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
24 * Usage of struct page flags:
25 * PG_private: identifies the first component page
26 * PG_private2: identifies the last component page
27 * PG_owner_priv_1: indentifies 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/bitops.h>
37 #include <linux/errno.h>
38 #include <linux/highmem.h>
39 #include <linux/string.h>
40 #include <linux/slab.h>
41 #include <asm/tlbflush.h>
42 #include <asm/pgtable.h>
43 #include <linux/cpumask.h>
44 #include <linux/cpu.h>
45 #include <linux/vmalloc.h>
46 #include <linux/preempt.h>
47 #include <linux/spinlock.h>
48 #include <linux/types.h>
49 #include <linux/debugfs.h>
50 #include <linux/zsmalloc.h>
51 #include <linux/zpool.h>
52 #include <linux/mount.h>
53 #include <linux/migrate.h>
54 #include <linux/pagemap.h>
56 #define ZSPAGE_MAGIC 0x58
59 * This must be power of 2 and greater than of equal to sizeof(link_free).
60 * These two conditions ensure that any 'struct link_free' itself doesn't
61 * span more than 1 page which avoids complex case of mapping 2 pages simply
62 * to restore link_free pointer values.
67 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
68 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
70 #define ZS_MAX_ZSPAGE_ORDER 2
71 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
73 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
76 * Object location (<PFN>, <obj_idx>) is encoded as
77 * as single (unsigned long) handle value.
79 * Note that object index <obj_idx> starts from 0.
81 * This is made more complicated by various memory models and PAE.
84 #ifndef MAX_PHYSMEM_BITS
85 #ifdef CONFIG_HIGHMEM64G
86 #define MAX_PHYSMEM_BITS 36
87 #else /* !CONFIG_HIGHMEM64G */
89 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
92 #define MAX_PHYSMEM_BITS BITS_PER_LONG
95 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
98 * Memory for allocating for handle keeps object position by
99 * encoding <page, obj_idx> and the encoded value has a room
100 * in least bit(ie, look at obj_to_location).
101 * We use the bit to synchronize between object access by
102 * user and migration.
104 #define HANDLE_PIN_BIT 0
107 * Head in allocated object should have OBJ_ALLOCATED_TAG
108 * to identify the object was allocated or not.
109 * It's okay to add the status bit in the least bit because
110 * header keeps handle which is 4byte-aligned address so we
111 * have room for two bit at least.
113 #define OBJ_ALLOCATED_TAG 1
114 #define OBJ_TAG_BITS 1
115 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
116 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
118 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
119 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
120 #define ZS_MIN_ALLOC_SIZE \
121 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
122 /* each chunk includes extra space to keep handle */
123 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
126 * On systems with 4K page size, this gives 255 size classes! There is a
128 * - Large number of size classes is potentially wasteful as free page are
129 * spread across these classes
130 * - Small number of size classes causes large internal fragmentation
131 * - Probably its better to use specific size classes (empirically
132 * determined). NOTE: all those class sizes must be set as multiple of
133 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
135 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
138 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
141 * We do not maintain any list for completely empty or full pages
143 enum fullness_group {
161 struct zs_size_stat {
162 unsigned long objs[NR_ZS_STAT_TYPE];
165 #ifdef CONFIG_ZSMALLOC_STAT
166 static struct dentry *zs_stat_root;
169 #ifdef CONFIG_COMPACTION
170 static struct vfsmount *zsmalloc_mnt;
174 * number of size_classes
176 static int zs_size_classes;
179 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
181 * n = number of allocated objects
182 * N = total number of objects zspage can store
183 * f = fullness_threshold_frac
185 * Similarly, we assign zspage to:
186 * ZS_ALMOST_FULL when n > N / f
187 * ZS_EMPTY when n == 0
188 * ZS_FULL when n == N
190 * (see: fix_fullness_group())
192 static const int fullness_threshold_frac = 4;
196 struct list_head fullness_list[NR_ZS_FULLNESS];
198 * Size of objects stored in this class. Must be multiple
203 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
204 int pages_per_zspage;
207 struct zs_size_stat stats;
210 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
211 static void SetPageHugeObject(struct page *page)
213 SetPageOwnerPriv1(page);
216 static void ClearPageHugeObject(struct page *page)
218 ClearPageOwnerPriv1(page);
221 static int PageHugeObject(struct page *page)
223 return PageOwnerPriv1(page);
227 * Placed within free objects to form a singly linked list.
228 * For every zspage, zspage->freeobj gives head of this list.
230 * This must be power of 2 and less than or equal to ZS_ALIGN
236 * It's valid for non-allocated object
240 * Handle of allocated object.
242 unsigned long handle;
249 struct size_class **size_class;
250 struct kmem_cache *handle_cachep;
251 struct kmem_cache *zspage_cachep;
253 atomic_long_t pages_allocated;
255 struct zs_pool_stats stats;
257 /* Compact classes */
258 struct shrinker shrinker;
260 * To signify that register_shrinker() was successful
261 * and unregister_shrinker() will not Oops.
263 bool shrinker_enabled;
264 #ifdef CONFIG_ZSMALLOC_STAT
265 struct dentry *stat_dentry;
267 #ifdef CONFIG_COMPACTION
269 struct work_struct free_work;
274 * A zspage's class index and fullness group
275 * are encoded in its (first)page->mapping
277 #define FULLNESS_BITS 2
279 #define ISOLATED_BITS 3
280 #define MAGIC_VAL_BITS 8
284 unsigned int fullness:FULLNESS_BITS;
285 unsigned int class:CLASS_BITS;
286 unsigned int isolated:ISOLATED_BITS;
287 unsigned int magic:MAGIC_VAL_BITS;
290 unsigned int freeobj;
291 struct page *first_page;
292 struct list_head list; /* fullness list */
293 #ifdef CONFIG_COMPACTION
298 struct mapping_area {
299 #ifdef CONFIG_PGTABLE_MAPPING
300 struct vm_struct *vm; /* vm area for mapping object that span pages */
302 char *vm_buf; /* copy buffer for objects that span pages */
304 char *vm_addr; /* address of kmap_atomic()'ed pages */
305 enum zs_mapmode vm_mm; /* mapping mode */
308 #ifdef CONFIG_COMPACTION
309 static int zs_register_migration(struct zs_pool *pool);
310 static void zs_unregister_migration(struct zs_pool *pool);
311 static void migrate_lock_init(struct zspage *zspage);
312 static void migrate_read_lock(struct zspage *zspage);
313 static void migrate_read_unlock(struct zspage *zspage);
314 static void kick_deferred_free(struct zs_pool *pool);
315 static void init_deferred_free(struct zs_pool *pool);
316 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
318 static int zsmalloc_mount(void) { return 0; }
319 static void zsmalloc_unmount(void) {}
320 static int zs_register_migration(struct zs_pool *pool) { return 0; }
321 static void zs_unregister_migration(struct zs_pool *pool) {}
322 static void migrate_lock_init(struct zspage *zspage) {}
323 static void migrate_read_lock(struct zspage *zspage) {}
324 static void migrate_read_unlock(struct zspage *zspage) {}
325 static void kick_deferred_free(struct zs_pool *pool) {}
326 static void init_deferred_free(struct zs_pool *pool) {}
327 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
330 static int create_cache(struct zs_pool *pool)
332 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
334 if (!pool->handle_cachep)
337 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
339 if (!pool->zspage_cachep) {
340 kmem_cache_destroy(pool->handle_cachep);
341 pool->handle_cachep = NULL;
348 static void destroy_cache(struct zs_pool *pool)
350 kmem_cache_destroy(pool->handle_cachep);
351 kmem_cache_destroy(pool->zspage_cachep);
354 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
356 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
357 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
360 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
362 kmem_cache_free(pool->handle_cachep, (void *)handle);
365 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
367 return kmem_cache_alloc(pool->zspage_cachep,
368 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
371 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
373 kmem_cache_free(pool->zspage_cachep, zspage);
376 static void record_obj(unsigned long handle, unsigned long obj)
379 * lsb of @obj represents handle lock while other bits
380 * represent object value the handle is pointing so
381 * updating shouldn't do store tearing.
