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)
114 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
117 #define FULLNESS_BITS 4
119 #define MAGIC_VAL_BITS 8
121 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
123 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
125 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
126 #define ZS_MIN_ALLOC_SIZE \
127 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
128 /* each chunk includes extra space to keep handle */
129 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
132 * On systems with 4K page size, this gives 255 size classes! There is a
134 * - Large number of size classes is potentially wasteful as free page are
135 * spread across these classes
136 * - Small number of size classes causes large internal fragmentation
137 * - Probably its better to use specific size classes (empirically
138 * determined). NOTE: all those class sizes must be set as multiple of
139 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
141 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
144 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
145 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
146 ZS_SIZE_CLASS_DELTA) + 1)
149 * Pages are distinguished by the ratio of used memory (that is the ratio
150 * of ->inuse objects to all objects that page can store). For example,
151 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
153 * The number of fullness groups is not random. It allows us to keep
154 * difference between the least busy page in the group (minimum permitted
155 * number of ->inuse objects) and the most busy page (maximum permitted
156 * number of ->inuse objects) at a reasonable value.
158 enum fullness_group
{
161 /* NOTE: 8 more fullness groups here */
162 ZS_INUSE_RATIO_99
= 10,
167 enum class_stat_type
{
168 /* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
169 ZS_OBJS_ALLOCATED
= NR_FULLNESS_GROUPS
,
174 struct zs_size_stat
{
175 unsigned long objs
[NR_CLASS_STAT_TYPES
];
178 #ifdef CONFIG_ZSMALLOC_STAT
179 static struct dentry
*zs_stat_root
;
182 static size_t huge_class_size
;
185 struct list_head fullness_list
[NR_FULLNESS_GROUPS
];
187 * Size of objects stored in this class. Must be multiple
192 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
193 int pages_per_zspage
;
196 struct zs_size_stat stats
;
200 * Placed within free objects to form a singly linked list.
201 * For every zspage, zspage->freeobj gives head of this list.
203 * This must be power of 2 and less than or equal to ZS_ALIGN
209 * It's valid for non-allocated object
213 * Handle of allocated object.
215 unsigned long handle
;
222 struct size_class
*size_class
[ZS_SIZE_CLASSES
];
223 struct kmem_cache
*handle_cachep
;
224 struct kmem_cache
*zspage_cachep
;
226 atomic_long_t pages_allocated
;
228 struct zs_pool_stats stats
;
230 /* Compact classes */
231 struct shrinker
*shrinker
;
233 #ifdef CONFIG_ZSMALLOC_STAT
234 struct dentry
*stat_dentry
;
236 #ifdef CONFIG_COMPACTION
237 struct work_struct free_work
;
240 atomic_t compaction_in_progress
;
245 unsigned int huge
:HUGE_BITS
;
246 unsigned int fullness
:FULLNESS_BITS
;
247 unsigned int class:CLASS_BITS
+ 1;
248 unsigned int magic
:MAGIC_VAL_BITS
;
251 unsigned int freeobj
;
252 struct page
*first_page
;
253 struct list_head list
; /* fullness list */
254 struct zs_pool
*pool
;
258 struct mapping_area
{
260 char *vm_buf
; /* copy buffer for objects that span pages */
261 char *vm_addr
; /* address of kmap_atomic()'ed pages */
262 enum zs_mapmode vm_mm
; /* mapping mode */
265 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
266 static void SetZsHugePage(struct zspage
*zspage
)
271 static bool ZsHugePage(struct zspage
*zspage
)
276 static void migrate_lock_init(struct zspage
*zspage
);
277 static void migrate_read_lock(struct zspage
*zspage
);
278 static void migrate_read_unlock(struct zspage
*zspage
);
279 static void migrate_write_lock(struct zspage
*zspage
);
280 static void migrate_write_unlock(struct zspage
*zspage
);
282 #ifdef CONFIG_COMPACTION
283 static void kick_deferred_free(struct zs_pool
*pool
);
284 static void init_deferred_free(struct zs_pool
*pool
);
285 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
);
287 static void kick_deferred_free(struct zs_pool
*pool
) {}
288 static void init_deferred_free(struct zs_pool
*pool
) {}
289 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
) {}
292 static int create_cache(struct zs_pool
*pool
)
294 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
296 if (!pool
->handle_cachep
)
299 pool
->zspage_cachep
= kmem_cache_create("zspage", sizeof(struct zspage
),
301 if (!pool
->zspage_cachep
) {
302 kmem_cache_destroy(pool
->handle_cachep
);
303 pool
->handle_cachep
= NULL
;
310 static void destroy_cache(struct zs_pool
*pool
)
312 kmem_cache_destroy(pool
->handle_cachep
);
313 kmem_cache_destroy(pool
->zspage_cachep
);
316 static unsigned long cache_alloc_handle(struct zs_pool
*pool
, gfp_t gfp
)
318 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
319 gfp
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
322 static void cache_free_handle(struct zs_pool
*pool
, unsigned long handle
)
324 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
327 static struct zspage
*cache_alloc_zspage(struct zs_pool
*pool
, gfp_t flags
)
329 return kmem_cache_zalloc(pool
->zspage_cachep
,
330 flags
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
333 static void cache_free_zspage(struct zs_pool
*pool
, struct zspage
*zspage
)
335 kmem_cache_free(pool
->zspage_cachep
, zspage
);
338 /* pool->lock(which owns the handle) synchronizes races */
339 static void record_obj(unsigned long handle
, unsigned long obj
)
341 *(unsigned long *)handle
= obj
;
348 static void *zs_zpool_create(const char *name
, gfp_t gfp
)
351 * Ignore global gfp flags: zs_malloc() may be invoked from
352 * different contexts and its caller must provide a valid
355 return zs_create_pool(name
);
358 static void zs_zpool_destroy(void *pool
)
360 zs_destroy_pool(pool
);
363 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
364 unsigned long *handle
)
366 *handle
= zs_malloc(pool
, size
, gfp
);
368 if (IS_ERR_VALUE(*handle
))
369 return PTR_ERR((void *)*handle
);
372 static void zs_zpool_free(void *pool
, unsigned long handle
)
374 zs_free(pool
, handle
);
377 static void *zs_zpool_map(void *pool
, unsigned long handle
,
378 enum zpool_mapmode mm
)
380 enum zs_mapmode zs_mm
;
395 return zs_map_object(pool
, handle
, zs_mm
);
397 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
399 zs_unmap_object(pool
, handle
);
402 static u64
zs_zpool_total_pages(void *pool
)
404 return zs_get_total_pages(pool
);
407 static struct zpool_driver zs_zpool_driver
= {
409 .owner
= THIS_MODULE
,
410 .create
= zs_zpool_create
,
411 .destroy
= zs_zpool_destroy
,
412 .malloc_support_movable
= true,
413 .malloc
= zs_zpool_malloc
,
414 .free
= zs_zpool_free
,
416 .unmap
= zs_zpool_unmap
,
417 .total_pages
= zs_zpool_total_pages
,
420 MODULE_ALIAS("zpool-zsmalloc");
421 #endif /* CONFIG_ZPOOL */
423 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
424 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
) = {
425 .lock
= INIT_LOCAL_LOCK(lock
),
428 static __maybe_unused
int is_first_page(struct page
*page
)
430 return PagePrivate(page
);
433 /* Protected by pool->lock */
434 static inline int get_zspage_inuse(struct zspage
*zspage
)
436 return zspage
->inuse
;
440 static inline void mod_zspage_inuse(struct zspage
*zspage
, int val
)
442 zspage
->inuse
+= val
;
445 static inline struct page
*get_first_page(struct zspage
*zspage
)
447 struct page
*first_page
= zspage
->first_page
;
449 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
453 static inline unsigned int get_first_obj_offset(struct page
*page
)
455 return page
->page_type
;
458 static inline void set_first_obj_offset(struct page
*page
, unsigned int offset
)
460 page
->page_type
= offset
;
463 static inline unsigned int get_freeobj(struct zspage
*zspage
)
465 return zspage
->freeobj
;
468 static inline void set_freeobj(struct zspage
*zspage
, unsigned int obj
)
470 zspage
->freeobj
= obj
;
473 static struct size_class
*zspage_class(struct zs_pool
*pool
,
474 struct zspage
*zspage
)
476 return pool
->size_class
[zspage
->class];
480 * zsmalloc divides the pool into various size classes where each
481 * class maintains a list of zspages where each zspage is divided
482 * into equal sized chunks. Each allocation falls into one of these
483 * classes depending on its size. This function returns index of the
484 * size class which has chunk size big enough to hold the given size.
