2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->private: points to 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: refers to the component page after the first page
30 * If the page is first_page for huge object, it stores handle.
31 * Look at size_class->huge.
32 * page->freelist: points to the first free object in zspage.
33 * Free objects are linked together using in-place
35 * page->objects: maximum number of objects we can store in this
36 * zspage (class->zspage_order * PAGE_SIZE / class->size)
37 * page->lru: links together first pages of various zspages.
38 * Basically forming list of zspages in a fullness group.
39 * page->mapping: class index and fullness group of the zspage
40 * page->inuse: the number of objects that are used in this 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/preempt.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
,
171 #ifdef CONFIG_ZSMALLOC_STAT
172 #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1)
174 #define NR_ZS_STAT_TYPE (OBJ_USED + 1)
177 struct zs_size_stat
{
178 unsigned long objs
[NR_ZS_STAT_TYPE
];
181 #ifdef CONFIG_ZSMALLOC_STAT
182 static struct dentry
*zs_stat_root
;
186 * number of size_classes
188 static int zs_size_classes
;
191 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
193 * n = number of allocated objects
194 * N = total number of objects zspage can store
195 * f = fullness_threshold_frac
197 * Similarly, we assign zspage to:
198 * ZS_ALMOST_FULL when n > N / f
199 * ZS_EMPTY when n == 0
200 * ZS_FULL when n == N
202 * (see: fix_fullness_group())
204 static const int fullness_threshold_frac
= 4;
208 struct page
*fullness_list
[_ZS_NR_FULLNESS_GROUPS
];
210 * Size of objects stored in this class. Must be multiple
216 struct zs_size_stat stats
;
218 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
219 int pages_per_zspage
;
220 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
225 * Placed within free objects to form a singly linked list.
226 * For every zspage, first_page->freelist gives head of this list.
228 * This must be power of 2 and less than or equal to ZS_ALIGN
233 * Position of next free chunk (encodes <PFN, obj_idx>)
234 * It's valid for non-allocated object
238 * Handle of allocated object.
240 unsigned long handle
;
247 struct size_class
**size_class
;
248 struct kmem_cache
*handle_cachep
;
250 gfp_t flags
; /* allocation flags used when growing pool */
251 atomic_long_t pages_allocated
;
253 struct zs_pool_stats stats
;
255 /* Compact classes */
256 struct shrinker shrinker
;
258 * To signify that register_shrinker() was successful
259 * and unregister_shrinker() will not Oops.
261 bool shrinker_enabled
;
262 #ifdef CONFIG_ZSMALLOC_STAT
263 struct dentry
*stat_dentry
;
268 * A zspage's class index and fullness group
269 * are encoded in its (first)page->mapping
271 #define CLASS_IDX_BITS 28
272 #define FULLNESS_BITS 4
273 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
274 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
276 struct mapping_area
{
277 #ifdef CONFIG_PGTABLE_MAPPING
278 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
280 char *vm_buf
; /* copy buffer for objects that span pages */
282 char *vm_addr
; /* address of kmap_atomic()'ed pages */
283 enum zs_mapmode vm_mm
; /* mapping mode */
286 static int create_handle_cache(struct zs_pool
*pool
)
288 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
290 return pool
->handle_cachep
? 0 : 1;
293 static void destroy_handle_cache(struct zs_pool
*pool
)
295 kmem_cache_destroy(pool
->handle_cachep
);
298 static unsigned long alloc_handle(struct zs_pool
*pool
)
300 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
301 pool
->flags
& ~__GFP_HIGHMEM
);
304 static void free_handle(struct zs_pool
*pool
, unsigned long handle
)
306 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
309 static void record_obj(unsigned long handle
, unsigned long obj
)
312 * lsb of @obj represents handle lock while other bits
313 * represent object value the handle is pointing so
314 * updating shouldn't do store tearing.
316 WRITE_ONCE(*(unsigned long *)handle
, obj
);
323 static void *zs_zpool_create(const char *name
, gfp_t gfp
,
324 const struct zpool_ops
*zpool_ops
,
327 return zs_create_pool(name
, gfp
);
330 static void zs_zpool_destroy(void *pool
)
332 zs_destroy_pool(pool
);
335 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
336 unsigned long *handle
)
338 *handle
= zs_malloc(pool
, size
);
339 return *handle
? 0 : -1;
341 static void zs_zpool_free(void *pool
, unsigned long handle
)
343 zs_free(pool
, handle
);
346 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
347 unsigned int *reclaimed
)
352 static void *zs_zpool_map(void *pool
, unsigned long handle
,
353 enum zpool_mapmode mm
)
355 enum zs_mapmode zs_mm
;
364 case ZPOOL_MM_RW
: /* fallthru */
370 return zs_map_object(pool
, handle
, zs_mm
);
372 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
374 zs_unmap_object(pool
, handle
);
377 static u64
zs_zpool_total_size(void *pool
)
379 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
382 static struct zpool_driver zs_zpool_driver
= {
384 .owner
= THIS_MODULE
,
385 .create
= zs_zpool_create
,
386 .destroy
= zs_zpool_destroy
,
387 .malloc
= zs_zpool_malloc
,
388 .free
= zs_zpool_free
,
389 .shrink
= zs_zpool_shrink
,
391 .unmap
= zs_zpool_unmap
,
392 .total_size
= zs_zpool_total_size
,
395 MODULE_ALIAS("zpool-zsmalloc");
396 #endif /* CONFIG_ZPOOL */
398 static unsigned int get_maxobj_per_zspage(int size
, int pages_per_zspage
)
400 return pages_per_zspage
* PAGE_SIZE
/ size
;
403 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
404 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
406 static int is_first_page(struct page
*page
)
408 return PagePrivate(page
);
411 static int is_last_page(struct page
*page
)
413 return PagePrivate2(page
);
416 static void get_zspage_mapping(struct page
*first_page
,
417 unsigned int *class_idx
,
418 enum fullness_group
*fullness
)
421 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
423 m
= (unsigned long)first_page
->mapping
;
424 *fullness
= m
& FULLNESS_MASK
;
425 *class_idx
= (m
>> FULLNESS_BITS
) & CLASS_IDX_MASK
;
428 static void set_zspage_mapping(struct page
*first_page
,
429 unsigned int class_idx
,
430 enum fullness_group fullness
)
433 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
435 m
= ((class_idx
& CLASS_IDX_MASK
) << FULLNESS_BITS
) |
436 (fullness
& FULLNESS_MASK
);
437 first_page
->mapping
= (struct address_space
*)m
;
441 * zsmalloc divides the pool into various size classes where each
442 * class maintains a list of zspages where each zspage is divided
443 * into equal sized chunks. Each allocation falls into one of these
444 * classes depending on its size. This function returns index of the
445 * size class which has chunk size big enough to hold the give size.
