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)
125 #define FULLNESS_BITS 2
127 #define ISOLATED_BITS 3
128 #define MAGIC_VAL_BITS 8
130 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
131 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
132 #define ZS_MIN_ALLOC_SIZE \
133 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
134 /* each chunk includes extra space to keep handle */
135 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
138 * On systems with 4K page size, this gives 255 size classes! There is a
140 * - Large number of size classes is potentially wasteful as free page are
141 * spread across these classes
142 * - Small number of size classes causes large internal fragmentation
143 * - Probably its better to use specific size classes (empirically
144 * determined). NOTE: all those class sizes must be set as multiple of
145 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
147 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
150 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> CLASS_BITS)
151 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
152 ZS_SIZE_CLASS_DELTA) + 1)
154 enum fullness_group {
162 enum class_stat_type {
172 struct zs_size_stat {
173 unsigned long objs[NR_ZS_STAT_TYPE];
176 #ifdef CONFIG_ZSMALLOC_STAT
177 static struct dentry *zs_stat_root;
180 #ifdef CONFIG_COMPACTION
181 static struct vfsmount *zsmalloc_mnt;
185 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
187 * n = number of allocated objects
188 * N = total number of objects zspage can store
189 * f = fullness_threshold_frac
191 * Similarly, we assign zspage to:
192 * ZS_ALMOST_FULL when n > N / f
193 * ZS_EMPTY when n == 0
194 * ZS_FULL when n == N
196 * (see: fix_fullness_group())
198 static const int fullness_threshold_frac = 4;
199 static size_t huge_class_size;
203 struct list_head fullness_list[NR_ZS_FULLNESS];
205 * Size of objects stored in this class. Must be multiple
210 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
211 int pages_per_zspage;
214 struct zs_size_stat stats;
218 * Placed within free objects to form a singly linked list.
219 * For every zspage, zspage->freeobj gives head of this list.
221 * This must be power of 2 and less than or equal to ZS_ALIGN
227 * It's valid for non-allocated object
231 * Handle of allocated object.
233 unsigned long handle;
240 struct size_class *size_class[ZS_SIZE_CLASSES];
241 struct kmem_cache *handle_cachep;
242 struct kmem_cache *zspage_cachep;
244 atomic_long_t pages_allocated;
246 struct zs_pool_stats stats;
248 /* Compact classes */
249 struct shrinker shrinker;
251 #ifdef CONFIG_ZSMALLOC_STAT
252 struct dentry *stat_dentry;
254 #ifdef CONFIG_COMPACTION
256 struct work_struct free_work;
262 unsigned int huge:HUGE_BITS;
263 unsigned int fullness:FULLNESS_BITS;
264 unsigned int class:CLASS_BITS + 1;
265 unsigned int isolated:ISOLATED_BITS;
266 unsigned int magic:MAGIC_VAL_BITS;
269 unsigned int freeobj;
270 struct page *first_page;
271 struct list_head list; /* fullness list */
272 #ifdef CONFIG_COMPACTION
277 struct mapping_area {
278 char *vm_buf; /* copy buffer for objects that span pages */
279 char *vm_addr; /* address of kmap_atomic()'ed pages */
280 enum zs_mapmode vm_mm; /* mapping mode */
283 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
284 static void SetZsHugePage(struct zspage *zspage)
289 static bool ZsHugePage(struct zspage *zspage)
294 #ifdef CONFIG_COMPACTION
295 static int zs_register_migration(struct zs_pool *pool);
296 static void zs_unregister_migration(struct zs_pool *pool);
297 static void migrate_lock_init(struct zspage *zspage);
298 static void migrate_read_lock(struct zspage *zspage);
299 static void migrate_read_unlock(struct zspage *zspage);
300 static void kick_deferred_free(struct zs_pool *pool);
301 static void init_deferred_free(struct zs_pool *pool);
302 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
304 static int zsmalloc_mount(void) { return 0; }
305 static void zsmalloc_unmount(void) {}
306 static int zs_register_migration(struct zs_pool *pool) { return 0; }
307 static void zs_unregister_migration(struct zs_pool *pool) {}
308 static void migrate_lock_init(struct zspage *zspage) {}
309 static void migrate_read_lock(struct zspage *zspage) {}
310 static void migrate_read_unlock(struct zspage *zspage) {}
311 static void kick_deferred_free(struct zs_pool *pool) {}
312 static void init_deferred_free(struct zs_pool *pool) {}
313 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
316 static int create_cache(struct zs_pool *pool)
318 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
320 if (!pool->handle_cachep)
323 pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
325 if (!pool->zspage_cachep) {
326 kmem_cache_destroy(pool->handle_cachep);
327 pool->handle_cachep = NULL;
334 static void destroy_cache(struct zs_pool *pool)
336 kmem_cache_destroy(pool->handle_cachep);
337 kmem_cache_destroy(pool->zspage_cachep);
340 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
342 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
343 gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
346 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
348 kmem_cache_free(pool->handle_cachep, (void *)handle);
351 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
353 return kmem_cache_zalloc(pool->zspage_cachep,
354 flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
357 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
359 kmem_cache_free(pool->zspage_cachep, zspage);
362 static void record_obj(unsigned long handle, unsigned long obj)
365 * lsb of @obj represents handle lock while other bits
366 * represent object value the handle is pointing so
367 * updating shouldn't do store tearing.
