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->private: 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: refers to the component page after the first page
30 * If the page is first_page for huge object, it stores handle.
31 * Look at size_class->huge.
32 * page->freelist: points to the first free object in zspage.
33 * Free objects are linked together using in-place
35 * page->objects: maximum number of objects we can store in this
36 * zspage (class->zspage_order * PAGE_SIZE / class->size)
37 * page->lru: links together first pages of various zspages.
38 * Basically forming list of zspages in a fullness group.
39 * page->mapping: class index and fullness group of the zspage
40 * page->inuse: the number of objects that are used in this 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/preempt.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>
67 #include <linux/locallock.h>
70 * This must be power of 2 and greater than of equal to sizeof(link_free).
71 * These two conditions ensure that any 'struct link_free' itself doesn't
72 * span more than 1 page which avoids complex case of mapping 2 pages simply
73 * to restore link_free pointer values.
78 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
79 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
81 #define ZS_MAX_ZSPAGE_ORDER 2
82 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
84 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
87 * Object location (<PFN>, <obj_idx>) is encoded as
88 * as single (unsigned long) handle value.
90 * Note that object index <obj_idx> is relative to system
91 * page <PFN> it is stored in, so for each sub-page belonging
92 * to a zspage, obj_idx starts with 0.
94 * This is made more complicated by various memory models and PAE.
97 #ifndef MAX_PHYSMEM_BITS
98 #ifdef CONFIG_HIGHMEM64G
99 #define MAX_PHYSMEM_BITS 36
100 #else /* !CONFIG_HIGHMEM64G */
102 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
105 #define MAX_PHYSMEM_BITS BITS_PER_LONG
108 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
111 * Memory for allocating for handle keeps object position by
112 * encoding <page, obj_idx> and the encoded value has a room
113 * in least bit(ie, look at obj_to_location).
114 * We use the bit to synchronize between object access by
115 * user and migration.
117 #define HANDLE_PIN_BIT 0
120 * Head in allocated object should have OBJ_ALLOCATED_TAG
121 * to identify the object was allocated or not.
122 * It's okay to add the status bit in the least bit because
123 * header keeps handle which is 4byte-aligned address so we
124 * have room for two bit at least.
126 #define OBJ_ALLOCATED_TAG 1
127 #define OBJ_TAG_BITS 1
128 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
129 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
131 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
132 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
133 #define ZS_MIN_ALLOC_SIZE \
134 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
135 /* each chunk includes extra space to keep handle */
136 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
139 * On systems with 4K page size, this gives 255 size classes! There is a
141 * - Large number of size classes is potentially wasteful as free page are
142 * spread across these classes
143 * - Small number of size classes causes large internal fragmentation
144 * - Probably its better to use specific size classes (empirically
145 * determined). NOTE: all those class sizes must be set as multiple of
146 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
148 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
151 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
154 * We do not maintain any list for completely empty or full pages
156 enum fullness_group {
159 _ZS_NR_FULLNESS_GROUPS,
172 #ifdef CONFIG_ZSMALLOC_STAT
173 #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1)
175 #define NR_ZS_STAT_TYPE (OBJ_USED + 1)
178 struct zs_size_stat {
179 unsigned long objs[NR_ZS_STAT_TYPE];
182 #ifdef CONFIG_ZSMALLOC_STAT
183 static struct dentry *zs_stat_root;
187 * number of size_classes
189 static int zs_size_classes;
192 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
194 * n = number of allocated objects
195 * N = total number of objects zspage can store
196 * f = fullness_threshold_frac
198 * Similarly, we assign zspage to:
199 * ZS_ALMOST_FULL when n > N / f
200 * ZS_EMPTY when n == 0
201 * ZS_FULL when n == N
203 * (see: fix_fullness_group())
205 static const int fullness_threshold_frac = 4;
209 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
211 * Size of objects stored in this class. Must be multiple
217 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
218 int pages_per_zspage;
219 struct zs_size_stat stats;
221 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
226 * Placed within free objects to form a singly linked list.
227 * For every zspage, first_page->freelist gives head of this list.
229 * This must be power of 2 and less than or equal to ZS_ALIGN
234 * Position of next free chunk (encodes <PFN, obj_idx>)
235 * It's valid for non-allocated object
239 * Handle of allocated object.
241 unsigned long handle;
248 struct size_class **size_class;
249 struct kmem_cache *handle_cachep;
251 gfp_t flags; /* allocation flags used when growing pool */
252 atomic_long_t pages_allocated;
254 struct zs_pool_stats stats;
256 /* Compact classes */
257 struct shrinker shrinker;
259 * To signify that register_shrinker() was successful
260 * and unregister_shrinker() will not Oops.
262 bool shrinker_enabled;
263 #ifdef CONFIG_ZSMALLOC_STAT
264 struct dentry *stat_dentry;
269 * A zspage's class index and fullness group
270 * are encoded in its (first)page->mapping
272 #define CLASS_IDX_BITS 28
273 #define FULLNESS_BITS 4
274 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
275 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
277 struct mapping_area {
278 #ifdef CONFIG_PGTABLE_MAPPING
279 struct vm_struct *vm; /* vm area for mapping object that span pages */
281 char *vm_buf; /* copy buffer for objects that span pages */
283 char *vm_addr; /* address of kmap_atomic()'ed pages */
284 enum zs_mapmode vm_mm; /* mapping mode */
288 static int create_handle_cache(struct zs_pool *pool)
290 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
292 return pool->handle_cachep ? 0 : 1;
295 static void destroy_handle_cache(struct zs_pool *pool)
297 kmem_cache_destroy(pool->handle_cachep);
300 static unsigned long alloc_handle(struct zs_pool *pool)
302 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
303 pool->flags & ~__GFP_HIGHMEM);
306 static void free_handle(struct zs_pool *pool, unsigned long handle)
308 kmem_cache_free(pool->handle_cachep, (void *)handle);
311 static void record_obj(unsigned long handle, unsigned long obj)
314 * lsb of @obj represents handle lock while other bits
315 * represent object value the handle is pointing so
316 * updating shouldn't do store tearing.
