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 = 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
;
202 struct page
*fullness_list
[_ZS_NR_FULLNESS_GROUPS
];
206 * Placed within free objects to form a singly linked list.
207 * For every zspage, first_page->freelist gives head of this list.
209 * This must be power of 2 and less than or equal to ZS_ALIGN
212 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
217 struct size_class
*size_class
[ZS_SIZE_CLASSES
];
219 gfp_t flags
; /* allocation flags used when growing pool */
220 atomic_long_t pages_allocated
;
224 * A zspage's class index and fullness group
225 * are encoded in its (first)page->mapping
227 #define CLASS_IDX_BITS 28
228 #define FULLNESS_BITS 4
229 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
230 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
232 struct mapping_area
{
233 #ifdef CONFIG_PGTABLE_MAPPING
234 struct vm_struct
*vm
; /* vm area for mapping object that span pages */
236 char *vm_buf
; /* copy buffer for objects that span pages */
238 char *vm_addr
; /* address of kmap_atomic()'ed pages */
239 enum zs_mapmode vm_mm
; /* mapping mode */
246 static void *zs_zpool_create(gfp_t gfp
, struct zpool_ops
*zpool_ops
)
248 return zs_create_pool(gfp
);
251 static void zs_zpool_destroy(void *pool
)
253 zs_destroy_pool(pool
);
256 static int zs_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
257 unsigned long *handle
)
259 *handle
= zs_malloc(pool
, size
);
260 return *handle
? 0 : -1;
262 static void zs_zpool_free(void *pool
, unsigned long handle
)
264 zs_free(pool
, handle
);
267 static int zs_zpool_shrink(void *pool
, unsigned int pages
,
268 unsigned int *reclaimed
)
273 static void *zs_zpool_map(void *pool
, unsigned long handle
,
274 enum zpool_mapmode mm
)
276 enum zs_mapmode zs_mm
;
285 case ZPOOL_MM_RW
: /* fallthru */
291 return zs_map_object(pool
, handle
, zs_mm
);
293 static void zs_zpool_unmap(void *pool
, unsigned long handle
)
295 zs_unmap_object(pool
, handle
);
298 static u64
zs_zpool_total_size(void *pool
)
300 return zs_get_total_pages(pool
) << PAGE_SHIFT
;
303 static struct zpool_driver zs_zpool_driver
= {
305 .owner
= THIS_MODULE
,
306 .create
= zs_zpool_create
,
307 .destroy
= zs_zpool_destroy
,
308 .malloc
= zs_zpool_malloc
,
309 .free
= zs_zpool_free
,
310 .shrink
= zs_zpool_shrink
,
312 .unmap
= zs_zpool_unmap
,
313 .total_size
= zs_zpool_total_size
,
316 MODULE_ALIAS("zpool-zsmalloc");
317 #endif /* CONFIG_ZPOOL */
319 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
320 static DEFINE_PER_CPU(struct mapping_area
, zs_map_area
);
322 static int is_first_page(struct page
*page
)
324 return PagePrivate(page
);
327 static int is_last_page(struct page
*page
)
329 return PagePrivate2(page
);
332 static void get_zspage_mapping(struct page
*page
, unsigned int *class_idx
,
333 enum fullness_group
*fullness
)
336 BUG_ON(!is_first_page(page
));
338 m
= (unsigned long)page
->mapping
;
339 *fullness
= m
& FULLNESS_MASK
;
340 *class_idx
= (m
>> FULLNESS_BITS
) & CLASS_IDX_MASK
;
343 static void set_zspage_mapping(struct page
*page
, unsigned int class_idx
,
344 enum fullness_group fullness
)
347 BUG_ON(!is_first_page(page
));
349 m
= ((class_idx
& CLASS_IDX_MASK
) << FULLNESS_BITS
) |
350 (fullness
& FULLNESS_MASK
);
351 page
->mapping
= (struct address_space
*)m
;
355 * zsmalloc divides the pool into various size classes where each
356 * class maintains a list of zspages where each zspage is divided
357 * into equal sized chunks. Each allocation falls into one of these
358 * classes depending on its size. This function returns index of the
359 * size class which has chunk size big enough to hold the give size.
