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>
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,
171 #ifdef CONFIG_ZSMALLOC_STAT
172 #define NR_ZS_STAT_TYPE (CLASS_ALMOST_EMPTY + 1)
174 #define NR_ZS_STAT_TYPE (OBJ_USED + 1)
177 struct zs_size_stat {
178 unsigned long objs[NR_ZS_STAT_TYPE];
181 #ifdef CONFIG_ZSMALLOC_STAT
182 static struct dentry *zs_stat_root;
186 * number of size_classes
188 static int zs_size_classes;
191 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
193 * n = number of allocated objects
194 * N = total number of objects zspage can store
195 * f = fullness_threshold_frac
197 * Similarly, we assign zspage to:
198 * ZS_ALMOST_FULL when n > N / f
199 * ZS_EMPTY when n == 0
200 * ZS_FULL when n == N
202 * (see: fix_fullness_group())
204 static const int fullness_threshold_frac = 4;
208 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
210 * Size of objects stored in this class. Must be multiple
216 struct zs_size_stat stats;
218 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
219 int pages_per_zspage;
220 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
225 * Placed within free objects to form a singly linked list.
226 * For every zspage, first_page->freelist gives head of this list.
228 * This must be power of 2 and less than or equal to ZS_ALIGN
233 * Position of next free chunk (encodes <PFN, obj_idx>)
234 * It's valid for non-allocated object
238 * Handle of allocated object.
240 unsigned long handle;
247 struct size_class **size_class;
248 struct kmem_cache *handle_cachep;
250 gfp_t flags; /* allocation flags used when growing pool */
251 atomic_long_t pages_allocated;
253 struct zs_pool_stats stats;
255 /* Compact classes */
256 struct shrinker shrinker;
258 * To signify that register_shrinker() was successful
259 * and unregister_shrinker() will not Oops.
261 bool shrinker_enabled;
262 #ifdef CONFIG_ZSMALLOC_STAT
263 struct dentry *stat_dentry;
268 * A zspage's class index and fullness group
269 * are encoded in its (first)page->mapping
271 #define CLASS_IDX_BITS 28
272 #define FULLNESS_BITS 4
273 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
274 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
276 struct mapping_area {
277 #ifdef CONFIG_PGTABLE_MAPPING
278 struct vm_struct *vm; /* vm area for mapping object that span pages */
280 char *vm_buf; /* copy buffer for objects that span pages */
282 char *vm_addr; /* address of kmap_atomic()'ed pages */
283 enum zs_mapmode vm_mm; /* mapping mode */
286 static int create_handle_cache(struct zs_pool *pool)
288 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
290 return pool->handle_cachep ? 0 : 1;
293 static void destroy_handle_cache(struct zs_pool *pool)
295 kmem_cache_destroy(pool->handle_cachep);
298 static unsigned long alloc_handle(struct zs_pool *pool)
300 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
301 pool->flags & ~__GFP_HIGHMEM);
304 static void free_handle(struct zs_pool *pool, unsigned long handle)
306 kmem_cache_free(pool->handle_cachep, (void *)handle);
309 static void record_obj(unsigned long handle, unsigned long obj)
312 * lsb of @obj represents handle lock while other bits
313 * represent object value the handle is pointing so
314 * updating shouldn't do store tearing.
316 WRITE_ONCE(*(unsigned long *)handle, obj);
323 static void *zs_zpool_create(const char *name, gfp_t gfp,
324 const struct zpool_ops *zpool_ops,
327 return zs_create_pool(name, gfp);
330 static void zs_zpool_destroy(void *pool)
332 zs_destroy_pool(pool);
335 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
336 unsigned long *handle)
338 *handle = zs_malloc(pool, size);
339 return *handle ? 0 : -1;
341 static void zs_zpool_free(void *pool, unsigned long handle)
343 zs_free(pool, handle);
346 static int zs_zpool_shrink(void *pool, unsigned int pages,
347 unsigned int *reclaimed)
352 static void *zs_zpool_map(void *pool, unsigned long handle,
353 enum zpool_mapmode mm)
355 enum zs_mapmode zs_mm;
364 case ZPOOL_MM_RW: /* fallthru */
370 return zs_map_object(pool, handle, zs_mm);
372 static void zs_zpool_unmap(void *pool, unsigned long handle)
374 zs_unmap_object(pool, handle);
377 static u64 zs_zpool_total_size(void *pool)
379 return zs_get_total_pages(pool) << PAGE_SHIFT;
382 static struct zpool_driver zs_zpool_driver = {
384 .owner = THIS_MODULE,
385 .create = zs_zpool_create,
386 .destroy = zs_zpool_destroy,
387 .malloc = zs_zpool_malloc,
388 .free = zs_zpool_free,
389 .shrink = zs_zpool_shrink,
391 .unmap = zs_zpool_unmap,
392 .total_size = zs_zpool_total_size,
395 MODULE_ALIAS("zpool-zsmalloc");
396 #endif /* CONFIG_ZPOOL */
398 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
400 return pages_per_zspage * PAGE_SIZE / size;
403 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
404 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
406 static int is_first_page(struct page *page)
408 return PagePrivate(page);
411 static int is_last_page(struct page *page)
413 return PagePrivate2(page);
416 static void get_zspage_mapping(struct page *first_page,
417 unsigned int *class_idx,
418 enum fullness_group *fullness)
421 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
423 m = (unsigned long)first_page->mapping;
424 *fullness = m & FULLNESS_MASK;
425 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
428 static void set_zspage_mapping(struct page *first_page,
429 unsigned int class_idx,
430 enum fullness_group fullness)
433 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
435 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
436 (fullness & FULLNESS_MASK);
437 first_page->mapping = (struct address_space *)m;
441 * zsmalloc divides the pool into various size classes where each
442 * class maintains a list of zspages where each zspage is divided
443 * into equal sized chunks. Each allocation falls into one of these
444 * classes depending on its size. This function returns index of the
445 * size class which has chunk size big enough to hold the give size.