369 WRITE_ONCE(*(unsigned long *)handle, obj);
376 static void *zs_zpool_create(const char *name, gfp_t gfp,
377 const struct zpool_ops *zpool_ops,
381 * Ignore global gfp flags: zs_malloc() may be invoked from
382 * different contexts and its caller must provide a valid
385 return zs_create_pool(name);
388 static void zs_zpool_destroy(void *pool)
390 zs_destroy_pool(pool);
393 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
394 unsigned long *handle)
396 *handle = zs_malloc(pool, size, gfp);
397 return *handle ? 0 : -1;
399 static void zs_zpool_free(void *pool, unsigned long handle)
401 zs_free(pool, handle);
404 static void *zs_zpool_map(void *pool, unsigned long handle,
405 enum zpool_mapmode mm)
407 enum zs_mapmode zs_mm;
422 return zs_map_object(pool, handle, zs_mm);
424 static void zs_zpool_unmap(void *pool, unsigned long handle)
426 zs_unmap_object(pool, handle);
429 static u64 zs_zpool_total_size(void *pool)
431 return zs_get_total_pages(pool) << PAGE_SHIFT;
434 static struct zpool_driver zs_zpool_driver = {
436 .owner = THIS_MODULE,
437 .create = zs_zpool_create,
438 .destroy = zs_zpool_destroy,
439 .malloc_support_movable = true,
440 .malloc = zs_zpool_malloc,
441 .free = zs_zpool_free,
443 .unmap = zs_zpool_unmap,
444 .total_size = zs_zpool_total_size,
447 MODULE_ALIAS("zpool-zsmalloc");
448 #endif /* CONFIG_ZPOOL */
450 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
451 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
453 static __maybe_unused int is_first_page(struct page *page)
455 return PagePrivate(page);
458 /* Protected by class->lock */
459 static inline int get_zspage_inuse(struct zspage *zspage)
461 return zspage->inuse;
465 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
467 zspage->inuse += val;
470 static inline struct page *get_first_page(struct zspage *zspage)
472 struct page *first_page = zspage->first_page;
474 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
478 static inline int get_first_obj_offset(struct page *page)
480 return page->page_type;
483 static inline void set_first_obj_offset(struct page *page, int offset)
485 page->page_type = offset;
488 static inline unsigned int get_freeobj(struct zspage *zspage)
490 return zspage->freeobj;
493 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
495 zspage->freeobj = obj;
498 static void get_zspage_mapping(struct zspage *zspage,
499 unsigned int *class_idx,
500 enum fullness_group *fullness)
502 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
504 *fullness = zspage->fullness;
505 *class_idx = zspage->class;
508 static struct size_class *zspage_class(struct zs_pool *pool,
509 struct zspage *zspage)
511 return pool->size_class[zspage->class];
514 static void set_zspage_mapping(struct zspage *zspage,
515 unsigned int class_idx,
516 enum fullness_group fullness)
518 zspage->class = class_idx;
519 zspage->fullness = fullness;
523 * zsmalloc divides the pool into various size classes where each
524 * class maintains a list of zspages where each zspage is divided
525 * into equal sized chunks. Each allocation falls into one of these
526 * classes depending on its size. This function returns index of the
527 * size class which has chunk size big enough to hold the given size.
