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 <linux/pgtable.h>
43 #include <asm/tlbflush.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/pseudo_fs.h>
56 #include <linux/migrate.h>
57 #include <linux/wait.h>
58 #include <linux/pagemap.h>
61 #define ZSPAGE_MAGIC 0x58
64 * This must be power of 2 and greater than of equal to sizeof(link_free).
65 * These two conditions ensure that any 'struct link_free' itself doesn't
66 * span more than 1 page which avoids complex case of mapping 2 pages simply
67 * to restore link_free pointer values.
72 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
73 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
75 #define ZS_MAX_ZSPAGE_ORDER 2
76 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
78 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
81 * Object location (<PFN>, <obj_idx>) is encoded as
82 * a single (unsigned long) handle value.
84 * Note that object index <obj_idx> starts from 0.
86 * This is made more complicated by various memory models and PAE.
89 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
90 #ifdef MAX_PHYSMEM_BITS
91 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
94 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
97 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
101 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
104 * Memory for allocating for handle keeps object position by
105 * encoding <page, obj_idx> and the encoded value has a room
106 * in least bit(ie, look at obj_to_location).
107 * We use the bit to synchronize between object access by
108 * user and migration.
110 #define HANDLE_PIN_BIT 0
113 * Head in allocated object should have OBJ_ALLOCATED_TAG
114 * to identify the object was allocated or not.
115 * It's okay to add the status bit in the least bit because
116 * header keeps handle which is 4byte-aligned address so we
117 * have room for two bit at least.
119 #define OBJ_ALLOCATED_TAG 1
120 #define OBJ_TAG_BITS 1
121 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
122 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
124 #define FULLNESS_BITS 2
126 #define ISOLATED_BITS 3
127 #define MAGIC_VAL_BITS 8
129 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
130 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
131 #define ZS_MIN_ALLOC_SIZE \
132 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
133 /* each chunk includes extra space to keep handle */
134 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
137 * On systems with 4K page size, this gives 255 size classes! There is a
139 * - Large number of size classes is potentially wasteful as free page are
140 * spread across these classes
141 * - Small number of size classes causes large internal fragmentation
142 * - Probably its better to use specific size classes (empirically
143 * determined). NOTE: all those class sizes must be set as multiple of
144 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
146 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
149 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
150 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
151 ZS_SIZE_CLASS_DELTA) + 1)
153 enum fullness_group {
171 struct zs_size_stat {
172 unsigned long objs[NR_ZS_STAT_TYPE];
175 #ifdef CONFIG_ZSMALLOC_STAT
176 static struct dentry *zs_stat_root;
179 #ifdef CONFIG_COMPACTION
180 static struct vfsmount *zsmalloc_mnt;
184 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
186 * n = number of allocated objects
187 * N = total number of objects zspage can store
188 * f = fullness_threshold_frac
190 * Similarly, we assign zspage to:
191 * ZS_ALMOST_FULL when n > N / f
192 * ZS_EMPTY when n == 0
193 * ZS_FULL when n == N
195 * (see: fix_fullness_group())
197 static const int fullness_threshold_frac = 4;
198 static size_t huge_class_size;
202 struct list_head fullness_list[NR_ZS_FULLNESS];
204 * Size of objects stored in this class. Must be multiple
209 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
210 int pages_per_zspage;
213 struct zs_size_stat stats;
216 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
217 static void SetPageHugeObject(struct page *page)
219 SetPageOwnerPriv1(page);
222 static void ClearPageHugeObject(struct page *page)
224 ClearPageOwnerPriv1(page);
227 static int PageHugeObject(struct page *page)
229 return PageOwnerPriv1(page);
233 * Placed within free objects to form a singly linked list.
234 * For every zspage, zspage->freeobj gives head of this list.
236 * This must be power of 2 and less than or equal to ZS_ALIGN
242 * It's valid for non-allocated object
246 * Handle of allocated object.
248 unsigned long handle;
255 struct size_class *size_class[ZS_SIZE_CLASSES];
256 struct kmem_cache *handle_cachep;
257 struct kmem_cache *zspage_cachep;
259 atomic_long_t pages_allocated;
261 struct zs_pool_stats stats;
263 /* Compact classes */
264 struct shrinker shrinker;
266 #ifdef CONFIG_ZSMALLOC_STAT
267 struct dentry *stat_dentry;
269 #ifdef CONFIG_COMPACTION
271 struct work_struct free_work;
272 /* A wait queue for when migration races with async_free_zspage() */
273 struct wait_queue_head migration_wait;
274 atomic_long_t isolated_pages;
281 unsigned int fullness:FULLNESS_BITS;
282 unsigned int class:CLASS_BITS + 1;
283 unsigned int isolated:ISOLATED_BITS;
284 unsigned int magic:MAGIC_VAL_BITS;
287 unsigned int freeobj;
288 struct page *first_page;
289 struct list_head list; /* fullness list */
290 #ifdef CONFIG_COMPACTION
295 struct mapping_area {
296 char *vm_buf; /* copy buffer for objects that span pages */
297 char *vm_addr; /* address of kmap_atomic()'ed pages */
298 enum zs_mapmode vm_mm; /* mapping mode */
301 #ifdef CONFIG_COMPACTION
302 static int zs_register_migration(struct zs_pool *pool);
303 static void zs_unregister_migration(struct zs_pool *pool);
304 static void migrate_lock_init(struct zspage *zspage);
305 static void migrate_read_lock(struct zspage *zspage);
306 static void migrate_read_unlock(struct zspage *zspage);
307 static void kick_deferred_free(struct zs_pool *pool);
308 static void init_deferred_free(struct zs_pool *pool);
309 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
311 static int zsmalloc_mount(void) { return 0; }
312 static void zsmalloc_unmount(void) {}
313 static int zs_register_migration(struct zs_pool *pool) { return 0; }
314 static void zs_unregister_migration(struct zs_pool *pool) {}
315 static void migrate_lock_init(struct zspage *zspage) {}
316 static void migrate_read_lock(struct zspage *zspage) {}
317 static void migrate_read_unlock(struct zspage *zspage) {}
318 static void kick_deferred_free(struct zs_pool *pool) {}
319 static void init_deferred_free(struct zs_pool *pool) {}
320 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
323 static int create_cache(struct zs_pool *pool)
325 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
327 if (!pool->handle_cachep)
330 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
332 if (!pool->zspage_cachep) {
333 kmem_cache_destroy(pool->handle_cachep);
334 pool->handle_cachep = NULL;
341 static void destroy_cache(struct zs_pool *pool)
343 kmem_cache_destroy(pool->handle_cachep);
344 kmem_cache_destroy(pool->zspage_cachep);
347 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
349 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
350 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
353 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
355 kmem_cache_free(pool->handle_cachep, (void *)handle);
358 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
360 return kmem_cache_zalloc(pool->zspage_cachep,
361 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
364 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
366 kmem_cache_free(pool->zspage_cachep, zspage);
369 static void record_obj(unsigned long handle, unsigned long obj)
372 * lsb of @obj represents handle lock while other bits
373 * represent object value the handle is pointing so
374 * updating shouldn't do store tearing.
