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: PG_zsmalloc, lower 16 bit locate the first object
24 * offset in a subpage of a zspage
26 * Usage of struct page flags:
27 * PG_private: identifies the first component page
28 * PG_owner_priv_1: identifies the huge component page
32 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
42 #include <linux/module.h>
43 #include <linux/kernel.h>
44 #include <linux/sched.h>
45 #include <linux/bitops.h>
46 #include <linux/errno.h>
47 #include <linux/highmem.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/pgtable.h>
51 #include <asm/tlbflush.h>
52 #include <linux/cpumask.h>
53 #include <linux/cpu.h>
54 #include <linux/vmalloc.h>
55 #include <linux/preempt.h>
56 #include <linux/spinlock.h>
57 #include <linux/sprintf.h>
58 #include <linux/shrinker.h>
59 #include <linux/types.h>
60 #include <linux/debugfs.h>
61 #include <linux/zsmalloc.h>
62 #include <linux/zpool.h>
63 #include <linux/migrate.h>
64 #include <linux/wait.h>
65 #include <linux/pagemap.h>
67 #include <linux/local_lock.h>
69 #define ZSPAGE_MAGIC 0x58
72 * This must be power of 2 and greater than or equal to sizeof(link_free).
73 * These two conditions ensure that any 'struct link_free' itself doesn't
74 * span more than 1 page which avoids complex case of mapping 2 pages simply
75 * to restore link_free pointer values.
79 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
82 * Object location (<PFN>, <obj_idx>) is encoded as
83 * a single (unsigned long) handle value.
85 * Note that object index <obj_idx> starts from 0.
87 * This is made more complicated by various memory models and PAE.
90 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
91 #ifdef MAX_PHYSMEM_BITS
92 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
95 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
98 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
102 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
105 * Head in allocated object should have OBJ_ALLOCATED_TAG
106 * to identify the object was allocated or not.
107 * It's okay to add the status bit in the least bit because
108 * header keeps handle which is 4byte-aligned address so we
109 * have room for two bit at least.
111 #define OBJ_ALLOCATED_TAG 1
113 #define OBJ_TAG_BITS 1
114 #define OBJ_TAG_MASK OBJ_ALLOCATED_TAG
116 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
117 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
120 #define FULLNESS_BITS 4
122 #define MAGIC_VAL_BITS 8
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;
187 struct list_head fullness_list[NR_FULLNESS_GROUPS];
189 * Size of objects stored in this class. Must be multiple
194 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
195 int pages_per_zspage;
198 struct zs_size_stat stats;
202 * Placed within free objects to form a singly linked list.
203 * For every zspage, zspage->freeobj gives head of this list.
205 * This must be power of 2 and less than or equal to ZS_ALIGN
211 * It's valid for non-allocated object
215 * Handle of allocated object.
217 unsigned long handle;
224 struct size_class *size_class[ZS_SIZE_CLASSES];
225 struct kmem_cache *handle_cachep;
226 struct kmem_cache *zspage_cachep;
228 atomic_long_t pages_allocated;
230 struct zs_pool_stats stats;
232 /* Compact classes */
233 struct shrinker *shrinker;
235 #ifdef CONFIG_ZSMALLOC_STAT
236 struct dentry *stat_dentry;
238 #ifdef CONFIG_COMPACTION
239 struct work_struct free_work;
241 /* protect page/zspage migration */
242 rwlock_t migrate_lock;
243 atomic_t compaction_in_progress;
248 unsigned int huge:HUGE_BITS;
249 unsigned int fullness:FULLNESS_BITS;
250 unsigned int class:CLASS_BITS + 1;
251 unsigned int magic:MAGIC_VAL_BITS;
254 unsigned int freeobj;
255 struct page *first_page;
256 struct list_head list; /* fullness list */
257 struct zs_pool *pool;
261 struct mapping_area {
263 char *vm_buf; /* copy buffer for objects that span pages */
264 char *vm_addr; /* address of kmap_atomic()'ed pages */
265 enum zs_mapmode vm_mm; /* mapping mode */
268 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
269 static void SetZsHugePage(struct zspage *zspage)
274 static bool ZsHugePage(struct zspage *zspage)
279 static void migrate_lock_init(struct zspage *zspage);
280 static void migrate_read_lock(struct zspage *zspage);
281 static void migrate_read_unlock(struct zspage *zspage);
282 static void migrate_write_lock(struct zspage *zspage);
283 static void migrate_write_unlock(struct zspage *zspage);
285 #ifdef CONFIG_COMPACTION
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 kick_deferred_free(struct zs_pool *pool) {}
291 static void init_deferred_free(struct zs_pool *pool) {}
292 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
295 static int create_cache(struct zs_pool *pool)
299 name = kasprintf(GFP_KERNEL, "zs_handle-%s", pool->name);
302 pool->handle_cachep = kmem_cache_create(name, ZS_HANDLE_SIZE,
305 if (!pool->handle_cachep)
308 name = kasprintf(GFP_KERNEL, "zspage-%s", pool->name);
311 pool->zspage_cachep = kmem_cache_create(name, sizeof(struct zspage),
314 if (!pool->zspage_cachep) {
315 kmem_cache_destroy(pool->handle_cachep);
316 pool->handle_cachep = NULL;
323 static void destroy_cache(struct zs_pool *pool)
325 kmem_cache_destroy(pool->handle_cachep);
326 kmem_cache_destroy(pool->zspage_cachep);
329 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
331 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
332 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
335 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
337 kmem_cache_free(pool->handle_cachep, (void *)handle);
340 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
342 return kmem_cache_zalloc(pool->zspage_cachep,
343 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
346 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
348 kmem_cache_free(pool->zspage_cachep, zspage);
351 /* class->lock(which owns the handle) synchronizes races */
352 static void record_obj(unsigned long handle, unsigned long obj)
354 *(unsigned long *)handle = obj;
361 static void *zs_zpool_create(const char *name, gfp_t gfp)
364 * Ignore global gfp flags: zs_malloc() may be invoked from
365 * different contexts and its caller must provide a valid
368 return zs_create_pool(name);
371 static void zs_zpool_destroy(void *pool)
373 zs_destroy_pool(pool);
376 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
377 unsigned long *handle)
379 *handle = zs_malloc(pool, size, gfp);
381 if (IS_ERR_VALUE(*handle))
382 return PTR_ERR((void *)*handle);