486 static int get_size_class_index(int size
)
490 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
491 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
492 ZS_SIZE_CLASS_DELTA
);
494 return min_t(int, ZS_SIZE_CLASSES
- 1, idx
);
497 static inline void class_stat_inc(struct size_class
*class,
498 int type
, unsigned long cnt
)
500 class->stats
.objs
[type
] += cnt
;
503 static inline void class_stat_dec(struct size_class
*class,
504 int type
, unsigned long cnt
)
506 class->stats
.objs
[type
] -= cnt
;
509 static inline unsigned long zs_stat_get(struct size_class
*class, int type
)
511 return class->stats
.objs
[type
];
514 #ifdef CONFIG_ZSMALLOC_STAT
516 static void __init
zs_stat_init(void)
518 if (!debugfs_initialized()) {
519 pr_warn("debugfs not available, stat dir not created\n");
523 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
526 static void __exit
zs_stat_exit(void)
528 debugfs_remove_recursive(zs_stat_root
);
531 static unsigned long zs_can_compact(struct size_class
*class);
533 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
536 struct zs_pool
*pool
= s
->private;
537 struct size_class
*class;
539 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
540 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
541 unsigned long total_freeable
= 0;
542 unsigned long inuse_totals
[NR_FULLNESS_GROUPS
] = {0, };
544 seq_printf(s
, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
545 "class", "size", "10%", "20%", "30%", "40%",
546 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
547 "obj_allocated", "obj_used", "pages_used",
548 "pages_per_zspage", "freeable");
550 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
552 class = pool
->size_class
[i
];
554 if (class->index
!= i
)
557 spin_lock(&pool
->lock
);
559 seq_printf(s
, " %5u %5u ", i
, class->size
);
560 for (fg
= ZS_INUSE_RATIO_10
; fg
< NR_FULLNESS_GROUPS
; fg
++) {
561 inuse_totals
[fg
] += zs_stat_get(class, fg
);
562 seq_printf(s
, "%9lu ", zs_stat_get(class, fg
));
565 obj_allocated
= zs_stat_get(class, ZS_OBJS_ALLOCATED
);
566 obj_used
= zs_stat_get(class, ZS_OBJS_INUSE
);
567 freeable
= zs_can_compact(class);
568 spin_unlock(&pool
->lock
);
570 objs_per_zspage
= class->objs_per_zspage
;
571 pages_used
= obj_allocated
/ objs_per_zspage
*
572 class->pages_per_zspage
;
574 seq_printf(s
, "%13lu %10lu %10lu %16d %8lu\n",
575 obj_allocated
, obj_used
, pages_used
,
576 class->pages_per_zspage
, freeable
);
578 total_objs
+= obj_allocated
;
579 total_used_objs
+= obj_used
;
580 total_pages
+= pages_used
;
581 total_freeable
+= freeable
;
585 seq_printf(s
, " %5s %5s ", "Total", "");
587 for (fg
= ZS_INUSE_RATIO_10
; fg
< NR_FULLNESS_GROUPS
; fg
++)
588 seq_printf(s
, "%9lu ", inuse_totals
[fg
]);
590 seq_printf(s
, "%13lu %10lu %10lu %16s %8lu\n",
591 total_objs
, total_used_objs
, total_pages
, "",
596 DEFINE_SHOW_ATTRIBUTE(zs_stats_size
);
598 static void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
601 pr_warn("no root stat dir, not creating <%s> stat dir\n", name
);
605 pool
->stat_dentry
= debugfs_create_dir(name
, zs_stat_root
);
607 debugfs_create_file("classes", S_IFREG
| 0444, pool
->stat_dentry
, pool
,
608 &zs_stats_size_fops
);
611 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
613 debugfs_remove_recursive(pool
->stat_dentry
);
616 #else /* CONFIG_ZSMALLOC_STAT */
617 static void __init
zs_stat_init(void)
621 static void __exit
zs_stat_exit(void)
625 static inline void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
629 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
636 * For each size class, zspages are divided into different groups
637 * depending on their usage ratio. This function returns fullness
638 * status of the given page.
640 static int get_fullness_group(struct size_class
*class, struct zspage
*zspage
)
642 int inuse
, objs_per_zspage
, ratio
;
644 inuse
= get_zspage_inuse(zspage
);
645 objs_per_zspage
= class->objs_per_zspage
;
648 return ZS_INUSE_RATIO_0
;
649 if (inuse
== objs_per_zspage
)
650 return ZS_INUSE_RATIO_100
;
652 ratio
= 100 * inuse
/ objs_per_zspage
;
654 * Take integer division into consideration: a page with one inuse
655 * object out of 127 possible, will end up having 0 usage ratio,
656 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
658 return ratio
/ 10 + 1;
662 * Each size class maintains various freelists and zspages are assigned
663 * to one of these freelists based on the number of live objects they
664 * have. This functions inserts the given zspage into the freelist
665 * identified by <class, fullness_group>.
667 static void insert_zspage(struct size_class
*class,
668 struct zspage
*zspage
,
671 class_stat_inc(class, fullness
, 1);
672 list_add(&zspage
->list
, &class->fullness_list
[fullness
]);
673 zspage
->fullness
= fullness
;
677 * This function removes the given zspage from the freelist identified
678 * by <class, fullness_group>.
