2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->freelist(index): links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
23 * page->units: first object offset in a subpage of zspage
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_owner_priv_1: identifies the huge component page
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33 #include <linux/module.h>
34 #include <linux/kernel.h>
35 #include <linux/sched.h>
36 #include <linux/magic.h>
37 #include <linux/bitops.h>
38 #include <linux/errno.h>
39 #include <linux/highmem.h>
40 #include <linux/string.h>
41 #include <linux/slab.h>
42 #include <asm/tlbflush.h>
43 #include <asm/pgtable.h>
44 #include <linux/cpumask.h>
45 #include <linux/cpu.h>
46 #include <linux/vmalloc.h>
47 #include <linux/preempt.h>
48 #include <linux/spinlock.h>
49 #include <linux/shrinker.h>
50 #include <linux/types.h>
51 #include <linux/debugfs.h>
52 #include <linux/zsmalloc.h>
53 #include <linux/zpool.h>
54 #include <linux/mount.h>
55 #include <linux/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 * as 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 #ifdef CONFIG_PGTABLE_MAPPING
297 struct vm_struct *vm; /* vm area for mapping object that span pages */
299 char *vm_buf; /* copy buffer for objects that span pages */
301 char *vm_addr; /* address of kmap_atomic()'ed pages */
302 enum zs_mapmode vm_mm; /* mapping mode */
305 #ifdef CONFIG_COMPACTION
306 static int zs_register_migration(struct zs_pool *pool);
307 static void zs_unregister_migration(struct zs_pool *pool);
308 static void migrate_lock_init(struct zspage *zspage);
309 static void migrate_read_lock(struct zspage *zspage);
310 static void migrate_read_unlock(struct zspage *zspage);
311 static void kick_deferred_free(struct zs_pool *pool);
312 static void init_deferred_free(struct zs_pool *pool);
313 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
315 static int zsmalloc_mount(void) { return 0; }
316 static void zsmalloc_unmount(void) {}
317 static int zs_register_migration(struct zs_pool *pool) { return 0; }
318 static void zs_unregister_migration(struct zs_pool *pool) {}
319 static void migrate_lock_init(struct zspage *zspage) {}
320 static void migrate_read_lock(struct zspage *zspage) {}
321 static void migrate_read_unlock(struct zspage *zspage) {}
322 static void kick_deferred_free(struct zs_pool *pool) {}
323 static void init_deferred_free(struct zs_pool *pool) {}
324 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
327 static int create_cache(struct zs_pool *pool)
329 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
331 if (!pool->handle_cachep)
334 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
336 if (!pool->zspage_cachep) {
337 kmem_cache_destroy(pool->handle_cachep);
338 pool->handle_cachep = NULL;
345 static void destroy_cache(struct zs_pool *pool)
347 kmem_cache_destroy(pool->handle_cachep);
348 kmem_cache_destroy(pool->zspage_cachep);
351 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
353 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
354 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
357 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
359 kmem_cache_free(pool->handle_cachep, (void *)handle);
362 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
364 return kmem_cache_alloc(pool->zspage_cachep,
365 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
368 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
370 kmem_cache_free(pool->zspage_cachep, zspage);
373 static void record_obj(unsigned long handle, unsigned long obj)
376 * lsb of @obj represents handle lock while other bits
377 * represent object value the handle is pointing so
378 * updating shouldn't do store tearing.
380 WRITE_ONCE(*(unsigned long *)handle, obj);
387 static void *zs_zpool_create(const char *name, gfp_t gfp,
388 const struct zpool_ops *zpool_ops,
392 * Ignore global gfp flags: zs_malloc() may be invoked from
393 * different contexts and its caller must provide a valid
396 return zs_create_pool(name);
399 static void zs_zpool_destroy(void *pool)
401 zs_destroy_pool(pool);
404 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
405 unsigned long *handle)
407 *handle = zs_malloc(pool, size, gfp);
408 return *handle ? 0 : -1;
410 static void zs_zpool_free(void *pool, unsigned long handle)
412 zs_free(pool, handle);
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: /* fall through */
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_support_movable = true,
451 .malloc = zs_zpool_malloc,
452 .free = zs_zpool_free,
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);
587 static void __exit zs_stat_exit(void)
589 debugfs_remove_recursive(zs_stat_root);
592 static unsigned long zs_can_compact(struct size_class *class);
594 static int zs_stats_size_show(struct seq_file *s, void *v)
597 struct zs_pool *pool = s->private;
598 struct size_class *class;
600 unsigned long class_almost_full, class_almost_empty;
601 unsigned long obj_allocated, obj_used, pages_used, freeable;
602 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
603 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
604 unsigned long total_freeable = 0;
606 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
607 "class", "size", "almost_full", "almost_empty",
608 "obj_allocated", "obj_used", "pages_used",
609 "pages_per_zspage", "freeable");
611 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
612 class = pool->size_class[i];
614 if (class->index != i)
617 spin_lock(&class->lock);
618 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
619 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
620 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
621 obj_used = zs_stat_get(class, OBJ_USED);
622 freeable = zs_can_compact(class);
623 spin_unlock(&class->lock);
625 objs_per_zspage = class->objs_per_zspage;
626 pages_used = obj_allocated / objs_per_zspage *
627 class->pages_per_zspage;
629 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
630 " %10lu %10lu %16d %8lu\n",
631 i, class->size, class_almost_full, class_almost_empty,
632 obj_allocated, obj_used, pages_used,
633 class->pages_per_zspage, freeable);
635 total_class_almost_full += class_almost_full;
636 total_class_almost_empty += class_almost_empty;
637 total_objs += obj_allocated;
638 total_used_objs += obj_used;
639 total_pages += pages_used;
640 total_freeable += freeable;
644 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
645 "Total", "", total_class_almost_full,
646 total_class_almost_empty, total_objs,
647 total_used_objs, total_pages, "", total_freeable);
651 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
653 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
656 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
660 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
662 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
663 &zs_stats_size_fops);
666 static void zs_pool_stat_destroy(struct zs_pool *pool)
668 debugfs_remove_recursive(pool->stat_dentry);
671 #else /* CONFIG_ZSMALLOC_STAT */
672 static void __init zs_stat_init(void)
676 static void __exit zs_stat_exit(void)
680 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
684 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
691 * For each size class, zspages are divided into different groups
692 * depending on how "full" they are. This was done so that we could
693 * easily find empty or nearly empty zspages when we try to shrink
694 * the pool (not yet implemented). This function returns fullness
695 * status of the given page.
