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
41 * page->inuse: the number of objects that are used in this zspage
43 * Usage of struct page flags:
44 * PG_private: identifies the first component page
45 * PG_private2: identifies the last component page
49 #include <linux/module.h>
50 #include <linux/kernel.h>
51 #include <linux/sched.h>
52 #include <linux/bitops.h>
53 #include <linux/errno.h>
54 #include <linux/highmem.h>
55 #include <linux/string.h>
56 #include <linux/slab.h>
57 #include <asm/tlbflush.h>
58 #include <asm/pgtable.h>
59 #include <linux/cpumask.h>
60 #include <linux/cpu.h>
61 #include <linux/vmalloc.h>
62 #include <linux/preempt.h>
63 #include <linux/spinlock.h>
64 #include <linux/types.h>
65 #include <linux/debugfs.h>
66 #include <linux/zsmalloc.h>
67 #include <linux/zpool.h>
70 * This must be power of 2 and greater than of equal to sizeof(link_free).
71 * These two conditions ensure that any 'struct link_free' itself doesn't
72 * span more than 1 page which avoids complex case of mapping 2 pages simply
73 * to restore link_free pointer values.
78 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
79 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
81 #define ZS_MAX_ZSPAGE_ORDER 2
82 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
84 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
87 * Object location (<PFN>, <obj_idx>) is encoded as
88 * as single (unsigned long) handle value.
90 * Note that object index <obj_idx> is relative to system
91 * page <PFN> it is stored in, so for each sub-page belonging
92 * to a zspage, obj_idx starts with 0.
94 * This is made more complicated by various memory models and PAE.
97 #ifndef MAX_PHYSMEM_BITS
98 #ifdef CONFIG_HIGHMEM64G
99 #define MAX_PHYSMEM_BITS 36
100 #else /* !CONFIG_HIGHMEM64G */
102 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
105 #define MAX_PHYSMEM_BITS BITS_PER_LONG
108 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
111 * Memory for allocating for handle keeps object position by
112 * encoding <page, obj_idx> and the encoded value has a room
113 * in least bit(ie, look at obj_to_location).
114 * We use the bit to synchronize between object access by
115 * user and migration.
117 #define HANDLE_PIN_BIT 0
120 * Head in allocated object should have OBJ_ALLOCATED_TAG
121 * to identify the object was allocated or not.
122 * It's okay to add the status bit in the least bit because
123 * header keeps handle which is 4byte-aligned address so we
124 * have room for two bit at least.
126 #define OBJ_ALLOCATED_TAG 1
127 #define OBJ_TAG_BITS 1
128 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
129 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
131 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
132 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
133 #define ZS_MIN_ALLOC_SIZE \
134 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
135 /* each chunk includes extra space to keep handle */
136 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
139 * On systems with 4K page size, this gives 255 size classes! There is a
141 * - Large number of size classes is potentially wasteful as free page are
142 * spread across these classes
143 * - Small number of size classes causes large internal fragmentation
144 * - Probably its better to use specific size classes (empirically
145 * determined). NOTE: all those class sizes must be set as multiple of
146 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
148 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
151 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
154 * We do not maintain any list for completely empty or full pages
156 enum fullness_group
{
159 _ZS_NR_FULLNESS_GROUPS
,
173 struct zs_size_stat
{
174 unsigned long objs
[NR_ZS_STAT_TYPE
];
177 #ifdef CONFIG_ZSMALLOC_STAT
178 static struct dentry
*zs_stat_root
;
182 * number of size_classes
184 static int zs_size_classes
;
187 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
189 * n = number of allocated objects
190 * N = total number of objects zspage can store
191 * f = fullness_threshold_frac
193 * Similarly, we assign zspage to:
194 * ZS_ALMOST_FULL when n > N / f
195 * ZS_EMPTY when n == 0
196 * ZS_FULL when n == N
198 * (see: fix_fullness_group())
200 static const int fullness_threshold_frac
= 4;
204 struct page
*fullness_list
[_ZS_NR_FULLNESS_GROUPS
];
206 * Size of objects stored in this class. Must be multiple
212 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
213 int pages_per_zspage
;
214 struct zs_size_stat stats
;
216 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
221 * Placed within free objects to form a singly linked list.
222 * For every zspage, first_page->freelist gives head of this list.
224 * This must be power of 2 and less than or equal to ZS_ALIGN
229 * Position of next free chunk (encodes <PFN, obj_idx>)
230 * It's valid for non-allocated object
234 * Handle of allocated object.
236 unsigned long handle
;
243 struct size_class
**size_class
;
244 struct kmem_cache
*handle_cachep
;
246 gfp_t flags
; /* allocation flags used when growing pool */
247 atomic_long_t pages_allocated
;
249 struct zs_pool_stats stats
;
251 /* Compact classes */
252 struct shrinker shrinker
;
254 * To signify that register_shrinker() was successful
255 * and unregister_shrinker() will not Oops.
