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 #include <linux/module.h>
49 #include <linux/kernel.h>
50 #include <linux/sched.h>
51 #include <linux/bitops.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/string.h>
55 #include <linux/slab.h>
56 #include <asm/tlbflush.h>
57 #include <asm/pgtable.h>
58 #include <linux/cpumask.h>
59 #include <linux/cpu.h>
60 #include <linux/vmalloc.h>
61 #include <linux/hardirq.h>
62 #include <linux/spinlock.h>
63 #include <linux/types.h>
64 #include <linux/debugfs.h>
65 #include <linux/zsmalloc.h>
66 #include <linux/zpool.h>
69 * This must be power of 2 and greater than of equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
80 #define ZS_MAX_ZSPAGE_ORDER 2
81 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
83 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
86 * Object location (<PFN>, <obj_idx>) is encoded as
87 * as single (unsigned long) handle value.
89 * Note that object index <obj_idx> is relative to system
90 * page <PFN> it is stored in, so for each sub-page belonging
91 * to a zspage, obj_idx starts with 0.
93 * This is made more complicated by various memory models and PAE.
96 #ifndef MAX_PHYSMEM_BITS
97 #ifdef CONFIG_HIGHMEM64G
98 #define MAX_PHYSMEM_BITS 36
99 #else /* !CONFIG_HIGHMEM64G */
101 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
104 #define MAX_PHYSMEM_BITS BITS_PER_LONG
107 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
110 * Memory for allocating for handle keeps object position by
111 * encoding <page, obj_idx> and the encoded value has a room
112 * in least bit(ie, look at obj_to_location).
113 * We use the bit to synchronize between object access by
114 * user and migration.
116 #define HANDLE_PIN_BIT 0
119 * Head in allocated object should have OBJ_ALLOCATED_TAG
120 * to identify the object was allocated or not.
121 * It's okay to add the status bit in the least bit because
122 * header keeps handle which is 4byte-aligned address so we
123 * have room for two bit at least.
125 #define OBJ_ALLOCATED_TAG 1
126 #define OBJ_TAG_BITS 1
127 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
128 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
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 >> 8)
153 * We do not maintain any list for completely empty or full pages
155 enum fullness_group
{
158 _ZS_NR_FULLNESS_GROUPS
,
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
;
181 * number of size_classes
183 static int zs_size_classes
;
186 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
188 * n = number of allocated objects
189 * N = total number of objects zspage can store
190 * f = fullness_threshold_frac
192 * Similarly, we assign zspage to:
193 * ZS_ALMOST_FULL when n > N / f
194 * ZS_EMPTY when n == 0
195 * ZS_FULL when n == N
197 * (see: fix_fullness_group())
199 static const int fullness_threshold_frac
= 4;
203 struct page
*fullness_list
[_ZS_NR_FULLNESS_GROUPS
];
205 * Size of objects stored in this class. Must be multiple
211 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
212 int pages_per_zspage
;
213 struct zs_size_stat stats
;
215 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
220 * Placed within free objects to form a singly linked list.
221 * For every zspage, first_page->freelist gives head of this list.
223 * This must be power of 2 and less than or equal to ZS_ALIGN
228 * Position of next free chunk (encodes <PFN, obj_idx>)
229 * It's valid for non-allocated object
233 * Handle of allocated object.
235 unsigned long handle
;
242 struct size_class
**size_class
;
243 struct kmem_cache
*handle_cachep
;
245 gfp_t flags
; /* allocation flags used when growing pool */
246 atomic_long_t pages_allocated
;
248 struct zs_pool_stats stats
;
250 /* Compact classes */
251 struct shrinker shrinker
;
253 * To signify that register_shrinker() was successful
254 * and unregister_shrinker() will not Oops.
256 bool shrinker_enabled
;
257 #ifdef CONFIG_ZSMALLOC_STAT
258 struct dentry
*stat_dentry
;
263 * A zspage's class index and fullness group
264 * are encoded in its (first)page->mapping
266 #define CLASS_IDX_BITS 28
267 #define FULLNESS_BITS 4
268 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
269 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
271 struct mapping_area
{
272 #ifdef CONFIG_PGTABLE_MAPPING
273 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
275 char *vm_buf
; /* copy buffer for objects that span pages */
277 char *vm_addr
; /* address of kmap_atomic()'ed pages */
278 enum zs_mapmode vm_mm
; /* mapping mode */
282 static int create_handle_cache(struct zs_pool
*pool
)
284 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
286 return pool
->handle_cachep
? 0 : 1;
289 static void destroy_handle_cache(struct zs_pool
*pool
)
291 kmem_cache_destroy(pool
->handle_cachep
);
294 static unsigned long alloc_handle(struct zs_pool
*pool
)
296 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
297 pool
->flags
& ~__GFP_HIGHMEM
);
300 static void free_handle(struct zs_pool
*pool
, unsigned long handle
)
302 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
305 static void record_obj(unsigned long handle
, unsigned long obj
)
307 *(unsigned long *)handle
= obj
;
314 static void *zs_zpool_create(char *name
, gfp_t gfp
, struct zpool_ops
*zpool_ops
,
317 return zs_create_pool(name
, gfp
);
320 static void zs_zpool_destroy(void *pool
)
322 zs_destroy_pool(pool
);
325 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
326 unsigned long *handle
)
328 *handle
= zs_malloc(pool
, size
);
329 return *handle
? 0 : -1;
331 static void zs_zpool_free(void *pool
, unsigned long handle
)
333 zs_free(pool
, handle
);
336 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
337 unsigned int *reclaimed
)
342 static void *zs_zpool_map(void *pool
, unsigned long handle
,
343 enum zpool_mapmode mm
)
345 enum zs_mapmode zs_mm
;
354 case ZPOOL_MM_RW
: /* fallthru */
