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->first_page: points to the first component (0-order) page
20 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
27 * For _first_ page only:
29 * page->private (union with page->first_page): refers to the
30 * component page after the first page
31 * If the page is first_page for huge object, it stores handle.
32 * Look at size_class->huge.
33 * page->freelist: points to the first free object in zspage.
34 * Free objects are linked together using in-place
36 * page->objects: maximum number of objects we can store in this
37 * zspage (class->zspage_order * PAGE_SIZE / class->size)
38 * page->lru: links together first pages of various zspages.
39 * Basically forming list of zspages in a fullness group.
40 * page->mapping: class index and fullness group of the zspage
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
48 #ifdef CONFIG_ZSMALLOC_DEBUG
52 #include <linux/module.h>
53 #include <linux/kernel.h>
54 #include <linux/sched.h>
55 #include <linux/bitops.h>
56 #include <linux/errno.h>
57 #include <linux/highmem.h>
58 #include <linux/string.h>
59 #include <linux/slab.h>
60 #include <asm/tlbflush.h>
61 #include <asm/pgtable.h>
62 #include <linux/cpumask.h>
63 #include <linux/cpu.h>
64 #include <linux/vmalloc.h>
65 #include <linux/hardirq.h>
66 #include <linux/spinlock.h>
67 #include <linux/types.h>
68 #include <linux/debugfs.h>
69 #include <linux/zsmalloc.h>
70 #include <linux/zpool.h>
73 * This must be power of 2 and greater than of 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 * as single (unsigned long) handle value.
93 * Note that object index <obj_idx> is relative to system
94 * page <PFN> it is stored in, so for each sub-page belonging
95 * to a zspage, obj_idx starts with 0.
97 * This is made more complicated by various memory models and PAE.
100 #ifndef MAX_PHYSMEM_BITS
101 #ifdef CONFIG_HIGHMEM64G
102 #define MAX_PHYSMEM_BITS 36
103 #else /* !CONFIG_HIGHMEM64G */
105 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
108 #define MAX_PHYSMEM_BITS BITS_PER_LONG
111 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
114 * Memory for allocating for handle keeps object position by
115 * encoding <page, obj_idx> and the encoded value has a room
116 * in least bit(ie, look at obj_to_location).
117 * We use the bit to synchronize between object access by
118 * user and migration.
120 #define HANDLE_PIN_BIT 0
123 * Head in allocated object should have OBJ_ALLOCATED_TAG
124 * to identify the object was allocated or not.
125 * It's okay to add the status bit in the least bit because
126 * header keeps handle which is 4byte-aligned address so we
127 * have room for two bit at least.
129 #define OBJ_ALLOCATED_TAG 1
130 #define OBJ_TAG_BITS 1
131 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
132 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
134 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
135 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
136 #define ZS_MIN_ALLOC_SIZE \
137 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
138 /* each chunk includes extra space to keep handle */
139 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
142 * On systems with 4K page size, this gives 255 size classes! There is a
144 * - Large number of size classes is potentially wasteful as free page are
145 * spread across these classes
146 * - Small number of size classes causes large internal fragmentation
147 * - Probably its better to use specific size classes (empirically
148 * determined). NOTE: all those class sizes must be set as multiple of
149 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
151 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
154 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
157 * We do not maintain any list for completely empty or full pages
159 enum fullness_group
{
162 _ZS_NR_FULLNESS_GROUPS
,
176 #ifdef CONFIG_ZSMALLOC_STAT
178 static struct dentry
*zs_stat_root
;
180 struct zs_size_stat
{
181 unsigned long objs
[NR_ZS_STAT_TYPE
];
187 * number of size_classes
189 static int zs_size_classes
;
192 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
194 * n = number of allocated objects
195 * N = total number of objects zspage can store
196 * f = fullness_threshold_frac
198 * Similarly, we assign zspage to:
199 * ZS_ALMOST_FULL when n > N / f
200 * ZS_EMPTY when n == 0
201 * ZS_FULL when n == N
203 * (see: fix_fullness_group())
205 static const int fullness_threshold_frac
= 4;
209 * Size of objects stored in this class. Must be multiple
215 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
216 int pages_per_zspage
;
217 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
220 #ifdef CONFIG_ZSMALLOC_STAT
221 struct zs_size_stat stats
;
226 struct page
*fullness_list
[_ZS_NR_FULLNESS_GROUPS
];
230 * Placed within free objects to form a singly linked list.
231 * For every zspage, first_page->freelist gives head of this list.
233 * This must be power of 2 and less than or equal to ZS_ALIGN
238 * Position of next free chunk (encodes <PFN, obj_idx>)
239 * It's valid for non-allocated object
243 * Handle of allocated object.
