2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->index: links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
23 * page->page_type: first object offset in a subpage of zspage
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_owner_priv_1: identifies the huge component page
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
40 #include <linux/module.h>
41 #include <linux/kernel.h>
42 #include <linux/sched.h>
43 #include <linux/bitops.h>
44 #include <linux/errno.h>
45 #include <linux/highmem.h>
46 #include <linux/string.h>
47 #include <linux/slab.h>
48 #include <linux/pgtable.h>
49 #include <asm/tlbflush.h>
50 #include <linux/cpumask.h>
51 #include <linux/cpu.h>
52 #include <linux/vmalloc.h>
53 #include <linux/preempt.h>
54 #include <linux/spinlock.h>
55 #include <linux/shrinker.h>
56 #include <linux/types.h>
57 #include <linux/debugfs.h>
58 #include <linux/zsmalloc.h>
59 #include <linux/zpool.h>
60 #include <linux/migrate.h>
61 #include <linux/wait.h>
62 #include <linux/pagemap.h>
64 #include <linux/local_lock.h>
66 #define ZSPAGE_MAGIC 0x58
69 * This must be power of 2 and greater than or 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.
76 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
79 * Object location (<PFN>, <obj_idx>) is encoded as
80 * a single (unsigned long) handle value.
82 * Note that object index <obj_idx> starts from 0.
84 * This is made more complicated by various memory models and PAE.
87 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
88 #ifdef MAX_PHYSMEM_BITS
89 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
92 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
95 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
99 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
102 * Head in allocated object should have OBJ_ALLOCATED_TAG
103 * to identify the object was allocated or not.
104 * It's okay to add the status bit in the least bit because
105 * header keeps handle which is 4byte-aligned address so we
106 * have room for two bit at least.
108 #define OBJ_ALLOCATED_TAG 1
110 #define OBJ_TAG_BITS 1
111 #define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
113 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
114 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
117 #define FULLNESS_BITS 4
119 #define ISOLATED_BITS 5
120 #define MAGIC_VAL_BITS 8
122 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
124 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
126 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
127 #define ZS_MIN_ALLOC_SIZE \
128 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
129 /* each chunk includes extra space to keep handle */
130 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
133 * On systems with 4K page size, this gives 255 size classes! There is a
135 * - Large number of size classes is potentially wasteful as free page are
136 * spread across these classes
137 * - Small number of size classes causes large internal fragmentation
138 * - Probably its better to use specific size classes (empirically
139 * determined). NOTE: all those class sizes must be set as multiple of
140 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
142 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
145 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
146 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
147 ZS_SIZE_CLASS_DELTA) + 1)
150 * Pages are distinguished by the ratio of used memory (that is the ratio
151 * of ->inuse objects to all objects that page can store). For example,
152 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
154 * The number of fullness groups is not random. It allows us to keep
155 * difference between the least busy page in the group (minimum permitted
156 * number of ->inuse objects) and the most busy page (maximum permitted
157 * number of ->inuse objects) at a reasonable value.
159 enum fullness_group
{
162 /* NOTE: 8 more fullness groups here */
163 ZS_INUSE_RATIO_99
= 10,
168 enum class_stat_type
{
169 /* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
170 ZS_OBJS_ALLOCATED
= NR_FULLNESS_GROUPS
,
175 struct zs_size_stat
{
176 unsigned long objs
[NR_CLASS_STAT_TYPES
];
179 #ifdef CONFIG_ZSMALLOC_STAT
180 static struct dentry
*zs_stat_root
;
183 static size_t huge_class_size
;
186 struct list_head fullness_list
[NR_FULLNESS_GROUPS
];
188 * Size of objects stored in this class. Must be multiple
193 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
194 int pages_per_zspage
;
197 struct zs_size_stat stats
;
201 * Placed within free objects to form a singly linked list.
202 * For every zspage, zspage->freeobj gives head of this list.
204 * This must be power of 2 and less than or equal to ZS_ALIGN
210 * It's valid for non-allocated object
214 * Handle of allocated object.
216 unsigned long handle
;
223 struct size_class
*size_class
[ZS_SIZE_CLASSES
];
224 struct kmem_cache
*handle_cachep
;
225 struct kmem_cache
*zspage_cachep
;
227 atomic_long_t pages_allocated
;
229 struct zs_pool_stats stats
;
231 /* Compact classes */
232 struct shrinker
*shrinker
;
234 #ifdef CONFIG_ZSMALLOC_STAT
235 struct dentry
*stat_dentry
;
237 #ifdef CONFIG_COMPACTION
238 struct work_struct free_work
;
241 atomic_t compaction_in_progress
;
246 unsigned int huge
:HUGE_BITS
;
247 unsigned int fullness
:FULLNESS_BITS
;
248 unsigned int class:CLASS_BITS
+ 1;
249 unsigned int isolated
:ISOLATED_BITS
;
250 unsigned int magic
:MAGIC_VAL_BITS
;
253 unsigned int freeobj
;
254 struct page
*first_page
;
255 struct list_head list
; /* fullness list */
256 struct zs_pool
*pool
;
260 struct mapping_area
{
262 char *vm_buf
; /* copy buffer for objects that span pages */
263 char *vm_addr
; /* address of kmap_atomic()'ed pages */
264 enum zs_mapmode vm_mm
; /* mapping mode */
267 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
268 static void SetZsHugePage(struct zspage
*zspage
)
273 static bool ZsHugePage(struct zspage
*zspage
)
278 static void migrate_lock_init(struct zspage
*zspage
);
279 static void migrate_read_lock(struct zspage
*zspage
);
280 static void migrate_read_unlock(struct zspage
*zspage
);
282 #ifdef CONFIG_COMPACTION
283 static void migrate_write_lock(struct zspage
*zspage
);
284 static void migrate_write_lock_nested(struct zspage
*zspage
);
285 static void migrate_write_unlock(struct zspage
*zspage
);
286 static void kick_deferred_free(struct zs_pool
*pool
);
287 static void init_deferred_free(struct zs_pool
*pool
);
288 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
);
290 static void migrate_write_lock(struct zspage
*zspage
) {}
291 static void migrate_write_lock_nested(struct zspage
*zspage
) {}
292 static void migrate_write_unlock(struct zspage
*zspage
) {}
293 static void kick_deferred_free(struct zs_pool
*pool
) {}
294 static void init_deferred_free(struct zs_pool
*pool
) {}
295 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
) {}
298 static int create_cache(struct zs_pool
*pool
)
300 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
302 if (!pool
->handle_cachep
)
305 pool
->zspage_cachep
= kmem_cache_create("zspage", sizeof(struct zspage
),
307 if (!pool
->zspage_cachep
) {
308 kmem_cache_destroy(pool
->handle_cachep
);
309 pool
->handle_cachep
= NULL
;
316 static void destroy_cache(struct zs_pool
*pool
)
318 kmem_cache_destroy(pool
->handle_cachep
);
319 kmem_cache_destroy(pool
->zspage_cachep
);
322 static unsigned long cache_alloc_handle(struct zs_pool
*pool
, gfp_t gfp
)
324 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
325 gfp
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
328 static void cache_free_handle(struct zs_pool
