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
41 #include <linux/module.h>
42 #include <linux/kernel.h>
43 #include <linux/sched.h>
44 #include <linux/magic.h>
45 #include <linux/bitops.h>
46 #include <linux/errno.h>
47 #include <linux/highmem.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/pgtable.h>
51 #include <asm/tlbflush.h>
52 #include <linux/cpumask.h>
53 #include <linux/cpu.h>
54 #include <linux/vmalloc.h>
55 #include <linux/preempt.h>
56 #include <linux/spinlock.h>
57 #include <linux/shrinker.h>
58 #include <linux/types.h>
59 #include <linux/debugfs.h>
60 #include <linux/zsmalloc.h>
61 #include <linux/zpool.h>
62 #include <linux/mount.h>
63 #include <linux/pseudo_fs.h>
64 #include <linux/migrate.h>
65 #include <linux/wait.h>
66 #include <linux/pagemap.h>
68 #include <linux/local_lock.h>
70 #define ZSPAGE_MAGIC 0x58
73 * This must be power of 2 and greater than or equal to sizeof(link_free).
74 * These two conditions ensure that any 'struct link_free' itself doesn't
75 * span more than 1 page which avoids complex case of mapping 2 pages simply
76 * to restore link_free pointer values.
81 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
82 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
84 #define ZS_MAX_ZSPAGE_ORDER 2
85 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
87 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
90 * Object location (<PFN>, <obj_idx>) is encoded as
91 * a single (unsigned long) handle value.
93 * Note that object index <obj_idx> starts from 0.
95 * This is made more complicated by various memory models and PAE.
98 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
99 #ifdef MAX_PHYSMEM_BITS
100 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
103 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
106 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
110 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
113 * Head in allocated object should have OBJ_ALLOCATED_TAG
114 * to identify the object was allocated or not.
115 * It's okay to add the status bit in the least bit because
116 * header keeps handle which is 4byte-aligned address so we
117 * have room for two bit at least.
119 #define OBJ_ALLOCATED_TAG 1
120 #define OBJ_TAG_BITS 1
121 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
122 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
125 #define FULLNESS_BITS 2
127 #define ISOLATED_BITS 3
128 #define MAGIC_VAL_BITS 8
130 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
131 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132 #define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
134 /* each chunk includes extra space to keep handle */
135 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
138 * On systems with 4K page size, this gives 255 size classes! There is a
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
150 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
151 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
152 ZS_SIZE_CLASS_DELTA) + 1)
154 enum fullness_group
{
162 enum class_stat_type
{
172 struct zs_size_stat
{
173 unsigned long objs
[NR_ZS_STAT_TYPE
];
176 #ifdef CONFIG_ZSMALLOC_STAT
177 static struct dentry
*zs_stat_root
;
180 #ifdef CONFIG_COMPACTION
181 static struct vfsmount
*zsmalloc_mnt
;
185 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
187 * n = number of allocated objects
188 * N = total number of objects zspage can store
189 * f = fullness_threshold_frac
191 * Similarly, we assign zspage to:
192 * ZS_ALMOST_FULL when n > N / f
193 * ZS_EMPTY when n == 0
194 * ZS_FULL when n == N
196 * (see: fix_fullness_group())
198 static const int fullness_threshold_frac
= 4;
199 static size_t huge_class_size
;
203 struct list_head fullness_list
[NR_ZS_FULLNESS
];
205 * Size of objects stored in this class. Must be multiple
210 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
211 int pages_per_zspage
;
214 struct zs_size_stat stats
;
218 * Placed within free objects to form a singly linked list.
219 * For every zspage, zspage->freeobj gives head of this list.
221 * This must be power of 2 and less than or equal to ZS_ALIGN
227 * It's valid for non-allocated object
231 * Handle of allocated object.
233 unsigned long handle
;
240 struct size_class
*size_class
[ZS_SIZE_CLASSES
];
241 struct kmem_cache
*handle_cachep
;
242 struct kmem_cache
*zspage_cachep
;
244 atomic_long_t pages_allocated
;
246 struct zs_pool_stats stats
;
248 /* Compact classes */
249 struct shrinker shrinker
;
251 #ifdef CONFIG_ZSMALLOC_STAT
252 struct dentry
*stat_dentry
;
254 #ifdef CONFIG_COMPACTION
256 struct work_struct free_work
;
258 /* protect page/zspage migration */
259 rwlock_t migrate_lock
;
264 unsigned int huge
:HUGE_BITS
;
265 unsigned int fullness
:FULLNESS_BITS
;
266 unsigned int class:CLASS_BITS
+ 1;
267 unsigned int isolated
:ISOLATED_BITS
;
268 unsigned int magic
:MAGIC_VAL_BITS
;
271 unsigned int freeobj
;
272 struct page
*first_page
;
273 struct list_head list
; /* fullness list */
274 #ifdef CONFIG_COMPACTION
279 struct mapping_area
{
281 char *vm_buf
; /* copy buffer for objects that span pages */
282 char *vm_addr
; /* address of kmap_atomic()'ed pages */
283 enum zs_mapmode vm_mm
; /* mapping mode */
286 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
287 static void SetZsHugePage(struct zspage
*zspage
)
292 static bool ZsHugePage(struct zspage
*zspage
)
297 #ifdef CONFIG_COMPACTION
298 static int zs_register_migration(struct zs_pool
*pool
);
299 static void zs_unregister_migration(struct zs_pool
*pool
);
300 static void migrate_lock_init(struct zspage
*zspage
);
301 static void migrate_read_lock(struct zspage
*zspage
);
302 static void migrate_read_unlock(struct zspage
*zspage
);
303 static void migrate_write_lock(struct zspage
*zspage
);
304 static void migrate_write_lock_nested(struct zspage
*zspage
);
305 static void migrate_write_unlock(struct zspage
*zspage
);
306 static void kick_deferred_free(struct zs_pool
*pool
);
307 static void init_deferred_free(struct zs_pool
*pool
);
308 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
);
310 static int zsmalloc_mount(void) { return 0; }
311 static void zsmalloc_unmount(void) {}
312 static int zs_register_migration(struct zs_pool
*pool
) { return 0; }
313 static void zs_unregister_migration(struct zs_pool
*pool
) {}
314 static void migrate_lock_init(struct zspage
*zspage
) {}
315 static void migrate_read_lock(struct zspage
*zspage
) {}
316 static void migrate_read_unlock(struct zspage
*zspage
) {}
317 static void migrate_write_lock(struct zspage
*zspage
) {}
318 static void migrate_write_lock_nested(struct zspage
*zspage
) {}
319 static void migrate_write_unlock(struct zspage
*zspage
) {}
320 static void kick_deferred_free(struct zs_pool
*pool
) {}
321 static void init_deferred_free(struct zs_pool
*pool
) {}
322 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
) {}
325 static int create_cache(struct zs_pool
*pool
)
327 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
329 if (!pool
->handle_cachep
)
332 pool
->zspage_cachep
= kmem_cache_create("zspage", sizeof(struct zspage
),
334 if (!pool
->zspage_cachep
) {
335 kmem_cache_destroy(pool
->handle_cachep
);
336 pool
->handle_cachep
= NULL
;
343 static void destroy_cache(struct zs_pool
*pool
)
345 kmem_cache_destroy(pool
->handle_cachep
);
346 kmem_cache_destroy(pool
->zspage_cachep
);
349 static unsigned long cache_alloc_handle(struct zs_pool
*pool
, gfp_t gfp
)
351 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
352 gfp
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
355 static void cache_free_handle(struct zs_pool
*pool
, unsigned long handle
)
357 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
360 static struct zspage
*cache_alloc_zspage(struct zs_pool
*pool
, gfp_t flags
)
362 return kmem_cache_zalloc(pool
