2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
5 * Copyright (C) 2012, 2013 Minchan Kim
7 * This code is released using a dual license strategy: BSD/GPL
8 * You can choose the license that better fits your requirements.
10 * Released under the terms of 3-clause BSD License
11 * Released under the terms of GNU General Public License Version 2.0
15 * Following is how we use various fields and flags of underlying
16 * struct page(s) to form a zspage.
18 * Usage of struct page fields:
19 * page->private: points to zspage
20 * page->index: links together all component pages of a zspage
21 * For the huge page, this is always 0, so we use this field
23 * page->page_type: first object offset in a subpage of zspage
25 * Usage of struct page flags:
26 * PG_private: identifies the first component page
27 * PG_owner_priv_1: identifies the huge component page
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
33 #include <linux/module.h>
34 #include <linux/kernel.h>
35 #include <linux/sched.h>
36 #include <linux/magic.h>
37 #include <linux/bitops.h>
38 #include <linux/errno.h>
39 #include <linux/highmem.h>
40 #include <linux/string.h>
41 #include <linux/slab.h>
42 #include <linux/pgtable.h>
43 #include <asm/tlbflush.h>
44 #include <linux/cpumask.h>
45 #include <linux/cpu.h>
46 #include <linux/vmalloc.h>
47 #include <linux/preempt.h>
48 #include <linux/spinlock.h>
49 #include <linux/shrinker.h>
50 #include <linux/types.h>
51 #include <linux/debugfs.h>
52 #include <linux/zsmalloc.h>
53 #include <linux/zpool.h>
54 #include <linux/mount.h>
55 #include <linux/pseudo_fs.h>
56 #include <linux/migrate.h>
57 #include <linux/wait.h>
58 #include <linux/pagemap.h>
61 #define ZSPAGE_MAGIC 0x58
64 * This must be power of 2 and greater than or equal to sizeof(link_free).
65 * These two conditions ensure that any 'struct link_free' itself doesn't
66 * span more than 1 page which avoids complex case of mapping 2 pages simply
67 * to restore link_free pointer values.
72 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
73 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
75 #define ZS_MAX_ZSPAGE_ORDER 2
76 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
78 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
81 * Object location (<PFN>, <obj_idx>) is encoded as
82 * a single (unsigned long) handle value.
84 * Note that object index <obj_idx> starts from 0.
86 * This is made more complicated by various memory models and PAE.
89 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
90 #ifdef MAX_PHYSMEM_BITS
91 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
94 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
97 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
101 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
104 * Memory for allocating for handle keeps object position by
105 * encoding <page, obj_idx> and the encoded value has a room
106 * in least bit(ie, look at obj_to_location).
107 * We use the bit to synchronize between object access by
108 * user and migration.
110 #define HANDLE_PIN_BIT 0
113 * Head in allocated object should have OBJ_ALLOCATED_TAG
114 * to identify the object was allocated or not.
115 * It's okay to add the status bit in the least bit because
116 * header keeps handle which is 4byte-aligned address so we
117 * have room for two bit at least.
119 #define OBJ_ALLOCATED_TAG 1
120 #define OBJ_TAG_BITS 1
121 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
122 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
124 #define FULLNESS_BITS 2
126 #define ISOLATED_BITS 3
127 #define MAGIC_VAL_BITS 8
129 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
130 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
131 #define ZS_MIN_ALLOC_SIZE \
132 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
133 /* each chunk includes extra space to keep handle */
134 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
137 * On systems with 4K page size, this gives 255 size classes! There is a
139 * - Large number of size classes is potentially wasteful as free page are
140 * spread across these classes
141 * - Small number of size classes causes large internal fragmentation
142 * - Probably its better to use specific size classes (empirically
143 * determined). NOTE: all those class sizes must be set as multiple of
144 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
146 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
149 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
150 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
151 ZS_SIZE_CLASS_DELTA) + 1)
153 enum fullness_group {
161 enum class_stat_type {
171 struct zs_size_stat {
172 unsigned long objs[NR_ZS_STAT_TYPE];
175 #ifdef CONFIG_ZSMALLOC_STAT
176 static struct dentry *zs_stat_root;
179 #ifdef CONFIG_COMPACTION
180 static struct vfsmount *zsmalloc_mnt;
184 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
186 * n = number of allocated objects
187 * N = total number of objects zspage can store
188 * f = fullness_threshold_frac
190 * Similarly, we assign zspage to:
191 * ZS_ALMOST_FULL when n > N / f
192 * ZS_EMPTY when n == 0
193 * ZS_FULL when n == N
195 * (see: fix_fullness_group())
197 static const int fullness_threshold_frac = 4;
198 static size_t huge_class_size;
202 struct list_head fullness_list[NR_ZS_FULLNESS];
204 * Size of objects stored in this class. Must be multiple
209 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
210 int pages_per_zspage;
213 struct zs_size_stat stats;
216 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
217 static void SetPageHugeObject(struct page *page)
219 SetPageOwnerPriv1(page);
222 static void ClearPageHugeObject(struct page *page)
224 ClearPageOwnerPriv1(page);
227 static int PageHugeObject(struct page *page)
229 return PageOwnerPriv1(page);
233 * Placed within free objects to form a singly linked list.
234 * For every zspage, zspage->freeobj gives head of this list.
236 * This must be power of 2 and less than or equal to ZS_ALIGN
242 * It's valid for non-allocated object
246 * Handle of allocated object.
248 unsigned long handle;
255 struct size_class *size_class[ZS_SIZE_CLASSES];
256 struct kmem_cache *handle_cachep;
257 struct kmem_cache *zspage_cachep;
259 atomic_long_t pages_allocated;
261 struct zs_pool_stats stats;
263 /* Compact classes */
264 struct shrinker shrinker;
266 #ifdef CONFIG_ZSMALLOC_STAT
267 struct dentry *stat_dentry;
269 #ifdef CONFIG_COMPACTION
271 struct work_struct free_work;
272 /* A wait queue for when migration races with async_free_zspage() */
273 struct wait_queue_head migration_wait;
274 atomic_long_t isolated_pages;
281 unsigned int fullness:FULLNESS_BITS;
282 unsigned int class:CLASS_BITS + 1;
283 unsigned int isolated:ISOLATED_BITS;
284 unsigned int magic:MAGIC_VAL_BITS;
287 unsigned int freeobj;
288 struct page *first_page;
289 struct list_head list; /* fullness list */
290 #ifdef CONFIG_COMPACTION
295 struct mapping_area {
296 char *vm_buf; /* copy buffer for objects that span pages */
297 char *vm_addr; /* address of kmap_atomic()'ed pages */
298 enum zs_mapmode vm_mm; /* mapping mode */
301 #ifdef CONFIG_COMPACTION
302 static int zs_register_migration(struct zs_pool *pool);
303 static void zs_unregister_migration(struct zs_pool *pool);
304 static void migrate_lock_init(struct zspage *zspage);
305 static void migrate_read_lock(struct zspage *zspage);
306 static void migrate_read_unlock(struct zspage *zspage);
307 static void kick_deferred_free(struct zs_pool *pool);
308 static void init_deferred_free(struct zs_pool *pool);
309 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
311 static int zsmalloc_mount(void) { return 0; }
312 static void zsmalloc_unmount(void) {}
313 static int zs_register_migration(struct zs_pool *pool) { return 0; }
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) {}
323 static int create_cache(struct zs_pool *pool)
325 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
327 if (!pool->handle_cachep)
330 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
332 if (!pool->zspage_cachep) {
333 kmem_cache_destroy(pool->handle_cachep);
334 pool->handle_cachep = NULL;
341 static void destroy_cache(struct zs_pool *pool)
343 kmem_cache_destroy(pool->handle_cachep);
344 kmem_cache_destroy(pool->zspage_cachep);
347 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
349 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
350 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
353 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
355 kmem_cache_free(pool->handle_cachep, (void *)handle);
358 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
360 return kmem_cache_zalloc(pool->zspage_cachep,
361 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
364 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
366 kmem_cache_free(pool->zspage_cachep, zspage);
369 static void record_obj(unsigned long handle, unsigned long obj)
372 * lsb of @obj represents handle lock while other bits
373 * represent object value the handle is pointing so
374 * updating shouldn't do store tearing.
