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 zspage
20 * page->freelist(index): links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
23 * page->units: first object offset in a subpage of zspage
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_owner_priv_1: identifies the huge component page
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33 #include <linux/module.h>
34 #include <linux/kernel.h>
35 #include <linux/sched.h>
36 #include <linux/magic.h>
37 #include <linux/bitops.h>
38 #include <linux/errno.h>
39 #include <linux/highmem.h>
40 #include <linux/string.h>
41 #include <linux/slab.h>
42 #include <asm/tlbflush.h>
43 #include <asm/pgtable.h>
44 #include <linux/cpumask.h>
45 #include <linux/cpu.h>
46 #include <linux/vmalloc.h>
47 #include <linux/preempt.h>
48 #include <linux/spinlock.h>
49 #include <linux/shrinker.h>
50 #include <linux/types.h>
51 #include <linux/debugfs.h>
52 #include <linux/zsmalloc.h>
53 #include <linux/zpool.h>
54 #include <linux/mount.h>
55 #include <linux/migrate.h>
56 #include <linux/pagemap.h>
59 #define ZSPAGE_MAGIC 0x58
62 * This must be power of 2 and greater than of equal to sizeof(link_free).
63 * These two conditions ensure that any 'struct link_free' itself doesn't
64 * span more than 1 page which avoids complex case of mapping 2 pages simply
65 * to restore link_free pointer values.
70 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
71 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
73 #define ZS_MAX_ZSPAGE_ORDER 2
74 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
76 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
79 * Object location (<PFN>, <obj_idx>) is encoded as
80 * as single (unsigned long) handle value.
82 * Note that object index <obj_idx> starts from 0.
84 * This is made more complicated by various memory models and PAE.
87 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
88 #ifdef MAX_PHYSMEM_BITS
89 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
92 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
95 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
99 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
102 * Memory for allocating for handle keeps object position by
103 * encoding <page, obj_idx> and the encoded value has a room
104 * in least bit(ie, look at obj_to_location).
105 * We use the bit to synchronize between object access by
106 * user and migration.
108 #define HANDLE_PIN_BIT 0
111 * Head in allocated object should have OBJ_ALLOCATED_TAG
112 * to identify the object was allocated or not.
113 * It's okay to add the status bit in the least bit because
114 * header keeps handle which is 4byte-aligned address so we
115 * have room for two bit at least.
117 #define OBJ_ALLOCATED_TAG 1
118 #define OBJ_TAG_BITS 1
119 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
120 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
122 #define FULLNESS_BITS 2
124 #define ISOLATED_BITS 3
125 #define MAGIC_VAL_BITS 8
127 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
128 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
129 #define ZS_MIN_ALLOC_SIZE \
130 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
131 /* each chunk includes extra space to keep handle */
132 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
135 * On systems with 4K page size, this gives 255 size classes! There is a
137 * - Large number of size classes is potentially wasteful as free page are
138 * spread across these classes
139 * - Small number of size classes causes large internal fragmentation
140 * - Probably its better to use specific size classes (empirically
141 * determined). NOTE: all those class sizes must be set as multiple of
142 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
144 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
147 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
148 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
149 ZS_SIZE_CLASS_DELTA) + 1)
151 enum fullness_group {
169 struct zs_size_stat {
170 unsigned long objs[NR_ZS_STAT_TYPE];
173 #ifdef CONFIG_ZSMALLOC_STAT
174 static struct dentry *zs_stat_root;
177 #ifdef CONFIG_COMPACTION
178 static struct vfsmount *zsmalloc_mnt;
182 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
184 * n = number of allocated objects
185 * N = total number of objects zspage can store
186 * f = fullness_threshold_frac
188 * Similarly, we assign zspage to:
189 * ZS_ALMOST_FULL when n > N / f
190 * ZS_EMPTY when n == 0
191 * ZS_FULL when n == N
193 * (see: fix_fullness_group())
195 static const int fullness_threshold_frac = 4;
196 static size_t huge_class_size;
200 struct list_head fullness_list[NR_ZS_FULLNESS];
202 * Size of objects stored in this class. Must be multiple
207 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
208 int pages_per_zspage;
211 struct zs_size_stat stats;
214 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
215 static void SetPageHugeObject(struct page *page)
217 SetPageOwnerPriv1(page);
220 static void ClearPageHugeObject(struct page *page)
222 ClearPageOwnerPriv1(page);
225 static int PageHugeObject(struct page *page)
227 return PageOwnerPriv1(page);
231 * Placed within free objects to form a singly linked list.
232 * For every zspage, zspage->freeobj gives head of this list.
234 * This must be power of 2 and less than or equal to ZS_ALIGN
240 * It's valid for non-allocated object
244 * Handle of allocated object.
246 unsigned long handle;
253 struct size_class *size_class[ZS_SIZE_CLASSES];
254 struct kmem_cache *handle_cachep;
255 struct kmem_cache *zspage_cachep;
257 atomic_long_t pages_allocated;
259 struct zs_pool_stats stats;
261 /* Compact classes */
262 struct shrinker shrinker;
264 #ifdef CONFIG_ZSMALLOC_STAT
265 struct dentry *stat_dentry;
267 #ifdef CONFIG_COMPACTION
269 struct work_struct free_work;
275 unsigned int fullness:FULLNESS_BITS;
276 unsigned int class:CLASS_BITS + 1;
277 unsigned int isolated:ISOLATED_BITS;
278 unsigned int magic:MAGIC_VAL_BITS;
281 unsigned int freeobj;
282 struct page *first_page;
283 struct list_head list; /* fullness list */
284 #ifdef CONFIG_COMPACTION
289 struct mapping_area {
290 #ifdef CONFIG_PGTABLE_MAPPING
291 struct vm_struct *vm; /* vm area for mapping object that span pages */
293 char *vm_buf; /* copy buffer for objects that span pages */
295 char *vm_addr; /* address of kmap_atomic()'ed pages */
296 enum zs_mapmode vm_mm; /* mapping mode */
299 #ifdef CONFIG_COMPACTION
300 static int zs_register_migration(struct zs_pool *pool);
301 static void zs_unregister_migration(struct zs_pool *pool);
302 static void migrate_lock_init(struct zspage *zspage);
303 static void migrate_read_lock(struct zspage *zspage);
304 static void migrate_read_unlock(struct zspage *zspage);
305 static void kick_deferred_free(struct zs_pool *pool);
306 static void init_deferred_free(struct zs_pool *pool);
307 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
309 static int zsmalloc_mount(void) { return 0; }
310 static void zsmalloc_unmount(void) {}
311 static int zs_register_migration(struct zs_pool *pool) { return 0; }
312 static void zs_unregister_migration(struct zs_pool *pool) {}
313 static void migrate_lock_init(struct zspage *zspage) {}
314 static void migrate_read_lock(struct zspage *zspage) {}
315 static void migrate_read_unlock(struct zspage *zspage) {}
316 static void kick_deferred_free(struct zs_pool *pool) {}
317 static void init_deferred_free(struct zs_pool *pool) {}
318 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
321 static int create_cache(struct zs_pool *pool)
323 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
325 if (!pool->handle_cachep)
328 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
330 if (!pool->zspage_cachep) {
331 kmem_cache_destroy(pool->handle_cachep);
332 pool->handle_cachep = NULL;
339 static void destroy_cache(struct zs_pool *pool)
341 kmem_cache_destroy(pool->handle_cachep);
342 kmem_cache_destroy(pool->zspage_cachep);
345 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
347 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
348 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
351 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
353 kmem_cache_free(pool->handle_cachep, (void *)handle);
356 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
358 return kmem_cache_alloc(pool->zspage_cachep,
359 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
362 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
364 kmem_cache_free(pool->zspage_cachep, zspage);
367 static void record_obj(unsigned long handle, unsigned long obj)
370 * lsb of @obj represents handle lock while other bits
371 * represent object value the handle is pointing so
372 * updating shouldn't do store tearing.
