2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->freelist(index): links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
23 * page->units: first object offset in a subpage of zspage
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_owner_priv_1: identifies the huge component page
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33 #include <linux/module.h>
34 #include <linux/kernel.h>
35 #include <linux/sched.h>
36 #include <linux/magic.h>
37 #include <linux/bitops.h>
38 #include <linux/errno.h>
39 #include <linux/highmem.h>
40 #include <linux/string.h>
41 #include <linux/slab.h>
42 #include <linux/pgtable.h>
43 #include <asm/tlbflush.h>
44 #include <linux/cpumask.h>
45 #include <linux/cpu.h>
46 #include <linux/vmalloc.h>
47 #include <linux/preempt.h>
48 #include <linux/spinlock.h>
49 #include <linux/shrinker.h>
50 #include <linux/types.h>
51 #include <linux/debugfs.h>
52 #include <linux/zsmalloc.h>
53 #include <linux/zpool.h>
54 #include <linux/mount.h>
55 #include <linux/pseudo_fs.h>
56 #include <linux/migrate.h>
57 #include <linux/wait.h>
58 #include <linux/pagemap.h>
61 #define ZSPAGE_MAGIC 0x58
64 * This must be power of 2 and greater than of equal to sizeof(link_free).
65 * These two conditions ensure that any 'struct link_free' itself doesn't
66 * span more than 1 page which avoids complex case of mapping 2 pages simply
67 * to restore link_free pointer values.
72 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
73 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
75 #define ZS_MAX_ZSPAGE_ORDER 2
76 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
78 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
81 * Object location (<PFN>, <obj_idx>) is encoded as
82 * a single (unsigned long) handle value.
84 * Note that object index <obj_idx> starts from 0.
86 * This is made more complicated by various memory models and PAE.
89 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
90 #ifdef MAX_PHYSMEM_BITS
91 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
94 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
97 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
101 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
104 * Memory for allocating for handle keeps object position by
105 * encoding <page, obj_idx> and the encoded value has a room
106 * in least bit(ie, look at obj_to_location).
107 * We use the bit to synchronize between object access by
108 * user and migration.
110 #define HANDLE_PIN_BIT 0
113 * Head in allocated object should have OBJ_ALLOCATED_TAG
114 * to identify the object was allocated or not.
115 * It's okay to add the status bit in the least bit because
116 * header keeps handle which is 4byte-aligned address so we
117 * have room for two bit at least.
119 #define OBJ_ALLOCATED_TAG 1
120 #define OBJ_TAG_BITS 1
121 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
122 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
124 #define FULLNESS_BITS 2
126 #define ISOLATED_BITS 3
127 #define MAGIC_VAL_BITS 8
129 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
130 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
131 #define ZS_MIN_ALLOC_SIZE \
132 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
133 /* each chunk includes extra space to keep handle */
134 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
137 * On systems with 4K page size, this gives 255 size classes! There is a
139 * - Large number of size classes is potentially wasteful as free page are
140 * spread across these classes
141 * - Small number of size classes causes large internal fragmentation
142 * - Probably its better to use specific size classes (empirically
143 * determined). NOTE: all those class sizes must be set as multiple of
144 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
146 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
149 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
150 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
151 ZS_SIZE_CLASS_DELTA) + 1)
153 enum fullness_group
{
171 struct zs_size_stat
{
172 unsigned long objs
[NR_ZS_STAT_TYPE
];
175 #ifdef CONFIG_ZSMALLOC_STAT
176 static struct dentry
*zs_stat_root
;
179 #ifdef CONFIG_COMPACTION
180 static struct vfsmount
*zsmalloc_mnt
;
184 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
186 * n = number of allocated objects
187 * N = total number of objects zspage can store
188 * f = fullness_threshold_frac
190 * Similarly, we assign zspage to:
191 * ZS_ALMOST_FULL when n > N / f
192 * ZS_EMPTY when n == 0
193 * ZS_FULL when n == N
195 * (see: fix_fullness_group())
197 static const int fullness_threshold_frac
= 4;
198 static size_t huge_class_size
;
202 struct list_head fullness_list
[NR_ZS_FULLNESS
];
204 * Size of objects stored in this class. Must be multiple
209 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
210 int pages_per_zspage
;
213 struct zs_size_stat stats
;
216 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
217 static void SetPageHugeObject(struct page
*page
)
219 SetPageOwnerPriv1(page
);
222 static void ClearPageHugeObject(struct page
*page
)
224 ClearPageOwnerPriv1(page
);
227 static int PageHugeObject(struct page
*page
)
229 return PageOwnerPriv1(page
);
233 * Placed within free objects to form a singly linked list.
234 * For every zspage, zspage->freeobj gives head of this list.
236 * This must be power of 2 and less than or equal to ZS_ALIGN
242 * It's valid for non-allocated object
246 * Handle of allocated object.
248 unsigned long handle
;
255 struct size_class
*size_class
[ZS_SIZE_CLASSES
];
256 struct kmem_cache
*handle_cachep
;
257 struct kmem_cache
*zspage_cachep
;
259 atomic_long_t pages_allocated
;
261 struct zs_pool_stats stats
;
263 /* Compact classes */
264 struct shrinker shrinker
;
266 #ifdef CONFIG_ZSMALLOC_STAT
267 struct dentry
*stat_dentry
;
269 #ifdef CONFIG_COMPACTION
271 struct work_struct free_work
;
272 /* A wait queue for when migration races with async_free_zspage() */
273 struct wait_queue_head migration_wait
;
274 atomic_long_t isolated_pages
;
281 unsigned int fullness
:FULLNESS_BITS
;
282 unsigned int class:CLASS_BITS
+ 1;
283 unsigned int isolated
:ISOLATED_BITS
;
284 unsigned int magic
:MAGIC_VAL_BITS
;
287 unsigned int freeobj
;
288 struct page
*first_page
;
289 struct list_head list
; /* fullness list */
290 #ifdef CONFIG_COMPACTION
295 struct mapping_area
{
296 char *vm_buf
; /* copy buffer for objects that span pages */
297 char *vm_addr
; /* address of kmap_atomic()'ed pages */
298 enum zs_mapmode vm_mm
; /* mapping mode */
301 #ifdef CONFIG_COMPACTION
302 static int zs_register_migration(struct zs_pool
*pool
);
303 static void zs_unregister_migration(struct zs_pool
*pool
);
304 static void migrate_lock_init(struct zspage
*zspage
);
305 static void migrate_read_lock(struct zspage
*zspage
);
306 static void migrate_read_unlock(struct zspage
*zspage
);
307 static void kick_deferred_free(struct zs_pool
*pool
);
308 static void init_deferred_free(struct zs_pool
*pool
);
309 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
);
311 static int zsmalloc_mount(void) { return 0; }
312 static void zsmalloc_unmount(void) {}
313 static int zs_register_migration(struct zs_pool
*pool
) { return 0; }
314 static void zs_unregister_migration(struct zs_pool
*pool
) {}
315 static void migrate_lock_init(struct zspage
*zspage
) {}
316 static void migrate_read_lock(struct zspage
*zspage
) {}
317 static void migrate_read_unlock(struct zspage
*zspage
) {}
318 static void kick_deferred_free(struct zs_pool
*pool
) {}
319 static void init_deferred_free(struct zs_pool
*pool
) {}
320 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
) {}
323 static int create_cache(struct zs_pool
*pool
)
325 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
327 if (!pool
->handle_cachep
)
330 pool
->zspage_cachep
= kmem_cache_create("zspage", sizeof(struct zspage
),
332 if (!pool
->zspage_cachep
) {
333 kmem_cache_destroy(pool
->handle_cachep
);
334 pool
->handle_cachep
= NULL
;
341 static void destroy_cache(struct zs_pool
*pool
)
343 kmem_cache_destroy(pool
->handle_cachep
);
344 kmem_cache_destroy(pool
->zspage_cachep
);
347 static unsigned long cache_alloc_handle(struct zs_pool
*pool
, gfp_t gfp
)
349 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
350 gfp
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
353 static void cache_free_handle(struct zs_pool
*pool
, unsigned long handle
)
355 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
358 static struct zspage
*cache_alloc_zspage(struct zs_pool
*pool
, gfp_t flags
)
360 return kmem_cache_alloc(pool
->zspage_cachep
,
361 flags
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
364 static void cache_free_zspage(struct zs_pool
*pool
, struct zspage
*zspage
)
366 kmem_cache_free(pool
->zspage_cachep
, zspage
);
369 static void record_obj(unsigned long handle
, unsigned long obj
)
372 * lsb of @obj represents handle lock while other bits
373 * represent object value the handle is pointing so
374 * updating shouldn't do store tearing.