680 static void remove_zspage(struct size_class
*class, struct zspage
*zspage
)
682 int fullness
= zspage
->fullness
;
684 VM_BUG_ON(list_empty(&class->fullness_list
[fullness
]));
686 list_del_init(&zspage
->list
);
687 class_stat_dec(class, fullness
, 1);
691 * Each size class maintains zspages in different fullness groups depending
692 * on the number of live objects they contain. When allocating or freeing
693 * objects, the fullness status of the page can change, for instance, from
694 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
695 * checks if such a status change has occurred for the given page and
696 * accordingly moves the page from the list of the old fullness group to that
697 * of the new fullness group.
699 static int fix_fullness_group(struct size_class
*class, struct zspage
*zspage
)
703 newfg
= get_fullness_group(class, zspage
);
704 if (newfg
== zspage
->fullness
)
707 remove_zspage(class, zspage
);
708 insert_zspage(class, zspage
, newfg
);
713 static struct zspage
*get_zspage(struct page
*page
)
715 struct zspage
*zspage
= (struct zspage
*)page_private(page
);
717 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
721 static struct page
*get_next_page(struct page
*page
)
723 struct zspage
*zspage
= get_zspage(page
);
725 if (unlikely(ZsHugePage(zspage
)))
728 return (struct page
*)page
->index
;
732 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
733 * @obj: the encoded object value
734 * @page: page object resides in zspage
735 * @obj_idx: object index
737 static void obj_to_location(unsigned long obj
, struct page
**page
,
738 unsigned int *obj_idx
)
740 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
741 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
744 static void obj_to_page(unsigned long obj
, struct page
**page
)
746 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
750 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
751 * @page: page object resides in zspage
752 * @obj_idx: object index
754 static unsigned long location_to_obj(struct page
*page
, unsigned int obj_idx
)
758 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
759 obj
|= obj_idx
& OBJ_INDEX_MASK
;
764 static unsigned long handle_to_obj(unsigned long handle
)
766 return *(unsigned long *)handle
;
769 static inline bool obj_allocated(struct page
*page
, void *obj
,
770 unsigned long *phandle
)
772 unsigned long handle
;
773 struct zspage
*zspage
= get_zspage(page
);
775 if (unlikely(ZsHugePage(zspage
))) {
776 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
777 handle
= page
->index
;
779 handle
= *(unsigned long *)obj
;
781 if (!(handle
& OBJ_ALLOCATED_TAG
))
784 /* Clear all tags before returning the handle */
785 *phandle
= handle
& ~OBJ_TAG_MASK
;
789 static void reset_page(struct page
*page
)
791 __ClearPageMovable(page
);
792 ClearPagePrivate(page
);
793 set_page_private(page
, 0);
794 page_mapcount_reset(page
);
798 static int trylock_zspage(struct zspage
*zspage
)
800 struct page
*cursor
, *fail
;
802 for (cursor
= get_first_page(zspage
); cursor
!= NULL
; cursor
=
803 get_next_page(cursor
)) {
804 if (!trylock_page(cursor
)) {
812 for (cursor
= get_first_page(zspage
); cursor
!= fail
; cursor
=
813 get_next_page(cursor
))
819 static void __free_zspage(struct zs_pool
*pool
, struct size_class
*class,
820 struct zspage
*zspage
)
822 struct page
*page
, *next
;
824 assert_spin_locked(&pool
->lock
);
826 VM_BUG_ON(get_zspage_inuse(zspage
));
827 VM_BUG_ON(zspage
->fullness
!= ZS_INUSE_RATIO_0
);
829 next
= page
= get_first_page(zspage
);
831 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
832 next
= get_next_page(page
);
835 dec_zone_page_state(page
, NR_ZSPAGES
);
838 } while (page
!= NULL
);
840 cache_free_zspage(pool
, zspage
);
842 class_stat_dec(class, ZS_OBJS_ALLOCATED
, class->objs_per_zspage
);
843 atomic_long_sub(class->pages_per_zspage
, &pool
->pages_allocated
);
846 static void free_zspage(struct zs_pool
*pool
, struct size_class
*class,
847 struct zspage
*zspage
)
849 VM_BUG_ON(get_zspage_inuse(zspage
));
850 VM_BUG_ON(list_empty(&zspage
->list
));
853 * Since zs_free couldn't be sleepable, this function cannot call
854 * lock_page. The page locks trylock_zspage got will be released
857 if (!trylock_zspage(zspage
)) {
858 kick_deferred_free(pool
);
862 remove_zspage(class, zspage
);
863 __free_zspage(pool
, class, zspage
);
866 /* Initialize a newly allocated zspage */
867 static void init_zspage(struct size_class
*class, struct zspage
*zspage
)
869 unsigned int freeobj
= 1;
870 unsigned long off
= 0;
871 struct page
*page
= get_first_page(zspage
);
874 struct page
*next_page
;
875 struct link_free
*link
;
878 set_first_obj_offset(page
, off
);
880 vaddr
= kmap_atomic(page
);
881 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
883 while ((off
+= class->size
) < PAGE_SIZE
) {
884 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
885 link
+= class->size
/ sizeof(*link
);
889 * We now come to the last (full or partial) object on this
890 * page, which must point to the first object on the next
893 next_page
= get_next_page(page
);
895 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
898 * Reset OBJ_TAG_BITS bit to last link to tell
899 * whether it's allocated object or not.
901 link
->next
= -1UL << OBJ_TAG_BITS
;
903 kunmap_atomic(vaddr
);
908 set_freeobj(zspage
, 0);
911 static void create_page_chain(struct size_class
*class, struct zspage
*zspage
,
912 struct page
*pages
[])
916 struct page
*prev_page
= NULL
;
917 int nr_pages
= class->pages_per_zspage
;
920 * Allocate individual pages and link them together as:
921 * 1. all pages are linked together using page->index
922 * 2. each sub-page point to zspage using page->private
924 * we set PG_private to identify the first page (i.e. no other sub-page
925 * has this flag set).