374 WRITE_ONCE(*(unsigned long *)handle, obj);
381 static void *zs_zpool_create(const char *name, gfp_t gfp,
382 const struct zpool_ops *zpool_ops,
386 * Ignore global gfp flags: zs_malloc() may be invoked from
387 * different contexts and its caller must provide a valid
390 return zs_create_pool(name);
393 static void zs_zpool_destroy(void *pool)
395 zs_destroy_pool(pool);
398 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
399 unsigned long *handle)
401 *handle = zs_malloc(pool, size, gfp);
402 return *handle ? 0 : -1;
404 static void zs_zpool_free(void *pool, unsigned long handle)
406 zs_free(pool, handle);
409 static void *zs_zpool_map(void *pool, unsigned long handle,
410 enum zpool_mapmode mm)
412 enum zs_mapmode zs_mm;
421 case ZPOOL_MM_RW: /* fallthru */
427 return zs_map_object(pool, handle, zs_mm);
429 static void zs_zpool_unmap(void *pool, unsigned long handle)
431 zs_unmap_object(pool, handle);
434 static u64 zs_zpool_total_size(void *pool)
436 return zs_get_total_pages(pool) << PAGE_SHIFT;
439 static struct zpool_driver zs_zpool_driver = {
441 .owner = THIS_MODULE,
442 .create = zs_zpool_create,
443 .destroy = zs_zpool_destroy,
444 .malloc = zs_zpool_malloc,
445 .free = zs_zpool_free,
447 .unmap = zs_zpool_unmap,
448 .total_size = zs_zpool_total_size,
451 MODULE_ALIAS("zpool-zsmalloc");
452 #endif /* CONFIG_ZPOOL */
454 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
455 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
457 static bool is_zspage_isolated(struct zspage *zspage)
459 return zspage->isolated;
462 static __maybe_unused int is_first_page(struct page *page)
464 return PagePrivate(page);
467 /* Protected by class->lock */
468 static inline int get_zspage_inuse(struct zspage *zspage)
470 return zspage->inuse;
473 static inline void set_zspage_inuse(struct zspage *zspage, int val)
478 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
480 zspage->inuse += val;
483 static inline struct page *get_first_page(struct zspage *zspage)
485 struct page *first_page = zspage->first_page;
487 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
491 static inline int get_first_obj_offset(struct page *page)
496 static inline void set_first_obj_offset(struct page *page, int offset)
498 page->units = offset;
501 static inline unsigned int get_freeobj(struct zspage *zspage)
503 return zspage->freeobj;
506 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
508 zspage->freeobj = obj;
511 static void get_zspage_mapping(struct zspage *zspage,
512 unsigned int *class_idx,
513 enum fullness_group *fullness)
515 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
517 *fullness = zspage->fullness;
518 *class_idx = zspage->class;
521 static void set_zspage_mapping(struct zspage *zspage,
522 unsigned int class_idx,
523 enum fullness_group fullness)
525 zspage->class = class_idx;
526 zspage->fullness = fullness;
530 * zsmalloc divides the pool into various size classes where each
531 * class maintains a list of zspages where each zspage is divided
532 * into equal sized chunks. Each allocation falls into one of these
533 * classes depending on its size. This function returns index of the
534 * size class which has chunk size big enough to hold the give size.
536 static int get_size_class_index(int size)
540 if (likely(size > ZS_MIN_ALLOC_SIZE))
541 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
542 ZS_SIZE_CLASS_DELTA);
544 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
547 /* type can be of enum type zs_stat_type or fullness_group */
548 static inline void zs_stat_inc(struct size_class *class,
549 int type, unsigned long cnt)
551 class->stats.objs[type] += cnt;
554 /* type can be of enum type zs_stat_type or fullness_group */
555 static inline void zs_stat_dec(struct size_class *class,
556 int type, unsigned long cnt)
558 class->stats.objs[type] -= cnt;
561 /* type can be of enum type zs_stat_type or fullness_group */
562 static inline unsigned long zs_stat_get(struct size_class *class,
565 return class->stats.objs[type];
568 #ifdef CONFIG_ZSMALLOC_STAT
570 static void __init zs_stat_init(void)
572 if (!debugfs_initialized()) {
573 pr_warn("debugfs not available, stat dir not created\n");
577 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
579 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
582 static void __exit zs_stat_exit(void)
584 debugfs_remove_recursive(zs_stat_root);
587 static unsigned long zs_can_compact(struct size_class *class);
589 static int zs_stats_size_show(struct seq_file *s, void *v)
592 struct zs_pool *pool = s->private;
593 struct size_class *class;
595 unsigned long class_almost_full, class_almost_empty;
596 unsigned long obj_allocated, obj_used, pages_used, freeable;
597 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
598 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
599 unsigned long total_freeable = 0;
601 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
602 "class", "size", "almost_full", "almost_empty",
603 "obj_allocated", "obj_used", "pages_used",
604 "pages_per_zspage", "freeable");
606 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
607 class = pool->size_class[i];
609 if (class->index != i)
612 spin_lock(&class->lock);
613 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
614 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
615 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
616 obj_used = zs_stat_get(class, OBJ_USED);
617 freeable = zs_can_compact(class);
618 spin_unlock(&class->lock);
620 objs_per_zspage = class->objs_per_zspage;
621 pages_used = obj_allocated / objs_per_zspage *
622 class->pages_per_zspage;
624 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
625 " %10lu %10lu %16d %8lu\n",
626 i, class->size, class_almost_full, class_almost_empty,
627 obj_allocated, obj_used, pages_used,
628 class->pages_per_zspage, freeable);
630 total_class_almost_full += class_almost_full;
631 total_class_almost_empty += class_almost_empty;
632 total_objs += obj_allocated;
633 total_used_objs += obj_used;
634 total_pages += pages_used;
635 total_freeable += freeable;
639 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
640 "Total", "", total_class_almost_full,
641 total_class_almost_empty, total_objs,
642 total_used_objs, total_pages, "", total_freeable);
646 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
648 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
650 struct dentry *entry;
653 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
657 entry = debugfs_create_dir(name, zs_stat_root);
659 pr_warn("debugfs dir <%s> creation failed\n", name);
662 pool->stat_dentry = entry;
664 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
665 pool->stat_dentry, pool, &zs_stats_size_fops);
667 pr_warn("%s: debugfs file entry <%s> creation failed\n",
669 debugfs_remove_recursive(pool->stat_dentry);
670 pool->stat_dentry = NULL;
674 static void zs_pool_stat_destroy(struct zs_pool *pool)
676 debugfs_remove_recursive(pool->stat_dentry);
679 #else /* CONFIG_ZSMALLOC_STAT */
680 static void __init zs_stat_init(void)
684 static void __exit zs_stat_exit(void)
688 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
692 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
699 * For each size class, zspages are divided into different groups
700 * depending on how "full" they are. This was done so that we could
701 * easily find empty or nearly empty zspages when we try to shrink
702 * the pool (not yet implemented). This function returns fullness
703 * status of the given page.