376 WRITE_ONCE(*(unsigned long *)handle
, obj
);
383 static void *zs_zpool_create(const char *name
, gfp_t gfp
,
384 const struct zpool_ops
*zpool_ops
,
388 * Ignore global gfp flags: zs_malloc() may be invoked from
389 * different contexts and its caller must provide a valid
392 return zs_create_pool(name
);
395 static void zs_zpool_destroy(void *pool
)
397 zs_destroy_pool(pool
);
400 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
401 unsigned long *handle
)
403 *handle
= zs_malloc(pool
, size
, gfp
);
404 return *handle
? 0 : -1;
406 static void zs_zpool_free(void *pool
, unsigned long handle
)
408 zs_free(pool
, handle
);
411 static void *zs_zpool_map(void *pool
, unsigned long handle
,
412 enum zpool_mapmode mm
)
414 enum zs_mapmode zs_mm
;
429 return zs_map_object(pool
, handle
, zs_mm
);
431 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
433 zs_unmap_object(pool
, handle
);
436 static u64
zs_zpool_total_size(void *pool
)
438 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
441 static struct zpool_driver zs_zpool_driver
= {
443 .owner
= THIS_MODULE
,
444 .create
= zs_zpool_create
,
445 .destroy
= zs_zpool_destroy
,
446 .malloc_support_movable
= true,
447 .malloc
= zs_zpool_malloc
,
448 .free
= zs_zpool_free
,
450 .unmap
= zs_zpool_unmap
,
451 .total_size
= zs_zpool_total_size
,
454 MODULE_ALIAS("zpool-zsmalloc");
455 #endif /* CONFIG_ZPOOL */
457 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
458 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
460 static bool is_zspage_isolated(struct zspage
*zspage
)
462 return zspage
->isolated
;
465 static __maybe_unused
int is_first_page(struct page
*page
)
467 return PagePrivate(page
);
470 /* Protected by class->lock */
471 static inline int get_zspage_inuse(struct zspage
*zspage
)
473 return zspage
->inuse
;
477 static inline void mod_zspage_inuse(struct zspage
*zspage
, int val
)
479 zspage
->inuse
+= val
;
482 static inline struct page
*get_first_page(struct zspage
*zspage
)
484 struct page
*first_page
= zspage
->first_page
;
486 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
490 static inline int get_first_obj_offset(struct page
*page
)
495 static inline void set_first_obj_offset(struct page
*page
, int offset
)
497 page
->units
= offset
;
500 static inline unsigned int get_freeobj(struct zspage
*zspage
)
502 return zspage
->freeobj
;
505 static inline void set_freeobj(struct zspage
*zspage
, unsigned int obj
)
507 zspage
->freeobj
= obj
;
510 static void get_zspage_mapping(struct zspage
*zspage
,
511 unsigned int *class_idx
,
512 enum fullness_group
*fullness
)
514 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
516 *fullness
= zspage
->fullness
;
517 *class_idx
= zspage
->class;
520 static void set_zspage_mapping(struct zspage
*zspage
,
521 unsigned int class_idx
,
522 enum fullness_group fullness
)
524 zspage
->class = class_idx
;
525 zspage
->fullness
= fullness
;
529 * zsmalloc divides the pool into various size classes where each
530 * class maintains a list of zspages where each zspage is divided
531 * into equal sized chunks. Each allocation falls into one of these
532 * classes depending on its size. This function returns index of the
533 * size class which has chunk size big enough to hold the give size.
535 static int get_size_class_index(int size
)
539 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
540 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
541 ZS_SIZE_CLASS_DELTA
);
543 return min_t(int, ZS_SIZE_CLASSES
- 1, idx
);
546 /* type can be of enum type zs_stat_type or fullness_group */
547 static inline void zs_stat_inc(struct size_class
*class,
548 int type
, unsigned long cnt
)
550 class->stats
.objs
[type
] += cnt
;
553 /* type can be of enum type zs_stat_type or fullness_group */
554 static inline void zs_stat_dec(struct size_class
*class,
555 int type
, unsigned long cnt
)
557 class->stats
.objs
[type
] -= cnt
;
560 /* type can be of enum type zs_stat_type or fullness_group */
561 static inline unsigned long zs_stat_get(struct size_class
*class,
564 return class->stats
.objs
[type
];
567 #ifdef CONFIG_ZSMALLOC_STAT
569 static void __init
zs_stat_init(void)
571 if (!debugfs_initialized()) {
572 pr_warn("debugfs not available, stat dir not created\n");
576 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
579 static void __exit
zs_stat_exit(void)
581 debugfs_remove_recursive(zs_stat_root
);
584 static unsigned long zs_can_compact(struct size_class
*class);
586 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
589 struct zs_pool
*pool
= s
->private;
590 struct size_class
*class;
592 unsigned long class_almost_full
, class_almost_empty
;
593 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
594 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
595 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
596 unsigned long total_freeable
= 0;
598 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
599 "class", "size", "almost_full", "almost_empty",
600 "obj_allocated", "obj_used", "pages_used",
601 "pages_per_zspage", "freeable");
603 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
604 class = pool
->size_class
[i
];
606 if (class->index
!= i
)
609 spin_lock(&class->lock
);
610 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
611 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
612 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
613 obj_used
= zs_stat_get(class, OBJ_USED
);
614 freeable
= zs_can_compact(class);
615 spin_unlock(&class->lock
);
617 objs_per_zspage
= class->objs_per_zspage
;
618 pages_used
= obj_allocated
/ objs_per_zspage
*
619 class->pages_per_zspage
;
621 seq_printf(s
, " %5u %5u %11lu %12lu %13lu"
622 " %10lu %10lu %16d %8lu\n",
623 i
, class->size
, class_almost_full
, class_almost_empty
,
624 obj_allocated
, obj_used
, pages_used
,
625 class->pages_per_zspage
, freeable
);
627 total_class_almost_full
+= class_almost_full
;
628 total_class_almost_empty
+= class_almost_empty
;
629 total_objs
+= obj_allocated
;
630 total_used_objs
+= obj_used
;
631 total_pages
+= pages_used
;
632 total_freeable
+= freeable
;
636 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
637 "Total", "", total_class_almost_full
,
638 total_class_almost_empty
, total_objs
,
639 total_used_objs
, total_pages
, "", total_freeable
);
643 DEFINE_SHOW_ATTRIBUTE(zs_stats_size
);
645 static void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
648 pr_warn("no root stat dir, not creating <%s> stat dir\n", name
);
652 pool
->stat_dentry
= debugfs_create_dir(name
, zs_stat_root
);
654 debugfs_create_file("classes", S_IFREG
| 0444, pool
->stat_dentry
, pool
,
655 &zs_stats_size_fops
);
658 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
660 debugfs_remove_recursive(pool
->stat_dentry
);
663 #else /* CONFIG_ZSMALLOC_STAT */
664 static void __init
zs_stat_init(void)
668 static void __exit
zs_stat_exit(void)
672 static inline void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
676 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
683 * For each size class, zspages are divided into different groups
684 * depending on how "full" they are. This was done so that we could
685 * easily find empty or nearly empty zspages when we try to shrink
686 * the pool (not yet implemented). This function returns fullness
687 * status of the given page.
689 static enum fullness_group
get_fullness_group(struct size_class
*class,
690 struct zspage
*zspage
)
692 int inuse
, objs_per_zspage
;
693 enum fullness_group fg
;
695 inuse
= get_zspage_inuse(zspage
);
696 objs_per_zspage
= class->objs_per_zspage
;
700 else if (inuse
== objs_per_zspage
)
702 else if (inuse
<= 3 * objs_per_zspage
/ fullness_threshold_frac
)
703 fg
= ZS_ALMOST_EMPTY
;
711 * Each size class maintains various freelists and zspages are assigned
712 * to one of these freelists based on the number of live objects they
713 * have. This functions inserts the given zspage into the freelist
714 * identified by <class, fullness_group>.