->zspage_cachep
,
363 flags
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
366 static void cache_free_zspage(struct zs_pool
*pool
, struct zspage
*zspage
)
368 kmem_cache_free(pool
->zspage_cachep
, zspage
);
371 /* class->lock(which owns the handle) synchronizes races */
372 static void record_obj(unsigned long handle
, unsigned long obj
)
374 *(unsigned long *)handle
= obj
;
381 static void *zs_zpool_create(const char *name
, gfp_t gfp
,
382 const struct zpool_ops
*zpool_ops
,
386 * Ignore global gfp flags: zs_malloc() may be invoked from
387 * different contexts and its caller must provide a valid
390 return zs_create_pool(name
);
393 static void zs_zpool_destroy(void *pool
)
395 zs_destroy_pool(pool
);
398 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
399 unsigned long *handle
)
401 *handle
= zs_malloc(pool
, size
, gfp
);
402 return *handle
? 0 : -1;
404 static void zs_zpool_free(void *pool
, unsigned long handle
)
406 zs_free(pool
, handle
);
409 static void *zs_zpool_map(void *pool
, unsigned long handle
,
410 enum zpool_mapmode mm
)
412 enum zs_mapmode zs_mm
;
427 return zs_map_object(pool
, handle
, zs_mm
);
429 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
431 zs_unmap_object(pool
, handle
);
434 static u64
zs_zpool_total_size(void *pool
)
436 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
439 static struct zpool_driver zs_zpool_driver
= {
441 .owner
= THIS_MODULE
,
442 .create
= zs_zpool_create
,
443 .destroy
= zs_zpool_destroy
,
444 .malloc_support_movable
= true,
445 .malloc
= zs_zpool_malloc
,
446 .free
= zs_zpool_free
,
448 .unmap
= zs_zpool_unmap
,
449 .total_size
= zs_zpool_total_size
,
452 MODULE_ALIAS("zpool-zsmalloc");
453 #endif /* CONFIG_ZPOOL */
455 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
456 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
) = {
457 .lock
= INIT_LOCAL_LOCK(lock
),
460 static __maybe_unused
int is_first_page(struct page
*page
)
462 return PagePrivate(page
);
465 /* Protected by class->lock */
466 static inline int get_zspage_inuse(struct zspage
*zspage
)
468 return zspage
->inuse
;
472 static inline void mod_zspage_inuse(struct zspage
*zspage
, int val
)
474 zspage
->inuse
+= val
;
477 static inline struct page
*get_first_page(struct zspage
*zspage
)
479 struct page
*first_page
= zspage
->first_page
;
481 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
485 static inline int get_first_obj_offset(struct page
*page
)
487 return page
->page_type
;
490 static inline void set_first_obj_offset(struct page
*page
, int offset
)
492 page
->page_type
= offset
;
495 static inline unsigned int get_freeobj(struct zspage
*zspage
)
497 return zspage
->freeobj
;
500 static inline void set_freeobj(struct zspage
*zspage
, unsigned int obj
)
502 zspage
->freeobj
= obj
;
505 static void get_zspage_mapping(struct zspage
*zspage
,
506 unsigned int *class_idx
,
507 enum fullness_group
*fullness
)
509 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
511 *fullness
= zspage
->fullness
;
512 *class_idx
= zspage
->class;
515 static struct size_class
*zspage_class(struct zs_pool
*pool
,
516 struct zspage
*zspage
)
518 return pool
->size_class
[zspage
->class];
521 static void set_zspage_mapping(struct zspage
*zspage
,
522 unsigned int class_idx
,
523 enum fullness_group fullness
)
525 zspage
->class = class_idx
;
526 zspage
->fullness
= fullness
;
530 * zsmalloc divides the pool into various size classes where each
531 * class maintains a list of zspages where each zspage is divided
532 * into equal sized chunks. Each allocation falls into one of these
533 * classes depending on its size. This function returns index of the
534 * size class which has chunk size big enough to hold the given size.
536 static int get_size_class_index(int size
)
540 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
541 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
542 ZS_SIZE_CLASS_DELTA
);
544 return min_t(int, ZS_SIZE_CLASSES
- 1, idx
);
547 /* type can be of enum type class_stat_type or fullness_group */
548 static inline void class_stat_inc(struct size_class
*class,
549 int type
, unsigned long cnt
)
551 class->stats
.objs
[type
] += cnt
;
554 /* type can be of enum type class_stat_type or fullness_group */
555 static inline void class_stat_dec(struct size_class
*class,
556 int type
, unsigned long cnt
)
558 class->stats
.objs
[type
] -= cnt
;
561 /* type can be of enum type class_stat_type or fullness_group */
562 static inline unsigned long zs_stat_get(struct size_class
*class,
565 return class->stats
.objs
[type
];
568 #ifdef CONFIG_ZSMALLOC_STAT
570 static void __init
zs_stat_init(void)
572 if (!debugfs_initialized()) {
573 pr_warn("debugfs not available, stat dir not created\n");
577 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
580 static void __exit
zs_stat_exit(void)
582 debugfs_remove_recursive(zs_stat_root
);
585 static unsigned long zs_can_compact(struct size_class
*class);
587 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
590 struct zs_pool
*pool
= s
->private;
591 struct size_class
*class;
593 unsigned long class_almost_full
, class_almost_empty
;
594 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
595 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
596 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
597 unsigned long total_freeable
= 0;
599 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
600 "class", "size", "almost_full", "almost_empty",
601 "obj_allocated", "obj_used", "pages_used",
602 "pages_per_zspage", "freeable");
604 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
605 class = pool
->size_class
[i
];
607 if (class->index
!= i
)
610 spin_lock(&class->lock
);
611 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
612 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
613 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
614 obj_used
= zs_stat_get(class, OBJ_USED
);
615 freeable
= zs_can_compact(class);
616 spin_unlock(&class->lock
);
618 objs_per_zspage
= class->objs_per_zspage
;
619 pages_used
= obj_allocated
/ objs_per_zspage
*
620 class->pages_per_zspage
;
622 seq_printf(s
, " %5u %5u %11lu %12lu %13lu"
623 " %10lu %10lu %16d %8lu\n",
624 i
, class->size
, class_almost_full
, class_almost_empty
,
625 obj_allocated
, obj_used
, pages_used
,
626 class->pages_per_zspage
, freeable
);
628 total_class_almost_full
+= class_almost_full
;
629 total_class_almost_empty
+= class_almost_empty
;
630 total_objs
+= obj_allocated
;
631 total_used_objs
+= obj_used
;
632 total_pages
+= pages_used
;
633 total_freeable
+= freeable
;
637 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
638 "Total", "", total_class_almost_full
,
639 total_class_almost_empty
, total_objs
,
640 total_used_objs
, total_pages
, "", total_freeable
);
644 DEFINE_SHOW_ATTRIBUTE(zs_stats_size
);
646 static void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
649 pr_warn("no root stat dir, not creating <%s> stat dir\n", name
);
653 pool
->stat_dentry
= debugfs_create_dir(name
, zs_stat_root
);
655 debugfs_create_file("classes", S_IFREG
| 0444, pool
->stat_dentry
, pool
,
656 &zs_stats_size_fops
);
659 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
661 debugfs_remove_recursive(pool
->stat_dentry
);
664 #else /* CONFIG_ZSMALLOC_STAT */
665 static void __init
zs_stat_init(void)
669 static void __exit
zs_stat_exit(void)
673 static inline void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
677 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
684 * For each size class, zspages are divided into different groups
685 * depending on how "full" they are. This was done so that we could
686 * easily find empty or nearly empty zspages when we try to shrink
687 * the pool (not yet implemented). This function returns fullness
688 * status of the given page.