376 WRITE_ONCE(*(unsigned long *)handle, obj);
383 static void *zs_zpool_create(const char *name, gfp_t gfp,
384 const struct zpool_ops *zpool_ops,
388 * Ignore global gfp flags: zs_malloc() may be invoked from
389 * different contexts and its caller must provide a valid
392 return zs_create_pool(name);
395 static void zs_zpool_destroy(void *pool)
397 zs_destroy_pool(pool);
400 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
401 unsigned long *handle)
403 *handle = zs_malloc(pool, size, gfp);
404 return *handle ? 0 : -1;
406 static void zs_zpool_free(void *pool, unsigned long handle)
408 zs_free(pool, handle);
411 static void *zs_zpool_map(void *pool, unsigned long handle,
412 enum zpool_mapmode mm)
414 enum zs_mapmode zs_mm;
429 return zs_map_object(pool, handle, zs_mm);
431 static void zs_zpool_unmap(void *pool, unsigned long handle)
433 zs_unmap_object(pool, handle);
436 static u64 zs_zpool_total_size(void *pool)
438 return zs_get_total_pages(pool) << PAGE_SHIFT;
441 static struct zpool_driver zs_zpool_driver = {
443 .owner = THIS_MODULE,
444 .create = zs_zpool_create,
445 .destroy = zs_zpool_destroy,
446 .malloc_support_movable = true,
447 .malloc = zs_zpool_malloc,
448 .free = zs_zpool_free,
450 .unmap = zs_zpool_unmap,
451 .total_size = zs_zpool_total_size,
454 MODULE_ALIAS("zpool-zsmalloc");
455 #endif /* CONFIG_ZPOOL */
457 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
458 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
460 static bool is_zspage_isolated(struct zspage *zspage)
462 return zspage->isolated;
465 static __maybe_unused int is_first_page(struct page *page)
467 return PagePrivate(page);
470 /* Protected by class->lock */
471 static inline int get_zspage_inuse(struct zspage *zspage)
473 return zspage->inuse;
477 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
479 zspage->inuse += val;
482 static inline struct page *get_first_page(struct zspage *zspage)
484 struct page *first_page = zspage->first_page;
486 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
490 static inline int get_first_obj_offset(struct page *page)
495 static inline void set_first_obj_offset(struct page *page, int offset)
497 page->units = offset;
500 static inline unsigned int get_freeobj(struct zspage *zspage)
502 return zspage->freeobj;
505 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
507 zspage->freeobj = obj;
510 static void get_zspage_mapping(struct zspage *zspage,
511 unsigned int *class_idx,
512 enum fullness_group *fullness)
514 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
516 *fullness = zspage->fullness;
517 *class_idx = zspage->class;
520 static void set_zspage_mapping(struct zspage *zspage,
521 unsigned int class_idx,
522 enum fullness_group fullness)
524 zspage->class = class_idx;
525 zspage->fullness = fullness;
529 * zsmalloc divides the pool into various size classes where each
530 * class maintains a list of zspages where each zspage is divided
531 * into equal sized chunks. Each allocation falls into one of these
532 * classes depending on its size. This function returns index of the
533 * size class which has chunk size big enough to hold the give size.
535 static int get_size_class_index(int size)
539 if (likely(size > ZS_MIN_ALLOC_SIZE))
540 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
541 ZS_SIZE_CLASS_DELTA);
543 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
546 /* type can be of enum type zs_stat_type or fullness_group */
547 static inline void zs_stat_inc(struct size_class *class,
548 int type, unsigned long cnt)
550 class->stats.objs[type] += cnt;
553 /* type can be of enum type zs_stat_type or fullness_group */
554 static inline void zs_stat_dec(struct size_class *class,
555 int type, unsigned long cnt)
557 class->stats.objs[type] -= cnt;
560 /* type can be of enum type zs_stat_type or fullness_group */
561 static inline unsigned long zs_stat_get(struct size_class *class,
564 return class->stats.objs[type];
567 #ifdef CONFIG_ZSMALLOC_STAT
569 static void __init zs_stat_init(void)
571 if (!debugfs_initialized()) {
572 pr_warn("debugfs not available, stat dir not created\n");
576 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
579 static void __exit zs_stat_exit(void)
581 debugfs_remove_recursive(zs_stat_root);
584 static unsigned long zs_can_compact(struct size_class *class);
586 static int zs_stats_size_show(struct seq_file *s, void *v)
589 struct zs_pool *pool = s->private;
590 struct size_class *class;
592 unsigned long class_almost_full, class_almost_empty;
593 unsigned long obj_allocated, obj_used, pages_used, freeable;
594 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
595 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
596 unsigned long total_freeable = 0;
598 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
599 "class", "size", "almost_full", "almost_empty",
600 "obj_allocated", "obj_used", "pages_used",
601 "pages_per_zspage", "freeable");
603 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
604 class = pool->size_class[i];
606 if (class->index != i)
609 spin_lock(&class->lock);
610 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
611 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
612 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
613 obj_used = zs_stat_get(class, OBJ_USED);
614 freeable = zs_can_compact(class);
615 spin_unlock(&class->lock);
617 objs_per_zspage = class->objs_per_zspage;
618 pages_used = obj_allocated / objs_per_zspage *
619 class->pages_per_zspage;
621 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
622 " %10lu %10lu %16d %8lu\n",
623 i, class->size, class_almost_full, class_almost_empty,
624 obj_allocated, obj_used, pages_used,
625 class->pages_per_zspage, freeable);
627 total_class_almost_full += class_almost_full;
628 total_class_almost_empty += class_almost_empty;
629 total_objs += obj_allocated;
630 total_used_objs += obj_used;
631 total_pages += pages_used;
632 total_freeable += freeable;
636 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
637 "Total", "", total_class_almost_full,
638 total_class_almost_empty, total_objs,
639 total_used_objs, total_pages, "", total_freeable);
643 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
645 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
648 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
652 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
654 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
655 &zs_stats_size_fops);
658 static void zs_pool_stat_destroy(struct zs_pool *pool)
660 debugfs_remove_recursive(pool->stat_dentry);
663 #else /* CONFIG_ZSMALLOC_STAT */
664 static void __init zs_stat_init(void)
668 static void __exit zs_stat_exit(void)
672 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
676 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
683 * For each size class, zspages are divided into different groups
684 * depending on how "full" they are. This was done so that we could
685 * easily find empty or nearly empty zspages when we try to shrink
686 * the pool (not yet implemented). This function returns fullness
687 * status of the given page.
689 static enum fullness_group get_fullness_group(struct size_class *class,
690 struct zspage *zspage)
692 int inuse, objs_per_zspage;
693 enum fullness_group fg;
695 inuse = get_zspage_inuse(zspage);
696 objs_per_zspage = class->objs_per_zspage;
700 else if (inuse == objs_per_zspage)
702 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
703 fg = ZS_ALMOST_EMPTY;
711 * Each size class maintains various freelists and zspages are assigned
712 * to one of these freelists based on the number of live objects they
713 * have. This functions inserts the given zspage into the freelist
714 * identified by <class, fullness_group>.
