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 #ifdef CONFIG_ZSMALLOC_STAT
174 static struct dentry *zs_stat_root;
176 struct zs_size_stat {
177 unsigned long objs[NR_ZS_STAT_TYPE];
183 * number of size_classes
185 static int zs_size_classes;
188 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
190 * n = number of allocated objects
191 * N = total number of objects zspage can store
192 * f = fullness_threshold_frac
194 * Similarly, we assign zspage to:
195 * ZS_ALMOST_FULL when n > N / f
196 * ZS_EMPTY when n == 0
197 * ZS_FULL when n == N
199 * (see: fix_fullness_group())
201 static const int fullness_threshold_frac = 4;
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 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
216 #ifdef CONFIG_ZSMALLOC_STAT
217 struct zs_size_stat stats;
222 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
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 #ifdef CONFIG_ZSMALLOC_STAT
255 struct dentry *stat_dentry;
260 * A zspage's class index and fullness group
261 * are encoded in its (first)page->mapping
263 #define CLASS_IDX_BITS 28
264 #define FULLNESS_BITS 4
265 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
266 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
268 struct mapping_area {
269 #ifdef CONFIG_PGTABLE_MAPPING
270 struct vm_struct *vm; /* vm area for mapping object that span pages */
272 char *vm_buf; /* copy buffer for objects that span pages */
274 char *vm_addr; /* address of kmap_atomic()'ed pages */
275 enum zs_mapmode vm_mm; /* mapping mode */
279 static int create_handle_cache(struct zs_pool *pool)
281 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
283 return pool->handle_cachep ? 0 : 1;
286 static void destroy_handle_cache(struct zs_pool *pool)
288 if (pool->handle_cachep)
289 kmem_cache_destroy(pool->handle_cachep);
292 static unsigned long alloc_handle(struct zs_pool *pool)
294 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
295 pool->flags & ~__GFP_HIGHMEM);
298 static void free_handle(struct zs_pool *pool, unsigned long handle)
300 kmem_cache_free(pool->handle_cachep, (void *)handle);
303 static void record_obj(unsigned long handle, unsigned long obj)
305 *(unsigned long *)handle = obj;
312 static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops)
314 return zs_create_pool(name, gfp);
317 static void zs_zpool_destroy(void *pool)
319 zs_destroy_pool(pool);
322 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
323 unsigned long *handle)
325 *handle = zs_malloc(pool, size);
326 return *handle ? 0 : -1;
328 static void zs_zpool_free(void *pool, unsigned long handle)
330 zs_free(pool, handle);
333 static int zs_zpool_shrink(void *pool, unsigned int pages,
334 unsigned int *reclaimed)
339 static void *zs_zpool_map(void *pool, unsigned long handle,
340 enum zpool_mapmode mm)
342 enum zs_mapmode zs_mm;
351 case ZPOOL_MM_RW: /* fallthru */
357 return zs_map_object(pool, handle, zs_mm);
359 static void zs_zpool_unmap(void *pool, unsigned long handle)
361 zs_unmap_object(pool, handle);
364 static u64 zs_zpool_total_size(void *pool)
366 return zs_get_total_pages(pool) << PAGE_SHIFT;
369 static struct zpool_driver zs_zpool_driver = {
371 .owner = THIS_MODULE,
372 .create = zs_zpool_create,
373 .destroy = zs_zpool_destroy,
374 .malloc = zs_zpool_malloc,
375 .free = zs_zpool_free,
376 .shrink = zs_zpool_shrink,
378 .unmap = zs_zpool_unmap,
379 .total_size = zs_zpool_total_size,
382 MODULE_ALIAS("zpool-zsmalloc");
383 #endif /* CONFIG_ZPOOL */
385 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
387 return pages_per_zspage * PAGE_SIZE / size;
390 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
391 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
393 static int is_first_page(struct page *page)
395 return PagePrivate(page);
398 static int is_last_page(struct page *page)
400 return PagePrivate2(page);
403 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
404 enum fullness_group *fullness)
407 BUG_ON(!is_first_page(page));
409 m = (unsigned long)page->mapping;
410 *fullness = m & FULLNESS_MASK;
411 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
414 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
415 enum fullness_group fullness)
418 BUG_ON(!is_first_page(page));
420 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
421 (fullness & FULLNESS_MASK);
422 page->mapping = (struct address_space *)m;
426 * zsmalloc divides the pool into various size classes where each
427 * class maintains a list of zspages where each zspage is divided
428 * into equal sized chunks. Each allocation falls into one of these
429 * classes depending on its size. This function returns index of the
430 * size class which has chunk size big enough to hold the give size.