447 static int get_size_class_index(int size
)
451 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
452 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
453 ZS_SIZE_CLASS_DELTA
);
455 return min(zs_size_classes
- 1, idx
);
458 static inline void zs_stat_inc(struct size_class
*class,
459 enum zs_stat_type type
, unsigned long cnt
)
461 if (type
< NR_ZS_STAT_TYPE
)
462 class->stats
.objs
[type
] += cnt
;
465 static inline void zs_stat_dec(struct size_class
*class,
466 enum zs_stat_type type
, unsigned long cnt
)
468 if (type
< NR_ZS_STAT_TYPE
)
469 class->stats
.objs
[type
] -= cnt
;
472 static inline unsigned long zs_stat_get(struct size_class
*class,
473 enum zs_stat_type type
)
475 if (type
< NR_ZS_STAT_TYPE
)
476 return class->stats
.objs
[type
];
480 #ifdef CONFIG_ZSMALLOC_STAT
482 static int __init
zs_stat_init(void)
484 if (!debugfs_initialized())
487 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
494 static void __exit
zs_stat_exit(void)
496 debugfs_remove_recursive(zs_stat_root
);
499 static unsigned long zs_can_compact(struct size_class
*class);
501 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
504 struct zs_pool
*pool
= s
->private;
505 struct size_class
*class;
507 unsigned long class_almost_full
, class_almost_empty
;
508 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
509 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
510 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
511 unsigned long total_freeable
= 0;
513 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
514 "class", "size", "almost_full", "almost_empty",
515 "obj_allocated", "obj_used", "pages_used",
516 "pages_per_zspage", "freeable");
518 for (i
= 0; i
< zs_size_classes
; i
++) {
519 class = pool
->size_class
[i
];
521 if (class->index
!= i
)
524 spin_lock(&class->lock
);
525 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
526 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
527 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
528 obj_used
= zs_stat_get(class, OBJ_USED
);
529 freeable
= zs_can_compact(class);
530 spin_unlock(&class->lock
);
532 objs_per_zspage
= get_maxobj_per_zspage(class->size
,
533 class->pages_per_zspage
);
534 pages_used
= obj_allocated
/ objs_per_zspage
*
535 class->pages_per_zspage
;
537 seq_printf(s
, " %5u %5u %11lu %12lu %13lu"
538 " %10lu %10lu %16d %8lu\n",
539 i
, class->size
, class_almost_full
, class_almost_empty
,
540 obj_allocated
, obj_used
, pages_used
,
541 class->pages_per_zspage
, freeable
);
543 total_class_almost_full
+= class_almost_full
;
544 total_class_almost_empty
+= class_almost_empty
;
545 total_objs
+= obj_allocated
;
546 total_used_objs
+= obj_used
;
547 total_pages
+= pages_used
;
548 total_freeable
+= freeable
;
552 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
553 "Total", "", total_class_almost_full
,
554 total_class_almost_empty
, total_objs
,
555 total_used_objs
, total_pages
, "", total_freeable
);
560 static int zs_stats_size_open(struct inode
*inode
, struct file
*file
)
562 return single_open(file
, zs_stats_size_show
, inode
->i_private
);
565 static const struct file_operations zs_stat_size_ops
= {
566 .open
= zs_stats_size_open
,
569 .release
= single_release
,
572 static int zs_pool_stat_create(const char *name
, struct zs_pool
*pool
)
574 struct dentry
*entry
;
579 entry
= debugfs_create_dir(name
, zs_stat_root
);
581 pr_warn("debugfs dir <%s> creation failed\n", name
);
584 pool
->stat_dentry
= entry
;
586 entry
= debugfs_create_file("classes", S_IFREG
| S_IRUGO
,
587 pool
->stat_dentry
, pool
, &zs_stat_size_ops
);
589 pr_warn("%s: debugfs file entry <%s> creation failed\n",
597 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
599 debugfs_remove_recursive(pool
->stat_dentry
);
602 #else /* CONFIG_ZSMALLOC_STAT */
603 static int __init
zs_stat_init(void)
608 static void __exit
zs_stat_exit(void)
612 static inline int zs_pool_stat_create(const char *name
, struct zs_pool
*pool
)
617 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
624 * For each size class, zspages are divided into different groups
625 * depending on how "full" they are. This was done so that we could
626 * easily find empty or nearly empty zspages when we try to shrink
627 * the pool (not yet implemented). This function returns fullness
628 * status of the given page.