529 static int get_size_class_index(int size)
533 if (likely(size > ZS_MIN_ALLOC_SIZE))
534 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
535 ZS_SIZE_CLASS_DELTA);
537 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
540 /* type can be of enum type class_stat_type or fullness_group */
541 static inline void class_stat_inc(struct size_class *class,
542 int type, unsigned long cnt)
544 class->stats.objs[type] += cnt;
547 /* type can be of enum type class_stat_type or fullness_group */
548 static inline void class_stat_dec(struct size_class *class,
549 int type, unsigned long cnt)
551 class->stats.objs[type] -= cnt;
554 /* type can be of enum type class_stat_type or fullness_group */
555 static inline unsigned long zs_stat_get(struct size_class *class,
558 return class->stats.objs[type];
561 #ifdef CONFIG_ZSMALLOC_STAT
563 static void __init zs_stat_init(void)
565 if (!debugfs_initialized()) {
566 pr_warn("debugfs not available, stat dir not created\n");
570 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
573 static void __exit zs_stat_exit(void)
575 debugfs_remove_recursive(zs_stat_root);
578 static unsigned long zs_can_compact(struct size_class *class);
580 static int zs_stats_size_show(struct seq_file *s, void *v)
583 struct zs_pool *pool = s->private;
584 struct size_class *class;
586 unsigned long class_almost_full, class_almost_empty;
587 unsigned long obj_allocated, obj_used, pages_used, freeable;
588 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
589 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
590 unsigned long total_freeable = 0;
592 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
593 "class", "size", "almost_full", "almost_empty",
594 "obj_allocated", "obj_used", "pages_used",
595 "pages_per_zspage", "freeable");
597 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
598 class = pool->size_class[i];
600 if (class->index != i)
603 spin_lock(&class->lock);
604 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
605 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
606 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
607 obj_used = zs_stat_get(class, OBJ_USED);
608 freeable = zs_can_compact(class);
609 spin_unlock(&class->lock);
611 objs_per_zspage = class->objs_per_zspage;
612 pages_used = obj_allocated / objs_per_zspage *
613 class->pages_per_zspage;
615 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
616 " %10lu %10lu %16d %8lu\n",
617 i, class->size, class_almost_full, class_almost_empty,
618 obj_allocated, obj_used, pages_used,
619 class->pages_per_zspage, freeable);
621 total_class_almost_full += class_almost_full;
622 total_class_almost_empty += class_almost_empty;
623 total_objs += obj_allocated;
624 total_used_objs += obj_used;
625 total_pages += pages_used;
626 total_freeable += freeable;
630 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
631 "Total", "", total_class_almost_full,
632 total_class_almost_empty, total_objs,
633 total_used_objs, total_pages, "", total_freeable);
637 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
639 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
642 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
646 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
648 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
649 &zs_stats_size_fops);
652 static void zs_pool_stat_destroy(struct zs_pool *pool)
654 debugfs_remove_recursive(pool->stat_dentry);
657 #else /* CONFIG_ZSMALLOC_STAT */
658 static void __init zs_stat_init(void)
662 static void __exit zs_stat_exit(void)
666 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
670 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
677 * For each size class, zspages are divided into different groups
678 * depending on how "full" they are. This was done so that we could
679 * easily find empty or nearly empty zspages when we try to shrink
680 * the pool (not yet implemented). This function returns fullness
681 * status of the given page.
683 static enum fullness_group get_fullness_group(struct size_class *class,
684 struct zspage *zspage)
686 int inuse, objs_per_zspage;
687 enum fullness_group fg;
689 inuse = get_zspage_inuse(zspage);
690 objs_per_zspage = class->objs_per_zspage;
694 else if (inuse == objs_per_zspage)
696 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
697 fg = ZS_ALMOST_EMPTY;
705 * Each size class maintains various freelists and zspages are assigned
706 * to one of these freelists based on the number of live objects they
707 * have. This functions inserts the given zspage into the freelist
708 * identified by <class, fullness_group>.
710 static void insert_zspage(struct size_class *class,
711 struct zspage *zspage,
712 enum fullness_group fullness)
716 class_stat_inc(class, fullness, 1);
717 head = list_first_entry_or_null(&class->fullness_list[fullness],
718 struct zspage, list);
720 * We want to see more ZS_FULL pages and less almost empty/full.
721 * Put pages with higher ->inuse first.
723 if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head))
724 list_add(&zspage->list, &head->list);
726 list_add(&zspage->list, &class->fullness_list[fullness]);
730 * This function removes the given zspage from the freelist identified
731 * by <class, fullness_group>.