432 static int get_size_class_index(int size)
436 if (likely(size > ZS_MIN_ALLOC_SIZE))
437 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
438 ZS_SIZE_CLASS_DELTA);
440 return min(zs_size_classes - 1, idx);
443 #ifdef CONFIG_ZSMALLOC_STAT
445 static inline void zs_stat_inc(struct size_class *class,
446 enum zs_stat_type type, unsigned long cnt)
448 class->stats.objs[type] += cnt;
451 static inline void zs_stat_dec(struct size_class *class,
452 enum zs_stat_type type, unsigned long cnt)
454 class->stats.objs[type] -= cnt;
457 static inline unsigned long zs_stat_get(struct size_class *class,
458 enum zs_stat_type type)
460 return class->stats.objs[type];
463 static int __init zs_stat_init(void)
465 if (!debugfs_initialized())
468 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
475 static void __exit zs_stat_exit(void)
477 debugfs_remove_recursive(zs_stat_root);
480 static int zs_stats_size_show(struct seq_file *s, void *v)
483 struct zs_pool *pool = s->private;
484 struct size_class *class;
486 unsigned long class_almost_full, class_almost_empty;
487 unsigned long obj_allocated, obj_used, pages_used;
488 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
489 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
491 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
492 "class", "size", "almost_full", "almost_empty",
493 "obj_allocated", "obj_used", "pages_used",
496 for (i = 0; i < zs_size_classes; i++) {
497 class = pool->size_class[i];
499 if (class->index != i)
502 spin_lock(&class->lock);
503 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
504 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
505 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
506 obj_used = zs_stat_get(class, OBJ_USED);
507 spin_unlock(&class->lock);
509 objs_per_zspage = get_maxobj_per_zspage(class->size,
510 class->pages_per_zspage);
511 pages_used = obj_allocated / objs_per_zspage *
512 class->pages_per_zspage;
514 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
515 i, class->size, class_almost_full, class_almost_empty,
516 obj_allocated, obj_used, pages_used,
517 class->pages_per_zspage);
519 total_class_almost_full += class_almost_full;
520 total_class_almost_empty += class_almost_empty;
521 total_objs += obj_allocated;
522 total_used_objs += obj_used;
523 total_pages += pages_used;
527 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
528 "Total", "", total_class_almost_full,
529 total_class_almost_empty, total_objs,
530 total_used_objs, total_pages);
535 static int zs_stats_size_open(struct inode *inode, struct file *file)
537 return single_open(file, zs_stats_size_show, inode->i_private);
540 static const struct file_operations zs_stat_size_ops = {
541 .open = zs_stats_size_open,
544 .release = single_release,
547 static int zs_pool_stat_create(char *name, struct zs_pool *pool)
549 struct dentry *entry;
554 entry = debugfs_create_dir(name, zs_stat_root);
556 pr_warn("debugfs dir <%s> creation failed\n", name);
559 pool->stat_dentry = entry;
561 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
562 pool->stat_dentry, pool, &zs_stat_size_ops);
564 pr_warn("%s: debugfs file entry <%s> creation failed\n",
572 static void zs_pool_stat_destroy(struct zs_pool *pool)
574 debugfs_remove_recursive(pool->stat_dentry);
577 #else /* CONFIG_ZSMALLOC_STAT */
579 static inline void zs_stat_inc(struct size_class *class,
580 enum zs_stat_type type, unsigned long cnt)
584 static inline void zs_stat_dec(struct size_class *class,
585 enum zs_stat_type type, unsigned long cnt)
589 static inline unsigned long zs_stat_get(struct size_class *class,
590 enum zs_stat_type type)
595 static int __init zs_stat_init(void)
600 static void __exit zs_stat_exit(void)
604 static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
609 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
617 * For each size class, zspages are divided into different groups
618 * depending on how "full" they are. This was done so that we could
619 * easily find empty or nearly empty zspages when we try to shrink
620 * the pool (not yet implemented). This function returns fullness
621 * status of the given page.
623 static enum fullness_group get_fullness_group(struct page *page)
625 int inuse, max_objects;
626 enum fullness_group fg;
627 BUG_ON(!is_first_page(page));
630 max_objects = page->objects;
634 else if (inuse == max_objects)
636 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
637 fg = ZS_ALMOST_EMPTY;
645 * Each size class maintains various freelists and zspages are assigned
646 * to one of these freelists based on the number of live objects they
647 * have. This functions inserts the given zspage into the freelist
648 * identified by <class, fullness_group>.
650 static void insert_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 list_add_tail(&page->lru, &(*head)->lru);
665 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
666 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
670 * This function removes the given zspage from the freelist identified
671 * by <class, fullness_group>.