318 WRITE_ONCE(*(unsigned long *)handle, obj);
325 static void *zs_zpool_create(const char *name, gfp_t gfp,
326 const struct zpool_ops *zpool_ops,
329 return zs_create_pool(name, gfp);
332 static void zs_zpool_destroy(void *pool)
334 zs_destroy_pool(pool);
337 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
338 unsigned long *handle)
340 *handle = zs_malloc(pool, size);
341 return *handle ? 0 : -1;
343 static void zs_zpool_free(void *pool, unsigned long handle)
345 zs_free(pool, handle);
348 static int zs_zpool_shrink(void *pool, unsigned int pages,
349 unsigned int *reclaimed)
354 static void *zs_zpool_map(void *pool, unsigned long handle,
355 enum zpool_mapmode mm)
357 enum zs_mapmode zs_mm;
366 case ZPOOL_MM_RW: /* fallthru */
372 return zs_map_object(pool, handle, zs_mm);
374 static void zs_zpool_unmap(void *pool, unsigned long handle)
376 zs_unmap_object(pool, handle);
379 static u64 zs_zpool_total_size(void *pool)
381 return zs_get_total_pages(pool) << PAGE_SHIFT;
384 static struct zpool_driver zs_zpool_driver = {
386 .owner = THIS_MODULE,
387 .create = zs_zpool_create,
388 .destroy = zs_zpool_destroy,
389 .malloc = zs_zpool_malloc,
390 .free = zs_zpool_free,
391 .shrink = zs_zpool_shrink,
393 .unmap = zs_zpool_unmap,
394 .total_size = zs_zpool_total_size,
397 MODULE_ALIAS("zpool-zsmalloc");
398 #endif /* CONFIG_ZPOOL */
400 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
402 return pages_per_zspage * PAGE_SIZE / size;
405 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
406 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
407 static DEFINE_LOCAL_IRQ_LOCK(zs_map_area_lock);
409 static int is_first_page(struct page *page)
411 return PagePrivate(page);
414 static int is_last_page(struct page *page)
416 return PagePrivate2(page);
419 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
420 enum fullness_group *fullness)
423 BUG_ON(!is_first_page(page));
425 m = (unsigned long)page->mapping;
426 *fullness = m & FULLNESS_MASK;
427 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
430 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
431 enum fullness_group fullness)
434 BUG_ON(!is_first_page(page));
436 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
437 (fullness & FULLNESS_MASK);
438 page->mapping = (struct address_space *)m;
442 * zsmalloc divides the pool into various size classes where each
443 * class maintains a list of zspages where each zspage is divided
444 * into equal sized chunks. Each allocation falls into one of these
445 * classes depending on its size. This function returns index of the
446 * size class which has chunk size big enough to hold the give size.
448 static int get_size_class_index(int size)
452 if (likely(size > ZS_MIN_ALLOC_SIZE))
453 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
454 ZS_SIZE_CLASS_DELTA);
456 return min(zs_size_classes - 1, idx);
459 static inline void zs_stat_inc(struct size_class *class,
460 enum zs_stat_type type, unsigned long cnt)
462 if (type < NR_ZS_STAT_TYPE)
463 class->stats.objs[type] += cnt;
466 static inline void zs_stat_dec(struct size_class *class,
467 enum zs_stat_type type, unsigned long cnt)
469 if (type < NR_ZS_STAT_TYPE)
470 class->stats.objs[type] -= cnt;
473 static inline unsigned long zs_stat_get(struct size_class *class,
474 enum zs_stat_type type)
476 if (type < NR_ZS_STAT_TYPE)
477 return class->stats.objs[type];
481 #ifdef CONFIG_ZSMALLOC_STAT
483 static int __init zs_stat_init(void)
485 if (!debugfs_initialized())
488 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
495 static void __exit zs_stat_exit(void)
497 debugfs_remove_recursive(zs_stat_root);
500 static int zs_stats_size_show(struct seq_file *s, void *v)
503 struct zs_pool *pool = s->private;
504 struct size_class *class;
506 unsigned long class_almost_full, class_almost_empty;
507 unsigned long obj_allocated, obj_used, pages_used;
508 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
509 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
511 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
512 "class", "size", "almost_full", "almost_empty",
513 "obj_allocated", "obj_used", "pages_used",
516 for (i = 0; i < zs_size_classes; i++) {
517 class = pool->size_class[i];
519 if (class->index != i)
522 spin_lock(&class->lock);
523 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
524 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
525 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
526 obj_used = zs_stat_get(class, OBJ_USED);
527 spin_unlock(&class->lock);
529 objs_per_zspage = get_maxobj_per_zspage(class->size,
530 class->pages_per_zspage);
531 pages_used = obj_allocated / objs_per_zspage *
532 class->pages_per_zspage;
534 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
535 i, class->size, class_almost_full, class_almost_empty,
536 obj_allocated, obj_used, pages_used,
537 class->pages_per_zspage);
539 total_class_almost_full += class_almost_full;
540 total_class_almost_empty += class_almost_empty;
541 total_objs += obj_allocated;
542 total_used_objs += obj_used;
543 total_pages += pages_used;
547 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
548 "Total", "", total_class_almost_full,
549 total_class_almost_empty, total_objs,
550 total_used_objs, total_pages);
555 static int zs_stats_size_open(struct inode *inode, struct file *file)
557 return single_open(file, zs_stats_size_show, inode->i_private);
560 static const struct file_operations zs_stat_size_ops = {
561 .open = zs_stats_size_open,
564 .release = single_release,
567 static int zs_pool_stat_create(const char *name, struct zs_pool *pool)
569 struct dentry *entry;
574 entry = debugfs_create_dir(name, zs_stat_root);
576 pr_warn("debugfs dir <%s> creation failed\n", name);
579 pool->stat_dentry = entry;
581 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
582 pool->stat_dentry, pool, &zs_stat_size_ops);
584 pr_warn("%s: debugfs file entry <%s> creation failed\n",
592 static void zs_pool_stat_destroy(struct zs_pool *pool)
594 debugfs_remove_recursive(pool->stat_dentry);
597 #else /* CONFIG_ZSMALLOC_STAT */
598 static int __init zs_stat_init(void)
603 static void __exit zs_stat_exit(void)
607 static inline int zs_pool_stat_create(const char *name, struct zs_pool *pool)
612 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
619 * For each size class, zspages are divided into different groups
620 * depending on how "full" they are. This was done so that we could
621 * easily find empty or nearly empty zspages when we try to shrink
622 * the pool (not yet implemented). This function returns fullness
623 * status of the given page.