*pool
, unsigned long handle
)
330 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
333 static struct zspage
*cache_alloc_zspage(struct zs_pool
*pool
, gfp_t flags
)
335 return kmem_cache_zalloc(pool
->zspage_cachep
,
336 flags
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
339 static void cache_free_zspage(struct zs_pool
*pool
, struct zspage
*zspage
)
341 kmem_cache_free(pool
->zspage_cachep
, zspage
);
344 /* pool->lock(which owns the handle) synchronizes races */
345 static void record_obj(unsigned long handle
, unsigned long obj
)
347 *(unsigned long *)handle
= obj
;
354 static void *zs_zpool_create(const char *name
, gfp_t gfp
)
357 * Ignore global gfp flags: zs_malloc() may be invoked from
358 * different contexts and its caller must provide a valid
361 return zs_create_pool(name
);
364 static void zs_zpool_destroy(void *pool
)
366 zs_destroy_pool(pool
);
369 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
370 unsigned long *handle
)
372 *handle
= zs_malloc(pool
, size
, gfp
);
374 if (IS_ERR_VALUE(*handle
))
375 return PTR_ERR((void *)*handle
);
378 static void zs_zpool_free(void *pool
, unsigned long handle
)
380 zs_free(pool
, handle
);
383 static void *zs_zpool_map(void *pool
, unsigned long handle
,
384 enum zpool_mapmode mm
)
386 enum zs_mapmode zs_mm
;
401 return zs_map_object(pool
, handle
, zs_mm
);
403 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
405 zs_unmap_object(pool
, handle
);
408 static u64
zs_zpool_total_size(void *pool
)
410 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
413 static struct zpool_driver zs_zpool_driver
= {
415 .owner
= THIS_MODULE
,
416 .create
= zs_zpool_create
,
417 .destroy
= zs_zpool_destroy
,
418 .malloc_support_movable
= true,
419 .malloc
= zs_zpool_malloc
,
420 .free
= zs_zpool_free
,
422 .unmap
= zs_zpool_unmap
,
423 .total_size
= zs_zpool_total_size
,
426 MODULE_ALIAS("zpool-zsmalloc");
427 #endif /* CONFIG_ZPOOL */
429 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
430 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
) = {
431 .lock
= INIT_LOCAL_LOCK(lock
),
434 static __maybe_unused
int is_first_page(struct page
*page
)
436 return PagePrivate(page
);
439 /* Protected by pool->lock */
440 static inline int get_zspage_inuse(struct zspage
*zspage
)
442 return zspage
->inuse
;
446 static inline void mod_zspage_inuse(struct zspage
*zspage
, int val
)
448 zspage
->inuse
+= val
;
451 static inline struct page
*get_first_page(struct zspage
*zspage
)
453 struct page
*first_page
= zspage
->first_page
;
455 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
459 static inline unsigned int get_first_obj_offset(struct page
*page
)
461 return page
->page_type
;
464 static inline void set_first_obj_offset(struct page
*page
, unsigned int offset
)
466 page
->page_type
= offset
;
469 static inline unsigned int get_freeobj(struct zspage
*zspage
)
471 return zspage
->freeobj
;
474 static inline void set_freeobj(struct zspage
*zspage
, unsigned int obj
)
476 zspage
->freeobj
= obj
;
479 static void get_zspage_mapping(struct zspage
*zspage
,
480 unsigned int *class_idx
,
483 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
485 *fullness
= zspage
->fullness
;
486 *class_idx
= zspage
->class;
489 static struct size_class
*zspage_class(struct zs_pool
*pool
,
490 struct zspage
*zspage
)
492 return pool
->size_class
[zspage
->class];
495 static void set_zspage_mapping(struct zspage
*zspage
,
496 unsigned int class_idx
,
499 zspage
->class = class_idx
;
500 zspage
->fullness
= fullness
;
504 * zsmalloc divides the pool into various size classes where each
505 * class maintains a list of zspages where each zspage is divided
506 * into equal sized chunks. Each allocation falls into one of these
507 * classes depending on its size. This function returns index of the
508 * size class which has chunk size big enough to hold the given size.
510 static int get_size_class_index(int size
)
514 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
515 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
516 ZS_SIZE_CLASS_DELTA
);
518 return min_t(int, ZS_SIZE_CLASSES
- 1, idx
);
521 static inline void class_stat_inc(struct size_class
*class,
522 int type
, unsigned long cnt
)
524 class->stats
.objs
[type
] += cnt
;
527 static inline void class_stat_dec(struct size_class
*class,
528 int type
, unsigned long cnt
)
530 class->stats
.objs
[type
] -= cnt
;
533 static inline unsigned long zs_stat_get(struct size_class
*class, int type
)
535 return class->stats
.objs
[type
];
538 #ifdef CONFIG_ZSMALLOC_STAT
540 static void __init
zs_stat_init(void)
542 if (!debugfs_initialized()) {
543 pr_warn("debugfs not available, stat dir not created\n");
547 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
550 static void __exit
zs_stat_exit(void)
552 debugfs_remove_recursive(zs_stat_root
);
555 static unsigned long zs_can_compact(struct size_class
*class);
557 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
560 struct zs_pool
*pool
= s
->private;
561 struct size_class
*class;
563 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
564 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
565 unsigned long total_freeable
= 0;
566 unsigned long inuse_totals
[NR_FULLNESS_GROUPS
] = {0, };
568 seq_printf(s
, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
569 "class", "size", "10%", "20%", "30%", "40%",
570 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
571 "obj_allocated", "obj_used", "pages_used",
572 "pages_per_zspage", "freeable");
574 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
576 class = pool
->size_class
[i
];
578 if (class->index
!= i
)
581 spin_lock(&pool
->lock
);
583 seq_printf(s
, " %5u %5u ", i
, class->size
);
584 for (fg
= ZS_INUSE_RATIO_10
; fg
< NR_FULLNESS_GROUPS
; fg
++) {
585 inuse_totals
[fg
] += zs_stat_get(class, fg
);
586 seq_printf(s
, "%9lu ", zs_stat_get(class, fg
));
589 obj_allocated
= zs_stat_get(class, ZS_OBJS_ALLOCATED
);
590 obj_used
= zs_stat_get(class, ZS_OBJS_INUSE
);
591 freeable
= zs_can_compact(class);
592 spin_unlock(&pool
->lock
);
594 objs_per_zspage
= class->objs_per_zspage
;
595 pages_used
= obj_allocated
/ objs_per_zspage
*
596 class->pages_per_zspage
;
598 seq_printf(s
, "%13lu %10lu %10lu %16d %8lu\n",
599 obj_allocated
, obj_used
, pages_used
,
600 class->pages_per_zspage
, freeable
);
602 total_objs
+= obj_allocated
;
603 total_used_objs
+= obj_used
;
604 total_pages
+= pages_used
;
605 total_freeable
+= freeable
;
609 seq_printf(s
, " %5s %5s ", "Total", "");
611 for (fg
= ZS_INUSE_RATIO_10
; fg
< NR_FULLNESS_GROUPS
; fg
++)
612 seq_printf(s
, "%9lu ", inuse_totals
[fg
]);
614 seq_printf(s
, "%13lu %10lu %10lu %16s %8lu\n",
615 total_objs
, total_used_objs
, total_pages
, "",
620 DEFINE_SHOW_ATTRIBUTE(zs_stats_size
);
622 static void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
625 pr_warn("no root stat dir, not creating <%s> stat dir\n", name
);
629 pool
->stat_dentry
= debugfs_create_dir(name
, zs_stat_root
);
631 debugfs_create_file("classes", S_IFREG
| 0444, pool
->stat_dentry
, pool
,
632 &zs_stats_size_fops
);
635 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
637 debugfs_remove_recursive(pool
->stat_dentry
);
640 #else /* CONFIG_ZSMALLOC_STAT */
641 static void __init
zs_stat_init(void)
645 static void __exit
zs_stat_exit(void)
649 static inline void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
653 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
660 * For each size class, zspages are divided into different groups
661 * depending on their usage ratio. This function returns fullness
662 * status of the given page.
