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/types.h>
50 #include <linux/debugfs.h>
51 #include <linux/zsmalloc.h>
52 #include <linux/zpool.h>
53 #include <linux/mount.h>
54 #include <linux/migrate.h>
55 #include <linux/pagemap.h>
57 #define ZSPAGE_MAGIC 0x58
60 * This must be power of 2 and greater than of equal to sizeof(link_free).
61 * These two conditions ensure that any 'struct link_free' itself doesn't
62 * span more than 1 page which avoids complex case of mapping 2 pages simply
63 * to restore link_free pointer values.
68 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
69 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
71 #define ZS_MAX_ZSPAGE_ORDER 2
72 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
74 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
77 * Object location (<PFN>, <obj_idx>) is encoded as
78 * as single (unsigned long) handle value.
80 * Note that object index <obj_idx> starts from 0.
82 * This is made more complicated by various memory models and PAE.
85 #ifndef MAX_PHYSMEM_BITS
86 #ifdef CONFIG_HIGHMEM64G
87 #define MAX_PHYSMEM_BITS 36
88 #else /* !CONFIG_HIGHMEM64G */
90 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
93 #define MAX_PHYSMEM_BITS BITS_PER_LONG
96 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
99 * Memory for allocating for handle keeps object position by
100 * encoding <page, obj_idx> and the encoded value has a room
101 * in least bit(ie, look at obj_to_location).
102 * We use the bit to synchronize between object access by
103 * user and migration.
105 #define HANDLE_PIN_BIT 0
108 * Head in allocated object should have OBJ_ALLOCATED_TAG
109 * to identify the object was allocated or not.
110 * It's okay to add the status bit in the least bit because
111 * header keeps handle which is 4byte-aligned address so we
112 * have room for two bit at least.
114 #define OBJ_ALLOCATED_TAG 1
115 #define OBJ_TAG_BITS 1
116 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
117 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
119 #define FULLNESS_BITS 2
121 #define ISOLATED_BITS 3
122 #define MAGIC_VAL_BITS 8
124 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
125 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
126 #define ZS_MIN_ALLOC_SIZE \
127 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
128 /* each chunk includes extra space to keep handle */
129 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
132 * On systems with 4K page size, this gives 255 size classes! There is a
134 * - Large number of size classes is potentially wasteful as free page are
135 * spread across these classes
136 * - Small number of size classes causes large internal fragmentation
137 * - Probably its better to use specific size classes (empirically
138 * determined). NOTE: all those class sizes must be set as multiple of
139 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
141 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
144 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
145 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
146 ZS_SIZE_CLASS_DELTA) + 1)
148 enum fullness_group {
166 struct zs_size_stat {
167 unsigned long objs[NR_ZS_STAT_TYPE];
170 #ifdef CONFIG_ZSMALLOC_STAT
171 static struct dentry *zs_stat_root;
174 #ifdef CONFIG_COMPACTION
175 static struct vfsmount *zsmalloc_mnt;
179 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
181 * n = number of allocated objects
182 * N = total number of objects zspage can store
183 * f = fullness_threshold_frac
185 * Similarly, we assign zspage to:
186 * ZS_ALMOST_FULL when n > N / f
187 * ZS_EMPTY when n == 0
188 * ZS_FULL when n == N
190 * (see: fix_fullness_group())
192 static const int fullness_threshold_frac = 4;
196 struct list_head fullness_list[NR_ZS_FULLNESS];
198 * Size of objects stored in this class. Must be multiple
203 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
204 int pages_per_zspage;
207 struct zs_size_stat stats;
210 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
211 static void SetPageHugeObject(struct page *page)
213 SetPageOwnerPriv1(page);
216 static void ClearPageHugeObject(struct page *page)
218 ClearPageOwnerPriv1(page);
221 static int PageHugeObject(struct page *page)
223 return PageOwnerPriv1(page);
227 * Placed within free objects to form a singly linked list.
228 * For every zspage, zspage->freeobj gives head of this list.
230 * This must be power of 2 and less than or equal to ZS_ALIGN
236 * It's valid for non-allocated object
240 * Handle of allocated object.
242 unsigned long handle;
249 struct size_class *size_class[ZS_SIZE_CLASSES];
250 struct kmem_cache *handle_cachep;
251 struct kmem_cache *zspage_cachep;
253 atomic_long_t pages_allocated;
255 struct zs_pool_stats stats;
257 /* Compact classes */
258 struct shrinker shrinker;
260 * To signify that register_shrinker() was successful
261 * and unregister_shrinker() will not Oops.
263 bool shrinker_enabled;
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 int zs_zpool_shrink(void *pool, unsigned int pages,
410 unsigned int *reclaimed)
415 static void *zs_zpool_map(void *pool, unsigned long handle,
416 enum zpool_mapmode mm)
418 enum zs_mapmode zs_mm;
427 case ZPOOL_MM_RW: /* fallthru */
433 return zs_map_object(pool, handle, zs_mm);
435 static void zs_zpool_unmap(void *pool, unsigned long handle)
437 zs_unmap_object(pool, handle);
440 static u64 zs_zpool_total_size(void *pool)
442 return zs_get_total_pages(pool) << PAGE_SHIFT;
445 static struct zpool_driver zs_zpool_driver = {
447 .owner = THIS_MODULE,
448 .create = zs_zpool_create,
449 .destroy = zs_zpool_destroy,
450 .malloc = zs_zpool_malloc,
451 .free = zs_zpool_free,
452 .shrink = zs_zpool_shrink,
454 .unmap = zs_zpool_unmap,
455 .total_size = zs_zpool_total_size,
458 MODULE_ALIAS("zpool-zsmalloc");
459 #endif /* CONFIG_ZPOOL */
461 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
462 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
464 static bool is_zspage_isolated(struct zspage *zspage)
466 return zspage->isolated;
469 static __maybe_unused int is_first_page(struct page *page)
471 return PagePrivate(page);
474 /* Protected by class->lock */
475 static inline int get_zspage_inuse(struct zspage *zspage)
477 return zspage->inuse;
480 static inline void set_zspage_inuse(struct zspage *zspage, int val)
485 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
487 zspage->inuse += val;
490 static inline struct page *get_first_page(struct zspage *zspage)
492 struct page *first_page = zspage->first_page;
494 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
498 static inline int get_first_obj_offset(struct page *page)
503 static inline void set_first_obj_offset(struct page *page, int offset)
505 page->units = offset;
508 static inline unsigned int get_freeobj(struct zspage *zspage)
510 return zspage->freeobj;
513 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
515 zspage->freeobj = obj;
518 static void get_zspage_mapping(struct zspage *zspage,
519 unsigned int *class_idx,
520 enum fullness_group *fullness)
522 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
524 *fullness = zspage->fullness;
525 *class_idx = zspage->class;
528 static void set_zspage_mapping(struct zspage *zspage,
529 unsigned int class_idx,
530 enum fullness_group fullness)
532 zspage->class = class_idx;
533 zspage->fullness = fullness;
537 * zsmalloc divides the pool into various size classes where each
538 * class maintains a list of zspages where each zspage is divided
539 * into equal sized chunks. Each allocation falls into one of these
540 * classes depending on its size. This function returns index of the
541 * size class which has chunk size big enough to hold the give size.