360 return zs_map_object(pool
, handle
, zs_mm
);
362 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
364 zs_unmap_object(pool
, handle
);
367 static u64
zs_zpool_total_size(void *pool
)
369 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
372 static struct zpool_driver zs_zpool_driver
= {
374 .owner
= THIS_MODULE
,
375 .create
= zs_zpool_create
,
376 .destroy
= zs_zpool_destroy
,
377 .malloc
= zs_zpool_malloc
,
378 .free
= zs_zpool_free
,
379 .shrink
= zs_zpool_shrink
,
381 .unmap
= zs_zpool_unmap
,
382 .total_size
= zs_zpool_total_size
,
385 MODULE_ALIAS("zpool-zsmalloc");
386 #endif /* CONFIG_ZPOOL */
388 static unsigned int get_maxobj_per_zspage(int size
, int pages_per_zspage
)
390 return pages_per_zspage
* PAGE_SIZE
/ size
;
393 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
394 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
396 static int is_first_page(struct page
*page
)
398 return PagePrivate(page
);
401 static int is_last_page(struct page
*page
)
403 return PagePrivate2(page
);
406 static void get_zspage_mapping(struct page
*page
, unsigned int *class_idx
,
407 enum fullness_group
*fullness
)
410 BUG_ON(!is_first_page(page
));
412 m
= (unsigned long)page
->mapping
;
413 *fullness
= m
& FULLNESS_MASK
;
414 *class_idx
= (m
>> FULLNESS_BITS
) & CLASS_IDX_MASK
;
417 static void set_zspage_mapping(struct page
*page
, unsigned int class_idx
,
418 enum fullness_group fullness
)
421 BUG_ON(!is_first_page(page
));
423 m
= ((class_idx
& CLASS_IDX_MASK
) << FULLNESS_BITS
) |
424 (fullness
& FULLNESS_MASK
);
425 page
->mapping
= (struct address_space
*)m
;
429 * zsmalloc divides the pool into various size classes where each
430 * class maintains a list of zspages where each zspage is divided
431 * into equal sized chunks. Each allocation falls into one of these
432 * classes depending on its size. This function returns index of the
433 * size class which has chunk size big enough to hold the give size.
435 static int get_size_class_index(int size
)
439 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
440 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
441 ZS_SIZE_CLASS_DELTA
);
443 return min(zs_size_classes
- 1, idx
);
446 static inline void zs_stat_inc(struct size_class
*class,
447 enum zs_stat_type type
, unsigned long cnt
)
449 class->stats
.objs
[type
] += cnt
;
452 static inline void zs_stat_dec(struct size_class
*class,
453 enum zs_stat_type type
, unsigned long cnt
)
455 class->stats
.objs
[type
] -= cnt
;
458 static inline unsigned long zs_stat_get(struct size_class
*class,
459 enum zs_stat_type type
)
461 return class->stats
.objs
[type
];
464 #ifdef CONFIG_ZSMALLOC_STAT
466 static int __init
zs_stat_init(void)
468 if (!debugfs_initialized())
471 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
478 static void __exit
zs_stat_exit(void)
480 debugfs_remove_recursive(zs_stat_root
);
483 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
486 struct zs_pool
*pool
= s
->private;
487 struct size_class
*class;
489 unsigned long class_almost_full
, class_almost_empty
;
490 unsigned long obj_allocated
, obj_used
, pages_used
;
491 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
492 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
494 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
495 "class", "size", "almost_full", "almost_empty",
496 "obj_allocated", "obj_used", "pages_used",
499 for (i
= 0; i
< zs_size_classes
; i
++) {
500 class = pool
->size_class
[i
];
502 if (class->index
!= i
)
505 spin_lock(&class->lock
);
506 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
507 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
508 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
509 obj_used
= zs_stat_get(class, OBJ_USED
);
510 spin_unlock(&class->lock
);
512 objs_per_zspage
= get_maxobj_per_zspage(class->size
,
513 class->pages_per_zspage
);
514 pages_used
= obj_allocated
/ objs_per_zspage
*
515 class->pages_per_zspage
;
517 seq_printf(s
, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
518 i
, class->size
, class_almost_full
, class_almost_empty
,
519 obj_allocated
, obj_used
, pages_used
,
520 class->pages_per_zspage
);
522 total_class_almost_full
+= class_almost_full
;
523 total_class_almost_empty
+= class_almost_empty
;
524 total_objs
+= obj_allocated
;
525 total_used_objs
+= obj_used
;
526 total_pages
+= pages_used
;
530 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
531 "Total", "", total_class_almost_full
,
532 total_class_almost_empty
, total_objs
,
533 total_used_objs
, total_pages
);
538 static int zs_stats_size_open(struct inode
*inode
, struct file
*file
)
540 return single_open(file
, zs_stats_size_show
, inode
->i_private
);
543 static const struct file_operations zs_stat_size_ops
= {
544 .open
= zs_stats_size_open
,
547 .release
= single_release
,
550 static int zs_pool_stat_create(char *name
, struct zs_pool
*pool
)
552 struct dentry
*entry
;
557 entry
= debugfs_create_dir(name
, zs_stat_root
);
559 pr_warn("debugfs dir <%s> creation failed\n", name
);
562 pool
->stat_dentry
= entry
;
564 entry
= debugfs_create_file("classes", S_IFREG
| S_IRUGO
,
565 pool
->stat_dentry
, pool
, &zs_stat_size_ops
);
567 pr_warn("%s: debugfs file entry <%s> creation failed\n",
575 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
577 debugfs_remove_recursive(pool
->stat_dentry
);
580 #else /* CONFIG_ZSMALLOC_STAT */
581 static int __init
zs_stat_init(void)
586 static void __exit
zs_stat_exit(void)
590 static inline int zs_pool_stat_create(char *name
, struct zs_pool
*pool
)
595 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
602 * For each size class, zspages are divided into different groups
603 * depending on how "full" they are. This was done so that we could
604 * easily find empty or nearly empty zspages when we try to shrink
605 * the pool (not yet implemented). This function returns fullness
606 * status of the given page.
