2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->first_page: points to the first component (0-order) page
20 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
27 * For _first_ page only:
29 * page->private (union with page->first_page): refers to the
30 * component page after the first page
31 * If the page is first_page for huge object, it stores handle.
32 * Look at size_class->huge.
33 * page->freelist: points to the first free object in zspage.
34 * Free objects are linked together using in-place
36 * page->objects: maximum number of objects we can store in this
37 * zspage (class->zspage_order * PAGE_SIZE / class->size)
38 * page->lru: links together first pages of various zspages.
39 * Basically forming list of zspages in a fullness group.
40 * page->mapping: class index and fullness group of the zspage
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
48 #include <linux/module.h>
49 #include <linux/kernel.h>
50 #include <linux/sched.h>
51 #include <linux/bitops.h>
52 #include <linux/errno.h>
53 #include <linux/highmem.h>
54 #include <linux/string.h>
55 #include <linux/slab.h>
56 #include <asm/tlbflush.h>
57 #include <asm/pgtable.h>
58 #include <linux/cpumask.h>
59 #include <linux/cpu.h>
60 #include <linux/vmalloc.h>
61 #include <linux/hardirq.h>
62 #include <linux/spinlock.h>
63 #include <linux/types.h>
64 #include <linux/debugfs.h>
65 #include <linux/zsmalloc.h>
66 #include <linux/zpool.h>
69 * This must be power of 2 and greater than of equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
77 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
78 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
80 #define ZS_MAX_ZSPAGE_ORDER 2
81 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
83 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
86 * Object location (<PFN>, <obj_idx>) is encoded as
87 * as single (unsigned long) handle value.
89 * Note that object index <obj_idx> is relative to system
90 * page <PFN> it is stored in, so for each sub-page belonging
91 * to a zspage, obj_idx starts with 0.
93 * This is made more complicated by various memory models and PAE.
96 #ifndef MAX_PHYSMEM_BITS
97 #ifdef CONFIG_HIGHMEM64G
98 #define MAX_PHYSMEM_BITS 36
99 #else /* !CONFIG_HIGHMEM64G */
101 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
104 #define MAX_PHYSMEM_BITS BITS_PER_LONG
107 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
110 * Memory for allocating for handle keeps object position by
111 * encoding <page, obj_idx> and the encoded value has a room
112 * in least bit(ie, look at obj_to_location).
113 * We use the bit to synchronize between object access by
114 * user and migration.
116 #define HANDLE_PIN_BIT 0
119 * Head in allocated object should have OBJ_ALLOCATED_TAG
120 * to identify the object was allocated or not.
121 * It's okay to add the status bit in the least bit because
122 * header keeps handle which is 4byte-aligned address so we
123 * have room for two bit at least.
125 #define OBJ_ALLOCATED_TAG 1
126 #define OBJ_TAG_BITS 1
127 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
128 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
130 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
131 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132 #define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
134 /* each chunk includes extra space to keep handle */
135 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
138 * On systems with 4K page size, this gives 255 size classes! There is a
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
150 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
153 * We do not maintain any list for completely empty or full pages
155 enum fullness_group {
158 _ZS_NR_FULLNESS_GROUPS,
172 struct zs_size_stat {
173 unsigned long objs[NR_ZS_STAT_TYPE];
176 #ifdef CONFIG_ZSMALLOC_STAT
177 static struct dentry *zs_stat_root;
181 * number of size_classes
183 static int zs_size_classes;
186 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
188 * n = number of allocated objects
189 * N = total number of objects zspage can store
190 * f = fullness_threshold_frac
192 * Similarly, we assign zspage to:
193 * ZS_ALMOST_FULL when n > N / f
194 * ZS_EMPTY when n == 0
195 * ZS_FULL when n == N
197 * (see: fix_fullness_group())
199 static const int fullness_threshold_frac = 4;
203 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
205 * Size of objects stored in this class. Must be multiple
211 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
212 int pages_per_zspage;
213 struct zs_size_stat stats;
215 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
220 * Placed within free objects to form a singly linked list.
221 * For every zspage, first_page->freelist gives head of this list.
223 * This must be power of 2 and less than or equal to ZS_ALIGN
228 * Position of next free chunk (encodes <PFN, obj_idx>)
229 * It's valid for non-allocated object
233 * Handle of allocated object.