383 WRITE_ONCE(*(unsigned long *)handle, obj);
390 static void *zs_zpool_create(const char *name, gfp_t gfp,
391 const struct zpool_ops *zpool_ops,
395 * Ignore global gfp flags: zs_malloc() may be invoked from
396 * different contexts and its caller must provide a valid
399 return zs_create_pool(name);
402 static void zs_zpool_destroy(void *pool)
404 zs_destroy_pool(pool);
407 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
408 unsigned long *handle)
410 *handle = zs_malloc(pool, size, gfp);
411 return *handle ? 0 : -1;
413 static void zs_zpool_free(void *pool, unsigned long handle)
415 zs_free(pool, handle);
418 static int zs_zpool_shrink(void *pool, unsigned int pages,
419 unsigned int *reclaimed)
424 static void *zs_zpool_map(void *pool, unsigned long handle,
425 enum zpool_mapmode mm)
427 enum zs_mapmode zs_mm;
436 case ZPOOL_MM_RW: /* fallthru */
442 return zs_map_object(pool, handle, zs_mm);
444 static void zs_zpool_unmap(void *pool, unsigned long handle)
446 zs_unmap_object(pool, handle);
449 static u64 zs_zpool_total_size(void *pool)
451 return zs_get_total_pages(pool) << PAGE_SHIFT;
454 static struct zpool_driver zs_zpool_driver = {
456 .owner = THIS_MODULE,
457 .create = zs_zpool_create,
458 .destroy = zs_zpool_destroy,
459 .malloc = zs_zpool_malloc,
460 .free = zs_zpool_free,
461 .shrink = zs_zpool_shrink,
463 .unmap = zs_zpool_unmap,
464 .total_size = zs_zpool_total_size,
467 MODULE_ALIAS("zpool-zsmalloc");
468 #endif /* CONFIG_ZPOOL */
470 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
472 return pages_per_zspage * PAGE_SIZE / size;
475 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
476 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
478 static bool is_zspage_isolated(struct zspage *zspage)
480 return zspage->isolated;
483 static int is_first_page(struct page *page)
485 return PagePrivate(page);
488 /* Protected by class->lock */
489 static inline int get_zspage_inuse(struct zspage *zspage)
491 return zspage->inuse;
494 static inline void set_zspage_inuse(struct zspage *zspage, int val)
499 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
501 zspage->inuse += val;
504 static inline struct page *get_first_page(struct zspage *zspage)
506 struct page *first_page = zspage->first_page;
508 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
512 static inline int get_first_obj_offset(struct page *page)
517 static inline void set_first_obj_offset(struct page *page, int offset)
519 page->units = offset;
522 static inline unsigned int get_freeobj(struct zspage *zspage)
524 return zspage->freeobj;
527 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
529 zspage->freeobj = obj;
532 static void get_zspage_mapping(struct zspage *zspage,
533 unsigned int *class_idx,
534 enum fullness_group *fullness)
536 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
538 *fullness = zspage->fullness;
539 *class_idx = zspage->class;
542 static void set_zspage_mapping(struct zspage *zspage,
543 unsigned int class_idx,
544 enum fullness_group fullness)
546 zspage->class = class_idx;
547 zspage->fullness = fullness;
551 * zsmalloc divides the pool into various size classes where each
552 * class maintains a list of zspages where each zspage is divided
553 * into equal sized chunks. Each allocation falls into one of these
554 * classes depending on its size. This function returns index of the
555 * size class which has chunk size big enough to hold the give size.
557 static int get_size_class_index(int size)
561 if (likely(size > ZS_MIN_ALLOC_SIZE))
562 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
563 ZS_SIZE_CLASS_DELTA);
565 return min(zs_size_classes - 1, idx);
568 static inline void zs_stat_inc(struct size_class *class,
569 enum zs_stat_type type, unsigned long cnt)
571 class->stats.objs[type] += cnt;
574 static inline void zs_stat_dec(struct size_class *class,
575 enum zs_stat_type type, unsigned long cnt)
577 class->stats.objs[type] -= cnt;
580 static inline unsigned long zs_stat_get(struct size_class *class,
581 enum zs_stat_type type)
583 return class->stats.objs[type];
586 #ifdef CONFIG_ZSMALLOC_STAT
588 static void __init zs_stat_init(void)
590 if (!debugfs_initialized()) {
591 pr_warn("debugfs not available, stat dir not created\n");
595 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
597 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
600 static void __exit zs_stat_exit(void)
602 debugfs_remove_recursive(zs_stat_root);
605 static unsigned long zs_can_compact(struct size_class *class);
607 static int zs_stats_size_show(struct seq_file *s, void *v)
610 struct zs_pool *pool = s->private;
611 struct size_class *class;
613 unsigned long class_almost_full, class_almost_empty;
614 unsigned long obj_allocated, obj_used, pages_used, freeable;
615 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
616 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
617 unsigned long total_freeable = 0;
619 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
620 "class", "size", "almost_full", "almost_empty",
621 "obj_allocated", "obj_used", "pages_used",
622 "pages_per_zspage", "freeable");
624 for (i = 0; i < zs_size_classes; i++) {
625 class = pool->size_class[i];
627 if (class->index != i)
630 spin_lock(&class->lock);
631 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
632 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
633 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
634 obj_used = zs_stat_get(class, OBJ_USED);
635 freeable = zs_can_compact(class);
636 spin_unlock(&class->lock);
638 objs_per_zspage = class->objs_per_zspage;
639 pages_used = obj_allocated / objs_per_zspage *
640 class->pages_per_zspage;
642 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
643 " %10lu %10lu %16d %8lu\n",
644 i, class->size, class_almost_full, class_almost_empty,
645 obj_allocated, obj_used, pages_used,
646 class->pages_per_zspage, freeable);
648 total_class_almost_full += class_almost_full;
649 total_class_almost_empty += class_almost_empty;
650 total_objs += obj_allocated;
651 total_used_objs += obj_used;
652 total_pages += pages_used;
653 total_freeable += freeable;
657 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
658 "Total", "", total_class_almost_full,
659 total_class_almost_empty, total_objs,
660 total_used_objs, total_pages, "", total_freeable);
665 static int zs_stats_size_open(struct inode *inode, struct file *file)
667 return single_open(file, zs_stats_size_show, inode->i_private);
670 static const struct file_operations zs_stat_size_ops = {
671 .open = zs_stats_size_open,
674 .release = single_release,
677 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
679 struct dentry *entry;
682 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
686 entry = debugfs_create_dir(name, zs_stat_root);
688 pr_warn("debugfs dir <%s> creation failed\n", name);
691 pool->stat_dentry = entry;
693 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
694 pool->stat_dentry, pool, &zs_stat_size_ops);
696 pr_warn("%s: debugfs file entry <%s> creation failed\n",
698 debugfs_remove_recursive(pool->stat_dentry);
699 pool->stat_dentry = NULL;
703 static void zs_pool_stat_destroy(struct zs_pool *pool)
705 debugfs_remove_recursive(pool->stat_dentry);
708 #else /* CONFIG_ZSMALLOC_STAT */
709 static void __init zs_stat_init(void)
713 static void __exit zs_stat_exit(void)
717 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
721 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
728 * For each size class, zspages are divided into different groups
729 * depending on how "full" they are. This was done so that we could
730 * easily find empty or nearly empty zspages when we try to shrink
731 * the pool (not yet implemented). This function returns fullness
732 * status of the given page.