673 static void remove_zspage(struct page *page, struct size_class *class,
674 enum fullness_group fullness)
678 BUG_ON(!is_first_page(page));
680 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
683 head = &class->fullness_list[fullness];
685 if (list_empty(&(*head)->lru))
687 else if (*head == page)
688 *head = (struct page *)list_entry((*head)->lru.next,
691 list_del_init(&page->lru);
692 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
693 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
697 * Each size class maintains zspages in different fullness groups depending
698 * on the number of live objects they contain. When allocating or freeing
699 * objects, the fullness status of the page can change, say, from ALMOST_FULL
700 * to ALMOST_EMPTY when freeing an object. This function checks if such
701 * a status change has occurred for the given page and accordingly moves the
702 * page from the freelist of the old fullness group to that of the new
705 static enum fullness_group fix_fullness_group(struct size_class *class,
709 enum fullness_group currfg, newfg;
711 BUG_ON(!is_first_page(page));
713 get_zspage_mapping(page, &class_idx, &currfg);
714 newfg = get_fullness_group(page);
718 remove_zspage(page, class, currfg);
719 insert_zspage(page, class, newfg);
720 set_zspage_mapping(page, class_idx, newfg);
727 * We have to decide on how many pages to link together
728 * to form a zspage for each size class. This is important
729 * to reduce wastage due to unusable space left at end of
730 * each zspage which is given as:
731 * wastage = Zp % class_size
732 * usage = Zp - wastage
733 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
735 * For example, for size class of 3/8 * PAGE_SIZE, we should
736 * link together 3 PAGE_SIZE sized pages to form a zspage
737 * since then we can perfectly fit in 8 such objects.
739 static int get_pages_per_zspage(int class_size)
741 int i, max_usedpc = 0;
742 /* zspage order which gives maximum used size per KB */
743 int max_usedpc_order = 1;
745 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
749 zspage_size = i * PAGE_SIZE;
750 waste = zspage_size % class_size;
751 usedpc = (zspage_size - waste) * 100 / zspage_size;
753 if (usedpc > max_usedpc) {
755 max_usedpc_order = i;
759 return max_usedpc_order;
763 * A single 'zspage' is composed of many system pages which are
764 * linked together using fields in struct page. This function finds
765 * the first/head page, given any component page of a zspage.
767 static struct page *get_first_page(struct page *page)
769 if (is_first_page(page))
772 return page->first_page;
775 static struct page *get_next_page(struct page *page)
779 if (is_last_page(page))
781 else if (is_first_page(page))
782 next = (struct page *)page_private(page);
784 next = list_entry(page->lru.next, struct page, lru);
790 * Encode <page, obj_idx> as a single handle value.
791 * We use the least bit of handle for tagging.
793 static void *location_to_obj(struct page *page, unsigned long obj_idx)
802 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
803 obj |= ((obj_idx) & OBJ_INDEX_MASK);
804 obj <<= OBJ_TAG_BITS;
810 * Decode <page, obj_idx> pair from the given object handle. We adjust the
811 * decoded obj_idx back to its original value since it was adjusted in
814 static void obj_to_location(unsigned long obj, struct page **page,
815 unsigned long *obj_idx)
817 obj >>= OBJ_TAG_BITS;
818 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
819 *obj_idx = (obj & OBJ_INDEX_MASK);
822 static unsigned long handle_to_obj(unsigned long handle)
824 return *(unsigned long *)handle;
827 static unsigned long obj_to_head(struct size_class *class, struct page *page,
831 VM_BUG_ON(!is_first_page(page));
832 return *(unsigned long *)page_private(page);
834 return *(unsigned long *)obj;
837 static unsigned long obj_idx_to_offset(struct page *page,
838 unsigned long obj_idx, int class_size)
840 unsigned long off = 0;
842 if (!is_first_page(page))
845 return off + obj_idx * class_size;
848 static inline int trypin_tag(unsigned long handle)
850 unsigned long *ptr = (unsigned long *)handle;
852 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
855 static void pin_tag(unsigned long handle)
857 while (!trypin_tag(handle));
860 static void unpin_tag(unsigned long handle)
862 unsigned long *ptr = (unsigned long *)handle;
864 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
867 static void reset_page(struct page *page)
869 clear_bit(PG_private, &page->flags);
870 clear_bit(PG_private_2, &page->flags);
871 set_page_private(page, 0);
872 page->mapping = NULL;
873 page->freelist = NULL;
874 page_mapcount_reset(page);
877 static void free_zspage(struct page *first_page)
879 struct page *nextp, *tmp, *head_extra;
881 BUG_ON(!is_first_page(first_page));
882 BUG_ON(first_page->inuse);
884 head_extra = (struct page *)page_private(first_page);
886 reset_page(first_page);
887 __free_page(first_page);
889 /* zspage with only 1 system page */
893 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
894 list_del(&nextp->lru);
898 reset_page(head_extra);
899 __free_page(head_extra);
902 /* Initialize a newly allocated zspage */
903 static void init_zspage(struct page *first_page, struct size_class *class)
905 unsigned long off = 0;
906 struct page *page = first_page;
908 BUG_ON(!is_first_page(first_page));
910 struct page *next_page;
911 struct link_free *link;
916 * page->index stores offset of first object starting
917 * in the page. For the first page, this is always 0,
918 * so we use first_page->index (aka ->freelist) to store
919 * head of corresponding zspage's freelist.