447 static int get_size_class_index(int size)
451 if (likely(size > ZS_MIN_ALLOC_SIZE))
452 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
453 ZS_SIZE_CLASS_DELTA);
455 return min(zs_size_classes - 1, idx);
458 static inline void zs_stat_inc(struct size_class *class,
459 enum zs_stat_type type, unsigned long cnt)
461 if (type < NR_ZS_STAT_TYPE)
462 class->stats.objs[type] += cnt;
465 static inline void zs_stat_dec(struct size_class *class,
466 enum zs_stat_type type, unsigned long cnt)
468 if (type < NR_ZS_STAT_TYPE)
469 class->stats.objs[type] -= cnt;
472 static inline unsigned long zs_stat_get(struct size_class *class,
473 enum zs_stat_type type)
475 if (type < NR_ZS_STAT_TYPE)
476 return class->stats.objs[type];
480 #ifdef CONFIG_ZSMALLOC_STAT
482 static int __init zs_stat_init(void)
484 if (!debugfs_initialized())
487 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
494 static void __exit zs_stat_exit(void)
496 debugfs_remove_recursive(zs_stat_root);
499 static unsigned long zs_can_compact(struct size_class *class);
501 static int zs_stats_size_show(struct seq_file *s, void *v)
504 struct zs_pool *pool = s->private;
505 struct size_class *class;
507 unsigned long class_almost_full, class_almost_empty;
508 unsigned long obj_allocated, obj_used, pages_used, freeable;
509 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
510 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
511 unsigned long total_freeable = 0;
513 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
514 "class", "size", "almost_full", "almost_empty",
515 "obj_allocated", "obj_used", "pages_used",
516 "pages_per_zspage", "freeable");
518 for (i = 0; i < zs_size_classes; i++) {
519 class = pool->size_class[i];
521 if (class->index != i)
524 spin_lock(&class->lock);
525 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
526 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
527 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
528 obj_used = zs_stat_get(class, OBJ_USED);
529 freeable = zs_can_compact(class);
530 spin_unlock(&class->lock);
532 objs_per_zspage = get_maxobj_per_zspage(class->size,
533 class->pages_per_zspage);
534 pages_used = obj_allocated / objs_per_zspage *
535 class->pages_per_zspage;
537 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
538 " %10lu %10lu %16d %8lu\n",
539 i, class->size, class_almost_full, class_almost_empty,
540 obj_allocated, obj_used, pages_used,
541 class->pages_per_zspage, freeable);
543 total_class_almost_full += class_almost_full;
544 total_class_almost_empty += class_almost_empty;
545 total_objs += obj_allocated;
546 total_used_objs += obj_used;
547 total_pages += pages_used;
548 total_freeable += freeable;
552 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
553 "Total", "", total_class_almost_full,
554 total_class_almost_empty, total_objs,
555 total_used_objs, total_pages, "", total_freeable);
560 static int zs_stats_size_open(struct inode *inode, struct file *file)
562 return single_open(file, zs_stats_size_show, inode->i_private);
565 static const struct file_operations zs_stat_size_ops = {
566 .open = zs_stats_size_open,
569 .release = single_release,
572 static int zs_pool_stat_create(struct zs_pool *pool, const char *name)
574 struct dentry *entry;
579 entry = debugfs_create_dir(name, zs_stat_root);
581 pr_warn("debugfs dir <%s> creation failed\n", name);
584 pool->stat_dentry = entry;
586 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
587 pool->stat_dentry, pool, &zs_stat_size_ops);
589 pr_warn("%s: debugfs file entry <%s> creation failed\n",
597 static void zs_pool_stat_destroy(struct zs_pool *pool)
599 debugfs_remove_recursive(pool->stat_dentry);
602 #else /* CONFIG_ZSMALLOC_STAT */
603 static int __init zs_stat_init(void)
608 static void __exit zs_stat_exit(void)
612 static inline int zs_pool_stat_create(struct zs_pool *pool, const char *name)
617 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
624 * For each size class, zspages are divided into different groups
625 * depending on how "full" they are. This was done so that we could
626 * easily find empty or nearly empty zspages when we try to shrink
627 * the pool (not yet implemented). This function returns fullness
628 * status of the given page.