716 static void insert_zspage(struct size_class
*class,
717 struct zspage
*zspage
,
718 enum fullness_group fullness
)
722 zs_stat_inc(class, fullness
, 1);
723 head
= list_first_entry_or_null(&class->fullness_list
[fullness
],
724 struct zspage
, list
);
726 * We want to see more ZS_FULL pages and less almost empty/full.
727 * Put pages with higher ->inuse first.
730 if (get_zspage_inuse(zspage
) < get_zspage_inuse(head
)) {
731 list_add(&zspage
->list
, &head
->list
);
735 list_add(&zspage
->list
, &class->fullness_list
[fullness
]);
739 * This function removes the given zspage from the freelist identified
740 * by <class, fullness_group>.
742 static void remove_zspage(struct size_class
*class,
743 struct zspage
*zspage
,
744 enum fullness_group fullness
)
746 VM_BUG_ON(list_empty(&class->fullness_list
[fullness
]));
747 VM_BUG_ON(is_zspage_isolated(zspage
));
749 list_del_init(&zspage
->list
);
750 zs_stat_dec(class, fullness
, 1);
754 * Each size class maintains zspages in different fullness groups depending
755 * on the number of live objects they contain. When allocating or freeing
756 * objects, the fullness status of the page can change, say, from ALMOST_FULL
757 * to ALMOST_EMPTY when freeing an object. This function checks if such
758 * a status change has occurred for the given page and accordingly moves the
759 * page from the freelist of the old fullness group to that of the new
762 static enum fullness_group
fix_fullness_group(struct size_class
*class,
763 struct zspage
*zspage
)
766 enum fullness_group currfg
, newfg
;
768 get_zspage_mapping(zspage
, &class_idx
, &currfg
);
769 newfg
= get_fullness_group(class, zspage
);
773 if (!is_zspage_isolated(zspage
)) {
774 remove_zspage(class, zspage
, currfg
);
775 insert_zspage(class, zspage
, newfg
);
778 set_zspage_mapping(zspage
, class_idx
, newfg
);
785 * We have to decide on how many pages to link together
786 * to form a zspage for each size class. This is important
787 * to reduce wastage due to unusable space left at end of
788 * each zspage which is given as:
789 * wastage = Zp % class_size
790 * usage = Zp - wastage
791 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
793 * For example, for size class of 3/8 * PAGE_SIZE, we should
794 * link together 3 PAGE_SIZE sized pages to form a zspage
795 * since then we can perfectly fit in 8 such objects.
797 static int get_pages_per_zspage(int class_size
)
799 int i
, max_usedpc
= 0;
800 /* zspage order which gives maximum used size per KB */
801 int max_usedpc_order
= 1;
803 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
807 zspage_size
= i
* PAGE_SIZE
;
808 waste
= zspage_size
% class_size
;
809 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
811 if (usedpc
> max_usedpc
) {
813 max_usedpc_order
= i
;
817 return max_usedpc_order
;
820 static struct zspage
*get_zspage(struct page
*page
)
822 struct zspage
*zspage
= (struct zspage
*)page
->private;
824 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
828 static struct page
*get_next_page(struct page
*page
)
830 if (unlikely(PageHugeObject(page
)))
833 return page
->freelist
;
837 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
838 * @obj: the encoded object value
839 * @page: page object resides in zspage
840 * @obj_idx: object index
842 static void obj_to_location(unsigned long obj
, struct page
**page
,
843 unsigned int *obj_idx
)
845 obj
>>= OBJ_TAG_BITS
;
846 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
847 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
851 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
852 * @page: page object resides in zspage
853 * @obj_idx: object index
855 static unsigned long location_to_obj(struct page
*page
, unsigned int obj_idx
)
859 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
860 obj
|= obj_idx
& OBJ_INDEX_MASK
;
861 obj
<<= OBJ_TAG_BITS
;
866 static unsigned long handle_to_obj(unsigned long handle
)
868 return *(unsigned long *)handle
;
871 static unsigned long obj_to_head(struct page
*page
, void *obj
)
873 if (unlikely(PageHugeObject(page
))) {
874 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
877 return *(unsigned long *)obj
;
880 static inline int testpin_tag(unsigned long handle
)
882 return bit_spin_is_locked(HANDLE_PIN_BIT
, (unsigned long *)handle
);
885 static inline int trypin_tag(unsigned long handle
)
887 return bit_spin_trylock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
890 static void pin_tag(unsigned long handle
) __acquires(bitlock
)
892 bit_spin_lock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
895 static void unpin_tag(unsigned long handle
) __releases(bitlock
)
897 bit_spin_unlock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
900 static void reset_page(struct page
*page
)
902 __ClearPageMovable(page
);
903 ClearPagePrivate(page
);
904 set_page_private(page
, 0);
905 page_mapcount_reset(page
);
906 ClearPageHugeObject(page
);
907 page
->freelist
= NULL
;
910 static int trylock_zspage(struct zspage
*zspage
)
912 struct page
*cursor
, *fail
;
914 for (cursor
= get_first_page(zspage
); cursor
!= NULL
; cursor
=
915 get_next_page(cursor
)) {
916 if (!trylock_page(cursor
)) {
924 for (cursor
= get_first_page(zspage
); cursor
!= fail
; cursor
=
925 get_next_page(cursor
))
931 static void __free_zspage(struct zs_pool
*pool
, struct size_class
*class,
932 struct zspage
*zspage
)
934 struct page
*page
, *next
;
935 enum fullness_group fg
;
936 unsigned int class_idx
;
938 get_zspage_mapping(zspage
, &class_idx
, &fg
);
940 assert_spin_locked(&class->lock
);
942 VM_BUG_ON(get_zspage_inuse(zspage
));
943 VM_BUG_ON(fg
!= ZS_EMPTY
);
945 next
= page
= get_first_page(zspage
);
947 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
948 next
= get_next_page(page
);
951 dec_zone_page_state(page
, NR_ZSPAGES
);
954 } while (page
!= NULL
);
956 cache_free_zspage(pool
, zspage
);
958 zs_stat_dec(class, OBJ_ALLOCATED
, class->objs_per_zspage
);
959 atomic_long_sub(class->pages_per_zspage
,
960 &pool
->pages_allocated
);
963 static void free_zspage(struct zs_pool
*pool
, struct size_class
*class,
964 struct zspage
*zspage
)
966 VM_BUG_ON(get_zspage_inuse(zspage
));
967 VM_BUG_ON(list_empty(&zspage
->list
));
969 if (!trylock_zspage(zspage
)) {
970 kick_deferred_free(pool
);
974 remove_zspage(class, zspage
, ZS_EMPTY
);
975 __free_zspage(pool
, class, zspage
);
978 /* Initialize a newly allocated zspage */
979 static void init_zspage(struct size_class
*class, struct zspage
*zspage
)
981 unsigned int freeobj
= 1;
982 unsigned long off
= 0;
983 struct page
*page
= get_first_page(zspage
);
986 struct page
*next_page
;
987 struct link_free
*link
;
990 set_first_obj_offset(page
, off
);
992 vaddr
= kmap_atomic(page
);
993 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
995 while ((off
+= class->size
) < PAGE_SIZE
) {
996 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
997 link
+= class->size
/ sizeof(*link
);
1001 * We now come to the last (full or partial) object on this
1002 * page, which must point to the first object on the next
1005 next_page
= get_next_page(page
);
1007 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
1010 * Reset OBJ_TAG_BITS bit to last link to tell
1011 * whether it's allocated object or not.
1013 link
->next
= -1UL << OBJ_TAG_BITS
;
1015 kunmap_atomic(vaddr
);
1020 set_freeobj(zspage
, 0);
1023 static void create_page_chain(struct size_class
*class, struct zspage
*zspage
,
1024 struct page
*pages
[])
1028 struct page
*prev_page
= NULL
;
1029 int nr_pages
= class->pages_per_zspage
;
1032 * Allocate individual pages and link them together as:
1033 * 1. all pages are linked together using page->freelist
1034 * 2. each sub-page point to zspage using page->private
1036 * we set PG_private to identify the first page (i.e. no other sub-page
1037 * has this flag set).