608 static enum fullness_group
get_fullness_group(struct page
*page
)
610 int inuse
, max_objects
;
611 enum fullness_group fg
;
612 BUG_ON(!is_first_page(page
));
615 max_objects
= page
->objects
;
619 else if (inuse
== max_objects
)
621 else if (inuse
<= 3 * max_objects
/ fullness_threshold_frac
)
622 fg
= ZS_ALMOST_EMPTY
;
630 * Each size class maintains various freelists and zspages are assigned
631 * to one of these freelists based on the number of live objects they
632 * have. This functions inserts the given zspage into the freelist
633 * identified by <class, fullness_group>.
635 static void insert_zspage(struct page
*page
, struct size_class
*class,
636 enum fullness_group fullness
)
640 BUG_ON(!is_first_page(page
));
642 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
645 zs_stat_inc(class, fullness
== ZS_ALMOST_EMPTY
?
646 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
648 head
= &class->fullness_list
[fullness
];
655 * We want to see more ZS_FULL pages and less almost
656 * empty/full. Put pages with higher ->inuse first.
658 list_add_tail(&page
->lru
, &(*head
)->lru
);
659 if (page
->inuse
>= (*head
)->inuse
)
664 * This function removes the given zspage from the freelist identified
665 * by <class, fullness_group>.
667 static void remove_zspage(struct page
*page
, struct size_class
*class,
668 enum fullness_group fullness
)
672 BUG_ON(!is_first_page(page
));
674 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
677 head
= &class->fullness_list
[fullness
];
679 if (list_empty(&(*head
)->lru
))
681 else if (*head
== page
)
682 *head
= (struct page
*)list_entry((*head
)->lru
.next
,
685 list_del_init(&page
->lru
);
686 zs_stat_dec(class, fullness
== ZS_ALMOST_EMPTY
?
687 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
691 * Each size class maintains zspages in different fullness groups depending
692 * on the number of live objects they contain. When allocating or freeing
693 * objects, the fullness status of the page can change, say, from ALMOST_FULL
694 * to ALMOST_EMPTY when freeing an object. This function checks if such
695 * a status change has occurred for the given page and accordingly moves the
696 * page from the freelist of the old fullness group to that of the new
699 static enum fullness_group
fix_fullness_group(struct size_class
*class,
703 enum fullness_group currfg
, newfg
;
705 BUG_ON(!is_first_page(page
));
707 get_zspage_mapping(page
, &class_idx
, &currfg
);
708 newfg
= get_fullness_group(page
);
712 remove_zspage(page
, class, currfg
);
713 insert_zspage(page
, class, newfg
);
714 set_zspage_mapping(page
, class_idx
, newfg
);
721 * We have to decide on how many pages to link together
722 * to form a zspage for each size class. This is important
723 * to reduce wastage due to unusable space left at end of
724 * each zspage which is given as:
725 * wastage = Zp % class_size
726 * usage = Zp - wastage
727 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
729 * For example, for size class of 3/8 * PAGE_SIZE, we should
730 * link together 3 PAGE_SIZE sized pages to form a zspage
731 * since then we can perfectly fit in 8 such objects.
733 static int get_pages_per_zspage(int class_size
)
735 int i
, max_usedpc
= 0;
736 /* zspage order which gives maximum used size per KB */
737 int max_usedpc_order
= 1;
739 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
743 zspage_size
= i
* PAGE_SIZE
;
744 waste
= zspage_size
% class_size
;
745 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
747 if (usedpc
> max_usedpc
) {
749 max_usedpc_order
= i
;
753 return max_usedpc_order
;
757 * A single 'zspage' is composed of many system pages which are
758 * linked together using fields in struct page. This function finds
759 * the first/head page, given any component page of a zspage.
761 static struct page
*get_first_page(struct page
*page
)
763 if (is_first_page(page
))
766 return page
->first_page
;
769 static struct page
*get_next_page(struct page
*page
)
773 if (is_last_page(page
))
775 else if (is_first_page(page
))
776 next
= (struct page
*)page_private(page
);
778 next
= list_entry(page
->lru
.next
, struct page
, lru
);
784 * Encode <page, obj_idx> as a single handle value.
785 * We use the least bit of handle for tagging.