235 unsigned long handle;
242 struct size_class **size_class;
243 struct kmem_cache *handle_cachep;
245 gfp_t flags; /* allocation flags used when growing pool */
246 atomic_long_t pages_allocated;
248 #ifdef CONFIG_ZSMALLOC_STAT
249 struct dentry *stat_dentry;
254 * A zspage's class index and fullness group
255 * are encoded in its (first)page->mapping
257 #define CLASS_IDX_BITS 28
258 #define FULLNESS_BITS 4
259 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
260 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
262 struct mapping_area {
263 #ifdef CONFIG_PGTABLE_MAPPING
264 struct vm_struct *vm; /* vm area for mapping object that span pages */
266 char *vm_buf; /* copy buffer for objects that span pages */
268 char *vm_addr; /* address of kmap_atomic()'ed pages */
269 enum zs_mapmode vm_mm; /* mapping mode */
273 static int create_handle_cache(struct zs_pool *pool)
275 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
277 return pool->handle_cachep ? 0 : 1;
280 static void destroy_handle_cache(struct zs_pool *pool)
282 if (pool->handle_cachep)
283 kmem_cache_destroy(pool->handle_cachep);
286 static unsigned long alloc_handle(struct zs_pool *pool)
288 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
289 pool->flags & ~__GFP_HIGHMEM);
292 static void free_handle(struct zs_pool *pool, unsigned long handle)
294 kmem_cache_free(pool->handle_cachep, (void *)handle);
297 static void record_obj(unsigned long handle, unsigned long obj)
299 *(unsigned long *)handle = obj;
306 static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops,
309 return zs_create_pool(name, gfp);
312 static void zs_zpool_destroy(void *pool)
314 zs_destroy_pool(pool);
317 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
318 unsigned long *handle)
320 *handle = zs_malloc(pool, size);
321 return *handle ? 0 : -1;
323 static void zs_zpool_free(void *pool, unsigned long handle)
325 zs_free(pool, handle);
328 static int zs_zpool_shrink(void *pool, unsigned int pages,
329 unsigned int *reclaimed)
334 static void *zs_zpool_map(void *pool, unsigned long handle,
335 enum zpool_mapmode mm)
337 enum zs_mapmode zs_mm;
346 case ZPOOL_MM_RW: /* fallthru */
352 return zs_map_object(pool, handle, zs_mm);
354 static void zs_zpool_unmap(void *pool, unsigned long handle)
356 zs_unmap_object(pool, handle);
359 static u64 zs_zpool_total_size(void *pool)
361 return zs_get_total_pages(pool) << PAGE_SHIFT;
364 static struct zpool_driver zs_zpool_driver = {
366 .owner = THIS_MODULE,
367 .create = zs_zpool_create,
368 .destroy = zs_zpool_destroy,
369 .malloc = zs_zpool_malloc,
370 .free = zs_zpool_free,
371 .shrink = zs_zpool_shrink,
373 .unmap = zs_zpool_unmap,
374 .total_size = zs_zpool_total_size,
377 MODULE_ALIAS("zpool-zsmalloc");
378 #endif /* CONFIG_ZPOOL */
380 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
382 return pages_per_zspage * PAGE_SIZE / size;
385 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
386 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
388 static int is_first_page(struct page *page)
390 return PagePrivate(page);
393 static int is_last_page(struct page *page)
395 return PagePrivate2(page);
398 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
399 enum fullness_group *fullness)
402 BUG_ON(!is_first_page(page));
404 m = (unsigned long)page->mapping;
405 *fullness = m & FULLNESS_MASK;
406 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
409 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
410 enum fullness_group fullness)
413 BUG_ON(!is_first_page(page));
415 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
416 (fullness & FULLNESS_MASK);
417 page->mapping = (struct address_space *)m;
421 * zsmalloc divides the pool into various size classes where each
422 * class maintains a list of zspages where each zspage is divided
423 * into equal sized chunks. Each allocation falls into one of these
424 * classes depending on its size. This function returns index of the
425 * size class which has chunk size big enough to hold the give size.
427 static int get_size_class_index(int size)
431 if (likely(size > ZS_MIN_ALLOC_SIZE))
432 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
433 ZS_SIZE_CLASS_DELTA);
435 return min(zs_size_classes - 1, idx);
438 static inline void zs_stat_inc(struct size_class *class,
439 enum zs_stat_type type, unsigned long cnt)
441 class->stats.objs[type] += cnt;
444 static inline void zs_stat_dec(struct size_class *class,
445 enum zs_stat_type type, unsigned long cnt)
447 class->stats.objs[type] -= cnt;
450 static inline unsigned long zs_stat_get(struct size_class *class,
451 enum zs_stat_type type)
453 return class->stats.objs[type];
456 #ifdef CONFIG_ZSMALLOC_STAT
458 static int __init zs_stat_init(void)
460 if (!debugfs_initialized())
463 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
470 static void __exit zs_stat_exit(void)
472 debugfs_remove_recursive(zs_stat_root);
475 static int zs_stats_size_show(struct seq_file *s, void *v)
478 struct zs_pool *pool = s->private;
479 struct size_class *class;
481 unsigned long class_almost_full, class_almost_empty;
482 unsigned long obj_allocated, obj_used, pages_used;
483 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
484 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
486 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
487 "class", "size", "almost_full", "almost_empty",
488 "obj_allocated", "obj_used", "pages_used",
491 for (i = 0; i < zs_size_classes; i++) {
492 class = pool->size_class[i];
494 if (class->index != i)
497 spin_lock(&class->lock);
498 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
499 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
500 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
501 obj_used = zs_stat_get(class, OBJ_USED);
502 spin_unlock(&class->lock);
504 objs_per_zspage = get_maxobj_per_zspage(class->size,
505 class->pages_per_zspage);
