2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->freelist(index): links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
23 * page->units: first object offset in a subpage of zspage
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_private2: identifies the last component page
28 * PG_owner_priv_1: indentifies the huge component page
32 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
34 #include <linux/module.h>
35 #include <linux/kernel.h>
36 #include <linux/sched.h>
37 #include <linux/bitops.h>
38 #include <linux/errno.h>
39 #include <linux/highmem.h>
40 #include <linux/string.h>
41 #include <linux/slab.h>
42 #include <asm/tlbflush.h>
43 #include <asm/pgtable.h>
44 #include <linux/cpumask.h>
45 #include <linux/cpu.h>
46 #include <linux/vmalloc.h>
47 #include <linux/preempt.h>
48 #include <linux/spinlock.h>
49 #include <linux/types.h>
50 #include <linux/debugfs.h>
51 #include <linux/zsmalloc.h>
52 #include <linux/zpool.h>
53 #include <linux/mount.h>
54 #include <linux/migrate.h>
55 #include <linux/wait.h>
56 #include <linux/pagemap.h>
58 #define ZSPAGE_MAGIC 0x58
61 * This must be power of 2 and greater than of equal to sizeof(link_free).
62 * These two conditions ensure that any 'struct link_free' itself doesn't
63 * span more than 1 page which avoids complex case of mapping 2 pages simply
64 * to restore link_free pointer values.
69 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
70 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
72 #define ZS_MAX_ZSPAGE_ORDER 2
73 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
75 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
78 * Object location (<PFN>, <obj_idx>) is encoded as
79 * as single (unsigned long) handle value.
81 * Note that object index <obj_idx> starts from 0.
83 * This is made more complicated by various memory models and PAE.
86 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
87 #ifdef MAX_PHYSMEM_BITS
88 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
91 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
94 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
98 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
101 * Memory for allocating for handle keeps object position by
102 * encoding <page, obj_idx> and the encoded value has a room
103 * in least bit(ie, look at obj_to_location).
104 * We use the bit to synchronize between object access by
105 * user and migration.
107 #define HANDLE_PIN_BIT 0
110 * Head in allocated object should have OBJ_ALLOCATED_TAG
111 * to identify the object was allocated or not.
112 * It's okay to add the status bit in the least bit because
113 * header keeps handle which is 4byte-aligned address so we
114 * have room for two bit at least.
116 #define OBJ_ALLOCATED_TAG 1
117 #define OBJ_TAG_BITS 1
118 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
119 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
121 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
122 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
123 #define ZS_MIN_ALLOC_SIZE \
124 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
125 /* each chunk includes extra space to keep handle */
126 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
129 * On systems with 4K page size, this gives 255 size classes! There is a
131 * - Large number of size classes is potentially wasteful as free page are
132 * spread across these classes
133 * - Small number of size classes causes large internal fragmentation
134 * - Probably its better to use specific size classes (empirically
135 * determined). NOTE: all those class sizes must be set as multiple of
136 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
138 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
141 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
143 enum fullness_group {
161 struct zs_size_stat {
162 unsigned long objs[NR_ZS_STAT_TYPE];
165 #ifdef CONFIG_ZSMALLOC_STAT
166 static struct dentry *zs_stat_root;
169 #ifdef CONFIG_COMPACTION
170 static struct vfsmount *zsmalloc_mnt;
174 * number of size_classes
176 static int zs_size_classes;
179 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
181 * n = number of allocated objects
182 * N = total number of objects zspage can store
183 * f = fullness_threshold_frac
185 * Similarly, we assign zspage to:
186 * ZS_ALMOST_FULL when n > N / f
187 * ZS_EMPTY when n == 0
188 * ZS_FULL when n == N
190 * (see: fix_fullness_group())
192 static const int fullness_threshold_frac = 4;
196 struct list_head fullness_list[NR_ZS_FULLNESS];
198 * Size of objects stored in this class. Must be multiple
203 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
204 int pages_per_zspage;
207 struct zs_size_stat stats;
210 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
211 static void SetPageHugeObject(struct page *page)
213 SetPageOwnerPriv1(page);
216 static void ClearPageHugeObject(struct page *page)
218 ClearPageOwnerPriv1(page);
221 static int PageHugeObject(struct page *page)
223 return PageOwnerPriv1(page);
227 * Placed within free objects to form a singly linked list.
228 * For every zspage, zspage->freeobj gives head of this list.
230 * This must be power of 2 and less than or equal to ZS_ALIGN
236 * It's valid for non-allocated object
240 * Handle of allocated object.
242 unsigned long handle;
249 struct size_class **size_class;
250 struct kmem_cache *handle_cachep;
251 struct kmem_cache *zspage_cachep;
253 atomic_long_t pages_allocated;
255 struct zs_pool_stats stats;
257 /* Compact classes */
258 struct shrinker shrinker;
260 * To signify that register_shrinker() was successful
261 * and unregister_shrinker() will not Oops.
263 bool shrinker_enabled;
264 #ifdef CONFIG_ZSMALLOC_STAT
265 struct dentry *stat_dentry;
267 #ifdef CONFIG_COMPACTION
269 struct work_struct free_work;
270 /* A wait queue for when migration races with async_free_zspage() */
271 wait_queue_head_t migration_wait;
272 atomic_long_t isolated_pages;
278 * A zspage's class index and fullness group
279 * are encoded in its (first)page->mapping
281 #define FULLNESS_BITS 2
283 #define ISOLATED_BITS 3
284 #define MAGIC_VAL_BITS 8
288 unsigned int fullness:FULLNESS_BITS;
289 unsigned int class:CLASS_BITS + 1;
290 unsigned int isolated:ISOLATED_BITS;
291 unsigned int magic:MAGIC_VAL_BITS;
294 unsigned int freeobj;
295 struct page *first_page;
296 struct list_head list; /* fullness list */
297 #ifdef CONFIG_COMPACTION
302 struct mapping_area {
303 #ifdef CONFIG_PGTABLE_MAPPING
304 struct vm_struct *vm; /* vm area for mapping object that span pages */
306 char *vm_buf; /* copy buffer for objects that span pages */
308 char *vm_addr; /* address of kmap_atomic()'ed pages */
309 enum zs_mapmode vm_mm; /* mapping mode */
312 #ifdef CONFIG_COMPACTION
313 static int zs_register_migration(struct zs_pool *pool);
314 static void zs_unregister_migration(struct zs_pool *pool);
315 static void migrate_lock_init(struct zspage *zspage);
316 static void migrate_read_lock(struct zspage *zspage);
317 static void migrate_read_unlock(struct zspage *zspage);
318 static void kick_deferred_free(struct zs_pool *pool);
319 static void init_deferred_free(struct zs_pool *pool);
320 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
322 static int zsmalloc_mount(void) { return 0; }
323 static void zsmalloc_unmount(void) {}
324 static int zs_register_migration(struct zs_pool *pool) { return 0; }
325 static void zs_unregister_migration(struct zs_pool *pool) {}
326 static void migrate_lock_init(struct zspage *zspage) {}
327 static void migrate_read_lock(struct zspage *zspage) {}
328 static void migrate_read_unlock(struct zspage *zspage) {}
329 static void kick_deferred_free(struct zs_pool *pool) {}
330 static void init_deferred_free(struct zs_pool *pool) {}
331 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
334 static int create_cache(struct zs_pool *pool)
336 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
338 if (!pool->handle_cachep)
341 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
343 if (!pool->zspage_cachep) {
344 kmem_cache_destroy(pool->handle_cachep);
345 pool->handle_cachep = NULL;
352 static void destroy_cache(struct zs_pool *pool)
354 kmem_cache_destroy(pool->handle_cachep);
355 kmem_cache_destroy(pool->zspage_cachep);
358 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
360 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
361 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
364 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
366 kmem_cache_free(pool->handle_cachep, (void *)handle);
369 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
371 return kmem_cache_alloc(pool->zspage_cachep,
372 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
375 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
377 kmem_cache_free(pool->zspage_cachep, zspage);
380 static void record_obj(unsigned long handle, unsigned long obj)
383 * lsb of @obj represents handle lock while other bits
384 * represent object value the handle is pointing so
385 * updating shouldn't do store tearing.
