blob: 2d7c4c11fc635d3822d64e8b2d796414945305f7 [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:
Kirill A. Shutemov32e7ba12015-11-06 16:29:47 -080019 * page->private: points to the first component (0-order) page
Nitin Gupta2db51da2012-06-09 17:41:14 -070020 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
25 * of a zspage
26 *
27 * For _first_ page only:
28 *
Kirill A. Shutemov32e7ba12015-11-06 16:29:47 -080029 * page->private: refers to the component page after the first page
Minchan Kim7b60a682015-04-15 16:15:39 -070030 * If the page is first_page for huge object, it stores handle.
31 * Look at size_class->huge.
Nitin Gupta2db51da2012-06-09 17:41:14 -070032 * page->freelist: points to the first free object in zspage.
33 * Free objects are linked together using in-place
34 * metadata.
35 * page->objects: maximum number of objects we can store in this
36 * zspage (class->zspage_order * PAGE_SIZE / class->size)
37 * page->lru: links together first pages of various zspages.
38 * Basically forming list of zspages in a fullness group.
39 * page->mapping: class index and fullness group of the zspage
Hui Zhu8f958c92015-11-06 16:29:23 -080040 * page->inuse: the number of objects that are used in this zspage
Nitin Gupta2db51da2012-06-09 17:41:14 -070041 *
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
45 *
46 */
47
Nitin Gupta61989a82012-01-09 16:51:56 -060048#include <linux/module.h>
49#include <linux/kernel.h>
Minchan Kim312fcae2015-04-15 16:15:30 -070050#include <linux/sched.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060051#include <linux/bitops.h>
52#include <linux/errno.h>
53#include <linux/highmem.h>
Nitin Gupta61989a82012-01-09 16:51:56 -060054#include <linux/string.h>
55#include <linux/slab.h>
56#include <asm/tlbflush.h>
57#include <asm/pgtable.h>
58#include <linux/cpumask.h>
59#include <linux/cpu.h>
Seth Jennings0cbb6132012-02-13 08:47:49 -060060#include <linux/vmalloc.h>
Sergey Senozhatsky759b26b2015-11-06 16:29:29 -080061#include <linux/preempt.h>
Seth Jennings0959c632012-08-08 15:12:17 +090062#include <linux/spinlock.h>
63#include <linux/types.h>
Ganesh Mahendran0f050d92015-02-12 15:00:54 -080064#include <linux/debugfs.h>
Minchan Kimbcf16472014-01-30 15:45:50 -080065#include <linux/zsmalloc.h>
Dan Streetmanc7957792014-08-06 16:08:38 -070066#include <linux/zpool.h>
Seth Jennings0959c632012-08-08 15:12:17 +090067
68/*
69 * This must be power of 2 and greater than of equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
73 */
74#define ZS_ALIGN 8
75
76/*
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
79 */
80#define ZS_MAX_ZSPAGE_ORDER 2
81#define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
82
Minchan Kim2e40e162015-04-15 16:15:23 -070083#define ZS_HANDLE_SIZE (sizeof(unsigned long))
84
Seth Jennings0959c632012-08-08 15:12:17 +090085/*
86 * Object location (<PFN>, <obj_idx>) is encoded as
Nitin Cuptac3e3e882013-12-11 11:04:37 +090087 * as single (unsigned long) handle value.
Seth Jennings0959c632012-08-08 15:12:17 +090088 *
89 * Note that object index <obj_idx> is relative to system
90 * page <PFN> it is stored in, so for each sub-page belonging
91 * to a zspage, obj_idx starts with 0.
92 *
93 * This is made more complicated by various memory models and PAE.
94 */
95
96#ifndef MAX_PHYSMEM_BITS
97#ifdef CONFIG_HIGHMEM64G
98#define MAX_PHYSMEM_BITS 36
99#else /* !CONFIG_HIGHMEM64G */
100/*
101 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
102 * be PAGE_SHIFT
103 */
104#define MAX_PHYSMEM_BITS BITS_PER_LONG
105#endif
106#endif
107#define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
Minchan Kim312fcae2015-04-15 16:15:30 -0700108
109/*
110 * Memory for allocating for handle keeps object position by
111 * encoding <page, obj_idx> and the encoded value has a room
112 * in least bit(ie, look at obj_to_location).
113 * We use the bit to synchronize between object access by
114 * user and migration.
115 */
116#define HANDLE_PIN_BIT 0
117
118/*
119 * Head in allocated object should have OBJ_ALLOCATED_TAG
120 * to identify the object was allocated or not.
121 * It's okay to add the status bit in the least bit because
122 * header keeps handle which is 4byte-aligned address so we
123 * have room for two bit at least.
124 */
125#define OBJ_ALLOCATED_TAG 1
126#define OBJ_TAG_BITS 1
127#define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
Seth Jennings0959c632012-08-08 15:12:17 +0900128#define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
129
130#define MAX(a, b) ((a) >= (b) ? (a) : (b))
131/* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132#define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
Minchan Kim2e40e162015-04-15 16:15:23 -0700134/* each chunk includes extra space to keep handle */
Minchan Kim7b60a682015-04-15 16:15:39 -0700135#define ZS_MAX_ALLOC_SIZE PAGE_SIZE
Seth Jennings0959c632012-08-08 15:12:17 +0900136
137/*
Weijie Yang7eb52512014-06-04 16:11:08 -0700138 * On systems with 4K page size, this gives 255 size classes! There is a
Seth Jennings0959c632012-08-08 15:12:17 +0900139 * trader-off here:
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
146 *
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
148 * (reason above)
149 */
Seth Jenningsd662b8e2013-01-25 11:46:18 -0600150#define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
Seth Jennings0959c632012-08-08 15:12:17 +0900151
152/*
153 * We do not maintain any list for completely empty or full pages
154 */
155enum fullness_group {
156 ZS_ALMOST_FULL,
157 ZS_ALMOST_EMPTY,
158 _ZS_NR_FULLNESS_GROUPS,
159
160 ZS_EMPTY,
161 ZS_FULL
162};
163
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800164enum zs_stat_type {
165 OBJ_ALLOCATED,
166 OBJ_USED,
Minchan Kim248ca1b2015-04-15 16:15:42 -0700167 CLASS_ALMOST_FULL,
168 CLASS_ALMOST_EMPTY,
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800169};
170
Sergey Senozhatsky6fe51862015-11-06 16:29:38 -0800171#ifdef CONFIG_ZSMALLOC_STAT
172#define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1)
173#else
174#define NR_ZS_STAT_TYPE (OBJ_USED + 1)
175#endif
176
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800177struct zs_size_stat {
178 unsigned long objs[NR_ZS_STAT_TYPE];
179};
180
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700181#ifdef CONFIG_ZSMALLOC_STAT
182static struct dentry *zs_stat_root;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -0800183#endif
184
Seth Jennings0959c632012-08-08 15:12:17 +0900185/*
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -0800186 * number of size_classes
187 */
188static int zs_size_classes;
189
190/*
Seth Jennings0959c632012-08-08 15:12:17 +0900191 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
192 * n <= N / f, where
193 * n = number of allocated objects
194 * N = total number of objects zspage can store
Wang Sheng-Hui6dd97372014-10-09 15:29:59 -0700195 * f = fullness_threshold_frac
Seth Jennings0959c632012-08-08 15:12:17 +0900196 *
197 * Similarly, we assign zspage to:
198 * ZS_ALMOST_FULL when n > N / f
199 * ZS_EMPTY when n == 0
200 * ZS_FULL when n == N
201 *
202 * (see: fix_fullness_group())
203 */
204static const int fullness_threshold_frac = 4;
205
206struct size_class {
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700207 spinlock_t lock;
208 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
Seth Jennings0959c632012-08-08 15:12:17 +0900209 /*
210 * Size of objects stored in this class. Must be multiple
211 * of ZS_ALIGN.
212 */
213 int size;
214 unsigned int index;
215
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700216 struct zs_size_stat stats;
217
Weijie Yang7dfa4612016-01-14 15:22:40 -0800218 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
219 int pages_per_zspage;
Minchan Kim7b60a682015-04-15 16:15:39 -0700220 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
221 bool huge;
Seth Jennings0959c632012-08-08 15:12:17 +0900222};
223
224/*
225 * Placed within free objects to form a singly linked list.
226 * For every zspage, first_page->freelist gives head of this list.
227 *
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 /*
233 * Position of next free chunk (encodes <PFN, obj_idx>)
234 * It's valid for non-allocated object
235 */
236 void *next;
237 /*
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;
Seth Jennings0959c632012-08-08 15:12:17 +0900249
250 gfp_t flags; /* allocation flags used when growing pool */
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
Seth Jennings0959c632012-08-08 15:12:17 +0900265};
Nitin Gupta61989a82012-01-09 16:51:56 -0600266
267/*
268 * A zspage's class index and fullness group
269 * are encoded in its (first)page->mapping
270 */
271#define CLASS_IDX_BITS 28
272#define FULLNESS_BITS 4
273#define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
274#define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
275
Seth Jenningsf5536462012-07-18 11:55:56 -0500276struct mapping_area {
Minchan Kim1b945ae2013-12-11 11:04:36 +0900277#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -0500278 struct vm_struct *vm; /* vm area for mapping object that span pages */
279#else
280 char *vm_buf; /* copy buffer for objects that span pages */
281#endif
282 char *vm_addr; /* address of kmap_atomic()'ed pages */
283 enum zs_mapmode vm_mm; /* mapping mode */
Minchan Kim7b60a682015-04-15 16:15:39 -0700284 bool huge;
Seth Jenningsf5536462012-07-18 11:55:56 -0500285};
286
Minchan Kim2e40e162015-04-15 16:15:23 -0700287static int create_handle_cache(struct zs_pool *pool)
288{
289 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
290 0, 0, NULL);
291 return pool->handle_cachep ? 0 : 1;
292}
293
294static void destroy_handle_cache(struct zs_pool *pool)
295{
Sergey Senozhatskycd10add2015-09-08 15:04:55 -0700296 kmem_cache_destroy(pool->handle_cachep);
Minchan Kim2e40e162015-04-15 16:15:23 -0700297}
298
299static unsigned long alloc_handle(struct zs_pool *pool)
300{
301 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
302 pool->flags & ~__GFP_HIGHMEM);
303}
304
305static void free_handle(struct zs_pool *pool, unsigned long handle)
306{
307 kmem_cache_free(pool->handle_cachep, (void *)handle);
308}
309
310static void record_obj(unsigned long handle, unsigned long obj)
311{
Junil Leec102f072016-01-20 14:58:18 -0800312 /*
313 * lsb of @obj represents handle lock while other bits
314 * represent object value the handle is pointing so
315 * updating shouldn't do store tearing.