257 bool shrinker_enabled
;
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 kmem_cache_destroy(pool
->handle_cachep
);
295 static unsigned long alloc_handle(struct zs_pool
*pool
)
297 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
298 pool
->flags
& ~__GFP_HIGHMEM
);
301 static void free_handle(struct zs_pool
*pool
, unsigned long handle
)
303 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
306 static void record_obj(unsigned long handle
, unsigned long obj
)
308 *(unsigned long *)handle
= obj
;
315 static void *zs_zpool_create(const char *name
, gfp_t gfp
,
316 const struct zpool_ops
*zpool_ops
,
319 return zs_create_pool(name
, gfp
);
322 static void zs_zpool_destroy(void *pool
)
324 zs_destroy_pool(pool
);
327 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
328 unsigned long *handle
)
330 *handle
= zs_malloc(pool
, size
);
331 return *handle
? 0 : -1;
333 static void zs_zpool_free(void *pool
, unsigned long handle
)
335 zs_free(pool
, handle
);
338 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
339 unsigned int *reclaimed
)
344 static void *zs_zpool_map(void *pool
, unsigned long handle
,
345 enum zpool_mapmode mm
)
347 enum zs_mapmode zs_mm
;
356 case ZPOOL_MM_RW
: /* fallthru */
362 return zs_map_object(pool
, handle
, zs_mm
);
364 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
366 zs_unmap_object(pool
, handle
);
369 static u64
zs_zpool_total_size(void *pool
)
371 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
374 static struct zpool_driver zs_zpool_driver
= {
376 .owner
= THIS_MODULE
,
377 .create
= zs_zpool_create
,
378 .destroy
= zs_zpool_destroy
,
379 .malloc
= zs_zpool_malloc
,
380 .free
= zs_zpool_free
,
381 .shrink
= zs_zpool_shrink
,
383 .unmap
= zs_zpool_unmap
,
384 .total_size
= zs_zpool_total_size
,
387 MODULE_ALIAS("zpool-zsmalloc");
388 #endif /* CONFIG_ZPOOL */
390 static unsigned int get_maxobj_per_zspage(int size
, int pages_per_zspage
)
392 return pages_per_zspage
* PAGE_SIZE
/ size
;
395 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
396 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
398 static int is_first_page(struct page
*page
)
400 return PagePrivate(page
);
403 static int is_last_page(struct page
*page
)
405 return PagePrivate2(page
);
408 static void get_zspage_mapping(struct page
*page
, unsigned int *class_idx
,
409 enum fullness_group
*fullness
)
412 BUG_ON(!is_first_page(page
));
414 m
= (unsigned long)page
->mapping
;
415 *fullness
= m
& FULLNESS_MASK
;
416 *class_idx
= (m
>> FULLNESS_BITS
) & CLASS_IDX_MASK
;
419 static void set_zspage_mapping(struct page
*page
, unsigned int class_idx
,
420 enum fullness_group fullness
)
423 BUG_ON(!is_first_page(page
));
425 m
= ((class_idx
& CLASS_IDX_MASK
) << FULLNESS_BITS
) |
426 (fullness
& FULLNESS_MASK
);
427 page
->mapping
= (struct address_space
*)m
;
431 * zsmalloc divides the pool into various size classes where each
432 * class maintains a list of zspages where each zspage is divided
433 * into equal sized chunks. Each allocation falls into one of these
434 * classes depending on its size. This function returns index of the
435 * size class which has chunk size big enough to hold the give size.
437 static int get_size_class_index(int size
)
441 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
442 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
443 ZS_SIZE_CLASS_DELTA
);
445 return min(zs_size_classes
- 1, idx
);
448 static inline void zs_stat_inc(struct size_class
*class,
449 enum zs_stat_type type
, unsigned long cnt
)
451 class->stats
.objs
[type
] += cnt
;
454 static inline void zs_stat_dec(struct size_class
*class,
455 enum zs_stat_type type
, unsigned long cnt
)
457 class->stats
.objs
[type
] -= cnt
;
460 static inline unsigned long zs_stat_get(struct size_class
*class,
461 enum zs_stat_type type
)
463 return class->stats
.objs
[type
];
466 #ifdef CONFIG_ZSMALLOC_STAT
468 static int __init
zs_stat_init(void)
470 if (!debugfs_initialized())
473 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
480 static void __exit
zs_stat_exit(void)
482 debugfs_remove_recursive(zs_stat_root
);
485 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
488 struct zs_pool
*pool
= s
->private;
489 struct size_class
*class;
491 unsigned long class_almost_full
, class_almost_empty
;
492 unsigned long obj_allocated
, obj_used
, pages_used
;
493 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
494 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
496 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
497 "class", "size", "almost_full", "almost_empty",
498 "obj_allocated", "obj_used", "pages_used",
501 for (i
= 0; i
< zs_size_classes
; i
++) {
502 class = pool
->size_class
[i
];
504 if (class->index
!= i
)
507 spin_lock(&class->lock
);
508 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
509 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
510 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
511 obj_used
= zs_stat_get(class, OBJ_USED
);
512 spin_unlock(&class->lock
);
514 objs_per_zspage
= get_maxobj_per_zspage(class->size
,
515 class->pages_per_zspage
);
516 pages_used
= obj_allocated
/ objs_per_zspage
*
517 class->pages_per_zspage
;
519 seq_printf(s
, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
520 i
, class->size
, class_almost_full
, class_almost_empty
,
521 obj_allocated
, obj_used
, pages_used
,
522 class->pages_per_zspage
);
524 total_class_almost_full
+= class_almost_full
;
525 total_class_almost_empty
+= class_almost_empty
;
526 total_objs
+= obj_allocated
;
527 total_used_objs
+= obj_used
;
528 total_pages
+= pages_used
;
532 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
533 "Total", "", total_class_almost_full
,
534 total_class_almost_empty
, total_objs
,
535 total_used_objs
, total_pages
);
540 static int zs_stats_size_open(struct inode
*inode
, struct file
*file
)
542 return single_open(file
, zs_stats_size_show
, inode
->i_private
);
545 static const struct file_operations zs_stat_size_ops
= {
546 .open
= zs_stats_size_open
,
549 .release
= single_release
,
552 static int zs_pool_stat_create(const char *name
, struct zs_pool
*pool
)
554 struct dentry
*entry
;
559 entry
= debugfs_create_dir(name
, zs_stat_root
);
561 pr_warn("debugfs dir <%s> creation failed\n", name
);
564 pool
->stat_dentry
= entry
;
566 entry
= debugfs_create_file("classes", S_IFREG
| S_IRUGO
,
567 pool
->stat_dentry
, pool
, &zs_stat_size_ops
);
569 pr_warn("%s: debugfs file entry <%s> creation failed\n",
577 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
579 debugfs_remove_recursive(pool
->stat_dentry
);
582 #else /* CONFIG_ZSMALLOC_STAT */
583 static int __init
zs_stat_init(void)
588 static void __exit
zs_stat_exit(void)
592 static inline int zs_pool_stat_create(const char *name
, struct zs_pool
*pool
)
597 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
604 * For each size class, zspages are divided into different groups
605 * depending on how "full" they are. This was done so that we could
606 * easily find empty or nearly empty zspages when we try to shrink
607 * the pool (not yet implemented). This function returns fullness
608 * status of the given page.