245 unsigned long handle
;
252 struct size_class
**size_class
;
253 struct kmem_cache
*handle_cachep
;
255 gfp_t flags
; /* allocation flags used when growing pool */
256 atomic_long_t pages_allocated
;
258 #ifdef CONFIG_ZSMALLOC_STAT
259 struct dentry
*stat_dentry
;
264 * A zspage's class index and fullness group
265 * are encoded in its (first)page->mapping
267 #define CLASS_IDX_BITS 28
268 #define FULLNESS_BITS 4
269 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
270 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
272 struct mapping_area
{
273 #ifdef CONFIG_PGTABLE_MAPPING
274 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
276 char *vm_buf
; /* copy buffer for objects that span pages */
278 char *vm_addr
; /* address of kmap_atomic()'ed pages */
279 enum zs_mapmode vm_mm
; /* mapping mode */
283 static int create_handle_cache(struct zs_pool
*pool
)
285 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
287 return pool
->handle_cachep
? 0 : 1;
290 static void destroy_handle_cache(struct zs_pool
*pool
)
292 if (pool
->handle_cachep
)
293 kmem_cache_destroy(pool
->handle_cachep
);
296 static unsigned long alloc_handle(struct zs_pool
*pool
)
298 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
299 pool
->flags
& ~__GFP_HIGHMEM
);
302 static void free_handle(struct zs_pool
*pool
, unsigned long handle
)
304 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
307 static void record_obj(unsigned long handle
, unsigned long obj
)
309 *(unsigned long *)handle
= obj
;
316 static void *zs_zpool_create(char *name
, gfp_t gfp
, struct zpool_ops
*zpool_ops
)
318 return zs_create_pool(name
, gfp
);
321 static void zs_zpool_destroy(void *pool
)
323 zs_destroy_pool(pool
);
326 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
327 unsigned long *handle
)
329 *handle
= zs_malloc(pool
, size
);
330 return *handle
? 0 : -1;
332 static void zs_zpool_free(void *pool
, unsigned long handle
)
334 zs_free(pool
, handle
);
337 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
338 unsigned int *reclaimed
)
343 static void *zs_zpool_map(void *pool
, unsigned long handle
,
344 enum zpool_mapmode mm
)
346 enum zs_mapmode zs_mm
;
355 case ZPOOL_MM_RW
: /* fallthru */
361 return zs_map_object(pool
, handle
, zs_mm
);
363 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
365 zs_unmap_object(pool
, handle
);
368 static u64
zs_zpool_total_size(void *pool
)
370 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
373 static struct zpool_driver zs_zpool_driver
= {
375 .owner
= THIS_MODULE
,
376 .create
= zs_zpool_create
,
377 .destroy
= zs_zpool_destroy
,
378 .malloc
= zs_zpool_malloc
,
379 .free
= zs_zpool_free
,
380 .shrink
= zs_zpool_shrink
,
382 .unmap
= zs_zpool_unmap
,
383 .total_size
= zs_zpool_total_size
,
386 MODULE_ALIAS("zpool-zsmalloc");
387 #endif /* CONFIG_ZPOOL */
389 static unsigned int get_maxobj_per_zspage(int size
, int pages_per_zspage
)
391 return pages_per_zspage
* PAGE_SIZE
/ size
;
394 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
395 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
397 static int is_first_page(struct page
*page
)
399 return PagePrivate(page
);
402 static int is_last_page(struct page
*page
)
404 return PagePrivate2(page
);
407 static void get_zspage_mapping(struct page
*page
, unsigned int *class_idx
,
408 enum fullness_group
*fullness
)
411 BUG_ON(!is_first_page(page
));
413 m
= (unsigned long)page
->mapping
;
414 *fullness
= m
& FULLNESS_MASK
;
415 *class_idx
= (m
>> FULLNESS_BITS
) & CLASS_IDX_MASK
;
418 static void set_zspage_mapping(struct page
*page
, unsigned int class_idx
,
419 enum fullness_group fullness
)
422 BUG_ON(!is_first_page(page
));
424 m
= ((class_idx
& CLASS_IDX_MASK
) << FULLNESS_BITS
) |
425 (fullness
& FULLNESS_MASK
);
426 page
->mapping
= (struct address_space
*)m
;
430 * zsmalloc divides the pool into various size classes where each
431 * class maintains a list of zspages where each zspage is divided
432 * into equal sized chunks. Each allocation falls into one of these
433 * classes depending on its size. This function returns index of the
434 * size class which has chunk size big enough to hold the give size.
436 static int get_size_class_index(int size
)
440 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
441 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
442 ZS_SIZE_CLASS_DELTA
);
444 return min(zs_size_classes
- 1, idx
);
447 #ifdef CONFIG_ZSMALLOC_STAT
449 static inline void zs_stat_inc(struct size_class
*class,
450 enum zs_stat_type type
, unsigned long cnt
)
452 class->stats
.objs
[type
] += cnt
;
455 static inline void zs_stat_dec(struct size_class
*class,
456 enum zs_stat_type type
, unsigned long cnt
)
458 class->stats
.objs
[type
] -= cnt
;
461 static inline unsigned long zs_stat_get(struct size_class
*class,
462 enum zs_stat_type type
)
464 return class->stats
.objs
[type
];
467 static int __init
zs_stat_init(void)
469 if (!debugfs_initialized())
472 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
479 static void __exit
zs_stat_exit(void)
481 debugfs_remove_recursive(zs_stat_root
);
484 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
487 struct zs_pool
*pool
= s
->private;
488 struct size_class
*class;
490 unsigned long class_almost_full
, class_almost_empty
;
491 unsigned long obj_allocated
, obj_used
, pages_used
;
492 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
493 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
495 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
496 "class", "size", "almost_full", "almost_empty",
497 "obj_allocated", "obj_used", "pages_used",
500 for (i
= 0; i
< zs_size_classes
; i
++) {
501 class = pool
->size_class
[i
];
503 if (class->index
!= i
)
506 spin_lock(&class->lock
);
507 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
508 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
509 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
510 obj_used
= zs_stat_get(class, OBJ_USED
);
511 spin_unlock(&class->lock
);
513 objs_per_zspage
= get_maxobj_per_zspage(class->size