927 for (i
= 0; i
< nr_pages
; i
++) {
929 set_page_private(page
, (unsigned long)zspage
);
932 zspage
->first_page
= page
;
933 SetPagePrivate(page
);
934 if (unlikely(class->objs_per_zspage
== 1 &&
935 class->pages_per_zspage
== 1))
936 SetZsHugePage(zspage
);
938 prev_page
->index
= (unsigned long)page
;
945 * Allocate a zspage for the given size class
947 static struct zspage
*alloc_zspage(struct zs_pool
*pool
,
948 struct size_class
*class,
952 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
];
953 struct zspage
*zspage
= cache_alloc_zspage(pool
, gfp
);
958 zspage
->magic
= ZSPAGE_MAGIC
;
959 migrate_lock_init(zspage
);
961 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
964 page
= alloc_page(gfp
);
967 dec_zone_page_state(pages
[i
], NR_ZSPAGES
);
968 __free_page(pages
[i
]);
970 cache_free_zspage(pool
, zspage
);
974 inc_zone_page_state(page
, NR_ZSPAGES
);
978 create_page_chain(class, zspage
, pages
);
979 init_zspage(class, zspage
);
981 zspage
->class = class->index
;
986 static struct zspage
*find_get_zspage(struct size_class
*class)
989 struct zspage
*zspage
;
991 for (i
= ZS_INUSE_RATIO_99
; i
>= ZS_INUSE_RATIO_0
; i
--) {
992 zspage
= list_first_entry_or_null(&class->fullness_list
[i
],
993 struct zspage
, list
);
1001 static inline int __zs_cpu_up(struct mapping_area
*area
)
1004 * Make sure we don't leak memory if a cpu UP notification
1005 * and zs_init() race and both call zs_cpu_up() on the same cpu
1009 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1015 static inline void __zs_cpu_down(struct mapping_area
*area
)
1017 kfree(area
->vm_buf
);
1018 area
->vm_buf
= NULL
;
1021 static void *__zs_map_object(struct mapping_area
*area
,
1022 struct page
*pages
[2], int off
, int size
)
1026 char *buf
= area
->vm_buf
;
1028 /* disable page faults to match kmap_atomic() return conditions */
1029 pagefault_disable();
1031 /* no read fastpath */
1032 if (area
->vm_mm
== ZS_MM_WO
)
1035 sizes
[0] = PAGE_SIZE
- off
;
1036 sizes
[1] = size
- sizes
[0];
1038 /* copy object to per-cpu buffer */
1039 addr
= kmap_atomic(pages
[0]);
1040 memcpy(buf
, addr
+ off
, sizes
[0]);
1041 kunmap_atomic(addr
);
1042 addr
= kmap_atomic(pages
[1]);
1043 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1044 kunmap_atomic(addr
);
1046 return area
->vm_buf
;
1049 static void __zs_unmap_object(struct mapping_area
*area
,
1050 struct page
*pages
[2], int off
, int size
)
1056 /* no write fastpath */
1057 if (area
->vm_mm
== ZS_MM_RO
)
1061 buf
= buf
+ ZS_HANDLE_SIZE
;
1062 size
-= ZS_HANDLE_SIZE
;
1063 off
+= ZS_HANDLE_SIZE
;
1065 sizes
[0] = PAGE_SIZE
- off
;
1066 sizes
[1] = size
- sizes
[0];
1068 /* copy per-cpu buffer to object */
1069 addr
= kmap_atomic(pages
[0]);
1070 memcpy(addr
+ off
, buf
, sizes
[0]);
1071 kunmap_atomic(addr
);
1072 addr
= kmap_atomic(pages
[1]);
1073 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1074 kunmap_atomic(addr
);
1077 /* enable page faults to match kunmap_atomic() return conditions */
1081 static int zs_cpu_prepare(unsigned int cpu
)
1083 struct mapping_area
*area
;
1085 area
= &per_cpu(zs_map_area
, cpu
);
1086 return __zs_cpu_up(area
);
1089 static int zs_cpu_dead(unsigned int cpu
)
1091 struct mapping_area
*area
;
1093 area
= &per_cpu(zs_map_area
, cpu
);
1094 __zs_cpu_down(area
);
1098 static bool can_merge(struct size_class
*prev
, int pages_per_zspage
,
1099 int objs_per_zspage
)
1101 if (prev
->pages_per_zspage
== pages_per_zspage
&&
1102 prev
->objs_per_zspage
== objs_per_zspage
)
1108 static bool zspage_full(struct size_class
*class, struct zspage
*zspage
)
1110 return get_zspage_inuse(zspage
) == class->objs_per_zspage
;
1113 static bool zspage_empty(struct zspage
*zspage
)
1115 return get_zspage_inuse(zspage
) == 0;
1119 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1120 * that hold objects of the provided size.
1121 * @pool: zsmalloc pool to use
1122 * @size: object size
1124 * Context: Any context.
1126 * Return: the index of the zsmalloc &size_class that hold objects of the
1129 unsigned int zs_lookup_class_index(struct zs_pool
*pool
, unsigned int size
)
1131 struct size_class
*class;
1133 class = pool
->size_class
[get_size_class_index(size
)];
1135 return class->index
;
1137 EXPORT_SYMBOL_GPL(zs_lookup_class_index
);
1139 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1141 return atomic_long_read(&pool
->pages_allocated
);
1143 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1146 * zs_map_object - get address of allocated object from handle.
1147 * @pool: pool from which the object was allocated
1148 * @handle: handle returned from zs_malloc
1149 * @mm: mapping mode to use
1151 * Before using an object allocated from zs_malloc, it must be mapped using
1152 * this function. When done with the object, it must be unmapped using
1155 * Only one object can be mapped per cpu at a time. There is no protection
1156 * against nested mappings.
1158 * This function returns with preemption and page faults disabled.
1160 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1163 struct zspage
*zspage
;
1165 unsigned long obj
, off
;
1166 unsigned int obj_idx
;
1168 struct size_class
*class;
1169 struct mapping_area
*area
;
1170 struct page
*pages
[2];
1174 * Because we use per-cpu mapping areas shared among the
1175 * pools/users, we can't allow mapping in interrupt context
1176 * because it can corrupt another users mappings.
1178 BUG_ON(in_interrupt());
1180 /* It guarantees it can get zspage from handle safely */
1181 spin_lock(&pool
->lock
);
1182 obj
= handle_to_obj(handle
);
1183 obj_to_location(obj
, &page
, &obj_idx
);
1184 zspage
= get_zspage(page
);
1187 * migration cannot move any zpages in this zspage. Here, pool->lock
1188 * is too heavy since callers would take some time until they calls
1189 * zs_unmap_object API so delegate the locking from class to zspage
1190 * which is smaller granularity.
1192 migrate_read_lock(zspage
);
1193 spin_unlock(&pool
->lock
);
1195 class = zspage_class(pool
, zspage
);
1196 off
= offset_in_page(class->size
* obj_idx
);
1198 local_lock(&zs_map_area
.lock
);
1199 area
= this_cpu_ptr(&zs_map_area
);
1201 if (off
+ class->size
<= PAGE_SIZE
) {
1202 /* this object is contained entirely within a page */
1203 area
->vm_addr
= kmap_atomic(page
);
1204 ret
= area
->vm_addr
+ off
;
1208 /* this object spans two pages */
1210 pages
[1] = get_next_page(page
);
1213 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1215 if (likely(!ZsHugePage(zspage
)))
1216 ret
+= ZS_HANDLE_SIZE
;
1220 EXPORT_SYMBOL_GPL(zs_map_object
);
1222 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1224 struct zspage
*zspage
;
1226 unsigned long obj
, off
;
1227 unsigned int obj_idx
;
1229 struct size_class
*class;
1230 struct mapping_area
*area
;
1232 obj
= handle_to_obj(handle
);
1233 obj_to_location(obj
, &page
, &obj_idx
);
1234 zspage
= get_zspage(page
);
1235 class = zspage_class(pool
, zspage
);
1236 off
= offset_in_page(class->size
* obj_idx
);
1238 area
= this_cpu_ptr(&zs_map_area
);
1239 if (off
+ class->size
<= PAGE_SIZE
)
1240 kunmap_atomic(area
->vm_addr
);
1242 struct page
*pages
[2];
1245 pages
[1] = get_next_page(page
);
1248 __zs_unmap_object(area
, pages
, off
, class->size
);
1250 local_unlock(&zs_map_area
.lock
);
1252 migrate_read_unlock(zspage
);
1254 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1257 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1258 * zsmalloc &size_class.
