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->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->page_type: 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
40 #include <linux/module.h>
41 #include <linux/kernel.h>
42 #include <linux/sched.h>
43 #include <linux/bitops.h>
44 #include <linux/errno.h>
45 #include <linux/highmem.h>
46 #include <linux/string.h>
47 #include <linux/slab.h>
48 #include <linux/pgtable.h>
49 #include <asm/tlbflush.h>
50 #include <linux/cpumask.h>
51 #include <linux/cpu.h>
52 #include <linux/vmalloc.h>
53 #include <linux/preempt.h>
54 #include <linux/spinlock.h>
55 #include <linux/shrinker.h>
56 #include <linux/types.h>
57 #include <linux/debugfs.h>
58 #include <linux/zsmalloc.h>
59 #include <linux/zpool.h>
60 #include <linux/migrate.h>
61 #include <linux/wait.h>
62 #include <linux/pagemap.h>
64 #include <linux/local_lock.h>
66 #define ZSPAGE_MAGIC 0x58
69 * This must be power of 2 and greater than or equal to sizeof(link_free).
70 * These two conditions ensure that any 'struct link_free' itself doesn't
71 * span more than 1 page which avoids complex case of mapping 2 pages simply
72 * to restore link_free pointer values.
76 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
79 * Object location (<PFN>, <obj_idx>) is encoded as
80 * a single (unsigned long) handle value.
82 * Note that object index <obj_idx> starts from 0.
84 * This is made more complicated by various memory models and PAE.
87 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
88 #ifdef MAX_PHYSMEM_BITS
89 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
92 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
95 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
99 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
102 * Head in allocated object should have OBJ_ALLOCATED_TAG
103 * to identify the object was allocated or not.
104 * It's okay to add the status bit in the least bit because
105 * header keeps handle which is 4byte-aligned address so we
106 * have room for two bit at least.
108 #define OBJ_ALLOCATED_TAG 1
110 #define OBJ_TAG_BITS 1
111 #define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
113 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
114 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
117 #define FULLNESS_BITS 4
119 #define ISOLATED_BITS 5
120 #define MAGIC_VAL_BITS 8
122 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
124 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(CONFIG_ZSMALLOC_CHAIN_SIZE, UL))
126 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
127 #define ZS_MIN_ALLOC_SIZE \
128 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
129 /* each chunk includes extra space to keep handle */
130 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
133 * On systems with 4K page size, this gives 255 size classes! There is a
135 * - Large number of size classes is potentially wasteful as free page are
136 * spread across these classes
137 * - Small number of size classes causes large internal fragmentation
138 * - Probably its better to use specific size classes (empirically
139 * determined). NOTE: all those class sizes must be set as multiple of
140 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
142 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
145 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
146 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
147 ZS_SIZE_CLASS_DELTA) + 1)
150 * Pages are distinguished by the ratio of used memory (that is the ratio
151 * of ->inuse objects to all objects that page can store). For example,
152 * INUSE_RATIO_10 means that the ratio of used objects is > 0% and <= 10%.
154 * The number of fullness groups is not random. It allows us to keep
155 * difference between the least busy page in the group (minimum permitted
156 * number of ->inuse objects) and the most busy page (maximum permitted
157 * number of ->inuse objects) at a reasonable value.
159 enum fullness_group {
162 /* NOTE: 8 more fullness groups here */
163 ZS_INUSE_RATIO_99 = 10,
168 enum class_stat_type {
169 /* NOTE: stats for 12 fullness groups here: from inuse 0 to 100 */
170 ZS_OBJS_ALLOCATED = NR_FULLNESS_GROUPS,
175 struct zs_size_stat {
176 unsigned long objs[NR_CLASS_STAT_TYPES];
179 #ifdef CONFIG_ZSMALLOC_STAT
180 static struct dentry *zs_stat_root;
183 static size_t huge_class_size;
186 struct list_head fullness_list[NR_FULLNESS_GROUPS];
188 * Size of objects stored in this class. Must be multiple
193 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
194 int pages_per_zspage;
197 struct zs_size_stat stats;
201 * Placed within free objects to form a singly linked list.
202 * For every zspage, zspage->freeobj gives head of this list.
204 * This must be power of 2 and less than or equal to ZS_ALIGN
210 * It's valid for non-allocated object
214 * Handle of allocated object.
216 unsigned long handle;
223 struct size_class *size_class[ZS_SIZE_CLASSES];
224 struct kmem_cache *handle_cachep;
225 struct kmem_cache *zspage_cachep;
227 atomic_long_t pages_allocated;
229 struct zs_pool_stats stats;
231 /* Compact classes */
232 struct shrinker shrinker;
234 #ifdef CONFIG_ZSMALLOC_STAT
235 struct dentry *stat_dentry;
237 #ifdef CONFIG_COMPACTION
238 struct work_struct free_work;
241 atomic_t compaction_in_progress;
246 unsigned int huge:HUGE_BITS;
247 unsigned int fullness:FULLNESS_BITS;
248 unsigned int class:CLASS_BITS + 1;
249 unsigned int isolated:ISOLATED_BITS;
250 unsigned int magic:MAGIC_VAL_BITS;
253 unsigned int freeobj;
254 struct page *first_page;
255 struct list_head list; /* fullness list */
256 struct zs_pool *pool;
260 struct mapping_area {
262 char *vm_buf; /* copy buffer for objects that span pages */
263 char *vm_addr; /* address of kmap_atomic()'ed pages */
264 enum zs_mapmode vm_mm; /* mapping mode */
267 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
268 static void SetZsHugePage(struct zspage *zspage)
273 static bool ZsHugePage(struct zspage *zspage)
278 static void migrate_lock_init(struct zspage *zspage);
279 static void migrate_read_lock(struct zspage *zspage);
280 static void migrate_read_unlock(struct zspage *zspage);
282 #ifdef CONFIG_COMPACTION
283 static void migrate_write_lock(struct zspage *zspage);
284 static void migrate_write_lock_nested(struct zspage *zspage);
285 static void migrate_write_unlock(struct zspage *zspage);
286 static void kick_deferred_free(struct zs_pool *pool);
287 static void init_deferred_free(struct zs_pool *pool);
288 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
290 static void migrate_write_lock(struct zspage *zspage) {}
291 static void migrate_write_lock_nested(struct zspage *zspage) {}
292 static void migrate_write_unlock(struct zspage *zspage) {}
293 static void kick_deferred_free(struct zs_pool *pool) {}
294 static void init_deferred_free(struct zs_pool *pool) {}
295 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
298 static int create_cache(struct zs_pool *pool)
300 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
302 if (!pool->handle_cachep)
305 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
307 if (!pool->zspage_cachep) {
308 kmem_cache_destroy(pool->handle_cachep);
309 pool->handle_cachep = NULL;
316 static void destroy_cache(struct zs_pool *pool)
318 kmem_cache_destroy(pool->handle_cachep);
319 kmem_cache_destroy(pool->zspage_cachep);
322 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