664 static int get_fullness_group(struct size_class
*class, struct zspage
*zspage
)
666 int inuse
, objs_per_zspage
, ratio
;
668 inuse
= get_zspage_inuse(zspage
);
669 objs_per_zspage
= class->objs_per_zspage
;
672 return ZS_INUSE_RATIO_0
;
673 if (inuse
== objs_per_zspage
)
674 return ZS_INUSE_RATIO_100
;
676 ratio
= 100 * inuse
/ objs_per_zspage
;
678 * Take integer division into consideration: a page with one inuse
679 * object out of 127 possible, will end up having 0 usage ratio,
680 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
682 return ratio
/ 10 + 1;
686 * Each size class maintains various freelists and zspages are assigned
687 * to one of these freelists based on the number of live objects they
688 * have. This functions inserts the given zspage into the freelist
689 * identified by <class, fullness_group>.
691 static void insert_zspage(struct size_class
*class,
692 struct zspage
*zspage
,
695 class_stat_inc(class, fullness
, 1);
696 list_add(&zspage
->list
, &class->fullness_list
[fullness
]);
700 * This function removes the given zspage from the freelist identified
701 * by <class, fullness_group>.
703 static void remove_zspage(struct size_class
*class,
704 struct zspage
*zspage
,
707 VM_BUG_ON(list_empty(&class->fullness_list
[fullness
]));
709 list_del_init(&zspage
->list
);
710 class_stat_dec(class, fullness
, 1);
714 * Each size class maintains zspages in different fullness groups depending
715 * on the number of live objects they contain. When allocating or freeing
716 * objects, the fullness status of the page can change, for instance, from
717 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
718 * checks if such a status change has occurred for the given page and
719 * accordingly moves the page from the list of the old fullness group to that
720 * of the new fullness group.
722 static int fix_fullness_group(struct size_class
*class, struct zspage
*zspage
)
727 get_zspage_mapping(zspage
, &class_idx
, &currfg
);
728 newfg
= get_fullness_group(class, zspage
);
732 remove_zspage(class, zspage
, currfg
);
733 insert_zspage(class, zspage
, newfg
);
734 set_zspage_mapping(zspage
, class_idx
, newfg
);
739 static struct zspage
*get_zspage(struct page
*page
)
741 struct zspage
*zspage
= (struct zspage
*)page_private(page
);
743 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
747 static struct page
*get_next_page(struct page
*page
)
749 struct zspage
*zspage
= get_zspage(page
);
751 if (unlikely(ZsHugePage(zspage
)))
754 return (struct page
*)page
->index
;
758 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
759 * @obj: the encoded object value
760 * @page: page object resides in zspage
761 * @obj_idx: object index
763 static void obj_to_location(unsigned long obj
, struct page
**page
,
764 unsigned int *obj_idx
)
766 obj
>>= OBJ_TAG_BITS
;
767 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
768 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
771 static void obj_to_page(unsigned long obj
, struct page
**page
)
773 obj
>>= OBJ_TAG_BITS
;
774 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
778 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
779 * @page: page object resides in zspage
780 * @obj_idx: object index
782 static unsigned long location_to_obj(struct page
*page
, unsigned int obj_idx
)
786 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
787 obj
|= obj_idx
& OBJ_INDEX_MASK
;
788 obj
<<= OBJ_TAG_BITS
;
793 static unsigned long handle_to_obj(unsigned long handle
)
795 return *(unsigned long *)handle
;
798 static inline bool obj_allocated(struct page
*page
, void *obj
,
799 unsigned long *phandle
)
801 unsigned long handle
;
802 struct zspage
*zspage
= get_zspage(page
);
804 if (unlikely(ZsHugePage(zspage
))) {
805 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
806 handle
= page
->index
;
808 handle
= *(unsigned long *)obj
;
810 if (!(handle
& OBJ_ALLOCATED_TAG
))
813 /* Clear all tags before returning the handle */
814 *phandle
= handle
& ~OBJ_TAG_MASK
;
818 static void reset_page(struct page
*page
)
820 __ClearPageMovable(page
);
821 ClearPagePrivate(page
);
822 set_page_private(page
, 0);
823 page_mapcount_reset(page
);
827 static int trylock_zspage(struct zspage
*zspage
)
829 struct page
*cursor
, *fail
;
831 for (cursor
= get_first_page(zspage
); cursor
!= NULL
; cursor
=
832 get_next_page(cursor
)) {
833 if (!trylock_page(cursor
)) {
841 for (cursor
= get_first_page(zspage
); cursor
!= fail
; cursor
=
842 get_next_page(cursor
))
848 static void __free_zspage(struct zs_pool
*pool
, struct size_class
*class,
849 struct zspage
*zspage
)
851 struct page
*page
, *next
;
853 unsigned int class_idx
;
855 get_zspage_mapping(zspage
, &class_idx
, &fg
);
857 assert_spin_locked(&pool
->lock
);
859 VM_BUG_ON(get_zspage_inuse(zspage
));
860 VM_BUG_ON(fg
!= ZS_INUSE_RATIO_0
);
862 next
= page
= get_first_page(zspage
);
864 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
865 next
= get_next_page(page
);
868 dec_zone_page_state(page
, NR_ZSPAGES
);
871 } while (page
!= NULL
);
873 cache_free_zspage(pool
, zspage
);
875 class_stat_dec(class, ZS_OBJS_ALLOCATED
, class->objs_per_zspage
);
876 atomic_long_sub(class->pages_per_zspage
, &pool
->pages_allocated
);
879 static void free_zspage(struct zs_pool
*pool
, struct size_class
*class,
880 struct zspage
*zspage
)
882 VM_BUG_ON(get_zspage_inuse(zspage
));
883 VM_BUG_ON(list_empty(&zspage
->list
));
886 * Since zs_free couldn't be sleepable, this function cannot call
887 * lock_page. The page locks trylock_zspage got will be released
890 if (!trylock_zspage(zspage
)) {
891 kick_deferred_free(pool
);
895 remove_zspage(class, zspage
, ZS_INUSE_RATIO_0
);
896 __free_zspage(pool
, class, zspage
);
899 /* Initialize a newly allocated zspage */
900 static void init_zspage(struct size_class
*class, struct zspage
*zspage
)
902 unsigned int freeobj
= 1;
903 unsigned long off
= 0;
904 struct page
*page
= get_first_page(zspage
);
907 struct page
*next_page
;
908 struct link_free
*link
;
911 set_first_obj_offset(page
, off
);
913 vaddr
= kmap_atomic(page
);
914 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
916 while ((off
+= class->size
) < PAGE_SIZE
) {
917 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
918 link
+= class->size
/ sizeof(*link
);
922 * We now come to the last (full or partial) object on this
923 * page, which must point to the first object on the next
926 next_page
= get_next_page(page
);
928 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
931 * Reset OBJ_TAG_BITS bit to last link to tell
932 * whether it's allocated object or not.
934 link
->next
= -1UL << OBJ_TAG_BITS
;
936 kunmap_atomic(vaddr
);
941 set_freeobj(zspage
, 0);
944 static void create_page_chain(struct size_class
*class, struct zspage
*zspage
,
945 struct page
*pages
[])
949 struct page
*prev_page
= NULL
;
950 int nr_pages
= class->pages_per_zspage
;
953 * Allocate individual pages and link them together as:
954 * 1. all pages are linked together using page->index
955 * 2. each sub-page point to zspage using page->private
957 * we set PG_private to identify the first page (i.e. no other sub-page
958 * has this flag set).