543 static int get_size_class_index(int size)
547 if (likely(size > ZS_MIN_ALLOC_SIZE))
548 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
549 ZS_SIZE_CLASS_DELTA);
551 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
554 /* type can be of enum type zs_stat_type or fullness_group */
555 static inline void zs_stat_inc(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 void zs_stat_dec(struct size_class *class,
563 int type, unsigned long cnt)
565 class->stats.objs[type] -= cnt;
568 /* type can be of enum type zs_stat_type or fullness_group */
569 static inline unsigned long zs_stat_get(struct size_class *class,
572 return class->stats.objs[type];
575 #ifdef CONFIG_ZSMALLOC_STAT
577 static void __init zs_stat_init(void)
579 if (!debugfs_initialized()) {
580 pr_warn("debugfs not available, stat dir not created\n");
584 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
586 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
589 static void __exit zs_stat_exit(void)
591 debugfs_remove_recursive(zs_stat_root);
594 static unsigned long zs_can_compact(struct size_class *class);
596 static int zs_stats_size_show(struct seq_file *s, void *v)
599 struct zs_pool *pool = s->private;
600 struct size_class *class;
602 unsigned long class_almost_full, class_almost_empty;
603 unsigned long obj_allocated, obj_used, pages_used, freeable;
604 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
605 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
606 unsigned long total_freeable = 0;
608 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
609 "class", "size", "almost_full", "almost_empty",
610 "obj_allocated", "obj_used", "pages_used",
611 "pages_per_zspage", "freeable");
613 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
614 class = pool->size_class[i];
616 if (class->index != i)
619 spin_lock(&class->lock);
620 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
621 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
622 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
623 obj_used = zs_stat_get(class, OBJ_USED);
624 freeable = zs_can_compact(class);
625 spin_unlock(&class->lock);
627 objs_per_zspage = class->objs_per_zspage;
628 pages_used = obj_allocated / objs_per_zspage *
629 class->pages_per_zspage;
631 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
632 " %10lu %10lu %16d %8lu\n",
633 i, class->size, class_almost_full, class_almost_empty,
634 obj_allocated, obj_used, pages_used,
635 class->pages_per_zspage, freeable);
637 total_class_almost_full += class_almost_full;
638 total_class_almost_empty += class_almost_empty;
639 total_objs += obj_allocated;
640 total_used_objs += obj_used;
641 total_pages += pages_used;
642 total_freeable += freeable;
646 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
647 "Total", "", total_class_almost_full,
648 total_class_almost_empty, total_objs,
649 total_used_objs, total_pages, "", total_freeable);
654 static int zs_stats_size_open(struct inode *inode, struct file *file)
656 return single_open(file, zs_stats_size_show, inode->i_private);
659 static const struct file_operations zs_stat_size_ops = {
660 .open = zs_stats_size_open,
663 .release = single_release,
666 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
668 struct dentry *entry;
671 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
675 entry = debugfs_create_dir(name, zs_stat_root);
677 pr_warn("debugfs dir <%s> creation failed\n", name);
680 pool->stat_dentry = entry;
682 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
683 pool->stat_dentry, pool, &zs_stat_size_ops);
685 pr_warn("%s: debugfs file entry <%s> creation failed\n",
687 debugfs_remove_recursive(pool->stat_dentry);
688 pool->stat_dentry = NULL;
692 static void zs_pool_stat_destroy(struct zs_pool *pool)
694 debugfs_remove_recursive(pool->stat_dentry);
697 #else /* CONFIG_ZSMALLOC_STAT */
698 static void __init zs_stat_init(void)
702 static void __exit zs_stat_exit(void)
706 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
710 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
717 * For each size class, zspages are divided into different groups
718 * depending on how "full" they are. This was done so that we could
719 * easily find empty or nearly empty zspages when we try to shrink
720 * the pool (not yet implemented). This function returns fullness
721 * status of the given page.
723 static enum fullness_group get_fullness_group(struct size_class *class,
724 struct zspage *zspage)
726 int inuse, objs_per_zspage;
727 enum fullness_group fg;
729 inuse = get_zspage_inuse(zspage);
730 objs_per_zspage = class->objs_per_zspage;
734 else if (inuse == objs_per_zspage)
736 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
737 fg = ZS_ALMOST_EMPTY;
745 * Each size class maintains various freelists and zspages are assigned
746 * to one of these freelists based on the number of live objects they
747 * have. This functions inserts the given zspage into the freelist
748 * identified by <class, fullness_group>.
750 static void insert_zspage(struct size_class *class,
751 struct zspage *zspage,
752 enum fullness_group fullness)
756 zs_stat_inc(class, fullness, 1);
757 head = list_first_entry_or_null(&class->fullness_list[fullness],
758 struct zspage, list);
760 * We want to see more ZS_FULL pages and less almost empty/full.
761 * Put pages with higher ->inuse first.
764 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
765 list_add(&zspage->list, &head->list);
769 list_add(&zspage->list, &class->fullness_list[fullness]);
773 * This function removes the given zspage from the freelist identified
774 * by <class, fullness_group>.
776 static void remove_zspage(struct size_class *class,
777 struct zspage *zspage,
778 enum fullness_group fullness)
780 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
781 VM_BUG_ON(is_zspage_isolated(zspage));
783 list_del_init(&zspage->list);
784 zs_stat_dec(class, fullness, 1);
788 * Each size class maintains zspages in different fullness groups depending
789 * on the number of live objects they contain. When allocating or freeing
790 * objects, the fullness status of the page can change, say, from ALMOST_FULL
791 * to ALMOST_EMPTY when freeing an object. This function checks if such
792 * a status change has occurred for the given page and accordingly moves the
793 * page from the freelist of the old fullness group to that of the new
796 static enum fullness_group fix_fullness_group(struct size_class *class,
797 struct zspage *zspage)
800 enum fullness_group currfg, newfg;
802 get_zspage_mapping(zspage, &class_idx, &currfg);
803 newfg = get_fullness_group(class, zspage);
807 if (!is_zspage_isolated(zspage)) {
808 remove_zspage(class, zspage, currfg);
809 insert_zspage(class, zspage, newfg);
812 set_zspage_mapping(zspage, class_idx, newfg);
819 * We have to decide on how many pages to link together
820 * to form a zspage for each size class. This is important
821 * to reduce wastage due to unusable space left at end of
822 * each zspage which is given as:
823 * wastage = Zp % class_size
824 * usage = Zp - wastage
825 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
827 * For example, for size class of 3/8 * PAGE_SIZE, we should
828 * link together 3 PAGE_SIZE sized pages to form a zspage
829 * since then we can perfectly fit in 8 such objects.