697 static enum fullness_group get_fullness_group(struct size_class *class,
698 struct zspage *zspage)
700 int inuse, objs_per_zspage;
701 enum fullness_group fg;
703 inuse = get_zspage_inuse(zspage);
704 objs_per_zspage = class->objs_per_zspage;
708 else if (inuse == objs_per_zspage)
710 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
711 fg = ZS_ALMOST_EMPTY;
719 * Each size class maintains various freelists and zspages are assigned
720 * to one of these freelists based on the number of live objects they
721 * have. This functions inserts the given zspage into the freelist
722 * identified by <class, fullness_group>.
724 static void insert_zspage(struct size_class *class,
725 struct zspage *zspage,
726 enum fullness_group fullness)
730 zs_stat_inc(class, fullness, 1);
731 head = list_first_entry_or_null(&class->fullness_list[fullness],
732 struct zspage, list);
734 * We want to see more ZS_FULL pages and less almost empty/full.
735 * Put pages with higher ->inuse first.
738 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
739 list_add(&zspage->list, &head->list);
743 list_add(&zspage->list, &class->fullness_list[fullness]);
747 * This function removes the given zspage from the freelist identified
748 * by <class, fullness_group>.
750 static void remove_zspage(struct size_class *class,
751 struct zspage *zspage,
752 enum fullness_group fullness)
754 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
755 VM_BUG_ON(is_zspage_isolated(zspage));
757 list_del_init(&zspage->list);
758 zs_stat_dec(class, fullness, 1);
762 * Each size class maintains zspages in different fullness groups depending
763 * on the number of live objects they contain. When allocating or freeing
764 * objects, the fullness status of the page can change, say, from ALMOST_FULL
765 * to ALMOST_EMPTY when freeing an object. This function checks if such
766 * a status change has occurred for the given page and accordingly moves the
767 * page from the freelist of the old fullness group to that of the new
770 static enum fullness_group fix_fullness_group(struct size_class *class,
771 struct zspage *zspage)
774 enum fullness_group currfg, newfg;
776 get_zspage_mapping(zspage, &class_idx, &currfg);
777 newfg = get_fullness_group(class, zspage);
781 if (!is_zspage_isolated(zspage)) {
782 remove_zspage(class, zspage, currfg);
783 insert_zspage(class, zspage, newfg);
786 set_zspage_mapping(zspage, class_idx, newfg);
793 * We have to decide on how many pages to link together
794 * to form a zspage for each size class. This is important
795 * to reduce wastage due to unusable space left at end of
796 * each zspage which is given as:
797 * wastage = Zp % class_size
798 * usage = Zp - wastage
799 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
801 * For example, for size class of 3/8 * PAGE_SIZE, we should
802 * link together 3 PAGE_SIZE sized pages to form a zspage
803 * since then we can perfectly fit in 8 such objects.
805 static int get_pages_per_zspage(int class_size)
807 int i, max_usedpc = 0;
808 /* zspage order which gives maximum used size per KB */
809 int max_usedpc_order = 1;
811 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
815 zspage_size = i * PAGE_SIZE;
816 waste = zspage_size % class_size;
817 usedpc = (zspage_size - waste) * 100 / zspage_size;
819 if (usedpc > max_usedpc) {
821 max_usedpc_order = i;
825 return max_usedpc_order;
828 static struct zspage *get_zspage(struct page *page)
830 struct zspage *zspage = (struct zspage *)page->private;
832 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
836 static struct page *get_next_page(struct page *page)
838 if (unlikely(PageHugeObject(page)))
841 return page->freelist;
845 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
846 * @obj: the encoded object value
847 * @page: page object resides in zspage
848 * @obj_idx: object index
850 static void obj_to_location(unsigned long obj, struct page **page,
851 unsigned int *obj_idx)
853 obj >>= OBJ_TAG_BITS;
854 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
855 *obj_idx = (obj & OBJ_INDEX_MASK);
859 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
860 * @page: page object resides in zspage
861 * @obj_idx: object index
863 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
867 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
868 obj |= obj_idx & OBJ_INDEX_MASK;
869 obj <<= OBJ_TAG_BITS;
874 static unsigned long handle_to_obj(unsigned long handle)
876 return *(unsigned long *)handle;
879 static unsigned long obj_to_head(struct page *page, void *obj)
881 if (unlikely(PageHugeObject(page))) {
882 VM_BUG_ON_PAGE(!is_first_page(page), page);
885 return *(unsigned long *)obj;
888 static inline int testpin_tag(unsigned long handle)
890 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
893 static inline int trypin_tag(unsigned long handle)
895 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
898 static void pin_tag(unsigned long handle)
900 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
903 static void unpin_tag(unsigned long handle)
905 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
908 static void reset_page(struct page *page)
910 __ClearPageMovable(page);
911 ClearPagePrivate(page);
912 set_page_private(page, 0);
913 page_mapcount_reset(page);
914 ClearPageHugeObject(page);
915 page->freelist = NULL;
918 static int trylock_zspage(struct zspage *zspage)
920 struct page *cursor, *fail;
922 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
923 get_next_page(cursor)) {
924 if (!trylock_page(cursor)) {
932 for (cursor = get_first_page(zspage); cursor != fail; cursor =
933 get_next_page(cursor))
939 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
940 struct zspage *zspage)
942 struct page *page, *next;
943 enum fullness_group fg;
944 unsigned int class_idx;
946 get_zspage_mapping(zspage, &class_idx, &fg);
948 assert_spin_locked(&class->lock);
950 VM_BUG_ON(get_zspage_inuse(zspage));
951 VM_BUG_ON(fg != ZS_EMPTY);
953 next = page = get_first_page(zspage);
955 VM_BUG_ON_PAGE(!PageLocked(page), page);
956 next = get_next_page(page);
959 dec_zone_page_state(page, NR_ZSPAGES);
962 } while (page != NULL);
964 cache_free_zspage(pool, zspage);
966 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
967 atomic_long_sub(class->pages_per_zspage,
968 &pool->pages_allocated);
971 static void free_zspage(struct zs_pool *pool, struct size_class *class,
972 struct zspage *zspage)
974 VM_BUG_ON(get_zspage_inuse(zspage));
975 VM_BUG_ON(list_empty(&zspage->list));
977 if (!trylock_zspage(zspage)) {
978 kick_deferred_free(pool);
982 remove_zspage(class, zspage, ZS_EMPTY);
983 __free_zspage(pool, class, zspage);
986 /* Initialize a newly allocated zspage */
987 static void init_zspage(struct size_class *class, struct zspage *zspage)
989 unsigned int freeobj = 1;
990 unsigned long off = 0;
991 struct page *page = get_first_page(zspage);
994 struct page *next_page;
995 struct link_free *link;
998 set_first_obj_offset(page, off);
1000 vaddr = kmap_atomic(page);
1001 link = (struct link_free *)vaddr + off / sizeof(*link);
1003 while ((off += class->size) < PAGE_SIZE) {
1004 link->next = freeobj++ << OBJ_TAG_BITS;
1005 link += class->size / sizeof(*link);
1009 * We now come to the last (full or partial) object on this
1010 * page, which must point to the first object on the next
1013 next_page = get_next_page(page);
1015 link->next = freeobj++ << OBJ_TAG_BITS;
1018 * Reset OBJ_TAG_BITS bit to last link to tell
1019 * whether it's allocated object or not.