690 static enum fullness_group
get_fullness_group(struct size_class
*class,
691 struct zspage
*zspage
)
693 int inuse
, objs_per_zspage
;
694 enum fullness_group fg
;
696 inuse
= get_zspage_inuse(zspage
);
697 objs_per_zspage
= class->objs_per_zspage
;
701 else if (inuse
== objs_per_zspage
)
703 else if (inuse
<= 3 * objs_per_zspage
/ fullness_threshold_frac
)
704 fg
= ZS_ALMOST_EMPTY
;
712 * Each size class maintains various freelists and zspages are assigned
713 * to one of these freelists based on the number of live objects they
714 * have. This functions inserts the given zspage into the freelist
715 * identified by <class, fullness_group>.
717 static void insert_zspage(struct size_class
*class,
718 struct zspage
*zspage
,
719 enum fullness_group fullness
)
723 class_stat_inc(class, fullness
, 1);
724 head
= list_first_entry_or_null(&class->fullness_list
[fullness
],
725 struct zspage
, list
);
727 * We want to see more ZS_FULL pages and less almost empty/full.
728 * Put pages with higher ->inuse first.
730 if (head
&& get_zspage_inuse(zspage
) < get_zspage_inuse(head
))
731 list_add(&zspage
->list
, &head
->list
);
733 list_add(&zspage
->list
, &class->fullness_list
[fullness
]);
737 * This function removes the given zspage from the freelist identified
738 * by <class, fullness_group>.
740 static void remove_zspage(struct size_class
*class,
741 struct zspage
*zspage
,
742 enum fullness_group fullness
)
744 VM_BUG_ON(list_empty(&class->fullness_list
[fullness
]));
746 list_del_init(&zspage
->list
);
747 class_stat_dec(class, fullness
, 1);
751 * Each size class maintains zspages in different fullness groups depending
752 * on the number of live objects they contain. When allocating or freeing
753 * objects, the fullness status of the page can change, say, from ALMOST_FULL
754 * to ALMOST_EMPTY when freeing an object. This function checks if such
755 * a status change has occurred for the given page and accordingly moves the
756 * page from the freelist of the old fullness group to that of the new
759 static enum fullness_group
fix_fullness_group(struct size_class
*class,
760 struct zspage
*zspage
)
763 enum fullness_group currfg
, newfg
;
765 get_zspage_mapping(zspage
, &class_idx
, &currfg
);
766 newfg
= get_fullness_group(class, zspage
);
770 remove_zspage(class, zspage
, currfg
);
771 insert_zspage(class, zspage
, newfg
);
772 set_zspage_mapping(zspage
, class_idx
, newfg
);
778 * We have to decide on how many pages to link together
779 * to form a zspage for each size class. This is important
780 * to reduce wastage due to unusable space left at end of
781 * each zspage which is given as:
782 * wastage = Zp % class_size
783 * usage = Zp - wastage
784 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
786 * For example, for size class of 3/8 * PAGE_SIZE, we should
787 * link together 3 PAGE_SIZE sized pages to form a zspage
788 * since then we can perfectly fit in 8 such objects.
790 static int get_pages_per_zspage(int class_size
)
792 int i
, max_usedpc
= 0;
793 /* zspage order which gives maximum used size per KB */
794 int max_usedpc_order
= 1;
796 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
800 zspage_size
= i
* PAGE_SIZE
;
801 waste
= zspage_size
% class_size
;
802 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
804 if (usedpc
> max_usedpc
) {
806 max_usedpc_order
= i
;
810 return max_usedpc_order
;
813 static struct zspage
*get_zspage(struct page
*page
)
815 struct zspage
*zspage
= (struct zspage
*)page_private(page
);
817 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
821 static struct page
*get_next_page(struct page
*page
)
823 struct zspage
*zspage
= get_zspage(page
);
825 if (unlikely(ZsHugePage(zspage
)))
828 return (struct page
*)page
->index
;
832 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
833 * @obj: the encoded object value
834 * @page: page object resides in zspage
835 * @obj_idx: object index
837 static void obj_to_location(unsigned long obj
, struct page
**page
,
838 unsigned int *obj_idx
)
840 obj
>>= OBJ_TAG_BITS
;
841 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
842 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
845 static void obj_to_page(unsigned long obj
, struct page
**page
)
847 obj
>>= OBJ_TAG_BITS
;
848 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
852 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
853 * @page: page object resides in zspage
854 * @obj_idx: object index
856 static unsigned long location_to_obj(struct page
*page
, unsigned int obj_idx
)
860 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
861 obj
|= obj_idx
& OBJ_INDEX_MASK
;
862 obj
<<= OBJ_TAG_BITS
;
867 static unsigned long handle_to_obj(unsigned long handle
)
869 return *(unsigned long *)handle
;
872 static bool obj_allocated(struct page
*page
, void *obj
, unsigned long *phandle
)
874 unsigned long handle
;
875 struct zspage
*zspage
= get_zspage(page
);
877 if (unlikely(ZsHugePage(zspage
))) {
878 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
879 handle
= page
->index
;
881 handle
= *(unsigned long *)obj
;
883 if (!(handle
& OBJ_ALLOCATED_TAG
))
886 *phandle
= handle
& ~OBJ_ALLOCATED_TAG
;
890 static void reset_page(struct page
*page
)
892 __ClearPageMovable(page
);
893 ClearPagePrivate(page
);
894 set_page_private(page
, 0);
895 page_mapcount_reset(page
);
899 static int trylock_zspage(struct zspage
*zspage
)
901 struct page
*cursor
, *fail
;
903 for (cursor
= get_first_page(zspage
); cursor
!= NULL
; cursor
=
904 get_next_page(cursor
)) {
905 if (!trylock_page(cursor
)) {
913 for (cursor
= get_first_page(zspage
); cursor
!= fail
; cursor
=
914 get_next_page(cursor
))
920 static void __free_zspage(struct zs_pool
*pool
, struct size_class
*class,
921 struct zspage
*zspage
)
923 struct page
*page
, *next
;
924 enum fullness_group fg
;
925 unsigned int class_idx
;
927 get_zspage_mapping(zspage
, &class_idx
, &fg
);
929 assert_spin_locked(&class->lock
);
931 VM_BUG_ON(get_zspage_inuse(zspage
));
932 VM_BUG_ON(fg
!= ZS_EMPTY
);
934 next
= page
= get_first_page(zspage
);
936 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
937 next
= get_next_page(page
);
940 dec_zone_page_state(page
, NR_ZSPAGES
);
943 } while (page
!= NULL
);
945 cache_free_zspage(pool
, zspage
);
947 class_stat_dec(class, OBJ_ALLOCATED
, class->objs_per_zspage
);
948 atomic_long_sub(class->pages_per_zspage
,
949 &pool
->pages_allocated
);
952 static void free_zspage(struct zs_pool
*pool
, struct size_class
*class,
953 struct zspage
*zspage
)
955 VM_BUG_ON(get_zspage_inuse(zspage
));
956 VM_BUG_ON(list_empty(&zspage
->list
));
959 * Since zs_free couldn't be sleepable, this function cannot call
960 * lock_page. The page locks trylock_zspage got will be released
963 if (!trylock_zspage(zspage
)) {
964 kick_deferred_free(pool
);
968 remove_zspage(class, zspage
, ZS_EMPTY
);
969 __free_zspage(pool
, class, zspage
);
972 /* Initialize a newly allocated zspage */
973 static void init_zspage(struct size_class
*class, struct zspage
*zspage
)
975 unsigned int freeobj
= 1;
976 unsigned long off
= 0;
977 struct page
*page
= get_first_page(zspage
);
980 struct page
*next_page
;
981 struct link_free
*link
;
984 set_first_obj_offset(page
, off
);
986 vaddr
= kmap_atomic(page
);
987 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
989 while ((off
+= class->size
) < PAGE_SIZE
) {
990 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
991 link
+= class->size
/ sizeof(*link
);
995 * We now come to the last (full or partial) object on this
996 * page, which must point to the first object on the next
999 next_page
= get_next_page(page
);
1001 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
1004 * Reset OBJ_TAG_BITS bit to last link to tell
1005 * whether it's allocated object or not.