,
514 class->pages_per_zspage
);
515 pages_used
= obj_allocated
/ objs_per_zspage
*
516 class->pages_per_zspage
;
518 seq_printf(s
, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
519 i
, class->size
, class_almost_full
, class_almost_empty
,
520 obj_allocated
, obj_used
, pages_used
,
521 class->pages_per_zspage
);
523 total_class_almost_full
+= class_almost_full
;
524 total_class_almost_empty
+= class_almost_empty
;
525 total_objs
+= obj_allocated
;
526 total_used_objs
+= obj_used
;
527 total_pages
+= pages_used
;
531 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
532 "Total", "", total_class_almost_full
,
533 total_class_almost_empty
, total_objs
,
534 total_used_objs
, total_pages
);
539 static int zs_stats_size_open(struct inode
*inode
, struct file
*file
)
541 return single_open(file
, zs_stats_size_show
, inode
->i_private
);
544 static const struct file_operations zs_stat_size_ops
= {
545 .open
= zs_stats_size_open
,
548 .release
= single_release
,
551 static int zs_pool_stat_create(char *name
, struct zs_pool
*pool
)
553 struct dentry
*entry
;
558 entry
= debugfs_create_dir(name
, zs_stat_root
);
560 pr_warn("debugfs dir <%s> creation failed\n", name
);
563 pool
->stat_dentry
= entry
;
565 entry
= debugfs_create_file("classes", S_IFREG
| S_IRUGO
,
566 pool
->stat_dentry
, pool
, &zs_stat_size_ops
);
568 pr_warn("%s: debugfs file entry <%s> creation failed\n",
576 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
578 debugfs_remove_recursive(pool
->stat_dentry
);
581 #else /* CONFIG_ZSMALLOC_STAT */
583 static inline void zs_stat_inc(struct size_class
*class,
584 enum zs_stat_type type
, unsigned long cnt
)
588 static inline void zs_stat_dec(struct size_class
*class,
589 enum zs_stat_type type
, unsigned long cnt
)
593 static inline unsigned long zs_stat_get(struct size_class
*class,
594 enum zs_stat_type type
)
599 static int __init
zs_stat_init(void)
604 static void __exit
zs_stat_exit(void)
608 static inline int zs_pool_stat_create(char *name
, struct zs_pool
*pool
)
613 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
621 * For each size class, zspages are divided into different groups
622 * depending on how "full" they are. This was done so that we could
623 * easily find empty or nearly empty zspages when we try to shrink
624 * the pool (not yet implemented). This function returns fullness
625 * status of the given page.
627 static enum fullness_group
get_fullness_group(struct page
*page
)
629 int inuse
, max_objects
;
630 enum fullness_group fg
;
631 BUG_ON(!is_first_page(page
));
634 max_objects
= page
->objects
;
638 else if (inuse
== max_objects
)
640 else if (inuse
<= 3 * max_objects
/ fullness_threshold_frac
)
641 fg
= ZS_ALMOST_EMPTY
;
649 * Each size class maintains various freelists and zspages are assigned
650 * to one of these freelists based on the number of live objects they
651 * have. This functions inserts the given zspage into the freelist
652 * identified by <class, fullness_group>.
654 static void insert_zspage(struct page
*page
, struct size_class
*class,
655 enum fullness_group fullness
)
659 BUG_ON(!is_first_page(page
));
661 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
664 head
= &class->fullness_list
[fullness
];
666 list_add_tail(&page
->lru
, &(*head
)->lru
);
669 zs_stat_inc(class, fullness
== ZS_ALMOST_EMPTY
?
670 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
674 * This function removes the given zspage from the freelist identified
675 * by <class, fullness_group>.
677 static void remove_zspage(struct page
*page
, struct size_class
*class,
678 enum fullness_group fullness
)
682 BUG_ON(!is_first_page(page
));
684 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
687 head
= &class->fullness_list
[fullness
];
689 if (list_empty(&(*head
)->lru
))
691 else if (*head
== page
)
692 *head
= (struct page
*)list_entry((*head
)->lru
.next
,
695 list_del_init(&page
->lru
);
696 zs_stat_dec(class, fullness
== ZS_ALMOST_EMPTY
?
697 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
701 * Each size class maintains zspages in different fullness groups depending
702 * on the number of live objects they contain. When allocating or freeing
703 * objects, the fullness status of the page can change, say, from ALMOST_FULL
704 * to ALMOST_EMPTY when freeing an object. This function checks if such
705 * a status change has occurred for the given page and accordingly moves the
706 * page from the freelist of the old fullness group to that of the new
709 static enum fullness_group
fix_fullness_group(struct size_class
*class,
713 enum fullness_group currfg
, newfg
;
715 BUG_ON(!is_first_page(page
));
717 get_zspage_mapping(page
, &class_idx
, &currfg
);
718 newfg
= get_fullness_group(page
);
722 remove_zspage(page
, class, currfg
);
723 insert_zspage(page
, class, newfg
);
724 set_zspage_mapping(page
, class_idx
, newfg
);
731 * We have to decide on how many pages to link together
732 * to form a zspage for each size class. This is important
733 * to reduce wastage due to unusable space left at end of
734 * each zspage which is given as:
735 * wastage = Zp % class_size
736 * usage = Zp - wastage
737 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
739 * For example, for size class of 3/8 * PAGE_SIZE, we should
740 * link together 3 PAGE_SIZE sized pages to form a zspage
741 * since then we can perfectly fit in 8 such objects.
743 static int get_pages_per_zspage(int class_size
)
745 int i
, max_usedpc
= 0;
746 /* zspage order which gives maximum used size per KB */
747 int max_usedpc_order
= 1;
749 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
753 zspage_size
= i
* PAGE_SIZE
;
754 waste
= zspage_size
% class_size
;
755 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
757 if (usedpc
> max_usedpc
) {
759 max_usedpc_order
= i
;
763 return max_usedpc_order
;
767 * A single 'zspage' is composed of many system pages which are
768 * linked together using fields in struct page. This function finds
769 * the first/head page, given any component page of a zspage.
