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
24 * Usage of struct page flags:
25 * PG_private: identifies the first component page
26 * PG_private2: identifies the last component page
27 * PG_owner_priv_1: indentifies 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/bitops.h>
37 #include <linux/errno.h>
38 #include <linux/highmem.h>
39 #include <linux/string.h>
40 #include <linux/slab.h>
41 #include <asm/tlbflush.h>
42 #include <asm/pgtable.h>
43 #include <linux/cpumask.h>
44 #include <linux/cpu.h>
45 #include <linux/vmalloc.h>
46 #include <linux/preempt.h>
47 #include <linux/spinlock.h>
48 #include <linux/types.h>
49 #include <linux/debugfs.h>
50 #include <linux/zsmalloc.h>
51 #include <linux/zpool.h>
52 #include <linux/mount.h>
53 #include <linux/compaction.h>
54 #include <linux/pagemap.h>
56 #define ZSPAGE_MAGIC 0x58
59 * This must be power of 2 and greater than of equal to sizeof(link_free).
60 * These two conditions ensure that any 'struct link_free' itself doesn't
61 * span more than 1 page which avoids complex case of mapping 2 pages simply
62 * to restore link_free pointer values.
67 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
68 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
70 #define ZS_MAX_ZSPAGE_ORDER 2
71 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
73 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
76 * Object location (<PFN>, <obj_idx>) is encoded as
77 * as single (unsigned long) handle value.
79 * Note that object index <obj_idx> starts from 0.
81 * This is made more complicated by various memory models and PAE.
84 #ifndef MAX_PHYSMEM_BITS
85 #ifdef CONFIG_HIGHMEM64G
86 #define MAX_PHYSMEM_BITS 36
87 #else /* !CONFIG_HIGHMEM64G */
89 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
92 #define MAX_PHYSMEM_BITS BITS_PER_LONG
95 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
98 * Memory for allocating for handle keeps object position by
99 * encoding <page, obj_idx> and the encoded value has a room
100 * in least bit(ie, look at obj_to_location).
101 * We use the bit to synchronize between object access by
102 * user and migration.
104 #define HANDLE_PIN_BIT 0
107 * Head in allocated object should have OBJ_ALLOCATED_TAG
108 * to identify the object was allocated or not.
109 * It's okay to add the status bit in the least bit because
110 * header keeps handle which is 4byte-aligned address so we
111 * have room for two bit at least.
113 #define OBJ_ALLOCATED_TAG 1
114 #define OBJ_TAG_BITS 1
115 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
116 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
118 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
119 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
120 #define ZS_MIN_ALLOC_SIZE \
121 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
122 /* each chunk includes extra space to keep handle */
123 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
126 * On systems with 4K page size, this gives 255 size classes! There is a
128 * - Large number of size classes is potentially wasteful as free page are
129 * spread across these classes
130 * - Small number of size classes causes large internal fragmentation
131 * - Probably its better to use specific size classes (empirically
132 * determined). NOTE: all those class sizes must be set as multiple of
133 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
135 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
138 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
141 * We do not maintain any list for completely empty or full pages
143 enum fullness_group
{
161 struct zs_size_stat
{
162 unsigned long objs
[NR_ZS_STAT_TYPE
];
165 #ifdef CONFIG_ZSMALLOC_STAT
166 static struct dentry
*zs_stat_root
;
169 #ifdef CONFIG_COMPACTION
170 static struct vfsmount
*zsmalloc_mnt
;
174 * number of size_classes
176 static int zs_size_classes
;
179 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
181 * n = number of allocated objects
182 * N = total number of objects zspage can store
183 * f = fullness_threshold_frac
185 * Similarly, we assign zspage to:
186 * ZS_ALMOST_FULL when n > N / f
187 * ZS_EMPTY when n == 0
188 * ZS_FULL when n == N
190 * (see: fix_fullness_group())
192 static const int fullness_threshold_frac
= 4;
196 struct list_head fullness_list
[NR_ZS_FULLNESS
];
198 * Size of objects stored in this class. Must be multiple
203 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
204 int pages_per_zspage
;
207 struct zs_size_stat stats
;
210 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
211 static void SetPageHugeObject(struct page
*page
)
213 SetPageOwnerPriv1(page
);
216 static void ClearPageHugeObject(struct page
*page
)
218 ClearPageOwnerPriv1(page
);
221 static int PageHugeObject(struct page
*page
)
223 return PageOwnerPriv1(page
);
227 * Placed within free objects to form a singly linked list.
228 * For every zspage, zspage->freeobj gives head of this list.
230 * This must be power of 2 and less than or equal to ZS_ALIGN
236 * It's valid for non-allocated object
240 * Handle of allocated object.
242 unsigned long handle
;
249 struct size_class
**size_class
;
250 struct kmem_cache
*handle_cachep
;
251 struct kmem_cache
*zspage_cachep
;
253 atomic_long_t pages_allocated
;
255 struct zs_pool_stats stats
;
257 /* Compact classes */
258 struct shrinker shrinker
;
260 * To signify that register_shrinker() was successful
261 * and unregister_shrinker() will not Oops.
263 bool shrinker_enabled
;
264 #ifdef CONFIG_ZSMALLOC_STAT
265 struct dentry
*stat_dentry
;
267 #ifdef CONFIG_COMPACTION
269 struct work_struct free_work
;
274 * A zspage's class index and fullness group
275 * are encoded in its (first)page->mapping
277 #define FULLNESS_BITS 2
279 #define ISOLATED_BITS 3
280 #define MAGIC_VAL_BITS 8
284 unsigned int fullness
:FULLNESS_BITS
;
285 unsigned int class:CLASS_BITS
;
286 unsigned int isolated
:ISOLATED_BITS
;
287 unsigned int magic
:MAGIC_VAL_BITS
;
290 unsigned int freeobj
;
291 struct page
*first_page
;
292 struct list_head list
; /* fullness list */
293 #ifdef CONFIG_COMPACTION
298 struct mapping_area
{
299 #ifdef CONFIG_PGTABLE_MAPPING
300 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
302 char *vm_buf
; /* copy buffer for objects that span pages */
304 char *vm_addr
; /* address of kmap_atomic()'ed pages */
305 enum zs_mapmode vm_mm
; /* mapping mode */
308 #ifdef CONFIG_COMPACTION
309 static int zs_register_migration(struct zs_pool
*pool
);
310 static void zs_unregister_migration(struct zs_pool
*pool
);
311 static void migrate_lock_init(struct zspage
*zspage
);
312 static void migrate_read_lock(struct zspage
*zspage
);
313 static void migrate_read_unlock(struct zspage
*zspage
);
314 static void kick_deferred_free(struct zs_pool
*pool
);
315 static void init_deferred_free(struct zs_pool
*pool
);
316 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
);
318 static int zsmalloc_mount(void) { return 0; }
319 static void zsmalloc_unmount(void) {}
320 static int zs_register_migration(struct zs_pool
*pool
) { return 0; }
321 static void zs_unregister_migration(struct zs_pool
*pool
) {}
322 static void migrate_lock_init(struct zspage
*zspage
) {}
323 static void migrate_read_lock(struct zspage
*zspage
) {}
324 static void migrate_read_unlock(struct zspage
*zspage
) {}
325 static void kick_deferred_free(struct zs_pool
*pool
) {}
326 static void init_deferred_free(struct zs_pool
*pool
) {}
327 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
) {}
330 static int create_cache(struct zs_pool
*pool
)
332 pool
->handle_cachep
= kmem_cache_create("zs_handle", ZS_HANDLE_SIZE
,
334 if (!pool
->handle_cachep
)
337 pool
->zspage_cachep
= kmem_cache_create("zspage", sizeof(struct zspage
),
339 if (!pool
->zspage_cachep
) {
340 kmem_cache_destroy(pool
->handle_cachep
);
341 pool
->handle_cachep
= NULL
;
348 static void destroy_cache(struct zs_pool
*pool
)
350 kmem_cache_destroy(pool
->handle_cachep
);
351 kmem_cache_destroy(pool
->zspage_cachep
);
354 static unsigned long cache_alloc_handle(struct zs_pool
*pool
, gfp_t gfp
)
356 return (unsigned long)kmem_cache_alloc(pool
->handle_cachep
,
357 gfp
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
360 static void cache_free_handle(struct zs_pool
*pool
, unsigned long handle
)
362 kmem_cache_free(pool
->handle_cachep
, (void *)handle
);
365 static struct zspage
*cache_alloc_zspage(struct zs_pool
*pool
, gfp_t flags
)
367 return kmem_cache_alloc(pool
->zspage_cachep
,
368 flags
& ~(__GFP_HIGHMEM
|__GFP_MOVABLE
));
371 static void cache_free_zspage(struct zs_pool
*pool
, struct zspage
*zspage
)
373 kmem_cache_free(pool
->zspage_cachep
, zspage
);
376 static void record_obj(unsigned long handle
, unsigned long obj
)
379 * lsb of @obj represents handle lock while other bits
380 * represent object value the handle is pointing so
381 * updating shouldn't do store tearing.