1007 link
->next
= -1UL << OBJ_TAG_BITS
;
1009 kunmap_atomic(vaddr
);
1014 set_freeobj(zspage
, 0);
1017 static void create_page_chain(struct size_class
*class, struct zspage
*zspage
,
1018 struct page
*pages
[])
1022 struct page
*prev_page
= NULL
;
1023 int nr_pages
= class->pages_per_zspage
;
1026 * Allocate individual pages and link them together as:
1027 * 1. all pages are linked together using page->index
1028 * 2. each sub-page point to zspage using page->private
1030 * we set PG_private to identify the first page (i.e. no other sub-page
1031 * has this flag set).
1033 for (i
= 0; i
< nr_pages
; i
++) {
1035 set_page_private(page
, (unsigned long)zspage
);
1038 zspage
->first_page
= page
;
1039 SetPagePrivate(page
);
1040 if (unlikely(class->objs_per_zspage
== 1 &&
1041 class->pages_per_zspage
== 1))
1042 SetZsHugePage(zspage
);
1044 prev_page
->index
= (unsigned long)page
;
1051 * Allocate a zspage for the given size class
1053 static struct zspage
*alloc_zspage(struct zs_pool
*pool
,
1054 struct size_class
*class,
1058 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
];
1059 struct zspage
*zspage
= cache_alloc_zspage(pool
, gfp
);
1064 zspage
->magic
= ZSPAGE_MAGIC
;
1065 migrate_lock_init(zspage
);
1067 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
1070 page
= alloc_page(gfp
);
1073 dec_zone_page_state(pages
[i
], NR_ZSPAGES
);
1074 __free_page(pages
[i
]);
1076 cache_free_zspage(pool
, zspage
);
1080 inc_zone_page_state(page
, NR_ZSPAGES
);
1084 create_page_chain(class, zspage
, pages
);
1085 init_zspage(class, zspage
);
1090 static struct zspage
*find_get_zspage(struct size_class
*class)
1093 struct zspage
*zspage
;
1095 for (i
= ZS_ALMOST_FULL
; i
>= ZS_EMPTY
; i
--) {
1096 zspage
= list_first_entry_or_null(&class->fullness_list
[i
],
1097 struct zspage
, list
);
1105 static inline int __zs_cpu_up(struct mapping_area
*area
)
1108 * Make sure we don't leak memory if a cpu UP notification
1109 * and zs_init() race and both call zs_cpu_up() on the same cpu
1113 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1119 static inline void __zs_cpu_down(struct mapping_area
*area
)
1121 kfree(area
->vm_buf
);
1122 area
->vm_buf
= NULL
;
1125 static void *__zs_map_object(struct mapping_area
*area
,
1126 struct page
*pages
[2], int off
, int size
)
1130 char *buf
= area
->vm_buf
;
1132 /* disable page faults to match kmap_atomic() return conditions */
1133 pagefault_disable();
1135 /* no read fastpath */
1136 if (area
->vm_mm
== ZS_MM_WO
)
1139 sizes
[0] = PAGE_SIZE
- off
;
1140 sizes
[1] = size
- sizes
[0];
1142 /* copy object to per-cpu buffer */
1143 addr
= kmap_atomic(pages
[0]);
1144 memcpy(buf
, addr
+ off
, sizes
[0]);
1145 kunmap_atomic(addr
);
1146 addr
= kmap_atomic(pages
[1]);
1147 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1148 kunmap_atomic(addr
);
1150 return area
->vm_buf
;
1153 static void __zs_unmap_object(struct mapping_area
*area
,
1154 struct page
*pages
[2], int off
, int size
)
1160 /* no write fastpath */
1161 if (area
->vm_mm
== ZS_MM_RO
)
1165 buf
= buf
+ ZS_HANDLE_SIZE
;
1166 size
-= ZS_HANDLE_SIZE
;
1167 off
+= ZS_HANDLE_SIZE
;
1169 sizes
[0] = PAGE_SIZE
- off
;
1170 sizes
[1] = size
- sizes
[0];
1172 /* copy per-cpu buffer to object */
1173 addr
= kmap_atomic(pages
[0]);
1174 memcpy(addr
+ off
, buf
, sizes
[0]);
1175 kunmap_atomic(addr
);
1176 addr
= kmap_atomic(pages
[1]);
1177 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1178 kunmap_atomic(addr
);
1181 /* enable page faults to match kunmap_atomic() return conditions */
1185 static int zs_cpu_prepare(unsigned int cpu
)
1187 struct mapping_area
*area
;
1189 area
= &per_cpu(zs_map_area
, cpu
);
1190 return __zs_cpu_up(area
);
1193 static int zs_cpu_dead(unsigned int cpu
)
1195 struct mapping_area
*area
;
1197 area
= &per_cpu(zs_map_area
, cpu
);
1198 __zs_cpu_down(area
);
1202 static bool can_merge(struct size_class
*prev
, int pages_per_zspage
,
1203 int objs_per_zspage
)
1205 if (prev
->pages_per_zspage
== pages_per_zspage
&&
1206 prev
->objs_per_zspage
== objs_per_zspage
)
1212 static bool zspage_full(struct size_class
*class, struct zspage
*zspage
)
1214 return get_zspage_inuse(zspage
) == class->objs_per_zspage
;
1217 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1219 return atomic_long_read(&pool
->pages_allocated
);
1221 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1224 * zs_map_object - get address of allocated object from handle.
1225 * @pool: pool from which the object was allocated
1226 * @handle: handle returned from zs_malloc
1227 * @mm: mapping mode to use
1229 * Before using an object allocated from zs_malloc, it must be mapped using
1230 * this function. When done with the object, it must be unmapped using
1233 * Only one object can be mapped per cpu at a time. There is no protection
1234 * against nested mappings.
1236 * This function returns with preemption and page faults disabled.
1238 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1241 struct zspage
*zspage
;
1243 unsigned long obj
, off
;
1244 unsigned int obj_idx
;
1246 struct size_class
*class;
1247 struct mapping_area
*area
;
1248 struct page
*pages
[2];
1252 * Because we use per-cpu mapping areas shared among the
1253 * pools/users, we can't allow mapping in interrupt context
1254 * because it can corrupt another users mappings.
1256 BUG_ON(in_interrupt());
1258 /* It guarantees it can get zspage from handle safely */
1259 read_lock(&pool
->migrate_lock
);
1260 obj
= handle_to_obj(handle
);
1261 obj_to_location(obj
, &page
, &obj_idx
);
1262 zspage
= get_zspage(page
);
1265 * migration cannot move any zpages in this zspage. Here, class->lock
1266 * is too heavy since callers would take some time until they calls
1267 * zs_unmap_object API so delegate the locking from class to zspage
1268 * which is smaller granularity.