716 static void insert_zspage(struct size_class *class,
717 struct zspage *zspage,
718 enum fullness_group fullness)
722 zs_stat_inc(class, fullness, 1);
723 head = list_first_entry_or_null(&class->fullness_list[fullness],
724 struct zspage, list);
726 * We want to see more ZS_FULL pages and less almost empty/full.
727 * Put pages with higher ->inuse first.
729 if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head))
730 list_add(&zspage->list, &head->list);
732 list_add(&zspage->list, &class->fullness_list[fullness]);
736 * This function removes the given zspage from the freelist identified
737 * by <class, fullness_group>.
739 static void remove_zspage(struct size_class *class,
740 struct zspage *zspage,
741 enum fullness_group fullness)
743 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
744 VM_BUG_ON(is_zspage_isolated(zspage));
746 list_del_init(&zspage->list);
747 zs_stat_dec(class, fullness, 1);
751 * Each size class maintains zspages in different fullness groups depending
752 * on the number of live objects they contain. When allocating or freeing
753 * objects, the fullness status of the page can change, say, from ALMOST_FULL
754 * to ALMOST_EMPTY when freeing an object. This function checks if such
755 * a status change has occurred for the given page and accordingly moves the
756 * page from the freelist of the old fullness group to that of the new
759 static enum fullness_group fix_fullness_group(struct size_class *class,
760 struct zspage *zspage)
763 enum fullness_group currfg, newfg;
765 get_zspage_mapping(zspage, &class_idx, &currfg);
766 newfg = get_fullness_group(class, zspage);
770 if (!is_zspage_isolated(zspage)) {
771 remove_zspage(class, zspage, currfg);
772 insert_zspage(class, zspage, newfg);
775 set_zspage_mapping(zspage, class_idx, newfg);
782 * We have to decide on how many pages to link together
783 * to form a zspage for each size class. This is important
784 * to reduce wastage due to unusable space left at end of
785 * each zspage which is given as:
786 * wastage = Zp % class_size
787 * usage = Zp - wastage
788 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
790 * For example, for size class of 3/8 * PAGE_SIZE, we should
791 * link together 3 PAGE_SIZE sized pages to form a zspage
792 * since then we can perfectly fit in 8 such objects.
794 static int get_pages_per_zspage(int class_size)
796 int i, max_usedpc = 0;
797 /* zspage order which gives maximum used size per KB */
798 int max_usedpc_order = 1;
800 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
804 zspage_size = i * PAGE_SIZE;
805 waste = zspage_size % class_size;
806 usedpc = (zspage_size - waste) * 100 / zspage_size;
808 if (usedpc > max_usedpc) {
810 max_usedpc_order = i;
814 return max_usedpc_order;
817 static struct zspage *get_zspage(struct page *page)
819 struct zspage *zspage = (struct zspage *)page_private(page);
821 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
825 static struct page *get_next_page(struct page *page)
827 if (unlikely(PageHugeObject(page)))
830 return page->freelist;
834 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
835 * @obj: the encoded object value
836 * @page: page object resides in zspage
837 * @obj_idx: object index
839 static void obj_to_location(unsigned long obj, struct page **page,
840 unsigned int *obj_idx)
842 obj >>= OBJ_TAG_BITS;
843 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
844 *obj_idx = (obj & OBJ_INDEX_MASK);
848 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
849 * @page: page object resides in zspage
850 * @obj_idx: object index
852 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
856 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
857 obj |= obj_idx & OBJ_INDEX_MASK;
858 obj <<= OBJ_TAG_BITS;
863 static unsigned long handle_to_obj(unsigned long handle)
865 return *(unsigned long *)handle;
868 static unsigned long obj_to_head(struct page *page, void *obj)
870 if (unlikely(PageHugeObject(page))) {
871 VM_BUG_ON_PAGE(!is_first_page(page), page);
874 return *(unsigned long *)obj;
877 static inline int testpin_tag(unsigned long handle)
879 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
882 static inline int trypin_tag(unsigned long handle)
884 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
887 static void pin_tag(unsigned long handle) __acquires(bitlock)
889 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
892 static void unpin_tag(unsigned long handle) __releases(bitlock)
894 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
897 static void reset_page(struct page *page)
899 __ClearPageMovable(page);
900 ClearPagePrivate(page);
901 set_page_private(page, 0);
902 page_mapcount_reset(page);
903 ClearPageHugeObject(page);
904 page->freelist = NULL;
907 static int trylock_zspage(struct zspage *zspage)
909 struct page *cursor, *fail;
911 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
912 get_next_page(cursor)) {
913 if (!trylock_page(cursor)) {
921 for (cursor = get_first_page(zspage); cursor != fail; cursor =
922 get_next_page(cursor))
928 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
929 struct zspage *zspage)
931 struct page *page, *next;
932 enum fullness_group fg;
933 unsigned int class_idx;
935 get_zspage_mapping(zspage, &class_idx, &fg);
937 assert_spin_locked(&class->lock);
939 VM_BUG_ON(get_zspage_inuse(zspage));
940 VM_BUG_ON(fg != ZS_EMPTY);
942 next = page = get_first_page(zspage);
944 VM_BUG_ON_PAGE(!PageLocked(page), page);
945 next = get_next_page(page);
948 dec_zone_page_state(page, NR_ZSPAGES);
951 } while (page != NULL);
953 cache_free_zspage(pool, zspage);
955 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
956 atomic_long_sub(class->pages_per_zspage,
957 &pool->pages_allocated);
960 static void free_zspage(struct zs_pool *pool, struct size_class *class,
961 struct zspage *zspage)
963 VM_BUG_ON(get_zspage_inuse(zspage));
964 VM_BUG_ON(list_empty(&zspage->list));
966 if (!trylock_zspage(zspage)) {
967 kick_deferred_free(pool);
971 remove_zspage(class, zspage, ZS_EMPTY);
972 __free_zspage(pool, class, zspage);
975 /* Initialize a newly allocated zspage */
976 static void init_zspage(struct size_class *class, struct zspage *zspage)
978 unsigned int freeobj = 1;
979 unsigned long off = 0;
980 struct page *page = get_first_page(zspage);
983 struct page *next_page;
984 struct link_free *link;
987 set_first_obj_offset(page, off);
989 vaddr = kmap_atomic(page);
990 link = (struct link_free *)vaddr + off / sizeof(*link);
992 while ((off += class->size) < PAGE_SIZE) {
993 link->next = freeobj++ << OBJ_TAG_BITS;
994 link += class->size / sizeof(*link);
998 * We now come to the last (full or partial) object on this
999 * page, which must point to the first object on the next
1002 next_page = get_next_page(page);
1004 link->next = freeobj++ << OBJ_TAG_BITS;
1007 * Reset OBJ_TAG_BITS bit to last link to tell
1008 * whether it's allocated object or not.
1010 link->next = -1UL << OBJ_TAG_BITS;
1012 kunmap_atomic(vaddr);
1017 set_freeobj(zspage, 0);
1020 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1021 struct page *pages[])
1025 struct page *prev_page = NULL;
1026 int nr_pages = class->pages_per_zspage;
1029 * Allocate individual pages and link them together as:
1030 * 1. all pages are linked together using page->freelist
1031 * 2. each sub-page point to zspage using page->private
1033 * we set PG_private to identify the first page (i.e. no other sub-page
1034 * has this flag set).