316 */
317 WRITE_ONCE(*(unsigned long *)handle, obj);
Minchan Kim2e40e162015-04-15 16:15:23 -0700318}
319
Dan Streetmanc7957792014-08-06 16:08:38 -0700320/* zpool driver */
321
322#ifdef CONFIG_ZPOOL
323
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800324static void *zs_zpool_create(const char *name, gfp_t gfp,
Krzysztof Kozlowski78672772015-09-08 15:05:03 -0700325 const struct zpool_ops *zpool_ops,
Dan Streetman479305f2015-06-25 15:00:40 -0700326 struct zpool *zpool)
Dan Streetmanc7957792014-08-06 16:08:38 -0700327{
Ganesh Mahendran3eba0c62015-02-12 15:00:51 -0800328 return zs_create_pool(name, gfp);
Dan Streetmanc7957792014-08-06 16:08:38 -0700329}
330
331static void zs_zpool_destroy(void *pool)
332{
333 zs_destroy_pool(pool);
334}
335
336static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
337 unsigned long *handle)
338{
339 *handle = zs_malloc(pool, size);
340 return *handle ? 0 : -1;
341}
342static void zs_zpool_free(void *pool, unsigned long handle)
343{
344 zs_free(pool, handle);
345}
346
347static int zs_zpool_shrink(void *pool, unsigned int pages,
348 unsigned int *reclaimed)
349{
350 return -EINVAL;
351}
352
353static void *zs_zpool_map(void *pool, unsigned long handle,
354 enum zpool_mapmode mm)
355{
356 enum zs_mapmode zs_mm;
357
358 switch (mm) {
359 case ZPOOL_MM_RO:
360 zs_mm = ZS_MM_RO;
361 break;
362 case ZPOOL_MM_WO:
363 zs_mm = ZS_MM_WO;
364 break;
365 case ZPOOL_MM_RW: /* fallthru */
366 default:
367 zs_mm = ZS_MM_RW;
368 break;
369 }
370
371 return zs_map_object(pool, handle, zs_mm);
372}
373static void zs_zpool_unmap(void *pool, unsigned long handle)
374{
375 zs_unmap_object(pool, handle);
376}
377
378static u64 zs_zpool_total_size(void *pool)
379{
Minchan Kim722cdc12014-10-09 15:29:50 -0700380 return zs_get_total_pages(pool) << PAGE_SHIFT;
Dan Streetmanc7957792014-08-06 16:08:38 -0700381}
382
383static struct zpool_driver zs_zpool_driver = {
384 .type = "zsmalloc",
385 .owner = THIS_MODULE,
386 .create = zs_zpool_create,
387 .destroy = zs_zpool_destroy,
388 .malloc = zs_zpool_malloc,
389 .free = zs_zpool_free,
390 .shrink = zs_zpool_shrink,
391 .map = zs_zpool_map,
392 .unmap = zs_zpool_unmap,
393 .total_size = zs_zpool_total_size,
394};
395
Kees Cook137f8cf2014-08-29 15:18:40 -0700396MODULE_ALIAS("zpool-zsmalloc");
Dan Streetmanc7957792014-08-06 16:08:38 -0700397#endif /* CONFIG_ZPOOL */
398
Minchan Kim248ca1b2015-04-15 16:15:42 -0700399static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
400{
401 return pages_per_zspage * PAGE_SIZE / size;
402}
403
Nitin Gupta61989a82012-01-09 16:51:56 -0600404/* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
405static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
406
407static int is_first_page(struct page *page)
408{
Minchan Kima27545bf2012-04-25 15:23:09 +0900409 return PagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600410}
411
412static int is_last_page(struct page *page)
413{
Minchan Kima27545bf2012-04-25 15:23:09 +0900414 return PagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600415}
416
417static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
418 enum fullness_group *fullness)
419{
420 unsigned long m;
421 BUG_ON(!is_first_page(page));
422
423 m = (unsigned long)page->mapping;
424 *fullness = m & FULLNESS_MASK;
425 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
426}
427
428static void set_zspage_mapping(struct page *page, unsigned int class_idx,
429 enum fullness_group fullness)
430{
431 unsigned long m;
432 BUG_ON(!is_first_page(page));
433
434 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
435 (fullness & FULLNESS_MASK);
436 page->mapping = (struct address_space *)m;
437}
438
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900439/*
440 * zsmalloc divides the pool into various size classes where each
441 * class maintains a list of zspages where each zspage is divided
442 * into equal sized chunks. Each allocation falls into one of these
443 * classes depending on its size. This function returns index of the
444 * size class which has chunk size big enough to hold the give size.
445 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600446static int get_size_class_index(int size)
447{
448 int idx = 0;
449
450 if (likely(size > ZS_MIN_ALLOC_SIZE))
451 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
452 ZS_SIZE_CLASS_DELTA);
453
Minchan Kim7b60a682015-04-15 16:15:39 -0700454 return min(zs_size_classes - 1, idx);
Nitin Gupta61989a82012-01-09 16:51:56 -0600455}
456
Minchan Kim248ca1b2015-04-15 16:15:42 -0700457static inline void zs_stat_inc(struct size_class *class,
458 enum zs_stat_type type, unsigned long cnt)
459{
Sergey Senozhatsky6fe51862015-11-06 16:29:38 -0800460 if (type < NR_ZS_STAT_TYPE)
461 class->stats.objs[type] += cnt;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700462}
463
464static inline void zs_stat_dec(struct size_class *class,
465 enum zs_stat_type type, unsigned long cnt)
466{
Sergey Senozhatsky6fe51862015-11-06 16:29:38 -0800467 if (type < NR_ZS_STAT_TYPE)
468 class->stats.objs[type] -= cnt;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700469}
470
471static inline unsigned long zs_stat_get(struct size_class *class,
472 enum zs_stat_type type)
473{
Sergey Senozhatsky6fe51862015-11-06 16:29:38 -0800474 if (type < NR_ZS_STAT_TYPE)
475 return class->stats.objs[type];
476 return 0;
Minchan Kim248ca1b2015-04-15 16:15:42 -0700477}
478
Sergey Senozhatsky57244592015-09-08 15:04:27 -0700479#ifdef CONFIG_ZSMALLOC_STAT
480
Minchan Kim248ca1b2015-04-15 16:15:42 -0700481static int __init zs_stat_init(void)
482{
483 if (!debugfs_initialized())
484 return -ENODEV;
485
486 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
487 if (!zs_stat_root)
488 return -ENOMEM;
489
490 return 0;
491}
492
493static void __exit zs_stat_exit(void)
494{
495 debugfs_remove_recursive(zs_stat_root);
496}
497
498static int zs_stats_size_show(struct seq_file *s, void *v)
499{
500 int i;
501 struct zs_pool *pool = s->private;
502 struct size_class *class;
503 int objs_per_zspage;
504 unsigned long class_almost_full, class_almost_empty;
505 unsigned long obj_allocated, obj_used, pages_used;
506 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
507 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
508
509 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
510 "class", "size", "almost_full", "almost_empty",
511 "obj_allocated", "obj_used", "pages_used",
512 "pages_per_zspage");
513
514 for (i = 0; i < zs_size_classes; i++) {
515 class = pool->size_class[i];
516
517 if (class->index != i)
518 continue;
519
520 spin_lock(&class->lock);
521 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
522 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
523 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
524 obj_used = zs_stat_get(class, OBJ_USED);
525 spin_unlock(&class->lock);
526
527 objs_per_zspage = get_maxobj_per_zspage(class->size,
528 class->pages_per_zspage);
529 pages_used = obj_allocated / objs_per_zspage *
530 class->pages_per_zspage;
531
532 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
533 i, class->size, class_almost_full, class_almost_empty,
534 obj_allocated, obj_used, pages_used,
535 class->pages_per_zspage);
536
537 total_class_almost_full += class_almost_full;
538 total_class_almost_empty += class_almost_empty;
539 total_objs += obj_allocated;
540 total_used_objs += obj_used;
541 total_pages += pages_used;
542 }
543
544 seq_puts(s, "\n");
545 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
546 "Total", "", total_class_almost_full,
547 total_class_almost_empty, total_objs,
548 total_used_objs, total_pages);
549
550 return 0;
551}
552
553static int zs_stats_size_open(struct inode *inode, struct file *file)
554{
555 return single_open(file, zs_stats_size_show, inode->i_private);
556}
557
558static const struct file_operations zs_stat_size_ops = {
559 .open = zs_stats_size_open,
560 .read = seq_read,
561 .llseek = seq_lseek,
562 .release = single_release,
563};
564
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800565static int zs_pool_stat_create(const char *name, struct zs_pool *pool)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700566{
567 struct dentry *entry;
568
569 if (!zs_stat_root)
570 return -ENODEV;
571
572 entry = debugfs_create_dir(name, zs_stat_root);
573 if (!entry) {
574 pr_warn("debugfs dir <%s> creation failed\n", name);
575 return -ENOMEM;
576 }
577 pool->stat_dentry = entry;
578
579 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
580 pool->stat_dentry, pool, &zs_stat_size_ops);
581 if (!entry) {
582 pr_warn("%s: debugfs file entry <%s> creation failed\n",
583 name, "classes");
584 return -ENOMEM;
585 }
586
587 return 0;
588}
589
590static void zs_pool_stat_destroy(struct zs_pool *pool)
591{
592 debugfs_remove_recursive(pool->stat_dentry);
593}
594
595#else /* CONFIG_ZSMALLOC_STAT */
Minchan Kim248ca1b2015-04-15 16:15:42 -0700596static int __init zs_stat_init(void)
597{
598 return 0;
599}
600
601static void __exit zs_stat_exit(void)
602{
603}
604
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -0800605static inline int zs_pool_stat_create(const char *name, struct zs_pool *pool)
Minchan Kim248ca1b2015-04-15 16:15:42 -0700606{
607 return 0;
608}
609
610static inline void zs_pool_stat_destroy(struct zs_pool *pool)
611{
612}
Minchan Kim248ca1b2015-04-15 16:15:42 -0700613#endif
614
615
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900616/*
617 * For each size class, zspages are divided into different groups
618 * depending on how "full" they are. This was done so that we could
619 * easily find empty or nearly empty zspages when we try to shrink
620 * the pool (not yet implemented). This function returns fullness
621 * status of the given page.
