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
16 * This allocator is designed for use with zcache and zram. Thus, the
17 * allocator is supposed to work well under low memory conditions. In
18 * particular, it never attempts higher order page allocation which is
19 * very likely to fail under memory pressure. On the other hand, if we
20 * just use single (0-order) pages, it would suffer from very high
21 * fragmentation -- any object of size PAGE_SIZE/2 or larger would occupy
22 * an entire page. This was one of the major issues with its predecessor
25 * To overcome these issues, zsmalloc allocates a bunch of 0-order pages
26 * and links them together using various 'struct page' fields. These linked
27 * pages act as a single higher-order page i.e. an object can span 0-order
28 * page boundaries. The code refers to these linked pages as a single entity
31 * Following is how we use various fields and flags of underlying
32 * struct page(s) to form a zspage.
34 * Usage of struct page fields:
35 * page->first_page: points to the first component (0-order) page
36 * page->index (union with page->freelist): offset of the first object
37 * starting in this page. For the first page, this is
38 * always 0, so we use this field (aka freelist) to point
39 * to the first free object in zspage.
40 * page->lru: links together all component pages (except the first page)
43 * For _first_ page only:
45 * page->private (union with page->first_page): refers to the
46 * component page after the first page
47 * page->freelist: points to the first free object in zspage.
48 * Free objects are linked together using in-place
50 * page->objects: maximum number of objects we can store in this
51 * zspage (class->zspage_order * PAGE_SIZE / class->size)
52 * page->lru: links together first pages of various zspages.
53 * Basically forming list of zspages in a fullness group.
54 * page->mapping: class index and fullness group of the zspage
56 * Usage of struct page flags:
57 * PG_private: identifies the first component page
58 * PG_private2: identifies the last component page
62 #ifdef CONFIG_ZSMALLOC_DEBUG
66 #include <linux/module.h>
67 #include <linux/kernel.h>
68 #include <linux/bitops.h>
69 #include <linux/errno.h>
70 #include <linux/highmem.h>
71 #include <linux/init.h>
72 #include <linux/string.h>
73 #include <linux/slab.h>
74 #include <asm/tlbflush.h>
75 #include <asm/pgtable.h>
76 #include <linux/cpumask.h>
77 #include <linux/cpu.h>
78 #include <linux/vmalloc.h>
79 #include <linux/hardirq.h>
80 #include <linux/spinlock.h>
81 #include <linux/types.h>
82 #include <linux/zsmalloc.h>
83 #include <linux/zpool.h>
86 * This must be power of 2 and greater than of equal to sizeof(link_free).
87 * These two conditions ensure that any 'struct link_free' itself doesn't
88 * span more than 1 page which avoids complex case of mapping 2 pages simply
89 * to restore link_free pointer values.
94 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
95 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
97 #define ZS_MAX_ZSPAGE_ORDER 2
98 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
101 * Object location (<PFN>, <obj_idx>) is encoded as
102 * as single (void *) handle value.
104 * Note that object index <obj_idx> is relative to system
105 * page <PFN> it is stored in, so for each sub-page belonging
106 * to a zspage, obj_idx starts with 0.
108 * This is made more complicated by various memory models and PAE.
111 #ifndef MAX_PHYSMEM_BITS
112 #ifdef CONFIG_HIGHMEM64G
113 #define MAX_PHYSMEM_BITS 36
114 #else /* !CONFIG_HIGHMEM64G */
116 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
119 #define MAX_PHYSMEM_BITS BITS_PER_LONG
122 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
123 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
124 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
126 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
127 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
128 #define ZS_MIN_ALLOC_SIZE \
129 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
130 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
133 * On systems with 4K page size, this gives 255 size classes! There is a
135 * - Large number of size classes is potentially wasteful as free page are
136 * spread across these classes
137 * - Small number of size classes causes large internal fragmentation
138 * - Probably its better to use specific size classes (empirically
139 * determined). NOTE: all those class sizes must be set as multiple of
140 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
142 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
145 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
146 #define ZS_SIZE_CLASSES ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
147 ZS_SIZE_CLASS_DELTA + 1)
150 * We do not maintain any list for completely empty or full pages
152 enum fullness_group {
155 _ZS_NR_FULLNESS_GROUPS,
162 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
164 * n = number of allocated objects
165 * N = total number of objects zspage can store
166 * f = 1/fullness_threshold_frac
168 * Similarly, we assign zspage to:
169 * ZS_ALMOST_FULL when n > N / f
170 * ZS_EMPTY when n == 0
171 * ZS_FULL when n == N
173 * (see: fix_fullness_group())
175 static const int fullness_threshold_frac = 4;
179 * Size of objects stored in this class. Must be multiple
185 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
186 int pages_per_zspage;
193 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
197 * Placed within free objects to form a singly linked list.