376 WRITE_ONCE(*(unsigned long *)handle, obj);
383 static void *zs_zpool_create(const char *name, gfp_t gfp,
384 const struct zpool_ops *zpool_ops,
388 * Ignore global gfp flags: zs_malloc() may be invoked from
389 * different contexts and its caller must provide a valid
392 return zs_create_pool(name);
395 static void zs_zpool_destroy(void *pool)
397 zs_destroy_pool(pool);
400 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
401 unsigned long *handle)
403 *handle = zs_malloc(pool, size, gfp);
404 return *handle ? 0 : -1;
406 static void zs_zpool_free(void *pool, unsigned long handle)
408 zs_free(pool, handle);
411 static void *zs_zpool_map(void *pool, unsigned long handle,
412 enum zpool_mapmode mm)
414 enum zs_mapmode zs_mm;
429 return zs_map_object(pool, handle, zs_mm);
431 static void zs_zpool_unmap(void *pool, unsigned long handle)
433 zs_unmap_object(pool, handle);
436 static u64 zs_zpool_total_size(void *pool)
438 return zs_get_total_pages(pool) << PAGE_SHIFT;
441 static struct zpool_driver zs_zpool_driver = {
443 .owner = THIS_MODULE,
444 .create = zs_zpool_create,
445 .destroy = zs_zpool_destroy,
446 .malloc_support_movable = true,
447 .malloc = zs_zpool_malloc,
448 .free = zs_zpool_free,
450 .unmap = zs_zpool_unmap,
451 .total_size = zs_zpool_total_size,
454 MODULE_ALIAS("zpool-zsmalloc");
455 #endif /* CONFIG_ZPOOL */
457 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
458 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
460 static bool is_zspage_isolated(struct zspage *zspage)
462 return zspage->isolated;
465 static __maybe_unused int is_first_page(struct page *page)
467 return PagePrivate(page);
470 /* Protected by class->lock */
471 static inline int get_zspage_inuse(struct zspage *zspage)
473 return zspage->inuse;
477 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
479 zspage->inuse += val;
482 static inline struct page *get_first_page(struct zspage *zspage)
484 struct page *first_page = zspage->first_page;
486 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
490 static inline int get_first_obj_offset(struct page *page)
492 return page->page_type;
495 static inline void set_first_obj_offset(struct page *page, int offset)
497 page->page_type = offset;
500 static inline unsigned int get_freeobj(struct zspage *zspage)
502 return zspage->freeobj;
505 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
507 zspage->freeobj = obj;
510 static void get_zspage_mapping(struct zspage *zspage,
511 unsigned int *class_idx,
512 enum fullness_group *fullness)
514 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
516 *fullness = zspage->fullness;
517 *class_idx = zspage->class;
520 static struct size_class *zspage_class(struct zs_pool *pool,
521 struct zspage *zspage)
523 return pool->size_class[zspage->class];
526 static void set_zspage_mapping(struct zspage *zspage,
527 unsigned int class_idx,
528 enum fullness_group fullness)
530 zspage->class = class_idx;
531 zspage->fullness = fullness;
535 * zsmalloc divides the pool into various size classes where each
536 * class maintains a list of zspages where each zspage is divided
537 * into equal sized chunks. Each allocation falls into one of these
538 * classes depending on its size. This function returns index of the
539 * size class which has chunk size big enough to hold the given size.
541 static int get_size_class_index(int size)
545 if (likely(size > ZS_MIN_ALLOC_SIZE))
546 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
547 ZS_SIZE_CLASS_DELTA);
549 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
552 /* type can be of enum type class_stat_type or fullness_group */
553 static inline void class_stat_inc(struct size_class *class,
554 int type, unsigned long cnt)
556 class->stats.objs[type] += cnt;
559 /* type can be of enum type class_stat_type or fullness_group */
560 static inline void class_stat_dec(struct size_class *class,
561 int type, unsigned long cnt)
563 class->stats.objs[type] -= cnt;
566 /* type can be of enum type class_stat_type or fullness_group */
567 static inline unsigned long zs_stat_get(struct size_class *class,
570 return class->stats.objs[type];
573 #ifdef CONFIG_ZSMALLOC_STAT
575 static void __init zs_stat_init(void)
577 if (!debugfs_initialized()) {
578 pr_warn("debugfs not available, stat dir not created\n");
582 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
585 static void __exit zs_stat_exit(void)
587 debugfs_remove_recursive(zs_stat_root);
590 static unsigned long zs_can_compact(struct size_class *class);
592 static int zs_stats_size_show(struct seq_file *s, void *v)
595 struct zs_pool *pool = s->private;
596 struct size_class *class;
598 unsigned long class_almost_full, class_almost_empty;
599 unsigned long obj_allocated, obj_used, pages_used, freeable;
600 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
601 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
602 unsigned long total_freeable = 0;
604 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
605 "class", "size", "almost_full", "almost_empty",
606 "obj_allocated", "obj_used", "pages_used",
607 "pages_per_zspage", "freeable");
609 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
610 class = pool->size_class[i];
612 if (class->index != i)
615 spin_lock(&class->lock);
616 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
617 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
618 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
619 obj_used = zs_stat_get(class, OBJ_USED);
620 freeable = zs_can_compact(class);
621 spin_unlock(&class->lock);
623 objs_per_zspage = class->objs_per_zspage;
624 pages_used = obj_allocated / objs_per_zspage *
625 class->pages_per_zspage;
627 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
628 " %10lu %10lu %16d %8lu\n",
629 i, class->size, class_almost_full, class_almost_empty,
630 obj_allocated, obj_used, pages_used,
631 class->pages_per_zspage, freeable);
633 total_class_almost_full += class_almost_full;
634 total_class_almost_empty += class_almost_empty;
635 total_objs += obj_allocated;
636 total_used_objs += obj_used;
637 total_pages += pages_used;
638 total_freeable += freeable;
642 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
643 "Total", "", total_class_almost_full,
644 total_class_almost_empty, total_objs,
645 total_used_objs, total_pages, "", total_freeable);
649 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
651 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
654 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
658 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
660 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
661 &zs_stats_size_fops);
664 static void zs_pool_stat_destroy(struct zs_pool *pool)
666 debugfs_remove_recursive(pool->stat_dentry);
669 #else /* CONFIG_ZSMALLOC_STAT */
670 static void __init zs_stat_init(void)
674 static void __exit zs_stat_exit(void)
678 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
682 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
689 * For each size class, zspages are divided into different groups
690 * depending on how "full" they are. This was done so that we could
691 * easily find empty or nearly empty zspages when we try to shrink
692 * the pool (not yet implemented). This function returns fullness
693 * status of the given page.
