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 * This allocator is designed for use with zram. Thus, the allocator is
16 * supposed to work well under low memory conditions. In particular, it
17 * never attempts higher order page allocation which is very likely to
18 * fail under memory pressure. On the other hand, if we just use single
19 * (0-order) pages, it would suffer from very high fragmentation --
20 * any object of size PAGE_SIZE/2 or larger would occupy an entire page.
21 * This was one of the major issues with its predecessor (xvmalloc).
23 * To overcome these issues, zsmalloc allocates a bunch of 0-order pages
24 * and links them together using various 'struct page' fields. These linked
25 * pages act as a single higher-order page i.e. an object can span 0-order
26 * page boundaries. The code refers to these linked pages as a single entity
29 * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE
30 * since this satisfies the requirements of all its current users (in the
31 * worst case, page is incompressible and is thus stored "as-is" i.e. in
32 * uncompressed form). For allocation requests larger than this size, failure
33 * is returned (see zs_malloc).
35 * Additionally, zs_malloc() does not return a dereferenceable pointer.
36 * Instead, it returns an opaque handle (unsigned long) which encodes actual
37 * location of the allocated object. The reason for this indirection is that
38 * zsmalloc does not keep zspages permanently mapped since that would cause
39 * issues on 32-bit systems where the VA region for kernel space mappings
40 * is very small. So, before using the allocating memory, the object has to
41 * be mapped using zs_map_object() to get a usable pointer and subsequently
42 * unmapped using zs_unmap_object().
44 * Following is how we use various fields and flags of underlying
45 * struct page(s) to form a zspage.
47 * Usage of struct page fields:
48 * page->first_page: points to the first component (0-order) page
49 * page->index (union with page->freelist): offset of the first object
50 * starting in this page. For the first page, this is
51 * always 0, so we use this field (aka freelist) to point
52 * to the first free object in zspage.
53 * page->lru: links together all component pages (except the first page)
56 * For _first_ page only:
58 * page->private (union with page->first_page): refers to the
59 * component page after the first page
60 * page->freelist: points to the first free object in zspage.
61 * Free objects are linked together using in-place
63 * page->objects: maximum number of objects we can store in this
64 * zspage (class->zspage_order * PAGE_SIZE / class->size)
65 * page->lru: links together first pages of various zspages.
66 * Basically forming list of zspages in a fullness group.
67 * page->mapping: class index and fullness group of the zspage
69 * Usage of struct page flags:
70 * PG_private: identifies the first component page
71 * PG_private2: identifies the last component page
75 #ifdef CONFIG_ZSMALLOC_DEBUG
79 #include <linux/module.h>
80 #include <linux/kernel.h>
81 #include <linux/bitops.h>
82 #include <linux/errno.h>
83 #include <linux/highmem.h>
84 #include <linux/string.h>
85 #include <linux/slab.h>
86 #include <asm/tlbflush.h>
87 #include <asm/pgtable.h>
88 #include <linux/cpumask.h>
89 #include <linux/cpu.h>
90 #include <linux/vmalloc.h>
91 #include <linux/hardirq.h>
92 #include <linux/spinlock.h>
93 #include <linux/types.h>
94 #include <linux/zsmalloc.h>
95 #include <linux/zpool.h>
98 * This must be power of 2 and greater than of equal to sizeof(link_free).
99 * These two conditions ensure that any 'struct link_free' itself doesn't
100 * span more than 1 page which avoids complex case of mapping 2 pages simply
101 * to restore link_free pointer values.
106 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
107 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
109 #define ZS_MAX_ZSPAGE_ORDER 2
110 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
113 * Object location (<PFN>, <obj_idx>) is encoded as
114 * as single (unsigned long) handle value.
116 * Note that object index <obj_idx> is relative to system
117 * page <PFN> it is stored in, so for each sub-page belonging
118 * to a zspage, obj_idx starts with 0.
120 * This is made more complicated by various memory models and PAE.