831 static int get_pages_per_zspage(int class_size)
833 int i, max_usedpc = 0;
834 /* zspage order which gives maximum used size per KB */
835 int max_usedpc_order = 1;
837 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
841 zspage_size = i * PAGE_SIZE;
842 waste = zspage_size % class_size;
843 usedpc = (zspage_size - waste) * 100 / zspage_size;
845 if (usedpc > max_usedpc) {
847 max_usedpc_order = i;
851 return max_usedpc_order;
854 static struct zspage *get_zspage(struct page *page)
856 struct zspage *zspage = (struct zspage *)page->private;
858 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
862 static struct page *get_next_page(struct page *page)
864 if (unlikely(PageHugeObject(page)))
867 return page->freelist;
871 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
872 * @page: page object resides in zspage
873 * @obj_idx: object index
875 static void obj_to_location(unsigned long obj, struct page **page,
876 unsigned int *obj_idx)
878 obj >>= OBJ_TAG_BITS;
879 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
880 *obj_idx = (obj & OBJ_INDEX_MASK);
884 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
885 * @page: page object resides in zspage
886 * @obj_idx: object index
888 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
892 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
893 obj |= obj_idx & OBJ_INDEX_MASK;
894 obj <<= OBJ_TAG_BITS;
899 static unsigned long handle_to_obj(unsigned long handle)
901 return *(unsigned long *)handle;
904 static unsigned long obj_to_head(struct page *page, void *obj)
906 if (unlikely(PageHugeObject(page))) {
907 VM_BUG_ON_PAGE(!is_first_page(page), page);
910 return *(unsigned long *)obj;
913 static inline int testpin_tag(unsigned long handle)
915 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
918 static inline int trypin_tag(unsigned long handle)
920 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
923 static void pin_tag(unsigned long handle)
925 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
928 static void unpin_tag(unsigned long handle)
930 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
933 static void reset_page(struct page *page)
935 __ClearPageMovable(page);
936 ClearPagePrivate(page);
937 set_page_private(page, 0);
938 page_mapcount_reset(page);
939 ClearPageHugeObject(page);
940 page->freelist = NULL;
944 * To prevent zspage destroy during migration, zspage freeing should
945 * hold locks of all pages in the zspage.
947 void lock_zspage(struct zspage *zspage)
949 struct page *page = get_first_page(zspage);
953 } while ((page = get_next_page(page)) != NULL);
956 int trylock_zspage(struct zspage *zspage)
958 struct page *cursor, *fail;
960 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
961 get_next_page(cursor)) {
962 if (!trylock_page(cursor)) {
970 for (cursor = get_first_page(zspage); cursor != fail; cursor =
971 get_next_page(cursor))
977 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
978 struct zspage *zspage)
980 struct page *page, *next;
981 enum fullness_group fg;
982 unsigned int class_idx;
984 get_zspage_mapping(zspage, &class_idx, &fg);
986 assert_spin_locked(&class->lock);
988 VM_BUG_ON(get_zspage_inuse(zspage));
989 VM_BUG_ON(fg != ZS_EMPTY);
991 next = page = get_first_page(zspage);
993 VM_BUG_ON_PAGE(!PageLocked(page), page);
994 next = get_next_page(page);
997 dec_zone_page_state(page, NR_ZSPAGES);
1000 } while (page != NULL);
1002 cache_free_zspage(pool, zspage);
1004 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1005 atomic_long_sub(class->pages_per_zspage,
1006 &pool->pages_allocated);
1009 static void free_zspage(struct zs_pool *pool, struct size_class *class,
1010 struct zspage *zspage)
1012 VM_BUG_ON(get_zspage_inuse(zspage));
1013 VM_BUG_ON(list_empty(&zspage->list));
1015 if (!trylock_zspage(zspage)) {
1016 kick_deferred_free(pool);
1020 remove_zspage(class, zspage, ZS_EMPTY);
1021 __free_zspage(pool, class, zspage);
1024 /* Initialize a newly allocated zspage */
1025 static void init_zspage(struct size_class *class, struct zspage *zspage)
1027 unsigned int freeobj = 1;
1028 unsigned long off = 0;
1029 struct page *page = get_first_page(zspage);
1032 struct page *next_page;
1033 struct link_free *link;
1036 set_first_obj_offset(page, off);
1038 vaddr = kmap_atomic(page);
1039 link = (struct link_free *)vaddr + off / sizeof(*link);
1041 while ((off += class->size) < PAGE_SIZE) {
1042 link->next = freeobj++ << OBJ_TAG_BITS;
1043 link += class->size / sizeof(*link);
1047 * We now come to the last (full or partial) object on this
1048 * page, which must point to the first object on the next
1051 next_page = get_next_page(page);
1053 link->next = freeobj++ << OBJ_TAG_BITS;
1056 * Reset OBJ_TAG_BITS bit to last link to tell
1057 * whether it's allocated object or not.
1059 link->next = -1 << OBJ_TAG_BITS;
1061 kunmap_atomic(vaddr);
1066 set_freeobj(zspage, 0);
1069 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1070 struct page *pages[])
1074 struct page *prev_page = NULL;
1075 int nr_pages = class->pages_per_zspage;
1078 * Allocate individual pages and link them together as:
1079 * 1. all pages are linked together using page->freelist
1080 * 2. each sub-page point to zspage using page->private
1082 * we set PG_private to identify the first page (i.e. no other sub-page
1083 * has this flag set).