361 static int get_size_class_index(int size
)
365 if (likely(size
> ZS_MIN_ALLOC_SIZE
))
366 idx
= DIV_ROUND_UP(size
- ZS_MIN_ALLOC_SIZE
,
367 ZS_SIZE_CLASS_DELTA
);
373 * For each size class, zspages are divided into different groups
374 * depending on how "full" they are. This was done so that we could
375 * easily find empty or nearly empty zspages when we try to shrink
376 * the pool (not yet implemented). This function returns fullness
377 * status of the given page.
379 static enum fullness_group
get_fullness_group(struct page
*page
)
381 int inuse
, max_objects
;
382 enum fullness_group fg
;
383 BUG_ON(!is_first_page(page
));
386 max_objects
= page
->objects
;
390 else if (inuse
== max_objects
)
392 else if (inuse
<= max_objects
/ fullness_threshold_frac
)
393 fg
= ZS_ALMOST_EMPTY
;
401 * Each size class maintains various freelists and zspages are assigned
402 * to one of these freelists based on the number of live objects they
403 * have. This functions inserts the given zspage into the freelist
404 * identified by <class, fullness_group>.
406 static void insert_zspage(struct page
*page
, struct size_class
*class,
407 enum fullness_group fullness
)
411 BUG_ON(!is_first_page(page
));
413 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
416 head
= &class->fullness_list
[fullness
];
418 list_add_tail(&page
->lru
, &(*head
)->lru
);
424 * This function removes the given zspage from the freelist identified
425 * by <class, fullness_group>.
427 static void remove_zspage(struct page
*page
, struct size_class
*class,
428 enum fullness_group fullness
)
432 BUG_ON(!is_first_page(page
));
434 if (fullness
>= _ZS_NR_FULLNESS_GROUPS
)
437 head
= &class->fullness_list
[fullness
];
439 if (list_empty(&(*head
)->lru
))
441 else if (*head
== page
)
442 *head
= (struct page
*)list_entry((*head
)->lru
.next
,
445 list_del_init(&page
->lru
);
449 * Each size class maintains zspages in different fullness groups depending
450 * on the number of live objects they contain. When allocating or freeing
451 * objects, the fullness status of the page can change, say, from ALMOST_FULL
452 * to ALMOST_EMPTY when freeing an object. This function checks if such
453 * a status change has occurred for the given page and accordingly moves the
454 * page from the freelist of the old fullness group to that of the new
457 static enum fullness_group
fix_fullness_group(struct zs_pool
*pool
,
461 struct size_class
*class;
462 enum fullness_group currfg
, newfg
;
464 BUG_ON(!is_first_page(page
));
466 get_zspage_mapping(page
, &class_idx
, &currfg
);
467 newfg
= get_fullness_group(page
);
471 class = pool
->size_class
[class_idx
];
472 remove_zspage(page
, class, currfg
);
473 insert_zspage(page
, class, newfg
);
474 set_zspage_mapping(page
, class_idx
, newfg
);
481 * We have to decide on how many pages to link together
482 * to form a zspage for each size class. This is important
483 * to reduce wastage due to unusable space left at end of
484 * each zspage which is given as:
485 * wastage = Zp - Zp % size_class
486 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
488 * For example, for size class of 3/8 * PAGE_SIZE, we should
489 * link together 3 PAGE_SIZE sized pages to form a zspage
490 * since then we can perfectly fit in 8 such objects.
492 static int get_pages_per_zspage(int class_size
)
494 int i
, max_usedpc
= 0;
495 /* zspage order which gives maximum used size per KB */
496 int max_usedpc_order
= 1;
498 for (i
= 1; i
<= ZS_MAX_PAGES_PER_ZSPAGE
; i
++) {
502 zspage_size
= i
* PAGE_SIZE
;
503 waste
= zspage_size
% class_size
;
504 usedpc
= (zspage_size
- waste
) * 100 / zspage_size
;
506 if (usedpc
> max_usedpc
) {
508 max_usedpc_order
= i
;
512 return max_usedpc_order
;
516 * A single 'zspage' is composed of many system pages which are
517 * linked together using fields in struct page. This function finds
518 * the first/head page, given any component page of a zspage.