385 static void zs_zpool_free(void *pool, unsigned long handle)
387 zs_free(pool, handle);
390 static void *zs_zpool_map(void *pool, unsigned long handle,
391 enum zpool_mapmode mm)
393 enum zs_mapmode zs_mm;
408 return zs_map_object(pool, handle, zs_mm);
410 static void zs_zpool_unmap(void *pool, unsigned long handle)
412 zs_unmap_object(pool, handle);
415 static u64 zs_zpool_total_pages(void *pool)
417 return zs_get_total_pages(pool);
420 static struct zpool_driver zs_zpool_driver = {
422 .owner = THIS_MODULE,
423 .create = zs_zpool_create,
424 .destroy = zs_zpool_destroy,
425 .malloc_support_movable = true,
426 .malloc = zs_zpool_malloc,
427 .free = zs_zpool_free,
429 .unmap = zs_zpool_unmap,
430 .total_pages = zs_zpool_total_pages,
433 MODULE_ALIAS("zpool-zsmalloc");
434 #endif /* CONFIG_ZPOOL */
436 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
437 static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = {
438 .lock = INIT_LOCAL_LOCK(lock),
441 static __maybe_unused int is_first_page(struct page *page)
443 return PagePrivate(page);
446 /* Protected by class->lock */
447 static inline int get_zspage_inuse(struct zspage *zspage)
449 return zspage->inuse;
453 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
455 zspage->inuse += val;
458 static inline struct page *get_first_page(struct zspage *zspage)
460 struct page *first_page = zspage->first_page;
462 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
466 #define FIRST_OBJ_PAGE_TYPE_MASK 0xffff
468 static inline void reset_first_obj_offset(struct page *page)
470 VM_WARN_ON_ONCE(!PageZsmalloc(page));
471 page->page_type |= FIRST_OBJ_PAGE_TYPE_MASK;
474 static inline unsigned int get_first_obj_offset(struct page *page)
476 VM_WARN_ON_ONCE(!PageZsmalloc(page));
477 return page->page_type & FIRST_OBJ_PAGE_TYPE_MASK;
480 static inline void set_first_obj_offset(struct page *page, unsigned int offset)
482 /* With 16 bit available, we can support offsets into 64 KiB pages. */
483 BUILD_BUG_ON(PAGE_SIZE > SZ_64K);
484 VM_WARN_ON_ONCE(!PageZsmalloc(page));
485 VM_WARN_ON_ONCE(offset & ~FIRST_OBJ_PAGE_TYPE_MASK);
486 page->page_type &= ~FIRST_OBJ_PAGE_TYPE_MASK;
487 page->page_type |= offset & FIRST_OBJ_PAGE_TYPE_MASK;
490 static inline unsigned int get_freeobj(struct zspage *zspage)
492 return zspage->freeobj;
495 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
497 zspage->freeobj = obj;
500 static struct size_class *zspage_class(struct zs_pool *pool,
501 struct zspage *zspage)
503 return pool->size_class[zspage->class];
507 * zsmalloc divides the pool into various size classes where each
508 * class maintains a list of zspages where each zspage is divided
509 * into equal sized chunks. Each allocation falls into one of these
510 * classes depending on its size. This function returns index of the
511 * size class which has chunk size big enough to hold the given size.
513 static int get_size_class_index(int size)
517 if (likely(size > ZS_MIN_ALLOC_SIZE))
518 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
519 ZS_SIZE_CLASS_DELTA);
521 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
524 static inline void class_stat_add(struct size_class *class, int type,
527 class->stats.objs[type] += cnt;
530 static inline void class_stat_sub(struct size_class *class, int type,
533 class->stats.objs[type] -= cnt;
536 static inline unsigned long class_stat_read(struct size_class *class, int type)
538 return class->stats.objs[type];
541 #ifdef CONFIG_ZSMALLOC_STAT
543 static void __init zs_stat_init(void)
545 if (!debugfs_initialized()) {
546 pr_warn("debugfs not available, stat dir not created\n");
550 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
553 static void __exit zs_stat_exit(void)
555 debugfs_remove_recursive(zs_stat_root);
558 static unsigned long zs_can_compact(struct size_class *class);
560 static int zs_stats_size_show(struct seq_file *s, void *v)
563 struct zs_pool *pool = s->private;
564 struct size_class *class;
566 unsigned long obj_allocated, obj_used, pages_used, freeable;
567 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
568 unsigned long total_freeable = 0;
569 unsigned long inuse_totals[NR_FULLNESS_GROUPS] = {0, };
571 seq_printf(s, " %5s %5s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %9s %13s %10s %10s %16s %8s\n",
572 "class", "size", "10%", "20%", "30%", "40%",
573 "50%", "60%", "70%", "80%", "90%", "99%", "100%",
574 "obj_allocated", "obj_used", "pages_used",
575 "pages_per_zspage", "freeable");
577 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
579 class = pool->size_class[i];
581 if (class->index != i)
584 spin_lock(&class->lock);
586 seq_printf(s, " %5u %5u ", i, class->size);
587 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++) {
588 inuse_totals[fg] += class_stat_read(class, fg);
589 seq_printf(s, "%9lu ", class_stat_read(class, fg));
592 obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
593 obj_used = class_stat_read(class, ZS_OBJS_INUSE);
594 freeable = zs_can_compact(class);
595 spin_unlock(&class->lock);
597 objs_per_zspage = class->objs_per_zspage;
598 pages_used = obj_allocated / objs_per_zspage *
599 class->pages_per_zspage;
601 seq_printf(s, "%13lu %10lu %10lu %16d %8lu\n",
602 obj_allocated, obj_used, pages_used,
603 class->pages_per_zspage, freeable);
605 total_objs += obj_allocated;
606 total_used_objs += obj_used;
607 total_pages += pages_used;
608 total_freeable += freeable;
612 seq_printf(s, " %5s %5s ", "Total", "");
614 for (fg = ZS_INUSE_RATIO_10; fg < NR_FULLNESS_GROUPS; fg++)
615 seq_printf(s, "%9lu ", inuse_totals[fg]);
617 seq_printf(s, "%13lu %10lu %10lu %16s %8lu\n",
618 total_objs, total_used_objs, total_pages, "",
623 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
625 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
628 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
632 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
634 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
635 &zs_stats_size_fops);
638 static void zs_pool_stat_destroy(struct zs_pool *pool)
640 debugfs_remove_recursive(pool->stat_dentry);
643 #else /* CONFIG_ZSMALLOC_STAT */
644 static void __init zs_stat_init(void)
648 static void __exit zs_stat_exit(void)
652 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
656 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
663 * For each size class, zspages are divided into different groups
664 * depending on their usage ratio. This function returns fullness
665 * status of the given page.