622 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600623static enum fullness_group get_fullness_group(struct page *page)
624{
625 int inuse, max_objects;
626 enum fullness_group fg;
627 BUG_ON(!is_first_page(page));
628
629 inuse = page->inuse;
630 max_objects = page->objects;
631
632 if (inuse == 0)
633 fg = ZS_EMPTY;
634 else if (inuse == max_objects)
635 fg = ZS_FULL;
Minchan Kimd3d07c92015-04-15 16:15:33 -0700636 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
Nitin Gupta61989a82012-01-09 16:51:56 -0600637 fg = ZS_ALMOST_EMPTY;
638 else
639 fg = ZS_ALMOST_FULL;
640
641 return fg;
642}
643
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900644/*
645 * Each size class maintains various freelists and zspages are assigned
646 * to one of these freelists based on the number of live objects they
647 * have. This functions inserts the given zspage into the freelist
648 * identified by <class, fullness_group>.
649 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600650static void insert_zspage(struct page *page, struct size_class *class,
651 enum fullness_group fullness)
652{
653 struct page **head;
654
655 BUG_ON(!is_first_page(page));
656
657 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
658 return;
659
Minchan Kim248ca1b2015-04-15 16:15:42 -0700660 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
661 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Sergey Senozhatsky58f17112015-09-08 15:04:44 -0700662
663 head = &class->fullness_list[fullness];
664 if (!*head) {
665 *head = page;
666 return;
667 }
668
669 /*
670 * We want to see more ZS_FULL pages and less almost
671 * empty/full. Put pages with higher ->inuse first.
672 */
673 list_add_tail(&page->lru, &(*head)->lru);
674 if (page->inuse >= (*head)->inuse)
675 *head = page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600676}
677
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900678/*
679 * This function removes the given zspage from the freelist identified
680 * by <class, fullness_group>.
681 */
Nitin Gupta61989a82012-01-09 16:51:56 -0600682static void remove_zspage(struct page *page, struct size_class *class,
683 enum fullness_group fullness)
684{
685 struct page **head;
686
687 BUG_ON(!is_first_page(page));
688
689 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
690 return;
691
692 head = &class->fullness_list[fullness];
693 BUG_ON(!*head);
694 if (list_empty(&(*head)->lru))
695 *head = NULL;
696 else if (*head == page)
697 *head = (struct page *)list_entry((*head)->lru.next,
698 struct page, lru);
699
700 list_del_init(&page->lru);
Minchan Kim248ca1b2015-04-15 16:15:42 -0700701 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
702 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
Nitin Gupta61989a82012-01-09 16:51:56 -0600703}
704
Nitin Cuptac3e3e882013-12-11 11:04:37 +0900705/*
706 * Each size class maintains zspages in different fullness groups depending
707 * on the number of live objects they contain. When allocating or freeing
708 * objects, the fullness status of the page can change, say, from ALMOST_FULL
709 * to ALMOST_EMPTY when freeing an object. This function checks if such
710 * a status change has occurred for the given page and accordingly moves the
711 * page from the freelist of the old fullness group to that of the new
712 * fullness group.
713 */
Minchan Kimc7806262015-04-15 16:15:26 -0700714static enum fullness_group fix_fullness_group(struct size_class *class,
Nitin Gupta61989a82012-01-09 16:51:56 -0600715 struct page *page)
716{
717 int class_idx;
Nitin Gupta61989a82012-01-09 16:51:56 -0600718 enum fullness_group currfg, newfg;
719
720 BUG_ON(!is_first_page(page));
721
722 get_zspage_mapping(page, &class_idx, &currfg);
723 newfg = get_fullness_group(page);
724 if (newfg == currfg)
725 goto out;
726
Nitin Gupta61989a82012-01-09 16:51:56 -0600727 remove_zspage(page, class, currfg);
728 insert_zspage(page, class, newfg);
729 set_zspage_mapping(page, class_idx, newfg);
730
731out:
732 return newfg;
733}
734
735/*
736 * We have to decide on how many pages to link together
737 * to form a zspage for each size class. This is important
738 * to reduce wastage due to unusable space left at end of
739 * each zspage which is given as:
Yinghao Xie888fa3742015-04-15 16:15:49 -0700740 * wastage = Zp % class_size
741 * usage = Zp - wastage
Nitin Gupta61989a82012-01-09 16:51:56 -0600742 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
743 *
744 * For example, for size class of 3/8 * PAGE_SIZE, we should
745 * link together 3 PAGE_SIZE sized pages to form a zspage
746 * since then we can perfectly fit in 8 such objects.
747 */
Minchan Kim2e3b6152012-05-03 15:40:39 +0900748static int get_pages_per_zspage(int class_size)
Nitin Gupta61989a82012-01-09 16:51:56 -0600749{
750 int i, max_usedpc = 0;
751 /* zspage order which gives maximum used size per KB */
752 int max_usedpc_order = 1;
753
Seth Jennings84d4faa2012-03-05 11:33:21 -0600754 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600755 int zspage_size;
756 int waste, usedpc;
757
758 zspage_size = i * PAGE_SIZE;
759 waste = zspage_size % class_size;
760 usedpc = (zspage_size - waste) * 100 / zspage_size;
761
762 if (usedpc > max_usedpc) {
763 max_usedpc = usedpc;
764 max_usedpc_order = i;
765 }
766 }
767
768 return max_usedpc_order;
769}
770
771/*
772 * A single 'zspage' is composed of many system pages which are
773 * linked together using fields in struct page. This function finds
774 * the first/head page, given any component page of a zspage.
775 */
776static struct page *get_first_page(struct page *page)
777{
778 if (is_first_page(page))
779 return page;
780 else
Kirill A. Shutemov32e7ba12015-11-06 16:29:47 -0800781 return (struct page *)page_private(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600782}
783
784static struct page *get_next_page(struct page *page)
785{
786 struct page *next;
787
788 if (is_last_page(page))
789 next = NULL;
790 else if (is_first_page(page))
Sunghan Suhe842b972013-07-12 16:08:13 +0900791 next = (struct page *)page_private(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600792 else
793 next = list_entry(page->lru.next, struct page, lru);
794
795 return next;
796}
797
Olav Haugan67296872013-11-22 09:30:41 -0800798/*
799 * Encode <page, obj_idx> as a single handle value.
Minchan Kim312fcae2015-04-15 16:15:30 -0700800 * We use the least bit of handle for tagging.
Olav Haugan67296872013-11-22 09:30:41 -0800801 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700802static void *location_to_obj(struct page *page, unsigned long obj_idx)
Nitin Gupta61989a82012-01-09 16:51:56 -0600803{
Minchan Kim312fcae2015-04-15 16:15:30 -0700804 unsigned long obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600805
806 if (!page) {
807 BUG_ON(obj_idx);
808 return NULL;
809 }
810
Minchan Kim312fcae2015-04-15 16:15:30 -0700811 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
812 obj |= ((obj_idx) & OBJ_INDEX_MASK);
813 obj <<= OBJ_TAG_BITS;
Nitin Gupta61989a82012-01-09 16:51:56 -0600814
Minchan Kim312fcae2015-04-15 16:15:30 -0700815 return (void *)obj;
Nitin Gupta61989a82012-01-09 16:51:56 -0600816}
817
Olav Haugan67296872013-11-22 09:30:41 -0800818/*
819 * Decode <page, obj_idx> pair from the given object handle. We adjust the
820 * decoded obj_idx back to its original value since it was adjusted in
Minchan Kim312fcae2015-04-15 16:15:30 -0700821 * location_to_obj().