610 static enum fullness_group
get_fullness_group(struct page
*page
)
612 int inuse
, max_objects
;
613 enum fullness_group fg
;
614 BUG_ON(!is_first_page(page
));
617 max_objects
= page
->objects
;
621 else if (inuse
== max_objects
)
623 else if (inuse
<= 3 * max_objects
/ fullness_threshold_frac
)
624 fg
= ZS_ALMOST_EMPTY
;
632 * Each size class maintains various freelists and zspages are assigned
633 * to one of these freelists based on the number of live objects they
634 * have. This functions inserts the given zspage into the freelist
635 * identified by <class, fullness_group>.
637 static void insert_zspage(struct page
*page
, struct size_class
*class,
638 enum fullness_group fullness
)
642 BUG_ON(!is_first_page(page
));
644 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
647 zs_stat_inc(class, fullness
== ZS_ALMOST_EMPTY
?
648 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
650 head
= &class->fullness_list
[fullness
];
657 * We want to see more ZS_FULL pages and less almost
658 * empty/full. Put pages with higher ->inuse first.
660 list_add_tail(&page
->lru
, &(*head
)->lru
);
661 if (page
->inuse
>= (*head
)->inuse
)
666 * This function removes the given zspage from the freelist identified
667 * by <class, fullness_group>.
669 static void remove_zspage(struct page
*page
, struct size_class
*class,
670 enum fullness_group fullness
)
674 BUG_ON(!is_first_page(page
));
676 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
679 head
= &class->fullness_list
[fullness
];
681 if (list_empty(&(*head
)->lru
))
683 else if (*head
== page
)
684 *head
= (struct page
*)list_entry((*head
)->lru
.next
,
687 list_del_init(&page
->lru
);
688 zs_stat_dec(class, fullness
== ZS_ALMOST_EMPTY
?
689 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
693 * Each size class maintains zspages in different fullness groups depending
694 * on the number of live objects they contain. When allocating or freeing
695 * objects, the fullness status of the page can change, say, from ALMOST_FULL
696 * to ALMOST_EMPTY when freeing an object. This function checks if such
697 * a status change has occurred for the given page and accordingly moves the
698 * page from the freelist of the old fullness group to that of the new
701 static enum fullness_group
fix_fullness_group(struct size_class
*class,
705 enum fullness_group currfg
, newfg
;
707 BUG_ON(!is_first_page(page
));
709 get_zspage_mapping(page
, &class_idx
, &currfg
);
710 newfg
= get_fullness_group(page
);
714 remove_zspage(page
, class, currfg
);
715 insert_zspage(page
, class, newfg
);
716 set_zspage_mapping(page
, class_idx
, newfg
);
723 * We have to decide on how many pages to link together
724 * to form a zspage for each size class. This is important
725 * to reduce wastage due to unusable space left at end of
726 * each zspage which is given as:
727 * wastage = Zp % class_size
728 * usage = Zp - wastage
729 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
731 * For example, for size class of 3/8 * PAGE_SIZE, we should
732 * link together 3 PAGE_SIZE sized pages to form a zspage
733 * since then we can perfectly fit in 8 such objects.
735 static int get_pages_per_zspage(int class_size
)
737 int i
, max_usedpc
= 0;
738 /* zspage order which gives maximum used size per KB */
739 int max_usedpc_order
= 1;
741 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
745 zspage_size
= i
* PAGE_SIZE
;
746 waste
= zspage_size
% class_size
;
747 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
749 if (usedpc
> max_usedpc
) {
751 max_usedpc_order
= i
;
755 return max_usedpc_order
;
759 * A single 'zspage' is composed of many system pages which are
760 * linked together using fields in struct page. This function finds
761 * the first/head page, given any component page of a zspage.
763 static struct page
*get_first_page(struct page
*page
)
765 if (is_first_page(page
))
768 return page
->first_page
;
771 static struct page
*get_next_page(struct page
*page
)
775 if (is_last_page(page
))
777 else if (is_first_page(page
))
778 next
= (struct page
*)page_private(page
);
780 next
= list_entry(page
->lru
.next
, struct page
, lru
);
786 * Encode <page, obj_idx> as a single handle value.
787 * We use the least bit of handle for tagging.