1039 for (i
= 0; i
< nr_pages
; i
++) {
1041 set_page_private(page
, (unsigned long)zspage
);
1042 page
->freelist
= NULL
;
1044 zspage
->first_page
= page
;
1045 SetPagePrivate(page
);
1046 if (unlikely(class->objs_per_zspage
== 1 &&
1047 class->pages_per_zspage
== 1))
1048 SetPageHugeObject(page
);
1050 prev_page
->freelist
= page
;
1057 * Allocate a zspage for the given size class
1059 static struct zspage
*alloc_zspage(struct zs_pool
*pool
,
1060 struct size_class
*class,
1064 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
];
1065 struct zspage
*zspage
= cache_alloc_zspage(pool
, gfp
);
1070 memset(zspage
, 0, sizeof(struct zspage
));
1071 zspage
->magic
= ZSPAGE_MAGIC
;
1072 migrate_lock_init(zspage
);
1074 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
1077 page
= alloc_page(gfp
);
1080 dec_zone_page_state(pages
[i
], NR_ZSPAGES
);
1081 __free_page(pages
[i
]);
1083 cache_free_zspage(pool
, zspage
);
1087 inc_zone_page_state(page
, NR_ZSPAGES
);
1091 create_page_chain(class, zspage
, pages
);
1092 init_zspage(class, zspage
);
1097 static struct zspage
*find_get_zspage(struct size_class
*class)
1100 struct zspage
*zspage
;
1102 for (i
= ZS_ALMOST_FULL
; i
>= ZS_EMPTY
; i
--) {
1103 zspage
= list_first_entry_or_null(&class->fullness_list
[i
],
1104 struct zspage
, list
);
1112 static inline int __zs_cpu_up(struct mapping_area
*area
)
1115 * Make sure we don't leak memory if a cpu UP notification
1116 * and zs_init() race and both call zs_cpu_up() on the same cpu
1120 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1126 static inline void __zs_cpu_down(struct mapping_area
*area
)
1128 kfree(area
->vm_buf
);
1129 area
->vm_buf
= NULL
;
1132 static void *__zs_map_object(struct mapping_area
*area
,
1133 struct page
*pages
[2], int off
, int size
)
1137 char *buf
= area
->vm_buf
;
1139 /* disable page faults to match kmap_atomic() return conditions */
1140 pagefault_disable();
1142 /* no read fastpath */
1143 if (area
->vm_mm
== ZS_MM_WO
)
1146 sizes
[0] = PAGE_SIZE
- off
;
1147 sizes
[1] = size
- sizes
[0];
1149 /* copy object to per-cpu buffer */
1150 addr
= kmap_atomic(pages
[0]);
1151 memcpy(buf
, addr
+ off
, sizes
[0]);
1152 kunmap_atomic(addr
);
1153 addr
= kmap_atomic(pages
[1]);
1154 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1155 kunmap_atomic(addr
);
1157 return area
->vm_buf
;
1160 static void __zs_unmap_object(struct mapping_area
*area
,
1161 struct page
*pages
[2], int off
, int size
)
1167 /* no write fastpath */
1168 if (area
->vm_mm
== ZS_MM_RO
)
1172 buf
= buf
+ ZS_HANDLE_SIZE
;
1173 size
-= ZS_HANDLE_SIZE
;
1174 off
+= ZS_HANDLE_SIZE
;
1176 sizes
[0] = PAGE_SIZE
- off
;
1177 sizes
[1] = size
- sizes
[0];
1179 /* copy per-cpu buffer to object */
1180 addr
= kmap_atomic(pages
[0]);
1181 memcpy(addr
+ off
, buf
, sizes
[0]);
1182 kunmap_atomic(addr
);
1183 addr
= kmap_atomic(pages
[1]);
1184 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1185 kunmap_atomic(addr
);
1188 /* enable page faults to match kunmap_atomic() return conditions */
1192 static int zs_cpu_prepare(unsigned int cpu
)
1194 struct mapping_area
*area
;
1196 area
= &per_cpu(zs_map_area
, cpu
);
1197 return __zs_cpu_up(area
);
1200 static int zs_cpu_dead(unsigned int cpu
)
1202 struct mapping_area
*area
;
1204 area
= &per_cpu(zs_map_area
, cpu
);
1205 __zs_cpu_down(area
);
1209 static bool can_merge(struct size_class
*prev
, int pages_per_zspage
,
1210 int objs_per_zspage
)
1212 if (prev
->pages_per_zspage
== pages_per_zspage
&&
1213 prev
->objs_per_zspage
== objs_per_zspage
)
1219 static bool zspage_full(struct size_class
*class, struct zspage
*zspage
)
1221 return get_zspage_inuse(zspage
) == class->objs_per_zspage
;
1224 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1226 return atomic_long_read(&pool
->pages_allocated
);
1228 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1231 * zs_map_object - get address of allocated object from handle.
1232 * @pool: pool from which the object was allocated
1233 * @handle: handle returned from zs_malloc
1234 * @mm: maping mode to use
1236 * Before using an object allocated from zs_malloc, it must be mapped using
1237 * this function. When done with the object, it must be unmapped using
1240 * Only one object can be mapped per cpu at a time. There is no protection
1241 * against nested mappings.
1243 * This function returns with preemption and page faults disabled.
1245 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1248 struct zspage
*zspage
;
1250 unsigned long obj
, off
;
1251 unsigned int obj_idx
;
1253 unsigned int class_idx
;
1254 enum fullness_group fg
;
1255 struct size_class
*class;
1256 struct mapping_area
*area
;
1257 struct page
*pages
[2];
1261 * Because we use per-cpu mapping areas shared among the
1262 * pools/users, we can't allow mapping in interrupt context
1263 * because it can corrupt another users mappings.
1265 BUG_ON(in_interrupt());
1267 /* From now on, migration cannot move the object */
1270 obj
= handle_to_obj(handle
);
1271 obj_to_location(obj
, &page
, &obj_idx
);
1272 zspage
= get_zspage(page
);
1274 /* migration cannot move any subpage in this zspage */
1275 migrate_read_lock(zspage
);
1277 get_zspage_mapping(zspage
, &class_idx
, &fg
);
1278 class = pool
->size_class
[class_idx
];
1279 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1281 area
= &get_cpu_var(zs_map_area
);
1283 if (off
+ class->size
<= PAGE_SIZE
) {
1284 /* this object is contained entirely within a page */
1285 area
->vm_addr
= kmap_atomic(page
);
1286 ret
= area
->vm_addr
+ off
;
1290 /* this object spans two pages */
1292 pages
[1] = get_next_page(page
);
1295 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1297 if (likely(!PageHugeObject(page
)))
1298 ret
+= ZS_HANDLE_SIZE
;
1302 EXPORT_SYMBOL_GPL(zs_map_object
);
1304 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1306 struct zspage
*zspage
;
1308 unsigned long obj
, off
;
1309 unsigned int obj_idx
;
1311 unsigned int class_idx
;
1312 enum fullness_group fg
;
1313 struct size_class
*class;
1314 struct mapping_area
*area
;
1316 obj
= handle_to_obj(handle
);
1317 obj_to_location(obj
, &page
, &obj_idx
);
1318 zspage
= get_zspage(page
);
1319 get_zspage_mapping(zspage
, &class_idx
, &fg
);
1320 class = pool
->size_class
[class_idx
];
1321 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1323 area
= this_cpu_ptr(&zs_map_area
);
1324 if (off
+ class->size
<= PAGE_SIZE
)
1325 kunmap_atomic(area
->vm_addr
);
1327 struct page
*pages
[2];
1330 pages
[1] = get_next_page(page
);
1333 __zs_unmap_object(area
, pages
, off
, class->size
);
1335 put_cpu_var(zs_map_area
);
1337 migrate_read_unlock(zspage
);
1340 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1343 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1344 * zsmalloc &size_class.