1270 migrate_read_lock(zspage
);
1271 read_unlock(&pool
->migrate_lock
);
1273 class = zspage_class(pool
, zspage
);
1274 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1276 local_lock(&zs_map_area
.lock
);
1277 area
= this_cpu_ptr(&zs_map_area
);
1279 if (off
+ class->size
<= PAGE_SIZE
) {
1280 /* this object is contained entirely within a page */
1281 area
->vm_addr
= kmap_atomic(page
);
1282 ret
= area
->vm_addr
+ off
;
1286 /* this object spans two pages */
1288 pages
[1] = get_next_page(page
);
1291 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1293 if (likely(!ZsHugePage(zspage
)))
1294 ret
+= ZS_HANDLE_SIZE
;
1298 EXPORT_SYMBOL_GPL(zs_map_object
);
1300 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1302 struct zspage
*zspage
;
1304 unsigned long obj
, off
;
1305 unsigned int obj_idx
;
1307 struct size_class
*class;
1308 struct mapping_area
*area
;
1310 obj
= handle_to_obj(handle
);
1311 obj_to_location(obj
, &page
, &obj_idx
);
1312 zspage
= get_zspage(page
);
1313 class = zspage_class(pool
, zspage
);
1314 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1316 area
= this_cpu_ptr(&zs_map_area
);
1317 if (off
+ class->size
<= PAGE_SIZE
)
1318 kunmap_atomic(area
->vm_addr
);
1320 struct page
*pages
[2];
1323 pages
[1] = get_next_page(page
);
1326 __zs_unmap_object(area
, pages
, off
, class->size
);
1328 local_unlock(&zs_map_area
.lock
);
1330 migrate_read_unlock(zspage
);
1332 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1335 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1336 * zsmalloc &size_class.
1337 * @pool: zsmalloc pool to use
1339 * The function returns the size of the first huge class - any object of equal
1340 * or bigger size will be stored in zspage consisting of a single physical
1343 * Context: Any context.
1345 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1347 size_t zs_huge_class_size(struct zs_pool
*pool
)
1349 return huge_class_size
;
1351 EXPORT_SYMBOL_GPL(zs_huge_class_size
);
1353 static unsigned long obj_malloc(struct zs_pool
*pool
,
1354 struct zspage
*zspage
, unsigned long handle
)
1356 int i
, nr_page
, offset
;
1358 struct link_free
*link
;
1359 struct size_class
*class;
1361 struct page
*m_page
;
1362 unsigned long m_offset
;
1365 class = pool
->size_class
[zspage
->class];
1366 handle
|= OBJ_ALLOCATED_TAG
;
1367 obj
= get_freeobj(zspage
);
1369 offset
= obj
* class->size
;
1370 nr_page
= offset
>> PAGE_SHIFT
;
1371 m_offset
= offset
& ~PAGE_MASK
;
1372 m_page
= get_first_page(zspage
);
1374 for (i
= 0; i
< nr_page
; i
++)
1375 m_page
= get_next_page(m_page
);
1377 vaddr
= kmap_atomic(m_page
);
1378 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1379 set_freeobj(zspage
, link
->next
>> OBJ_TAG_BITS
);
1380 if (likely(!ZsHugePage(zspage
)))
1381 /* record handle in the header of allocated chunk */
1382 link
->handle
= handle
;
1384 /* record handle to page->index */
1385 zspage
->first_page
->index
= handle
;
1387 kunmap_atomic(vaddr
);
1388 mod_zspage_inuse(zspage
, 1);
1390 obj
= location_to_obj(m_page
, obj
);
1397 * zs_malloc - Allocate block of given size from pool.
1398 * @pool: pool to allocate from
1399 * @size: size of block to allocate
1400 * @gfp: gfp flags when allocating object
1402 * On success, handle to the allocated object is returned,
1404 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1406 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
, gfp_t gfp
)
1408 unsigned long handle
, obj
;
1409 struct size_class
*class;
1410 enum fullness_group newfg
;
1411 struct zspage
*zspage
;
1413 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1416 handle
= cache_alloc_handle(pool
, gfp
);
1420 /* extra space in chunk to keep the handle */
1421 size
+= ZS_HANDLE_SIZE
;
1422 class = pool
->size_class
[get_size_class_index(size
)];
1424 /* class->lock effectively protects the zpage migration */
1425 spin_lock(&class->lock
);
1426 zspage
= find_get_zspage(class);
1427 if (likely(zspage
)) {
1428 obj
= obj_malloc(pool
, zspage
, handle
);
1429 /* Now move the zspage to another fullness group, if required */
1430 fix_fullness_group(class, zspage
);
1431 record_obj(handle
, obj
);
1432 class_stat_inc(class, OBJ_USED
, 1);
1433 spin_unlock(&class->lock
);
1438 spin_unlock(&class->lock
);
1440 zspage
= alloc_zspage(pool
, class, gfp
);
1442 cache_free_handle(pool
, handle
);
1446 spin_lock(&class->lock
);
1447 obj
= obj_malloc(pool
, zspage
, handle
);
1448 newfg
= get_fullness_group(class, zspage
);
1449 insert_zspage(class, zspage
, newfg
);
1450 set_zspage_mapping(zspage
, class->index
, newfg
);
1451 record_obj(handle
, obj
);
1452 atomic_long_add(class->pages_per_zspage
,
1453 &pool
->pages_allocated
);
1454 class_stat_inc(class, OBJ_ALLOCATED
, class->objs_per_zspage
);
1455 class_stat_inc(class, OBJ_USED
, 1);
1457 /* We completely set up zspage so mark them as movable */
1458 SetZsPageMovable(pool
, zspage
);
1459 spin_unlock(&class->lock
);
1463 EXPORT_SYMBOL_GPL(zs_malloc
);
1465 static void obj_free(int class_size
, unsigned long obj
)
1467 struct link_free
*link
;
1468 struct zspage
*zspage
;
1469 struct page
*f_page
;
1470 unsigned long f_offset
;
1471 unsigned int f_objidx
;
1474 obj_to_location(obj
, &f_page
, &f_objidx
);
1475 f_offset
= (class_size
* f_objidx
) & ~PAGE_MASK
;
1476 zspage
= get_zspage(f_page
);
1478 vaddr
= kmap_atomic(f_page
);
1480 /* Insert this object in containing zspage's freelist */
1481 link
= (struct link_free
*)(vaddr
+ f_offset
);
1482 if (likely(!ZsHugePage(zspage
)))
1483 link
->next
= get_freeobj(zspage
) << OBJ_TAG_BITS
;
1486 kunmap_atomic(vaddr
);
1487 set_freeobj(zspage
, f_objidx
);
1488 mod_zspage_inuse(zspage
, -1);
1491 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1493 struct zspage
*zspage
;
1494 struct page
*f_page
;
1496 struct size_class
*class;
1497 enum fullness_group fullness
;
1499 if (unlikely(!handle
))
1503 * The pool->migrate_lock protects the race with zpage's migration
1504 * so it's safe to get the page from handle.