734 static enum fullness_group get_fullness_group(struct size_class *class,
735 struct zspage *zspage)
737 int inuse, objs_per_zspage;
738 enum fullness_group fg;
740 inuse = get_zspage_inuse(zspage);
741 objs_per_zspage = class->objs_per_zspage;
745 else if (inuse == objs_per_zspage)
747 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
748 fg = ZS_ALMOST_EMPTY;
756 * Each size class maintains various freelists and zspages are assigned
757 * to one of these freelists based on the number of live objects they
758 * have. This functions inserts the given zspage into the freelist
759 * identified by <class, fullness_group>.
761 static void insert_zspage(struct size_class *class,
762 struct zspage *zspage,
763 enum fullness_group fullness)
767 zs_stat_inc(class, fullness, 1);
768 head = list_first_entry_or_null(&class->fullness_list[fullness],
769 struct zspage, list);
771 * We want to see more ZS_FULL pages and less almost empty/full.
772 * Put pages with higher ->inuse first.
775 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
776 list_add(&zspage->list, &head->list);
780 list_add(&zspage->list, &class->fullness_list[fullness]);
784 * This function removes the given zspage from the freelist identified
785 * by <class, fullness_group>.
787 static void remove_zspage(struct size_class *class,
788 struct zspage *zspage,
789 enum fullness_group fullness)
791 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
792 VM_BUG_ON(is_zspage_isolated(zspage));
794 list_del_init(&zspage->list);
795 zs_stat_dec(class, fullness, 1);
799 * Each size class maintains zspages in different fullness groups depending
800 * on the number of live objects they contain. When allocating or freeing
801 * objects, the fullness status of the page can change, say, from ALMOST_FULL
802 * to ALMOST_EMPTY when freeing an object. This function checks if such
803 * a status change has occurred for the given page and accordingly moves the
804 * page from the freelist of the old fullness group to that of the new
807 static enum fullness_group fix_fullness_group(struct size_class *class,
808 struct zspage *zspage)
811 enum fullness_group currfg, newfg;
813 get_zspage_mapping(zspage, &class_idx, &currfg);
814 newfg = get_fullness_group(class, zspage);
818 if (!is_zspage_isolated(zspage)) {
819 remove_zspage(class, zspage, currfg);
820 insert_zspage(class, zspage, newfg);
823 set_zspage_mapping(zspage, class_idx, newfg);
830 * We have to decide on how many pages to link together
831 * to form a zspage for each size class. This is important
832 * to reduce wastage due to unusable space left at end of
833 * each zspage which is given as:
834 * wastage = Zp % class_size
835 * usage = Zp - wastage
836 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
838 * For example, for size class of 3/8 * PAGE_SIZE, we should
839 * link together 3 PAGE_SIZE sized pages to form a zspage
840 * since then we can perfectly fit in 8 such objects.
842 static int get_pages_per_zspage(int class_size)
844 int i, max_usedpc = 0;
845 /* zspage order which gives maximum used size per KB */
846 int max_usedpc_order = 1;
848 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
852 zspage_size = i * PAGE_SIZE;
853 waste = zspage_size % class_size;
854 usedpc = (zspage_size - waste) * 100 / zspage_size;
856 if (usedpc > max_usedpc) {
858 max_usedpc_order = i;
862 return max_usedpc_order;
865 static struct zspage *get_zspage(struct page *page)
867 struct zspage *zspage = (struct zspage *)page->private;
869 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
873 static struct page *get_next_page(struct page *page)
875 if (unlikely(PageHugeObject(page)))
878 return page->freelist;
882 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
883 * @page: page object resides in zspage
884 * @obj_idx: object index
886 static void obj_to_location(unsigned long obj, struct page **page,
887 unsigned int *obj_idx)
889 obj >>= OBJ_TAG_BITS;
890 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
891 *obj_idx = (obj & OBJ_INDEX_MASK);
895 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
896 * @page: page object resides in zspage
897 * @obj_idx: object index
899 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
903 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
904 obj |= obj_idx & OBJ_INDEX_MASK;
905 obj <<= OBJ_TAG_BITS;
910 static unsigned long handle_to_obj(unsigned long handle)
912 return *(unsigned long *)handle;
915 static unsigned long obj_to_head(struct page *page, void *obj)
917 if (unlikely(PageHugeObject(page))) {
918 VM_BUG_ON_PAGE(!is_first_page(page), page);
921 return *(unsigned long *)obj;
924 static inline int testpin_tag(unsigned long handle)
926 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
929 static inline int trypin_tag(unsigned long handle)
931 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
934 static void pin_tag(unsigned long handle)
936 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
939 static void unpin_tag(unsigned long handle)
941 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
944 static void reset_page(struct page *page)
946 __ClearPageMovable(page);
947 clear_bit(PG_private, &page->flags);
948 clear_bit(PG_private_2, &page->flags);
949 set_page_private(page, 0);
950 page_mapcount_reset(page);
951 ClearPageHugeObject(page);
952 page->freelist = NULL;
956 * To prevent zspage destroy during migration, zspage freeing should
957 * hold locks of all pages in the zspage.
959 void lock_zspage(struct zspage *zspage)
961 struct page *page = get_first_page(zspage);
965 } while ((page = get_next_page(page)) != NULL);
968 int trylock_zspage(struct zspage *zspage)
970 struct page *cursor, *fail;
972 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
973 get_next_page(cursor)) {
974 if (!trylock_page(cursor)) {
982 for (cursor = get_first_page(zspage); cursor != fail; cursor =
983 get_next_page(cursor))
989 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
990 struct zspage *zspage)
992 struct page *page, *next;
993 enum fullness_group fg;
994 unsigned int class_idx;
996 get_zspage_mapping(zspage, &class_idx, &fg);
998 assert_spin_locked(&class->lock);
1000 VM_BUG_ON(get_zspage_inuse(zspage));
1001 VM_BUG_ON(fg != ZS_EMPTY);
1003 next = page = get_first_page(zspage);
1005 VM_BUG_ON_PAGE(!PageLocked(page), page);
1006 next = get_next_page(page);
1009 dec_zone_page_state(page, NR_ZSPAGES);
1012 } while (page != NULL);
1014 cache_free_zspage(pool, zspage);
1016 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1017 atomic_long_sub(class->pages_per_zspage,
1018 &pool->pages_allocated);
1021 static void free_zspage(struct zs_pool *pool, struct size_class *class,
1022 struct zspage *zspage)
1024 VM_BUG_ON(get_zspage_inuse(zspage));
1025 VM_BUG_ON(list_empty(&zspage->list));
1027 if (!trylock_zspage(zspage)) {
1028 kick_deferred_free(pool);
1032 remove_zspage(class, zspage, ZS_EMPTY);
1033 __free_zspage(pool, class, zspage);
1036 /* Initialize a newly allocated zspage */
1037 static void init_zspage(struct size_class *class, struct zspage *zspage)
1039 unsigned int freeobj = 1;
1040 unsigned long off = 0;
1041 struct page *page = get_first_page(zspage);
1044 struct page *next_page;
1045 struct link_free *link;
1048 set_first_obj_offset(page, off);
1050 vaddr = kmap_atomic(page);
1051 link = (struct link_free *)vaddr + off / sizeof(*link);
1053 while ((off += class->size) < PAGE_SIZE) {
1054 link->next = freeobj++ << OBJ_TAG_BITS;
1055 link += class->size / sizeof(*link);
1059 * We now come to the last (full or partial) object on this
1060 * page, which must point to the first object on the next
1063 next_page = get_next_page(page);
1065 link->next = freeobj++ << OBJ_TAG_BITS;
1068 * Reset OBJ_TAG_BITS bit to last link to tell
1069 * whether it's allocated object or not.