733 static void remove_zspage(struct size_class *class,
734 struct zspage *zspage,
735 enum fullness_group fullness)
737 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
739 list_del_init(&zspage->list);
740 class_stat_dec(class, fullness, 1);
744 * Each size class maintains zspages in different fullness groups depending
745 * on the number of live objects they contain. When allocating or freeing
746 * objects, the fullness status of the page can change, say, from ALMOST_FULL
747 * to ALMOST_EMPTY when freeing an object. This function checks if such
748 * a status change has occurred for the given page and accordingly moves the
749 * page from the freelist of the old fullness group to that of the new
752 static enum fullness_group fix_fullness_group(struct size_class *class,
753 struct zspage *zspage)
756 enum fullness_group currfg, newfg;
758 get_zspage_mapping(zspage, &class_idx, &currfg);
759 newfg = get_fullness_group(class, zspage);
763 remove_zspage(class, zspage, currfg);
764 insert_zspage(class, zspage, newfg);
765 set_zspage_mapping(zspage, class_idx, newfg);
771 * We have to decide on how many pages to link together
772 * to form a zspage for each size class. This is important
773 * to reduce wastage due to unusable space left at end of
774 * each zspage which is given as:
775 * wastage = Zp % class_size
776 * usage = Zp - wastage
777 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
779 * For example, for size class of 3/8 * PAGE_SIZE, we should
780 * link together 3 PAGE_SIZE sized pages to form a zspage
781 * since then we can perfectly fit in 8 such objects.
783 static int get_pages_per_zspage(int class_size)
785 int i, max_usedpc = 0;
786 /* zspage order which gives maximum used size per KB */
787 int max_usedpc_order = 1;
789 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
793 zspage_size = i * PAGE_SIZE;
794 waste = zspage_size % class_size;
795 usedpc = (zspage_size - waste) * 100 / zspage_size;
797 if (usedpc > max_usedpc) {
799 max_usedpc_order = i;
803 return max_usedpc_order;
806 static struct zspage *get_zspage(struct page *page)
808 struct zspage *zspage = (struct zspage *)page_private(page);
810 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
814 static struct page *get_next_page(struct page *page)
816 struct zspage *zspage = get_zspage(page);
818 if (unlikely(ZsHugePage(zspage)))
821 return (struct page *)page->index;
825 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
826 * @obj: the encoded object value
827 * @page: page object resides in zspage
828 * @obj_idx: object index
830 static void obj_to_location(unsigned long obj, struct page **page,
831 unsigned int *obj_idx)
833 obj >>= OBJ_TAG_BITS;
834 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
835 *obj_idx = (obj & OBJ_INDEX_MASK);
838 static void obj_to_page(unsigned long obj, struct page **page)
840 obj >>= OBJ_TAG_BITS;
841 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
845 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
846 * @page: page object resides in zspage
847 * @obj_idx: object index
849 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
853 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
854 obj |= obj_idx & OBJ_INDEX_MASK;
855 obj <<= OBJ_TAG_BITS;
860 static unsigned long handle_to_obj(unsigned long handle)
862 return *(unsigned long *)handle;
865 static bool obj_allocated(struct page *page, void *obj, unsigned long *phandle)
867 unsigned long handle;
868 struct zspage *zspage = get_zspage(page);
870 if (unlikely(ZsHugePage(zspage))) {
871 VM_BUG_ON_PAGE(!is_first_page(page), page);
872 handle = page->index;
874 handle = *(unsigned long *)obj;
876 if (!(handle & OBJ_ALLOCATED_TAG))
879 *phandle = handle & ~OBJ_ALLOCATED_TAG;
883 static inline int testpin_tag(unsigned long handle)
885 return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
888 static inline int trypin_tag(unsigned long handle)
890 return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
893 static void pin_tag(unsigned long handle) __acquires(bitlock)
895 bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
898 static void unpin_tag(unsigned long handle) __releases(bitlock)
900 bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
903 static void reset_page(struct page *page)
905 __ClearPageMovable(page);
906 ClearPagePrivate(page);
907 set_page_private(page, 0);
908 page_mapcount_reset(page);
912 static int trylock_zspage(struct zspage *zspage)
