15 16:15:30 -07001826{
1827 int i;
Minchan Kim37836892016-07-26 15:23:23 -07001828 struct zspage *zspage;
1829 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
Minchan Kim312fcae2015-04-15 16:15:30 -07001830
Minchan Kim37836892016-07-26 15:23:23 -07001831 if (!source) {
1832 fg[0] = ZS_ALMOST_FULL;
1833 fg[1] = ZS_ALMOST_EMPTY;
1834 }
1835
1836 for (i = 0; i < 2; i++) {
1837 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1838 struct zspage, list);
1839 if (zspage) {
Minchan Kim48b48002016-07-26 15:23:31 -07001840 VM_BUG_ON(is_zspage_isolated(zspage));
Minchan Kim37836892016-07-26 15:23:23 -07001841 remove_zspage(class, zspage, fg[i]);
1842 return zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001843 }
1844 }
1845
Minchan Kim37836892016-07-26 15:23:23 -07001846 return zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001847}
1848
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001849/*
Minchan Kim37836892016-07-26 15:23:23 -07001850 * putback_zspage - add @zspage into right class's fullness list
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001851 * @class: destination class
Minchan Kim37836892016-07-26 15:23:23 -07001852 * @zspage: target page
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001853 *
Minchan Kim37836892016-07-26 15:23:23 -07001854 * Return @zspage's fullness_group
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001855 */
Minchan Kim4aa409c2016-07-26 15:23:26 -07001856static enum fullness_group putback_zspage(struct size_class *class,
Minchan Kim37836892016-07-26 15:23:23 -07001857 struct zspage *zspage)
Minchan Kim312fcae2015-04-15 16:15:30 -07001858{
Minchan Kim312fcae2015-04-15 16:15:30 -07001859 enum fullness_group fullness;
1860
Minchan Kim48b48002016-07-26 15:23:31 -07001861 VM_BUG_ON(is_zspage_isolated(zspage));
1862
Minchan Kim37836892016-07-26 15:23:23 -07001863 fullness = get_fullness_group(class, zspage);
1864 insert_zspage(class, zspage, fullness);
1865 set_zspage_mapping(zspage, class->index, fullness);
Minchan Kim839373e2015-04-15 16:16:18 -07001866
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001867 return fullness;
Minchan Kim312fcae2015-04-15 16:15:30 -07001868}
1869
Minchan Kim48b48002016-07-26 15:23:31 -07001870#ifdef CONFIG_COMPACTION
1871static struct dentry *zs_mount(struct file_system_type *fs_type,
1872 int flags, const char *dev_name, void *data)
1873{
1874 static const struct dentry_operations ops = {
1875 .d_dname = simple_dname,
1876 };
1877
1878 return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1879}
1880
1881static struct file_system_type zsmalloc_fs = {
1882 .name = "zsmalloc",
1883 .mount = zs_mount,
1884 .kill_sb = kill_anon_super,
1885};
1886
1887static int zsmalloc_mount(void)
1888{
1889 int ret = 0;
1890
1891 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1892 if (IS_ERR(zsmalloc_mnt))
1893 ret = PTR_ERR(zsmalloc_mnt);
1894
1895 return ret;
1896}
1897
1898static void zsmalloc_unmount(void)
1899{
1900 kern_unmount(zsmalloc_mnt);
1901}
1902
1903static void migrate_lock_init(struct zspage *zspage)
1904{
1905 rwlock_init(&zspage->lock);
1906}
1907
1908static void migrate_read_lock(struct zspage *zspage)
1909{
1910 read_lock(&zspage->lock);
1911}
1912
1913static void migrate_read_unlock(struct zspage *zspage)
1914{
1915 read_unlock(&zspage->lock);
1916}
1917
1918static void migrate_write_lock(struct zspage *zspage)
1919{
1920 write_lock(&zspage->lock);
1921}
1922
1923static void migrate_write_unlock(struct zspage *zspage)
1924{
1925 write_unlock(&zspage->lock);
1926}
1927
1928/* Number of isolated subpage for *page migration* in this zspage */
1929static void inc_zspage_isolation(struct zspage *zspage)
1930{
1931 zspage->isolated++;
1932}
1933
1934static void dec_zspage_isolation(struct zspage *zspage)
1935{
1936 zspage->isolated--;
1937}
1938
1939static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1940 struct page *newpage, struct page *oldpage)
1941{
1942 struct page *page;
1943 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1944 int idx = 0;
1945
1946 page = get_first_page(zspage);
1947 do {
1948 if (page == oldpage)
1949 pages[idx] = newpage;
1950 else
1951 pages[idx] = page;
1952 idx++;
1953 } while ((page = get_next_page(page)) != NULL);
1954
1955 create_page_chain(class, zspage, pages);
1956 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1957 if (unlikely(PageHugeObject(oldpage)))
1958 newpage->index = oldpage->index;
1959 __SetPageMovable(newpage, page_mapping(oldpage));
1960}
1961
1962bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1963{
1964 struct zs_pool *pool;
1965 struct size_class *class;
1966 int class_idx;
1967 enum fullness_group fullness;
1968 struct zspage *zspage;
1969 struct address_space *mapping;
1970
1971 /*
1972 * Page is locked so zspage couldn't be destroyed. For detail, look at
1973 * lock_zspage in free_zspage.