1071 link->next = -1 << OBJ_TAG_BITS;
1073 kunmap_atomic(vaddr);
1078 set_freeobj(zspage, 0);
1081 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1082 struct page *pages[])
1086 struct page *prev_page = NULL;
1087 int nr_pages = class->pages_per_zspage;
1090 * Allocate individual pages and link them together as:
1091 * 1. all pages are linked together using page->freelist
1092 * 2. each sub-page point to zspage using page->private
1094 * we set PG_private to identify the first page (i.e. no other sub-page
1095 * has this flag set) and PG_private_2 to identify the last page.
1097 for (i = 0; i < nr_pages; i++) {
1099 set_page_private(page, (unsigned long)zspage);
1100 page->freelist = NULL;
1102 zspage->first_page = page;
1103 SetPagePrivate(page);
1104 if (unlikely(class->objs_per_zspage == 1 &&
1105 class->pages_per_zspage == 1))
1106 SetPageHugeObject(page);
1108 prev_page->freelist = page;
1110 if (i == nr_pages - 1)
1111 SetPagePrivate2(page);
1117 * Allocate a zspage for the given size class
1119 static struct zspage *alloc_zspage(struct zs_pool *pool,
1120 struct size_class *class,
1124 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1125 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1130 memset(zspage, 0, sizeof(struct zspage));
1131 zspage->magic = ZSPAGE_MAGIC;
1132 migrate_lock_init(zspage);
1134 for (i = 0; i < class->pages_per_zspage; i++) {
1137 page = alloc_page(gfp);
1140 dec_zone_page_state(pages[i], NR_ZSPAGES);
1141 __free_page(pages[i]);
1143 cache_free_zspage(pool, zspage);
1147 inc_zone_page_state(page, NR_ZSPAGES);
1151 create_page_chain(class, zspage, pages);
1152 init_zspage(class, zspage);
1157 static struct zspage *find_get_zspage(struct size_class *class)
1160 struct zspage *zspage;
1162 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1163 zspage = list_first_entry_or_null(&class->fullness_list[i],
1164 struct zspage, list);
1172 #ifdef CONFIG_PGTABLE_MAPPING
1173 static inline int __zs_cpu_up(struct mapping_area *area)
1176 * Make sure we don't leak memory if a cpu UP notification
1177 * and zs_init() race and both call zs_cpu_up() on the same cpu
1181 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1187 static inline void __zs_cpu_down(struct mapping_area *area)
1190 free_vm_area(area->vm);
1194 static inline void *__zs_map_object(struct mapping_area *area,
1195 struct page *pages[2], int off, int size)
1197 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1198 area->vm_addr = area->vm->addr;
1199 return area->vm_addr + off;
1202 static inline void __zs_unmap_object(struct mapping_area *area,
1203 struct page *pages[2], int off, int size)
1205 unsigned long addr = (unsigned long)area->vm_addr;
1207 unmap_kernel_range(addr, PAGE_SIZE * 2);
1210 #else /* CONFIG_PGTABLE_MAPPING */
1212 static inline int __zs_cpu_up(struct mapping_area *area)
1215 * Make sure we don't leak memory if a cpu UP notification
1216 * and zs_init() race and both call zs_cpu_up() on the same cpu
1220 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1226 static inline void __zs_cpu_down(struct mapping_area *area)
1228 kfree(area->vm_buf);
1229 area->vm_buf = NULL;
1232 static void *__zs_map_object(struct mapping_area *area,
1233 struct page *pages[2], int off, int size)
1237 char *buf = area->vm_buf;
1239 /* disable page faults to match kmap_atomic() return conditions */
1240 pagefault_disable();
1242 /* no read fastpath */
1243 if (area->vm_mm == ZS_MM_WO)
1246 sizes[0] = PAGE_SIZE - off;
1247 sizes[1] = size - sizes[0];
1249 /* copy object to per-cpu buffer */
1250 addr = kmap_atomic(pages[0]);
1251 memcpy(buf, addr + off, sizes[0]);
1252 kunmap_atomic(addr);
1253 addr = kmap_atomic(pages[1]);
1254 memcpy(buf + sizes[0], addr, sizes[1]);
1255 kunmap_atomic(addr);
1257 return area->vm_buf;
1260 static void __zs_unmap_object(struct mapping_area *area,
1261 struct page *pages[2], int off, int size)
1267 /* no write fastpath */
1268 if (area->vm_mm == ZS_MM_RO)
1272 buf = buf + ZS_HANDLE_SIZE;
1273 size -= ZS_HANDLE_SIZE;
1274 off += ZS_HANDLE_SIZE;
1276 sizes[0] = PAGE_SIZE - off;
1277 sizes[1] = size - sizes[0];
1279 /* copy per-cpu buffer to object */
1280 addr = kmap_atomic(pages[0]);
1281 memcpy(addr + off, buf, sizes[0]);
1282 kunmap_atomic(addr);
1283 addr = kmap_atomic(pages[1]);
1284 memcpy(addr, buf + sizes[0], sizes[1]);
1285 kunmap_atomic(addr);
1288 /* enable page faults to match kunmap_atomic() return conditions */
1292 #endif /* CONFIG_PGTABLE_MAPPING */
1294 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1297 int ret, cpu = (long)pcpu;
1298 struct mapping_area *area;
1301 case CPU_UP_PREPARE:
1302 area = &per_cpu(zs_map_area, cpu);
1303 ret = __zs_cpu_up(area);
1305 return notifier_from_errno(ret);
1308 case CPU_UP_CANCELED:
1309 area = &per_cpu(zs_map_area, cpu);
1310 __zs_cpu_down(area);
1317 static struct notifier_block zs_cpu_nb = {
1318 .notifier_call = zs_cpu_notifier
1321 static int zs_register_cpu_notifier(void)
1323 int cpu, uninitialized_var(ret);
1325 cpu_notifier_register_begin();
1327 __register_cpu_notifier(&zs_cpu_nb);
1328 for_each_online_cpu(cpu) {
1329 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1330 if (notifier_to_errno(ret))
1334 cpu_notifier_register_done();
1335 return notifier_to_errno(ret);
1338 static void zs_unregister_cpu_notifier(void)
1342 cpu_notifier_register_begin();
1344 for_each_online_cpu(cpu)
1345 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1346 __unregister_cpu_notifier(&zs_cpu_nb);
1348 cpu_notifier_register_done();
1351 static void init_zs_size_classes(void)
1355 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1356 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1359 zs_size_classes = nr;
1362 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1364 if (prev->pages_per_zspage != pages_per_zspage)
1367 if (prev->objs_per_zspage
1368 != get_maxobj_per_zspage(size, pages_per_zspage))
1374 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1376 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1379 unsigned long zs_get_total_pages(struct zs_pool *pool)
1381 return atomic_long_read(&pool->pages_allocated);
1383 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1386 * zs_map_object - get address of allocated object from handle.