1345 * @pool: zsmalloc pool to use
1347 * The function returns the size of the first huge class - any object of equal
1348 * or bigger size will be stored in zspage consisting of a single physical
1351 * Context: Any context.
1353 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1355 size_t zs_huge_class_size(struct zs_pool
*pool
)
1357 return huge_class_size
;
1359 EXPORT_SYMBOL_GPL(zs_huge_class_size
);
1361 static unsigned long obj_malloc(struct size_class
*class,
1362 struct zspage
*zspage
, unsigned long handle
)
1364 int i
, nr_page
, offset
;
1366 struct link_free
*link
;
1368 struct page
*m_page
;
1369 unsigned long m_offset
;
1372 handle
|= OBJ_ALLOCATED_TAG
;
1373 obj
= get_freeobj(zspage
);
1375 offset
= obj
* class->size
;
1376 nr_page
= offset
>> PAGE_SHIFT
;
1377 m_offset
= offset
& ~PAGE_MASK
;
1378 m_page
= get_first_page(zspage
);
1380 for (i
= 0; i
< nr_page
; i
++)
1381 m_page
= get_next_page(m_page
);
1383 vaddr
= kmap_atomic(m_page
);
1384 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1385 set_freeobj(zspage
, link
->next
>> OBJ_TAG_BITS
);
1386 if (likely(!PageHugeObject(m_page
)))
1387 /* record handle in the header of allocated chunk */
1388 link
->handle
= handle
;
1390 /* record handle to page->index */
1391 zspage
->first_page
->index
= handle
;
1393 kunmap_atomic(vaddr
);
1394 mod_zspage_inuse(zspage
, 1);
1395 zs_stat_inc(class, OBJ_USED
, 1);
1397 obj
= location_to_obj(m_page
, obj
);
1404 * zs_malloc - Allocate block of given size from pool.
1405 * @pool: pool to allocate from
1406 * @size: size of block to allocate
1407 * @gfp: gfp flags when allocating object
1409 * On success, handle to the allocated object is returned,
1411 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1413 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
, gfp_t gfp
)
1415 unsigned long handle
, obj
;
1416 struct size_class
*class;
1417 enum fullness_group newfg
;
1418 struct zspage
*zspage
;
1420 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1423 handle
= cache_alloc_handle(pool
, gfp
);
1427 /* extra space in chunk to keep the handle */
1428 size
+= ZS_HANDLE_SIZE
;
1429 class = pool
->size_class
[get_size_class_index(size
)];
1431 spin_lock(&class->lock
);
1432 zspage
= find_get_zspage(class);
1433 if (likely(zspage
)) {
1434 obj
= obj_malloc(class, zspage
, handle
);
1435 /* Now move the zspage to another fullness group, if required */
1436 fix_fullness_group(class, zspage
);
1437 record_obj(handle
, obj
);
1438 spin_unlock(&class->lock
);
1443 spin_unlock(&class->lock
);
1445 zspage
= alloc_zspage(pool
, class, gfp
);
1447 cache_free_handle(pool
, handle
);
1451 spin_lock(&class->lock
);
1452 obj
= obj_malloc(class, zspage
, handle
);
1453 newfg
= get_fullness_group(class, zspage
);
1454 insert_zspage(class, zspage
, newfg
);
1455 set_zspage_mapping(zspage
, class->index
, newfg
);
1456 record_obj(handle
, obj
);
1457 atomic_long_add(class->pages_per_zspage
,
1458 &pool
->pages_allocated
);
1459 zs_stat_inc(class, OBJ_ALLOCATED
, class->objs_per_zspage
);
1461 /* We completely set up zspage so mark them as movable */
1462 SetZsPageMovable(pool
, zspage
);
1463 spin_unlock(&class->lock
);
1467 EXPORT_SYMBOL_GPL(zs_malloc
);
1469 static void obj_free(struct size_class
*class, unsigned long obj
)
1471 struct link_free
*link
;
1472 struct zspage
*zspage
;
1473 struct page
*f_page
;
1474 unsigned long f_offset
;
1475 unsigned int f_objidx
;
1478 obj
&= ~OBJ_ALLOCATED_TAG
;
1479 obj_to_location(obj
, &f_page
, &f_objidx
);
1480 f_offset
= (class->size
* f_objidx
) & ~PAGE_MASK
;
1481 zspage
= get_zspage(f_page
);
1483 vaddr
= kmap_atomic(f_page
);
1485 /* Insert this object in containing zspage's freelist */
1486 link
= (struct link_free
*)(vaddr
+ f_offset
);
1487 link
->next
= get_freeobj(zspage
) << OBJ_TAG_BITS
;
1488 kunmap_atomic(vaddr
);
1489 set_freeobj(zspage
, f_objidx
);
1490 mod_zspage_inuse(zspage
, -1);
1491 zs_stat_dec(class, OBJ_USED
, 1);
1494 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1496 struct zspage
*zspage
;
1497 struct page
*f_page
;
1499 unsigned int f_objidx
;
1501 struct size_class
*class;
1502 enum fullness_group fullness
;
1505 if (unlikely(!handle
))
1509 obj
= handle_to_obj(handle
);
1510 obj_to_location(obj
, &f_page
, &f_objidx
);
1511 zspage
= get_zspage(f_page
);
1513 migrate_read_lock(zspage
);
1515 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1516 class = pool
->size_class
[class_idx
];
1518 spin_lock(&class->lock
);
1519 obj_free(class, obj
);
1520 fullness
= fix_fullness_group(class, zspage
);
1521 if (fullness
!= ZS_EMPTY
) {
1522 migrate_read_unlock(zspage
);
1526 isolated
= is_zspage_isolated(zspage
);
1527 migrate_read_unlock(zspage
);
1528 /* If zspage is isolated, zs_page_putback will free the zspage */
1529 if (likely(!isolated
))
1530 free_zspage(pool
, class, zspage
);
1533 spin_unlock(&class->lock
);
1535 cache_free_handle(pool
, handle
);
1537 EXPORT_SYMBOL_GPL(zs_free
);
1539 static void zs_object_copy(struct size_class
*class, unsigned long dst
,
1542 struct page
*s_page
, *d_page
;
1543 unsigned int s_objidx
, d_objidx
;
1544 unsigned long s_off
, d_off
;
1545 void *s_addr
, *d_addr
;
1546 int s_size
, d_size
, size
;
1549 s_size
= d_size
= class->size
;
1551 obj_to_location(src
, &s_page
, &s_objidx
);
1552 obj_to_location(dst
, &d_page
, &d_objidx
);
1554 s_off
= (class->size
* s_objidx
) & ~PAGE_MASK
;
1555 d_off
= (class->size
* d_objidx
) & ~PAGE_MASK
;
1557 if (s_off
+ class->size
> PAGE_SIZE
)
1558 s_size
= PAGE_SIZE
- s_off
;
1560 if (d_off
+ class->size
> PAGE_SIZE
)
1561 d_size
= PAGE_SIZE
- d_off
;
1563 s_addr
= kmap_atomic(s_page
);
1564 d_addr
= kmap_atomic(d_page
);
1567 size
= min(s_size
, d_size
);
1568 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1571 if (written
== class->size
)
1579 if (s_off
>= PAGE_SIZE
) {
1580 kunmap_atomic(d_addr
);
1581 kunmap_atomic(s_addr
);
1582 s_page
= get_next_page(s_page
);
1583 s_addr
= kmap_atomic(s_page
);
1584 d_addr
= kmap_atomic(d_page
);
1585 s_size
= class->size
- written
;
1589 if (d_off
>= PAGE_SIZE
) {
1590 kunmap_atomic(d_addr
);
1591 d_page
= get_next_page(d_page
);
1592 d_addr
= kmap_atomic(d_page
);
1593 d_size
= class->size
- written
;
1598 kunmap_atomic(d_addr
);
1599 kunmap_atomic(s_addr
);
1603 * Find alloced object in zspage from index object and
1606 static unsigned long find_alloced_obj(struct size_class
*class,
1607 struct page
*page
, int *obj_idx
)
1611 int index
= *obj_idx
;
1612 unsigned long handle
= 0;
1613 void *addr
= kmap_atomic(page
);
1615 offset
= get_first_obj_offset(page
);
1616 offset
+= class->size
* index
;
1618 while (offset
< PAGE_SIZE
) {
1619 head
= obj_to_head(page
, addr
+ offset
);
1620 if (head
& OBJ_ALLOCATED_TAG
) {
1621 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1622 if (trypin_tag(handle
))
1627 offset
+= class->size
;
1631 kunmap_atomic(addr
);
1638 struct zs_compact_control
{
1639 /* Source spage for migration which could be a subpage of zspage */
1640 struct page
*s_page
;
1641 /* Destination page for migration which should be a first page
1643 struct page
*d_page
;
1644 /* Starting object index within @s_page which used for live object
1645 * in the subpage. */
1649 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1650 struct zs_compact_control
*cc
)
1652 unsigned long used_obj
, free_obj
;
1653 unsigned long handle
;
1654 struct page
*s_page
= cc
->s_page
;
1655 struct page
*d_page
= cc
->d_page
;
1656 int obj_idx
= cc
->obj_idx
;
1660 handle
= find_alloced_obj(class, s_page
, &obj_idx
);
1662 s_page
= get_next_page(s_page
);
1669 /* Stop if there is no more space */
1670 if (zspage_full(class, get_zspage(d_page
))) {
1676 used_obj
= handle_to_obj(handle
);
1677 free_obj
= obj_malloc(class, get_zspage(d_page
), handle
);
1678 zs_object_copy(class, free_obj
, used_obj
);
1681 * record_obj updates handle's value to free_obj and it will
1682 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1683 * breaks synchronization using pin_tag(e,g, zs_free) so
1684 * let's keep the lock bit.