1974 */
1975 VM_BUG_ON_PAGE(!PageMovable(page), page);
1976 VM_BUG_ON_PAGE(PageIsolated(page), page);
1977
1978 zspage = get_zspage(page);
1979
1980 /*
1981 * Without class lock, fullness could be stale while class_idx is okay
1982 * because class_idx is constant unless page is freed so we should get
1983 * fullness again under class lock.
1984 */
1985 get_zspage_mapping(zspage, &class_idx, &fullness);
1986 mapping = page_mapping(page);
1987 pool = mapping->private_data;
1988 class = pool->size_class[class_idx];
1989
1990 spin_lock(&class->lock);
1991 if (get_zspage_inuse(zspage) == 0) {
1992 spin_unlock(&class->lock);
1993 return false;
1994 }
1995
1996 /* zspage is isolated for object migration */
1997 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1998 spin_unlock(&class->lock);
1999 return false;
2000 }
2001
2002 /*
2003 * If this is first time isolation for the zspage, isolate zspage from
2004 * size_class to prevent further object allocation from the zspage.
2005 */
2006 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
2007 get_zspage_mapping(zspage, &class_idx, &fullness);
2008 remove_zspage(class, zspage, fullness);
2009 }
2010
2011 inc_zspage_isolation(zspage);
2012 spin_unlock(&class->lock);
2013
2014 return true;
2015}
2016
2017int zs_page_migrate(struct address_space *mapping, struct page *newpage,
2018 struct page *page, enum migrate_mode mode)
2019{
2020 struct zs_pool *pool;
2021 struct size_class *class;
2022 int class_idx;
2023 enum fullness_group fullness;
2024 struct zspage *zspage;
2025 struct page *dummy;
2026 void *s_addr, *d_addr, *addr;
2027 int offset, pos;
2028 unsigned long handle, head;
2029 unsigned long old_obj, new_obj;
2030 unsigned int obj_idx;
2031 int ret = -EAGAIN;
2032
2033 VM_BUG_ON_PAGE(!PageMovable(page), page);
2034 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2035
2036 zspage = get_zspage(page);
2037
2038 /* Concurrent compactor cannot migrate any subpage in zspage */
2039 migrate_write_lock(zspage);
2040 get_zspage_mapping(zspage, &class_idx, &fullness);
2041 pool = mapping->private_data;
2042 class = pool->size_class[class_idx];
2043 offset = get_first_obj_offset(page);
2044
2045 spin_lock(&class->lock);
2046 if (!get_zspage_inuse(zspage)) {
2047 ret = -EBUSY;
2048 goto unlock_class;
2049 }
2050
2051 pos = offset;
2052 s_addr = kmap_atomic(page);
2053 while (pos < PAGE_SIZE) {
2054 head = obj_to_head(page, s_addr + pos);
2055 if (head & OBJ_ALLOCATED_TAG) {
2056 handle = head & ~OBJ_ALLOCATED_TAG;
2057 if (!trypin_tag(handle))
2058 goto unpin_objects;
2059 }
2060 pos += class->size;
2061 }
2062
2063 /*
2064 * Here, any user cannot access all objects in the zspage so let's move.