198 * For every zspage, first_page->freelist gives head of this list.
200 * This must be power of 2 and less than or equal to ZS_ALIGN
203 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
208 struct size_class size_class[ZS_SIZE_CLASSES];
210 gfp_t flags; /* allocation flags used when growing pool */
214 * A zspage's class index and fullness group
215 * are encoded in its (first)page->mapping
217 #define CLASS_IDX_BITS 28
218 #define FULLNESS_BITS 4
219 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
220 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
223 * By default, zsmalloc uses a copy-based object mapping method to access
224 * allocations that span two pages. However, if a particular architecture
225 * performs VM mapping faster than copying, then it should be added here
226 * so that USE_PGTABLE_MAPPING is defined. This causes zsmalloc to use
227 * page table mapping rather than copying for object mapping.
229 #if defined(CONFIG_ARM) && !defined(MODULE)
230 #define USE_PGTABLE_MAPPING
233 struct mapping_area {
234 #ifdef USE_PGTABLE_MAPPING
235 struct vm_struct *vm; /* vm area for mapping object that span pages */
237 char *vm_buf; /* copy buffer for objects that span pages */
239 char *vm_addr; /* address of kmap_atomic()'ed pages */
240 enum zs_mapmode vm_mm; /* mapping mode */
247 static void *zs_zpool_create(gfp_t gfp, struct zpool_ops *zpool_ops)
249 return zs_create_pool(gfp);
252 static void zs_zpool_destroy(void *pool)
254 zs_destroy_pool(pool);
257 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
258 unsigned long *handle)
260 *handle = zs_malloc(pool, size);
261 return *handle ? 0 : -1;
263 static void zs_zpool_free(void *pool, unsigned long handle)
265 zs_free(pool, handle);
268 static int zs_zpool_shrink(void *pool, unsigned int pages,
269 unsigned int *reclaimed)
274 static void *zs_zpool_map(void *pool, unsigned long handle,
275 enum zpool_mapmode mm)
277 enum zs_mapmode zs_mm;
286 case ZPOOL_MM_RW: /* fallthru */
292 return zs_map_object(pool, handle, zs_mm);
294 static void zs_zpool_unmap(void *pool, unsigned long handle)
296 zs_unmap_object(pool, handle);
299 static u64 zs_zpool_total_size(void *pool)
301 return zs_get_total_size_bytes(pool);
304 static struct zpool_driver zs_zpool_driver = {
306 .owner = THIS_MODULE,
307 .create = zs_zpool_create,
308 .destroy = zs_zpool_destroy,
309 .malloc = zs_zpool_malloc,
310 .free = zs_zpool_free,
311 .shrink = zs_zpool_shrink,
313 .unmap = zs_zpool_unmap,
314 .total_size = zs_zpool_total_size,
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;
354 static int get_size_class_index(int size)
358 if (likely(size > ZS_MIN_ALLOC_SIZE))
359 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
360 ZS_SIZE_CLASS_DELTA);
365 static enum fullness_group get_fullness_group(struct page *page)
367 int inuse, max_objects;
368 enum fullness_group fg;
369 BUG_ON(!is_first_page(page));
372 max_objects = page->objects;
376 else if (inuse == max_objects)
378 else if (inuse <= max_objects / fullness_threshold_frac)
379 fg = ZS_ALMOST_EMPTY;
386 static void insert_zspage(struct page *page, struct size_class *class,
387 enum fullness_group fullness)
391 BUG_ON(!is_first_page(page));
393 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
396 head = &class->fullness_list[fullness];
398 list_add_tail(&page->lru, &(*head)->lru);
403 static void remove_zspage(struct page *page, struct size_class *class,
404 enum fullness_group fullness)
408 BUG_ON(!is_first_page(page));
410 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
413 head = &class->fullness_list[fullness];
415 if (list_empty(&(*head)->lru))
417 else if (*head == page)
418 *head = (struct page *)list_entry((*head)->lru.next,
421 list_del_init(&page->lru);
424 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
428 struct size_class *class;
429 enum fullness_group currfg, newfg;
431 BUG_ON(!is_first_page(page));
433 get_zspage_mapping(page, &class_idx, &currfg);
434 newfg = get_fullness_group(page);
438 class = &pool->size_class[class_idx];
439 remove_zspage(page, class, currfg);
440 insert_zspage(page, class, newfg);
441 set_zspage_mapping(page, class_idx, newfg);
448 * We have to decide on how many pages to link together
449 * to form a zspage for each size class. This is important
450 * to reduce wastage due to unusable space left at end of
451 * each zspage which is given as:
452 * wastage = Zp - Zp % size_class
453 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
455 * For example, for size class of 3/8 * PAGE_SIZE, we should
456 * link together 3 PAGE_SIZE sized pages to form a zspage
457 * since then we can perfectly fit in 8 such objects.