705 static enum fullness_group get_fullness_group(struct size_class *class,
706 struct zspage *zspage)
708 int inuse, objs_per_zspage;
709 enum fullness_group fg;
711 inuse = get_zspage_inuse(zspage);
712 objs_per_zspage = class->objs_per_zspage;
716 else if (inuse == objs_per_zspage)
718 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
719 fg = ZS_ALMOST_EMPTY;
727 * Each size class maintains various freelists and zspages are assigned
728 * to one of these freelists based on the number of live objects they
729 * have. This functions inserts the given zspage into the freelist
730 * identified by <class, fullness_group>.
732 static void insert_zspage(struct size_class *class,
733 struct zspage *zspage,
734 enum fullness_group fullness)
738 zs_stat_inc(class, fullness, 1);
739 head = list_first_entry_or_null(&class->fullness_list[fullness],
740 struct zspage, list);
742 * We want to see more ZS_FULL pages and less almost empty/full.
743 * Put pages with higher ->inuse first.
746 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
747 list_add(&zspage->list, &head->list);
751 list_add(&zspage->list, &class->fullness_list[fullness]);
755 * This function removes the given zspage from the freelist identified
756 * by <class, fullness_group>.
758 static void remove_zspage(struct size_class *class,
759 struct zspage *zspage,
760 enum fullness_group fullness)
762 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
763 VM_BUG_ON(is_zspage_isolated(zspage));
765 list_del_init(&zspage->list);
766 zs_stat_dec(class, fullness, 1);
770 * Each size class maintains zspages in different fullness groups depending
771 * on the number of live objects they contain. When allocating or freeing
772 * objects, the fullness status of the page can change, say, from ALMOST_FULL
773 * to ALMOST_EMPTY when freeing an object. This function checks if such
774 * a status change has occurred for the given page and accordingly moves the
775 * page from the freelist of the old fullness group to that of the new
778 static enum fullness_group fix_fullness_group(struct size_class *class,
779 struct zspage *zspage)
782 enum fullness_group currfg, newfg;
784 get_zspage_mapping(zspage, &class_idx, &currfg);
785 newfg = get_fullness_group(class, zspage);
789 if (!is_zspage_isolated(zspage)) {
790 remove_zspage(class, zspage, currfg);
791 insert_zspage(class, zspage, newfg);
794 set_zspage_mapping(zspage, class_idx, newfg);
801 * We have to decide on how many pages to link together
802 * to form a zspage for each size class. This is important
803 * to reduce wastage due to unusable space left at end of
804 * each zspage which is given as:
805 * wastage = Zp % class_size
806 * usage = Zp - wastage
807 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
809 * For example, for size class of 3/8 * PAGE_SIZE, we should
810 * link together 3 PAGE_SIZE sized pages to form a zspage
811 * since then we can perfectly fit in 8 such objects.
813 static int get_pages_per_zspage(int class_size)
815 int i, max_usedpc = 0;
816 /* zspage order which gives maximum used size per KB */
817 int max_usedpc_order = 1;
819 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
823 zspage_size = i * PAGE_SIZE;
824 waste = zspage_size % class_size;
825 usedpc = (zspage_size - waste) * 100 / zspage_size;
827 if (usedpc > max_usedpc) {
829 max_usedpc_order = i;
833 return max_usedpc_order;
836 static struct zspage *get_zspage(struct page *page)
838 struct zspage *zspage = (struct zspage *)page->private;
840 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
844 static struct page *get_next_page(struct page *page)
846 if (unlikely(PageHugeObject(page)))
849 return page->freelist;
853 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
854 * @page: page object resides in zspage
855 * @obj_idx: object index
857 static void obj_to_location(unsigned long obj, struct page **page,
858 unsigned int *obj_idx)
860 obj >>= OBJ_TAG_BITS;
861 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
862 *obj_idx = (obj & OBJ_INDEX_MASK);
866 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
867 * @page: page object resides in zspage
868 * @obj_idx: object index
870 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
874 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
875 obj |= obj_idx & OBJ_INDEX_MASK;
876 obj <<= OBJ_TAG_BITS;
881 static unsigned long handle_to_obj(unsigned long handle)
883 return *(unsigned long *)handle;
886 static unsigned long obj_to_head(struct page *page, void *obj)
888 if (unlikely(PageHugeObject(page))) {
889 VM_BUG_ON_PAGE(!is_first_page(page), page);
892 return *(unsigned long *)obj;
895 static inline int testpin_tag(unsigned long handle)
897 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
900 static inline int trypin_tag(unsigned long handle)
902 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
905 static void pin_tag(unsigned long handle)
907 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
910 static void unpin_tag(unsigned long handle)
912 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
915 static void reset_page(struct page *page)
917 __ClearPageMovable(page);
918 ClearPagePrivate(page);
919 set_page_private(page, 0);
920 page_mapcount_reset(page);
921 ClearPageHugeObject(page);
922 page->freelist = NULL;
926 * To prevent zspage destroy during migration, zspage freeing should
927 * hold locks of all pages in the zspage.
929 void lock_zspage(struct zspage *zspage)
931 struct page *page = get_first_page(zspage);
935 } while ((page = get_next_page(page)) != NULL);
938 int trylock_zspage(struct zspage *zspage)
940 struct page *cursor, *fail;
942 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
943 get_next_page(cursor)) {
944 if (!trylock_page(cursor)) {
952 for (cursor = get_first_page(zspage); cursor != fail; cursor =
953 get_next_page(cursor))
959 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
960 struct zspage *zspage)
962 struct page *page, *next;
963 enum fullness_group fg;
964 unsigned int class_idx;
966 get_zspage_mapping(zspage, &class_idx, &fg);
968 assert_spin_locked(&class->lock);
970 VM_BUG_ON(get_zspage_inuse(zspage));
971 VM_BUG_ON(fg != ZS_EMPTY);
973 next = page = get_first_page(zspage);
975 VM_BUG_ON_PAGE(!PageLocked(page), page);
976 next = get_next_page(page);
979 dec_zone_page_state(page, NR_ZSPAGES);
982 } while (page != NULL);
984 cache_free_zspage(pool, zspage);
986 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
987 atomic_long_sub(class->pages_per_zspage,
988 &pool->pages_allocated);
991 static void free_zspage(struct zs_pool *pool, struct size_class *class,
992 struct zspage *zspage)
994 VM_BUG_ON(get_zspage_inuse(zspage));
995 VM_BUG_ON(list_empty(&zspage->list));
997 if (!trylock_zspage(zspage)) {
998 kick_deferred_free(pool);
1002 remove_zspage(class, zspage, ZS_EMPTY);
1003 __free_zspage(pool, class, zspage);
1006 /* Initialize a newly allocated zspage */
1007 static void init_zspage(struct size_class *class, struct zspage *zspage)
1009 unsigned int freeobj = 1;
1010 unsigned long off = 0;
1011 struct page *page = get_first_page(zspage);
1014 struct page *next_page;
1015 struct link_free *link;
1018 set_first_obj_offset(page, off);
1020 vaddr = kmap_atomic(page);
1021 link = (struct link_free *)vaddr + off / sizeof(*link);
1023 while ((off += class->size) < PAGE_SIZE) {
1024 link->next = freeobj++ << OBJ_TAG_BITS;
1025 link += class->size / sizeof(*link);
1029 * We now come to the last (full or partial) object on this
1030 * page, which must point to the first object on the next
1033 next_page = get_next_page(page);
1035 link->next = freeobj++ << OBJ_TAG_BITS;
1038 * Reset OBJ_TAG_BITS bit to last link to tell
1039 * whether it's allocated object or not.