1036 for (i = 0; i < nr_pages; i++) {
1038 set_page_private(page, (unsigned long)zspage);
1039 page->freelist = NULL;
1041 zspage->first_page = page;
1042 SetPagePrivate(page);
1043 if (unlikely(class->objs_per_zspage == 1 &&
1044 class->pages_per_zspage == 1))
1045 SetPageHugeObject(page);
1047 prev_page->freelist = page;
1054 * Allocate a zspage for the given size class
1056 static struct zspage *alloc_zspage(struct zs_pool *pool,
1057 struct size_class *class,
1061 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1062 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1067 zspage->magic = ZSPAGE_MAGIC;
1068 migrate_lock_init(zspage);
1070 for (i = 0; i < class->pages_per_zspage; i++) {
1073 page = alloc_page(gfp);
1076 dec_zone_page_state(pages[i], NR_ZSPAGES);
1077 __free_page(pages[i]);
1079 cache_free_zspage(pool, zspage);
1083 inc_zone_page_state(page, NR_ZSPAGES);
1087 create_page_chain(class, zspage, pages);
1088 init_zspage(class, zspage);
1093 static struct zspage *find_get_zspage(struct size_class *class)
1096 struct zspage *zspage;
1098 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1099 zspage = list_first_entry_or_null(&class->fullness_list[i],
1100 struct zspage, list);
1108 static inline int __zs_cpu_up(struct mapping_area *area)
1111 * Make sure we don't leak memory if a cpu UP notification
1112 * and zs_init() race and both call zs_cpu_up() on the same cpu
1116 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1122 static inline void __zs_cpu_down(struct mapping_area *area)
1124 kfree(area->vm_buf);
1125 area->vm_buf = NULL;
1128 static void *__zs_map_object(struct mapping_area *area,
1129 struct page *pages[2], int off, int size)
1133 char *buf = area->vm_buf;
1135 /* disable page faults to match kmap_atomic() return conditions */
1136 pagefault_disable();
1138 /* no read fastpath */
1139 if (area->vm_mm == ZS_MM_WO)
1142 sizes[0] = PAGE_SIZE - off;
1143 sizes[1] = size - sizes[0];
1145 /* copy object to per-cpu buffer */
1146 addr = kmap_atomic(pages[0]);
1147 memcpy(buf, addr + off, sizes[0]);
1148 kunmap_atomic(addr);
1149 addr = kmap_atomic(pages[1]);
1150 memcpy(buf + sizes[0], addr, sizes[1]);
1151 kunmap_atomic(addr);
1153 return area->vm_buf;
1156 static void __zs_unmap_object(struct mapping_area *area,
1157 struct page *pages[2], int off, int size)
1163 /* no write fastpath */
1164 if (area->vm_mm == ZS_MM_RO)
1168 buf = buf + ZS_HANDLE_SIZE;
1169 size -= ZS_HANDLE_SIZE;
1170 off += ZS_HANDLE_SIZE;
1172 sizes[0] = PAGE_SIZE - off;
1173 sizes[1] = size - sizes[0];
1175 /* copy per-cpu buffer to object */
1176 addr = kmap_atomic(pages[0]);
1177 memcpy(addr + off, buf, sizes[0]);
1178 kunmap_atomic(addr);
1179 addr = kmap_atomic(pages[1]);
1180 memcpy(addr, buf + sizes[0], sizes[1]);
1181 kunmap_atomic(addr);
1184 /* enable page faults to match kunmap_atomic() return conditions */
1188 static int zs_cpu_prepare(unsigned int cpu)
1190 struct mapping_area *area;
1192 area = &per_cpu(zs_map_area, cpu);
1193 return __zs_cpu_up(area);
1196 static int zs_cpu_dead(unsigned int cpu)
1198 struct mapping_area *area;
1200 area = &per_cpu(zs_map_area, cpu);
1201 __zs_cpu_down(area);
1205 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1206 int objs_per_zspage)
1208 if (prev->pages_per_zspage == pages_per_zspage &&
1209 prev->objs_per_zspage == objs_per_zspage)
1215 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1217 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1220 unsigned long zs_get_total_pages(struct zs_pool *pool)
1222 return atomic_long_read(&pool->pages_allocated);
1224 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1227 * zs_map_object - get address of allocated object from handle.
1228 * @pool: pool from which the object was allocated
1229 * @handle: handle returned from zs_malloc
1230 * @mm: maping mode to use
1232 * Before using an object allocated from zs_malloc, it must be mapped using
1233 * this function. When done with the object, it must be unmapped using
1236 * Only one object can be mapped per cpu at a time. There is no protection
1237 * against nested mappings.
1239 * This function returns with preemption and page faults disabled.
1241 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1244 struct zspage *zspage;
1246 unsigned long obj, off;
1247 unsigned int obj_idx;
1249 unsigned int class_idx;
1250 enum fullness_group fg;
1251 struct size_class *class;
1252 struct mapping_area *area;
1253 struct page *pages[2];
1257 * Because we use per-cpu mapping areas shared among the
1258 * pools/users, we can't allow mapping in interrupt context
1259 * because it can corrupt another users mappings.
1261 BUG_ON(in_interrupt());
1263 /* From now on, migration cannot move the object */
1266 obj = handle_to_obj(handle);
1267 obj_to_location(obj, &page, &obj_idx);
1268 zspage = get_zspage(page);
1270 /* migration cannot move any subpage in this zspage */
1271 migrate_read_lock(zspage);
1273 get_zspage_mapping(zspage, &class_idx, &fg);
1274 class = pool->size_class[class_idx];
1275 off = (class->size * obj_idx) & ~PAGE_MASK;
1277 area = &get_cpu_var(zs_map_area);
1279 if (off + class->size <= PAGE_SIZE) {
1280 /* this object is contained entirely within a page */
1281 area->vm_addr = kmap_atomic(page);
1282 ret = area->vm_addr + off;
1286 /* this object spans two pages */
1288 pages[1] = get_next_page(page);
1291 ret = __zs_map_object(area, pages, off, class->size);
1293 if (likely(!PageHugeObject(page)))
1294 ret += ZS_HANDLE_SIZE;
1298 EXPORT_SYMBOL_GPL(zs_map_object);
1300 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1302 struct zspage *zspage;
1304 unsigned long obj, off;
1305 unsigned int obj_idx;
1307 unsigned int class_idx;
1308 enum fullness_group fg;
1309 struct size_class *class;
1310 struct mapping_area *area;
1312 obj = handle_to_obj(handle);
1313 obj_to_location(obj, &page, &obj_idx);
1314 zspage = get_zspage(page);
1315 get_zspage_mapping(zspage, &class_idx, &fg);
1316 class = pool->size_class[class_idx];
1317 off = (class->size * obj_idx) & ~PAGE_MASK;
1319 area = this_cpu_ptr(&zs_map_area);
1320 if (off + class->size <= PAGE_SIZE)
1321 kunmap_atomic(area->vm_addr);
1323 struct page *pages[2];
1326 pages[1] = get_next_page(page);
1329 __zs_unmap_object(area, pages, off, class->size);
1331 put_cpu_var(zs_map_area);
1333 migrate_read_unlock(zspage);
1336 EXPORT_SYMBOL_GPL(zs_unmap_object);
1339 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1340 * zsmalloc &size_class.