960 for (i
= 0; i
< nr_pages
; i
++) {
962 set_page_private(page
, (unsigned long)zspage
);
965 zspage
->first_page
= page
;
966 SetPagePrivate(page
);
967 if (unlikely(class->objs_per_zspage
== 1 &&
968 class->pages_per_zspage
== 1))
969 SetZsHugePage(zspage
);
971 prev_page
->index
= (unsigned long)page
;
978 * Allocate a zspage for the given size class
980 static struct zspage
*alloc_zspage(struct zs_pool
*pool
,
981 struct size_class
*class,
985 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
];
986 struct zspage
*zspage
= cache_alloc_zspage(pool
, gfp
);
991 zspage
->magic
= ZSPAGE_MAGIC
;
992 migrate_lock_init(zspage
);
994 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
997 page
= alloc_page(gfp
);
1000 dec_zone_page_state(pages
[i
], NR_ZSPAGES
);
1001 __free_page(pages
[i
]);
1003 cache_free_zspage(pool
, zspage
);
1007 inc_zone_page_state(page
, NR_ZSPAGES
);
1011 create_page_chain(class, zspage
, pages
);
1012 init_zspage(class, zspage
);
1013 zspage
->pool
= pool
;
1018 static struct zspage
*find_get_zspage(struct size_class
*class)
1021 struct zspage
*zspage
;
1023 for (i
= ZS_INUSE_RATIO_99
; i
>= ZS_INUSE_RATIO_0
; i
--) {
1024 zspage
= list_first_entry_or_null(&class->fullness_list
[i
],
1025 struct zspage
, list
);
1033 static inline int __zs_cpu_up(struct mapping_area
*area
)
1036 * Make sure we don't leak memory if a cpu UP notification
1037 * and zs_init() race and both call zs_cpu_up() on the same cpu
1041 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1047 static inline void __zs_cpu_down(struct mapping_area
*area
)
1049 kfree(area
->vm_buf
);
1050 area
->vm_buf
= NULL
;
1053 static void *__zs_map_object(struct mapping_area
*area
,
1054 struct page
*pages
[2], int off
, int size
)
1058 char *buf
= area
->vm_buf
;
1060 /* disable page faults to match kmap_atomic() return conditions */
1061 pagefault_disable();
1063 /* no read fastpath */
1064 if (area
->vm_mm
== ZS_MM_WO
)
1067 sizes
[0] = PAGE_SIZE
- off
;
1068 sizes
[1] = size
- sizes
[0];
1070 /* copy object to per-cpu buffer */
1071 addr
= kmap_atomic(pages
[0]);
1072 memcpy(buf
, addr
+ off
, sizes
[0]);
1073 kunmap_atomic(addr
);
1074 addr
= kmap_atomic(pages
[1]);
1075 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1076 kunmap_atomic(addr
);
1078 return area
->vm_buf
;
1081 static void __zs_unmap_object(struct mapping_area
*area
,
1082 struct page
*pages
[2], int off
, int size
)
1088 /* no write fastpath */
1089 if (area
->vm_mm
== ZS_MM_RO
)
1093 buf
= buf
+ ZS_HANDLE_SIZE
;
1094 size
-= ZS_HANDLE_SIZE
;
1095 off
+= ZS_HANDLE_SIZE
;
1097 sizes
[0] = PAGE_SIZE
- off
;
1098 sizes
[1] = size
- sizes
[0];
1100 /* copy per-cpu buffer to object */
1101 addr
= kmap_atomic(pages
[0]);
1102 memcpy(addr
+ off
, buf
, sizes
[0]);
1103 kunmap_atomic(addr
);
1104 addr
= kmap_atomic(pages
[1]);
1105 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1106 kunmap_atomic(addr
);
1109 /* enable page faults to match kunmap_atomic() return conditions */
1113 static int zs_cpu_prepare(unsigned int cpu
)
1115 struct mapping_area
*area
;
1117 area
= &per_cpu(zs_map_area
, cpu
);
1118 return __zs_cpu_up(area
);
1121 static int zs_cpu_dead(unsigned int cpu
)
1123 struct mapping_area
*area
;
1125 area
= &per_cpu(zs_map_area
, cpu
);
1126 __zs_cpu_down(area
);
1130 static bool can_merge(struct size_class
*prev
, int pages_per_zspage
,
1131 int objs_per_zspage
)
1133 if (prev
->pages_per_zspage
== pages_per_zspage
&&
1134 prev
->objs_per_zspage
== objs_per_zspage
)
1140 static bool zspage_full(struct size_class
*class, struct zspage
*zspage
)
1142 return get_zspage_inuse(zspage
) == class->objs_per_zspage
;
1145 static bool zspage_empty(struct zspage
*zspage
)
1147 return get_zspage_inuse(zspage
) == 0;
1151 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1152 * that hold objects of the provided size.
1153 * @pool: zsmalloc pool to use
1154 * @size: object size
1156 * Context: Any context.
1158 * Return: the index of the zsmalloc &size_class that hold objects of the
1161 unsigned int zs_lookup_class_index(struct zs_pool
*pool
, unsigned int size
)
1163 struct size_class
*class;
1165 class = pool
->size_class
[get_size_class_index(size
)];
1167 return class->index
;
1169 EXPORT_SYMBOL_GPL(zs_lookup_class_index
);
1171 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1173 return atomic_long_read(&pool
->pages_allocated
);
1175 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1178 * zs_map_object - get address of allocated object from handle.
1179 * @pool: pool from which the object was allocated
1180 * @handle: handle returned from zs_malloc
1181 * @mm: mapping mode to use
1183 * Before using an object allocated from zs_malloc, it must be mapped using
1184 * this function. When done with the object, it must be unmapped using
1187 * Only one object can be mapped per cpu at a time. There is no protection
1188 * against nested mappings.
1190 * This function returns with preemption and page faults disabled.
1192 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1195 struct zspage
*zspage
;
1197 unsigned long obj
, off
;
1198 unsigned int obj_idx
;
1200 struct size_class
*class;
1201 struct mapping_area
*area
;
1202 struct page
*pages
[2];
1206 * Because we use per-cpu mapping areas shared among the
1207 * pools/users, we can't allow mapping in interrupt context
1208 * because it can corrupt another users mappings.
1210 BUG_ON(in_interrupt());
1212 /* It guarantees it can get zspage from handle safely */
1213 spin_lock(&pool
->lock
);
1214 obj
= handle_to_obj(handle
);
1215 obj_to_location(obj
, &page
, &obj_idx
);
1216 zspage
= get_zspage(page
);
1219 * migration cannot move any zpages in this zspage. Here, pool->lock
1220 * is too heavy since callers would take some time until they calls
1221 * zs_unmap_object API so delegate the locking from class to zspage
1222 * which is smaller granularity.