787 static void *location_to_obj(struct page
*page
, unsigned long obj_idx
)
796 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
797 obj
|= ((obj_idx
) & OBJ_INDEX_MASK
);
798 obj
<<= OBJ_TAG_BITS
;
804 * Decode <page, obj_idx> pair from the given object handle. We adjust the
805 * decoded obj_idx back to its original value since it was adjusted in
808 static void obj_to_location(unsigned long obj
, struct page
**page
,
809 unsigned long *obj_idx
)
811 obj
>>= OBJ_TAG_BITS
;
812 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
813 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
816 static unsigned long handle_to_obj(unsigned long handle
)
818 return *(unsigned long *)handle
;
821 static unsigned long obj_to_head(struct size_class
*class, struct page
*page
,
825 VM_BUG_ON(!is_first_page(page
));
826 return *(unsigned long *)page_private(page
);
828 return *(unsigned long *)obj
;
831 static unsigned long obj_idx_to_offset(struct page
*page
,
832 unsigned long obj_idx
, int class_size
)
834 unsigned long off
= 0;
836 if (!is_first_page(page
))
839 return off
+ obj_idx
* class_size
;
842 static inline int trypin_tag(unsigned long handle
)
844 unsigned long *ptr
= (unsigned long *)handle
;
846 return !test_and_set_bit_lock(HANDLE_PIN_BIT
, ptr
);
849 static void pin_tag(unsigned long handle
)
851 while (!trypin_tag(handle
));
854 static void unpin_tag(unsigned long handle
)
856 unsigned long *ptr
= (unsigned long *)handle
;
858 clear_bit_unlock(HANDLE_PIN_BIT
, ptr
);
861 static void reset_page(struct page
*page
)
863 clear_bit(PG_private
, &page
->flags
);
864 clear_bit(PG_private_2
, &page
->flags
);
865 set_page_private(page
, 0);
866 page
->mapping
= NULL
;
867 page
->freelist
= NULL
;
868 page_mapcount_reset(page
);
871 static void free_zspage(struct page
*first_page
)
873 struct page
*nextp
, *tmp
, *head_extra
;
875 BUG_ON(!is_first_page(first_page
));
876 BUG_ON(first_page
->inuse
);
878 head_extra
= (struct page
*)page_private(first_page
);
880 reset_page(first_page
);
881 __free_page(first_page
);
883 /* zspage with only 1 system page */
887 list_for_each_entry_safe(nextp
, tmp
, &head_extra
->lru
, lru
) {
888 list_del(&nextp
->lru
);
892 reset_page(head_extra
);
893 __free_page(head_extra
);
896 /* Initialize a newly allocated zspage */
897 static void init_zspage(struct page
*first_page
, struct size_class
*class)
899 unsigned long off
= 0;
900 struct page
*page
= first_page
;
902 BUG_ON(!is_first_page(first_page
));
904 struct page
*next_page
;
905 struct link_free
*link
;
910 * page->index stores offset of first object starting
911 * in the page. For the first page, this is always 0,
912 * so we use first_page->index (aka ->freelist) to store
913 * head of corresponding zspage's freelist.
915 if (page
!= first_page
)
918 vaddr
= kmap_atomic(page
);
919 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
921 while ((off
+= class->size
) < PAGE_SIZE
) {
922 link
->next
= location_to_obj(page
, i
++);
923 link
+= class->size
/ sizeof(*link
);
927 * We now come to the last (full or partial) object on this
928 * page, which must point to the first object on the next
931 next_page
= get_next_page(page
);
932 link
->next
= location_to_obj(next_page
, 0);
933 kunmap_atomic(vaddr
);
940 * Allocate a zspage for the given size class
942 static struct page
*alloc_zspage(struct size_class
*class, gfp_t flags
)
945 struct page
*first_page
= NULL
, *uninitialized_var(prev_page
);
948 * Allocate individual pages and link them together as:
949 * 1. first page->private = first sub-page
950 * 2. all sub-pages are linked together using page->lru
951 * 3. each sub-page is linked to the first page using page->first_page
953 * For each size class, First/Head pages are linked together using
954 * page->lru. Also, we set PG_private to identify the first page
955 * (i.e. no other sub-page has this flag set) and PG_private_2 to
956 * identify the last page.
959 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
962 page
= alloc_page(flags
);
966 INIT_LIST_HEAD(&page
->lru
);
967 if (i
== 0) { /* first page */
968 SetPagePrivate(page
);
969 set_page_private(page
, 0);
971 first_page
->inuse
= 0;
974 set_page_private(first_page
, (unsigned long)page
);
976 page
->first_page
= first_page
;
978 list_add(&page
->lru
, &prev_page
->lru
);
979 if (i
== class->pages_per_zspage
- 1) /* last page */
980 SetPagePrivate2(page
);
984 init_zspage(first_page
, class);
986 first_page
->freelist
= location_to_obj(first_page
, 0);
987 /* Maximum number of objects we can store in this zspage */
988 first_page
->objects
= class->pages_per_zspage
* PAGE_SIZE
/ class->size
;
990 error
= 0; /* Success */
993 if (unlikely(error
) && first_page
) {
994 free_zspage(first_page
);
1001 static struct page
*find_get_zspage(struct size_class
*class)
1006 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1007 page
= class->fullness_list
[i
];
1015 #ifdef CONFIG_PGTABLE_MAPPING
1016 static inline int __zs_cpu_up(struct mapping_area
*area
)
1019 * Make sure we don't leak memory if a cpu UP notification
1020 * and zs_init() race and both call zs_cpu_up() on the same cpu
1024 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
1030 static inline void __zs_cpu_down(struct mapping_area
*area
)
1033 free_vm_area(area
->vm
);
1037 static inline void *__zs_map_object(struct mapping_area
*area
,
1038 struct page
*pages
[2], int off
, int size
)
1040 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
1041 area
->vm_addr
= area
->vm
->addr
;
1042 return area
->vm_addr
+ off
;
1045 static inline void __zs_unmap_object(struct mapping_area
*area
,
1046 struct page
*pages
[2], int off
, int size
)
1048 unsigned long addr
= (unsigned long)area
->vm_addr
;
1050 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
1053 #else /* CONFIG_PGTABLE_MAPPING */
1055 static inline int __zs_cpu_up(struct mapping_area
*area
)
1058 * Make sure we don't leak memory if a cpu UP notification
1059 * and zs_init() race and both call zs_cpu_up() on the same cpu
1063 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1069 static inline void __zs_cpu_down(struct mapping_area
*area
)
1071 kfree(area
->vm_buf
);
1072 area
->vm_buf
= NULL
;
1075 static void *__zs_map_object(struct mapping_area
*area
,
1076 struct page
*pages
[2], int off