383 WRITE_ONCE(*(unsigned long *)handle
, obj
);
390 static void *zs_zpool_create(const char *name
, gfp_t gfp
,
391 const struct zpool_ops
*zpool_ops
,
395 * Ignore global gfp flags: zs_malloc() may be invoked from
396 * different contexts and its caller must provide a valid
399 return zs_create_pool(name
);
402 static void zs_zpool_destroy(void *pool
)
404 zs_destroy_pool(pool
);
407 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
408 unsigned long *handle
)
410 *handle
= zs_malloc(pool
, size
, gfp
);
411 return *handle
? 0 : -1;
413 static void zs_zpool_free(void *pool
, unsigned long handle
)
415 zs_free(pool
, handle
);
418 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
419 unsigned int *reclaimed
)
424 static void *zs_zpool_map(void *pool
, unsigned long handle
,
425 enum zpool_mapmode mm
)
427 enum zs_mapmode zs_mm
;
436 case ZPOOL_MM_RW
: /* fallthru */
442 return zs_map_object(pool
, handle
, zs_mm
);
444 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
446 zs_unmap_object(pool
, handle
);
449 static u64
zs_zpool_total_size(void *pool
)
451 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
454 static struct zpool_driver zs_zpool_driver
= {
456 .owner
= THIS_MODULE
,
457 .create
= zs_zpool_create
,
458 .destroy
= zs_zpool_destroy
,
459 .malloc
= zs_zpool_malloc
,
460 .free
= zs_zpool_free
,
461 .shrink
= zs_zpool_shrink
,
463 .unmap
= zs_zpool_unmap
,
464 .total_size
= zs_zpool_total_size
,
467 MODULE_ALIAS("zpool-zsmalloc");
468 #endif /* CONFIG_ZPOOL */
470 static unsigned int get_maxobj_per_zspage(int size
, int pages_per_zspage
)
472 return pages_per_zspage
* PAGE_SIZE
/ size
;
475 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
476 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
478 static bool is_zspage_isolated(struct zspage
*zspage
)
480 return zspage
->isolated
;
483 static int is_first_page(struct page
*page
)
485 return PagePrivate(page
);
488 /* Protected by class->lock */
489 static inline int get_zspage_inuse(struct zspage
*zspage
)
491 return zspage
->inuse
;
494 static inline void set_zspage_inuse(struct zspage
*zspage
, int val
)
499 static inline void mod_zspage_inuse(struct zspage
*zspage
, int val
)
501 zspage
->inuse
+= val
;
504 static inline struct page
*get_first_page(struct zspage
*zspage
)
506 struct page
*first_page
= zspage
->first_page
;
508 VM_BUG_ON_PAGE(!is_first_page(first_page
), first_page
);
512 static inline int get_first_obj_offset(struct page
*page
)
517 static inline void set_first_obj_offset(struct page
*page
, int offset
)
519 page
->units
= offset
;
522 static inline unsigned int get_freeobj(struct zspage
*zspage
)
524 return zspage
->freeobj
;
527 static inline void set_freeobj(struct zspage
*zspage
, unsigned int obj
)
529 zspage
->freeobj
= obj
;
532 static void get_zspage_mapping(struct zspage
*zspage
,
533 unsigned int *class_idx
,
534 enum fullness_group
*fullness
)
536 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
538 *fullness
= zspage
->fullness
;
539 *class_idx
= zspage
->class;
542 static void set_zspage_mapping(struct zspage
*zspage
,
543 unsigned int class_idx
,
544 enum fullness_group fullness
)
546 zspage
->class = class_idx
;
547 zspage
->fullness
= fullness
;
551 * zsmalloc divides the pool into various size classes where each
552 * class maintains a list of zspages where each zspage is divided
553 * into equal sized chunks. Each allocation falls into one of these
554 * classes depending on its size. This function returns index of the
555 * size class which has chunk size big enough to hold the give size.
557 static int get_size_class_index(int size
)
561 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
562 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
563 ZS_SIZE_CLASS_DELTA
);
565 return min(zs_size_classes
- 1, idx
);
568 static inline void zs_stat_inc(struct size_class
*class,
569 enum zs_stat_type type
, unsigned long cnt
)
571 class->stats
.objs
[type
] += cnt
;
574 static inline void zs_stat_dec(struct size_class
*class,
575 enum zs_stat_type type
, unsigned long cnt
)
577 class->stats
.objs
[type
] -= cnt
;
580 static inline unsigned long zs_stat_get(struct size_class
*class,
581 enum zs_stat_type type
)
583 return class->stats
.objs
[type
];
586 #ifdef CONFIG_ZSMALLOC_STAT
588 static void __init
zs_stat_init(void)
590 if (!debugfs_initialized()) {
591 pr_warn("debugfs not available, stat dir not created\n");
595 zs_stat_root
= debugfs_create_dir("zsmalloc", NULL
);
597 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
600 static void __exit
zs_stat_exit(void)
602 debugfs_remove_recursive(zs_stat_root
);
605 static unsigned long zs_can_compact(struct size_class
*class);
607 static int zs_stats_size_show(struct seq_file
*s
, void *v
)
610 struct zs_pool
*pool
= s
->private;
611 struct size_class
*class;
613 unsigned long class_almost_full
, class_almost_empty
;
614 unsigned long obj_allocated
, obj_used
, pages_used
, freeable
;
615 unsigned long total_class_almost_full
= 0, total_class_almost_empty
= 0;
616 unsigned long total_objs
= 0, total_used_objs
= 0, total_pages
= 0;
617 unsigned long total_freeable
= 0;
619 seq_printf(s
, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
620 "class", "size", "almost_full", "almost_empty",
621 "obj_allocated", "obj_used", "pages_used",
622 "pages_per_zspage", "freeable");
624 for (i
= 0; i
< zs_size_classes
; i
++) {
625 class = pool
->size_class
[i
];
627 if (class->index
!= i
)
630 spin_lock(&class->lock
);
631 class_almost_full
= zs_stat_get(class, CLASS_ALMOST_FULL
);
632 class_almost_empty
= zs_stat_get(class, CLASS_ALMOST_EMPTY
);
633 obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
634 obj_used
= zs_stat_get(class, OBJ_USED
);
635 freeable
= zs_can_compact(class);
636 spin_unlock(&class->lock
);
638 objs_per_zspage
= get_maxobj_per_zspage(class->size
,
639 class->pages_per_zspage
);
640 pages_used
= obj_allocated
/ objs_per_zspage
*
641 class->pages_per_zspage
;
643 seq_printf(s
, " %5u %5u %11lu %12lu %13lu"
644 " %10lu %10lu %16d %8lu\n",
645 i
, class->size
, class_almost_full
, class_almost_empty
,
646 obj_allocated
, obj_used
, pages_used
,
647 class->pages_per_zspage
, freeable
);
649 total_class_almost_full
+= class_almost_full
;
650 total_class_almost_empty
+= class_almost_empty
;
651 total_objs
+= obj_allocated
;
652 total_used_objs
+= obj_used
;
653 total_pages
+= pages_used
;
654 total_freeable
+= freeable
;
658 seq_printf(s
, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
659 "Total", "", total_class_almost_full
,
660 total_class_almost_empty
, total_objs
,
661 total_used_objs
, total_pages
, "", total_freeable
);
666 static int zs_stats_size_open(struct inode
*inode
, struct file
*file
)
668 return single_open(file
, zs_stats_size_show
, inode
->i_private
);
671 static const struct file_operations zs_stat_size_ops
= {
672 .open
= zs_stats_size_open
,
675 .release
= single_release
,
678 static void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
680 struct dentry
*entry
;
683 pr_warn("no root stat dir, not creating <%s> stat dir\n", name
);
687 entry
= debugfs_create_dir(name
, zs_stat_root
);
689 pr_warn("debugfs dir <%s> creation failed\n", name
);
692 pool
->stat_dentry
= entry
;
694 entry
= debugfs_create_file("classes", S_IFREG
| S_IRUGO
,
695 pool
->stat_dentry
, pool
, &zs_stat_size_ops
);
697 pr_warn("%s: debugfs file entry <%s> creation failed\n",
699 debugfs_remove_recursive(pool
->stat_dentry
);
700 pool
->stat_dentry
= NULL
;
704 static void zs_pool_stat_destroy(struct zs_pool
*pool
)
706 debugfs_remove_recursive(pool
->stat_dentry
);
709 #else /* CONFIG_ZSMALLOC_STAT */
710 static void __init
zs_stat_init(void)
714 static void __exit
zs_stat_exit(void)
718 static inline void zs_pool_stat_create(struct zs_pool
*pool
, const char *name
)
722 static inline void zs_pool_stat_destroy(struct zs_pool
*pool
)
729 * For each size class, zspages are divided into different groups
730 * depending on how "full" they are. This was done so that we could
731 * easily find empty or nearly empty zspages when we try to shrink
732 * the pool (not yet implemented). This function returns fullness
733 * status of the given page.