1686 free_obj
|= BIT(HANDLE_PIN_BIT
);
1687 record_obj(handle
, free_obj
);
1689 obj_free(class, used_obj
);
1692 /* Remember last position in this iteration */
1693 cc
->s_page
= s_page
;
1694 cc
->obj_idx
= obj_idx
;
1699 static struct zspage
*isolate_zspage(struct size_class
*class, bool source
)
1702 struct zspage
*zspage
;
1703 enum fullness_group fg
[2] = {ZS_ALMOST_EMPTY
, ZS_ALMOST_FULL
};
1706 fg
[0] = ZS_ALMOST_FULL
;
1707 fg
[1] = ZS_ALMOST_EMPTY
;
1710 for (i
= 0; i
< 2; i
++) {
1711 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
[i
]],
1712 struct zspage
, list
);
1714 VM_BUG_ON(is_zspage_isolated(zspage
));
1715 remove_zspage(class, zspage
, fg
[i
]);
1724 * putback_zspage - add @zspage into right class's fullness list
1725 * @class: destination class
1726 * @zspage: target page
1728 * Return @zspage's fullness_group
1730 static enum fullness_group
putback_zspage(struct size_class
*class,
1731 struct zspage
*zspage
)
1733 enum fullness_group fullness
;
1735 VM_BUG_ON(is_zspage_isolated(zspage
));
1737 fullness
= get_fullness_group(class, zspage
);
1738 insert_zspage(class, zspage
, fullness
);
1739 set_zspage_mapping(zspage
, class->index
, fullness
);
1744 #ifdef CONFIG_COMPACTION
1746 * To prevent zspage destroy during migration, zspage freeing should
1747 * hold locks of all pages in the zspage.
1749 static void lock_zspage(struct zspage
*zspage
)
1751 struct page
*page
= get_first_page(zspage
);
1755 } while ((page
= get_next_page(page
)) != NULL
);
1758 static int zs_init_fs_context(struct fs_context
*fc
)
1760 return init_pseudo(fc
, ZSMALLOC_MAGIC
) ? 0 : -ENOMEM
;
1763 static struct file_system_type zsmalloc_fs
= {
1765 .init_fs_context
= zs_init_fs_context
,
1766 .kill_sb
= kill_anon_super
,
1769 static int zsmalloc_mount(void)
1773 zsmalloc_mnt
= kern_mount(&zsmalloc_fs
);
1774 if (IS_ERR(zsmalloc_mnt
))
1775 ret
= PTR_ERR(zsmalloc_mnt
);
1780 static void zsmalloc_unmount(void)
1782 kern_unmount(zsmalloc_mnt
);
1785 static void migrate_lock_init(struct zspage
*zspage
)
1787 rwlock_init(&zspage
->lock
);
1790 static void migrate_read_lock(struct zspage
*zspage
) __acquires(&zspage
->lock
)
1792 read_lock(&zspage
->lock
);
1795 static void migrate_read_unlock(struct zspage
*zspage
) __releases(&zspage
->lock
)
1797 read_unlock(&zspage
->lock
);
1800 static void migrate_write_lock(struct zspage
*zspage
)
1802 write_lock(&zspage
->lock
);
1805 static void migrate_write_unlock(struct zspage
*zspage
)
1807 write_unlock(&zspage
->lock
);
1810 /* Number of isolated subpage for *page migration* in this zspage */
1811 static void inc_zspage_isolation(struct zspage
*zspage
)
1816 static void dec_zspage_isolation(struct zspage
*zspage
)
1821 static void putback_zspage_deferred(struct zs_pool
*pool
,
1822 struct size_class
*class,
1823 struct zspage
*zspage
)
1825 enum fullness_group fg
;
1827 fg
= putback_zspage(class, zspage
);
1829 schedule_work(&pool
->free_work
);
1833 static inline void zs_pool_dec_isolated(struct zs_pool
*pool
)
1835 VM_BUG_ON(atomic_long_read(&pool
->isolated_pages
) <= 0);
1836 atomic_long_dec(&pool
->isolated_pages
);
1838 * There's no possibility of racing, since wait_for_isolated_drain()
1839 * checks the isolated count under &class->lock after enqueuing
1840 * on migration_wait.
1842 if (atomic_long_read(&pool
->isolated_pages
) == 0 && pool
->destroying
)
1843 wake_up_all(&pool
->migration_wait
);
1846 static void replace_sub_page(struct size_class
*class, struct zspage
*zspage
,
1847 struct page
*newpage
, struct page
*oldpage
)
1850 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
] = {NULL
, };
1853 page
= get_first_page(zspage
);
1855 if (page
== oldpage
)
1856 pages
[idx
] = newpage
;
1860 } while ((page
= get_next_page(page
)) != NULL
);
1862 create_page_chain(class, zspage
, pages
);
1863 set_first_obj_offset(newpage
, get_first_obj_offset(oldpage
));
1864 if (unlikely(PageHugeObject(oldpage
)))
1865 newpage
->index
= oldpage
->index
;
1866 __SetPageMovable(newpage
, page_mapping(oldpage
));
1869 static bool zs_page_isolate(struct page
*page
, isolate_mode_t mode
)
1871 struct zs_pool
*pool
;
1872 struct size_class
*class;
1874 enum fullness_group fullness
;
1875 struct zspage
*zspage
;
1876 struct address_space
*mapping
;
1879 * Page is locked so zspage couldn't be destroyed. For detail, look at
1880 * lock_zspage in free_zspage.
1882 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1883 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1885 zspage
= get_zspage(page
);
1888 * Without class lock, fullness could be stale while class_idx is okay
1889 * because class_idx is constant unless page is freed so we should get
1890 * fullness again under class lock.
1892 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1893 mapping
= page_mapping(page
);
1894 pool
= mapping
->private_data
;
1895 class = pool
->size_class
[class_idx
];
1897 spin_lock(&class->lock
);
1898 if (get_zspage_inuse(zspage
) == 0) {
1899 spin_unlock(&class->lock
);
1903 /* zspage is isolated for object migration */
1904 if (list_empty(&zspage
->list
) && !is_zspage_isolated(zspage
)) {
1905 spin_unlock(&class->lock
);
1910 * If this is first time isolation for the zspage, isolate zspage from
1911 * size_class to prevent further object allocation from the zspage.