1021 link->next = -1UL << OBJ_TAG_BITS;
1023 kunmap_atomic(vaddr);
1028 set_freeobj(zspage, 0);
1031 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1032 struct page *pages[])
1036 struct page *prev_page = NULL;
1037 int nr_pages = class->pages_per_zspage;
1040 * Allocate individual pages and link them together as:
1041 * 1. all pages are linked together using page->freelist
1042 * 2. each sub-page point to zspage using page->private
1044 * we set PG_private to identify the first page (i.e. no other sub-page
1045 * has this flag set).
1047 for (i = 0; i < nr_pages; i++) {
1049 set_page_private(page, (unsigned long)zspage);
1050 page->freelist = NULL;
1052 zspage->first_page = page;
1053 SetPagePrivate(page);
1054 if (unlikely(class->objs_per_zspage == 1 &&
1055 class->pages_per_zspage == 1))
1056 SetPageHugeObject(page);
1058 prev_page->freelist = page;
1065 * Allocate a zspage for the given size class
1067 static struct zspage *alloc_zspage(struct zs_pool *pool,
1068 struct size_class *class,
1072 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1073 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1078 memset(zspage, 0, sizeof(struct zspage));
1079 zspage->magic = ZSPAGE_MAGIC;
1080 migrate_lock_init(zspage);
1082 for (i = 0; i < class->pages_per_zspage; i++) {
1085 page = alloc_page(gfp);
1088 dec_zone_page_state(pages[i], NR_ZSPAGES);
1089 __free_page(pages[i]);
1091 cache_free_zspage(pool, zspage);
1095 inc_zone_page_state(page, NR_ZSPAGES);
1099 create_page_chain(class, zspage, pages);
1100 init_zspage(class, zspage);
1105 static struct zspage *find_get_zspage(struct size_class *class)
1108 struct zspage *zspage;
1110 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1111 zspage = list_first_entry_or_null(&class->fullness_list[i],
1112 struct zspage, list);
1120 #ifdef CONFIG_PGTABLE_MAPPING
1121 static inline int __zs_cpu_up(struct mapping_area *area)
1124 * Make sure we don't leak memory if a cpu UP notification
1125 * and zs_init() race and both call zs_cpu_up() on the same cpu
1129 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1135 static inline void __zs_cpu_down(struct mapping_area *area)
1138 free_vm_area(area->vm);
1142 static inline void *__zs_map_object(struct mapping_area *area,
1143 struct page *pages[2], int off, int size)
1145 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1146 area->vm_addr = area->vm->addr;
1147 return area->vm_addr + off;
1150 static inline void __zs_unmap_object(struct mapping_area *area,
1151 struct page *pages[2], int off, int size)
1153 unsigned long addr = (unsigned long)area->vm_addr;
1155 unmap_kernel_range(addr, PAGE_SIZE * 2);
1158 #else /* CONFIG_PGTABLE_MAPPING */
1160 static inline int __zs_cpu_up(struct mapping_area *area)
1163 * Make sure we don't leak memory if a cpu UP notification
1164 * and zs_init() race and both call zs_cpu_up() on the same cpu
1168 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1174 static inline void __zs_cpu_down(struct mapping_area *area)
1176 kfree(area->vm_buf);
1177 area->vm_buf = NULL;
1180 static void *__zs_map_object(struct mapping_area *area,
1181 struct page *pages[2], int off, int size)
1185 char *buf = area->vm_buf;
1187 /* disable page faults to match kmap_atomic() return conditions */
1188 pagefault_disable();
1190 /* no read fastpath */
1191 if (area->vm_mm == ZS_MM_WO)
1194 sizes[0] = PAGE_SIZE - off;
1195 sizes[1] = size - sizes[0];
1197 /* copy object to per-cpu buffer */
1198 addr = kmap_atomic(pages[0]);
1199 memcpy(buf, addr + off, sizes[0]);
1200 kunmap_atomic(addr);
1201 addr = kmap_atomic(pages[1]);
1202 memcpy(buf + sizes[0], addr, sizes[1]);
1203 kunmap_atomic(addr);
1205 return area->vm_buf;
1208 static void __zs_unmap_object(struct mapping_area *area,
1209 struct page *pages[2], int off, int size)
1215 /* no write fastpath */
1216 if (area->vm_mm == ZS_MM_RO)
1220 buf = buf + ZS_HANDLE_SIZE;
1221 size -= ZS_HANDLE_SIZE;
1222 off += ZS_HANDLE_SIZE;
1224 sizes[0] = PAGE_SIZE - off;
1225 sizes[1] = size - sizes[0];
1227 /* copy per-cpu buffer to object */
1228 addr = kmap_atomic(pages[0]);
1229 memcpy(addr + off, buf, sizes[0]);
1230 kunmap_atomic(addr);
1231 addr = kmap_atomic(pages[1]);
1232 memcpy(addr, buf + sizes[0], sizes[1]);
1233 kunmap_atomic(addr);
1236 /* enable page faults to match kunmap_atomic() return conditions */
1240 #endif /* CONFIG_PGTABLE_MAPPING */
1242 static int zs_cpu_prepare(unsigned int cpu)
1244 struct mapping_area *area;
1246 area = &per_cpu(zs_map_area, cpu);
1247 return __zs_cpu_up(area);
1250 static int zs_cpu_dead(unsigned int cpu)
1252 struct mapping_area *area;
1254 area = &per_cpu(zs_map_area, cpu);
1255 __zs_cpu_down(area);
1259 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1260 int objs_per_zspage)
1262 if (prev->pages_per_zspage == pages_per_zspage &&
1263 prev->objs_per_zspage == objs_per_zspage)
1269 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1271 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1274 unsigned long zs_get_total_pages(struct zs_pool *pool)
1276 return atomic_long_read(&pool->pages_allocated);
1278 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1281 * zs_map_object - get address of allocated object from handle.