459 static int get_pages_per_zspage(int class_size)
461 int i, max_usedpc = 0;
462 /* zspage order which gives maximum used size per KB */
463 int max_usedpc_order = 1;
465 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
469 zspage_size = i * PAGE_SIZE;
470 waste = zspage_size % class_size;
471 usedpc = (zspage_size - waste) * 100 / zspage_size;
473 if (usedpc > max_usedpc) {
475 max_usedpc_order = i;
479 return max_usedpc_order;
483 * A single 'zspage' is composed of many system pages which are
484 * linked together using fields in struct page. This function finds
485 * the first/head page, given any component page of a zspage.
487 static struct page *get_first_page(struct page *page)
489 if (is_first_page(page))
492 return page->first_page;
495 static struct page *get_next_page(struct page *page)
499 if (is_last_page(page))
501 else if (is_first_page(page))
502 next = (struct page *)page->private;
504 next = list_entry(page->lru.next, struct page, lru);
510 * Encode <page, obj_idx> as a single handle value.
511 * On hardware platforms with physical memory starting at 0x0 the pfn
512 * could be 0 so we ensure that the handle will never be 0 by adjusting the
513 * encoded obj_idx value before encoding.
515 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
517 unsigned long handle;
524 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
525 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
527 return (void *)handle;
531 * Decode <page, obj_idx> pair from the given object handle. We adjust the
532 * decoded obj_idx back to its original value since it was adjusted in
533 * obj_location_to_handle().
535 static void obj_handle_to_location(unsigned long handle, struct page **page,
536 unsigned long *obj_idx)
538 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
539 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
542 static unsigned long obj_idx_to_offset(struct page *page,
543 unsigned long obj_idx, int class_size)
545 unsigned long off = 0;
547 if (!is_first_page(page))
550 return off + obj_idx * class_size;
553 static void reset_page(struct page *page)
555 clear_bit(PG_private, &page->flags);
556 clear_bit(PG_private_2, &page->flags);
557 set_page_private(page, 0);
558 page->mapping = NULL;
559 page->freelist = NULL;
560 page_mapcount_reset(page);
563 static void free_zspage(struct page *first_page)
565 struct page *nextp, *tmp, *head_extra;
567 BUG_ON(!is_first_page(first_page));
568 BUG_ON(first_page->inuse);
570 head_extra = (struct page *)page_private(first_page);
572 reset_page(first_page);
573 __free_page(first_page);
575 /* zspage with only 1 system page */
579 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
580 list_del(&nextp->lru);
584 reset_page(head_extra);
585 __free_page(head_extra);
588 /* Initialize a newly allocated zspage */
589 static void init_zspage(struct page *first_page, struct size_class *class)
591 unsigned long off = 0;
592 struct page *page = first_page;
594 BUG_ON(!is_first_page(first_page));
596 struct page *next_page;
597 struct link_free *link;
598 unsigned int i, objs_on_page;
601 * page->index stores offset of first object starting
602 * in the page. For the first page, this is always 0,
603 * so we use first_page->index (aka ->freelist) to store
604 * head of corresponding zspage's freelist.