667 static int get_fullness_group(struct size_class *class, struct zspage *zspage)
669 int inuse, objs_per_zspage, ratio;
671 inuse = get_zspage_inuse(zspage);
672 objs_per_zspage = class->objs_per_zspage;
675 return ZS_INUSE_RATIO_0;
676 if (inuse == objs_per_zspage)
677 return ZS_INUSE_RATIO_100;
679 ratio = 100 * inuse / objs_per_zspage;
681 * Take integer division into consideration: a page with one inuse
682 * object out of 127 possible, will end up having 0 usage ratio,
683 * which is wrong as it belongs in ZS_INUSE_RATIO_10 fullness group.
685 return ratio / 10 + 1;
689 * Each size class maintains various freelists and zspages are assigned
690 * to one of these freelists based on the number of live objects they
691 * have. This functions inserts the given zspage into the freelist
692 * identified by <class, fullness_group>.
694 static void insert_zspage(struct size_class *class,
695 struct zspage *zspage,
698 class_stat_add(class, fullness, 1);
699 list_add(&zspage->list, &class->fullness_list[fullness]);
700 zspage->fullness = fullness;
704 * This function removes the given zspage from the freelist identified
705 * by <class, fullness_group>.
707 static void remove_zspage(struct size_class *class, struct zspage *zspage)
709 int fullness = zspage->fullness;
711 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
713 list_del_init(&zspage->list);
714 class_stat_sub(class, fullness, 1);
718 * Each size class maintains zspages in different fullness groups depending
719 * on the number of live objects they contain. When allocating or freeing
720 * objects, the fullness status of the page can change, for instance, from
721 * INUSE_RATIO_80 to INUSE_RATIO_70 when freeing an object. This function
722 * checks if such a status change has occurred for the given page and
723 * accordingly moves the page from the list of the old fullness group to that
724 * of the new fullness group.
726 static int fix_fullness_group(struct size_class *class, struct zspage *zspage)
730 newfg = get_fullness_group(class, zspage);
731 if (newfg == zspage->fullness)
734 remove_zspage(class, zspage);
735 insert_zspage(class, zspage, newfg);
740 static struct zspage *get_zspage(struct page *page)
742 struct zspage *zspage = (struct zspage *)page_private(page);
744 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
748 static struct page *get_next_page(struct page *page)
750 struct zspage *zspage = get_zspage(page);
752 if (unlikely(ZsHugePage(zspage)))
755 return (struct page *)page->index;
759 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
760 * @obj: the encoded object value
761 * @page: page object resides in zspage
762 * @obj_idx: object index
764 static void obj_to_location(unsigned long obj, struct page **page,
765 unsigned int *obj_idx)
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 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
777 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
778 * @page: page object resides in zspage
779 * @obj_idx: object index
781 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
785 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
786 obj |= obj_idx & OBJ_INDEX_MASK;
791 static unsigned long handle_to_obj(unsigned long handle)
793 return *(unsigned long *)handle;
796 static inline bool obj_allocated(struct page *page, void *obj,
797 unsigned long *phandle)
799 unsigned long handle;
800 struct zspage *zspage = get_zspage(page);
802 if (unlikely(ZsHugePage(zspage))) {
803 VM_BUG_ON_PAGE(!is_first_page(page), page);
804 handle = page->index;
806 handle = *(unsigned long *)obj;
808 if (!(handle & OBJ_ALLOCATED_TAG))
811 /* Clear all tags before returning the handle */
812 *phandle = handle & ~OBJ_TAG_MASK;
816 static void reset_page(struct page *page)
818 __ClearPageMovable(page);
819 ClearPagePrivate(page);
820 set_page_private(page, 0);
822 reset_first_obj_offset(page);
823 __ClearPageZsmalloc(page);
826 static int trylock_zspage(struct zspage *zspage)
828 struct page *cursor, *fail;
830 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
831 get_next_page(cursor)) {
832 if (!trylock_page(cursor)) {
840 for (cursor = get_first_page(zspage); cursor != fail; cursor =
841 get_next_page(cursor))
847 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
848 struct zspage *zspage)
850 struct page *page, *next;
852 assert_spin_locked(&class->lock);
854 VM_BUG_ON(get_zspage_inuse(zspage));
855 VM_BUG_ON(zspage->fullness != ZS_INUSE_RATIO_0);
857 next = page = get_first_page(zspage);
859 VM_BUG_ON_PAGE(!PageLocked(page), page);
860 next = get_next_page(page);
863 dec_zone_page_state(page, NR_ZSPAGES);
866 } while (page != NULL);
868 cache_free_zspage(pool, zspage);
870 class_stat_sub(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
871 atomic_long_sub(class->pages_per_zspage, &pool->pages_allocated);
874 static void free_zspage(struct zs_pool *pool, struct size_class *class,
875 struct zspage *zspage)
877 VM_BUG_ON(get_zspage_inuse(zspage));
878 VM_BUG_ON(list_empty(&zspage->list));
881 * Since zs_free couldn't be sleepable, this function cannot call
882 * lock_page. The page locks trylock_zspage got will be released
885 if (!trylock_zspage(zspage)) {
886 kick_deferred_free(pool);
890 remove_zspage(class, zspage);
891 __free_zspage(pool, class, zspage);
894 /* Initialize a newly allocated zspage */
895 static void init_zspage(struct size_class *class, struct zspage *zspage)
897 unsigned int freeobj = 1;
898 unsigned long off = 0;
899 struct page *page = get_first_page(zspage);
902 struct page *next_page;
903 struct link_free *link;
906 set_first_obj_offset(page, off);
908 vaddr = kmap_atomic(page);
909 link = (struct link_free *)vaddr + off / sizeof(*link);
911 while ((off += class->size) < PAGE_SIZE) {
912 link->next = freeobj++ << OBJ_TAG_BITS;
913 link += class->size / sizeof(*link);
917 * We now come to the last (full or partial) object on this
918 * page, which must point to the first object on the next
921 next_page = get_next_page(page);
923 link->next = freeobj++ << OBJ_TAG_BITS;
926 * Reset OBJ_TAG_BITS bit to last link to tell
927 * whether it's allocated object or not.
929 link->next = -1UL << OBJ_TAG_BITS;
931 kunmap_atomic(vaddr);
936 set_freeobj(zspage, 0);
939 static void create_page_chain(struct size_class *class, struct zspage *zspage,
940 struct page *pages[])
944 struct page *prev_page = NULL;
945 int nr_pages = class->pages_per_zspage;
948 * Allocate individual pages and link them together as:
949 * 1. all pages are linked together using page->index
950 * 2. each sub-page point to zspage using page->private
952 * we set PG_private to identify the first page (i.e. no other sub-page
953 * has this flag set).