1282 * @pool: pool from which the object was allocated
1283 * @handle: handle returned from zs_malloc
1284 * @mm: maping mode to use
1286 * Before using an object allocated from zs_malloc, it must be mapped using
1287 * this function. When done with the object, it must be unmapped using
1290 * Only one object can be mapped per cpu at a time. There is no protection
1291 * against nested mappings.
1293 * This function returns with preemption and page faults disabled.
1295 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1298 struct zspage *zspage;
1300 unsigned long obj, off;
1301 unsigned int obj_idx;
1303 unsigned int class_idx;
1304 enum fullness_group fg;
1305 struct size_class *class;
1306 struct mapping_area *area;
1307 struct page *pages[2];
1311 * Because we use per-cpu mapping areas shared among the
1312 * pools/users, we can't allow mapping in interrupt context
1313 * because it can corrupt another users mappings.
1315 BUG_ON(in_interrupt());
1317 /* From now on, migration cannot move the object */
1320 obj = handle_to_obj(handle);
1321 obj_to_location(obj, &page, &obj_idx);
1322 zspage = get_zspage(page);
1324 /* migration cannot move any subpage in this zspage */
1325 migrate_read_lock(zspage);
1327 get_zspage_mapping(zspage, &class_idx, &fg);
1328 class = pool->size_class[class_idx];
1329 off = (class->size * obj_idx) & ~PAGE_MASK;
1331 area = &get_cpu_var(zs_map_area);
1333 if (off + class->size <= PAGE_SIZE) {
1334 /* this object is contained entirely within a page */
1335 area->vm_addr = kmap_atomic(page);
1336 ret = area->vm_addr + off;
1340 /* this object spans two pages */
1342 pages[1] = get_next_page(page);
1345 ret = __zs_map_object(area, pages, off, class->size);
1347 if (likely(!PageHugeObject(page)))
1348 ret += ZS_HANDLE_SIZE;
1352 EXPORT_SYMBOL_GPL(zs_map_object);
1354 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1356 struct zspage *zspage;
1358 unsigned long obj, off;
1359 unsigned int obj_idx;
1361 unsigned int class_idx;
1362 enum fullness_group fg;
1363 struct size_class *class;
1364 struct mapping_area *area;
1366 obj = handle_to_obj(handle);
1367 obj_to_location(obj, &page, &obj_idx);
1368 zspage = get_zspage(page);
1369 get_zspage_mapping(zspage, &class_idx, &fg);
1370 class = pool->size_class[class_idx];
1371 off = (class->size * obj_idx) & ~PAGE_MASK;
1373 area = this_cpu_ptr(&zs_map_area);
1374 if (off + class->size <= PAGE_SIZE)
1375 kunmap_atomic(area->vm_addr);
1377 struct page *pages[2];
1380 pages[1] = get_next_page(page);
1383 __zs_unmap_object(area, pages, off, class->size);
1385 put_cpu_var(zs_map_area);
1387 migrate_read_unlock(zspage);
1390 EXPORT_SYMBOL_GPL(zs_unmap_object);
1393 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1394 * zsmalloc &size_class.
1395 * @pool: zsmalloc pool to use
1397 * The function returns the size of the first huge class - any object of equal
1398 * or bigger size will be stored in zspage consisting of a single physical
1401 * Context: Any context.
1403 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1405 size_t zs_huge_class_size(struct zs_pool *pool)
1407 return huge_class_size;
1409 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1411 static unsigned long obj_malloc(struct size_class *class,
1412 struct zspage *zspage, unsigned long handle)
1414 int i, nr_page, offset;
1416 struct link_free *link;
1418 struct page *m_page;
1419 unsigned long m_offset;
1422 handle |= OBJ_ALLOCATED_TAG;
1423 obj = get_freeobj(zspage);
1425 offset = obj * class->size;
1426 nr_page = offset >> PAGE_SHIFT;
1427 m_offset = offset & ~PAGE_MASK;
1428 m_page = get_first_page(zspage);
1430 for (i = 0; i < nr_page; i++)
1431 m_page = get_next_page(m_page);
1433 vaddr = kmap_atomic(m_page);
1434 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1435 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1436 if (likely(!PageHugeObject(m_page)))
1437 /* record handle in the header of allocated chunk */
1438 link->handle = handle;
1440 /* record handle to page->index */
1441 zspage->first_page->index = handle;
1443 kunmap_atomic(vaddr);
1444 mod_zspage_inuse(zspage, 1);
1445 zs_stat_inc(class, OBJ_USED, 1);
1447 obj = location_to_obj(m_page, obj);
1454 * zs_malloc - Allocate block of given size from pool.