606 if (page != first_page)
609 link = (struct link_free *)kmap_atomic(page) +
611 objs_on_page = (PAGE_SIZE - off) / class->size;
613 for (i = 1; i <= objs_on_page; i++) {
615 if (off < PAGE_SIZE) {
616 link->next = obj_location_to_handle(page, i);
617 link += class->size / sizeof(*link);
622 * We now come to the last (full or partial) object on this
623 * page, which must point to the first object on the next
626 next_page = get_next_page(page);
627 link->next = obj_location_to_handle(next_page, 0);
630 off = (off + class->size) % PAGE_SIZE;
635 * Allocate a zspage for the given size class
637 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
640 struct page *first_page = NULL, *uninitialized_var(prev_page);
643 * Allocate individual pages and link them together as:
644 * 1. first page->private = first sub-page
645 * 2. all sub-pages are linked together using page->lru
646 * 3. each sub-page is linked to the first page using page->first_page
648 * For each size class, First/Head pages are linked together using
649 * page->lru. Also, we set PG_private to identify the first page
650 * (i.e. no other sub-page has this flag set) and PG_private_2 to
651 * identify the last page.
654 for (i = 0; i < class->pages_per_zspage; i++) {
657 page = alloc_page(flags);
661 INIT_LIST_HEAD(&page->lru);
662 if (i == 0) { /* first page */
663 SetPagePrivate(page);
664 set_page_private(page, 0);
666 first_page->inuse = 0;
669 first_page->private = (unsigned long)page;
671 page->first_page = first_page;
673 list_add(&page->lru, &prev_page->lru);
674 if (i == class->pages_per_zspage - 1) /* last page */
675 SetPagePrivate2(page);
679 init_zspage(first_page, class);
681 first_page->freelist = obj_location_to_handle(first_page, 0);
682 /* Maximum number of objects we can store in this zspage */
683 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
685 error = 0; /* Success */
688 if (unlikely(error) && first_page) {
689 free_zspage(first_page);
696 static struct page *find_get_zspage(struct size_class *class)
701 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
702 page = class->fullness_list[i];
710 #ifdef USE_PGTABLE_MAPPING
711 static inline int __zs_cpu_up(struct mapping_area *area)
714 * Make sure we don't leak memory if a cpu UP notification
715 * and zs_init() race and both call zs_cpu_up() on the same cpu
719 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
725 static inline void __zs_cpu_down(struct mapping_area *area)
728 free_vm_area(area->vm);
732 static inline void *__zs_map_object(struct mapping_area *area,
733 struct page *pages[2], int off, int size)
735 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, &pages));
736 area->vm_addr = area->vm->addr;
737 return area->vm_addr + off;
740 static inline void __zs_unmap_object(struct mapping_area *area,
741 struct page *pages[2], int off, int size)
743 unsigned long addr = (unsigned long)area->vm_addr;
745 unmap_kernel_range(addr, PAGE_SIZE * 2);
748 #else /* USE_PGTABLE_MAPPING */
750 static inline int __zs_cpu_up(struct mapping_area *area)
753 * Make sure we don't leak memory if a cpu UP notification
754 * and zs_init() race and both call zs_cpu_up() on the same cpu
758 area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
764 static inline void __zs_cpu_down(struct mapping_area *area)
767 free_page((unsigned long)area->vm_buf);