387 WRITE_ONCE(*(unsigned long *)handle, obj);
394 static void *zs_zpool_create(const char *name, gfp_t gfp,
395 const struct zpool_ops *zpool_ops,
399 * Ignore global gfp flags: zs_malloc() may be invoked from
400 * different contexts and its caller must provide a valid
403 return zs_create_pool(name);
406 static void zs_zpool_destroy(void *pool)
408 zs_destroy_pool(pool);
411 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
412 unsigned long *handle)
414 *handle = zs_malloc(pool, size, gfp);
415 return *handle ? 0 : -1;
417 static void zs_zpool_free(void *pool, unsigned long handle)
419 zs_free(pool, handle);
422 static int zs_zpool_shrink(void *pool, unsigned int pages,
423 unsigned int *reclaimed)
428 static void *zs_zpool_map(void *pool, unsigned long handle,
429 enum zpool_mapmode mm)
431 enum zs_mapmode zs_mm;
440 case ZPOOL_MM_RW: /* fallthru */
446 return zs_map_object(pool, handle, zs_mm);
448 static void zs_zpool_unmap(void *pool, unsigned long handle)
450 zs_unmap_object(pool, handle);
453 static u64 zs_zpool_total_size(void *pool)
455 return zs_get_total_pages(pool) << PAGE_SHIFT;
458 static struct zpool_driver zs_zpool_driver = {
460 .owner = THIS_MODULE,
461 .create = zs_zpool_create,
462 .destroy = zs_zpool_destroy,
463 .malloc = zs_zpool_malloc,
464 .free = zs_zpool_free,
465 .shrink = zs_zpool_shrink,
467 .unmap = zs_zpool_unmap,
468 .total_size = zs_zpool_total_size,
471 MODULE_ALIAS("zpool-zsmalloc");
472 #endif /* CONFIG_ZPOOL */
474 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
475 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
477 static bool is_zspage_isolated(struct zspage *zspage)
479 return zspage->isolated;
482 static __maybe_unused int is_first_page(struct page *page)
484 return PagePrivate(page);
487 /* Protected by class->lock */
488 static inline int get_zspage_inuse(struct zspage *zspage)
490 return zspage->inuse;
493 static inline void set_zspage_inuse(struct zspage *zspage, int val)
498 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
500 zspage->inuse += val;
503 static inline struct page *get_first_page(struct zspage *zspage)
505 struct page *first_page = zspage->first_page;
507 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
511 static inline int get_first_obj_offset(struct page *page)
516 static inline void set_first_obj_offset(struct page *page, int offset)
518 page->units = offset;
521 static inline unsigned int get_freeobj(struct zspage *zspage)
523 return zspage->freeobj;
526 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
528 zspage->freeobj = obj;
531 static void get_zspage_mapping(struct zspage *zspage,
532 unsigned int *class_idx,
533 enum fullness_group *fullness)
535 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
537 *fullness = zspage->fullness;
538 *class_idx = zspage->class;
541 static void set_zspage_mapping(struct zspage *zspage,
542 unsigned int class_idx,
543 enum fullness_group fullness)
545 zspage->class = class_idx;
546 zspage->fullness = fullness;
550 * zsmalloc divides the pool into various size classes where each
551 * class maintains a list of zspages where each zspage is divided
552 * into equal sized chunks. Each allocation falls into one of these
553 * classes depending on its size. This function returns index of the
554 * size class which has chunk size big enough to hold the give size.
556 static int get_size_class_index(int size)
560 if (likely(size > ZS_MIN_ALLOC_SIZE))
561 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
562 ZS_SIZE_CLASS_DELTA);
564 return min(zs_size_classes - 1, idx);
567 /* type can be of enum type zs_stat_type or fullness_group */
568 static inline void zs_stat_inc(struct size_class *class,
569 int type, unsigned long cnt)
571 class->stats.objs[type] += cnt;
574 /* type can be of enum type zs_stat_type or fullness_group */
575 static inline void zs_stat_dec(struct size_class *class,
576 int type, unsigned long cnt)
578 class->stats.objs[type] -= cnt;
581 /* type can be of enum type zs_stat_type or fullness_group */
582 static inline unsigned long zs_stat_get(struct size_class *class,
585 return class->stats.objs[type];
588 #ifdef CONFIG_ZSMALLOC_STAT
590 static void __init zs_stat_init(void)
592 if (!debugfs_initialized()) {
593 pr_warn("debugfs not available, stat dir not created\n");
597 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
599 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
602 static void __exit zs_stat_exit(void)
604 debugfs_remove_recursive(zs_stat_root);
607 static unsigned long zs_can_compact(struct size_class *class);
609 static int zs_stats_size_show(struct seq_file *s, void *v)
612 struct zs_pool *pool = s->private;
613 struct size_class *class;
615 unsigned long class_almost_full, class_almost_empty;
616 unsigned long obj_allocated, obj_used, pages_used, freeable;
617 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
618 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
619 unsigned long total_freeable = 0;
621 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
622 "class", "size", "almost_full", "almost_empty",
623 "obj_allocated", "obj_used", "pages_used",
624 "pages_per_zspage", "freeable");
626 for (i = 0; i < zs_size_classes; i++) {
627 class = pool->size_class[i];
629 if (class->index != i)
632 spin_lock(&class->lock);
633 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
634 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
635 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
636 obj_used = zs_stat_get(class, OBJ_USED);
637 freeable = zs_can_compact(class);
638 spin_unlock(&class->lock);
640 objs_per_zspage = class->objs_per_zspage;
641 pages_used = obj_allocated / objs_per_zspage *
642 class->pages_per_zspage;
644 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
645 " %10lu %10lu %16d %8lu\n",
646 i, class->size, class_almost_full, class_almost_empty,
647 obj_allocated, obj_used, pages_used,
648 class->pages_per_zspage, freeable);
650 total_class_almost_full += class_almost_full;
651 total_class_almost_empty += class_almost_empty;
652 total_objs += obj_allocated;
653 total_used_objs += obj_used;
654 total_pages += pages_used;
655 total_freeable += freeable;
659 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
660 "Total", "", total_class_almost_full,
661 total_class_almost_empty, total_objs,
662 total_used_objs, total_pages, "", total_freeable);
667 static int zs_stats_size_open(struct inode *inode, struct file *file)
669 return single_open(file, zs_stats_size_show, inode->i_private);
672 static const struct file_operations zs_stat_size_ops = {
673 .open = zs_stats_size_open,
676 .release = single_release,
679 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
681 struct dentry *entry;
684 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
688 entry = debugfs_create_dir(name, zs_stat_root);
690 pr_warn("debugfs dir <%s> creation failed\n", name);
693 pool->stat_dentry = entry;
695 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
696 pool->stat_dentry, pool, &zs_stat_size_ops);
698 pr_warn("%s: debugfs file entry <%s> creation failed\n",
700 debugfs_remove_recursive(pool->stat_dentry);
701 pool->stat_dentry = NULL;
705 static void zs_pool_stat_destroy(struct zs_pool *pool)
707 debugfs_remove_recursive(pool->stat_dentry);
710 #else /* CONFIG_ZSMALLOC_STAT */
711 static void __init zs_stat_init(void)
715 static void __exit zs_stat_exit(void)
719 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
723 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
730 * For each size class, zspages are divided into different groups
731 * depending on how "full" they are. This was done so that we could
732 * easily find empty or nearly empty zspages when we try to shrink
733 * the pool (not yet implemented). This function returns fullness
734 * status of the given page.
