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_private2: identifies the last component page
28 * PG_owner_priv_1: indentifies the huge component page
32 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
34 #define CREATE_TRACE_POINTS
36 #include <linux/module.h>
37 #include <linux/kernel.h>
38 #include <linux/sched.h>
39 #include <linux/bitops.h>
40 #include <linux/errno.h>
41 #include <linux/highmem.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
44 #include <asm/tlbflush.h>
45 #include <asm/pgtable.h>
46 #include <linux/cpumask.h>
47 #include <linux/cpu.h>
48 #include <linux/vmalloc.h>
49 #include <linux/preempt.h>
50 #include <linux/spinlock.h>
51 #include <linux/types.h>
52 #include <linux/debugfs.h>
53 #include <linux/zsmalloc.h>
54 #include <linux/zpool.h>
55 #include <linux/mount.h>
56 #include <linux/migrate.h>
57 #include <linux/pagemap.h>
58 #include <trace/events/zsmalloc.h>
60 #define ZSPAGE_MAGIC 0x58
63 * This must be power of 2 and greater than of equal to sizeof(link_free).
64 * These two conditions ensure that any 'struct link_free' itself doesn't
65 * span more than 1 page which avoids complex case of mapping 2 pages simply
66 * to restore link_free pointer values.
71 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
72 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
74 #define ZS_MAX_ZSPAGE_ORDER 2
75 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
77 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
80 * Object location (<PFN>, <obj_idx>) is encoded as
81 * as single (unsigned long) handle value.
83 * Note that object index <obj_idx> starts from 0.
85 * This is made more complicated by various memory models and PAE.
88 #ifndef MAX_PHYSMEM_BITS
89 #ifdef CONFIG_HIGHMEM64G
90 #define MAX_PHYSMEM_BITS 36
91 #else /* !CONFIG_HIGHMEM64G */
93 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
96 #define MAX_PHYSMEM_BITS BITS_PER_LONG
99 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
102 * Memory for allocating for handle keeps object position by
103 * encoding <page, obj_idx> and the encoded value has a room
104 * in least bit(ie, look at obj_to_location).
105 * We use the bit to synchronize between object access by
106 * user and migration.
108 #define HANDLE_PIN_BIT 0
111 * Head in allocated object should have OBJ_ALLOCATED_TAG
112 * to identify the object was allocated or not.
113 * It's okay to add the status bit in the least bit because
114 * header keeps handle which is 4byte-aligned address so we
115 * have room for two bit at least.
117 #define OBJ_ALLOCATED_TAG 1
118 #define OBJ_TAG_BITS 1
119 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
120 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
122 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
123 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
124 #define ZS_MIN_ALLOC_SIZE \
125 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
126 /* each chunk includes extra space to keep handle */
127 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
130 * On systems with 4K page size, this gives 255 size classes! There is a
132 * - Large number of size classes is potentially wasteful as free page are
133 * spread across these classes
134 * - Small number of size classes causes large internal fragmentation
135 * - Probably its better to use specific size classes (empirically
136 * determined). NOTE: all those class sizes must be set as multiple of
137 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
139 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
142 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
144 enum fullness_group
{
162 struct zs_size_stat
{
163 unsigned long objs
[NR_ZS_STAT_TYPE
];
166 #ifdef CONFIG_ZSMALLOC_STAT
167 static struct dentry
*zs_stat_root
;
170 #ifdef CONFIG_COMPACTION
171 static struct vfsmount
*zsmalloc_mnt
;
175 * number of size_classes
177 static int zs_size_classes
;
180 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
182 * n = number of allocated objects
183 * N = total number of objects zspage can store
184 * f = fullness_threshold_frac
186 * Similarly, we assign zspage to:
187 * ZS_ALMOST_FULL when n > N / f
188 * ZS_EMPTY when n == 0
189 * ZS_FULL when n == N
191 * (see: fix_fullness_group())
193 static const int fullness_threshold_frac
= 4;
197 struct list_head fullness_list
[NR_ZS_FULLNESS
];
199 * Size of objects stored in this class. Must be multiple
204 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
205 int pages_per_zspage
;
208 struct zs_size_stat stats
;
211 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
212 static void SetPageHugeObject(struct page
*page
)
214 SetPageOwnerPriv1(page
);
217 static void ClearPageHugeObject(struct page
*page
)
219 ClearPageOwnerPriv1(page
);
222 static int PageHugeObject(struct page
*page
)
224 return PageOwnerPriv1(page
);
228 * Placed within free objects to form a singly linked list.
229 * For every zspage, zspage->freeobj gives head of this list.
231 * This must be power of 2 and less than or equal to ZS_ALIGN
237 * It's valid for non-allocated object
241 * Handle of allocated object.
243 unsigned long handle
;
250 struct size_class
**size_class
;
251 struct kmem_cache
*handle_cachep
;
252 struct kmem_cache
*zspage_cachep
;
254 atomic_long_t pages_allocated
;
256 struct zs_pool_stats stats
;
258 /* Compact classes */
259 struct shrinker shrinker
;
261 * To signify that register_shrinker() was successful
262 * and unregister_shrinker() will not Oops.
264 bool shrinker_enabled
;
265 #ifdef CONFIG_ZSMALLOC_STAT
266 struct dentry
*stat_dentry
;
268 #ifdef CONFIG_COMPACTION
270 struct work_struct free_work
;
275 * A zspage's class index and fullness group
276 * are encoded in its (first)page->mapping
278 #define FULLNESS_BITS 2
280 #define ISOLATED_BITS 3
281 #define MAGIC_VAL_BITS 8
285 unsigned int fullness
:FULLNESS_BITS
;
286 unsigned int class:CLASS_BITS
;
287 unsigned int isolated
:ISOLATED_BITS
;
288 unsigned int magic
:MAGIC_VAL_BITS
;
291 unsigned int freeobj
;
292 struct page
*first_page
;
293 struct list_head list
; /* fullness list */
294 #ifdef CONFIG_COMPACTION
299 struct mapping_area
{
300 #ifdef CONFIG_PGTABLE_MAPPING
301 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
303 char *vm_buf
; /* copy buffer for objects that span pages */
305 char *vm_addr
; /* address of kmap_atomic()'ed pages */
306 enum zs_mapmode vm_mm
; /* mapping mode */
309 #ifdef CONFIG_COMPACTION
310 static int zs_register_migration(struct zs_pool
*pool
);
311 static void zs_unregister_migration(struct zs_pool
*pool
);
312 static void migrate_lock_init(struct zspage
*zspage
);
313 static void migrate_read_lock(struct zspage
*zspage
);
314 static void migrate_read_unlock(struct zspage
*zspage
);
315 static void kick_deferred_free(struct zs_pool
*pool
);
316 static void init_deferred_free(struct zs_pool
*pool
);
317 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
);
319 static int zsmalloc_mount(void) { return 0; }
320 static void zsmalloc_unmount(void) {}
321 static int zs_register_migration(struct zs_pool
*pool
) { return 0; }
322 static void zs_unregister_migration(struct zs_pool
*pool
) {}
323 static void migrate_lock_init(struct zspage
*zspage
) {}
324 static void migrate_read_lock(struct zspage
*zspage
) {}
325 static void migrate_read_unlock(struct zspage
*zspage
) {}
326 static void kick_deferred_free(struct zs_pool
*pool
) {}
327 static void init_deferred_free(struct zs_pool
*pool
) {}
328 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
) {}
331 static int create_cache(struct zs_pool
*pool
)
333 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
335 if (!pool
->handle_cachep
)
338 pool
->zspage_cachep
= kmem_cache_create("zspage", sizeof(struct zspage
),
340 if (!pool
->zspage_cachep
) {
341 kmem_cache_destroy(pool
->handle_cachep
);
342 pool
->handle_cachep
= NULL
;
349 static void destroy_cache(struct zs_pool
*pool
)
351 kmem_cache_destroy(pool
->handle_cachep
);
352 kmem_cache_destroy(pool
->zspage_cachep
);
355 static unsigned long cache_alloc_handle(struct zs_pool
*pool
, gfp_t gfp
)
357 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
358 gfp
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
361 static void cache_free_handle(struct zs_pool
*pool
, unsigned long handle
)
363 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
366 static struct zspage
*cache_alloc_zspage(struct zs_pool
*pool
, gfp_t flags
)
368 return kmem_cache_alloc(pool
->zspage_cachep
,
369 flags
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
372 static void cache_free_zspage(struct zs_pool
*pool
, struct zspage
*zspage
)
374 kmem_cache_free(pool
->zspage_cachep
, zspage
);
377 static void record_obj(unsigned long handle
, unsigned long obj
)
380 * lsb of @obj represents handle lock while other bits
381 * represent object value the handle is pointing so
382 * updating shouldn't do store tearing.