735 static enum fullness_group
get_fullness_group(struct size_class
*class,
736 struct zspage
*zspage
)
738 int inuse
, objs_per_zspage
;
739 enum fullness_group fg
;
741 inuse
= get_zspage_inuse(zspage
);
742 objs_per_zspage
= class->objs_per_zspage
;
746 else if (inuse
== objs_per_zspage
)
748 else if (inuse
<= 3 * objs_per_zspage
/ fullness_threshold_frac
)
749 fg
= ZS_ALMOST_EMPTY
;
757 * Each size class maintains various freelists and zspages are assigned
758 * to one of these freelists based on the number of live objects they
759 * have. This functions inserts the given zspage into the freelist
760 * identified by <class, fullness_group>.
762 static void insert_zspage(struct size_class
*class,
763 struct zspage
*zspage
,
764 enum fullness_group fullness
)
768 zs_stat_inc(class, fullness
, 1);
769 head
= list_first_entry_or_null(&class->fullness_list
[fullness
],
770 struct zspage
, list
);
772 * We want to see more ZS_FULL pages and less almost empty/full.
773 * Put pages with higher ->inuse first.
776 if (get_zspage_inuse(zspage
) < get_zspage_inuse(head
)) {
777 list_add(&zspage
->list
, &head
->list
);
781 list_add(&zspage
->list
, &class->fullness_list
[fullness
]);
785 * This function removes the given zspage from the freelist identified
786 * by <class, fullness_group>.
788 static void remove_zspage(struct size_class
*class,
789 struct zspage
*zspage
,
790 enum fullness_group fullness
)
792 VM_BUG_ON(list_empty(&class->fullness_list
[fullness
]));
793 VM_BUG_ON(is_zspage_isolated(zspage
));
795 list_del_init(&zspage
->list
);
796 zs_stat_dec(class, fullness
, 1);
800 * Each size class maintains zspages in different fullness groups depending
801 * on the number of live objects they contain. When allocating or freeing
802 * objects, the fullness status of the page can change, say, from ALMOST_FULL
803 * to ALMOST_EMPTY when freeing an object. This function checks if such
804 * a status change has occurred for the given page and accordingly moves the
805 * page from the freelist of the old fullness group to that of the new
808 static enum fullness_group
fix_fullness_group(struct size_class
*class,
809 struct zspage
*zspage
)
812 enum fullness_group currfg
, newfg
;
814 get_zspage_mapping(zspage
, &class_idx
, &currfg
);
815 newfg
= get_fullness_group(class, zspage
);
819 if (!is_zspage_isolated(zspage
)) {
820 remove_zspage(class, zspage
, currfg
);
821 insert_zspage(class, zspage
, newfg
);
824 set_zspage_mapping(zspage
, class_idx
, newfg
);
831 * We have to decide on how many pages to link together
832 * to form a zspage for each size class. This is important
833 * to reduce wastage due to unusable space left at end of
834 * each zspage which is given as:
835 * wastage = Zp % class_size
836 * usage = Zp - wastage
837 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
839 * For example, for size class of 3/8 * PAGE_SIZE, we should
840 * link together 3 PAGE_SIZE sized pages to form a zspage
841 * since then we can perfectly fit in 8 such objects.
843 static int get_pages_per_zspage(int class_size
)
845 int i
, max_usedpc
= 0;
846 /* zspage order which gives maximum used size per KB */
847 int max_usedpc_order
= 1;
849 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
853 zspage_size
= i
* PAGE_SIZE
;
854 waste
= zspage_size
% class_size
;
855 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
857 if (usedpc
> max_usedpc
) {
859 max_usedpc_order
= i
;
863 return max_usedpc_order
;
866 static struct zspage
*get_zspage(struct page
*page
)
868 struct zspage
*zspage
= (struct zspage
*)page
->private;
870 BUG_ON(zspage
->magic
!= ZSPAGE_MAGIC
);
874 static struct page
*get_next_page(struct page
*page
)
876 if (unlikely(PageHugeObject(page
)))
879 return page
->freelist
;
883 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
884 * @page: page object resides in zspage
885 * @obj_idx: object index
887 static void obj_to_location(unsigned long obj
, struct page
**page
,
888 unsigned int *obj_idx
)
890 obj
>>= OBJ_TAG_BITS
;
891 *page
= pfn_to_page(obj
>> OBJ_INDEX_BITS
);
892 *obj_idx
= (obj
& OBJ_INDEX_MASK
);
896 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
897 * @page: page object resides in zspage
898 * @obj_idx: object index
900 static unsigned long location_to_obj(struct page
*page
, unsigned int obj_idx
)
904 obj
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
905 obj
|= obj_idx
& OBJ_INDEX_MASK
;
906 obj
<<= OBJ_TAG_BITS
;
911 static unsigned long handle_to_obj(unsigned long handle
)
913 return *(unsigned long *)handle
;
916 static unsigned long obj_to_head(struct page
*page
, void *obj
)
918 if (unlikely(PageHugeObject(page
))) {
919 VM_BUG_ON_PAGE(!is_first_page(page
), page
);
922 return *(unsigned long *)obj
;
925 static inline int testpin_tag(unsigned long handle
)
927 return bit_spin_is_locked(HANDLE_PIN_BIT
, (unsigned long *)handle
);
930 static inline int trypin_tag(unsigned long handle
)
932 return bit_spin_trylock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
935 static void pin_tag(unsigned long handle
)
937 bit_spin_lock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
940 static void unpin_tag(unsigned long handle
)
942 bit_spin_unlock(HANDLE_PIN_BIT
, (unsigned long *)handle
);
945 static void reset_page(struct page
*page
)
947 __ClearPageMovable(page
);
948 clear_bit(PG_private
, &page
->flags
);
949 clear_bit(PG_private_2
, &page
->flags
);
950 set_page_private(page
, 0);
951 page_mapcount_reset(page
);
952 ClearPageHugeObject(page
);
953 page
->freelist
= NULL
;
957 * To prevent zspage destroy during migration, zspage freeing should
958 * hold locks of all pages in the zspage.