955 for (i = 0; i < nr_pages; i++) {
957 set_page_private(page, (unsigned long)zspage);
960 zspage->first_page = page;
961 SetPagePrivate(page);
962 if (unlikely(class->objs_per_zspage == 1 &&
963 class->pages_per_zspage == 1))
964 SetZsHugePage(zspage);
966 prev_page->index = (unsigned long)page;
973 * Allocate a zspage for the given size class
975 static struct zspage *alloc_zspage(struct zs_pool *pool,
976 struct size_class *class,
980 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
981 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
986 zspage->magic = ZSPAGE_MAGIC;
987 migrate_lock_init(zspage);
989 for (i = 0; i < class->pages_per_zspage; i++) {
992 page = alloc_page(gfp);
995 dec_zone_page_state(pages[i], NR_ZSPAGES);
996 __ClearPageZsmalloc(pages[i]);
997 __free_page(pages[i]);
999 cache_free_zspage(pool, zspage);
1002 __SetPageZsmalloc(page);
1004 inc_zone_page_state(page, NR_ZSPAGES);
1008 create_page_chain(class, zspage, pages);
1009 init_zspage(class, zspage);
1010 zspage->pool = pool;
1011 zspage->class = class->index;
1016 static struct zspage *find_get_zspage(struct size_class *class)
1019 struct zspage *zspage;
1021 for (i = ZS_INUSE_RATIO_99; i >= ZS_INUSE_RATIO_0; i--) {
1022 zspage = list_first_entry_or_null(&class->fullness_list[i],
1023 struct zspage, list);
1031 static inline int __zs_cpu_up(struct mapping_area *area)
1034 * Make sure we don't leak memory if a cpu UP notification
1035 * and zs_init() race and both call zs_cpu_up() on the same cpu
1039 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1045 static inline void __zs_cpu_down(struct mapping_area *area)
1047 kfree(area->vm_buf);
1048 area->vm_buf = NULL;
1051 static void *__zs_map_object(struct mapping_area *area,
1052 struct page *pages[2], int off, int size)
1056 char *buf = area->vm_buf;
1058 /* disable page faults to match kmap_atomic() return conditions */
1059 pagefault_disable();
1061 /* no read fastpath */
1062 if (area->vm_mm == ZS_MM_WO)
1065 sizes[0] = PAGE_SIZE - off;
1066 sizes[1] = size - sizes[0];
1068 /* copy object to per-cpu buffer */
1069 addr = kmap_atomic(pages[0]);
1070 memcpy(buf, addr + off, sizes[0]);
1071 kunmap_atomic(addr);
1072 addr = kmap_atomic(pages[1]);
1073 memcpy(buf + sizes[0], addr, sizes[1]);
1074 kunmap_atomic(addr);
1076 return area->vm_buf;
1079 static void __zs_unmap_object(struct mapping_area *area,
1080 struct page *pages[2], int off, int size)
1086 /* no write fastpath */
1087 if (area->vm_mm == ZS_MM_RO)
1091 buf = buf + ZS_HANDLE_SIZE;
1092 size -= ZS_HANDLE_SIZE;
1093 off += ZS_HANDLE_SIZE;
1095 sizes[0] = PAGE_SIZE - off;
1096 sizes[1] = size - sizes[0];
1098 /* copy per-cpu buffer to object */
1099 addr = kmap_atomic(pages[0]);
1100 memcpy(addr + off, buf, sizes[0]);
1101 kunmap_atomic(addr);
1102 addr = kmap_atomic(pages[1]);
1103 memcpy(addr, buf + sizes[0], sizes[1]);
1104 kunmap_atomic(addr);
1107 /* enable page faults to match kunmap_atomic() return conditions */
1111 static int zs_cpu_prepare(unsigned int cpu)
1113 struct mapping_area *area;
1115 area = &per_cpu(zs_map_area, cpu);
1116 return __zs_cpu_up(area);
1119 static int zs_cpu_dead(unsigned int cpu)
1121 struct mapping_area *area;
1123 area = &per_cpu(zs_map_area, cpu);
1124 __zs_cpu_down(area);
1128 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1129 int objs_per_zspage)
1131 if (prev->pages_per_zspage == pages_per_zspage &&
1132 prev->objs_per_zspage == objs_per_zspage)
1138 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1140 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1143 static bool zspage_empty(struct zspage *zspage)
1145 return get_zspage_inuse(zspage) == 0;
1149 * zs_lookup_class_index() - Returns index of the zsmalloc &size_class
1150 * that hold objects of the provided size.
1151 * @pool: zsmalloc pool to use
1152 * @size: object size
1154 * Context: Any context.
1156 * Return: the index of the zsmalloc &size_class that hold objects of the
1159 unsigned int zs_lookup_class_index(struct zs_pool *pool, unsigned int size)
1161 struct size_class *class;
1163 class = pool->size_class[get_size_class_index(size)];
1165 return class->index;
1167 EXPORT_SYMBOL_GPL(zs_lookup_class_index);
1169 unsigned long zs_get_total_pages(struct zs_pool *pool)
1171 return atomic_long_read(&pool->pages_allocated);
1173 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1176 * zs_map_object - get address of allocated object from handle.
1177 * @pool: pool from which the object was allocated
1178 * @handle: handle returned from zs_malloc
1179 * @mm: mapping mode to use
1181 * Before using an object allocated from zs_malloc, it must be mapped using
1182 * this function. When done with the object, it must be unmapped using
1185 * Only one object can be mapped per cpu at a time. There is no protection
1186 * against nested mappings.
1188 * This function returns with preemption and page faults disabled.
1190 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1193 struct zspage *zspage;
1195 unsigned long obj, off;
1196 unsigned int obj_idx;
1198 struct size_class *class;
1199 struct mapping_area *area;
1200 struct page *pages[2];
1204 * Because we use per-cpu mapping areas shared among the
1205 * pools/users, we can't allow mapping in interrupt context
1206 * because it can corrupt another users mappings.
1208 BUG_ON(in_interrupt());
1210 /* It guarantees it can get zspage from handle safely */
1211 read_lock(&pool->migrate_lock);
1212 obj = handle_to_obj(handle);
1213 obj_to_location(obj, &page, &obj_idx);
1214 zspage = get_zspage(page);
1217 * migration cannot move any zpages in this zspage. Here, class->lock
1218 * is too heavy since callers would take some time until they calls
1219 * zs_unmap_object API so delegate the locking from class to zspage
1220 * which is smaller granularity.