914 struct page *cursor, *fail;
916 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
917 get_next_page(cursor)) {
918 if (!trylock_page(cursor)) {
926 for (cursor = get_first_page(zspage); cursor != fail; cursor =
927 get_next_page(cursor))
933 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
934 struct zspage *zspage)
936 struct page *page, *next;
937 enum fullness_group fg;
938 unsigned int class_idx;
940 get_zspage_mapping(zspage, &class_idx, &fg);
942 assert_spin_locked(&class->lock);
944 VM_BUG_ON(get_zspage_inuse(zspage));
945 VM_BUG_ON(fg != ZS_EMPTY);
947 next = page = get_first_page(zspage);
949 VM_BUG_ON_PAGE(!PageLocked(page), page);
950 next = get_next_page(page);
953 dec_zone_page_state(page, NR_ZSPAGES);
956 } while (page != NULL);
958 cache_free_zspage(pool, zspage);
960 class_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
961 atomic_long_sub(class->pages_per_zspage,
962 &pool->pages_allocated);
965 static void free_zspage(struct zs_pool *pool, struct size_class *class,
966 struct zspage *zspage)
968 VM_BUG_ON(get_zspage_inuse(zspage));
969 VM_BUG_ON(list_empty(&zspage->list));
971 if (!trylock_zspage(zspage)) {
972 kick_deferred_free(pool);
976 remove_zspage(class, zspage, ZS_EMPTY);
977 __free_zspage(pool, class, zspage);
980 /* Initialize a newly allocated zspage */
981 static void init_zspage(struct size_class *class, struct zspage *zspage)
983 unsigned int freeobj = 1;
984 unsigned long off = 0;
985 struct page *page = get_first_page(zspage);
988 struct page *next_page;
989 struct link_free *link;
992 set_first_obj_offset(page, off);
994 vaddr = kmap_atomic(page);
995 link = (struct link_free *)vaddr + off / sizeof(*link);
997 while ((off += class->size) < PAGE_SIZE) {
998 link->next = freeobj++ << OBJ_TAG_BITS;
999 link += class->size / sizeof(*link);
1003 * We now come to the last (full or partial) object on this
1004 * page, which must point to the first object on the next
1007 next_page = get_next_page(page);
1009 link->next = freeobj++ << OBJ_TAG_BITS;
1012 * Reset OBJ_TAG_BITS bit to last link to tell
1013 * whether it's allocated object or not.
1015 link->next = -1UL << OBJ_TAG_BITS;
1017 kunmap_atomic(vaddr);
1022 set_freeobj(zspage, 0);
1025 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1026 struct page *pages[])
1030 struct page *prev_page = NULL;
1031 int nr_pages = class->pages_per_zspage;
1034 * Allocate individual pages and link them together as:
1035 * 1. all pages are linked together using page->index
1036 * 2. each sub-page point to zspage using page->private
1038 * we set PG_private to identify the first page (i.e. no other sub-page
1039 * has this flag set).
1041 for (i = 0; i < nr_pages; i++) {
1043 set_page_private(page, (unsigned long)zspage);
1046 zspage->first_page = page;
1047 SetPagePrivate(page);
1048 if (unlikely(class->objs_per_zspage == 1 &&
1049 class->pages_per_zspage == 1))
1050 SetZsHugePage(zspage);
1052 prev_page->index = (unsigned long)page;
1059 * Allocate a zspage for the given size class
1061 static struct zspage *alloc_zspage(struct zs_pool *pool,
1062 struct size_class *class,
1066 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1067 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1072 zspage->magic = ZSPAGE_MAGIC;
1073 migrate_lock_init(zspage);
1075 for (i = 0; i < class->pages_per_zspage; i++) {
1078 page = alloc_page(gfp);
1081 dec_zone_page_state(pages[i], NR_ZSPAGES);
1082 __free_page(pages[i]);
1084 cache_free_zspage(pool, zspage);
1088 inc_zone_page_state(page, NR_ZSPAGES);
1092 create_page_chain(class, zspage, pages);
1093 init_zspage(class, zspage);
1098 static struct zspage *find_get_zspage(struct size_class *class)
1101 struct zspage *zspage;
1103 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1104 zspage = list_first_entry_or_null(&class->fullness_list[i],
1105 struct zspage, list);
1113 static inline int __zs_cpu_up(struct mapping_area *area)
1116 * Make sure we don't leak memory if a cpu UP notification
1117 * and zs_init() race and both call zs_cpu_up() on the same cpu
1121 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1127 static inline void __zs_cpu_down(struct mapping_area *area)
1129 kfree(area->vm_buf);
1130 area->vm_buf = NULL;
1133 static void *__zs_map_object(struct mapping_area *area,
1134 struct page *pages[2], int off, int size)
1138 char *buf = area->vm_buf;
1140 /* disable page faults to match kmap_atomic() return conditions */
