2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->first_page: points to the first component (0-order) page
20 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
27 * For _first_ page only:
29 * page->private (union with page->first_page): refers to the
30 * component page after the first page
31 * If the page is first_page for huge object, it stores handle.
32 * Look at size_class->huge.
33 * page->freelist: points to the first free object in zspage.
34 * Free objects are linked together using in-place
36 * page->objects: maximum number of objects we can store in this
37 * zspage (class->zspage_order * PAGE_SIZE / class->size)
38 * page->lru: links together first pages of various zspages.
39 * Basically forming list of zspages in a fullness group.
40 * page->mapping: class index and fullness group of the zspage
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
48 #include <linux/module.h>
49 #include <linux/kernel.h>
50 #include <linux/sched.h>
51 #include <linux/bitops.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/string.h>
55 #include <linux/slab.h>
56 #include <asm/tlbflush.h>
57 #include <asm/pgtable.h>
58 #include <linux/cpumask.h>
59 #include <linux/cpu.h>
60 #include <linux/vmalloc.h>
61 #include <linux/hardirq.h>
62 #include <linux/spinlock.h>
63 #include <linux/types.h>
64 #include <linux/debugfs.h>
65 #include <linux/zsmalloc.h>
66 #include <linux/zpool.h>
69 * This must be power of 2 and greater than of equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
80 #define ZS_MAX_ZSPAGE_ORDER 2
81 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
83 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
86 * Object location (<PFN>, <obj_idx>) is encoded as
87 * as single (unsigned long) handle value.
89 * Note that object index <obj_idx> is relative to system
90 * page <PFN> it is stored in, so for each sub-page belonging
91 * to a zspage, obj_idx starts with 0.
93 * This is made more complicated by various memory models and PAE.
96 #ifndef MAX_PHYSMEM_BITS
97 #ifdef CONFIG_HIGHMEM64G
98 #define MAX_PHYSMEM_BITS 36
99 #else /* !CONFIG_HIGHMEM64G */
101 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
104 #define MAX_PHYSMEM_BITS BITS_PER_LONG
107 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
110 * Memory for allocating for handle keeps object position by
111 * encoding <page, obj_idx> and the encoded value has a room
112 * in least bit(ie, look at obj_to_location).
113 * We use the bit to synchronize between object access by
114 * user and migration.
116 #define HANDLE_PIN_BIT 0
119 * Head in allocated object should have OBJ_ALLOCATED_TAG
120 * to identify the object was allocated or not.
121 * It's okay to add the status bit in the least bit because
122 * header keeps handle which is 4byte-aligned address so we
123 * have room for two bit at least.
125 #define OBJ_ALLOCATED_TAG 1
126 #define OBJ_TAG_BITS 1
127 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
128 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
130 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
131 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132 #define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
134 /* each chunk includes extra space to keep handle */
135 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
138 * On systems with 4K page size, this gives 255 size classes! There is a
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
150 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
153 * We do not maintain any list for completely empty or full pages
155 enum fullness_group {
158 _ZS_NR_FULLNESS_GROUPS,
172 struct zs_size_stat {
173 unsigned long objs[NR_ZS_STAT_TYPE];
176 #ifdef CONFIG_ZSMALLOC_STAT
177 static struct dentry *zs_stat_root;
181 * number of size_classes
183 static int zs_size_classes;
186 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
188 * n = number of allocated objects
189 * N = total number of objects zspage can store
190 * f = fullness_threshold_frac
192 * Similarly, we assign zspage to:
193 * ZS_ALMOST_FULL when n > N / f
194 * ZS_EMPTY when n == 0
195 * ZS_FULL when n == N
197 * (see: fix_fullness_group())
199 static const int fullness_threshold_frac = 4;
203 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
205 * Size of objects stored in this class. Must be multiple
211 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
212 int pages_per_zspage;
213 struct zs_size_stat stats;
215 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
220 * Placed within free objects to form a singly linked list.
221 * For every zspage, first_page->freelist gives head of this list.
223 * This must be power of 2 and less than or equal to ZS_ALIGN
228 * Position of next free chunk (encodes <PFN, obj_idx>)
229 * It's valid for non-allocated object
233 * Handle of allocated object.
235 unsigned long handle;
242 struct size_class **size_class;
243 struct kmem_cache *handle_cachep;
245 gfp_t flags; /* allocation flags used when growing pool */
246 atomic_long_t pages_allocated;
248 struct zs_pool_stats stats;
250 /* Compact classes */
251 struct shrinker shrinker;
253 * To signify that register_shrinker() was successful
254 * and unregister_shrinker() will not Oops.