123 #ifndef MAX_PHYSMEM_BITS
124 #ifdef CONFIG_HIGHMEM64G
125 #define MAX_PHYSMEM_BITS 36
126 #else /* !CONFIG_HIGHMEM64G */
128 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
131 #define MAX_PHYSMEM_BITS BITS_PER_LONG
134 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
135 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
136 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
138 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
139 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
140 #define ZS_MIN_ALLOC_SIZE \
141 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
142 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
145 * On systems with 4K page size, this gives 255 size classes! There is a
147 * - Large number of size classes is potentially wasteful as free page are
148 * spread across these classes
149 * - Small number of size classes causes large internal fragmentation
150 * - Probably its better to use specific size classes (empirically
151 * determined). NOTE: all those class sizes must be set as multiple of
152 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
154 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
157 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
158 #define ZS_SIZE_CLASSES ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
159 ZS_SIZE_CLASS_DELTA + 1)
162 * We do not maintain any list for completely empty or full pages
164 enum fullness_group
{
167 _ZS_NR_FULLNESS_GROUPS
,
174 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
176 * n = number of allocated objects
177 * N = total number of objects zspage can store
178 * f = 1/fullness_threshold_frac
180 * Similarly, we assign zspage to:
181 * ZS_ALMOST_FULL when n > N / f
182 * ZS_EMPTY when n == 0
183 * ZS_FULL when n == N
185 * (see: fix_fullness_group())
187 static const int fullness_threshold_frac
= 4;
191 * Size of objects stored in this class. Must be multiple
197 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
198 int pages_per_zspage
;
205 struct page
*fullness_list
[_ZS_NR_FULLNESS_GROUPS
];
209 * Placed within free objects to form a singly linked list.
210 * For every zspage, first_page->freelist gives head of this list.
212 * This must be power of 2 and less than or equal to ZS_ALIGN
215 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
220 struct size_class size_class
[ZS_SIZE_CLASSES
];
222 gfp_t flags
; /* allocation flags used when growing pool */
226 * A zspage's class index and fullness group
227 * are encoded in its (first)page->mapping
229 #define CLASS_IDX_BITS 28
230 #define FULLNESS_BITS 4
231 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
232 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
234 struct mapping_area
{
235 #ifdef CONFIG_PGTABLE_MAPPING
236 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
238 char *vm_buf
; /* copy buffer for objects that span pages */
240 char *vm_addr
; /* address of kmap_atomic()'ed pages */
241 enum zs_mapmode vm_mm
; /* mapping mode */
248 static void *zs_zpool_create(gfp_t gfp
, struct zpool_ops
*zpool_ops
)
250 return zs_create_pool(gfp
);
253 static void zs_zpool_destroy(void *pool
)
255 zs_destroy_pool(pool
);
258 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
259 unsigned long *handle
)
261 *handle
= zs_malloc(pool
, size
);
262 return *handle
? 0 : -1;
264 static void zs_zpool_free(void *pool
, unsigned long handle
)
266 zs_free(pool
, handle
);
269 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
270 unsigned int *reclaimed
)
275 static void *zs_zpool_map(void *pool
, unsigned long handle
,
276 enum zpool_mapmode mm
)
278 enum zs_mapmode zs_mm
;
287 case ZPOOL_MM_RW
: /* fallthru */
293 return zs_map_object(pool
, handle
, zs_mm
);
295 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
297 zs_unmap_object(pool
, handle
);
300 static u64
zs_zpool_total_size(void *pool
)
302 return zs_get_total_size_bytes(pool
);
305 static struct zpool_driver zs_zpool_driver
= {
307 .owner
= THIS_MODULE
,
308 .create
= zs_zpool_create
,
309 .destroy
= zs_zpool_destroy
,
310 .malloc
= zs_zpool_malloc
,
311 .free
= zs_zpool_free
,
312 .shrink
= zs_zpool_shrink
,
314 .unmap
= zs_zpool_unmap
,
315 .total_size
= zs_zpool_total_size
,
318 #endif /* CONFIG_ZPOOL */
320 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
321 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
323 static int is_first_page(struct page
*page
)
325 return PagePrivate(page
);
328 static int is_last_page(struct page
*page
)
330 return PagePrivate2(page
);
333 static void get_zspage_mapping(struct page
*page
, unsigned int *class_idx
,
334 enum fullness_group
*fullness
)
337 BUG_ON(!is_first_page(page
));
339 m
= (unsigned long)page
->mapping
;
340 *fullness
= m
& FULLNESS_MASK
;
341 *class_idx
= (m
>> FULLNESS_BITS
) & CLASS_IDX_MASK
;
344 static void set_zspage_mapping(struct page
*page
, unsigned int class_idx
,
345 enum fullness_group fullness
)
348 BUG_ON(!is_first_page(page
));
350 m
= ((class_idx
& CLASS_IDX_MASK
) << FULLNESS_BITS
) |
351 (fullness
& FULLNESS_MASK
);
352 page
->mapping
= (struct address_space
*)m
;
356 * zsmalloc divides the pool into various size classes where each
357 * class maintains a list of zspages where each zspage is divided
358 * into equal sized chunks. Each allocation falls into one of these
359 * classes depending on its size. This function returns index of the
360 * size class which has chunk size big enough to hold the give size.