1085 for (i = 0; i < nr_pages; i++) {
1087 set_page_private(page, (unsigned long)zspage);
1088 page->freelist = NULL;
1090 zspage->first_page = page;
1091 SetPagePrivate(page);
1092 if (unlikely(class->objs_per_zspage == 1 &&
1093 class->pages_per_zspage == 1))
1094 SetPageHugeObject(page);
1096 prev_page->freelist = page;
1103 * Allocate a zspage for the given size class
1105 static struct zspage *alloc_zspage(struct zs_pool *pool,
1106 struct size_class *class,
1110 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1111 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1116 memset(zspage, 0, sizeof(struct zspage));
1117 zspage->magic = ZSPAGE_MAGIC;
1118 migrate_lock_init(zspage);
1120 for (i = 0; i < class->pages_per_zspage; i++) {
1123 page = alloc_page(gfp);
1126 dec_zone_page_state(pages[i], NR_ZSPAGES);
1127 __free_page(pages[i]);
1129 cache_free_zspage(pool, zspage);
1133 inc_zone_page_state(page, NR_ZSPAGES);
1137 create_page_chain(class, zspage, pages);
1138 init_zspage(class, zspage);
1143 static struct zspage *find_get_zspage(struct size_class *class)
1146 struct zspage *zspage;
1148 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1149 zspage = list_first_entry_or_null(&class->fullness_list[i],
1150 struct zspage, list);
1158 #ifdef CONFIG_PGTABLE_MAPPING
1159 static inline int __zs_cpu_up(struct mapping_area *area)
1162 * Make sure we don't leak memory if a cpu UP notification
1163 * and zs_init() race and both call zs_cpu_up() on the same cpu
1167 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1173 static inline void __zs_cpu_down(struct mapping_area *area)
1176 free_vm_area(area->vm);
1180 static inline void *__zs_map_object(struct mapping_area *area,
1181 struct page *pages[2], int off, int size)
1183 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1184 area->vm_addr = area->vm->addr;
1185 return area->vm_addr + off;
1188 static inline void __zs_unmap_object(struct mapping_area *area,
1189 struct page *pages[2], int off, int size)
1191 unsigned long addr = (unsigned long)area->vm_addr;
1193 unmap_kernel_range(addr, PAGE_SIZE * 2);
1196 #else /* CONFIG_PGTABLE_MAPPING */
1198 static inline int __zs_cpu_up(struct mapping_area *area)
1201 * Make sure we don't leak memory if a cpu UP notification
1202 * and zs_init() race and both call zs_cpu_up() on the same cpu
1206 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1212 static inline void __zs_cpu_down(struct mapping_area *area)
1214 kfree(area->vm_buf);
1215 area->vm_buf = NULL;
1218 static void *__zs_map_object(struct mapping_area *area,
1219 struct page *pages[2], int off, int size)
1223 char *buf = area->vm_buf;
1225 /* disable page faults to match kmap_atomic() return conditions */
1226 pagefault_disable();
1228 /* no read fastpath */
1229 if (area->vm_mm == ZS_MM_WO)
1232 sizes[0] = PAGE_SIZE - off;
1233 sizes[1] = size - sizes[0];
1235 /* copy object to per-cpu buffer */
1236 addr = kmap_atomic(pages[0]);
1237 memcpy(buf, addr + off, sizes[0]);
1238 kunmap_atomic(addr);
1239 addr = kmap_atomic(pages[1]);
1240 memcpy(buf + sizes[0], addr, sizes[1]);
1241 kunmap_atomic(addr);
1243 return area->vm_buf;
1246 static void __zs_unmap_object(struct mapping_area *area,
1247 struct page *pages[2], int off, int size)
1253 /* no write fastpath */
1254 if (area->vm_mm == ZS_MM_RO)
1258 buf = buf + ZS_HANDLE_SIZE;
1259 size -= ZS_HANDLE_SIZE;
1260 off += ZS_HANDLE_SIZE;
1262 sizes[0] = PAGE_SIZE - off;
1263 sizes[1] = size - sizes[0];
1265 /* copy per-cpu buffer to object */
1266 addr = kmap_atomic(pages[0]);
1267 memcpy(addr + off, buf, sizes[0]);
1268 kunmap_atomic(addr);
1269 addr = kmap_atomic(pages[1]);
1270 memcpy(addr, buf + sizes[0], sizes[1]);
1271 kunmap_atomic(addr);
1274 /* enable page faults to match kunmap_atomic() return conditions */
1278 #endif /* CONFIG_PGTABLE_MAPPING */
1280 static int zs_cpu_prepare(unsigned int cpu)
1282 struct mapping_area *area;
1284 area = &per_cpu(zs_map_area, cpu);
1285 return __zs_cpu_up(area);
1288 static int zs_cpu_dead(unsigned int cpu)
1290 struct mapping_area *area;
1292 area = &per_cpu(zs_map_area, cpu);
1293 __zs_cpu_down(area);
1297 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1298 int objs_per_zspage)
1300 if (prev->pages_per_zspage == pages_per_zspage &&
1301 prev->objs_per_zspage == objs_per_zspage)
1307 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1309 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1312 unsigned long zs_get_total_pages(struct zs_pool *pool)
1314 return atomic_long_read(&pool->pages_allocated);
1316 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1319 * zs_map_object - get address of allocated object from handle.
1320 * @pool: pool from which the object was allocated
1321 * @handle: handle returned from zs_malloc
1323 * Before using an object allocated from zs_malloc, it must be mapped using
1324 * this function. When done with the object, it must be unmapped using
1327 * Only one object can be mapped per cpu at a time. There is no protection
1328 * against nested mappings.
1330 * This function returns with preemption and page faults disabled.
1332 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1335 struct zspage *zspage;
1337 unsigned long obj, off;
1338 unsigned int obj_idx;
1340 unsigned int class_idx;
1341 enum fullness_group fg;
1342 struct size_class *class;
1343 struct mapping_area *area;
1344 struct page *pages[2];
1348 * Because we use per-cpu mapping areas shared among the
1349 * pools/users, we can't allow mapping in interrupt context
1350 * because it can corrupt another users mappings.