771 static void *__zs_map_object(struct mapping_area *area,
772 struct page *pages[2], int off, int size)
776 char *buf = area->vm_buf;
778 /* disable page faults to match kmap_atomic() return conditions */
781 /* no read fastpath */
782 if (area->vm_mm == ZS_MM_WO)
785 sizes[0] = PAGE_SIZE - off;
786 sizes[1] = size - sizes[0];
788 /* copy object to per-cpu buffer */
789 addr = kmap_atomic(pages[0]);
790 memcpy(buf, addr + off, sizes[0]);
792 addr = kmap_atomic(pages[1]);
793 memcpy(buf + sizes[0], addr, sizes[1]);
799 static void __zs_unmap_object(struct mapping_area *area,
800 struct page *pages[2], int off, int size)
804 char *buf = area->vm_buf;
806 /* no write fastpath */
807 if (area->vm_mm == ZS_MM_RO)
810 sizes[0] = PAGE_SIZE - off;
811 sizes[1] = size - sizes[0];
813 /* copy per-cpu buffer to object */
814 addr = kmap_atomic(pages[0]);
815 memcpy(addr + off, buf, sizes[0]);
817 addr = kmap_atomic(pages[1]);
818 memcpy(addr, buf + sizes[0], sizes[1]);
822 /* enable page faults to match kunmap_atomic() return conditions */
826 #endif /* USE_PGTABLE_MAPPING */
828 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
831 int ret, cpu = (long)pcpu;
832 struct mapping_area *area;
836 area = &per_cpu(zs_map_area, cpu);
837 ret = __zs_cpu_up(area);
839 return notifier_from_errno(ret);
842 case CPU_UP_CANCELED:
843 area = &per_cpu(zs_map_area, cpu);
851 static struct notifier_block zs_cpu_nb = {
852 .notifier_call = zs_cpu_notifier
855 static void zs_exit(void)
860 zpool_unregister_driver(&zs_zpool_driver);
863 cpu_notifier_register_begin();
865 for_each_online_cpu(cpu)
866 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
867 __unregister_cpu_notifier(&zs_cpu_nb);
869 cpu_notifier_register_done();
872 static int zs_init(void)
876 cpu_notifier_register_begin();
878 __register_cpu_notifier(&zs_cpu_nb);
879 for_each_online_cpu(cpu) {
880 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
881 if (notifier_to_errno(ret)) {
882 cpu_notifier_register_done();
887 cpu_notifier_register_done();
890 zpool_register_driver(&zs_zpool_driver);
896 return notifier_to_errno(ret);
900 * zs_create_pool - Creates an allocation pool to work from.
901 * @flags: allocation flags used to allocate pool metadata
903 * This function must be called before anything when using
904 * the zsmalloc allocator.
906 * On success, a pointer to the newly created pool is returned,
909 struct zs_pool *zs_create_pool(gfp_t flags)
912 struct zs_pool *pool;
914 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
915 pool = kzalloc(ovhd_size, GFP_KERNEL);
919 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
921 struct size_class *class;
923 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
924 if (size > ZS_MAX_ALLOC_SIZE)
925 size = ZS_MAX_ALLOC_SIZE;
927 class = &pool->size_class[i];
930 spin_lock_init(&class->lock);
931 class->pages_per_zspage = get_pages_per_zspage(size);
939 EXPORT_SYMBOL_GPL(zs_create_pool);
941 void zs_destroy_pool(struct zs_pool *pool)
945 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
947 struct size_class *class = &pool->size_class[i];
949 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
950 if (class->fullness_list[fg]) {
951 pr_info("Freeing non-empty class with size "
952 "%db, fullness group %d\n",
959 EXPORT_SYMBOL_GPL(zs_destroy_pool);
962 * zs_malloc - Allocate block of given size from pool.