921 if (page != first_page)
924 vaddr = kmap_atomic(page);
925 link = (struct link_free *)vaddr + off / sizeof(*link);
927 while ((off += class->size) < PAGE_SIZE) {
928 link->next = location_to_obj(page, i++);
929 link += class->size / sizeof(*link);
933 * We now come to the last (full or partial) object on this
934 * page, which must point to the first object on the next
937 next_page = get_next_page(page);
938 link->next = location_to_obj(next_page, 0);
939 kunmap_atomic(vaddr);
946 * Allocate a zspage for the given size class
948 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
951 struct page *first_page = NULL, *uninitialized_var(prev_page);
954 * Allocate individual pages and link them together as:
955 * 1. first page->private = first sub-page
956 * 2. all sub-pages are linked together using page->lru
957 * 3. each sub-page is linked to the first page using page->first_page
959 * For each size class, First/Head pages are linked together using
960 * page->lru. Also, we set PG_private to identify the first page
961 * (i.e. no other sub-page has this flag set) and PG_private_2 to
962 * identify the last page.
965 for (i = 0; i < class->pages_per_zspage; i++) {
968 page = alloc_page(flags);
972 INIT_LIST_HEAD(&page->lru);
973 if (i == 0) { /* first page */
974 SetPagePrivate(page);
975 set_page_private(page, 0);
977 first_page->inuse = 0;
980 set_page_private(first_page, (unsigned long)page);
982 page->first_page = first_page;
984 list_add(&page->lru, &prev_page->lru);
985 if (i == class->pages_per_zspage - 1) /* last page */
986 SetPagePrivate2(page);
990 init_zspage(first_page, class);
992 first_page->freelist = location_to_obj(first_page, 0);
993 /* Maximum number of objects we can store in this zspage */
994 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
996 error = 0; /* Success */
999 if (unlikely(error) && first_page) {
1000 free_zspage(first_page);
1007 static struct page *find_get_zspage(struct size_class *class)
1012 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1013 page = class->fullness_list[i];
1021 #ifdef CONFIG_PGTABLE_MAPPING
1022 static inline int __zs_cpu_up(struct mapping_area *area)
1025 * Make sure we don't leak memory if a cpu UP notification
1026 * and zs_init() race and both call zs_cpu_up() on the same cpu
1030 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1036 static inline void __zs_cpu_down(struct mapping_area *area)
1039 free_vm_area(area->vm);
1043 static inline void *__zs_map_object(struct mapping_area *area,
1044 struct page *pages[2], int off, int size)
1046 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1047 area->vm_addr = area->vm->addr;
1048 return area->vm_addr + off;
1051 static inline void __zs_unmap_object(struct mapping_area *area,
1052 struct page *pages[2], int off, int size)
1054 unsigned long addr = (unsigned long)area->vm_addr;
1056 unmap_kernel_range(addr, PAGE_SIZE * 2);
1059 #else /* CONFIG_PGTABLE_MAPPING */
1061 static inline int __zs_cpu_up(struct mapping_area *area)
1064 * Make sure we don't leak memory if a cpu UP notification
1065 * and zs_init() race and both call zs_cpu_up() on the same cpu
1069 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1075 static inline void __zs_cpu_down(struct mapping_area *area)
1077 kfree(area->vm_buf);
1078 area->vm_buf = NULL;
1081 static void *__zs_map_object(struct mapping_area *area,
1082 struct page *pages[2], int off, int size)
1086 char *buf = area->vm_buf;
1088 /* disable page faults to match kmap_atomic() return conditions */
1089 pagefault_disable();
1091 /* no read fastpath */
1092 if (area->vm_mm == ZS_MM_WO)
1095 sizes[0] = PAGE_SIZE - off;
1096 sizes[1] = size - sizes[0];