324 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
325 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
328 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
330 kmem_cache_free(pool->handle_cachep, (void *)handle);
333 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
335 return kmem_cache_zalloc(pool->zspage_cachep,
336 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
339 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
341 kmem_cache_free(pool->zspage_cachep, zspage);
344 /* pool->lock(which owns the handle) synchronizes races */
345 static void record_obj(unsigned long handle, unsigned long obj)
347 *(unsigned long *)handle = obj;
354 static void *zs_zpool_create(const char *name, gfp_t gfp)
357 * Ignore global gfp flags: zs_malloc() may be invoked from
358 * different contexts and its caller must provide a valid
361 return zs_create_pool(name);
364 static void zs_zpool_destroy(void *pool)
366 zs_destroy_pool(pool);
369 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
370 unsigned long *handle)
372 *handle = zs_malloc(pool, size, gfp);
374 if (IS_ERR_VALUE(*handle))
375 return PTR_ERR((void *)*handle);
378 static void zs_zpool_free(void *pool, unsigned long handle)
380 zs_free(pool, handle);
383 static void *zs_zpool_map(void *pool, unsigned long handle,
384 enum zpool_mapmode mm)
386 enum zs_mapmode zs_mm;
401 return zs_map_object(pool, handle, zs_mm);
403 static void zs_zpool_unmap(void *pool, unsigned long handle)
405 zs_unmap_object(pool, handle);
408 static u64 zs_zpool_total_size(void *pool)
410 return zs_get_total_pages(pool) << PAGE_SHIFT;
413 static struct zpool_driver zs_zpool_driver = {
415 .owner = THIS_MODULE,
416 .create = zs_zpool_create,
417 .destroy = zs_zpool_destroy,
418 .malloc_support_movable = true,
419 .malloc = zs_zpool_malloc,
420 .free = zs_zpool_free,
422 .unmap = zs_zpool_unmap,
423 .total_size = zs_zpool_total_size,
426 MODULE_ALIAS("zpool-zsmalloc");
427 #endif /* CONFIG_ZPOOL */
429 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
430 static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
431 .lock = INIT_LOCAL_LOCK(lock),
434 static __maybe_unused int is_first_page(struct page *page)
436 return PagePrivate(page);
439 /* Protected by pool->lock */
440 static inline int get_zspage_inuse(struct zspage *zspage)
442 return zspage->inuse;
446 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
448 zspage->inuse += val;
451 static inline struct page *get_first_page(struct zspage *zspage)
453 struct page *first_page = zspage->first_page;
455 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
459 static inline unsigned int get_first_obj_offset(struct page *page)
461 return page->page_type;
464 static inline void set_first_obj_offset(struct page *page, unsigned int offset)
466 page->page_type = offset;
469 static inline unsigned int get_freeobj(struct zspage *zspage)
471 return zspage->freeobj;
474 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
476 zspage->freeobj = obj;
479 static void get_zspage_mapping(struct zspage *zspage,
480 unsigned int *class_idx,
483 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
485 *fullness = zspage->fullness;
486 *class_idx = zspage->class;
489 static struct size_class *zspage_class(struct zs_pool *pool,
490 struct zspage *zspage)
492 return pool->size_class[zspage->class];
495 static void set_zspage_mapping(struct zspage *zspage,
496 unsigned int class_idx,
499 zspage->class = class_idx;
500 zspage->fullness = fullness;
504 * zsmalloc divides the pool into various size classes where each
505 * class maintains a list of zspages where each zspage is divided
506 * into equal sized chunks. Each allocation falls into one of these
507 * classes depending on its size. This function returns index of the
508 * size class which has chunk size big enough to hold the given size.
510 static int get_size_class_index(int size)
514 if (likely(size > ZS_MIN_ALLOC_SIZE))
515 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
516 ZS_SIZE_CLASS_DELTA);
518 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
521 static inline void class_stat_inc(struct size_class *class,
522 int type, unsigned long cnt)
524 class->stats.objs[type] += cnt;
527 static inline void class_stat_dec(struct size_class *class,
528 int type, unsigned long cnt)
530 class->stats.objs[type] -= cnt;
533 static inline unsigned long zs_stat_get(struct size_class *class, int type)
535 return class->stats.objs[type];
538 #ifdef CONFIG_ZSMALLOC_STAT
540 static void __init zs_stat_init(void)
542 if (!debugfs_initialized()) {
543 pr_warn("debugfs not available, stat dir not created\n");
547 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
550 static void __exit zs_stat_exit(void)
552 debugfs_remove_recursive(zs_stat_root);
555 static unsigned long zs_can_compact(struct size_class *class);
557 static int zs_stats_size_show(struct seq_file *s, void *v)
560 struct zs_pool *pool = s->private;
561 struct size_class *class;
563 unsigned long obj_allocated, obj_used, pages_used, freeable;
564 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
565 unsigned long total_freeable = 0;
566 unsigned long inuse_totals[NR_FULLNESS_GROUPS] = {0, };
568 seq_printf(s, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
569 "class", "size", "10%", "20%", "30%", "40%",
570 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
571 "obj_allocated", "obj_used", "pages_used",
572 "pages_per_zspage", "freeable");
574 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
576 class = pool->size_class[i];
578 if (class->index != i)
581 spin_lock(&pool->lock);
583 seq_printf(s, " %5u %5u ", i, class->size);
584 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++) {
585 inuse_totals[fg] += zs_stat_get(class, fg);
586 seq_printf(s, "%9lu ", zs_stat_get(class, fg));
589 obj_allocated = zs_stat_get(class, ZS_OBJS_ALLOCATED);
590 obj_used = zs_stat_get(class, ZS_OBJS_INUSE);
591 freeable = zs_can_compact(class);
592 spin_unlock(&pool->lock);
594 objs_per_zspage = class->objs_per_zspage;
595 pages_used = obj_allocated / objs_per_zspage *
596 class->pages_per_zspage;
598 seq_printf(s, "%13lu %10lu %10lu %16d %8lu\n",
599 obj_allocated, obj_used, pages_used,
600 class->pages_per_zspage, freeable);
602 total_objs += obj_allocated;
603 total_used_objs += obj_used;
604 total_pages += pages_used;
605 total_freeable += freeable;
609 seq_printf(s, " %5s %5s ", "Total", "");
611 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++)
612 seq_printf(s, "%9lu ", inuse_totals[fg]);
614 seq_printf(s, "%13lu %10lu %10lu %16s %8lu\n",
615 total_objs, total_used_objs, total_pages, "",
620 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
622 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
625 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
629 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
631 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
632 &zs_stats_size_fops);
635 static void zs_pool_stat_destroy(struct zs_pool *pool)
637 debugfs_remove_recursive(pool->stat_dentry);
640 #else /* CONFIG_ZSMALLOC_STAT */
641 static void __init zs_stat_init(void)
645 static void __exit zs_stat_exit(void)
649 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
653 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
660 * For each size class, zspages are divided into different groups
661 * depending on their usage ratio. This function returns fullness
662 * status of the given page.