1352 BUG_ON(in_interrupt());
1354 /* From now on, migration cannot move the object */
1357 obj = handle_to_obj(handle);
1358 obj_to_location(obj, &page, &obj_idx);
1359 zspage = get_zspage(page);
1361 /* migration cannot move any subpage in this zspage */
1362 migrate_read_lock(zspage);
1364 get_zspage_mapping(zspage, &class_idx, &fg);
1365 class = pool->size_class[class_idx];
1366 off = (class->size * obj_idx) & ~PAGE_MASK;
1368 area = &get_cpu_var(zs_map_area);
1370 if (off + class->size <= PAGE_SIZE) {
1371 /* this object is contained entirely within a page */
1372 area->vm_addr = kmap_atomic(page);
1373 ret = area->vm_addr + off;
1377 /* this object spans two pages */
1379 pages[1] = get_next_page(page);
1382 ret = __zs_map_object(area, pages, off, class->size);
1384 if (likely(!PageHugeObject(page)))
1385 ret += ZS_HANDLE_SIZE;
1389 EXPORT_SYMBOL_GPL(zs_map_object);
1391 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1393 struct zspage *zspage;
1395 unsigned long obj, off;
1396 unsigned int obj_idx;
1398 unsigned int class_idx;
1399 enum fullness_group fg;
1400 struct size_class *class;
1401 struct mapping_area *area;
1403 obj = handle_to_obj(handle);
1404 obj_to_location(obj, &page, &obj_idx);
1405 zspage = get_zspage(page);
1406 get_zspage_mapping(zspage, &class_idx, &fg);
1407 class = pool->size_class[class_idx];
1408 off = (class->size * obj_idx) & ~PAGE_MASK;
1410 area = this_cpu_ptr(&zs_map_area);
1411 if (off + class->size <= PAGE_SIZE)
1412 kunmap_atomic(area->vm_addr);
1414 struct page *pages[2];
1417 pages[1] = get_next_page(page);
1420 __zs_unmap_object(area, pages, off, class->size);
1422 put_cpu_var(zs_map_area);
1424 migrate_read_unlock(zspage);
1427 EXPORT_SYMBOL_GPL(zs_unmap_object);
1429 static unsigned long obj_malloc(struct size_class *class,
1430 struct zspage *zspage, unsigned long handle)
1432 int i, nr_page, offset;
1434 struct link_free *link;
1436 struct page *m_page;
1437 unsigned long m_offset;
1440 handle |= OBJ_ALLOCATED_TAG;
1441 obj = get_freeobj(zspage);
1443 offset = obj * class->size;
1444 nr_page = offset >> PAGE_SHIFT;
1445 m_offset = offset & ~PAGE_MASK;
1446 m_page = get_first_page(zspage);
1448 for (i = 0; i < nr_page; i++)
1449 m_page = get_next_page(m_page);
1451 vaddr = kmap_atomic(m_page);
1452 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1453 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1454 if (likely(!PageHugeObject(m_page)))
1455 /* record handle in the header of allocated chunk */
1456 link->handle = handle;
1458 /* record handle to page->index */
1459 zspage->first_page->index = handle;
1461 kunmap_atomic(vaddr);
1462 mod_zspage_inuse(zspage, 1);
1463 zs_stat_inc(class, OBJ_USED, 1);
1465 obj = location_to_obj(m_page, obj);
1472 * zs_malloc - Allocate block of given size from pool.
1473 * @pool: pool to allocate from
1474 * @size: size of block to allocate
1475 * @gfp: gfp flags when allocating object
1477 * On success, handle to the allocated object is returned,
1479 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1481 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1483 unsigned long handle, obj;
1484 struct size_class *class;
1485 enum fullness_group newfg;
1486 struct zspage *zspage;
1488 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1491 handle = cache_alloc_handle(pool, gfp);
1495 /* extra space in chunk to keep the handle */
1496 size += ZS_HANDLE_SIZE;
1497 class = pool->size_class[get_size_class_index(size)];
1499 spin_lock(&class->lock);
1500 zspage = find_get_zspage(class);
1501 if (likely(zspage)) {
1502 obj = obj_malloc(class, zspage, handle);
1503 /* Now move the zspage to another fullness group, if required */
1504 fix_fullness_group(class, zspage);
1505 record_obj(handle, obj);
1506 spin_unlock(&class->lock);
1511 spin_unlock(&class->lock);
1513 zspage = alloc_zspage(pool, class, gfp);
1515 cache_free_handle(pool, handle);
1519 spin_lock(&class->lock);
1520 obj = obj_malloc(class, zspage, handle);
1521 newfg = get_fullness_group(class, zspage);
1522 insert_zspage(class, zspage, newfg);
1523 set_zspage_mapping(zspage, class->index, newfg);
1524 record_obj(handle, obj);
1525 atomic_long_add(class->pages_per_zspage,
1526 &pool->pages_allocated);
1527 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1529 /* We completely set up zspage so mark them as movable */
1530 SetZsPageMovable(pool, zspage);
1531 spin_unlock(&class->lock);
1535 EXPORT_SYMBOL_GPL(zs_malloc);
1537 static void obj_free(struct size_class *class, unsigned long obj)
1539 struct link_free *link;
1540 struct zspage *zspage;
1541 struct page *f_page;
1542 unsigned long f_offset;
1543 unsigned int f_objidx;
1546 obj &= ~OBJ_ALLOCATED_TAG;
1547 obj_to_location(obj, &f_page, &f_objidx);
1548 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1549 zspage = get_zspage(f_page);
1551 vaddr = kmap_atomic(f_page);
1553 /* Insert this object in containing zspage's freelist */
1554 link = (struct link_free *)(vaddr + f_offset);
1555 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1556 kunmap_atomic(vaddr);
1557 set_freeobj(zspage, f_objidx);
1558 mod_zspage_inuse(zspage, -1);
1559 zs_stat_dec(class, OBJ_USED, 1);
1562 void zs_free(struct zs_pool *pool, unsigned long handle)
1564 struct zspage *zspage;
1565 struct page *f_page;
1567 unsigned int f_objidx;
1569 struct size_class *class;
1570 enum fullness_group fullness;
1573 if (unlikely(!handle))
1577 obj = handle_to_obj(handle);
1578 obj_to_location(obj, &f_page, &f_objidx);
1579 zspage = get_zspage(f_page);
1581 migrate_read_lock(zspage);
1583 get_zspage_mapping(zspage, &class_idx, &fullness);
1584 class = pool->size_class[class_idx];
1586 spin_lock(&class->lock);
1587 obj_free(class, obj);
1588 fullness = fix_fullness_group(class, zspage);
1589 if (fullness != ZS_EMPTY) {
1590 migrate_read_unlock(zspage);
1594 isolated = is_zspage_isolated(zspage);
1595 migrate_read_unlock(zspage);
1596 /* If zspage is isolated, zs_page_putback will free the zspage */
1597 if (likely(!isolated))
1598 free_zspage(pool, class, zspage);
1601 spin_unlock(&class->lock);
1603 cache_free_handle(pool, handle);
1605 EXPORT_SYMBOL_GPL(zs_free);
1607 static void zs_object_copy(struct size_class *class, unsigned long dst,
1610 struct page *s_page, *d_page;
1611 unsigned int s_objidx, d_objidx;
1612 unsigned long s_off, d_off;
1613 void *s_addr, *d_addr;
