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 MODULE_ALIAS("zpool-zsmalloc");
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;
355 static int get_size_class_index(int size)
359 if (likely(size > ZS_MIN_ALLOC_SIZE))
360 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
361 ZS_SIZE_CLASS_DELTA);
366 static enum fullness_group get_fullness_group(struct page *page)
368 int inuse, max_objects;
369 enum fullness_group fg;
370 BUG_ON(!is_first_page(page));
373 max_objects = page->objects;
377 else if (inuse == max_objects)
379 else if (inuse <= max_objects / fullness_threshold_frac)
380 fg = ZS_ALMOST_EMPTY;
387 static void insert_zspage(struct page *page, struct size_class *class,
388 enum fullness_group fullness)
392 BUG_ON(!is_first_page(page));
394 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
397 head = &class->fullness_list[fullness];
399 list_add_tail(&page->lru, &(*head)->lru);
404 static void remove_zspage(struct page *page, struct size_class *class,
405 enum fullness_group fullness)
409 BUG_ON(!is_first_page(page));
411 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
414 head = &class->fullness_list[fullness];
416 if (list_empty(&(*head)->lru))
418 else if (*head == page)
419 *head = (struct page *)list_entry((*head)->lru.next,
422 list_del_init(&page->lru);
425 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
429 struct size_class *class;
430 enum fullness_group currfg, newfg;
432 BUG_ON(!is_first_page(page));
434 get_zspage_mapping(page, &class_idx, &currfg);
435 newfg = get_fullness_group(page);
439 class = &pool->size_class[class_idx];
440 remove_zspage(page, class, currfg);
441 insert_zspage(page, class, newfg);
442 set_zspage_mapping(page, class_idx, newfg);
449 * We have to decide on how many pages to link together
450 * to form a zspage for each size class. This is important
451 * to reduce wastage due to unusable space left at end of
452 * each zspage which is given as:
453 * wastage = Zp - Zp % size_class
454 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
456 * For example, for size class of 3/8 * PAGE_SIZE, we should
457 * link together 3 PAGE_SIZE sized pages to form a zspage
458 * since then we can perfectly fit in 8 such objects.
460 static int get_pages_per_zspage(int class_size)
462 int i, max_usedpc = 0;
463 /* zspage order which gives maximum used size per KB */
464 int max_usedpc_order = 1;
466 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
470 zspage_size = i * PAGE_SIZE;
471 waste = zspage_size % class_size;
472 usedpc = (zspage_size - waste) * 100 / zspage_size;
474 if (usedpc > max_usedpc) {
476 max_usedpc_order = i;
480 return max_usedpc_order;
484 * A single 'zspage' is composed of many system pages which are
485 * linked together using fields in struct page. This function finds
486 * the first/head page, given any component page of a zspage.
488 static struct page *get_first_page(struct page *page)
490 if (is_first_page(page))
493 return page->first_page;
496 static struct page *get_next_page(struct page *page)
500 if (is_last_page(page))
502 else if (is_first_page(page))
503 next = (struct page *)page->private;
505 next = list_entry(page->lru.next, struct page, lru);
511 * Encode <page, obj_idx> as a single handle value.
512 * On hardware platforms with physical memory starting at 0x0 the pfn
513 * could be 0 so we ensure that the handle will never be 0 by adjusting the
514 * encoded obj_idx value before encoding.
516 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
518 unsigned long handle;
525 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
526 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
528 return (void *)handle;
532 * Decode <page, obj_idx> pair from the given object handle. We adjust the
533 * decoded obj_idx back to its original value since it was adjusted in
534 * obj_location_to_handle().
536 static void obj_handle_to_location(unsigned long handle, struct page **page,
537 unsigned long *obj_idx)
539 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
540 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
543 static unsigned long obj_idx_to_offset(struct page *page,
544 unsigned long obj_idx, int class_size)
546 unsigned long off = 0;
548 if (!is_first_page(page))
551 return off + obj_idx * class_size;
554 static void reset_page(struct page *page)
556 clear_bit(PG_private, &page->flags);
557 clear_bit(PG_private_2, &page->flags);
558 set_page_private(page, 0);
559 page->mapping = NULL;
560 page->freelist = NULL;
561 page_mapcount_reset(page);
564 static void free_zspage(struct page *first_page)
566 struct page *nextp, *tmp, *head_extra;
568 BUG_ON(!is_first_page(first_page));
569 BUG_ON(first_page->inuse);
571 head_extra = (struct page *)page_private(first_page);
573 reset_page(first_page);
574 __free_page(first_page);
576 /* zspage with only 1 system page */
580 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
581 list_del(&nextp->lru);
585 reset_page(head_extra);
586 __free_page(head_extra);
589 /* Initialize a newly allocated zspage */
590 static void init_zspage(struct page *first_page, struct size_class *class)
592 unsigned long off = 0;
593 struct page *page = first_page;
595 BUG_ON(!is_first_page(first_page));
597 struct page *next_page;
598 struct link_free *link;
599 unsigned int i, objs_on_page;
602 * page->index stores offset of first object starting
603 * in the page. For the first page, this is always 0,
604 * so we use first_page->index (aka ->freelist) to store
605 * head of corresponding zspage's freelist.