695 static enum fullness_group get_fullness_group(struct size_class *class,
696 struct zspage *zspage)
698 int inuse, objs_per_zspage;
699 enum fullness_group fg;
701 inuse = get_zspage_inuse(zspage);
702 objs_per_zspage = class->objs_per_zspage;
706 else if (inuse == objs_per_zspage)
708 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
709 fg = ZS_ALMOST_EMPTY;
717 * Each size class maintains various freelists and zspages are assigned
718 * to one of these freelists based on the number of live objects they
719 * have. This functions inserts the given zspage into the freelist
720 * identified by <class, fullness_group>.
722 static void insert_zspage(struct size_class *class,
723 struct zspage *zspage,
724 enum fullness_group fullness)
728 class_stat_inc(class, fullness, 1);
729 head = list_first_entry_or_null(&class->fullness_list[fullness],
730 struct zspage, list);
732 * We want to see more ZS_FULL pages and less almost empty/full.
733 * Put pages with higher ->inuse first.
735 if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head))
736 list_add(&zspage->list, &head->list);
738 list_add(&zspage->list, &class->fullness_list[fullness]);
742 * This function removes the given zspage from the freelist identified
743 * by <class, fullness_group>.
745 static void remove_zspage(struct size_class *class,
746 struct zspage *zspage,
747 enum fullness_group fullness)
749 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
750 VM_BUG_ON(is_zspage_isolated(zspage));
752 list_del_init(&zspage->list);
753 class_stat_dec(class, fullness, 1);
757 * Each size class maintains zspages in different fullness groups depending
758 * on the number of live objects they contain. When allocating or freeing
759 * objects, the fullness status of the page can change, say, from ALMOST_FULL
760 * to ALMOST_EMPTY when freeing an object. This function checks if such
761 * a status change has occurred for the given page and accordingly moves the
762 * page from the freelist of the old fullness group to that of the new
765 static enum fullness_group fix_fullness_group(struct size_class *class,
766 struct zspage *zspage)
769 enum fullness_group currfg, newfg;
771 get_zspage_mapping(zspage, &class_idx, &currfg);
772 newfg = get_fullness_group(class, zspage);
776 if (!is_zspage_isolated(zspage)) {
777 remove_zspage(class, zspage, currfg);
778 insert_zspage(class, zspage, newfg);
781 set_zspage_mapping(zspage, class_idx, newfg);
788 * We have to decide on how many pages to link together
789 * to form a zspage for each size class. This is important
790 * to reduce wastage due to unusable space left at end of
791 * each zspage which is given as:
792 * wastage = Zp % class_size
793 * usage = Zp - wastage
794 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
796 * For example, for size class of 3/8 * PAGE_SIZE, we should
797 * link together 3 PAGE_SIZE sized pages to form a zspage
798 * since then we can perfectly fit in 8 such objects.
800 static int get_pages_per_zspage(int class_size)
802 int i, max_usedpc = 0;
803 /* zspage order which gives maximum used size per KB */
804 int max_usedpc_order = 1;
806 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
810 zspage_size = i * PAGE_SIZE;
811 waste = zspage_size % class_size;
812 usedpc = (zspage_size - waste) * 100 / zspage_size;
814 if (usedpc > max_usedpc) {
816 max_usedpc_order = i;
820 return max_usedpc_order;
823 static struct zspage *get_zspage(struct page *page)
825 struct zspage *zspage = (struct zspage *)page_private(page);
827 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
831 static struct page *get_next_page(struct page *page)
833 if (unlikely(PageHugeObject(page)))
836 return (struct page *)page->index;
840 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
841 * @obj: the encoded object value
842 * @page: page object resides in zspage
843 * @obj_idx: object index
845 static void obj_to_location(unsigned long obj, struct page **page,
846 unsigned int *obj_idx)
848 obj >>= OBJ_TAG_BITS;
849 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
850 *obj_idx = (obj & OBJ_INDEX_MASK);
853 static void obj_to_page(unsigned long obj, struct page **page)
855 obj >>= OBJ_TAG_BITS;
856 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
860 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
861 * @page: page object resides in zspage
862 * @obj_idx: object index
864 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
868 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
869 obj |= obj_idx & OBJ_INDEX_MASK;
870 obj <<= OBJ_TAG_BITS;
875 static unsigned long handle_to_obj(unsigned long handle)
877 return *(unsigned long *)handle;
880 static unsigned long obj_to_head(struct page *page, void *obj)
882 if (unlikely(PageHugeObject(page))) {
883 VM_BUG_ON_PAGE(!is_first_page(page), page);
886 return *(unsigned long *)obj;
889 static inline int testpin_tag(unsigned long handle)
891 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
894 static inline int trypin_tag(unsigned long handle)
896 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
899 static void pin_tag(unsigned long handle) __acquires(bitlock)
901 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
904 static void unpin_tag(unsigned long handle) __releases(bitlock)
906 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
909 static void reset_page(struct page *page)
911 __ClearPageMovable(page);
912 ClearPagePrivate(page);
913 set_page_private(page, 0);
914 page_mapcount_reset(page);
915 ClearPageHugeObject(page);
919 static int trylock_zspage(struct zspage *zspage)
921 struct page *cursor, *fail;
923 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
924 get_next_page(cursor)) {
925 if (!trylock_page(cursor)) {
933 for (cursor = get_first_page(zspage); cursor != fail; cursor =
934 get_next_page(cursor))
940 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
941 struct zspage *zspage)
943 struct page *page, *next;
944 enum fullness_group fg;
945 unsigned int class_idx;
947 get_zspage_mapping(zspage, &class_idx, &fg);
949 assert_spin_locked(&class->lock);
951 VM_BUG_ON(get_zspage_inuse(zspage));
952 VM_BUG_ON(fg != ZS_EMPTY);
954 next = page = get_first_page(zspage);
956 VM_BUG_ON_PAGE(!PageLocked(page), page);
957 next = get_next_page(page);
960 dec_zone_page_state(page, NR_ZSPAGES);
963 } while (page != NULL);
965 cache_free_zspage(pool, zspage);
967 class_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
968 atomic_long_sub(class->pages_per_zspage,
969 &pool->pages_allocated);
972 static void free_zspage(struct zs_pool *pool, struct size_class *class,
973 struct zspage *zspage)
975 VM_BUG_ON(get_zspage_inuse(zspage));
976 VM_BUG_ON(list_empty(&zspage->list));
978 if (!trylock_zspage(zspage)) {
979 kick_deferred_free(pool);
983 remove_zspage(class, zspage, ZS_EMPTY);
984 __free_zspage(pool, class, zspage);
987 /* Initialize a newly allocated zspage */
988 static void init_zspage(struct size_class *class, struct zspage *zspage)
990 unsigned int freeobj = 1;
991 unsigned long off = 0;
992 struct page *page = get_first_page(zspage);
995 struct page *next_page;
996 struct link_free *link;
999 set_first_obj_offset(page, off);
1001 vaddr = kmap_atomic(page);
1002 link = (struct link_free *)vaddr + off / sizeof(*link);
1004 while ((off += class->size) < PAGE_SIZE) {
1005 link->next = freeobj++ << OBJ_TAG_BITS;
1006 link += class->size / sizeof(*link);
1010 * We now come to the last (full or partial) object on this
1011 * page, which must point to the first object on the next
1014 next_page = get_next_page(page);
1016 link->next = freeobj++ << OBJ_TAG_BITS;
1019 * Reset OBJ_TAG_BITS bit to last link to tell
1020 * whether it's allocated object or not.