960 void lock_zspage(struct zspage
*zspage
)
962 struct page
*page
= get_first_page(zspage
);
966 } while ((page
= get_next_page(page
)) != NULL
);
969 int trylock_zspage(struct zspage
*zspage
)
971 struct page
*cursor
, *fail
;
973 for (cursor
= get_first_page(zspage
); cursor
!= NULL
; cursor
=
974 get_next_page(cursor
)) {
975 if (!trylock_page(cursor
)) {
983 for (cursor
= get_first_page(zspage
); cursor
!= fail
; cursor
=
984 get_next_page(cursor
))
990 static void __free_zspage(struct zs_pool
*pool
, struct size_class
*class,
991 struct zspage
*zspage
)
993 struct page
*page
, *next
;
994 enum fullness_group fg
;
995 unsigned int class_idx
;
997 get_zspage_mapping(zspage
, &class_idx
, &fg
);
999 assert_spin_locked(&class->lock
);
1001 VM_BUG_ON(get_zspage_inuse(zspage
));
1002 VM_BUG_ON(fg
!= ZS_EMPTY
);
1004 next
= page
= get_first_page(zspage
);
1006 VM_BUG_ON_PAGE(!PageLocked(page
), page
);
1007 next
= get_next_page(page
);
1012 } while (page
!= NULL
);
1014 cache_free_zspage(pool
, zspage
);
1016 zs_stat_dec(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1017 class->size
, class->pages_per_zspage
));
1018 atomic_long_sub(class->pages_per_zspage
,
1019 &pool
->pages_allocated
);
1022 static void free_zspage(struct zs_pool
*pool
, struct size_class
*class,
1023 struct zspage
*zspage
)
1025 VM_BUG_ON(get_zspage_inuse(zspage
));
1026 VM_BUG_ON(list_empty(&zspage
->list
));
1028 if (!trylock_zspage(zspage
)) {
1029 kick_deferred_free(pool
);
1033 remove_zspage(class, zspage
, ZS_EMPTY
);
1034 __free_zspage(pool
, class, zspage
);
1037 /* Initialize a newly allocated zspage */
1038 static void init_zspage(struct size_class
*class, struct zspage
*zspage
)
1040 unsigned int freeobj
= 1;
1041 unsigned long off
= 0;
1042 struct page
*page
= get_first_page(zspage
);
1045 struct page
*next_page
;
1046 struct link_free
*link
;
1049 set_first_obj_offset(page
, off
);
1051 vaddr
= kmap_atomic(page
);
1052 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
1054 while ((off
+= class->size
) < PAGE_SIZE
) {
1055 link
->next
= freeobj
++ << OBJ_ALLOCATED_TAG
;
1056 link
+= class->size
/ sizeof(*link
);
1060 * We now come to the last (full or partial) object on this
1061 * page, which must point to the first object on the next
1064 next_page
= get_next_page(page
);
1066 link
->next
= freeobj
++ << OBJ_ALLOCATED_TAG
;
1069 * Reset OBJ_ALLOCATED_TAG bit to last link to tell
1070 * whether it's allocated object or not.
1072 link
->next
= -1 << OBJ_ALLOCATED_TAG
;
1074 kunmap_atomic(vaddr
);
1079 set_freeobj(zspage
, 0);
1082 static void create_page_chain(struct size_class
*class, struct zspage
*zspage
,
1083 struct page
*pages
[])
1087 struct page
*prev_page
= NULL
;
1088 int nr_pages
= class->pages_per_zspage
;
1091 * Allocate individual pages and link them together as:
1092 * 1. all pages are linked together using page->freelist
1093 * 2. each sub-page point to zspage using page->private
1095 * we set PG_private to identify the first page (i.e. no other sub-page
1096 * has this flag set) and PG_private_2 to identify the last page.
1098 for (i
= 0; i
< nr_pages
; i
++) {
1100 set_page_private(page
, (unsigned long)zspage
);
1101 page
->freelist
= NULL
;
1103 zspage
->first_page
= page
;
1104 SetPagePrivate(page
);
1105 if (unlikely(class->objs_per_zspage
== 1 &&
1106 class->pages_per_zspage
== 1))
1107 SetPageHugeObject(page
);
1109 prev_page
->freelist
= page
;
1111 if (i
== nr_pages
- 1)
1112 SetPagePrivate2(page
);
1118 * Allocate a zspage for the given size class
1120 static struct zspage
*alloc_zspage(struct zs_pool
*pool
,
1121 struct size_class
*class,
1125 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
];
1126 struct zspage
*zspage
= cache_alloc_zspage(pool
, gfp
);
1131 memset(zspage
, 0, sizeof(struct zspage
));
1132 zspage
->magic
= ZSPAGE_MAGIC
;
1133 migrate_lock_init(zspage
);
1135 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
1138 page
= alloc_page(gfp
);
1141 __free_page(pages
[i
]);
1142 cache_free_zspage(pool
, zspage
);
1148 create_page_chain(class, zspage
, pages
);
1149 init_zspage(class, zspage
);
1154 static struct zspage
*find_get_zspage(struct size_class
*class)
1157 struct zspage
*zspage
;
1159 for (i
= ZS_ALMOST_FULL
; i
>= ZS_EMPTY
; i
--) {
1160 zspage
= list_first_entry_or_null(&class->fullness_list
[i
],
1161 struct zspage
, list
);
1169 #ifdef CONFIG_PGTABLE_MAPPING
1170 static inline int __zs_cpu_up(struct mapping_area
*area
)
1173 * Make sure we don't leak memory if a cpu UP notification
1174 * and zs_init() race and both call zs_cpu_up() on the same cpu
1178 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
1184 static inline void __zs_cpu_down(struct mapping_area
*area
)
1187 free_vm_area(area
->vm
);
1191 static inline void *__zs_map_object(struct mapping_area
*area
,
1192 struct page
*pages
[2], int off
, int size
)
1194 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
1195 area
->vm_addr
= area
->vm
->addr
;
1196 return area
->vm_addr
+ off
;
1199 static inline void __zs_unmap_object(struct mapping_area
*area
,
1200 struct page
*pages
[2], int off
, int size
)
1202 unsigned long addr
= (unsigned long)area
->vm_addr
;
1204 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
1207 #else /* CONFIG_PGTABLE_MAPPING */
1209 static inline int __zs_cpu_up(struct mapping_area
*area
)
1212 * Make sure we don't leak memory if a cpu UP notification
1213 * and zs_init() race and both call zs_cpu_up() on the same cpu
1217 area
->vm_buf
= kmalloc(ZS_MAX_ALLOC_SIZE
, GFP_KERNEL
);
1223 static inline void __zs_cpu_down(struct mapping_area
*area
)
1225 kfree(area
->vm_buf
);
1226 area
->vm_buf
= NULL
;
1229 static void *__zs_map_object(struct mapping_area
*area
,
1230 struct page
*pages
[2], int off
, int size
)
1234 char *buf
= area
->vm_buf
;
1236 /* disable page faults to match kmap_atomic() return conditions */
1237 pagefault_disable();
1239 /* no read fastpath */
1240 if (area
->vm_mm
== ZS_MM_WO
)
1243 sizes
[0] = PAGE_SIZE
- off
;
1244 sizes
[1] = size
- sizes
[0];
1246 /* copy object to per-cpu buffer */
1247 addr
= kmap_atomic(pages
[0]);
1248 memcpy(buf
, addr
+ off
, sizes
[0]);
1249 kunmap_atomic(addr
);
1250 addr
= kmap_atomic(pages
[1]);
1251 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
1252 kunmap_atomic(addr
);
1254 return area
->vm_buf
;
1257 static void __zs_unmap_object(struct mapping_area
*area
,
1258 struct page
*pages
[2], int off
, int size
)
1264 /* no write fastpath */
1265 if (area
->vm_mm
== ZS_MM_RO
)
1269 buf
= buf
+ ZS_HANDLE_SIZE
;
1270 size
-= ZS_HANDLE_SIZE
;
1271 off
+= ZS_HANDLE_SIZE
;
1273 sizes
[0] = PAGE_SIZE
- off
;
1274 sizes
[1] = size
- sizes
[0];
1276 /* copy per-cpu buffer to object */
1277 addr
= kmap_atomic(pages
[0]);
1278 memcpy(addr
+ off
, buf
, sizes
[0]);
1279 kunmap_atomic(addr
);
1280 addr
= kmap_atomic(pages
[1]);
1281 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
1282 kunmap_atomic(addr
);
1285 /* enable page faults to match kunmap_atomic() return conditions */
1289 #endif /* CONFIG_PGTABLE_MAPPING */
1291 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
1294 int ret
, cpu
= (long)pcpu
;
1295 struct mapping_area
*area
;
1298 case CPU_UP_PREPARE
:
1299 area
= &per_cpu(zs_map_area
, cpu
);
1300 ret
= __zs_cpu_up(area
);
1302 return notifier_from_errno(ret
);
1305 case CPU_UP_CANCELED
:
1306 area
= &per_cpu(zs_map_area
, cpu
);
1307 __zs_cpu_down(area
);
1314 static struct notifier_block zs_cpu_nb
= {
1315 .notifier_call
= zs_cpu_notifier
1318 static int zs_register_cpu_notifier(void)
1320 int cpu
, uninitialized_var(ret
);
1322 cpu_notifier_register_begin();
1324 __register_cpu_notifier(&zs_cpu_nb
);
1325 for_each_online_cpu(cpu
) {
1326 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
1327 if (notifier_to_errno(ret
))
1331 cpu_notifier_register_done();
1332 return notifier_to_errno(ret
);
1335 static void zs_unregister_cpu_notifier(void)
1339 cpu_notifier_register_begin();
1341 for_each_online_cpu(cpu
)
1342 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
1343 __unregister_cpu_notifier(&zs_cpu_nb
);
1345 cpu_notifier_register_done();
1348 static void init_zs_size_classes(void)
1352 nr
= (ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) / ZS_SIZE_CLASS_DELTA
+ 1;
1353 if ((ZS_MAX_ALLOC_SIZE
- ZS_MIN_ALLOC_SIZE
) % ZS_SIZE_CLASS_DELTA
)
1356 zs_size_classes
= nr
;
1359 static bool can_merge(struct size_class
*prev
, int size
, int pages_per_zspage
)
1361 if (prev
->pages_per_zspage
!= pages_per_zspage
)
1364 if (get_maxobj_per_zspage(prev
->size
, prev
->pages_per_zspage
)
1365 != get_maxobj_per_zspage(size
, pages_per_zspage
))
1371 static bool zspage_full(struct size_class
*class, struct zspage
*zspage
)
1373 return get_zspage_inuse(zspage
) == class->objs_per_zspage
;
1376 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1378 return atomic_long_read(&pool
->pages_allocated
);
1380 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1383 * zs_map_object - get address of allocated object from handle.