1098 /* copy object to per-cpu buffer */
1099 addr = kmap_atomic(pages[0]);
1100 memcpy(buf, addr + off, sizes[0]);
1101 kunmap_atomic(addr);
1102 addr = kmap_atomic(pages[1]);
1103 memcpy(buf + sizes[0], addr, sizes[1]);
1104 kunmap_atomic(addr);
1106 return area->vm_buf;
1109 static void __zs_unmap_object(struct mapping_area *area,
1110 struct page *pages[2], int off, int size)
1116 /* no write fastpath */
1117 if (area->vm_mm == ZS_MM_RO)
1122 buf = buf + ZS_HANDLE_SIZE;
1123 size -= ZS_HANDLE_SIZE;
1124 off += ZS_HANDLE_SIZE;
1127 sizes[0] = PAGE_SIZE - off;
1128 sizes[1] = size - sizes[0];
1130 /* copy per-cpu buffer to object */
1131 addr = kmap_atomic(pages[0]);
1132 memcpy(addr + off, buf, sizes[0]);
1133 kunmap_atomic(addr);
1134 addr = kmap_atomic(pages[1]);
1135 memcpy(addr, buf + sizes[0], sizes[1]);
1136 kunmap_atomic(addr);
1139 /* enable page faults to match kunmap_atomic() return conditions */
1143 #endif /* CONFIG_PGTABLE_MAPPING */
1145 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1148 int ret, cpu = (long)pcpu;
1149 struct mapping_area *area;
1152 case CPU_UP_PREPARE:
1153 area = &per_cpu(zs_map_area, cpu);
1154 ret = __zs_cpu_up(area);
1156 return notifier_from_errno(ret);
1159 case CPU_UP_CANCELED:
1160 area = &per_cpu(zs_map_area, cpu);
1161 __zs_cpu_down(area);
1168 static struct notifier_block zs_cpu_nb = {
1169 .notifier_call = zs_cpu_notifier
1172 static int zs_register_cpu_notifier(void)
1174 int cpu, uninitialized_var(ret);
1176 cpu_notifier_register_begin();
1178 __register_cpu_notifier(&zs_cpu_nb);
1179 for_each_online_cpu(cpu) {
1180 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1181 if (notifier_to_errno(ret))
1185 cpu_notifier_register_done();
1186 return notifier_to_errno(ret);
1189 static void zs_unregister_cpu_notifier(void)
1193 cpu_notifier_register_begin();
1195 for_each_online_cpu(cpu)
1196 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1197 __unregister_cpu_notifier(&zs_cpu_nb);
1199 cpu_notifier_register_done();
1202 static void init_zs_size_classes(void)
1206 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1207 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1210 zs_size_classes = nr;
1213 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1215 if (prev->pages_per_zspage != pages_per_zspage)
1218 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1219 != get_maxobj_per_zspage(size, pages_per_zspage))
1225 static bool zspage_full(struct page *page)
1227 BUG_ON(!is_first_page(page));
1229 return page->inuse == page->objects;
1232 unsigned long zs_get_total_pages(struct zs_pool *pool)
1234 return atomic_long_read(&pool->pages_allocated);
1236 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1239 * zs_map_object - get address of allocated object from handle.
1240 * @pool: pool from which the object was allocated
1241 * @handle: handle returned from zs_malloc
1243 * Before using an object allocated from zs_malloc, it must be mapped using
1244 * this function. When done with the object, it must be unmapped using
1247 * Only one object can be mapped per cpu at a time. There is no protection
1248 * against nested mappings.
1250 * This function returns with preemption and page faults disabled.
1252 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1256 unsigned long obj, obj_idx, off;
1258 unsigned int class_idx;
1259 enum fullness_group fg;
1260 struct size_class *class;
1261 struct mapping_area *area;
1262 struct page *pages[2];
1268 * Because we use per-cpu mapping areas shared among the
1269 * pools/users, we can't allow mapping in interrupt context
1270 * because it can corrupt another users mappings.