2065 */
2066 d_addr = kmap_atomic(newpage);
2067 memcpy(d_addr, s_addr, PAGE_SIZE);
2068 kunmap_atomic(d_addr);
2069
2070 for (addr = s_addr + offset; addr < s_addr + pos;
2071 addr += class->size) {
2072 head = obj_to_head(page, addr);
2073 if (head & OBJ_ALLOCATED_TAG) {
2074 handle = head & ~OBJ_ALLOCATED_TAG;
2075 if (!testpin_tag(handle))
2076 BUG();
2077
2078 old_obj = handle_to_obj(handle);
2079 obj_to_location(old_obj, &dummy, &obj_idx);
2080 new_obj = (unsigned long)location_to_obj(newpage,
2081 obj_idx);
2082 new_obj |= BIT(HANDLE_PIN_BIT);
2083 record_obj(handle, new_obj);
2084 }
2085 }
2086
2087 replace_sub_page(class, zspage, newpage, page);
2088 get_page(newpage);
2089
2090 dec_zspage_isolation(zspage);
2091
2092 /*
2093 * Page migration is done so let's putback isolated zspage to
2094 * the list if @page is final isolated subpage in the zspage.
2095 */
2096 if (!is_zspage_isolated(zspage))
2097 putback_zspage(class, zspage);
2098
2099 reset_page(page);
2100 put_page(page);
2101 page = newpage;
2102
Minchan Kimdd4123f2016-07-26 15:26:50 -07002103 ret = MIGRATEPAGE_SUCCESS;
Minchan Kim48b48002016-07-26 15:23:31 -07002104unpin_objects:
2105 for (addr = s_addr + offset; addr < s_addr + pos;
2106 addr += class->size) {
2107 head = obj_to_head(page, addr);
2108 if (head & OBJ_ALLOCATED_TAG) {
2109 handle = head & ~OBJ_ALLOCATED_TAG;
2110 if (!testpin_tag(handle))
2111 BUG();
2112 unpin_tag(handle);
2113 }
2114 }
2115 kunmap_atomic(s_addr);
2116unlock_class:
2117 spin_unlock(&class->lock);
2118 migrate_write_unlock(zspage);
2119
2120 return ret;
2121}
2122
2123void zs_page_putback(struct page *page)
2124{
2125 struct zs_pool *pool;
2126 struct size_class *class;
2127 int class_idx;
2128 enum fullness_group fg;
2129 struct address_space *mapping;
2130 struct zspage *zspage;
2131
2132 VM_BUG_ON_PAGE(!PageMovable(page), page);
2133 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2134
2135 zspage = get_zspage(page);
2136 get_zspage_mapping(zspage, &class_idx, &fg);
2137 mapping = page_mapping(page);
2138 pool = mapping->private_data;
2139 class = pool->size_class[class_idx];
2140
2141 spin_lock(&class->lock);
2142 dec_zspage_isolation(zspage);
2143 if (!is_zspage_isolated(zspage)) {
2144 fg = putback_zspage(class, zspage);
2145 /*
2146 * Due to page_lock, we cannot free zspage immediately
2147 * so let's defer.
2148 */
2149 if (fg == ZS_EMPTY)
2150 schedule_work(&pool->free_work);
2151 }
2152 spin_unlock(&class->lock);
2153}
2154
2155const struct address_space_operations zsmalloc_aops = {
2156 .isolate_page = zs_page_isolate,
2157 .migratepage = zs_page_migrate,
2158 .putback_page = zs_page_putback,
2159};
2160
2161static int zs_register_migration(struct zs_pool *pool)
2162{
2163 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2164 if (IS_ERR(pool->inode)) {
2165 pool->inode = NULL;
2166 return 1;
2167 }
2168
2169 pool->inode->i_mapping->private_data = pool;
2170 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2171 return 0;
2172}
2173
2174static void zs_unregister_migration(struct zs_pool *pool)
2175{
2176 flush_work(&pool->free_work);
Markus Elfringc3491ec2016-07-28 15:48:59 -07002177 iput(pool->inode);
Minchan Kim48b48002016-07-26 15:23:31 -07002178}
2179
2180/*
2181 * Caller should hold page_lock of all pages in the zspage
2182 * In here, we cannot use zspage meta data.