1022 link->next = -1UL << OBJ_TAG_BITS;
1024 kunmap_atomic(vaddr);
1029 set_freeobj(zspage, 0);
1032 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1033 struct page *pages[])
1037 struct page *prev_page = NULL;
1038 int nr_pages = class->pages_per_zspage;
1041 * Allocate individual pages and link them together as:
1042 * 1. all pages are linked together using page->index
1043 * 2. each sub-page point to zspage using page->private
1045 * we set PG_private to identify the first page (i.e. no other sub-page
1046 * has this flag set).
1048 for (i = 0; i < nr_pages; i++) {
1050 set_page_private(page, (unsigned long)zspage);
1053 zspage->first_page = page;
1054 SetPagePrivate(page);
1055 if (unlikely(class->objs_per_zspage == 1 &&
1056 class->pages_per_zspage == 1))
1057 SetPageHugeObject(page);
1059 prev_page->index = (unsigned long)page;
1066 * Allocate a zspage for the given size class
1068 static struct zspage *alloc_zspage(struct zs_pool *pool,
1069 struct size_class *class,
1073 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1074 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1079 zspage->magic = ZSPAGE_MAGIC;
1080 migrate_lock_init(zspage);
1082 for (i = 0; i < class->pages_per_zspage; i++) {
1085 page = alloc_page(gfp);
1088 dec_zone_page_state(pages[i], NR_ZSPAGES);
1089 __free_page(pages[i]);
1091 cache_free_zspage(pool, zspage);
1095 inc_zone_page_state(page, NR_ZSPAGES);
1099 create_page_chain(class, zspage, pages);
1100 init_zspage(class, zspage);
1105 static struct zspage *find_get_zspage(struct size_class *class)
1108 struct zspage *zspage;
1110 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1111 zspage = list_first_entry_or_null(&class->fullness_list[i],
1112 struct zspage, list);
1120 static inline int __zs_cpu_up(struct mapping_area *area)
1123 * Make sure we don't leak memory if a cpu UP notification
1124 * and zs_init() race and both call zs_cpu_up() on the same cpu
1128 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1134 static inline void __zs_cpu_down(struct mapping_area *area)
1136 kfree(area->vm_buf);
1137 area->vm_buf = NULL;
1140 static void *__zs_map_object(struct mapping_area *area,
1141 struct page *pages[2], int off, int size)
1145 char *buf = area->vm_buf;
1147 /* disable page faults to match kmap_atomic() return conditions */
1148 pagefault_disable();
1150 /* no read fastpath */
1151 if (area->vm_mm == ZS_MM_WO)
1154 sizes[0] = PAGE_SIZE - off;
1155 sizes[1] = size - sizes[0];
1157 /* copy object to per-cpu buffer */
1158 addr = kmap_atomic(pages[0]);
1159 memcpy(buf, addr + off, sizes[0]);
1160 kunmap_atomic(addr);
1161 addr = kmap_atomic(pages[1]);
1162 memcpy(buf + sizes[0], addr, sizes[1]);
1163 kunmap_atomic(addr);
1165 return area->vm_buf;
1168 static void __zs_unmap_object(struct mapping_area *area,
1169 struct page *pages[2], int off, int size)
1175 /* no write fastpath */
1176 if (area->vm_mm == ZS_MM_RO)
1180 buf = buf + ZS_HANDLE_SIZE;
1181 size -= ZS_HANDLE_SIZE;
1182 off += ZS_HANDLE_SIZE;
1184 sizes[0] = PAGE_SIZE - off;
1185 sizes[1] = size - sizes[0];
1187 /* copy per-cpu buffer to object */
1188 addr = kmap_atomic(pages[0]);
1189 memcpy(addr + off, buf, sizes[0]);
1190 kunmap_atomic(addr);
1191 addr = kmap_atomic(pages[1]);
1192 memcpy(addr, buf + sizes[0], sizes[1]);
1193 kunmap_atomic(addr);
1196 /* enable page faults to match kunmap_atomic() return conditions */
1200 static int zs_cpu_prepare(unsigned int cpu)
1202 struct mapping_area *area;
1204 area = &per_cpu(zs_map_area, cpu);
1205 return __zs_cpu_up(area);
1208 static int zs_cpu_dead(unsigned int cpu)
1210 struct mapping_area *area;
1212 area = &per_cpu(zs_map_area, cpu);
1213 __zs_cpu_down(area);
1217 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1218 int objs_per_zspage)
1220 if (prev->pages_per_zspage == pages_per_zspage &&
1221 prev->objs_per_zspage == objs_per_zspage)
1227 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1229 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1232 unsigned long zs_get_total_pages(struct zs_pool *pool)
1234 return atomic_long_read(&pool->pages_allocated);
1236 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1239 * zs_map_object - get address of allocated object from handle.
1240 * @pool: pool from which the object was allocated
1241 * @handle: handle returned from zs_malloc
1242 * @mm: mapping mode to use
1244 * Before using an object allocated from zs_malloc, it must be mapped using
1245 * this function. When done with the object, it must be unmapped using
1248 * Only one object can be mapped per cpu at a time. There is no protection
1249 * against nested mappings.
1251 * This function returns with preemption and page faults disabled.
1253 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1256 struct zspage *zspage;
1258 unsigned long obj, off;
1259 unsigned int obj_idx;
1261 struct size_class *class;
1262 struct mapping_area *area;
1263 struct page *pages[2];
1267 * Because we use per-cpu mapping areas shared among the
1268 * pools/users, we can't allow mapping in interrupt context
1269 * because it can corrupt another users mappings.
1271 BUG_ON(in_interrupt());
1273 /* From now on, migration cannot move the object */
1276 obj = handle_to_obj(handle);
1277 obj_to_location(obj, &page, &obj_idx);
1278 zspage = get_zspage(page);
1280 /* migration cannot move any subpage in this zspage */
1281 migrate_read_lock(zspage);
1283 class = zspage_class(pool, zspage);
1284 off = (class->size * obj_idx) & ~PAGE_MASK;
1286 area = &get_cpu_var(zs_map_area);
1288 if (off + class->size <= PAGE_SIZE) {
1289 /* this object is contained entirely within a page */
1290 area->vm_addr = kmap_atomic(page);
1291 ret = area->vm_addr + off;
1295 /* this object spans two pages */
1297 pages[1] = get_next_page(page);
1300 ret = __zs_map_object(area, pages, off, class->size);
1302 if (likely(!PageHugeObject(page)))
1303 ret += ZS_HANDLE_SIZE;
1307 EXPORT_SYMBOL_GPL(zs_map_object);
1309 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1311 struct zspage *zspage;
1313 unsigned long obj, off;
1314 unsigned int obj_idx;
1316 struct size_class *class;
1317 struct mapping_area *area;
1319 obj = handle_to_obj(handle);
1320 obj_to_location(obj, &page, &obj_idx);
1321 zspage = get_zspage(page);
1322 class = zspage_class(pool, zspage);
1323 off = (class->size * obj_idx) & ~PAGE_MASK;
1325 area = this_cpu_ptr(&zs_map_area);
1326 if (off + class->size <= PAGE_SIZE)
1327 kunmap_atomic(area->vm_addr);
1329 struct page *pages[2];
1332 pages[1] = get_next_page(page);
1335 __zs_unmap_object(area, pages, off, class->size);
1337 put_cpu_var(zs_map_area);
1339 migrate_read_unlock(zspage);
1342 EXPORT_SYMBOL_GPL(zs_unmap_object);
1345 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1346 * zsmalloc &size_class.