771 static struct page
*get_first_page(struct page
*page
)
773 if (is_first_page(page
))
776 return page
->first_page
;
779 static struct page
*get_next_page(struct page
*page
)
783 if (is_last_page(page
))
785 else if (is_first_page(page
))
786 next
= (struct page
*)page_private(page
);
788 next
= list_entry(page
->lru
.next
, struct page
, lru
);
794 * Encode <page, obj_idx> as a single handle value.
795 * We use the least bit of handle for tagging.
797 static void *location_to_obj(struct page
*page
, unsigned long obj_idx
)
806 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
807 obj
|= ((obj_idx
) & OBJ_INDEX_MASK
);
808 obj
<<= OBJ_TAG_BITS
;
814 * Decode <page, obj_idx> pair from the given object handle. We adjust the
815 * decoded obj_idx back to its original value since it was adjusted in
818 static void obj_to_location(unsigned long obj
, struct page
**page
,
819 unsigned long *obj_idx
)
821 obj
>>= OBJ_TAG_BITS
;
822 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
823 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
826 static unsigned long handle_to_obj(unsigned long handle
)
828 return *(unsigned long *)handle
;
831 static unsigned long obj_to_head(struct size_class
*class, struct page
*page
,
835 VM_BUG_ON(!is_first_page(page
));
836 return *(unsigned long *)page_private(page
);
838 return *(unsigned long *)obj
;
841 static unsigned long obj_idx_to_offset(struct page
*page
,
842 unsigned long obj_idx
, int class_size
)
844 unsigned long off
= 0;
846 if (!is_first_page(page
))
849 return off
+ obj_idx
* class_size
;
852 static inline int trypin_tag(unsigned long handle
)
854 unsigned long *ptr
= (unsigned long *)handle
;
856 return !test_and_set_bit_lock(HANDLE_PIN_BIT
, ptr
);
859 static void pin_tag(unsigned long handle
)
861 while (!trypin_tag(handle
));
864 static void unpin_tag(unsigned long handle
)
866 unsigned long *ptr
= (unsigned long *)handle
;
868 clear_bit_unlock(HANDLE_PIN_BIT
, ptr
);
871 static void reset_page(struct page
*page
)
873 clear_bit(PG_private
, &page
->flags
);
874 clear_bit(PG_private_2
, &page
->flags
);
875 set_page_private(page
, 0);
876 page
->mapping
= NULL
;
877 page
->freelist
= NULL
;
878 page_mapcount_reset(page
);
881 static void free_zspage(struct page
*first_page
)
883 struct page
*nextp
, *tmp
, *head_extra
;
885 BUG_ON(!is_first_page(first_page
));
886 BUG_ON(first_page
->inuse
);
888 head_extra
= (struct page
*)page_private(first_page
);
890 reset_page(first_page
);
891 __free_page(first_page
);
893 /* zspage with only 1 system page */
897 list_for_each_entry_safe(nextp
, tmp
, &head_extra
->lru
, lru
) {
898 list_del(&nextp
->lru
);
902 reset_page(head_extra
);
903 __free_page(head_extra
);
906 /* Initialize a newly allocated zspage */
907 static void init_zspage(struct page
*first_page
, struct size_class
*class)
909 unsigned long off
= 0;
910 struct page
*page
= first_page
;
912 BUG_ON(!is_first_page(first_page
));
914 struct page
*next_page
;
915 struct link_free
*link
;
920 * page->index stores offset of first object starting
921 * in the page. For the first page, this is always 0,
922 * so we use first_page->index (aka ->freelist) to store
923 * head of corresponding zspage's freelist.
925 if (page
!= first_page
)
928 vaddr
= kmap_atomic(page
);
929 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
931 while ((off
+= class->size
) < PAGE_SIZE
) {
932 link
->next
= location_to_obj(page
, i
++);
933 link
+= class->size
/ sizeof(*link
);
937 * We now come to the last (full or partial) object on this
938 * page, which must point to the first object on the next
941 next_page
= get_next_page(page
);
942 link
->next
= location_to_obj(next_page
, 0);
943 kunmap_atomic(vaddr
);
950 * Allocate a zspage for the given size class
952 static struct page
*alloc_zspage(struct size_class
*class, gfp_t flags
)
955 struct page
*first_page
= NULL
, *uninitialized_var(prev_page
);
958 * Allocate individual pages and link them together as:
959 * 1. first page->private = first sub-page
960 * 2. all sub-pages are linked together using page->lru
961 * 3. each sub-page is linked to the first page using page->first_page
963 * For each size class, First/Head pages are linked together using
964 * page->lru. Also, we set PG_private to identify the first page
965 * (i.e. no other sub-page has this flag set) and PG_private_2 to
966 * identify the last page.