Olav Haugan67296872013-11-22 09:30:41 -0800822 */
Minchan Kim312fcae2015-04-15 16:15:30 -0700823static void obj_to_location(unsigned long obj, struct page **page,
Nitin Gupta61989a82012-01-09 16:51:56 -0600824 unsigned long *obj_idx)
825{
Minchan Kim312fcae2015-04-15 16:15:30 -0700826 obj >>= OBJ_TAG_BITS;
827 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
828 *obj_idx = (obj & OBJ_INDEX_MASK);
Nitin Gupta61989a82012-01-09 16:51:56 -0600829}
830
Minchan Kim2e40e162015-04-15 16:15:23 -0700831static unsigned long handle_to_obj(unsigned long handle)
832{
833 return *(unsigned long *)handle;
834}
835
Minchan Kim7b60a682015-04-15 16:15:39 -0700836static unsigned long obj_to_head(struct size_class *class, struct page *page,
837 void *obj)
Minchan Kim312fcae2015-04-15 16:15:30 -0700838{
Minchan Kim7b60a682015-04-15 16:15:39 -0700839 if (class->huge) {
840 VM_BUG_ON(!is_first_page(page));
Hui Zhu12a7bfa2015-11-06 16:29:26 -0800841 return page_private(page);
Minchan Kim7b60a682015-04-15 16:15:39 -0700842 } else
843 return *(unsigned long *)obj;
Minchan Kim312fcae2015-04-15 16:15:30 -0700844}
845
Nitin Gupta61989a82012-01-09 16:51:56 -0600846static unsigned long obj_idx_to_offset(struct page *page,
847 unsigned long obj_idx, int class_size)
848{
849 unsigned long off = 0;
850
851 if (!is_first_page(page))
852 off = page->index;
853
854 return off + obj_idx * class_size;
855}
856
Minchan Kim312fcae2015-04-15 16:15:30 -0700857static inline int trypin_tag(unsigned long handle)
858{
859 unsigned long *ptr = (unsigned long *)handle;
860
861 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
862}
863
864static void pin_tag(unsigned long handle)
865{
866 while (!trypin_tag(handle));
867}
868
869static void unpin_tag(unsigned long handle)
870{
871 unsigned long *ptr = (unsigned long *)handle;
872
873 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
874}
875
Nitin Guptaf4477e92012-04-02 09:13:56 -0500876static void reset_page(struct page *page)
877{
878 clear_bit(PG_private, &page->flags);
879 clear_bit(PG_private_2, &page->flags);
880 set_page_private(page, 0);
881 page->mapping = NULL;
882 page->freelist = NULL;
Mel Gorman22b751c2013-02-22 16:34:59 -0800883 page_mapcount_reset(page);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500884}
885
Nitin Gupta61989a82012-01-09 16:51:56 -0600886static void free_zspage(struct page *first_page)
887{
Nitin Guptaf4477e92012-04-02 09:13:56 -0500888 struct page *nextp, *tmp, *head_extra;
Nitin Gupta61989a82012-01-09 16:51:56 -0600889
890 BUG_ON(!is_first_page(first_page));
891 BUG_ON(first_page->inuse);
892
Nitin Guptaf4477e92012-04-02 09:13:56 -0500893 head_extra = (struct page *)page_private(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600894
Nitin Guptaf4477e92012-04-02 09:13:56 -0500895 reset_page(first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600896 __free_page(first_page);
897
898 /* zspage with only 1 system page */
Nitin Guptaf4477e92012-04-02 09:13:56 -0500899 if (!head_extra)
Nitin Gupta61989a82012-01-09 16:51:56 -0600900 return;
901
Nitin Guptaf4477e92012-04-02 09:13:56 -0500902 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
Nitin Gupta61989a82012-01-09 16:51:56 -0600903 list_del(&nextp->lru);
Nitin Guptaf4477e92012-04-02 09:13:56 -0500904 reset_page(nextp);
Nitin Gupta61989a82012-01-09 16:51:56 -0600905 __free_page(nextp);
906 }
Nitin Guptaf4477e92012-04-02 09:13:56 -0500907 reset_page(head_extra);
908 __free_page(head_extra);
Nitin Gupta61989a82012-01-09 16:51:56 -0600909}
910
911/* Initialize a newly allocated zspage */
912static void init_zspage(struct page *first_page, struct size_class *class)
913{
914 unsigned long off = 0;
915 struct page *page = first_page;
916
917 BUG_ON(!is_first_page(first_page));
918 while (page) {
919 struct page *next_page;
920 struct link_free *link;
Dan Streetman5538c562014-10-09 15:30:01 -0700921 unsigned int i = 1;
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800922 void *vaddr;
Nitin Gupta61989a82012-01-09 16:51:56 -0600923
924 /*
925 * page->index stores offset of first object starting
926 * in the page. For the first page, this is always 0,
927 * so we use first_page->index (aka ->freelist) to store
928 * head of corresponding zspage's freelist.
929 */
930 if (page != first_page)
931 page->index = off;
932
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800933 vaddr = kmap_atomic(page);
934 link = (struct link_free *)vaddr + off / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600935
Dan Streetman5538c562014-10-09 15:30:01 -0700936 while ((off += class->size) < PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -0700937 link->next = location_to_obj(page, i++);
Dan Streetman5538c562014-10-09 15:30:01 -0700938 link += class->size / sizeof(*link);
Nitin Gupta61989a82012-01-09 16:51:56 -0600939 }
940
941 /*
942 * We now come to the last (full or partial) object on this
943 * page, which must point to the first object on the next
944 * page (if present)
945 */
946 next_page = get_next_page(page);
Minchan Kim312fcae2015-04-15 16:15:30 -0700947 link->next = location_to_obj(next_page, 0);
Minchan Kimaf4ee5e2014-12-12 16:56:58 -0800948 kunmap_atomic(vaddr);
Nitin Gupta61989a82012-01-09 16:51:56 -0600949 page = next_page;
Dan Streetman5538c562014-10-09 15:30:01 -0700950 off %= PAGE_SIZE;
Nitin Gupta61989a82012-01-09 16:51:56 -0600951 }
952}
953
954/*
955 * Allocate a zspage for the given size class
956 */
957static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
958{
959 int i, error;
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500960 struct page *first_page = NULL, *uninitialized_var(prev_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600961
962 /*
963 * Allocate individual pages and link them together as:
964 * 1. first page->private = first sub-page
965 * 2. all sub-pages are linked together using page->lru
Kirill A. Shutemov32e7ba12015-11-06 16:29:47 -0800966 * 3. each sub-page is linked to the first page using page->private
Nitin Gupta61989a82012-01-09 16:51:56 -0600967 *
968 * For each size class, First/Head pages are linked together using
969 * page->lru. Also, we set PG_private to identify the first page
970 * (i.e. no other sub-page has this flag set) and PG_private_2 to
971 * identify the last page.
972 */
973 error = -ENOMEM;
Minchan Kim2e3b6152012-05-03 15:40:39 +0900974 for (i = 0; i < class->pages_per_zspage; i++) {
Seth Jenningsb4b700c2012-06-13 16:03:42 -0500975 struct page *page;
Nitin Gupta61989a82012-01-09 16:51:56 -0600976
977 page = alloc_page(flags);
978 if (!page)
979 goto cleanup;
980
981 INIT_LIST_HEAD(&page->lru);
982 if (i == 0) { /* first page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900983 SetPagePrivate(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600984 set_page_private(page, 0);
985 first_page = page;
986 first_page->inuse = 0;
987 }
988 if (i == 1)
Sunghan Suhe842b972013-07-12 16:08:13 +0900989 set_page_private(first_page, (unsigned long)page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600990 if (i >= 1)
Kirill A. Shutemov32e7ba12015-11-06 16:29:47 -0800991 set_page_private(page, (unsigned long)first_page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600992 if (i >= 2)
993 list_add(&page->lru, &prev_page->lru);
Minchan Kim2e3b6152012-05-03 15:40:39 +0900994 if (i == class->pages_per_zspage - 1) /* last page */
Minchan Kima27545bf2012-04-25 15:23:09 +0900995 SetPagePrivate2(page);
Nitin Gupta61989a82012-01-09 16:51:56 -0600996 prev_page = page;
997 }
998
999 init_zspage(first_page, class);
1000
Minchan Kim312fcae2015-04-15 16:15:30 -07001001 first_page->freelist = location_to_obj(first_page, 0);
Nitin Gupta61989a82012-01-09 16:51:56 -06001002 /* Maximum number of objects we can store in this zspage */
Minchan Kim2e3b6152012-05-03 15:40:39 +09001003 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
Nitin Gupta61989a82012-01-09 16:51:56 -06001004
1005 error = 0; /* Success */
1006
1007cleanup:
1008 if (unlikely(error) && first_page) {
1009 free_zspage(first_page);
1010 first_page = NULL;
1011 }
1012
1013 return first_page;
1014}
1015
1016static struct page *find_get_zspage(struct size_class *class)
1017{
1018 int i;
1019 struct page *page;
1020
1021 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1022 page = class->fullness_list[i];
1023 if (page)
1024 break;
1025 }
1026
1027 return page;
1028}
1029
Minchan Kim1b945ae2013-12-11 11:04:36 +09001030#ifdef CONFIG_PGTABLE_MAPPING
Seth Jenningsf5536462012-07-18 11:55:56 -05001031static inline int __zs_cpu_up(struct mapping_area *area)
Seth Jennings5f601902012-07-02 16:15:49 -05001032{
Seth Jenningsf5536462012-07-18 11:55:56 -05001033 /*
1034 * Make sure we don't leak memory if a cpu UP notification
1035 * and zs_init() race and both call zs_cpu_up() on the same cpu
1036 */
1037 if (area->vm)
1038 return 0;
1039 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1040 if (!area->vm)