1455 * @pool: pool to allocate from
1456 * @size: size of block to allocate
1457 * @gfp: gfp flags when allocating object
1459 * On success, handle to the allocated object is returned,
1461 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1463 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1465 unsigned long handle, obj;
1466 struct size_class *class;
1467 enum fullness_group newfg;
1468 struct zspage *zspage;
1470 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1473 handle = cache_alloc_handle(pool, gfp);
1477 /* extra space in chunk to keep the handle */
1478 size += ZS_HANDLE_SIZE;
1479 class = pool->size_class[get_size_class_index(size)];
1481 spin_lock(&class->lock);
1482 zspage = find_get_zspage(class);
1483 if (likely(zspage)) {
1484 obj = obj_malloc(class, zspage, handle);
1485 /* Now move the zspage to another fullness group, if required */
1486 fix_fullness_group(class, zspage);
1487 record_obj(handle, obj);
1488 spin_unlock(&class->lock);
1493 spin_unlock(&class->lock);
1495 zspage = alloc_zspage(pool, class, gfp);
1497 cache_free_handle(pool, handle);
1501 spin_lock(&class->lock);
1502 obj = obj_malloc(class, zspage, handle);
1503 newfg = get_fullness_group(class, zspage);
1504 insert_zspage(class, zspage, newfg);
1505 set_zspage_mapping(zspage, class->index, newfg);
1506 record_obj(handle, obj);
1507 atomic_long_add(class->pages_per_zspage,
1508 &pool->pages_allocated);
1509 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1511 /* We completely set up zspage so mark them as movable */
1512 SetZsPageMovable(pool, zspage);
1513 spin_unlock(&class->lock);
1517 EXPORT_SYMBOL_GPL(zs_malloc);
1519 static void obj_free(struct size_class *class, unsigned long obj)
1521 struct link_free *link;
1522 struct zspage *zspage;
1523 struct page *f_page;
1524 unsigned long f_offset;
1525 unsigned int f_objidx;
1528 obj &= ~OBJ_ALLOCATED_TAG;
1529 obj_to_location(obj, &f_page, &f_objidx);
1530 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1531 zspage = get_zspage(f_page);
1533 vaddr = kmap_atomic(f_page);
1535 /* Insert this object in containing zspage's freelist */
1536 link = (struct link_free *)(vaddr + f_offset);
1537 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1538 kunmap_atomic(vaddr);
1539 set_freeobj(zspage, f_objidx);
1540 mod_zspage_inuse(zspage, -1);
1541 zs_stat_dec(class, OBJ_USED, 1);
1544 void zs_free(struct zs_pool *pool, unsigned long handle)
1546 struct zspage *zspage;
1547 struct page *f_page;
1549 unsigned int f_objidx;
1551 struct size_class *class;
1552 enum fullness_group fullness;
1555 if (unlikely(!handle))
1559 obj = handle_to_obj(handle);
1560 obj_to_location(obj, &f_page, &f_objidx);
1561 zspage = get_zspage(f_page);
1563 migrate_read_lock(zspage);
1565 get_zspage_mapping(zspage, &class_idx, &fullness);
1566 class = pool->size_class[class_idx];
1568 spin_lock(&class->lock);
1569 obj_free(class, obj);
1570 fullness = fix_fullness_group(class, zspage);
1571 if (fullness != ZS_EMPTY) {
1572 migrate_read_unlock(zspage);
1576 isolated = is_zspage_isolated(zspage);
1577 migrate_read_unlock(zspage);
1578 /* If zspage is isolated, zs_page_putback will free the zspage */
1579 if (likely(!isolated))
1580 free_zspage(pool, class, zspage);
1583 spin_unlock(&class->lock);
1585 cache_free_handle(pool, handle);
1587 EXPORT_SYMBOL_GPL(zs_free);
1589 static void zs_object_copy(struct size_class *class, unsigned long dst,
1592 struct page *s_page, *d_page;
1593 unsigned int s_objidx, d_objidx;
1594 unsigned long s_off, d_off;
1595 void *s_addr, *d_addr;
1596 int s_size, d_size, size;
1599 s_size = d_size = class->size;
1601 obj_to_location(src, &s_page, &s_objidx);
1602 obj_to_location(dst, &d_page, &d_objidx);
1604 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1605 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1607 if (s_off + class->size > PAGE_SIZE)
1608 s_size = PAGE_SIZE - s_off;
1610 if (d_off + class->size > PAGE_SIZE)
1611 d_size = PAGE_SIZE - d_off;
1613 s_addr = kmap_atomic(s_page);
1614 d_addr = kmap_atomic(d_page);
1617 size = min(s_size, d_size);
1618 memcpy(d_addr + d_off, s_addr + s_off, size);
1621 if (written == class->size)
1629 if (s_off >= PAGE_SIZE) {
1630 kunmap_atomic(d_addr);
1631 kunmap_atomic(s_addr);
1632 s_page = get_next_page(s_page);
1633 s_addr = kmap_atomic(s_page);
1634 d_addr = kmap_atomic(d_page);
1635 s_size = class->size - written;
1639 if (d_off >= PAGE_SIZE) {
1640 kunmap_atomic(d_addr);
1641 d_page = get_next_page(d_page);
1642 d_addr = kmap_atomic(d_page);
1643 d_size = class->size - written;
1648 kunmap_atomic(d_addr);
1649 kunmap_atomic(s_addr);
1653 * Find alloced object in zspage from index object and
1656 static unsigned long find_alloced_obj(struct size_class *class,
1657 struct page *page, int *obj_idx)
1661 int index = *obj_idx;
1662 unsigned long handle = 0;
1663 void *addr = kmap_atomic(page);
1665 offset = get_first_obj_offset(page);
1666 offset += class->size * index;
1668 while (offset < PAGE_SIZE) {
1669 head = obj_to_head(page, addr + offset);
1670 if (head & OBJ_ALLOCATED_TAG) {
1671 handle = head & ~OBJ_ALLOCATED_TAG;
1672 if (trypin_tag(handle))
1677 offset += class->size;
1681 kunmap_atomic(addr);
1688 struct zs_compact_control {
1689 /* Source spage for migration which could be a subpage of zspage */
1690 struct page *s_page;
1691 /* Destination page for migration which should be a first page
1693 struct page *d_page;
1694 /* Starting object index within @s_page which used for live object
1695 * in the subpage. */
1699 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1700 struct zs_compact_control *cc)
1702 unsigned long used_obj, free_obj;
1703 unsigned long handle;
1704 struct page *s_page = cc->s_page;
1705 struct page *d_page = cc->d_page;
1706 int obj_idx = cc->obj_idx;
1710 handle = find_alloced_obj(class, s_page, &obj_idx);
1712 s_page = get_next_page(s_page);
1719 /* Stop if there is no more space */
1720 if (zspage_full(class, get_zspage(d_page))) {
1726 used_obj = handle_to_obj(handle);
1727 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1728 zs_object_copy(class, free_obj, used_obj);
1731 * record_obj updates handle's value to free_obj and it will
1732 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1733 * breaks synchronization using pin_tag(e,g, zs_free) so
1734 * let's keep the lock bit.