384 WRITE_ONCE(*(unsigned long *)handle
, obj
);
391 static void *zs_zpool_create(const char *name
, gfp_t gfp
,
392 const struct zpool_ops
*zpool_ops
,
396 * Ignore global gfp flags: zs_malloc() may be invoked from
397 * different contexts and its caller must provide a valid
400 return zs_create_pool(name
);
403 static void zs_zpool_destroy(void *pool
)
405 zs_destroy_pool(pool
);
408 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
409 unsigned long *handle
)
411 *handle
= zs_malloc(pool
, size
, gfp
);
412 return *handle
? 0 : -1;
414 static void zs_zpool_free(void *pool
, unsigned long handle
)
416 zs_free(pool
, handle
);
419 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
420 unsigned int *reclaimed
)
425 static void *zs_zpool_map(void *pool
, unsigned long handle
,
426 enum zpool_mapmode mm
)
428 enum zs_mapmode zs_mm
;
437 case ZPOOL_MM_RW
: /* fallthru */
443 return zs_map_object(pool
, handle
, zs_mm
);
445 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
447 zs_unmap_object(pool
, handle
);
450 static u64
zs_zpool_total_size(void *pool
)
452 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
455 static struct zpool_driver zs_zpool_driver
= {
457 .owner
= THIS_MODULE
,
458 .create
= zs_zpool_create
,
459 .destroy
= zs_zpool_destroy
,
460 .malloc
= zs_zpool_malloc
,
461 .free
= zs_zpool_free
,
462 .shrink
= zs_zpool_shrink
,
464 .unmap
= zs_zpool_unmap
,
465 .total_size
= zs_zpool_total_size
,
468 MODULE_ALIAS("zpool-zsmalloc");
469 #endif /* CONFIG_ZPOOL */
471 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
472 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
474 static bool is_zspage_isolated(struct zspage
*zspage
)
476 return zspage
->isolated
;
479 static int is_first_page(struct page
*page
)
481 return PagePrivate(page
);
484 /* Protected by class->lock */
485 static inline int get_zspage_inuse(struct zspage
*zspage
)
487 return zspage
->inuse
;
490 static inline void set_zspage_inuse(struct zspage
*zspage
, int val
)
495 static inline void mod_zspage_inuse(struct zspage
*zspage
, int val
)
497 zspage
->inuse
+= val
;
500 static inline struct page
*get_first_page(struct zspage
*zspage
)
502 struct page
*first_page
= zspage
->first_page
;
504 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
508 static inline int get_first_obj_offset(struct page
*page
)
513 static inline void set_first_obj_offset(struct page
*page
, int offset
)
515 page
->units
= offset
;
518 static inline unsigned int get_freeobj(struct zspage
*zspage
)
520 return zspage
->freeobj
;
523 static inline void set_freeobj(struct zspage
*zspage
, unsigned int obj
)
525 zspage
->freeobj
= obj
;
528 static void get_zspage_mapping(struct zspage
*zspage
,
529 unsigned int *class_idx
,
530 enum fullness_group
*fullness
)
532 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
534 *fullness
= zspage
->fullness
;
535 *class_idx
= zspage
->class;
538 static void set_zspage_mapping(struct zspage
*zspage
,
539 unsigned int class_idx
,
540 enum fullness_group fullness
)
542 zspage
->class = class_idx
;
543 zspage
->fullness
= fullness
;
547 * zsmalloc divides the pool into various size classes where each
548 * class maintains a list of zspages where each zspage is divided
549 * into equal sized chunks. Each allocation falls into one of these
550 * classes depending on its size. This function returns index of the
551 * size class which has chunk size big enough to hold the give size.
553 static int get_size_class_index(int size
)
557 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
558 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
559 ZS_SIZE_CLASS_DELTA
);
561 return min(zs_size_classes
- 1, idx
);
564 static inline void zs_stat_inc(struct size_class
*class,
565 enum zs_stat_type type
, unsigned long cnt
)
567 class->stats
.objs
[type
] += cnt
;
570 static inline void zs_stat_dec(struct size_class
*class,
571 enum zs_stat_type type
, unsigned long cnt
)
573 class->stats
.objs
[type
] -= cnt
;
576 static inline unsigned long zs_stat_get(struct size_class
*class,
577 enum zs_stat_type type
)
579 return class->stats
.objs
[type
];
582 #ifdef CONFIG_ZSMALLOC_STAT
584 static void __init
zs_stat_init(void)
586 if (!debugfs_initialized()) {
587 pr_warn("debugfs not available, stat dir not created\n");
591 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
593 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
596 static void __exit
zs_stat_exit(void)
598 debugfs_remove_recursive(zs_stat_root
);
601 static unsigned long zs_can_compact(struct size_class
*class);
603 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
606 struct zs_pool
*pool
= s
->private;
607 struct size_class
*class;
609 unsigned long class_almost_full
, class_almost_empty
;
610 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
611 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
612 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
613 unsigned long total_freeable
= 0;
615 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
616 "class", "size", "almost_full", "almost_empty",
617 "obj_allocated", "obj_used", "pages_used",
618 "pages_per_zspage", "freeable");
620 for (i
= 0; i
< zs_size_classes
; i
++) {
621 class = pool
->size_class
[i
];
623 if (class->index
!= i
)
626 spin_lock(&class->lock
);
627 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
628 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
629 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
630 obj_used
= zs_stat_get(class, OBJ_USED
);
631 freeable
= zs_can_compact(class);
632 spin_unlock(&class->lock
);
634 objs_per_zspage
= class->objs_per_zspage
;
635 pages_used
= obj_allocated
/ objs_per_zspage
*
636 class->pages_per_zspage
;
638 seq_printf(s
, " %5u %5u %11lu %12lu %13lu"
639 " %10lu %10lu %16d %8lu\n",
640 i
, class->size
, class_almost_full
, class_almost_empty
,
641 obj_allocated
, obj_used
, pages_used
,
642 class->pages_per_zspage
, freeable
);
644 total_class_almost_full
+= class_almost_full
;
645 total_class_almost_empty
+= class_almost_empty
;
646 total_objs
+= obj_allocated
;
647 total_used_objs
+= obj_used
;
648 total_pages
+= pages_used
;
649 total_freeable
+= freeable
;
653 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
654 "Total", "", total_class_almost_full
,
655 total_class_almost_empty
, total_objs
,
656 total_used_objs
, total_pages
, "", total_freeable
);
661 static int zs_stats_size_open(struct inode
*inode
, struct file
*file
)
663 return single_open(file
, zs_stats_size_show
, inode
->i_private
);
666 static const struct file_operations zs_stat_size_ops
= {
667 .open
= zs_stats_size_open
,
670 .release
= single_release
,
673 static void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
675 struct dentry
*entry
;
678 pr_warn("no root stat dir, not creating <%s> stat dir\n", name
);
682 entry
= debugfs_create_dir(name
, zs_stat_root
);
684 pr_warn("debugfs dir <%s> creation failed\n", name
);
687 pool
->stat_dentry
= entry
;
689 entry
= debugfs_create_file("classes", S_IFREG
| S_IRUGO
,
690 pool
->stat_dentry
, pool
, &zs_stat_size_ops
);
692 pr_warn("%s: debugfs file entry <%s> creation failed\n",
694 debugfs_remove_recursive(pool
->stat_dentry
);
695 pool
->stat_dentry
= NULL
;
699 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
701 debugfs_remove_recursive(pool
->stat_dentry
);
704 #else /* CONFIG_ZSMALLOC_STAT */
705 static void __init
zs_stat_init(void)
709 static void __exit
zs_stat_exit(void)
713 static inline void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
717 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
724 * For each size class, zspages are divided into different groups
725 * depending on how "full" they are. This was done so that we could
726 * easily find empty or nearly empty zspages when we try to shrink
727 * the pool (not yet implemented). This function returns fullness
728 * status of the given page.