2183 */
2184static void async_free_zspage(struct work_struct *work)
2185{
2186 int i;
2187 struct size_class *class;
2188 unsigned int class_idx;
2189 enum fullness_group fullness;
2190 struct zspage *zspage, *tmp;
2191 LIST_HEAD(free_pages);
2192 struct zs_pool *pool = container_of(work, struct zs_pool,
2193 free_work);
2194
2195 for (i = 0; i < zs_size_classes; i++) {
2196 class = pool->size_class[i];
2197 if (class->index != i)
2198 continue;
2199
2200 spin_lock(&class->lock);
2201 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2202 spin_unlock(&class->lock);
2203 }
2204
2205
2206 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2207 list_del(&zspage->list);
2208 lock_zspage(zspage);
2209
2210 get_zspage_mapping(zspage, &class_idx, &fullness);
2211 VM_BUG_ON(fullness != ZS_EMPTY);
2212 class = pool->size_class[class_idx];
2213 spin_lock(&class->lock);
2214 __free_zspage(pool, pool->size_class[class_idx], zspage);
2215 spin_unlock(&class->lock);
2216 }
2217};
2218
2219static void kick_deferred_free(struct zs_pool *pool)
2220{
2221 schedule_work(&pool->free_work);
2222}
2223
2224static void init_deferred_free(struct zs_pool *pool)
2225{
2226 INIT_WORK(&pool->free_work, async_free_zspage);
2227}
2228
2229static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2230{
2231 struct page *page = get_first_page(zspage);
2232
2233 do {
2234 WARN_ON(!trylock_page(page));
2235 __SetPageMovable(page, pool->inode->i_mapping);
2236 unlock_page(page);
2237 } while ((page = get_next_page(page)) != NULL);
2238}
2239#endif
2240
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002241/*
2242 *
2243 * Based on the number of unused allocated objects calculate
2244 * and return the number of pages that we can free.
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002245 */
2246static unsigned long zs_can_compact(struct size_class *class)
2247{
2248 unsigned long obj_wasted;
Sergey Senozhatsky44f43e92016-05-09 16:28:49 -07002249 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2250 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002251
Sergey Senozhatsky44f43e92016-05-09 16:28:49 -07002252 if (obj_allocated <= obj_used)
2253 return 0;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002254
Sergey Senozhatsky44f43e92016-05-09 16:28:49 -07002255 obj_wasted = obj_allocated - obj_used;
Ganesh Mahendranb4fd07a2016-07-28 15:47:49 -07002256 obj_wasted /= class->objs_per_zspage;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002257
Minchan Kim6cbf16b2015-09-08 15:04:49 -07002258 return obj_wasted * class->pages_per_zspage;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002259}
2260
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07002261static void __zs_compact(struct zs_pool *pool, struct size_class *class)
Minchan Kim312fcae2015-04-15 16:15:30 -07002262{
Minchan Kim312fcae2015-04-15 16:15:30 -07002263 struct zs_compact_control cc;
Minchan Kim37836892016-07-26 15:23:23 -07002264 struct zspage *src_zspage;
2265 struct zspage *dst_zspage = NULL;
Minchan Kim312fcae2015-04-15 16:15:30 -07002266
Minchan Kim312fcae2015-04-15 16:15:30 -07002267 spin_lock(&class->lock);
Minchan Kim37836892016-07-26 15:23:23 -07002268 while ((src_zspage = isolate_zspage(class, true))) {
Minchan Kim312fcae2015-04-15 16:15:30 -07002269
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07002270 if (!zs_can_compact(class))
2271 break;
2272
Ganesh Mahendran41b88e12016-07-28 15:47:43 -07002273 cc.obj_idx = 0;
Minchan Kim48b48002016-07-26 15:23:31 -07002274 cc.s_page = get_first_page(src_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002275
Minchan Kim37836892016-07-26 15:23:23 -07002276 while ((dst_zspage = isolate_zspage(class, false))) {
Minchan Kim48b48002016-07-26 15:23:31 -07002277 cc.d_page = get_first_page(dst_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002278 /*
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07002279 * If there is no more space in dst_page, resched
2280 * and see if anyone had allocated another zspage.