1222 migrate_read_lock(zspage);
1223 read_unlock(&pool->migrate_lock);
1225 class = zspage_class(pool, zspage);
1226 off = offset_in_page(class->size * obj_idx);
1228 local_lock(&zs_map_area.lock);
1229 area = this_cpu_ptr(&zs_map_area);
1231 if (off + class->size <= PAGE_SIZE) {
1232 /* this object is contained entirely within a page */
1233 area->vm_addr = kmap_atomic(page);
1234 ret = area->vm_addr + off;
1238 /* this object spans two pages */
1240 pages[1] = get_next_page(page);
1243 ret = __zs_map_object(area, pages, off, class->size);
1245 if (likely(!ZsHugePage(zspage)))
1246 ret += ZS_HANDLE_SIZE;
1250 EXPORT_SYMBOL_GPL(zs_map_object);
1252 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1254 struct zspage *zspage;
1256 unsigned long obj, off;
1257 unsigned int obj_idx;
1259 struct size_class *class;
1260 struct mapping_area *area;
1262 obj = handle_to_obj(handle);
1263 obj_to_location(obj, &page, &obj_idx);
1264 zspage = get_zspage(page);
1265 class = zspage_class(pool, zspage);
1266 off = offset_in_page(class->size * obj_idx);
1268 area = this_cpu_ptr(&zs_map_area);
1269 if (off + class->size <= PAGE_SIZE)
1270 kunmap_atomic(area->vm_addr);
1272 struct page *pages[2];
1275 pages[1] = get_next_page(page);
1278 __zs_unmap_object(area, pages, off, class->size);
1280 local_unlock(&zs_map_area.lock);
1282 migrate_read_unlock(zspage);
1284 EXPORT_SYMBOL_GPL(zs_unmap_object);
1287 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1288 * zsmalloc &size_class.
1289 * @pool: zsmalloc pool to use
1291 * The function returns the size of the first huge class - any object of equal
1292 * or bigger size will be stored in zspage consisting of a single physical
1295 * Context: Any context.
1297 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1299 size_t zs_huge_class_size(struct zs_pool *pool)
1301 return huge_class_size;
1303 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1305 static unsigned long obj_malloc(struct zs_pool *pool,
1306 struct zspage *zspage, unsigned long handle)
1308 int i, nr_page, offset;
1310 struct link_free *link;
1311 struct size_class *class;
1313 struct page *m_page;
1314 unsigned long m_offset;
1317 class = pool->size_class[zspage->class];
1318 obj = get_freeobj(zspage);
1320 offset = obj * class->size;
1321 nr_page = offset >> PAGE_SHIFT;
1322 m_offset = offset_in_page(offset);
1323 m_page = get_first_page(zspage);
1325 for (i = 0; i < nr_page; i++)
1326 m_page = get_next_page(m_page);
1328 vaddr = kmap_atomic(m_page);
1329 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1330 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1331 if (likely(!ZsHugePage(zspage)))
1332 /* record handle in the header of allocated chunk */
1333 link->handle = handle | OBJ_ALLOCATED_TAG;
1335 /* record handle to page->index */
1336 zspage->first_page->index = handle | OBJ_ALLOCATED_TAG;
1338 kunmap_atomic(vaddr);
1339 mod_zspage_inuse(zspage, 1);
1341 obj = location_to_obj(m_page, obj);
1342 record_obj(handle, obj);
1349 * zs_malloc - Allocate block of given size from pool.
1350 * @pool: pool to allocate from
1351 * @size: size of block to allocate
1352 * @gfp: gfp flags when allocating object
1354 * On success, handle to the allocated object is returned,
1355 * otherwise an ERR_PTR().
1356 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1358 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1360 unsigned long handle;
1361 struct size_class *class;
1363 struct zspage *zspage;
1365 if (unlikely(!size))
1366 return (unsigned long)ERR_PTR(-EINVAL);
1368 if (unlikely(size > ZS_MAX_ALLOC_SIZE))
1369 return (unsigned long)ERR_PTR(-ENOSPC);
1371 handle = cache_alloc_handle(pool, gfp);
1373 return (unsigned long)ERR_PTR(-ENOMEM);
1375 /* extra space in chunk to keep the handle */
1376 size += ZS_HANDLE_SIZE;
1377 class = pool->size_class[get_size_class_index(size)];
1379 /* class->lock effectively protects the zpage migration */
1380 spin_lock(&class->lock);
1381 zspage = find_get_zspage(class);
1382 if (likely(zspage)) {
1383 obj_malloc(pool, zspage, handle);
1384 /* Now move the zspage to another fullness group, if required */
1385 fix_fullness_group(class, zspage);
1386 class_stat_add(class, ZS_OBJS_INUSE, 1);
1391 spin_unlock(&class->lock);
1393 zspage = alloc_zspage(pool, class, gfp);
1395 cache_free_handle(pool, handle);
1396 return (unsigned long)ERR_PTR(-ENOMEM);
1399 spin_lock(&class->lock);
1400 obj_malloc(pool, zspage, handle);
1401 newfg = get_fullness_group(class, zspage);
1402 insert_zspage(class, zspage, newfg);
1403 atomic_long_add(class->pages_per_zspage, &pool->pages_allocated);
1404 class_stat_add(class, ZS_OBJS_ALLOCATED, class->objs_per_zspage);
1405 class_stat_add(class, ZS_OBJS_INUSE, 1);
1407 /* We completely set up zspage so mark them as movable */
1408 SetZsPageMovable(pool, zspage);
1410 spin_unlock(&class->lock);
1414 EXPORT_SYMBOL_GPL(zs_malloc);
1416 static void obj_free(int class_size, unsigned long obj)
1418 struct link_free *link;
1419 struct zspage *zspage;
1420 struct page *f_page;
1421 unsigned long f_offset;
1422 unsigned int f_objidx;
1425 obj_to_location(obj, &f_page, &f_objidx);
1426 f_offset = offset_in_page(class_size * f_objidx);
1427 zspage = get_zspage(f_page);
1429 vaddr = kmap_atomic(f_page);
1430 link = (struct link_free *)(vaddr + f_offset);
1432 /* Insert this object in containing zspage's freelist */
1433 if (likely(!ZsHugePage(zspage)))
1434 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1437 set_freeobj(zspage, f_objidx);
1439 kunmap_atomic(vaddr);
1440 mod_zspage_inuse(zspage, -1);
1443 void zs_free(struct zs_pool *pool, unsigned long handle)
1445 struct zspage *zspage;
1446 struct page *f_page;
1448 struct size_class *class;
1451 if (IS_ERR_OR_NULL((void *)handle))
1455 * The pool->migrate_lock protects the race with zpage's migration
1456 * so it's safe to get the page from handle.