1341 * @pool: zsmalloc pool to use
1343 * The function returns the size of the first huge class - any object of equal
1344 * or bigger size will be stored in zspage consisting of a single physical
1347 * Context: Any context.
1349 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1351 size_t zs_huge_class_size(struct zs_pool *pool)
1353 return huge_class_size;
1355 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1357 static unsigned long obj_malloc(struct size_class *class,
1358 struct zspage *zspage, unsigned long handle)
1360 int i, nr_page, offset;
1362 struct link_free *link;
1364 struct page *m_page;
1365 unsigned long m_offset;
1368 handle |= OBJ_ALLOCATED_TAG;
1369 obj = get_freeobj(zspage);
1371 offset = obj * class->size;
1372 nr_page = offset >> PAGE_SHIFT;
1373 m_offset = offset & ~PAGE_MASK;
1374 m_page = get_first_page(zspage);
1376 for (i = 0; i < nr_page; i++)
1377 m_page = get_next_page(m_page);
1379 vaddr = kmap_atomic(m_page);
1380 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1381 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1382 if (likely(!PageHugeObject(m_page)))
1383 /* record handle in the header of allocated chunk */
1384 link->handle = handle;
1386 /* record handle to page->index */
1387 zspage->first_page->index = handle;
1389 kunmap_atomic(vaddr);
1390 mod_zspage_inuse(zspage, 1);
1391 zs_stat_inc(class, OBJ_USED, 1);
1393 obj = location_to_obj(m_page, obj);
1400 * zs_malloc - Allocate block of given size from pool.
1401 * @pool: pool to allocate from
1402 * @size: size of block to allocate
1403 * @gfp: gfp flags when allocating object
1405 * On success, handle to the allocated object is returned,
1407 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1409 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1411 unsigned long handle, obj;
1412 struct size_class *class;
1413 enum fullness_group newfg;
1414 struct zspage *zspage;
1416 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1419 handle = cache_alloc_handle(pool, gfp);
1423 /* extra space in chunk to keep the handle */
1424 size += ZS_HANDLE_SIZE;
1425 class = pool->size_class[get_size_class_index(size)];
1427 spin_lock(&class->lock);
1428 zspage = find_get_zspage(class);
1429 if (likely(zspage)) {
1430 obj = obj_malloc(class, zspage, handle);
1431 /* Now move the zspage to another fullness group, if required */
1432 fix_fullness_group(class, zspage);
1433 record_obj(handle, obj);
1434 spin_unlock(&class->lock);
1439 spin_unlock(&class->lock);
1441 zspage = alloc_zspage(pool, class, gfp);
1443 cache_free_handle(pool, handle);
1447 spin_lock(&class->lock);
1448 obj = obj_malloc(class, zspage, handle);
1449 newfg = get_fullness_group(class, zspage);
1450 insert_zspage(class, zspage, newfg);
1451 set_zspage_mapping(zspage, class->index, newfg);
1452 record_obj(handle, obj);
1453 atomic_long_add(class->pages_per_zspage,
1454 &pool->pages_allocated);
1455 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1457 /* We completely set up zspage so mark them as movable */
1458 SetZsPageMovable(pool, zspage);
1459 spin_unlock(&class->lock);
1463 EXPORT_SYMBOL_GPL(zs_malloc);
1465 static void obj_free(struct size_class *class, unsigned long obj)
1467 struct link_free *link;
1468 struct zspage *zspage;
1469 struct page *f_page;
1470 unsigned long f_offset;
1471 unsigned int f_objidx;
1474 obj &= ~OBJ_ALLOCATED_TAG;
1475 obj_to_location(obj, &f_page, &f_objidx);
1476 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1477 zspage = get_zspage(f_page);
1479 vaddr = kmap_atomic(f_page);
1481 /* Insert this object in containing zspage's freelist */
1482 link = (struct link_free *)(vaddr + f_offset);
1483 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1484 kunmap_atomic(vaddr);
1485 set_freeobj(zspage, f_objidx);
1486 mod_zspage_inuse(zspage, -1);
1487 zs_stat_dec(class, OBJ_USED, 1);
1490 void zs_free(struct zs_pool *pool, unsigned long handle)
1492 struct zspage *zspage;
1493 struct page *f_page;
1495 unsigned int f_objidx;
1497 struct size_class *class;
1498 enum fullness_group fullness;
1501 if (unlikely(!handle))
1505 obj = handle_to_obj(handle);
1506 obj_to_location(obj, &f_page, &f_objidx);
1507 zspage = get_zspage(f_page);
1509 migrate_read_lock(zspage);
1511 get_zspage_mapping(zspage, &class_idx, &fullness);
1512 class = pool->size_class[class_idx];
1514 spin_lock(&class->lock);
1515 obj_free(class, obj);
1516 fullness = fix_fullness_group(class, zspage);
1517 if (fullness != ZS_EMPTY) {
1518 migrate_read_unlock(zspage);
1522 isolated = is_zspage_isolated(zspage);
1523 migrate_read_unlock(zspage);
1524 /* If zspage is isolated, zs_page_putback will free the zspage */
1525 if (likely(!isolated))
1526 free_zspage(pool, class, zspage);
1529 spin_unlock(&class->lock);
1531 cache_free_handle(pool, handle);
1533 EXPORT_SYMBOL_GPL(zs_free);
1535 static void zs_object_copy(struct size_class *class, unsigned long dst,
1538 struct page *s_page, *d_page;
1539 unsigned int s_objidx, d_objidx;
1540 unsigned long s_off, d_off;
1541 void *s_addr, *d_addr;
1542 int s_size, d_size, size;
1545 s_size = d_size = class->size;
1547 obj_to_location(src, &s_page, &s_objidx);
1548 obj_to_location(dst, &d_page, &d_objidx);
1550 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1551 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1553 if (s_off + class->size > PAGE_SIZE)
1554 s_size = PAGE_SIZE - s_off;
1556 if (d_off + class->size > PAGE_SIZE)
1557 d_size = PAGE_SIZE - d_off;
1559 s_addr = kmap_atomic(s_page);
1560 d_addr = kmap_atomic(d_page);
1563 size = min(s_size, d_size);
1564 memcpy(d_addr + d_off, s_addr + s_off, size);
1567 if (written == class->size)
1575 if (s_off >= PAGE_SIZE) {
1576 kunmap_atomic(d_addr);
1577 kunmap_atomic(s_addr);
1578 s_page = get_next_page(s_page);
1579 s_addr = kmap_atomic(s_page);
1580 d_addr = kmap_atomic(d_page);
1581 s_size = class->size - written;
1585 if (d_off >= PAGE_SIZE) {
1586 kunmap_atomic(d_addr);
1587 d_page = get_next_page(d_page);
1588 d_addr = kmap_atomic(d_page);
1589 d_size = class->size - written;
1594 kunmap_atomic(d_addr);
1595 kunmap_atomic(s_addr);
1599 * Find alloced object in zspage from index object and
1602 static unsigned long find_alloced_obj(struct size_class *class,
1603 struct page *page, int *obj_idx)
1607 int index = *obj_idx;
1608 unsigned long handle = 0;
1609 void *addr = kmap_atomic(page);
1611 offset = get_first_obj_offset(page);
1612 offset += class->size * index;
1614 while (offset < PAGE_SIZE) {
1615 head = obj_to_head(page, addr + offset);
1616 if (head & OBJ_ALLOCATED_TAG) {
1617 handle = head & ~OBJ_ALLOCATED_TAG;
1618 if (trypin_tag(handle))
1623 offset += class->size;
1627 kunmap_atomic(addr);
1634 struct zs_compact_control {
1635 /* Source spage for migration which could be a subpage of zspage */
1636 struct page *s_page;
1637 /* Destination page for migration which should be a first page
1639 struct page *d_page;
1640 /* Starting object index within @s_page which used for live object
1641 * in the subpage. */
1645 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1646 struct zs_compact_control *cc)
1648 unsigned long used_obj, free_obj;
1649 unsigned long handle;
1650 struct page *s_page = cc->s_page;
1651 struct page *d_page = cc->d_page;
1652 int obj_idx = cc->obj_idx;
1656 handle = find_alloced_obj(class, s_page, &obj_idx);
1658 s_page = get_next_page(s_page);
1665 /* Stop if there is no more space */
1666 if (zspage_full(class, get_zspage(d_page))) {
1672 used_obj = handle_to_obj(handle);
1673 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1674 zs_object_copy(class, free_obj, used_obj);
1677 * record_obj updates handle's value to free_obj and it will
1678 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1679 * breaks synchronization using pin_tag(e,g, zs_free) so
1680 * let's keep the lock bit.