1041 link->next = -1UL << OBJ_TAG_BITS;
1043 kunmap_atomic(vaddr);
1048 set_freeobj(zspage, 0);
1051 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1052 struct page *pages[])
1056 struct page *prev_page = NULL;
1057 int nr_pages = class->pages_per_zspage;
1060 * Allocate individual pages and link them together as:
1061 * 1. all pages are linked together using page->freelist
1062 * 2. each sub-page point to zspage using page->private
1064 * we set PG_private to identify the first page (i.e. no other sub-page
1065 * has this flag set).
1067 for (i = 0; i < nr_pages; i++) {
1069 set_page_private(page, (unsigned long)zspage);
1070 page->freelist = NULL;
1072 zspage->first_page = page;
1073 SetPagePrivate(page);
1074 if (unlikely(class->objs_per_zspage == 1 &&
1075 class->pages_per_zspage == 1))
1076 SetPageHugeObject(page);
1078 prev_page->freelist = page;
1085 * Allocate a zspage for the given size class
1087 static struct zspage *alloc_zspage(struct zs_pool *pool,
1088 struct size_class *class,
1092 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1093 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1098 memset(zspage, 0, sizeof(struct zspage));
1099 zspage->magic = ZSPAGE_MAGIC;
1100 migrate_lock_init(zspage);
1102 for (i = 0; i < class->pages_per_zspage; i++) {
1105 page = alloc_page(gfp);
1108 dec_zone_page_state(pages[i], NR_ZSPAGES);
1109 __free_page(pages[i]);
1111 cache_free_zspage(pool, zspage);
1115 inc_zone_page_state(page, NR_ZSPAGES);
1119 create_page_chain(class, zspage, pages);
1120 init_zspage(class, zspage);
1125 static struct zspage *find_get_zspage(struct size_class *class)
1128 struct zspage *zspage;
1130 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1131 zspage = list_first_entry_or_null(&class->fullness_list[i],
1132 struct zspage, list);
1140 #ifdef CONFIG_PGTABLE_MAPPING
1141 static inline int __zs_cpu_up(struct mapping_area *area)
1144 * Make sure we don't leak memory if a cpu UP notification
1145 * and zs_init() race and both call zs_cpu_up() on the same cpu
1149 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1155 static inline void __zs_cpu_down(struct mapping_area *area)
1158 free_vm_area(area->vm);
1162 static inline void *__zs_map_object(struct mapping_area *area,
1163 struct page *pages[2], int off, int size)
1165 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1166 area->vm_addr = area->vm->addr;
1167 return area->vm_addr + off;
1170 static inline void __zs_unmap_object(struct mapping_area *area,
1171 struct page *pages[2], int off, int size)
1173 unsigned long addr = (unsigned long)area->vm_addr;
1175 unmap_kernel_range(addr, PAGE_SIZE * 2);
1178 #else /* CONFIG_PGTABLE_MAPPING */
1180 static inline int __zs_cpu_up(struct mapping_area *area)
1183 * Make sure we don't leak memory if a cpu UP notification
1184 * and zs_init() race and both call zs_cpu_up() on the same cpu
1188 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1194 static inline void __zs_cpu_down(struct mapping_area *area)
1196 kfree(area->vm_buf);
1197 area->vm_buf = NULL;
1200 static void *__zs_map_object(struct mapping_area *area,
1201 struct page *pages[2], int off, int size)
1205 char *buf = area->vm_buf;
1207 /* disable page faults to match kmap_atomic() return conditions */
1208 pagefault_disable();
1210 /* no read fastpath */
1211 if (area->vm_mm == ZS_MM_WO)
1214 sizes[0] = PAGE_SIZE - off;
1215 sizes[1] = size - sizes[0];
1217 /* copy object to per-cpu buffer */
1218 addr = kmap_atomic(pages[0]);
1219 memcpy(buf, addr + off, sizes[0]);
1220 kunmap_atomic(addr);
1221 addr = kmap_atomic(pages[1]);
1222 memcpy(buf + sizes[0], addr, sizes[1]);
1223 kunmap_atomic(addr);
1225 return area->vm_buf;
1228 static void __zs_unmap_object(struct mapping_area *area,
1229 struct page *pages[2], int off, int size)
1235 /* no write fastpath */
1236 if (area->vm_mm == ZS_MM_RO)
1240 buf = buf + ZS_HANDLE_SIZE;
1241 size -= ZS_HANDLE_SIZE;
1242 off += ZS_HANDLE_SIZE;
1244 sizes[0] = PAGE_SIZE - off;
1245 sizes[1] = size - sizes[0];
1247 /* copy per-cpu buffer to object */
1248 addr = kmap_atomic(pages[0]);
1249 memcpy(addr + off, buf, sizes[0]);
1250 kunmap_atomic(addr);
1251 addr = kmap_atomic(pages[1]);
1252 memcpy(addr, buf + sizes[0], sizes[1]);
1253 kunmap_atomic(addr);
1256 /* enable page faults to match kunmap_atomic() return conditions */
1260 #endif /* CONFIG_PGTABLE_MAPPING */
1262 static int zs_cpu_prepare(unsigned int cpu)
1264 struct mapping_area *area;
1266 area = &per_cpu(zs_map_area, cpu);
1267 return __zs_cpu_up(area);
1270 static int zs_cpu_dead(unsigned int cpu)
1272 struct mapping_area *area;
1274 area = &per_cpu(zs_map_area, cpu);
1275 __zs_cpu_down(area);
1279 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1280 int objs_per_zspage)
1282 if (prev->pages_per_zspage == pages_per_zspage &&
1283 prev->objs_per_zspage == objs_per_zspage)
1289 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1291 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1294 unsigned long zs_get_total_pages(struct zs_pool *pool)
1296 return atomic_long_read(&pool->pages_allocated);
1298 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1301 * zs_map_object - get address of allocated object from handle.
1302 * @pool: pool from which the object was allocated
1303 * @handle: handle returned from zs_malloc
1305 * Before using an object allocated from zs_malloc, it must be mapped using
1306 * this function. When done with the object, it must be unmapped using
1309 * Only one object can be mapped per cpu at a time. There is no protection
1310 * against nested mappings.
1312 * This function returns with preemption and page faults disabled.