1141 pagefault_disable();
1143 /* no read fastpath */
1144 if (area->vm_mm == ZS_MM_WO)
1147 sizes[0] = PAGE_SIZE - off;
1148 sizes[1] = size - sizes[0];
1150 /* copy object to per-cpu buffer */
1151 addr = kmap_atomic(pages[0]);
1152 memcpy(buf, addr + off, sizes[0]);
1153 kunmap_atomic(addr);
1154 addr = kmap_atomic(pages[1]);
1155 memcpy(buf + sizes[0], addr, sizes[1]);
1156 kunmap_atomic(addr);
1158 return area->vm_buf;
1161 static void __zs_unmap_object(struct mapping_area *area,
1162 struct page *pages[2], int off, int size)
1168 /* no write fastpath */
1169 if (area->vm_mm == ZS_MM_RO)
1173 buf = buf + ZS_HANDLE_SIZE;
1174 size -= ZS_HANDLE_SIZE;
1175 off += ZS_HANDLE_SIZE;
1177 sizes[0] = PAGE_SIZE - off;
1178 sizes[1] = size - sizes[0];
1180 /* copy per-cpu buffer to object */
1181 addr = kmap_atomic(pages[0]);
1182 memcpy(addr + off, buf, sizes[0]);
1183 kunmap_atomic(addr);
1184 addr = kmap_atomic(pages[1]);
1185 memcpy(addr, buf + sizes[0], sizes[1]);
1186 kunmap_atomic(addr);
1189 /* enable page faults to match kunmap_atomic() return conditions */
1193 static int zs_cpu_prepare(unsigned int cpu)
1195 struct mapping_area *area;
1197 area = &per_cpu(zs_map_area, cpu);
1198 return __zs_cpu_up(area);
1201 static int zs_cpu_dead(unsigned int cpu)
1203 struct mapping_area *area;
1205 area = &per_cpu(zs_map_area, cpu);
1206 __zs_cpu_down(area);
1210 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1211 int objs_per_zspage)
1213 if (prev->pages_per_zspage == pages_per_zspage &&
1214 prev->objs_per_zspage == objs_per_zspage)
1220 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1222 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1225 unsigned long zs_get_total_pages(struct zs_pool *pool)
1227 return atomic_long_read(&pool->pages_allocated);
1229 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1232 * zs_map_object - get address of allocated object from handle.
1233 * @pool: pool from which the object was allocated
1234 * @handle: handle returned from zs_malloc
1235 * @mm: mapping mode to use
1237 * Before using an object allocated from zs_malloc, it must be mapped using
1238 * this function. When done with the object, it must be unmapped using
1241 * Only one object can be mapped per cpu at a time. There is no protection
1242 * against nested mappings.
1244 * This function returns with preemption and page faults disabled.
1246 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1249 struct zspage *zspage;
1251 unsigned long obj, off;
1252 unsigned int obj_idx;
1254 struct size_class *class;
1255 struct mapping_area *area;
1256 struct page *pages[2];
1260 * Because we use per-cpu mapping areas shared among the
1261 * pools/users, we can't allow mapping in interrupt context
1262 * because it can corrupt another users mappings.
1264 BUG_ON(in_interrupt());
1266 /* From now on, migration cannot move the object */
1269 obj = handle_to_obj(handle);
1270 obj_to_location(obj, &page, &obj_idx);
1271 zspage = get_zspage(page);
1273 /* migration cannot move any subpage in this zspage */
1274 migrate_read_lock(zspage);
1276 class = zspage_class(pool, zspage);
1277 off = (class->size * obj_idx) & ~PAGE_MASK;
1279 area = &get_cpu_var(zs_map_area);
1281 if (off + class->size <= PAGE_SIZE) {
1282 /* this object is contained entirely within a page */
1283 area->vm_addr = kmap_atomic(page);
1284 ret = area->vm_addr + off;
1288 /* this object spans two pages */
1290 pages[1] = get_next_page(page);
1293 ret = __zs_map_object(area, pages, off, class->size);
1295 if (likely(!ZsHugePage(zspage)))
1296 ret += ZS_HANDLE_SIZE;
1300 EXPORT_SYMBOL_GPL(zs_map_object);
1302 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1304 struct zspage *zspage;
1306 unsigned long obj, off;
1307 unsigned int obj_idx;
1309 struct size_class *class;
1310 struct mapping_area *area;
1312 obj = handle_to_obj(handle);
1313 obj_to_location(obj, &page, &obj_idx);
1314 zspage = get_zspage(page);
1315 class = zspage_class(pool, zspage);
1316 off = (class->size * obj_idx) & ~PAGE_MASK;
1318 area = this_cpu_ptr(&zs_map_area);
1319 if (off + class->size <= PAGE_SIZE)
1320 kunmap_atomic(area->vm_addr);
1322 struct page *pages[2];
1325 pages[1] = get_next_page(page);
1328 __zs_unmap_object(area, pages, off, class->size);
1330 put_cpu_var(zs_map_area);
1332 migrate_read_unlock(zspage);
1335 EXPORT_SYMBOL_GPL(zs_unmap_object);
1338 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1339 * zsmalloc &size_class.