625 static enum fullness_group get_fullness_group(struct page *page)
627 int inuse, max_objects;
628 enum fullness_group fg;
629 BUG_ON(!is_first_page(page));
632 max_objects = page->objects;
636 else if (inuse == max_objects)
638 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
639 fg = ZS_ALMOST_EMPTY;
647 * Each size class maintains various freelists and zspages are assigned
648 * to one of these freelists based on the number of live objects they
649 * have. This functions inserts the given zspage into the freelist
650 * identified by <class, fullness_group>.
652 static void insert_zspage(struct page *page, struct size_class *class,
653 enum fullness_group fullness)
657 BUG_ON(!is_first_page(page));
659 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
662 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
663 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
665 head = &class->fullness_list[fullness];
672 * We want to see more ZS_FULL pages and less almost
673 * empty/full. Put pages with higher ->inuse first.
675 list_add_tail(&page->lru, &(*head)->lru);
676 if (page->inuse >= (*head)->inuse)
681 * This function removes the given zspage from the freelist identified
682 * by <class, fullness_group>.
684 static void remove_zspage(struct page *page, struct size_class *class,
685 enum fullness_group fullness)
689 BUG_ON(!is_first_page(page));
691 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
694 head = &class->fullness_list[fullness];
696 if (list_empty(&(*head)->lru))
698 else if (*head == page)
699 *head = (struct page *)list_entry((*head)->lru.next,
702 list_del_init(&page->lru);
703 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
704 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
708 * Each size class maintains zspages in different fullness groups depending
709 * on the number of live objects they contain. When allocating or freeing
710 * objects, the fullness status of the page can change, say, from ALMOST_FULL
711 * to ALMOST_EMPTY when freeing an object. This function checks if such
712 * a status change has occurred for the given page and accordingly moves the
713 * page from the freelist of the old fullness group to that of the new
716 static enum fullness_group fix_fullness_group(struct size_class *class,
720 enum fullness_group currfg, newfg;
722 BUG_ON(!is_first_page(page));
724 get_zspage_mapping(page, &class_idx, &currfg);
725 newfg = get_fullness_group(page);
729 remove_zspage(page, class, currfg);
730 insert_zspage(page, class, newfg);
731 set_zspage_mapping(page, class_idx, newfg);
738 * We have to decide on how many pages to link together
739 * to form a zspage for each size class. This is important
740 * to reduce wastage due to unusable space left at end of
741 * each zspage which is given as:
742 * wastage = Zp % class_size
743 * usage = Zp - wastage
744 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
746 * For example, for size class of 3/8 * PAGE_SIZE, we should
747 * link together 3 PAGE_SIZE sized pages to form a zspage
748 * since then we can perfectly fit in 8 such objects.
750 static int get_pages_per_zspage(int class_size)
752 int i, max_usedpc = 0;
753 /* zspage order which gives maximum used size per KB */
754 int max_usedpc_order = 1;
756 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
760 zspage_size = i * PAGE_SIZE;
761 waste = zspage_size % class_size;
762 usedpc = (zspage_size - waste) * 100 / zspage_size;
764 if (usedpc > max_usedpc) {
766 max_usedpc_order = i;
770 return max_usedpc_order;
774 * A single 'zspage' is composed of many system pages which are
775 * linked together using fields in struct page. This function finds
776 * the first/head page, given any component page of a zspage.
778 static struct page *get_first_page(struct page *page)
780 if (is_first_page(page))
783 return (struct page *)page_private(page);
786 static struct page *get_next_page(struct page *page)
790 if (is_last_page(page))
792 else if (is_first_page(page))
793 next = (struct page *)page_private(page);
795 next = list_entry(page->lru.next, struct page, lru);
801 * Encode <page, obj_idx> as a single handle value.
802 * We use the least bit of handle for tagging.