1314 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1317 struct zspage *zspage;
1319 unsigned long obj, off;
1320 unsigned int obj_idx;
1322 unsigned int class_idx;
1323 enum fullness_group fg;
1324 struct size_class *class;
1325 struct mapping_area *area;
1326 struct page *pages[2];
1330 * Because we use per-cpu mapping areas shared among the
1331 * pools/users, we can't allow mapping in interrupt context
1332 * because it can corrupt another users mappings.
1334 BUG_ON(in_interrupt());
1336 /* From now on, migration cannot move the object */
1339 obj = handle_to_obj(handle);
1340 obj_to_location(obj, &page, &obj_idx);
1341 zspage = get_zspage(page);
1343 /* migration cannot move any subpage in this zspage */
1344 migrate_read_lock(zspage);
1346 get_zspage_mapping(zspage, &class_idx, &fg);
1347 class = pool->size_class[class_idx];
1348 off = (class->size * obj_idx) & ~PAGE_MASK;
1350 area = &get_cpu_var(zs_map_area);
1352 if (off + class->size <= PAGE_SIZE) {
1353 /* this object is contained entirely within a page */
1354 area->vm_addr = kmap_atomic(page);
1355 ret = area->vm_addr + off;
1359 /* this object spans two pages */
1361 pages[1] = get_next_page(page);
1364 ret = __zs_map_object(area, pages, off, class->size);
1366 if (likely(!PageHugeObject(page)))
1367 ret += ZS_HANDLE_SIZE;
1371 EXPORT_SYMBOL_GPL(zs_map_object);
1373 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1375 struct zspage *zspage;
1377 unsigned long obj, off;
1378 unsigned int obj_idx;
1380 unsigned int class_idx;
1381 enum fullness_group fg;
1382 struct size_class *class;
1383 struct mapping_area *area;
1385 obj = handle_to_obj(handle);
1386 obj_to_location(obj, &page, &obj_idx);
1387 zspage = get_zspage(page);
1388 get_zspage_mapping(zspage, &class_idx, &fg);
1389 class = pool->size_class[class_idx];
1390 off = (class->size * obj_idx) & ~PAGE_MASK;
1392 area = this_cpu_ptr(&zs_map_area);
1393 if (off + class->size <= PAGE_SIZE)
1394 kunmap_atomic(area->vm_addr);
1396 struct page *pages[2];
1399 pages[1] = get_next_page(page);
1402 __zs_unmap_object(area, pages, off, class->size);
1404 put_cpu_var(zs_map_area);
1406 migrate_read_unlock(zspage);
1409 EXPORT_SYMBOL_GPL(zs_unmap_object);
1412 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1413 * zsmalloc &size_class.
1414 * @pool: zsmalloc pool to use
1416 * The function returns the size of the first huge class - any object of equal
1417 * or bigger size will be stored in zspage consisting of a single physical
1420 * Context: Any context.
1422 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1424 size_t zs_huge_class_size(struct zs_pool *pool)
1426 return huge_class_size;
1428 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1430 static unsigned long obj_malloc(struct size_class *class,
1431 struct zspage *zspage, unsigned long handle)
1433 int i, nr_page, offset;
1435 struct link_free *link;
1437 struct page *m_page;
1438 unsigned long m_offset;
1441 handle |= OBJ_ALLOCATED_TAG;
1442 obj = get_freeobj(zspage);
1444 offset = obj * class->size;
1445 nr_page = offset >> PAGE_SHIFT;
1446 m_offset = offset & ~PAGE_MASK;
1447 m_page = get_first_page(zspage);
1449 for (i = 0; i < nr_page; i++)
1450 m_page = get_next_page(m_page);
1452 vaddr = kmap_atomic(m_page);
1453 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1454 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1455 if (likely(!PageHugeObject(m_page)))
1456 /* record handle in the header of allocated chunk */
1457 link->handle = handle;
1459 /* record handle to page->index */
1460 zspage->first_page->index = handle;
1462 kunmap_atomic(vaddr);
1463 mod_zspage_inuse(zspage, 1);
1464 zs_stat_inc(class, OBJ_USED, 1);
1466 obj = location_to_obj(m_page, obj);
1473 * zs_malloc - Allocate block of given size from pool.
1474 * @pool: pool to allocate from
1475 * @size: size of block to allocate
1476 * @gfp: gfp flags when allocating object
1478 * On success, handle to the allocated object is returned,
1480 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1482 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1484 unsigned long handle, obj;
1485 struct size_class *class;
1486 enum fullness_group newfg;
1487 struct zspage *zspage;
1489 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1492 handle = cache_alloc_handle(pool, gfp);
1496 /* extra space in chunk to keep the handle */
1497 size += ZS_HANDLE_SIZE;
1498 class = pool->size_class[get_size_class_index(size)];
1500 spin_lock(&class->lock);
1501 zspage = find_get_zspage(class);
1502 if (likely(zspage)) {
1503 obj = obj_malloc(class, zspage, handle);
1504 /* Now move the zspage to another fullness group, if required */
1505 fix_fullness_group(class, zspage);
1506 record_obj(handle, obj);
1507 spin_unlock(&class->lock);
1512 spin_unlock(&class->lock);
1514 zspage = alloc_zspage(pool, class, gfp);
1516 cache_free_handle(pool, handle);
1520 spin_lock(&class->lock);
1521 obj = obj_malloc(class, zspage, handle);
1522 newfg = get_fullness_group(class, zspage);
1523 insert_zspage(class, zspage, newfg);
1524 set_zspage_mapping(zspage, class->index, newfg);
1525 record_obj(handle, obj);
1526 atomic_long_add(class->pages_per_zspage,
1527 &pool->pages_allocated);
1528 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1530 /* We completely set up zspage so mark them as movable */
1531 SetZsPageMovable(pool, zspage);
1532 spin_unlock(&class->lock);
1536 EXPORT_SYMBOL_GPL(zs_malloc);
1538 static void obj_free(struct size_class *class, unsigned long obj)
1540 struct link_free *link;
1541 struct zspage *zspage;
1542 struct page *f_page;
1543 unsigned long f_offset;
1544 unsigned int f_objidx;
1547 obj &= ~OBJ_ALLOCATED_TAG;
1548 obj_to_location(obj, &f_page, &f_objidx);
1549 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1550 zspage = get_zspage(f_page);
1552 vaddr = kmap_atomic(f_page);
1554 /* Insert this object in containing zspage's freelist */
1555 link = (struct link_free *)(vaddr + f_offset);
1556 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1557 kunmap_atomic(vaddr);
1558 set_freeobj(zspage, f_objidx);
1559 mod_zspage_inuse(zspage, -1);
1560 zs_stat_dec(class, OBJ_USED, 1);
1563 void zs_free(struct zs_pool *pool, unsigned long handle)
1565 struct zspage *zspage;
1566 struct page *f_page;
1568 unsigned int f_objidx;
1570 struct size_class *class;
1571 enum fullness_group fullness;
1574 if (unlikely(!handle))
1578 obj = handle_to_obj(handle);
1579 obj_to_location(obj, &f_page, &f_objidx);
1580 zspage = get_zspage(f_page);
1582 migrate_read_lock(zspage);
1584 get_zspage_mapping(zspage, &class_idx, &fullness);
1585 class = pool->size_class[class_idx];
1587 spin_lock(&class->lock);
1588 obj_free(class, obj);