506 pages_used = obj_allocated / objs_per_zspage *
507 class->pages_per_zspage;
509 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
510 i, class->size, class_almost_full, class_almost_empty,
511 obj_allocated, obj_used, pages_used,
512 class->pages_per_zspage);
514 total_class_almost_full += class_almost_full;
515 total_class_almost_empty += class_almost_empty;
516 total_objs += obj_allocated;
517 total_used_objs += obj_used;
518 total_pages += pages_used;
522 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
523 "Total", "", total_class_almost_full,
524 total_class_almost_empty, total_objs,
525 total_used_objs, total_pages);
530 static int zs_stats_size_open(struct inode *inode, struct file *file)
532 return single_open(file, zs_stats_size_show, inode->i_private);
535 static const struct file_operations zs_stat_size_ops = {
536 .open = zs_stats_size_open,
539 .release = single_release,
542 static int zs_pool_stat_create(char *name, struct zs_pool *pool)
544 struct dentry *entry;
549 entry = debugfs_create_dir(name, zs_stat_root);
551 pr_warn("debugfs dir <%s> creation failed\n", name);
554 pool->stat_dentry = entry;
556 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
557 pool->stat_dentry, pool, &zs_stat_size_ops);
559 pr_warn("%s: debugfs file entry <%s> creation failed\n",
567 static void zs_pool_stat_destroy(struct zs_pool *pool)
569 debugfs_remove_recursive(pool->stat_dentry);
572 #else /* CONFIG_ZSMALLOC_STAT */
573 static int __init zs_stat_init(void)
578 static void __exit zs_stat_exit(void)
582 static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
587 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
594 * For each size class, zspages are divided into different groups
595 * depending on how "full" they are. This was done so that we could
596 * easily find empty or nearly empty zspages when we try to shrink
597 * the pool (not yet implemented). This function returns fullness
598 * status of the given page.
600 static enum fullness_group get_fullness_group(struct page *page)
602 int inuse, max_objects;
603 enum fullness_group fg;
604 BUG_ON(!is_first_page(page));
607 max_objects = page->objects;
611 else if (inuse == max_objects)
613 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
614 fg = ZS_ALMOST_EMPTY;
622 * Each size class maintains various freelists and zspages are assigned
623 * to one of these freelists based on the number of live objects they
624 * have. This functions inserts the given zspage into the freelist
625 * identified by <class, fullness_group>.
627 static void insert_zspage(struct page *page, struct size_class *class,
628 enum fullness_group fullness)
632 BUG_ON(!is_first_page(page));
634 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
637 head = &class->fullness_list[fullness];
639 list_add_tail(&page->lru, &(*head)->lru);
642 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
643 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
647 * This function removes the given zspage from the freelist identified
648 * by <class, fullness_group>.
650 static void remove_zspage(struct page *page, struct size_class *class,
651 enum fullness_group fullness)
655 BUG_ON(!is_first_page(page));
657 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
660 head = &class->fullness_list[fullness];
662 if (list_empty(&(*head)->lru))
664 else if (*head == page)
665 *head = (struct page *)list_entry((*head)->lru.next,
668 list_del_init(&page->lru);
669 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
670 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
674 * Each size class maintains zspages in different fullness groups depending
675 * on the number of live objects they contain. When allocating or freeing
676 * objects, the fullness status of the page can change, say, from ALMOST_FULL
677 * to ALMOST_EMPTY when freeing an object. This function checks if such
678 * a status change has occurred for the given page and accordingly moves the
679 * page from the freelist of the old fullness group to that of the new
682 static enum fullness_group fix_fullness_group(struct size_class *class,
686 enum fullness_group currfg, newfg;
688 BUG_ON(!is_first_page(page));
690 get_zspage_mapping(page, &class_idx, &currfg);
691 newfg = get_fullness_group(page);
695 remove_zspage(page, class, currfg);
696 insert_zspage(page, class, newfg);
697 set_zspage_mapping(page, class_idx, newfg);
704 * We have to decide on how many pages to link together
705 * to form a zspage for each size class. This is important
706 * to reduce wastage due to unusable space left at end of
707 * each zspage which is given as:
708 * wastage = Zp % class_size
709 * usage = Zp - wastage
710 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
712 * For example, for size class of 3/8 * PAGE_SIZE, we should
713 * link together 3 PAGE_SIZE sized pages to form a zspage
714 * since then we can perfectly fit in 8 such objects.
716 static int get_pages_per_zspage(int class_size)
718 int i, max_usedpc = 0;
719 /* zspage order which gives maximum used size per KB */
720 int max_usedpc_order = 1;
722 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
726 zspage_size = i * PAGE_SIZE;
727 waste = zspage_size % class_size;
728 usedpc = (zspage_size - waste) * 100 / zspage_size;
730 if (usedpc > max_usedpc) {
732 max_usedpc_order = i;
736 return max_usedpc_order;
740 * A single 'zspage' is composed of many system pages which are
741 * linked together using fields in struct page. This function finds
742 * the first/head page, given any component page of a zspage.
744 static struct page *get_first_page(struct page *page)
746 if (is_first_page(page))
749 return page->first_page;
752 static struct page *get_next_page(struct page *page)
756 if (is_last_page(page))
758 else if (is_first_page(page))
759 next = (struct page *)page_private(page);
761 next = list_entry(page->lru.next, struct page, lru);
767 * Encode <page, obj_idx> as a single handle value.