664 static int get_fullness_group(struct size_class *class, struct zspage *zspage)
666 int inuse, objs_per_zspage, ratio;
668 inuse = get_zspage_inuse(zspage);
669 objs_per_zspage = class->objs_per_zspage;
672 return ZS_INUSE_RATIO_0;
673 if (inuse == objs_per_zspage)
674 return ZS_INUSE_RATIO_100;
676 ratio = 100 * inuse / objs_per_zspage;
678 * Take integer division into consideration: a page with one inuse
679 * object out of 127 possible, will end up having 0 usage ratio,
680 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
682 return ratio / 10 + 1;
686 * Each size class maintains various freelists and zspages are assigned
687 * to one of these freelists based on the number of live objects they
688 * have. This functions inserts the given zspage into the freelist
689 * identified by <class, fullness_group>.
691 static void insert_zspage(struct size_class *class,
692 struct zspage *zspage,
695 class_stat_inc(class, fullness, 1);
696 list_add(&zspage->list, &class->fullness_list[fullness]);
700 * This function removes the given zspage from the freelist identified
701 * by <class, fullness_group>.
703 static void remove_zspage(struct size_class *class,
704 struct zspage *zspage,
707 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
709 list_del_init(&zspage->list);
710 class_stat_dec(class, fullness, 1);
714 * Each size class maintains zspages in different fullness groups depending
715 * on the number of live objects they contain. When allocating or freeing
716 * objects, the fullness status of the page can change, for instance, from
717 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
718 * checks if such a status change has occurred for the given page and
719 * accordingly moves the page from the list of the old fullness group to that
720 * of the new fullness group.
722 static int fix_fullness_group(struct size_class *class, struct zspage *zspage)
727 get_zspage_mapping(zspage, &class_idx, &currfg);
728 newfg = get_fullness_group(class, zspage);
732 remove_zspage(class, zspage, currfg);
733 insert_zspage(class, zspage, newfg);
734 set_zspage_mapping(zspage, class_idx, newfg);
739 static struct zspage *get_zspage(struct page *page)
741 struct zspage *zspage = (struct zspage *)page_private(page);
743 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
747 static struct page *get_next_page(struct page *page)
749 struct zspage *zspage = get_zspage(page);
751 if (unlikely(ZsHugePage(zspage)))
754 return (struct page *)page->index;
758 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
759 * @obj: the encoded object value
760 * @page: page object resides in zspage
761 * @obj_idx: object index
763 static void obj_to_location(unsigned long obj, struct page **page,
764 unsigned int *obj_idx)
766 obj >>= OBJ_TAG_BITS;
767 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
768 *obj_idx = (obj & OBJ_INDEX_MASK);
771 static void obj_to_page(unsigned long obj, struct page **page)
773 obj >>= OBJ_TAG_BITS;
774 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
778 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
779 * @page: page object resides in zspage
780 * @obj_idx: object index
782 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
786 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
787 obj |= obj_idx & OBJ_INDEX_MASK;
788 obj <<= OBJ_TAG_BITS;
793 static unsigned long handle_to_obj(unsigned long handle)
795 return *(unsigned long *)handle;
798 static bool obj_tagged(struct page *page, void *obj, unsigned long *phandle,
801 unsigned long handle;
802 struct zspage *zspage = get_zspage(page);
804 if (unlikely(ZsHugePage(zspage))) {
805 VM_BUG_ON_PAGE(!is_first_page(page), page);
806 handle = page->index;
808 handle = *(unsigned long *)obj;
813 /* Clear all tags before returning the handle */
814 *phandle = handle & ~OBJ_TAG_MASK;
818 static inline bool obj_allocated(struct page *page, void *obj, unsigned long *phandle)
820 return obj_tagged(page, obj, phandle, OBJ_ALLOCATED_TAG);
823 static void reset_page(struct page *page)
825 __ClearPageMovable(page);
826 ClearPagePrivate(page);
827 set_page_private(page, 0);
828 page_mapcount_reset(page);
832 static int trylock_zspage(struct zspage *zspage)
834 struct page *cursor, *fail;
836 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
837 get_next_page(cursor)) {
838 if (!trylock_page(cursor)) {
846 for (cursor = get_first_page(zspage); cursor != fail; cursor =
847 get_next_page(cursor))
853 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
854 struct zspage *zspage)
856 struct page *page, *next;
858 unsigned int class_idx;
860 get_zspage_mapping(zspage, &class_idx, &fg);
862 assert_spin_locked(&pool->lock);
864 VM_BUG_ON(get_zspage_inuse(zspage));
865 VM_BUG_ON(fg != ZS_INUSE_RATIO_0);
867 next = page = get_first_page(zspage);
869 VM_BUG_ON_PAGE(!PageLocked(page), page);
870 next = get_next_page(page);
873 dec_zone_page_state(page, NR_ZSPAGES);
876 } while (page != NULL);
878 cache_free_zspage(pool, zspage);
880 class_stat_dec(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
881 atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
884 static void free_zspage(struct zs_pool *pool, struct size_class *class,
885 struct zspage *zspage)
887 VM_BUG_ON(get_zspage_inuse(zspage));
888 VM_BUG_ON(list_empty(&zspage->list));
891 * Since zs_free couldn't be sleepable, this function cannot call
892 * lock_page. The page locks trylock_zspage got will be released
895 if (!trylock_zspage(zspage)) {
896 kick_deferred_free(pool);
900 remove_zspage(class, zspage, ZS_INUSE_RATIO_0);
901 __free_zspage(pool, class, zspage);
904 /* Initialize a newly allocated zspage */
905 static void init_zspage(struct size_class *class, struct zspage *zspage)
907 unsigned int freeobj = 1;
908 unsigned long off = 0;
909 struct page *page = get_first_page(zspage);
912 struct page *next_page;
913 struct link_free *link;
916 set_first_obj_offset(page, off);
918 vaddr = kmap_atomic(page);
919 link = (struct link_free *)vaddr + off / sizeof(*link);
921 while ((off += class->size) < PAGE_SIZE) {
922 link->next = freeobj++ << OBJ_TAG_BITS;
923 link += class->size / sizeof(*link);
927 * We now come to the last (full or partial) object on this
928 * page, which must point to the first object on the next
931 next_page = get_next_page(page);
933 link->next = freeobj++ << OBJ_TAG_BITS;
936 * Reset OBJ_TAG_BITS bit to last link to tell
937 * whether it's allocated object or not.
939 link->next = -1UL << OBJ_TAG_BITS;
941 kunmap_atomic(vaddr);
946 set_freeobj(zspage, 0);
949 static void create_page_chain(struct size_class *class, struct zspage *zspage,
950 struct page *pages[])
954 struct page *prev_page = NULL;
955 int nr_pages = class->pages_per_zspage;
958 * Allocate individual pages and link them together as:
959 * 1. all pages are linked together using page->index
960 * 2. each sub-page point to zspage using page->private
962 * we set PG_private to identify the first page (i.e. no other sub-page
963 * has this flag set).