630 static enum fullness_group
get_fullness_group(struct page
*first_page
)
632 int inuse
, max_objects
;
633 enum fullness_group fg
;
635 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
637 inuse
= first_page
->inuse
;
638 max_objects
= first_page
->objects
;
642 else if (inuse
== max_objects
)
644 else if (inuse
<= 3 * max_objects
/ fullness_threshold_frac
)
645 fg
= ZS_ALMOST_EMPTY
;
653 * Each size class maintains various freelists and zspages are assigned
654 * to one of these freelists based on the number of live objects they
655 * have. This functions inserts the given zspage into the freelist
656 * identified by <class, fullness_group>.
658 static void insert_zspage(struct page
*first_page
, struct size_class
*class,
659 enum fullness_group fullness
)
663 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
665 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
668 zs_stat_inc(class, fullness
== ZS_ALMOST_EMPTY
?
669 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
671 head
= &class->fullness_list
[fullness
];
678 * We want to see more ZS_FULL pages and less almost
679 * empty/full. Put pages with higher ->inuse first.
681 list_add_tail(&first_page
->lru
, &(*head
)->lru
);
682 if (first_page
->inuse
>= (*head
)->inuse
)
687 * This function removes the given zspage from the freelist identified
688 * by <class, fullness_group>.
690 static void remove_zspage(struct page
*first_page
, struct size_class
*class,
691 enum fullness_group fullness
)
695 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
697 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
700 head
= &class->fullness_list
[fullness
];
701 VM_BUG_ON_PAGE(!*head
, first_page
);
702 if (list_empty(&(*head
)->lru
))
704 else if (*head
== first_page
)
705 *head
= (struct page
*)list_entry((*head
)->lru
.next
,
708 list_del_init(&first_page
->lru
);
709 zs_stat_dec(class, fullness
== ZS_ALMOST_EMPTY
?
710 CLASS_ALMOST_EMPTY
: CLASS_ALMOST_FULL
, 1);
714 * Each size class maintains zspages in different fullness groups depending
715 * on the number of live objects they contain. When allocating or freeing
716 * objects, the fullness status of the page can change, say, from ALMOST_FULL
717 * to ALMOST_EMPTY when freeing an object. This function checks if such
718 * a status change has occurred for the given page and accordingly moves the
719 * page from the freelist of the old fullness group to that of the new
722 static enum fullness_group
fix_fullness_group(struct size_class
*class,
723 struct page
*first_page
)
726 enum fullness_group currfg
, newfg
;
728 get_zspage_mapping(first_page
, &class_idx
, &currfg
);
729 newfg
= get_fullness_group(first_page
);
733 remove_zspage(first_page
, class, currfg
);
734 insert_zspage(first_page
, class, newfg
);
735 set_zspage_mapping(first_page
, class_idx
, newfg
);
742 * We have to decide on how many pages to link together
743 * to form a zspage for each size class. This is important
744 * to reduce wastage due to unusable space left at end of
745 * each zspage which is given as:
746 * wastage = Zp % class_size
747 * usage = Zp - wastage
748 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
750 * For example, for size class of 3/8 * PAGE_SIZE, we should
751 * link together 3 PAGE_SIZE sized pages to form a zspage
752 * since then we can perfectly fit in 8 such objects.
754 static int get_pages_per_zspage(int class_size
)
756 int i
, max_usedpc
= 0;
757 /* zspage order which gives maximum used size per KB */
758 int max_usedpc_order
= 1;
760 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
764 zspage_size
= i
* PAGE_SIZE
;
765 waste
= zspage_size
% class_size
;
766 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
768 if (usedpc
> max_usedpc
) {
770 max_usedpc_order
= i
;
774 return max_usedpc_order
;
778 * A single 'zspage' is composed of many system pages which are
779 * linked together using fields in struct page. This function finds
780 * the first/head page, given any component page of a zspage.
782 static struct page
*get_first_page(struct page
*page
)
784 if (is_first_page(page
))
787 return (struct page
*)page_private(page
);
790 static struct page
*get_next_page(struct page
*page
)
794 if (is_last_page(page
))
796 else if (is_first_page(page
))
797 next
= (struct page
*)page_private(page
);
799 next
= list_entry(page
->lru
.next
, struct page
, lru
);
805 * Encode <page, obj_idx> as a single handle value.
806 * We use the least bit of handle for tagging.
808 static void *location_to_obj(struct page
*page
, unsigned long obj_idx
)
817 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
818 obj
|= ((obj_idx
) & OBJ_INDEX_MASK
);
819 obj
<<= OBJ_TAG_BITS
;
825 * Decode <page, obj_idx> pair from the given object handle. We adjust the
826 * decoded obj_idx back to its original value since it was adjusted in
829 static void obj_to_location(unsigned long obj
, struct page
**page
,
830 unsigned long *obj_idx
)
832 obj
>>= OBJ_TAG_BITS
;
833 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
834 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
837 static unsigned long handle_to_obj(unsigned long handle
)
839 return *(unsigned long *)handle
;
842 static unsigned long obj_to_head(struct size_class
*class, struct page
*page
,
846 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
847 return page_private(page
);
849 return *(unsigned long *)obj
;
852 static unsigned long obj_idx_to_offset(struct page
*page
,
853 unsigned long obj_idx
, int class_size
)
855 unsigned long off
= 0;
857 if (!is_first_page(page
))
860 return off
+ obj_idx
* class_size
;
863 static inline int trypin_tag(unsigned long handle
)
865 unsigned long *ptr
= (unsigned long *)handle
;
867 return !test_and_set_bit_lock(HANDLE_PIN_BIT
, ptr
);
870 static void pin_tag(unsigned long handle
)
872 while (!trypin_tag(handle
));
875 static void unpin_tag(unsigned long handle
)
877 unsigned long *ptr
= (unsigned long *)handle
;
879 clear_bit_unlock(HANDLE_PIN_BIT
, ptr
);
882 static void reset_page(struct page
*page
)
884 clear_bit(PG_private
, &page
->flags
);
885 clear_bit(PG_private_2
, &page
->flags
);
886 set_page_private(page
, 0);
887 page
->mapping
= NULL
;
888 page
->freelist
= NULL
;
889 page_mapcount_reset(page
);
892 static void free_zspage(struct page
*first_page
)
894 struct page
*nextp
, *tmp
, *head_extra
;
896 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
897 VM_BUG_ON_PAGE(first_page
->inuse
, first_page
);
899 head_extra
= (struct page
*)page_private(first_page
);
901 reset_page(first_page
);
902 __free_page(first_page
);
904 /* zspage with only 1 system page */
908 list_for_each_entry_safe(nextp
, tmp
, &head_extra
->lru
, lru
) {
909 list_del(&nextp
->lru
);
913 reset_page(head_extra
);
914 __free_page(head_extra
);
917 /* Initialize a newly allocated zspage */
918 static void init_zspage(struct page
*first_page
, struct size_class
*class)
920 unsigned long off
= 0;
921 struct page
*page
= first_page
;
923 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
926 struct page
*next_page
;
927 struct link_free
*link
;
932 * page->index stores offset of first object starting
933 * in the page. For the first page, this is always 0,
934 * so we use first_page->index (aka ->freelist) to store
935 * head of corresponding zspage's freelist.