1736 free_obj |= BIT(HANDLE_PIN_BIT);
1737 record_obj(handle, free_obj);
1739 obj_free(class, used_obj);
1742 /* Remember last position in this iteration */
1743 cc->s_page = s_page;
1744 cc->obj_idx = obj_idx;
1749 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1752 struct zspage *zspage;
1753 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1756 fg[0] = ZS_ALMOST_FULL;
1757 fg[1] = ZS_ALMOST_EMPTY;
1760 for (i = 0; i < 2; i++) {
1761 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1762 struct zspage, list);
1764 VM_BUG_ON(is_zspage_isolated(zspage));
1765 remove_zspage(class, zspage, fg[i]);
1774 * putback_zspage - add @zspage into right class's fullness list
1775 * @class: destination class
1776 * @zspage: target page
1778 * Return @zspage's fullness_group
1780 static enum fullness_group putback_zspage(struct size_class *class,
1781 struct zspage *zspage)
1783 enum fullness_group fullness;
1785 VM_BUG_ON(is_zspage_isolated(zspage));
1787 fullness = get_fullness_group(class, zspage);
1788 insert_zspage(class, zspage, fullness);
1789 set_zspage_mapping(zspage, class->index, fullness);
1794 #ifdef CONFIG_COMPACTION
1796 * To prevent zspage destroy during migration, zspage freeing should
1797 * hold locks of all pages in the zspage.
1799 static void lock_zspage(struct zspage *zspage)
1801 struct page *page = get_first_page(zspage);
1805 } while ((page = get_next_page(page)) != NULL);
1808 static int zs_init_fs_context(struct fs_context *fc)
1810 return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1813 static struct file_system_type zsmalloc_fs = {
1815 .init_fs_context = zs_init_fs_context,
1816 .kill_sb = kill_anon_super,
1819 static int zsmalloc_mount(void)
1823 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1824 if (IS_ERR(zsmalloc_mnt))
1825 ret = PTR_ERR(zsmalloc_mnt);
1830 static void zsmalloc_unmount(void)
1832 kern_unmount(zsmalloc_mnt);
1835 static void migrate_lock_init(struct zspage *zspage)
1837 rwlock_init(&zspage->lock);
1840 static void migrate_read_lock(struct zspage *zspage)
1842 read_lock(&zspage->lock);
1845 static void migrate_read_unlock(struct zspage *zspage)
1847 read_unlock(&zspage->lock);
1850 static void migrate_write_lock(struct zspage *zspage)
1852 write_lock(&zspage->lock);
1855 static void migrate_write_unlock(struct zspage *zspage)
1857 write_unlock(&zspage->lock);
1860 /* Number of isolated subpage for *page migration* in this zspage */
1861 static void inc_zspage_isolation(struct zspage *zspage)
1866 static void dec_zspage_isolation(struct zspage *zspage)
1871 static void putback_zspage_deferred(struct zs_pool *pool,
1872 struct size_class *class,
1873 struct zspage *zspage)
1875 enum fullness_group fg;
1877 fg = putback_zspage(class, zspage);
1879 schedule_work(&pool->free_work);
1883 static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1885 VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1886 atomic_long_dec(&pool->isolated_pages);
1888 * There's no possibility of racing, since wait_for_isolated_drain()
1889 * checks the isolated count under &class->lock after enqueuing
1890 * on migration_wait.
1892 if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1893 wake_up_all(&pool->migration_wait);
1896 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1897 struct page *newpage, struct page *oldpage)
1900 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1903 page = get_first_page(zspage);
1905 if (page == oldpage)
1906 pages[idx] = newpage;
1910 } while ((page = get_next_page(page)) != NULL);
1912 create_page_chain(class, zspage, pages);
1913 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1914 if (unlikely(PageHugeObject(oldpage)))
1915 newpage->index = oldpage->index;
1916 __SetPageMovable(newpage, page_mapping(oldpage));
1919 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1921 struct zs_pool *pool;
1922 struct size_class *class;
1924 enum fullness_group fullness;
1925 struct zspage *zspage;
1926 struct address_space *mapping;
1929 * Page is locked so zspage couldn't be destroyed. For detail, look at
1930 * lock_zspage in free_zspage.
1932 VM_BUG_ON_PAGE(!PageMovable(page), page);
1933 VM_BUG_ON_PAGE(PageIsolated(page), page);
1935 zspage = get_zspage(page);
1938 * Without class lock, fullness could be stale while class_idx is okay
1939 * because class_idx is constant unless page is freed so we should get
1940 * fullness again under class lock.
1942 get_zspage_mapping(zspage, &class_idx, &fullness);
1943 mapping = page_mapping(page);
1944 pool = mapping->private_data;
1945 class = pool->size_class[class_idx];
1947 spin_lock(&class->lock);
1948 if (get_zspage_inuse(zspage) == 0) {
1949 spin_unlock(&class->lock);
1953 /* zspage is isolated for object migration */
1954 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1955 spin_unlock(&class->lock);
1960 * If this is first time isolation for the zspage, isolate zspage from
1961 * size_class to prevent further object allocation from the zspage.
1963 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1964 get_zspage_mapping(zspage, &class_idx, &fullness);
1965 atomic_long_inc(&pool->isolated_pages);
1966 remove_zspage(class, zspage, fullness);
1969 inc_zspage_isolation(zspage);
1970 spin_unlock(&class->lock);
1975 static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1976 struct page *page, enum migrate_mode mode)
1978 struct zs_pool *pool;
1979 struct size_class *class;
1981 enum fullness_group fullness;
1982 struct zspage *zspage;
1984 void *s_addr, *d_addr, *addr;
1986 unsigned long handle, head;
1987 unsigned long old_obj, new_obj;
1988 unsigned int obj_idx;
1992 * We cannot support the _NO_COPY case here, because copy needs to
1993 * happen under the zs lock, which does not work with
1994 * MIGRATE_SYNC_NO_COPY workflow.
1996 if (mode == MIGRATE_SYNC_NO_COPY)
1999 VM_BUG_ON_PAGE(!PageMovable(page), page);
2000 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2002 zspage = get_zspage(page);
2004 /* Concurrent compactor cannot migrate any subpage in zspage */
2005 migrate_write_lock(zspage);
2006 get_zspage_mapping(zspage, &class_idx, &fullness);
2007 pool = mapping->private_data;
2008 class = pool->size_class[class_idx];
2009 offset = get_first_obj_offset(page);
2011 spin_lock(&class->lock);
2012 if (!get_zspage_inuse(zspage)) {
2014 * Set "offset" to end of the page so that every loops
2015 * skips unnecessary object scanning.
2021 s_addr = kmap_atomic(page);
2022 while (pos < PAGE_SIZE) {
2023 head = obj_to_head(page, s_addr + pos);
2024 if (head & OBJ_ALLOCATED_TAG) {
2025 handle = head & ~OBJ_ALLOCATED_TAG;
2026 if (!trypin_tag(handle))
2033 * Here, any user cannot access all objects in the zspage so let's move.