1347 * @pool: zsmalloc pool to use
1349 * The function returns the size of the first huge class - any object of equal
1350 * or bigger size will be stored in zspage consisting of a single physical
1353 * Context: Any context.
1355 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1357 size_t zs_huge_class_size(struct zs_pool *pool)
1359 return huge_class_size;
1361 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1363 static unsigned long obj_malloc(struct zs_pool *pool,
1364 struct zspage *zspage, unsigned long handle)
1366 int i, nr_page, offset;
1368 struct link_free *link;
1369 struct size_class *class;
1371 struct page *m_page;
1372 unsigned long m_offset;
1375 class = pool->size_class[zspage->class];
1376 handle |= OBJ_ALLOCATED_TAG;
1377 obj = get_freeobj(zspage);
1379 offset = obj * class->size;
1380 nr_page = offset >> PAGE_SHIFT;
1381 m_offset = offset & ~PAGE_MASK;
1382 m_page = get_first_page(zspage);
1384 for (i = 0; i < nr_page; i++)
1385 m_page = get_next_page(m_page);
1387 vaddr = kmap_atomic(m_page);
1388 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1389 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1390 if (likely(!PageHugeObject(m_page)))
1391 /* record handle in the header of allocated chunk */
1392 link->handle = handle;
1394 /* record handle to page->index */
1395 zspage->first_page->index = handle;
1397 kunmap_atomic(vaddr);
1398 mod_zspage_inuse(zspage, 1);
1400 obj = location_to_obj(m_page, obj);
1407 * zs_malloc - Allocate block of given size from pool.
1408 * @pool: pool to allocate from
1409 * @size: size of block to allocate
1410 * @gfp: gfp flags when allocating object
1412 * On success, handle to the allocated object is returned,
1414 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1416 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1418 unsigned long handle, obj;
1419 struct size_class *class;
1420 enum fullness_group newfg;
1421 struct zspage *zspage;
1423 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1426 handle = cache_alloc_handle(pool, gfp);
1430 /* extra space in chunk to keep the handle */
1431 size += ZS_HANDLE_SIZE;
1432 class = pool->size_class[get_size_class_index(size)];
1434 spin_lock(&class->lock);
1435 zspage = find_get_zspage(class);
1436 if (likely(zspage)) {
1437 obj = obj_malloc(pool, zspage, handle);
1438 /* Now move the zspage to another fullness group, if required */
1439 fix_fullness_group(class, zspage);
1440 record_obj(handle, obj);
1441 class_stat_inc(class, OBJ_USED, 1);
1442 spin_unlock(&class->lock);
1447 spin_unlock(&class->lock);
1449 zspage = alloc_zspage(pool, class, gfp);
1451 cache_free_handle(pool, handle);
1455 spin_lock(&class->lock);
1456 obj = obj_malloc(pool, zspage, handle);
1457 newfg = get_fullness_group(class, zspage);
1458 insert_zspage(class, zspage, newfg);
1459 set_zspage_mapping(zspage, class->index, newfg);
1460 record_obj(handle, obj);
1461 atomic_long_add(class->pages_per_zspage,
1462 &pool->pages_allocated);
1463 class_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1464 class_stat_inc(class, OBJ_USED, 1);
1466 /* We completely set up zspage so mark them as movable */
1467 SetZsPageMovable(pool, zspage);
1468 spin_unlock(&class->lock);
1472 EXPORT_SYMBOL_GPL(zs_malloc);
1474 static void obj_free(int class_size, unsigned long obj)
1476 struct link_free *link;
1477 struct zspage *zspage;
1478 struct page *f_page;
1479 unsigned long f_offset;
1480 unsigned int f_objidx;
1483 obj_to_location(obj, &f_page, &f_objidx);
1484 f_offset = (class_size * f_objidx) & ~PAGE_MASK;
1485 zspage = get_zspage(f_page);
1487 vaddr = kmap_atomic(f_page);
1489 /* Insert this object in containing zspage's freelist */
1490 link = (struct link_free *)(vaddr + f_offset);
1491 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1492 kunmap_atomic(vaddr);
1493 set_freeobj(zspage, f_objidx);
1494 mod_zspage_inuse(zspage, -1);
1497 void zs_free(struct zs_pool *pool, unsigned long handle)
1499 struct zspage *zspage;
1500 struct page *f_page;
1502 struct size_class *class;
1503 enum fullness_group fullness;
1506 if (unlikely(!handle))
1510 obj = handle_to_obj(handle);
1511 obj_to_page(obj, &f_page);
1512 zspage = get_zspage(f_page);
1514 migrate_read_lock(zspage);
1515 class = zspage_class(pool, zspage);
1517 spin_lock(&class->lock);
1518 obj_free(class->size, obj);
1519 class_stat_dec(class, OBJ_USED, 1);
1520 fullness = fix_fullness_group(class, zspage);
1521 if (fullness != ZS_EMPTY) {
1522 migrate_read_unlock(zspage);
1526 isolated = is_zspage_isolated(zspage);
1527 migrate_read_unlock(zspage);
1528 /* If zspage is isolated, zs_page_putback will free the zspage */
1529 if (likely(!isolated))
1530 free_zspage(pool, class, zspage);
1533 spin_unlock(&class->lock);
1535 cache_free_handle(pool, handle);
1537 EXPORT_SYMBOL_GPL(zs_free);
1539 static void zs_object_copy(struct size_class *class, unsigned long dst,
1542 struct page *s_page, *d_page;
1543 unsigned int s_objidx, d_objidx;
1544 unsigned long s_off, d_off;
1545 void *s_addr, *d_addr;
1546 int s_size, d_size, size;
1549 s_size = d_size = class->size;
1551 obj_to_location(src, &s_page, &s_objidx);
1552 obj_to_location(dst, &d_page, &d_objidx);
1554 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1555 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1557 if (s_off + class->size > PAGE_SIZE)
1558 s_size = PAGE_SIZE - s_off;
1560 if (d_off + class->size > PAGE_SIZE)
1561 d_size = PAGE_SIZE - d_off;
1563 s_addr = kmap_atomic(s_page);
1564 d_addr = kmap_atomic(d_page);
1567 size = min(s_size, d_size);
1568 memcpy(d_addr + d_off, s_addr + s_off, size);
1571 if (written == class->size)
1579 if (s_off >= PAGE_SIZE) {
1580 kunmap_atomic(d_addr);
1581 kunmap_atomic(s_addr);
1582 s_page = get_next_page(s_page);
1583 s_addr = kmap_atomic(s_page);
1584 d_addr = kmap_atomic(d_page);
1585 s_size = class->size - written;
1589 if (d_off >= PAGE_SIZE) {
1590 kunmap_atomic(d_addr);
1591 d_page = get_next_page(d_page);
1592 d_addr = kmap_atomic(d_page);
1593 d_size = class->size - written;
1598 kunmap_atomic(d_addr);
1599 kunmap_atomic(s_addr);
1603 * Find alloced object in zspage from index object and
1606 static unsigned long find_alloced_obj(struct size_class *class,
1607 struct page *page, int *obj_idx)