937 if (page
!= first_page
)
940 vaddr
= kmap_atomic(page
);
941 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
943 while ((off
+= class->size
) < PAGE_SIZE
) {
944 link
->next
= location_to_obj(page
, i
++);
945 link
+= class->size
/ sizeof(*link
);
949 * We now come to the last (full or partial) object on this
950 * page, which must point to the first object on the next
953 next_page
= get_next_page(page
);
954 link
->next
= location_to_obj(next_page
, 0);
955 kunmap_atomic(vaddr
);
962 * Allocate a zspage for the given size class
964 static struct page
*alloc_zspage(struct size_class
*class, gfp_t flags
)
967 struct page
*first_page
= NULL
, *uninitialized_var(prev_page
);
970 * Allocate individual pages and link them together as:
971 * 1. first page->private = first sub-page
972 * 2. all sub-pages are linked together using page->lru
973 * 3. each sub-page is linked to the first page using page->private
975 * For each size class, First/Head pages are linked together using
976 * page->lru. Also, we set PG_private to identify the first page
977 * (i.e. no other sub-page has this flag set) and PG_private_2 to
978 * identify the last page.
981 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
984 page
= alloc_page(flags
);
988 INIT_LIST_HEAD(&page
->lru
);
989 if (i
== 0) { /* first page */
990 SetPagePrivate(page
);
991 set_page_private(page
, 0);
993 first_page
->inuse
= 0;
996 set_page_private(first_page
, (unsigned long)page
);
998 set_page_private(page
, (unsigned long)first_page
);
1000 list_add(&page
->lru
, &prev_page
->lru
);
1001 if (i
== class->pages_per_zspage
- 1) /* last page */
1002 SetPagePrivate2(page
);
1006 init_zspage(first_page
, class);
1008 first_page
->freelist
= location_to_obj(first_page
, 0);
1009 /* Maximum number of objects we can store in this zspage */
1010 first_page
->objects
= class->pages_per_zspage
* PAGE_SIZE
/ class->size
;
1012 error
= 0; /* Success */
1015 if (unlikely(error
) && first_page
) {
1016 free_zspage(first_page
);
1023 static struct page
*find_get_zspage(struct size_class
*class)
1028 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1029 page
= class->fullness_list
[i
];
1037 #ifdef CONFIG_PGTABLE_MAPPING
1038 static inline int __zs_cpu_up(struct mapping_area
*area
)
1041 * Make sure we don't leak memory if a cpu UP notification
1042 * and zs_init() race and both call zs_cpu_up() on the same cpu
1046 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
1052 static inline void __zs_cpu_down(struct mapping_area
*area
)
1055 free_vm_area(area
->vm
);
1059 static inline void *__zs_map_object(struct mapping_area
*area
,
1060 struct page
*pages
[2], int off
, int size
)
1062 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
1063 area
->vm_addr
= area
->vm
->addr
;
1064 return area
->vm_addr
+ off
;
1067 static inline void __zs_unmap_object(struct mapping_area
*area
,
1068 struct page
*pages
[2], int off
, int size
)
1070 unsigned long addr
= (unsigned long)area
->vm_addr
;
1072 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
1075 #else /* CONFIG_PGTABLE_MAPPING */
1077 static inline int __zs_cpu_up(struct mapping_area
*area
)
1080 * Make sure we don't leak memory if a cpu UP notification
1081 * and zs_init() race and both call zs_cpu_up() on the same cpu
1085 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1091 static inline void __zs_cpu_down(struct mapping_area
*area
)
1093 kfree(area
->vm_buf
);
1094 area
->vm_buf
= NULL
;
1097 static void *__zs_map_object(struct mapping_area
*area
,
1098 struct page
*pages
[2], int off
, int size
)
1102 char *buf
= area
->vm_buf
;
1104 /* disable page faults to match kmap_atomic() return conditions */
1105 pagefault_disable();
1107 /* no read fastpath */
1108 if (area
->vm_mm
== ZS_MM_WO
)
1111 sizes
[0] = PAGE_SIZE
- off
;
1112 sizes
[1] = size
- sizes
[0];
1114 /* copy object to per-cpu buffer */
1115 addr
= kmap_atomic(pages
[0]);
1116 memcpy(buf
, addr
+ off
, sizes
[0]);
1117 kunmap_atomic(addr
);
1118 addr
= kmap_atomic(pages
[1]);
1119 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1120 kunmap_atomic(addr
);
1122 return area
->vm_buf
;
1125 static void __zs_unmap_object(struct mapping_area
*area
,
1126 struct page
*pages
[2], int off
, int size
)
1132 /* no write fastpath */
1133 if (area
->vm_mm
== ZS_MM_RO
)
1137 buf
= buf
+ ZS_HANDLE_SIZE
;
1138 size
-= ZS_HANDLE_SIZE
;
1139 off
+= ZS_HANDLE_SIZE
;
1141 sizes
[0] = PAGE_SIZE
- off
;
1142 sizes
[1] = size
- sizes
[0];
1144 /* copy per-cpu buffer to object */
1145 addr
= kmap_atomic(pages
[0]);
1146 memcpy(addr
+ off
, buf
, sizes
[0]);
1147 kunmap_atomic(addr
);
1148 addr
= kmap_atomic(pages
[1]);
1149 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1150 kunmap_atomic(addr
);
1153 /* enable page faults to match kunmap_atomic() return conditions */
1157 #endif /* CONFIG_PGTABLE_MAPPING */
1159 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
1162 int ret
, cpu
= (long)pcpu
;
1163 struct mapping_area
*area
;
1166 case CPU_UP_PREPARE
:
1167 area
= &per_cpu(zs_map_area
, cpu
);
1168 ret
= __zs_cpu_up(area
);