736 static enum fullness_group get_fullness_group(struct size_class *class,
737 struct zspage *zspage)
739 int inuse, objs_per_zspage;
740 enum fullness_group fg;
742 inuse = get_zspage_inuse(zspage);
743 objs_per_zspage = class->objs_per_zspage;
747 else if (inuse == objs_per_zspage)
749 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
750 fg = ZS_ALMOST_EMPTY;
758 * Each size class maintains various freelists and zspages are assigned
759 * to one of these freelists based on the number of live objects they
760 * have. This functions inserts the given zspage into the freelist
761 * identified by <class, fullness_group>.
763 static void insert_zspage(struct size_class *class,
764 struct zspage *zspage,
765 enum fullness_group fullness)
769 zs_stat_inc(class, fullness, 1);
770 head = list_first_entry_or_null(&class->fullness_list[fullness],
771 struct zspage, list);
773 * We want to see more ZS_FULL pages and less almost empty/full.
774 * Put pages with higher ->inuse first.
777 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
778 list_add(&zspage->list, &head->list);
782 list_add(&zspage->list, &class->fullness_list[fullness]);
786 * This function removes the given zspage from the freelist identified
787 * by <class, fullness_group>.
789 static void remove_zspage(struct size_class *class,
790 struct zspage *zspage,
791 enum fullness_group fullness)
793 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
794 VM_BUG_ON(is_zspage_isolated(zspage));
796 list_del_init(&zspage->list);
797 zs_stat_dec(class, fullness, 1);
801 * Each size class maintains zspages in different fullness groups depending
802 * on the number of live objects they contain. When allocating or freeing
803 * objects, the fullness status of the page can change, say, from ALMOST_FULL
804 * to ALMOST_EMPTY when freeing an object. This function checks if such
805 * a status change has occurred for the given page and accordingly moves the
806 * page from the freelist of the old fullness group to that of the new
809 static enum fullness_group fix_fullness_group(struct size_class *class,
810 struct zspage *zspage)
813 enum fullness_group currfg, newfg;
815 get_zspage_mapping(zspage, &class_idx, &currfg);
816 newfg = get_fullness_group(class, zspage);
820 if (!is_zspage_isolated(zspage)) {
821 remove_zspage(class, zspage, currfg);
822 insert_zspage(class, zspage, newfg);
825 set_zspage_mapping(zspage, class_idx, newfg);
832 * We have to decide on how many pages to link together
833 * to form a zspage for each size class. This is important
834 * to reduce wastage due to unusable space left at end of
835 * each zspage which is given as:
836 * wastage = Zp % class_size
837 * usage = Zp - wastage
838 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
840 * For example, for size class of 3/8 * PAGE_SIZE, we should
841 * link together 3 PAGE_SIZE sized pages to form a zspage
842 * since then we can perfectly fit in 8 such objects.
844 static int get_pages_per_zspage(int class_size)
846 int i, max_usedpc = 0;
847 /* zspage order which gives maximum used size per KB */
848 int max_usedpc_order = 1;
850 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
854 zspage_size = i * PAGE_SIZE;
855 waste = zspage_size % class_size;
856 usedpc = (zspage_size - waste) * 100 / zspage_size;
858 if (usedpc > max_usedpc) {
860 max_usedpc_order = i;
864 return max_usedpc_order;
867 static struct zspage *get_zspage(struct page *page)
869 struct zspage *zspage = (struct zspage *)page->private;
871 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
875 static struct page *get_next_page(struct page *page)
877 if (unlikely(PageHugeObject(page)))
880 return page->freelist;
884 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
885 * @page: page object resides in zspage
886 * @obj_idx: object index
888 static void obj_to_location(unsigned long obj, struct page **page,
889 unsigned int *obj_idx)
891 obj >>= OBJ_TAG_BITS;
892 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
893 *obj_idx = (obj & OBJ_INDEX_MASK);
897 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
898 * @page: page object resides in zspage
899 * @obj_idx: object index
901 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
905 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
906 obj |= obj_idx & OBJ_INDEX_MASK;
907 obj <<= OBJ_TAG_BITS;
912 static unsigned long handle_to_obj(unsigned long handle)
914 return *(unsigned long *)handle;
917 static unsigned long obj_to_head(struct page *page, void *obj)
919 if (unlikely(PageHugeObject(page))) {
920 VM_BUG_ON_PAGE(!is_first_page(page), page);
923 return *(unsigned long *)obj;
926 static inline int testpin_tag(unsigned long handle)
928 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
931 static inline int trypin_tag(unsigned long handle)
933 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
936 static void pin_tag(unsigned long handle)
938 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
941 static void unpin_tag(unsigned long handle)
943 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
946 static void reset_page(struct page *page)
948 __ClearPageMovable(page);
949 ClearPagePrivate(page);
950 ClearPagePrivate2(page);
951 set_page_private(page, 0);
952 page_mapcount_reset(page);
953 ClearPageHugeObject(page);
954 page->freelist = NULL;
958 * To prevent zspage destroy during migration, zspage freeing should
959 * hold locks of all pages in the zspage.
961 void lock_zspage(struct zspage *zspage)
963 struct page *page = get_first_page(zspage);
967 } while ((page = get_next_page(page)) != NULL);
970 int trylock_zspage(struct zspage *zspage)
972 struct page *cursor, *fail;
974 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
975 get_next_page(cursor)) {
976 if (!trylock_page(cursor)) {
984 for (cursor = get_first_page(zspage); cursor != fail; cursor =
985 get_next_page(cursor))
991 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
992 struct zspage *zspage)
994 struct page *page, *next;
995 enum fullness_group fg;
996 unsigned int class_idx;
998 get_zspage_mapping(zspage, &class_idx, &fg);
1000 assert_spin_locked(&class->lock);
1002 VM_BUG_ON(get_zspage_inuse(zspage));
1003 VM_BUG_ON(fg != ZS_EMPTY);
1005 next = page = get_first_page(zspage);
1007 VM_BUG_ON_PAGE(!PageLocked(page), page);
1008 next = get_next_page(page);
1011 dec_zone_page_state(page, NR_ZSPAGES);
1014 } while (page != NULL);
1016 cache_free_zspage(pool, zspage);
1018 zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1019 atomic_long_sub(class->pages_per_zspage,
1020 &pool->pages_allocated);
1023 static void free_zspage(struct zs_pool *pool, struct size_class *class,
1024 struct zspage *zspage)
1026 VM_BUG_ON(get_zspage_inuse(zspage));
1027 VM_BUG_ON(list_empty(&zspage->list));
1029 if (!trylock_zspage(zspage)) {
1030 kick_deferred_free(pool);
1034 remove_zspage(class, zspage, ZS_EMPTY);
1035 __free_zspage(pool, class, zspage);
1038 /* Initialize a newly allocated zspage */
1039 static void init_zspage(struct size_class *class, struct zspage *zspage)
1041 unsigned int freeobj = 1;
1042 unsigned long off = 0;
1043 struct page *page = get_first_page(zspage);
1046 struct page *next_page;
1047 struct link_free *link;
1050 set_first_obj_offset(page, off);
1052 vaddr = kmap_atomic(page);
1053 link = (struct link_free *)vaddr + off / sizeof(*link);
1055 while ((off += class->size) < PAGE_SIZE) {
1056 link->next = freeobj++ << OBJ_TAG_BITS;
1057 link += class->size / sizeof(*link);
1061 * We now come to the last (full or partial) object on this
1062 * page, which must point to the first object on the next
1065 next_page = get_next_page(page);
1067 link->next = freeobj++ << OBJ_TAG_BITS;
1070 * Reset OBJ_TAG_BITS bit to last link to tell
1071 * whether it's allocated object or not.