1682 free_obj |= BIT(HANDLE_PIN_BIT);
1683 record_obj(handle, free_obj);
1685 obj_free(class, used_obj);
1688 /* Remember last position in this iteration */
1689 cc->s_page = s_page;
1690 cc->obj_idx = obj_idx;
1695 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1698 struct zspage *zspage;
1699 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1702 fg[0] = ZS_ALMOST_FULL;
1703 fg[1] = ZS_ALMOST_EMPTY;
1706 for (i = 0; i < 2; i++) {
1707 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1708 struct zspage, list);
1710 VM_BUG_ON(is_zspage_isolated(zspage));
1711 remove_zspage(class, zspage, fg[i]);
1720 * putback_zspage - add @zspage into right class's fullness list
1721 * @class: destination class
1722 * @zspage: target page
1724 * Return @zspage's fullness_group
1726 static enum fullness_group putback_zspage(struct size_class *class,
1727 struct zspage *zspage)
1729 enum fullness_group fullness;
1731 VM_BUG_ON(is_zspage_isolated(zspage));
1733 fullness = get_fullness_group(class, zspage);
1734 insert_zspage(class, zspage, fullness);
1735 set_zspage_mapping(zspage, class->index, fullness);
1740 #ifdef CONFIG_COMPACTION
1742 * To prevent zspage destroy during migration, zspage freeing should
1743 * hold locks of all pages in the zspage.
1745 static void lock_zspage(struct zspage *zspage)
1747 struct page *page = get_first_page(zspage);
1751 } while ((page = get_next_page(page)) != NULL);
1754 static int zs_init_fs_context(struct fs_context *fc)
1756 return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1759 static struct file_system_type zsmalloc_fs = {
1761 .init_fs_context = zs_init_fs_context,
1762 .kill_sb = kill_anon_super,
1765 static int zsmalloc_mount(void)
1769 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1770 if (IS_ERR(zsmalloc_mnt))
1771 ret = PTR_ERR(zsmalloc_mnt);
1776 static void zsmalloc_unmount(void)
1778 kern_unmount(zsmalloc_mnt);
1781 static void migrate_lock_init(struct zspage *zspage)
1783 rwlock_init(&zspage->lock);
1786 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1788 read_lock(&zspage->lock);
1791 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1793 read_unlock(&zspage->lock);
1796 static void migrate_write_lock(struct zspage *zspage)
1798 write_lock(&zspage->lock);
1801 static void migrate_write_unlock(struct zspage *zspage)
1803 write_unlock(&zspage->lock);
1806 /* Number of isolated subpage for *page migration* in this zspage */
1807 static void inc_zspage_isolation(struct zspage *zspage)
1812 static void dec_zspage_isolation(struct zspage *zspage)
1817 static void putback_zspage_deferred(struct zs_pool *pool,
1818 struct size_class *class,
1819 struct zspage *zspage)
1821 enum fullness_group fg;
1823 fg = putback_zspage(class, zspage);
1825 schedule_work(&pool->free_work);
1829 static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1831 VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1832 atomic_long_dec(&pool->isolated_pages);
1834 * There's no possibility of racing, since wait_for_isolated_drain()
1835 * checks the isolated count under &class->lock after enqueuing
1836 * on migration_wait.
1838 if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1839 wake_up_all(&pool->migration_wait);
1842 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1843 struct page *newpage, struct page *oldpage)
1846 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1849 page = get_first_page(zspage);
1851 if (page == oldpage)
1852 pages[idx] = newpage;
1856 } while ((page = get_next_page(page)) != NULL);
1858 create_page_chain(class, zspage, pages);
1859 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1860 if (unlikely(PageHugeObject(oldpage)))
1861 newpage->index = oldpage->index;
1862 __SetPageMovable(newpage, page_mapping(oldpage));
1865 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1867 struct zs_pool *pool;
1868 struct size_class *class;
1870 enum fullness_group fullness;
1871 struct zspage *zspage;
1872 struct address_space *mapping;
1875 * Page is locked so zspage couldn't be destroyed. For detail, look at
1876 * lock_zspage in free_zspage.
1878 VM_BUG_ON_PAGE(!PageMovable(page), page);
1879 VM_BUG_ON_PAGE(PageIsolated(page), page);
1881 zspage = get_zspage(page);
1884 * Without class lock, fullness could be stale while class_idx is okay
1885 * because class_idx is constant unless page is freed so we should get
1886 * fullness again under class lock.
1888 get_zspage_mapping(zspage, &class_idx, &fullness);
1889 mapping = page_mapping(page);
1890 pool = mapping->private_data;
1891 class = pool->size_class[class_idx];
1893 spin_lock(&class->lock);
1894 if (get_zspage_inuse(zspage) == 0) {
1895 spin_unlock(&class->lock);
1899 /* zspage is isolated for object migration */
1900 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1901 spin_unlock(&class->lock);
1906 * If this is first time isolation for the zspage, isolate zspage from
1907 * size_class to prevent further object allocation from the zspage.
1909 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1910 get_zspage_mapping(zspage, &class_idx, &fullness);
1911 atomic_long_inc(&pool->isolated_pages);
1912 remove_zspage(class, zspage, fullness);
1915 inc_zspage_isolation(zspage);
1916 spin_unlock(&class->lock);
1921 static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1922 struct page *page, enum migrate_mode mode)
1924 struct zs_pool *pool;
1925 struct size_class *class;
1927 enum fullness_group fullness;
1928 struct zspage *zspage;
1930 void *s_addr, *d_addr, *addr;
1932 unsigned long handle, head;
1933 unsigned long old_obj, new_obj;
1934 unsigned int obj_idx;
1938 * We cannot support the _NO_COPY case here, because copy needs to
1939 * happen under the zs lock, which does not work with
1940 * MIGRATE_SYNC_NO_COPY workflow.