1170 return notifier_from_errno(ret
);
1173 case CPU_UP_CANCELED
:
1174 area
= &per_cpu(zs_map_area
, cpu
);
1175 __zs_cpu_down(area
);
1182 static struct notifier_block zs_cpu_nb
= {
1183 .notifier_call
= zs_cpu_notifier
1186 static int zs_register_cpu_notifier(void)
1188 int cpu
, uninitialized_var(ret
);
1190 cpu_notifier_register_begin();
1192 __register_cpu_notifier(&zs_cpu_nb
);
1193 for_each_online_cpu(cpu
) {
1194 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1195 if (notifier_to_errno(ret
))
1199 cpu_notifier_register_done();
1200 return notifier_to_errno(ret
);
1203 static void zs_unregister_cpu_notifier(void)
1207 cpu_notifier_register_begin();
1209 for_each_online_cpu(cpu
)
1210 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
1211 __unregister_cpu_notifier(&zs_cpu_nb
);
1213 cpu_notifier_register_done();
1216 static void init_zs_size_classes(void)
1220 nr
= (ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) / ZS_SIZE_CLASS_DELTA
+ 1;
1221 if ((ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) % ZS_SIZE_CLASS_DELTA
)
1224 zs_size_classes
= nr
;
1227 static bool can_merge(struct size_class
*prev
, int size
, int pages_per_zspage
)
1229 if (prev
->pages_per_zspage
!= pages_per_zspage
)
1232 if (get_maxobj_per_zspage(prev
->size
, prev
->pages_per_zspage
)
1233 != get_maxobj_per_zspage(size
, pages_per_zspage
))
1239 static bool zspage_full(struct page
*first_page
)
1241 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
1243 return first_page
->inuse
== first_page
->objects
;
1246 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1248 return atomic_long_read(&pool
->pages_allocated
);
1250 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1253 * zs_map_object - get address of allocated object from handle.
1254 * @pool: pool from which the object was allocated
1255 * @handle: handle returned from zs_malloc
1257 * Before using an object allocated from zs_malloc, it must be mapped using
1258 * this function. When done with the object, it must be unmapped using
1261 * Only one object can be mapped per cpu at a time. There is no protection
1262 * against nested mappings.
1264 * This function returns with preemption and page faults disabled.
1266 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1270 unsigned long obj
, obj_idx
, off
;
1272 unsigned int class_idx
;
1273 enum fullness_group fg
;
1274 struct size_class
*class;
1275 struct mapping_area
*area
;
1276 struct page
*pages
[2];
1280 * Because we use per-cpu mapping areas shared among the
1281 * pools/users, we can't allow mapping in interrupt context
1282 * because it can corrupt another users mappings.
1284 WARN_ON_ONCE(in_interrupt());
1286 /* From now on, migration cannot move the object */
1289 obj
= handle_to_obj(handle
);
1290 obj_to_location(obj
, &page
, &obj_idx
);
1291 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1292 class = pool
->size_class
[class_idx
];
1293 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1295 area
= &get_cpu_var(zs_map_area
);
1297 if (off
+ class->size
<= PAGE_SIZE
) {
1298 /* this object is contained entirely within a page */
1299 area
->vm_addr
= kmap_atomic(page
);
1300 ret
= area
->vm_addr
+ off
;
1304 /* this object spans two pages */
1306 pages
[1] = get_next_page(page
);
1309 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1312 ret
+= ZS_HANDLE_SIZE
;
1316 EXPORT_SYMBOL_GPL(zs_map_object
);
1318 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1321 unsigned long obj
, obj_idx
, off
;
1323 unsigned int class_idx
;
1324 enum fullness_group fg
;
1325 struct size_class
*class;
1326 struct mapping_area
*area
;
1328 obj
= handle_to_obj(handle
);
1329 obj_to_location(obj
, &page
, &obj_idx
);
1330 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1331 class = pool
->size_class
[class_idx
];
1332 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1334 area
= this_cpu_ptr(&zs_map_area
);
1335 if (off
+ class->size
<= PAGE_SIZE
)
1336 kunmap_atomic(area
->vm_addr
);
1338 struct page
*pages
[2];
1341 pages
[1] = get_next_page(page
);
1344 __zs_unmap_object(area
, pages
, off
, class->size
);
1346 put_cpu_var(zs_map_area
);
1349 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1351 static unsigned long obj_malloc(struct page
*first_page
,
1352 struct size_class
*class, unsigned long handle
)
1355 struct link_free
*link
;
1357 struct page
*m_page
;
1358 unsigned long m_objidx
, m_offset
;
1361 handle
|= OBJ_ALLOCATED_TAG
;
1362 obj
= (unsigned long)first_page
->freelist
;
1363 obj_to_location(obj
, &m_page
, &m_objidx
);
1364 m_offset
= obj_idx_to_offset(m_page
, m_objidx
, class->size
);
1366 vaddr
= kmap_atomic(m_page
);
1367 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1368 first_page
->freelist
= link
->next
;
1370 /* record handle in the header of allocated chunk */
1371 link
->handle
= handle
;
1373 /* record handle in first_page->private */
1374 set_page_private(first_page
, handle
);
1375 kunmap_atomic(vaddr
);
1376 first_page
->inuse
++;
1377 zs_stat_inc(class, OBJ_USED
, 1);
1384 * zs_malloc - Allocate block of given size from pool.