630 static enum fullness_group get_fullness_group(struct page *first_page)
632 int inuse, max_objects;
633 enum fullness_group fg;
635 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
637 inuse = first_page->inuse;
638 max_objects = first_page->objects;
642 else if (inuse == max_objects)
644 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
645 fg = ZS_ALMOST_EMPTY;
653 * Each size class maintains various freelists and zspages are assigned
654 * to one of these freelists based on the number of live objects they
655 * have. This functions inserts the given zspage into the freelist
656 * identified by <class, fullness_group>.
658 static void insert_zspage(struct size_class *class,
659 enum fullness_group fullness,
660 struct page *first_page)
664 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
666 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
669 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
670 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
672 head = &class->fullness_list[fullness];
679 * We want to see more ZS_FULL pages and less almost
680 * empty/full. Put pages with higher ->inuse first.
682 list_add_tail(&first_page->lru, &(*head)->lru);
683 if (first_page->inuse >= (*head)->inuse)
688 * This function removes the given zspage from the freelist identified
689 * by <class, fullness_group>.
691 static void remove_zspage(struct size_class *class,
692 enum fullness_group fullness,
693 struct page *first_page)
697 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
699 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
702 head = &class->fullness_list[fullness];
703 VM_BUG_ON_PAGE(!*head, first_page);
704 if (list_empty(&(*head)->lru))
706 else if (*head == first_page)
707 *head = (struct page *)list_entry((*head)->lru.next,
710 list_del_init(&first_page->lru);
711 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
712 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
716 * Each size class maintains zspages in different fullness groups depending
717 * on the number of live objects they contain. When allocating or freeing
718 * objects, the fullness status of the page can change, say, from ALMOST_FULL
719 * to ALMOST_EMPTY when freeing an object. This function checks if such
720 * a status change has occurred for the given page and accordingly moves the
721 * page from the freelist of the old fullness group to that of the new
724 static enum fullness_group fix_fullness_group(struct size_class *class,
725 struct page *first_page)
728 enum fullness_group currfg, newfg;
730 get_zspage_mapping(first_page, &class_idx, &currfg);
731 newfg = get_fullness_group(first_page);
735 remove_zspage(class, currfg, first_page);
736 insert_zspage(class, newfg, first_page);
737 set_zspage_mapping(first_page, class_idx, newfg);
744 * We have to decide on how many pages to link together
745 * to form a zspage for each size class. This is important
746 * to reduce wastage due to unusable space left at end of
747 * each zspage which is given as:
748 * wastage = Zp % class_size
749 * usage = Zp - wastage
750 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
752 * For example, for size class of 3/8 * PAGE_SIZE, we should
753 * link together 3 PAGE_SIZE sized pages to form a zspage
754 * since then we can perfectly fit in 8 such objects.
756 static int get_pages_per_zspage(int class_size)
758 int i, max_usedpc = 0;
759 /* zspage order which gives maximum used size per KB */
760 int max_usedpc_order = 1;
762 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
766 zspage_size = i * PAGE_SIZE;
767 waste = zspage_size % class_size;
768 usedpc = (zspage_size - waste) * 100 / zspage_size;
770 if (usedpc > max_usedpc) {
772 max_usedpc_order = i;
776 return max_usedpc_order;
780 * A single 'zspage' is composed of many system pages which are
781 * linked together using fields in struct page. This function finds
782 * the first/head page, given any component page of a zspage.
784 static struct page *get_first_page(struct page *page)
786 if (is_first_page(page))
789 return (struct page *)page_private(page);
792 static struct page *get_next_page(struct page *page)
796 if (is_last_page(page))
798 else if (is_first_page(page))
799 next = (struct page *)page_private(page);
801 next = list_entry(page->lru.next, struct page, lru);
807 * Encode <page, obj_idx> as a single handle value.
808 * We use the least bit of handle for tagging.
810 static void *location_to_obj(struct page *page, unsigned long obj_idx)
819 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
820 obj |= ((obj_idx) & OBJ_INDEX_MASK);
821 obj <<= OBJ_TAG_BITS;
827 * Decode <page, obj_idx> pair from the given object handle. We adjust the
828 * decoded obj_idx back to its original value since it was adjusted in
831 static void obj_to_location(unsigned long obj, struct page **page,
832 unsigned long *obj_idx)
834 obj >>= OBJ_TAG_BITS;
835 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
836 *obj_idx = (obj & OBJ_INDEX_MASK);
839 static unsigned long handle_to_obj(unsigned long handle)
841 return *(unsigned long *)handle;
844 static unsigned long obj_to_head(struct size_class *class, struct page *page,
848 VM_BUG_ON_PAGE(!is_first_page(page), page);
849 return page_private(page);
851 return *(unsigned long *)obj;
854 static unsigned long obj_idx_to_offset(struct page *page,
855 unsigned long obj_idx, int class_size)
857 unsigned long off = 0;
859 if (!is_first_page(page))
862 return off + obj_idx * class_size;
865 static inline int trypin_tag(unsigned long handle)
867 unsigned long *ptr = (unsigned long *)handle;
869 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
872 static void pin_tag(unsigned long handle)
874 while (!trypin_tag(handle));
877 static void unpin_tag(unsigned long handle)
879 unsigned long *ptr = (unsigned long *)handle;
881 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
884 static void reset_page(struct page *page)
886 clear_bit(PG_private, &page->flags);
887 clear_bit(PG_private_2, &page->flags);
888 set_page_private(page, 0);
889 page->mapping = NULL;
890 page->freelist = NULL;
891 page_mapcount_reset(page);
894 static void free_zspage(struct page *first_page)
896 struct page *nextp, *tmp, *head_extra;
898 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
899 VM_BUG_ON_PAGE(first_page->inuse, first_page);
901 head_extra = (struct page *)page_private(first_page);
903 reset_page(first_page);
904 __free_page(first_page);
906 /* zspage with only 1 system page */
910 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
911 list_del(&nextp->lru);
915 reset_page(head_extra);
916 __free_page(head_extra);
919 /* Initialize a newly allocated zspage */
920 static void init_zspage(struct size_class *class, struct page *first_page)
922 unsigned long off = 0;
923 struct page *page = first_page;
925 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
928 struct page *next_page;
929 struct link_free *link;
934 * page->index stores offset of first object starting
935 * in the page. For the first page, this is always 0,
936 * so we use first_page->index (aka ->freelist) to store
937 * head of corresponding zspage's freelist.