804 static void *location_to_obj(struct page *page, unsigned long obj_idx)
813 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
814 obj |= ((obj_idx) & OBJ_INDEX_MASK);
815 obj <<= OBJ_TAG_BITS;
821 * Decode <page, obj_idx> pair from the given object handle. We adjust the
822 * decoded obj_idx back to its original value since it was adjusted in
825 static void obj_to_location(unsigned long obj, struct page **page,
826 unsigned long *obj_idx)
828 obj >>= OBJ_TAG_BITS;
829 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
830 *obj_idx = (obj & OBJ_INDEX_MASK);
833 static unsigned long handle_to_obj(unsigned long handle)
835 return *(unsigned long *)handle;
838 static unsigned long obj_to_head(struct size_class *class, struct page *page,
842 VM_BUG_ON(!is_first_page(page));
843 return page_private(page);
845 return *(unsigned long *)obj;
848 static unsigned long obj_idx_to_offset(struct page *page,
849 unsigned long obj_idx, int class_size)
851 unsigned long off = 0;
853 if (!is_first_page(page))
856 return off + obj_idx * class_size;
859 static inline int trypin_tag(unsigned long handle)
861 unsigned long *ptr = (unsigned long *)handle;
863 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
866 static void pin_tag(unsigned long handle)
868 while (!trypin_tag(handle));
871 static void unpin_tag(unsigned long handle)
873 unsigned long *ptr = (unsigned long *)handle;
875 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
878 static void reset_page(struct page *page)
880 clear_bit(PG_private, &page->flags);
881 clear_bit(PG_private_2, &page->flags);
882 set_page_private(page, 0);
883 page->mapping = NULL;
884 page->freelist = NULL;
885 page_mapcount_reset(page);
888 static void free_zspage(struct page *first_page)
890 struct page *nextp, *tmp, *head_extra;
892 BUG_ON(!is_first_page(first_page));
893 BUG_ON(first_page->inuse);
895 head_extra = (struct page *)page_private(first_page);
897 reset_page(first_page);
898 __free_page(first_page);
900 /* zspage with only 1 system page */
904 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
905 list_del(&nextp->lru);
909 reset_page(head_extra);
910 __free_page(head_extra);
913 /* Initialize a newly allocated zspage */
914 static void init_zspage(struct page *first_page, struct size_class *class)
916 unsigned long off = 0;
917 struct page *page = first_page;
919 BUG_ON(!is_first_page(first_page));
921 struct page *next_page;
922 struct link_free *link;
927 * page->index stores offset of first object starting
928 * in the page. For the first page, this is always 0,
929 * so we use first_page->index (aka ->freelist) to store
930 * head of corresponding zspage's freelist.
932 if (page != first_page)
935 vaddr = kmap_atomic(page);
936 link = (struct link_free *)vaddr + off / sizeof(*link);
938 while ((off += class->size) < PAGE_SIZE) {
939 link->next = location_to_obj(page, i++);
940 link += class->size / sizeof(*link);
944 * We now come to the last (full or partial) object on this
945 * page, which must point to the first object on the next
948 next_page = get_next_page(page);
949 link->next = location_to_obj(next_page, 0);
950 kunmap_atomic(vaddr);
957 * Allocate a zspage for the given size class
959 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
962 struct page *first_page = NULL, *uninitialized_var(prev_page);
965 * Allocate individual pages and link them together as:
966 * 1. first page->private = first sub-page
967 * 2. all sub-pages are linked together using page->lru
968 * 3. each sub-page is linked to the first page using page->private
970 * For each size class, First/Head pages are linked together using
971 * page->lru. Also, we set PG_private to identify the first page
972 * (i.e. no other sub-page has this flag set) and PG_private_2 to
973 * identify the last page.
976 for (i = 0; i < class->pages_per_zspage; i++) {
979 page = alloc_page(flags);
983 INIT_LIST_HEAD(&page->lru);
984 if (i == 0) { /* first page */
985 SetPagePrivate(page);
986 set_page_private(page, 0);
988 first_page->inuse = 0;
991 set_page_private(first_page, (unsigned long)page);
993 set_page_private(page, (unsigned long)first_page);
995 list_add(&page->lru, &prev_page->lru);
996 if (i == class->pages_per_zspage - 1) /* last page */
997 SetPagePrivate2(page);
1001 init_zspage(first_page, class);
1003 first_page->freelist = location_to_obj(first_page, 0);
1004 /* Maximum number of objects we can store in this zspage */
1005 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
1007 error = 0; /* Success */
1010 if (unlikely(error) && first_page) {
1011 free_zspage(first_page);
1018 static struct page *find_get_zspage(struct size_class *class)
1023 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1024 page = class->fullness_list[i];
1032 #ifdef CONFIG_PGTABLE_MAPPING
1033 static inline int __zs_cpu_up(struct mapping_area *area)
1036 * Make sure we don't leak memory if a cpu UP notification
1037 * and zs_init() race and both call zs_cpu_up() on the same cpu