1259 * @pool: zsmalloc pool to use
1261 * The function returns the size of the first huge class - any object of equal
1262 * or bigger size will be stored in zspage consisting of a single physical
1265 * Context: Any context.
1267 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1269 size_t zs_huge_class_size(struct zs_pool
*pool
)
1271 return huge_class_size
;
1273 EXPORT_SYMBOL_GPL(zs_huge_class_size
);
1275 static unsigned long obj_malloc(struct zs_pool
*pool
,
1276 struct zspage
*zspage
, unsigned long handle
)
1278 int i
, nr_page
, offset
;
1280 struct link_free
*link
;
1281 struct size_class
*class;
1283 struct page
*m_page
;
1284 unsigned long m_offset
;
1287 class = pool
->size_class
[zspage
->class];
1288 handle
|= OBJ_ALLOCATED_TAG
;
1289 obj
= get_freeobj(zspage
);
1291 offset
= obj
* class->size
;
1292 nr_page
= offset
>> PAGE_SHIFT
;
1293 m_offset
= offset_in_page(offset
);
1294 m_page
= get_first_page(zspage
);
1296 for (i
= 0; i
< nr_page
; i
++)
1297 m_page
= get_next_page(m_page
);
1299 vaddr
= kmap_atomic(m_page
);
1300 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1301 set_freeobj(zspage
, link
->next
>> OBJ_TAG_BITS
);
1302 if (likely(!ZsHugePage(zspage
)))
1303 /* record handle in the header of allocated chunk */
1304 link
->handle
= handle
;
1306 /* record handle to page->index */
1307 zspage
->first_page
->index
= handle
;
1309 kunmap_atomic(vaddr
);
1310 mod_zspage_inuse(zspage
, 1);
1312 obj
= location_to_obj(m_page
, obj
);
1319 * zs_malloc - Allocate block of given size from pool.
1320 * @pool: pool to allocate from
1321 * @size: size of block to allocate
1322 * @gfp: gfp flags when allocating object
1324 * On success, handle to the allocated object is returned,
1325 * otherwise an ERR_PTR().
1326 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1328 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
, gfp_t gfp
)
1330 unsigned long handle
, obj
;
1331 struct size_class
*class;
1333 struct zspage
*zspage
;
1335 if (unlikely(!size
))
1336 return (unsigned long)ERR_PTR(-EINVAL
);
1338 if (unlikely(size
> ZS_MAX_ALLOC_SIZE
))
1339 return (unsigned long)ERR_PTR(-ENOSPC
);
1341 handle
= cache_alloc_handle(pool
, gfp
);
1343 return (unsigned long)ERR_PTR(-ENOMEM
);
1345 /* extra space in chunk to keep the handle */
1346 size
+= ZS_HANDLE_SIZE
;
1347 class = pool
->size_class
[get_size_class_index(size
)];
1349 /* pool->lock effectively protects the zpage migration */
1350 spin_lock(&pool
->lock
);
1351 zspage
= find_get_zspage(class);
1352 if (likely(zspage
)) {
1353 obj
= obj_malloc(pool
, zspage
, handle
);
1354 /* Now move the zspage to another fullness group, if required */
1355 fix_fullness_group(class, zspage
);
1356 record_obj(handle
, obj
);
1357 class_stat_inc(class, ZS_OBJS_INUSE
, 1);
1362 spin_unlock(&pool
->lock
);
1364 zspage
= alloc_zspage(pool
, class, gfp
);
1366 cache_free_handle(pool
, handle
);
1367 return (unsigned long)ERR_PTR(-ENOMEM
);
1370 spin_lock(&pool
->lock
);
1371 obj
= obj_malloc(pool
, zspage
, handle
);
1372 newfg
= get_fullness_group(class, zspage
);
1373 insert_zspage(class, zspage
, newfg
);
1374 record_obj(handle
, obj
);
1375 atomic_long_add(class->pages_per_zspage
, &pool
->pages_allocated
);
1376 class_stat_inc(class, ZS_OBJS_ALLOCATED
, class->objs_per_zspage
);
1377 class_stat_inc(class, ZS_OBJS_INUSE
, 1);
1379 /* We completely set up zspage so mark them as movable */
1380 SetZsPageMovable(pool
, zspage
);
1382 spin_unlock(&pool
->lock
);
1386 EXPORT_SYMBOL_GPL(zs_malloc
);
1388 static void obj_free(int class_size
, unsigned long obj
)
1390 struct link_free
*link
;
1391 struct zspage
*zspage
;
1392 struct page
*f_page
;
1393 unsigned long f_offset
;
1394 unsigned int f_objidx
;
1397 obj_to_location(obj
, &f_page
, &f_objidx
);
1398 f_offset
= offset_in_page(class_size
* f_objidx
);
1399 zspage
= get_zspage(f_page
);
1401 vaddr
= kmap_atomic(f_page
);
1402 link
= (struct link_free
*)(vaddr
+ f_offset
);
1404 /* Insert this object in containing zspage's freelist */
1405 if (likely(!ZsHugePage(zspage
)))
1406 link
->next
= get_freeobj(zspage
) << OBJ_TAG_BITS
;
1409 set_freeobj(zspage
, f_objidx
);
1411 kunmap_atomic(vaddr
);
1412 mod_zspage_inuse(zspage
, -1);
1415 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1417 struct zspage
*zspage
;
1418 struct page
*f_page
;
1420 struct size_class
*class;
1423 if (IS_ERR_OR_NULL((void *)handle
))
1427 * The pool->lock protects the race with zpage's migration
1428 * so it's safe to get the page from handle.
1430 spin_lock(&pool
->lock
);
1431 obj
= handle_to_obj(handle
);
1432 obj_to_page(obj
, &f_page
);
1433 zspage
= get_zspage(f_page
);
1434 class = zspage_class(pool
, zspage
);
1436 class_stat_dec(class, ZS_OBJS_INUSE
, 1);
1437 obj_free(class->size
, obj
);
1439 fullness
= fix_fullness_group(class, zspage
);
1440 if (fullness
== ZS_INUSE_RATIO_0
)
1441 free_zspage(pool
, class, zspage
);
1443 spin_unlock(&pool
->lock
);
1444 cache_free_handle(pool
, handle
);
1446 EXPORT_SYMBOL_GPL(zs_free
);
1448 static void zs_object_copy(struct size_class
*class, unsigned long dst
,
1451 struct page
*s_page
, *d_page
;
1452 unsigned int s_objidx
, d_objidx
;
1453 unsigned long s_off
, d_off
;
1454 void *s_addr
, *d_addr
;
1455 int s_size
, d_size
, size
;
1458 s_size
= d_size
= class->size
;
1460 obj_to_location(src
, &s_page
, &s_objidx
);
1461 obj_to_location(dst
, &d_page
, &d_objidx
);
1463 s_off
= offset_in_page(class->size
* s_objidx
);
1464 d_off
= offset_in_page(class->size
* d_objidx
);
1466 if (s_off
+ class->size
> PAGE_SIZE
)
1467 s_size
= PAGE_SIZE
- s_off
;
1469 if (d_off
+ class->size
> PAGE_SIZE
)
1470 d_size
= PAGE_SIZE
- d_off
;
1472 s_addr
= kmap_atomic(s_page
);
1473 d_addr
= kmap_atomic(d_page
);
1476 size
= min(s_size
, d_size
);
1477 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1480 if (written
== class->size
)
1489 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1490 * calls must occurs in reverse order of calls to kmap_atomic().