520 static struct page
*get_first_page(struct page
*page
)
522 if (is_first_page(page
))
525 return page
->first_page
;
528 static struct page
*get_next_page(struct page
*page
)
532 if (is_last_page(page
))
534 else if (is_first_page(page
))
535 next
= (struct page
*)page_private(page
);
537 next
= list_entry(page
->lru
.next
, struct page
, lru
);
543 * Encode <page, obj_idx> as a single handle value.
544 * On hardware platforms with physical memory starting at 0x0 the pfn
545 * could be 0 so we ensure that the handle will never be 0 by adjusting the
546 * encoded obj_idx value before encoding.
548 static void *obj_location_to_handle(struct page
*page
, unsigned long obj_idx
)
550 unsigned long handle
;
557 handle
= page_to_pfn(page
) << OBJ_INDEX_BITS
;
558 handle
|= ((obj_idx
+ 1) & OBJ_INDEX_MASK
);
560 return (void *)handle
;
564 * Decode <page, obj_idx> pair from the given object handle. We adjust the
565 * decoded obj_idx back to its original value since it was adjusted in
566 * obj_location_to_handle().
568 static void obj_handle_to_location(unsigned long handle
, struct page
**page
,
569 unsigned long *obj_idx
)
571 *page
= pfn_to_page(handle
>> OBJ_INDEX_BITS
);
572 *obj_idx
= (handle
& OBJ_INDEX_MASK
) - 1;
575 static unsigned long obj_idx_to_offset(struct page
*page
,
576 unsigned long obj_idx
, int class_size
)
578 unsigned long off
= 0;
580 if (!is_first_page(page
))
583 return off
+ obj_idx
* class_size
;
586 static void reset_page(struct page
*page
)
588 clear_bit(PG_private
, &page
->flags
);
589 clear_bit(PG_private_2
, &page
->flags
);
590 set_page_private(page
, 0);
591 page
->mapping
= NULL
;
592 page
->freelist
= NULL
;
593 page_mapcount_reset(page
);
596 static void free_zspage(struct page
*first_page
)
598 struct page
*nextp
, *tmp
, *head_extra
;
600 BUG_ON(!is_first_page(first_page
));
601 BUG_ON(first_page
->inuse
);
603 head_extra
= (struct page
*)page_private(first_page
);
605 reset_page(first_page
);
606 __free_page(first_page
);
608 /* zspage with only 1 system page */
612 list_for_each_entry_safe(nextp
, tmp
, &head_extra
->lru
, lru
) {
613 list_del(&nextp
->lru
);
617 reset_page(head_extra
);
618 __free_page(head_extra
);
621 /* Initialize a newly allocated zspage */
622 static void init_zspage(struct page
*first_page
, struct size_class
*class)
624 unsigned long off
= 0;
625 struct page
*page
= first_page
;
627 BUG_ON(!is_first_page(first_page
));
629 struct page
*next_page
;
630 struct link_free
*link
;
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 vaddr
= kmap_atomic(page
);
644 link
= (struct link_free
*)vaddr
+ off
/ sizeof(*link
);
646 while ((off
+= class->size
) < PAGE_SIZE
) {
647 link
->next
= obj_location_to_handle(page
, i
++);
648 link
+= class->size
/ sizeof(*link
);
652 * We now come to the last (full or partial) object on this
653 * page, which must point to the first object on the next
656 next_page
= get_next_page(page
);
657 link
->next
= obj_location_to_handle(next_page
, 0);
658 kunmap_atomic(vaddr
);
665 * Allocate a zspage for the given size class
667 static struct page
*alloc_zspage(struct size_class
*class, gfp_t flags
)
670 struct page
*first_page
= NULL
, *uninitialized_var(prev_page
);
673 * Allocate individual pages and link them together as:
674 * 1. first page->private = first sub-page
675 * 2. all sub-pages are linked together using page->lru
676 * 3. each sub-page is linked to the first page using page->first_page
678 * For each size class, First/Head pages are linked together using
679 * page->lru. Also, we set PG_private to identify the first page
680 * (i.e. no other sub-page has this flag set) and PG_private_2 to
681 * identify the last page.