768 * We use the least bit of handle for tagging.
770 static void *location_to_obj(struct page *page, unsigned long obj_idx)
779 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
780 obj |= ((obj_idx) & OBJ_INDEX_MASK);
781 obj <<= OBJ_TAG_BITS;
787 * Decode <page, obj_idx> pair from the given object handle. We adjust the
788 * decoded obj_idx back to its original value since it was adjusted in
791 static void obj_to_location(unsigned long obj, struct page **page,
792 unsigned long *obj_idx)
794 obj >>= OBJ_TAG_BITS;
795 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
796 *obj_idx = (obj & OBJ_INDEX_MASK);
799 static unsigned long handle_to_obj(unsigned long handle)
801 return *(unsigned long *)handle;
804 static unsigned long obj_to_head(struct size_class *class, struct page *page,
808 VM_BUG_ON(!is_first_page(page));
809 return *(unsigned long *)page_private(page);
811 return *(unsigned long *)obj;
814 static unsigned long obj_idx_to_offset(struct page *page,
815 unsigned long obj_idx, int class_size)
817 unsigned long off = 0;
819 if (!is_first_page(page))
822 return off + obj_idx * class_size;
825 static inline int trypin_tag(unsigned long handle)
827 unsigned long *ptr = (unsigned long *)handle;
829 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
832 static void pin_tag(unsigned long handle)
834 while (!trypin_tag(handle));
837 static void unpin_tag(unsigned long handle)
839 unsigned long *ptr = (unsigned long *)handle;
841 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
844 static void reset_page(struct page *page)
846 clear_bit(PG_private, &page->flags);
847 clear_bit(PG_private_2, &page->flags);
848 set_page_private(page, 0);
849 page->mapping = NULL;
850 page->freelist = NULL;
851 page_mapcount_reset(page);
854 static void free_zspage(struct page *first_page)
856 struct page *nextp, *tmp, *head_extra;
858 BUG_ON(!is_first_page(first_page));
859 BUG_ON(first_page->inuse);
861 head_extra = (struct page *)page_private(first_page);
863 reset_page(first_page);
864 __free_page(first_page);
866 /* zspage with only 1 system page */
870 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
871 list_del(&nextp->lru);
875 reset_page(head_extra);
876 __free_page(head_extra);
879 /* Initialize a newly allocated zspage */
880 static void init_zspage(struct page *first_page, struct size_class *class)
882 unsigned long off = 0;
883 struct page *page = first_page;
885 BUG_ON(!is_first_page(first_page));
887 struct page *next_page;
888 struct link_free *link;
893 * page->index stores offset of first object starting
894 * in the page. For the first page, this is always 0,
895 * so we use first_page->index (aka ->freelist) to store
896 * head of corresponding zspage's freelist.
898 if (page != first_page)
901 vaddr = kmap_atomic(page);
902 link = (struct link_free *)vaddr + off / sizeof(*link);
904 while ((off += class->size) < PAGE_SIZE) {
905 link->next = location_to_obj(page, i++);
906 link += class->size / sizeof(*link);
910 * We now come to the last (full or partial) object on this
911 * page, which must point to the first object on the next
914 next_page = get_next_page(page);
915 link->next = location_to_obj(next_page, 0);
916 kunmap_atomic(vaddr);
923 * Allocate a zspage for the given size class
925 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
928 struct page *first_page = NULL, *uninitialized_var(prev_page);
931 * Allocate individual pages and link them together as:
932 * 1. first page->private = first sub-page
933 * 2. all sub-pages are linked together using page->lru
934 * 3. each sub-page is linked to the first page using page->first_page
936 * For each size class, First/Head pages are linked together using
937 * page->lru. Also, we set PG_private to identify the first page
938 * (i.e. no other sub-page has this flag set) and PG_private_2 to
939 * identify the last page.