969 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
972 page
= alloc_page(flags
);
976 INIT_LIST_HEAD(&page
->lru
);
977 if (i
== 0) { /* first page */
978 SetPagePrivate(page
);
979 set_page_private(page
, 0);
981 first_page
->inuse
= 0;
984 set_page_private(first_page
, (unsigned long)page
);
986 page
->first_page
= first_page
;
988 list_add(&page
->lru
, &prev_page
->lru
);
989 if (i
== class->pages_per_zspage
- 1) /* last page */
990 SetPagePrivate2(page
);
994 init_zspage(first_page
, class);
996 first_page
->freelist
= location_to_obj(first_page
, 0);
997 /* Maximum number of objects we can store in this zspage */
998 first_page
->objects
= class->pages_per_zspage
* PAGE_SIZE
/ class->size
;
1000 error
= 0; /* Success */
1003 if (unlikely(error
) && first_page
) {
1004 free_zspage(first_page
);
1011 static struct page
*find_get_zspage(struct size_class
*class)
1016 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1017 page
= class->fullness_list
[i
];
1025 #ifdef CONFIG_PGTABLE_MAPPING
1026 static inline int __zs_cpu_up(struct mapping_area
*area
)
1029 * Make sure we don't leak memory if a cpu UP notification
1030 * and zs_init() race and both call zs_cpu_up() on the same cpu
1034 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
1040 static inline void __zs_cpu_down(struct mapping_area
*area
)
1043 free_vm_area(area
->vm
);
1047 static inline void *__zs_map_object(struct mapping_area
*area
,
1048 struct page
*pages
[2], int off
, int size
)
1050 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
1051 area
->vm_addr
= area
->vm
->addr
;
1052 return area
->vm_addr
+ off
;
1055 static inline void __zs_unmap_object(struct mapping_area
*area
,
1056 struct page
*pages
[2], int off
, int size
)
1058 unsigned long addr
= (unsigned long)area
->vm_addr
;
1060 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
1063 #else /* CONFIG_PGTABLE_MAPPING */
1065 static inline int __zs_cpu_up(struct mapping_area
*area
)
1068 * Make sure we don't leak memory if a cpu UP notification
1069 * and zs_init() race and both call zs_cpu_up() on the same cpu
1073 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1079 static inline void __zs_cpu_down(struct mapping_area
*area
)
1081 kfree(area
->vm_buf
);
1082 area
->vm_buf
= NULL
;
1085 static void *__zs_map_object(struct mapping_area
*area
,
1086 struct page
*pages
[2], int off
, int size
)
1090 char *buf
= area
->vm_buf
;
1092 /* disable page faults to match kmap_atomic() return conditions */
1093 pagefault_disable();
1095 /* no read fastpath */
1096 if (area
->vm_mm
== ZS_MM_WO
)
1099 sizes
[0] = PAGE_SIZE
- off
;
1100 sizes
[1] = size
- sizes
[0];
1102 /* copy object to per-cpu buffer */
1103 addr
= kmap_atomic(pages
[0]);
1104 memcpy(buf
, addr
+ off
, sizes
[0]);
1105 kunmap_atomic(addr
);
1106 addr
= kmap_atomic(pages
[1]);
1107 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1108 kunmap_atomic(addr
);
1110 return area
->vm_buf
;
1113 static void __zs_unmap_object(struct mapping_area
*area
,
1114 struct page
*pages
[2], int off
, int size
)
1120 /* no write fastpath */
1121 if (area
->vm_mm
== ZS_MM_RO
)
1126 buf
= buf
+ ZS_HANDLE_SIZE
;
1127 size
-= ZS_HANDLE_SIZE
;
1128 off
+= ZS_HANDLE_SIZE
;
1131 sizes
[0] = PAGE_SIZE
- off
;
1132 sizes
[1] = size
- sizes
[0];
1134 /* copy per-cpu buffer to object */
1135 addr
= kmap_atomic(pages
[0]);
1136 memcpy(addr
+ off
, buf
, sizes
[0]);
1137 kunmap_atomic(addr
);
1138 addr
= kmap_atomic(pages
[1]);
1139 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1140 kunmap_atomic(addr
);
1143 /* enable page faults to match kunmap_atomic() return conditions */
1147 #endif /* CONFIG_PGTABLE_MAPPING */
1149 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
1152 int ret
, cpu
= (long)pcpu
;
1153 struct mapping_area
*area
;
1156 case CPU_UP_PREPARE
:
1157 area
= &per_cpu(zs_map_area
, cpu
);
1158 ret
= __zs_cpu_up(area
);
1160 return notifier_from_errno(ret
);
1163 case CPU_UP_CANCELED
:
1164 area
= &per_cpu(zs_map_area
, cpu
);
1165 __zs_cpu_down(area
);
1172 static struct notifier_block zs_cpu_nb
= {
1173 .notifier_call
= zs_cpu_notifier
1176 static int zs_register_cpu_notifier(void)
1178 int cpu
, uninitialized_var(ret
);
1180 cpu_notifier_register_begin();
1182 __register_cpu_notifier(&zs_cpu_nb
);
1183 for_each_online_cpu(cpu
) {
1184 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1185 if (notifier_to_errno(ret
))
1189 cpu_notifier_register_done();
1190 return notifier_to_errno(ret
);
1193 static void zs_unregister_cpu_notifier(void)
1197 cpu_notifier_register_begin();
1199 for_each_online_cpu(cpu
)
1200 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
1201 __unregister_cpu_notifier(&zs_cpu_nb
);
1203 cpu_notifier_register_done();
1206 static void init_zs_size_classes(void)
1210 nr
= (ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) / ZS_SIZE_CLASS_DELTA
+ 1;
1211 if ((ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) % ZS_SIZE_CLASS_DELTA
)
1214 zs_size_classes
= nr
;
1217 static bool can_merge(struct size_class
*prev
, int size
, int pages_per_zspage
)
1219 if (prev
->pages_per_zspage
!= pages_per_zspage
)
1222 if (get_maxobj_per_zspage(prev
->size
, prev
->pages_per_zspage
)
1223 != get_maxobj_per_zspage(size
, pages_per_zspage
))
1229 static bool zspage_full(struct page
*page
)
1231 BUG_ON(!is_first_page(page
));
1233 return page
->inuse
== page
->objects
;
1236 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1238 return atomic_long_read(&pool
->pages_allocated
);
1240 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1243 * zs_map_object - get address of allocated object from handle.
1244 * @pool: pool from which the object was allocated
1245 * @handle: handle returned from zs_malloc
1247 * Before using an object allocated from zs_malloc, it must be mapped using
1248 * this function. When done with the object, it must be unmapped using
1251 * Only one object can be mapped per cpu at a time. There is no protection
1252 * against nested mappings.