684 for (i
= 0; i
< class->pages_per_zspage
; i
++) {
687 page
= alloc_page(flags
);
691 INIT_LIST_HEAD(&page
->lru
);
692 if (i
== 0) { /* first page */
693 SetPagePrivate(page
);
694 set_page_private(page
, 0);
696 first_page
->inuse
= 0;
699 set_page_private(first_page
, (unsigned long)page
);
701 page
->first_page
= first_page
;
703 list_add(&page
->lru
, &prev_page
->lru
);
704 if (i
== class->pages_per_zspage
- 1) /* last page */
705 SetPagePrivate2(page
);
709 init_zspage(first_page
, class);
711 first_page
->freelist
= obj_location_to_handle(first_page
, 0);
712 /* Maximum number of objects we can store in this zspage */
713 first_page
->objects
= class->pages_per_zspage
* PAGE_SIZE
/ class->size
;
715 error
= 0; /* Success */
718 if (unlikely(error
) && first_page
) {
719 free_zspage(first_page
);
726 static struct page
*find_get_zspage(struct size_class
*class)
731 for (i
= 0; i
< _ZS_NR_FULLNESS_GROUPS
; i
++) {
732 page
= class->fullness_list
[i
];
740 #ifdef CONFIG_PGTABLE_MAPPING
741 static inline int __zs_cpu_up(struct mapping_area
*area
)
744 * Make sure we don't leak memory if a cpu UP notification
745 * and zs_init() race and both call zs_cpu_up() on the same cpu
749 area
->vm
= alloc_vm_area(PAGE_SIZE
* 2, NULL
);
755 static inline void __zs_cpu_down(struct mapping_area
*area
)
758 free_vm_area(area
->vm
);
762 static inline void *__zs_map_object(struct mapping_area
*area
,
763 struct page
*pages
[2], int off
, int size
)
765 BUG_ON(map_vm_area(area
->vm
, PAGE_KERNEL
, pages
));
766 area
->vm_addr
= area
->vm
->addr
;
767 return area
->vm_addr
+ off
;
770 static inline void __zs_unmap_object(struct mapping_area
*area
,
771 struct page
*pages
[2], int off
, int size
)
773 unsigned long addr
= (unsigned long)area
->vm_addr
;
775 unmap_kernel_range(addr
, PAGE_SIZE
* 2);
778 #else /* CONFIG_PGTABLE_MAPPING */
780 static inline int __zs_cpu_up(struct mapping_area
*area
)
783 * Make sure we don't leak memory if a cpu UP notification
784 * and zs_init() race and both call zs_cpu_up() on the same cpu
788 area
->vm_buf
= (char *)__get_free_page(GFP_KERNEL
);
794 static inline void __zs_cpu_down(struct mapping_area
*area
)
797 free_page((unsigned long)area
->vm_buf
);
801 static void *__zs_map_object(struct mapping_area
*area
,
802 struct page
*pages
[2], int off
, int size
)
806 char *buf
= area
->vm_buf
;
808 /* disable page faults to match kmap_atomic() return conditions */
811 /* no read fastpath */
812 if (area
->vm_mm
== ZS_MM_WO
)
815 sizes
[0] = PAGE_SIZE
- off
;
816 sizes
[1] = size
- sizes
[0];
818 /* copy object to per-cpu buffer */
819 addr
= kmap_atomic(pages
[0]);
820 memcpy(buf
, addr
+ off
, sizes
[0]);
822 addr
= kmap_atomic(pages
[1]);
823 memcpy(buf
+ sizes
[0], addr
, sizes
[1]);
829 static void __zs_unmap_object(struct mapping_area
*area
,
830 struct page
*pages
[2], int off
, int size
)
834 char *buf
= area
->vm_buf
;
836 /* no write fastpath */
837 if (area
->vm_mm
== ZS_MM_RO
)
840 sizes
[0] = PAGE_SIZE
- off
;
841 sizes
[1] = size
- sizes
[0];
843 /* copy per-cpu buffer to object */
844 addr
= kmap_atomic(pages
[0]);
845 memcpy(addr
+ off
, buf
, sizes
[0]);
847 addr
= kmap_atomic(pages