362 static int get_size_class_index(int size
)
366 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
367 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
368 ZS_SIZE_CLASS_DELTA
);
374 * For each size class, zspages are divided into different groups
375 * depending on how "full" they are. This was done so that we could
376 * easily find empty or nearly empty zspages when we try to shrink
377 * the pool (not yet implemented). This function returns fullness
378 * status of the given page.
380 static enum fullness_group
get_fullness_group(struct page
*page
)
382 int inuse
, max_objects
;
383 enum fullness_group fg
;
384 BUG_ON(!is_first_page(page
));
387 max_objects
= page
->objects
;
391 else if (inuse
== max_objects
)
393 else if (inuse
<= max_objects
/ fullness_threshold_frac
)
394 fg
= ZS_ALMOST_EMPTY
;
402 * Each size class maintains various freelists and zspages are assigned
403 * to one of these freelists based on the number of live objects they
404 * have. This functions inserts the given zspage into the freelist
405 * identified by <class, fullness_group>.
407 static void insert_zspage(struct page
*page
, struct size_class
*class,
408 enum fullness_group fullness
)
412 BUG_ON(!is_first_page(page
));
414 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
417 head
= &class->fullness_list
[fullness
];
419 list_add_tail(&page
->lru
, &(*head
)->lru
);
425 * This function removes the given zspage from the freelist identified
426 * by <class, fullness_group>.
428 static void remove_zspage(struct page
*page
, struct size_class
*class,
429 enum fullness_group fullness
)
433 BUG_ON(!is_first_page(page
));
435 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
438 head
= &class->fullness_list
[fullness
];
440 if (list_empty(&(*head
)->lru
))
442 else if (*head
== page
)
443 *head
= (struct page
*)list_entry((*head
)->lru
.next
,
446 list_del_init(&page
->lru
);
450 * Each size class maintains zspages in different fullness groups depending
451 * on the number of live objects they contain. When allocating or freeing
452 * objects, the fullness status of the page can change, say, from ALMOST_FULL
453 * to ALMOST_EMPTY when freeing an object. This function checks if such
454 * a status change has occurred for the given page and accordingly moves the
455 * page from the freelist of the old fullness group to that of the new
458 static enum fullness_group
fix_fullness_group(struct zs_pool
*pool
,
462 struct size_class
*class;
463 enum fullness_group currfg
, newfg
;
465 BUG_ON(!is_first_page(page
));
467 get_zspage_mapping(page
, &class_idx
, &currfg
);
468 newfg
= get_fullness_group(page
);
472 class = &pool
->size_class
[class_idx
];
473 remove_zspage(page
, class, currfg
);
474 insert_zspage(page
, class, newfg
);
475 set_zspage_mapping(page
, class_idx
, newfg
);
482 * We have to decide on how many pages to link together
483 * to form a zspage for each size class. This is important
484 * to reduce wastage due to unusable space left at end of
485 * each zspage which is given as:
486 * wastage = Zp - Zp % size_class
487 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
489 * For example, for size class of 3/8 * PAGE_SIZE, we should
490 * link together 3 PAGE_SIZE sized pages to form a zspage
491 * since then we can perfectly fit in 8 such objects.
493 static int get_pages_per_zspage(int class_size
)
495 int i
, max_usedpc
= 0;
496 /* zspage order which gives maximum used size per KB */
497 int max_usedpc_order
= 1;
499 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
503 zspage_size
= i
* PAGE_SIZE
;
504 waste
= zspage_size
% class_size
;
505 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
507 if (usedpc
> max_usedpc
) {
509 max_usedpc_order
= i
;
513 return max_usedpc_order
;
517 * A single 'zspage' is composed of many system pages which are
518 * linked together using fields in struct page. This function finds
519 * the first/head page, given any component page of a zspage.