Minchan Kim312fcae2015-04-15 16:15:30 -07002281 */
2282 if (!migrate_zspage(pool, class, &cc))
2283 break;
2284
Minchan Kim4aa409c2016-07-26 15:23:26 -07002285 putback_zspage(class, dst_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002286 }
2287
2288 /* Stop if we couldn't find slot */
Minchan Kim37836892016-07-26 15:23:23 -07002289 if (dst_zspage == NULL)
Minchan Kim312fcae2015-04-15 16:15:30 -07002290 break;
2291
Minchan Kim4aa409c2016-07-26 15:23:26 -07002292 putback_zspage(class, dst_zspage);
2293 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
Minchan Kim48b48002016-07-26 15:23:31 -07002294 free_zspage(pool, class, src_zspage);
Minchan Kim6cbf16b2015-09-08 15:04:49 -07002295 pool->stats.pages_compacted += class->pages_per_zspage;
Minchan Kim4aa409c2016-07-26 15:23:26 -07002296 }
Minchan Kim312fcae2015-04-15 16:15:30 -07002297 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07002298 cond_resched();
2299 spin_lock(&class->lock);
2300 }
2301
Minchan Kim37836892016-07-26 15:23:23 -07002302 if (src_zspage)
Minchan Kim4aa409c2016-07-26 15:23:26 -07002303 putback_zspage(class, src_zspage);
Minchan Kim312fcae2015-04-15 16:15:30 -07002304
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07002305 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07002306}
2307
2308unsigned long zs_compact(struct zs_pool *pool)
2309{
2310 int i;
Minchan Kim312fcae2015-04-15 16:15:30 -07002311 struct size_class *class;
2312
2313 for (i = zs_size_classes - 1; i >= 0; i--) {
2314 class = pool->size_class[i];
2315 if (!class)
2316 continue;
2317 if (class->index != i)
2318 continue;
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07002319 __zs_compact(pool, class);
Minchan Kim312fcae2015-04-15 16:15:30 -07002320 }
2321
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07002322 return pool->stats.pages_compacted;
Minchan Kim312fcae2015-04-15 16:15:30 -07002323}
2324EXPORT_SYMBOL_GPL(zs_compact);
2325
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07002326void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2327{
2328 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2329}
2330EXPORT_SYMBOL_GPL(zs_pool_stats);
2331
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002332static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2333 struct shrink_control *sc)
2334{
2335 unsigned long pages_freed;
2336 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2337 shrinker);
2338
2339 pages_freed = pool->stats.pages_compacted;
2340 /*
2341 * Compact classes and calculate compaction delta.
2342 * Can run concurrently with a manually triggered
2343 * (by user) compaction.
2344 */
2345 pages_freed = zs_compact(pool) - pages_freed;
2346
2347 return pages_freed ? pages_freed : SHRINK_STOP;
2348}
2349
2350static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2351 struct shrink_control *sc)
2352{
2353 int i;
2354 struct size_class *class;
2355 unsigned long pages_to_free = 0;
2356 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2357 shrinker);
2358
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002359 for (i = zs_size_classes - 1; i >= 0; i--) {
2360 class = pool->size_class[i];
2361 if (!class)
2362 continue;
2363 if (class->index != i)
2364 continue;
2365
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002366 pages_to_free += zs_can_compact(class);
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002367 }
2368
2369 return pages_to_free;
2370}
2371
2372static void zs_unregister_shrinker(struct zs_pool *pool)
2373{
2374 if (pool->shrinker_enabled) {
2375 unregister_shrinker(&pool->shrinker);
2376 pool->shrinker_enabled = false;
2377 }
2378}
2379
2380static int zs_register_shrinker(struct zs_pool *pool)
2381{
2382 pool->shrinker.scan_objects = zs_shrinker_scan;
2383 pool->shrinker.count_objects = zs_shrinker_count;
2384 pool->shrinker.batch = 0;
2385 pool->shrinker.seeks = DEFAULT_SEEKS;
2386
2387 return register_shrinker(&pool->shrinker);
2388}
2389
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08002390/**
2391 * zs_create_pool - Creates an allocation pool to work from.