1073 link->next = -1 << OBJ_TAG_BITS;
1075 kunmap_atomic(vaddr);
1080 set_freeobj(zspage, 0);
1083 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1084 struct page *pages[])
1088 struct page *prev_page = NULL;
1089 int nr_pages = class->pages_per_zspage;
1092 * Allocate individual pages and link them together as:
1093 * 1. all pages are linked together using page->freelist
1094 * 2. each sub-page point to zspage using page->private
1096 * we set PG_private to identify the first page (i.e. no other sub-page
1097 * has this flag set) and PG_private_2 to identify the last page.
1099 for (i = 0; i < nr_pages; i++) {
1101 set_page_private(page, (unsigned long)zspage);
1102 page->freelist = NULL;
1104 zspage->first_page = page;
1105 SetPagePrivate(page);
1106 if (unlikely(class->objs_per_zspage == 1 &&
1107 class->pages_per_zspage == 1))
1108 SetPageHugeObject(page);
1110 prev_page->freelist = page;
1112 if (i == nr_pages - 1)
1113 SetPagePrivate2(page);
1119 * Allocate a zspage for the given size class
1121 static struct zspage *alloc_zspage(struct zs_pool *pool,
1122 struct size_class *class,
1126 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1127 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1132 memset(zspage, 0, sizeof(struct zspage));
1133 zspage->magic = ZSPAGE_MAGIC;
1134 migrate_lock_init(zspage);
1136 for (i = 0; i < class->pages_per_zspage; i++) {
1139 page = alloc_page(gfp);
1142 dec_zone_page_state(pages[i], NR_ZSPAGES);
1143 __free_page(pages[i]);
1145 cache_free_zspage(pool, zspage);
1149 inc_zone_page_state(page, NR_ZSPAGES);
1153 create_page_chain(class, zspage, pages);
1154 init_zspage(class, zspage);
1159 static struct zspage *find_get_zspage(struct size_class *class)
1162 struct zspage *zspage;
1164 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1165 zspage = list_first_entry_or_null(&class->fullness_list[i],
1166 struct zspage, list);
1174 #ifdef CONFIG_PGTABLE_MAPPING
1175 static inline int __zs_cpu_up(struct mapping_area *area)
1178 * Make sure we don't leak memory if a cpu UP notification
1179 * and zs_init() race and both call zs_cpu_up() on the same cpu
1183 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1189 static inline void __zs_cpu_down(struct mapping_area *area)
1192 free_vm_area(area->vm);
1196 static inline void *__zs_map_object(struct mapping_area *area,
1197 struct page *pages[2], int off, int size)
1199 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1200 area->vm_addr = area->vm->addr;
1201 return area->vm_addr + off;
1204 static inline void __zs_unmap_object(struct mapping_area *area,
1205 struct page *pages[2], int off, int size)
1207 unsigned long addr = (unsigned long)area->vm_addr;
1209 unmap_kernel_range(addr, PAGE_SIZE * 2);
1212 #else /* CONFIG_PGTABLE_MAPPING */
1214 static inline int __zs_cpu_up(struct mapping_area *area)
1217 * Make sure we don't leak memory if a cpu UP notification
1218 * and zs_init() race and both call zs_cpu_up() on the same cpu
1222 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1228 static inline void __zs_cpu_down(struct mapping_area *area)
1230 kfree(area->vm_buf);
1231 area->vm_buf = NULL;
1234 static void *__zs_map_object(struct mapping_area *area,
1235 struct page *pages[2], int off, int size)
1239 char *buf = area->vm_buf;
1241 /* disable page faults to match kmap_atomic() return conditions */
1242 pagefault_disable();
1244 /* no read fastpath */
1245 if (area->vm_mm == ZS_MM_WO)
1248 sizes[0] = PAGE_SIZE - off;
1249 sizes[1] = size - sizes[0];
1251 /* copy object to per-cpu buffer */
1252 addr = kmap_atomic(pages[0]);
1253 memcpy(buf, addr + off, sizes[0]);
1254 kunmap_atomic(addr);
1255 addr = kmap_atomic(pages[1]);
1256 memcpy(buf + sizes[0], addr, sizes[1]);
1257 kunmap_atomic(addr);
1259 return area->vm_buf;
1262 static void __zs_unmap_object(struct mapping_area *area,
1263 struct page *pages[2], int off, int size)
1269 /* no write fastpath */
1270 if (area->vm_mm == ZS_MM_RO)
1274 buf = buf + ZS_HANDLE_SIZE;
1275 size -= ZS_HANDLE_SIZE;
1276 off += ZS_HANDLE_SIZE;
1278 sizes[0] = PAGE_SIZE - off;
1279 sizes[1] = size - sizes[0];
1281 /* copy per-cpu buffer to object */
1282 addr = kmap_atomic(pages[0]);
1283 memcpy(addr + off, buf, sizes[0]);
1284 kunmap_atomic(addr);
1285 addr = kmap_atomic(pages[1]);
1286 memcpy(addr, buf + sizes[0], sizes[1]);
1287 kunmap_atomic(addr);
1290 /* enable page faults to match kunmap_atomic() return conditions */
1294 #endif /* CONFIG_PGTABLE_MAPPING */
1296 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1299 int ret, cpu = (long)pcpu;
1300 struct mapping_area *area;
1303 case CPU_UP_PREPARE:
1304 area = &per_cpu(zs_map_area, cpu);
1305 ret = __zs_cpu_up(area);
1307 return notifier_from_errno(ret);
1310 case CPU_UP_CANCELED:
1311 area = &per_cpu(zs_map_area, cpu);
1312 __zs_cpu_down(area);
1319 static struct notifier_block zs_cpu_nb = {
1320 .notifier_call = zs_cpu_notifier
1323 static int zs_register_cpu_notifier(void)
1325 int cpu, uninitialized_var(ret);
1327 cpu_notifier_register_begin();
1329 __register_cpu_notifier(&zs_cpu_nb);
1330 for_each_online_cpu(cpu) {
1331 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1332 if (notifier_to_errno(ret))
1336 cpu_notifier_register_done();
1337 return notifier_to_errno(ret);
1340 static void zs_unregister_cpu_notifier(void)
1344 cpu_notifier_register_begin();
1346 for_each_online_cpu(cpu)
1347 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1348 __unregister_cpu_notifier(&zs_cpu_nb);
1350 cpu_notifier_register_done();
1353 static void __init init_zs_size_classes(void)
1357 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1358 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1361 zs_size_classes = nr;
1364 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1365 int objs_per_zspage)
1367 if (prev->pages_per_zspage == pages_per_zspage &&
1368 prev->objs_per_zspage == objs_per_zspage)
1374 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1376 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1379 unsigned long zs_get_total_pages(struct zs_pool *pool)
1381 return atomic_long_read(&pool->pages_allocated);
1383 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1386 * zs_map_object - get address of allocated object from handle.