1942 if (mode == MIGRATE_SYNC_NO_COPY)
1945 VM_BUG_ON_PAGE(!PageMovable(page), page);
1946 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1948 zspage = get_zspage(page);
1950 /* Concurrent compactor cannot migrate any subpage in zspage */
1951 migrate_write_lock(zspage);
1952 get_zspage_mapping(zspage, &class_idx, &fullness);
1953 pool = mapping->private_data;
1954 class = pool->size_class[class_idx];
1955 offset = get_first_obj_offset(page);
1957 spin_lock(&class->lock);
1958 if (!get_zspage_inuse(zspage)) {
1960 * Set "offset" to end of the page so that every loops
1961 * skips unnecessary object scanning.
1967 s_addr = kmap_atomic(page);
1968 while (pos < PAGE_SIZE) {
1969 head = obj_to_head(page, s_addr + pos);
1970 if (head & OBJ_ALLOCATED_TAG) {
1971 handle = head & ~OBJ_ALLOCATED_TAG;
1972 if (!trypin_tag(handle))
1979 * Here, any user cannot access all objects in the zspage so let's move.
1981 d_addr = kmap_atomic(newpage);
1982 memcpy(d_addr, s_addr, PAGE_SIZE);
1983 kunmap_atomic(d_addr);
1985 for (addr = s_addr + offset; addr < s_addr + pos;
1986 addr += class->size) {
1987 head = obj_to_head(page, addr);
1988 if (head & OBJ_ALLOCATED_TAG) {
1989 handle = head & ~OBJ_ALLOCATED_TAG;
1990 if (!testpin_tag(handle))
1993 old_obj = handle_to_obj(handle);
1994 obj_to_location(old_obj, &dummy, &obj_idx);
1995 new_obj = (unsigned long)location_to_obj(newpage,
1997 new_obj |= BIT(HANDLE_PIN_BIT);
1998 record_obj(handle, new_obj);
2002 replace_sub_page(class, zspage, newpage, page);
2005 dec_zspage_isolation(zspage);
2008 * Page migration is done so let's putback isolated zspage to
2009 * the list if @page is final isolated subpage in the zspage.
2011 if (!is_zspage_isolated(zspage)) {
2013 * We cannot race with zs_destroy_pool() here because we wait
2014 * for isolation to hit zero before we start destroying.
2015 * Also, we ensure that everyone can see pool->destroying before
2018 putback_zspage_deferred(pool, class, zspage);
2019 zs_pool_dec_isolated(pool);
2022 if (page_zone(newpage) != page_zone(page)) {
2023 dec_zone_page_state(page, NR_ZSPAGES);
2024 inc_zone_page_state(newpage, NR_ZSPAGES);
2031 ret = MIGRATEPAGE_SUCCESS;
2033 for (addr = s_addr + offset; addr < s_addr + pos;
2034 addr += class->size) {
2035 head = obj_to_head(page, addr);
2036 if (head & OBJ_ALLOCATED_TAG) {
2037 handle = head & ~OBJ_ALLOCATED_TAG;
2038 if (!testpin_tag(handle))
2043 kunmap_atomic(s_addr);
2044 spin_unlock(&class->lock);
2045 migrate_write_unlock(zspage);
2050 static void zs_page_putback(struct page *page)
2052 struct zs_pool *pool;
2053 struct size_class *class;
2055 enum fullness_group fg;
2056 struct address_space *mapping;
2057 struct zspage *zspage;
2059 VM_BUG_ON_PAGE(!PageMovable(page), page);
2060 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2062 zspage = get_zspage(page);
2063 get_zspage_mapping(zspage, &class_idx, &fg);
2064 mapping = page_mapping(page);
2065 pool = mapping->private_data;
2066 class = pool->size_class[class_idx];
2068 spin_lock(&class->lock);
2069 dec_zspage_isolation(zspage);
2070 if (!is_zspage_isolated(zspage)) {
2072 * Due to page_lock, we cannot free zspage immediately
2075 putback_zspage_deferred(pool, class, zspage);
2076 zs_pool_dec_isolated(pool);
2078 spin_unlock(&class->lock);
2081 static const struct address_space_operations zsmalloc_aops = {
2082 .isolate_page = zs_page_isolate,
2083 .migratepage = zs_page_migrate,
2084 .putback_page = zs_page_putback,
2087 static int zs_register_migration(struct zs_pool *pool)
2089 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2090 if (IS_ERR(pool->inode)) {
2095 pool->inode->i_mapping->private_data = pool;
2096 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2100 static bool pool_isolated_are_drained(struct zs_pool *pool)
2102 return atomic_long_read(&pool->isolated_pages) == 0;
2105 /* Function for resolving migration */
2106 static void wait_for_isolated_drain(struct zs_pool *pool)
2110 * We're in the process of destroying the pool, so there are no
2111 * active allocations. zs_page_isolate() fails for completely free
2112 * zspages, so we need only wait for the zs_pool's isolated
2113 * count to hit zero.
2115 wait_event(pool->migration_wait,
2116 pool_isolated_are_drained(pool));
2119 static void zs_unregister_migration(struct zs_pool *pool)
2121 pool->destroying = true;
2123 * We need a memory barrier here to ensure global visibility of
2124 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2125 * case we don't care, or it will be > 0 and pool->destroying will
2126 * ensure that we wake up once isolation hits 0.
2129 wait_for_isolated_drain(pool); /* This can block */
2130 flush_work(&pool->free_work);
2135 * Caller should hold page_lock of all pages in the zspage
2136 * In here, we cannot use zspage meta data.
2138 static void async_free_zspage(struct work_struct *work)
2141 struct size_class *class;
2142 unsigned int class_idx;
2143 enum fullness_group fullness;
2144 struct zspage *zspage, *tmp;
2145 LIST_HEAD(free_pages);
2146 struct zs_pool *pool = container_of(work, struct zs_pool,
2149 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2150 class = pool->size_class[i];
2151 if (class->index != i)
2154 spin_lock(&class->lock);
2155 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2156 spin_unlock(&class->lock);
2160 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2161 list_del(&zspage->list);
2162 lock_zspage(zspage);
2164 get_zspage_mapping(zspage, &class_idx, &fullness);
2165 VM_BUG_ON(fullness != ZS_EMPTY);
2166 class = pool->size_class[class_idx];
2167 spin_lock(&class->lock);
2168 __free_zspage(pool, pool->size_class[class_idx], zspage);
2169 spin_unlock(&class->lock);
2173 static void kick_deferred_free(struct zs_pool *pool)
2175 schedule_work(&pool->free_work);
2178 static void init_deferred_free(struct zs_pool *pool)
2180 INIT_WORK(&pool->free_work, async_free_zspage);
2183 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2185 struct page *page = get_first_page(zspage);
2188 WARN_ON(!trylock_page(page));
2189 __SetPageMovable(page, pool->inode->i_mapping);
2191 } while ((page = get_next_page(page)) != NULL);
2197 * Based on the number of unused allocated objects calculate
2198 * and return the number of pages that we can free.