1506 read_lock(&pool
->migrate_lock
);
1507 obj
= handle_to_obj(handle
);
1508 obj_to_page(obj
, &f_page
);
1509 zspage
= get_zspage(f_page
);
1510 class = zspage_class(pool
, zspage
);
1511 spin_lock(&class->lock
);
1512 read_unlock(&pool
->migrate_lock
);
1514 obj_free(class->size
, obj
);
1515 class_stat_dec(class, OBJ_USED
, 1);
1516 fullness
= fix_fullness_group(class, zspage
);
1517 if (fullness
!= ZS_EMPTY
)
1520 free_zspage(pool
, class, zspage
);
1522 spin_unlock(&class->lock
);
1523 cache_free_handle(pool
, handle
);
1525 EXPORT_SYMBOL_GPL(zs_free
);
1527 static void zs_object_copy(struct size_class
*class, unsigned long dst
,
1530 struct page
*s_page
, *d_page
;
1531 unsigned int s_objidx
, d_objidx
;
1532 unsigned long s_off
, d_off
;
1533 void *s_addr
, *d_addr
;
1534 int s_size
, d_size
, size
;
1537 s_size
= d_size
= class->size
;
1539 obj_to_location(src
, &s_page
, &s_objidx
);
1540 obj_to_location(dst
, &d_page
, &d_objidx
);
1542 s_off
= (class->size
* s_objidx
) & ~PAGE_MASK
;
1543 d_off
= (class->size
* d_objidx
) & ~PAGE_MASK
;
1545 if (s_off
+ class->size
> PAGE_SIZE
)
1546 s_size
= PAGE_SIZE
- s_off
;
1548 if (d_off
+ class->size
> PAGE_SIZE
)
1549 d_size
= PAGE_SIZE
- d_off
;
1551 s_addr
= kmap_atomic(s_page
);
1552 d_addr
= kmap_atomic(d_page
);
1555 size
= min(s_size
, d_size
);
1556 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1559 if (written
== class->size
)
1567 if (s_off
>= PAGE_SIZE
) {
1568 kunmap_atomic(d_addr
);
1569 kunmap_atomic(s_addr
);
1570 s_page
= get_next_page(s_page
);
1571 s_addr
= kmap_atomic(s_page
);
1572 d_addr
= kmap_atomic(d_page
);
1573 s_size
= class->size
- written
;
1577 if (d_off
>= PAGE_SIZE
) {
1578 kunmap_atomic(d_addr
);
1579 d_page
= get_next_page(d_page
);
1580 d_addr
= kmap_atomic(d_page
);
1581 d_size
= class->size
- written
;
1586 kunmap_atomic(d_addr
);
1587 kunmap_atomic(s_addr
);
1591 * Find alloced object in zspage from index object and
1594 static unsigned long find_alloced_obj(struct size_class
*class,
1595 struct page
*page
, int *obj_idx
)
1598 int index
= *obj_idx
;
1599 unsigned long handle
= 0;
1600 void *addr
= kmap_atomic(page
);
1602 offset
= get_first_obj_offset(page
);
1603 offset
+= class->size
* index
;
1605 while (offset
< PAGE_SIZE
) {
1606 if (obj_allocated(page
, addr
+ offset
, &handle
))
1609 offset
+= class->size
;
1613 kunmap_atomic(addr
);
1620 struct zs_compact_control
{
1621 /* Source spage for migration which could be a subpage of zspage */
1622 struct page
*s_page
;
1623 /* Destination page for migration which should be a first page
1625 struct page
*d_page
;
1626 /* Starting object index within @s_page which used for live object
1627 * in the subpage. */
1631 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1632 struct zs_compact_control
*cc
)
1634 unsigned long used_obj
, free_obj
;
1635 unsigned long handle
;
1636 struct page
*s_page
= cc
->s_page
;
1637 struct page
*d_page
= cc
->d_page
;
1638 int obj_idx
= cc
->obj_idx
;
1642 handle
= find_alloced_obj(class, s_page
, &obj_idx
);
1644 s_page
= get_next_page(s_page
);
1651 /* Stop if there is no more space */
1652 if (zspage_full(class, get_zspage(d_page
))) {
1657 used_obj
= handle_to_obj(handle
);
1658 free_obj
= obj_malloc(pool
, get_zspage(d_page
), handle
);
1659 zs_object_copy(class, free_obj
, used_obj
);
1661 record_obj(handle
, free_obj
);
1662 obj_free(class->size
, used_obj
);
1665 /* Remember last position in this iteration */
1666 cc
->s_page
= s_page
;
1667 cc
->obj_idx
= obj_idx
;
1672 static struct zspage
*isolate_zspage(struct size_class
*class, bool source
)
1675 struct zspage
*zspage
;
1676 enum fullness_group fg
[2] = {ZS_ALMOST_EMPTY
, ZS_ALMOST_FULL
};
1679 fg
[0] = ZS_ALMOST_FULL
;
1680 fg
[1] = ZS_ALMOST_EMPTY
;
1683 for (i
= 0; i
< 2; i
++) {
1684 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
[i
]],
1685 struct zspage
, list
);
1687 remove_zspage(class, zspage
, fg
[i
]);
1696 * putback_zspage - add @zspage into right class's fullness list
1697 * @class: destination class
1698 * @zspage: target page
1700 * Return @zspage's fullness_group
1702 static enum fullness_group
putback_zspage(struct size_class
*class,
1703 struct zspage
*zspage
)
1705 enum fullness_group fullness
;
1707 fullness
= get_fullness_group(class, zspage
);
1708 insert_zspage(class, zspage
, fullness
);
1709 set_zspage_mapping(zspage
, class->index
, fullness
);
1714 #ifdef CONFIG_COMPACTION
1716 * To prevent zspage destroy during migration, zspage freeing should
1717 * hold locks of all pages in the zspage.
1719 static void lock_zspage(struct zspage
*zspage
)
1721 struct page
*page
= get_first_page(zspage
);
1725 } while ((page
= get_next_page(page
)) != NULL
);
1728 static int zs_init_fs_context(struct fs_context
*fc
)
1730 return init_pseudo(fc
, ZSMALLOC_MAGIC
) ? 0 : -ENOMEM
;
1733 static struct file_system_type zsmalloc_fs
= {
1735 .init_fs_context
= zs_init_fs_context
,
1736 .kill_sb
= kill_anon_super
,
1739 static int zsmalloc_mount(void)
1743 zsmalloc_mnt
= kern_mount(&zsmalloc_fs
);
1744 if (IS_ERR(zsmalloc_mnt
))
1745 ret
= PTR_ERR(zsmalloc_mnt
);
1750 static void zsmalloc_unmount(void)
1752 kern_unmount(zsmalloc_mnt
);
1755 static void migrate_lock_init(struct zspage
*zspage
)
1757 rwlock_init(&zspage
->lock
);
1760 static void migrate_read_lock(struct zspage
*zspage
) __acquires(&zspage
->lock
)
1762 read_lock(&zspage
->lock
);
1765 static void migrate_read_unlock(struct zspage
*zspage
) __releases(&zspage
->lock
)
1767 read_unlock(&zspage
->lock
);
1770 static void migrate_write_lock(struct zspage
*zspage
)
1772 write_lock(&zspage
->lock
);
1775 static void migrate_write_lock_nested(struct zspage
*zspage
)
1777 write_lock_nested(&zspage
->lock
, SINGLE_DEPTH_NESTING
);
1780 static void migrate_write_unlock(struct zspage
*zspage
)
1782 write_unlock(&zspage
->lock
);
1785 /* Number of isolated subpage for *page migration* in this zspage */
1786 static void inc_zspage_isolation(struct zspage
*zspage
)
1791 static void dec_zspage_isolation(struct zspage
*zspage
)
1793 VM_BUG_ON(zspage
->isolated
== 0);
1797 static void replace_sub_page(struct size_class
*class, struct zspage
*zspage
,
1798 struct page
*newpage
, struct page
*oldpage
)
1801 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
] = {NULL
, };
1804 page
= get_first_page(zspage
);
1806 if (page
== oldpage
)
1807 pages
[idx
] = newpage
;
1811 } while ((page
= get_next_page(page
)) != NULL
);
1813 create_page_chain(class, zspage
, pages
);
1814 set_first_obj_offset(newpage
, get_first_obj_offset(oldpage
));
1815 if (unlikely(ZsHugePage(zspage
)))
1816 newpage
->index
= oldpage
->index
;
1817 __SetPageMovable(newpage
, page_mapping(oldpage
));
1820 static bool zs_page_isolate(struct page
*page
, isolate_mode_t mode
)
1822 struct zspage
*zspage
;
1825 * Page is locked so zspage couldn't be destroyed. For detail, look at
1826 * lock_zspage in free_zspage.