789 static void *location_to_obj(struct page
*page
, unsigned long obj_idx
)
798 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
799 obj
|= ((obj_idx
) & OBJ_INDEX_MASK
);
800 obj
<<= OBJ_TAG_BITS
;
806 * Decode <page, obj_idx> pair from the given object handle. We adjust the
807 * decoded obj_idx back to its original value since it was adjusted in
810 static void obj_to_location(unsigned long obj
, struct page
**page
,
811 unsigned long *obj_idx
)
813 obj
>>= OBJ_TAG_BITS
;
814 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
815 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
818 static unsigned long handle_to_obj(unsigned long handle
)
820 return *(unsigned long *)handle
;
823 static unsigned long obj_to_head(struct size_class
*class, struct page
*page
,
827 VM_BUG_ON(!is_first_page(page
));
828 return page_private(page
);
830 return *(unsigned long *)obj
;
833 static unsigned long obj_idx_to_offset(struct page
*page
,
834 unsigned long obj_idx
, int class_size
)
836 unsigned long off
= 0;
838 if (!is_first_page(page
))
841 return off
+ obj_idx
* class_size
;
844 static inline int trypin_tag(unsigned long handle
)
846 unsigned long *ptr
= (unsigned long *)handle
;
848 return !test_and_set_bit_lock(HANDLE_PIN_BIT
, ptr
);
851 static void pin_tag(unsigned long handle
)
853 while (!trypin_tag(handle
));
856 static void unpin_tag(unsigned long handle
)
858 unsigned long *ptr
= (unsigned long *)handle
;
860 clear_bit_unlock(HANDLE_PIN_BIT
, ptr
);
863 static void reset_page(struct page
*page
)
865 clear_bit(PG_private
, &page
->flags
);
866 clear_bit(PG_private_2
, &page
->flags
);
867 set_page_private(page
, 0);
868 page
->mapping
= NULL
;
869 page
->freelist
= NULL
;
870 page_mapcount_reset(page
);
873 static void free_zspage(struct page
*first_page
)
875 struct page
*nextp
, *tmp
, *head_extra
;
877 BUG_ON(!is_first_page(first_page
));
878 BUG_ON(first_page
->inuse
);
880 head_extra
= (struct page
*)page_private(first_page
);
882 reset_page(first_page
);
883 __free_page(first_page
);
885 /* zspage with only 1 system page */
889 list_for_each_entry_safe(nextp
, tmp
, &head_extra
->lru
, lru
) {
890 list_del(&nextp
->lru
);
894 reset_page(head_extra
);
895 __free_page(head_extra
);
898 /* Initialize a newly allocated zspage */
899 static void init_zspage(struct page
*first_page
, struct size_class
*class)
901 unsigned long off
= 0;
902 struct page
*page
= first_page
;
904 BUG_ON(!is_first_page(first_page
));
906 struct page
*next_page
;
907 struct link_free
*link
;
912 * page->index stores offset of first object starting
913 * in the page. For the first page, this is always 0,
914 * so we use first_page->index (aka ->freelist) to store
915 * head of corresponding zspage's freelist.
917 if (page
!= first_page
)
920 vaddr
= kmap_atomic(page
);
921 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
923 while ((off
+= class->size
) < PAGE_SIZE
) {
924 link
->next
= location_to_obj(page
, i
++);
925 link
+= class->size
/ sizeof(*link
);
929 * We now come to the last (full or partial) object on this
930 * page, which must point to the first object on the next
933 next_page
= get_next_page(page
);
934 link
->next
= location_to_obj(next_page
, 0);
935 kunmap_atomic(vaddr
);
942 * Allocate a zspage for the given size class
944 static struct page
*alloc_zspage(struct size_class
*class, gfp_t flags
)
947 struct page
*first_page
= NULL
, *uninitialized_var(prev_page
);
950 * Allocate individual pages and link them together as:
951 * 1. first page->private = first sub-page
952 * 2. all sub-pages are linked together using page->lru
953 * 3. each sub-page is linked to the first page using page->first_page
955 * For each size class, First/Head pages are linked together using
956 * page->lru. Also, we set PG_private to identify the first page
957 * (i.e. no other sub-page has this flag set) and PG_private_2 to
958 * identify the last page.
961 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
964 page
= alloc_page(flags
);
968 INIT_LIST_HEAD(&page
->lru
);
969 if (i
== 0) { /* first page */
970 SetPagePrivate(page
);
971 set_page_private(page
, 0);
973 first_page
->inuse
= 0;
976 set_page_private(first_page
, (unsigned long)page
);
978 page
->first_page
= first_page
;
980 list_add(&page
->lru
, &prev_page
->lru
);
981 if (i
== class->pages_per_zspage
- 1) /* last page */
982 SetPagePrivate2(page
);
986 init_zspage(first_page
, class);
988 first_page
->freelist
= location_to_obj(first_page
, 0);
989 /* Maximum number of objects we can store in this zspage */
990 first_page
->objects
= class->pages_per_zspage
* PAGE_SIZE
/ class->size
;
992 error
= 0; /* Success */
995 if (unlikely(error
) && first_page
) {
996 free_zspage(first_page
);
1003 static struct page
*find_get_zspage(struct size_class
*class)
1008 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1009 page
= class->fullness_list
[i
];
1017 #ifdef CONFIG_PGTABLE_MAPPING
1018 static inline int __zs_cpu_up(struct mapping_area
*area
)
1021 * Make sure we don't leak memory if a cpu UP notification
1022 * and zs_init() race and both call zs_cpu_up() on the same cpu
1026 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