256 bool shrinker_enabled;
257 #ifdef CONFIG_ZSMALLOC_STAT
258 struct dentry *stat_dentry;
263 * A zspage's class index and fullness group
264 * are encoded in its (first)page->mapping
266 #define CLASS_IDX_BITS 28
267 #define FULLNESS_BITS 4
268 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
269 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
271 struct mapping_area {
272 #ifdef CONFIG_PGTABLE_MAPPING
273 struct vm_struct *vm; /* vm area for mapping object that span pages */
275 char *vm_buf; /* copy buffer for objects that span pages */
277 char *vm_addr; /* address of kmap_atomic()'ed pages */
278 enum zs_mapmode vm_mm; /* mapping mode */
282 static int create_handle_cache(struct zs_pool *pool)
284 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
286 return pool->handle_cachep ? 0 : 1;
289 static void destroy_handle_cache(struct zs_pool *pool)
291 if (pool->handle_cachep)
292 kmem_cache_destroy(pool->handle_cachep);
295 static unsigned long alloc_handle(struct zs_pool *pool)
297 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
298 pool->flags & ~__GFP_HIGHMEM);
301 static void free_handle(struct zs_pool *pool, unsigned long handle)
303 kmem_cache_free(pool->handle_cachep, (void *)handle);
306 static void record_obj(unsigned long handle, unsigned long obj)
308 *(unsigned long *)handle = obj;
315 static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops,
318 return zs_create_pool(name, gfp);
321 static void zs_zpool_destroy(void *pool)
323 zs_destroy_pool(pool);
326 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
327 unsigned long *handle)
329 *handle = zs_malloc(pool, size);
330 return *handle ? 0 : -1;
332 static void zs_zpool_free(void *pool, unsigned long handle)
334 zs_free(pool, handle);
337 static int zs_zpool_shrink(void *pool, unsigned int pages,
338 unsigned int *reclaimed)
343 static void *zs_zpool_map(void *pool, unsigned long handle,
344 enum zpool_mapmode mm)
346 enum zs_mapmode zs_mm;
355 case ZPOOL_MM_RW: /* fallthru */
361 return zs_map_object(pool, handle, zs_mm);
363 static void zs_zpool_unmap(void *pool, unsigned long handle)
365 zs_unmap_object(pool, handle);
368 static u64 zs_zpool_total_size(void *pool)
370 return zs_get_total_pages(pool) << PAGE_SHIFT;
373 static struct zpool_driver zs_zpool_driver = {
375 .owner = THIS_MODULE,
376 .create = zs_zpool_create,
377 .destroy = zs_zpool_destroy,
378 .malloc = zs_zpool_malloc,
379 .free = zs_zpool_free,
380 .shrink = zs_zpool_shrink,
382 .unmap = zs_zpool_unmap,
383 .total_size = zs_zpool_total_size,
386 MODULE_ALIAS("zpool-zsmalloc");
387 #endif /* CONFIG_ZPOOL */
389 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
391 return pages_per_zspage * PAGE_SIZE / size;
394 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
395 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
397 static int is_first_page(struct page *page)
399 return PagePrivate(page);
402 static int is_last_page(struct page *page)
404 return PagePrivate2(page);
407 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
408 enum fullness_group *fullness)
411 BUG_ON(!is_first_page(page));
413 m = (unsigned long)page->mapping;
414 *fullness = m & FULLNESS_MASK;
415 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
418 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
419 enum fullness_group fullness)
422 BUG_ON(!is_first_page(page));
424 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
425 (fullness & FULLNESS_MASK);
426 page->mapping = (struct address_space *)m;
430 * zsmalloc divides the pool into various size classes where each
431 * class maintains a list of zspages where each zspage is divided
432 * into equal sized chunks. Each allocation falls into one of these
433 * classes depending on its size. This function returns index of the
434 * size class which has chunk size big enough to hold the give size.
436 static int get_size_class_index(int size)
440 if (likely(size > ZS_MIN_ALLOC_SIZE))
441 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
442 ZS_SIZE_CLASS_DELTA);
444 return min(zs_size_classes - 1, idx);
447 static inline void zs_stat_inc(struct size_class *class,
448 enum zs_stat_type type, unsigned long cnt)
450 class->stats.objs[type] += cnt;
453 static inline void zs_stat_dec(struct size_class *class,
454 enum zs_stat_type type, unsigned long cnt)
456 class->stats.objs[type] -= cnt;
459 static inline unsigned long zs_stat_get(struct size_class *class,
460 enum zs_stat_type type)
462 return class->stats.objs[type];
465 #ifdef CONFIG_ZSMALLOC_STAT
467 static int __init zs_stat_init(void)
469 if (!debugfs_initialized())
472 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
479 static void __exit zs_stat_exit(void)
481 debugfs_remove_recursive(zs_stat_root);
484 static int zs_stats_size_show(struct seq_file *s, void *v)
487 struct zs_pool *pool = s->private;
488 struct size_class *class;
490 unsigned long class_almost_full, class_almost_empty;
491 unsigned long obj_allocated, obj_used, pages_used;
492 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
493 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
495 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
496 "class", "size", "almost_full", "almost_empty",
497 "obj_allocated", "obj_used", "pages_used",
500 for (i = 0; i < zs_size_classes; i++) {
501 class = pool->size_class[i];
503 if (class->index != i)
506 spin_lock(&class->lock);
507 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
508 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
509 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
510 obj_used = zs_stat_get(class, OBJ_USED);
511 spin_unlock(&class->lock);
513 objs_per_zspage = get_maxobj_per_zspage(class->size,
514 class->pages_per_zspage);
515 pages_used = obj_allocated / objs_per_zspage *
516 class->pages_per_zspage;
518 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
519 i, class->size, class_almost_full, class_almost_empty,
520 obj_allocated, obj_used, pages_used,
521 class->pages_per_zspage);
523 total_class_almost_full += class_almost_full;
524 total_class_almost_empty += class_almost_empty;
525 total_objs += obj_allocated;
526 total_used_objs += obj_used;
527 total_pages += pages_used;
531 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
532 "Total", "", total_class_almost_full,
533 total_class_almost_empty, total_objs,
534 total_used_objs, total_pages);
539 static int zs_stats_size_open(struct inode *inode, struct file *file)
541 return single_open(file, zs_stats_size_show, inode->i_private);
544 static const struct file_operations zs_stat_size_ops = {
545 .open = zs_stats_size_open,
548 .release = single_release,
551 static int zs_pool_stat_create(char *name, struct zs_pool *pool)
553 struct dentry *entry;
558 entry = debugfs_create_dir(name, zs_stat_root);
560 pr_warn("debugfs dir <%s> creation failed\n", name);
563 pool->stat_dentry = entry;
565 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
566 pool->stat_dentry, pool, &zs_stat_size_ops);
568 pr_warn("%s: debugfs file entry <%s> creation failed\n",
576 static void zs_pool_stat_destroy(struct zs_pool *pool)
578 debugfs_remove_recursive(pool->stat_dentry);
581 #else /* CONFIG_ZSMALLOC_STAT */
582 static int __init zs_stat_init(void)
587 static void __exit zs_stat_exit(void)
591 static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
596 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
603 * For each size class, zspages are divided into different groups
604 * depending on how "full" they are. This was done so that we could
605 * easily find empty or nearly empty zspages when we try to shrink
606 * the pool (not yet implemented). This function returns fullness
607 * status of the given page.