2035 d_addr = kmap_atomic(newpage);
2036 memcpy(d_addr, s_addr, PAGE_SIZE);
2037 kunmap_atomic(d_addr);
2039 for (addr = s_addr + offset; addr < s_addr + pos;
2040 addr += class->size) {
2041 head = obj_to_head(page, addr);
2042 if (head & OBJ_ALLOCATED_TAG) {
2043 handle = head & ~OBJ_ALLOCATED_TAG;
2044 if (!testpin_tag(handle))
2047 old_obj = handle_to_obj(handle);
2048 obj_to_location(old_obj, &dummy, &obj_idx);
2049 new_obj = (unsigned long)location_to_obj(newpage,
2051 new_obj |= BIT(HANDLE_PIN_BIT);
2052 record_obj(handle, new_obj);
2056 replace_sub_page(class, zspage, newpage, page);
2059 dec_zspage_isolation(zspage);
2062 * Page migration is done so let's putback isolated zspage to
2063 * the list if @page is final isolated subpage in the zspage.
2065 if (!is_zspage_isolated(zspage)) {
2067 * We cannot race with zs_destroy_pool() here because we wait
2068 * for isolation to hit zero before we start destroying.
2069 * Also, we ensure that everyone can see pool->destroying before
2072 putback_zspage_deferred(pool, class, zspage);
2073 zs_pool_dec_isolated(pool);
2080 ret = MIGRATEPAGE_SUCCESS;
2082 for (addr = s_addr + offset; addr < s_addr + pos;
2083 addr += class->size) {
2084 head = obj_to_head(page, addr);
2085 if (head & OBJ_ALLOCATED_TAG) {
2086 handle = head & ~OBJ_ALLOCATED_TAG;
2087 if (!testpin_tag(handle))
2092 kunmap_atomic(s_addr);
2093 spin_unlock(&class->lock);
2094 migrate_write_unlock(zspage);
2099 static void zs_page_putback(struct page *page)
2101 struct zs_pool *pool;
2102 struct size_class *class;
2104 enum fullness_group fg;
2105 struct address_space *mapping;
2106 struct zspage *zspage;
2108 VM_BUG_ON_PAGE(!PageMovable(page), page);
2109 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2111 zspage = get_zspage(page);
2112 get_zspage_mapping(zspage, &class_idx, &fg);
2113 mapping = page_mapping(page);
2114 pool = mapping->private_data;
2115 class = pool->size_class[class_idx];
2117 spin_lock(&class->lock);
2118 dec_zspage_isolation(zspage);
2119 if (!is_zspage_isolated(zspage)) {
2121 * Due to page_lock, we cannot free zspage immediately
2124 putback_zspage_deferred(pool, class, zspage);
2125 zs_pool_dec_isolated(pool);
2127 spin_unlock(&class->lock);
2130 static const struct address_space_operations zsmalloc_aops = {
2131 .isolate_page = zs_page_isolate,
2132 .migratepage = zs_page_migrate,
2133 .putback_page = zs_page_putback,
2136 static int zs_register_migration(struct zs_pool *pool)
2138 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2139 if (IS_ERR(pool->inode)) {
2144 pool->inode->i_mapping->private_data = pool;
2145 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2149 static bool pool_isolated_are_drained(struct zs_pool *pool)
2151 return atomic_long_read(&pool->isolated_pages) == 0;
2154 /* Function for resolving migration */
2155 static void wait_for_isolated_drain(struct zs_pool *pool)
2159 * We're in the process of destroying the pool, so there are no
2160 * active allocations. zs_page_isolate() fails for completely free
2161 * zspages, so we need only wait for the zs_pool's isolated
2162 * count to hit zero.
2164 wait_event(pool->migration_wait,
2165 pool_isolated_are_drained(pool));
2168 static void zs_unregister_migration(struct zs_pool *pool)
2170 pool->destroying = true;
2172 * We need a memory barrier here to ensure global visibility of
2173 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2174 * case we don't care, or it will be > 0 and pool->destroying will
2175 * ensure that we wake up once isolation hits 0.
2178 wait_for_isolated_drain(pool); /* This can block */
2179 flush_work(&pool->free_work);
2184 * Caller should hold page_lock of all pages in the zspage
2185 * In here, we cannot use zspage meta data.
2187 static void async_free_zspage(struct work_struct *work)
2190 struct size_class *class;
2191 unsigned int class_idx;
2192 enum fullness_group fullness;
2193 struct zspage *zspage, *tmp;
2194 LIST_HEAD(free_pages);
2195 struct zs_pool *pool = container_of(work, struct zs_pool,
2198 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2199 class = pool->size_class[i];
2200 if (class->index != i)
2203 spin_lock(&class->lock);
2204 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2205 spin_unlock(&class->lock);
2209 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2210 list_del(&zspage->list);
2211 lock_zspage(zspage);
2213 get_zspage_mapping(zspage, &class_idx, &fullness);
2214 VM_BUG_ON(fullness != ZS_EMPTY);
2215 class = pool->size_class[class_idx];
2216 spin_lock(&class->lock);
2217 __free_zspage(pool, pool->size_class[class_idx], zspage);
2218 spin_unlock(&class->lock);
2222 static void kick_deferred_free(struct zs_pool *pool)
2224 schedule_work(&pool->free_work);
2227 static void init_deferred_free(struct zs_pool *pool)
2229 INIT_WORK(&pool->free_work, async_free_zspage);
2232 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2234 struct page *page = get_first_page(zspage);
2237 WARN_ON(!trylock_page(page));
2238 __SetPageMovable(page, pool->inode->i_mapping);
2240 } while ((page = get_next_page(page)) != NULL);
2246 * Based on the number of unused allocated objects calculate
2247 * and return the number of pages that we can free.
2249 static unsigned long zs_can_compact(struct size_class *class)
2251 unsigned long obj_wasted;
2252 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2253 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2255 if (obj_allocated <= obj_used)
2258 obj_wasted = obj_allocated - obj_used;
2259 obj_wasted /= class->objs_per_zspage;
2261 return obj_wasted * class->pages_per_zspage;
2264 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2266 struct zs_compact_control cc;
2267 struct zspage *src_zspage;
2268 struct zspage *dst_zspage = NULL;
2270 spin_lock(&class->lock);
2271 while ((src_zspage = isolate_zspage(class, true))) {
2273 if (!zs_can_compact(class))
2277 cc.s_page = get_first_page(src_zspage);
2279 while ((dst_zspage = isolate_zspage(class, false))) {
2280 cc.d_page = get_first_page(dst_zspage);
2282 * If there is no more space in dst_page, resched
2283 * and see if anyone had allocated another zspage.