[1]);
848 memcpy(addr
, buf
+ sizes
[0], sizes
[1]);
852 /* enable page faults to match kunmap_atomic() return conditions */
856 #endif /* CONFIG_PGTABLE_MAPPING */
858 static int zs_cpu_notifier(struct notifier_block
*nb
, unsigned long action
,
861 int ret
, cpu
= (long)pcpu
;
862 struct mapping_area
*area
;
866 area
= &per_cpu(zs_map_area
, cpu
);
867 ret
= __zs_cpu_up(area
);
869 return notifier_from_errno(ret
);
872 case CPU_UP_CANCELED
:
873 area
= &per_cpu(zs_map_area
, cpu
);
881 static struct notifier_block zs_cpu_nb
= {
882 .notifier_call
= zs_cpu_notifier
885 static void zs_unregister_cpu_notifier(void)
889 cpu_notifier_register_begin();
891 for_each_online_cpu(cpu
)
892 zs_cpu_notifier(NULL
, CPU_DEAD
, (void *)(long)cpu
);
893 __unregister_cpu_notifier(&zs_cpu_nb
);
895 cpu_notifier_register_done();
898 static int zs_register_cpu_notifier(void)
900 int cpu
, uninitialized_var(ret
);
902 cpu_notifier_register_begin();
904 __register_cpu_notifier(&zs_cpu_nb
);
905 for_each_online_cpu(cpu
) {
906 ret
= zs_cpu_notifier(NULL
, CPU_UP_PREPARE
, (void *)(long)cpu
);
907 if (notifier_to_errno(ret
))
911 cpu_notifier_register_done();
912 return notifier_to_errno(ret
);
915 static void __exit
zs_exit(void)
918 zpool_unregister_driver(&zs_zpool_driver
);
920 zs_unregister_cpu_notifier();
923 static int __init
zs_init(void)
925 int ret
= zs_register_cpu_notifier();
928 zs_unregister_cpu_notifier();
933 zpool_register_driver(&zs_zpool_driver
);
938 static unsigned int get_maxobj_per_zspage(int size
, int pages_per_zspage
)
940 return pages_per_zspage
* PAGE_SIZE
/ size
;
943 static bool can_merge(struct size_class
*prev
, int size
, int pages_per_zspage
)
945 if (prev
->pages_per_zspage
!= pages_per_zspage
)
948 if (get_maxobj_per_zspage(prev
->size
, prev
->pages_per_zspage
)
949 != get_maxobj_per_zspage(size
, pages_per_zspage
))
956 * zs_create_pool - Creates an allocation pool to work from.
957 * @flags: allocation flags used to allocate pool metadata
959 * This function must be called before anything when using
960 * the zsmalloc allocator.
962 * On success, a pointer to the newly created pool is returned,
965 struct zs_pool
*zs_create_pool(gfp_t flags
)
968 struct zs_pool
*pool
;
970 ovhd_size
= roundup(sizeof(*pool
), PAGE_SIZE
);
971 pool
= kzalloc(ovhd_size
, GFP_KERNEL
);
976 * Iterate reversly, because, size of size_class that we want to use
977 * for merging should be larger or equal to current size.
979 for (i
= ZS_SIZE_CLASSES
- 1; i
>= 0; i
--) {
981 int pages_per_zspage
;
982 struct size_class
*class;
983 struct size_class
*prev_class
;
985 size
= ZS_MIN_ALLOC_SIZE
+ i
* ZS_SIZE_CLASS_DELTA
;
986 if (size
> ZS_MAX_ALLOC_SIZE
)
987 size
= ZS_MAX_ALLOC_SIZE
;
988 pages_per_zspage
= get_pages_per_zspage(size
);
991 * size_class is used for normal zsmalloc operation such
992 * as alloc/free for that size. Although it is natural that we
993 * have one size_class for each size, there is a chance that we
994 * can get more memory utilization if we use one size_class for
995 * many different sizes whose size_class have same
996 * characteristics. So, we makes size_class point to
997 * previous size_class if possible.