939 if (page != first_page)
942 vaddr = kmap_atomic(page);
943 link = (struct link_free *)vaddr + off / sizeof(*link);
945 while ((off += class->size) < PAGE_SIZE) {
946 link->next = location_to_obj(page, i++);
947 link += class->size / sizeof(*link);
951 * We now come to the last (full or partial) object on this
952 * page, which must point to the first object on the next
955 next_page = get_next_page(page);
956 link->next = location_to_obj(next_page, 0);
957 kunmap_atomic(vaddr);
964 * Allocate a zspage for the given size class
966 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
969 struct page *first_page = NULL, *uninitialized_var(prev_page);
972 * Allocate individual pages and link them together as:
973 * 1. first page->private = first sub-page
974 * 2. all sub-pages are linked together using page->lru
975 * 3. each sub-page is linked to the first page using page->private
977 * For each size class, First/Head pages are linked together using
978 * page->lru. Also, we set PG_private to identify the first page
979 * (i.e. no other sub-page has this flag set) and PG_private_2 to
980 * identify the last page.
983 for (i = 0; i < class->pages_per_zspage; i++) {
986 page = alloc_page(flags);
990 INIT_LIST_HEAD(&page->lru);
991 if (i == 0) { /* first page */
992 SetPagePrivate(page);
993 set_page_private(page, 0);
995 first_page->inuse = 0;
998 set_page_private(first_page, (unsigned long)page);
1000 set_page_private(page, (unsigned long)first_page);
1002 list_add(&page->lru, &prev_page->lru);
1003 if (i == class->pages_per_zspage - 1) /* last page */
1004 SetPagePrivate2(page);
1008 init_zspage(class, first_page);
1010 first_page->freelist = location_to_obj(first_page, 0);
1011 /* Maximum number of objects we can store in this zspage */
1012 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
1014 error = 0; /* Success */
1017 if (unlikely(error) && first_page) {
1018 free_zspage(first_page);
1025 static struct page *find_get_zspage(struct size_class *class)
1030 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1031 page = class->fullness_list[i];
1039 #ifdef CONFIG_PGTABLE_MAPPING
1040 static inline int __zs_cpu_up(struct mapping_area *area)
1043 * Make sure we don't leak memory if a cpu UP notification
1044 * and zs_init() race and both call zs_cpu_up() on the same cpu
1048 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1054 static inline void __zs_cpu_down(struct mapping_area *area)
1057 free_vm_area(area->vm);
1061 static inline void *__zs_map_object(struct mapping_area *area,
1062 struct page *pages[2], int off, int size)
1064 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1065 area->vm_addr = area->vm->addr;
1066 return area->vm_addr + off;
1069 static inline void __zs_unmap_object(struct mapping_area *area,
1070 struct page *pages[2], int off, int size)
1072 unsigned long addr = (unsigned long)area->vm_addr;
1074 unmap_kernel_range(addr, PAGE_SIZE * 2);
1077 #else /* CONFIG_PGTABLE_MAPPING */
1079 static inline int __zs_cpu_up(struct mapping_area *area)
1082 * Make sure we don't leak memory if a cpu UP notification
1083 * and zs_init() race and both call zs_cpu_up() on the same cpu
1087 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1093 static inline void __zs_cpu_down(struct mapping_area *area)
1095 kfree(area->vm_buf);
1096 area->vm_buf = NULL;
1099 static void *__zs_map_object(struct mapping_area *area,
1100 struct page *pages[2], int off, int size)
1104 char *buf = area->vm_buf;
1106 /* disable page faults to match kmap_atomic() return conditions */
1107 pagefault_disable();
1109 /* no read fastpath */
1110 if (area->vm_mm == ZS_MM_WO)
1113 sizes[0] = PAGE_SIZE - off;
1114 sizes[1] = size - sizes[0];
1116 /* copy object to per-cpu buffer */
1117 addr = kmap_atomic(pages[0]);
1118 memcpy(buf, addr + off, sizes[0]);
1119 kunmap_atomic(addr);
1120 addr = kmap_atomic(pages[1]);
1121 memcpy(buf + sizes[0], addr, sizes[1]);
1122 kunmap_atomic(addr);
1124 return area->vm_buf;
1127 static void __zs_unmap_object(struct mapping_area *area,
1128 struct page *pages[2], int off, int size)
1134 /* no write fastpath */
1135 if (area->vm_mm == ZS_MM_RO)
1139 buf = buf + ZS_HANDLE_SIZE;
1140 size -= ZS_HANDLE_SIZE;
1141 off += ZS_HANDLE_SIZE;
1143 sizes[0] = PAGE_SIZE - off;
1144 sizes[1] = size - sizes[0];
1146 /* copy per-cpu buffer to object */
1147 addr = kmap_atomic(pages[0]);
1148 memcpy(addr + off, buf, sizes[0]);
1149 kunmap_atomic(addr);
1150 addr = kmap_atomic(pages[1]);
1151 memcpy(addr, buf + sizes[0], sizes[1]);
1152 kunmap_atomic(addr);
1155 /* enable page faults to match kunmap_atomic() return conditions */
1159 #endif /* CONFIG_PGTABLE_MAPPING */
1161 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1164 int ret, cpu = (long)pcpu;
1165 struct mapping_area *area;
1168 case CPU_UP_PREPARE:
1169 area = &per_cpu(zs_map_area, cpu);
1170 ret = __zs_cpu_up(area);