, int size
)
1080 char *buf
= area
->vm_buf
;
1082 /* disable page faults to match kmap_atomic() return conditions */
1083 pagefault_disable();
1085 /* no read fastpath */
1086 if (area
->vm_mm
== ZS_MM_WO
)
1089 sizes
[0] = PAGE_SIZE
- off
;
1090 sizes
[1] = size
- sizes
[0];
1092 /* copy object to per-cpu buffer */
1093 addr
= kmap_atomic(pages
[0]);
1094 memcpy(buf
, addr
+ off
, sizes
[0]);
1095 kunmap_atomic(addr
);
1096 addr
= kmap_atomic(pages
[1]);
1097 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1098 kunmap_atomic(addr
);
1100 return area
->vm_buf
;
1103 static void __zs_unmap_object(struct mapping_area
*area
,
1104 struct page
*pages
[2], int off
, int size
)
1110 /* no write fastpath */
1111 if (area
->vm_mm
== ZS_MM_RO
)
1116 buf
= buf
+ ZS_HANDLE_SIZE
;
1117 size
-= ZS_HANDLE_SIZE
;
1118 off
+= ZS_HANDLE_SIZE
;
1121 sizes
[0] = PAGE_SIZE
- off
;
1122 sizes
[1] = size
- sizes
[0];
1124 /* copy per-cpu buffer to object */
1125 addr
= kmap_atomic(pages
[0]);
1126 memcpy(addr
+ off
, buf
, sizes
[0]);
1127 kunmap_atomic(addr
);
1128 addr
= kmap_atomic(pages
[1]);
1129 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1130 kunmap_atomic(addr
);
1133 /* enable page faults to match kunmap_atomic() return conditions */
1137 #endif /* CONFIG_PGTABLE_MAPPING */
1139 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
1142 int ret
, cpu
= (long)pcpu
;
1143 struct mapping_area
*area
;
1146 case CPU_UP_PREPARE
:
1147 area
= &per_cpu(zs_map_area
, cpu
);
1148 ret
= __zs_cpu_up(area
);
1150 return notifier_from_errno(ret
);
1153 case CPU_UP_CANCELED
:
1154 area
= &per_cpu(zs_map_area
, cpu
);
1155 __zs_cpu_down(area
);
1162 static struct notifier_block zs_cpu_nb
= {
1163 .notifier_call
= zs_cpu_notifier
1166 static int zs_register_cpu_notifier(void)
1168 int cpu
, uninitialized_var(ret
);
1170 cpu_notifier_register_begin();
1172 __register_cpu_notifier(&zs_cpu_nb
);
1173 for_each_online_cpu(cpu
) {
1174 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1175 if (notifier_to_errno(ret
))
1179 cpu_notifier_register_done();
1180 return notifier_to_errno(ret
);
1183 static void zs_unregister_cpu_notifier(void)
1187 cpu_notifier_register_begin();
1189 for_each_online_cpu(cpu
)
1190 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
1191 __unregister_cpu_notifier(&zs_cpu_nb
);
1193 cpu_notifier_register_done();
1196 static void init_zs_size_classes(void)
1200 nr
= (ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) / ZS_SIZE_CLASS_DELTA
+ 1;
1201 if ((ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) % ZS_SIZE_CLASS_DELTA
)
1204 zs_size_classes
= nr
;
1207 static bool can_merge(struct size_class
*prev
, int size
, int pages_per_zspage
)
1209 if (prev
->pages_per_zspage
!= pages_per_zspage
)
1212 if (get_maxobj_per_zspage(prev
->size
, prev
->pages_per_zspage
)
1213 != get_maxobj_per_zspage(size
, pages_per_zspage
))
1219 static bool zspage_full(struct page
*page
)
1221 BUG_ON(!is_first_page(page
));
1223 return page
->inuse
== page
->objects
;
1226 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1228 return atomic_long_read(&pool
->pages_allocated
);
1230 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1233 * zs_map_object - get address of allocated object from handle.
1234 * @pool: pool from which the object was allocated
1235 * @handle: handle returned from zs_malloc
1237 * Before using an object allocated from zs_malloc, it must be mapped using
1238 * this function. When done with the object, it must be unmapped using
1241 * Only one object can be mapped per cpu at a time. There is no protection
1242 * against nested mappings.
1244 * This function returns with preemption and page faults disabled.
1246 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1250 unsigned long obj
, obj_idx
, off
;
1252 unsigned int class_idx
;
1253 enum fullness_group fg
;
1254 struct size_class
*class;
1255 struct mapping_area
*area
;
1256 struct page
*pages
[2];
1262 * Because we use per-cpu mapping areas shared among the
1263 * pools/users, we can't allow mapping in interrupt context
1264 * because it can corrupt another users mappings.
1266 BUG_ON(in_interrupt());
1268 /* From now on, migration cannot move the object */
1271 obj
= handle_to_obj(handle
);
1272 obj_to_location(obj
, &page
, &obj_idx
);
1273 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1274 class = pool
->size_class
[class_idx
];
1275 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1277 area
= &get_cpu_var(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
);
1294 ret
+= ZS_HANDLE_SIZE
;
1298 EXPORT_SYMBOL_GPL(zs_map_object
);
1300 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1303 unsigned long obj
, obj_idx
, off
;
1305 unsigned int class_idx
;
1306 enum fullness_group fg
;
1307 struct size_class
*class;
1308 struct mapping_area
*area
;
1312 obj
= handle_to_obj(handle
);
1313 obj_to_location(obj
, &page
, &obj_idx
);
1314 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1315 class = pool
->size_class
[class_idx
];
1316 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1318 area
= this_cpu_ptr(&zs_map_area
);
1319 if (off
+ class->size
<= PAGE_SIZE
)
1320 kunmap_atomic(area
->vm_addr
);
1322 struct page
*pages
[2];
1325 pages
[1] = get_next_page(page
);
1328 __zs_unmap_object(area
, pages
, off
, class->size
);
1330 put_cpu_var(zs_map_area
);
1333 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1335 static unsigned long obj_malloc(struct page
*first_page
,
1336 struct size_class
*class, unsigned long handle
)
1339 struct link_free
*link
;
1341 struct page
*m_page
;
1342 unsigned long m_objidx
, m_offset
;
1345 handle
|= OBJ_ALLOCATED_TAG
;
1346 obj
= (unsigned long)first_page
->freelist
;
1347 obj_to_location(obj
, &m_page
, &m_objidx
);
1348 m_offset
= obj_idx_to_offset(m_page
, m_objidx
, class->size
);
1350 vaddr
= kmap_atomic(m_page
);
1351 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1352 first_page
->freelist
= link
->next
;
1354 /* record handle in the header of allocated chunk */
1355 link
->handle
= handle
;
1357 /* record handle in first_page->private */
1358 set_page_private(first_page
, handle
);
1359 kunmap_atomic(vaddr
);
1360 first_page
->inuse
++;
1361 zs_stat_inc(class, OBJ_USED
, 1);
1368 * zs_malloc - Allocate block of given size from pool.