1614 int s_size, d_size, size;
1617 s_size = d_size = class->size;
1619 obj_to_location(src, &s_page, &s_objidx);
1620 obj_to_location(dst, &d_page, &d_objidx);
1622 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1623 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1625 if (s_off + class->size > PAGE_SIZE)
1626 s_size = PAGE_SIZE - s_off;
1628 if (d_off + class->size > PAGE_SIZE)
1629 d_size = PAGE_SIZE - d_off;
1631 s_addr = kmap_atomic(s_page);
1632 d_addr = kmap_atomic(d_page);
1635 size = min(s_size, d_size);
1636 memcpy(d_addr + d_off, s_addr + s_off, size);
1639 if (written == class->size)
1647 if (s_off >= PAGE_SIZE) {
1648 kunmap_atomic(d_addr);
1649 kunmap_atomic(s_addr);
1650 s_page = get_next_page(s_page);
1651 s_addr = kmap_atomic(s_page);
1652 d_addr = kmap_atomic(d_page);
1653 s_size = class->size - written;
1657 if (d_off >= PAGE_SIZE) {
1658 kunmap_atomic(d_addr);
1659 d_page = get_next_page(d_page);
1660 d_addr = kmap_atomic(d_page);
1661 d_size = class->size - written;
1666 kunmap_atomic(d_addr);
1667 kunmap_atomic(s_addr);
1671 * Find alloced object in zspage from index object and
1674 static unsigned long find_alloced_obj(struct size_class *class,
1675 struct page *page, int *obj_idx)
1679 int index = *obj_idx;
1680 unsigned long handle = 0;
1681 void *addr = kmap_atomic(page);
1683 offset = get_first_obj_offset(page);
1684 offset += class->size * index;
1686 while (offset < PAGE_SIZE) {
1687 head = obj_to_head(page, addr + offset);
1688 if (head & OBJ_ALLOCATED_TAG) {
1689 handle = head & ~OBJ_ALLOCATED_TAG;
1690 if (trypin_tag(handle))
1695 offset += class->size;
1699 kunmap_atomic(addr);
1706 struct zs_compact_control {
1707 /* Source spage for migration which could be a subpage of zspage */
1708 struct page *s_page;
1709 /* Destination page for migration which should be a first page
1711 struct page *d_page;
1712 /* Starting object index within @s_page which used for live object
1713 * in the subpage. */
1717 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1718 struct zs_compact_control *cc)
1720 unsigned long used_obj, free_obj;
1721 unsigned long handle;
1722 struct page *s_page = cc->s_page;
1723 struct page *d_page = cc->d_page;
1724 int obj_idx = cc->obj_idx;
1728 handle = find_alloced_obj(class, s_page, &obj_idx);
1730 s_page = get_next_page(s_page);
1737 /* Stop if there is no more space */
1738 if (zspage_full(class, get_zspage(d_page))) {
1744 used_obj = handle_to_obj(handle);
1745 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1746 zs_object_copy(class, free_obj, used_obj);
1749 * record_obj updates handle's value to free_obj and it will
1750 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1751 * breaks synchronization using pin_tag(e,g, zs_free) so
1752 * let's keep the lock bit.
1754 free_obj |= BIT(HANDLE_PIN_BIT);
1755 record_obj(handle, free_obj);
1757 obj_free(class, used_obj);
1760 /* Remember last position in this iteration */
1761 cc->s_page = s_page;
1762 cc->obj_idx = obj_idx;
1767 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1770 struct zspage *zspage;
1771 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1774 fg[0] = ZS_ALMOST_FULL;
1775 fg[1] = ZS_ALMOST_EMPTY;
1778 for (i = 0; i < 2; i++) {
1779 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1780 struct zspage, list);
1782 VM_BUG_ON(is_zspage_isolated(zspage));
1783 remove_zspage(class, zspage, fg[i]);
1792 * putback_zspage - add @zspage into right class's fullness list
1793 * @class: destination class
1794 * @zspage: target page
1796 * Return @zspage's fullness_group
1798 static enum fullness_group putback_zspage(struct size_class *class,
1799 struct zspage *zspage)
1801 enum fullness_group fullness;
1803 VM_BUG_ON(is_zspage_isolated(zspage));
1805 fullness = get_fullness_group(class, zspage);
1806 insert_zspage(class, zspage, fullness);
1807 set_zspage_mapping(zspage, class->index, fullness);
1812 #ifdef CONFIG_COMPACTION
1813 static struct dentry *zs_mount(struct file_system_type *fs_type,
1814 int flags, const char *dev_name, void *data)
1816 static const struct dentry_operations ops = {
1817 .d_dname = simple_dname,
1820 return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1823 static struct file_system_type zsmalloc_fs = {
1826 .kill_sb = kill_anon_super,
1829 static int zsmalloc_mount(void)
1833 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1834 if (IS_ERR(zsmalloc_mnt))
1835 ret = PTR_ERR(zsmalloc_mnt);
1840 static void zsmalloc_unmount(void)
1842 kern_unmount(zsmalloc_mnt);
1845 static void migrate_lock_init(struct zspage *zspage)
1847 rwlock_init(&zspage->lock);
1850 static void migrate_read_lock(struct zspage *zspage)
1852 read_lock(&zspage->lock);
1855 static void migrate_read_unlock(struct zspage *zspage)
1857 read_unlock(&zspage->lock);
1860 static void migrate_write_lock(struct zspage *zspage)
1862 write_lock(&zspage->lock);
1865 static void migrate_write_unlock(struct zspage *zspage)
1867 write_unlock(&zspage->lock);
1870 /* Number of isolated subpage for *page migration* in this zspage */
1871 static void inc_zspage_isolation(struct zspage *zspage)
1876 static void dec_zspage_isolation(struct zspage *zspage)
1881 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1882 struct page *newpage, struct page *oldpage)
1885 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1888 page = get_first_page(zspage);
1890 if (page == oldpage)
1891 pages[idx] = newpage;
1895 } while ((page = get_next_page(page)) != NULL);
1897 create_page_chain(class, zspage, pages);
1898 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1899 if (unlikely(PageHugeObject(oldpage)))
1900 newpage->index = oldpage->index;
1901 __SetPageMovable(newpage, page_mapping(oldpage));
1904 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1906 struct zs_pool *pool;
1907 struct size_class *class;
1909 enum fullness_group fullness;
1910 struct zspage *zspage;
1911 struct address_space *mapping;
1914 * Page is locked so zspage couldn't be destroyed. For detail, look at
1915 * lock_zspage in free_zspage.
1917 VM_BUG_ON_PAGE(!PageMovable(page), page);
1918 VM_BUG_ON_PAGE(PageIsolated(page), page);
1920 zspage = get_zspage(page);
1923 * Without class lock, fullness could be stale while class_idx is okay
1924 * because class_idx is constant unless page is freed so we should get
1925 * fullness again under class lock.