1172 return notifier_from_errno(ret);
1175 case CPU_UP_CANCELED:
1176 area = &per_cpu(zs_map_area, cpu);
1177 __zs_cpu_down(area);
1184 static struct notifier_block zs_cpu_nb = {
1185 .notifier_call = zs_cpu_notifier
1188 static int zs_register_cpu_notifier(void)
1190 int cpu, uninitialized_var(ret);
1192 cpu_notifier_register_begin();
1194 __register_cpu_notifier(&zs_cpu_nb);
1195 for_each_online_cpu(cpu) {
1196 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1197 if (notifier_to_errno(ret))
1201 cpu_notifier_register_done();
1202 return notifier_to_errno(ret);
1205 static void zs_unregister_cpu_notifier(void)
1209 cpu_notifier_register_begin();
1211 for_each_online_cpu(cpu)
1212 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1213 __unregister_cpu_notifier(&zs_cpu_nb);
1215 cpu_notifier_register_done();
1218 static void init_zs_size_classes(void)
1222 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1223 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1226 zs_size_classes = nr;
1229 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1231 if (prev->pages_per_zspage != pages_per_zspage)
1234 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1235 != get_maxobj_per_zspage(size, pages_per_zspage))
1241 static bool zspage_full(struct page *first_page)
1243 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
1245 return first_page->inuse == first_page->objects;
1248 unsigned long zs_get_total_pages(struct zs_pool *pool)
1250 return atomic_long_read(&pool->pages_allocated);
1252 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1255 * zs_map_object - get address of allocated object from handle.
1256 * @pool: pool from which the object was allocated
1257 * @handle: handle returned from zs_malloc
1259 * Before using an object allocated from zs_malloc, it must be mapped using
1260 * this function. When done with the object, it must be unmapped using
1263 * Only one object can be mapped per cpu at a time. There is no protection
1264 * against nested mappings.
1266 * This function returns with preemption and page faults disabled.
1268 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1272 unsigned long obj, obj_idx, off;
1274 unsigned int class_idx;
1275 enum fullness_group fg;
1276 struct size_class *class;
1277 struct mapping_area *area;
1278 struct page *pages[2];
1282 * Because we use per-cpu mapping areas shared among the
1283 * pools/users, we can't allow mapping in interrupt context
1284 * because it can corrupt another users mappings.
1286 WARN_ON_ONCE(in_interrupt());
1288 /* From now on, migration cannot move the object */
1291 obj = handle_to_obj(handle);
1292 obj_to_location(obj, &page, &obj_idx);
1293 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1294 class = pool->size_class[class_idx];
1295 off = obj_idx_to_offset(page, obj_idx, class->size);
1297 area = &get_cpu_var(zs_map_area);
1299 if (off + class->size <= PAGE_SIZE) {
1300 /* this object is contained entirely within a page */
1301 area->vm_addr = kmap_atomic(page);
1302 ret = area->vm_addr + off;
1306 /* this object spans two pages */
1308 pages[1] = get_next_page(page);
1311 ret = __zs_map_object(area, pages, off, class->size);
1314 ret += ZS_HANDLE_SIZE;
1318 EXPORT_SYMBOL_GPL(zs_map_object);
1320 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1323 unsigned long obj, obj_idx, off;
1325 unsigned int class_idx;
1326 enum fullness_group fg;
1327 struct size_class *class;
1328 struct mapping_area *area;
1330 obj = handle_to_obj(handle);
1331 obj_to_location(obj, &page, &obj_idx);
1332 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1333 class = pool->size_class[class_idx];
1334 off = obj_idx_to_offset(page, obj_idx, class->size);
1336 area = this_cpu_ptr(&zs_map_area);
1337 if (off + class->size <= PAGE_SIZE)
1338 kunmap_atomic(area->vm_addr);
1340 struct page *pages[2];
1343 pages[1] = get_next_page(page);
1346 __zs_unmap_object(area, pages, off, class->size);
1348 put_cpu_var(zs_map_area);
1351 EXPORT_SYMBOL_GPL(zs_unmap_object);
1353 static unsigned long obj_malloc(struct size_class *class,
1354 struct page *first_page, unsigned long handle)
1357 struct link_free *link;
1359 struct page *m_page;
1360 unsigned long m_objidx, m_offset;
1363 handle |= OBJ_ALLOCATED_TAG;
1364 obj = (unsigned long)first_page->freelist;
1365 obj_to_location(obj, &m_page, &m_objidx);
1366 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1368 vaddr = kmap_atomic(m_page);
1369 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1370 first_page->freelist = link->next;
1372 /* record handle in the header of allocated chunk */
1373 link->handle = handle;
1375 /* record handle in first_page->private */
1376 set_page_private(first_page, handle);
1377 kunmap_atomic(vaddr);
1378 first_page->inuse++;
1379 zs_stat_inc(class, OBJ_USED, 1);
1386 * zs_malloc - Allocate block of given size from pool.