1032 static inline void __zs_cpu_down(struct mapping_area
*area
)
1035 free_vm_area(area
->vm
);
1039 static inline void *__zs_map_object(struct mapping_area
*area
,
1040 struct page
*pages
[2], int off
, int size
)
1042 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
1043 area
->vm_addr
= area
->vm
->addr
;
1044 return area
->vm_addr
+ off
;
1047 static inline void __zs_unmap_object(struct mapping_area
*area
,
1048 struct page
*pages
[2], int off
, int size
)
1050 unsigned long addr
= (unsigned long)area
->vm_addr
;
1052 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
1055 #else /* CONFIG_PGTABLE_MAPPING */
1057 static inline int __zs_cpu_up(struct mapping_area
*area
)
1060 * Make sure we don't leak memory if a cpu UP notification
1061 * and zs_init() race and both call zs_cpu_up() on the same cpu
1065 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1071 static inline void __zs_cpu_down(struct mapping_area
*area
)
1073 kfree(area
->vm_buf
);
1074 area
->vm_buf
= NULL
;
1077 static void *__zs_map_object(struct mapping_area
*area
,
1078 struct page
*pages
[2], int off
, int size
)
1082 char *buf
= area
->vm_buf
;
1084 /* disable page faults to match kmap_atomic() return conditions */
1085 pagefault_disable();
1087 /* no read fastpath */
1088 if (area
->vm_mm
== ZS_MM_WO
)
1091 sizes
[0] = PAGE_SIZE
- off
;
1092 sizes
[1] = size
- sizes
[0];
1094 /* copy object to per-cpu buffer */
1095 addr
= kmap_atomic(pages
[0]);
1096 memcpy(buf
, addr
+ off
, sizes
[0]);
1097 kunmap_atomic(addr
);
1098 addr
= kmap_atomic(pages
[1]);
1099 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1100 kunmap_atomic(addr
);
1102 return area
->vm_buf
;
1105 static void __zs_unmap_object(struct mapping_area
*area
,
1106 struct page
*pages
[2], int off
, int size
)
1112 /* no write fastpath */
1113 if (area
->vm_mm
== ZS_MM_RO
)
1118 buf
= buf
+ ZS_HANDLE_SIZE
;
1119 size
-= ZS_HANDLE_SIZE
;
1120 off
+= ZS_HANDLE_SIZE
;
1123 sizes
[0] = PAGE_SIZE
- off
;
1124 sizes
[1] = size
- sizes
[0];
1126 /* copy per-cpu buffer to object */
1127 addr
= kmap_atomic(pages
[0]);
1128 memcpy(addr
+ off
, buf
, sizes
[0]);
1129 kunmap_atomic(addr
);
1130 addr
= kmap_atomic(pages
[1]);
1131 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1132 kunmap_atomic(addr
);
1135 /* enable page faults to match kunmap_atomic() return conditions */
1139 #endif /* CONFIG_PGTABLE_MAPPING */
1141 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
1144 int ret
, cpu
= (long)pcpu
;
1145 struct mapping_area
*area
;
1148 case CPU_UP_PREPARE
:
1149 area
= &per_cpu(zs_map_area
, cpu
);
1150 ret
= __zs_cpu_up(area
);
1152 return notifier_from_errno(ret
);
1155 case CPU_UP_CANCELED
:
1156 area
= &per_cpu(zs_map_area
, cpu
);
1157 __zs_cpu_down(area
);
1164 static struct notifier_block zs_cpu_nb
= {
1165 .notifier_call
= zs_cpu_notifier
1168 static int zs_register_cpu_notifier(void)
1170 int cpu
, uninitialized_var(ret
);
1172 cpu_notifier_register_begin();
1174 __register_cpu_notifier(&zs_cpu_nb
);
1175 for_each_online_cpu(cpu
) {
1176 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1177 if (notifier_to_errno(ret
))
1181 cpu_notifier_register_done();
1182 return notifier_to_errno(ret
);
1185 static void zs_unregister_cpu_notifier(void)
1189 cpu_notifier_register_begin();
1191 for_each_online_cpu(cpu
)
1192 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
1193 __unregister_cpu_notifier(&zs_cpu_nb
);
1195 cpu_notifier_register_done();
1198 static void init_zs_size_classes(void)
1202 nr
= (ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) / ZS_SIZE_CLASS_DELTA
+ 1;
1203 if ((ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) % ZS_SIZE_CLASS_DELTA
)
1206 zs_size_classes
= nr
;
1209 static bool can_merge(struct size_class
*prev
, int size
, int pages_per_zspage
)
1211 if (prev
->pages_per_zspage
!= pages_per_zspage
)
1214 if (get_maxobj_per_zspage(prev
->size
, prev
->pages_per_zspage
)
1215 != get_maxobj_per_zspage(size
, pages_per_zspage
))
1221 static bool zspage_full(struct page
*page
)
1223 BUG_ON(!is_first_page(page
));
1225 return page
->inuse
== page
->objects
;
1228 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1230 return atomic_long_read(&pool
->pages_allocated
);
1232 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1235 * zs_map_object - get address of allocated object from handle.
1236 * @pool: pool from which the object was allocated
1237 * @handle: handle returned from zs_malloc
1239 * Before using an object allocated from zs_malloc, it must be mapped using
1240 * this function. When done with the object, it must be unmapped using
1243 * Only one object can be mapped per cpu at a time. There is no protection
1244 * against nested mappings.
1246 * This function returns with preemption and page faults disabled.
1248 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1252 unsigned long obj
, obj_idx
, off
;
1254 unsigned int class_idx
;
1255 enum fullness_group fg
;
1256 struct size_class
*class;
1257 struct mapping_area
*area
;
1258 struct page
*pages
[2];
1264 * Because we use per-cpu mapping areas shared among the
1265 * pools/users, we can't allow mapping in interrupt context
1266 * because it can corrupt another users mappings.