1387 * @pool: pool from which the object was allocated
1388 * @handle: handle returned from zs_malloc
1390 * Before using an object allocated from zs_malloc, it must be mapped using
1391 * this function. When done with the object, it must be unmapped using
1394 * Only one object can be mapped per cpu at a time. There is no protection
1395 * against nested mappings.
1397 * This function returns with preemption and page faults disabled.
1399 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1402 struct zspage *zspage;
1404 unsigned long obj, off;
1405 unsigned int obj_idx;
1407 unsigned int class_idx;
1408 enum fullness_group fg;
1409 struct size_class *class;
1410 struct mapping_area *area;
1411 struct page *pages[2];
1415 * Because we use per-cpu mapping areas shared among the
1416 * pools/users, we can't allow mapping in interrupt context
1417 * because it can corrupt another users mappings.
1419 BUG_ON(in_interrupt());
1421 /* From now on, migration cannot move the object */
1424 obj = handle_to_obj(handle);
1425 obj_to_location(obj, &page, &obj_idx);
1426 zspage = get_zspage(page);
1428 /* migration cannot move any subpage in this zspage */
1429 migrate_read_lock(zspage);
1431 get_zspage_mapping(zspage, &class_idx, &fg);
1432 class = pool->size_class[class_idx];
1433 off = (class->size * obj_idx) & ~PAGE_MASK;
1435 area = &get_cpu_var(zs_map_area);
1437 if (off + class->size <= PAGE_SIZE) {
1438 /* this object is contained entirely within a page */
1439 area->vm_addr = kmap_atomic(page);
1440 ret = area->vm_addr + off;
1444 /* this object spans two pages */
1446 pages[1] = get_next_page(page);
1449 ret = __zs_map_object(area, pages, off, class->size);
1451 if (likely(!PageHugeObject(page)))
1452 ret += ZS_HANDLE_SIZE;
1456 EXPORT_SYMBOL_GPL(zs_map_object);
1458 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1460 struct zspage *zspage;
1462 unsigned long obj, off;
1463 unsigned int obj_idx;
1465 unsigned int class_idx;
1466 enum fullness_group fg;
1467 struct size_class *class;
1468 struct mapping_area *area;
1470 obj = handle_to_obj(handle);
1471 obj_to_location(obj, &page, &obj_idx);
1472 zspage = get_zspage(page);
1473 get_zspage_mapping(zspage, &class_idx, &fg);
1474 class = pool->size_class[class_idx];
1475 off = (class->size * obj_idx) & ~PAGE_MASK;
1477 area = this_cpu_ptr(&zs_map_area);
1478 if (off + class->size <= PAGE_SIZE)
1479 kunmap_atomic(area->vm_addr);
1481 struct page *pages[2];
1484 pages[1] = get_next_page(page);
1487 __zs_unmap_object(area, pages, off, class->size);
1489 put_cpu_var(zs_map_area);
1491 migrate_read_unlock(zspage);
1494 EXPORT_SYMBOL_GPL(zs_unmap_object);
1496 static unsigned long obj_malloc(struct size_class *class,
1497 struct zspage *zspage, unsigned long handle)
1499 int i, nr_page, offset;
1501 struct link_free *link;
1503 struct page *m_page;
1504 unsigned long m_offset;
1507 handle |= OBJ_ALLOCATED_TAG;
1508 obj = get_freeobj(zspage);
1510 offset = obj * class->size;
1511 nr_page = offset >> PAGE_SHIFT;
1512 m_offset = offset & ~PAGE_MASK;
1513 m_page = get_first_page(zspage);
1515 for (i = 0; i < nr_page; i++)
1516 m_page = get_next_page(m_page);
1518 vaddr = kmap_atomic(m_page);
1519 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1520 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1521 if (likely(!PageHugeObject(m_page)))
1522 /* record handle in the header of allocated chunk */
1523 link->handle = handle;
1525 /* record handle to page->index */
1526 zspage->first_page->index = handle;
1528 kunmap_atomic(vaddr);
1529 mod_zspage_inuse(zspage, 1);
1530 zs_stat_inc(class, OBJ_USED, 1);
1532 obj = location_to_obj(m_page, obj);
1539 * zs_malloc - Allocate block of given size from pool.
1540 * @pool: pool to allocate from
1541 * @size: size of block to allocate
1542 * @gfp: gfp flags when allocating object
1544 * On success, handle to the allocated object is returned,
1546 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1548 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1550 unsigned long handle, obj;
1551 struct size_class *class;
1552 enum fullness_group newfg;
1553 struct zspage *zspage;
1555 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1558 handle = cache_alloc_handle(pool, gfp);
1562 /* extra space in chunk to keep the handle */
1563 size += ZS_HANDLE_SIZE;
1564 class = pool->size_class[get_size_class_index(size)];
1566 spin_lock(&class->lock);
1567 zspage = find_get_zspage(class);
1568 if (likely(zspage)) {
1569 obj = obj_malloc(class, zspage, handle);
1570 /* Now move the zspage to another fullness group, if required */
1571 fix_fullness_group(class, zspage);
1572 record_obj(handle, obj);
1573 spin_unlock(&class->lock);
1578 spin_unlock(&class->lock);
1580 zspage = alloc_zspage(pool, class, gfp);
1582 cache_free_handle(pool, handle);
1586 spin_lock(&class->lock);
1587 obj = obj_malloc(class, zspage, handle);
1588 newfg = get_fullness_group(class, zspage);
1589 insert_zspage(class, zspage, newfg);
1590 set_zspage_mapping(zspage, class->index, newfg);
1591 record_obj(handle, obj);
1592 atomic_long_add(class->pages_per_zspage,
1593 &pool->pages_allocated);
1594 zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1596 /* We completely set up zspage so mark them as movable */
1597 SetZsPageMovable(pool, zspage);
1598 spin_unlock(&class->lock);
1602 EXPORT_SYMBOL_GPL(zs_malloc);
1604 static void obj_free(struct size_class *class, unsigned long obj)
1606 struct link_free *link;
1607 struct zspage *zspage;
1608 struct page *f_page;
1609 unsigned long f_offset;
1610 unsigned int f_objidx;
1613 obj &= ~OBJ_ALLOCATED_TAG;
1614 obj_to_location(obj, &f_page, &f_objidx);
1615 f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1616 zspage = get_zspage(f_page);
1618 vaddr = kmap_atomic(f_page);
1620 /* Insert this object in containing zspage's freelist */
1621 link = (struct link_free *)(vaddr + f_offset);
1622 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1623 kunmap_atomic(vaddr);
1624 set_freeobj(zspage, f_objidx);
1625 mod_zspage_inuse(zspage, -1);
1626 zs_stat_dec(class, OBJ_USED, 1);
1629 void zs_free(struct zs_pool *pool, unsigned long handle)
1631 struct zspage *zspage;
1632 struct page *f_page;
1634 unsigned int f_objidx;
1636 struct size_class *class;
1637 enum fullness_group fullness;
1640 if (unlikely(!handle))
1644 obj = handle_to_obj(handle);
1645 obj_to_location(obj, &f_page, &f_objidx);
1646 zspage = get_zspage(f_page);
1648 migrate_read_lock(zspage);
1650 get_zspage_mapping(zspage, &class_idx, &fullness);
1651 class = pool->size_class[class_idx];
1653 spin_lock(&class->lock);
1654 obj_free(class, obj);
1655 fullness = fix_fullness_group(class, zspage);
1656 if (fullness != ZS_EMPTY) {
1657 migrate_read_unlock(zspage);
1661 isolated = is_zspage_isolated(zspage);
1662 migrate_read_unlock(zspage);
1663 /* If zspage is isolated, zs_page_putback will free the zspage */
1664 if (likely(!isolated))
1665 free_zspage(pool, class, zspage);
1668 spin_unlock(&class->lock);