521 static struct page
*get_first_page(struct page
*page
)
523 if (is_first_page(page
))
526 return page
->first_page
;
529 static struct page
*get_next_page(struct page
*page
)
533 if (is_last_page(page
))
535 else if (is_first_page(page
))
536 next
= (struct page
*)page_private(page
);
538 next
= list_entry(page
->lru
.next
, struct page
, lru
);
544 * Encode <page, obj_idx> as a single handle value.
545 * On hardware platforms with physical memory starting at 0x0 the pfn
546 * could be 0 so we ensure that the handle will never be 0 by adjusting the
547 * encoded obj_idx value before encoding.
549 static void *obj_location_to_handle(struct page
*page
, unsigned long obj_idx
)
551 unsigned long handle
;
558 handle
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
559 handle
|= ((obj_idx
+ 1) & OBJ_INDEX_MASK
);
561 return (void *)handle
;
565 * Decode <page, obj_idx> pair from the given object handle. We adjust the
566 * decoded obj_idx back to its original value since it was adjusted in
567 * obj_location_to_handle().
569 static void obj_handle_to_location(unsigned long handle
, struct page
**page
,
570 unsigned long *obj_idx
)
572 *page
= pfn_to_page(handle
>> OBJ_INDEX_BITS
);
573 *obj_idx
= (handle
& OBJ_INDEX_MASK
) - 1;
576 static unsigned long obj_idx_to_offset(struct page
*page
,
577 unsigned long obj_idx
, int class_size
)
579 unsigned long off
= 0;
581 if (!is_first_page(page
))
584 return off
+ obj_idx
* class_size
;
587 static void reset_page(struct page
*page
)
589 clear_bit(PG_private
, &page
->flags
);
590 clear_bit(PG_private_2
, &page
->flags
);
591 set_page_private(page
, 0);
592 page
->mapping
= NULL
;
593 page
->freelist
= NULL
;
594 page_mapcount_reset(page
);
597 static void free_zspage(struct page
*first_page
)
599 struct page
*nextp
, *tmp
, *head_extra
;
601 BUG_ON(!is_first_page(first_page
));
602 BUG_ON(first_page
->inuse
);
604 head_extra
= (struct page
*)page_private(first_page
);
606 reset_page(first_page
);
607 __free_page(first_page
);
609 /* zspage with only 1 system page */
613 list_for_each_entry_safe(nextp
, tmp
, &head_extra
->lru
, lru
) {
614 list_del(&nextp
->lru
);
618 reset_page(head_extra
);
619 __free_page(head_extra
);
622 /* Initialize a newly allocated zspage */
623 static void init_zspage(struct page
*first_page
, struct size_class
*class)
625 unsigned long off
= 0;
626 struct page
*page
= first_page
;
628 BUG_ON(!is_first_page(first_page
));
630 struct page
*next_page
;
631 struct link_free
*link
;
632 unsigned int i
, objs_on_page
;
635 * page->index stores offset of first object starting
636 * in the page. For the first page, this is always 0,
637 * so we use first_page->index (aka ->freelist) to store
638 * head of corresponding zspage's freelist.
640 if (page
!= first_page
)
643 link
= (struct link_free
*)kmap_atomic(page
) +
645 objs_on_page
= (PAGE_SIZE
- off
) / class->size
;
647 for (i
= 1; i
<= objs_on_page
; i
++) {
649 if (off
< PAGE_SIZE
) {
650 link
->next
= obj_location_to_handle(page
, i
);
651 link
+= class->size
/ sizeof(*link
);
656 * We now come to the last (full or partial) object on this
657 * page, which must point to the first object on the next
660 next_page
= get_next_page(page
);
661 link
->next
= obj_location_to_handle(next_page
, 0);
664 off
= (off
+ class->size
) % PAGE_SIZE
;
669 * Allocate a zspage for the given size class
671 static struct page
*alloc_zspage(struct size_class
*class, gfp_t flags
)
674 struct page
*first_page
= NULL
, *uninitialized_var(prev_page
);
677 * Allocate individual pages and link them together as:
678 * 1. first page->private = first sub-page
679 * 2. all sub-pages are linked together using page->lru
680 * 3. each sub-page is linked to the first page using page->first_page
682 * For each size class, First/Head pages are linked together using
683 * page->lru. Also, we set PG_private to identify the first page
684 * (i.e. no other sub-page has this flag set) and PG_private_2 to
685 * identify the last page.