1340 * @pool: zsmalloc pool to use
1342 * The function returns the size of the first huge class - any object of equal
1343 * or bigger size will be stored in zspage consisting of a single physical
1346 * Context: Any context.
1348 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1350 size_t zs_huge_class_size(struct zs_pool *pool)
1352 return huge_class_size;
1354 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1356 static unsigned long obj_malloc(struct zs_pool *pool,
1357 struct zspage *zspage, unsigned long handle)
1359 int i, nr_page, offset;
1361 struct link_free *link;
1362 struct size_class *class;
1364 struct page *m_page;
1365 unsigned long m_offset;
1368 class = pool->size_class[zspage->class];
1369 handle |= OBJ_ALLOCATED_TAG;
1370 obj = get_freeobj(zspage);
1372 offset = obj * class->size;
1373 nr_page = offset >> PAGE_SHIFT;
1374 m_offset = offset & ~PAGE_MASK;
1375 m_page = get_first_page(zspage);
1377 for (i = 0; i < nr_page; i++)
1378 m_page = get_next_page(m_page);
1380 vaddr = kmap_atomic(m_page);
1381 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1382 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1383 if (likely(!ZsHugePage(zspage)))
1384 /* record handle in the header of allocated chunk */
1385 link->handle = handle;
1387 /* record handle to page->index */
1388 zspage->first_page->index = handle;
1390 kunmap_atomic(vaddr);
1391 mod_zspage_inuse(zspage, 1);
1393 obj = location_to_obj(m_page, obj);
1400 * zs_malloc - Allocate block of given size from pool.
1401 * @pool: pool to allocate from
1402 * @size: size of block to allocate
1403 * @gfp: gfp flags when allocating object
1405 * On success, handle to the allocated object is returned,
1407 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1409 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1411 unsigned long handle, obj;
1412 struct size_class *class;
1413 enum fullness_group newfg;
1414 struct zspage *zspage;
1416 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1419 handle = cache_alloc_handle(pool, gfp);
1423 /* extra space in chunk to keep the handle */
1424 size += ZS_HANDLE_SIZE;
1425 class = pool->size_class[get_size_class_index(size)];
1427 spin_lock(&class->lock);
1428 zspage = find_get_zspage(class);
1429 if (likely(zspage)) {
1430 obj = obj_malloc(pool, zspage, handle);
1431 /* Now move the zspage to another fullness group, if required */
1432 fix_fullness_group(class, zspage);
1433 record_obj(handle, obj);
1434 class_stat_inc(class, OBJ_USED, 1);
1435 spin_unlock(&class->lock);
1440 spin_unlock(&class->lock);
1442 zspage = alloc_zspage(pool, class, gfp);
1444 cache_free_handle(pool, handle);
1448 spin_lock(&class->lock);
1449 obj = obj_malloc(pool, zspage, handle);
1450 newfg = get_fullness_group(class, zspage);
1451 insert_zspage(class, zspage, newfg);
1452 set_zspage_mapping(zspage, class->index, newfg);
1453 record_obj(handle, obj);
1454 atomic_long_add(class->pages_per_zspage,
1455 &pool->pages_allocated);
1456 class_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1457 class_stat_inc(class, OBJ_USED, 1);
1459 /* We completely set up zspage so mark them as movable */
1460 SetZsPageMovable(pool, zspage);
1461 spin_unlock(&class->lock);
1465 EXPORT_SYMBOL_GPL(zs_malloc);
1467 static void obj_free(int class_size, unsigned long obj)
1469 struct link_free *link;
1470 struct zspage *zspage;
1471 struct page *f_page;
1472 unsigned long f_offset;
1473 unsigned int f_objidx;
1476 obj_to_location(obj, &f_page, &f_objidx);
1477 f_offset = (class_size * f_objidx) & ~PAGE_MASK;
1478 zspage = get_zspage(f_page);
1480 vaddr = kmap_atomic(f_page);
1482 /* Insert this object in containing zspage's freelist */
1483 link = (struct link_free *)(vaddr + f_offset);
1484 if (likely(!ZsHugePage(zspage)))
1485 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1488 kunmap_atomic(vaddr);
1489 set_freeobj(zspage, f_objidx);
1490 mod_zspage_inuse(zspage, -1);
1493 void zs_free(struct zs_pool *pool, unsigned long handle)