609 static enum fullness_group get_fullness_group(struct page *page)
611 int inuse, max_objects;
612 enum fullness_group fg;
613 BUG_ON(!is_first_page(page));
616 max_objects = page->objects;
620 else if (inuse == max_objects)
622 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
623 fg = ZS_ALMOST_EMPTY;
631 * Each size class maintains various freelists and zspages are assigned
632 * to one of these freelists based on the number of live objects they
633 * have. This functions inserts the given zspage into the freelist
634 * identified by <class, fullness_group>.
636 static void insert_zspage(struct page *page, struct size_class *class,
637 enum fullness_group fullness)
641 BUG_ON(!is_first_page(page));
643 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
646 head = &class->fullness_list[fullness];
648 list_add_tail(&page->lru, &(*head)->lru);
651 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
652 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
656 * This function removes the given zspage from the freelist identified
657 * by <class, fullness_group>.
659 static void remove_zspage(struct page *page, struct size_class *class,
660 enum fullness_group fullness)
664 BUG_ON(!is_first_page(page));
666 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
669 head = &class->fullness_list[fullness];
671 if (list_empty(&(*head)->lru))
673 else if (*head == page)
674 *head = (struct page *)list_entry((*head)->lru.next,
677 list_del_init(&page->lru);
678 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
679 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
683 * Each size class maintains zspages in different fullness groups depending
684 * on the number of live objects they contain. When allocating or freeing
685 * objects, the fullness status of the page can change, say, from ALMOST_FULL
686 * to ALMOST_EMPTY when freeing an object. This function checks if such
687 * a status change has occurred for the given page and accordingly moves the
688 * page from the freelist of the old fullness group to that of the new
691 static enum fullness_group fix_fullness_group(struct size_class *class,
695 enum fullness_group currfg, newfg;
697 BUG_ON(!is_first_page(page));
699 get_zspage_mapping(page, &class_idx, &currfg);
700 newfg = get_fullness_group(page);
704 remove_zspage(page, class, currfg);
705 insert_zspage(page, class, newfg);
706 set_zspage_mapping(page, class_idx, newfg);
713 * We have to decide on how many pages to link together
714 * to form a zspage for each size class. This is important
715 * to reduce wastage due to unusable space left at end of
716 * each zspage which is given as:
717 * wastage = Zp % class_size
718 * usage = Zp - wastage
719 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
721 * For example, for size class of 3/8 * PAGE_SIZE, we should
722 * link together 3 PAGE_SIZE sized pages to form a zspage
723 * since then we can perfectly fit in 8 such objects.
725 static int get_pages_per_zspage(int class_size)
727 int i, max_usedpc = 0;
728 /* zspage order which gives maximum used size per KB */
729 int max_usedpc_order = 1;
731 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
735 zspage_size = i * PAGE_SIZE;
736 waste = zspage_size % class_size;
737 usedpc = (zspage_size - waste) * 100 / zspage_size;
739 if (usedpc > max_usedpc) {
741 max_usedpc_order = i;
745 return max_usedpc_order;
749 * A single 'zspage' is composed of many system pages which are
750 * linked together using fields in struct page. This function finds
751 * the first/head page, given any component page of a zspage.
753 static struct page *get_first_page(struct page *page)
755 if (is_first_page(page))
758 return page->first_page;
761 static struct page *get_next_page(struct page *page)
765 if (is_last_page(page))
767 else if (is_first_page(page))
768 next = (struct page *)page_private(page);
770 next = list_entry(page->lru.next, struct page, lru);
776 * Encode <page, obj_idx> as a single handle value.
777 * We use the least bit of handle for tagging.