1913 if (!list_empty(&zspage
->list
) && !is_zspage_isolated(zspage
)) {
1914 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1915 atomic_long_inc(&pool
->isolated_pages
);
1916 remove_zspage(class, zspage
, fullness
);
1919 inc_zspage_isolation(zspage
);
1920 spin_unlock(&class->lock
);
1925 static int zs_page_migrate(struct address_space
*mapping
, struct page
*newpage
,
1926 struct page
*page
, enum migrate_mode mode
)
1928 struct zs_pool
*pool
;
1929 struct size_class
*class;
1931 enum fullness_group fullness
;
1932 struct zspage
*zspage
;
1934 void *s_addr
, *d_addr
, *addr
;
1936 unsigned long handle
, head
;
1937 unsigned long old_obj
, new_obj
;
1938 unsigned int obj_idx
;
1942 * We cannot support the _NO_COPY case here, because copy needs to
1943 * happen under the zs lock, which does not work with
1944 * MIGRATE_SYNC_NO_COPY workflow.
1946 if (mode
== MIGRATE_SYNC_NO_COPY
)
1949 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1950 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1952 zspage
= get_zspage(page
);
1954 /* Concurrent compactor cannot migrate any subpage in zspage */
1955 migrate_write_lock(zspage
);
1956 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1957 pool
= mapping
->private_data
;
1958 class = pool
->size_class
[class_idx
];
1959 offset
= get_first_obj_offset(page
);
1961 spin_lock(&class->lock
);
1962 if (!get_zspage_inuse(zspage
)) {
1964 * Set "offset" to end of the page so that every loops
1965 * skips unnecessary object scanning.
1971 s_addr
= kmap_atomic(page
);
1972 while (pos
< PAGE_SIZE
) {
1973 head
= obj_to_head(page
, s_addr
+ pos
);
1974 if (head
& OBJ_ALLOCATED_TAG
) {
1975 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1976 if (!trypin_tag(handle
))
1983 * Here, any user cannot access all objects in the zspage so let's move.
1985 d_addr
= kmap_atomic(newpage
);
1986 memcpy(d_addr
, s_addr
, PAGE_SIZE
);
1987 kunmap_atomic(d_addr
);
1989 for (addr
= s_addr
+ offset
; addr
< s_addr
+ pos
;
1990 addr
+= class->size
) {
1991 head
= obj_to_head(page
, addr
);
1992 if (head
& OBJ_ALLOCATED_TAG
) {
1993 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1994 if (!testpin_tag(handle
))
1997 old_obj
= handle_to_obj(handle
);
1998 obj_to_location(old_obj
, &dummy
, &obj_idx
);
1999 new_obj
= (unsigned long)location_to_obj(newpage
,
2001 new_obj
|= BIT(HANDLE_PIN_BIT
);
2002 record_obj(handle
, new_obj
);
2006 replace_sub_page(class, zspage
, newpage
, page
);
2009 dec_zspage_isolation(zspage
);
2012 * Page migration is done so let's putback isolated zspage to
2013 * the list if @page is final isolated subpage in the zspage.
2015 if (!is_zspage_isolated(zspage
)) {
2017 * We cannot race with zs_destroy_pool() here because we wait
2018 * for isolation to hit zero before we start destroying.
2019 * Also, we ensure that everyone can see pool->destroying before
2022 putback_zspage_deferred(pool
, class, zspage
);
2023 zs_pool_dec_isolated(pool
);
2026 if (page_zone(newpage
) != page_zone(page
)) {
2027 dec_zone_page_state(page
, NR_ZSPAGES
);
2028 inc_zone_page_state(newpage
, NR_ZSPAGES
);
2035 ret
= MIGRATEPAGE_SUCCESS
;
2037 for (addr
= s_addr
+ offset
; addr
< s_addr
+ pos
;
2038 addr
+= class->size
) {
2039 head
= obj_to_head(page
, addr
);
2040 if (head
& OBJ_ALLOCATED_TAG
) {
2041 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2042 if (!testpin_tag(handle
))
2047 kunmap_atomic(s_addr
);
2048 spin_unlock(&class->lock
);
2049 migrate_write_unlock(zspage
);
2054 static void zs_page_putback(struct page
*page
)
2056 struct zs_pool
*pool
;
2057 struct size_class
*class;
2059 enum fullness_group fg
;
2060 struct address_space
*mapping
;
2061 struct zspage
*zspage
;
2063 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
2064 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
2066 zspage
= get_zspage(page
);
2067 get_zspage_mapping(zspage
, &class_idx
, &fg
);
2068 mapping
= page_mapping(page
);
2069 pool
= mapping
->private_data
;
2070 class = pool
->size_class
[class_idx
];
2072 spin_lock(&class->lock
);
2073 dec_zspage_isolation(zspage
);
2074 if (!is_zspage_isolated(zspage
)) {
2076 * Due to page_lock, we cannot free zspage immediately
2079 putback_zspage_deferred(pool
, class, zspage
);
2080 zs_pool_dec_isolated(pool
);
2082 spin_unlock(&class->lock
);
2085 static const struct address_space_operations zsmalloc_aops
= {
2086 .isolate_page
= zs_page_isolate
,
2087 .migratepage
= zs_page_migrate
,
2088 .putback_page
= zs_page_putback
,
2091 static int zs_register_migration(struct zs_pool
*pool
)
2093 pool
->inode
= alloc_anon_inode(zsmalloc_mnt
->mnt_sb
);
2094 if (IS_ERR(pool
->inode
)) {
2099 pool
->inode
->i_mapping
->private_data
= pool
;
2100 pool
->inode
->i_mapping
->a_ops
= &zsmalloc_aops
;
2104 static bool pool_isolated_are_drained(struct zs_pool
*pool
)
2106 return atomic_long_read(&pool
->isolated_pages
) == 0;
2109 /* Function for resolving migration */
2110 static void wait_for_isolated_drain(struct zs_pool
*pool
)
2114 * We're in the process of destroying the pool, so there are no
2115 * active allocations. zs_page_isolate() fails for completely free
2116 * zspages, so we need only wait for the zs_pool's isolated
2117 * count to hit zero.
2119 wait_event(pool
->migration_wait
,
2120 pool_isolated_are_drained(pool
));
2123 static void zs_unregister_migration(struct zs_pool
*pool
)
2125 pool
->destroying
= true;
2127 * We need a memory barrier here to ensure global visibility of
2128 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2129 * case we don't care, or it will be > 0 and pool->destroying will
2130 * ensure that we wake up once isolation hits 0.
2133 wait_for_isolated_drain(pool
); /* This can block */
2134 flush_work(&pool
->free_work
);
2139 * Caller should hold page_lock of all pages in the zspage
2140 * In here, we cannot use zspage meta data.
2142 static void async_free_zspage(struct work_struct
*work
)
2145 struct size_class
*class;
2146 unsigned int class_idx
;
2147 enum fullness_group fullness
;
2148 struct zspage
*zspage
, *tmp
;
2149 LIST_HEAD(free_pages
);
2150 struct zs_pool
*pool
= container_of(work
, struct zs_pool
,
2153 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
2154 class = pool
->size_class
[i
];
2155 if (class->index
!= i
)
2158 spin_lock(&class->lock
);
2159 list_splice_init(&class->fullness_list
[ZS_EMPTY
], &free_pages
);
2160 spin_unlock(&class->lock
);
2164 list_for_each_entry_safe(zspage
, tmp
, &free_pages
, list
) {
2165 list_del(&zspage
->list
);
2166 lock_zspage(zspage
);
2168 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2169 VM_BUG_ON(fullness
!= ZS_EMPTY
);
2170 class = pool
->size_class
[class_idx
];
2171 spin_lock(&class->lock
);
2172 __free_zspage(pool
, pool
->size_class
[class_idx
], zspage
);
2173 spin_unlock(&class->lock
);
2177 static void kick_deferred_free(struct zs_pool
*pool
)
2179 schedule_work(&pool
->free_work
);
2182 static void init_deferred_free(struct zs_pool
*pool
)
2184 INIT_WORK(&pool
->free_work
, async_free_zspage
);
2187 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
)
2189 struct page
*page
= get_first_page(zspage
);
2192 WARN_ON(!trylock_page(page
));
2193 __SetPageMovable(page
, pool
->inode
->i_mapping
);
2195 } while ((page
= get_next_page(page
)) != NULL
);
2201 * Based on the number of unused allocated objects calculate
2202 * and return the number of pages that we can free.