1384 * @pool: pool from which the object was allocated
1385 * @handle: handle returned from zs_malloc
1387 * Before using an object allocated from zs_malloc, it must be mapped using
1388 * this function. When done with the object, it must be unmapped using
1391 * Only one object can be mapped per cpu at a time. There is no protection
1392 * against nested mappings.
1394 * This function returns with preemption and page faults disabled.
1396 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1399 struct zspage
*zspage
;
1401 unsigned long obj
, off
;
1402 unsigned int obj_idx
;
1404 unsigned int class_idx
;
1405 enum fullness_group fg
;
1406 struct size_class
*class;
1407 struct mapping_area
*area
;
1408 struct page
*pages
[2];
1412 * Because we use per-cpu mapping areas shared among the
1413 * pools/users, we can't allow mapping in interrupt context
1414 * because it can corrupt another users mappings.
1416 WARN_ON_ONCE(in_interrupt());
1418 /* From now on, migration cannot move the object */
1421 obj
= handle_to_obj(handle
);
1422 obj_to_location(obj
, &page
, &obj_idx
);
1423 zspage
= get_zspage(page
);
1425 /* migration cannot move any subpage in this zspage */
1426 migrate_read_lock(zspage
);
1428 get_zspage_mapping(zspage
, &class_idx
, &fg
);
1429 class = pool
->size_class
[class_idx
];
1430 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1432 area
= &get_cpu_var(zs_map_area
);
1434 if (off
+ class->size
<= PAGE_SIZE
) {
1435 /* this object is contained entirely within a page */
1436 area
->vm_addr
= kmap_atomic(page
);
1437 ret
= area
->vm_addr
+ off
;
1441 /* this object spans two pages */
1443 pages
[1] = get_next_page(page
);
1446 ret
= __zs_map_object(area
, pages
, off
, class->size
);
1448 if (likely(!PageHugeObject(page
)))
1449 ret
+= ZS_HANDLE_SIZE
;
1453 EXPORT_SYMBOL_GPL(zs_map_object
);
1455 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1457 struct zspage
*zspage
;
1459 unsigned long obj
, off
;
1460 unsigned int obj_idx
;
1462 unsigned int class_idx
;
1463 enum fullness_group fg
;
1464 struct size_class
*class;
1465 struct mapping_area
*area
;
1467 obj
= handle_to_obj(handle
);
1468 obj_to_location(obj
, &page
, &obj_idx
);
1469 zspage
= get_zspage(page
);
1470 get_zspage_mapping(zspage
, &class_idx
, &fg
);
1471 class = pool
->size_class
[class_idx
];
1472 off
= (class->size
* obj_idx
) & ~PAGE_MASK
;
1474 area
= this_cpu_ptr(&zs_map_area
);
1475 if (off
+ class->size
<= PAGE_SIZE
)
1476 kunmap_atomic(area
->vm_addr
);
1478 struct page
*pages
[2];
1481 pages
[1] = get_next_page(page
);
1484 __zs_unmap_object(area
, pages
, off
, class->size
);
1486 put_cpu_var(zs_map_area
);
1488 migrate_read_unlock(zspage
);
1491 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1493 static unsigned long obj_malloc(struct size_class
*class,
1494 struct zspage
*zspage
, unsigned long handle
)
1496 int i
, nr_page
, offset
;
1498 struct link_free
*link
;
1500 struct page
*m_page
;
1501 unsigned long m_offset
;
1504 handle
|= OBJ_ALLOCATED_TAG
;
1505 obj
= get_freeobj(zspage
);
1507 offset
= obj
* class->size
;
1508 nr_page
= offset
>> PAGE_SHIFT
;
1509 m_offset
= offset
& ~PAGE_MASK
;
1510 m_page
= get_first_page(zspage
);
1512 for (i
= 0; i
< nr_page
; i
++)
1513 m_page
= get_next_page(m_page
);
1515 vaddr
= kmap_atomic(m_page
);
1516 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1517 set_freeobj(zspage
, link
->next
>> OBJ_ALLOCATED_TAG
);
1518 if (likely(!PageHugeObject(m_page
)))
1519 /* record handle in the header of allocated chunk */
1520 link
->handle
= handle
;
1522 /* record handle to page->index */
1523 zspage
->first_page
->index
= handle
;
1525 kunmap_atomic(vaddr
);
1526 mod_zspage_inuse(zspage
, 1);
1527 zs_stat_inc(class, OBJ_USED
, 1);
1529 obj
= location_to_obj(m_page
, obj
);
1536 * zs_malloc - Allocate block of given size from pool.
1537 * @pool: pool to allocate from
1538 * @size: size of block to allocate
1540 * On success, handle to the allocated object is returned,
1542 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1544 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
, gfp_t gfp
)
1546 unsigned long handle
, obj
;
1547 struct size_class
*class;
1548 enum fullness_group newfg
;
1549 struct zspage
*zspage
;
1551 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1554 handle
= cache_alloc_handle(pool
, gfp
);
1558 /* extra space in chunk to keep the handle */
1559 size
+= ZS_HANDLE_SIZE
;
1560 class = pool
->size_class
[get_size_class_index(size
)];
1562 spin_lock(&class->lock
);
1563 zspage
= find_get_zspage(class);
1564 if (likely(zspage
)) {
1565 obj
= obj_malloc(class, zspage
, handle
);
1566 /* Now move the zspage to another fullness group, if required */
1567 fix_fullness_group(class, zspage
);
1568 record_obj(handle
, obj
);
1569 spin_unlock(&class->lock
);
1574 spin_unlock(&class->lock
);
1576 zspage
= alloc_zspage(pool
, class, gfp
);
1578 cache_free_handle(pool
, handle
);
1582 spin_lock(&class->lock
);
1583 obj
= obj_malloc(class, zspage
, handle
);
1584 newfg
= get_fullness_group(class, zspage
);
1585 insert_zspage(class, zspage
, newfg
);
1586 set_zspage_mapping(zspage
, class->index
, newfg
);
1587 record_obj(handle
, obj
);
1588 atomic_long_add(class->pages_per_zspage
,
1589 &pool
->pages_allocated
);
1590 zs_stat_inc(class, OBJ_ALLOCATED
, get_maxobj_per_zspage(
1591 class->size
, class->pages_per_zspage
));
1593 /* We completely set up zspage so mark them as movable */
1594 SetZsPageMovable(pool
, zspage
);
1595 spin_unlock(&class->lock
);
1599 EXPORT_SYMBOL_GPL(zs_malloc
);
1601 static void obj_free(struct size_class
*class, unsigned long obj
)
1603 struct link_free
*link
;
1604 struct zspage
*zspage
;
1605 struct page
*f_page
;
1606 unsigned long f_offset
;
1607 unsigned int f_objidx
;
1610 obj
&= ~OBJ_ALLOCATED_TAG
;
1611 obj_to_location(obj
, &f_page
, &f_objidx
);
1612 f_offset
= (class->size
* f_objidx
) & ~PAGE_MASK
;
1613 zspage
= get_zspage(f_page
);
1615 vaddr
= kmap_atomic(f_page
);
1617 /* Insert this object in containing zspage's freelist */
1618 link
= (struct link_free
*)(vaddr
+ f_offset
);
1619 link
->next
= get_freeobj(zspage
) << OBJ_ALLOCATED_TAG
;
1620 kunmap_atomic(vaddr
);
1621 set_freeobj(zspage
, f_objidx
);
1622 mod_zspage_inuse(zspage
, -1);
1623 zs_stat_dec(class, OBJ_USED
, 1);
1626 void zs_free(struct zs_pool
*pool
, unsigned long handle
)
1628 struct zspage
*zspage
;
1629 struct page
*f_page
;
1631 unsigned int f_objidx
;
1633 struct size_class
*class;
1634 enum fullness_group fullness
;
1637 if (unlikely(!handle
))
1641 obj
= handle_to_obj(handle
);
1642 obj_to_location(obj
, &f_page
, &f_objidx
);
1643 zspage
= get_zspage(f_page
);
1645 migrate_read_lock(zspage
);
1647 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1648 class = pool
->size_class
[class_idx
];
1650 spin_lock(&class->lock
);
1651 obj_free(class, obj
);
1652 fullness
= fix_fullness_group(class, zspage