965 for (i = 0; i < nr_pages; i++) {
967 set_page_private(page, (unsigned long)zspage);
970 zspage->first_page = page;
971 SetPagePrivate(page);
972 if (unlikely(class->objs_per_zspage == 1 &&
973 class->pages_per_zspage == 1))
974 SetZsHugePage(zspage);
976 prev_page->index = (unsigned long)page;
983 * Allocate a zspage for the given size class
985 static struct zspage *alloc_zspage(struct zs_pool *pool,
986 struct size_class *class,
990 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
991 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
996 zspage->magic = ZSPAGE_MAGIC;
997 migrate_lock_init(zspage);
999 for (i = 0; i < class->pages_per_zspage; i++) {
1002 page = alloc_page(gfp);
1005 dec_zone_page_state(pages[i], NR_ZSPAGES);
1006 __free_page(pages[i]);
1008 cache_free_zspage(pool, zspage);
1012 inc_zone_page_state(page, NR_ZSPAGES);
1016 create_page_chain(class, zspage, pages);
1017 init_zspage(class, zspage);
1018 zspage->pool = pool;
1023 static struct zspage *find_get_zspage(struct size_class *class)
1026 struct zspage *zspage;
1028 for (i = ZS_INUSE_RATIO_99; i >= ZS_INUSE_RATIO_0; i--) {
1029 zspage = list_first_entry_or_null(&class->fullness_list[i],
1030 struct zspage, list);
1038 static inline int __zs_cpu_up(struct mapping_area *area)
1041 * Make sure we don't leak memory if a cpu UP notification
1042 * and zs_init() race and both call zs_cpu_up() on the same cpu
1046 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1052 static inline void __zs_cpu_down(struct mapping_area *area)
1054 kfree(area->vm_buf);
1055 area->vm_buf = NULL;
1058 static void *__zs_map_object(struct mapping_area *area,
1059 struct page *pages[2], int off, int size)
1063 char *buf = area->vm_buf;
1065 /* disable page faults to match kmap_atomic() return conditions */
1066 pagefault_disable();
1068 /* no read fastpath */
1069 if (area->vm_mm == ZS_MM_WO)
1072 sizes[0] = PAGE_SIZE - off;
1073 sizes[1] = size - sizes[0];
1075 /* copy object to per-cpu buffer */
1076 addr = kmap_atomic(pages[0]);
1077 memcpy(buf, addr + off, sizes[0]);
1078 kunmap_atomic(addr);
1079 addr = kmap_atomic(pages[1]);
1080 memcpy(buf + sizes[0], addr, sizes[1]);
1081 kunmap_atomic(addr);
1083 return area->vm_buf;
1086 static void __zs_unmap_object(struct mapping_area *area,
1087 struct page *pages[2], int off, int size)
1093 /* no write fastpath */
1094 if (area->vm_mm == ZS_MM_RO)
1098 buf = buf + ZS_HANDLE_SIZE;
1099 size -= ZS_HANDLE_SIZE;
1100 off += ZS_HANDLE_SIZE;
1102 sizes[0] = PAGE_SIZE - off;
1103 sizes[1] = size - sizes[0];
1105 /* copy per-cpu buffer to object */
1106 addr = kmap_atomic(pages[0]);
1107 memcpy(addr + off, buf, sizes[0]);
1108 kunmap_atomic(addr);
1109 addr = kmap_atomic(pages[1]);
1110 memcpy(addr, buf + sizes[0], sizes[1]);
1111 kunmap_atomic(addr);
1114 /* enable page faults to match kunmap_atomic() return conditions */
1118 static int zs_cpu_prepare(unsigned int cpu)
1120 struct mapping_area *area;
1122 area = &per_cpu(zs_map_area, cpu);
1123 return __zs_cpu_up(area);
1126 static int zs_cpu_dead(unsigned int cpu)
1128 struct mapping_area *area;
1130 area = &per_cpu(zs_map_area, cpu);
1131 __zs_cpu_down(area);
1135 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1136 int objs_per_zspage)
1138 if (prev->pages_per_zspage == pages_per_zspage &&
1139 prev->objs_per_zspage == objs_per_zspage)
1145 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1147 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1150 static bool zspage_empty(struct zspage *zspage)
1152 return get_zspage_inuse(zspage) == 0;
1156 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1157 * that hold objects of the provided size.
1158 * @pool: zsmalloc pool to use
1159 * @size: object size
1161 * Context: Any context.
1163 * Return: the index of the zsmalloc &size_class that hold objects of the
1166 unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1168 struct size_class *class;
1170 class = pool->size_class[get_size_class_index(size)];
1172 return class->index;
1174 EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1176 unsigned long zs_get_total_pages(struct zs_pool *pool)
1178 return atomic_long_read(&pool->pages_allocated);
1180 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1183 * zs_map_object - get address of allocated object from handle.
1184 * @pool: pool from which the object was allocated
1185 * @handle: handle returned from zs_malloc
1186 * @mm: mapping mode to use
1188 * Before using an object allocated from zs_malloc, it must be mapped using
1189 * this function. When done with the object, it must be unmapped using
1192 * Only one object can be mapped per cpu at a time. There is no protection
1193 * against nested mappings.
1195 * This function returns with preemption and page faults disabled.
1197 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1200 struct zspage *zspage;
1202 unsigned long obj, off;
1203 unsigned int obj_idx;
1205 struct size_class *class;
1206 struct mapping_area *area;
1207 struct page *pages[2];
1211 * Because we use per-cpu mapping areas shared among the
1212 * pools/users, we can't allow mapping in interrupt context
1213 * because it can corrupt another users mappings.
1215 BUG_ON(in_interrupt());
1217 /* It guarantees it can get zspage from handle safely */
1218 spin_lock(&pool->lock);
1219 obj = handle_to_obj(handle);
1220 obj_to_location(obj, &page, &obj_idx);
1221 zspage = get_zspage(page);
1224 * migration cannot move any zpages in this zspage. Here, pool->lock
1225 * is too heavy since callers would take some time until they calls
1226 * zs_unmap_object API so delegate the locking from class to zspage
1227 * which is smaller granularity.