1041 return -ENOMEM;
1042 return 0;
1043}
1044
1045static inline void __zs_cpu_down(struct mapping_area *area)
1046{
1047 if (area->vm)
1048 free_vm_area(area->vm);
1049 area->vm = NULL;
1050}
1051
1052static inline void *__zs_map_object(struct mapping_area *area,
1053 struct page *pages[2], int off, int size)
1054{
WANG Chaof6f8ed42014-08-06 16:06:58 -07001055 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
Seth Jenningsf5536462012-07-18 11:55:56 -05001056 area->vm_addr = area->vm->addr;
1057 return area->vm_addr + off;
1058}
1059
1060static inline void __zs_unmap_object(struct mapping_area *area,
1061 struct page *pages[2], int off, int size)
1062{
1063 unsigned long addr = (unsigned long)area->vm_addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001064
Joerg Roedeld95abbb2013-03-27 01:43:14 +01001065 unmap_kernel_range(addr, PAGE_SIZE * 2);
Seth Jenningsf5536462012-07-18 11:55:56 -05001066}
1067
Minchan Kim1b945ae2013-12-11 11:04:36 +09001068#else /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001069
1070static inline int __zs_cpu_up(struct mapping_area *area)
1071{
1072 /*
1073 * Make sure we don't leak memory if a cpu UP notification
1074 * and zs_init() race and both call zs_cpu_up() on the same cpu
1075 */
1076 if (area->vm_buf)
1077 return 0;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001078 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
Seth Jenningsf5536462012-07-18 11:55:56 -05001079 if (!area->vm_buf)
1080 return -ENOMEM;
1081 return 0;
1082}
1083
1084static inline void __zs_cpu_down(struct mapping_area *area)
1085{
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001086 kfree(area->vm_buf);
Seth Jenningsf5536462012-07-18 11:55:56 -05001087 area->vm_buf = NULL;
1088}
1089
1090static void *__zs_map_object(struct mapping_area *area,
1091 struct page *pages[2], int off, int size)
1092{
Seth Jennings5f601902012-07-02 16:15:49 -05001093 int sizes[2];
1094 void *addr;
Seth Jenningsf5536462012-07-18 11:55:56 -05001095 char *buf = area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001096
Seth Jenningsf5536462012-07-18 11:55:56 -05001097 /* disable page faults to match kmap_atomic() return conditions */
1098 pagefault_disable();
1099
1100 /* no read fastpath */
1101 if (area->vm_mm == ZS_MM_WO)
1102 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001103
1104 sizes[0] = PAGE_SIZE - off;
1105 sizes[1] = size - sizes[0];
1106
Seth Jennings5f601902012-07-02 16:15:49 -05001107 /* copy object to per-cpu buffer */
1108 addr = kmap_atomic(pages[0]);
1109 memcpy(buf, addr + off, sizes[0]);
1110 kunmap_atomic(addr);
1111 addr = kmap_atomic(pages[1]);
1112 memcpy(buf + sizes[0], addr, sizes[1]);
1113 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001114out:
1115 return area->vm_buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001116}
1117
Seth Jenningsf5536462012-07-18 11:55:56 -05001118static void __zs_unmap_object(struct mapping_area *area,
1119 struct page *pages[2], int off, int size)
Seth Jennings5f601902012-07-02 16:15:49 -05001120{
Seth Jennings5f601902012-07-02 16:15:49 -05001121 int sizes[2];
1122 void *addr;
Minchan Kim2e40e162015-04-15 16:15:23 -07001123 char *buf;
Seth Jennings5f601902012-07-02 16:15:49 -05001124
Seth Jenningsf5536462012-07-18 11:55:56 -05001125 /* no write fastpath */
1126 if (area->vm_mm == ZS_MM_RO)
1127 goto out;
Seth Jennings5f601902012-07-02 16:15:49 -05001128
Minchan Kim7b60a682015-04-15 16:15:39 -07001129 buf = area->vm_buf;
1130 if (!area->huge) {
1131 buf = buf + ZS_HANDLE_SIZE;
1132 size -= ZS_HANDLE_SIZE;
1133 off += ZS_HANDLE_SIZE;
1134 }
Minchan Kim2e40e162015-04-15 16:15:23 -07001135
Seth Jennings5f601902012-07-02 16:15:49 -05001136 sizes[0] = PAGE_SIZE - off;
1137 sizes[1] = size - sizes[0];
1138
1139 /* copy per-cpu buffer to object */
1140 addr = kmap_atomic(pages[0]);
1141 memcpy(addr + off, buf, sizes[0]);
1142 kunmap_atomic(addr);
1143 addr = kmap_atomic(pages[1]);
1144 memcpy(addr, buf + sizes[0], sizes[1]);
1145 kunmap_atomic(addr);
Seth Jenningsf5536462012-07-18 11:55:56 -05001146
1147out:
1148 /* enable page faults to match kunmap_atomic() return conditions */
1149 pagefault_enable();
Seth Jennings5f601902012-07-02 16:15:49 -05001150}
Nitin Gupta61989a82012-01-09 16:51:56 -06001151
Minchan Kim1b945ae2013-12-11 11:04:36 +09001152#endif /* CONFIG_PGTABLE_MAPPING */
Seth Jenningsf5536462012-07-18 11:55:56 -05001153
Nitin Gupta61989a82012-01-09 16:51:56 -06001154static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1155 void *pcpu)
1156{
Seth Jenningsf5536462012-07-18 11:55:56 -05001157 int ret, cpu = (long)pcpu;
Nitin Gupta61989a82012-01-09 16:51:56 -06001158 struct mapping_area *area;
1159
1160 switch (action) {
1161 case CPU_UP_PREPARE:
1162 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001163 ret = __zs_cpu_up(area);
1164 if (ret)
1165 return notifier_from_errno(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001166 break;
1167 case CPU_DEAD:
1168 case CPU_UP_CANCELED:
1169 area = &per_cpu(zs_map_area, cpu);
Seth Jenningsf5536462012-07-18 11:55:56 -05001170 __zs_cpu_down(area);
Nitin Gupta61989a82012-01-09 16:51:56 -06001171 break;
1172 }
1173
1174 return NOTIFY_OK;
1175}
1176
1177static struct notifier_block zs_cpu_nb = {
1178 .notifier_call = zs_cpu_notifier
1179};
1180
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001181static int zs_register_cpu_notifier(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06001182{
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001183 int cpu, uninitialized_var(ret);
Nitin Gupta61989a82012-01-09 16:51:56 -06001184
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301185 cpu_notifier_register_begin();
1186
1187 __register_cpu_notifier(&zs_cpu_nb);
Nitin Gupta61989a82012-01-09 16:51:56 -06001188 for_each_online_cpu(cpu) {
1189 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001190 if (notifier_to_errno(ret))
1191 break;
Nitin Gupta61989a82012-01-09 16:51:56 -06001192 }
Srivatsa S. Bhatf0e71fc2014-03-11 02:09:59 +05301193
1194 cpu_notifier_register_done();
Sergey Senozhatskyb1b00a52014-12-12 16:56:56 -08001195 return notifier_to_errno(ret);
1196}
1197
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001198static void zs_unregister_cpu_notifier(void)
1199{
1200 int cpu;
1201
1202 cpu_notifier_register_begin();
1203
1204 for_each_online_cpu(cpu)
1205 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1206 __unregister_cpu_notifier(&zs_cpu_nb);
1207
1208 cpu_notifier_register_done();
1209}
1210
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001211static void init_zs_size_classes(void)
1212{
1213 int nr;
1214
1215 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1216 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1217 nr += 1;
1218
1219 zs_size_classes = nr;
1220}
1221
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001222static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1223{
1224 if (prev->pages_per_zspage != pages_per_zspage)
1225 return false;
1226
1227 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1228 != get_maxobj_per_zspage(size, pages_per_zspage))
1229 return false;
1230
1231 return true;
1232}
1233
Minchan Kim312fcae2015-04-15 16:15:30 -07001234static bool zspage_full(struct page *page)
1235{
1236 BUG_ON(!is_first_page(page));
1237
1238 return page->inuse == page->objects;
1239}
1240
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001241unsigned long zs_get_total_pages(struct zs_pool *pool)
1242{
1243 return atomic_long_read(&pool->pages_allocated);
1244}
1245EXPORT_SYMBOL_GPL(zs_get_total_pages);
1246
1247/**
1248 * zs_map_object - get address of allocated object from handle.
1249 * @pool: pool from which the object was allocated
1250 * @handle: handle returned from zs_malloc
1251 *
1252 * Before using an object allocated from zs_malloc, it must be mapped using
1253 * this function. When done with the object, it must be unmapped using
1254 * zs_unmap_object.
1255 *
1256 * Only one object can be mapped per cpu at a time. There is no protection
1257 * against nested mappings.
1258 *
1259 * This function returns with preemption and page faults disabled.
1260 */
1261void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1262 enum zs_mapmode mm)
1263{
1264 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001265 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001266
1267 unsigned int class_idx;
1268 enum fullness_group fg;
1269 struct size_class *class;
1270 struct mapping_area *area;
1271 struct page *pages[2];
Minchan Kim2e40e162015-04-15 16:15:23 -07001272 void *ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001273
1274 BUG_ON(!handle);
1275
1276 /*
1277 * Because we use per-cpu mapping areas shared among the
1278 * pools/users, we can't allow mapping in interrupt context
1279 * because it can corrupt another users mappings.