1828 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1829 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1831 zspage
= get_zspage(page
);
1832 migrate_write_lock(zspage
);
1833 inc_zspage_isolation(zspage
);
1834 migrate_write_unlock(zspage
);
1839 static int zs_page_migrate(struct address_space
*mapping
, struct page
*newpage
,
1840 struct page
*page
, enum migrate_mode mode
)
1842 struct zs_pool
*pool
;
1843 struct size_class
*class;
1844 struct zspage
*zspage
;
1846 void *s_addr
, *d_addr
, *addr
;
1848 unsigned long handle
;
1849 unsigned long old_obj
, new_obj
;
1850 unsigned int obj_idx
;
1853 * We cannot support the _NO_COPY case here, because copy needs to
1854 * happen under the zs lock, which does not work with
1855 * MIGRATE_SYNC_NO_COPY workflow.
1857 if (mode
== MIGRATE_SYNC_NO_COPY
)
1860 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1861 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1863 pool
= mapping
->private_data
;
1866 * The pool migrate_lock protects the race between zpage migration
1869 write_lock(&pool
->migrate_lock
);
1870 zspage
= get_zspage(page
);
1871 class = zspage_class(pool
, zspage
);
1874 * the class lock protects zpage alloc/free in the zspage.
1876 spin_lock(&class->lock
);
1877 /* the migrate_write_lock protects zpage access via zs_map_object */
1878 migrate_write_lock(zspage
);
1880 offset
= get_first_obj_offset(page
);
1881 s_addr
= kmap_atomic(page
);
1884 * Here, any user cannot access all objects in the zspage so let's move.
1886 d_addr
= kmap_atomic(newpage
);
1887 memcpy(d_addr
, s_addr
, PAGE_SIZE
);
1888 kunmap_atomic(d_addr
);
1890 for (addr
= s_addr
+ offset
; addr
< s_addr
+ PAGE_SIZE
;
1891 addr
+= class->size
) {
1892 if (obj_allocated(page
, addr
, &handle
)) {
1894 old_obj
= handle_to_obj(handle
);
1895 obj_to_location(old_obj
, &dummy
, &obj_idx
);
1896 new_obj
= (unsigned long)location_to_obj(newpage
,
1898 record_obj(handle
, new_obj
);
1901 kunmap_atomic(s_addr
);
1903 replace_sub_page(class, zspage
, newpage
, page
);
1905 * Since we complete the data copy and set up new zspage structure,
1906 * it's okay to release migration_lock.
1908 write_unlock(&pool
->migrate_lock
);
1909 spin_unlock(&class->lock
);
1910 dec_zspage_isolation(zspage
);
1911 migrate_write_unlock(zspage
);
1914 if (page_zone(newpage
) != page_zone(page
)) {
1915 dec_zone_page_state(page
, NR_ZSPAGES
);
1916 inc_zone_page_state(newpage
, NR_ZSPAGES
);
1922 return MIGRATEPAGE_SUCCESS
;
1925 static void zs_page_putback(struct page
*page
)
1927 struct zspage
*zspage
;
1929 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1930 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1932 zspage
= get_zspage(page
);
1933 migrate_write_lock(zspage
);
1934 dec_zspage_isolation(zspage
);
1935 migrate_write_unlock(zspage
);
1938 static const struct address_space_operations zsmalloc_aops
= {
1939 .isolate_page
= zs_page_isolate
,
1940 .migratepage
= zs_page_migrate
,
1941 .putback_page
= zs_page_putback
,
1944 static int zs_register_migration(struct zs_pool
*pool
)
1946 pool
->inode
= alloc_anon_inode(zsmalloc_mnt
->mnt_sb
);
1947 if (IS_ERR(pool
->inode
)) {
1952 pool
->inode
->i_mapping
->private_data
= pool
;
1953 pool
->inode
->i_mapping
->a_ops
= &zsmalloc_aops
;
1957 static void zs_unregister_migration(struct zs_pool
*pool
)
1959 flush_work(&pool
->free_work
);
1964 * Caller should hold page_lock of all pages in the zspage
1965 * In here, we cannot use zspage meta data.
1967 static void async_free_zspage(struct work_struct
*work
)
1970 struct size_class
*class;
1971 unsigned int class_idx
;
1972 enum fullness_group fullness
;
1973 struct zspage
*zspage
, *tmp
;
1974 LIST_HEAD(free_pages
);
1975 struct zs_pool
*pool
= container_of(work
, struct zs_pool
,
1978 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
1979 class = pool
->size_class
[i
];
1980 if (class->index
!= i
)
1983 spin_lock(&class->lock
);
1984 list_splice_init(&class->fullness_list
[ZS_EMPTY
], &free_pages
);
1985 spin_unlock(&class->lock
);
1988 list_for_each_entry_safe(zspage
, tmp
, &free_pages
, list
) {
1989 list_del(&zspage
->list
);
1990 lock_zspage(zspage
);
1992 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1993 VM_BUG_ON(fullness
!= ZS_EMPTY
);
1994 class = pool
->size_class
[class_idx
];
1995 spin_lock(&class->lock
);
1996 __free_zspage(pool
, class, zspage
);
1997 spin_unlock(&class->lock
);
2001 static void kick_deferred_free(struct zs_pool
*pool
)
2003 schedule_work(&pool
->free_work
);
2006 static void init_deferred_free(struct zs_pool
*pool
)
2008 INIT_WORK(&pool
->free_work
, async_free_zspage
);
2011 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
)
2013 struct page
*page
= get_first_page(zspage
);
2016 WARN_ON(!trylock_page(page
));
2017 __SetPageMovable(page
, pool
->inode
->i_mapping
);
2019 } while ((page
= get_next_page(page
)) != NULL
);
2025 * Based on the number of unused allocated objects calculate
2026 * and return the number of pages that we can free.
2028 static unsigned long zs_can_compact(struct size_class
*class)
2030 unsigned long obj_wasted
;
2031 unsigned long obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
2032 unsigned long obj_used
= zs_stat_get(class, OBJ_USED
);
2034 if (obj_allocated
<= obj_used
)
2037 obj_wasted
= obj_allocated
- obj_used
;
2038 obj_wasted
/= class->objs_per_zspage
;
2040 return obj_wasted
* class->pages_per_zspage
;
2043 static unsigned long __zs_compact(struct zs_pool
*pool
,
2044 struct size_class
*class)
2046 struct zs_compact_control cc
;
2047 struct zspage
*src_zspage
;
2048 struct zspage
*dst_zspage
= NULL
;
2049 unsigned long pages_freed
= 0;
2051 /* protect the race between zpage migration and zs_free */
2052 write_lock(&pool
->migrate_lock
);
2053 /* protect zpage allocation/free */
2054 spin_lock(&class->lock
);
2055 while ((src_zspage
= isolate_zspage(class, true))) {
2056 /* protect someone accessing the zspage(i.e., zs_map_object) */
2057 migrate_write_lock(src_zspage
);
2059 if (!zs_can_compact(class))
2063 cc
.s_page
= get_first_page(src_zspage
);
2065 while ((dst_zspage
= isolate_zspage(class, false))) {
2066 migrate_write_lock_nested(dst_zspage
);
2068 cc
.d_page
= get_first_page(dst_zspage
);
2070 * If there is no more space in dst_page, resched
2071 * and see if anyone had allocated another zspage.