2200 static unsigned long zs_can_compact(struct size_class *class)
2202 unsigned long obj_wasted;
2203 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2204 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2206 if (obj_allocated <= obj_used)
2209 obj_wasted = obj_allocated - obj_used;
2210 obj_wasted /= class->objs_per_zspage;
2212 return obj_wasted * class->pages_per_zspage;
2215 static unsigned long __zs_compact(struct zs_pool *pool,
2216 struct size_class *class)
2218 struct zs_compact_control cc;
2219 struct zspage *src_zspage;
2220 struct zspage *dst_zspage = NULL;
2221 unsigned long pages_freed = 0;
2223 spin_lock(&class->lock);
2224 while ((src_zspage = isolate_zspage(class, true))) {
2226 if (!zs_can_compact(class))
2230 cc.s_page = get_first_page(src_zspage);
2232 while ((dst_zspage = isolate_zspage(class, false))) {
2233 cc.d_page = get_first_page(dst_zspage);
2235 * If there is no more space in dst_page, resched
2236 * and see if anyone had allocated another zspage.
2238 if (!migrate_zspage(pool, class, &cc))
2241 putback_zspage(class, dst_zspage);
2244 /* Stop if we couldn't find slot */
2245 if (dst_zspage == NULL)
2248 putback_zspage(class, dst_zspage);
2249 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2250 free_zspage(pool, class, src_zspage);
2251 pages_freed += class->pages_per_zspage;
2253 spin_unlock(&class->lock);
2255 spin_lock(&class->lock);
2259 putback_zspage(class, src_zspage);
2261 spin_unlock(&class->lock);
2266 unsigned long zs_compact(struct zs_pool *pool)
2269 struct size_class *class;
2270 unsigned long pages_freed = 0;
2272 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2273 class = pool->size_class[i];
2276 if (class->index != i)
2278 pages_freed += __zs_compact(pool, class);
2280 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2284 EXPORT_SYMBOL_GPL(zs_compact);
2286 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2288 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2290 EXPORT_SYMBOL_GPL(zs_pool_stats);
2292 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2293 struct shrink_control *sc)
2295 unsigned long pages_freed;
2296 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2300 * Compact classes and calculate compaction delta.
2301 * Can run concurrently with a manually triggered
2302 * (by user) compaction.
2304 pages_freed = zs_compact(pool);
2306 return pages_freed ? pages_freed : SHRINK_STOP;
2309 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2310 struct shrink_control *sc)
2313 struct size_class *class;
2314 unsigned long pages_to_free = 0;
2315 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2318 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2319 class = pool->size_class[i];
2322 if (class->index != i)
2325 pages_to_free += zs_can_compact(class);
2328 return pages_to_free;
2331 static void zs_unregister_shrinker(struct zs_pool *pool)
2333 unregister_shrinker(&pool->shrinker);
2336 static int zs_register_shrinker(struct zs_pool *pool)
2338 pool->shrinker.scan_objects = zs_shrinker_scan;
2339 pool->shrinker.count_objects = zs_shrinker_count;
2340 pool->shrinker.batch = 0;
2341 pool->shrinker.seeks = DEFAULT_SEEKS;
2343 return register_shrinker(&pool->shrinker);
2347 * zs_create_pool - Creates an allocation pool to work from.
2348 * @name: pool name to be created
2350 * This function must be called before anything when using
2351 * the zsmalloc allocator.
2353 * On success, a pointer to the newly created pool is returned,
2356 struct zs_pool *zs_create_pool(const char *name)
2359 struct zs_pool *pool;
2360 struct size_class *prev_class = NULL;
2362 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2366 init_deferred_free(pool);
2368 pool->name = kstrdup(name, GFP_KERNEL);
2372 #ifdef CONFIG_COMPACTION
2373 init_waitqueue_head(&pool->migration_wait);
2376 if (create_cache(pool))
2380 * Iterate reversely, because, size of size_class that we want to use
2381 * for merging should be larger or equal to current size.
2383 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2385 int pages_per_zspage;
2386 int objs_per_zspage;
2387 struct size_class *class;
2390 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2391 if (size > ZS_MAX_ALLOC_SIZE)
2392 size = ZS_MAX_ALLOC_SIZE;
2393 pages_per_zspage = get_pages_per_zspage(size);
2394 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2397 * We iterate from biggest down to smallest classes,
2398 * so huge_class_size holds the size of the first huge
2399 * class. Any object bigger than or equal to that will
2400 * endup in the huge class.
2402 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2404 huge_class_size = size;
2406 * The object uses ZS_HANDLE_SIZE bytes to store the
2407 * handle. We need to subtract it, because zs_malloc()
2408 * unconditionally adds handle size before it performs
2409 * size class search - so object may be smaller than
2410 * huge class size, yet it still can end up in the huge
2411 * class because it grows by ZS_HANDLE_SIZE extra bytes
2412 * right before class lookup.
2414 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2418 * size_class is used for normal zsmalloc operation such
2419 * as alloc/free for that size. Although it is natural that we
2420 * have one size_class for each size, there is a chance that we
2421 * can get more memory utilization if we use one size_class for
2422 * many different sizes whose size_class have same
2423 * characteristics. So, we makes size_class point to
2424 * previous size_class if possible.
2427 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2428 pool->size_class[i] = prev_class;
2433 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2439 class->pages_per_zspage = pages_per_zspage;
2440 class->objs_per_zspage = objs_per_zspage;
2441 spin_lock_init(&class->lock);
2442 pool->size_class[i] = class;
2443 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2445 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2450 /* debug only, don't abort if it fails */
2451 zs_pool_stat_create(pool, name);
2453 if (zs_register_migration(pool))
2457 * Not critical since shrinker is only used to trigger internal
2458 * defragmentation of the pool which is pretty optional thing. If
2459 * registration fails we still can use the pool normally and user can
2460 * trigger compaction manually. Thus, ignore return code.
2462 zs_register_shrinker(pool);
2467 zs_destroy_pool(pool);
2470 EXPORT_SYMBOL_GPL(zs_create_pool);
2472 void zs_destroy_pool(struct zs_pool *pool)
2476 zs_unregister_shrinker(pool);
2477 zs_unregister_migration(pool);
2478 zs_pool_stat_destroy(pool);
2480 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2482 struct size_class *class = pool->size_class[i];
2487 if (class->index != i)
2490 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2491 if (!list_empty(&class->fullness_list[fg])) {
2492 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2499 destroy_cache(pool);
2503 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2505 static int __init zs_init(void)
2509 ret = zsmalloc_mount();
2513 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2514 zs_cpu_prepare, zs_cpu_dead);
2519 zpool_register_driver(&zs_zpool_driver);
2532 static void __exit zs_exit(void)
2535 zpool_unregister_driver(&zs_zpool_driver);
2538 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2543 module_init(zs_init);
2544 module_exit(zs_exit);
2546 MODULE_LICENSE("Dual BSD/GPL");
2547 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");