1268 BUG_ON(in_interrupt());
1270 /* From now on, migration cannot move the object */
1273 obj
= handle_to_obj(handle
);
1274 obj_to_location(obj
, &page
, &obj_idx
);
1275 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1276 class = pool
->size_class
[class_idx
];
1277 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1279 area
= &get_cpu_var(zs_map_area
);
1281 if (off
+ class->size
<= PAGE_SIZE
) {
1282 /* this object is contained entirely within a page */
1283 area
->vm_addr
= kmap_atomic(page
);
1284 ret
= area
->vm_addr
+ off
;
1288 /* this object spans two pages */
1290 pages
[1] = get_next_page(page
);
1293 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1296 ret
+= ZS_HANDLE_SIZE
;
1300 EXPORT_SYMBOL_GPL(zs_map_object
);
1302 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1305 unsigned long obj
, obj_idx
, off
;
1307 unsigned int class_idx
;
1308 enum fullness_group fg
;
1309 struct size_class
*class;
1310 struct mapping_area
*area
;
1314 obj
= handle_to_obj(handle
);
1315 obj_to_location(obj
, &page
, &obj_idx
);
1316 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1317 class = pool
->size_class
[class_idx
];
1318 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1320 area
= this_cpu_ptr(&zs_map_area
);
1321 if (off
+ class->size
<= PAGE_SIZE
)
1322 kunmap_atomic(area
->vm_addr
);
1324 struct page
*pages
[2];
1327 pages
[1] = get_next_page(page
);
1330 __zs_unmap_object(area
, pages
, off
, class->size
);
1332 put_cpu_var(zs_map_area
);
1335 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1337 static unsigned long obj_malloc(struct page
*first_page
,
1338 struct size_class
*class, unsigned long handle
)
1341 struct link_free
*link
;
1343 struct page
*m_page
;
1344 unsigned long m_objidx
, m_offset
;
1347 handle
|= OBJ_ALLOCATED_TAG
;
1348 obj
= (unsigned long)first_page
->freelist
;
1349 obj_to_location(obj
, &m_page
, &m_objidx
);
1350 m_offset
= obj_idx_to_offset(m_page
, m_objidx
, class->size
);
1352 vaddr
= kmap_atomic(m_page
);
1353 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1354 first_page
->freelist
= link
->next
;
1356 /* record handle in the header of allocated chunk */
1357 link
->handle
= handle
;
1359 /* record handle in first_page->private */
1360 set_page_private(first_page
, handle
);
1361 kunmap_atomic(vaddr
);
1362 first_page
->inuse
++;
1363 zs_stat_inc(class, OBJ_USED
, 1);
1370 * zs_malloc - Allocate block of given size from pool.
1371 * @pool: pool to allocate from
1372 * @size: size of block to allocate
1374 * On success, handle to the allocated object is returned,
1376 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1378 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
)
1380 unsigned long handle
, obj
;
1381 struct size_class
*class;
1382 struct page
*first_page
;
1384 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1387 handle
= alloc_handle(pool
);
1391 /* extra space in chunk to keep the handle */
1392 size
+= ZS_HANDLE_SIZE
;
1393 class = pool
->size_class
[get_size_class_index(size
)];
1395 spin_lock(&class->lock
);
1396 first_page
= find_get_zspage(class);
1399 spin_unlock(&class->lock
);
1400 first_page
= alloc_zspage(class, pool
->flags
);
1401 if (unlikely(!first_page
)) {
1402 free_handle(pool
, handle
);
1406 set_zspage_mapping(first_page
, class->index
, ZS_EMPTY
);
1407 atomic_long_add(class->pages_per_zspage
,
1408 &pool
->pages_allocated
);
1410 spin_lock(&class->lock
);
1411 zs_stat_inc(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1412 class->size
, class->pages_per_zspage
));
1415 obj
= obj_malloc(first_page
, class, handle
);
1416 /* Now move the zspage to another fullness group, if required */
1417 fix_fullness_group(class, first_page
);
1418 record_obj(handle
, obj
);
1419 spin_unlock(&class->lock
);
1423 EXPORT_SYMBOL_GPL(zs_malloc
);
1425 static void obj_free(struct zs_pool
*pool
, struct size_class
*class,
1428 struct link_free
*link
;
1429 struct page
*first_page
, *f_page
;
1430 unsigned long f_objidx
, f_offset
;
1433 enum fullness_group fullness
;
1437 obj
&= ~OBJ_ALLOCATED_TAG
;
1438 obj_to_location(obj
, &f_page
, &f_objidx
);
1439 first_page
= get_first_page(f_page
);
1441 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1442 f_offset
= obj_idx_to_offset(f_page
, f_objidx
, class->size
);
1444 vaddr
= kmap_atomic(f_page
);
1446 /* Insert this object in containing zspage's freelist */
1447 link
= (struct link_free
*)(vaddr
+ f_offset
);
1448 link
->next
= first_page
->freelist
;
1450 set_page_private(first_page
, 0);
1451 kunmap_atomic(vaddr
);
1452 first_page
->freelist
= (void *)obj
;
1453 first_page
->inuse
--;
1454 zs_stat_dec(class, OBJ_USED
, 1);
1457 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1459 struct page
*first_page
, *f_page
;
1460 unsigned long obj
, f_objidx
;
1462 struct size_class
*class;
1463 enum fullness_group fullness
;
1465 if (unlikely(!handle
))
1469 obj
= handle_to_obj(handle
);
1470 obj_to_location(obj
, &f_page
, &f_objidx
);
1471 first_page
= get_first_page(f_page
);
1473 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1474 class = pool
->size_class
[class_idx
];
1476 spin_lock(&class->lock
);
1477 obj_free(pool
, class, obj
);
1478 fullness
= fix_fullness_group(class, first_page
);
1479 if (fullness
== ZS_EMPTY
) {
1480 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1481 class->size
, class->pages_per_zspage
));
1482 atomic_long_sub(class->pages_per_zspage
,
1483 &pool
->pages_allocated
);