1589 fullness = fix_fullness_group(class, zspage);
1590 if (fullness != ZS_EMPTY) {
1591 migrate_read_unlock(zspage);
1595 isolated = is_zspage_isolated(zspage);
1596 migrate_read_unlock(zspage);
1597 /* If zspage is isolated, zs_page_putback will free the zspage */
1598 if (likely(!isolated))
1599 free_zspage(pool, class, zspage);
1602 spin_unlock(&class->lock);
1604 cache_free_handle(pool, handle);
1606 EXPORT_SYMBOL_GPL(zs_free);
1608 static void zs_object_copy(struct size_class *class, unsigned long dst,
1611 struct page *s_page, *d_page;
1612 unsigned int s_objidx, d_objidx;
1613 unsigned long s_off, d_off;
1614 void *s_addr, *d_addr;
1615 int s_size, d_size, size;
1618 s_size = d_size = class->size;
1620 obj_to_location(src, &s_page, &s_objidx);
1621 obj_to_location(dst, &d_page, &d_objidx);
1623 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1624 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1626 if (s_off + class->size > PAGE_SIZE)
1627 s_size = PAGE_SIZE - s_off;
1629 if (d_off + class->size > PAGE_SIZE)
1630 d_size = PAGE_SIZE - d_off;
1632 s_addr = kmap_atomic(s_page);
1633 d_addr = kmap_atomic(d_page);
1636 size = min(s_size, d_size);
1637 memcpy(d_addr + d_off, s_addr + s_off, size);
1640 if (written == class->size)
1648 if (s_off >= PAGE_SIZE) {
1649 kunmap_atomic(d_addr);
1650 kunmap_atomic(s_addr);
1651 s_page = get_next_page(s_page);
1652 s_addr = kmap_atomic(s_page);
1653 d_addr = kmap_atomic(d_page);
1654 s_size = class->size - written;
1658 if (d_off >= PAGE_SIZE) {
1659 kunmap_atomic(d_addr);
1660 d_page = get_next_page(d_page);
1661 d_addr = kmap_atomic(d_page);
1662 d_size = class->size - written;
1667 kunmap_atomic(d_addr);
1668 kunmap_atomic(s_addr);
1672 * Find alloced object in zspage from index object and
1675 static unsigned long find_alloced_obj(struct size_class *class,
1676 struct page *page, int *obj_idx)
1680 int index = *obj_idx;
1681 unsigned long handle = 0;
1682 void *addr = kmap_atomic(page);
1684 offset = get_first_obj_offset(page);
1685 offset += class->size * index;
1687 while (offset < PAGE_SIZE) {
1688 head = obj_to_head(page, addr + offset);
1689 if (head & OBJ_ALLOCATED_TAG) {
1690 handle = head & ~OBJ_ALLOCATED_TAG;
1691 if (trypin_tag(handle))
1696 offset += class->size;
1700 kunmap_atomic(addr);
1707 struct zs_compact_control {
1708 /* Source spage for migration which could be a subpage of zspage */
1709 struct page *s_page;
1710 /* Destination page for migration which should be a first page
1712 struct page *d_page;
1713 /* Starting object index within @s_page which used for live object
1714 * in the subpage. */
1718 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1719 struct zs_compact_control *cc)
1721 unsigned long used_obj, free_obj;
1722 unsigned long handle;
1723 struct page *s_page = cc->s_page;
1724 struct page *d_page = cc->d_page;
1725 int obj_idx = cc->obj_idx;
1729 handle = find_alloced_obj(class, s_page, &obj_idx);
1731 s_page = get_next_page(s_page);
1738 /* Stop if there is no more space */
1739 if (zspage_full(class, get_zspage(d_page))) {
1745 used_obj = handle_to_obj(handle);
1746 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1747 zs_object_copy(class, free_obj, used_obj);
1750 * record_obj updates handle's value to free_obj and it will
1751 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1752 * breaks synchronization using pin_tag(e,g, zs_free) so
1753 * let's keep the lock bit.
1755 free_obj |= BIT(HANDLE_PIN_BIT);
1756 record_obj(handle, free_obj);
1758 obj_free(class, used_obj);
1761 /* Remember last position in this iteration */
1762 cc->s_page = s_page;
1763 cc->obj_idx = obj_idx;
1768 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1771 struct zspage *zspage;
1772 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1775 fg[0] = ZS_ALMOST_FULL;
1776 fg[1] = ZS_ALMOST_EMPTY;
1779 for (i = 0; i < 2; i++) {
1780 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1781 struct zspage, list);
1783 VM_BUG_ON(is_zspage_isolated(zspage));
1784 remove_zspage(class, zspage, fg[i]);
1793 * putback_zspage - add @zspage into right class's fullness list
1794 * @class: destination class
1795 * @zspage: target page
1797 * Return @zspage's fullness_group
1799 static enum fullness_group putback_zspage(struct size_class *class,
1800 struct zspage *zspage)
1802 enum fullness_group fullness;
1804 VM_BUG_ON(is_zspage_isolated(zspage));
1806 fullness = get_fullness_group(class, zspage);
1807 insert_zspage(class, zspage, fullness);
1808 set_zspage_mapping(zspage, class->index, fullness);
1813 #ifdef CONFIG_COMPACTION
1814 static struct dentry *zs_mount(struct file_system_type *fs_type,
1815 int flags, const char *dev_name, void *data)
1817 static const struct dentry_operations ops = {
1818 .d_dname = simple_dname,
1821 return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1824 static struct file_system_type zsmalloc_fs = {
1827 .kill_sb = kill_anon_super,
1830 static int zsmalloc_mount(void)
1834 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1835 if (IS_ERR(zsmalloc_mnt))
1836 ret = PTR_ERR(zsmalloc_mnt);
1841 static void zsmalloc_unmount(void)
1843 kern_unmount(zsmalloc_mnt);
1846 static void migrate_lock_init(struct zspage *zspage)
1848 rwlock_init(&zspage->lock);
1851 static void migrate_read_lock(struct zspage *zspage)
1853 read_lock(&zspage->lock);
1856 static void migrate_read_unlock(struct zspage *zspage)
1858 read_unlock(&zspage->lock);
1861 static void migrate_write_lock(struct zspage *zspage)
1863 write_lock(&zspage->lock);
1866 static void migrate_write_unlock(struct zspage *zspage)
1868 write_unlock(&zspage->lock);
1871 /* Number of isolated subpage for *page migration* in this zspage */
1872 static void inc_zspage_isolation(struct zspage *zspage)
1877 static void dec_zspage_isolation(struct zspage *zspage)
1882 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1883 struct page *newpage, struct page *oldpage)
1886 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1889 page = get_first_page(zspage);
1891 if (page == oldpage)
1892 pages[idx] = newpage;
1896 } while ((page = get_next_page(page)) != NULL);
1898 create_page_chain(class, zspage, pages);
1899 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1900 if (unlikely(PageHugeObject(oldpage)))
1901 newpage->index = oldpage->index;
1902 __SetPageMovable(newpage, page_mapping(oldpage));
1905 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1907 struct zs_pool *pool;
1908 struct size_class *class;
1910 enum fullness_group fullness;
1911 struct zspage *zspage;
1912 struct address_space *mapping;
1915 * Page is locked so zspage couldn't be destroyed. For detail, look at
1916 * lock_zspage in free_zspage.