1369 * @pool: pool to allocate from
1370 * @size: size of block to allocate
1372 * On success, handle to the allocated object is returned,
1374 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1376 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
)
1378 unsigned long handle
, obj
;
1379 struct size_class
*class;
1380 struct page
*first_page
;
1382 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1385 handle
= alloc_handle(pool
);
1389 /* extra space in chunk to keep the handle */
1390 size
+= ZS_HANDLE_SIZE
;
1391 class = pool
->size_class
[get_size_class_index(size
)];
1393 spin_lock(&class->lock
);
1394 first_page
= find_get_zspage(class);
1397 spin_unlock(&class->lock
);
1398 first_page
= alloc_zspage(class, pool
->flags
);
1399 if (unlikely(!first_page
)) {
1400 free_handle(pool
, handle
);
1404 set_zspage_mapping(first_page
, class->index
, ZS_EMPTY
);
1405 atomic_long_add(class->pages_per_zspage
,
1406 &pool
->pages_allocated
);
1408 spin_lock(&class->lock
);
1409 zs_stat_inc(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1410 class->size
, class->pages_per_zspage
));
1413 obj
= obj_malloc(first_page
, class, handle
);
1414 /* Now move the zspage to another fullness group, if required */
1415 fix_fullness_group(class, first_page
);
1416 record_obj(handle
, obj
);
1417 spin_unlock(&class->lock
);
1421 EXPORT_SYMBOL_GPL(zs_malloc
);
1423 static void obj_free(struct zs_pool
*pool
, struct size_class
*class,
1426 struct link_free
*link
;
1427 struct page
*first_page
, *f_page
;
1428 unsigned long f_objidx
, f_offset
;
1431 enum fullness_group fullness
;
1435 obj
&= ~OBJ_ALLOCATED_TAG
;
1436 obj_to_location(obj
, &f_page
, &f_objidx
);
1437 first_page
= get_first_page(f_page
);
1439 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1440 f_offset
= obj_idx_to_offset(f_page
, f_objidx
, class->size
);
1442 vaddr
= kmap_atomic(f_page
);
1444 /* Insert this object in containing zspage's freelist */
1445 link
= (struct link_free
*)(vaddr
+ f_offset
);
1446 link
->next
= first_page
->freelist
;
1448 set_page_private(first_page
, 0);
1449 kunmap_atomic(vaddr
);
1450 first_page
->freelist
= (void *)obj
;
1451 first_page
->inuse
--;
1452 zs_stat_dec(class, OBJ_USED
, 1);
1455 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1457 struct page
*first_page
, *f_page
;
1458 unsigned long obj
, f_objidx
;
1460 struct size_class
*class;
1461 enum fullness_group fullness
;
1463 if (unlikely(!handle
))
1467 obj
= handle_to_obj(handle
);
1468 obj_to_location(obj
, &f_page
, &f_objidx
);
1469 first_page
= get_first_page(f_page
);
1471 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1472 class = pool
->size_class
[class_idx
];
1474 spin_lock(&class->lock
);
1475 obj_free(pool
, class, obj
);
1476 fullness
= fix_fullness_group(class, first_page
);
1477 if (fullness
== ZS_EMPTY
) {
1478 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1479 class->size
, class->pages_per_zspage
));
1480 atomic_long_sub(class->pages_per_zspage
,
1481 &pool
->pages_allocated
);
1482 free_zspage(first_page
);
1484 spin_unlock(&class->lock
);
1487 free_handle(pool
, handle
);
1489 EXPORT_SYMBOL_GPL(zs_free
);
1491 static void zs_object_copy(unsigned long dst
, unsigned long src
,
1492 struct size_class
*class)
1494 struct page
*s_page
, *d_page
;
1495 unsigned long s_objidx
, d_objidx
;
1496 unsigned long s_off
, d_off
;
1497 void *s_addr
, *d_addr
;
1498 int s_size
, d_size
, size
;
1501 s_size
= d_size
= class->size
;
1503 obj_to_location(src
, &s_page
, &s_objidx
);
1504 obj_to_location(dst
, &d_page
, &d_objidx
);
1506 s_off
= obj_idx_to_offset(s_page
, s_objidx
, class->size
);
1507 d_off
= obj_idx_to_offset(d_page
, d_objidx
, class->size
);
1509 if (s_off
+ class->size
> PAGE_SIZE
)
1510 s_size
= PAGE_SIZE
- s_off
;
1512 if (d_off
+ class->size
> PAGE_SIZE
)
1513 d_size
= PAGE_SIZE
- d_off
;
1515 s_addr
= kmap_atomic(s_page
);
1516 d_addr
= kmap_atomic(d_page
);
1519 size
= min(s_size
, d_size
);
1520 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1523 if (written
== class->size
)
1531 if (s_off
>= PAGE_SIZE
) {
1532 kunmap_atomic(d_addr
);
1533 kunmap_atomic(s_addr
);
1534 s_page
= get_next_page(s_page
);
1536 s_addr
= kmap_atomic(s_page
);
1537 d_addr
= kmap_atomic(d_page