1385 * @pool: pool to allocate from
1386 * @size: size of block to allocate
1388 * On success, handle to the allocated object is returned,
1390 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1392 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
)
1394 unsigned long handle
, obj
;
1395 struct size_class
*class;
1396 struct page
*first_page
;
1398 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1401 handle
= alloc_handle(pool
);
1405 /* extra space in chunk to keep the handle */
1406 size
+= ZS_HANDLE_SIZE
;
1407 class = pool
->size_class
[get_size_class_index(size
)];
1409 spin_lock(&class->lock
);
1410 first_page
= find_get_zspage(class);
1413 spin_unlock(&class->lock
);
1414 first_page
= alloc_zspage(class, pool
->flags
);
1415 if (unlikely(!first_page
)) {
1416 free_handle(pool
, handle
);
1420 set_zspage_mapping(first_page
, class->index
, ZS_EMPTY
);
1421 atomic_long_add(class->pages_per_zspage
,
1422 &pool
->pages_allocated
);
1424 spin_lock(&class->lock
);
1425 zs_stat_inc(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1426 class->size
, class->pages_per_zspage
));
1429 obj
= obj_malloc(first_page
, class, handle
);
1430 /* Now move the zspage to another fullness group, if required */
1431 fix_fullness_group(class, first_page
);
1432 record_obj(handle
, obj
);
1433 spin_unlock(&class->lock
);
1437 EXPORT_SYMBOL_GPL(zs_malloc
);
1439 static void obj_free(struct zs_pool
*pool
, struct size_class
*class,
1442 struct link_free
*link
;
1443 struct page
*first_page
, *f_page
;
1444 unsigned long f_objidx
, f_offset
;
1447 obj
&= ~OBJ_ALLOCATED_TAG
;
1448 obj_to_location(obj
, &f_page
, &f_objidx
);
1449 first_page
= get_first_page(f_page
);
1451 f_offset
= obj_idx_to_offset(f_page
, f_objidx
, class->size
);
1453 vaddr
= kmap_atomic(f_page
);
1455 /* Insert this object in containing zspage's freelist */
1456 link
= (struct link_free
*)(vaddr
+ f_offset
);
1457 link
->next
= first_page
->freelist
;
1459 set_page_private(first_page
, 0);
1460 kunmap_atomic(vaddr
);
1461 first_page
->freelist
= (void *)obj
;
1462 first_page
->inuse
--;
1463 zs_stat_dec(class, OBJ_USED
, 1);
1466 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1468 struct page
*first_page
, *f_page
;
1469 unsigned long obj
, f_objidx
;
1471 struct size_class
*class;
1472 enum fullness_group fullness
;
1474 if (unlikely(!handle
))
1478 obj
= handle_to_obj(handle
);
1479 obj_to_location(obj
, &f_page
, &f_objidx
);
1480 first_page
= get_first_page(f_page
);
1482 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1483 class = pool
->size_class
[class_idx
];
1485 spin_lock(&class->lock
);
1486 obj_free(pool
, class, obj
);
1487 fullness
= fix_fullness_group(class, first_page
);
1488 if (fullness
== ZS_EMPTY
) {
1489 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1490 class->size
, class->pages_per_zspage
));
1491 atomic_long_sub(class->pages_per_zspage
,
1492 &pool
->pages_allocated
);
1493 free_zspage(first_page
);
1495 spin_unlock(&class->lock
);
1498 free_handle(pool
, handle
);
1500 EXPORT_SYMBOL_GPL(zs_free
);
1502 static void zs_object_copy(unsigned long dst
, unsigned long src
,
1503 struct size_class
*class)
1505 struct page
*s_page
, *d_page
;
1506 unsigned long s_objidx
, d_objidx
;
1507 unsigned long s_off
, d_off
;
1508 void *s_addr
, *d_addr
;
1509 int s_size
, d_size
, size
;
1512 s_size
= d_size
= class->size
;
1514 obj_to_location(src
, &s_page
, &s_objidx
);
1515 obj_to_location(dst
, &d_page
, &d_objidx
);
1517 s_off
= obj_idx_to_offset(s_page
, s_objidx
, class->size
);
1518 d_off
= obj_idx_to_offset(d_page
, d_objidx
, class->size
);
1520 if (s_off
+ class->size
> PAGE_SIZE
)
1521 s_size
= PAGE_SIZE
- s_off
;
1523 if (d_off
+ class->size
> PAGE_SIZE
)
1524 d_size
= PAGE_SIZE
- d_off
;
1526 s_addr
= kmap_atomic(s_page
);
1527 d_addr
= kmap_atomic(d_page
);
1530 size
= min(s_size
, d_size
);
1531 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1534 if (written
== class->size
)
1542 if (s_off
>= PAGE_SIZE
) {
1543 kunmap_atomic(d_addr
);
1544 kunmap_atomic(s_addr
);
1545 s_page
= get_next_page(s_page
);
1546 s_addr
= kmap_atomic(s_page
);
1547 d_addr
= kmap_atomic(d_page
);
1548 s_size
= class->size
- written
;
1552 if (d_off
>= PAGE_SIZE
) {
1553 kunmap_atomic(d_addr
);
1554 d_page
= get_next_page(d_page
);
1555 d_addr
= kmap_atomic(d_page
);
1556 d_size