1458 read_lock(&pool->migrate_lock);
1459 obj = handle_to_obj(handle);
1460 obj_to_page(obj, &f_page);
1461 zspage = get_zspage(f_page);
1462 class = zspage_class(pool, zspage);
1463 spin_lock(&class->lock);
1464 read_unlock(&pool->migrate_lock);
1466 class_stat_sub(class, ZS_OBJS_INUSE, 1);
1467 obj_free(class->size, obj);
1469 fullness = fix_fullness_group(class, zspage);
1470 if (fullness == ZS_INUSE_RATIO_0)
1471 free_zspage(pool, class, zspage);
1473 spin_unlock(&class->lock);
1474 cache_free_handle(pool, handle);
1476 EXPORT_SYMBOL_GPL(zs_free);
1478 static void zs_object_copy(struct size_class *class, unsigned long dst,
1481 struct page *s_page, *d_page;
1482 unsigned int s_objidx, d_objidx;
1483 unsigned long s_off, d_off;
1484 void *s_addr, *d_addr;
1485 int s_size, d_size, size;
1488 s_size = d_size = class->size;
1490 obj_to_location(src, &s_page, &s_objidx);
1491 obj_to_location(dst, &d_page, &d_objidx);
1493 s_off = offset_in_page(class->size * s_objidx);
1494 d_off = offset_in_page(class->size * d_objidx);
1496 if (s_off + class->size > PAGE_SIZE)
1497 s_size = PAGE_SIZE - s_off;
1499 if (d_off + class->size > PAGE_SIZE)
1500 d_size = PAGE_SIZE - d_off;
1502 s_addr = kmap_atomic(s_page);
1503 d_addr = kmap_atomic(d_page);
1506 size = min(s_size, d_size);
1507 memcpy(d_addr + d_off, s_addr + s_off, size);
1510 if (written == class->size)
1519 * Calling kunmap_atomic(d_addr) is necessary. kunmap_atomic()
1520 * calls must occurs in reverse order of calls to kmap_atomic().
1521 * So, to call kunmap_atomic(s_addr) we should first call
1522 * kunmap_atomic(d_addr). For more details see
1523 * Documentation/mm/highmem.rst.
1525 if (s_off >= PAGE_SIZE) {
1526 kunmap_atomic(d_addr);
1527 kunmap_atomic(s_addr);
1528 s_page = get_next_page(s_page);
1529 s_addr = kmap_atomic(s_page);
1530 d_addr = kmap_atomic(d_page);
1531 s_size = class->size - written;
1535 if (d_off >= PAGE_SIZE) {
1536 kunmap_atomic(d_addr);
1537 d_page = get_next_page(d_page);
1538 d_addr = kmap_atomic(d_page);
1539 d_size = class->size - written;
1544 kunmap_atomic(d_addr);
1545 kunmap_atomic(s_addr);
1549 * Find alloced object in zspage from index object and
1552 static unsigned long find_alloced_obj(struct size_class *class,
1553 struct page *page, int *obj_idx)
1555 unsigned int offset;
1556 int index = *obj_idx;
1557 unsigned long handle = 0;
1558 void *addr = kmap_atomic(page);
1560 offset = get_first_obj_offset(page);
1561 offset += class->size * index;
1563 while (offset < PAGE_SIZE) {
1564 if (obj_allocated(page, addr + offset, &handle))
1567 offset += class->size;
1571 kunmap_atomic(addr);
1578 static void migrate_zspage(struct zs_pool *pool, struct zspage *src_zspage,
1579 struct zspage *dst_zspage)
1581 unsigned long used_obj, free_obj;
1582 unsigned long handle;
1584 struct page *s_page = get_first_page(src_zspage);
1585 struct size_class *class = pool->size_class[src_zspage->class];
1588 handle = find_alloced_obj(class, s_page, &obj_idx);
1590 s_page = get_next_page(s_page);
1597 used_obj = handle_to_obj(handle);
1598 free_obj = obj_malloc(pool, dst_zspage, handle);
1599 zs_object_copy(class, free_obj, used_obj);
1601 obj_free(class->size, used_obj);
1603 /* Stop if there is no more space */
1604 if (zspage_full(class, dst_zspage))
1607 /* Stop if there are no more objects to migrate */
1608 if (zspage_empty(src_zspage))
1613 static struct zspage *isolate_src_zspage(struct size_class *class)
1615 struct zspage *zspage;
1618 for (fg = ZS_INUSE_RATIO_10; fg <= ZS_INUSE_RATIO_99; fg++) {
1619 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1620 struct zspage, list);
1622 remove_zspage(class, zspage);
1630 static struct zspage *isolate_dst_zspage(struct size_class *class)
1632 struct zspage *zspage;
1635 for (fg = ZS_INUSE_RATIO_99; fg >= ZS_INUSE_RATIO_10; fg--) {
1636 zspage = list_first_entry_or_null(&class->fullness_list[fg],
1637 struct zspage, list);
1639 remove_zspage(class, zspage);
1648 * putback_zspage - add @zspage into right class's fullness list
1649 * @class: destination class
1650 * @zspage: target page
1652 * Return @zspage's fullness status
1654 static int putback_zspage(struct size_class *class, struct zspage *zspage)
1658 fullness = get_fullness_group(class, zspage);
1659 insert_zspage(class, zspage, fullness);
1664 #ifdef CONFIG_COMPACTION
1666 * To prevent zspage destroy during migration, zspage freeing should
1667 * hold locks of all pages in the zspage.
1669 static void lock_zspage(struct zspage *zspage)
1671 struct page *curr_page, *page;
1674 * Pages we haven't locked yet can be migrated off the list while we're
1675 * trying to lock them, so we need to be careful and only attempt to
1676 * lock each page under migrate_read_lock(). Otherwise, the page we lock
1677 * may no longer belong to the zspage. This means that we may wait for
1678 * the wrong page to unlock, so we must take a reference to the page
1679 * prior to waiting for it to unlock outside migrate_read_lock().
1682 migrate_read_lock(zspage);
1683 page = get_first_page(zspage);
1684 if (trylock_page(page))
1687 migrate_read_unlock(zspage);
1688 wait_on_page_locked(page);
1693 while ((page = get_next_page(curr_page))) {
1694 if (trylock_page(page)) {
1698 migrate_read_unlock(zspage);
1699 wait_on_page_locked(page);
1701 migrate_read_lock(zspage);
1704 migrate_read_unlock(zspage);
1706 #endif /* CONFIG_COMPACTION */
1708 static void migrate_lock_init(struct zspage *zspage)
1710 rwlock_init(&zspage->lock);
1713 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1715 read_lock(&zspage->lock);
1718 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1720 read_unlock(&zspage->lock);
1723 static void migrate_write_lock(struct zspage *zspage)
1725 write_lock(&zspage->lock);
1728 static void migrate_write_unlock(struct zspage *zspage)
1730 write_unlock(&zspage->lock);
1733 #ifdef CONFIG_COMPACTION
1735 static const struct movable_operations zsmalloc_mops;
1737 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1738 struct page *newpage, struct page *oldpage)
1741 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1744 page = get_first_page(zspage);
1746 if (page == oldpage)
1747 pages[idx] = newpage;
1751 } while ((page = get_next_page(page)) != NULL);
1753 create_page_chain(class, zspage, pages);
1754 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1755 if (unlikely(ZsHugePage(zspage)))
1756 newpage->index = oldpage->index;
1757 __SetPageMovable(newpage, &zsmalloc_mops);
1760 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1763 * Page is locked so zspage couldn't be destroyed. For detail, look at
1764 * lock_zspage in free_zspage.