1918 VM_BUG_ON_PAGE(!PageMovable(page), page);
1919 VM_BUG_ON_PAGE(PageIsolated(page), page);
1921 zspage = get_zspage(page);
1924 * Without class lock, fullness could be stale while class_idx is okay
1925 * because class_idx is constant unless page is freed so we should get
1926 * fullness again under class lock.
1928 get_zspage_mapping(zspage, &class_idx, &fullness);
1929 mapping = page_mapping(page);
1930 pool = mapping->private_data;
1931 class = pool->size_class[class_idx];
1933 spin_lock(&class->lock);
1934 if (get_zspage_inuse(zspage) == 0) {
1935 spin_unlock(&class->lock);
1939 /* zspage is isolated for object migration */
1940 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1941 spin_unlock(&class->lock);
1946 * If this is first time isolation for the zspage, isolate zspage from
1947 * size_class to prevent further object allocation from the zspage.
1949 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1950 get_zspage_mapping(zspage, &class_idx, &fullness);
1951 remove_zspage(class, zspage, fullness);
1954 inc_zspage_isolation(zspage);
1955 spin_unlock(&class->lock);
1960 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1961 struct page *page, enum migrate_mode mode)
1963 struct zs_pool *pool;
1964 struct size_class *class;
1966 enum fullness_group fullness;
1967 struct zspage *zspage;
1969 void *s_addr, *d_addr, *addr;
1971 unsigned long handle, head;
1972 unsigned long old_obj, new_obj;
1973 unsigned int obj_idx;
1977 * We cannot support the _NO_COPY case here, because copy needs to
1978 * happen under the zs lock, which does not work with
1979 * MIGRATE_SYNC_NO_COPY workflow.
1981 if (mode == MIGRATE_SYNC_NO_COPY)
1984 VM_BUG_ON_PAGE(!PageMovable(page), page);
1985 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1987 zspage = get_zspage(page);
1989 /* Concurrent compactor cannot migrate any subpage in zspage */
1990 migrate_write_lock(zspage);
1991 get_zspage_mapping(zspage, &class_idx, &fullness);
1992 pool = mapping->private_data;
1993 class = pool->size_class[class_idx];
1994 offset = get_first_obj_offset(page);
1996 spin_lock(&class->lock);
1997 if (!get_zspage_inuse(zspage)) {
1999 * Set "offset" to end of the page so that every loops
2000 * skips unnecessary object scanning.
2006 s_addr = kmap_atomic(page);
2007 while (pos < PAGE_SIZE) {
2008 head = obj_to_head(page, s_addr + pos);
2009 if (head & OBJ_ALLOCATED_TAG) {
2010 handle = head & ~OBJ_ALLOCATED_TAG;
2011 if (!trypin_tag(handle))
2018 * Here, any user cannot access all objects in the zspage so let's move.
2020 d_addr = kmap_atomic(newpage);
2021 memcpy(d_addr, s_addr, PAGE_SIZE);
2022 kunmap_atomic(d_addr);
2024 for (addr = s_addr + offset; addr < s_addr + pos;
2025 addr += class->size) {
2026 head = obj_to_head(page, addr);
2027 if (head & OBJ_ALLOCATED_TAG) {
2028 handle = head & ~OBJ_ALLOCATED_TAG;
2029 if (!testpin_tag(handle))
2032 old_obj = handle_to_obj(handle);
2033 obj_to_location(old_obj, &dummy, &obj_idx);
2034 new_obj = (unsigned long)location_to_obj(newpage,
2036 new_obj |= BIT(HANDLE_PIN_BIT);
2037 record_obj(handle, new_obj);
2041 replace_sub_page(class, zspage, newpage, page);
2044 dec_zspage_isolation(zspage);
2047 * Page migration is done so let's putback isolated zspage to
2048 * the list if @page is final isolated subpage in the zspage.
2050 if (!is_zspage_isolated(zspage))
2051 putback_zspage(class, zspage);
2057 ret = MIGRATEPAGE_SUCCESS;
2059 for (addr = s_addr + offset; addr < s_addr + pos;
2060 addr += class->size) {
2061 head = obj_to_head(page, addr);
2062 if (head & OBJ_ALLOCATED_TAG) {
2063 handle = head & ~OBJ_ALLOCATED_TAG;
2064 if (!testpin_tag(handle))
2069 kunmap_atomic(s_addr);
2070 spin_unlock(&class->lock);
2071 migrate_write_unlock(zspage);
2076 void zs_page_putback(struct page *page)
2078 struct zs_pool *pool;
2079 struct size_class *class;
2081 enum fullness_group fg;
2082 struct address_space *mapping;
2083 struct zspage *zspage;
2085 VM_BUG_ON_PAGE(!PageMovable(page), page);
2086 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2088 zspage = get_zspage(page);
2089 get_zspage_mapping(zspage, &class_idx, &fg);
2090 mapping = page_mapping(page);
2091 pool = mapping->private_data;
2092 class = pool->size_class[class_idx];
2094 spin_lock(&class->lock);
2095 dec_zspage_isolation(zspage);
2096 if (!is_zspage_isolated(zspage)) {
2097 fg = putback_zspage(class, zspage);
2099 * Due to page_lock, we cannot free zspage immediately
2103 schedule_work(&pool->free_work);
2105 spin_unlock(&class->lock);
2108 const struct address_space_operations zsmalloc_aops = {
2109 .isolate_page = zs_page_isolate,
2110 .migratepage = zs_page_migrate,
2111 .putback_page = zs_page_putback,
2114 static int zs_register_migration(struct zs_pool *pool)
2116 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2117 if (IS_ERR(pool->inode)) {
2122 pool->inode->i_mapping->private_data = pool;
2123 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2127 static void zs_unregister_migration(struct zs_pool *pool)
2129 flush_work(&pool->free_work);
2134 * Caller should hold page_lock of all pages in the zspage
2135 * In here, we cannot use zspage meta data.
2137 static void async_free_zspage(struct work_struct *work)
2140 struct size_class *class;
2141 unsigned int class_idx;
2142 enum fullness_group fullness;
2143 struct zspage *zspage, *tmp;
2144 LIST_HEAD(free_pages);
2145 struct zs_pool *pool = container_of(work, struct zs_pool,
2148 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2149 class = pool->size_class[i];
2150 if (class->index != i)
2153 spin_lock(&class->lock);
2154 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2155 spin_unlock(&class->lock);
2159 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2160 list_del(&zspage->list);
2161 lock_zspage(zspage);
2163 get_zspage_mapping(zspage, &class_idx, &fullness);
2164 VM_BUG_ON(fullness != ZS_EMPTY);
2165 class = pool->size_class[class_idx];
2166 spin_lock(&class->lock);
2167 __free_zspage(pool, pool->size_class[class_idx], zspage);
2168 spin_unlock(&class->lock);
2172 static void kick_deferred_free(struct zs_pool *pool)
2174 schedule_work(&pool->free_work);
2177 static void init_deferred_free(struct zs_pool *pool)
2179 INIT_WORK(&pool->free_work, async_free_zspage);
2182 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2184 struct page *page = get_first_page(zspage);
2187 WARN_ON(!trylock_page(page));
2188 __SetPageMovable(page, pool->inode->i_mapping);
2190 } while ((page = get_next_page(page)) != NULL);
2196 * Based on the number of unused allocated objects calculate
2197 * and return the number of pages that we can free.