1670 cache_free_handle(pool, handle);
1672 EXPORT_SYMBOL_GPL(zs_free);
1674 static void zs_object_copy(struct size_class *class, unsigned long dst,
1677 struct page *s_page, *d_page;
1678 unsigned int s_objidx, d_objidx;
1679 unsigned long s_off, d_off;
1680 void *s_addr, *d_addr;
1681 int s_size, d_size, size;
1684 s_size = d_size = class->size;
1686 obj_to_location(src, &s_page, &s_objidx);
1687 obj_to_location(dst, &d_page, &d_objidx);
1689 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1690 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1692 if (s_off + class->size > PAGE_SIZE)
1693 s_size = PAGE_SIZE - s_off;
1695 if (d_off + class->size > PAGE_SIZE)
1696 d_size = PAGE_SIZE - d_off;
1698 s_addr = kmap_atomic(s_page);
1699 d_addr = kmap_atomic(d_page);
1702 size = min(s_size, d_size);
1703 memcpy(d_addr + d_off, s_addr + s_off, size);
1706 if (written == class->size)
1714 if (s_off >= PAGE_SIZE) {
1715 kunmap_atomic(d_addr);
1716 kunmap_atomic(s_addr);
1717 s_page = get_next_page(s_page);
1718 s_addr = kmap_atomic(s_page);
1719 d_addr = kmap_atomic(d_page);
1720 s_size = class->size - written;
1724 if (d_off >= PAGE_SIZE) {
1725 kunmap_atomic(d_addr);
1726 d_page = get_next_page(d_page);
1727 d_addr = kmap_atomic(d_page);
1728 d_size = class->size - written;
1733 kunmap_atomic(d_addr);
1734 kunmap_atomic(s_addr);
1738 * Find alloced object in zspage from index object and
1741 static unsigned long find_alloced_obj(struct size_class *class,
1742 struct page *page, int *obj_idx)
1746 int index = *obj_idx;
1747 unsigned long handle = 0;
1748 void *addr = kmap_atomic(page);
1750 offset = get_first_obj_offset(page);
1751 offset += class->size * index;
1753 while (offset < PAGE_SIZE) {
1754 head = obj_to_head(page, addr + offset);
1755 if (head & OBJ_ALLOCATED_TAG) {
1756 handle = head & ~OBJ_ALLOCATED_TAG;
1757 if (trypin_tag(handle))
1762 offset += class->size;
1766 kunmap_atomic(addr);
1773 struct zs_compact_control {
1774 /* Source spage for migration which could be a subpage of zspage */
1775 struct page *s_page;
1776 /* Destination page for migration which should be a first page
1778 struct page *d_page;
1779 /* Starting object index within @s_page which used for live object
1780 * in the subpage. */
1784 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1785 struct zs_compact_control *cc)
1787 unsigned long used_obj, free_obj;
1788 unsigned long handle;
1789 struct page *s_page = cc->s_page;
1790 struct page *d_page = cc->d_page;
1791 int obj_idx = cc->obj_idx;
1795 handle = find_alloced_obj(class, s_page, &obj_idx);
1797 s_page = get_next_page(s_page);
1804 /* Stop if there is no more space */
1805 if (zspage_full(class, get_zspage(d_page))) {
1811 used_obj = handle_to_obj(handle);
1812 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1813 zs_object_copy(class, free_obj, used_obj);
1816 * record_obj updates handle's value to free_obj and it will
1817 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1818 * breaks synchronization using pin_tag(e,g, zs_free) so
1819 * let's keep the lock bit.
1821 free_obj |= BIT(HANDLE_PIN_BIT);
1822 record_obj(handle, free_obj);
1824 obj_free(class, used_obj);
1827 /* Remember last position in this iteration */
1828 cc->s_page = s_page;
1829 cc->obj_idx = obj_idx;
1834 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1837 struct zspage *zspage;
1838 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1841 fg[0] = ZS_ALMOST_FULL;
1842 fg[1] = ZS_ALMOST_EMPTY;
1845 for (i = 0; i < 2; i++) {
1846 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1847 struct zspage, list);
1849 VM_BUG_ON(is_zspage_isolated(zspage));
1850 remove_zspage(class, zspage, fg[i]);
1859 * putback_zspage - add @zspage into right class's fullness list
1860 * @class: destination class
1861 * @zspage: target page
1863 * Return @zspage's fullness_group
1865 static enum fullness_group putback_zspage(struct size_class *class,
1866 struct zspage *zspage)
1868 enum fullness_group fullness;
1870 VM_BUG_ON(is_zspage_isolated(zspage));
1872 fullness = get_fullness_group(class, zspage);
1873 insert_zspage(class, zspage, fullness);
1874 set_zspage_mapping(zspage, class->index, fullness);
1879 #ifdef CONFIG_COMPACTION
1880 static struct dentry *zs_mount(struct file_system_type *fs_type,
1881 int flags, const char *dev_name, void *data)
1883 static const struct dentry_operations ops = {
1884 .d_dname = simple_dname,
1887 return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1890 static struct file_system_type zsmalloc_fs = {
1893 .kill_sb = kill_anon_super,
1896 static int zsmalloc_mount(void)
1900 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1901 if (IS_ERR(zsmalloc_mnt))
1902 ret = PTR_ERR(zsmalloc_mnt);
1907 static void zsmalloc_unmount(void)
1909 kern_unmount(zsmalloc_mnt);
1912 static void migrate_lock_init(struct zspage *zspage)
1914 rwlock_init(&zspage->lock);
1917 static void migrate_read_lock(struct zspage *zspage)
1919 read_lock(&zspage->lock);
1922 static void migrate_read_unlock(struct zspage *zspage)
1924 read_unlock(&zspage->lock);
1927 static void migrate_write_lock(struct zspage *zspage)
1929 write_lock(&zspage->lock);
1932 static void migrate_write_unlock(struct zspage *zspage)
1934 write_unlock(&zspage->lock);
1937 /* Number of isolated subpage for *page migration* in this zspage */
1938 static void inc_zspage_isolation(struct zspage *zspage)
1943 static void dec_zspage_isolation(struct zspage *zspage)
1948 static void putback_zspage_deferred(struct zs_pool *pool,
1949 struct size_class *class,
1950 struct zspage *zspage)
1952 enum fullness_group fg;
1954 fg = putback_zspage(class, zspage);
1956 schedule_work(&pool->free_work);
1960 static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1962 VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1963 atomic_long_dec(&pool->isolated_pages);
1965 * Checking pool->destroying must happen after atomic_long_dec()
1966 * for pool->isolated_pages above. Paired with the smp_mb() in
1967 * zs_unregister_migration().
1969 smp_mb__after_atomic();
1970 if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1971 wake_up_all(&pool->migration_wait);
1974 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1975 struct page *newpage, struct page *oldpage)
1978 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1981 page = get_first_page(zspage);
1983 if (page == oldpage)
1984 pages[idx] = newpage;
1988 } while ((page = get_next_page(page)) != NULL);
1990 create_page_chain(class, zspage, pages);
1991 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1992 if (unlikely(PageHugeObject(oldpage)))
1993 newpage->index = oldpage->index;
1994 __SetPageMovable(newpage, page_mapping(oldpage));
1997 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1999 struct zs_pool *pool;
2000 struct size_class *class;
2002 enum fullness_group fullness;
2003 struct zspage *zspage;
2004 struct address_space *mapping;
2007 * Page is locked so zspage couldn't be destroyed. For detail, look at
2008 * lock_zspage in free_zspage.