1041 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1047 static inline void __zs_cpu_down(struct mapping_area *area)
1050 free_vm_area(area->vm);
1054 static inline void *__zs_map_object(struct mapping_area *area,
1055 struct page *pages[2], int off, int size)
1057 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1058 area->vm_addr = area->vm->addr;
1059 return area->vm_addr + off;
1062 static inline void __zs_unmap_object(struct mapping_area *area,
1063 struct page *pages[2], int off, int size)
1065 unsigned long addr = (unsigned long)area->vm_addr;
1067 unmap_kernel_range(addr, PAGE_SIZE * 2);
1070 #else /* CONFIG_PGTABLE_MAPPING */
1072 static inline int __zs_cpu_up(struct mapping_area *area)
1075 * Make sure we don't leak memory if a cpu UP notification
1076 * and zs_init() race and both call zs_cpu_up() on the same cpu
1080 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1086 static inline void __zs_cpu_down(struct mapping_area *area)
1088 kfree(area->vm_buf);
1089 area->vm_buf = NULL;
1092 static void *__zs_map_object(struct mapping_area *area,
1093 struct page *pages[2], int off, int size)
1097 char *buf = area->vm_buf;
1099 /* disable page faults to match kmap_atomic() return conditions */
1100 pagefault_disable();
1102 /* no read fastpath */
1103 if (area->vm_mm == ZS_MM_WO)
1106 sizes[0] = PAGE_SIZE - off;
1107 sizes[1] = size - sizes[0];
1109 /* copy object to per-cpu buffer */
1110 addr = kmap_atomic(pages[0]);
1111 memcpy(buf, addr + off, sizes[0]);
1112 kunmap_atomic(addr);
1113 addr = kmap_atomic(pages[1]);
1114 memcpy(buf + sizes[0], addr, sizes[1]);
1115 kunmap_atomic(addr);
1117 return area->vm_buf;
1120 static void __zs_unmap_object(struct mapping_area *area,
1121 struct page *pages[2], int off, int size)
1127 /* no write fastpath */
1128 if (area->vm_mm == ZS_MM_RO)
1133 buf = buf + ZS_HANDLE_SIZE;
1134 size -= ZS_HANDLE_SIZE;
1135 off += ZS_HANDLE_SIZE;
1138 sizes[0] = PAGE_SIZE - off;
1139 sizes[1] = size - sizes[0];
1141 /* copy per-cpu buffer to object */
1142 addr = kmap_atomic(pages[0]);
1143 memcpy(addr + off, buf, sizes[0]);
1144 kunmap_atomic(addr);
1145 addr = kmap_atomic(pages[1]);
1146 memcpy(addr, buf + sizes[0], sizes[1]);
1147 kunmap_atomic(addr);
1150 /* enable page faults to match kunmap_atomic() return conditions */
1154 #endif /* CONFIG_PGTABLE_MAPPING */
1156 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1159 int ret, cpu = (long)pcpu;
1160 struct mapping_area *area;
1163 case CPU_UP_PREPARE:
1164 area = &per_cpu(zs_map_area, cpu);
1165 ret = __zs_cpu_up(area);
1167 return notifier_from_errno(ret);
1170 case CPU_UP_CANCELED:
1171 area = &per_cpu(zs_map_area, cpu);
1172 __zs_cpu_down(area);
1179 static struct notifier_block zs_cpu_nb = {
1180 .notifier_call = zs_cpu_notifier
1183 static int zs_register_cpu_notifier(void)
1185 int cpu, uninitialized_var(ret);
1187 cpu_notifier_register_begin();
1189 __register_cpu_notifier(&zs_cpu_nb);
1190 for_each_online_cpu(cpu) {
1191 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1192 if (notifier_to_errno(ret))
1196 cpu_notifier_register_done();
1197 return notifier_to_errno(ret);
1200 static void zs_unregister_cpu_notifier(void)
1204 cpu_notifier_register_begin();
1206 for_each_online_cpu(cpu)
1207 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1208 __unregister_cpu_notifier(&zs_cpu_nb);
1210 cpu_notifier_register_done();
1213 static void init_zs_size_classes(void)
1217 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1218 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1221 zs_size_classes = nr;
1224 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1226 if (prev->pages_per_zspage != pages_per_zspage)
1229 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1230 != get_maxobj_per_zspage(size, pages_per_zspage))
1236 static bool zspage_full(struct page *page)
1238 BUG_ON(!is_first_page(page));
1240 return page->inuse == page->objects;
1243 unsigned long zs_get_total_pages(struct zs_pool *pool)
1245 return atomic_long_read(&pool->pages_allocated);
1247 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1250 * zs_map_object - get address of allocated object from handle.
1251 * @pool: pool from which the object was allocated
1252 * @handle: handle returned from zs_malloc
1254 * Before using an object allocated from zs_malloc, it must be mapped using
1255 * this function. When done with the object, it must be unmapped using
1258 * Only one object can be mapped per cpu at a time. There is no protection
1259 * against nested mappings.
1261 * This function returns with preemption and page faults disabled.
1263 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1267 unsigned long obj, obj_idx, off;
1269 unsigned int class_idx;
1270 enum fullness_group fg;
1271 struct size_class *class;
1272 struct mapping_area *area;
1273 struct page *pages[2];
1279 * Because we use per-cpu mapping areas shared among the
1280 * pools/users, we can't allow mapping in interrupt context
1281 * because it can corrupt another users mappings.