730 static enum fullness_group
get_fullness_group(struct size_class
*class,
731 struct zspage
*zspage
)
733 int inuse
, objs_per_zspage
;
734 enum fullness_group fg
;
736 inuse
= get_zspage_inuse(zspage
);
737 objs_per_zspage
= class->objs_per_zspage
;
741 else if (inuse
== objs_per_zspage
)
743 else if (inuse
<= 3 * objs_per_zspage
/ fullness_threshold_frac
)
744 fg
= ZS_ALMOST_EMPTY
;
752 * Each size class maintains various freelists and zspages are assigned
753 * to one of these freelists based on the number of live objects they
754 * have. This functions inserts the given zspage into the freelist
755 * identified by <class, fullness_group>.
757 static void insert_zspage(struct size_class
*class,
758 struct zspage
*zspage
,
759 enum fullness_group fullness
)
763 zs_stat_inc(class, fullness
, 1);
764 head
= list_first_entry_or_null(&class->fullness_list
[fullness
],
765 struct zspage
, list
);
767 * We want to see more ZS_FULL pages and less almost empty/full.
768 * Put pages with higher ->inuse first.
771 if (get_zspage_inuse(zspage
) < get_zspage_inuse(head
)) {
772 list_add(&zspage
->list
, &head
->list
);
776 list_add(&zspage
->list
, &class->fullness_list
[fullness
]);
780 * This function removes the given zspage from the freelist identified
781 * by <class, fullness_group>.
783 static void remove_zspage(struct size_class
*class,
784 struct zspage
*zspage
,
785 enum fullness_group fullness
)
787 VM_BUG_ON(list_empty(&class->fullness_list
[fullness
]));
788 VM_BUG_ON(is_zspage_isolated(zspage
));
790 list_del_init(&zspage
->list
);
791 zs_stat_dec(class, fullness
, 1);
795 * Each size class maintains zspages in different fullness groups depending
796 * on the number of live objects they contain. When allocating or freeing
797 * objects, the fullness status of the page can change, say, from ALMOST_FULL
798 * to ALMOST_EMPTY when freeing an object. This function checks if such
799 * a status change has occurred for the given page and accordingly moves the
800 * page from the freelist of the old fullness group to that of the new
803 static enum fullness_group
fix_fullness_group(struct size_class
*class,
804 struct zspage
*zspage
)
807 enum fullness_group currfg
, newfg
;
809 get_zspage_mapping(zspage
, &class_idx
, &currfg
);
810 newfg
= get_fullness_group(class, zspage
);
814 if (!is_zspage_isolated(zspage
)) {
815 remove_zspage(class, zspage
, currfg
);
816 insert_zspage(class, zspage
, newfg
);
819 set_zspage_mapping(zspage
, class_idx
, newfg
);
826 * We have to decide on how many pages to link together
827 * to form a zspage for each size class. This is important
828 * to reduce wastage due to unusable space left at end of
829 * each zspage which is given as:
830 * wastage = Zp % class_size
831 * usage = Zp - wastage
832 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
834 * For example, for size class of 3/8 * PAGE_SIZE, we should
835 * link together 3 PAGE_SIZE sized pages to form a zspage
836 * since then we can perfectly fit in 8 such objects.
838 static int get_pages_per_zspage(int class_size
)
840 int i
, max_usedpc
= 0;
841 /* zspage order which gives maximum used size per KB */
842 int max_usedpc_order
= 1;
844 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
848 zspage_size
= i
* PAGE_SIZE
;
849 waste
= zspage_size
% class_size
;
850 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
852 if (usedpc
> max_usedpc
) {
854 max_usedpc_order
= i
;
858 return max_usedpc_order
;
861 static struct zspage
*get_zspage(struct page
*page
)
863 struct zspage
*zspage
= (struct zspage
*)page
->private;
865 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
869 static struct page
*get_next_page(struct page
*page
)
871 if (unlikely(PageHugeObject(page
)))
874 return page
->freelist
;
878 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
879 * @page: page object resides in zspage
880 * @obj_idx: object index
882 static void obj_to_location(unsigned long obj
, struct page
**page
,
883 unsigned int *obj_idx
)
885 obj
>>= OBJ_TAG_BITS
;
886 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
887 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
891 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
892 * @page: page object resides in zspage
893 * @obj_idx: object index
895 static unsigned long location_to_obj(struct page
*page
, unsigned int obj_idx
)
899 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
900 obj
|= obj_idx
& OBJ_INDEX_MASK
;
901 obj
<<= OBJ_TAG_BITS
;
906 static unsigned long handle_to_obj(unsigned long handle
)
908 return *(unsigned long *)handle
;
911 static unsigned long obj_to_head(struct page
*page
, void *obj
)
913 if (unlikely(PageHugeObject(page
))) {
914 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
917 return *(unsigned long *)obj
;
920 static inline int testpin_tag(unsigned long handle
)
922 return bit_spin_is_locked(HANDLE_PIN_BIT
, (unsigned long *)handle
);
925 static inline int trypin_tag(unsigned long handle
)
927 return bit_spin_trylock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
930 static void pin_tag(unsigned long handle
)
932 bit_spin_lock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
935 static void unpin_tag(unsigned long handle
)
937 bit_spin_unlock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
940 static void reset_page(struct page
*page
)
942 __ClearPageMovable(page
);
943 ClearPagePrivate(page
);
944 ClearPagePrivate2(page
);
945 set_page_private(page
, 0);
946 page_mapcount_reset(page
);
947 ClearPageHugeObject(page
);
948 page
->freelist
= NULL
;
952 * To prevent zspage destroy during migration, zspage freeing should
953 * hold locks of all pages in the zspage.
955 void lock_zspage(struct zspage
*zspage
)
957 struct page
*page
= get_first_page(zspage
);
961 } while ((page
= get_next_page(page
)) != NULL
);
964 int trylock_zspage(struct zspage
*zspage
)
966 struct page
*cursor
, *fail
;
968 for (cursor
= get_first_page(zspage
); cursor
!= NULL
; cursor
=
969 get_next_page(cursor
)) {
970 if (!trylock_page(cursor
)) {
978 for (cursor
= get_first_page(zspage
); cursor
!= fail
; cursor
=
979 get_next_page(cursor
))
985 static void __free_zspage(struct zs_pool
*pool
, struct size_class
*class,
986 struct zspage
*zspage
)
988 struct page
*page
, *next
;
989 enum fullness_group fg
;
990 unsigned int class_idx
;
992 get_zspage_mapping(zspage
, &class_idx
, &fg
);
994 assert_spin_locked(&class->lock
);
996 VM_BUG_ON(get_zspage_inuse(zspage
));
997 VM_BUG_ON(fg
!= ZS_EMPTY
);
999 next
= page
= get_first_page(zspage
);
1001 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1002 next
= get_next_page(page
);
1005 dec_zone_page_state(page
, NR_ZSPAGES
);
1008 } while (page
!= NULL
);
1010 cache_free_zspage(pool
, zspage
);
1012 zs_stat_dec(class, OBJ_ALLOCATED
, class->objs_per_zspage
);
1013 atomic_long_sub(class->pages_per_zspage
,
1014 &pool
->pages_allocated
);
1017 static void free_zspage(struct zs_pool
*pool
, struct size_class
*class,
1018 struct zspage
*zspage
)
1020 VM_BUG_ON(get_zspage_inuse(zspage
));
1021 VM_BUG_ON(list_empty(&zspage
->list
));
1023 if (!trylock_zspage(zspage
)) {
1024 kick_deferred_free(pool
);
1028 remove_zspage(class, zspage
, ZS_EMPTY
);
1029 __free_zspage(pool
, class, zspage
);
1032 /* Initialize a newly allocated zspage */
1033 static void init_zspage(struct size_class
*class, struct zspage
*zspage
)
1035 unsigned int freeobj
= 1;
1036 unsigned long off
= 0;
1037 struct page
*page
= get_first_page(zspage
);
1040 struct page
*next_page
;
1041 struct link_free
*link
;
1044 set_first_obj_offset(page
, off
);
1046 vaddr
= kmap_atomic(page
);
1047 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
1049 while ((off
+= class->size
) < PAGE_SIZE
) {
1050 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
1051 link
+= class->size
/ sizeof(*link
);
1055 * We now come to the last (full or partial) object on this
1056 * page, which must point to the first object on the next
1059 next_page
= get_next_page(page
);
1061 link
->next
= freeobj
++ << OBJ_TAG_BITS
;
1064 * Reset OBJ_TAG_BITS bit to last link to tell
1065 * whether it's allocated object or not.