1280 */
1281 BUG_ON(in_interrupt());
1282
Minchan Kim312fcae2015-04-15 16:15:30 -07001283 /* From now on, migration cannot move the object */
1284 pin_tag(handle);
1285
Minchan Kim2e40e162015-04-15 16:15:23 -07001286 obj = handle_to_obj(handle);
1287 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001288 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1289 class = pool->size_class[class_idx];
1290 off = obj_idx_to_offset(page, obj_idx, class->size);
1291
1292 area = &get_cpu_var(zs_map_area);
1293 area->vm_mm = mm;
1294 if (off + class->size <= PAGE_SIZE) {
1295 /* this object is contained entirely within a page */
1296 area->vm_addr = kmap_atomic(page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001297 ret = area->vm_addr + off;
1298 goto out;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001299 }
1300
1301 /* this object spans two pages */
1302 pages[0] = page;
1303 pages[1] = get_next_page(page);
1304 BUG_ON(!pages[1]);
1305
Minchan Kim2e40e162015-04-15 16:15:23 -07001306 ret = __zs_map_object(area, pages, off, class->size);
1307out:
Minchan Kim7b60a682015-04-15 16:15:39 -07001308 if (!class->huge)
1309 ret += ZS_HANDLE_SIZE;
1310
1311 return ret;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001312}
1313EXPORT_SYMBOL_GPL(zs_map_object);
1314
1315void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1316{
1317 struct page *page;
Minchan Kim2e40e162015-04-15 16:15:23 -07001318 unsigned long obj, obj_idx, off;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001319
1320 unsigned int class_idx;
1321 enum fullness_group fg;
1322 struct size_class *class;
1323 struct mapping_area *area;
1324
1325 BUG_ON(!handle);
1326
Minchan Kim2e40e162015-04-15 16:15:23 -07001327 obj = handle_to_obj(handle);
1328 obj_to_location(obj, &page, &obj_idx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001329 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1330 class = pool->size_class[class_idx];
1331 off = obj_idx_to_offset(page, obj_idx, class->size);
1332
1333 area = this_cpu_ptr(&zs_map_area);
1334 if (off + class->size <= PAGE_SIZE)
1335 kunmap_atomic(area->vm_addr);
1336 else {
1337 struct page *pages[2];
1338
1339 pages[0] = page;
1340 pages[1] = get_next_page(page);
1341 BUG_ON(!pages[1]);
1342
1343 __zs_unmap_object(area, pages, off, class->size);
1344 }
1345 put_cpu_var(zs_map_area);
Minchan Kim312fcae2015-04-15 16:15:30 -07001346 unpin_tag(handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001347}
1348EXPORT_SYMBOL_GPL(zs_unmap_object);
1349
Minchan Kimc7806262015-04-15 16:15:26 -07001350static unsigned long obj_malloc(struct page *first_page,
1351 struct size_class *class, unsigned long handle)
1352{
1353 unsigned long obj;
1354 struct link_free *link;
1355
1356 struct page *m_page;
1357 unsigned long m_objidx, m_offset;
1358 void *vaddr;
1359
Minchan Kim312fcae2015-04-15 16:15:30 -07001360 handle |= OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001361 obj = (unsigned long)first_page->freelist;
1362 obj_to_location(obj, &m_page, &m_objidx);
1363 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1364
1365 vaddr = kmap_atomic(m_page);
1366 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1367 first_page->freelist = link->next;
Minchan Kim7b60a682015-04-15 16:15:39 -07001368 if (!class->huge)
1369 /* record handle in the header of allocated chunk */
1370 link->handle = handle;
1371 else
1372 /* record handle in first_page->private */
1373 set_page_private(first_page, handle);
Minchan Kimc7806262015-04-15 16:15:26 -07001374 kunmap_atomic(vaddr);
1375 first_page->inuse++;
1376 zs_stat_inc(class, OBJ_USED, 1);
1377
1378 return obj;
1379}
1380
1381
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001382/**
1383 * zs_malloc - Allocate block of given size from pool.
1384 * @pool: pool to allocate from
1385 * @size: size of block to allocate
1386 *
1387 * On success, handle to the allocated object is returned,
1388 * otherwise 0.
1389 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1390 */
1391unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1392{
Minchan Kim2e40e162015-04-15 16:15:23 -07001393 unsigned long handle, obj;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001394 struct size_class *class;
Minchan Kimc7806262015-04-15 16:15:26 -07001395 struct page *first_page;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001396
Minchan Kim7b60a682015-04-15 16:15:39 -07001397 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001398 return 0;
1399
Minchan Kim2e40e162015-04-15 16:15:23 -07001400 handle = alloc_handle(pool);
1401 if (!handle)
1402 return 0;
1403
1404 /* extra space in chunk to keep the handle */
1405 size += ZS_HANDLE_SIZE;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001406 class = pool->size_class[get_size_class_index(size)];
1407
1408 spin_lock(&class->lock);
1409 first_page = find_get_zspage(class);
1410
1411 if (!first_page) {
1412 spin_unlock(&class->lock);
1413 first_page = alloc_zspage(class, pool->flags);
Minchan Kim2e40e162015-04-15 16:15:23 -07001414 if (unlikely(!first_page)) {
1415 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001416 return 0;
Minchan Kim2e40e162015-04-15 16:15:23 -07001417 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001418
1419 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1420 atomic_long_add(class->pages_per_zspage,
1421 &pool->pages_allocated);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001422
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001423 spin_lock(&class->lock);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001424 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1425 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001426 }
1427
Minchan Kimc7806262015-04-15 16:15:26 -07001428 obj = obj_malloc(first_page, class, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001429 /* Now move the zspage to another fullness group, if required */
Minchan Kimc7806262015-04-15 16:15:26 -07001430 fix_fullness_group(class, first_page);
Minchan Kim2e40e162015-04-15 16:15:23 -07001431 record_obj(handle, obj);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001432 spin_unlock(&class->lock);
1433
Minchan Kim2e40e162015-04-15 16:15:23 -07001434 return handle;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001435}
1436EXPORT_SYMBOL_GPL(zs_malloc);
1437
Minchan Kimc7806262015-04-15 16:15:26 -07001438static void obj_free(struct zs_pool *pool, struct size_class *class,
1439 unsigned long obj)
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001440{
1441 struct link_free *link;
1442 struct page *first_page, *f_page;
Minchan Kimc7806262015-04-15 16:15:26 -07001443 unsigned long f_objidx, f_offset;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001444 void *vaddr;
1445
Minchan Kimc7806262015-04-15 16:15:26 -07001446 BUG_ON(!obj);
1447
Minchan Kim312fcae2015-04-15 16:15:30 -07001448 obj &= ~OBJ_ALLOCATED_TAG;
Minchan Kimc7806262015-04-15 16:15:26 -07001449 obj_to_location(obj, &f_page, &f_objidx);
1450 first_page = get_first_page(f_page);
1451
Minchan Kimc7806262015-04-15 16:15:26 -07001452 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1453
1454 vaddr = kmap_atomic(f_page);
1455
1456 /* Insert this object in containing zspage's freelist */
1457 link = (struct link_free *)(vaddr + f_offset);
1458 link->next = first_page->freelist;
Minchan Kim7b60a682015-04-15 16:15:39 -07001459 if (class->huge)
1460 set_page_private(first_page, 0);
Minchan Kimc7806262015-04-15 16:15:26 -07001461 kunmap_atomic(vaddr);
1462 first_page->freelist = (void *)obj;
1463 first_page->inuse--;
1464 zs_stat_dec(class, OBJ_USED, 1);
1465}
1466
1467void zs_free(struct zs_pool *pool, unsigned long handle)
1468{
1469 struct page *first_page, *f_page;
1470 unsigned long obj, f_objidx;
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001471 int class_idx;
1472 struct size_class *class;
1473 enum fullness_group fullness;
1474
Minchan Kim2e40e162015-04-15 16:15:23 -07001475 if (unlikely(!handle))
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001476 return;
1477
Minchan Kim312fcae2015-04-15 16:15:30 -07001478 pin_tag(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001479 obj = handle_to_obj(handle);
Minchan Kim2e40e162015-04-15 16:15:23 -07001480 obj_to_location(obj, &f_page, &f_objidx);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001481 first_page = get_first_page(f_page);
1482
1483 get_zspage_mapping(first_page, &class_idx, &fullness);
1484 class = pool->size_class[class_idx];
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001485
1486 spin_lock(&class->lock);
Minchan Kimc7806262015-04-15 16:15:26 -07001487 obj_free(pool, class, obj);
1488 fullness = fix_fullness_group(class, first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001489 if (fullness == ZS_EMPTY) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001490 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1491 class->size, class->pages_per_zspage));
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001492 atomic_long_sub(class->pages_per_zspage,
1493 &pool->pages_allocated);
1494 free_zspage(first_page);
1495 }
Minchan Kim312fcae2015-04-15 16:15:30 -07001496 spin_unlock(&class->lock);
1497 unpin_tag(handle);
1498
1499 free_handle(pool, handle);
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08001500}
1501EXPORT_SYMBOL_GPL(zs_free);
1502
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001503static void zs_object_copy(unsigned long dst, unsigned long src,
Minchan Kim312fcae2015-04-15 16:15:30 -07001504 struct size_class *class)
1505{
1506 struct page *s_page, *d_page;
1507 unsigned long s_objidx, d_objidx;
1508 unsigned long s_off, d_off;
1509 void *s_addr, *d_addr;
1510 int s_size, d_size, size;
1511 int written = 0;
1512
1513 s_size = d_size = class->size;
1514
1515 obj_to_location(src, &s_page, &s_objidx);
1516 obj_to_location(dst, &d_page, &d_objidx);
1517
1518 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1519 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1520
1521 if (s_off + class->size > PAGE_SIZE)
1522 s_size = PAGE_SIZE - s_off;
1523
1524 if (d_off + class->size > PAGE_SIZE)
1525 d_size = PAGE_SIZE - d_off;
1526
1527 s_addr = kmap_atomic(s_page);
1528 d_addr = kmap_atomic(d_page);
1529
1530 while (1) {
1531 size = min(s_size, d_size);
1532 memcpy(d_addr + d_off, s_addr + s_off, size);
1533 written += size;
1534
1535 if (written == class->size)
1536 break;
1537
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001538 s_off += size;
1539 s_size -= size;
1540 d_off += size;
1541 d_size -= size;
1542
1543 if (s_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001544 kunmap_atomic(d_addr);
1545 kunmap_atomic(s_addr);
1546 s_page = get_next_page(s_page);
1547 BUG_ON(!s_page);
1548 s_addr = kmap_atomic(s_page);
1549 d_addr = kmap_atomic(d_page);
1550 s_size = class->size - written;
1551 s_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001552 }
1553
Sergey Senozhatsky495819e2015-04-15 16:16:15 -07001554 if (d_off >= PAGE_SIZE) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001555 kunmap_atomic(d_addr);
1556 d_page = get_next_page(d_page);
1557 BUG_ON(!d_page);
1558 d_addr = kmap_atomic(d_page);
1559 d_size = class->size - written;
1560 d_off = 0;
Minchan Kim312fcae2015-04-15 16:15:30 -07001561 }
1562 }
1563
1564 kunmap_atomic(d_addr);
1565 kunmap_atomic(s_addr);
1566}
1567
1568/*
1569 * Find alloced object in zspage from index object and
1570 * return handle.