);
1538 s_size
= class->size
- written
;
1542 if (d_off
>= PAGE_SIZE
) {
1543 kunmap_atomic(d_addr
);
1544 d_page
= get_next_page(d_page
);
1546 d_addr
= kmap_atomic(d_page
);
1547 d_size
= class->size
- written
;
1552 kunmap_atomic(d_addr
);
1553 kunmap_atomic(s_addr
);
1557 * Find alloced object in zspage from index object and
1560 static unsigned long find_alloced_obj(struct page
*page
, int index
,
1561 struct size_class
*class)
1565 unsigned long handle
= 0;
1566 void *addr
= kmap_atomic(page
);
1568 if (!is_first_page(page
))
1569 offset
= page
->index
;
1570 offset
+= class->size
* index
;
1572 while (offset
< PAGE_SIZE
) {
1573 head
= obj_to_head(class, page
, addr
+ offset
);
1574 if (head
& OBJ_ALLOCATED_TAG
) {
1575 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1576 if (trypin_tag(handle
))
1581 offset
+= class->size
;
1585 kunmap_atomic(addr
);
1589 struct zs_compact_control
{
1590 /* Source page for migration which could be a subpage of zspage. */
1591 struct page
*s_page
;
1592 /* Destination page for migration which should be a first page
1594 struct page
*d_page
;
1595 /* Starting object index within @s_page which used for live object
1596 * in the subpage. */
1600 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1601 struct zs_compact_control
*cc
)
1603 unsigned long used_obj
, free_obj
;
1604 unsigned long handle
;
1605 struct page
*s_page
= cc
->s_page
;
1606 struct page
*d_page
= cc
->d_page
;
1607 unsigned long index
= cc
->index
;
1611 handle
= find_alloced_obj(s_page
, index
, class);
1613 s_page
= get_next_page(s_page
);
1620 /* Stop if there is no more space */
1621 if (zspage_full(d_page
)) {
1627 used_obj
= handle_to_obj(handle
);
1628 free_obj
= obj_malloc(d_page
, class, handle
);
1629 zs_object_copy(free_obj
, used_obj
, class);
1631 record_obj(handle
, free_obj
);
1633 obj_free(pool
, class, used_obj
);
1636 /* Remember last position in this iteration */
1637 cc
->s_page
= s_page
;
1643 static struct page
*isolate_target_page(struct size_class
*class)
1648 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1649 page
= class->fullness_list
[i
];
1651 remove_zspage(page
, class, i
);
1660 * putback_zspage - add @first_page into right class's fullness list
1661 * @pool: target pool
1662 * @class: destination class
1663 * @first_page: target page
1665 * Return @fist_page's fullness_group
1667 static enum fullness_group
putback_zspage(struct zs_pool
*pool
,
1668 struct size_class
*class,
1669 struct page
*first_page
)
1671 enum fullness_group fullness
;
1673 BUG_ON(!is_first_page(first_page
));
1675 fullness
= get_fullness_group(first_page
);
1676 insert_zspage(first_page
, class, fullness
);
1677 set_zspage_mapping(first_page
, class->index
, fullness
);
1679 if (fullness
== ZS_EMPTY
) {
1680 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1681 class->size
, class->pages_per_zspage
));
1682 atomic_long_sub(class->pages_per_zspage
,
1683 &pool
->pages_allocated
);
1685 free_zspage(first_page
);
1691 static struct page
*isolate_source_page(struct size_class
*class)
1694 struct page
*page
= NULL
;
1696 for (i
= ZS_ALMOST_EMPTY
; i
>= ZS_ALMOST_FULL
; i
--) {
1697 page
= class->fullness_list
[i
];
1701 remove_zspage(page
, class, i
);
1710 * Based on the number of unused allocated objects calculate
1711 * and return the number of pages that we can free.
1713 static unsigned long zs_can_compact(struct size_class
*class)
1715 unsigned long obj_wasted
;
1717 obj_wasted
= zs_stat_get(class, OBJ_ALLOCATED
) -
1718 zs_stat_get(class, OBJ_USED
);
1720 obj_wasted
/= get_maxobj_per_zspage(class->size
,
1721 class->pages_per_zspage
);
1723 return obj_wasted
* class->pages_per_zspage
;
1726 static void __zs_compact(struct zs_pool
*pool
, struct size_class
*class)
1728 struct zs_compact_control cc
;
1729 struct page
*src_page
;
1730 struct page
*dst_page
= NULL
;
1732 spin_lock(&class->lock
);
1733 while ((src_page
= isolate_source_page(class))) {
1735 BUG_ON(!is_first_page(src_page
));
1737 if (!zs_can_compact(class))
1741 cc
.s_page
= src_page
;
1743 while ((dst_page
= isolate_target_page(class))) {
1744 cc
.d_page
= dst_page
;
1746 * If there is no more space in dst_page, resched
1747 * and see if anyone had allocated another zspage.