607 if (page != first_page)
610 link = (struct link_free *)kmap_atomic(page) +
612 objs_on_page = (PAGE_SIZE - off) / class->size;
614 for (i = 1; i <= objs_on_page; i++) {
616 if (off < PAGE_SIZE) {
617 link->next = obj_location_to_handle(page, i);
618 link += class->size / sizeof(*link);
623 * We now come to the last (full or partial) object on this
624 * page, which must point to the first object on the next
627 next_page = get_next_page(page);
628 link->next = obj_location_to_handle(next_page, 0);
631 off = (off + class->size) % PAGE_SIZE;
636 * Allocate a zspage for the given size class
638 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
641 struct page *first_page = NULL, *uninitialized_var(prev_page);
644 * Allocate individual pages and link them together as:
645 * 1. first page->private = first sub-page
646 * 2. all sub-pages are linked together using page->lru
647 * 3. each sub-page is linked to the first page using page->first_page
649 * For each size class, First/Head pages are linked together using
650 * page->lru. Also, we set PG_private to identify the first page
651 * (i.e. no other sub-page has this flag set) and PG_private_2 to
652 * identify the last page.
655 for (i = 0; i < class->pages_per_zspage; i++) {
658 page = alloc_page(flags);
662 INIT_LIST_HEAD(&page->lru);
663 if (i == 0) { /* first page */
664 SetPagePrivate(page);
665 set_page_private(page, 0);
667 first_page->inuse = 0;
670 first_page->private = (unsigned long)page;
672 page->first_page = first_page;
674 list_add(&page->lru, &prev_page->lru);
675 if (i == class->pages_per_zspage - 1) /* last page */
676 SetPagePrivate2(page);
680 init_zspage(first_page, class);
682 first_page->freelist = obj_location_to_handle(first_page, 0);
683 /* Maximum number of objects we can store in this zspage */
684 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
686 error = 0; /* Success */
689 if (unlikely(error) && first_page) {
690 free_zspage(first_page);
697 static struct page *find_get_zspage(struct size_class *class)
702 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
703 page = class->fullness_list[i];
711 #ifdef USE_PGTABLE_MAPPING
712 static inline int __zs_cpu_up(struct mapping_area *area)
715 * Make sure we don't leak memory if a cpu UP notification
716 * and zs_init() race and both call zs_cpu_up() on the same cpu
720 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
726 static inline void __zs_cpu_down(struct mapping_area *area)
729 free_vm_area(area->vm);
733 static inline void *__zs_map_object(struct mapping_area *area,
734 struct page *pages[2], int off, int size)
736 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, &pages));
737 area->vm_addr = area->vm->addr;
738 return area->vm_addr + off;
741 static inline void __zs_unmap_object(struct mapping_area *area,
742 struct page *pages[2], int off, int size)
744 unsigned long addr = (unsigned long)area->vm_addr;
746 unmap_kernel_range(addr, PAGE_SIZE * 2);
749 #else /* USE_PGTABLE_MAPPING */
751 static inline int __zs_cpu_up(struct mapping_area *area)
754 * Make sure we don't leak memory if a cpu UP notification
755 * and zs_init() race and both call zs_cpu_up() on the same cpu
759 area->vm_buf = (char *)__get_free_page(GFP_KERNEL);
765 static inline void __zs_cpu_down(struct mapping_area *area)
768 free_page((unsigned long)area->vm_buf);