1491 * So, to call kunmap_atomic(s_addr) we should first call
1492 * kunmap_atomic(d_addr). For more details see
1493 * Documentation/mm/highmem.rst.
1495 if (s_off
>= PAGE_SIZE
) {
1496 kunmap_atomic(d_addr
);
1497 kunmap_atomic(s_addr
);
1498 s_page
= get_next_page(s_page
);
1499 s_addr
= kmap_atomic(s_page
);
1500 d_addr
= kmap_atomic(d_page
);
1501 s_size
= class->size
- written
;
1505 if (d_off
>= PAGE_SIZE
) {
1506 kunmap_atomic(d_addr
);
1507 d_page
= get_next_page(d_page
);
1508 d_addr
= kmap_atomic(d_page
);
1509 d_size
= class->size
- written
;
1514 kunmap_atomic(d_addr
);
1515 kunmap_atomic(s_addr
);
1519 * Find alloced object in zspage from index object and
1522 static unsigned long find_alloced_obj(struct size_class
*class,
1523 struct page
*page
, int *obj_idx
)
1525 unsigned int offset
;
1526 int index
= *obj_idx
;
1527 unsigned long handle
= 0;
1528 void *addr
= kmap_atomic(page
);
1530 offset
= get_first_obj_offset(page
);
1531 offset
+= class->size
* index
;
1533 while (offset
< PAGE_SIZE
) {
1534 if (obj_allocated(page
, addr
+ offset
, &handle
))
1537 offset
+= class->size
;
1541 kunmap_atomic(addr
);
1548 static void migrate_zspage(struct zs_pool
*pool
, struct zspage
*src_zspage
,
1549 struct zspage
*dst_zspage
)
1551 unsigned long used_obj
, free_obj
;
1552 unsigned long handle
;
1554 struct page
*s_page
= get_first_page(src_zspage
);
1555 struct size_class
*class = pool
->size_class
[src_zspage
->class];
1558 handle
= find_alloced_obj(class, s_page
, &obj_idx
);
1560 s_page
= get_next_page(s_page
);
1567 used_obj
= handle_to_obj(handle
);
1568 free_obj
= obj_malloc(pool
, dst_zspage
, handle
);
1569 zs_object_copy(class, free_obj
, used_obj
);
1571 record_obj(handle
, free_obj
);
1572 obj_free(class->size
, used_obj
);
1574 /* Stop if there is no more space */
1575 if (zspage_full(class, dst_zspage
))
1578 /* Stop if there are no more objects to migrate */
1579 if (zspage_empty(src_zspage
))
1584 static struct zspage
*isolate_src_zspage(struct size_class
*class)
1586 struct zspage
*zspage
;
1589 for (fg
= ZS_INUSE_RATIO_10
; fg
<= ZS_INUSE_RATIO_99
; fg
++) {
1590 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
],
1591 struct zspage
, list
);
1593 remove_zspage(class, zspage
);
1601 static struct zspage
*isolate_dst_zspage(struct size_class
*class)
1603 struct zspage
*zspage
;
1606 for (fg
= ZS_INUSE_RATIO_99
; fg
>= ZS_INUSE_RATIO_10
; fg
--) {
1607 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
],
1608 struct zspage
, list
);
1610 remove_zspage(class, zspage
);
1619 * putback_zspage - add @zspage into right class's fullness list
1620 * @class: destination class
1621 * @zspage: target page
1623 * Return @zspage's fullness status
1625 static int putback_zspage(struct size_class
*class, struct zspage
*zspage
)
1629 fullness
= get_fullness_group(class, zspage
);
1630 insert_zspage(class, zspage
, fullness
);
1635 #ifdef CONFIG_COMPACTION
1637 * To prevent zspage destroy during migration, zspage freeing should
1638 * hold locks of all pages in the zspage.
1640 static void lock_zspage(struct zspage
*zspage
)
1642 struct page
*curr_page
, *page
;
1645 * Pages we haven't locked yet can be migrated off the list while we're
1646 * trying to lock them, so we need to be careful and only attempt to
1647 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1648 * may no longer belong to the zspage. This means that we may wait for
1649 * the wrong page to unlock, so we must take a reference to the page
1650 * prior to waiting for it to unlock outside migrate_read_lock().
1653 migrate_read_lock(zspage
);
1654 page
= get_first_page(zspage
);
1655 if (trylock_page(page
))
1658 migrate_read_unlock(zspage
);
1659 wait_on_page_locked(page
);
1664 while ((page
= get_next_page(curr_page
))) {
1665 if (trylock_page(page
)) {
1669 migrate_read_unlock(zspage
);
1670 wait_on_page_locked(page
);
1672 migrate_read_lock(zspage
);
1675 migrate_read_unlock(zspage
);
1677 #endif /* CONFIG_COMPACTION */
1679 static void migrate_lock_init(struct zspage
*zspage
)
1681 rwlock_init(&zspage
->lock
);
1684 static void migrate_read_lock(struct zspage
*zspage
) __acquires(&zspage
->lock
)
1686 read_lock(&zspage
->lock
);
1689 static void migrate_read_unlock(struct zspage
*zspage
) __releases(&zspage
->lock
)
1691 read_unlock(&zspage
->lock
);
1694 static void migrate_write_lock(struct zspage
*zspage
)
1696 write_lock(&zspage
->lock
);
1699 static void migrate_write_unlock(struct zspage
*zspage
)
1701 write_unlock(&zspage
->lock
);
1704 #ifdef CONFIG_COMPACTION
1706 static const struct movable_operations zsmalloc_mops
;
1708 static void replace_sub_page(struct size_class
*class, struct zspage
*zspage
,
1709 struct page
*newpage
, struct page
*oldpage
)
1712 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
] = {NULL
, };
1715 page
= get_first_page(zspage
);
1717 if (page
== oldpage
)
1718 pages
[idx
] = newpage
;
1722 } while ((page
= get_next_page(page
)) != NULL
);
1724 create_page_chain(class, zspage
, pages
);
1725 set_first_obj_offset(newpage
, get_first_obj_offset(oldpage
));
1726 if (unlikely(ZsHugePage(zspage
)))
1727 newpage
->index
= oldpage
->index
;
1728 __SetPageMovable(newpage
, &zsmalloc_mops
);
1731 static bool zs_page_isolate(struct page
*page
, isolate_mode_t mode
)
1734 * Page is locked so zspage couldn't be destroyed. For detail, look at
1735 * lock_zspage in free_zspage.