1229 migrate_read_lock(zspage);
1230 spin_unlock(&pool->lock);
1232 class = zspage_class(pool, zspage);
1233 off = offset_in_page(class->size * obj_idx);
1235 local_lock(&zs_map_area.lock);
1236 area = this_cpu_ptr(&zs_map_area);
1238 if (off + class->size <= PAGE_SIZE) {
1239 /* this object is contained entirely within a page */
1240 area->vm_addr = kmap_atomic(page);
1241 ret = area->vm_addr + off;
1245 /* this object spans two pages */
1247 pages[1] = get_next_page(page);
1250 ret = __zs_map_object(area, pages, off, class->size);
1252 if (likely(!ZsHugePage(zspage)))
1253 ret += ZS_HANDLE_SIZE;
1257 EXPORT_SYMBOL_GPL(zs_map_object);
1259 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1261 struct zspage *zspage;
1263 unsigned long obj, off;
1264 unsigned int obj_idx;
1266 struct size_class *class;
1267 struct mapping_area *area;
1269 obj = handle_to_obj(handle);
1270 obj_to_location(obj, &page, &obj_idx);
1271 zspage = get_zspage(page);
1272 class = zspage_class(pool, zspage);
1273 off = offset_in_page(class->size * obj_idx);
1275 area = this_cpu_ptr(&zs_map_area);
1276 if (off + class->size <= PAGE_SIZE)
1277 kunmap_atomic(area->vm_addr);
1279 struct page *pages[2];
1282 pages[1] = get_next_page(page);
1285 __zs_unmap_object(area, pages, off, class->size);
1287 local_unlock(&zs_map_area.lock);
1289 migrate_read_unlock(zspage);
1291 EXPORT_SYMBOL_GPL(zs_unmap_object);
1294 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1295 * zsmalloc &size_class.
1296 * @pool: zsmalloc pool to use
1298 * The function returns the size of the first huge class - any object of equal
1299 * or bigger size will be stored in zspage consisting of a single physical
1302 * Context: Any context.
1304 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1306 size_t zs_huge_class_size(struct zs_pool *pool)
1308 return huge_class_size;
1310 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1312 static unsigned long obj_malloc(struct zs_pool *pool,
1313 struct zspage *zspage, unsigned long handle)
1315 int i, nr_page, offset;
1317 struct link_free *link;
1318 struct size_class *class;
1320 struct page *m_page;
1321 unsigned long m_offset;
1324 class = pool->size_class[zspage->class];
1325 handle |= OBJ_ALLOCATED_TAG;
1326 obj = get_freeobj(zspage);
1328 offset = obj * class->size;
1329 nr_page = offset >> PAGE_SHIFT;
1330 m_offset = offset_in_page(offset);
1331 m_page = get_first_page(zspage);
1333 for (i = 0; i < nr_page; i++)
1334 m_page = get_next_page(m_page);
1336 vaddr = kmap_atomic(m_page);
1337 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1338 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1339 if (likely(!ZsHugePage(zspage)))
1340 /* record handle in the header of allocated chunk */
1341 link->handle = handle;
1343 /* record handle to page->index */
1344 zspage->first_page->index = handle;
1346 kunmap_atomic(vaddr);
1347 mod_zspage_inuse(zspage, 1);
1349 obj = location_to_obj(m_page, obj);
1356 * zs_malloc - Allocate block of given size from pool.
1357 * @pool: pool to allocate from
1358 * @size: size of block to allocate
1359 * @gfp: gfp flags when allocating object
1361 * On success, handle to the allocated object is returned,
1362 * otherwise an ERR_PTR().
1363 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1365 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1367 unsigned long handle, obj;
1368 struct size_class *class;
1370 struct zspage *zspage;
1372 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1373 return (unsigned long)ERR_PTR(-EINVAL);
1375 handle = cache_alloc_handle(pool, gfp);
1377 return (unsigned long)ERR_PTR(-ENOMEM);
1379 /* extra space in chunk to keep the handle */
1380 size += ZS_HANDLE_SIZE;
1381 class = pool->size_class[get_size_class_index(size)];
1383 /* pool->lock effectively protects the zpage migration */
1384 spin_lock(&pool->lock);
1385 zspage = find_get_zspage(class);
1386 if (likely(zspage)) {
1387 obj = obj_malloc(pool, zspage, handle);
1388 /* Now move the zspage to another fullness group, if required */
1389 fix_fullness_group(class, zspage);
1390 record_obj(handle, obj);
1391 class_stat_inc(class, ZS_OBJS_INUSE, 1);
1396 spin_unlock(&pool->lock);
1398 zspage = alloc_zspage(pool, class, gfp);
1400 cache_free_handle(pool, handle);
1401 return (unsigned long)ERR_PTR(-ENOMEM);
1404 spin_lock(&pool->lock);
1405 obj = obj_malloc(pool, zspage, handle);
1406 newfg = get_fullness_group(class, zspage);
1407 insert_zspage(class, zspage, newfg);
1408 set_zspage_mapping(zspage, class->index, newfg);
1409 record_obj(handle, obj);
1410 atomic_long_add(class->pages_per_zspage, &pool->pages_allocated);
1411 class_stat_inc(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
1412 class_stat_inc(class, ZS_OBJS_INUSE, 1);
1414 /* We completely set up zspage so mark them as movable */
1415 SetZsPageMovable(pool, zspage);
1417 spin_unlock(&pool->lock);
1421 EXPORT_SYMBOL_GPL(zs_malloc);
1423 static void obj_free(int class_size, unsigned long obj)
1425 struct link_free *link;
1426 struct zspage *zspage;
1427 struct page *f_page;
1428 unsigned long f_offset;
1429 unsigned int f_objidx;
1432 obj_to_location(obj, &f_page, &f_objidx);
1433 f_offset = offset_in_page(class_size * f_objidx);
1434 zspage = get_zspage(f_page);
1436 vaddr = kmap_atomic(f_page);
1437 link = (struct link_free *)(vaddr + f_offset);
1439 /* Insert this object in containing zspage's freelist */
1440 if (likely(!ZsHugePage(zspage)))
1441 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1444 set_freeobj(zspage, f_objidx);
1446 kunmap_atomic(vaddr);
1447 mod_zspage_inuse(zspage, -1);
1450 void zs_free(struct zs_pool *pool, unsigned long handle)
1452 struct zspage *zspage;
1453 struct page *f_page;
1455 struct size_class *class;
1458 if (IS_ERR_OR_NULL((void *)handle))
1462 * The pool->lock protects the race with zpage's migration
1463 * so it's safe to get the page from handle.
1465 spin_lock(&pool->lock);
1466 obj = handle_to_obj(handle);
1467 obj_to_page(obj, &f_page);
1468 zspage = get_zspage(f_page);
1469 class = zspage_class(pool, zspage);
1471 class_stat_dec(class, ZS_OBJS_INUSE, 1);
1472 obj_free(class->size, obj);
1474 fullness = fix_fullness_group(class, zspage);
1475 if (fullness == ZS_INUSE_RATIO_0)
1476 free_zspage(pool, class, zspage);
1478 spin_unlock(&pool->lock);
1479 cache_free_handle(pool, handle);
1481 EXPORT_SYMBOL_GPL(zs_free);
1483 static void zs_object_copy(struct size_class *class, unsigned long dst,
1486 struct page *s_page, *d_page;
1487 unsigned int s_objidx, d_objidx;
1488 unsigned long s_off, d_off;
1489 void *s_addr, *d_addr;
1490 int s_size, d_size, size;
1493 s_size = d_size = class->size;
1495 obj_to_location(src, &s_page, &s_objidx);
1496 obj_to_location(dst, &d_page, &d_objidx);
1498 s_off = offset_in_page(class->size * s_objidx);
1499 d_off = offset_in_page(class->size * d_objidx);
1501 if (s_off + class->size > PAGE_SIZE)
1502 s_size = PAGE_SIZE - s_off;
1504 if (d_off + class->size > PAGE_SIZE)
1505 d_size = PAGE_SIZE - d_off;
1507 s_addr = kmap_atomic(s_page);
1508 d_addr = kmap_atomic(d_page);
1511 size = min(s_size, d_size);
1512 memcpy(d_addr + d_off, s_addr + s_off, size);
1515 if (written == class->size)
1524 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1525 * calls must occurs in reverse order of calls to kmap_atomic().