1611 int index = *obj_idx;
1612 unsigned long handle = 0;
1613 void *addr = kmap_atomic(page);
1615 offset = get_first_obj_offset(page);
1616 offset += class->size * index;
1618 while (offset < PAGE_SIZE) {
1619 head = obj_to_head(page, addr + offset);
1620 if (head & OBJ_ALLOCATED_TAG) {
1621 handle = head & ~OBJ_ALLOCATED_TAG;
1622 if (trypin_tag(handle))
1627 offset += class->size;
1631 kunmap_atomic(addr);
1638 struct zs_compact_control {
1639 /* Source spage for migration which could be a subpage of zspage */
1640 struct page *s_page;
1641 /* Destination page for migration which should be a first page
1643 struct page *d_page;
1644 /* Starting object index within @s_page which used for live object
1645 * in the subpage. */
1649 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1650 struct zs_compact_control *cc)
1652 unsigned long used_obj, free_obj;
1653 unsigned long handle;
1654 struct page *s_page = cc->s_page;
1655 struct page *d_page = cc->d_page;
1656 int obj_idx = cc->obj_idx;
1660 handle = find_alloced_obj(class, s_page, &obj_idx);
1662 s_page = get_next_page(s_page);
1669 /* Stop if there is no more space */
1670 if (zspage_full(class, get_zspage(d_page))) {
1676 used_obj = handle_to_obj(handle);
1677 free_obj = obj_malloc(pool, get_zspage(d_page), handle);
1678 zs_object_copy(class, free_obj, used_obj);
1681 * record_obj updates handle's value to free_obj and it will
1682 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1683 * breaks synchronization using pin_tag(e,g, zs_free) so
1684 * let's keep the lock bit.
1686 free_obj |= BIT(HANDLE_PIN_BIT);
1687 record_obj(handle, free_obj);
1689 obj_free(class->size, used_obj);
1692 /* Remember last position in this iteration */
1693 cc->s_page = s_page;
1694 cc->obj_idx = obj_idx;
1699 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1702 struct zspage *zspage;
1703 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1706 fg[0] = ZS_ALMOST_FULL;
1707 fg[1] = ZS_ALMOST_EMPTY;
1710 for (i = 0; i < 2; i++) {
1711 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1712 struct zspage, list);
1714 VM_BUG_ON(is_zspage_isolated(zspage));
1715 remove_zspage(class, zspage, fg[i]);
1724 * putback_zspage - add @zspage into right class's fullness list
1725 * @class: destination class
1726 * @zspage: target page
1728 * Return @zspage's fullness_group
1730 static enum fullness_group putback_zspage(struct size_class *class,
1731 struct zspage *zspage)
1733 enum fullness_group fullness;
1735 VM_BUG_ON(is_zspage_isolated(zspage));
1737 fullness = get_fullness_group(class, zspage);
1738 insert_zspage(class, zspage, fullness);
1739 set_zspage_mapping(zspage, class->index, fullness);
1744 #ifdef CONFIG_COMPACTION
1746 * To prevent zspage destroy during migration, zspage freeing should
1747 * hold locks of all pages in the zspage.
1749 static void lock_zspage(struct zspage *zspage)
1751 struct page *page = get_first_page(zspage);
1755 } while ((page = get_next_page(page)) != NULL);
1758 static int zs_init_fs_context(struct fs_context *fc)
1760 return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1763 static struct file_system_type zsmalloc_fs = {
1765 .init_fs_context = zs_init_fs_context,
1766 .kill_sb = kill_anon_super,
1769 static int zsmalloc_mount(void)
1773 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1774 if (IS_ERR(zsmalloc_mnt))
1775 ret = PTR_ERR(zsmalloc_mnt);
1780 static void zsmalloc_unmount(void)
1782 kern_unmount(zsmalloc_mnt);
1785 static void migrate_lock_init(struct zspage *zspage)
1787 rwlock_init(&zspage->lock);
1790 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1792 read_lock(&zspage->lock);
1795 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1797 read_unlock(&zspage->lock);
1800 static void migrate_write_lock(struct zspage *zspage)
1802 write_lock(&zspage->lock);
1805 static void migrate_write_unlock(struct zspage *zspage)
1807 write_unlock(&zspage->lock);
1810 /* Number of isolated subpage for *page migration* in this zspage */
1811 static void inc_zspage_isolation(struct zspage *zspage)
1816 static void dec_zspage_isolation(struct zspage *zspage)
1821 static void putback_zspage_deferred(struct zs_pool *pool,
1822 struct size_class *class,
1823 struct zspage *zspage)
1825 enum fullness_group fg;
1827 fg = putback_zspage(class, zspage);
1829 schedule_work(&pool->free_work);
1833 static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1835 VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1836 atomic_long_dec(&pool->isolated_pages);
1838 * Checking pool->destroying must happen after atomic_long_dec()
1839 * for pool->isolated_pages above. Paired with the smp_mb() in
1840 * zs_unregister_migration().
1842 smp_mb__after_atomic();
1843 if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1844 wake_up_all(&pool->migration_wait);
1847 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1848 struct page *newpage, struct page *oldpage)
1851 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1854 page = get_first_page(zspage);
1856 if (page == oldpage)
1857 pages[idx] = newpage;
1861 } while ((page = get_next_page(page)) != NULL);
1863 create_page_chain(class, zspage, pages);
1864 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1865 if (unlikely(PageHugeObject(oldpage)))
1866 newpage->index = oldpage->index;
1867 __SetPageMovable(newpage, page_mapping(oldpage));
1870 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1872 struct zs_pool *pool;
1873 struct size_class *class;
1874 struct zspage *zspage;
1875 struct address_space *mapping;
1878 * Page is locked so zspage couldn't be destroyed. For detail, look at
1879 * lock_zspage in free_zspage.
1881 VM_BUG_ON_PAGE(!PageMovable(page), page);
1882 VM_BUG_ON_PAGE(PageIsolated(page), page);
1884 zspage = get_zspage(page);
1886 mapping = page_mapping(page);
1887 pool = mapping->private_data;
1889 class = zspage_class(pool, zspage);
1891 spin_lock(&class->lock);
1892 if (get_zspage_inuse(zspage) == 0) {
1893 spin_unlock(&class->lock);
1897 /* zspage is isolated for object migration */
1898 if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1899 spin_unlock(&class->lock);
1904 * If this is first time isolation for the zspage, isolate zspage from
1905 * size_class to prevent further object allocation from the zspage.