1387 * @pool: pool from which the object was allocated
1388 * @handle: handle returned from zs_malloc
1390 * Before using an object allocated from zs_malloc, it must be mapped using
1391 * this function. When done with the object, it must be unmapped using
1394 * Only one object can be mapped per cpu at a time. There is no protection
1395 * against nested mappings.
1397 * This function returns with preemption and page faults disabled.
1399 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1402 struct zspage *zspage;
1404 unsigned long obj, off;
1405 unsigned int obj_idx;
1407 unsigned int class_idx;
1408 enum fullness_group fg;
1409 struct size_class *class;
1410 struct mapping_area *area;
1411 struct page *pages[2];
1415 * Because we use per-cpu mapping areas shared among the
1416 * pools/users, we can't allow mapping in interrupt context
1417 * because it can corrupt another users mappings.
1419 WARN_ON_ONCE(in_interrupt());
1421 /* From now on, migration cannot move the object */
1424 obj = handle_to_obj(handle);
1425 obj_to_location(obj, &page, &obj_idx);
1426 zspage = get_zspage(page);
1428 /* migration cannot move any subpage in this zspage */
1429 migrate_read_lock(zspage);
1431 get_zspage_mapping(zspage, &class_idx, &fg);
1432 class = pool->size_class[class_idx];
1433 off = (class->size * obj_idx) & ~PAGE_MASK;
1435 area = &get_cpu_var(zs_map_area);
1437 if (off + class->size <= PAGE_SIZE) {
1438 /* this object is contained entirely within a page */
1439 area->vm_addr = kmap_atomic(page);
1440 ret = area->vm_addr + off;
1444 /* this object spans two pages */
1446 pages[1] = get_next_page(page);
1449 ret = __zs_map_object(area, pages, off, class->size);
1451 if (likely(!PageHugeObject(page)))
1452 ret += ZS_HANDLE_SIZE;
1456 EXPORT_SYMBOL_GPL(zs_map_object);
1458 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1460 struct zspage *zspage;
1462 unsigned long obj, off;
1463 unsigned int obj_idx;
1465 unsigned int class_idx;
1466 enum fullness_group fg;
1467 struct size_class *class;
1468 struct mapping_area *area;
1470 obj = handle_to_obj(handle);
1471 obj_to_location(obj, &page, &obj_idx);
1472 zspage = get_zspage(page);
1473 get_zspage_mapping(zspage, &class_idx, &fg);
1474 class = pool->size_class[class_idx];
1475 off = (class->size * obj_idx) & ~PAGE_MASK;
1477 area = this_cpu_ptr(&zs_map_area);
1478 if (off + class->size <= PAGE_SIZE)
1479 kunmap_atomic(area->vm_addr);
1481 struct page *pages[2];
1484 pages[1] = get_next_page(page);
1487 __zs_unmap_object(area, pages, off, class->size);
1489 put_cpu_var(zs_map_area);
1491 migrate_read_unlock(zspage);
1494 EXPORT_SYMBOL_GPL(zs_unmap_object);
1496 static unsigned long obj_malloc(struct size_class *class,
1497 struct zspage *zspage, unsigned long handle)
1499 int i, nr_page, offset;
1501 struct link_free *link;
1503 struct page *m_page;
1504 unsigned long m_offset;
1507 handle |= OBJ_ALLOCATED_TAG;
1508 obj = get_freeobj(zspage);
1510 offset = obj * class->size;
1511 nr_page = offset >> PAGE_SHIFT;
1512 m_offset = offset & ~PAGE_MASK;
1513 m_page = get_first_page(zspage);
1515 for (i = 0; i < nr_page; i++)
1516 m_page = get_next_page(m_page);
1518 vaddr = kmap_atomic(m_page);
1519 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1520 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1521 if (likely(!PageHugeObject(m_page)))
1522 /* record handle in the header of allocated chunk */
1523 link->handle = handle;
1525 /* record handle to page->index */
1526 zspage->first_page->index = handle;
1528 kunmap_atomic(vaddr);
1529 mod_zspage_inuse(zspage, 1);
1530 zs_stat_inc(class, OBJ_USED, 1);
1532 obj = location_to_obj(m_page, obj);
1539 * zs_malloc - Allocate block of given size from pool.
1540 * @pool: pool to allocate from
1541 * @size: size of block to allocate
1543 * On success, handle to the allocated object is returned,
1545 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1547 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1549 unsigned long handle, obj;
1550 struct size_class *class;
1551 enum fullness_group newfg;
1552 struct zspage *zspage;
1554 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1557 handle = cache_alloc_handle(pool, gfp);
1561 /* extra space in chunk to keep the handle */
1562 size += ZS_HANDLE_SIZE;
1563 class = pool->size_class[get_size_class_index(size)];
1565 spin_lock(&class->lock);
1566 zspage = find_get_zspage(class);
1567 if (likely(zspage)) {
1568 obj = obj_malloc(class, zspage, handle);
1569 /* Now move the zspage to another fullness group, if required */
1570 fix_fullness_group(class, zspage);
1571 record_obj(handle, obj);
1572 spin_unlock(&class->lock);
1577 spin_unlock(&class->lock);
1579 zspage = alloc_zspage(pool, class, gfp);
1581 cache_free_handle(pool, handle);
1585 spin_lock(&class->lock);
1586 obj = obj_malloc(class, zspage, handle);
1587 newfg = get_fullness_group(class, zspage);
1588 insert_zspage(class, zspage, newfg);
1589 set_zspage_mapping(zspage, class->index, newfg);
1590 record_obj(handle, obj);
1591 atomic_long_add(class->pages_per_zspage,
1592 &pool->pages_allocated);
1593 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1595 /* We completely set up zspage so mark them as movable */
1596 SetZsPageMovable(pool, zspage);
1597 spin_unlock(&class->lock);
1601 EXPORT_SYMBOL_GPL(zs_malloc);
1603 static void obj_free(struct size_class *class, unsigned long obj)
1605 struct link_free *link;
1606 struct zspage *zspage;
1607 struct page *f_page;
1608 unsigned long f_offset;
1609 unsigned int f_objidx;
1612 obj &= ~OBJ_ALLOCATED_TAG;
1613 obj_to_location(obj, &f_page, &f_objidx);
1614 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1615 zspage = get_zspage(f_page);
1617 vaddr = kmap_atomic(f_page);
1619 /* Insert this object in containing zspage's freelist */
1620 link = (struct link_free *)(vaddr + f_offset);
1621 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1622 kunmap_atomic(vaddr);
1623 set_freeobj(zspage, f_objidx);
1624 mod_zspage_inuse(zspage, -1);
1625 zs_stat_dec(class, OBJ_USED, 1);
1628 void zs_free(struct zs_pool *pool, unsigned long handle)
1630 struct zspage *zspage;
1631 struct page *f_page;
1633 unsigned int f_objidx;
1635 struct size_class *class;
1636 enum fullness_group fullness;
1639 if (unlikely(!handle))
1643 obj = handle_to_obj(handle);
1644 obj_to_location(obj, &f_page, &f_objidx);
1645 zspage = get_zspage(f_page);
1647 migrate_read_lock(zspage);
1649 get_zspage_mapping(zspage, &class_idx, &fullness);
1650 class = pool->size_class[class_idx];
1652 spin_lock(&class->lock);
1653 obj_free(class, obj);
1654 fullness = fix_fullness_group(class, zspage);