2010 VM_BUG_ON_PAGE(!PageMovable(page), page);
2011 VM_BUG_ON_PAGE(PageIsolated(page), page);
2013 zspage = get_zspage(page);
2016 * Without class lock, fullness could be stale while class_idx is okay
2017 * because class_idx is constant unless page is freed so we should get
2018 * fullness again under class lock.
2020 get_zspage_mapping(zspage, &class_idx, &fullness);
2021 mapping = page_mapping(page);
2022 pool = mapping->private_data;
2023 class = pool->size_class[class_idx];
2025 spin_lock(&class->lock);
2026 if (get_zspage_inuse(zspage) == 0) {
2027 spin_unlock(&class->lock);
2031 /* zspage is isolated for object migration */
2032 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
2033 spin_unlock(&class->lock);
2038 * If this is first time isolation for the zspage, isolate zspage from
2039 * size_class to prevent further object allocation from the zspage.
2041 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
2042 get_zspage_mapping(zspage, &class_idx, &fullness);
2043 atomic_long_inc(&pool->isolated_pages);
2044 remove_zspage(class, zspage, fullness);
2047 inc_zspage_isolation(zspage);
2048 spin_unlock(&class->lock);
2053 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
2054 struct page *page, enum migrate_mode mode)
2056 struct zs_pool *pool;
2057 struct size_class *class;
2059 enum fullness_group fullness;
2060 struct zspage *zspage;
2062 void *s_addr, *d_addr, *addr;
2064 unsigned long handle, head;
2065 unsigned long old_obj, new_obj;
2066 unsigned int obj_idx;
2069 VM_BUG_ON_PAGE(!PageMovable(page), page);
2070 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2072 zspage = get_zspage(page);
2074 /* Concurrent compactor cannot migrate any subpage in zspage */
2075 migrate_write_lock(zspage);
2076 get_zspage_mapping(zspage, &class_idx, &fullness);
2077 pool = mapping->private_data;
2078 class = pool->size_class[class_idx];
2079 offset = get_first_obj_offset(page);
2081 spin_lock(&class->lock);
2082 if (!get_zspage_inuse(zspage)) {
2088 s_addr = kmap_atomic(page);
2089 while (pos < PAGE_SIZE) {
2090 head = obj_to_head(page, s_addr + pos);
2091 if (head & OBJ_ALLOCATED_TAG) {
2092 handle = head & ~OBJ_ALLOCATED_TAG;
2093 if (!trypin_tag(handle))
2100 * Here, any user cannot access all objects in the zspage so let's move.
2102 d_addr = kmap_atomic(newpage);
2103 memcpy(d_addr, s_addr, PAGE_SIZE);
2104 kunmap_atomic(d_addr);
2106 for (addr = s_addr + offset; addr < s_addr + pos;
2107 addr += class->size) {
2108 head = obj_to_head(page, addr);
2109 if (head & OBJ_ALLOCATED_TAG) {
2110 handle = head & ~OBJ_ALLOCATED_TAG;
2111 if (!testpin_tag(handle))
2114 old_obj = handle_to_obj(handle);
2115 obj_to_location(old_obj, &dummy, &obj_idx);
2116 new_obj = (unsigned long)location_to_obj(newpage,
2118 new_obj |= BIT(HANDLE_PIN_BIT);
2119 record_obj(handle, new_obj);
2123 replace_sub_page(class, zspage, newpage, page);
2126 dec_zspage_isolation(zspage);
2129 * Page migration is done so let's putback isolated zspage to
2130 * the list if @page is final isolated subpage in the zspage.
2132 if (!is_zspage_isolated(zspage)) {
2134 * We cannot race with zs_destroy_pool() here because we wait
2135 * for isolation to hit zero before we start destroying.
2136 * Also, we ensure that everyone can see pool->destroying before
2139 putback_zspage_deferred(pool, class, zspage);
2140 zs_pool_dec_isolated(pool);
2143 if (page_zone(newpage) != page_zone(page)) {
2144 dec_zone_page_state(page, NR_ZSPAGES);
2145 inc_zone_page_state(newpage, NR_ZSPAGES);
2152 ret = MIGRATEPAGE_SUCCESS;
2154 for (addr = s_addr + offset; addr < s_addr + pos;
2155 addr += class->size) {
2156 head = obj_to_head(page, addr);
2157 if (head & OBJ_ALLOCATED_TAG) {
2158 handle = head & ~OBJ_ALLOCATED_TAG;
2159 if (!testpin_tag(handle))
2164 kunmap_atomic(s_addr);
2166 spin_unlock(&class->lock);
2167 migrate_write_unlock(zspage);
2172 void zs_page_putback(struct page *page)
2174 struct zs_pool *pool;
2175 struct size_class *class;
2177 enum fullness_group fg;
2178 struct address_space *mapping;
2179 struct zspage *zspage;
2181 VM_BUG_ON_PAGE(!PageMovable(page), page);
2182 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2184 zspage = get_zspage(page);
2185 get_zspage_mapping(zspage, &class_idx, &fg);
2186 mapping = page_mapping(page);
2187 pool = mapping->private_data;
2188 class = pool->size_class[class_idx];
2190 spin_lock(&class->lock);
2191 dec_zspage_isolation(zspage);
2192 if (!is_zspage_isolated(zspage)) {
2194 * Due to page_lock, we cannot free zspage immediately
2197 putback_zspage_deferred(pool, class, zspage);
2198 zs_pool_dec_isolated(pool);
2200 spin_unlock(&class->lock);
2203 const struct address_space_operations zsmalloc_aops = {
2204 .isolate_page = zs_page_isolate,
2205 .migratepage = zs_page_migrate,
2206 .putback_page = zs_page_putback,
2209 static int zs_register_migration(struct zs_pool *pool)
2211 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2212 if (IS_ERR(pool->inode)) {
2217 pool->inode->i_mapping->private_data = pool;
2218 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2222 static bool pool_isolated_are_drained(struct zs_pool *pool)
2224 return atomic_long_read(&pool->isolated_pages) == 0;
2227 /* Function for resolving migration */
2228 static void wait_for_isolated_drain(struct zs_pool *pool)
2232 * We're in the process of destroying the pool, so there are no
2233 * active allocations. zs_page_isolate() fails for completely free
2234 * zspages, so we need only wait for the zs_pool's isolated
2235 * count to hit zero.
2237 wait_event(pool->migration_wait,
2238 pool_isolated_are_drained(pool));
2241 static void zs_unregister_migration(struct zs_pool *pool)
2243 pool->destroying = true;
2245 * We need a memory barrier here to ensure global visibility of
2246 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2247 * case we don't care, or it will be > 0 and pool->destroying will
2248 * ensure that we wake up once isolation hits 0.
2251 wait_for_isolated_drain(pool); /* This can block */
2252 flush_work(&pool->free_work);
2257 * Caller should hold page_lock of all pages in the zspage
2258 * In here, we cannot use zspage meta data.