772 static void *__zs_map_object(struct mapping_area *area,
773 struct page *pages[2], int off, int size)
777 char *buf = area->vm_buf;
779 /* disable page faults to match kmap_atomic() return conditions */
782 /* no read fastpath */
783 if (area->vm_mm == ZS_MM_WO)
786 sizes[0] = PAGE_SIZE - off;
787 sizes[1] = size - sizes[0];
789 /* copy object to per-cpu buffer */
790 addr = kmap_atomic(pages[0]);
791 memcpy(buf, addr + off, sizes[0]);
793 addr = kmap_atomic(pages[1]);
794 memcpy(buf + sizes[0], addr, sizes[1]);
800 static void __zs_unmap_object(struct mapping_area *area,
801 struct page *pages[2], int off, int size)
805 char *buf = area->vm_buf;
807 /* no write fastpath */
808 if (area->vm_mm == ZS_MM_RO)
811 sizes[0] = PAGE_SIZE - off;
812 sizes[1] = size - sizes[0];
814 /* copy per-cpu buffer to object */
815 addr = kmap_atomic(pages[0]);
816 memcpy(addr + off, buf, sizes[0]);
818 addr = kmap_atomic(pages[1]);
819 memcpy(addr, buf + sizes[0], sizes[1]);
823 /* enable page faults to match kunmap_atomic() return conditions */
827 #endif /* USE_PGTABLE_MAPPING */
829 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
832 int ret, cpu = (long)pcpu;
833 struct mapping_area *area;
837 area = &per_cpu(zs_map_area, cpu);
838 ret = __zs_cpu_up(area);
840 return notifier_from_errno(ret);
843 case CPU_UP_CANCELED:
844 area = &per_cpu(zs_map_area, cpu);
852 static struct notifier_block zs_cpu_nb = {
853 .notifier_call = zs_cpu_notifier
856 static void zs_exit(void)
861 zpool_unregister_driver(&zs_zpool_driver);
864 cpu_notifier_register_begin();
866 for_each_online_cpu(cpu)
867 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
868 __unregister_cpu_notifier(&zs_cpu_nb);
870 cpu_notifier_register_done();
873 static int zs_init(void)
877 cpu_notifier_register_begin();
879 __register_cpu_notifier(&zs_cpu_nb);
880 for_each_online_cpu(cpu) {
881 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
882 if (notifier_to_errno(ret)) {
883 cpu_notifier_register_done();
888 cpu_notifier_register_done();
891 zpool_register_driver(&zs_zpool_driver);
897 return notifier_to_errno(ret);
901 * zs_create_pool - Creates an allocation pool to work from.
902 * @flags: allocation flags used to allocate pool metadata
904 * This function must be called before anything when using
905 * the zsmalloc allocator.
907 * On success, a pointer to the newly created pool is returned,
910 struct zs_pool *zs_create_pool(gfp_t flags)
913 struct zs_pool *pool;
915 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
916 pool = kzalloc(ovhd_size, GFP_KERNEL);
920 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
922 struct size_class *class;
924 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
925 if (size > ZS_MAX_ALLOC_SIZE)
926 size = ZS_MAX_ALLOC_SIZE;
928 class = &pool->size_class[i];
931 spin_lock_init(&class->lock);
932 class->pages_per_zspage = get_pages_per_zspage(size);
940 EXPORT_SYMBOL_GPL(zs_create_pool);
942 void zs_destroy_pool(struct zs_pool *pool)
946 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
948 struct size_class *class = &pool->size_class[i];
950 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
951 if (class->fullness_list[fg]) {
952 pr_info("Freeing non-empty class with size "
953 "%db, fullness group %d\n",
960 EXPORT_SYMBOL_GPL(zs_destroy_pool);
963 * zs_malloc - Allocate block of given size from pool.