1655 if (fullness != ZS_EMPTY) {
1656 migrate_read_unlock(zspage);
1660 isolated = is_zspage_isolated(zspage);
1661 migrate_read_unlock(zspage);
1662 /* If zspage is isolated, zs_page_putback will free the zspage */
1663 if (likely(!isolated))
1664 free_zspage(pool, class, zspage);
1667 spin_unlock(&class->lock);
1669 cache_free_handle(pool, handle);
1671 EXPORT_SYMBOL_GPL(zs_free);
1673 static void zs_object_copy(struct size_class *class, unsigned long dst,
1676 struct page *s_page, *d_page;
1677 unsigned int s_objidx, d_objidx;
1678 unsigned long s_off, d_off;
1679 void *s_addr, *d_addr;
1680 int s_size, d_size, size;
1683 s_size = d_size = class->size;
1685 obj_to_location(src, &s_page, &s_objidx);
1686 obj_to_location(dst, &d_page, &d_objidx);
1688 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1689 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1691 if (s_off + class->size > PAGE_SIZE)
1692 s_size = PAGE_SIZE - s_off;
1694 if (d_off + class->size > PAGE_SIZE)
1695 d_size = PAGE_SIZE - d_off;
1697 s_addr = kmap_atomic(s_page);
1698 d_addr = kmap_atomic(d_page);
1701 size = min(s_size, d_size);
1702 memcpy(d_addr + d_off, s_addr + s_off, size);
1705 if (written == class->size)
1713 if (s_off >= PAGE_SIZE) {
1714 kunmap_atomic(d_addr);
1715 kunmap_atomic(s_addr);
1716 s_page = get_next_page(s_page);
1717 s_addr = kmap_atomic(s_page);
1718 d_addr = kmap_atomic(d_page);
1719 s_size = class->size - written;
1723 if (d_off >= PAGE_SIZE) {
1724 kunmap_atomic(d_addr);
1725 d_page = get_next_page(d_page);
1726 d_addr = kmap_atomic(d_page);
1727 d_size = class->size - written;
1732 kunmap_atomic(d_addr);
1733 kunmap_atomic(s_addr);
1737 * Find alloced object in zspage from index object and
1740 static unsigned long find_alloced_obj(struct size_class *class,
1741 struct page *page, int *obj_idx)
1745 int index = *obj_idx;
1746 unsigned long handle = 0;
1747 void *addr = kmap_atomic(page);
1749 offset = get_first_obj_offset(page);
1750 offset += class->size * index;
1752 while (offset < PAGE_SIZE) {
1753 head = obj_to_head(page, addr + offset);
1754 if (head & OBJ_ALLOCATED_TAG) {
1755 handle = head & ~OBJ_ALLOCATED_TAG;
1756 if (trypin_tag(handle))
1761 offset += class->size;
1765 kunmap_atomic(addr);
1772 struct zs_compact_control {
1773 /* Source spage for migration which could be a subpage of zspage */
1774 struct page *s_page;
1775 /* Destination page for migration which should be a first page
1777 struct page *d_page;
1778 /* Starting object index within @s_page which used for live object
1779 * in the subpage. */
1783 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1784 struct zs_compact_control *cc)
1786 unsigned long used_obj, free_obj;
1787 unsigned long handle;
1788 struct page *s_page = cc->s_page;
1789 struct page *d_page = cc->d_page;
1790 int obj_idx = cc->obj_idx;
1794 handle = find_alloced_obj(class, s_page, &obj_idx);
1796 s_page = get_next_page(s_page);
1803 /* Stop if there is no more space */
1804 if (zspage_full(class, get_zspage(d_page))) {
1810 used_obj = handle_to_obj(handle);
1811 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1812 zs_object_copy(class, free_obj, used_obj);
1815 * record_obj updates handle's value to free_obj and it will
1816 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1817 * breaks synchronization using pin_tag(e,g, zs_free) so
1818 * let's keep the lock bit.
1820 free_obj |= BIT(HANDLE_PIN_BIT);
1821 record_obj(handle, free_obj);
1823 obj_free(class, used_obj);
1826 /* Remember last position in this iteration */
1827 cc->s_page = s_page;
1828 cc->obj_idx = obj_idx;
1833 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1836 struct zspage *zspage;
1837 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1840 fg[0] = ZS_ALMOST_FULL;
1841 fg[1] = ZS_ALMOST_EMPTY;
1844 for (i = 0; i < 2; i++) {
1845 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1846 struct zspage, list);
1848 VM_BUG_ON(is_zspage_isolated(zspage));
1849 remove_zspage(class, zspage, fg[i]);
1858 * putback_zspage - add @zspage into right class's fullness list
1859 * @class: destination class
1860 * @zspage: target page
1862 * Return @zspage's fullness_group
1864 static enum fullness_group putback_zspage(struct size_class *class,
1865 struct zspage *zspage)
1867 enum fullness_group fullness;
1869 VM_BUG_ON(is_zspage_isolated(zspage));
1871 fullness = get_fullness_group(class, zspage);
1872 insert_zspage(class, zspage, fullness);
1873 set_zspage_mapping(zspage, class->index, fullness);
1878 #ifdef CONFIG_COMPACTION
1879 static struct dentry *zs_mount(struct file_system_type *fs_type,
1880 int flags, const char *dev_name, void *data)
1882 static const struct dentry_operations ops = {
1883 .d_dname = simple_dname,
1886 return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1889 static struct file_system_type zsmalloc_fs = {
1892 .kill_sb = kill_anon_super,
1895 static int zsmalloc_mount(void)
1899 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1900 if (IS_ERR(zsmalloc_mnt))
1901 ret = PTR_ERR(zsmalloc_mnt);
1906 static void zsmalloc_unmount(void)
1908 kern_unmount(zsmalloc_mnt);
1911 static void migrate_lock_init(struct zspage *zspage)
1913 rwlock_init(&zspage->lock);
1916 static void migrate_read_lock(struct zspage *zspage)
1918 read_lock(&zspage->lock);
1921 static void migrate_read_unlock(struct zspage *zspage)
1923 read_unlock(&zspage->lock);
1926 static void migrate_write_lock(struct zspage *zspage)
1928 write_lock(&zspage->lock);
1931 static void migrate_write_unlock(struct zspage *zspage)
1933 write_unlock(&zspage->lock);
1936 /* Number of isolated subpage for *page migration* in this zspage */
1937 static void inc_zspage_isolation(struct zspage *zspage)
1942 static void dec_zspage_isolation(struct zspage *zspage)
1947 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1948 struct page *newpage, struct page *oldpage)
1951 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1954 page = get_first_page(zspage);
1956 if (page == oldpage)
1957 pages[idx] = newpage;
1961 } while ((page = get_next_page(page)) != NULL);
1963 create_page_chain(class, zspage, pages);
1964 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1965 if (unlikely(PageHugeObject(oldpage)))
1966 newpage->index = oldpage->index;
1967 __SetPageMovable(newpage, page_mapping(oldpage));
1970 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1972 struct zs_pool *pool;
1973 struct size_class *class;
1975 enum fullness_group fullness;
1976 struct zspage *zspage;
1977 struct address_space *mapping;
1980 * Page is locked so zspage couldn't be destroyed. For detail, look at
1981 * lock_zspage in free_zspage.