2199 static unsigned long zs_can_compact(struct size_class *class)
2201 unsigned long obj_wasted;
2202 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2203 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2205 if (obj_allocated <= obj_used)
2208 obj_wasted = obj_allocated - obj_used;
2209 obj_wasted /= class->objs_per_zspage;
2211 return obj_wasted * class->pages_per_zspage;
2214 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2216 struct zs_compact_control cc;
2217 struct zspage *src_zspage;
2218 struct zspage *dst_zspage = NULL;
2220 spin_lock(&class->lock);
2221 while ((src_zspage = isolate_zspage(class, true))) {
2223 if (!zs_can_compact(class))
2227 cc.s_page = get_first_page(src_zspage);
2229 while ((dst_zspage = isolate_zspage(class, false))) {
2230 cc.d_page = get_first_page(dst_zspage);
2232 * If there is no more space in dst_page, resched
2233 * and see if anyone had allocated another zspage.
2235 if (!migrate_zspage(pool, class, &cc))
2238 putback_zspage(class, dst_zspage);
2241 /* Stop if we couldn't find slot */
2242 if (dst_zspage == NULL)
2245 putback_zspage(class, dst_zspage);
2246 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2247 free_zspage(pool, class, src_zspage);
2248 pool->stats.pages_compacted += class->pages_per_zspage;
2250 spin_unlock(&class->lock);
2252 spin_lock(&class->lock);
2256 putback_zspage(class, src_zspage);
2258 spin_unlock(&class->lock);
2261 unsigned long zs_compact(struct zs_pool *pool)
2264 struct size_class *class;
2266 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2267 class = pool->size_class[i];
2270 if (class->index != i)
2272 __zs_compact(pool, class);
2275 return pool->stats.pages_compacted;
2277 EXPORT_SYMBOL_GPL(zs_compact);
2279 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2281 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2283 EXPORT_SYMBOL_GPL(zs_pool_stats);
2285 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2286 struct shrink_control *sc)
2288 unsigned long pages_freed;
2289 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2292 pages_freed = pool->stats.pages_compacted;
2294 * Compact classes and calculate compaction delta.
2295 * Can run concurrently with a manually triggered
2296 * (by user) compaction.
2298 pages_freed = zs_compact(pool) - pages_freed;
2300 return pages_freed ? pages_freed : SHRINK_STOP;
2303 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2304 struct shrink_control *sc)
2307 struct size_class *class;
2308 unsigned long pages_to_free = 0;
2309 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2312 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2313 class = pool->size_class[i];
2316 if (class->index != i)
2319 pages_to_free += zs_can_compact(class);
2322 return pages_to_free;
2325 static void zs_unregister_shrinker(struct zs_pool *pool)
2327 unregister_shrinker(&pool->shrinker);
2330 static int zs_register_shrinker(struct zs_pool *pool)
2332 pool->shrinker.scan_objects = zs_shrinker_scan;
2333 pool->shrinker.count_objects = zs_shrinker_count;
2334 pool->shrinker.batch = 0;
2335 pool->shrinker.seeks = DEFAULT_SEEKS;
2337 return register_shrinker(&pool->shrinker);
2341 * zs_create_pool - Creates an allocation pool to work from.
2342 * @name: pool name to be created
2344 * This function must be called before anything when using
2345 * the zsmalloc allocator.
2347 * On success, a pointer to the newly created pool is returned,
2350 struct zs_pool *zs_create_pool(const char *name)
2353 struct zs_pool *pool;
2354 struct size_class *prev_class = NULL;
2356 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2360 init_deferred_free(pool);
2362 pool->name = kstrdup(name, GFP_KERNEL);
2366 if (create_cache(pool))
2370 * Iterate reversely, because, size of size_class that we want to use
2371 * for merging should be larger or equal to current size.
2373 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2375 int pages_per_zspage;
2376 int objs_per_zspage;
2377 struct size_class *class;
2380 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2381 if (size > ZS_MAX_ALLOC_SIZE)
2382 size = ZS_MAX_ALLOC_SIZE;
2383 pages_per_zspage = get_pages_per_zspage(size);
2384 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2387 * We iterate from biggest down to smallest classes,
2388 * so huge_class_size holds the size of the first huge
2389 * class. Any object bigger than or equal to that will
2390 * endup in the huge class.
2392 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2394 huge_class_size = size;
2396 * The object uses ZS_HANDLE_SIZE bytes to store the
2397 * handle. We need to subtract it, because zs_malloc()
2398 * unconditionally adds handle size before it performs
2399 * size class search - so object may be smaller than
2400 * huge class size, yet it still can end up in the huge
2401 * class because it grows by ZS_HANDLE_SIZE extra bytes
2402 * right before class lookup.
2404 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2408 * size_class is used for normal zsmalloc operation such
2409 * as alloc/free for that size. Although it is natural that we
2410 * have one size_class for each size, there is a chance that we
2411 * can get more memory utilization if we use one size_class for
2412 * many different sizes whose size_class have same
2413 * characteristics. So, we makes size_class point to
2414 * previous size_class if possible.
2417 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2418 pool->size_class[i] = prev_class;
2423 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2429 class->pages_per_zspage = pages_per_zspage;
2430 class->objs_per_zspage = objs_per_zspage;
2431 spin_lock_init(&class->lock);
2432 pool->size_class[i] = class;
2433 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2435 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2440 /* debug only, don't abort if it fails */
2441 zs_pool_stat_create(pool, name);
2443 if (zs_register_migration(pool))
2447 * Not critical since shrinker is only used to trigger internal
2448 * defragmentation of the pool which is pretty optional thing. If
2449 * registration fails we still can use the pool normally and user can
2450 * trigger compaction manually. Thus, ignore return code.
2452 zs_register_shrinker(pool);
2457 zs_destroy_pool(pool);
2460 EXPORT_SYMBOL_GPL(zs_create_pool);
2462 void zs_destroy_pool(struct zs_pool *pool)
2466 zs_unregister_shrinker(pool);
2467 zs_unregister_migration(pool);
2468 zs_pool_stat_destroy(pool);
2470 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2472 struct size_class *class = pool->size_class[i];
2477 if (class->index != i)
2480 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2481 if (!list_empty(&class->fullness_list[fg])) {
2482 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2489 destroy_cache(pool);
2493 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2495 static int __init zs_init(void)
2499 ret = zsmalloc_mount();
2503 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2504 zs_cpu_prepare, zs_cpu_dead);
2509 zpool_register_driver(&zs_zpool_driver);
2522 static void __exit zs_exit(void)
2525 zpool_unregister_driver(&zs_zpool_driver);
2528 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2533 module_init(zs_init);
2534 module_exit(zs_exit);
2536 MODULE_LICENSE("Dual BSD/GPL");
2537 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");