1526 * So, to call kunmap_atomic(s_addr) we should first call
1527 * kunmap_atomic(d_addr). For more details see
1528 * Documentation/mm/highmem.rst.
1530 if (s_off >= PAGE_SIZE) {
1531 kunmap_atomic(d_addr);
1532 kunmap_atomic(s_addr);
1533 s_page = get_next_page(s_page);
1534 s_addr = kmap_atomic(s_page);
1535 d_addr = kmap_atomic(d_page);
1536 s_size = class->size - written;
1540 if (d_off >= PAGE_SIZE) {
1541 kunmap_atomic(d_addr);
1542 d_page = get_next_page(d_page);
1543 d_addr = kmap_atomic(d_page);
1544 d_size = class->size - written;
1549 kunmap_atomic(d_addr);
1550 kunmap_atomic(s_addr);
1554 * Find object with a certain tag in zspage from index object and
1557 static unsigned long find_tagged_obj(struct size_class *class,
1558 struct page *page, int *obj_idx, int tag)
1560 unsigned int offset;
1561 int index = *obj_idx;
1562 unsigned long handle = 0;
1563 void *addr = kmap_atomic(page);
1565 offset = get_first_obj_offset(page);
1566 offset += class->size * index;
1568 while (offset < PAGE_SIZE) {
1569 if (obj_tagged(page, addr + offset, &handle, tag))
1572 offset += class->size;
1576 kunmap_atomic(addr);
1584 * Find alloced object in zspage from index object and
1587 static unsigned long find_alloced_obj(struct size_class *class,
1588 struct page *page, int *obj_idx)
1590 return find_tagged_obj(class, page, obj_idx, OBJ_ALLOCATED_TAG);
1593 struct zs_compact_control {
1594 /* Source spage for migration which could be a subpage of zspage */
1595 struct page *s_page;
1596 /* Destination page for migration which should be a first page
1598 struct page *d_page;
1599 /* Starting object index within @s_page which used for live object
1600 * in the subpage. */
1604 static void migrate_zspage(struct zs_pool *pool, struct size_class *class,
1605 struct zs_compact_control *cc)
1607 unsigned long used_obj, free_obj;
1608 unsigned long handle;
1609 struct page *s_page = cc->s_page;
1610 struct page *d_page = cc->d_page;
1611 int obj_idx = cc->obj_idx;
1614 handle = find_alloced_obj(class, s_page, &obj_idx);
1616 s_page = get_next_page(s_page);
1623 used_obj = handle_to_obj(handle);
1624 free_obj = obj_malloc(pool, get_zspage(d_page), handle);
1625 zs_object_copy(class, free_obj, used_obj);
1627 record_obj(handle, free_obj);
1628 obj_free(class->size, used_obj);
1630 /* Stop if there is no more space */
1631 if (zspage_full(class, get_zspage(d_page)))
1634 /* Stop if there are no more objects to migrate */
1635 if (zspage_empty(get_zspage(s_page)))
1639 /* Remember last position in this iteration */
1640 cc->s_page = s_page;
1641 cc->obj_idx = obj_idx;
1644 static struct zspage *isolate_src_zspage(struct size_class *class)
1646 struct zspage *zspage;
1649 for (fg = ZS_INUSE_RATIO_10; fg <= ZS_INUSE_RATIO_99; fg++) {
1650 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1651 struct zspage, list);
1653 remove_zspage(class, zspage, fg);
1661 static struct zspage *isolate_dst_zspage(struct size_class *class)
1663 struct zspage *zspage;
1666 for (fg = ZS_INUSE_RATIO_99; fg >= ZS_INUSE_RATIO_10; fg--) {
1667 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1668 struct zspage, list);
1670 remove_zspage(class, zspage, fg);
1679 * putback_zspage - add @zspage into right class's fullness list
1680 * @class: destination class
1681 * @zspage: target page
1683 * Return @zspage's fullness status
1685 static int putback_zspage(struct size_class *class, struct zspage *zspage)
1689 fullness = get_fullness_group(class, zspage);
1690 insert_zspage(class, zspage, fullness);
1691 set_zspage_mapping(zspage, class->index, fullness);
1696 #ifdef CONFIG_COMPACTION
1698 * To prevent zspage destroy during migration, zspage freeing should
1699 * hold locks of all pages in the zspage.
1701 static void lock_zspage(struct zspage *zspage)
1703 struct page *curr_page, *page;
1706 * Pages we haven't locked yet can be migrated off the list while we're
1707 * trying to lock them, so we need to be careful and only attempt to
1708 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1709 * may no longer belong to the zspage. This means that we may wait for
1710 * the wrong page to unlock, so we must take a reference to the page
1711 * prior to waiting for it to unlock outside migrate_read_lock().
1714 migrate_read_lock(zspage);
1715 page = get_first_page(zspage);
1716 if (trylock_page(page))
1719 migrate_read_unlock(zspage);
1720 wait_on_page_locked(page);
1725 while ((page = get_next_page(curr_page))) {
1726 if (trylock_page(page)) {
1730 migrate_read_unlock(zspage);
1731 wait_on_page_locked(page);
1733 migrate_read_lock(zspage);
1736 migrate_read_unlock(zspage);
1738 #endif /* CONFIG_COMPACTION */
1740 static void migrate_lock_init(struct zspage *zspage)
1742 rwlock_init(&zspage->lock);
1745 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1747 read_lock(&zspage->lock);
1750 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1752 read_unlock(&zspage->lock);
1755 #ifdef CONFIG_COMPACTION
1756 static void migrate_write_lock(struct zspage *zspage)
1758 write_lock(&zspage->lock);
1761 static void migrate_write_lock_nested(struct zspage *zspage)
1763 write_lock_nested(&zspage->lock, SINGLE_DEPTH_NESTING);
1766 static void migrate_write_unlock(struct zspage *zspage)
1768 write_unlock(&zspage->lock);
1771 /* Number of isolated subpage for *page migration* in this zspage */
1772 static void inc_zspage_isolation(struct zspage *zspage)
1777 static void dec_zspage_isolation(struct zspage *zspage)
1779 VM_BUG_ON(zspage->isolated == 0);
1783 static const struct movable_operations zsmalloc_mops;
1785 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1786 struct page *newpage, struct page *oldpage)
1789 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1792 page = get_first_page(zspage);
1794 if (page == oldpage)
1795 pages[idx] = newpage;
1799 } while ((page = get_next_page(page)) != NULL);
1801 create_page_chain(class, zspage, pages);
1802 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1803 if (unlikely(ZsHugePage(zspage)))
1804 newpage->index = oldpage->index;
1805 __SetPageMovable(newpage, &zsmalloc_mops);
1808 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1810 struct zspage *zspage;
1813 * Page is locked so zspage couldn't be destroyed. For detail, look at
1814 * lock_zspage in free_zspage.