1224 migrate_read_lock(zspage
);
1225 spin_unlock(&pool
->lock
);
1227 class = zspage_class(pool
, zspage
);
1228 off
= offset_in_page(class->size
* obj_idx
);
1230 local_lock(&zs_map_area
.lock
);
1231 area
= this_cpu_ptr(&zs_map_area
);
1233 if (off
+ class->size
<= PAGE_SIZE
) {
1234 /* this object is contained entirely within a page */
1235 area
->vm_addr
= kmap_atomic(page
);
1236 ret
= area
->vm_addr
+ off
;
1240 /* this object spans two pages */
1242 pages
[1] = get_next_page(page
);
1245 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1247 if (likely(!ZsHugePage(zspage
)))
1248 ret
+= ZS_HANDLE_SIZE
;
1252 EXPORT_SYMBOL_GPL(zs_map_object
);
1254 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1256 struct zspage
*zspage
;
1258 unsigned long obj
, off
;
1259 unsigned int obj_idx
;
1261 struct size_class
*class;
1262 struct mapping_area
*area
;
1264 obj
= handle_to_obj(handle
);
1265 obj_to_location(obj
, &page
, &obj_idx
);
1266 zspage
= get_zspage(page
);
1267 class = zspage_class(pool
, zspage
);
1268 off
= offset_in_page(class->size
* obj_idx
);
1270 area
= this_cpu_ptr(&zs_map_area
);
1271 if (off
+ class->size
<= PAGE_SIZE
)
1272 kunmap_atomic(area
->vm_addr
);
1274 struct page
*pages
[2];
1277 pages
[1] = get_next_page(page
);
1280 __zs_unmap_object(area
, pages
, off
, class->size
);
1282 local_unlock(&zs_map_area
.lock
);
1284 migrate_read_unlock(zspage
);
1286 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1289 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1290 * zsmalloc &size_class.
1291 * @pool: zsmalloc pool to use
1293 * The function returns the size of the first huge class - any object of equal
1294 * or bigger size will be stored in zspage consisting of a single physical
1297 * Context: Any context.
1299 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1301 size_t zs_huge_class_size(struct zs_pool
*pool
)
1303 return huge_class_size
;
1305 EXPORT_SYMBOL_GPL(zs_huge_class_size
);
1307 static unsigned long obj_malloc(struct zs_pool
*pool
,
1308 struct zspage
*zspage
, unsigned long handle
)
1310 int i
, nr_page
, offset
;
1312 struct link_free
*link
;
1313 struct size_class
*class;
1315 struct page
*m_page
;
1316 unsigned long m_offset
;
1319 class = pool
->size_class
[zspage
->class];
1320 handle
|= OBJ_ALLOCATED_TAG
;
1321 obj
= get_freeobj(zspage
);
1323 offset
= obj
* class->size
;
1324 nr_page
= offset
>> PAGE_SHIFT
;
1325 m_offset
= offset_in_page(offset
);
1326 m_page
= get_first_page(zspage
);
1328 for (i
= 0; i
< nr_page
; i
++)
1329 m_page
= get_next_page(m_page
);
1331 vaddr
= kmap_atomic(m_page
);
1332 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1333 set_freeobj(zspage
, link
->next
>> OBJ_TAG_BITS
);
1334 if (likely(!ZsHugePage(zspage
)))
1335 /* record handle in the header of allocated chunk */
1336 link
->handle
= handle
;
1338 /* record handle to page->index */
1339 zspage
->first_page
->index
= handle
;
1341 kunmap_atomic(vaddr
);
1342 mod_zspage_inuse(zspage
, 1);
1344 obj
= location_to_obj(m_page
, obj
);
1351 * zs_malloc - Allocate block of given size from pool.
1352 * @pool: pool to allocate from
1353 * @size: size of block to allocate
1354 * @gfp: gfp flags when allocating object
1356 * On success, handle to the allocated object is returned,
1357 * otherwise an ERR_PTR().
1358 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1360 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
, gfp_t gfp
)
1362 unsigned long handle
, obj
;
1363 struct size_class
*class;
1365 struct zspage
*zspage
;
1367 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1368 return (unsigned long)ERR_PTR(-EINVAL
);
1370 handle
= cache_alloc_handle(pool
, gfp
);
1372 return (unsigned long)ERR_PTR(-ENOMEM
);
1374 /* extra space in chunk to keep the handle */
1375 size
+= ZS_HANDLE_SIZE
;
1376 class = pool
->size_class
[get_size_class_index(size
)];
1378 /* pool->lock effectively protects the zpage migration */
1379 spin_lock(&pool
->lock
);
1380 zspage
= find_get_zspage(class);
1381 if (likely(zspage
)) {
1382 obj
= obj_malloc(pool
, zspage
, handle
);
1383 /* Now move the zspage to another fullness group, if required */
1384 fix_fullness_group(class, zspage
);
1385 record_obj(handle
, obj
);
1386 class_stat_inc(class, ZS_OBJS_INUSE
, 1);
1391 spin_unlock(&pool
->lock
);
1393 zspage
= alloc_zspage(pool
, class, gfp
);
1395 cache_free_handle(pool
, handle
);
1396 return (unsigned long)ERR_PTR(-ENOMEM
);
1399 spin_lock(&pool
->lock
);
1400 obj
= obj_malloc(pool
, zspage
, handle
);
1401 newfg
= get_fullness_group(class, zspage
);
1402 insert_zspage(class, zspage
, newfg
);
1403 set_zspage_mapping(zspage
, class->index
, newfg
);
1404 record_obj(handle
, obj
);
1405 atomic_long_add(class->pages_per_zspage
, &pool
->pages_allocated
);
1406 class_stat_inc(class, ZS_OBJS_ALLOCATED
, class->objs_per_zspage
);
1407 class_stat_inc(class, ZS_OBJS_INUSE
, 1);
1409 /* We completely set up zspage so mark them as movable */
1410 SetZsPageMovable(pool
, zspage
);
1412 spin_unlock(&pool
->lock
);
1416 EXPORT_SYMBOL_GPL(zs_malloc
);
1418 static void obj_free(int class_size
, unsigned long obj
)
1420 struct link_free
*link
;
1421 struct zspage
*zspage
;
1422 struct page
*f_page
;
1423 unsigned long f_offset
;
1424 unsigned int f_objidx
;
1427 obj_to_location(obj
, &f_page
, &f_objidx
);
1428 f_offset
= offset_in_page(class_size
* f_objidx
);
1429 zspage
= get_zspage(f_page
);
1431 vaddr
= kmap_atomic(f_page
);
1432 link
= (struct link_free
*)(vaddr
+ f_offset
);
1434 /* Insert this object in containing zspage's freelist */
1435 if (likely(!ZsHugePage(zspage
)))
1436 link
->next
= get_freeobj(zspage
) << OBJ_TAG_BITS
;
1439 set_freeobj(zspage
, f_objidx
);
1441 kunmap_atomic(vaddr
);
1442 mod_zspage_inuse(zspage
, -1);
1445 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1447 struct zspage
*zspage
;
1448 struct page
*f_page
;
1450 struct size_class
*class;
1453 if (IS_ERR_OR_NULL((void *)handle
))
1457 * The pool->lock protects the race with zpage's migration
1458 * so it's safe to get the page from handle.
1460 spin_lock(&pool
->lock
);
1461 obj
= handle_to_obj(handle
);
1462 obj_to_page(obj
, &f_page
);
1463 zspage
= get_zspage(f_page
);
1464 class = zspage_class(pool
, zspage
);
1466 class_stat_dec(class, ZS_OBJS_INUSE
, 1);
1467 obj_free(class->size
, obj
);
1469 fullness
= fix_fullness_group(class, zspage
);
1470 if (fullness
== ZS_INUSE_RATIO_0
)
1471 free_zspage(pool
, class, zspage
);
1473 spin_unlock(&pool
->lock
);
1474 cache_free_handle(pool
, handle
);
1476 EXPORT_SYMBOL_GPL(zs_free
);
1478 static void zs_object_copy(struct size_class
*class, unsigned long dst
,
1481 struct page
*s_page
, *d_page
;
1482 unsigned int s_objidx
, d_objidx
;
1483 unsigned long s_off
, d_off
;
1484 void *s_addr
, *d_addr
;
1485 int s_size
, d_size
, size
;
1488 s_size
= d_size
= class->size
;
1490 obj_to_location(src
, &s_page
, &s_objidx
);
1491 obj_to_location(dst
, &d_page
, &d_objidx
);
1493 s_off
= offset_in_page(class->size
* s_objidx
);
1494 d_off
= offset_in_page(class->size
* d_objidx
);
1496 if (s_off
+ class->size
> PAGE_SIZE
)
1497 s_size
= PAGE_SIZE
- s_off
;
1499 if (d_off
+ class->size
> PAGE_SIZE
)
1500 d_size
= PAGE_SIZE
- d_off
;
1502 s_addr
= kmap_atomic(s_page
);
1503 d_addr
= kmap_atomic(d_page
);
1506 size
= min(s_size
, d_size
);
1507 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1510 if (written
== class->size
)
1519 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1520 * calls must occurs in reverse order of calls to kmap_atomic().