1387 * @pool: pool to allocate from
1388 * @size: size of block to allocate
1390 * On success, handle to the allocated object is returned,
1392 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1394 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1396 unsigned long handle, obj;
1397 struct size_class *class;
1398 struct page *first_page;
1400 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1403 handle = alloc_handle(pool);
1407 /* extra space in chunk to keep the handle */
1408 size += ZS_HANDLE_SIZE;
1409 class = pool->size_class[get_size_class_index(size)];
1411 spin_lock(&class->lock);
1412 first_page = find_get_zspage(class);
1415 spin_unlock(&class->lock);
1416 first_page = alloc_zspage(class, pool->flags);
1417 if (unlikely(!first_page)) {
1418 free_handle(pool, handle);
1422 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1423 atomic_long_add(class->pages_per_zspage,
1424 &pool->pages_allocated);
1426 spin_lock(&class->lock);
1427 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1428 class->size, class->pages_per_zspage));
1431 obj = obj_malloc(class, first_page, handle);
1432 /* Now move the zspage to another fullness group, if required */
1433 fix_fullness_group(class, first_page);
1434 record_obj(handle, obj);
1435 spin_unlock(&class->lock);
1439 EXPORT_SYMBOL_GPL(zs_malloc);
1441 static void obj_free(struct zs_pool *pool, struct size_class *class,
1444 struct link_free *link;
1445 struct page *first_page, *f_page;
1446 unsigned long f_objidx, f_offset;
1449 obj &= ~OBJ_ALLOCATED_TAG;
1450 obj_to_location(obj, &f_page, &f_objidx);
1451 first_page = get_first_page(f_page);
1453 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1455 vaddr = kmap_atomic(f_page);
1457 /* Insert this object in containing zspage's freelist */
1458 link = (struct link_free *)(vaddr + f_offset);
1459 link->next = first_page->freelist;
1461 set_page_private(first_page, 0);
1462 kunmap_atomic(vaddr);
1463 first_page->freelist = (void *)obj;
1464 first_page->inuse--;
1465 zs_stat_dec(class, OBJ_USED, 1);
1468 void zs_free(struct zs_pool *pool, unsigned long handle)
1470 struct page *first_page, *f_page;
1471 unsigned long obj, f_objidx;
1473 struct size_class *class;
1474 enum fullness_group fullness;
1476 if (unlikely(!handle))
1480 obj = handle_to_obj(handle);
1481 obj_to_location(obj, &f_page, &f_objidx);
1482 first_page = get_first_page(f_page);
1484 get_zspage_mapping(first_page, &class_idx, &fullness);
1485 class = pool->size_class[class_idx];
1487 spin_lock(&class->lock);
1488 obj_free(pool, class, obj);
1489 fullness = fix_fullness_group(class, first_page);
1490 if (fullness == ZS_EMPTY) {
1491 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1492 class->size, class->pages_per_zspage));
1493 atomic_long_sub(class->pages_per_zspage,
1494 &pool->pages_allocated);
1495 free_zspage(first_page);
1497 spin_unlock(&class->lock);
1500 free_handle(pool, handle);
1502 EXPORT_SYMBOL_GPL(zs_free);
1504 static void zs_object_copy(struct size_class *class, unsigned long dst,
1507 struct page *s_page, *d_page;
1508 unsigned long s_objidx, d_objidx;
1509 unsigned long s_off, d_off;
1510 void *s_addr, *d_addr;
1511 int s_size, d_size, size;
1514 s_size = d_size = class->size;
1516 obj_to_location(src, &s_page, &s_objidx);
1517 obj_to_location(dst, &d_page, &d_objidx);
1519 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1520 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1522 if (s_off + class->size > PAGE_SIZE)
1523 s_size = PAGE_SIZE - s_off;
1525 if (d_off + class->size > PAGE_SIZE)
1526 d_size = PAGE_SIZE - d_off;
1528 s_addr = kmap_atomic(s_page);
1529 d_addr = kmap_atomic(d_page);
1532 size = min(s_size, d_size);
1533 memcpy(d_addr + d_off, s_addr + s_off, size);
1536 if (written == class->size)
1544 if (s_off >= PAGE_SIZE) {
1545 kunmap_atomic(d_addr);
1546 kunmap_atomic(s_addr);
1547 s_page = get_next_page(s_page);
1548 s_addr = kmap_atomic(s_page);
1549 d_addr = kmap_atomic(d_page);
1550 s_size = class->size - written;
1554 if (d_off >= PAGE_SIZE) {
1555 kunmap_atomic(d_addr);
1556 d_page = get_next_page(d_page);
1557 d_addr = kmap_atomic(d_page);
1558 d_size = class->size - written;