1983 VM_BUG_ON_PAGE(!PageMovable(page), page);
1984 VM_BUG_ON_PAGE(PageIsolated(page), page);
1986 zspage = get_zspage(page);
1989 * Without class lock, fullness could be stale while class_idx is okay
1990 * because class_idx is constant unless page is freed so we should get
1991 * fullness again under class lock.
1993 get_zspage_mapping(zspage, &class_idx, &fullness);
1994 mapping = page_mapping(page);
1995 pool = mapping->private_data;
1996 class = pool->size_class[class_idx];
1998 spin_lock(&class->lock);
1999 if (get_zspage_inuse(zspage) == 0) {
2000 spin_unlock(&class->lock);
2004 /* zspage is isolated for object migration */
2005 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
2006 spin_unlock(&class->lock);
2011 * If this is first time isolation for the zspage, isolate zspage from
2012 * size_class to prevent further object allocation from the zspage.
2014 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
2015 get_zspage_mapping(zspage, &class_idx, &fullness);
2016 remove_zspage(class, zspage, fullness);
2019 inc_zspage_isolation(zspage);
2020 spin_unlock(&class->lock);
2025 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
2026 struct page *page, enum migrate_mode mode)
2028 struct zs_pool *pool;
2029 struct size_class *class;
2031 enum fullness_group fullness;
2032 struct zspage *zspage;
2034 void *s_addr, *d_addr, *addr;
2036 unsigned long handle, head;
2037 unsigned long old_obj, new_obj;
2038 unsigned int obj_idx;
2041 VM_BUG_ON_PAGE(!PageMovable(page), page);
2042 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2044 zspage = get_zspage(page);
2046 /* Concurrent compactor cannot migrate any subpage in zspage */
2047 migrate_write_lock(zspage);
2048 get_zspage_mapping(zspage, &class_idx, &fullness);
2049 pool = mapping->private_data;
2050 class = pool->size_class[class_idx];
2051 offset = get_first_obj_offset(page);
2053 spin_lock(&class->lock);
2054 if (!get_zspage_inuse(zspage)) {
2060 s_addr = kmap_atomic(page);
2061 while (pos < PAGE_SIZE) {
2062 head = obj_to_head(page, s_addr + pos);
2063 if (head & OBJ_ALLOCATED_TAG) {
2064 handle = head & ~OBJ_ALLOCATED_TAG;
2065 if (!trypin_tag(handle))
2072 * Here, any user cannot access all objects in the zspage so let's move.
2074 d_addr = kmap_atomic(newpage);
2075 memcpy(d_addr, s_addr, PAGE_SIZE);
2076 kunmap_atomic(d_addr);
2078 for (addr = s_addr + offset; addr < s_addr + pos;
2079 addr += class->size) {
2080 head = obj_to_head(page, addr);
2081 if (head & OBJ_ALLOCATED_TAG) {
2082 handle = head & ~OBJ_ALLOCATED_TAG;
2083 if (!testpin_tag(handle))
2086 old_obj = handle_to_obj(handle);
2087 obj_to_location(old_obj, &dummy, &obj_idx);
2088 new_obj = (unsigned long)location_to_obj(newpage,
2090 new_obj |= BIT(HANDLE_PIN_BIT);
2091 record_obj(handle, new_obj);
2095 replace_sub_page(class, zspage, newpage, page);
2098 dec_zspage_isolation(zspage);
2101 * Page migration is done so let's putback isolated zspage to
2102 * the list if @page is final isolated subpage in the zspage.
2104 if (!is_zspage_isolated(zspage))
2105 putback_zspage(class, zspage);
2111 ret = MIGRATEPAGE_SUCCESS;
2113 for (addr = s_addr + offset; addr < s_addr + pos;
2114 addr += class->size) {
2115 head = obj_to_head(page, addr);
2116 if (head & OBJ_ALLOCATED_TAG) {
2117 handle = head & ~OBJ_ALLOCATED_TAG;
2118 if (!testpin_tag(handle))
2123 kunmap_atomic(s_addr);
2125 spin_unlock(&class->lock);
2126 migrate_write_unlock(zspage);
2131 void zs_page_putback(struct page *page)
2133 struct zs_pool *pool;
2134 struct size_class *class;
2136 enum fullness_group fg;
2137 struct address_space *mapping;
2138 struct zspage *zspage;
2140 VM_BUG_ON_PAGE(!PageMovable(page), page);
2141 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2143 zspage = get_zspage(page);
2144 get_zspage_mapping(zspage, &class_idx, &fg);
2145 mapping = page_mapping(page);
2146 pool = mapping->private_data;
2147 class = pool->size_class[class_idx];
2149 spin_lock(&class->lock);
2150 dec_zspage_isolation(zspage);
2151 if (!is_zspage_isolated(zspage)) {
2152 fg = putback_zspage(class, zspage);
2154 * Due to page_lock, we cannot free zspage immediately
2158 schedule_work(&pool->free_work);
2160 spin_unlock(&class->lock);
2163 const struct address_space_operations zsmalloc_aops = {
2164 .isolate_page = zs_page_isolate,
2165 .migratepage = zs_page_migrate,
2166 .putback_page = zs_page_putback,
2169 static int zs_register_migration(struct zs_pool *pool)
2171 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2172 if (IS_ERR(pool->inode)) {
2177 pool->inode->i_mapping->private_data = pool;
2178 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2182 static void zs_unregister_migration(struct zs_pool *pool)
2184 flush_work(&pool->free_work);
2190 * Caller should hold page_lock of all pages in the zspage
2191 * In here, we cannot use zspage meta data.
2193 static void async_free_zspage(struct work_struct *work)
2196 struct size_class *class;
2197 unsigned int class_idx;
2198 enum fullness_group fullness;
2199 struct zspage *zspage, *tmp;
2200 LIST_HEAD(free_pages);
2201 struct zs_pool *pool = container_of(work, struct zs_pool,
2204 for (i = 0; i < zs_size_classes; i++) {
2205 class = pool->size_class[i];
2206 if (class->index != i)
2209 spin_lock(&class->lock);
2210 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2211 spin_unlock(&class->lock);
2215 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2216 list_del(&zspage->list);
2217 lock_zspage(zspage);
2219 get_zspage_mapping(zspage, &class_idx, &fullness);
2220 VM_BUG_ON(fullness != ZS_EMPTY);
2221 class = pool->size_class[class_idx];
2222 spin_lock(&class->lock);
2223 __free_zspage(pool, pool->size_class[class_idx], zspage);
2224 spin_unlock(&class->lock);
2228 static void kick_deferred_free(struct zs_pool *pool)
2230 schedule_work(&pool->free_work);
2233 static void init_deferred_free(struct zs_pool *pool)
2235 INIT_WORK(&pool->free_work, async_free_zspage);
2238 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2240 struct page *page = get_first_page(zspage);
2243 WARN_ON(!trylock_page(page));
2244 __SetPageMovable(page, pool->inode->i_mapping);
2246 } while ((page = get_next_page(page)) != NULL);
2252 * Based on the number of unused allocated objects calculate
2253 * and return the number of pages that we can free.