1254 * This function returns with preemption and page faults disabled.
1256 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1260 unsigned long obj
, obj_idx
, off
;
1262 unsigned int class_idx
;
1263 enum fullness_group fg
;
1264 struct size_class
*class;
1265 struct mapping_area
*area
;
1266 struct page
*pages
[2];
1272 * Because we use per-cpu mapping areas shared among the
1273 * pools/users, we can't allow mapping in interrupt context
1274 * because it can corrupt another users mappings.
1276 BUG_ON(in_interrupt());
1278 /* From now on, migration cannot move the object */
1281 obj
= handle_to_obj(handle
);
1282 obj_to_location(obj
, &page
, &obj_idx
);
1283 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1284 class = pool
->size_class
[class_idx
];
1285 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1287 area
= &get_cpu_var(zs_map_area
);
1289 if (off
+ class->size
<= PAGE_SIZE
) {
1290 /* this object is contained entirely within a page */
1291 area
->vm_addr
= kmap_atomic(page
);
1292 ret
= area
->vm_addr
+ off
;
1296 /* this object spans two pages */
1298 pages
[1] = get_next_page(page
);
1301 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1304 ret
+= ZS_HANDLE_SIZE
;
1308 EXPORT_SYMBOL_GPL(zs_map_object
);
1310 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1313 unsigned long obj
, obj_idx
, off
;
1315 unsigned int class_idx
;
1316 enum fullness_group fg
;
1317 struct size_class
*class;
1318 struct mapping_area
*area
;
1322 obj
= handle_to_obj(handle
);
1323 obj_to_location(obj
, &page
, &obj_idx
);
1324 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1325 class = pool
->size_class
[class_idx
];
1326 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1328 area
= this_cpu_ptr(&zs_map_area
);
1329 if (off
+ class->size
<= PAGE_SIZE
)
1330 kunmap_atomic(area
->vm_addr
);
1332 struct page
*pages
[2];
1335 pages
[1] = get_next_page(page
);
1338 __zs_unmap_object(area
, pages
, off
, class->size
);
1340 put_cpu_var(zs_map_area
);
1343 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1345 static unsigned long obj_malloc(struct page
*first_page
,
1346 struct size_class
*class, unsigned long handle
)
1349 struct link_free
*link
;
1351 struct page
*m_page
;
1352 unsigned long m_objidx
, m_offset
;
1355 handle
|= OBJ_ALLOCATED_TAG
;
1356 obj
= (unsigned long)first_page
->freelist
;
1357 obj_to_location(obj
, &m_page
, &m_objidx
);
1358 m_offset
= obj_idx_to_offset(m_page
, m_objidx
, class->size
);
1360 vaddr
= kmap_atomic(m_page
);
1361 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1362 first_page
->freelist
= link
->next
;
1364 /* record handle in the header of allocated chunk */
1365 link
->handle
= handle
;
1367 /* record handle in first_page->private */
1368 set_page_private(first_page
, handle
);
1369 kunmap_atomic(vaddr
);
1370 first_page
->inuse
++;
1371 zs_stat_inc(class, OBJ_USED
, 1);
1378 * zs_malloc - Allocate block of given size from pool.
1379 * @pool: pool to allocate from
1380 * @size: size of block to allocate
1382 * On success, handle to the allocated object is returned,
1384 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1386 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
)
1388 unsigned long handle
, obj
;
1389 struct size_class
*class;
1390 struct page
*first_page
;
1392 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1395 handle
= alloc_handle(pool
);
1399 /* extra space in chunk to keep the handle */
1400 size
+= ZS_HANDLE_SIZE
;
1401 class = pool
->size_class
[get_size_class_index(size
)];
1403 spin_lock(&class->lock
);
1404 first_page
= find_get_zspage(class);
1407 spin_unlock(&class->lock
);
1408 first_page
= alloc_zspage(class, pool
->flags
);
1409 if (unlikely(!first_page
)) {
1410 free_handle(pool
, handle
);
1414 set_zspage_mapping(first_page
, class->index
, ZS_EMPTY
);
1415 atomic_long_add(class->pages_per_zspage
,
1416 &pool
->pages_allocated
);
1418 spin_lock(&class->lock
);
1419 zs_stat_inc(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1420 class->size
, class->pages_per_zspage
));
1423 obj
= obj_malloc(first_page
, class, handle
);
1424 /* Now move the zspage to another fullness group, if required */
1425 fix_fullness_group(class, first_page
);
1426 record_obj(handle
, obj
);
1427 spin_unlock(&class->lock
);
1431 EXPORT_SYMBOL_GPL(zs_malloc
);
1433 static void obj_free(struct zs_pool
*pool
, struct size_class
*class,
1436 struct link_free
*link
;
1437 struct page
*first_page
, *f_page
;
1438 unsigned long f_objidx
, f_offset
;
1441 enum fullness_group fullness
;
1445 obj
&= ~OBJ_ALLOCATED_TAG
;
1446 obj_to_location(obj
, &f_page
, &f_objidx
);
1447 first_page
= get_first_page(f_page
);
1449 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1450 f_offset
= obj_idx_to_offset(f_page
, f_objidx
, class->size
);
1452 vaddr
= kmap_atomic(f_page
);
1454 /* Insert this object in containing zspage's freelist */
1455 link
= (struct link_free
*)(vaddr
+ f_offset
);
1456 link
->next
= first_page
->freelist
;
1458 set_page_private(first_page
, 0);
1459 kunmap_atomic(vaddr
);
1460 first_page
->freelist
= (void *)obj
;
1461 first_page
->inuse
--;
1462 zs_stat_dec(class, OBJ_USED
, 1);
1465 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1467 struct page
*first_page
, *f_page
;
1468 unsigned long obj
, f_objidx
;
1470 struct size_class
*class;
1471 enum fullness_group fullness
;
1473 if (unlikely(!handle
))
1477 obj
= handle_to_obj(handle
);
1478 obj_to_location(obj