1816 VM_BUG_ON_PAGE(PageIsolated(page), page);
1818 zspage = get_zspage(page);
1819 migrate_write_lock(zspage);
1820 inc_zspage_isolation(zspage);
1821 migrate_write_unlock(zspage);
1826 static int zs_page_migrate(struct page *newpage, struct page *page,
1827 enum migrate_mode mode)
1829 struct zs_pool *pool;
1830 struct size_class *class;
1831 struct zspage *zspage;
1833 void *s_addr, *d_addr, *addr;
1834 unsigned int offset;
1835 unsigned long handle;
1836 unsigned long old_obj, new_obj;
1837 unsigned int obj_idx;
1840 * We cannot support the _NO_COPY case here, because copy needs to
1841 * happen under the zs lock, which does not work with
1842 * MIGRATE_SYNC_NO_COPY workflow.
1844 if (mode == MIGRATE_SYNC_NO_COPY)
1847 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1849 /* The page is locked, so this pointer must remain valid */
1850 zspage = get_zspage(page);
1851 pool = zspage->pool;
1854 * The pool's lock protects the race between zpage migration
1857 spin_lock(&pool->lock);
1858 class = zspage_class(pool, zspage);
1860 /* the migrate_write_lock protects zpage access via zs_map_object */
1861 migrate_write_lock(zspage);
1863 offset = get_first_obj_offset(page);
1864 s_addr = kmap_atomic(page);
1867 * Here, any user cannot access all objects in the zspage so let's move.
1869 d_addr = kmap_atomic(newpage);
1870 memcpy(d_addr, s_addr, PAGE_SIZE);
1871 kunmap_atomic(d_addr);
1873 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1874 addr += class->size) {
1875 if (obj_allocated(page, addr, &handle)) {
1877 old_obj = handle_to_obj(handle);
1878 obj_to_location(old_obj, &dummy, &obj_idx);
1879 new_obj = (unsigned long)location_to_obj(newpage,
1881 record_obj(handle, new_obj);
1884 kunmap_atomic(s_addr);
1886 replace_sub_page(class, zspage, newpage, page);
1888 * Since we complete the data copy and set up new zspage structure,
1889 * it's okay to release the pool's lock.
1891 spin_unlock(&pool->lock);
1892 dec_zspage_isolation(zspage);
1893 migrate_write_unlock(zspage);
1896 if (page_zone(newpage) != page_zone(page)) {
1897 dec_zone_page_state(page, NR_ZSPAGES);
1898 inc_zone_page_state(newpage, NR_ZSPAGES);
1904 return MIGRATEPAGE_SUCCESS;
1907 static void zs_page_putback(struct page *page)
1909 struct zspage *zspage;
1911 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1913 zspage = get_zspage(page);
1914 migrate_write_lock(zspage);
1915 dec_zspage_isolation(zspage);
1916 migrate_write_unlock(zspage);
1919 static const struct movable_operations zsmalloc_mops = {
1920 .isolate_page = zs_page_isolate,
1921 .migrate_page = zs_page_migrate,
1922 .putback_page = zs_page_putback,
1926 * Caller should hold page_lock of all pages in the zspage
1927 * In here, we cannot use zspage meta data.
1929 static void async_free_zspage(struct work_struct *work)
1932 struct size_class *class;
1933 unsigned int class_idx;
1935 struct zspage *zspage, *tmp;
1936 LIST_HEAD(free_pages);
1937 struct zs_pool *pool = container_of(work, struct zs_pool,
1940 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1941 class = pool->size_class[i];
1942 if (class->index != i)
1945 spin_lock(&pool->lock);
1946 list_splice_init(&class->fullness_list[ZS_INUSE_RATIO_0],
1948 spin_unlock(&pool->lock);
1951 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1952 list_del(&zspage->list);
1953 lock_zspage(zspage);
1955 get_zspage_mapping(zspage, &class_idx, &fullness);
1956 VM_BUG_ON(fullness != ZS_INUSE_RATIO_0);
1957 class = pool->size_class[class_idx];
1958 spin_lock(&pool->lock);
1959 __free_zspage(pool, class, zspage);
1960 spin_unlock(&pool->lock);
1964 static void kick_deferred_free(struct zs_pool *pool)
1966 schedule_work(&pool->free_work);
1969 static void zs_flush_migration(struct zs_pool *pool)
1971 flush_work(&pool->free_work);
1974 static void init_deferred_free(struct zs_pool *pool)
1976 INIT_WORK(&pool->free_work, async_free_zspage);
1979 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
1981 struct page *page = get_first_page(zspage);
1984 WARN_ON(!trylock_page(page));
1985 __SetPageMovable(page, &zsmalloc_mops);
1987 } while ((page = get_next_page(page)) != NULL);
1990 static inline void zs_flush_migration(struct zs_pool *pool) { }
1995 * Based on the number of unused allocated objects calculate
1996 * and return the number of pages that we can free.
1998 static unsigned long zs_can_compact(struct size_class *class)
2000 unsigned long obj_wasted;
2001 unsigned long obj_allocated = zs_stat_get(class, ZS_OBJS_ALLOCATED);
2002 unsigned long obj_used = zs_stat_get(class, ZS_OBJS_INUSE);
2004 if (obj_allocated <= obj_used)
2007 obj_wasted = obj_allocated - obj_used;
2008 obj_wasted /= class->objs_per_zspage;
2010 return obj_wasted * class->pages_per_zspage;
2013 static unsigned long __zs_compact(struct zs_pool *pool,
2014 struct size_class *class)
2016 struct zs_compact_control cc;
2017 struct zspage *src_zspage = NULL;
2018 struct zspage *dst_zspage = NULL;
2019 unsigned long pages_freed = 0;
2022 * protect the race between zpage migration and zs_free
2023 * as well as zpage allocation/free
2025 spin_lock(&pool->lock);
2026 while (zs_can_compact(class)) {
2030 dst_zspage = isolate_dst_zspage(class);
2033 migrate_write_lock(dst_zspage);
2034 cc.d_page = get_first_page(dst_zspage);
2037 src_zspage = isolate_src_zspage(class);
2041 migrate_write_lock_nested(src_zspage);
2044 cc.s_page = get_first_page(src_zspage);
2045 migrate_zspage(pool, class, &cc);
2046 fg = putback_zspage(class, src_zspage);
2047 migrate_write_unlock(src_zspage);
2049 if (fg == ZS_INUSE_RATIO_0) {
2050 free_zspage(pool, class, src_zspage);
2051 pages_freed += class->pages_per_zspage;
2055 if (get_fullness_group(class, dst_zspage) == ZS_INUSE_RATIO_100
2056 || spin_is_contended(&pool->lock)) {
2057 putback_zspage(class, dst_zspage);
2058 migrate_write_unlock(dst_zspage);
2061 spin_unlock(&pool->lock);
2063 spin_lock(&pool->lock);
2068 putback_zspage(class, src_zspage);
2069 migrate_write_unlock(src_zspage);
2073 putback_zspage(class, dst_zspage);
2074 migrate_write_unlock(dst_zspage);
2076 spin_unlock(&pool->lock);
2081 unsigned long zs_compact(struct zs_pool *pool)
2084 struct size_class *class;
2085 unsigned long pages_freed = 0;
2088 * Pool compaction is performed under pool->lock so it is basically
2089 * single-threaded. Having more than one thread in __zs_compact()
2090 * will increase pool->lock contention, which will impact other
2091 * zsmalloc operations that need pool->lock.