2255 static unsigned long zs_can_compact(struct size_class *class)
2257 unsigned long obj_wasted;
2258 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2259 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2261 if (obj_allocated <= obj_used)
2264 obj_wasted = obj_allocated - obj_used;
2265 obj_wasted /= class->objs_per_zspage;
2267 return obj_wasted * class->pages_per_zspage;
2270 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2272 struct zs_compact_control cc;
2273 struct zspage *src_zspage;
2274 struct zspage *dst_zspage = NULL;
2276 spin_lock(&class->lock);
2277 while ((src_zspage = isolate_zspage(class, true))) {
2279 if (!zs_can_compact(class))
2283 cc.s_page = get_first_page(src_zspage);
2285 while ((dst_zspage = isolate_zspage(class, false))) {
2286 cc.d_page = get_first_page(dst_zspage);
2288 * If there is no more space in dst_page, resched
2289 * and see if anyone had allocated another zspage.
2291 if (!migrate_zspage(pool, class, &cc))
2294 putback_zspage(class, dst_zspage);
2297 /* Stop if we couldn't find slot */
2298 if (dst_zspage == NULL)
2301 putback_zspage(class, dst_zspage);
2302 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2303 free_zspage(pool, class, src_zspage);
2304 pool->stats.pages_compacted += class->pages_per_zspage;
2306 spin_unlock(&class->lock);
2308 spin_lock(&class->lock);
2312 putback_zspage(class, src_zspage);
2314 spin_unlock(&class->lock);
2317 unsigned long zs_compact(struct zs_pool *pool)
2320 struct size_class *class;
2322 for (i = zs_size_classes - 1; i >= 0; i--) {
2323 class = pool->size_class[i];
2326 if (class->index != i)
2328 __zs_compact(pool, class);
2331 return pool->stats.pages_compacted;
2333 EXPORT_SYMBOL_GPL(zs_compact);
2335 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2337 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2339 EXPORT_SYMBOL_GPL(zs_pool_stats);
2341 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2342 struct shrink_control *sc)
2344 unsigned long pages_freed;
2345 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2348 pages_freed = pool->stats.pages_compacted;
2350 * Compact classes and calculate compaction delta.
2351 * Can run concurrently with a manually triggered
2352 * (by user) compaction.
2354 pages_freed = zs_compact(pool) - pages_freed;
2356 return pages_freed ? pages_freed : SHRINK_STOP;
2359 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2360 struct shrink_control *sc)
2363 struct size_class *class;
2364 unsigned long pages_to_free = 0;
2365 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2368 for (i = zs_size_classes - 1; i >= 0; i--) {
2369 class = pool->size_class[i];
2372 if (class->index != i)
2375 pages_to_free += zs_can_compact(class);
2378 return pages_to_free;
2381 static void zs_unregister_shrinker(struct zs_pool *pool)
2383 if (pool->shrinker_enabled) {
2384 unregister_shrinker(&pool->shrinker);
2385 pool->shrinker_enabled = false;
2389 static int zs_register_shrinker(struct zs_pool *pool)
2391 pool->shrinker.scan_objects = zs_shrinker_scan;
2392 pool->shrinker.count_objects = zs_shrinker_count;
2393 pool->shrinker.batch = 0;
2394 pool->shrinker.seeks = DEFAULT_SEEKS;
2396 return register_shrinker(&pool->shrinker);
2400 * zs_create_pool - Creates an allocation pool to work from.
2401 * @flags: allocation flags used to allocate pool metadata
2403 * This function must be called before anything when using
2404 * the zsmalloc allocator.
2406 * On success, a pointer to the newly created pool is returned,
2409 struct zs_pool *zs_create_pool(const char *name)
2412 struct zs_pool *pool;
2413 struct size_class *prev_class = NULL;
2415 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2419 init_deferred_free(pool);
2420 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
2422 if (!pool->size_class) {
2427 pool->name = kstrdup(name, GFP_KERNEL);
2431 if (create_cache(pool))
2435 * Iterate reversly, because, size of size_class that we want to use
2436 * for merging should be larger or equal to current size.
2438 for (i = zs_size_classes - 1; i >= 0; i--) {
2440 int pages_per_zspage;
2441 struct size_class *class;
2444 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2445 if (size > ZS_MAX_ALLOC_SIZE)
2446 size = ZS_MAX_ALLOC_SIZE;
2447 pages_per_zspage = get_pages_per_zspage(size);
2450 * size_class is used for normal zsmalloc operation such
2451 * as alloc/free for that size. Although it is natural that we
2452 * have one size_class for each size, there is a chance that we
2453 * can get more memory utilization if we use one size_class for
2454 * many different sizes whose size_class have same
2455 * characteristics. So, we makes size_class point to
2456 * previous size_class if possible.
2459 if (can_merge(prev_class, size, pages_per_zspage)) {
2460 pool->size_class[i] = prev_class;
2465 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2471 class->pages_per_zspage = pages_per_zspage;
2472 class->objs_per_zspage = get_maxobj_per_zspage(class->size,
2473 class->pages_per_zspage);
2474 spin_lock_init(&class->lock);
2475 pool->size_class[i] = class;
2476 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2478 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2483 /* debug only, don't abort if it fails */
2484 zs_pool_stat_create(pool, name);
2486 if (zs_register_migration(pool))
2490 * Not critical, we still can use the pool
2491 * and user can trigger compaction manually.
2493 if (zs_register_shrinker(pool) == 0)
2494 pool->shrinker_enabled = true;
2498 zs_destroy_pool(pool);
2501 EXPORT_SYMBOL_GPL(zs_create_pool);
2503 void zs_destroy_pool(struct zs_pool *pool)
2507 zs_unregister_shrinker(pool);
2508 zs_unregister_migration(pool);
2509 zs_pool_stat_destroy(pool);
2511 for (i = 0; i < zs_size_classes; i++) {
2513 struct size_class *class = pool->size_class[i];
2518 if (class->index != i)
2521 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2522 if (!list_empty(&class->fullness_list[fg])) {
2523 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2530 destroy_cache(pool);
2531 kfree(pool->size_class);
2535 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2537 static int __init zs_init(void)
2541 ret = zsmalloc_mount();
2545 ret = zs_register_cpu_notifier();
2550 init_zs_size_classes();
2553 zpool_register_driver(&zs_zpool_driver);
2561 zs_unregister_cpu_notifier();
2567 static void __exit zs_exit(void)
2570 zpool_unregister_driver(&zs_zpool_driver);
2573 zs_unregister_cpu_notifier();
2578 module_init(zs_init);
2579 module_exit(zs_exit);
2581 MODULE_LICENSE("Dual BSD/GPL");
2582 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");