942 for (i = 0; i < class->pages_per_zspage; i++) {
945 page = alloc_page(flags);
949 INIT_LIST_HEAD(&page->lru);
950 if (i == 0) { /* first page */
951 SetPagePrivate(page);
952 set_page_private(page, 0);
954 first_page->inuse = 0;
957 set_page_private(first_page, (unsigned long)page);
959 page->first_page = first_page;
961 list_add(&page->lru, &prev_page->lru);
962 if (i == class->pages_per_zspage - 1) /* last page */
963 SetPagePrivate2(page);
967 init_zspage(first_page, class);
969 first_page->freelist = location_to_obj(first_page, 0);
970 /* Maximum number of objects we can store in this zspage */
971 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
973 error = 0; /* Success */
976 if (unlikely(error) && first_page) {
977 free_zspage(first_page);
984 static struct page *find_get_zspage(struct size_class *class)
989 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
990 page = class->fullness_list[i];
998 #ifdef CONFIG_PGTABLE_MAPPING
999 static inline int __zs_cpu_up(struct mapping_area *area)
1002 * Make sure we don't leak memory if a cpu UP notification
1003 * and zs_init() race and both call zs_cpu_up() on the same cpu
1007 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1013 static inline void __zs_cpu_down(struct mapping_area *area)
1016 free_vm_area(area->vm);
1020 static inline void *__zs_map_object(struct mapping_area *area,
1021 struct page *pages[2], int off, int size)
1023 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1024 area->vm_addr = area->vm->addr;
1025 return area->vm_addr + off;
1028 static inline void __zs_unmap_object(struct mapping_area *area,
1029 struct page *pages[2], int off, int size)
1031 unsigned long addr = (unsigned long)area->vm_addr;
1033 unmap_kernel_range(addr, PAGE_SIZE * 2);
1036 #else /* CONFIG_PGTABLE_MAPPING */
1038 static inline int __zs_cpu_up(struct mapping_area *area)
1041 * Make sure we don't leak memory if a cpu UP notification
1042 * and zs_init() race and both call zs_cpu_up() on the same cpu
1046 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1052 static inline void __zs_cpu_down(struct mapping_area *area)
1054 kfree(area->vm_buf);
1055 area->vm_buf = NULL;
1058 static void *__zs_map_object(struct mapping_area *area,
1059 struct page *pages[2], int off, int size)
1063 char *buf = area->vm_buf;
1065 /* disable page faults to match kmap_atomic() return conditions */
1066 pagefault_disable();
1068 /* no read fastpath */
1069 if (area->vm_mm == ZS_MM_WO)
1072 sizes[0] = PAGE_SIZE - off;
1073 sizes[1] = size - sizes[0];
1075 /* copy object to per-cpu buffer */
1076 addr = kmap_atomic(pages[0]);
1077 memcpy(buf, addr + off, sizes[0]);
1078 kunmap_atomic(addr);
1079 addr = kmap_atomic(pages[1]);
1080 memcpy(buf + sizes[0], addr, sizes[1]);
1081 kunmap_atomic(addr);
1083 return area->vm_buf;
1086 static void __zs_unmap_object(struct mapping_area *area,
1087 struct page *pages[2], int off, int size)
1093 /* no write fastpath */
1094 if (area->vm_mm == ZS_MM_RO)
1099 buf = buf + ZS_HANDLE_SIZE;
1100 size -= ZS_HANDLE_SIZE;
1101 off += ZS_HANDLE_SIZE;
1104 sizes[0] = PAGE_SIZE - off;
1105 sizes[1] = size - sizes[0];
1107 /* copy per-cpu buffer to object */
1108 addr = kmap_atomic(pages[0]);
1109 memcpy(addr + off, buf, sizes[0]);
1110 kunmap_atomic(addr);
1111 addr = kmap_atomic(pages[1]);
1112 memcpy(addr, buf + sizes[0], sizes[1]);
1113 kunmap_atomic(addr);
1116 /* enable page faults to match kunmap_atomic() return conditions */
1120 #endif /* CONFIG_PGTABLE_MAPPING */
1122 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1125 int ret, cpu = (long)pcpu;
1126 struct mapping_area *area;
1129 case CPU_UP_PREPARE:
1130 area = &per_cpu(zs_map_area, cpu);
1131 ret = __zs_cpu_up(area);
1133 return notifier_from_errno(ret);
1136 case CPU_UP_CANCELED:
1137 area = &per_cpu(zs_map_area, cpu);
1138 __zs_cpu_down(area);
1145 static struct notifier_block zs_cpu_nb = {
1146 .notifier_call = zs_cpu_notifier
1149 static int zs_register_cpu_notifier(void)
1151 int cpu, uninitialized_var(ret);
1153 cpu_notifier_register_begin();
1155 __register_cpu_notifier(&zs_cpu_nb);
1156 for_each_online_cpu(cpu) {
1157 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1158 if (notifier_to_errno(ret))
1162 cpu_notifier_register_done();
1163 return notifier_to_errno(ret);
1166 static void zs_unregister_cpu_notifier(void)
1170 cpu_notifier_register_begin();
1172 for_each_online_cpu(cpu)
1173 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1174 __unregister_cpu_notifier(&zs_cpu_nb);
1176 cpu_notifier_register_done();
1179 static void init_zs_size_classes(void)
1183 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1184 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1187 zs_size_classes = nr;
1190 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1192 if (prev->pages_per_zspage != pages_per_zspage)
1195 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1196 != get_maxobj_per_zspage(size, pages_per_zspage))
1202 static bool zspage_full(struct page *page)
1204 BUG_ON(!is_first_page(page));
1206 return page->inuse == page->objects;
1209 unsigned long zs_get_total_pages(struct zs_pool *pool)
1211 return atomic_long_read(&pool->pages_allocated);
1213 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1216 * zs_map_object - get address of allocated object from handle.
1217 * @pool: pool from which the object was allocated
1218 * @handle: handle returned from zs_malloc
1220 * Before using an object allocated from zs_malloc, it must be mapped using
1221 * this function. When done with the object, it must be unmapped using
1224 * Only one object can be mapped per cpu at a time. There is no protection
1225 * against nested mappings.
1227 * This function returns with preemption and page faults disabled.