1766 VM_BUG_ON_PAGE(PageIsolated(page), page);
1771 static int zs_page_migrate(struct page *newpage, struct page *page,
1772 enum migrate_mode mode)
1774 struct zs_pool *pool;
1775 struct size_class *class;
1776 struct zspage *zspage;
1778 void *s_addr, *d_addr, *addr;
1779 unsigned int offset;
1780 unsigned long handle;
1781 unsigned long old_obj, new_obj;
1782 unsigned int obj_idx;
1784 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1786 /* We're committed, tell the world that this is a Zsmalloc page. */
1787 __SetPageZsmalloc(newpage);
1789 /* The page is locked, so this pointer must remain valid */
1790 zspage = get_zspage(page);
1791 pool = zspage->pool;
1794 * The pool migrate_lock protects the race between zpage migration
1797 write_lock(&pool->migrate_lock);
1798 class = zspage_class(pool, zspage);
1801 * the class lock protects zpage alloc/free in the zspage.
1803 spin_lock(&class->lock);
1804 /* the migrate_write_lock protects zpage access via zs_map_object */
1805 migrate_write_lock(zspage);
1807 offset = get_first_obj_offset(page);
1808 s_addr = kmap_atomic(page);
1811 * Here, any user cannot access all objects in the zspage so let's move.
1813 d_addr = kmap_atomic(newpage);
1814 copy_page(d_addr, s_addr);
1815 kunmap_atomic(d_addr);
1817 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1818 addr += class->size) {
1819 if (obj_allocated(page, addr, &handle)) {
1821 old_obj = handle_to_obj(handle);
1822 obj_to_location(old_obj, &dummy, &obj_idx);
1823 new_obj = (unsigned long)location_to_obj(newpage,
1825 record_obj(handle, new_obj);
1828 kunmap_atomic(s_addr);
1830 replace_sub_page(class, zspage, newpage, page);
1832 * Since we complete the data copy and set up new zspage structure,
1833 * it's okay to release migration_lock.
1835 write_unlock(&pool->migrate_lock);
1836 spin_unlock(&class->lock);
1837 migrate_write_unlock(zspage);
1840 if (page_zone(newpage) != page_zone(page)) {
1841 dec_zone_page_state(page, NR_ZSPAGES);
1842 inc_zone_page_state(newpage, NR_ZSPAGES);
1848 return MIGRATEPAGE_SUCCESS;
1851 static void zs_page_putback(struct page *page)
1853 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1856 static const struct movable_operations zsmalloc_mops = {
1857 .isolate_page = zs_page_isolate,
1858 .migrate_page = zs_page_migrate,
1859 .putback_page = zs_page_putback,
1863 * Caller should hold page_lock of all pages in the zspage
1864 * In here, we cannot use zspage meta data.
1866 static void async_free_zspage(struct work_struct *work)
1869 struct size_class *class;
1870 struct zspage *zspage, *tmp;
1871 LIST_HEAD(free_pages);
1872 struct zs_pool *pool = container_of(work, struct zs_pool,
1875 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1876 class = pool->size_class[i];
1877 if (class->index != i)
1880 spin_lock(&class->lock);
1881 list_splice_init(&class->fullness_list[ZS_INUSE_RATIO_0],
1883 spin_unlock(&class->lock);
1886 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1887 list_del(&zspage->list);
1888 lock_zspage(zspage);
1890 class = zspage_class(pool, zspage);
1891 spin_lock(&class->lock);
1892 class_stat_sub(class, ZS_INUSE_RATIO_0, 1);
1893 __free_zspage(pool, class, zspage);
1894 spin_unlock(&class->lock);
1898 static void kick_deferred_free(struct zs_pool *pool)
1900 schedule_work(&pool->free_work);
1903 static void zs_flush_migration(struct zs_pool *pool)
1905 flush_work(&pool->free_work);
1908 static void init_deferred_free(struct zs_pool *pool)
1910 INIT_WORK(&pool->free_work, async_free_zspage);
1913 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
1915 struct page *page = get_first_page(zspage);
1918 WARN_ON(!trylock_page(page));
1919 __SetPageMovable(page, &zsmalloc_mops);
1921 } while ((page = get_next_page(page)) != NULL);
1924 static inline void zs_flush_migration(struct zs_pool *pool) { }
1929 * Based on the number of unused allocated objects calculate
1930 * and return the number of pages that we can free.
1932 static unsigned long zs_can_compact(struct size_class *class)
1934 unsigned long obj_wasted;
1935 unsigned long obj_allocated = class_stat_read(class, ZS_OBJS_ALLOCATED);
1936 unsigned long obj_used = class_stat_read(class, ZS_OBJS_INUSE);
1938 if (obj_allocated <= obj_used)
1941 obj_wasted = obj_allocated - obj_used;
1942 obj_wasted /= class->objs_per_zspage;
1944 return obj_wasted * class->pages_per_zspage;
1947 static unsigned long __zs_compact(struct zs_pool *pool,
1948 struct size_class *class)
1950 struct zspage *src_zspage = NULL;
1951 struct zspage *dst_zspage = NULL;
1952 unsigned long pages_freed = 0;
1955 * protect the race between zpage migration and zs_free
1956 * as well as zpage allocation/free
1958 write_lock(&pool->migrate_lock);
1959 spin_lock(&class->lock);
1960 while (zs_can_compact(class)) {
1964 dst_zspage = isolate_dst_zspage(class);
1969 src_zspage = isolate_src_zspage(class);
1973 migrate_write_lock(src_zspage);
1974 migrate_zspage(pool, src_zspage, dst_zspage);
1975 migrate_write_unlock(src_zspage);
1977 fg = putback_zspage(class, src_zspage);
1978 if (fg == ZS_INUSE_RATIO_0) {
1979 free_zspage(pool, class, src_zspage);
1980 pages_freed += class->pages_per_zspage;
1984 if (get_fullness_group(class, dst_zspage) == ZS_INUSE_RATIO_100
1985 || rwlock_is_contended(&pool->migrate_lock)) {
1986 putback_zspage(class, dst_zspage);
1989 spin_unlock(&class->lock);
1990 write_unlock(&pool->migrate_lock);
1992 write_lock(&pool->migrate_lock);
1993 spin_lock(&class->lock);
1998 putback_zspage(class, src_zspage);
2001 putback_zspage(class, dst_zspage);
2003 spin_unlock(&class->lock);
2004 write_unlock(&pool->migrate_lock);
2009 unsigned long zs_compact(struct zs_pool *pool)
2012 struct size_class *class;
2013 unsigned long pages_freed = 0;
2016 * Pool compaction is performed under pool->migrate_lock so it is basically
2017 * single-threaded. Having more than one thread in __zs_compact()
2018 * will increase pool->migrate_lock contention, which will impact other
2019 * zsmalloc operations that need pool->migrate_lock.