2093 if (atomic_xchg(&pool->compaction_in_progress, 1))
2096 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2097 class = pool->size_class[i];
2098 if (class->index != i)
2100 pages_freed += __zs_compact(pool, class);
2102 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2103 atomic_set(&pool->compaction_in_progress, 0);
2107 EXPORT_SYMBOL_GPL(zs_compact);
2109 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2111 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2113 EXPORT_SYMBOL_GPL(zs_pool_stats);
2115 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2116 struct shrink_control *sc)
2118 unsigned long pages_freed;
2119 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2123 * Compact classes and calculate compaction delta.
2124 * Can run concurrently with a manually triggered
2125 * (by user) compaction.
2127 pages_freed = zs_compact(pool);
2129 return pages_freed ? pages_freed : SHRINK_STOP;
2132 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2133 struct shrink_control *sc)
2136 struct size_class *class;
2137 unsigned long pages_to_free = 0;
2138 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2141 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2142 class = pool->size_class[i];
2143 if (class->index != i)
2146 pages_to_free += zs_can_compact(class);
2149 return pages_to_free;
2152 static void zs_unregister_shrinker(struct zs_pool *pool)
2154 unregister_shrinker(&pool->shrinker);
2157 static int zs_register_shrinker(struct zs_pool *pool)
2159 pool->shrinker.scan_objects = zs_shrinker_scan;
2160 pool->shrinker.count_objects = zs_shrinker_count;
2161 pool->shrinker.batch = 0;
2162 pool->shrinker.seeks = DEFAULT_SEEKS;
2164 return register_shrinker(&pool->shrinker, "mm-zspool:%s",
2168 static int calculate_zspage_chain_size(int class_size)
2170 int i, min_waste = INT_MAX;
2173 if (is_power_of_2(class_size))
2176 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
2179 waste = (i * PAGE_SIZE) % class_size;
2180 if (waste < min_waste) {
2190 * zs_create_pool - Creates an allocation pool to work from.
2191 * @name: pool name to be created
2193 * This function must be called before anything when using
2194 * the zsmalloc allocator.
2196 * On success, a pointer to the newly created pool is returned,
2199 struct zs_pool *zs_create_pool(const char *name)
2202 struct zs_pool *pool;
2203 struct size_class *prev_class = NULL;
2205 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2209 init_deferred_free(pool);
2210 spin_lock_init(&pool->lock);
2211 atomic_set(&pool->compaction_in_progress, 0);
2213 pool->name = kstrdup(name, GFP_KERNEL);
2217 if (create_cache(pool))
2221 * Iterate reversely, because, size of size_class that we want to use
2222 * for merging should be larger or equal to current size.
2224 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2226 int pages_per_zspage;
2227 int objs_per_zspage;
2228 struct size_class *class;
2231 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2232 if (size > ZS_MAX_ALLOC_SIZE)
2233 size = ZS_MAX_ALLOC_SIZE;
2234 pages_per_zspage = calculate_zspage_chain_size(size);
2235 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2238 * We iterate from biggest down to smallest classes,
2239 * so huge_class_size holds the size of the first huge
2240 * class. Any object bigger than or equal to that will
2241 * endup in the huge class.
2243 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2245 huge_class_size = size;
2247 * The object uses ZS_HANDLE_SIZE bytes to store the
2248 * handle. We need to subtract it, because zs_malloc()
2249 * unconditionally adds handle size before it performs
2250 * size class search - so object may be smaller than
2251 * huge class size, yet it still can end up in the huge
2252 * class because it grows by ZS_HANDLE_SIZE extra bytes
2253 * right before class lookup.
2255 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2259 * size_class is used for normal zsmalloc operation such
2260 * as alloc/free for that size. Although it is natural that we
2261 * have one size_class for each size, there is a chance that we
2262 * can get more memory utilization if we use one size_class for
2263 * many different sizes whose size_class have same
2264 * characteristics. So, we makes size_class point to
2265 * previous size_class if possible.
2268 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2269 pool->size_class[i] = prev_class;
2274 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2280 class->pages_per_zspage = pages_per_zspage;
2281 class->objs_per_zspage = objs_per_zspage;
2282 pool->size_class[i] = class;
2284 fullness = ZS_INUSE_RATIO_0;
2285 while (fullness < NR_FULLNESS_GROUPS) {
2286 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2293 /* debug only, don't abort if it fails */
2294 zs_pool_stat_create(pool, name);
2297 * Not critical since shrinker is only used to trigger internal
2298 * defragmentation of the pool which is pretty optional thing. If
2299 * registration fails we still can use the pool normally and user can
2300 * trigger compaction manually. Thus, ignore return code.
2302 zs_register_shrinker(pool);
2307 zs_destroy_pool(pool);
2310 EXPORT_SYMBOL_GPL(zs_create_pool);
2312 void zs_destroy_pool(struct zs_pool *pool)
2316 zs_unregister_shrinker(pool);
2317 zs_flush_migration(pool);
2318 zs_pool_stat_destroy(pool);
2320 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2322 struct size_class *class = pool->size_class[i];
2327 if (class->index != i)
2330 for (fg = ZS_INUSE_RATIO_0; fg < NR_FULLNESS_GROUPS; fg++) {
2331 if (list_empty(&class->fullness_list[fg]))
2334 pr_err("Class-%d fullness group %d is not empty\n",
2340 destroy_cache(pool);
2344 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2346 static int __init zs_init(void)
2350 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2351 zs_cpu_prepare, zs_cpu_dead);
2356 zpool_register_driver(&zs_zpool_driver);
2367 static void __exit zs_exit(void)
2370 zpool_unregister_driver(&zs_zpool_driver);
2372 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2377 module_init(zs_init);
2378 module_exit(zs_exit);
2380 MODULE_LICENSE("Dual BSD/GPL");
2381 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");