1927 get_zspage_mapping(zspage, &class_idx, &fullness);
1928 mapping = page_mapping(page);
1929 pool = mapping->private_data;
1930 class = pool->size_class[class_idx];
1932 spin_lock(&class->lock);
1933 if (get_zspage_inuse(zspage) == 0) {
1934 spin_unlock(&class->lock);
1938 /* zspage is isolated for object migration */
1939 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1940 spin_unlock(&class->lock);
1945 * If this is first time isolation for the zspage, isolate zspage from
1946 * size_class to prevent further object allocation from the zspage.
1948 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1949 get_zspage_mapping(zspage, &class_idx, &fullness);
1950 remove_zspage(class, zspage, fullness);
1953 inc_zspage_isolation(zspage);
1954 spin_unlock(&class->lock);
1959 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1960 struct page *page, enum migrate_mode mode)
1962 struct zs_pool *pool;
1963 struct size_class *class;
1965 enum fullness_group fullness;
1966 struct zspage *zspage;
1968 void *s_addr, *d_addr, *addr;
1970 unsigned long handle, head;
1971 unsigned long old_obj, new_obj;
1972 unsigned int obj_idx;
1976 * We cannot support the _NO_COPY case here, because copy needs to
1977 * happen under the zs lock, which does not work with
1978 * MIGRATE_SYNC_NO_COPY workflow.
1980 if (mode == MIGRATE_SYNC_NO_COPY)
1983 VM_BUG_ON_PAGE(!PageMovable(page), page);
1984 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1986 zspage = get_zspage(page);
1988 /* Concurrent compactor cannot migrate any subpage in zspage */
1989 migrate_write_lock(zspage);
1990 get_zspage_mapping(zspage, &class_idx, &fullness);
1991 pool = mapping->private_data;
1992 class = pool->size_class[class_idx];
1993 offset = get_first_obj_offset(page);
1995 spin_lock(&class->lock);
1996 if (!get_zspage_inuse(zspage)) {
1998 * Set "offset" to end of the page so that every loops
1999 * skips unnecessary object scanning.
2005 s_addr = kmap_atomic(page);
2006 while (pos < PAGE_SIZE) {
2007 head = obj_to_head(page, s_addr + pos);
2008 if (head & OBJ_ALLOCATED_TAG) {
2009 handle = head & ~OBJ_ALLOCATED_TAG;
2010 if (!trypin_tag(handle))
2017 * Here, any user cannot access all objects in the zspage so let's move.
2019 d_addr = kmap_atomic(newpage);
2020 memcpy(d_addr, s_addr, PAGE_SIZE);
2021 kunmap_atomic(d_addr);
2023 for (addr = s_addr + offset; addr < s_addr + pos;
2024 addr += class->size) {
2025 head = obj_to_head(page, addr);
2026 if (head & OBJ_ALLOCATED_TAG) {
2027 handle = head & ~OBJ_ALLOCATED_TAG;
2028 if (!testpin_tag(handle))
2031 old_obj = handle_to_obj(handle);
2032 obj_to_location(old_obj, &dummy, &obj_idx);
2033 new_obj = (unsigned long)location_to_obj(newpage,
2035 new_obj |= BIT(HANDLE_PIN_BIT);
2036 record_obj(handle, new_obj);
2040 replace_sub_page(class, zspage, newpage, page);
2043 dec_zspage_isolation(zspage);
2046 * Page migration is done so let's putback isolated zspage to
2047 * the list if @page is final isolated subpage in the zspage.
2049 if (!is_zspage_isolated(zspage))
2050 putback_zspage(class, zspage);
2056 ret = MIGRATEPAGE_SUCCESS;
2058 for (addr = s_addr + offset; addr < s_addr + pos;
2059 addr += class->size) {
2060 head = obj_to_head(page, addr);
2061 if (head & OBJ_ALLOCATED_TAG) {
2062 handle = head & ~OBJ_ALLOCATED_TAG;
2063 if (!testpin_tag(handle))
2068 kunmap_atomic(s_addr);
2069 spin_unlock(&class->lock);
2070 migrate_write_unlock(zspage);
2075 void zs_page_putback(struct page *page)
2077 struct zs_pool *pool;
2078 struct size_class *class;
2080 enum fullness_group fg;
2081 struct address_space *mapping;
2082 struct zspage *zspage;
2084 VM_BUG_ON_PAGE(!PageMovable(page), page);
2085 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2087 zspage = get_zspage(page);
2088 get_zspage_mapping(zspage, &class_idx, &fg);
2089 mapping = page_mapping(page);
2090 pool = mapping->private_data;
2091 class = pool->size_class[class_idx];
2093 spin_lock(&class->lock);
2094 dec_zspage_isolation(zspage);
2095 if (!is_zspage_isolated(zspage)) {
2096 fg = putback_zspage(class, zspage);
2098 * Due to page_lock, we cannot free zspage immediately
2102 schedule_work(&pool->free_work);
2104 spin_unlock(&class->lock);
2107 const struct address_space_operations zsmalloc_aops = {
2108 .isolate_page = zs_page_isolate,
2109 .migratepage = zs_page_migrate,
2110 .putback_page = zs_page_putback,
2113 static int zs_register_migration(struct zs_pool *pool)
2115 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2116 if (IS_ERR(pool->inode)) {
2121 pool->inode->i_mapping->private_data = pool;
2122 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2126 static void zs_unregister_migration(struct zs_pool *pool)
2128 flush_work(&pool->free_work);
2133 * Caller should hold page_lock of all pages in the zspage
2134 * In here, we cannot use zspage meta data.
2136 static void async_free_zspage(struct work_struct *work)
2139 struct size_class *class;
2140 unsigned int class_idx;
2141 enum fullness_group fullness;
2142 struct zspage *zspage, *tmp;
2143 LIST_HEAD(free_pages);
2144 struct zs_pool *pool = container_of(work, struct zs_pool,
2147 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2148 class = pool->size_class[i];
2149 if (class->index != i)
2152 spin_lock(&class->lock);
2153 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2154 spin_unlock(&class->lock);
2158 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2159 list_del(&zspage->list);
2160 lock_zspage(zspage);
2162 get_zspage_mapping(zspage, &class_idx, &fullness);
2163 VM_BUG_ON(fullness != ZS_EMPTY);
2164 class = pool->size_class[class_idx];
2165 spin_lock(&class->lock);
2166 __free_zspage(pool, pool->size_class[class_idx], zspage);
2167 spin_unlock(&class->lock);
2171 static void kick_deferred_free(struct zs_pool *pool)
2173 schedule_work(&pool->free_work);
2176 static void init_deferred_free(struct zs_pool *pool)
2178 INIT_WORK(&pool->free_work, async_free_zspage);
2181 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2183 struct page *page = get_first_page(zspage);
2186 WARN_ON(!trylock_page(page));
2187 __SetPageMovable(page, pool->inode->i_mapping);
2189 } while ((page = get_next_page(page)) != NULL);
2195 * Based on the number of unused allocated objects calculate
2196 * and return the number of pages that we can free.