1907 if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1908 enum fullness_group fullness;
1909 unsigned int class_idx;
1911 get_zspage_mapping(zspage, &class_idx, &fullness);
1912 atomic_long_inc(&pool->isolated_pages);
1913 remove_zspage(class, zspage, fullness);
1916 inc_zspage_isolation(zspage);
1917 spin_unlock(&class->lock);
1922 static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1923 struct page *page, enum migrate_mode mode)
1925 struct zs_pool *pool;
1926 struct size_class *class;
1927 struct zspage *zspage;
1929 void *s_addr, *d_addr, *addr;
1931 unsigned long handle, head;
1932 unsigned long old_obj, new_obj;
1933 unsigned int obj_idx;
1937 * We cannot support the _NO_COPY case here, because copy needs to
1938 * happen under the zs lock, which does not work with
1939 * MIGRATE_SYNC_NO_COPY workflow.
1941 if (mode == MIGRATE_SYNC_NO_COPY)
1944 VM_BUG_ON_PAGE(!PageMovable(page), page);
1945 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1947 zspage = get_zspage(page);
1949 /* Concurrent compactor cannot migrate any subpage in zspage */
1950 migrate_write_lock(zspage);
1951 pool = mapping->private_data;
1952 class = zspage_class(pool, zspage);
1953 offset = get_first_obj_offset(page);
1955 spin_lock(&class->lock);
1956 if (!get_zspage_inuse(zspage)) {
1958 * Set "offset" to end of the page so that every loops
1959 * skips unnecessary object scanning.
1965 s_addr = kmap_atomic(page);
1966 while (pos < PAGE_SIZE) {
1967 head = obj_to_head(page, s_addr + pos);
1968 if (head & OBJ_ALLOCATED_TAG) {
1969 handle = head & ~OBJ_ALLOCATED_TAG;
1970 if (!trypin_tag(handle))
1977 * Here, any user cannot access all objects in the zspage so let's move.
1979 d_addr = kmap_atomic(newpage);
1980 memcpy(d_addr, s_addr, PAGE_SIZE);
1981 kunmap_atomic(d_addr);
1983 for (addr = s_addr + offset; addr < s_addr + pos;
1984 addr += class->size) {
1985 head = obj_to_head(page, addr);
1986 if (head & OBJ_ALLOCATED_TAG) {
1987 handle = head & ~OBJ_ALLOCATED_TAG;
1988 BUG_ON(!testpin_tag(handle));
1990 old_obj = handle_to_obj(handle);
1991 obj_to_location(old_obj, &dummy, &obj_idx);
1992 new_obj = (unsigned long)location_to_obj(newpage,
1994 new_obj |= BIT(HANDLE_PIN_BIT);
1995 record_obj(handle, new_obj);
1999 replace_sub_page(class, zspage, newpage, page);
2002 dec_zspage_isolation(zspage);
2005 * Page migration is done so let's putback isolated zspage to
2006 * the list if @page is final isolated subpage in the zspage.
2008 if (!is_zspage_isolated(zspage)) {
2010 * We cannot race with zs_destroy_pool() here because we wait
2011 * for isolation to hit zero before we start destroying.
2012 * Also, we ensure that everyone can see pool->destroying before
2015 putback_zspage_deferred(pool, class, zspage);
2016 zs_pool_dec_isolated(pool);
2019 if (page_zone(newpage) != page_zone(page)) {
2020 dec_zone_page_state(page, NR_ZSPAGES);
2021 inc_zone_page_state(newpage, NR_ZSPAGES);
2028 ret = MIGRATEPAGE_SUCCESS;
2030 for (addr = s_addr + offset; addr < s_addr + pos;
2031 addr += class->size) {
2032 head = obj_to_head(page, addr);
2033 if (head & OBJ_ALLOCATED_TAG) {
2034 handle = head & ~OBJ_ALLOCATED_TAG;
2035 BUG_ON(!testpin_tag(handle));
2039 kunmap_atomic(s_addr);
2040 spin_unlock(&class->lock);
2041 migrate_write_unlock(zspage);
2046 static void zs_page_putback(struct page *page)
2048 struct zs_pool *pool;
2049 struct size_class *class;
2050 struct address_space *mapping;
2051 struct zspage *zspage;
2053 VM_BUG_ON_PAGE(!PageMovable(page), page);
2054 VM_BUG_ON_PAGE(!PageIsolated(page), page);
2056 zspage = get_zspage(page);
2057 mapping = page_mapping(page);
2058 pool = mapping->private_data;
2059 class = zspage_class(pool, zspage);
2061 spin_lock(&class->lock);
2062 dec_zspage_isolation(zspage);
2063 if (!is_zspage_isolated(zspage)) {
2065 * Due to page_lock, we cannot free zspage immediately
2068 putback_zspage_deferred(pool, class, zspage);
2069 zs_pool_dec_isolated(pool);
2071 spin_unlock(&class->lock);
2074 static const struct address_space_operations zsmalloc_aops = {
2075 .isolate_page = zs_page_isolate,
2076 .migratepage = zs_page_migrate,
2077 .putback_page = zs_page_putback,
2080 static int zs_register_migration(struct zs_pool *pool)
2082 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2083 if (IS_ERR(pool->inode)) {
2088 pool->inode->i_mapping->private_data = pool;
2089 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2093 static bool pool_isolated_are_drained(struct zs_pool *pool)
2095 return atomic_long_read(&pool->isolated_pages) == 0;
2098 /* Function for resolving migration */
2099 static void wait_for_isolated_drain(struct zs_pool *pool)
2103 * We're in the process of destroying the pool, so there are no
2104 * active allocations. zs_page_isolate() fails for completely free
2105 * zspages, so we need only wait for the zs_pool's isolated
2106 * count to hit zero.
2108 wait_event(pool->migration_wait,
2109 pool_isolated_are_drained(pool));
2112 static void zs_unregister_migration(struct zs_pool *pool)
2114 pool->destroying = true;
2116 * We need a memory barrier here to ensure global visibility of
2117 * pool->destroying. Thus pool->isolated pages will either be 0 in which
2118 * case we don't care, or it will be > 0 and pool->destroying will
2119 * ensure that we wake up once isolation hits 0.
2122 wait_for_isolated_drain(pool); /* This can block */
2123 flush_work(&pool->free_work);
2128 * Caller should hold page_lock of all pages in the zspage
2129 * In here, we cannot use zspage meta data.
2131 static void async_free_zspage(struct work_struct *work)
2134 struct size_class *class;
2135 unsigned int class_idx;
2136 enum fullness_group fullness;
2137 struct zspage *zspage, *tmp;
2138 LIST_HEAD(free_pages);
2139 struct zs_pool *pool = container_of(work, struct zs_pool,
2142 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2143 class = pool->size_class[i];
2144 if (class->index != i)
2147 spin_lock(&class->lock);
2148 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2149 spin_unlock(&class->lock);
2153 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2154 list_del(&zspage->list);
2155 lock_zspage(zspage);
2157 get_zspage_mapping(zspage, &class_idx, &fullness);
2158 VM_BUG_ON(fullness != ZS_EMPTY);
2159 class = pool->size_class[class_idx];
2160 spin_lock(&class->lock);
2161 __free_zspage(pool, class, zspage);
2162 spin_unlock(&class->lock);
2166 static void kick_deferred_free(struct zs_pool *pool)
2168 schedule_work(&pool->free_work);
2171 static void init_deferred_free(struct zs_pool *pool)
2173 INIT_WORK(&pool->free_work, async_free_zspage);
2176 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2178 struct page *page = get_first_page(zspage);
2181 WARN_ON(!trylock_page(page));
2182 __SetPageMovable(page, pool->inode->i_mapping);
2184 } while ((page = get_next_page(page)) != NULL);
2190 * Based on the number of unused allocated objects calculate
2191 * and return the number of pages that we can free.