779 static void *location_to_obj(struct page *page, unsigned long obj_idx)
788 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
789 obj |= ((obj_idx) & OBJ_INDEX_MASK);
790 obj <<= OBJ_TAG_BITS;
796 * Decode <page, obj_idx> pair from the given object handle. We adjust the
797 * decoded obj_idx back to its original value since it was adjusted in
800 static void obj_to_location(unsigned long obj, struct page **page,
801 unsigned long *obj_idx)
803 obj >>= OBJ_TAG_BITS;
804 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
805 *obj_idx = (obj & OBJ_INDEX_MASK);
808 static unsigned long handle_to_obj(unsigned long handle)
810 return *(unsigned long *)handle;
813 static unsigned long obj_to_head(struct size_class *class, struct page *page,
817 VM_BUG_ON(!is_first_page(page));
818 return *(unsigned long *)page_private(page);
820 return *(unsigned long *)obj;
823 static unsigned long obj_idx_to_offset(struct page *page,
824 unsigned long obj_idx, int class_size)
826 unsigned long off = 0;
828 if (!is_first_page(page))
831 return off + obj_idx * class_size;
834 static inline int trypin_tag(unsigned long handle)
836 unsigned long *ptr = (unsigned long *)handle;
838 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
841 static void pin_tag(unsigned long handle)
843 while (!trypin_tag(handle));
846 static void unpin_tag(unsigned long handle)
848 unsigned long *ptr = (unsigned long *)handle;
850 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
853 static void reset_page(struct page *page)
855 clear_bit(PG_private, &page->flags);
856 clear_bit(PG_private_2, &page->flags);
857 set_page_private(page, 0);
858 page->mapping = NULL;
859 page->freelist = NULL;
860 page_mapcount_reset(page);
863 static void free_zspage(struct page *first_page)
865 struct page *nextp, *tmp, *head_extra;
867 BUG_ON(!is_first_page(first_page));
868 BUG_ON(first_page->inuse);
870 head_extra = (struct page *)page_private(first_page);
872 reset_page(first_page);
873 __free_page(first_page);
875 /* zspage with only 1 system page */
879 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
880 list_del(&nextp->lru);
884 reset_page(head_extra);
885 __free_page(head_extra);
888 /* Initialize a newly allocated zspage */
889 static void init_zspage(struct page *first_page, struct size_class *class)
891 unsigned long off = 0;
892 struct page *page = first_page;
894 BUG_ON(!is_first_page(first_page));
896 struct page *next_page;
897 struct link_free *link;
902 * page->index stores offset of first object starting
903 * in the page. For the first page, this is always 0,
904 * so we use first_page->index (aka ->freelist) to store
905 * head of corresponding zspage's freelist.
907 if (page != first_page)
910 vaddr = kmap_atomic(page);
911 link = (struct link_free *)vaddr + off / sizeof(*link);
913 while ((off += class->size) < PAGE_SIZE) {
914 link->next = location_to_obj(page, i++);
915 link += class->size / sizeof(*link);
919 * We now come to the last (full or partial) object on this
920 * page, which must point to the first object on the next
923 next_page = get_next_page(page);
924 link->next = location_to_obj(next_page, 0);
925 kunmap_atomic(vaddr);
932 * Allocate a zspage for the given size class
934 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
937 struct page *first_page = NULL, *uninitialized_var(prev_page);
940 * Allocate individual pages and link them together as:
941 * 1. first page->private = first sub-page
942 * 2. all sub-pages are linked together using page->lru
943 * 3. each sub-page is linked to the first page using page->first_page
945 * For each size class, First/Head pages are linked together using
946 * page->lru. Also, we set PG_private to identify the first page
947 * (i.e. no other sub-page has this flag set) and PG_private_2 to
948 * identify the last page.
951 for (i = 0; i < class->pages_per_zspage; i++) {
954 page = alloc_page(flags);
958 INIT_LIST_HEAD(&page->lru);
959 if (i == 0) { /* first page */
960 SetPagePrivate(page);
961 set_page_private(page, 0);
963 first_page->inuse = 0;
966 set_page_private(first_page, (unsigned long)page);
968 page->first_page = first_page;
970 list_add(&page->lru, &prev_page->lru);
971 if (i == class->pages_per_zspage - 1) /* last page */
972 SetPagePrivate2(page);
976 init_zspage(first_page, class);
978 first_page->freelist = location_to_obj(first_page, 0);
979 /* Maximum number of objects we can store in this zspage */
980 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
982 error = 0; /* Success */
985 if (unlikely(error) && first_page) {
986 free_zspage(first_page);
993 static struct page *find_get_zspage(struct size_class *class)
998 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
999 page = class->fullness_list[i];
1007 #ifdef CONFIG_PGTABLE_MAPPING
1008 static inline int __zs_cpu_up(struct mapping_area *area)
1011 * Make sure we don't leak memory if a cpu UP notification
1012 * and zs_init() race and both call zs_cpu_up() on the same cpu