688 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
691 page
= alloc_page(flags
);
695 INIT_LIST_HEAD(&page
->lru
);
696 if (i
== 0) { /* first page */
697 SetPagePrivate(page
);
698 set_page_private(page
, 0);
700 first_page
->inuse
= 0;
703 set_page_private(first_page
, (unsigned long)page
);
705 page
->first_page
= first_page
;
707 list_add(&page
->lru
, &prev_page
->lru
);
708 if (i
== class->pages_per_zspage
- 1) /* last page */
709 SetPagePrivate2(page
);
713 init_zspage(first_page
, class);
715 first_page
->freelist
= obj_location_to_handle(first_page
, 0);
716 /* Maximum number of objects we can store in this zspage */
717 first_page
->objects
= class->pages_per_zspage
* PAGE_SIZE
/ class->size
;
719 error
= 0; /* Success */
722 if (unlikely(error
) && first_page
) {
723 free_zspage(first_page
);
730 static struct page
*find_get_zspage(struct size_class
*class)
735 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
736 page
= class->fullness_list
[i
];
744 #ifdef CONFIG_PGTABLE_MAPPING
745 static inline int __zs_cpu_up(struct mapping_area
*area
)
748 * Make sure we don't leak memory if a cpu UP notification
749 * and zs_init() race and both call zs_cpu_up() on the same cpu
753 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
759 static inline void __zs_cpu_down(struct mapping_area
*area
)
762 free_vm_area(area
->vm
);
766 static inline void *__zs_map_object(struct mapping_area
*area
,
767 struct page
*pages
[2], int off
, int size
)
769 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
770 area
->vm_addr
= area
->vm
->addr
;
771 return area
->vm_addr
+ off
;
774 static inline void __zs_unmap_object(struct mapping_area
*area
,
775 struct page
*pages
[2], int off
, int size
)
777 unsigned long addr
= (unsigned long)area
->vm_addr
;
779 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
782 #else /* CONFIG_PGTABLE_MAPPING */
784 static inline int __zs_cpu_up(struct mapping_area
*area
)
787 * Make sure we don't leak memory if a cpu UP notification
788 * and zs_init() race and both call zs_cpu_up() on the same cpu
792 area
->vm_buf
= (char *)__get_free_page(GFP_KERNEL
);
798 static inline void __zs_cpu_down(struct mapping_area
*area
)
801 free_page((unsigned long)area
->vm_buf
);
805 static void *__zs_map_object(struct mapping_area
*area
,
806 struct page
*pages
[2], int off
, int size
)
810 char *buf
= area
->vm_buf
;
812 /* disable page faults to match kmap_atomic() return conditions */
815 /* no read fastpath */
816 if (area
->vm_mm
== ZS_MM_WO
)
819 sizes
[0] = PAGE_SIZE
- off
;
820 sizes
[1] = size
- sizes
[0];
822 /* copy object to per-cpu buffer */
823 addr
= kmap_atomic(pages
[0]);
824 memcpy(buf
, addr
+ off
, sizes
[0]);
826 addr
= kmap_atomic(pages
[1]);
827 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
833 static void __zs_unmap_object(struct mapping_area
*area
,
834 struct page
*pages
[2], int off
, int size
)
838 char *buf
= area
->vm_buf
;
840 /* no write fastpath */
841 if (area
->vm_mm
== ZS_MM_RO
)
844 sizes
[0] = PAGE_SIZE
- off
;
845 sizes
[1] = size
- sizes
[0];
847 /* copy per-cpu buffer to object */
848 addr
= kmap_atomic(pages
[0]);
849 memcpy(addr
+ off
, buf
, sizes
[0]);
851 addr
= kmap_atomic(pages
[1]);
852 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
856 /* enable page faults to match kunmap_atomic() return conditions */
860 #endif /* CONFIG_PGTABLE_MAPPING */
862 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
865 int ret
, cpu
= (long)pcpu
;
866 struct mapping_area
*area
;
870 area
= &per_cpu(zs_map_area
, cpu
);
871 ret
= __zs_cpu_up(area
);
873 return notifier_from_errno(ret
);
876 case CPU_UP_CANCELED
:
877 area
= &per_cpu(zs_map_area
, cpu
);