2285 if (!migrate_zspage(pool, class, &cc))
2288 putback_zspage(class, dst_zspage);
2291 /* Stop if we couldn't find slot */
2292 if (dst_zspage == NULL)
2295 putback_zspage(class, dst_zspage);
2296 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2297 free_zspage(pool, class, src_zspage);
2298 pool->stats.pages_compacted += class->pages_per_zspage;
2300 spin_unlock(&class->lock);
2302 spin_lock(&class->lock);
2306 putback_zspage(class, src_zspage);
2308 spin_unlock(&class->lock);
2311 unsigned long zs_compact(struct zs_pool *pool)
2314 struct size_class *class;
2316 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2317 class = pool->size_class[i];
2320 if (class->index != i)
2322 __zs_compact(pool, class);
2325 return pool->stats.pages_compacted;
2327 EXPORT_SYMBOL_GPL(zs_compact);
2329 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2331 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2333 EXPORT_SYMBOL_GPL(zs_pool_stats);
2335 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2336 struct shrink_control *sc)
2338 unsigned long pages_freed;
2339 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2342 pages_freed = pool->stats.pages_compacted;
2344 * Compact classes and calculate compaction delta.
2345 * Can run concurrently with a manually triggered
2346 * (by user) compaction.
2348 pages_freed = zs_compact(pool) - pages_freed;
2350 return pages_freed ? pages_freed : SHRINK_STOP;
2353 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2354 struct shrink_control *sc)
2357 struct size_class *class;
2358 unsigned long pages_to_free = 0;
2359 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2362 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2363 class = pool->size_class[i];
2366 if (class->index != i)
2369 pages_to_free += zs_can_compact(class);
2372 return pages_to_free;
2375 static void zs_unregister_shrinker(struct zs_pool *pool)
2377 unregister_shrinker(&pool->shrinker);
2380 static int zs_register_shrinker(struct zs_pool *pool)
2382 pool->shrinker.scan_objects = zs_shrinker_scan;
2383 pool->shrinker.count_objects = zs_shrinker_count;
2384 pool->shrinker.batch = 0;
2385 pool->shrinker.seeks = DEFAULT_SEEKS;
2387 return register_shrinker(&pool->shrinker);
2391 * zs_create_pool - Creates an allocation pool to work from.
2392 * @name: pool name to be created
2394 * This function must be called before anything when using
2395 * the zsmalloc allocator.
2397 * On success, a pointer to the newly created pool is returned,
2400 struct zs_pool *zs_create_pool(const char *name)
2403 struct zs_pool *pool;
2404 struct size_class *prev_class = NULL;
2406 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2410 init_deferred_free(pool);
2412 pool->name = kstrdup(name, GFP_KERNEL);
2416 #ifdef CONFIG_COMPACTION
2417 init_waitqueue_head(&pool->migration_wait);
2420 if (create_cache(pool))
2424 * Iterate reversely, because, size of size_class that we want to use
2425 * for merging should be larger or equal to current size.
2427 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2429 int pages_per_zspage;
2430 int objs_per_zspage;
2431 struct size_class *class;
2434 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2435 if (size > ZS_MAX_ALLOC_SIZE)
2436 size = ZS_MAX_ALLOC_SIZE;
2437 pages_per_zspage = get_pages_per_zspage(size);
2438 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2441 * We iterate from biggest down to smallest classes,
2442 * so huge_class_size holds the size of the first huge
2443 * class. Any object bigger than or equal to that will
2444 * endup in the huge class.
2446 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2448 huge_class_size = size;
2450 * The object uses ZS_HANDLE_SIZE bytes to store the
2451 * handle. We need to subtract it, because zs_malloc()
2452 * unconditionally adds handle size before it performs
2453 * size class search - so object may be smaller than
2454 * huge class size, yet it still can end up in the huge
2455 * class because it grows by ZS_HANDLE_SIZE extra bytes
2456 * right before class lookup.
2458 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2462 * size_class is used for normal zsmalloc operation such
2463 * as alloc/free for that size. Although it is natural that we
2464 * have one size_class for each size, there is a chance that we
2465 * can get more memory utilization if we use one size_class for
2466 * many different sizes whose size_class have same
2467 * characteristics. So, we makes size_class point to
2468 * previous size_class if possible.
2471 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2472 pool->size_class[i] = prev_class;
2477 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2483 class->pages_per_zspage = pages_per_zspage;
2484 class->objs_per_zspage = objs_per_zspage;
2485 spin_lock_init(&class->lock);
2486 pool->size_class[i] = class;
2487 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2489 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2494 /* debug only, don't abort if it fails */
2495 zs_pool_stat_create(pool, name);
2497 if (zs_register_migration(pool))
2501 * Not critical since shrinker is only used to trigger internal
2502 * defragmentation of the pool which is pretty optional thing. If
2503 * registration fails we still can use the pool normally and user can
2504 * trigger compaction manually. Thus, ignore return code.
2506 zs_register_shrinker(pool);
2511 zs_destroy_pool(pool);
2514 EXPORT_SYMBOL_GPL(zs_create_pool);
2516 void zs_destroy_pool(struct zs_pool *pool)
2520 zs_unregister_shrinker(pool);
2521 zs_unregister_migration(pool);
2522 zs_pool_stat_destroy(pool);
2524 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2526 struct size_class *class = pool->size_class[i];
2531 if (class->index != i)
2534 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2535 if (!list_empty(&class->fullness_list[fg])) {
2536 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2543 destroy_cache(pool);
2547 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2549 static int __init zs_init(void)
2553 ret = zsmalloc_mount();
2557 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2558 zs_cpu_prepare, zs_cpu_dead);
2563 zpool_register_driver(&zs_zpool_driver);
2576 static void __exit zs_exit(void)
2579 zpool_unregister_driver(&zs_zpool_driver);
2582 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2587 module_init(zs_init);
2588 module_exit(zs_exit);
2590 MODULE_LICENSE("Dual BSD/GPL");
2591 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");