);
1653 if (fullness
!= ZS_EMPTY
) {
1654 migrate_read_unlock(zspage
);
1658 isolated
= is_zspage_isolated(zspage
);
1659 migrate_read_unlock(zspage
);
1660 /* If zspage is isolated, zs_page_putback will free the zspage */
1661 if (likely(!isolated
))
1662 free_zspage(pool
, class, zspage
);
1665 spin_unlock(&class->lock
);
1667 cache_free_handle(pool
, handle
);
1669 EXPORT_SYMBOL_GPL(zs_free
);
1671 static void zs_object_copy(struct size_class
*class, unsigned long dst
,
1674 struct page
*s_page
, *d_page
;
1675 unsigned int s_objidx
, d_objidx
;
1676 unsigned long s_off
, d_off
;
1677 void *s_addr
, *d_addr
;
1678 int s_size
, d_size
, size
;
1681 s_size
= d_size
= class->size
;
1683 obj_to_location(src
, &s_page
, &s_objidx
);
1684 obj_to_location(dst
, &d_page
, &d_objidx
);
1686 s_off
= (class->size
* s_objidx
) & ~PAGE_MASK
;
1687 d_off
= (class->size
* d_objidx
) & ~PAGE_MASK
;
1689 if (s_off
+ class->size
> PAGE_SIZE
)
1690 s_size
= PAGE_SIZE
- s_off
;
1692 if (d_off
+ class->size
> PAGE_SIZE
)
1693 d_size
= PAGE_SIZE
- d_off
;
1695 s_addr
= kmap_atomic(s_page
);
1696 d_addr
= kmap_atomic(d_page
);
1699 size
= min(s_size
, d_size
);
1700 memcpy(d_addr
+ d_off
, s_addr
+ s_off
, size
);
1703 if (written
== class->size
)
1711 if (s_off
>= PAGE_SIZE
) {
1712 kunmap_atomic(d_addr
);
1713 kunmap_atomic(s_addr
);
1714 s_page
= get_next_page(s_page
);
1715 s_addr
= kmap_atomic(s_page
);
1716 d_addr
= kmap_atomic(d_page
);
1717 s_size
= class->size
- written
;
1721 if (d_off
>= PAGE_SIZE
) {
1722 kunmap_atomic(d_addr
);
1723 d_page
= get_next_page(d_page
);
1724 d_addr
= kmap_atomic(d_page
);
1725 d_size
= class->size
- written
;
1730 kunmap_atomic(d_addr
);
1731 kunmap_atomic(s_addr
);
1735 * Find alloced object in zspage from index object and
1738 static unsigned long find_alloced_obj(struct size_class
*class,
1739 struct page
*page
, int index
)
1743 unsigned long handle
= 0;
1744 void *addr
= kmap_atomic(page
);
1746 offset
= get_first_obj_offset(page
);
1747 offset
+= class->size
* index
;
1749 while (offset
< PAGE_SIZE
) {
1750 head
= obj_to_head(page
, addr
+ offset
);
1751 if (head
& OBJ_ALLOCATED_TAG
) {
1752 handle
= head
& ~OBJ_ALLOCATED_TAG
;
1753 if (trypin_tag(handle
))
1758 offset
+= class->size
;
1762 kunmap_atomic(addr
);
1766 struct zs_compact_control
{
1767 /* Source spage for migration which could be a subpage of zspage */
1768 struct page
*s_page
;
1769 /* Destination page for migration which should be a first page
1771 struct page
*d_page
;
1772 /* Starting object index within @s_page which used for live object
1773 * in the subpage. */
1777 static int migrate_zspage(struct zs_pool
*pool
, struct size_class
*class,
1778 struct zs_compact_control
*cc
)
1780 unsigned long used_obj
, free_obj
;
1781 unsigned long handle
;
1782 struct page
*s_page
= cc
->s_page
;
1783 struct page
*d_page
= cc
->d_page
;
1784 unsigned long index
= cc
->index
;
1788 handle
= find_alloced_obj(class, s_page
, index
);
1790 s_page
= get_next_page(s_page
);
1797 /* Stop if there is no more space */
1798 if (zspage_full(class, get_zspage(d_page
))) {
1804 used_obj
= handle_to_obj(handle
);
1805 free_obj
= obj_malloc(class, get_zspage(d_page
), handle
);
1806 zs_object_copy(class, free_obj
, used_obj
);
1809 * record_obj updates handle's value to free_obj and it will
1810 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1811 * breaks synchronization using pin_tag(e,g, zs_free) so
1812 * let's keep the lock bit.
1814 free_obj
|= BIT(HANDLE_PIN_BIT
);
1815 record_obj(handle
, free_obj
);
1817 obj_free(class, used_obj
);
1820 /* Remember last position in this iteration */
1821 cc
->s_page
= s_page
;
1827 static struct zspage
*isolate_zspage(struct size_class
*class, bool source
)
1830 struct zspage
*zspage
;
1831 enum fullness_group fg
[2] = {ZS_ALMOST_EMPTY
, ZS_ALMOST_FULL
};
1834 fg
[0] = ZS_ALMOST_FULL
;
1835 fg
[1] = ZS_ALMOST_EMPTY
;
1838 for (i
= 0; i
< 2; i
++) {
1839 zspage
= list_first_entry_or_null(&class->fullness_list
[fg
[i
]],
1840 struct zspage
, list
);
1842 VM_BUG_ON(is_zspage_isolated(zspage
));
1843 remove_zspage(class, zspage
, fg
[i
]);
1852 * putback_zspage - add @zspage into right class's fullness list
1853 * @class: destination class
1854 * @zspage: target page
1856 * Return @zspage's fullness_group
1858 static enum fullness_group
putback_zspage(struct size_class
*class,
1859 struct zspage
*zspage
)
1861 enum fullness_group fullness
;
1863 VM_BUG_ON(is_zspage_isolated(zspage
));
1865 fullness
= get_fullness_group(class, zspage
);
1866 insert_zspage(class, zspage
, fullness
);
1867 set_zspage_mapping(zspage
, class->index
, fullness
);
1872 #ifdef CONFIG_COMPACTION
1873 static struct dentry
*zs_mount(struct file_system_type
*fs_type
,
1874 int flags
, const char *dev_name
, void *data
)
1876 static const struct dentry_operations ops
= {
1877 .d_dname
= simple_dname
,
1880 return mount_pseudo(fs_type
, "zsmalloc:", NULL
, &ops
, ZSMALLOC_MAGIC
);
1883 static struct file_system_type zsmalloc_fs
= {
1886 .kill_sb
= kill_anon_super
,
1889 static int zsmalloc_mount(void)
1893 zsmalloc_mnt
= kern_mount(&zsmalloc_fs
);
1894 if (IS_ERR(zsmalloc_mnt
))
1895 ret
= PTR_ERR(zsmalloc_mnt
);
1900 static void zsmalloc_unmount(void)
1902 kern_unmount(zsmalloc_mnt
);
1905 static void migrate_lock_init(struct zspage
*zspage
)
1907 rwlock_init(&zspage
->lock
);
1910 static void migrate_read_lock(struct zspage
*zspage
)
1912 read_lock(&zspage
->lock
);
1915 static void migrate_read_unlock(struct zspage
*zspage
)
1917 read_unlock(&zspage
->lock
);
1920 static void migrate_write_lock(struct zspage
*zspage
)
1922 write_lock(&zspage
->lock
);
1925 static void migrate_write_unlock(struct zspage
*zspage
)
1927 write_unlock(&zspage
->lock
);
1930 /* Number of isolated subpage for *page migration* in this zspage */
1931 static void inc_zspage_isolation(struct zspage
*zspage
)
1936 static void dec_zspage_isolation(struct zspage
*zspage
)
1941 static void replace_sub_page(struct size_class
*class, struct zspage
*zspage
,
1942 struct page
*newpage
, struct page
*oldpage
)
1945 struct page
*pages
[ZS_MAX_PAGES_PER_ZSPAGE
] = {NULL
, };
1948 page
= get_first_page(zspage
);
1950 if (page
== oldpage
)
1951 pages
[idx
] = newpage
;
1955 } while ((page
= get_next_page(page
)) != NULL
);
1957 create_page_chain(class, zspage
, pages
);
1958 set_first_obj_offset(newpage
, get_first_obj_offset(oldpage
));
1959 if (unlikely(PageHugeObject(oldpage
)))
1960 newpage
->index
= oldpage
->index
;
1961 __SetPageMovable(newpage
, page_mapping(oldpage
));
1964 bool zs_page_isolate(struct page
*page
, isolate_mode_t mode
)
1966 struct zs_pool
*pool
;
1967 struct size_class
*class;
1969 enum fullness_group fullness
;
1970 struct zspage
*zspage
;
1971 struct address_space
*mapping
;
1974 * Page is locked so zspage couldn't be destroyed. For detail, look at
1975 * lock_zspage in free_zspage.