885 static struct notifier_block zs_cpu_nb
= {
886 .notifier_call
= zs_cpu_notifier
889 static void zs_exit(void)
894 zpool_unregister_driver(&zs_zpool_driver
);
897 cpu_notifier_register_begin();
899 for_each_online_cpu(cpu
)
900 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
901 __unregister_cpu_notifier(&zs_cpu_nb
);
903 cpu_notifier_register_done();
906 static int zs_init(void)
910 cpu_notifier_register_begin();
912 __register_cpu_notifier(&zs_cpu_nb
);
913 for_each_online_cpu(cpu
) {
914 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
915 if (notifier_to_errno(ret
)) {
916 cpu_notifier_register_done();
921 cpu_notifier_register_done();
924 zpool_register_driver(&zs_zpool_driver
);
930 return notifier_to_errno(ret
);
934 * zs_create_pool - Creates an allocation pool to work from.
935 * @flags: allocation flags used to allocate pool metadata
937 * This function must be called before anything when using
938 * the zsmalloc allocator.
940 * On success, a pointer to the newly created pool is returned,
943 struct zs_pool
*zs_create_pool(gfp_t flags
)
946 struct zs_pool
*pool
;
948 ovhd_size
= roundup(sizeof(*pool
), PAGE_SIZE
);
949 pool
= kzalloc(ovhd_size
, GFP_KERNEL
);
953 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
955 struct size_class
*class;
957 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
958 if (size
> ZS_MAX_ALLOC_SIZE
)
959 size
= ZS_MAX_ALLOC_SIZE
;
961 class = &pool
->size_class
[i
];
964 spin_lock_init(&class->lock
);
965 class->pages_per_zspage
= get_pages_per_zspage(size
);
973 EXPORT_SYMBOL_GPL(zs_create_pool
);
975 void zs_destroy_pool(struct zs_pool
*pool
)
979 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
981 struct size_class
*class = &pool
->size_class
[i
];
983 for (fg
= 0; fg
< _ZS_NR_FULLNESS_GROUPS
; fg
++) {
984 if (class->fullness_list
[fg
]) {
985 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
992 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
995 * zs_malloc - Allocate block of given size from pool.
996 * @pool: pool to allocate from
997 * @size: size of block to allocate
999 * On success, handle to the allocated object is returned,
1001 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1003 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
)
1006 struct link_free
*link
;
1008 struct size_class
*class;
1010 struct page
*first_page
, *m_page
;
1011 unsigned long m_objidx
, m_offset
;
1013 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1016 class_idx
= get_size_class_index(size
);
1017 class = &pool
->size_class
[class_idx
];
1018 BUG_ON(class_idx
!= class->index
);
1020 spin_lock(&class->lock
);
1021 first_page
= find_get_zspage(class);
1024 spin_unlock(&class->lock
);
1025 first_page
= alloc_zspage(class, pool
->flags
);
1026 if (unlikely(!first_page
))
1029 set_zspage_mapping(first_page
, class->index
, ZS_EMPTY
);
1030 spin_lock(&class->lock
);
1031 class->pages_allocated
+= class->pages_per_zspage
;
1034 obj
= (unsigned long)first_page
->freelist
;
1035 obj_handle_to_location(obj
, &m_page
, &m_objidx
);
1036 m_offset
= obj_idx_to_offset(m_page
, m_objidx
, class->size
);
1038 link
= (struct link_free
*)kmap_atomic(m_page
) +
1039 m_offset
/ sizeof(*link
);
1040 first_page
->freelist
= link
->next
;
1041 memset(link
, POISON_INUSE
, sizeof(*link
));
1042 kunmap_atomic(link
);
1044 first_page
->inuse
++;
1045 /* Now move the zspage to another fullness group, if required */
1046 fix_fullness_group(pool
, first_page
);
1047 spin_unlock(&class->lock
);
1051 EXPORT_SYMBOL_GPL(zs_malloc
);
1053 void zs_free(struct zs_pool
*pool
, unsigned long obj
)
1055 struct link_free
*link
;
1056 struct page
*first_page
, *f_page
;
1057 unsigned long f_objidx
, f_offset