2260 static void async_free_zspage(struct work_struct *work)
2263 struct size_class *class;
2264 unsigned int class_idx;
2265 enum fullness_group fullness;
2266 struct zspage *zspage, *tmp;
2267 LIST_HEAD(free_pages);
2268 struct zs_pool *pool = container_of(work, struct zs_pool,
2271 for (i = 0; i < zs_size_classes; i++) {
2272 class = pool->size_class[i];
2273 if (class->index != i)
2276 spin_lock(&class->lock);
2277 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2278 spin_unlock(&class->lock);
2282 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2283 list_del(&zspage->list);
2284 lock_zspage(zspage);
2286 get_zspage_mapping(zspage, &class_idx, &fullness);
2287 VM_BUG_ON(fullness != ZS_EMPTY);
2288 class = pool->size_class[class_idx];
2289 spin_lock(&class->lock);
2290 __free_zspage(pool, pool->size_class[class_idx], zspage);
2291 spin_unlock(&class->lock);
2295 static void kick_deferred_free(struct zs_pool *pool)
2297 schedule_work(&pool->free_work);
2300 static void init_deferred_free(struct zs_pool *pool)
2302 INIT_WORK(&pool->free_work, async_free_zspage);
2305 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2307 struct page *page = get_first_page(zspage);
2310 WARN_ON(!trylock_page(page));
2311 __SetPageMovable(page, pool->inode->i_mapping);
2313 } while ((page = get_next_page(page)) != NULL);
2319 * Based on the number of unused allocated objects calculate
2320 * and return the number of pages that we can free.
2322 static unsigned long zs_can_compact(struct size_class *class)
2324 unsigned long obj_wasted;
2325 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2326 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2328 if (obj_allocated <= obj_used)
2331 obj_wasted = obj_allocated - obj_used;
2332 obj_wasted /= class->objs_per_zspage;
2334 return obj_wasted * class->pages_per_zspage;
2337 static unsigned long __zs_compact(struct zs_pool *pool,
2338 struct size_class *class)
2340 struct zs_compact_control cc;
2341 struct zspage *src_zspage;
2342 struct zspage *dst_zspage = NULL;
2343 unsigned long pages_freed = 0;
2345 spin_lock(&class->lock);
2346 while ((src_zspage = isolate_zspage(class, true))) {
2348 if (!zs_can_compact(class))
2352 cc.s_page = get_first_page(src_zspage);
2354 while ((dst_zspage = isolate_zspage(class, false))) {
2355 cc.d_page = get_first_page(dst_zspage);
2357 * If there is no more space in dst_page, resched
2358 * and see if anyone had allocated another zspage.
2360 if (!migrate_zspage(pool, class, &cc))
2363 putback_zspage(class, dst_zspage);
2366 /* Stop if we couldn't find slot */
2367 if (dst_zspage == NULL)
2370 putback_zspage(class, dst_zspage);
2371 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2372 free_zspage(pool, class, src_zspage);
2373 pages_freed += class->pages_per_zspage;
2375 spin_unlock(&class->lock);
2377 spin_lock(&class->lock);
2381 putback_zspage(class, src_zspage);
2383 spin_unlock(&class->lock);
2388 unsigned long zs_compact(struct zs_pool *pool)
2391 struct size_class *class;
2392 unsigned long pages_freed = 0;
2394 for (i = zs_size_classes - 1; i >= 0; i--) {
2395 class = pool->size_class[i];
2398 if (class->index != i)
2400 pages_freed += __zs_compact(pool, class);
2402 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2406 EXPORT_SYMBOL_GPL(zs_compact);
2408 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2410 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2412 EXPORT_SYMBOL_GPL(zs_pool_stats);
2414 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2415 struct shrink_control *sc)
2417 unsigned long pages_freed;
2418 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2422 * Compact classes and calculate compaction delta.
2423 * Can run concurrently with a manually triggered
2424 * (by user) compaction.
2426 pages_freed = zs_compact(pool);
2428 return pages_freed ? pages_freed : SHRINK_STOP;
2431 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2432 struct shrink_control *sc)
2435 struct size_class *class;
2436 unsigned long pages_to_free = 0;
2437 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2440 for (i = zs_size_classes - 1; i >= 0; i--) {
2441 class = pool->size_class[i];
2444 if (class->index != i)
2447 pages_to_free += zs_can_compact(class);
2450 return pages_to_free;
2453 static void zs_unregister_shrinker(struct zs_pool *pool)
2455 if (pool->shrinker_enabled) {
2456 unregister_shrinker(&pool->shrinker);
2457 pool->shrinker_enabled = false;
2461 static int zs_register_shrinker(struct zs_pool *pool)
2463 pool->shrinker.scan_objects = zs_shrinker_scan;
2464 pool->shrinker.count_objects = zs_shrinker_count;
2465 pool->shrinker.batch = 0;
2466 pool->shrinker.seeks = DEFAULT_SEEKS;
2468 return register_shrinker(&pool->shrinker);
2472 * zs_create_pool - Creates an allocation pool to work from.
2473 * @name: pool name to be created
2475 * This function must be called before anything when using
2476 * the zsmalloc allocator.
2478 * On success, a pointer to the newly created pool is returned,
2481 struct zs_pool *zs_create_pool(const char *name)
2484 struct zs_pool *pool;
2485 struct size_class *prev_class = NULL;
2487 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2491 init_deferred_free(pool);
2492 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
2494 if (!pool->size_class) {
2499 pool->name = kstrdup(name, GFP_KERNEL);
2503 #ifdef CONFIG_COMPACTION
2504 init_waitqueue_head(&pool->migration_wait);
2507 if (create_cache(pool))
2511 * Iterate reversly, because, size of size_class that we want to use
2512 * for merging should be larger or equal to current size.
2514 for (i = zs_size_classes - 1; i >= 0; i--) {
2516 int pages_per_zspage;
2517 int objs_per_zspage;
2518 struct size_class *class;
2521 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2522 if (size > ZS_MAX_ALLOC_SIZE)
2523 size = ZS_MAX_ALLOC_SIZE;
2524 pages_per_zspage = get_pages_per_zspage(size);
2525 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2528 * size_class is used for normal zsmalloc operation such
2529 * as alloc/free for that size. Although it is natural that we
2530 * have one size_class for each size, there is a chance that we
2531 * can get more memory utilization if we use one size_class for
2532 * many different sizes whose size_class have same
2533 * characteristics. So, we makes size_class point to
2534 * previous size_class if possible.
2537 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2538 pool->size_class[i] = prev_class;
2543 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2549 class->pages_per_zspage = pages_per_zspage;
2550 class->objs_per_zspage = objs_per_zspage;
2551 spin_lock_init(&class->lock);
2552 pool->size_class[i] = class;
2553 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2555 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2560 /* debug only, don't abort if it fails */
2561 zs_pool_stat_create(pool, name);
2563 if (zs_register_migration(pool))
2567 * Not critical, we still can use the pool
2568 * and user can trigger compaction manually.
2570 if (zs_register_shrinker(pool) == 0)
2571 pool->shrinker_enabled = true;
2575 zs_destroy_pool(pool);
2578 EXPORT_SYMBOL_GPL(zs_create_pool);
2580 void zs_destroy_pool(struct zs_pool *pool)
2584 zs_unregister_shrinker(pool);
2585 zs_unregister_migration(pool);
2586 zs_pool_stat_destroy(pool);
2588 for (i = 0; i < zs_size_classes; i++) {
2590 struct size_class *class = pool->size_class[i];
2595 if (class->index != i)
2598 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2599 if (!list_empty(&class->fullness_list[fg])) {
2600 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2607 destroy_cache(pool);
2608 kfree(pool->size_class);
2612 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2614 static int __init zs_init(void)
2618 ret = zsmalloc_mount();
2622 ret = zs_register_cpu_notifier();
2627 init_zs_size_classes();
2630 zpool_register_driver(&zs_zpool_driver);
2638 zs_unregister_cpu_notifier();
2644 static void __exit zs_exit(void)
2647 zpool_unregister_driver(&zs_zpool_driver);
2650 zs_unregister_cpu_notifier();
2655 module_init(zs_init);
2656 module_exit(zs_exit);
2658 MODULE_LICENSE("Dual BSD/GPL");
2659 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");