1283 BUG_ON(in_interrupt());
1285 /* From now on, migration cannot move the object */
1288 obj = handle_to_obj(handle);
1289 obj_to_location(obj, &page, &obj_idx);
1290 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1291 class = pool->size_class[class_idx];
1292 off = obj_idx_to_offset(page, obj_idx, class->size);
1294 area = &get_locked_var(zs_map_area_lock, zs_map_area);
1296 if (off + class->size <= PAGE_SIZE) {
1297 /* this object is contained entirely within a page */
1298 area->vm_addr = kmap_atomic(page);
1299 ret = area->vm_addr + off;
1303 /* this object spans two pages */
1305 pages[1] = get_next_page(page);
1308 ret = __zs_map_object(area, pages, off, class->size);
1311 ret += ZS_HANDLE_SIZE;
1315 EXPORT_SYMBOL_GPL(zs_map_object);
1317 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1320 unsigned long obj, obj_idx, off;
1322 unsigned int class_idx;
1323 enum fullness_group fg;
1324 struct size_class *class;
1325 struct mapping_area *area;
1329 obj = handle_to_obj(handle);
1330 obj_to_location(obj, &page, &obj_idx);
1331 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1332 class = pool->size_class[class_idx];
1333 off = obj_idx_to_offset(page, obj_idx, class->size);
1335 area = this_cpu_ptr(&zs_map_area);
1336 if (off + class->size <= PAGE_SIZE)
1337 kunmap_atomic(area->vm_addr);
1339 struct page *pages[2];
1342 pages[1] = get_next_page(page);
1345 __zs_unmap_object(area, pages, off, class->size);
1347 put_locked_var(zs_map_area_lock, zs_map_area);
1350 EXPORT_SYMBOL_GPL(zs_unmap_object);
1352 static unsigned long obj_malloc(struct page *first_page,
1353 struct size_class *class, unsigned long handle)
1356 struct link_free *link;
1358 struct page *m_page;
1359 unsigned long m_objidx, m_offset;
1362 handle |= OBJ_ALLOCATED_TAG;
1363 obj = (unsigned long)first_page->freelist;
1364 obj_to_location(obj, &m_page, &m_objidx);
1365 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1367 vaddr = kmap_atomic(m_page);
1368 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1369 first_page->freelist = link->next;
1371 /* record handle in the header of allocated chunk */
1372 link->handle = handle;
1374 /* record handle in first_page->private */
1375 set_page_private(first_page, handle);
1376 kunmap_atomic(vaddr);
1377 first_page->inuse++;
1378 zs_stat_inc(class, OBJ_USED, 1);
1385 * zs_malloc - Allocate block of given size from pool.
1386 * @pool: pool to allocate from
1387 * @size: size of block to allocate
1389 * On success, handle to the allocated object is returned,
1391 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1393 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1395 unsigned long handle, obj;
1396 struct size_class *class;
1397 struct page *first_page;
1399 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1402 handle = alloc_handle(pool);
1406 /* extra space in chunk to keep the handle */
1407 size += ZS_HANDLE_SIZE;
1408 class = pool->size_class[get_size_class_index(size)];
1410 spin_lock(&class->lock);
1411 first_page = find_get_zspage(class);
1414 spin_unlock(&class->lock);
1415 first_page = alloc_zspage(class, pool->flags);
1416 if (unlikely(!first_page)) {
1417 free_handle(pool, handle);
1421 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1422 atomic_long_add(class->pages_per_zspage,
1423 &pool->pages_allocated);
1425 spin_lock(&class->lock);
1426 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1427 class->size, class->pages_per_zspage));
1430 obj = obj_malloc(first_page, class, handle);
1431 /* Now move the zspage to another fullness group, if required */
1432 fix_fullness_group(class, first_page);
1433 record_obj(handle, obj);
1434 spin_unlock(&class->lock);
1438 EXPORT_SYMBOL_GPL(zs_malloc);
1440 static void obj_free(struct zs_pool *pool, struct size_class *class,
1443 struct link_free *link;
1444 struct page *first_page, *f_page;
1445 unsigned long f_objidx, f_offset;
1450 obj &= ~OBJ_ALLOCATED_TAG;
1451 obj_to_location(obj, &f_page, &f_objidx);
1452 first_page = get_first_page(f_page);
1454 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1456 vaddr = kmap_atomic(f_page);
1458 /* Insert this object in containing zspage's freelist */
1459 link = (struct link_free *)(vaddr + f_offset);
1460 link->next = first_page->freelist;
1462 set_page_private(first_page, 0);
1463 kunmap_atomic(vaddr);
1464 first_page->freelist = (void *)obj;
1465 first_page->inuse--;
1466 zs_stat_dec(class, OBJ_USED, 1);
1469 void zs_free(struct zs_pool *pool, unsigned long handle)
1471 struct page *first_page, *f_page;
1472 unsigned long obj, f_objidx;
1474 struct size_class *class;
1475 enum fullness_group fullness;
1477 if (unlikely(!handle))
1481 obj = handle_to_obj(handle);
1482 obj_to_location(obj, &f_page, &f_objidx);
1483 first_page = get_first_page(f_page);
1485 get_zspage_mapping(first_page, &class_idx, &fullness);
1486 class = pool->size_class[class_idx];
1488 spin_lock(&class->lock);
1489 obj_free(pool, class, obj);
1490 fullness = fix_fullness_group(class, first_page);
1491 if (fullness == ZS_EMPTY) {
1492 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1493 class->size, class->pages_per_zspage));
1494 atomic_long_sub(class->pages_per_zspage,
1495 &pool->pages_allocated);
1496 free_zspage(first_page);
1498 spin_unlock(&class->lock);
1501 free_handle(pool, handle);
1503 EXPORT_SYMBOL_GPL(zs_free);
1505 static void zs_object_copy(unsigned long dst, unsigned long src,