;
1060 struct size_class
*class;
1061 enum fullness_group fullness
;
1066 obj_handle_to_location(obj
, &f_page
, &f_objidx
);
1067 first_page
= get_first_page(f_page
);
1069 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1070 class = &pool
->size_class
[class_idx
];
1071 f_offset
= obj_idx_to_offset(f_page
, f_objidx
, class->size
);
1073 spin_lock(&class->lock
);
1075 /* Insert this object in containing zspage's freelist */
1076 link
= (struct link_free
*)((unsigned char *)kmap_atomic(f_page
)
1078 link
->next
= first_page
->freelist
;
1079 kunmap_atomic(link
);
1080 first_page
->freelist
= (void *)obj
;
1082 first_page
->inuse
--;
1083 fullness
= fix_fullness_group(pool
, first_page
);
1085 if (fullness
== ZS_EMPTY
)
1086 class->pages_allocated
-= class->pages_per_zspage
;
1088 spin_unlock(&class->lock
);
1090 if (fullness
== ZS_EMPTY
)
1091 free_zspage(first_page
);
1093 EXPORT_SYMBOL_GPL(zs_free
);
1096 * zs_map_object - get address of allocated object from handle.
1097 * @pool: pool from which the object was allocated
1098 * @handle: handle returned from zs_malloc
1100 * Before using an object allocated from zs_malloc, it must be mapped using
1101 * this function. When done with the object, it must be unmapped using
1104 * Only one object can be mapped per cpu at a time. There is no protection
1105 * against nested mappings.
1107 * This function returns with preemption and page faults disabled.
1109 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1113 unsigned long obj_idx
, off
;
1115 unsigned int class_idx
;
1116 enum fullness_group fg
;
1117 struct size_class
*class;
1118 struct mapping_area
*area
;
1119 struct page
*pages
[2];
1124 * Because we use per-cpu mapping areas shared among the
1125 * pools/users, we can't allow mapping in interrupt context
1126 * because it can corrupt another users mappings.
1128 BUG_ON(in_interrupt());
1130 obj_handle_to_location(handle
, &page
, &obj_idx
);
1131 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1132 class = &pool
->size_class
[class_idx
];
1133 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1135 area
= &get_cpu_var(zs_map_area
);
1137 if (off
+ class->size
<= PAGE_SIZE
) {
1138 /* this object is contained entirely within a page */
1139 area
->vm_addr
= kmap_atomic(page
);
1140 return area
->vm_addr
+ off
;
1143 /* this object spans two pages */
1145 pages
[1] = get_next_page(page
);
1148 return __zs_map_object(area
, pages
, off
, class->size
);
1150 EXPORT_SYMBOL_GPL(zs_map_object
);
1152 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1155 unsigned long obj_idx
, off
;
1157 unsigned int class_idx
;
1158 enum fullness_group fg
;
1159 struct size_class
*class;
1160 struct mapping_area
*area
;
1164 obj_handle_to_location(handle
, &page
, &obj_idx
);
1165 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1166 class = &pool
->size_class
[class_idx
];
1167 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1169 area
= this_cpu_ptr(&zs_map_area
);
1170 if (off
+ class->size
<= PAGE_SIZE
)
1171 kunmap_atomic(area
->vm_addr
);
1173 struct page
*pages
[2];
1176 pages
[1] = get_next_page(page
);
1179 __zs_unmap_object(area
, pages
, off
, class->size
);
1181 put_cpu_var(zs_map_area
);
1183 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1185 u64
zs_get_total_size_bytes(struct zs_pool
*pool
)
1190 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++)
1191 npages
+= pool
->size_class
[i
].pages_allocated
;
1193 return npages
<< PAGE_SHIFT
;
1195 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes
);
1197 module_init(zs_init
);
1198 module_exit(zs_exit
);
1200 MODULE_LICENSE("Dual BSD/GPL");
1201 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");