1016 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1022 static inline void __zs_cpu_down(struct mapping_area *area)
1025 free_vm_area(area->vm);
1029 static inline void *__zs_map_object(struct mapping_area *area,
1030 struct page *pages[2], int off, int size)
1032 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1033 area->vm_addr = area->vm->addr;
1034 return area->vm_addr + off;
1037 static inline void __zs_unmap_object(struct mapping_area *area,
1038 struct page *pages[2], int off, int size)
1040 unsigned long addr = (unsigned long)area->vm_addr;
1042 unmap_kernel_range(addr, PAGE_SIZE * 2);
1045 #else /* CONFIG_PGTABLE_MAPPING */
1047 static inline int __zs_cpu_up(struct mapping_area *area)
1050 * Make sure we don't leak memory if a cpu UP notification
1051 * and zs_init() race and both call zs_cpu_up() on the same cpu
1055 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1061 static inline void __zs_cpu_down(struct mapping_area *area)
1063 kfree(area->vm_buf);
1064 area->vm_buf = NULL;
1067 static void *__zs_map_object(struct mapping_area *area,
1068 struct page *pages[2], int off, int size)
1072 char *buf = area->vm_buf;
1074 /* disable page faults to match kmap_atomic() return conditions */
1075 pagefault_disable();
1077 /* no read fastpath */
1078 if (area->vm_mm == ZS_MM_WO)
1081 sizes[0] = PAGE_SIZE - off;
1082 sizes[1] = size - sizes[0];
1084 /* copy object to per-cpu buffer */
1085 addr = kmap_atomic(pages[0]);
1086 memcpy(buf, addr + off, sizes[0]);
1087 kunmap_atomic(addr);
1088 addr = kmap_atomic(pages[1]);
1089 memcpy(buf + sizes[0], addr, sizes[1]);
1090 kunmap_atomic(addr);
1092 return area->vm_buf;
1095 static void __zs_unmap_object(struct mapping_area *area,
1096 struct page *pages[2], int off, int size)
1102 /* no write fastpath */
1103 if (area->vm_mm == ZS_MM_RO)
1108 buf = buf + ZS_HANDLE_SIZE;
1109 size -= ZS_HANDLE_SIZE;
1110 off += ZS_HANDLE_SIZE;
1113 sizes[0] = PAGE_SIZE - off;
1114 sizes[1] = size - sizes[0];
1116 /* copy per-cpu buffer to object */
1117 addr = kmap_atomic(pages[0]);
1118 memcpy(addr + off, buf, sizes[0]);
1119 kunmap_atomic(addr);
1120 addr = kmap_atomic(pages[1]);
1121 memcpy(addr, buf + sizes[0], sizes[1]);
1122 kunmap_atomic(addr);
1125 /* enable page faults to match kunmap_atomic() return conditions */
1129 #endif /* CONFIG_PGTABLE_MAPPING */
1131 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1134 int ret, cpu = (long)pcpu;
1135 struct mapping_area *area;
1138 case CPU_UP_PREPARE:
1139 area = &per_cpu(zs_map_area, cpu);
1140 ret = __zs_cpu_up(area);
1142 return notifier_from_errno(ret);
1145 case CPU_UP_CANCELED:
1146 area = &per_cpu(zs_map_area, cpu);
1147 __zs_cpu_down(area);
1154 static struct notifier_block zs_cpu_nb = {
1155 .notifier_call = zs_cpu_notifier
1158 static int zs_register_cpu_notifier(void)
1160 int cpu, uninitialized_var(ret);
1162 cpu_notifier_register_begin();
1164 __register_cpu_notifier(&zs_cpu_nb);
1165 for_each_online_cpu(cpu) {
1166 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1167 if (notifier_to_errno(ret))
1171 cpu_notifier_register_done();
1172 return notifier_to_errno(ret);
1175 static void zs_unregister_cpu_notifier(void)
1179 cpu_notifier_register_begin();
1181 for_each_online_cpu(cpu)
1182 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1183 __unregister_cpu_notifier(&zs_cpu_nb);
1185 cpu_notifier_register_done();
1188 static void init_zs_size_classes(void)
1192 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1193 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1196 zs_size_classes = nr;
1199 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1201 if (prev->pages_per_zspage != pages_per_zspage)
1204 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1205 != get_maxobj_per_zspage(size, pages_per_zspage))
1211 static bool zspage_full(struct page *page)
1213 BUG_ON(!is_first_page(page));
1215 return page->inuse == page->objects;
1218 unsigned long zs_get_total_pages(struct zs_pool *pool)
1220 return atomic_long_read(&pool->pages_allocated);
1222 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1225 * zs_map_object - get address of allocated object from handle.
1226 * @pool: pool from which the object was allocated
1227 * @handle: handle returned from zs_malloc
1229 * Before using an object allocated from zs_malloc, it must be mapped using
1230 * this function. When done with the object, it must be unmapped using
1233 * Only one object can be mapped per cpu at a time. There is no protection
1234 * against nested mappings.
1236 * This function returns with preemption and page faults disabled.
1238 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1242 unsigned long obj, obj_idx, off;
1244 unsigned int class_idx;
1245 enum fullness_group fg;
1246 struct size_class *class;
1247 struct mapping_area *area;
1248 struct page *pages[2];
1254 * Because we use per-cpu mapping areas shared among the
1255 * pools/users, we can't allow mapping in interrupt context
1256 * because it can corrupt another users mappings.