1506 struct size_class *class)
1508 struct page *s_page, *d_page;
1509 unsigned long s_objidx, d_objidx;
1510 unsigned long s_off, d_off;
1511 void *s_addr, *d_addr;
1512 int s_size, d_size, size;
1515 s_size = d_size = class->size;
1517 obj_to_location(src, &s_page, &s_objidx);
1518 obj_to_location(dst, &d_page, &d_objidx);
1520 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1521 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1523 if (s_off + class->size > PAGE_SIZE)
1524 s_size = PAGE_SIZE - s_off;
1526 if (d_off + class->size > PAGE_SIZE)
1527 d_size = PAGE_SIZE - d_off;
1529 s_addr = kmap_atomic(s_page);
1530 d_addr = kmap_atomic(d_page);
1533 size = min(s_size, d_size);
1534 memcpy(d_addr + d_off, s_addr + s_off, size);
1537 if (written == class->size)
1545 if (s_off >= PAGE_SIZE) {
1546 kunmap_atomic(d_addr);
1547 kunmap_atomic(s_addr);
1548 s_page = get_next_page(s_page);
1550 s_addr = kmap_atomic(s_page);
1551 d_addr = kmap_atomic(d_page);
1552 s_size = class->size - written;
1556 if (d_off >= PAGE_SIZE) {
1557 kunmap_atomic(d_addr);
1558 d_page = get_next_page(d_page);
1560 d_addr = kmap_atomic(d_page);
1561 d_size = class->size - written;
1566 kunmap_atomic(d_addr);
1567 kunmap_atomic(s_addr);
1571 * Find alloced object in zspage from index object and
1574 static unsigned long find_alloced_obj(struct page *page, int index,
1575 struct size_class *class)
1579 unsigned long handle = 0;
1580 void *addr = kmap_atomic(page);
1582 if (!is_first_page(page))
1583 offset = page->index;
1584 offset += class->size * index;
1586 while (offset < PAGE_SIZE) {
1587 head = obj_to_head(class, page, addr + offset);
1588 if (head & OBJ_ALLOCATED_TAG) {
1589 handle = head & ~OBJ_ALLOCATED_TAG;
1590 if (trypin_tag(handle))
1595 offset += class->size;
1599 kunmap_atomic(addr);
1603 struct zs_compact_control {
1604 /* Source page for migration which could be a subpage of zspage. */
1605 struct page *s_page;
1606 /* Destination page for migration which should be a first page
1608 struct page *d_page;
1609 /* Starting object index within @s_page which used for live object
1610 * in the subpage. */
1614 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1615 struct zs_compact_control *cc)
1617 unsigned long used_obj, free_obj;
1618 unsigned long handle;
1619 struct page *s_page = cc->s_page;
1620 struct page *d_page = cc->d_page;
1621 unsigned long index = cc->index;
1625 handle = find_alloced_obj(s_page, index, class);
1627 s_page = get_next_page(s_page);
1634 /* Stop if there is no more space */
1635 if (zspage_full(d_page)) {
1641 used_obj = handle_to_obj(handle);
1642 free_obj = obj_malloc(d_page, class, handle);
1643 zs_object_copy(free_obj, used_obj, class);
1646 * record_obj updates handle's value to free_obj and it will
1647 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1648 * breaks synchronization using pin_tag(e,g, zs_free) so
1649 * let's keep the lock bit.
1651 free_obj |= BIT(HANDLE_PIN_BIT);
1652 record_obj(handle, free_obj);
1654 obj_free(pool, class, used_obj);
1657 /* Remember last position in this iteration */
1658 cc->s_page = s_page;
1664 static struct page *isolate_target_page(struct size_class *class)
1669 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1670 page = class->fullness_list[i];
1672 remove_zspage(page, class, i);
1681 * putback_zspage - add @first_page into right class's fullness list
1682 * @pool: target pool
1683 * @class: destination class
1684 * @first_page: target page
1686 * Return @fist_page's fullness_group
1688 static enum fullness_group putback_zspage(struct zs_pool *pool,
1689 struct size_class *class,
1690 struct page *first_page)
1692 enum fullness_group fullness;
1694 BUG_ON(!is_first_page(first_page));
1696 fullness = get_fullness_group(first_page);
1697 insert_zspage(first_page, class, fullness);
1698 set_zspage_mapping(first_page, class->index, fullness);
1700 if (fullness == ZS_EMPTY) {
1701 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1702 class->size, class->pages_per_zspage));
1703 atomic_long_sub(class->pages_per_zspage,
1704 &pool->pages_allocated);
1706 free_zspage(first_page);
1712 static struct page *isolate_source_page(struct size_class *class)
1715 struct page *page = NULL;
1717 for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) {
1718 page = class->fullness_list[i];
1722 remove_zspage(page, class, i);
1731 * Based on the number of unused allocated objects calculate
1732 * and return the number of pages that we can free.
1734 static unsigned long zs_can_compact(struct size_class *class)
1736 unsigned long obj_wasted;
1737 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
1738 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
1740 if (obj_allocated <= obj_used)
1743 obj_wasted = obj_allocated - obj_used;
1744 obj_wasted /= get_maxobj_per_zspage(class->size,
1745 class->pages_per_zspage);
1747 return obj_wasted * class->pages_per_zspage;
1750 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
1752 struct zs_compact_control cc;
1753 struct page *src_page;
1754 struct page *dst_page = NULL;
1756 spin_lock(&class->lock);
1757 while ((src_page = isolate_source_page(class))) {
1759 BUG_ON(!is_first_page(src_page));
1761 if (!zs_can_compact(class))
1765 cc.s_page = src_page;
1767 while ((dst_page = isolate_target_page(class))) {
1768 cc.d_page = dst_page;
1770 * If there is no more space in dst_page, resched
1771 * and see if anyone had allocated another zspage.