1067 link
->next
= -1 << OBJ_TAG_BITS
;
1069 kunmap_atomic(vaddr
);
1074 set_freeobj(zspage
, 0);
1077 static void create_page_chain(struct size_class
*class, struct zspage
*zspage
,
1078 struct page
*pages
[])
1082 struct page
*prev_page
= NULL
;
1083 int nr_pages
= class->pages_per_zspage
;
1086 * Allocate individual pages and link them together as:
1087 * 1. all pages are linked together using page->freelist
1088 * 2. each sub-page point to zspage using page->private
1090 * we set PG_private to identify the first page (i.e. no other sub-page
1091 * has this flag set) and PG_private_2 to identify the last page.
1093 for (i
= 0; i
< nr_pages
; i
++) {
1095 set_page_private(page
, (unsigned long)zspage
);
1096 page
->freelist
= NULL
;
1098 zspage
->first_page
= page
;
1099 SetPagePrivate(page
);
1100 if (unlikely(class->objs_per_zspage
== 1 &&
1101 class->pages_per_zspage
== 1))
1102 SetPageHugeObject(page
);
1104 prev_page
->freelist
= page
;
1106 if (i
== nr_pages
- 1)
1107 SetPagePrivate2(page
);
1113 * Allocate a zspage for the given size class
1115 static struct zspage
*alloc_zspage(struct zs_pool
*pool
,
1116 struct size_class
*class,
1120 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
];
1121 struct zspage
*zspage
= cache_alloc_zspage(pool
, gfp
);
1126 memset(zspage
, 0, sizeof(struct zspage
));
1127 zspage
->magic
= ZSPAGE_MAGIC
;
1128 migrate_lock_init(zspage
);
1130 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
1133 page
= alloc_page(gfp
);
1136 dec_zone_page_state(pages
[i
], NR_ZSPAGES
);
1137 __free_page(pages
[i
]);
1139 cache_free_zspage(pool
, zspage
);
1143 inc_zone_page_state(page
, NR_ZSPAGES
);
1147 create_page_chain(class, zspage
, pages
);
1148 init_zspage(class, zspage
);
1153 static struct zspage
*find_get_zspage(struct size_class
*class)
1156 struct zspage
*zspage
;
1158 for (i
= ZS_ALMOST_FULL
; i
>= ZS_EMPTY
; i
--) {
1159 zspage
= list_first_entry_or_null(&class->fullness_list
[i
],
1160 struct zspage
, list
);
1168 #ifdef CONFIG_PGTABLE_MAPPING
1169 static inline int __zs_cpu_up(struct mapping_area
*area
)
1172 * Make sure we don't leak memory if a cpu UP notification
1173 * and zs_init() race and both call zs_cpu_up() on the same cpu
1177 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
1183 static inline void __zs_cpu_down(struct mapping_area
*area
)
1186 free_vm_area(area
->vm
);
1190 static inline void *__zs_map_object(struct mapping_area
*area
,
1191 struct page
*pages
[2], int off
, int size
)
1193 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
1194 area
->vm_addr
= area
->vm
->addr
;
1195 return area
->vm_addr
+ off
;
1198 static inline void __zs_unmap_object(struct mapping_area
*area
,
1199 struct page
*pages
[2], int off
, int size
)
1201 unsigned long addr
= (unsigned long)area
->vm_addr
;
1203 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
1206 #else /* CONFIG_PGTABLE_MAPPING */
1208 static inline int __zs_cpu_up(struct mapping_area
*area
)
1211 * Make sure we don't leak memory if a cpu UP notification
1212 * and zs_init() race and both call zs_cpu_up() on the same cpu
1216 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1222 static inline void __zs_cpu_down(struct mapping_area
*area
)
1224 kfree(area
->vm_buf
);
1225 area
->vm_buf
= NULL
;
1228 static void *__zs_map_object(struct mapping_area
*area
,
1229 struct page
*pages
[2], int off
, int size
)
1233 char *buf
= area
->vm_buf
;
1235 /* disable page faults to match kmap_atomic() return conditions */
1236 pagefault_disable();
1238 /* no read fastpath */
1239 if (area
->vm_mm
== ZS_MM_WO
)
1242 sizes
[0] = PAGE_SIZE
- off
;
1243 sizes
[1] = size
- sizes
[0];
1245 /* copy object to per-cpu buffer */
1246 addr
= kmap_atomic(pages
[0]);
1247 memcpy(buf
, addr
+ off
, sizes
[0]);
1248 kunmap_atomic(addr
);
1249 addr
= kmap_atomic(pages
[1]);
1250 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1251 kunmap_atomic(addr
);
1253 return area
->vm_buf
;
1256 static void __zs_unmap_object(struct mapping_area
*area
,
1257 struct page
*pages
[2], int off
, int size
)
1263 /* no write fastpath */
1264 if (area
->vm_mm
== ZS_MM_RO
)
1268 buf
= buf
+ ZS_HANDLE_SIZE
;
1269 size
-= ZS_HANDLE_SIZE
;
1270 off
+= ZS_HANDLE_SIZE
;
1272 sizes
[0] = PAGE_SIZE
- off
;
1273 sizes
[1] = size
- sizes
[0];
1275 /* copy per-cpu buffer to object */
1276 addr
= kmap_atomic(pages
[0]);
1277 memcpy(addr
+ off
, buf
, sizes
[0]);
1278 kunmap_atomic(addr
);
1279 addr
= kmap_atomic(pages
[1]);
1280 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1281 kunmap_atomic(addr
);
1284 /* enable page faults to match kunmap_atomic() return conditions */
1288 #endif /* CONFIG_PGTABLE_MAPPING */
1290 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
1293 int ret
, cpu
= (long)pcpu
;
1294 struct mapping_area
*area
;
1297 case CPU_UP_PREPARE
:
1298 area
= &per_cpu(zs_map_area
, cpu
);
1299 ret
= __zs_cpu_up(area
);
1301 return notifier_from_errno(ret
);
1304 case CPU_UP_CANCELED
:
1305 area
= &per_cpu(zs_map_area
, cpu
);
1306 __zs_cpu_down(area
);
1313 static struct notifier_block zs_cpu_nb
= {
1314 .notifier_call
= zs_cpu_notifier
1317 static int zs_register_cpu_notifier(void)
1319 int cpu
, uninitialized_var(ret
);
1321 cpu_notifier_register_begin();
1323 __register_cpu_notifier(&zs_cpu_nb
);
1324 for_each_online_cpu(cpu
) {
1325 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1326 if (notifier_to_errno(ret
))
1330 cpu_notifier_register_done();
1331 return notifier_to_errno(ret
);
1334 static void zs_unregister_cpu_notifier(void)
1338 cpu_notifier_register_begin();
1340 for_each_online_cpu(cpu
)
1341 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
1342 __unregister_cpu_notifier(&zs_cpu_nb
);
1344 cpu_notifier_register_done();
1347 static void __init
init_zs_size_classes(void)
1351 nr
= (ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) / ZS_SIZE_CLASS_DELTA
+ 1;
1352 if ((ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) % ZS_SIZE_CLASS_DELTA
)
1355 zs_size_classes
= nr
;
1358 static bool can_merge(struct size_class
*prev
, int pages_per_zspage
,
1359 int objs_per_zspage
)
1361 if (prev
->pages_per_zspage
== pages_per_zspage
&&
1362 prev
->objs_per_zspage
== objs_per_zspage
)
1368 static bool zspage_full(struct size_class
*class, struct zspage
*zspage
)
1370 return get_zspage_inuse(zspage
) == class->objs_per_zspage
;
1373 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1375 return atomic_long_read(&pool
->pages_allocated
);
1377 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1380 * zs_map_object - get address of allocated object from handle.