964 * @pool: pool to allocate from
965 * @size: size of block to allocate
967 * On success, handle to the allocated object is returned,
969 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
971 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
974 struct link_free *link;
976 struct size_class *class;
978 struct page *first_page, *m_page;
979 unsigned long m_objidx, m_offset;
981 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
984 class_idx = get_size_class_index(size);
985 class = &pool->size_class[class_idx];
986 BUG_ON(class_idx != class->index);
988 spin_lock(&class->lock);
989 first_page = find_get_zspage(class);
992 spin_unlock(&class->lock);
993 first_page = alloc_zspage(class, pool->flags);
994 if (unlikely(!first_page))
997 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
998 spin_lock(&class->lock);
999 class->pages_allocated += class->pages_per_zspage;
1002 obj = (unsigned long)first_page->freelist;
1003 obj_handle_to_location(obj, &m_page, &m_objidx);
1004 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1006 link = (struct link_free *)kmap_atomic(m_page) +
1007 m_offset / sizeof(*link);
1008 first_page->freelist = link->next;
1009 memset(link, POISON_INUSE, sizeof(*link));
1010 kunmap_atomic(link);
1012 first_page->inuse++;
1013 /* Now move the zspage to another fullness group, if required */
1014 fix_fullness_group(pool, first_page);
1015 spin_unlock(&class->lock);
1019 EXPORT_SYMBOL_GPL(zs_malloc);
1021 void zs_free(struct zs_pool *pool, unsigned long obj)
1023 struct link_free *link;
1024 struct page *first_page, *f_page;
1025 unsigned long f_objidx, f_offset;
1028 struct size_class *class;
1029 enum fullness_group fullness;
1034 obj_handle_to_location(obj, &f_page, &f_objidx);
1035 first_page = get_first_page(f_page);
1037 get_zspage_mapping(first_page, &class_idx, &fullness);
1038 class = &pool->size_class[class_idx];
1039 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1041 spin_lock(&class->lock);
1043 /* Insert this object in containing zspage's freelist */
1044 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
1046 link->next = first_page->freelist;
1047 kunmap_atomic(link);
1048 first_page->freelist = (void *)obj;
1050 first_page->inuse--;
1051 fullness = fix_fullness_group(pool, first_page);
1053 if (fullness == ZS_EMPTY)
1054 class->pages_allocated -= class->pages_per_zspage;
1056 spin_unlock(&class->lock);
1058 if (fullness == ZS_EMPTY)
1059 free_zspage(first_page);
1061 EXPORT_SYMBOL_GPL(zs_free);
1064 * zs_map_object - get address of allocated object from handle.
1065 * @pool: pool from which the object was allocated
1066 * @handle: handle returned from zs_malloc
1068 * Before using an object allocated from zs_malloc, it must be mapped using
1069 * this function. When done with the object, it must be unmapped using
1072 * Only one object can be mapped per cpu at a time. There is no protection
1073 * against nested mappings.
1075 * This function returns with preemption and page faults disabled.
1077 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1081 unsigned long obj_idx, off;
1083 unsigned int class_idx;
1084 enum fullness_group fg;
1085 struct size_class *class;
1086 struct mapping_area *area;
1087 struct page *pages[2];
1092 * Because we use per-cpu mapping areas shared among the
1093 * pools/users, we can't allow mapping in interrupt context
1094 * because it can corrupt another users mappings.
1096 BUG_ON(in_interrupt());
1098 obj_handle_to_location(handle, &page, &obj_idx);
1099 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1100 class = &pool->size_class[class_idx];
1101 off = obj_idx_to_offset(page, obj_idx, class->size);
1103 area = &get_cpu_var(zs_map_area);
1105 if (off + class->size <= PAGE_SIZE) {
1106 /* this object is contained entirely within a page */
1107 area->vm_addr = kmap_atomic(page);
1108 return area->vm_addr + off;
1111 /* this object spans two pages */
1113 pages[1] = get_next_page(page);
1116 return __zs_map_object(area, pages, off, class->size);
1118 EXPORT_SYMBOL_GPL(zs_map_object);
1120 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1123 unsigned long obj_idx, off;
1125 unsigned int class_idx;
1126 enum fullness_group fg;
1127 struct size_class *class;
1128 struct mapping_area *area;
1132 obj_handle_to_location(handle, &page, &obj_idx);
1133 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1134 class = &pool->size_class[class_idx];
1135 off = obj_idx_to_offset(page, obj_idx, class->size);
1137 area = &__get_cpu_var(zs_map_area);
1138 if (off + class->size <= PAGE_SIZE)
1139 kunmap_atomic(area->vm_addr);
1141 struct page *pages[2];
1144 pages[1] = get_next_page(page);
1147 __zs_unmap_object(area, pages, off, class->size);
1149 put_cpu_var(zs_map_area);
1151 EXPORT_SYMBOL_GPL(zs_unmap_object);
1153 u64 zs_get_total_size_bytes(struct zs_pool *pool)
1158 for (i = 0; i < ZS_SIZE_CLASSES; i++)
1159 npages += pool->size_class[i].pages_allocated;
1161 return npages << PAGE_SHIFT;
1163 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);
1165 module_init(zs_init);
1166 module_exit(zs_exit);
1168 MODULE_LICENSE("Dual BSD/GPL");
1169 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");