1229 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1233 unsigned long obj, obj_idx, off;
1235 unsigned int class_idx;
1236 enum fullness_group fg;
1237 struct size_class *class;
1238 struct mapping_area *area;
1239 struct page *pages[2];
1245 * Because we use per-cpu mapping areas shared among the
1246 * pools/users, we can't allow mapping in interrupt context
1247 * because it can corrupt another users mappings.
1249 BUG_ON(in_interrupt());
1251 /* From now on, migration cannot move the object */
1254 obj = handle_to_obj(handle);
1255 obj_to_location(obj, &page, &obj_idx);
1256 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1257 class = pool->size_class[class_idx];
1258 off = obj_idx_to_offset(page, obj_idx, class->size);
1260 area = &get_cpu_var(zs_map_area);
1262 if (off + class->size <= PAGE_SIZE) {
1263 /* this object is contained entirely within a page */
1264 area->vm_addr = kmap_atomic(page);
1265 ret = area->vm_addr + off;
1269 /* this object spans two pages */
1271 pages[1] = get_next_page(page);
1274 ret = __zs_map_object(area, pages, off, class->size);
1277 ret += ZS_HANDLE_SIZE;
1281 EXPORT_SYMBOL_GPL(zs_map_object);
1283 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1286 unsigned long obj, obj_idx, off;
1288 unsigned int class_idx;
1289 enum fullness_group fg;
1290 struct size_class *class;
1291 struct mapping_area *area;
1295 obj = handle_to_obj(handle);
1296 obj_to_location(obj, &page, &obj_idx);
1297 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1298 class = pool->size_class[class_idx];
1299 off = obj_idx_to_offset(page, obj_idx, class->size);
1301 area = this_cpu_ptr(&zs_map_area);
1302 if (off + class->size <= PAGE_SIZE)
1303 kunmap_atomic(area->vm_addr);
1305 struct page *pages[2];
1308 pages[1] = get_next_page(page);
1311 __zs_unmap_object(area, pages, off, class->size);
1313 put_cpu_var(zs_map_area);
1316 EXPORT_SYMBOL_GPL(zs_unmap_object);
1318 static unsigned long obj_malloc(struct page *first_page,
1319 struct size_class *class, unsigned long handle)
1322 struct link_free *link;
1324 struct page *m_page;
1325 unsigned long m_objidx, m_offset;
1328 handle |= OBJ_ALLOCATED_TAG;
1329 obj = (unsigned long)first_page->freelist;
1330 obj_to_location(obj, &m_page, &m_objidx);
1331 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1333 vaddr = kmap_atomic(m_page);
1334 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1335 first_page->freelist = link->next;
1337 /* record handle in the header of allocated chunk */
1338 link->handle = handle;
1340 /* record handle in first_page->private */
1341 set_page_private(first_page, handle);
1342 kunmap_atomic(vaddr);
1343 first_page->inuse++;
1344 zs_stat_inc(class, OBJ_USED, 1);
1351 * zs_malloc - Allocate block of given size from pool.
1352 * @pool: pool to allocate from
1353 * @size: size of block to allocate
1355 * On success, handle to the allocated object is returned,
1357 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1359 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1361 unsigned long handle, obj;
1362 struct size_class *class;
1363 struct page *first_page;
1365 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1368 handle = alloc_handle(pool);
1372 /* extra space in chunk to keep the handle */
1373 size += ZS_HANDLE_SIZE;
1374 class = pool->size_class[get_size_class_index(size)];
1376 spin_lock(&class->lock);
1377 first_page = find_get_zspage(class);
1380 spin_unlock(&class->lock);
1381 first_page = alloc_zspage(class, pool->flags);
1382 if (unlikely(!first_page)) {
1383 free_handle(pool, handle);
1387 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1388 atomic_long_add(class->pages_per_zspage,
1389 &pool->pages_allocated);
1391 spin_lock(&class->lock);
1392 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1393 class->size, class->pages_per_zspage));
1396 obj = obj_malloc(first_page, class, handle);
1397 /* Now move the zspage to another fullness group, if required */
1398 fix_fullness_group(class, first_page);
1399 record_obj(handle, obj);
1400 spin_unlock(&class->lock);
1404 EXPORT_SYMBOL_GPL(zs_malloc);
1406 static void obj_free(struct zs_pool *pool, struct size_class *class,
1409 struct link_free *link;
1410 struct page *first_page, *f_page;
1411 unsigned long f_objidx, f_offset;
1414 enum fullness_group fullness;
1418 obj &= ~OBJ_ALLOCATED_TAG;
1419 obj_to_location(obj, &f_page, &f_objidx);
1420 first_page = get_first_page(f_page);
1422 get_zspage_mapping(first_page, &class_idx, &fullness);
1423 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1425 vaddr = kmap_atomic(f_page);
1427 /* Insert this object in containing zspage's freelist */
1428 link = (struct link_free *)(vaddr + f_offset);
1429 link->next = first_page->freelist;
1431 set_page_private(first_page, 0);
1432 kunmap_atomic(vaddr);
1433 first_page->freelist = (void *)obj;
1434 first_page->inuse--;
1435 zs_stat_dec(class, OBJ_USED, 1);
1438 void zs_free(struct zs_pool *pool, unsigned long handle)
1440 struct page *first_page, *f_page;
1441 unsigned long obj, f_objidx;
1443 struct size_class *class;
1444 enum fullness_group fullness;
1446 if (unlikely(!handle))
1450 obj = handle_to_obj(handle);