1773 if (!migrate_zspage(pool, class, &cc))
1776 putback_zspage(pool, class, dst_page);
1779 /* Stop if we couldn't find slot */
1780 if (dst_page == NULL)
1783 putback_zspage(pool, class, dst_page);
1784 if (putback_zspage(pool, class, src_page) == ZS_EMPTY)
1785 pool->stats.pages_compacted += class->pages_per_zspage;
1786 spin_unlock(&class->lock);
1788 spin_lock(&class->lock);
1792 putback_zspage(pool, class, src_page);
1794 spin_unlock(&class->lock);
1797 unsigned long zs_compact(struct zs_pool *pool)
1800 struct size_class *class;
1802 for (i = zs_size_classes - 1; i >= 0; i--) {
1803 class = pool->size_class[i];
1806 if (class->index != i)
1808 __zs_compact(pool, class);
1811 return pool->stats.pages_compacted;
1813 EXPORT_SYMBOL_GPL(zs_compact);
1815 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
1817 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
1819 EXPORT_SYMBOL_GPL(zs_pool_stats);
1821 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
1822 struct shrink_control *sc)
1824 unsigned long pages_freed;
1825 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1828 pages_freed = pool->stats.pages_compacted;
1830 * Compact classes and calculate compaction delta.
1831 * Can run concurrently with a manually triggered
1832 * (by user) compaction.
1834 pages_freed = zs_compact(pool) - pages_freed;
1836 return pages_freed ? pages_freed : SHRINK_STOP;
1839 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
1840 struct shrink_control *sc)
1843 struct size_class *class;
1844 unsigned long pages_to_free = 0;
1845 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1848 for (i = zs_size_classes - 1; i >= 0; i--) {
1849 class = pool->size_class[i];
1852 if (class->index != i)
1855 pages_to_free += zs_can_compact(class);
1858 return pages_to_free;
1861 static void zs_unregister_shrinker(struct zs_pool *pool)
1863 if (pool->shrinker_enabled) {
1864 unregister_shrinker(&pool->shrinker);
1865 pool->shrinker_enabled = false;
1869 static int zs_register_shrinker(struct zs_pool *pool)
1871 pool->shrinker.scan_objects = zs_shrinker_scan;
1872 pool->shrinker.count_objects = zs_shrinker_count;
1873 pool->shrinker.batch = 0;
1874 pool->shrinker.seeks = DEFAULT_SEEKS;
1876 return register_shrinker(&pool->shrinker);
1880 * zs_create_pool - Creates an allocation pool to work from.
1881 * @flags: allocation flags used to allocate pool metadata
1883 * This function must be called before anything when using
1884 * the zsmalloc allocator.
1886 * On success, a pointer to the newly created pool is returned,
1889 struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
1892 struct zs_pool *pool;
1893 struct size_class *prev_class = NULL;
1895 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1899 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1901 if (!pool->size_class) {
1906 pool->name = kstrdup(name, GFP_KERNEL);
1910 if (create_handle_cache(pool))
1914 * Iterate reversly, because, size of size_class that we want to use
1915 * for merging should be larger or equal to current size.
1917 for (i = zs_size_classes - 1; i >= 0; i--) {
1919 int pages_per_zspage;
1920 struct size_class *class;
1922 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1923 if (size > ZS_MAX_ALLOC_SIZE)
1924 size = ZS_MAX_ALLOC_SIZE;
1925 pages_per_zspage = get_pages_per_zspage(size);
1928 * size_class is used for normal zsmalloc operation such
1929 * as alloc/free for that size. Although it is natural that we
1930 * have one size_class for each size, there is a chance that we
1931 * can get more memory utilization if we use one size_class for
1932 * many different sizes whose size_class have same
1933 * characteristics. So, we makes size_class point to
1934 * previous size_class if possible.
1937 if (can_merge(prev_class, size, pages_per_zspage)) {
1938 pool->size_class[i] = prev_class;
1943 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1949 class->pages_per_zspage = pages_per_zspage;
1950 if (pages_per_zspage == 1 &&
1951 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1953 spin_lock_init(&class->lock);
1954 pool->size_class[i] = class;
1959 pool->flags = flags;
1961 if (zs_pool_stat_create(name, pool))
1965 * Not critical, we still can use the pool
1966 * and user can trigger compaction manually.
1968 if (zs_register_shrinker(pool) == 0)
1969 pool->shrinker_enabled = true;
1973 zs_destroy_pool(pool);
1976 EXPORT_SYMBOL_GPL(zs_create_pool);
1978 void zs_destroy_pool(struct zs_pool *pool)
1982 zs_unregister_shrinker(pool);
1983 zs_pool_stat_destroy(pool);
1985 for (i = 0; i < zs_size_classes; i++) {
1987 struct size_class *class = pool->size_class[i];
1992 if (class->index != i)
1995 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1996 if (class->fullness_list[fg]) {
1997 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2004 destroy_handle_cache(pool);
2005 kfree(pool->size_class);
2009 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2011 static int __init zs_init(void)
2013 int ret = zs_register_cpu_notifier();
2018 init_zs_size_classes();
2021 zpool_register_driver(&zs_zpool_driver);
2024 ret = zs_stat_init();
2026 pr_err("zs stat initialization failed\n");
2033 zpool_unregister_driver(&zs_zpool_driver);
2036 zs_unregister_cpu_notifier();
2041 static void __exit zs_exit(void)
2044 zpool_unregister_driver(&zs_zpool_driver);
2046 zs_unregister_cpu_notifier();
2051 module_init(zs_init);
2052 module_exit(zs_exit);
2054 MODULE_LICENSE("Dual BSD/GPL");
2055 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");