Ganesh Mahendranfd854462016-07-28 15:47:54 -07002392 * @name: pool name to be created
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08002393 *
2394 * This function must be called before anything when using
2395 * the zsmalloc allocator.
2396 *
2397 * On success, a pointer to the newly created pool is returned,
2398 * otherwise NULL.
2399 */
Sergey Senozhatskyd0d8da22016-05-20 16:59:48 -07002400struct zs_pool *zs_create_pool(const char *name)
Nitin Gupta61989a82012-01-09 16:51:56 -06002401{
Ganesh Mahendran18136652014-12-12 16:57:10 -08002402 int i;
Nitin Gupta61989a82012-01-09 16:51:56 -06002403 struct zs_pool *pool;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08002404 struct size_class *prev_class = NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06002405
Ganesh Mahendran18136652014-12-12 16:57:10 -08002406 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
Nitin Gupta61989a82012-01-09 16:51:56 -06002407 if (!pool)
2408 return NULL;
2409
Minchan Kim48b48002016-07-26 15:23:31 -07002410 init_deferred_free(pool);
Minchan Kim2e40e162015-04-15 16:15:23 -07002411 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
2412 GFP_KERNEL);
2413 if (!pool->size_class) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002414 kfree(pool);
2415 return NULL;
2416 }
2417
Minchan Kim2e40e162015-04-15 16:15:23 -07002418 pool->name = kstrdup(name, GFP_KERNEL);
2419 if (!pool->name)
2420 goto err;
2421
Minchan Kim37836892016-07-26 15:23:23 -07002422 if (create_cache(pool))
Minchan Kim2e40e162015-04-15 16:15:23 -07002423 goto err;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002424
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002425 /*
2426 * Iterate reversly, because, size of size_class that we want to use
2427 * for merging should be larger or equal to current size.
2428 */
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002429 for (i = zs_size_classes - 1; i >= 0; i--) {
Nitin Gupta61989a82012-01-09 16:51:56 -06002430 int size;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002431 int pages_per_zspage;
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002432 int objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06002433 struct size_class *class;
Minchan Kim37836892016-07-26 15:23:23 -07002434 int fullness = 0;
Nitin Gupta61989a82012-01-09 16:51:56 -06002435
2436 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2437 if (size > ZS_MAX_ALLOC_SIZE)
2438 size = ZS_MAX_ALLOC_SIZE;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002439 pages_per_zspage = get_pages_per_zspage(size);
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002440 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
Nitin Gupta61989a82012-01-09 16:51:56 -06002441
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002442 /*
2443 * size_class is used for normal zsmalloc operation such
2444 * as alloc/free for that size. Although it is natural that we
2445 * have one size_class for each size, there is a chance that we
2446 * can get more memory utilization if we use one size_class for
2447 * many different sizes whose size_class have same
2448 * characteristics. So, we makes size_class point to
2449 * previous size_class if possible.
2450 */
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08002451 if (prev_class) {
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002452 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002453 pool->size_class[i] = prev_class;
2454 continue;
2455 }
2456 }
2457
2458 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2459 if (!class)
2460 goto err;
2461
Nitin Gupta61989a82012-01-09 16:51:56 -06002462 class->size = size;
2463 class->index = i;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002464 class->pages_per_zspage = pages_per_zspage;
Ganesh Mahendran64d90462016-07-28 15:47:51 -07002465 class->objs_per_zspage = objs_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06002466 spin_lock_init(&class->lock);
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002467 pool->size_class[i] = class;
Minchan Kim48b48002016-07-26 15:23:31 -07002468 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2469 fullness++)
Minchan Kim37836892016-07-26 15:23:23 -07002470 INIT_LIST_HEAD(&class->fullness_list[fullness]);
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08002471
2472 prev_class = class;
Nitin Gupta61989a82012-01-09 16:51:56 -06002473 }
2474
Dan Streetmand34f6152016-05-20 16:59:56 -07002475 /* debug only, don't abort if it fails */
2476 zs_pool_stat_create(pool, name);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002477
Minchan Kim48b48002016-07-26 15:23:31 -07002478 if (zs_register_migration(pool))
2479 goto err;
2480
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002481 /*
2482 * Not critical, we still can use the pool
2483 * and user can trigger compaction manually.