1258 BUG_ON(in_interrupt());
1260 /* From now on, migration cannot move the object */
1263 obj = handle_to_obj(handle);
1264 obj_to_location(obj, &page, &obj_idx);
1265 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1266 class = pool->size_class[class_idx];
1267 off = obj_idx_to_offset(page, obj_idx, class->size);
1269 area = &get_cpu_var(zs_map_area);
1271 if (off + class->size <= PAGE_SIZE) {
1272 /* this object is contained entirely within a page */
1273 area->vm_addr = kmap_atomic(page);
1274 ret = area->vm_addr + off;
1278 /* this object spans two pages */
1280 pages[1] = get_next_page(page);
1283 ret = __zs_map_object(area, pages, off, class->size);
1286 ret += ZS_HANDLE_SIZE;
1290 EXPORT_SYMBOL_GPL(zs_map_object);
1292 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1295 unsigned long obj, obj_idx, off;
1297 unsigned int class_idx;
1298 enum fullness_group fg;
1299 struct size_class *class;
1300 struct mapping_area *area;
1304 obj = handle_to_obj(handle);
1305 obj_to_location(obj, &page, &obj_idx);
1306 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1307 class = pool->size_class[class_idx];
1308 off = obj_idx_to_offset(page, obj_idx, class->size);
1310 area = this_cpu_ptr(&zs_map_area);
1311 if (off + class->size <= PAGE_SIZE)
1312 kunmap_atomic(area->vm_addr);
1314 struct page *pages[2];
1317 pages[1] = get_next_page(page);
1320 __zs_unmap_object(area, pages, off, class->size);
1322 put_cpu_var(zs_map_area);
1325 EXPORT_SYMBOL_GPL(zs_unmap_object);
1327 static unsigned long obj_malloc(struct page *first_page,
1328 struct size_class *class, unsigned long handle)
1331 struct link_free *link;
1333 struct page *m_page;
1334 unsigned long m_objidx, m_offset;
1337 handle |= OBJ_ALLOCATED_TAG;
1338 obj = (unsigned long)first_page->freelist;
1339 obj_to_location(obj, &m_page, &m_objidx);
1340 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1342 vaddr = kmap_atomic(m_page);
1343 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1344 first_page->freelist = link->next;
1346 /* record handle in the header of allocated chunk */
1347 link->handle = handle;
1349 /* record handle in first_page->private */
1350 set_page_private(first_page, handle);
1351 kunmap_atomic(vaddr);
1352 first_page->inuse++;
1353 zs_stat_inc(class, OBJ_USED, 1);
1360 * zs_malloc - Allocate block of given size from pool.
1361 * @pool: pool to allocate from
1362 * @size: size of block to allocate
1364 * On success, handle to the allocated object is returned,
1366 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1368 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1370 unsigned long handle, obj;
1371 struct size_class *class;
1372 struct page *first_page;
1374 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1377 handle = alloc_handle(pool);
1381 /* extra space in chunk to keep the handle */
1382 size += ZS_HANDLE_SIZE;
1383 class = pool->size_class[get_size_class_index(size)];
1385 spin_lock(&class->lock);
1386 first_page = find_get_zspage(class);
1389 spin_unlock(&class->lock);
1390 first_page = alloc_zspage(class, pool->flags);
1391 if (unlikely(!first_page)) {
1392 free_handle(pool, handle);
1396 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1397 atomic_long_add(class->pages_per_zspage,
1398 &pool->pages_allocated);
1400 spin_lock(&class->lock);
1401 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1402 class->size, class->pages_per_zspage));
1405 obj = obj_malloc(first_page, class, handle);
1406 /* Now move the zspage to another fullness group, if required */
1407 fix_fullness_group(class, first_page);
1408 record_obj(handle, obj);
1409 spin_unlock(&class->lock);
1413 EXPORT_SYMBOL_GPL(zs_malloc);
1415 static void obj_free(struct zs_pool *pool, struct size_class *class,
1418 struct link_free *link;
1419 struct page *first_page, *f_page;
1420 unsigned long f_objidx, f_offset;
1423 enum fullness_group fullness;
1427 obj &= ~OBJ_ALLOCATED_TAG;
1428 obj_to_location(obj, &f_page, &f_objidx);
1429 first_page = get_first_page(f_page);
1431 get_zspage_mapping(first_page, &class_idx, &fullness);
1432 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1434 vaddr = kmap_atomic(f_page);
1436 /* Insert this object in containing zspage's freelist */
1437 link = (struct link_free *)(vaddr + f_offset);
1438 link->next = first_page->freelist;
1440 set_page_private(first_page, 0);
1441 kunmap_atomic(vaddr);
1442 first_page->freelist = (void *)obj;
1443 first_page->inuse--;
1444 zs_stat_dec(class, OBJ_USED, 1);
1447 void zs_free(struct zs_pool *pool, unsigned long handle)
1449 struct page *first_page, *f_page;
1450 unsigned long obj, f_objidx;
1452 struct size_class *class;
1453 enum fullness_group fullness;
1455 if (unlikely(!handle))
1459 obj = handle_to_obj(handle);
1460 obj_to_location(obj, &f_page, &f_objidx);
1461 first_page = get_first_page(f_page);
1463 get_zspage_mapping(first_page, &class_idx, &fullness);
1464 class = pool->size_class[class_idx];
1466 spin_lock(&class->lock);
1467 obj_free(pool, class, obj);
1468 fullness = fix_fullness_group(class, first_page);
1469 if (fullness == ZS_EMPTY) {
1470 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1471 class->size, class->pages_per_zspage));
1472 atomic_long_sub(class->pages_per_zspage,
1473 &pool->pages_allocated);
1474 free_zspage(first_page);