2198 static unsigned long zs_can_compact(struct size_class *class)
2200 unsigned long obj_wasted;
2201 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2202 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2204 if (obj_allocated <= obj_used)
2207 obj_wasted = obj_allocated - obj_used;
2208 obj_wasted /= class->objs_per_zspage;
2210 return obj_wasted * class->pages_per_zspage;
2213 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2215 struct zs_compact_control cc;
2216 struct zspage *src_zspage;
2217 struct zspage *dst_zspage = NULL;
2219 spin_lock(&class->lock);
2220 while ((src_zspage = isolate_zspage(class, true))) {
2222 if (!zs_can_compact(class))
2226 cc.s_page = get_first_page(src_zspage);
2228 while ((dst_zspage = isolate_zspage(class, false))) {
2229 cc.d_page = get_first_page(dst_zspage);
2231 * If there is no more space in dst_page, resched
2232 * and see if anyone had allocated another zspage.
2234 if (!migrate_zspage(pool, class, &cc))
2237 putback_zspage(class, dst_zspage);
2240 /* Stop if we couldn't find slot */
2241 if (dst_zspage == NULL)
2244 putback_zspage(class, dst_zspage);
2245 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2246 free_zspage(pool, class, src_zspage);
2247 pool->stats.pages_compacted += class->pages_per_zspage;
2249 spin_unlock(&class->lock);
2251 spin_lock(&class->lock);
2255 putback_zspage(class, src_zspage);
2257 spin_unlock(&class->lock);
2260 unsigned long zs_compact(struct zs_pool *pool)
2263 struct size_class *class;
2265 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2266 class = pool->size_class[i];
2269 if (class->index != i)
2271 __zs_compact(pool, class);
2274 return pool->stats.pages_compacted;
2276 EXPORT_SYMBOL_GPL(zs_compact);
2278 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2280 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2282 EXPORT_SYMBOL_GPL(zs_pool_stats);
2284 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2285 struct shrink_control *sc)
2287 unsigned long pages_freed;
2288 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2291 pages_freed = pool->stats.pages_compacted;
2293 * Compact classes and calculate compaction delta.
2294 * Can run concurrently with a manually triggered
2295 * (by user) compaction.
2297 pages_freed = zs_compact(pool) - pages_freed;
2299 return pages_freed ? pages_freed : SHRINK_STOP;
2302 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2303 struct shrink_control *sc)
2306 struct size_class *class;
2307 unsigned long pages_to_free = 0;
2308 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2311 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2312 class = pool->size_class[i];
2315 if (class->index != i)
2318 pages_to_free += zs_can_compact(class);
2321 return pages_to_free;
2324 static void zs_unregister_shrinker(struct zs_pool *pool)
2326 if (pool->shrinker_enabled) {
2327 unregister_shrinker(&pool->shrinker);
2328 pool->shrinker_enabled = false;
2332 static int zs_register_shrinker(struct zs_pool *pool)
2334 pool->shrinker.scan_objects = zs_shrinker_scan;
2335 pool->shrinker.count_objects = zs_shrinker_count;
2336 pool->shrinker.batch = 0;
2337 pool->shrinker.seeks = DEFAULT_SEEKS;
2339 return register_shrinker(&pool->shrinker);
2343 * zs_create_pool - Creates an allocation pool to work from.
2344 * @name: pool name to be created
2346 * This function must be called before anything when using
2347 * the zsmalloc allocator.
2349 * On success, a pointer to the newly created pool is returned,
2352 struct zs_pool *zs_create_pool(const char *name)
2355 struct zs_pool *pool;
2356 struct size_class *prev_class = NULL;
2358 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2362 init_deferred_free(pool);
2364 pool->name = kstrdup(name, GFP_KERNEL);
2368 if (create_cache(pool))
2372 * Iterate reversely, because, size of size_class that we want to use
2373 * for merging should be larger or equal to current size.
2375 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2377 int pages_per_zspage;
2378 int objs_per_zspage;
2379 struct size_class *class;
2382 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2383 if (size > ZS_MAX_ALLOC_SIZE)
2384 size = ZS_MAX_ALLOC_SIZE;
2385 pages_per_zspage = get_pages_per_zspage(size);
2386 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2389 * size_class is used for normal zsmalloc operation such
2390 * as alloc/free for that size. Although it is natural that we
2391 * have one size_class for each size, there is a chance that we
2392 * can get more memory utilization if we use one size_class for
2393 * many different sizes whose size_class have same
2394 * characteristics. So, we makes size_class point to
2395 * previous size_class if possible.
2398 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2399 pool->size_class[i] = prev_class;
2404 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2410 class->pages_per_zspage = pages_per_zspage;
2411 class->objs_per_zspage = objs_per_zspage;
2412 spin_lock_init(&class->lock);
2413 pool->size_class[i] = class;
2414 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2416 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2421 /* debug only, don't abort if it fails */
2422 zs_pool_stat_create(pool, name);
2424 if (zs_register_migration(pool))
2428 * Not critical, we still can use the pool
2429 * and user can trigger compaction manually.
2431 if (zs_register_shrinker(pool) == 0)
2432 pool->shrinker_enabled = true;
2436 zs_destroy_pool(pool);
2439 EXPORT_SYMBOL_GPL(zs_create_pool);
2441 void zs_destroy_pool(struct zs_pool *pool)
2445 zs_unregister_shrinker(pool);
2446 zs_unregister_migration(pool);
2447 zs_pool_stat_destroy(pool);
2449 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2451 struct size_class *class = pool->size_class[i];
2456 if (class->index != i)
2459 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2460 if (!list_empty(&class->fullness_list[fg])) {
2461 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2468 destroy_cache(pool);
2472 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2474 static int __init zs_init(void)
2478 ret = zsmalloc_mount();
2482 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2483 zs_cpu_prepare, zs_cpu_dead);
2488 zpool_register_driver(&zs_zpool_driver);
2501 static void __exit zs_exit(void)
2504 zpool_unregister_driver(&zs_zpool_driver);
2507 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2512 module_init(zs_init);
2513 module_exit(zs_exit);
2515 MODULE_LICENSE("Dual BSD/GPL");
2516 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");