1571 */
1572static unsigned long find_alloced_obj(struct page *page, int index,
1573 struct size_class *class)
1574{
1575 unsigned long head;
1576 int offset = 0;
1577 unsigned long handle = 0;
1578 void *addr = kmap_atomic(page);
1579
1580 if (!is_first_page(page))
1581 offset = page->index;
1582 offset += class->size * index;
1583
1584 while (offset < PAGE_SIZE) {
Minchan Kim7b60a682015-04-15 16:15:39 -07001585 head = obj_to_head(class, page, addr + offset);
Minchan Kim312fcae2015-04-15 16:15:30 -07001586 if (head & OBJ_ALLOCATED_TAG) {
1587 handle = head & ~OBJ_ALLOCATED_TAG;
1588 if (trypin_tag(handle))
1589 break;
1590 handle = 0;
1591 }
1592
1593 offset += class->size;
1594 index++;
1595 }
1596
1597 kunmap_atomic(addr);
1598 return handle;
1599}
1600
1601struct zs_compact_control {
1602 /* Source page for migration which could be a subpage of zspage. */
1603 struct page *s_page;
1604 /* Destination page for migration which should be a first page
1605 * of zspage. */
1606 struct page *d_page;
1607 /* Starting object index within @s_page which used for live object
1608 * in the subpage. */
1609 int index;
Minchan Kim312fcae2015-04-15 16:15:30 -07001610};
1611
1612static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1613 struct zs_compact_control *cc)
1614{
1615 unsigned long used_obj, free_obj;
1616 unsigned long handle;
1617 struct page *s_page = cc->s_page;
1618 struct page *d_page = cc->d_page;
1619 unsigned long index = cc->index;
Minchan Kim312fcae2015-04-15 16:15:30 -07001620 int ret = 0;
1621
1622 while (1) {
1623 handle = find_alloced_obj(s_page, index, class);
1624 if (!handle) {
1625 s_page = get_next_page(s_page);
1626 if (!s_page)
1627 break;
1628 index = 0;
1629 continue;
1630 }
1631
1632 /* Stop if there is no more space */
1633 if (zspage_full(d_page)) {
1634 unpin_tag(handle);
1635 ret = -ENOMEM;
1636 break;
1637 }
1638
1639 used_obj = handle_to_obj(handle);
1640 free_obj = obj_malloc(d_page, class, handle);
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001641 zs_object_copy(free_obj, used_obj, class);
Minchan Kim312fcae2015-04-15 16:15:30 -07001642 index++;
Junil Leec102f072016-01-20 14:58:18 -08001643 /*
1644 * record_obj updates handle's value to free_obj and it will
1645 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1646 * breaks synchronization using pin_tag(e,g, zs_free) so
1647 * let's keep the lock bit.
1648 */
1649 free_obj |= BIT(HANDLE_PIN_BIT);
Minchan Kim312fcae2015-04-15 16:15:30 -07001650 record_obj(handle, free_obj);
1651 unpin_tag(handle);
1652 obj_free(pool, class, used_obj);
Minchan Kim312fcae2015-04-15 16:15:30 -07001653 }
1654
1655 /* Remember last position in this iteration */
1656 cc->s_page = s_page;
1657 cc->index = index;
Minchan Kim312fcae2015-04-15 16:15:30 -07001658
1659 return ret;
1660}
1661
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001662static struct page *isolate_target_page(struct size_class *class)
Minchan Kim312fcae2015-04-15 16:15:30 -07001663{
1664 int i;
1665 struct page *page;
1666
1667 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1668 page = class->fullness_list[i];
1669 if (page) {
1670 remove_zspage(page, class, i);
1671 break;
1672 }
1673 }
1674
1675 return page;
1676}
1677
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001678/*
1679 * putback_zspage - add @first_page into right class's fullness list
1680 * @pool: target pool
1681 * @class: destination class
1682 * @first_page: target page
1683 *
1684 * Return @fist_page's fullness_group
1685 */
1686static enum fullness_group putback_zspage(struct zs_pool *pool,
1687 struct size_class *class,
1688 struct page *first_page)
Minchan Kim312fcae2015-04-15 16:15:30 -07001689{
Minchan Kim312fcae2015-04-15 16:15:30 -07001690 enum fullness_group fullness;
1691
1692 BUG_ON(!is_first_page(first_page));
1693
Minchan Kim839373e2015-04-15 16:16:18 -07001694 fullness = get_fullness_group(first_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001695 insert_zspage(first_page, class, fullness);
Minchan Kim839373e2015-04-15 16:16:18 -07001696 set_zspage_mapping(first_page, class->index, fullness);
1697
Minchan Kim312fcae2015-04-15 16:15:30 -07001698 if (fullness == ZS_EMPTY) {
1699 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1700 class->size, class->pages_per_zspage));
1701 atomic_long_sub(class->pages_per_zspage,
1702 &pool->pages_allocated);
1703
1704 free_zspage(first_page);
1705 }
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001706
1707 return fullness;
Minchan Kim312fcae2015-04-15 16:15:30 -07001708}
1709
1710static struct page *isolate_source_page(struct size_class *class)
1711{
Minchan Kimad9d5e12015-09-08 15:04:47 -07001712 int i;
1713 struct page *page = NULL;
Minchan Kim312fcae2015-04-15 16:15:30 -07001714
Minchan Kimad9d5e12015-09-08 15:04:47 -07001715 for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) {
1716 page = class->fullness_list[i];
1717 if (!page)
1718 continue;
1719
1720 remove_zspage(page, class, i);
1721 break;
1722 }
Minchan Kim312fcae2015-04-15 16:15:30 -07001723
1724 return page;
1725}
1726
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001727/*
1728 *
1729 * Based on the number of unused allocated objects calculate
1730 * and return the number of pages that we can free.
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001731 */
1732static unsigned long zs_can_compact(struct size_class *class)
1733{
1734 unsigned long obj_wasted;
1735
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001736 obj_wasted = zs_stat_get(class, OBJ_ALLOCATED) -
1737 zs_stat_get(class, OBJ_USED);
1738
1739 obj_wasted /= get_maxobj_per_zspage(class->size,
1740 class->pages_per_zspage);
1741
Minchan Kim6cbf16b2015-09-08 15:04:49 -07001742 return obj_wasted * class->pages_per_zspage;
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001743}
1744
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001745static void __zs_compact(struct zs_pool *pool, struct size_class *class)
Minchan Kim312fcae2015-04-15 16:15:30 -07001746{
Minchan Kim312fcae2015-04-15 16:15:30 -07001747 struct zs_compact_control cc;
1748 struct page *src_page;
1749 struct page *dst_page = NULL;
Minchan Kim312fcae2015-04-15 16:15:30 -07001750
Minchan Kim312fcae2015-04-15 16:15:30 -07001751 spin_lock(&class->lock);
1752 while ((src_page = isolate_source_page(class))) {
1753
1754 BUG_ON(!is_first_page(src_page));
1755
Sergey Senozhatsky04f05902015-09-08 15:04:30 -07001756 if (!zs_can_compact(class))
1757 break;
1758
Minchan Kim312fcae2015-04-15 16:15:30 -07001759 cc.index = 0;
1760 cc.s_page = src_page;
1761
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001762 while ((dst_page = isolate_target_page(class))) {
Minchan Kim312fcae2015-04-15 16:15:30 -07001763 cc.d_page = dst_page;
1764 /*
Sergey Senozhatsky0dc63d42015-09-08 15:04:33 -07001765 * If there is no more space in dst_page, resched
1766 * and see if anyone had allocated another zspage.
Minchan Kim312fcae2015-04-15 16:15:30 -07001767 */
1768 if (!migrate_zspage(pool, class, &cc))
1769 break;
1770
1771 putback_zspage(pool, class, dst_page);
Minchan Kim312fcae2015-04-15 16:15:30 -07001772 }
1773
1774 /* Stop if we couldn't find slot */
1775 if (dst_page == NULL)
1776 break;
1777
1778 putback_zspage(pool, class, dst_page);
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001779 if (putback_zspage(pool, class, src_page) == ZS_EMPTY)
Minchan Kim6cbf16b2015-09-08 15:04:49 -07001780 pool->stats.pages_compacted += class->pages_per_zspage;
Minchan Kim312fcae2015-04-15 16:15:30 -07001781 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07001782 cond_resched();
1783 spin_lock(&class->lock);
1784 }
1785
1786 if (src_page)
1787 putback_zspage(pool, class, src_page);
1788
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001789 spin_unlock(&class->lock);
Minchan Kim312fcae2015-04-15 16:15:30 -07001790}
1791
1792unsigned long zs_compact(struct zs_pool *pool)
1793{
1794 int i;
Minchan Kim312fcae2015-04-15 16:15:30 -07001795 struct size_class *class;
1796
1797 for (i = zs_size_classes - 1; i >= 0; i--) {
1798 class = pool->size_class[i];
1799 if (!class)
1800 continue;
1801 if (class->index != i)
1802 continue;
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001803 __zs_compact(pool, class);
Minchan Kim312fcae2015-04-15 16:15:30 -07001804 }
1805
Sergey Senozhatsky860c7072015-09-08 15:04:38 -07001806 return pool->stats.pages_compacted;
Minchan Kim312fcae2015-04-15 16:15:30 -07001807}
1808EXPORT_SYMBOL_GPL(zs_compact);
1809
Sergey Senozhatsky7d3f3932015-09-08 15:04:35 -07001810void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
1811{
1812 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
1813}
1814EXPORT_SYMBOL_GPL(zs_pool_stats);
1815
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001816static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
1817 struct shrink_control *sc)
1818{
1819 unsigned long pages_freed;
1820 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1821 shrinker);
1822
1823 pages_freed = pool->stats.pages_compacted;
1824 /*
1825 * Compact classes and calculate compaction delta.