1476 spin_unlock(&class->lock);
1479 free_handle(pool, handle);
1481 EXPORT_SYMBOL_GPL(zs_free);
1483 static void zs_object_copy(unsigned long dst, unsigned long src,
1484 struct size_class *class)
1486 struct page *s_page, *d_page;
1487 unsigned long s_objidx, d_objidx;
1488 unsigned long s_off, d_off;
1489 void *s_addr, *d_addr;
1490 int s_size, d_size, size;
1493 s_size = d_size = class->size;
1495 obj_to_location(src, &s_page, &s_objidx);
1496 obj_to_location(dst, &d_page, &d_objidx);
1498 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1499 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1501 if (s_off + class->size > PAGE_SIZE)
1502 s_size = PAGE_SIZE - s_off;
1504 if (d_off + class->size > PAGE_SIZE)
1505 d_size = PAGE_SIZE - d_off;
1507 s_addr = kmap_atomic(s_page);
1508 d_addr = kmap_atomic(d_page);
1511 size = min(s_size, d_size);
1512 memcpy(d_addr + d_off, s_addr + s_off, size);
1515 if (written == class->size)
1523 if (s_off >= PAGE_SIZE) {
1524 kunmap_atomic(d_addr);
1525 kunmap_atomic(s_addr);
1526 s_page = get_next_page(s_page);
1528 s_addr = kmap_atomic(s_page);
1529 d_addr = kmap_atomic(d_page);
1530 s_size = class->size - written;
1534 if (d_off >= PAGE_SIZE) {
1535 kunmap_atomic(d_addr);
1536 d_page = get_next_page(d_page);
1538 d_addr = kmap_atomic(d_page);
1539 d_size = class->size - written;
1544 kunmap_atomic(d_addr);
1545 kunmap_atomic(s_addr);
1549 * Find alloced object in zspage from index object and
1552 static unsigned long find_alloced_obj(struct page *page, int index,
1553 struct size_class *class)
1557 unsigned long handle = 0;
1558 void *addr = kmap_atomic(page);
1560 if (!is_first_page(page))
1561 offset = page->index;
1562 offset += class->size * index;
1564 while (offset < PAGE_SIZE) {
1565 head = obj_to_head(class, page, addr + offset);
1566 if (head & OBJ_ALLOCATED_TAG) {
1567 handle = head & ~OBJ_ALLOCATED_TAG;
1568 if (trypin_tag(handle))
1573 offset += class->size;
1577 kunmap_atomic(addr);
1581 struct zs_compact_control {
1582 /* Source page for migration which could be a subpage of zspage. */
1583 struct page *s_page;
1584 /* Destination page for migration which should be a first page
1586 struct page *d_page;
1587 /* Starting object index within @s_page which used for live object
1588 * in the subpage. */
1592 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1593 struct zs_compact_control *cc)
1595 unsigned long used_obj, free_obj;
1596 unsigned long handle;
1597 struct page *s_page = cc->s_page;
1598 struct page *d_page = cc->d_page;
1599 unsigned long index = cc->index;
1603 handle = find_alloced_obj(s_page, index, class);
1605 s_page = get_next_page(s_page);
1612 /* Stop if there is no more space */
1613 if (zspage_full(d_page)) {
1619 used_obj = handle_to_obj(handle);
1620 free_obj = obj_malloc(d_page, class, handle);
1621 zs_object_copy(free_obj, used_obj, class);
1623 record_obj(handle, free_obj);
1625 obj_free(pool, class, used_obj);
1628 /* Remember last position in this iteration */
1629 cc->s_page = s_page;
1635 static struct page *isolate_target_page(struct size_class *class)
1640 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1641 page = class->fullness_list[i];
1643 remove_zspage(page, class, i);
1652 * putback_zspage - add @first_page into right class's fullness list
1653 * @pool: target pool
1654 * @class: destination class
1655 * @first_page: target page
1657 * Return @fist_page's fullness_group
1659 static enum fullness_group putback_zspage(struct zs_pool *pool,
1660 struct size_class *class,
1661 struct page *first_page)
1663 enum fullness_group fullness;
1665 BUG_ON(!is_first_page(first_page));
1667 fullness = get_fullness_group(first_page);
1668 insert_zspage(first_page, class, fullness);
1669 set_zspage_mapping(first_page, class->index, fullness);
1671 if (fullness == ZS_EMPTY) {
1672 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1673 class->size, class->pages_per_zspage));
1674 atomic_long_sub(class->pages_per_zspage,
1675 &pool->pages_allocated);
1677 free_zspage(first_page);
1683 static struct page *isolate_source_page(struct size_class *class)
1687 page = class->fullness_list[ZS_ALMOST_EMPTY];
1689 remove_zspage(page, class, ZS_ALMOST_EMPTY);
1696 * Based on the number of unused allocated objects calculate
1697 * and return the number of pages that we can free.
1699 * Should be called under class->lock.
1701 static unsigned long zs_can_compact(struct size_class *class)
1703 unsigned long obj_wasted;
1705 if (!zs_stat_get(class, CLASS_ALMOST_EMPTY))
1708 obj_wasted = zs_stat_get(class, OBJ_ALLOCATED) -
1709 zs_stat_get(class, OBJ_USED);
1711 obj_wasted /= get_maxobj_per_zspage(class->size,
1712 class->pages_per_zspage);
1714 return obj_wasted * get_pages_per_zspage(class->size);
1717 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
1719 struct zs_compact_control cc;
1720 struct page *src_page;
1721 struct page *dst_page = NULL;
1723 spin_lock(&class->lock);
1724 while ((src_page = isolate_source_page(class))) {
1726 BUG_ON(!is_first_page(src_page));
1728 if (!zs_can_compact(class))
1732 cc.s_page = src_page;
1734 while ((dst_page = isolate_target_page(class))) {
1735 cc.d_page = dst_page;
1737 * If there is no more space in dst_page, resched
1738 * and see if anyone had allocated another zspage.