, &f_page
, &f_objidx
);
1479 first_page
= get_first_page(f_page
);
1481 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1482 class = pool
->size_class
[class_idx
];
1484 spin_lock(&class->lock
);
1485 obj_free(pool
, class, obj
);
1486 fullness
= fix_fullness_group(class, first_page
);
1487 if (fullness
== ZS_EMPTY
) {
1488 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1489 class->size
, class->pages_per_zspage
));
1490 atomic_long_sub(class->pages_per_zspage
,
1491 &pool
->pages_allocated
);
1492 free_zspage(first_page
);
1494 spin_unlock(&class->lock
);
1497 free_handle(pool
, handle
);
1499 EXPORT_SYMBOL_GPL(zs_free
);
1501 static void zs_object_copy(unsigned long src
, unsigned long dst
,
1502 struct size_class
*class)
1504 struct page
*s_page
, *d_page
;
1505 unsigned long s_objidx
, d_objidx
;
1506 unsigned long s_off
, d_off
;
1507 void *s_addr
, *d_addr
;
1508 int s_size
, d_size
, size
;
1511 s_size
= d_size
= class->size
;
1513 obj_to_location(src
, &s_page
, &s_objidx
);
1514 obj_to_location(dst
, &d_page
, &d_objidx
);
1516 s_off
= obj_idx_to_offset(s_page
, s_objidx
, class->size
);
1517 d_off
= obj_idx_to_offset(d_page
, d_objidx
, class->size
);
1519 if (s_off
+ class->size
> PAGE_SIZE
)
1520 s_size
= PAGE_SIZE
- s_off
;
1522 if (d_off
+ class->size
> PAGE_SIZE
)
1523 d_size
= PAGE_SIZE
- d_off
;
1525 s_addr
= kmap_atomic(s_page
);
1526 d_addr
= kmap_atomic(d_page
);
1529 size
= min(s_size
, d_size
);
1530 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1533 if (written
== class->size
)
1541 if (s_off
>= PAGE_SIZE
) {
1542 kunmap_atomic(d_addr
);
1543 kunmap_atomic(s_addr
);
1544 s_page
= get_next_page(s_page
);
1546 s_addr
= kmap_atomic(s_page
);
1547 d_addr
= kmap_atomic(d_page
);
1548 s_size
= class->size
- written
;
1552 if (d_off
>= PAGE_SIZE
) {
1553 kunmap_atomic(d_addr
);
1554 d_page
= get_next_page(d_page
);
1556 d_addr
= kmap_atomic(d_page
);
1557 d_size
= class->size
- written
;
1562 kunmap_atomic(d_addr
);
1563 kunmap_atomic(s_addr
);
1567 * Find alloced object in zspage from index object and
1570 static unsigned long find_alloced_obj(struct page
*page
, int index
,
1571 struct size_class
*class)
1575 unsigned long handle
= 0;
1576 void *addr
= kmap_atomic(page
);
1578 if (!is_first_page(page
))
1579 offset
= page
->index
;
1580 offset
+= class->size
* index
;
1582 while (offset
< PAGE_SIZE
) {
1583 head
= obj_to_head(class, page
, addr
+ offset
);
1584 if (head
& OBJ_ALLOCATED_TAG
) {
1585 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1586 if (trypin_tag(handle
))
1591 offset
+= class->size
;
1595 kunmap_atomic(addr
);
1599 struct zs_compact_control
{
1600 /* Source page for migration which could be a subpage of zspage. */
1601 struct page
*s_page
;
1602 /* Destination page for migration which should be a first page
1604 struct page
*d_page
;
1605 /* Starting object index within @s_page which used for live object
1606 * in the subpage. */
1608 /* how many of objects are migrated */
1612 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1613 struct zs_compact_control
*cc
)
1615 unsigned long used_obj
, free_obj
;
1616 unsigned long handle
;
1617 struct page
*s_page
= cc
->s_page
;
1618 struct page
*d_page
= cc
->d_page
;
1619 unsigned long index
= cc
->index
;
1620 int nr_migrated
= 0;
1624 handle
= find_alloced_obj(s_page
, index
, class);
1626 s_page
= get_next_page(s_page
);
1633 /* Stop if there is no more space */
1634 if (zspage_full(d_page
)) {
1640 used_obj
= handle_to_obj(handle
);
1641 free_obj
= obj_malloc(d_page
, class, handle
);
1642 zs_object_copy(used_obj
, free_obj
, class);
1644 record_obj(handle
, free_obj
);
1646 obj_free(pool
, class, used_obj
);
1650 /* Remember last position in this iteration */
1651 cc
->s_page
= s_page
;
1653 cc
->nr_migrated
= nr_migrated
;
1658 static struct page
*alloc_target_page(struct size_class
*class)
1663 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1664 page
= class->fullness_list
[i
];
1666 remove_zspage(page
, class, i
);
1674 static void putback_zspage(struct zs_pool
*pool
, struct size_class
*class,
1675 struct page
*first_page
)
1677 enum fullness_group fullness
;
1679 BUG_ON(!is_first_page(first_page
));
1681 fullness
= get_fullness_group(first_page
);
1682 insert_zspage(first_page
, class, fullness
);
1683 set_zspage_mapping(first_page
, class->index
, fullness
);
1685 if (fullness
== ZS_EMPTY
) {
1686 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1687 class->size
, class->pages_per_zspage
));
1688 atomic_long_sub(class->pages_per_zspage
,
1689 &pool
->pages_allocated
);
1691 free_zspage(first_page
);
1695 static struct page
*isolate_source_page(struct size_class
*class)
1699 page
= class->fullness_list
[ZS_ALMOST_EMPTY
];
1701 remove_zspage(page
, class, ZS_ALMOST_EMPTY
);
1706 static unsigned long __zs_compact(struct zs_pool
*pool
,
1707 struct size_class
*class)
1710 struct zs_compact_control cc
;
1711 struct page
*src_page
;
1712 struct page
*dst_page
= NULL
;
1713 unsigned long nr_total_migrated
= 0;
1715 spin_lock(&class->lock
);
1716 while ((src_page
= isolate_source_page(class))) {
1718 BUG_ON(!is_first_page(src_page
));
1720 /* The goal is to migrate all live objects in source page */
1721 nr_to_migrate
= src_page
->inuse
;
1723 cc
.s_page
= src_page
;
1725 while ((dst_page
= alloc_target_page(class))) {
1726 cc
.d_page
= dst_page
;
1728 * If there is no more space in dst_page, try to
1729 * allocate another zspage.