1977 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1978 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1980 zspage
= get_zspage(page
);
1983 * Without class lock, fullness could be stale while class_idx is okay
1984 * because class_idx is constant unless page is freed so we should get
1985 * fullness again under class lock.
1987 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
1988 mapping
= page_mapping(page
);
1989 pool
= mapping
->private_data
;
1990 class = pool
->size_class
[class_idx
];
1992 spin_lock(&class->lock
);
1993 if (get_zspage_inuse(zspage
) == 0) {
1994 spin_unlock(&class->lock
);
1998 /* zspage is isolated for object migration */
1999 if (list_empty(&zspage
->list
) && !is_zspage_isolated(zspage
)) {
2000 spin_unlock(&class->lock
);
2005 * If this is first time isolation for the zspage, isolate zspage from
2006 * size_class to prevent further object allocation from the zspage.
2008 if (!list_empty(&zspage
->list
) && !is_zspage_isolated(zspage
)) {
2009 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2010 remove_zspage(class, zspage
, fullness
);
2013 inc_zspage_isolation(zspage
);
2014 spin_unlock(&class->lock
);
2019 int zs_page_migrate(struct address_space
*mapping
, struct page
*newpage
,
2020 struct page
*page
, enum migrate_mode mode
)
2022 struct zs_pool
*pool
;
2023 struct size_class
*class;
2025 enum fullness_group fullness
;
2026 struct zspage
*zspage
;
2028 void *s_addr
, *d_addr
, *addr
;
2030 unsigned long handle
, head
;
2031 unsigned long old_obj
, new_obj
;
2032 unsigned int obj_idx
;
2035 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
2036 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
2038 zspage
= get_zspage(page
);
2040 /* Concurrent compactor cannot migrate any subpage in zspage */
2041 migrate_write_lock(zspage
);
2042 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2043 pool
= mapping
->private_data
;
2044 class = pool
->size_class
[class_idx
];
2045 offset
= get_first_obj_offset(page
);
2047 spin_lock(&class->lock
);
2048 if (!get_zspage_inuse(zspage
)) {
2054 s_addr
= kmap_atomic(page
);
2055 while (pos
< PAGE_SIZE
) {
2056 head
= obj_to_head(page
, s_addr
+ pos
);
2057 if (head
& OBJ_ALLOCATED_TAG
) {
2058 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2059 if (!trypin_tag(handle
))
2066 * Here, any user cannot access all objects in the zspage so let's move.
2068 d_addr
= kmap_atomic(newpage
);
2069 memcpy(d_addr
, s_addr
, PAGE_SIZE
);
2070 kunmap_atomic(d_addr
);
2072 for (addr
= s_addr
+ offset
; addr
< s_addr
+ pos
;
2073 addr
+= class->size
) {
2074 head
= obj_to_head(page
, addr
);
2075 if (head
& OBJ_ALLOCATED_TAG
) {
2076 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2077 if (!testpin_tag(handle
))
2080 old_obj
= handle_to_obj(handle
);
2081 obj_to_location(old_obj
, &dummy
, &obj_idx
);
2082 new_obj
= (unsigned long)location_to_obj(newpage
,
2084 new_obj
|= BIT(HANDLE_PIN_BIT
);
2085 record_obj(handle
, new_obj
);
2089 replace_sub_page(class, zspage
, newpage
, page
);
2092 dec_zspage_isolation(zspage
);
2095 * Page migration is done so let's putback isolated zspage to
2096 * the list if @page is final isolated subpage in the zspage.
2098 if (!is_zspage_isolated(zspage
))
2099 putback_zspage(class, zspage
);
2107 for (addr
= s_addr
+ offset
; addr
< s_addr
+ pos
;
2108 addr
+= class->size
) {
2109 head
= obj_to_head(page
, addr
);
2110 if (head
& OBJ_ALLOCATED_TAG
) {
2111 handle
= head
& ~OBJ_ALLOCATED_TAG
;
2112 if (!testpin_tag(handle
))
2117 kunmap_atomic(s_addr
);
2119 spin_unlock(&class->lock
);
2120 migrate_write_unlock(zspage
);
2125 void zs_page_putback(struct page
*page
)
2127 struct zs_pool
*pool
;
2128 struct size_class
*class;
2130 enum fullness_group fg
;
2131 struct address_space
*mapping
;
2132 struct zspage
*zspage
;
2134 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
2135 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
2137 zspage
= get_zspage(page
);
2138 get_zspage_mapping(zspage
, &class_idx
, &fg
);
2139 mapping
= page_mapping(page
);
2140 pool
= mapping
->private_data
;
2141 class = pool
->size_class
[class_idx
];
2143 spin_lock(&class->lock
);
2144 dec_zspage_isolation(zspage
);
2145 if (!is_zspage_isolated(zspage
)) {
2146 fg
= putback_zspage(class, zspage
);
2148 * Due to page_lock, we cannot free zspage immediately
2152 schedule_work(&pool
->free_work
);
2154 spin_unlock(&class->lock
);
2157 const struct address_space_operations zsmalloc_aops
= {
2158 .isolate_page
= zs_page_isolate
,
2159 .migratepage
= zs_page_migrate
,
2160 .putback_page
= zs_page_putback
,
2163 static int zs_register_migration(struct zs_pool
*pool
)
2165 pool
->inode
= alloc_anon_inode(zsmalloc_mnt
->mnt_sb
);
2166 if (IS_ERR(pool
->inode
)) {
2171 pool
->inode
->i_mapping
->private_data
= pool
;
2172 pool
->inode
->i_mapping
->a_ops
= &zsmalloc_aops
;
2176 static void zs_unregister_migration(struct zs_pool
*pool
)
2178 flush_work(&pool
->free_work
);
2184 * Caller should hold page_lock of all pages in the zspage
2185 * In here, we cannot use zspage meta data.
2187 static void async_free_zspage(struct work_struct
*work
)
2190 struct size_class
*class;
2191 unsigned int class_idx
;
2192 enum fullness_group fullness
;
2193 struct zspage
*zspage
, *tmp
;
2194 LIST_HEAD(free_pages
);
2195 struct zs_pool
*pool
= container_of(work
, struct zs_pool
,
2198 for (i
= 0; i
< zs_size_classes
; i
++) {
2199 class = pool
->size_class
[i
];
2200 if (class->index
!= i
)
2203 spin_lock(&class->lock
);
2204 list_splice_init(&class->fullness_list
[ZS_EMPTY
], &free_pages
);
2205 spin_unlock(&class->lock
);
2209 list_for_each_entry_safe(zspage
, tmp
, &free_pages
, list
) {
2210 list_del(&zspage
->list
);
2211 lock_zspage(zspage
);
2213 get_zspage_mapping(zspage
, &class_idx
, &fullness
);
2214 VM_BUG_ON(fullness
!= ZS_EMPTY
);
2215 class = pool
->size_class
[class_idx
];
2216 spin_lock(&class->lock
);
2217 __free_zspage(pool
, pool
->size_class
[class_idx
], zspage
);
2218 spin_unlock(&class->lock
);
2222 static void kick_deferred_free(struct zs_pool
*pool
)
2224 schedule_work(&pool
->free_work
);
2227 static void init_deferred_free(struct zs_pool
*pool
)
2229 INIT_WORK(&pool
->free_work
, async_free_zspage
);
2232 static void SetZsPageMovable(struct zs_pool
*pool
, struct zspage
*zspage
)
2234 struct page
*page
= get_first_page(zspage
);
2237 WARN_ON(!trylock_page(page
));
2238 __SetPageMovable(page
, pool
->inode
->i_mapping
);
2240 } while ((page
= get_next_page(page
)) != NULL
);
2246 * Based on the number of unused allocated objects calculate
2247 * and return the number of pages that we can free.