1272 BUG_ON(in_interrupt());
1274 /* From now on, migration cannot move the object */
1277 obj = handle_to_obj(handle);
1278 obj_to_location(obj, &page, &obj_idx);
1279 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1280 class = pool->size_class[class_idx];
1281 off = obj_idx_to_offset(page, obj_idx, class->size);
1283 area = &get_cpu_var(zs_map_area);
1285 if (off + class->size <= PAGE_SIZE) {
1286 /* this object is contained entirely within a page */
1287 area->vm_addr = kmap_atomic(page);
1288 ret = area->vm_addr + off;
1292 /* this object spans two pages */
1294 pages[1] = get_next_page(page);
1297 ret = __zs_map_object(area, pages, off, class->size);
1300 ret += ZS_HANDLE_SIZE;
1304 EXPORT_SYMBOL_GPL(zs_map_object);
1306 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1309 unsigned long obj, obj_idx, off;
1311 unsigned int class_idx;
1312 enum fullness_group fg;
1313 struct size_class *class;
1314 struct mapping_area *area;
1318 obj = handle_to_obj(handle);
1319 obj_to_location(obj, &page, &obj_idx);
1320 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1321 class = pool->size_class[class_idx];
1322 off = obj_idx_to_offset(page, obj_idx, class->size);
1324 area = this_cpu_ptr(&zs_map_area);
1325 if (off + class->size <= PAGE_SIZE)
1326 kunmap_atomic(area->vm_addr);
1328 struct page *pages[2];
1331 pages[1] = get_next_page(page);
1334 __zs_unmap_object(area, pages, off, class->size);
1336 put_cpu_var(zs_map_area);
1339 EXPORT_SYMBOL_GPL(zs_unmap_object);
1341 static unsigned long obj_malloc(struct page *first_page,
1342 struct size_class *class, unsigned long handle)
1345 struct link_free *link;
1347 struct page *m_page;
1348 unsigned long m_objidx, m_offset;
1351 handle |= OBJ_ALLOCATED_TAG;
1352 obj = (unsigned long)first_page->freelist;
1353 obj_to_location(obj, &m_page, &m_objidx);
1354 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1356 vaddr = kmap_atomic(m_page);
1357 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1358 first_page->freelist = link->next;
1360 /* record handle in the header of allocated chunk */
1361 link->handle = handle;
1363 /* record handle in first_page->private */
1364 set_page_private(first_page, handle);
1365 kunmap_atomic(vaddr);
1366 first_page->inuse++;
1367 zs_stat_inc(class, OBJ_USED, 1);
1374 * zs_malloc - Allocate block of given size from pool.
1375 * @pool: pool to allocate from
1376 * @size: size of block to allocate
1378 * On success, handle to the allocated object is returned,
1380 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1382 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1384 unsigned long handle, obj;
1385 struct size_class *class;
1386 struct page *first_page;
1388 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1391 handle = alloc_handle(pool);
1395 /* extra space in chunk to keep the handle */
1396 size += ZS_HANDLE_SIZE;
1397 class = pool->size_class[get_size_class_index(size)];
1399 spin_lock(&class->lock);
1400 first_page = find_get_zspage(class);
1403 spin_unlock(&class->lock);
1404 first_page = alloc_zspage(class, pool->flags);
1405 if (unlikely(!first_page)) {
1406 free_handle(pool, handle);
1410 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1411 atomic_long_add(class->pages_per_zspage,
1412 &pool->pages_allocated);
1414 spin_lock(&class->lock);
1415 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1416 class->size, class->pages_per_zspage));
1419 obj = obj_malloc(first_page, class, handle);
1420 /* Now move the zspage to another fullness group, if required */
1421 fix_fullness_group(class, first_page);
1422 record_obj(handle, obj);
1423 spin_unlock(&class->lock);
1427 EXPORT_SYMBOL_GPL(zs_malloc);
1429 static void obj_free(struct zs_pool *pool, struct size_class *class,
1432 struct link_free *link;
1433 struct page *first_page, *f_page;
1434 unsigned long f_objidx, f_offset;
1437 enum fullness_group fullness;
1441 obj &= ~OBJ_ALLOCATED_TAG;
1442 obj_to_location(obj, &f_page, &f_objidx);
1443 first_page = get_first_page(f_page);
1445 get_zspage_mapping(first_page, &class_idx, &fullness);
1446 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1448 vaddr = kmap_atomic(f_page);
1450 /* Insert this object in containing zspage's freelist */
1451 link = (struct link_free *)(vaddr + f_offset);
1452 link->next = first_page->freelist;
1454 set_page_private(first_page, 0);
1455 kunmap_atomic(vaddr);
1456 first_page->freelist = (void *)obj;
1457 first_page->inuse--;
1458 zs_stat_dec(class, OBJ_USED, 1);
1461 void zs_free(struct zs_pool *pool, unsigned long handle)
1463 struct page *first_page, *f_page;
1464 unsigned long obj, f_objidx;
1466 struct size_class *class;
1467 enum fullness_group fullness;
1469 if (unlikely(!handle))
1473 obj = handle_to_obj(handle);
1474 obj_to_location(obj, &f_page, &f_objidx);
1475 first_page = get_first_page(f_page);
1477 get_zspage_mapping(first_page, &class_idx, &fullness);
1478 class = pool->size_class[class_idx];
1480 spin_lock(&class->lock);
1481 obj_free(pool, class, obj);