1826 * Can run concurrently with a manually triggered
1827 * (by user) compaction.
1828 */
1829 pages_freed = zs_compact(pool) - pages_freed;
1830
1831 return pages_freed ? pages_freed : SHRINK_STOP;
1832}
1833
1834static unsigned long zs_shrinker_count(struct shrinker *shrinker,
1835 struct shrink_control *sc)
1836{
1837 int i;
1838 struct size_class *class;
1839 unsigned long pages_to_free = 0;
1840 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1841 shrinker);
1842
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001843 for (i = zs_size_classes - 1; i >= 0; i--) {
1844 class = pool->size_class[i];
1845 if (!class)
1846 continue;
1847 if (class->index != i)
1848 continue;
1849
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001850 pages_to_free += zs_can_compact(class);
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001851 }
1852
1853 return pages_to_free;
1854}
1855
1856static void zs_unregister_shrinker(struct zs_pool *pool)
1857{
1858 if (pool->shrinker_enabled) {
1859 unregister_shrinker(&pool->shrinker);
1860 pool->shrinker_enabled = false;
1861 }
1862}
1863
1864static int zs_register_shrinker(struct zs_pool *pool)
1865{
1866 pool->shrinker.scan_objects = zs_shrinker_scan;
1867 pool->shrinker.count_objects = zs_shrinker_count;
1868 pool->shrinker.batch = 0;
1869 pool->shrinker.seeks = DEFAULT_SEEKS;
1870
1871 return register_shrinker(&pool->shrinker);
1872}
1873
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001874/**
1875 * zs_create_pool - Creates an allocation pool to work from.
Seth Jennings0d145a52013-01-30 09:36:52 -06001876 * @flags: allocation flags used to allocate pool metadata
Davidlohr Bueso4bbc0bc2013-01-04 12:14:00 -08001877 *
1878 * This function must be called before anything when using
1879 * the zsmalloc allocator.
1880 *
1881 * On success, a pointer to the newly created pool is returned,
1882 * otherwise NULL.
1883 */
Sergey SENOZHATSKY6f3526d2015-11-06 16:29:21 -08001884struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
Nitin Gupta61989a82012-01-09 16:51:56 -06001885{
Ganesh Mahendran18136652014-12-12 16:57:10 -08001886 int i;
Nitin Gupta61989a82012-01-09 16:51:56 -06001887 struct zs_pool *pool;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001888 struct size_class *prev_class = NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001889
Ganesh Mahendran18136652014-12-12 16:57:10 -08001890 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
Nitin Gupta61989a82012-01-09 16:51:56 -06001891 if (!pool)
1892 return NULL;
1893
Minchan Kim2e40e162015-04-15 16:15:23 -07001894 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1895 GFP_KERNEL);
1896 if (!pool->size_class) {
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001897 kfree(pool);
1898 return NULL;
1899 }
1900
Minchan Kim2e40e162015-04-15 16:15:23 -07001901 pool->name = kstrdup(name, GFP_KERNEL);
1902 if (!pool->name)
1903 goto err;
1904
1905 if (create_handle_cache(pool))
1906 goto err;
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001907
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001908 /*
1909 * Iterate reversly, because, size of size_class that we want to use
1910 * for merging should be larger or equal to current size.
1911 */
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001912 for (i = zs_size_classes - 1; i >= 0; i--) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001913 int size;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001914 int pages_per_zspage;
Nitin Gupta61989a82012-01-09 16:51:56 -06001915 struct size_class *class;
1916
1917 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1918 if (size > ZS_MAX_ALLOC_SIZE)
1919 size = ZS_MAX_ALLOC_SIZE;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001920 pages_per_zspage = get_pages_per_zspage(size);
Nitin Gupta61989a82012-01-09 16:51:56 -06001921
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001922 /*
1923 * size_class is used for normal zsmalloc operation such
1924 * as alloc/free for that size. Although it is natural that we
1925 * have one size_class for each size, there is a chance that we
1926 * can get more memory utilization if we use one size_class for
1927 * many different sizes whose size_class have same
1928 * characteristics. So, we makes size_class point to
1929 * previous size_class if possible.
1930 */
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001931 if (prev_class) {
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001932 if (can_merge(prev_class, size, pages_per_zspage)) {
1933 pool->size_class[i] = prev_class;
1934 continue;
1935 }
1936 }
1937
1938 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1939 if (!class)
1940 goto err;
1941
Nitin Gupta61989a82012-01-09 16:51:56 -06001942 class->size = size;
1943 class->index = i;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001944 class->pages_per_zspage = pages_per_zspage;
Minchan Kim7b60a682015-04-15 16:15:39 -07001945 if (pages_per_zspage == 1 &&
1946 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1947 class->huge = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06001948 spin_lock_init(&class->lock);
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001949 pool->size_class[i] = class;
Ganesh Mahendrandf8b5bb2014-12-12 16:57:07 -08001950
1951 prev_class = class;
Nitin Gupta61989a82012-01-09 16:51:56 -06001952 }
1953
Nitin Gupta61989a82012-01-09 16:51:56 -06001954 pool->flags = flags;
Nitin Gupta61989a82012-01-09 16:51:56 -06001955
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001956 if (zs_pool_stat_create(name, pool))
1957 goto err;
1958
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001959 /*
1960 * Not critical, we still can use the pool
1961 * and user can trigger compaction manually.
1962 */
1963 if (zs_register_shrinker(pool) == 0)
1964 pool->shrinker_enabled = true;
Nitin Gupta61989a82012-01-09 16:51:56 -06001965 return pool;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001966
1967err:
1968 zs_destroy_pool(pool);
1969 return NULL;
Nitin Gupta61989a82012-01-09 16:51:56 -06001970}
1971EXPORT_SYMBOL_GPL(zs_create_pool);
1972
1973void zs_destroy_pool(struct zs_pool *pool)
1974{
1975 int i;
1976
Sergey Senozhatskyab9d3062015-09-08 15:04:41 -07001977 zs_unregister_shrinker(pool);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08001978 zs_pool_stat_destroy(pool);
1979
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001980 for (i = 0; i < zs_size_classes; i++) {
Nitin Gupta61989a82012-01-09 16:51:56 -06001981 int fg;
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001982 struct size_class *class = pool->size_class[i];
1983
1984 if (!class)
1985 continue;
1986
1987 if (class->index != i)
1988 continue;
Nitin Gupta61989a82012-01-09 16:51:56 -06001989
1990 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1991 if (class->fullness_list[fg]) {
Marlies Ruck93ad5ab2013-05-15 16:56:49 -04001992 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
Nitin Gupta61989a82012-01-09 16:51:56 -06001993 class->size, fg);
1994 }
1995 }
Joonsoo Kim9eec4cd2014-12-12 16:56:44 -08001996 kfree(class);
Nitin Gupta61989a82012-01-09 16:51:56 -06001997 }
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08001998
Minchan Kim2e40e162015-04-15 16:15:23 -07001999 destroy_handle_cache(pool);
Mahendran Ganesh40f9fb82014-12-12 16:57:01 -08002000 kfree(pool->size_class);
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002001 kfree(pool->name);
Nitin Gupta61989a82012-01-09 16:51:56 -06002002 kfree(pool);
2003}
2004EXPORT_SYMBOL_GPL(zs_destroy_pool);
2005
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002006static int __init zs_init(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002007{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002008 int ret = zs_register_cpu_notifier();
Nitin Gupta61989a82012-01-09 16:51:56 -06002009
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002010 if (ret)
2011 goto notifier_fail;
Nitin Gupta61989a82012-01-09 16:51:56 -06002012
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002013 init_zs_size_classes();
Nitin Gupta61989a82012-01-09 16:51:56 -06002014
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002015#ifdef CONFIG_ZPOOL
2016 zpool_register_driver(&zs_zpool_driver);
2017#endif
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002018
2019 ret = zs_stat_init();
2020 if (ret) {
2021 pr_err("zs stat initialization failed\n");
2022 goto stat_fail;
2023 }
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002024 return 0;
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002025
2026stat_fail:
2027#ifdef CONFIG_ZPOOL
2028 zpool_unregister_driver(&zs_zpool_driver);
2029#endif
2030notifier_fail:
2031 zs_unregister_cpu_notifier();
2032
2033 return ret;
Nitin Gupta61989a82012-01-09 16:51:56 -06002034}
Nitin Gupta61989a82012-01-09 16:51:56 -06002035
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002036static void __exit zs_exit(void)
Nitin Gupta61989a82012-01-09 16:51:56 -06002037{
Ganesh Mahendran66cdef62014-12-18 16:17:40 -08002038#ifdef CONFIG_ZPOOL
2039 zpool_unregister_driver(&zs_zpool_driver);
2040#endif
2041 zs_unregister_cpu_notifier();
Ganesh Mahendran0f050d92015-02-12 15:00:54 -08002042
2043 zs_stat_exit();
Nitin Gupta61989a82012-01-09 16:51:56 -06002044}
Ben Hutchings069f1012012-06-20 02:31:11 +01002045
2046module_init(zs_init);
2047module_exit(zs_exit);
2048
2049MODULE_LICENSE("Dual BSD/GPL");
2050MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");