963 * @pool: pool to allocate from
964 * @size: size of block to allocate
966 * On success, handle to the allocated object is returned,
968 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
970 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
973 struct link_free *link;
975 struct size_class *class;
977 struct page *first_page, *m_page;
978 unsigned long m_objidx, m_offset;
980 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
983 class_idx = get_size_class_index(size);
984 class = &pool->size_class[class_idx];
985 BUG_ON(class_idx != class->index);
987 spin_lock(&class->lock);
988 first_page = find_get_zspage(class);
991 spin_unlock(&class->lock);
992 first_page = alloc_zspage(class, pool->flags);
993 if (unlikely(!first_page))
996 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
997 spin_lock(&class->lock);
998 class->pages_allocated += class->pages_per_zspage;
1001 obj = (unsigned long)first_page->freelist;
1002 obj_handle_to_location(obj, &m_page, &m_objidx);
1003 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1005 link = (struct link_free *)kmap_atomic(m_page) +
1006 m_offset / sizeof(*link);
1007 first_page->freelist = link->next;
1008 memset(link, POISON_INUSE, sizeof(*link));
1009 kunmap_atomic(link);
1011 first_page->inuse++;
1012 /* Now move the zspage to another fullness group, if required */
1013 fix_fullness_group(pool, first_page);
1014 spin_unlock(&class->lock);
1018 EXPORT_SYMBOL_GPL(zs_malloc);
1020 void zs_free(struct zs_pool *pool, unsigned long obj)
1022 struct link_free *link;
1023 struct page *first_page, *f_page;
1024 unsigned long f_objidx, f_offset;
1027 struct size_class *class;
1028 enum fullness_group fullness;
1033 obj_handle_to_location(obj, &f_page, &f_objidx);
1034 first_page = get_first_page(f_page);
1036 get_zspage_mapping(first_page, &class_idx, &fullness);
1037 class = &pool->size_class[class_idx];
1038 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1040 spin_lock(&class->lock);
1042 /* Insert this object in containing zspage's freelist */
1043 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
1045 link->next = first_page->freelist;
1046 kunmap_atomic(link);
1047 first_page->freelist = (void *)obj;
1049 first_page->inuse--;
1050 fullness = fix_fullness_group(pool, first_page);
1052 if (fullness == ZS_EMPTY)
1053 class->pages_allocated -= class->pages_per_zspage;
1055 spin_unlock(&class->lock);
1057 if (fullness == ZS_EMPTY)
1058 free_zspage(first_page);
1060 EXPORT_SYMBOL_GPL(zs_free);
1063 * zs_map_object - get address of allocated object from handle.
1064 * @pool: pool from which the object was allocated
1065 * @handle: handle returned from zs_malloc
1067 * Before using an object allocated from zs_malloc, it must be mapped using
1068 * this function. When done with the object, it must be unmapped using
1071 * Only one object can be mapped per cpu at a time. There is no protection
1072 * against nested mappings.
1074 * This function returns with preemption and page faults disabled.
1076 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1080 unsigned long obj_idx, off;
1082 unsigned int class_idx;
1083 enum fullness_group fg;
1084 struct size_class *class;
1085 struct mapping_area *area;
1086 struct page *pages[2];
1091 * Because we use per-cpu mapping areas shared among the
1092 * pools/users, we can't allow mapping in interrupt context
1093 * because it can corrupt another users mappings.
1095 BUG_ON(in_interrupt());
1097 obj_handle_to_location(handle, &page, &obj_idx);
1098 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1099 class = &pool->size_class[class_idx];
1100 off = obj_idx_to_offset(page, obj_idx, class->size);
1102 area = &get_cpu_var(zs_map_area);
1104 if (off + class->size <= PAGE_SIZE) {
1105 /* this object is contained entirely within a page */
1106 area->vm_addr = kmap_atomic(page);
1107 return area->vm_addr + off;
1110 /* this object spans two pages */
1112 pages[1] = get_next_page(page);
1115 return __zs_map_object(area, pages, off, class->size);
1117 EXPORT_SYMBOL_GPL(zs_map_object);
1119 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1122 unsigned long obj_idx, off;
1124 unsigned int class_idx;
1125 enum fullness_group fg;
1126 struct size_class *class;
1127 struct mapping_area *area;
1131 obj_handle_to_location(handle, &page, &obj_idx);
1132 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1133 class = &pool->size_class[class_idx];
1134 off = obj_idx_to_offset(page, obj_idx, class->size);
1136 area = &__get_cpu_var(zs_map_area);
1137 if (off + class->size <= PAGE_SIZE)
1138 kunmap_atomic(area->vm_addr);
1140 struct page *pages[2];
1143 pages[1] = get_next_page(page);
1146 __zs_unmap_object(area, pages, off, class->size);
1148 put_cpu_var(zs_map_area);
1150 EXPORT_SYMBOL_GPL(zs_unmap_object);
1152 u64 zs_get_total_size_bytes(struct zs_pool *pool)
1157 for (i = 0; i < ZS_SIZE_CLASSES; i++)
1158 npages += pool->size_class[i].pages_allocated;
1160 return npages << PAGE_SHIFT;
1162 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
1164 module_init(zs_init);
1165 module_exit(zs_exit);
1167 MODULE_LICENSE("Dual BSD/GPL");
1168 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");