1563 kunmap_atomic(d_addr);
1564 kunmap_atomic(s_addr);
1568 * Find alloced object in zspage from index object and
1571 static unsigned long find_alloced_obj(struct size_class *class,
1572 struct page *page, int index)
1576 unsigned long handle = 0;
1577 void *addr = kmap_atomic(page);
1579 if (!is_first_page(page))
1580 offset = page->index;
1581 offset += class->size * index;
1583 while (offset < PAGE_SIZE) {
1584 head = obj_to_head(class, page, addr + offset);
1585 if (head & OBJ_ALLOCATED_TAG) {
1586 handle = head & ~OBJ_ALLOCATED_TAG;
1587 if (trypin_tag(handle))
1592 offset += class->size;
1596 kunmap_atomic(addr);
1600 struct zs_compact_control {
1601 /* Source page for migration which could be a subpage of zspage. */
1602 struct page *s_page;
1603 /* Destination page for migration which should be a first page
1605 struct page *d_page;
1606 /* Starting object index within @s_page which used for live object
1607 * in the subpage. */
1611 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1612 struct zs_compact_control *cc)
1614 unsigned long used_obj, free_obj;
1615 unsigned long handle;
1616 struct page *s_page = cc->s_page;
1617 struct page *d_page = cc->d_page;
1618 unsigned long index = cc->index;
1622 handle = find_alloced_obj(class, s_page, index);
1624 s_page = get_next_page(s_page);
1631 /* Stop if there is no more space */
1632 if (zspage_full(d_page)) {
1638 used_obj = handle_to_obj(handle);
1639 free_obj = obj_malloc(class, d_page, handle);
1640 zs_object_copy(class, free_obj, used_obj);
1643 * record_obj updates handle's value to free_obj and it will
1644 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1645 * breaks synchronization using pin_tag(e,g, zs_free) so
1646 * let's keep the lock bit.
1648 free_obj |= BIT(HANDLE_PIN_BIT);
1649 record_obj(handle, free_obj);
1651 obj_free(pool, class, used_obj);
1654 /* Remember last position in this iteration */
1655 cc->s_page = s_page;
1661 static struct page *isolate_target_page(struct size_class *class)
1666 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1667 page = class->fullness_list[i];
1669 remove_zspage(class, i, page);
1678 * putback_zspage - add @first_page into right class's fullness list
1679 * @pool: target pool
1680 * @class: destination class
1681 * @first_page: target page
1683 * Return @fist_page's fullness_group
1685 static enum fullness_group putback_zspage(struct zs_pool *pool,
1686 struct size_class *class,
1687 struct page *first_page)
1689 enum fullness_group fullness;
1691 fullness = get_fullness_group(first_page);
1692 insert_zspage(class, fullness, first_page);
1693 set_zspage_mapping(first_page, class->index, fullness);
1695 if (fullness == ZS_EMPTY) {
1696 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1697 class->size, class->pages_per_zspage));
1698 atomic_long_sub(class->pages_per_zspage,
1699 &pool->pages_allocated);
1701 free_zspage(first_page);
1707 static struct page *isolate_source_page(struct size_class *class)
1710 struct page *page = NULL;
1712 for (i = ZS_ALMOST_EMPTY; i >= ZS_ALMOST_FULL; i--) {
1713 page = class->fullness_list[i];
1717 remove_zspage(class, i, page);
1726 * Based on the number of unused allocated objects calculate
1727 * and return the number of pages that we can free.
1729 static unsigned long zs_can_compact(struct size_class *class)
1731 unsigned long obj_wasted;
1732 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
1733 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
1735 if (obj_allocated <= obj_used)
1738 obj_wasted = obj_allocated - obj_used;
1739 obj_wasted /= get_maxobj_per_zspage(class->size,
1740 class->pages_per_zspage);
1742 return obj_wasted * class->pages_per_zspage;
1745 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
1747 struct zs_compact_control cc;
1748 struct page *src_page;
1749 struct page *dst_page = NULL;
1751 spin_lock(&class->lock);
1752 while ((src_page = isolate_source_page(class))) {
1754 if (!zs_can_compact(class))
1758 cc.s_page = src_page;
1760 while ((dst_page = isolate_target_page(class))) {
1761 cc.d_page = dst_page;
1763 * If there is no more space in dst_page, resched
1764 * and see if anyone had allocated another zspage.