999 if (i
< ZS_SIZE_CLASSES
- 1) {
1000 prev_class
= pool
->size_class
[i
+ 1];
1001 if (can_merge(prev_class
, size
, pages_per_zspage
)) {
1002 pool
->size_class
[i
] = prev_class
;
1007 class = kzalloc(sizeof(struct size_class
), GFP_KERNEL
);
1013 class->pages_per_zspage
= pages_per_zspage
;
1014 spin_lock_init(&class->lock
);
1015 pool
->size_class
[i
] = class;
1018 pool
->flags
= flags
;
1023 zs_destroy_pool(pool
);
1026 EXPORT_SYMBOL_GPL(zs_create_pool
);
1028 void zs_destroy_pool(struct zs_pool
*pool
)
1032 for (i
= 0; i
< ZS_SIZE_CLASSES
; i
++) {
1034 struct size_class
*class = pool
->size_class
[i
];
1039 if (class->index
!= i
)
1042 for (fg
= 0; fg
< _ZS_NR_FULLNESS_GROUPS
; fg
++) {
1043 if (class->fullness_list
[fg
]) {
1044 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1052 EXPORT_SYMBOL_GPL(zs_destroy_pool
);
1055 * zs_malloc - Allocate block of given size from pool.
1056 * @pool: pool to allocate from
1057 * @size: size of block to allocate
1059 * On success, handle to the allocated object is returned,
1061 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1063 unsigned long zs_malloc(struct zs_pool
*pool
, size_t size
)
1066 struct link_free
*link
;
1067 struct size_class
*class;
1070 struct page
*first_page
, *m_page
;
1071 unsigned long m_objidx
, m_offset
;
1073 if (unlikely(!size
|| size
> ZS_MAX_ALLOC_SIZE
))
1076 class = pool
->size_class
[get_size_class_index(size
)];
1078 spin_lock(&class->lock
);
1079 first_page
= find_get_zspage(class);
1082 spin_unlock(&class->lock
);
1083 first_page
= alloc_zspage(class, pool
->flags
);
1084 if (unlikely(!first_page
))
1087 set_zspage_mapping(first_page
, class->index
, ZS_EMPTY
);
1088 atomic_long_add(class->pages_per_zspage
,
1089 &pool
->pages_allocated
);
1090 spin_lock(&class->lock
);
1093 obj
= (unsigned long)first_page
->freelist
;
1094 obj_handle_to_location(obj
, &m_page
, &m_objidx
);
1095 m_offset
= obj_idx_to_offset(m_page
, m_objidx
, class->size
);
1097 vaddr
= kmap_atomic(m_page
);
1098 link
= (struct link_free
*)vaddr
+ m_offset
/ sizeof(*link
);
1099 first_page
->freelist
= link
->next
;
1100 memset(link
, POISON_INUSE
, sizeof(*link
));
1101 kunmap_atomic(vaddr
);
1103 first_page
->inuse
++;
1104 /* Now move the zspage to another fullness group, if required */
1105 fix_fullness_group(pool
, first_page
);
1106 spin_unlock(&class->lock
);
1110 EXPORT_SYMBOL_GPL(zs_malloc
);
1112 void zs_free(struct zs_pool
*pool
, unsigned long obj
)
1114 struct link_free
*link
;
1115 struct page
*first_page
, *f_page
;
1116 unsigned long f_objidx
, f_offset
;
1120 struct size_class
*class;
1121 enum fullness_group fullness
;
1126 obj_handle_to_location(obj
, &f_page
, &f_objidx
);
1127 first_page
= get_first_page(f_page
);
1129 get_zspage_mapping(first_page
, &class_idx
, &fullness
);
1130 class = pool
->size_class
[class_idx
];
1131 f_offset
= obj_idx_to_offset(f_page
, f_objidx
, class->size
);
1133 spin_lock(&class->lock
);
1135 /* Insert this object in containing zspage's freelist */
1136 vaddr
= kmap_atomic(f_page
);
1137 link
= (struct link_free
*)(vaddr
+ f_offset
);
1138 link
->next
= first_page
->freelist
;
1139 kunmap_atomic(vaddr
);
1140 first_page
->freelist
= (void *)obj
;
1142 first_page
->inuse
--;
1143 fullness
= fix_fullness_group(pool
, first_page
);
1144 spin_unlock(&class->lock
);
1146 if (fullness
== ZS_EMPTY
) {
1147 atomic_long_sub(class->pages_per_zspage
,
1148 &pool
->pages_allocated
);
1149 free_zspage(first_page
);
1152 EXPORT_SYMBOL_GPL(zs_free
);
1155 * zs_map_object - get address of allocated object from handle.