1521 * So, to call kunmap_atomic(s_addr) we should first call
1522 * kunmap_atomic(d_addr). For more details see
1523 * Documentation/mm/highmem.rst.
1525 if (s_off
>= PAGE_SIZE
) {
1526 kunmap_atomic(d_addr
);
1527 kunmap_atomic(s_addr
);
1528 s_page
= get_next_page(s_page
);
1529 s_addr
= kmap_atomic(s_page
);
1530 d_addr
= kmap_atomic(d_page
);
1531 s_size
= class->size
- written
;
1535 if (d_off
>= PAGE_SIZE
) {
1536 kunmap_atomic(d_addr
);
1537 d_page
= get_next_page(d_page
);
1538 d_addr
= kmap_atomic(d_page
);
1539 d_size
= class->size
- written
;
1544 kunmap_atomic(d_addr
);
1545 kunmap_atomic(s_addr
);
1549 * Find alloced object in zspage from index object and
1552 static unsigned long find_alloced_obj(struct size_class
*class,
1553 struct page
*page
, int *obj_idx
)
1555 unsigned int offset
;
1556 int index
= *obj_idx
;
1557 unsigned long handle
= 0;
1558 void *addr
= kmap_atomic(page
);
1560 offset
= get_first_obj_offset(page
);
1561 offset
+= class->size
* index
;
1563 while (offset
< PAGE_SIZE
) {
1564 if (obj_allocated(page
, addr
+ offset
, &handle
))
1567 offset
+= class->size
;
1571 kunmap_atomic(addr
);
1578 static void migrate_zspage(struct zs_pool
*pool
, struct zspage
*src_zspage
,
1579 struct zspage
*dst_zspage
)
1581 unsigned long used_obj
, free_obj
;
1582 unsigned long handle
;
1584 struct page
*s_page
= get_first_page(src_zspage
);
1585 struct size_class
*class = pool
->size_class
[src_zspage
->class];
1588 handle
= find_alloced_obj(class, s_page
, &obj_idx
);
1590 s_page
= get_next_page(s_page
);
1597 used_obj
= handle_to_obj(handle
);
1598 free_obj
= obj_malloc(pool
, dst_zspage
, handle
);
1599 zs_object_copy(class, free_obj
, used_obj
);
1601 record_obj(handle
, free_obj
);
1602 obj_free(class->size
, used_obj
);
1604 /* Stop if there is no more space */
1605 if (zspage_full(class, dst_zspage
))
1608 /* Stop if there are no more objects to migrate */
1609 if (zspage_empty(src_zspage
))
1614 static struct zspage
*isolate_src_zspage(struct size_class
*class)
1616 struct zspage
*zspage
;
1619 for (fg
= ZS_INUSE_RATIO_10
; fg
<= ZS_INUSE_RATIO_99
; fg
++) {
1620 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
],
1621 struct zspage
, list
);
1623 remove_zspage(class, zspage
, fg
);
1631 static struct zspage
*isolate_dst_zspage(struct size_class
*class)
1633 struct zspage
*zspage
;
1636 for (fg
= ZS_INUSE_RATIO_99
; fg
>= ZS_INUSE_RATIO_10
; fg
--) {
1637 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
],
1638 struct zspage
, list
);
1640 remove_zspage(class, zspage
, fg
);
1649 * putback_zspage - add @zspage into right class's fullness list
1650 * @class: destination class
1651 * @zspage: target page
1653 * Return @zspage's fullness status
1655 static int putback_zspage(struct size_class
*class, struct zspage
*zspage
)
1659 fullness
= get_fullness_group(class, zspage
);
1660 insert_zspage(class, zspage
, fullness
);
1661 set_zspage_mapping(zspage
, class->index
, fullness
);
1666 #ifdef CONFIG_COMPACTION
1668 * To prevent zspage destroy during migration, zspage freeing should
1669 * hold locks of all pages in the zspage.
1671 static void lock_zspage(struct zspage
*zspage
)
1673 struct page
*curr_page
, *page
;
1676 * Pages we haven't locked yet can be migrated off the list while we're
1677 * trying to lock them, so we need to be careful and only attempt to
1678 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1679 * may no longer belong to the zspage. This means that we may wait for
1680 * the wrong page to unlock, so we must take a reference to the page
1681 * prior to waiting for it to unlock outside migrate_read_lock().
1684 migrate_read_lock(zspage
);
1685 page
= get_first_page(zspage
);
1686 if (trylock_page(page
))
1689 migrate_read_unlock(zspage
);
1690 wait_on_page_locked(page
);
1695 while ((page
= get_next_page(curr_page
))) {
1696 if (trylock_page(page
)) {
1700 migrate_read_unlock(zspage
);
1701 wait_on_page_locked(page
);
1703 migrate_read_lock(zspage
);
1706 migrate_read_unlock(zspage
);
1708 #endif /* CONFIG_COMPACTION */
1710 static void migrate_lock_init(struct zspage
*zspage
)
1712 rwlock_init(&zspage
->lock
);
1715 static void migrate_read_lock(struct zspage
*zspage
) __acquires(&zspage
->lock
)
1717 read_lock(&zspage
->lock
);
1720 static void migrate_read_unlock(struct zspage
*zspage
) __releases(&zspage
->lock
)
1722 read_unlock(&zspage
->lock
);
1725 #ifdef CONFIG_COMPACTION
1726 static void migrate_write_lock(struct zspage
*zspage
)
1728 write_lock(&zspage
->lock
);
1731 static void migrate_write_lock_nested(struct zspage
*zspage
)
1733 write_lock_nested(&zspage
->lock
, SINGLE_DEPTH_NESTING
);
1736 static void migrate_write_unlock(struct zspage
*zspage
)
1738 write_unlock(&zspage
->lock
);
1741 /* Number of isolated subpage for *page migration* in this zspage */
1742 static void inc_zspage_isolation(struct zspage
*zspage
)
1747 static void dec_zspage_isolation(struct zspage
*zspage
)
1749 VM_BUG_ON(zspage
->isolated
== 0);
1753 static const struct movable_operations zsmalloc_mops
;
1755 static void replace_sub_page(struct size_class
*class, struct zspage
*zspage
,
1756 struct page
*newpage
, struct page
*oldpage
)
1759 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
] = {NULL
, };
1762 page
= get_first_page(zspage
);
1764 if (page
== oldpage
)
1765 pages
[idx
] = newpage
;
1769 } while ((page
= get_next_page(page
)) != NULL
);
1771 create_page_chain(class, zspage
, pages
);
1772 set_first_obj_offset(newpage
, get_first_obj_offset(oldpage
));
1773 if (unlikely(ZsHugePage(zspage
)))
1774 newpage
->index
= oldpage
->index
;
1775 __SetPageMovable(newpage
, &zsmalloc_mops
);
1778 static bool zs_page_isolate(struct page
*page
, isolate_mode_t mode
)
1780 struct zs_pool
*pool
;
1781 struct zspage
*zspage
;
1784 * Page is locked so zspage couldn't be destroyed. For detail, look at
1785 * lock_zspage in free_zspage.