2193 static unsigned long zs_can_compact(struct size_class *class)
2195 unsigned long obj_wasted;
2196 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2197 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2199 if (obj_allocated <= obj_used)
2202 obj_wasted = obj_allocated - obj_used;
2203 obj_wasted /= class->objs_per_zspage;
2205 return obj_wasted * class->pages_per_zspage;
2208 static unsigned long __zs_compact(struct zs_pool *pool,
2209 struct size_class *class)
2211 struct zs_compact_control cc;
2212 struct zspage *src_zspage;
2213 struct zspage *dst_zspage = NULL;
2214 unsigned long pages_freed = 0;
2216 spin_lock(&class->lock);
2217 while ((src_zspage = isolate_zspage(class, true))) {
2219 if (!zs_can_compact(class))
2223 cc.s_page = get_first_page(src_zspage);
2225 while ((dst_zspage = isolate_zspage(class, false))) {
2226 cc.d_page = get_first_page(dst_zspage);
2228 * If there is no more space in dst_page, resched
2229 * and see if anyone had allocated another zspage.
2231 if (!migrate_zspage(pool, class, &cc))
2234 putback_zspage(class, dst_zspage);
2237 /* Stop if we couldn't find slot */
2238 if (dst_zspage == NULL)
2241 putback_zspage(class, dst_zspage);
2242 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2243 free_zspage(pool, class, src_zspage);
2244 pages_freed += class->pages_per_zspage;
2246 spin_unlock(&class->lock);
2248 spin_lock(&class->lock);
2252 putback_zspage(class, src_zspage);
2254 spin_unlock(&class->lock);
2259 unsigned long zs_compact(struct zs_pool *pool)
2262 struct size_class *class;
2263 unsigned long pages_freed = 0;
2265 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2266 class = pool->size_class[i];
2269 if (class->index != i)
2271 pages_freed += __zs_compact(pool, class);
2273 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2277 EXPORT_SYMBOL_GPL(zs_compact);
2279 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2281 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2283 EXPORT_SYMBOL_GPL(zs_pool_stats);
2285 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2286 struct shrink_control *sc)
2288 unsigned long pages_freed;
2289 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2293 * Compact classes and calculate compaction delta.
2294 * Can run concurrently with a manually triggered
2295 * (by user) compaction.
2297 pages_freed = zs_compact(pool);
2299 return pages_freed ? pages_freed : SHRINK_STOP;
2302 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2303 struct shrink_control *sc)
2306 struct size_class *class;
2307 unsigned long pages_to_free = 0;
2308 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2311 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2312 class = pool->size_class[i];
2315 if (class->index != i)
2318 pages_to_free += zs_can_compact(class);
2321 return pages_to_free;
2324 static void zs_unregister_shrinker(struct zs_pool *pool)
2326 unregister_shrinker(&pool->shrinker);
2329 static int zs_register_shrinker(struct zs_pool *pool)
2331 pool->shrinker.scan_objects = zs_shrinker_scan;
2332 pool->shrinker.count_objects = zs_shrinker_count;
2333 pool->shrinker.batch = 0;
2334 pool->shrinker.seeks = DEFAULT_SEEKS;
2336 return register_shrinker(&pool->shrinker);
2340 * zs_create_pool - Creates an allocation pool to work from.
2341 * @name: pool name to be created
2343 * This function must be called before anything when using
2344 * the zsmalloc allocator.
2346 * On success, a pointer to the newly created pool is returned,
2349 struct zs_pool *zs_create_pool(const char *name)
2352 struct zs_pool *pool;
2353 struct size_class *prev_class = NULL;
2355 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2359 init_deferred_free(pool);
2361 pool->name = kstrdup(name, GFP_KERNEL);
2365 #ifdef CONFIG_COMPACTION
2366 init_waitqueue_head(&pool->migration_wait);
2369 if (create_cache(pool))
2373 * Iterate reversely, because, size of size_class that we want to use
2374 * for merging should be larger or equal to current size.
2376 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2378 int pages_per_zspage;
2379 int objs_per_zspage;
2380 struct size_class *class;
2383 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2384 if (size > ZS_MAX_ALLOC_SIZE)
2385 size = ZS_MAX_ALLOC_SIZE;
2386 pages_per_zspage = get_pages_per_zspage(size);
2387 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2390 * We iterate from biggest down to smallest classes,
2391 * so huge_class_size holds the size of the first huge
2392 * class. Any object bigger than or equal to that will
2393 * endup in the huge class.
2395 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2397 huge_class_size = size;
2399 * The object uses ZS_HANDLE_SIZE bytes to store the
2400 * handle. We need to subtract it, because zs_malloc()
2401 * unconditionally adds handle size before it performs
2402 * size class search - so object may be smaller than
2403 * huge class size, yet it still can end up in the huge
2404 * class because it grows by ZS_HANDLE_SIZE extra bytes
2405 * right before class lookup.
2407 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2411 * size_class is used for normal zsmalloc operation such
2412 * as alloc/free for that size. Although it is natural that we
2413 * have one size_class for each size, there is a chance that we
2414 * can get more memory utilization if we use one size_class for
2415 * many different sizes whose size_class have same
2416 * characteristics. So, we makes size_class point to
2417 * previous size_class if possible.
2420 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2421 pool->size_class[i] = prev_class;
2426 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2432 class->pages_per_zspage = pages_per_zspage;
2433 class->objs_per_zspage = objs_per_zspage;
2434 spin_lock_init(&class->lock);
2435 pool->size_class[i] = class;
2436 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2438 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2443 /* debug only, don't abort if it fails */
2444 zs_pool_stat_create(pool, name);
2446 if (zs_register_migration(pool))
2450 * Not critical since shrinker is only used to trigger internal
2451 * defragmentation of the pool which is pretty optional thing. If
2452 * registration fails we still can use the pool normally and user can
2453 * trigger compaction manually. Thus, ignore return code.
2455 zs_register_shrinker(pool);
2460 zs_destroy_pool(pool);
2463 EXPORT_SYMBOL_GPL(zs_create_pool);
2465 void zs_destroy_pool(struct zs_pool *pool)
2469 zs_unregister_shrinker(pool);
2470 zs_unregister_migration(pool);
2471 zs_pool_stat_destroy(pool);
2473 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2475 struct size_class *class = pool->size_class[i];
2480 if (class->index != i)
2483 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2484 if (!list_empty(&class->fullness_list[fg])) {
2485 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2492 destroy_cache(pool);
2496 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2498 static int __init zs_init(void)
2502 ret = zsmalloc_mount();
2506 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2507 zs_cpu_prepare, zs_cpu_dead);
2512 zpool_register_driver(&zs_zpool_driver);
2525 static void __exit zs_exit(void)
2528 zpool_unregister_driver(&zs_zpool_driver);
2531 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2536 module_init(zs_init);
2537 module_exit(zs_exit);
2539 MODULE_LICENSE("Dual BSD/GPL");
2540 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");