1731 if (!migrate_zspage(pool
, class, &cc
))
1734 putback_zspage(pool
, class, dst_page
);
1735 nr_total_migrated
+= cc
.nr_migrated
;
1736 nr_to_migrate
-= cc
.nr_migrated
;
1739 /* Stop if we couldn't find slot */
1740 if (dst_page
== NULL
)
1743 putback_zspage(pool
, class, dst_page
);
1744 putback_zspage(pool
, class, src_page
);
1745 spin_unlock(&class->lock
);
1746 nr_total_migrated
+= cc
.nr_migrated
;
1748 spin_lock(&class->lock
);
1752 putback_zspage(pool
, class, src_page
);
1754 spin_unlock(&class->lock
);
1756 return nr_total_migrated
;
1759 unsigned long zs_compact(struct zs_pool
*pool
)
1762 unsigned long nr_migrated
= 0;
1763 struct size_class
*class;
1765 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1766 class = pool
->size_class
[i
];
1769 if (class->index
!= i
)
1771 nr_migrated
+= __zs_compact(pool
, class);
1776 EXPORT_SYMBOL_GPL(zs_compact
);
1779 * zs_create_pool - Creates an allocation pool to work from.
1780 * @flags: allocation flags used to allocate pool metadata
1782 * This function must be called before anything when using
1783 * the zsmalloc allocator.
1785 * On success, a pointer to the newly created pool is returned,
1788 struct zs_pool
*zs_create_pool(char *name
, gfp_t flags
)
1791 struct zs_pool
*pool
;
1792 struct size_class
*prev_class
= NULL
;
1794 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
1798 pool
->size_class
= kcalloc(zs_size_classes
, sizeof(struct size_class
*),
1800 if (!pool
->size_class
) {
1805 pool
->name
= kstrdup(name
, GFP_KERNEL
);
1809 if (create_handle_cache(pool
))
1813 * Iterate reversly, because, size of size_class that we want to use
1814 * for merging should be larger or equal to current size.
1816 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1818 int pages_per_zspage
;
1819 struct size_class
*class;
1821 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
1822 if (size
> ZS_MAX_ALLOC_SIZE
)
1823 size
= ZS_MAX_ALLOC_SIZE
;
1824 pages_per_zspage
= get_pages_per_zspage(size
);
1827 * size_class is used for normal zsmalloc operation such
1828 * as alloc/free for that size. Although it is natural that we
1829 * have one size_class for each size, there is a chance that we
1830 * can get more memory utilization if we use one size_class for
1831 * many different sizes whose size_class have same
1832 * characteristics. So, we makes size_class point to
1833 * previous size_class if possible.
1836 if (can_merge(prev_class
, size
, pages_per_zspage
)) {
1837 pool
->size_class
[i
] = prev_class
;
1842 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
1848 class->pages_per_zspage
= pages_per_zspage
;
1849 if (pages_per_zspage
== 1 &&
1850 get_maxobj_per_zspage(size
, pages_per_zspage
) == 1)
1852 spin_lock_init(&class->lock
);
1853 pool
->size_class
[i
] = class;
1858 pool
->flags
= flags
;
1860 if (zs_pool_stat_create(name
, pool
))
1866 zs_destroy_pool(pool
);
1869 EXPORT_SYMBOL_GPL(zs_create_pool
);
1871 void zs_destroy_pool(struct zs_pool
*pool
)
1875 zs_pool_stat_destroy(pool
);
1877 for (i
= 0; i
< zs_size_classes
; i
++) {
1879 struct size_class
*class = pool
->size_class
[i
];
1884 if (class->index
!= i
)
1887 for (fg
= 0; fg
< _ZS_NR_FULLNESS_GROUPS
; fg
++) {
1888 if (class->fullness_list
[fg
]) {
1889 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1896 destroy_handle_cache(pool
);
1897 kfree(pool
->size_class
);
1901 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
1903 static int __init
zs_init(void)
1905 int ret
= zs_register_cpu_notifier();
1910 init_zs_size_classes();
1913 zpool_register_driver(&zs_zpool_driver
);
1916 ret
= zs_stat_init();
1918 pr_err("zs stat initialization failed\n");
1925 zpool_unregister_driver(&zs_zpool_driver
);
1928 zs_unregister_cpu_notifier();
1933 static void __exit
zs_exit(void)
1936 zpool_unregister_driver(&zs_zpool_driver
);
1938 zs_unregister_cpu_notifier();
1943 module_init(zs_init
);
1944 module_exit(zs_exit
);
1946 MODULE_LICENSE("Dual BSD/GPL");
1947 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");