1737 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1742 static int zs_page_migrate(struct page
*newpage
, struct page
*page
,
1743 enum migrate_mode mode
)
1745 struct zs_pool
*pool
;
1746 struct size_class
*class;
1747 struct zspage
*zspage
;
1749 void *s_addr
, *d_addr
, *addr
;
1750 unsigned int offset
;
1751 unsigned long handle
;
1752 unsigned long old_obj
, new_obj
;
1753 unsigned int obj_idx
;
1756 * We cannot support the _NO_COPY case here, because copy needs to
1757 * happen under the zs lock, which does not work with
1758 * MIGRATE_SYNC_NO_COPY workflow.
1760 if (mode
== MIGRATE_SYNC_NO_COPY
)
1763 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1765 /* The page is locked, so this pointer must remain valid */
1766 zspage
= get_zspage(page
);
1767 pool
= zspage
->pool
;
1770 * The pool's lock protects the race between zpage migration
1773 spin_lock(&pool
->lock
);
1774 class = zspage_class(pool
, zspage
);
1776 /* the migrate_write_lock protects zpage access via zs_map_object */
1777 migrate_write_lock(zspage
);
1779 offset
= get_first_obj_offset(page
);
1780 s_addr
= kmap_atomic(page
);
1783 * Here, any user cannot access all objects in the zspage so let's move.
1785 d_addr
= kmap_atomic(newpage
);
1786 copy_page(d_addr
, s_addr
);
1787 kunmap_atomic(d_addr
);
1789 for (addr
= s_addr
+ offset
; addr
< s_addr
+ PAGE_SIZE
;
1790 addr
+= class->size
) {
1791 if (obj_allocated(page
, addr
, &handle
)) {
1793 old_obj
= handle_to_obj(handle
);
1794 obj_to_location(old_obj
, &dummy
, &obj_idx
);
1795 new_obj
= (unsigned long)location_to_obj(newpage
,
1797 record_obj(handle
, new_obj
);
1800 kunmap_atomic(s_addr
);
1802 replace_sub_page(class, zspage
, newpage
, page
);
1804 * Since we complete the data copy and set up new zspage structure,
1805 * it's okay to release the pool's lock.
1807 spin_unlock(&pool
->lock
);
1808 migrate_write_unlock(zspage
);
1811 if (page_zone(newpage
) != page_zone(page
)) {
1812 dec_zone_page_state(page
, NR_ZSPAGES
);
1813 inc_zone_page_state(newpage
, NR_ZSPAGES
);
1819 return MIGRATEPAGE_SUCCESS
;
1822 static void zs_page_putback(struct page
*page
)
1824 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1827 static const struct movable_operations zsmalloc_mops
= {
1828 .isolate_page
= zs_page_isolate
,
1829 .migrate_page
= zs_page_migrate
,
1830 .putback_page
= zs_page_putback
,
1834 * Caller should hold page_lock of all pages in the zspage
1835 * In here, we cannot use zspage meta data.
1837 static void async_free_zspage(struct work_struct
*work
)
1840 struct size_class
*class;
1841 struct zspage
*zspage
, *tmp
;
1842 LIST_HEAD(free_pages
);
1843 struct zs_pool
*pool
= container_of(work
, struct zs_pool
,
1846 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
1847 class = pool
->size_class
[i
];
1848 if (class->index
!= i
)
1851 spin_lock(&pool
->lock
);
1852 list_splice_init(&class->fullness_list
[ZS_INUSE_RATIO_0
],
1854 spin_unlock(&pool
->lock
);
1857 list_for_each_entry_safe(zspage
, tmp
, &free_pages
, list
) {
1858 list_del(&zspage
->list
);
1859 lock_zspage(zspage
);
1861 spin_lock(&pool
->lock
);
1862 class = zspage_class(pool
, zspage
);
1863 __free_zspage(pool
, class, zspage
);
1864 spin_unlock(&pool
->lock
);
1868 static void kick_deferred_free(struct zs_pool
*pool
)
1870 schedule_work(&pool
->free_work
);
1873 static void zs_flush_migration(struct zs_pool
*pool
)
1875 flush_work(&pool
->free_work
);
1878 static void init_deferred_free(struct zs_pool
*pool
)
1880 INIT_WORK(&pool
->free_work
, async_free_zspage
);
1883 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
)
1885 struct page
*page
= get_first_page(zspage
);
1888 WARN_ON(!trylock_page(page
));
1889 __SetPageMovable(page
, &zsmalloc_mops
);
1891 } while ((page
= get_next_page(page
)) != NULL
);
1894 static inline void zs_flush_migration(struct zs_pool
*pool
) { }
1899 * Based on the number of unused allocated objects calculate
1900 * and return the number of pages that we can free.
1902 static unsigned long zs_can_compact(struct size_class
*class)
1904 unsigned long obj_wasted
;
1905 unsigned long obj_allocated
= zs_stat_get(class, ZS_OBJS_ALLOCATED
);
1906 unsigned long obj_used
= zs_stat_get(class, ZS_OBJS_INUSE
);
1908 if (obj_allocated
<= obj_used
)
1911 obj_wasted
= obj_allocated
- obj_used
;
1912 obj_wasted
/= class->objs_per_zspage
;
1914 return obj_wasted
* class->pages_per_zspage
;
1917 static unsigned long __zs_compact(struct zs_pool
*pool
,
1918 struct size_class
*class)
1920 struct zspage
*src_zspage
= NULL
;
1921 struct zspage
*dst_zspage
= NULL
;
1922 unsigned long pages_freed
= 0;
1925 * protect the race between zpage migration and zs_free
1926 * as well as zpage allocation/free
1928 spin_lock(&pool
->lock
);
1929 while (zs_can_compact(class)) {
1933 dst_zspage
= isolate_dst_zspage(class);
1938 src_zspage
= isolate_src_zspage(class);
1942 migrate_write_lock(src_zspage
);
1943 migrate_zspage(pool
, src_zspage
, dst_zspage
);
1944 migrate_write_unlock(src_zspage
);
1946 fg
= putback_zspage(class, src_zspage
);
1947 if (fg
== ZS_INUSE_RATIO_0
) {
1948 free_zspage(pool
, class, src_zspage
);
1949 pages_freed
+= class->pages_per_zspage
;
1953 if (get_fullness_group(class, dst_zspage
) == ZS_INUSE_RATIO_100
1954 || spin_is_contended(&pool
->lock
)) {
1955 putback_zspage(class, dst_zspage
);
1958 spin_unlock(&pool
->lock
);
1960 spin_lock(&pool
->lock
);
1965 putback_zspage(class, src_zspage
);
1968 putback_zspage(class, dst_zspage
);
1970 spin_unlock(&pool
->lock
);
1975 unsigned long zs_compact(struct zs_pool
*pool
)
1978 struct size_class
*class;
1979 unsigned long pages_freed
= 0;
1982 * Pool compaction is performed under pool->lock so it is basically
1983 * single-threaded. Having more than one thread in __zs_compact()
1984 * will increase pool->lock contention, which will impact other
1985 * zsmalloc operations that need pool->lock.