1156 * @pool: pool from which the object was allocated
1157 * @handle: handle returned from zs_malloc
1159 * Before using an object allocated from zs_malloc, it must be mapped using
1160 * this function. When done with the object, it must be unmapped using
1163 * Only one object can be mapped per cpu at a time. There is no protection
1164 * against nested mappings.
1166 * This function returns with preemption and page faults disabled.
1168 void *zs_map_object(struct zs_pool
*pool
, unsigned long handle
,
1172 unsigned long obj_idx
, off
;
1174 unsigned int class_idx
;
1175 enum fullness_group fg
;
1176 struct size_class
*class;
1177 struct mapping_area
*area
;
1178 struct page
*pages
[2];
1183 * Because we use per-cpu mapping areas shared among the
1184 * pools/users, we can't allow mapping in interrupt context
1185 * because it can corrupt another users mappings.
1187 BUG_ON(in_interrupt());
1189 obj_handle_to_location(handle
, &page
, &obj_idx
);
1190 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1191 class = pool
->size_class
[class_idx
];
1192 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1194 area
= &get_cpu_var(zs_map_area
);
1196 if (off
+ class->size
<= PAGE_SIZE
) {
1197 /* this object is contained entirely within a page */
1198 area
->vm_addr
= kmap_atomic(page
);
1199 return area
->vm_addr
+ off
;
1202 /* this object spans two pages */
1204 pages
[1] = get_next_page(page
);
1207 return __zs_map_object(area
, pages
, off
, class->size
);
1209 EXPORT_SYMBOL_GPL(zs_map_object
);
1211 void zs_unmap_object(struct zs_pool
*pool
, unsigned long handle
)
1214 unsigned long obj_idx
, off
;
1216 unsigned int class_idx
;
1217 enum fullness_group fg
;
1218 struct size_class
*class;
1219 struct mapping_area
*area
;
1223 obj_handle_to_location(handle
, &page
, &obj_idx
);
1224 get_zspage_mapping(get_first_page(page
), &class_idx
, &fg
);
1225 class = pool
->size_class
[class_idx
];
1226 off
= obj_idx_to_offset(page
, obj_idx
, class->size
);
1228 area
= this_cpu_ptr(&zs_map_area
);
1229 if (off
+ class->size
<= PAGE_SIZE
)
1230 kunmap_atomic(area
->vm_addr
);
1232 struct page
*pages
[2];
1235 pages
[1] = get_next_page(page
);
1238 __zs_unmap_object(area
, pages
, off
, class->size
);
1240 put_cpu_var(zs_map_area
);
1242 EXPORT_SYMBOL_GPL(zs_unmap_object
);
1244 unsigned long zs_get_total_pages(struct zs_pool
*pool
)
1246 return atomic_long_read(&pool
->pages_allocated
);
1248 EXPORT_SYMBOL_GPL(zs_get_total_pages
);
1250 module_init(zs_init
);
1251 module_exit(zs_exit
);
1253 MODULE_LICENSE("Dual BSD/GPL");
1254 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");