1381 * @pool: pool from which the object was allocated
1382 * @handle: handle returned from zs_malloc
1384 * Before using an object allocated from zs_malloc, it must be mapped using
1385 * this function. When done with the object, it must be unmapped using
1388 * Only one object can be mapped per cpu at a time. There is no protection
1389 * against nested mappings.
1391 * This function returns with preemption and page faults disabled.
1393 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1396 struct zspage
*zspage
;
1398 unsigned long obj
, off
;
1399 unsigned int obj_idx
;
1401 unsigned int class_idx
;
1402 enum fullness_group fg
;
1403 struct size_class
*class;
1404 struct mapping_area
*area
;
1405 struct page
*pages
[2];
1409 * Because we use per-cpu mapping areas shared among the
1410 * pools/users, we can't allow mapping in interrupt context
1411 * because it can corrupt another users mappings.
1413 WARN_ON_ONCE(in_interrupt());
1415 /* From now on, migration cannot move the object */
1418 obj
= handle_to_obj(handle
);
1419 obj_to_location(obj
, &page
, &obj_idx
);
1420 zspage
= get_zspage(page
);
1422 /* migration cannot move any subpage in this zspage */
1423 migrate_read_lock(zspage
);
1425 get_zspage_mapping(zspage
, &class_idx
, &fg
);
1426 class = pool
->size_class
[class_idx
];
1427 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1429 area
= &get_cpu_var(zs_map_area
);
1431 if (off
+ class->size
<= PAGE_SIZE
) {
1432 /* this object is contained entirely within a page */
1433 area
->vm_addr
= kmap_atomic(page
);
1434 ret
= area
->vm_addr
+ off
;
1438 /* this object spans two pages */
1440 pages
[1] = get_next_page(page
);
1443 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1445 if (likely(!PageHugeObject(page
)))
1446 ret
+= ZS_HANDLE_SIZE
;
1450 EXPORT_SYMBOL_GPL(zs_map_object
);
1452 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1454 struct zspage
*zspage
;
1456 unsigned long obj
, off
;
1457 unsigned int obj_idx
;
1459 unsigned int class_idx
;
1460 enum fullness_group fg
;
1461 struct size_class
*class;
1462 struct mapping_area
*area
;
1464 obj
= handle_to_obj(handle
);
1465 obj_to_location(obj
, &page
, &obj_idx
);
1466 zspage
= get_zspage(page
);
1467 get_zspage_mapping(zspage
, &class_idx
, &fg
);
1468 class = pool
->size_class
[class_idx
];
1469 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1471 area
= this_cpu_ptr(&zs_map_area
);
1472 if (off
+ class->size
<= PAGE_SIZE
)
1473 kunmap_atomic(area
->vm_addr
);
1475 struct page
*pages
[2];
1478 pages
[1] = get_next_page(page
);
1481 __zs_unmap_object(area
, pages
, off
, class->size
);
1483 put_cpu_var(zs_map_area
);
1485 migrate_read_unlock(zspage
);
1488 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1490 static unsigned long obj_malloc(struct size_class
*class,
1491 struct zspage
*zspage
, unsigned long handle
)
1493 int i
, nr_page
, offset
;
1495 struct link_free
*link
;
1497 struct page
*m_page
;
1498 unsigned long m_offset
;
1501 handle
|= OBJ_ALLOCATED_TAG
;
1502 obj
= get_freeobj(zspage
);
1504 offset
= obj
* class->size
;
1505 nr_page
= offset
>> PAGE_SHIFT
;
1506 m_offset
= offset
& ~PAGE_MASK
;
1507 m_page
= get_first_page(zspage
);
1509 for (i
= 0; i
< nr_page
; i
++)
1510 m_page
= get_next_page(m_page
);
1512 vaddr
= kmap_atomic(m_page
);
1513 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1514 set_freeobj(zspage
, link
->next
>> OBJ_TAG_BITS
);
1515 if (likely(!PageHugeObject(m_page
)))
1516 /* record handle in the header of allocated chunk */
1517 link
->handle
= handle
;
1519 /* record handle to page->index */
1520 zspage
->first_page
->index
= handle
;
1522 kunmap_atomic(vaddr
);
1523 mod_zspage_inuse(zspage
, 1);
1524 zs_stat_inc(class, OBJ_USED
, 1);
1526 obj
= location_to_obj(m_page
, obj
);
1533 * zs_malloc - Allocate block of given size from pool.
1534 * @pool: pool to allocate from
1535 * @size: size of block to allocate
1536 * @gfp: gfp flags when allocating object
1538 * On success, handle to the allocated object is returned,
1540 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1542 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
, gfp_t gfp
)
1544 unsigned long handle
, obj
;
1545 struct size_class
*class;
1546 enum fullness_group newfg
;
1547 struct zspage
*zspage
;
1549 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1552 handle
= cache_alloc_handle(pool
, gfp
);
1556 /* extra space in chunk to keep the handle */
1557 size
+= ZS_HANDLE_SIZE
;
1558 class = pool
->size_class
[get_size_class_index(size
)];
1560 spin_lock(&class->lock
);
1561 zspage
= find_get_zspage(class);
1562 if (likely(zspage
)) {
1563 obj
= obj_malloc(class, zspage
, handle
);
1564 /* Now move the zspage to another fullness group, if required */
1565 fix_fullness_group(class, zspage
);
1566 record_obj(handle
, obj
);
1567 spin_unlock(&class->lock
);
1572 spin_unlock(&class->lock
);
1574 zspage
= alloc_zspage(pool
, class, gfp
);
1576 cache_free_handle(pool
, handle
);
1580 spin_lock(&class->lock
);
1581 obj
= obj_malloc(class, zspage
, handle
);
1582 newfg
= get_fullness_group(class, zspage
);
1583 insert_zspage(class, zspage
, newfg
);
1584 set_zspage_mapping(zspage
, class->index
, newfg
);
1585 record_obj(handle
, obj
);
1586 atomic_long_add(class->pages_per_zspage
,
1587 &pool
->pages_allocated
);
1588 zs_stat_inc(class, OBJ_ALLOCATED
, class->objs_per_zspage
);
1590 /* We completely set up zspage so mark them as movable */
1591 SetZsPageMovable(pool
, zspage
);
1592 spin_unlock(&class->lock
);
1596 EXPORT_SYMBOL_GPL(zs_malloc
);
1598 static void obj_free(struct size_class
*class, unsigned long obj
)
1600 struct link_free
*link
;
1601 struct zspage
*zspage
;
1602 struct page
*f_page
;
1603 unsigned long f_offset
;
1604 unsigned int f_objidx
;
1607 obj
&= ~OBJ_ALLOCATED_TAG
;
1608 obj_to_location(obj
, &f_page
, &f_objidx
);
1609 f_offset
= (class->size
* f_objidx
) & ~PAGE_MASK
;
1610 zspage
= get_zspage(f_page
);
1612 vaddr
= kmap_atomic(f_page
);
1614 /* Insert this object in containing zspage's freelist */
1615 link
= (struct link_free
*)(vaddr
+ f_offset
);
1616 link
->next
= get_freeobj(zspage
) << OBJ_TAG_BITS
;
1617 kunmap_atomic(vaddr
);
1618 set_freeobj(zspage
, f_objidx
);
1619 mod_zspage_inuse(zspage
, -1);
1620 zs_stat_dec(class, OBJ_USED
, 1);
1623 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1625 struct zspage
*zspage
;
1626 struct page
*f_page
;
1628 unsigned int f_objidx
;
1630 struct size_class
*class;
1631 enum fullness_group fullness
;
1634 if (unlikely(!handle
))
1638 obj
= handle_to_obj(handle
);
1639 obj_to_location(obj
, &f_page
, &f_objidx
);
1640 zspage
= get_zspage(f_page
);
1642 migrate_read_lock(zspage
);
1644 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1645 class = pool
->size_class
[class_idx
];
1647 spin_lock(&class->lock
);
1648 obj_free(class, obj
);
1649 fullness
= fix_fullness_group(class, zspage
);
1650 if (fullness
!= ZS_EMPTY
) {
1651 migrate_read_unlock(zspage
);