2249 static unsigned long zs_can_compact(struct size_class
*class)
2251 unsigned long obj_wasted
;
2252 unsigned long obj_allocated
= zs_stat_get(class, OBJ_ALLOCATED
);
2253 unsigned long obj_used
= zs_stat_get(class, OBJ_USED
);
2255 if (obj_allocated
<= obj_used
)
2258 obj_wasted
= obj_allocated
- obj_used
;
2259 obj_wasted
/= get_maxobj_per_zspage(class->size
,
2260 class->pages_per_zspage
);
2262 return obj_wasted
* class->pages_per_zspage
;
2265 static void __zs_compact(struct zs_pool
*pool
, struct size_class
*class)
2267 struct zs_compact_control cc
;
2268 struct zspage
*src_zspage
;
2269 struct zspage
*dst_zspage
= NULL
;
2271 spin_lock(&class->lock
);
2272 while ((src_zspage
= isolate_zspage(class, true))) {
2274 if (!zs_can_compact(class))
2278 cc
.s_page
= get_first_page(src_zspage
);
2280 while ((dst_zspage
= isolate_zspage(class, false))) {
2281 cc
.d_page
= get_first_page(dst_zspage
);
2283 * If there is no more space in dst_page, resched
2284 * and see if anyone had allocated another zspage.
2286 if (!migrate_zspage(pool
, class, &cc
))
2289 putback_zspage(class, dst_zspage
);
2292 /* Stop if we couldn't find slot */
2293 if (dst_zspage
== NULL
)
2296 putback_zspage(class, dst_zspage
);
2297 if (putback_zspage(class, src_zspage
) == ZS_EMPTY
) {
2298 free_zspage(pool
, class, src_zspage
);
2299 pool
->stats
.pages_compacted
+= class->pages_per_zspage
;
2301 spin_unlock(&class->lock
);
2303 spin_lock(&class->lock
);
2307 putback_zspage(class, src_zspage
);
2309 spin_unlock(&class->lock
);
2312 unsigned long zs_compact(struct zs_pool
*pool
)
2315 struct size_class
*class;
2317 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
2318 class = pool
->size_class
[i
];
2321 if (class->index
!= i
)
2323 __zs_compact(pool
, class);
2326 return pool
->stats
.pages_compacted
;
2328 EXPORT_SYMBOL_GPL(zs_compact
);
2330 void zs_pool_stats(struct zs_pool
*pool
, struct zs_pool_stats
*stats
)
2332 memcpy(stats
, &pool
->stats
, sizeof(struct zs_pool_stats
));
2334 EXPORT_SYMBOL_GPL(zs_pool_stats
);
2336 static unsigned long zs_shrinker_scan(struct shrinker
*shrinker
,
2337 struct shrink_control
*sc
)
2339 unsigned long pages_freed
;
2340 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2343 pages_freed
= pool
->stats
.pages_compacted
;
2345 * Compact classes and calculate compaction delta.
2346 * Can run concurrently with a manually triggered
2347 * (by user) compaction.
2349 pages_freed
= zs_compact(pool
) - pages_freed
;
2351 return pages_freed
? pages_freed
: SHRINK_STOP
;
2354 static unsigned long zs_shrinker_count(struct shrinker
*shrinker
,
2355 struct shrink_control
*sc
)
2358 struct size_class
*class;
2359 unsigned long pages_to_free
= 0;
2360 struct zs_pool
*pool
= container_of(shrinker
, struct zs_pool
,
2363 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
2364 class = pool
->size_class
[i
];
2367 if (class->index
!= i
)
2370 pages_to_free
+= zs_can_compact(class);
2373 return pages_to_free
;
2376 static void zs_unregister_shrinker(struct zs_pool
*pool
)
2378 if (pool
->shrinker_enabled
) {
2379 unregister_shrinker(&pool
->shrinker
);
2380 pool
->shrinker_enabled
= false;
2384 static int zs_register_shrinker(struct zs_pool
*pool
)
2386 pool
->shrinker
.scan_objects
= zs_shrinker_scan
;
2387 pool
->shrinker
.count_objects
= zs_shrinker_count
;
2388 pool
->shrinker
.batch
= 0;
2389 pool
->shrinker
.seeks
= DEFAULT_SEEKS
;
2391 return register_shrinker(&pool
->shrinker
);
2395 * zs_create_pool - Creates an allocation pool to work from.
2396 * @flags: allocation flags used to allocate pool metadata
2398 * This function must be called before anything when using
2399 * the zsmalloc allocator.
2401 * On success, a pointer to the newly created pool is returned,
2404 struct zs_pool
*zs_create_pool(const char *name
)
2407 struct zs_pool
*pool
;
2408 struct size_class
*prev_class
= NULL
;
2410 pool
= kzalloc(sizeof(*pool
), GFP_KERNEL
);
2414 init_deferred_free(pool
);
2415 pool
->size_class
= kcalloc(zs_size_classes
, sizeof(struct size_class
*),
2417 if (!pool
->size_class
) {
2422 pool
->name
= kstrdup(name
, GFP_KERNEL
);
2426 if (create_cache(pool
))
2430 * Iterate reversly, because, size of size_class that we want to use
2431 * for merging should be larger or equal to current size.
2433 for (i
= zs_size_classes
- 1; i
>= 0; i
--) {
2435 int pages_per_zspage
;
2436 struct size_class
*class;
2439 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
2440 if (size
> ZS_MAX_ALLOC_SIZE
)
2441 size
= ZS_MAX_ALLOC_SIZE
;
2442 pages_per_zspage
= get_pages_per_zspage(size
);
2445 * size_class is used for normal zsmalloc operation such
2446 * as alloc/free for that size. Although it is natural that we
2447 * have one size_class for each size, there is a chance that we
2448 * can get more memory utilization if we use one size_class for
2449 * many different sizes whose size_class have same
2450 * characteristics. So, we makes size_class point to
2451 * previous size_class if possible.
2454 if (can_merge(prev_class
, size
, pages_per_zspage
)) {
2455 pool
->size_class
[i
] = prev_class
;
2460 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
2466 class->pages_per_zspage
= pages_per_zspage
;
2467 class->objs_per_zspage
= class->pages_per_zspage
*
2468 PAGE_SIZE
/ class->size
;
2469 spin_lock_init(&class->lock
);
2470 pool
->size_class
[i
] = class;
2471 for (fullness
= ZS_EMPTY
; fullness
< NR_ZS_FULLNESS
;
2473 INIT_LIST_HEAD(&class->fullness_list
[fullness
]);
2478 /* debug only, don't abort if it fails */
2479 zs_pool_stat_create(pool
, name
);
2481 if (zs_register_migration(pool
))
2485 * Not critical, we still can use the pool
2486 * and user can trigger compaction manually.
2488 if (zs_register_shrinker(pool
) == 0)
2489 pool
->shrinker_enabled
= true;
2493 zs_destroy_pool(pool
);
2496 EXPORT_SYMBOL_GPL(zs_create_pool
);
2498 void zs_destroy_pool(struct zs_pool
*pool
)
2502 zs_unregister_shrinker(pool
);
2503 zs_unregister_migration(pool
);
2504 zs_pool_stat_destroy(pool
);
2506 for (i
= 0; i
< zs_size_classes
; i
++) {
2508 struct size_class
*class = pool
->size_class
[i
];
2513 if (class->index
!= i
)
2516 for (fg
= ZS_EMPTY
; fg
< NR_ZS_FULLNESS
; fg
++) {
2517 if (!list_empty(&class->fullness_list
[fg
])) {
2518 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2525 destroy_cache(pool
);
2526 kfree(pool
->size_class
);
2530 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
2532 static int __init
zs_init(void)
2536 ret
= zsmalloc_mount();
2540 ret
= zs_register_cpu_notifier();
2545 init_zs_size_classes();
2548 zpool_register_driver(&zs_zpool_driver
);
2556 zs_unregister_cpu_notifier();
2562 static void __exit
zs_exit(void)
2565 zpool_unregister_driver(&zs_zpool_driver
);
2568 zs_unregister_cpu_notifier();
2573 module_init(zs_init
);
2574 module_exit(zs_exit
);
2576 MODULE_LICENSE("Dual BSD/GPL");
2577 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");