2021 if (atomic_xchg(&pool->compaction_in_progress, 1))
2024 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2025 class = pool->size_class[i];
2026 if (class->index != i)
2028 pages_freed += __zs_compact(pool, class);
2030 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2031 atomic_set(&pool->compaction_in_progress, 0);
2035 EXPORT_SYMBOL_GPL(zs_compact);
2037 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2039 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2041 EXPORT_SYMBOL_GPL(zs_pool_stats);
2043 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2044 struct shrink_control *sc)
2046 unsigned long pages_freed;
2047 struct zs_pool *pool = shrinker->private_data;
2050 * Compact classes and calculate compaction delta.
2051 * Can run concurrently with a manually triggered
2052 * (by user) compaction.
2054 pages_freed = zs_compact(pool);
2056 return pages_freed ? pages_freed : SHRINK_STOP;
2059 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2060 struct shrink_control *sc)
2063 struct size_class *class;
2064 unsigned long pages_to_free = 0;
2065 struct zs_pool *pool = shrinker->private_data;
2067 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2068 class = pool->size_class[i];
2069 if (class->index != i)
2072 pages_to_free += zs_can_compact(class);
2075 return pages_to_free;
2078 static void zs_unregister_shrinker(struct zs_pool *pool)
2080 shrinker_free(pool->shrinker);
2083 static int zs_register_shrinker(struct zs_pool *pool)
2085 pool->shrinker = shrinker_alloc(0, "mm-zspool:%s", pool->name);
2086 if (!pool->shrinker)
2089 pool->shrinker->scan_objects = zs_shrinker_scan;
2090 pool->shrinker->count_objects = zs_shrinker_count;
2091 pool->shrinker->batch = 0;
2092 pool->shrinker->private_data = pool;
2094 shrinker_register(pool->shrinker);
2099 static int calculate_zspage_chain_size(int class_size)
2101 int i, min_waste = INT_MAX;
2104 if (is_power_of_2(class_size))
2107 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
2110 waste = (i * PAGE_SIZE) % class_size;
2111 if (waste < min_waste) {
2121 * zs_create_pool - Creates an allocation pool to work from.
2122 * @name: pool name to be created
2124 * This function must be called before anything when using
2125 * the zsmalloc allocator.
2127 * On success, a pointer to the newly created pool is returned,
2130 struct zs_pool *zs_create_pool(const char *name)
2133 struct zs_pool *pool;
2134 struct size_class *prev_class = NULL;
2136 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2140 init_deferred_free(pool);
2141 rwlock_init(&pool->migrate_lock);
2142 atomic_set(&pool->compaction_in_progress, 0);
2144 pool->name = kstrdup(name, GFP_KERNEL);
2148 if (create_cache(pool))
2152 * Iterate reversely, because, size of size_class that we want to use
2153 * for merging should be larger or equal to current size.
2155 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2157 int pages_per_zspage;
2158 int objs_per_zspage;
2159 struct size_class *class;
2162 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2163 if (size > ZS_MAX_ALLOC_SIZE)
2164 size = ZS_MAX_ALLOC_SIZE;
2165 pages_per_zspage = calculate_zspage_chain_size(size);
2166 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2169 * We iterate from biggest down to smallest classes,
2170 * so huge_class_size holds the size of the first huge
2171 * class. Any object bigger than or equal to that will
2172 * endup in the huge class.
2174 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2176 huge_class_size = size;
2178 * The object uses ZS_HANDLE_SIZE bytes to store the
2179 * handle. We need to subtract it, because zs_malloc()
2180 * unconditionally adds handle size before it performs
2181 * size class search - so object may be smaller than
2182 * huge class size, yet it still can end up in the huge
2183 * class because it grows by ZS_HANDLE_SIZE extra bytes
2184 * right before class lookup.
2186 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2190 * size_class is used for normal zsmalloc operation such
2191 * as alloc/free for that size. Although it is natural that we
2192 * have one size_class for each size, there is a chance that we
2193 * can get more memory utilization if we use one size_class for
2194 * many different sizes whose size_class have same
2195 * characteristics. So, we makes size_class point to
2196 * previous size_class if possible.
2199 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2200 pool->size_class[i] = prev_class;
2205 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2211 class->pages_per_zspage = pages_per_zspage;
2212 class->objs_per_zspage = objs_per_zspage;
2213 spin_lock_init(&class->lock);
2214 pool->size_class[i] = class;
2216 fullness = ZS_INUSE_RATIO_0;
2217 while (fullness < NR_FULLNESS_GROUPS) {
2218 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2225 /* debug only, don't abort if it fails */
2226 zs_pool_stat_create(pool, name);
2229 * Not critical since shrinker is only used to trigger internal
2230 * defragmentation of the pool which is pretty optional thing. If
2231 * registration fails we still can use the pool normally and user can
2232 * trigger compaction manually. Thus, ignore return code.
2234 zs_register_shrinker(pool);
2239 zs_destroy_pool(pool);
2242 EXPORT_SYMBOL_GPL(zs_create_pool);
2244 void zs_destroy_pool(struct zs_pool *pool)
2248 zs_unregister_shrinker(pool);
2249 zs_flush_migration(pool);
2250 zs_pool_stat_destroy(pool);
2252 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2254 struct size_class *class = pool->size_class[i];
2259 if (class->index != i)
2262 for (fg = ZS_INUSE_RATIO_0; fg < NR_FULLNESS_GROUPS; fg++) {
2263 if (list_empty(&class->fullness_list[fg]))
2266 pr_err("Class-%d fullness group %d is not empty\n",
2272 destroy_cache(pool);
2276 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2278 static int __init zs_init(void)
2282 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2283 zs_cpu_prepare, zs_cpu_dead);
2288 zpool_register_driver(&zs_zpool_driver);
2299 static void __exit zs_exit(void)
2302 zpool_unregister_driver(&zs_zpool_driver);
2304 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2309 module_init(zs_init);
2310 module_exit(zs_exit);
2312 MODULE_LICENSE("Dual BSD/GPL");
2313 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
2314 MODULE_DESCRIPTION("zsmalloc memory allocator");