2484 */
2485 if (zs_register_shrinker(pool) == 0)
2486 pool->shrinker_enabled = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06002487 return pool;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002488
2489err:
2490 zs_destroy_pool(pool);
2491 return NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06002492}
2493EXPORT_SYMBOL_GPL(zs_create_pool);
2494
2495void zs_destroy_pool(struct zs_pool *pool)
2496{
2497 int i;
2498
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07002499 zs_unregister_shrinker(pool);
Minchan Kim48b48002016-07-26 15:23:31 -07002500 zs_unregister_migration(pool);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002501 zs_pool_stat_destroy(pool);
2502
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002503 for (i = 0; i < zs_size_classes; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -06002504 int fg;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002505 struct size_class *class = pool->size_class[i];
2506
2507 if (!class)
2508 continue;
2509
2510 if (class->index != i)
2511 continue;
Nitin Gupta61989a82012-01-09 16:51:56 -06002512
Minchan Kim48b48002016-07-26 15:23:31 -07002513 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
Minchan Kim37836892016-07-26 15:23:23 -07002514 if (!list_empty(&class->fullness_list[fg])) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -04002515 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -06002516 class->size, fg);
2517 }
2518 }
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08002519 kfree(class);
Nitin Gupta61989a82012-01-09 16:51:56 -06002520 }
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002521
Minchan Kim37836892016-07-26 15:23:23 -07002522 destroy_cache(pool);
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002523 kfree(pool->size_class);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002524 kfree(pool->name);
Nitin Gupta61989a82012-01-09 16:51:56 -06002525 kfree(pool);
2526}
2527EXPORT_SYMBOL_GPL(zs_destroy_pool);
2528
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002529static int __init zs_init(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002530{
Minchan Kim48b48002016-07-26 15:23:31 -07002531 int ret;
2532
2533 ret = zsmalloc_mount();
2534 if (ret)
2535 goto out;
2536
2537 ret = zs_register_cpu_notifier();
Nitin Gupta61989a82012-01-09 16:51:56 -06002538
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002539 if (ret)
2540 goto notifier_fail;
Nitin Gupta61989a82012-01-09 16:51:56 -06002541
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002542 init_zs_size_classes();
Nitin Gupta61989a82012-01-09 16:51:56 -06002543
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002544#ifdef CONFIG_ZPOOL
2545 zpool_register_driver(&zs_zpool_driver);
2546#endif
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002547
Dan Streetman4abaac92016-05-26 15:16:27 -07002548 zs_stat_init();
2549
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002550 return 0;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002551
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002552notifier_fail:
2553 zs_unregister_cpu_notifier();
Minchan Kim48b48002016-07-26 15:23:31 -07002554 zsmalloc_unmount();
2555out:
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002556 return ret;
Nitin Gupta61989a82012-01-09 16:51:56 -06002557}
Nitin Gupta61989a82012-01-09 16:51:56 -06002558
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002559static void __exit zs_exit(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002560{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002561#ifdef CONFIG_ZPOOL
2562 zpool_unregister_driver(&zs_zpool_driver);
2563#endif
Minchan Kim48b48002016-07-26 15:23:31 -07002564 zsmalloc_unmount();
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002565 zs_unregister_cpu_notifier();
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002566
2567 zs_stat_exit();
Nitin Gupta61989a82012-01-09 16:51:56 -06002568}
Ben Hutchings069f1012012-06-20 02:31:11 +01002569
2570module_init(zs_init);
2571module_exit(zs_exit);
2572
2573MODULE_LICENSE("Dual BSD/GPL");
2574MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");