2073 if (!migrate_zspage(pool
, class, &cc
))
2076 putback_zspage(class, dst_zspage
);
2077 migrate_write_unlock(dst_zspage
);
2079 if (rwlock_is_contended(&pool
->migrate_lock
))
2083 /* Stop if we couldn't find slot */
2084 if (dst_zspage
== NULL
)
2087 putback_zspage(class, dst_zspage
);
2088 migrate_write_unlock(dst_zspage
);
2090 if (putback_zspage(class, src_zspage
) == ZS_EMPTY
) {
2091 migrate_write_unlock(src_zspage
);
2092 free_zspage(pool
, class, src_zspage
);
2093 pages_freed
+= class->pages_per_zspage
;
2095 migrate_write_unlock(src_zspage
);
2096 spin_unlock(&class->lock
);
2097 write_unlock(&pool
->migrate_lock
);
2099 write_lock(&pool
->migrate_lock
);
2100 spin_lock(&class->lock
);
2104 putback_zspage(class, src_zspage
);
2105 migrate_write_unlock(src_zspage
);
2108 spin_unlock(&class->lock
);
2109 write_unlock(&pool
->migrate_lock
);
2114 unsigned long zs_compact(struct zs_pool
*pool
)
2117 struct size_class
*class;
2118 unsigned long pages_freed
= 0;
2120 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2121 class = pool
->size_class
[i
];
2124 if (class->index
!= i
)
2126 pages_freed
+= __zs_compact(pool
, class);
2128 atomic_long_add(pages_freed
, &pool
->stats
.pages_compacted
);
2132 EXPORT_SYMBOL_GPL(zs_compact
);
2134 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
2136 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
2138 EXPORT_SYMBOL_GPL(zs_pool_stats
);
2140 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
2141 struct shrink_control
*sc
)
2143 unsigned long pages_freed
;
2144 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2148 * Compact classes and calculate compaction delta.
2149 * Can run concurrently with a manually triggered
2150 * (by user) compaction.
2152 pages_freed
= zs_compact(pool
);
2154 return pages_freed
? pages_freed
: SHRINK_STOP
;
2157 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
2158 struct shrink_control
*sc
)
2161 struct size_class
*class;
2162 unsigned long pages_to_free
= 0;
2163 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2166 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2167 class = pool
->size_class
[i
];
2170 if (class->index
!= i
)
2173 pages_to_free
+= zs_can_compact(class);
2176 return pages_to_free
;
2179 static void zs_unregister_shrinker(struct zs_pool
*pool
)
2181 unregister_shrinker(&pool
->shrinker
);
2184 static int zs_register_shrinker(struct zs_pool
*pool
)
2186 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
2187 pool
->shrinker
.count_objects
= zs_shrinker_count
;
2188 pool
->shrinker
.batch
= 0;
2189 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
2191 return register_shrinker(&pool
->shrinker
);
2195 * zs_create_pool - Creates an allocation pool to work from.
2196 * @name: pool name to be created
2198 * This function must be called before anything when using
2199 * the zsmalloc allocator.
2201 * On success, a pointer to the newly created pool is returned,
2204 struct zs_pool
*zs_create_pool(const char *name
)
2207 struct zs_pool
*pool
;
2208 struct size_class
*prev_class
= NULL
;
2210 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
2214 init_deferred_free(pool
);
2215 rwlock_init(&pool
->migrate_lock
);
2217 pool
->name
= kstrdup(name
, GFP_KERNEL
);
2221 if (create_cache(pool
))
2225 * Iterate reversely, because, size of size_class that we want to use
2226 * for merging should be larger or equal to current size.
2228 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2230 int pages_per_zspage
;
2231 int objs_per_zspage
;
2232 struct size_class
*class;
2235 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
2236 if (size
> ZS_MAX_ALLOC_SIZE
)
2237 size
= ZS_MAX_ALLOC_SIZE
;
2238 pages_per_zspage
= get_pages_per_zspage(size
);
2239 objs_per_zspage
= pages_per_zspage
* PAGE_SIZE
/ size
;
2242 * We iterate from biggest down to smallest classes,
2243 * so huge_class_size holds the size of the first huge
2244 * class. Any object bigger than or equal to that will
2245 * endup in the huge class.
2247 if (pages_per_zspage
!= 1 && objs_per_zspage
!= 1 &&
2249 huge_class_size
= size
;
2251 * The object uses ZS_HANDLE_SIZE bytes to store the
2252 * handle. We need to subtract it, because zs_malloc()
2253 * unconditionally adds handle size before it performs
2254 * size class search - so object may be smaller than
2255 * huge class size, yet it still can end up in the huge
2256 * class because it grows by ZS_HANDLE_SIZE extra bytes
2257 * right before class lookup.
2259 huge_class_size
-= (ZS_HANDLE_SIZE
- 1);
2263 * size_class is used for normal zsmalloc operation such
2264 * as alloc/free for that size. Although it is natural that we
2265 * have one size_class for each size, there is a chance that we
2266 * can get more memory utilization if we use one size_class for
2267 * many different sizes whose size_class have same
2268 * characteristics. So, we makes size_class point to
2269 * previous size_class if possible.
2272 if (can_merge(prev_class
, pages_per_zspage
, objs_per_zspage
)) {
2273 pool
->size_class
[i
] = prev_class
;
2278 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
2284 class->pages_per_zspage
= pages_per_zspage
;
2285 class->objs_per_zspage
= objs_per_zspage
;
2286 spin_lock_init(&class->lock
);
2287 pool
->size_class
[i
] = class;
2288 for (fullness
= ZS_EMPTY
; fullness
< NR_ZS_FULLNESS
;
2290 INIT_LIST_HEAD(&class->fullness_list
[fullness
]);
2295 /* debug only, don't abort if it fails */
2296 zs_pool_stat_create(pool
, name
);
2298 if (zs_register_migration(pool
))
2302 * Not critical since shrinker is only used to trigger internal
2303 * defragmentation of the pool which is pretty optional thing. If
2304 * registration fails we still can use the pool normally and user can
2305 * trigger compaction manually. Thus, ignore return code.
2307 zs_register_shrinker(pool
);
2312 zs_destroy_pool(pool
);
2315 EXPORT_SYMBOL_GPL(zs_create_pool
);
2317 void zs_destroy_pool(struct zs_pool
*pool
)
2321 zs_unregister_shrinker(pool
);
2322 zs_unregister_migration(pool
);
2323 zs_pool_stat_destroy(pool
);
2325 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
2327 struct size_class
*class = pool
->size_class
[i
];
2332 if (class->index
!= i
)
2335 for (fg
= ZS_EMPTY
; fg
< NR_ZS_FULLNESS
; fg
++) {
2336 if (!list_empty(&class->fullness_list
[fg
])) {
2337 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2344 destroy_cache(pool
);
2348 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2350 static int __init
zs_init(void)
2354 ret
= zsmalloc_mount();
2358 ret
= cpuhp_setup_state(CPUHP_MM_ZS_PREPARE
, "mm/zsmalloc:prepare",
2359 zs_cpu_prepare
, zs_cpu_dead
);
2364 zpool_register_driver(&zs_zpool_driver
);
2377 static void __exit
zs_exit(void)
2380 zpool_unregister_driver(&zs_zpool_driver
);
2383 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE
);
2388 module_init(zs_init
);
2389 module_exit(zs_exit
);
2391 MODULE_LICENSE("Dual BSD/GPL");
2392 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");