= class->size
- written
;
1561 kunmap_atomic(d_addr
);
1562 kunmap_atomic(s_addr
);
1566 * Find alloced object in zspage from index object and
1569 static unsigned long find_alloced_obj(struct page
*page
, int index
,
1570 struct size_class
*class)
1574 unsigned long handle
= 0;
1575 void *addr
= kmap_atomic(page
);
1577 if (!is_first_page(page
))
1578 offset
= page
->index
;
1579 offset
+= class->size
* index
;
1581 while (offset
< PAGE_SIZE
) {
1582 head
= obj_to_head(class, page
, addr
+ offset
);
1583 if (head
& OBJ_ALLOCATED_TAG
) {
1584 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1585 if (trypin_tag(handle
))
1590 offset
+= class->size
;
1594 kunmap_atomic(addr
);
1598 struct zs_compact_control
{
1599 /* Source page for migration which could be a subpage of zspage. */
1600 struct page
*s_page
;
1601 /* Destination page for migration which should be a first page
1603 struct page
*d_page
;
1604 /* Starting object index within @s_page which used for live object
1605 * in the subpage. */
1609 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1610 struct zs_compact_control
*cc
)
1612 unsigned long used_obj
, free_obj
;
1613 unsigned long handle
;
1614 struct page
*s_page
= cc
->s_page
;
1615 struct page
*d_page
= cc
->d_page
;
1616 unsigned long index
= cc
->index
;
1620 handle
= find_alloced_obj(s_page
, index
, class);
1622 s_page
= get_next_page(s_page
);
1629 /* Stop if there is no more space */
1630 if (zspage_full(d_page
)) {
1636 used_obj
= handle_to_obj(handle
);
1637 free_obj
= obj_malloc(d_page
, class, handle
);
1638 zs_object_copy(free_obj
, used_obj
, class);
1641 * record_obj updates handle's value to free_obj and it will
1642 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1643 * breaks synchronization using pin_tag(e,g, zs_free) so
1644 * let's keep the lock bit.
1646 free_obj
|= BIT(HANDLE_PIN_BIT
);
1647 record_obj(handle
, free_obj
);
1649 obj_free(pool
, class, used_obj
);
1652 /* Remember last position in this iteration */
1653 cc
->s_page
= s_page
;
1659 static struct page
*isolate_target_page(struct size_class
*class)
1664 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
1665 page
= class->fullness_list
[i
];
1667 remove_zspage(page
, class, i
);
1676 * putback_zspage - add @first_page into right class's fullness list
1677 * @pool: target pool
1678 * @class: destination class
1679 * @first_page: target page
1681 * Return @fist_page's fullness_group
1683 static enum fullness_group
putback_zspage(struct zs_pool
*pool
,
1684 struct size_class
*class,
1685 struct page
*first_page
)
1687 enum fullness_group fullness
;
1689 fullness
= get_fullness_group(first_page
);
1690 insert_zspage(first_page
, class, fullness
);
1691 set_zspage_mapping(first_page
, class->index
, fullness
);
1693 if (fullness
== ZS_EMPTY
) {
1694 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1695 class->size
, class->pages_per_zspage
));
1696 atomic_long_sub(class->pages_per_zspage
,
1697 &pool
->pages_allocated
);
1699 free_zspage(first_page
);
1705 static struct page
*isolate_source_page(struct size_class
*class)
1708 struct page
*page
= NULL
;
1710 for (i
= ZS_ALMOST_EMPTY
; i
>= ZS_ALMOST_FULL
; i
--) {
1711 page
= class->fullness_list
[i
];
1715 remove_zspage(page
, class, i
);
1724 * Based on the number of unused allocated objects calculate
1725 * and return the number of pages that we can free.
1727 static unsigned long zs_can_compact(struct size_class
*class)
1729 unsigned long obj_wasted
;
1730 unsigned long obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
1731 unsigned long obj_used
= zs_stat_get(class, OBJ_USED
);
1733 if (obj_allocated
<= obj_used
)
1736 obj_wasted
= obj_allocated
- obj_used
;
1737 obj_wasted
/= get_maxobj_per_zspage(class->size
,
1738 class->pages_per_zspage
);
1740 return obj_wasted
* class->pages_per_zspage
;
1743 static void __zs_compact(struct zs_pool
*pool
, struct size_class
*class)
1745 struct zs_compact_control cc
;
1746 struct page
*src_page
;
1747 struct page
*dst_page
= NULL
;
1749 spin_lock(&class->lock
);
1750 while ((src_page
= isolate_source_page(class))) {
1752 if (!zs_can_compact(class))
1756 cc
.s_page
= src_page
;
1758 while ((dst_page
= isolate_target_page(class))) {
1759 cc
.d_page
= dst_page
;
1761 * If there is no more space in dst_page, resched
1762 * and see if anyone had allocated another zspage.