1766 if (!migrate_zspage(pool, class, &cc))
1769 putback_zspage(pool, class, dst_page);
1772 /* Stop if we couldn't find slot */
1773 if (dst_page == NULL)
1776 putback_zspage(pool, class, dst_page);
1777 if (putback_zspage(pool, class, src_page) == ZS_EMPTY)
1778 pool->stats.pages_compacted += class->pages_per_zspage;
1779 spin_unlock(&class->lock);
1781 spin_lock(&class->lock);
1785 putback_zspage(pool, class, src_page);
1787 spin_unlock(&class->lock);
1790 unsigned long zs_compact(struct zs_pool *pool)
1793 struct size_class *class;
1795 for (i = zs_size_classes - 1; i >= 0; i--) {
1796 class = pool->size_class[i];
1799 if (class->index != i)
1801 __zs_compact(pool, class);
1804 return pool->stats.pages_compacted;
1806 EXPORT_SYMBOL_GPL(zs_compact);
1808 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
1810 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
1812 EXPORT_SYMBOL_GPL(zs_pool_stats);
1814 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
1815 struct shrink_control *sc)
1817 unsigned long pages_freed;
1818 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1821 pages_freed = pool->stats.pages_compacted;
1823 * Compact classes and calculate compaction delta.
1824 * Can run concurrently with a manually triggered
1825 * (by user) compaction.
1827 pages_freed = zs_compact(pool) - pages_freed;
1829 return pages_freed ? pages_freed : SHRINK_STOP;
1832 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
1833 struct shrink_control *sc)
1836 struct size_class *class;
1837 unsigned long pages_to_free = 0;
1838 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
1841 for (i = zs_size_classes - 1; i >= 0; i--) {
1842 class = pool->size_class[i];
1845 if (class->index != i)
1848 pages_to_free += zs_can_compact(class);
1851 return pages_to_free;
1854 static void zs_unregister_shrinker(struct zs_pool *pool)
1856 if (pool->shrinker_enabled) {
1857 unregister_shrinker(&pool->shrinker);
1858 pool->shrinker_enabled = false;
1862 static int zs_register_shrinker(struct zs_pool *pool)
1864 pool->shrinker.scan_objects = zs_shrinker_scan;
1865 pool->shrinker.count_objects = zs_shrinker_count;
1866 pool->shrinker.batch = 0;
1867 pool->shrinker.seeks = DEFAULT_SEEKS;
1869 return register_shrinker(&pool->shrinker);
1873 * zs_create_pool - Creates an allocation pool to work from.
1874 * @flags: allocation flags used to allocate pool metadata
1876 * This function must be called before anything when using
1877 * the zsmalloc allocator.
1879 * On success, a pointer to the newly created pool is returned,
1882 struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
1885 struct zs_pool *pool;
1886 struct size_class *prev_class = NULL;
1888 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1892 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1894 if (!pool->size_class) {
1899 pool->name = kstrdup(name, GFP_KERNEL);
1903 if (create_handle_cache(pool))
1907 * Iterate reversly, because, size of size_class that we want to use
1908 * for merging should be larger or equal to current size.
1910 for (i = zs_size_classes - 1; i >= 0; i--) {
1912 int pages_per_zspage;
1913 struct size_class *class;
1915 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1916 if (size > ZS_MAX_ALLOC_SIZE)
1917 size = ZS_MAX_ALLOC_SIZE;
1918 pages_per_zspage = get_pages_per_zspage(size);
1921 * size_class is used for normal zsmalloc operation such
1922 * as alloc/free for that size. Although it is natural that we
1923 * have one size_class for each size, there is a chance that we
1924 * can get more memory utilization if we use one size_class for
1925 * many different sizes whose size_class have same
1926 * characteristics. So, we makes size_class point to
1927 * previous size_class if possible.
1930 if (can_merge(prev_class, size, pages_per_zspage)) {
1931 pool->size_class[i] = prev_class;
1936 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1942 class->pages_per_zspage = pages_per_zspage;
1943 if (pages_per_zspage == 1 &&
1944 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1946 spin_lock_init(&class->lock);
1947 pool->size_class[i] = class;
1952 pool->flags = flags;
1954 if (zs_pool_stat_create(pool, name))
1958 * Not critical, we still can use the pool
1959 * and user can trigger compaction manually.
1961 if (zs_register_shrinker(pool) == 0)
1962 pool->shrinker_enabled = true;
1966 zs_destroy_pool(pool);
1969 EXPORT_SYMBOL_GPL(zs_create_pool);
1971 void zs_destroy_pool(struct zs_pool *pool)
1975 zs_unregister_shrinker(pool);
1976 zs_pool_stat_destroy(pool);
1978 for (i = 0; i < zs_size_classes; i++) {
1980 struct size_class *class = pool->size_class[i];
1985 if (class->index != i)
1988 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1989 if (class->fullness_list[fg]) {
1990 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1997 destroy_handle_cache(pool);
1998 kfree(pool->size_class);
2002 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2004 static int __init zs_init(void)
2006 int ret = zs_register_cpu_notifier();
2011 init_zs_size_classes();
2014 zpool_register_driver(&zs_zpool_driver);
2017 ret = zs_stat_init();
2019 pr_err("zs stat initialization failed\n");
2026 zpool_unregister_driver(&zs_zpool_driver);
2029 zs_unregister_cpu_notifier();
2034 static void __exit zs_exit(void)
2037 zpool_unregister_driver(&zs_zpool_driver);
2039 zs_unregister_cpu_notifier();
2044 module_init(zs_init);
2045 module_exit(zs_exit);
2047 MODULE_LICENSE("Dual BSD/GPL");
2048 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");