1655 isolated
= is_zspage_isolated(zspage
);
1656 migrate_read_unlock(zspage
);
1657 /* If zspage is isolated, zs_page_putback will free the zspage */
1658 if (likely(!isolated
))
1659 free_zspage(pool
, class, zspage
);
1662 spin_unlock(&class->lock
);
1664 cache_free_handle(pool
, handle
);
1666 EXPORT_SYMBOL_GPL(zs_free
);
1668 static void zs_object_copy(struct size_class
*class, unsigned long dst
,
1671 struct page
*s_page
, *d_page
;
1672 unsigned int s_objidx
, d_objidx
;
1673 unsigned long s_off
, d_off
;
1674 void *s_addr
, *d_addr
;
1675 int s_size
, d_size
, size
;
1678 s_size
= d_size
= class->size
;
1680 obj_to_location(src
, &s_page
, &s_objidx
);
1681 obj_to_location(dst
, &d_page
, &d_objidx
);
1683 s_off
= (class->size
* s_objidx
) & ~PAGE_MASK
;
1684 d_off
= (class->size
* d_objidx
) & ~PAGE_MASK
;
1686 if (s_off
+ class->size
> PAGE_SIZE
)
1687 s_size
= PAGE_SIZE
- s_off
;
1689 if (d_off
+ class->size
> PAGE_SIZE
)
1690 d_size
= PAGE_SIZE
- d_off
;
1692 s_addr
= kmap_atomic(s_page
);
1693 d_addr
= kmap_atomic(d_page
);
1696 size
= min(s_size
, d_size
);
1697 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1700 if (written
== class->size
)
1708 if (s_off
>= PAGE_SIZE
) {
1709 kunmap_atomic(d_addr
);
1710 kunmap_atomic(s_addr
);
1711 s_page
= get_next_page(s_page
);
1712 s_addr
= kmap_atomic(s_page
);
1713 d_addr
= kmap_atomic(d_page
);
1714 s_size
= class->size
- written
;
1718 if (d_off
>= PAGE_SIZE
) {
1719 kunmap_atomic(d_addr
);
1720 d_page
= get_next_page(d_page
);
1721 d_addr
= kmap_atomic(d_page
);
1722 d_size
= class->size
- written
;
1727 kunmap_atomic(d_addr
);
1728 kunmap_atomic(s_addr
);
1732 * Find alloced object in zspage from index object and
1735 static unsigned long find_alloced_obj(struct size_class
*class,
1736 struct page
*page
, int *obj_idx
)
1740 int index
= *obj_idx
;
1741 unsigned long handle
= 0;
1742 void *addr
= kmap_atomic(page
);
1744 offset
= get_first_obj_offset(page
);
1745 offset
+= class->size
* index
;
1747 while (offset
< PAGE_SIZE
) {
1748 head
= obj_to_head(page
, addr
+ offset
);
1749 if (head
& OBJ_ALLOCATED_TAG
) {
1750 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1751 if (trypin_tag(handle
))
1756 offset
+= class->size
;
1760 kunmap_atomic(addr
);
1767 struct zs_compact_control
{
1768 /* Source spage for migration which could be a subpage of zspage */
1769 struct page
*s_page
;
1770 /* Destination page for migration which should be a first page
1772 struct page
*d_page
;
1773 /* Starting object index within @s_page which used for live object
1774 * in the subpage. */
1777 unsigned long nr_migrated_obj
;
1778 unsigned long nr_freed_pages
;
1781 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1782 struct zs_compact_control
*cc
)
1784 unsigned long used_obj
, free_obj
;
1785 unsigned long handle
;
1786 struct page
*s_page
= cc
->s_page
;
1787 struct page
*d_page
= cc
->d_page
;
1788 int obj_idx
= cc
->obj_idx
;
1792 handle
= find_alloced_obj(class, s_page
, &obj_idx
);
1794 s_page
= get_next_page(s_page
);
1801 /* Stop if there is no more space */
1802 if (zspage_full(class, get_zspage(d_page
))) {
1808 used_obj
= handle_to_obj(handle
);
1809 free_obj
= obj_malloc(class, get_zspage(d_page
), handle
);
1810 zs_object_copy(class, free_obj
, used_obj
);
1812 cc
->nr_migrated_obj
++;
1814 * record_obj updates handle's value to free_obj and it will
1815 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1816 * breaks synchronization using pin_tag(e,g, zs_free) so
1817 * let's keep the lock bit.
1819 free_obj
|= BIT(HANDLE_PIN_BIT
);
1820 record_obj(handle
, free_obj
);
1822 obj_free(class, used_obj
);
1825 /* Remember last position in this iteration */
1826 cc
->s_page
= s_page
;
1827 cc
->obj_idx
= obj_idx
;
1832 static struct zspage
*isolate_zspage(struct size_class
*class, bool source
)
1835 struct zspage
*zspage
;
1836 enum fullness_group fg
[2] = {ZS_ALMOST_EMPTY
, ZS_ALMOST_FULL
};
1839 fg
[0] = ZS_ALMOST_FULL
;
1840 fg
[1] = ZS_ALMOST_EMPTY
;
1843 for (i
= 0; i
< 2; i
++) {
1844 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
[i
]],
1845 struct zspage
, list
);
1847 VM_BUG_ON(is_zspage_isolated(zspage
));
1848 remove_zspage(class, zspage
, fg
[i
]);
1857 * putback_zspage - add @zspage into right class's fullness list
1858 * @class: destination class
1859 * @zspage: target page
1861 * Return @zspage's fullness_group
1863 static enum fullness_group
putback_zspage(struct size_class
*class,
1864 struct zspage
*zspage
)
1866 enum fullness_group fullness
;
1868 VM_BUG_ON(is_zspage_isolated(zspage
));
1870 fullness
= get_fullness_group(class, zspage
);
1871 insert_zspage(class, zspage
, fullness
);
1872 set_zspage_mapping(zspage
, class->index
, fullness
);
1877 #ifdef CONFIG_COMPACTION
1878 static struct dentry
*zs_mount(struct file_system_type
*fs_type
,
1879 int flags
, const char *dev_name
, void *data
)
1881 static const struct dentry_operations ops
= {
1882 .d_dname
= simple_dname
,
1885 return mount_pseudo(fs_type
, "zsmalloc:", NULL
, &ops
, ZSMALLOC_MAGIC
);
1888 static struct file_system_type zsmalloc_fs
= {
1891 .kill_sb
= kill_anon_super
,
1894 static int zsmalloc_mount(void)
1898 zsmalloc_mnt
= kern_mount(&zsmalloc_fs
);
1899 if (IS_ERR(zsmalloc_mnt
))
1900 ret
= PTR_ERR(zsmalloc_mnt
);
1905 static void zsmalloc_unmount(void)
1907 kern_unmount(zsmalloc_mnt
);
1910 static void migrate_lock_init(struct zspage
*zspage
)
1912 rwlock_init(&zspage
->lock
);
1915 static void migrate_read_lock(struct zspage
*zspage
)
1917 read_lock(&zspage
->lock
);
1920 static void migrate_read_unlock(struct zspage
*zspage
)
1922 read_unlock(&zspage
->lock
);
1925 static void migrate_write_lock(struct zspage
*zspage
)
1927 write_lock(&zspage
->lock
);
1930 static void migrate_write_unlock(struct zspage
*zspage
)
1932 write_unlock(&zspage
->lock
);
1935 /* Number of isolated subpage for *page migration* in this zspage */
1936 static void inc_zspage_isolation(struct zspage
*zspage
)
1941 static void dec_zspage_isolation(struct zspage
*zspage
)
1946 static void replace_sub_page(struct size_class
*class, struct zspage
*zspage
,
1947 struct page
*newpage
, struct page
*oldpage
)
1950 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
] = {NULL
, };
1953 page
= get_first_page(zspage
);
1955 if (page
== oldpage
)
1956 pages
[idx
] = newpage
;
1960 } while ((page
= get_next_page(page
)) != NULL
);
1962 create_page_chain(class, zspage
, pages
);
1963 set_first_obj_offset(newpage
, get_first_obj_offset(oldpage
));
1964 if (unlikely(PageHugeObject(oldpage
)))
1965 newpage
->index
= oldpage
->index
;
1966 __SetPageMovable(newpage
, page_mapping(oldpage
));
1969 bool zs_page_isolate(struct page
*page
, isolate_mode_t mode
)
1971 struct zs_pool
*pool
;
1972 struct size_class
*class;
1974 enum fullness_group fullness
;
1975 struct zspage
*zspage
;
1976 struct address_space
*mapping
;
1979 * Page is locked so zspage couldn't be destroyed. For detail, look at
1980 * lock_zspage in free_zspage.