1787 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1789 zspage
= get_zspage(page
);
1790 pool
= zspage
->pool
;
1791 spin_lock(&pool
->lock
);
1792 inc_zspage_isolation(zspage
);
1793 spin_unlock(&pool
->lock
);
1798 static int zs_page_migrate(struct page
*newpage
, struct page
*page
,
1799 enum migrate_mode mode
)
1801 struct zs_pool
*pool
;
1802 struct size_class
*class;
1803 struct zspage
*zspage
;
1805 void *s_addr
, *d_addr
, *addr
;
1806 unsigned int offset
;
1807 unsigned long handle
;
1808 unsigned long old_obj
, new_obj
;
1809 unsigned int obj_idx
;
1812 * We cannot support the _NO_COPY case here, because copy needs to
1813 * happen under the zs lock, which does not work with
1814 * MIGRATE_SYNC_NO_COPY workflow.
1816 if (mode
== MIGRATE_SYNC_NO_COPY
)
1819 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1821 /* The page is locked, so this pointer must remain valid */
1822 zspage
= get_zspage(page
);
1823 pool
= zspage
->pool
;
1826 * The pool's lock protects the race between zpage migration
1829 spin_lock(&pool
->lock
);
1830 class = zspage_class(pool
, zspage
);
1832 /* the migrate_write_lock protects zpage access via zs_map_object */
1833 migrate_write_lock(zspage
);
1835 offset
= get_first_obj_offset(page
);
1836 s_addr
= kmap_atomic(page
);
1839 * Here, any user cannot access all objects in the zspage so let's move.
1841 d_addr
= kmap_atomic(newpage
);
1842 copy_page(d_addr
, s_addr
);
1843 kunmap_atomic(d_addr
);
1845 for (addr
= s_addr
+ offset
; addr
< s_addr
+ PAGE_SIZE
;
1846 addr
+= class->size
) {
1847 if (obj_allocated(page
, addr
, &handle
)) {
1849 old_obj
= handle_to_obj(handle
);
1850 obj_to_location(old_obj
, &dummy
, &obj_idx
);
1851 new_obj
= (unsigned long)location_to_obj(newpage
,
1853 record_obj(handle
, new_obj
);
1856 kunmap_atomic(s_addr
);
1858 replace_sub_page(class, zspage
, newpage
, page
);
1859 dec_zspage_isolation(zspage
);
1861 * Since we complete the data copy and set up new zspage structure,
1862 * it's okay to release the pool's lock.
1864 spin_unlock(&pool
->lock
);
1865 migrate_write_unlock(zspage
);
1868 if (page_zone(newpage
) != page_zone(page
)) {
1869 dec_zone_page_state(page
, NR_ZSPAGES
);
1870 inc_zone_page_state(newpage
, NR_ZSPAGES
);
1876 return MIGRATEPAGE_SUCCESS
;
1879 static void zs_page_putback(struct page
*page
)
1881 struct zs_pool
*pool
;
1882 struct zspage
*zspage
;
1884 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1886 zspage
= get_zspage(page
);
1887 pool
= zspage
->pool
;
1888 spin_lock(&pool
->lock
);
1889 dec_zspage_isolation(zspage
);
1890 spin_unlock(&pool
->lock
);
1893 static const struct movable_operations zsmalloc_mops
= {
1894 .isolate_page
= zs_page_isolate
,
1895 .migrate_page
= zs_page_migrate
,
1896 .putback_page
= zs_page_putback
,
1900 * Caller should hold page_lock of all pages in the zspage
1901 * In here, we cannot use zspage meta data.
1903 static void async_free_zspage(struct work_struct
*work
)
1906 struct size_class
*class;
1907 unsigned int class_idx
;
1909 struct zspage
*zspage
, *tmp
;
1910 LIST_HEAD(free_pages
);
1911 struct zs_pool
*pool
= container_of(work
, struct zs_pool
,
1914 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
1915 class = pool
->size_class
[i
];
1916 if (class->index
!= i
)
1919 spin_lock(&pool
->lock
);
1920 list_splice_init(&class->fullness_list
[ZS_INUSE_RATIO_0
],
1922 spin_unlock(&pool
->lock
);
1925 list_for_each_entry_safe(zspage
, tmp
, &free_pages
, list
) {
1926 list_del(&zspage
->list
);
1927 lock_zspage(zspage
);
1929 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1930 VM_BUG_ON(fullness
!= ZS_INUSE_RATIO_0
);
1931 class = pool
->size_class
[class_idx
];
1932 spin_lock(&pool
->lock
);
1933 __free_zspage(pool
, class, zspage
);
1934 spin_unlock(&pool
->lock
);
1938 static void kick_deferred_free(struct zs_pool
*pool
)
1940 schedule_work(&pool
->free_work
);
1943 static void zs_flush_migration(struct zs_pool
*pool
)
1945 flush_work(&pool
->free_work
);
1948 static void init_deferred_free(struct zs_pool
*pool
)
1950 INIT_WORK(&pool
->free_work
, async_free_zspage
);
1953 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
)
1955 struct page
*page
= get_first_page(zspage
);
1958 WARN_ON(!trylock_page(page
));
1959 __SetPageMovable(page
, &zsmalloc_mops
);
1961 } while ((page
= get_next_page(page
)) != NULL
);
1964 static inline void zs_flush_migration(struct zs_pool
*pool
) { }
1969 * Based on the number of unused allocated objects calculate
1970 * and return the number of pages that we can free.
1972 static unsigned long zs_can_compact(struct size_class
*class)
1974 unsigned long obj_wasted
;
1975 unsigned long obj_allocated
= zs_stat_get(class, ZS_OBJS_ALLOCATED
);
1976 unsigned long obj_used
= zs_stat_get(class, ZS_OBJS_INUSE
);
1978 if (obj_allocated
<= obj_used
)
1981 obj_wasted
= obj_allocated
- obj_used
;
1982 obj_wasted
/= class->objs_per_zspage
;
1984 return obj_wasted
* class->pages_per_zspage
;
1987 static unsigned long __zs_compact(struct zs_pool
*pool
,
1988 struct size_class
*class)
1990 struct zspage
*src_zspage
= NULL
;
1991 struct zspage
*dst_zspage
= NULL
;
1992 unsigned long pages_freed
= 0;
1995 * protect the race between zpage migration and zs_free
1996 * as well as zpage allocation/free
1998 spin_lock(&pool
->lock
);
1999 while (zs_can_compact(class)) {
2003 dst_zspage
= isolate_dst_zspage(class);
2006 migrate_write_lock(dst_zspage
);
2009 src_zspage
= isolate_src_zspage(class);
2013 migrate_write_lock_nested(src_zspage
);
2015 migrate_zspage(pool
, src_zspage
, dst_zspage
);
2016 fg
= putback_zspage(class, src_zspage
);
2017 migrate_write_unlock(src_zspage
);
2019 if (fg
== ZS_INUSE_RATIO_0
) {
2020 free_zspage(pool
, class, src_zspage
);
2021 pages_freed
+= class->pages_per_zspage
;
2025 if (get_fullness_group(class, dst_zspage
) == ZS_INUSE_RATIO_100
2026 || spin_is_contended(&pool
->lock
)) {
2027 putback_zspage(class, dst_zspage
);
2028 migrate_write_unlock(dst_zspage
);
2031 spin_unlock(&pool
->lock
);
2033 spin_lock(&pool
->lock
);
2038 putback_zspage(class, src_zspage
);
2039 migrate_write_unlock(src_zspage
);
2043 putback_zspage(class, dst_zspage
);
2044 migrate_write_unlock(dst_zspage
);
2046 spin_unlock(&pool
->lock
);
2051 unsigned long zs_compact(struct zs_pool
*pool
)
2054 struct size_class
*class;
2055 unsigned long pages_freed
= 0;
2058 * Pool compaction is performed under pool->lock so it is basically
2059 * single-threaded. Having more than one thread in __zs_compact()
2060 * will increase pool->lock contention, which will impact other
2061 * zsmalloc operations that need pool->lock.