1482 fullness = fix_fullness_group(class, first_page);
1483 if (fullness == ZS_EMPTY) {
1484 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1485 class->size, class->pages_per_zspage));
1486 atomic_long_sub(class->pages_per_zspage,
1487 &pool->pages_allocated);
1488 free_zspage(first_page);
1490 spin_unlock(&class->lock);
1493 free_handle(pool, handle);
1495 EXPORT_SYMBOL_GPL(zs_free);
1497 static void zs_object_copy(unsigned long src, unsigned long dst,
1498 struct size_class *class)
1500 struct page *s_page, *d_page;
1501 unsigned long s_objidx, d_objidx;
1502 unsigned long s_off, d_off;
1503 void *s_addr, *d_addr;
1504 int s_size, d_size, size;
1507 s_size = d_size = class->size;
1509 obj_to_location(src, &s_page, &s_objidx);
1510 obj_to_location(dst, &d_page, &d_objidx);
1512 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1513 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1515 if (s_off + class->size > PAGE_SIZE)
1516 s_size = PAGE_SIZE - s_off;
1518 if (d_off + class->size > PAGE_SIZE)
1519 d_size = PAGE_SIZE - d_off;
1521 s_addr = kmap_atomic(s_page);
1522 d_addr = kmap_atomic(d_page);
1525 size = min(s_size, d_size);
1526 memcpy(d_addr + d_off, s_addr + s_off, size);
1529 if (written == class->size)
1537 if (s_off >= PAGE_SIZE) {
1538 kunmap_atomic(d_addr);
1539 kunmap_atomic(s_addr);
1540 s_page = get_next_page(s_page);
1542 s_addr = kmap_atomic(s_page);
1543 d_addr = kmap_atomic(d_page);
1544 s_size = class->size - written;
1548 if (d_off >= PAGE_SIZE) {
1549 kunmap_atomic(d_addr);
1550 d_page = get_next_page(d_page);
1552 d_addr = kmap_atomic(d_page);
1553 d_size = class->size - written;
1558 kunmap_atomic(d_addr);
1559 kunmap_atomic(s_addr);
1563 * Find alloced object in zspage from index object and
1566 static unsigned long find_alloced_obj(struct page *page, int index,
1567 struct size_class *class)
1571 unsigned long handle = 0;
1572 void *addr = kmap_atomic(page);
1574 if (!is_first_page(page))
1575 offset = page->index;
1576 offset += class->size * index;
1578 while (offset < PAGE_SIZE) {
1579 head = obj_to_head(class, page, addr + offset);
1580 if (head & OBJ_ALLOCATED_TAG) {
1581 handle = head & ~OBJ_ALLOCATED_TAG;
1582 if (trypin_tag(handle))
1587 offset += class->size;
1591 kunmap_atomic(addr);
1595 struct zs_compact_control {
1596 /* Source page for migration which could be a subpage of zspage. */
1597 struct page *s_page;
1598 /* Destination page for migration which should be a first page
1600 struct page *d_page;
1601 /* Starting object index within @s_page which used for live object
1602 * in the subpage. */
1604 /* how many of objects are migrated */
1608 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1609 struct zs_compact_control *cc)
1611 unsigned long used_obj, free_obj;
1612 unsigned long handle;
1613 struct page *s_page = cc->s_page;
1614 struct page *d_page = cc->d_page;
1615 unsigned long index = cc->index;
1616 int nr_migrated = 0;
1620 handle = find_alloced_obj(s_page, index, class);
1622 s_page = get_next_page(s_page);
1629 /* Stop if there is no more space */
1630 if (zspage_full(d_page)) {
1636 used_obj = handle_to_obj(handle);
1637 free_obj = obj_malloc(d_page, class, handle);
1638 zs_object_copy(used_obj, free_obj, class);
1640 record_obj(handle, free_obj);
1642 obj_free(pool, class, used_obj);
1646 /* Remember last position in this iteration */
1647 cc->s_page = s_page;
1649 cc->nr_migrated = nr_migrated;
1654 static struct page *alloc_target_page(struct size_class *class)
1659 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1660 page = class->fullness_list[i];
1662 remove_zspage(page, class, i);
1670 static void putback_zspage(struct zs_pool *pool, struct size_class *class,
1671 struct page *first_page)
1673 enum fullness_group fullness;
1675 BUG_ON(!is_first_page(first_page));
1677 fullness = get_fullness_group(first_page);
1678 insert_zspage(first_page, class, fullness);
1679 set_zspage_mapping(first_page, class->index, fullness);
1681 if (fullness == ZS_EMPTY) {
1682 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1683 class->size, class->pages_per_zspage));
1684 atomic_long_sub(class->pages_per_zspage,
1685 &pool->pages_allocated);
1687 free_zspage(first_page);
1691 static struct page *isolate_source_page(struct size_class *class)
1695 page = class->fullness_list[ZS_ALMOST_EMPTY];
1697 remove_zspage(page, class, ZS_ALMOST_EMPTY);
1702 static unsigned long __zs_compact(struct zs_pool *pool,
1703 struct size_class *class)
1706 struct zs_compact_control cc;
1707 struct page *src_page;
1708 struct page *dst_page = NULL;
1709 unsigned long nr_total_migrated = 0;
1711 spin_lock(&class->lock);
1712 while ((src_page = isolate_source_page(class))) {
1714 BUG_ON(!is_first_page(src_page));
1716 /* The goal is to migrate all live objects in source page */
1717 nr_to_migrate = src_page->inuse;
1719 cc.s_page = src_page;
1721 while ((dst_page = alloc_target_page(class))) {
1722 cc.d_page = dst_page;
1724 * If there is no more space in dst_page, try to
1725 * allocate another zspage.