1987 if (atomic_xchg(&pool
->compaction_in_progress
, 1))
1990 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
1991 class = pool
->size_class
[i
];
1992 if (class->index
!= i
)
1994 pages_freed
+= __zs_compact(pool
, class);
1996 atomic_long_add(pages_freed
, &pool
->stats
.pages_compacted
);
1997 atomic_set(&pool
->compaction_in_progress
, 0);
2001 EXPORT_SYMBOL_GPL(zs_compact
);
2003 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
2005 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
2007 EXPORT_SYMBOL_GPL(zs_pool_stats
);
2009 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
2010 struct shrink_control
*sc
)
2012 unsigned long pages_freed
;
2013 struct zs_pool
*pool
= shrinker
->private_data
;
2016 * Compact classes and calculate compaction delta.
2017 * Can run concurrently with a manually triggered
2018 * (by user) compaction.
2020 pages_freed
= zs_compact(pool
);
2022 return pages_freed
? pages_freed
: SHRINK_STOP
;
2025 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
2026 struct shrink_control
*sc
)
2029 struct size_class
*class;
2030 unsigned long pages_to_free
= 0;
2031 struct zs_pool
*pool
= shrinker
->private_data
;
2033 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2034 class = pool
->size_class
[i
];
2035 if (class->index
!= i
)
2038 pages_to_free
+= zs_can_compact(class);
2041 return pages_to_free
;
2044 static void zs_unregister_shrinker(struct zs_pool
*pool
)
2046 shrinker_free(pool
->shrinker
);
2049 static int zs_register_shrinker(struct zs_pool
*pool
)
2051 pool
->shrinker
= shrinker_alloc(0, "mm-zspool:%s", pool
->name
);
2052 if (!pool
->shrinker
)
2055 pool
->shrinker
->scan_objects
= zs_shrinker_scan
;
2056 pool
->shrinker
->count_objects
= zs_shrinker_count
;
2057 pool
->shrinker
->batch
= 0;
2058 pool
->shrinker
->private_data
= pool
;
2060 shrinker_register(pool
->shrinker
);
2065 static int calculate_zspage_chain_size(int class_size
)
2067 int i
, min_waste
= INT_MAX
;
2070 if (is_power_of_2(class_size
))
2073 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
2076 waste
= (i
* PAGE_SIZE
) % class_size
;
2077 if (waste
< min_waste
) {
2087 * zs_create_pool - Creates an allocation pool to work from.
2088 * @name: pool name to be created
2090 * This function must be called before anything when using
2091 * the zsmalloc allocator.
2093 * On success, a pointer to the newly created pool is returned,
2096 struct zs_pool
*zs_create_pool(const char *name
)
2099 struct zs_pool
*pool
;
2100 struct size_class
*prev_class
= NULL
;
2102 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
2106 init_deferred_free(pool
);
2107 spin_lock_init(&pool
->lock
);
2108 atomic_set(&pool
->compaction_in_progress
, 0);
2110 pool
->name
= kstrdup(name
, GFP_KERNEL
);
2114 if (create_cache(pool
))
2118 * Iterate reversely, because, size of size_class that we want to use
2119 * for merging should be larger or equal to current size.
2121 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2123 int pages_per_zspage
;
2124 int objs_per_zspage
;
2125 struct size_class
*class;
2128 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
2129 if (size
> ZS_MAX_ALLOC_SIZE
)
2130 size
= ZS_MAX_ALLOC_SIZE
;
2131 pages_per_zspage
= calculate_zspage_chain_size(size
);
2132 objs_per_zspage
= pages_per_zspage
* PAGE_SIZE
/ size
;
2135 * We iterate from biggest down to smallest classes,
2136 * so huge_class_size holds the size of the first huge
2137 * class. Any object bigger than or equal to that will
2138 * endup in the huge class.
2140 if (pages_per_zspage
!= 1 && objs_per_zspage
!= 1 &&
2142 huge_class_size
= size
;
2144 * The object uses ZS_HANDLE_SIZE bytes to store the
2145 * handle. We need to subtract it, because zs_malloc()
2146 * unconditionally adds handle size before it performs
2147 * size class search - so object may be smaller than
2148 * huge class size, yet it still can end up in the huge
2149 * class because it grows by ZS_HANDLE_SIZE extra bytes
2150 * right before class lookup.
2152 huge_class_size
-= (ZS_HANDLE_SIZE
- 1);
2156 * size_class is used for normal zsmalloc operation such
2157 * as alloc/free for that size. Although it is natural that we
2158 * have one size_class for each size, there is a chance that we
2159 * can get more memory utilization if we use one size_class for
2160 * many different sizes whose size_class have same
2161 * characteristics. So, we makes size_class point to
2162 * previous size_class if possible.
2165 if (can_merge(prev_class
, pages_per_zspage
, objs_per_zspage
)) {
2166 pool
->size_class
[i
] = prev_class
;
2171 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
2177 class->pages_per_zspage
= pages_per_zspage
;
2178 class->objs_per_zspage
= objs_per_zspage
;
2179 pool
->size_class
[i
] = class;
2181 fullness
= ZS_INUSE_RATIO_0
;
2182 while (fullness
< NR_FULLNESS_GROUPS
) {
2183 INIT_LIST_HEAD(&class->fullness_list
[fullness
]);
2190 /* debug only, don't abort if it fails */
2191 zs_pool_stat_create(pool
, name
);
2194 * Not critical since shrinker is only used to trigger internal
2195 * defragmentation of the pool which is pretty optional thing. If
2196 * registration fails we still can use the pool normally and user can
2197 * trigger compaction manually. Thus, ignore return code.
2199 zs_register_shrinker(pool
);
2204 zs_destroy_pool(pool
);
2207 EXPORT_SYMBOL_GPL(zs_create_pool
);
2209 void zs_destroy_pool(struct zs_pool
*pool
)
2213 zs_unregister_shrinker(pool
);
2214 zs_flush_migration(pool
);
2215 zs_pool_stat_destroy(pool
);
2217 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
2219 struct size_class
*class = pool
->size_class
[i
];
2224 if (class->index
!= i
)
2227 for (fg
= ZS_INUSE_RATIO_0
; fg
< NR_FULLNESS_GROUPS
; fg
++) {
2228 if (list_empty(&class->fullness_list
[fg
]))
2231 pr_err("Class-%d fullness group %d is not empty\n",
2237 destroy_cache(pool
);
2241 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2243 static int __init
zs_init(void)
2247 ret
= cpuhp_setup_state(CPUHP_MM_ZS_PREPARE
, "mm/zsmalloc:prepare",
2248 zs_cpu_prepare
, zs_cpu_dead
);
2253 zpool_register_driver(&zs_zpool_driver
);
2264 static void __exit
zs_exit(void)
2267 zpool_unregister_driver(&zs_zpool_driver
);
2269 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE
);
2274 module_init(zs_init
);
2275 module_exit(zs_exit
);
2277 MODULE_LICENSE("Dual BSD/GPL");
2278 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");