1484 free_zspage(first_page
);
1486 spin_unlock(&class->lock
);
1489 free_handle(pool
, handle
);
1491 EXPORT_SYMBOL_GPL(zs_free
);
1493 static void zs_object_copy(unsigned long dst
, unsigned long src
,
1494 struct size_class
*class)
1496 struct page
*s_page
, *d_page
;
1497 unsigned long s_objidx
, d_objidx
;
1498 unsigned long s_off
, d_off
;
1499 void *s_addr
, *d_addr
;
1500 int s_size
, d_size
, size
;
1503 s_size
= d_size
= class->size
;
1505 obj_to_location(src
, &s_page
, &s_objidx
);
1506 obj_to_location(dst
, &d_page
, &d_objidx
);
1508 s_off
= obj_idx_to_offset(s_page
, s_objidx
, class->size
);
1509 d_off
= obj_idx_to_offset(d_page
, d_objidx
, class->size
);
1511 if (s_off
+ class->size
> PAGE_SIZE
)
1512 s_size
= PAGE_SIZE
- s_off
;
1514 if (d_off
+ class->size
> PAGE_SIZE
)
1515 d_size
= PAGE_SIZE
- d_off
;
1517 s_addr
= kmap_atomic(s_page
);
1518 d_addr
= kmap_atomic(d_page
);
1521 size
= min(s_size
, d_size
);
1522 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1525 if (written
== class->size
)
1533 if (s_off
>= PAGE_SIZE
) {
1534 kunmap_atomic(d_addr
);
1535 kunmap_atomic(s_addr
);
1536 s_page
= get_next_page(s_page
);
1538 s_addr
= kmap_atomic(s_page
);
1539 d_addr
= kmap_atomic(d_page
);
1540 s_size
= class->size
- written
;
1544 if (d_off
>= PAGE_SIZE
) {
1545 kunmap_atomic(d_addr
);
1546 d_page
= get_next_page(d_page
);
1548 d_addr
= kmap_atomic(d_page
);
1549 d_size
= class->size
- written
;
1554 kunmap_atomic(d_addr
);
1555 kunmap_atomic(s_addr
);
1559 * Find alloced object in zspage from index object and
1562 static unsigned long find_alloced_obj(struct page
*page
, int index
,
1563 struct size_class
*class)
1567 unsigned long handle
= 0;
1568 void *addr
= kmap_atomic(page
);
1570 if (!is_first_page(page
))
1571 offset
= page
->index
;
1572 offset
+= class->size
* index
;
1574 while (offset
< PAGE_SIZE
) {
1575 head
= obj_to_head(class, page
, addr
+ offset
);
1576 if (head
& OBJ_ALLOCATED_TAG
) {
1577 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1578 if (trypin_tag(handle
))
1583 offset
+= class->size
;
1587 kunmap_atomic(addr
);
1591 struct zs_compact_control
{
1592 /* Source page for migration which could be a subpage of zspage. */
1593 struct page
*s_page
;
1594 /* Destination page for migration which should be a first page
1596 struct page
*d_page
;
1597 /* Starting object index within @s_page which used for live object
1598 * in the subpage. */
1602 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1603 struct zs_compact_control
*cc
)
1605 unsigned long used_obj
, free_obj
;
1606 unsigned long handle
;
1607 struct page
*s_page
= cc
->s_page
;
1608 struct page
*d_page
= cc
->d_page
;
1609 unsigned long index
= cc
->index
;
1613 handle
= find_alloced_obj(s_page
, index
, class);
1615 s_page
= get_next_page(s_page
);
1622 /* Stop if there is no more space */
1623 if (zspage_full(d_page
)) {
1629 used_obj
= handle_to_obj(handle
);
1630 free_obj
= obj_malloc(d_page
, class, handle
);
1631 zs_object_copy(free_obj
, used_obj
, class);
1633 record_obj(handle
, free_obj
);
1635 obj_free(pool
, class, used_obj
);
1638 /* Remember last position in this iteration */
1639 cc
->s_page
= s_page
;
1645 static struct page
*isolate_target_page(struct size_class
*class)
1650 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1651 page
= class->fullness_list
[i
];
1653 remove_zspage(page
, class, i
);
1662 * putback_zspage - add @first_page into right class's fullness list
1663 * @pool: target pool
1664 * @class: destination class
1665 * @first_page: target page
1667 * Return @fist_page's fullness_group
1669 static enum fullness_group
putback_zspage(struct zs_pool
*pool
,
1670 struct size_class
*class,
1671 struct page
*first_page
)
1673 enum fullness_group fullness
;
1675 BUG_ON(!is_first_page(first_page
));
1677 fullness
= get_fullness_group(first_page
);
1678 insert_zspage(first_page
, class, fullness
);
1679 set_zspage_mapping(first_page
, class->index
, fullness
);
1681 if (fullness
== ZS_EMPTY
) {
1682 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1683 class->size
, class->pages_per_zspage
));
1684 atomic_long_sub(class->pages_per_zspage
,
1685 &pool
->pages_allocated
);
1687 free_zspage(first_page
);
1693 static struct page
*isolate_source_page(struct size_class
*class)
1696 struct page
*page
= NULL
;
1698 for (i
= ZS_ALMOST_EMPTY
; i
>= ZS_ALMOST_FULL
; i
--) {
1699 page
= class->fullness_list
[i
];
1703 remove_zspage(page
, class, i
);
1712 * Based on the number of unused allocated objects calculate
1713 * and return the number of pages that we can free.
1715 static unsigned long zs_can_compact(struct size_class
*class)
1717 unsigned long obj_wasted
;
1719 obj_wasted
= zs_stat_get(class, OBJ_ALLOCATED
) -
1720 zs_stat_get(class, OBJ_USED
);
1722 obj_wasted
/= get_maxobj_per_zspage(class->size
,
1723 class->pages_per_zspage
);
1725 return obj_wasted
* class->pages_per_zspage
;
1728 static void __zs_compact(struct zs_pool
*pool
, struct size_class
*class)
1730 struct zs_compact_control cc
;
1731 struct page
*src_page
;
1732 struct page
*dst_page
= NULL
;
1734 spin_lock(&class->lock
);
1735 while ((src_page
= isolate_source_page(class))) {
1737 BUG_ON(!is_first_page(src_page
));
1739 if (!zs_can_compact(class))
1743 cc
.s_page
= src_page
;
1745 while ((dst_page
= isolate_target_page(class))) {
1746 cc
.d_page
= dst_page
;
1748 * If there is no more space in dst_page, resched
1749 * and see if anyone had allocated another zspage.