1727 if (!migrate_zspage(pool, class, &cc))
1730 putback_zspage(pool, class, dst_page);
1731 nr_total_migrated += cc.nr_migrated;
1732 nr_to_migrate -= cc.nr_migrated;
1735 /* Stop if we couldn't find slot */
1736 if (dst_page == NULL)
1739 putback_zspage(pool, class, dst_page);
1740 putback_zspage(pool, class, src_page);
1741 spin_unlock(&class->lock);
1742 nr_total_migrated += cc.nr_migrated;
1744 spin_lock(&class->lock);
1748 putback_zspage(pool, class, src_page);
1750 spin_unlock(&class->lock);
1752 return nr_total_migrated;
1755 unsigned long zs_compact(struct zs_pool *pool)
1758 unsigned long nr_migrated = 0;
1759 struct size_class *class;
1761 for (i = zs_size_classes - 1; i >= 0; i--) {
1762 class = pool->size_class[i];
1765 if (class->index != i)
1767 nr_migrated += __zs_compact(pool, class);
1772 EXPORT_SYMBOL_GPL(zs_compact);
1775 * zs_create_pool - Creates an allocation pool to work from.
1776 * @flags: allocation flags used to allocate pool metadata
1778 * This function must be called before anything when using
1779 * the zsmalloc allocator.
1781 * On success, a pointer to the newly created pool is returned,
1784 struct zs_pool *zs_create_pool(char *name, gfp_t flags)
1787 struct zs_pool *pool;
1788 struct size_class *prev_class = NULL;
1790 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1794 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1796 if (!pool->size_class) {
1801 pool->name = kstrdup(name, GFP_KERNEL);
1805 if (create_handle_cache(pool))
1809 * Iterate reversly, because, size of size_class that we want to use
1810 * for merging should be larger or equal to current size.
1812 for (i = zs_size_classes - 1; i >= 0; i--) {
1814 int pages_per_zspage;
1815 struct size_class *class;
1817 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1818 if (size > ZS_MAX_ALLOC_SIZE)
1819 size = ZS_MAX_ALLOC_SIZE;
1820 pages_per_zspage = get_pages_per_zspage(size);
1823 * size_class is used for normal zsmalloc operation such
1824 * as alloc/free for that size. Although it is natural that we
1825 * have one size_class for each size, there is a chance that we
1826 * can get more memory utilization if we use one size_class for
1827 * many different sizes whose size_class have same
1828 * characteristics. So, we makes size_class point to
1829 * previous size_class if possible.
1832 if (can_merge(prev_class, size, pages_per_zspage)) {
1833 pool->size_class[i] = prev_class;
1838 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1844 class->pages_per_zspage = pages_per_zspage;
1845 if (pages_per_zspage == 1 &&
1846 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1848 spin_lock_init(&class->lock);
1849 pool->size_class[i] = class;
1854 pool->flags = flags;
1856 if (zs_pool_stat_create(name, pool))
1862 zs_destroy_pool(pool);
1865 EXPORT_SYMBOL_GPL(zs_create_pool);
1867 void zs_destroy_pool(struct zs_pool *pool)
1871 zs_pool_stat_destroy(pool);
1873 for (i = 0; i < zs_size_classes; i++) {
1875 struct size_class *class = pool->size_class[i];
1880 if (class->index != i)
1883 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1884 if (class->fullness_list[fg]) {
1885 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1892 destroy_handle_cache(pool);
1893 kfree(pool->size_class);
1897 EXPORT_SYMBOL_GPL(zs_destroy_pool);
1899 static int __init zs_init(void)
1901 int ret = zs_register_cpu_notifier();
1906 init_zs_size_classes();
1909 zpool_register_driver(&zs_zpool_driver);
1912 ret = zs_stat_init();
1914 pr_err("zs stat initialization failed\n");
1921 zpool_unregister_driver(&zs_zpool_driver);
1924 zs_unregister_cpu_notifier();
1929 static void __exit zs_exit(void)
1932 zpool_unregister_driver(&zs_zpool_driver);
1934 zs_unregister_cpu_notifier();
1939 module_init(zs_init);
1940 module_exit(zs_exit);
1942 MODULE_LICENSE("Dual BSD/GPL");
1943 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");