1740 if (!migrate_zspage(pool, class, &cc))
1743 putback_zspage(pool, class, dst_page);
1746 /* Stop if we couldn't find slot */
1747 if (dst_page == NULL)
1750 putback_zspage(pool, class, dst_page);
1751 if (putback_zspage(pool, class, src_page) == ZS_EMPTY)
1752 pool->stats.pages_compacted +=
1753 get_pages_per_zspage(class->size);
1754 spin_unlock(&class->lock);
1756 spin_lock(&class->lock);
1760 putback_zspage(pool, class, src_page);
1762 spin_unlock(&class->lock);
1765 unsigned long zs_compact(struct zs_pool *pool)
1768 struct size_class *class;
1770 for (i = zs_size_classes - 1; i >= 0; i--) {
1771 class = pool->size_class[i];
1774 if (class->index != i)
1776 __zs_compact(pool, class);
1779 return pool->stats.pages_compacted;
1781 EXPORT_SYMBOL_GPL(zs_compact);
1783 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
1785 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
1787 EXPORT_SYMBOL_GPL(zs_pool_stats);
1789 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
1790 struct shrink_control *sc)
1792 unsigned long pages_freed;
1793 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1796 pages_freed = pool->stats.pages_compacted;
1798 * Compact classes and calculate compaction delta.
1799 * Can run concurrently with a manually triggered
1800 * (by user) compaction.
1802 pages_freed = zs_compact(pool) - pages_freed;
1804 return pages_freed ? pages_freed : SHRINK_STOP;
1807 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
1808 struct shrink_control *sc)
1811 struct size_class *class;
1812 unsigned long pages_to_free = 0;
1813 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1816 if (!pool->shrinker_enabled)
1819 for (i = zs_size_classes - 1; i >= 0; i--) {
1820 class = pool->size_class[i];
1823 if (class->index != i)
1826 spin_lock(&class->lock);
1827 pages_to_free += zs_can_compact(class);
1828 spin_unlock(&class->lock);
1831 return pages_to_free;
1834 static void zs_unregister_shrinker(struct zs_pool *pool)
1836 if (pool->shrinker_enabled) {
1837 unregister_shrinker(&pool->shrinker);
1838 pool->shrinker_enabled = false;
1842 static int zs_register_shrinker(struct zs_pool *pool)
1844 pool->shrinker.scan_objects = zs_shrinker_scan;
1845 pool->shrinker.count_objects = zs_shrinker_count;
1846 pool->shrinker.batch = 0;
1847 pool->shrinker.seeks = DEFAULT_SEEKS;
1849 return register_shrinker(&pool->shrinker);
1853 * zs_create_pool - Creates an allocation pool to work from.
1854 * @flags: allocation flags used to allocate pool metadata
1856 * This function must be called before anything when using
1857 * the zsmalloc allocator.
1859 * On success, a pointer to the newly created pool is returned,
1862 struct zs_pool *zs_create_pool(char *name, gfp_t flags)
1865 struct zs_pool *pool;
1866 struct size_class *prev_class = NULL;
1868 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1872 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1874 if (!pool->size_class) {
1879 pool->name = kstrdup(name, GFP_KERNEL);
1883 if (create_handle_cache(pool))
1887 * Iterate reversly, because, size of size_class that we want to use
1888 * for merging should be larger or equal to current size.
1890 for (i = zs_size_classes - 1; i >= 0; i--) {
1892 int pages_per_zspage;
1893 struct size_class *class;
1895 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1896 if (size > ZS_MAX_ALLOC_SIZE)
1897 size = ZS_MAX_ALLOC_SIZE;
1898 pages_per_zspage = get_pages_per_zspage(size);
1901 * size_class is used for normal zsmalloc operation such
1902 * as alloc/free for that size. Although it is natural that we
1903 * have one size_class for each size, there is a chance that we
1904 * can get more memory utilization if we use one size_class for
1905 * many different sizes whose size_class have same
1906 * characteristics. So, we makes size_class point to
1907 * previous size_class if possible.
1910 if (can_merge(prev_class, size, pages_per_zspage)) {
1911 pool->size_class[i] = prev_class;
1916 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1922 class->pages_per_zspage = pages_per_zspage;
1923 if (pages_per_zspage == 1 &&
1924 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1926 spin_lock_init(&class->lock);
1927 pool->size_class[i] = class;
1932 pool->flags = flags;
1934 if (zs_pool_stat_create(name, pool))
1938 * Not critical, we still can use the pool
1939 * and user can trigger compaction manually.
1941 if (zs_register_shrinker(pool) == 0)
1942 pool->shrinker_enabled = true;
1946 zs_destroy_pool(pool);
1949 EXPORT_SYMBOL_GPL(zs_create_pool);
1951 void zs_destroy_pool(struct zs_pool *pool)
1955 zs_unregister_shrinker(pool);
1956 zs_pool_stat_destroy(pool);
1958 for (i = 0; i < zs_size_classes; i++) {
1960 struct size_class *class = pool->size_class[i];
1965 if (class->index != i)
1968 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1969 if (class->fullness_list[fg]) {
1970 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1977 destroy_handle_cache(pool);
1978 kfree(pool->size_class);
1982 EXPORT_SYMBOL_GPL(zs_destroy_pool);
1984 static int __init zs_init(void)
1986 int ret = zs_register_cpu_notifier();
1991 init_zs_size_classes();
1994 zpool_register_driver(&zs_zpool_driver);
1997 ret = zs_stat_init();
1999 pr_err("zs stat initialization failed\n");
2006 zpool_unregister_driver(&zs_zpool_driver);
2009 zs_unregister_cpu_notifier();
2014 static void __exit zs_exit(void)
2017 zpool_unregister_driver(&zs_zpool_driver);
2019 zs_unregister_cpu_notifier();
2024 module_init(zs_init);
2025 module_exit(zs_exit);
2027 MODULE_LICENSE("Dual BSD/GPL");
2028 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");