1751 if (!migrate_zspage(pool
, class, &cc
))
1754 putback_zspage(pool
, class, dst_page
);
1757 /* Stop if we couldn't find slot */
1758 if (dst_page
== NULL
)
1761 putback_zspage(pool
, class, dst_page
);
1762 if (putback_zspage(pool
, class, src_page
) == ZS_EMPTY
)
1763 pool
->stats
.pages_compacted
+= class->pages_per_zspage
;
1764 spin_unlock(&class->lock
);
1766 spin_lock(&class->lock
);
1770 putback_zspage(pool
, class, src_page
);
1772 spin_unlock(&class->lock
);
1775 unsigned long zs_compact(struct zs_pool
*pool
)
1778 struct size_class
*class;
1780 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1781 class = pool
->size_class
[i
];
1784 if (class->index
!= i
)
1786 __zs_compact(pool
, class);
1789 return pool
->stats
.pages_compacted
;
1791 EXPORT_SYMBOL_GPL(zs_compact
);
1793 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
1795 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
1797 EXPORT_SYMBOL_GPL(zs_pool_stats
);
1799 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
1800 struct shrink_control
*sc
)
1802 unsigned long pages_freed
;
1803 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
1806 pages_freed
= pool
->stats
.pages_compacted
;
1808 * Compact classes and calculate compaction delta.
1809 * Can run concurrently with a manually triggered
1810 * (by user) compaction.
1812 pages_freed
= zs_compact(pool
) - pages_freed
;
1814 return pages_freed
? pages_freed
: SHRINK_STOP
;
1817 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
1818 struct shrink_control
*sc
)
1821 struct size_class
*class;
1822 unsigned long pages_to_free
= 0;
1823 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
1826 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1827 class = pool
->size_class
[i
];
1830 if (class->index
!= i
)
1833 pages_to_free
+= zs_can_compact(class);
1836 return pages_to_free
;
1839 static void zs_unregister_shrinker(struct zs_pool
*pool
)
1841 if (pool
->shrinker_enabled
) {
1842 unregister_shrinker(&pool
->shrinker
);
1843 pool
->shrinker_enabled
= false;
1847 static int zs_register_shrinker(struct zs_pool
*pool
)
1849 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
1850 pool
->shrinker
.count_objects
= zs_shrinker_count
;
1851 pool
->shrinker
.batch
= 0;
1852 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
1854 return register_shrinker(&pool
->shrinker
);
1858 * zs_create_pool - Creates an allocation pool to work from.
1859 * @flags: allocation flags used to allocate pool metadata
1861 * This function must be called before anything when using
1862 * the zsmalloc allocator.
1864 * On success, a pointer to the newly created pool is returned,
1867 struct zs_pool
*zs_create_pool(const char *name
, gfp_t flags
)
1870 struct zs_pool
*pool
;
1871 struct size_class
*prev_class
= NULL
;
1873 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
1877 pool
->size_class
= kcalloc(zs_size_classes
, sizeof(struct size_class
*),
1879 if (!pool
->size_class
) {
1884 pool
->name
= kstrdup(name
, GFP_KERNEL
);
1888 if (create_handle_cache(pool
))
1892 * Iterate reversly, because, size of size_class that we want to use
1893 * for merging should be larger or equal to current size.
1895 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1897 int pages_per_zspage
;
1898 struct size_class
*class;
1900 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
1901 if (size
> ZS_MAX_ALLOC_SIZE
)
1902 size
= ZS_MAX_ALLOC_SIZE
;
1903 pages_per_zspage
= get_pages_per_zspage(size
);
1906 * size_class is used for normal zsmalloc operation such
1907 * as alloc/free for that size. Although it is natural that we
1908 * have one size_class for each size, there is a chance that we
1909 * can get more memory utilization if we use one size_class for
1910 * many different sizes whose size_class have same
1911 * characteristics. So, we makes size_class point to
1912 * previous size_class if possible.
1915 if (can_merge(prev_class
, size
, pages_per_zspage
)) {
1916 pool
->size_class
[i
] = prev_class
;
1921 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
1927 class->pages_per_zspage
= pages_per_zspage
;
1928 if (pages_per_zspage
== 1 &&
1929 get_maxobj_per_zspage(size
, pages_per_zspage
) == 1)
1931 spin_lock_init(&class->lock
);
1932 pool
->size_class
[i
] = class;
1937 pool
->flags
= flags
;
1939 if (zs_pool_stat_create(name
, pool
))
1943 * Not critical, we still can use the pool
1944 * and user can trigger compaction manually.
1946 if (zs_register_shrinker(pool
) == 0)
1947 pool
->shrinker_enabled
= true;
1951 zs_destroy_pool(pool
);
1954 EXPORT_SYMBOL_GPL(zs_create_pool
);
1956 void zs_destroy_pool(struct zs_pool
*pool
)
1960 zs_unregister_shrinker(pool
);
1961 zs_pool_stat_destroy(pool
);
1963 for (i
= 0; i
< zs_size_classes
; i
++) {
1965 struct size_class
*class = pool
->size_class
[i
];
1970 if (class->index
!= i
)
1973 for (fg
= 0; fg
< _ZS_NR_FULLNESS_GROUPS
; fg
++) {
1974 if (class->fullness_list
[fg
]) {
1975 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1982 destroy_handle_cache(pool
);
1983 kfree(pool
->size_class
);
1987 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
1989 static int __init
zs_init(void)
1991 int ret
= zs_register_cpu_notifier();
1996 init_zs_size_classes();
1999 zpool_register_driver(&zs_zpool_driver
);
2002 ret
= zs_stat_init();
2004 pr_err("zs stat initialization failed\n");
2011 zpool_unregister_driver(&zs_zpool_driver
);
2014 zs_unregister_cpu_notifier();
2019 static void __exit
zs_exit(void)
2022 zpool_unregister_driver(&zs_zpool_driver
);
2024 zs_unregister_cpu_notifier();
2029 module_init(zs_init
);
2030 module_exit(zs_exit
);
2032 MODULE_LICENSE("Dual BSD/GPL");
2033 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");