1451 obj_to_location(obj, &f_page, &f_objidx);
1452 first_page = get_first_page(f_page);
1454 get_zspage_mapping(first_page, &class_idx, &fullness);
1455 class = pool->size_class[class_idx];
1457 spin_lock(&class->lock);
1458 obj_free(pool, class, obj);
1459 fullness = fix_fullness_group(class, first_page);
1460 if (fullness == ZS_EMPTY) {
1461 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1462 class->size, class->pages_per_zspage));
1463 atomic_long_sub(class->pages_per_zspage,
1464 &pool->pages_allocated);
1465 free_zspage(first_page);
1467 spin_unlock(&class->lock);
1470 free_handle(pool, handle);
1472 EXPORT_SYMBOL_GPL(zs_free);
1474 static void zs_object_copy(unsigned long src, unsigned long dst,
1475 struct size_class *class)
1477 struct page *s_page, *d_page;
1478 unsigned long s_objidx, d_objidx;
1479 unsigned long s_off, d_off;
1480 void *s_addr, *d_addr;
1481 int s_size, d_size, size;
1484 s_size = d_size = class->size;
1486 obj_to_location(src, &s_page, &s_objidx);
1487 obj_to_location(dst, &d_page, &d_objidx);
1489 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1490 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1492 if (s_off + class->size > PAGE_SIZE)
1493 s_size = PAGE_SIZE - s_off;
1495 if (d_off + class->size > PAGE_SIZE)
1496 d_size = PAGE_SIZE - d_off;
1498 s_addr = kmap_atomic(s_page);
1499 d_addr = kmap_atomic(d_page);
1502 size = min(s_size, d_size);
1503 memcpy(d_addr + d_off, s_addr + s_off, size);
1506 if (written == class->size)
1514 if (s_off >= PAGE_SIZE) {
1515 kunmap_atomic(d_addr);
1516 kunmap_atomic(s_addr);
1517 s_page = get_next_page(s_page);
1519 s_addr = kmap_atomic(s_page);
1520 d_addr = kmap_atomic(d_page);
1521 s_size = class->size - written;
1525 if (d_off >= PAGE_SIZE) {
1526 kunmap_atomic(d_addr);
1527 d_page = get_next_page(d_page);
1529 d_addr = kmap_atomic(d_page);
1530 d_size = class->size - written;
1535 kunmap_atomic(d_addr);
1536 kunmap_atomic(s_addr);
1540 * Find alloced object in zspage from index object and
1543 static unsigned long find_alloced_obj(struct page *page, int index,
1544 struct size_class *class)
1548 unsigned long handle = 0;
1549 void *addr = kmap_atomic(page);
1551 if (!is_first_page(page))
1552 offset = page->index;
1553 offset += class->size * index;
1555 while (offset < PAGE_SIZE) {
1556 head = obj_to_head(class, page, addr + offset);
1557 if (head & OBJ_ALLOCATED_TAG) {
1558 handle = head & ~OBJ_ALLOCATED_TAG;
1559 if (trypin_tag(handle))
1564 offset += class->size;
1568 kunmap_atomic(addr);
1572 struct zs_compact_control {
1573 /* Source page for migration which could be a subpage of zspage. */
1574 struct page *s_page;
1575 /* Destination page for migration which should be a first page
1577 struct page *d_page;
1578 /* Starting object index within @s_page which used for live object
1579 * in the subpage. */
1581 /* how many of objects are migrated */
1585 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1586 struct zs_compact_control *cc)
1588 unsigned long used_obj, free_obj;
1589 unsigned long handle;
1590 struct page *s_page = cc->s_page;
1591 struct page *d_page = cc->d_page;
1592 unsigned long index = cc->index;
1593 int nr_migrated = 0;
1597 handle = find_alloced_obj(s_page, index, class);
1599 s_page = get_next_page(s_page);
1606 /* Stop if there is no more space */
1607 if (zspage_full(d_page)) {
1613 used_obj = handle_to_obj(handle);
1614 free_obj = obj_malloc(d_page, class, handle);
1615 zs_object_copy(used_obj, free_obj, class);
1617 record_obj(handle, free_obj);
1619 obj_free(pool, class, used_obj);
1623 /* Remember last position in this iteration */
1624 cc->s_page = s_page;
1626 cc->nr_migrated = nr_migrated;
1631 static struct page *alloc_target_page(struct size_class *class)
1636 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1637 page = class->fullness_list[i];
1639 remove_zspage(page, class, i);
1647 static void putback_zspage(struct zs_pool *pool, struct size_class *class,
1648 struct page *first_page)
1650 enum fullness_group fullness;
1652 BUG_ON(!is_first_page(first_page));
1654 fullness = get_fullness_group(first_page);
1655 insert_zspage(first_page, class, fullness);
1656 set_zspage_mapping(first_page, class->index, fullness);
1658 if (fullness == ZS_EMPTY) {
1659 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1660 class->size, class->pages_per_zspage));
1661 atomic_long_sub(class->pages_per_zspage,
1662 &pool->pages_allocated);
1664 free_zspage(first_page);
1668 static struct page *isolate_source_page(struct size_class *class)
1672 page = class->fullness_list[ZS_ALMOST_EMPTY];
1674 remove_zspage(page, class, ZS_ALMOST_EMPTY);
1679 static unsigned long __zs_compact(struct zs_pool *pool,
1680 struct size_class *class)
1682 struct zs_compact_control cc;
1683 struct page *src_page;
1684 struct page *dst_page = NULL;
1685 unsigned long nr_total_migrated = 0;
1687 spin_lock(&class->lock);
1688 while ((src_page = isolate_source_page(class))) {
1690 BUG_ON(!is_first_page(src_page));
1693 cc.s_page = src_page;
1695 while ((dst_page = alloc_target_page(class))) {
1696 cc.d_page = dst_page;
1698 * If there is no more space in dst_page, try to
1699 * allocate another zspage.