1749 if (!migrate_zspage(pool
, class, &cc
))
1752 putback_zspage(pool
, class, dst_page
);
1755 /* Stop if we couldn't find slot */
1756 if (dst_page
== NULL
)
1759 putback_zspage(pool
, class, dst_page
);
1760 if (putback_zspage(pool
, class, src_page
) == ZS_EMPTY
)
1761 pool
->stats
.pages_compacted
+= class->pages_per_zspage
;
1762 spin_unlock(&class->lock
);
1764 spin_lock(&class->lock
);
1768 putback_zspage(pool
, class, src_page
);
1770 spin_unlock(&class->lock
);
1773 unsigned long zs_compact(struct zs_pool
*pool
)
1776 struct size_class
*class;
1778 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1779 class = pool
->size_class
[i
];
1782 if (class->index
!= i
)
1784 __zs_compact(pool
, class);
1787 return pool
->stats
.pages_compacted
;
1789 EXPORT_SYMBOL_GPL(zs_compact
);
1791 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
1793 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
1795 EXPORT_SYMBOL_GPL(zs_pool_stats
);
1797 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
1798 struct shrink_control
*sc
)
1800 unsigned long pages_freed
;
1801 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
1804 pages_freed
= pool
->stats
.pages_compacted
;
1806 * Compact classes and calculate compaction delta.
1807 * Can run concurrently with a manually triggered
1808 * (by user) compaction.
1810 pages_freed
= zs_compact(pool
) - pages_freed
;
1812 return pages_freed
? pages_freed
: SHRINK_STOP
;
1815 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
1816 struct shrink_control
*sc
)
1819 struct size_class
*class;
1820 unsigned long pages_to_free
= 0;
1821 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
1824 if (!pool
->shrinker_enabled
)
1827 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1828 class = pool
->size_class
[i
];
1831 if (class->index
!= i
)
1834 pages_to_free
+= zs_can_compact(class);
1837 return pages_to_free
;
1840 static void zs_unregister_shrinker(struct zs_pool
*pool
)
1842 if (pool
->shrinker_enabled
) {
1843 unregister_shrinker(&pool
->shrinker
);
1844 pool
->shrinker_enabled
= false;
1848 static int zs_register_shrinker(struct zs_pool
*pool
)
1850 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
1851 pool
->shrinker
.count_objects
= zs_shrinker_count
;
1852 pool
->shrinker
.batch
= 0;
1853 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
1855 return register_shrinker(&pool
->shrinker
);
1859 * zs_create_pool - Creates an allocation pool to work from.
1860 * @flags: allocation flags used to allocate pool metadata
1862 * This function must be called before anything when using
1863 * the zsmalloc allocator.
1865 * On success, a pointer to the newly created pool is returned,
1868 struct zs_pool
*zs_create_pool(char *name
, gfp_t flags
)
1871 struct zs_pool
*pool
;
1872 struct size_class
*prev_class
= NULL
;
1874 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
1878 pool
->size_class
= kcalloc(zs_size_classes
, sizeof(struct size_class
*),
1880 if (!pool
->size_class
) {
1885 pool
->name
= kstrdup(name
, GFP_KERNEL
);
1889 if (create_handle_cache(pool
))
1893 * Iterate reversly, because, size of size_class that we want to use
1894 * for merging should be larger or equal to current size.
1896 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1898 int pages_per_zspage
;
1899 struct size_class
*class;
1901 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
1902 if (size
> ZS_MAX_ALLOC_SIZE
)
1903 size
= ZS_MAX_ALLOC_SIZE
;
1904 pages_per_zspage
= get_pages_per_zspage(size
);
1907 * size_class is used for normal zsmalloc operation such
1908 * as alloc/free for that size. Although it is natural that we
1909 * have one size_class for each size, there is a chance that we
1910 * can get more memory utilization if we use one size_class for
1911 * many different sizes whose size_class have same
1912 * characteristics. So, we makes size_class point to
1913 * previous size_class if possible.
1916 if (can_merge(prev_class
, size
, pages_per_zspage
)) {
1917 pool
->size_class
[i
] = prev_class
;
1922 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
1928 class->pages_per_zspage
= pages_per_zspage
;
1929 if (pages_per_zspage
== 1 &&
1930 get_maxobj_per_zspage(size
, pages_per_zspage
) == 1)
1932 spin_lock_init(&class->lock
);
1933 pool
->size_class
[i
] = class;
1938 pool
->flags
= flags
;
1940 if (zs_pool_stat_create(name
, pool
))
1944 * Not critical, we still can use the pool
1945 * and user can trigger compaction manually.
1947 if (zs_register_shrinker(pool
) == 0)
1948 pool
->shrinker_enabled
= true;
1952 zs_destroy_pool(pool
);
1955 EXPORT_SYMBOL_GPL(zs_create_pool
);
1957 void zs_destroy_pool(struct zs_pool
*pool
)
1961 zs_unregister_shrinker(pool
);
1962 zs_pool_stat_destroy(pool
);
1964 for (i
= 0; i
< zs_size_classes
; i
++) {
1966 struct size_class
*class = pool
->size_class
[i
];
1971 if (class->index
!= i
)
1974 for (fg
= 0; fg
< _ZS_NR_FULLNESS_GROUPS
; fg
++) {
1975 if (class->fullness_list
[fg
]) {
1976 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1983 destroy_handle_cache(pool
);
1984 kfree(pool
->size_class
);
1988 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
1990 static int __init
zs_init(void)
1992 int ret
= zs_register_cpu_notifier();
1997 init_zs_size_classes();
2000 zpool_register_driver(&zs_zpool_driver
);
2003 ret
= zs_stat_init();
2005 pr_err("zs stat initialization failed\n");
2012 zpool_unregister_driver(&zs_zpool_driver
);
2015 zs_unregister_cpu_notifier();
2020 static void __exit
zs_exit(void)
2023 zpool_unregister_driver(&zs_zpool_driver
);
2025 zs_unregister_cpu_notifier();
2030 module_init(zs_init
);
2031 module_exit(zs_exit
);
2033 MODULE_LICENSE("Dual BSD/GPL");
2034 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");