1982 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1983 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1985 zspage
= get_zspage(page
);
1988 * Without class lock, fullness could be stale while class_idx is okay
1989 * because class_idx is constant unless page is freed so we should get
1990 * fullness again under class lock.
1992 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1993 mapping
= page_mapping(page
);
1994 pool
= mapping
->private_data
;
1995 class = pool
->size_class
[class_idx
];
1997 spin_lock(&class->lock
);
1998 if (get_zspage_inuse(zspage
) == 0) {
1999 spin_unlock(&class->lock
);
2003 /* zspage is isolated for object migration */
2004 if (list_empty(&zspage
->list
) && !is_zspage_isolated(zspage
)) {
2005 spin_unlock(&class->lock
);
2010 * If this is first time isolation for the zspage, isolate zspage from
2011 * size_class to prevent further object allocation from the zspage.
2013 if (!list_empty(&zspage
->list
) && !is_zspage_isolated(zspage
)) {
2014 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2015 remove_zspage(class, zspage
, fullness
);
2018 inc_zspage_isolation(zspage
);
2019 spin_unlock(&class->lock
);
2024 int zs_page_migrate(struct address_space
*mapping
, struct page
*newpage
,
2025 struct page
*page
, enum migrate_mode mode
)
2027 struct zs_pool
*pool
;
2028 struct size_class
*class;
2030 enum fullness_group fullness
;
2031 struct zspage
*zspage
;
2033 void *s_addr
, *d_addr
, *addr
;
2035 unsigned long handle
, head
;
2036 unsigned long old_obj
, new_obj
;
2037 unsigned int obj_idx
;
2040 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
2041 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
2043 zspage
= get_zspage(page
);
2045 /* Concurrent compactor cannot migrate any subpage in zspage */
2046 migrate_write_lock(zspage
);
2047 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2048 pool
= mapping
->private_data
;
2049 class = pool
->size_class
[class_idx
];
2050 offset
= get_first_obj_offset(page
);
2052 spin_lock(&class->lock
);
2053 if (!get_zspage_inuse(zspage
)) {
2059 s_addr
= kmap_atomic(page
);
2060 while (pos
< PAGE_SIZE
) {
2061 head
= obj_to_head(page
, s_addr
+ pos
);
2062 if (head
& OBJ_ALLOCATED_TAG
) {
2063 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2064 if (!trypin_tag(handle
))
2071 * Here, any user cannot access all objects in the zspage so let's move.
2073 d_addr
= kmap_atomic(newpage
);
2074 memcpy(d_addr
, s_addr
, PAGE_SIZE
);
2075 kunmap_atomic(d_addr
);
2077 for (addr
= s_addr
+ offset
; addr
< s_addr
+ pos
;
2078 addr
+= class->size
) {
2079 head
= obj_to_head(page
, addr
);
2080 if (head
& OBJ_ALLOCATED_TAG
) {
2081 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2082 if (!testpin_tag(handle
))
2085 old_obj
= handle_to_obj(handle
);
2086 obj_to_location(old_obj
, &dummy
, &obj_idx
);
2087 new_obj
= (unsigned long)location_to_obj(newpage
,
2089 new_obj
|= BIT(HANDLE_PIN_BIT
);
2090 record_obj(handle
, new_obj
);
2094 replace_sub_page(class, zspage
, newpage
, page
);
2097 dec_zspage_isolation(zspage
);
2100 * Page migration is done so let's putback isolated zspage to
2101 * the list if @page is final isolated subpage in the zspage.
2103 if (!is_zspage_isolated(zspage
))
2104 putback_zspage(class, zspage
);
2110 ret
= MIGRATEPAGE_SUCCESS
;
2112 for (addr
= s_addr
+ offset
; addr
< s_addr
+ pos
;
2113 addr
+= class->size
) {
2114 head
= obj_to_head(page
, addr
);
2115 if (head
& OBJ_ALLOCATED_TAG
) {
2116 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2117 if (!testpin_tag(handle
))
2122 kunmap_atomic(s_addr
);
2124 spin_unlock(&class->lock
);
2125 migrate_write_unlock(zspage
);
2130 void zs_page_putback(struct page
*page
)
2132 struct zs_pool
*pool
;
2133 struct size_class
*class;
2135 enum fullness_group fg
;
2136 struct address_space
*mapping
;
2137 struct zspage
*zspage
;
2139 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
2140 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
2142 zspage
= get_zspage(page
);
2143 get_zspage_mapping(zspage
, &class_idx
, &fg
);
2144 mapping
= page_mapping(page
);
2145 pool
= mapping
->private_data
;
2146 class = pool
->size_class
[class_idx
];
2148 spin_lock(&class->lock
);
2149 dec_zspage_isolation(zspage
);
2150 if (!is_zspage_isolated(zspage
)) {
2151 fg
= putback_zspage(class, zspage
);
2153 * Due to page_lock, we cannot free zspage immediately
2157 schedule_work(&pool
->free_work
);
2159 spin_unlock(&class->lock
);
2162 const struct address_space_operations zsmalloc_aops
= {
2163 .isolate_page
= zs_page_isolate
,
2164 .migratepage
= zs_page_migrate
,
2165 .putback_page
= zs_page_putback
,
2168 static int zs_register_migration(struct zs_pool
*pool
)
2170 pool
->inode
= alloc_anon_inode(zsmalloc_mnt
->mnt_sb
);
2171 if (IS_ERR(pool
->inode
)) {
2176 pool
->inode
->i_mapping
->private_data
= pool
;
2177 pool
->inode
->i_mapping
->a_ops
= &zsmalloc_aops
;
2181 static void zs_unregister_migration(struct zs_pool
*pool
)
2183 flush_work(&pool
->free_work
);
2188 * Caller should hold page_lock of all pages in the zspage
2189 * In here, we cannot use zspage meta data.
2191 static void async_free_zspage(struct work_struct
*work
)
2194 struct size_class
*class;
2195 unsigned int class_idx
;
2196 enum fullness_group fullness
;
2197 struct zspage
*zspage
, *tmp
;
2198 LIST_HEAD(free_pages
);
2199 struct zs_pool
*pool
= container_of(work
, struct zs_pool
,
2202 for (i
= 0; i
< zs_size_classes
; i
++) {
2203 class = pool
->size_class
[i
];
2204 if (class->index
!= i
)
2207 spin_lock(&class->lock
);
2208 list_splice_init(&class->fullness_list
[ZS_EMPTY
], &free_pages
);
2209 spin_unlock(&class->lock
);
2213 list_for_each_entry_safe(zspage
, tmp
, &free_pages
, list
) {
2214 list_del(&zspage
->list
);
2215 lock_zspage(zspage
);
2217 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2218 VM_BUG_ON(fullness
!= ZS_EMPTY
);
2219 class = pool
->size_class
[class_idx
];
2220 spin_lock(&class->lock
);
2221 __free_zspage(pool
, pool
->size_class
[class_idx
], zspage
);
2222 spin_unlock(&class->lock
);
2226 static void kick_deferred_free(struct zs_pool
*pool
)
2228 schedule_work(&pool
->free_work
);
2231 static void init_deferred_free(struct zs_pool
*pool
)
2233 INIT_WORK(&pool
->free_work
, async_free_zspage
);
2236 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
)
2238 struct page
*page
= get_first_page(zspage
);
2241 WARN_ON(!trylock_page(page
));
2242 __SetPageMovable(page
, pool
->inode
->i_mapping
);
2244 } while ((page
= get_next_page(page
)) != NULL
);
2250 * Based on the number of unused allocated objects calculate
2251 * and return the number of pages that we can free.