1701 if (!migrate_zspage(pool, class, &cc))
1704 putback_zspage(pool, class, dst_page);
1705 nr_total_migrated += cc.nr_migrated;
1708 /* Stop if we couldn't find slot */
1709 if (dst_page == NULL)
1712 putback_zspage(pool, class, dst_page);
1713 putback_zspage(pool, class, src_page);
1714 spin_unlock(&class->lock);
1715 nr_total_migrated += cc.nr_migrated;
1717 spin_lock(&class->lock);
1721 putback_zspage(pool, class, src_page);
1723 spin_unlock(&class->lock);
1725 return nr_total_migrated;
1728 unsigned long zs_compact(struct zs_pool *pool)
1731 unsigned long nr_migrated = 0;
1732 struct size_class *class;
1734 for (i = zs_size_classes - 1; i >= 0; i--) {
1735 class = pool->size_class[i];
1738 if (class->index != i)
1740 nr_migrated += __zs_compact(pool, class);
1745 EXPORT_SYMBOL_GPL(zs_compact);
1748 * zs_create_pool - Creates an allocation pool to work from.
1749 * @flags: allocation flags used to allocate pool metadata
1751 * This function must be called before anything when using
1752 * the zsmalloc allocator.
1754 * On success, a pointer to the newly created pool is returned,
1757 struct zs_pool *zs_create_pool(char *name, gfp_t flags)
1760 struct zs_pool *pool;
1761 struct size_class *prev_class = NULL;
1763 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1767 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1769 if (!pool->size_class) {
1774 pool->name = kstrdup(name, GFP_KERNEL);
1778 if (create_handle_cache(pool))
1782 * Iterate reversly, because, size of size_class that we want to use
1783 * for merging should be larger or equal to current size.
1785 for (i = zs_size_classes - 1; i >= 0; i--) {
1787 int pages_per_zspage;
1788 struct size_class *class;
1790 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1791 if (size > ZS_MAX_ALLOC_SIZE)
1792 size = ZS_MAX_ALLOC_SIZE;
1793 pages_per_zspage = get_pages_per_zspage(size);
1796 * size_class is used for normal zsmalloc operation such
1797 * as alloc/free for that size. Although it is natural that we
1798 * have one size_class for each size, there is a chance that we
1799 * can get more memory utilization if we use one size_class for
1800 * many different sizes whose size_class have same
1801 * characteristics. So, we makes size_class point to
1802 * previous size_class if possible.
1805 if (can_merge(prev_class, size, pages_per_zspage)) {
1806 pool->size_class[i] = prev_class;
1811 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1817 class->pages_per_zspage = pages_per_zspage;
1818 if (pages_per_zspage == 1 &&
1819 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1821 spin_lock_init(&class->lock);
1822 pool->size_class[i] = class;
1827 pool->flags = flags;
1829 if (zs_pool_stat_create(name, pool))
1835 zs_destroy_pool(pool);
1838 EXPORT_SYMBOL_GPL(zs_create_pool);
1840 void zs_destroy_pool(struct zs_pool *pool)
1844 zs_pool_stat_destroy(pool);
1846 for (i = 0; i < zs_size_classes; i++) {
1848 struct size_class *class = pool->size_class[i];
1853 if (class->index != i)
1856 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1857 if (class->fullness_list[fg]) {
1858 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1865 destroy_handle_cache(pool);
1866 kfree(pool->size_class);
1870 EXPORT_SYMBOL_GPL(zs_destroy_pool);
1872 static int __init zs_init(void)
1874 int ret = zs_register_cpu_notifier();
1879 init_zs_size_classes();
1882 zpool_register_driver(&zs_zpool_driver);
1885 ret = zs_stat_init();
1887 pr_err("zs stat initialization failed\n");
1894 zpool_unregister_driver(&zs_zpool_driver);
1897 zs_unregister_cpu_notifier();
1902 static void __exit zs_exit(void)
1905 zpool_unregister_driver(&zs_zpool_driver);
1907 zs_unregister_cpu_notifier();
1912 module_init(zs_init);
1913 module_exit(zs_exit);
1915 MODULE_LICENSE("Dual BSD/GPL");
1916 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");