2063 if (atomic_xchg(&pool
->compaction_in_progress
, 1))
2066 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2067 class = pool
->size_class
[i
];
2068 if (class->index
!= i
)
2070 pages_freed
+= __zs_compact(pool
, class);
2072 atomic_long_add(pages_freed
, &pool
->stats
.pages_compacted
);
2073 atomic_set(&pool
->compaction_in_progress
, 0);
2077 EXPORT_SYMBOL_GPL(zs_compact
);
2079 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
2081 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
2083 EXPORT_SYMBOL_GPL(zs_pool_stats
);
2085 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
2086 struct shrink_control
*sc
)
2088 unsigned long pages_freed
;
2089 struct zs_pool
*pool
= shrinker
->private_data
;
2092 * Compact classes and calculate compaction delta.
2093 * Can run concurrently with a manually triggered
2094 * (by user) compaction.
2096 pages_freed
= zs_compact(pool
);
2098 return pages_freed
? pages_freed
: SHRINK_STOP
;
2101 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
2102 struct shrink_control
*sc
)
2105 struct size_class
*class;
2106 unsigned long pages_to_free
= 0;
2107 struct zs_pool
*pool
= shrinker
->private_data
;
2109 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2110 class = pool
->size_class
[i
];
2111 if (class->index
!= i
)
2114 pages_to_free
+= zs_can_compact(class);
2117 return pages_to_free
;
2120 static void zs_unregister_shrinker(struct zs_pool
*pool
)
2122 shrinker_free(pool
->shrinker
);
2125 static int zs_register_shrinker(struct zs_pool
*pool
)
2127 pool
->shrinker
= shrinker_alloc(0, "mm-zspool:%s", pool
->name
);
2128 if (!pool
->shrinker
)
2131 pool
->shrinker
->scan_objects
= zs_shrinker_scan
;
2132 pool
->shrinker
->count_objects
= zs_shrinker_count
;
2133 pool
->shrinker
->batch
= 0;
2134 pool
->shrinker
->private_data
= pool
;
2136 shrinker_register(pool
->shrinker
);
2141 static int calculate_zspage_chain_size(int class_size
)
2143 int i
, min_waste
= INT_MAX
;
2146 if (is_power_of_2(class_size
))
2149 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
2152 waste
= (i
* PAGE_SIZE
) % class_size
;
2153 if (waste
< min_waste
) {
2163 * zs_create_pool - Creates an allocation pool to work from.
2164 * @name: pool name to be created
2166 * This function must be called before anything when using
2167 * the zsmalloc allocator.
2169 * On success, a pointer to the newly created pool is returned,
2172 struct zs_pool
*zs_create_pool(const char *name
)
2175 struct zs_pool
*pool
;
2176 struct size_class
*prev_class
= NULL
;
2178 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
2182 init_deferred_free(pool
);
2183 spin_lock_init(&pool
->lock
);
2184 atomic_set(&pool
->compaction_in_progress
, 0);
2186 pool
->name
= kstrdup(name
, GFP_KERNEL
);
2190 if (create_cache(pool
))
2194 * Iterate reversely, because, size of size_class that we want to use
2195 * for merging should be larger or equal to current size.
2197 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2199 int pages_per_zspage
;
2200 int objs_per_zspage
;
2201 struct size_class
*class;
2204 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
2205 if (size
> ZS_MAX_ALLOC_SIZE
)
2206 size
= ZS_MAX_ALLOC_SIZE
;
2207 pages_per_zspage
= calculate_zspage_chain_size(size
);
2208 objs_per_zspage
= pages_per_zspage
* PAGE_SIZE
/ size
;
2211 * We iterate from biggest down to smallest classes,
2212 * so huge_class_size holds the size of the first huge
2213 * class. Any object bigger than or equal to that will
2214 * endup in the huge class.
2216 if (pages_per_zspage
!= 1 && objs_per_zspage
!= 1 &&
2218 huge_class_size
= size
;
2220 * The object uses ZS_HANDLE_SIZE bytes to store the
2221 * handle. We need to subtract it, because zs_malloc()
2222 * unconditionally adds handle size before it performs
2223 * size class search - so object may be smaller than
2224 * huge class size, yet it still can end up in the huge
2225 * class because it grows by ZS_HANDLE_SIZE extra bytes
2226 * right before class lookup.
2228 huge_class_size
-= (ZS_HANDLE_SIZE
- 1);
2232 * size_class is used for normal zsmalloc operation such
2233 * as alloc/free for that size. Although it is natural that we
2234 * have one size_class for each size, there is a chance that we
2235 * can get more memory utilization if we use one size_class for
2236 * many different sizes whose size_class have same
2237 * characteristics. So, we makes size_class point to
2238 * previous size_class if possible.
2241 if (can_merge(prev_class
, pages_per_zspage
, objs_per_zspage
)) {
2242 pool
->size_class
[i
] = prev_class
;
2247 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
2253 class->pages_per_zspage
= pages_per_zspage
;
2254 class->objs_per_zspage
= objs_per_zspage
;
2255 pool
->size_class
[i
] = class;
2257 fullness
= ZS_INUSE_RATIO_0
;
2258 while (fullness
< NR_FULLNESS_GROUPS
) {
2259 INIT_LIST_HEAD(&class->fullness_list
[fullness
]);
2266 /* debug only, don't abort if it fails */
2267 zs_pool_stat_create(pool
, name
);
2270 * Not critical since shrinker is only used to trigger internal
2271 * defragmentation of the pool which is pretty optional thing. If
2272 * registration fails we still can use the pool normally and user can
2273 * trigger compaction manually. Thus, ignore return code.
2275 zs_register_shrinker(pool
);
2280 zs_destroy_pool(pool
);
2283 EXPORT_SYMBOL_GPL(zs_create_pool
);
2285 void zs_destroy_pool(struct zs_pool
*pool
)
2289 zs_unregister_shrinker(pool
);
2290 zs_flush_migration(pool
);
2291 zs_pool_stat_destroy(pool
);
2293 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
2295 struct size_class
*class = pool
->size_class
[i
];
2300 if (class->index
!= i
)
2303 for (fg
= ZS_INUSE_RATIO_0
; fg
< NR_FULLNESS_GROUPS
; fg
++) {
2304 if (list_empty(&class->fullness_list
[fg
]))
2307 pr_err("Class-%d fullness group %d is not empty\n",
2313 destroy_cache(pool
);
2317 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2319 static int __init
zs_init(void)
2323 ret
= cpuhp_setup_state(CPUHP_MM_ZS_PREPARE
, "mm/zsmalloc:prepare",
2324 zs_cpu_prepare
, zs_cpu_dead
);
2329 zpool_register_driver(&zs_zpool_driver
);
2340 static void __exit
zs_exit(void)
2343 zpool_unregister_driver(&zs_zpool_driver
);
2345 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE
);
2350 module_init(zs_init
);
2351 module_exit(zs_exit
);
2353 MODULE_LICENSE("Dual BSD/GPL");
2354 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");