2253 static unsigned long zs_can_compact(struct size_class
*class)
2255 unsigned long obj_wasted
;
2256 unsigned long obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
2257 unsigned long obj_used
= zs_stat_get(class, OBJ_USED
);
2259 if (obj_allocated
<= obj_used
)
2262 obj_wasted
= obj_allocated
- obj_used
;
2263 obj_wasted
/= class->objs_per_zspage
;
2265 return obj_wasted
* class->pages_per_zspage
;
2268 static void __zs_compact(struct zs_pool
*pool
, struct size_class
*class)
2270 struct zs_compact_control cc
= {
2271 .nr_migrated_obj
= 0,
2272 .nr_freed_pages
= 0,
2274 struct zspage
*src_zspage
;
2275 struct zspage
*dst_zspage
= NULL
;
2277 spin_lock(&class->lock
);
2278 while ((src_zspage
= isolate_zspage(class, true))) {
2280 if (!zs_can_compact(class))
2284 cc
.s_page
= get_first_page(src_zspage
);
2286 while ((dst_zspage
= isolate_zspage(class, false))) {
2287 cc
.d_page
= get_first_page(dst_zspage
);
2289 * If there is no more space in dst_page, resched
2290 * and see if anyone had allocated another zspage.
2292 if (!migrate_zspage(pool
, class, &cc
))
2295 putback_zspage(class, dst_zspage
);
2298 /* Stop if we couldn't find slot */
2299 if (dst_zspage
== NULL
)
2302 putback_zspage(class, dst_zspage
);
2303 if (putback_zspage(class, src_zspage
) == ZS_EMPTY
) {
2304 free_zspage(pool
, class, src_zspage
);
2305 cc
.nr_freed_pages
+= class->pages_per_zspage
;
2307 spin_unlock(&class->lock
);
2309 spin_lock(&class->lock
);
2313 putback_zspage(class, src_zspage
);
2315 spin_unlock(&class->lock
);
2317 pool
->stats
.pages_compacted
+= cc
.nr_freed_pages
;
2318 trace_zs_compact(class->index
, cc
.nr_migrated_obj
, cc
.nr_freed_pages
);
2321 unsigned long zs_compact(struct zs_pool
*pool
)
2324 struct size_class
*class;
2325 unsigned long pages_compacted_before
= pool
->stats
.pages_compacted
;
2327 trace_zs_compact_start(pool
->name
);
2329 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
2330 class = pool
->size_class
[i
];
2333 if (class->index
!= i
)
2335 __zs_compact(pool
, class);
2338 trace_zs_compact_end(pool
->name
,
2339 pool
->stats
.pages_compacted
- pages_compacted_before
);
2341 return pool
->stats
.pages_compacted
;
2343 EXPORT_SYMBOL_GPL(zs_compact
);
2345 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
2347 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
2349 EXPORT_SYMBOL_GPL(zs_pool_stats
);
2351 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
2352 struct shrink_control
*sc
)
2354 unsigned long pages_freed
;
2355 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2358 pages_freed
= pool
->stats
.pages_compacted
;
2360 * Compact classes and calculate compaction delta.
2361 * Can run concurrently with a manually triggered
2362 * (by user) compaction.
2364 pages_freed
= zs_compact(pool
) - pages_freed
;
2366 return pages_freed
? pages_freed
: SHRINK_STOP
;
2369 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
2370 struct shrink_control
*sc
)
2373 struct size_class
*class;
2374 unsigned long pages_to_free
= 0;
2375 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2378 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
2379 class = pool
->size_class
[i
];
2382 if (class->index
!= i
)
2385 pages_to_free
+= zs_can_compact(class);
2388 return pages_to_free
;
2391 static void zs_unregister_shrinker(struct zs_pool
*pool
)
2393 if (pool
->shrinker_enabled
) {
2394 unregister_shrinker(&pool
->shrinker
);
2395 pool
->shrinker_enabled
= false;
2399 static int zs_register_shrinker(struct zs_pool
*pool
)
2401 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
2402 pool
->shrinker
.count_objects
= zs_shrinker_count
;
2403 pool
->shrinker
.batch
= 0;
2404 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
2406 return register_shrinker(&pool
->shrinker
);
2410 * zs_create_pool - Creates an allocation pool to work from.
2411 * @name: pool name to be created
2413 * This function must be called before anything when using
2414 * the zsmalloc allocator.
2416 * On success, a pointer to the newly created pool is returned,
2419 struct zs_pool
*zs_create_pool(const char *name
)
2422 struct zs_pool
*pool
;
2423 struct size_class
*prev_class
= NULL
;
2425 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
2429 init_deferred_free(pool
);
2430 pool
->size_class
= kcalloc(zs_size_classes
, sizeof(struct size_class
*),
2432 if (!pool
->size_class
) {
2437 pool
->name
= kstrdup(name
, GFP_KERNEL
);
2441 if (create_cache(pool
))
2445 * Iterate reversly, because, size of size_class that we want to use
2446 * for merging should be larger or equal to current size.
2448 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
2450 int pages_per_zspage
;
2451 int objs_per_zspage
;
2452 struct size_class
*class;
2455 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
2456 if (size
> ZS_MAX_ALLOC_SIZE
)
2457 size
= ZS_MAX_ALLOC_SIZE
;
2458 pages_per_zspage
= get_pages_per_zspage(size
);
2459 objs_per_zspage
= pages_per_zspage
* PAGE_SIZE
/ size
;
2462 * size_class is used for normal zsmalloc operation such
2463 * as alloc/free for that size. Although it is natural that we
2464 * have one size_class for each size, there is a chance that we
2465 * can get more memory utilization if we use one size_class for
2466 * many different sizes whose size_class have same
2467 * characteristics. So, we makes size_class point to
2468 * previous size_class if possible.
2471 if (can_merge(prev_class
, pages_per_zspage
, objs_per_zspage
)) {
2472 pool
->size_class
[i
] = prev_class
;
2477 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
2483 class->pages_per_zspage
= pages_per_zspage
;
2484 class->objs_per_zspage
= objs_per_zspage
;
2485 spin_lock_init(&class->lock
);
2486 pool
->size_class
[i
] = class;
2487 for (fullness
= ZS_EMPTY
; fullness
< NR_ZS_FULLNESS
;
2489 INIT_LIST_HEAD(&class->fullness_list
[fullness
]);
2494 /* debug only, don't abort if it fails */
2495 zs_pool_stat_create(pool
, name
);
2497 if (zs_register_migration(pool
))
2501 * Not critical, we still can use the pool
2502 * and user can trigger compaction manually.
2504 if (zs_register_shrinker(pool
) == 0)
2505 pool
->shrinker_enabled
= true;
2509 zs_destroy_pool(pool
);
2512 EXPORT_SYMBOL_GPL(zs_create_pool
);
2514 void zs_destroy_pool(struct zs_pool
*pool
)
2518 zs_unregister_shrinker(pool
);
2519 zs_unregister_migration(pool
);
2520 zs_pool_stat_destroy(pool
);
2522 for (i
= 0; i
< zs_size_classes
; i
++) {
2524 struct size_class
*class = pool
->size_class
[i
];
2529 if (class->index
!= i
)
2532 for (fg
= ZS_EMPTY
; fg
< NR_ZS_FULLNESS
; fg
++) {
2533 if (!list_empty(&class->fullness_list
[fg
])) {
2534 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2541 destroy_cache(pool
);
2542 kfree(pool
->size_class
);
2546 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2548 static int __init
zs_init(void)
2552 ret
= zsmalloc_mount();
2556 ret
= zs_register_cpu_notifier();
2561 init_zs_size_classes();
2564 zpool_register_driver(&zs_zpool_driver
);
2572 zs_unregister_cpu_notifier();
2578 static void __exit
zs_exit(void)
2581 zpool_unregister_driver(&zs_zpool_driver
);
2584 zs_unregister_cpu_notifier();
2589 module_init(zs_init
);
2590 module_exit(zs_exit
);
2592 MODULE_LICENSE("Dual BSD/GPL");
2593 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");