1764 if (!migrate_zspage(pool
, class, &cc
))
1767 putback_zspage(pool
, class, dst_page
);
1770 /* Stop if we couldn't find slot */
1771 if (dst_page
== NULL
)
1774 putback_zspage(pool
, class, dst_page
);
1775 if (putback_zspage(pool
, class, src_page
) == ZS_EMPTY
)
1776 pool
->stats
.pages_compacted
+= class->pages_per_zspage
;
1777 spin_unlock(&class->lock
);
1779 spin_lock(&class->lock
);
1783 putback_zspage(pool
, class, src_page
);
1785 spin_unlock(&class->lock
);
1788 unsigned long zs_compact(struct zs_pool
*pool
)
1791 struct size_class
*class;
1793 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1794 class = pool
->size_class
[i
];
1797 if (class->index
!= i
)
1799 __zs_compact(pool
, class);
1802 return pool
->stats
.pages_compacted
;
1804 EXPORT_SYMBOL_GPL(zs_compact
);
1806 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
1808 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
1810 EXPORT_SYMBOL_GPL(zs_pool_stats
);
1812 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
1813 struct shrink_control
*sc
)
1815 unsigned long pages_freed
;
1816 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
1819 pages_freed
= pool
->stats
.pages_compacted
;
1821 * Compact classes and calculate compaction delta.
1822 * Can run concurrently with a manually triggered
1823 * (by user) compaction.
1825 pages_freed
= zs_compact(pool
) - pages_freed
;
1827 return pages_freed
? pages_freed
: SHRINK_STOP
;
1830 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
1831 struct shrink_control
*sc
)
1834 struct size_class
*class;
1835 unsigned long pages_to_free
= 0;
1836 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
1839 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1840 class = pool
->size_class
[i
];
1843 if (class->index
!= i
)
1846 pages_to_free
+= zs_can_compact(class);
1849 return pages_to_free
;
1852 static void zs_unregister_shrinker(struct zs_pool
*pool
)
1854 if (pool
->shrinker_enabled
) {
1855 unregister_shrinker(&pool
->shrinker
);
1856 pool
->shrinker_enabled
= false;
1860 static int zs_register_shrinker(struct zs_pool
*pool
)
1862 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
1863 pool
->shrinker
.count_objects
= zs_shrinker_count
;
1864 pool
->shrinker
.batch
= 0;
1865 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
1867 return register_shrinker(&pool
->shrinker
);
1871 * zs_create_pool - Creates an allocation pool to work from.
1872 * @flags: allocation flags used to allocate pool metadata
1874 * This function must be called before anything when using
1875 * the zsmalloc allocator.
1877 * On success, a pointer to the newly created pool is returned,
1880 struct zs_pool
*zs_create_pool(const char *name
, gfp_t flags
)
1883 struct zs_pool
*pool
;
1884 struct size_class
*prev_class
= NULL
;
1886 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
1890 pool
->size_class
= kcalloc(zs_size_classes
, sizeof(struct size_class
*),
1892 if (!pool
->size_class
) {
1897 pool
->name
= kstrdup(name
, GFP_KERNEL
);
1901 if (create_handle_cache(pool
))
1905 * Iterate reversly, because, size of size_class that we want to use
1906 * for merging should be larger or equal to current size.
1908 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
1910 int pages_per_zspage
;
1911 struct size_class
*class;
1913 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
1914 if (size
> ZS_MAX_ALLOC_SIZE
)
1915 size
= ZS_MAX_ALLOC_SIZE
;
1916 pages_per_zspage
= get_pages_per_zspage(size
);
1919 * size_class is used for normal zsmalloc operation such
1920 * as alloc/free for that size. Although it is natural that we
1921 * have one size_class for each size, there is a chance that we
1922 * can get more memory utilization if we use one size_class for
1923 * many different sizes whose size_class have same
1924 * characteristics. So, we makes size_class point to
1925 * previous size_class if possible.
1928 if (can_merge(prev_class
, size
, pages_per_zspage
)) {
1929 pool
->size_class
[i
] = prev_class
;
1934 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
1940 class->pages_per_zspage
= pages_per_zspage
;
1941 if (pages_per_zspage
== 1 &&
1942 get_maxobj_per_zspage(size
, pages_per_zspage
) == 1)
1944 spin_lock_init(&class->lock
);
1945 pool
->size_class
[i
] = class;
1950 pool
->flags
= flags
;
1952 if (zs_pool_stat_create(name
, pool
))
1956 * Not critical, we still can use the pool
1957 * and user can trigger compaction manually.
1959 if (zs_register_shrinker(pool
) == 0)
1960 pool
->shrinker_enabled
= true;
1964 zs_destroy_pool(pool
);
1967 EXPORT_SYMBOL_GPL(zs_create_pool
);
1969 void zs_destroy_pool(struct zs_pool
*pool
)
1973 zs_unregister_shrinker(pool
);
1974 zs_pool_stat_destroy(pool
);
1976 for (i
= 0; i
< zs_size_classes
; i
++) {
1978 struct size_class
*class = pool
->size_class
[i
];
1983 if (class->index
!= i
)
1986 for (fg
= 0; fg
< _ZS_NR_FULLNESS_GROUPS
; fg
++) {
1987 if (class->fullness_list
[fg
]) {
1988 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1995 destroy_handle_cache(pool
);
1996 kfree(pool
->size_class
);
2000 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2002 static int __init
zs_init(void)
2004 int ret
= zs_register_cpu_notifier();
2009 init_zs_size_classes();
2012 zpool_register_driver(&zs_zpool_driver
);
2015 ret
= zs_stat_init();
2017 pr_err("zs stat initialization failed\n");
2024 zpool_unregister_driver(&zs_zpool_driver
);
2027 zs_unregister_cpu_notifier();
2032 static void __exit
zs_exit(void)
2035 zpool_unregister_driver(&zs_zpool_driver
);
2037 zs_unregister_cpu_notifier();
2042 module_init(zs_init
);
2043 module_exit(zs_exit
);
2045 MODULE_LICENSE("Dual BSD/GPL");
2046 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");