2204 static unsigned long zs_can_compact(struct size_class
*class)
2206 unsigned long obj_wasted
;
2207 unsigned long obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
2208 unsigned long obj_used
= zs_stat_get(class, OBJ_USED
);
2210 if (obj_allocated
<= obj_used
)
2213 obj_wasted
= obj_allocated
- obj_used
;
2214 obj_wasted
/= class->objs_per_zspage
;
2216 return obj_wasted
* class->pages_per_zspage
;
2219 static void __zs_compact(struct zs_pool
*pool
, struct size_class
*class)
2221 struct zs_compact_control cc
;
2222 struct zspage
*src_zspage
;
2223 struct zspage
*dst_zspage
= NULL
;
2225 spin_lock(&class->lock
);
2226 while ((src_zspage
= isolate_zspage(class, true))) {
2228 if (!zs_can_compact(class))
2232 cc
.s_page
= get_first_page(src_zspage
);
2234 while ((dst_zspage
= isolate_zspage(class, false))) {
2235 cc
.d_page
= get_first_page(dst_zspage
);
2237 * If there is no more space in dst_page, resched
2238 * and see if anyone had allocated another zspage.
2240 if (!migrate_zspage(pool
, class, &cc
))
2243 putback_zspage(class, dst_zspage
);
2246 /* Stop if we couldn't find slot */
2247 if (dst_zspage
== NULL
)
2250 putback_zspage(class, dst_zspage
);
2251 if (putback_zspage(class, src_zspage
) == ZS_EMPTY
) {
2252 free_zspage(pool
, class, src_zspage
);
2253 pool
->stats
.pages_compacted
+= class->pages_per_zspage
;
2255 spin_unlock(&class->lock
);
2257 spin_lock(&class->lock
);
2261 putback_zspage(class, src_zspage
);
2263 spin_unlock(&class->lock
);
2266 unsigned long zs_compact(struct zs_pool
*pool
)
2269 struct size_class
*class;
2271 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2272 class = pool
->size_class
[i
];
2275 if (class->index
!= i
)
2277 __zs_compact(pool
, class);
2280 return pool
->stats
.pages_compacted
;
2282 EXPORT_SYMBOL_GPL(zs_compact
);
2284 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
2286 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
2288 EXPORT_SYMBOL_GPL(zs_pool_stats
);
2290 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
2291 struct shrink_control
*sc
)
2293 unsigned long pages_freed
;
2294 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2297 pages_freed
= pool
->stats
.pages_compacted
;
2299 * Compact classes and calculate compaction delta.
2300 * Can run concurrently with a manually triggered
2301 * (by user) compaction.
2303 pages_freed
= zs_compact(pool
) - pages_freed
;
2305 return pages_freed
? pages_freed
: SHRINK_STOP
;
2308 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
2309 struct shrink_control
*sc
)
2312 struct size_class
*class;
2313 unsigned long pages_to_free
= 0;
2314 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2317 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2318 class = pool
->size_class
[i
];
2321 if (class->index
!= i
)
2324 pages_to_free
+= zs_can_compact(class);
2327 return pages_to_free
;
2330 static void zs_unregister_shrinker(struct zs_pool
*pool
)
2332 unregister_shrinker(&pool
->shrinker
);
2335 static int zs_register_shrinker(struct zs_pool
*pool
)
2337 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
2338 pool
->shrinker
.count_objects
= zs_shrinker_count
;
2339 pool
->shrinker
.batch
= 0;
2340 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
2342 return register_shrinker(&pool
->shrinker
);
2346 * zs_create_pool - Creates an allocation pool to work from.
2347 * @name: pool name to be created
2349 * This function must be called before anything when using
2350 * the zsmalloc allocator.
2352 * On success, a pointer to the newly created pool is returned,
2355 struct zs_pool
*zs_create_pool(const char *name
)
2358 struct zs_pool
*pool
;
2359 struct size_class
*prev_class
= NULL
;
2361 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
2365 init_deferred_free(pool
);
2367 pool
->name
= kstrdup(name
, GFP_KERNEL
);
2371 #ifdef CONFIG_COMPACTION
2372 init_waitqueue_head(&pool
->migration_wait
);
2375 if (create_cache(pool
))
2379 * Iterate reversely, because, size of size_class that we want to use
2380 * for merging should be larger or equal to current size.
2382 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
2384 int pages_per_zspage
;
2385 int objs_per_zspage
;
2386 struct size_class
*class;
2389 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
2390 if (size
> ZS_MAX_ALLOC_SIZE
)
2391 size
= ZS_MAX_ALLOC_SIZE
;
2392 pages_per_zspage
= get_pages_per_zspage(size
);
2393 objs_per_zspage
= pages_per_zspage
* PAGE_SIZE
/ size
;
2396 * We iterate from biggest down to smallest classes,
2397 * so huge_class_size holds the size of the first huge
2398 * class. Any object bigger than or equal to that will
2399 * endup in the huge class.
2401 if (pages_per_zspage
!= 1 && objs_per_zspage
!= 1 &&
2403 huge_class_size
= size
;
2405 * The object uses ZS_HANDLE_SIZE bytes to store the
2406 * handle. We need to subtract it, because zs_malloc()
2407 * unconditionally adds handle size before it performs
2408 * size class search - so object may be smaller than
2409 * huge class size, yet it still can end up in the huge
2410 * class because it grows by ZS_HANDLE_SIZE extra bytes
2411 * right before class lookup.
2413 huge_class_size
-= (ZS_HANDLE_SIZE
- 1);
2417 * size_class is used for normal zsmalloc operation such
2418 * as alloc/free for that size. Although it is natural that we
2419 * have one size_class for each size, there is a chance that we
2420 * can get more memory utilization if we use one size_class for
2421 * many different sizes whose size_class have same
2422 * characteristics. So, we makes size_class point to
2423 * previous size_class if possible.
2426 if (can_merge(prev_class
, pages_per_zspage
, objs_per_zspage
)) {
2427 pool
->size_class
[i
] = prev_class
;
2432 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
2438 class->pages_per_zspage
= pages_per_zspage
;
2439 class->objs_per_zspage
= objs_per_zspage
;
2440 spin_lock_init(&class->lock
);
2441 pool
->size_class
[i
] = class;
2442 for (fullness
= ZS_EMPTY
; fullness
< NR_ZS_FULLNESS
;
2444 INIT_LIST_HEAD(&class->fullness_list
[fullness
]);
2449 /* debug only, don't abort if it fails */
2450 zs_pool_stat_create(pool
, name
);
2452 if (zs_register_migration(pool
))
2456 * Not critical since shrinker is only used to trigger internal
2457 * defragmentation of the pool which is pretty optional thing. If
2458 * registration fails we still can use the pool normally and user can
2459 * trigger compaction manually. Thus, ignore return code.
2461 zs_register_shrinker(pool
);
2466 zs_destroy_pool(pool
);
2469 EXPORT_SYMBOL_GPL(zs_create_pool
);
2471 void zs_destroy_pool(struct zs_pool
*pool
)
2475 zs_unregister_shrinker(pool
);
2476 zs_unregister_migration(pool
);
2477 zs_pool_stat_destroy(pool
);
2479 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
2481 struct size_class
*class = pool
->size_class
[i
];
2486 if (class->index
!= i
)
2489 for (fg
= ZS_EMPTY
; fg
< NR_ZS_FULLNESS
; fg
++) {
2490 if (!list_empty(&class->fullness_list
[fg
])) {
2491 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2498 destroy_cache(pool
);
2502 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2504 static int __init
zs_init(void)
2508 ret
= zsmalloc_mount();
2512 ret
= cpuhp_setup_state(CPUHP_MM_ZS_PREPARE
, "mm/zsmalloc:prepare",
2513 zs_cpu_prepare
, zs_cpu_dead
);
2518 zpool_register_driver(&zs_zpool_driver
);
2531 static void __exit
zs_exit(void)
2534 zpool_unregister_driver(&zs_zpool_driver
);
2537 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE
);
2542 module_init(zs_init
);
2543 module_exit(zs_exit
);
2545 MODULE_LICENSE("Dual BSD/GPL");
2546 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");