1495 struct zspage *zspage;
1496 struct page *f_page;
1498 struct size_class *class;
1499 enum fullness_group fullness;
1501 if (unlikely(!handle))
1505 obj = handle_to_obj(handle);
1506 obj_to_page(obj, &f_page);
1507 zspage = get_zspage(f_page);
1509 migrate_read_lock(zspage);
1510 class = zspage_class(pool, zspage);
1512 spin_lock(&class->lock);
1513 obj_free(class->size, obj);
1514 class_stat_dec(class, OBJ_USED, 1);
1515 fullness = fix_fullness_group(class, zspage);
1516 if (fullness != ZS_EMPTY) {
1517 migrate_read_unlock(zspage);
1521 migrate_read_unlock(zspage);
1522 /* If zspage is isolated, zs_page_putback will free the zspage */
1523 free_zspage(pool, class, zspage);
1526 spin_unlock(&class->lock);
1528 cache_free_handle(pool, handle);
1530 EXPORT_SYMBOL_GPL(zs_free);
1532 static void zs_object_copy(struct size_class *class, unsigned long dst,
1535 struct page *s_page, *d_page;
1536 unsigned int s_objidx, d_objidx;
1537 unsigned long s_off, d_off;
1538 void *s_addr, *d_addr;
1539 int s_size, d_size, size;
1542 s_size = d_size = class->size;
1544 obj_to_location(src, &s_page, &s_objidx);
1545 obj_to_location(dst, &d_page, &d_objidx);
1547 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1548 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1550 if (s_off + class->size > PAGE_SIZE)
1551 s_size = PAGE_SIZE - s_off;
1553 if (d_off + class->size > PAGE_SIZE)
1554 d_size = PAGE_SIZE - d_off;
1556 s_addr = kmap_atomic(s_page);
1557 d_addr = kmap_atomic(d_page);
1560 size = min(s_size, d_size);
1561 memcpy(d_addr + d_off, s_addr + s_off, size);
1564 if (written == class->size)
1572 if (s_off >= PAGE_SIZE) {
1573 kunmap_atomic(d_addr);
1574 kunmap_atomic(s_addr);
1575 s_page = get_next_page(s_page);
1576 s_addr = kmap_atomic(s_page);
1577 d_addr = kmap_atomic(d_page);
1578 s_size = class->size - written;
1582 if (d_off >= PAGE_SIZE) {
1583 kunmap_atomic(d_addr);
1584 d_page = get_next_page(d_page);
1585 d_addr = kmap_atomic(d_page);
1586 d_size = class->size - written;
1591 kunmap_atomic(d_addr);
1592 kunmap_atomic(s_addr);
1596 * Find alloced object in zspage from index object and
1599 static unsigned long find_alloced_obj(struct size_class *class,
1600 struct page *page, int *obj_idx)
1603 int index = *obj_idx;
1604 unsigned long handle = 0;
1605 void *addr = kmap_atomic(page);
1607 offset = get_first_obj_offset(page);
1608 offset += class->size * index;
1610 while (offset < PAGE_SIZE) {
1611 if (obj_allocated(page, addr + offset, &handle)) {
1612 if (trypin_tag(handle))
1617 offset += class->size;
1621 kunmap_atomic(addr);
1628 struct zs_compact_control {
1629 /* Source spage for migration which could be a subpage of zspage */
1630 struct page *s_page;
1631 /* Destination page for migration which should be a first page
1633 struct page *d_page;
1634 /* Starting object index within @s_page which used for live object
1635 * in the subpage. */
1639 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1640 struct zs_compact_control *cc)
1642 unsigned long used_obj, free_obj;
1643 unsigned long handle;
1644 struct page *s_page = cc->s_page;
1645 struct page *d_page = cc->d_page;
1646 int obj_idx = cc->obj_idx;
1650 handle = find_alloced_obj(class, s_page, &obj_idx);
1652 s_page = get_next_page(s_page);
1659 /* Stop if there is no more space */
1660 if (zspage_full(class, get_zspage(d_page))) {
1666 used_obj = handle_to_obj(handle);
1667 free_obj = obj_malloc(pool, get_zspage(d_page), handle);
1668 zs_object_copy(class, free_obj, used_obj);
1671 * record_obj updates handle's value to free_obj and it will
1672 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1673 * breaks synchronization using pin_tag(e,g, zs_free) so
1674 * let's keep the lock bit.
1676 free_obj |= BIT(HANDLE_PIN_BIT);
1677 record_obj(handle, free_obj);
1679 obj_free(class->size, used_obj);
1682 /* Remember last position in this iteration */
1683 cc->s_page = s_page;
1684 cc->obj_idx = obj_idx;
1689 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1692 struct zspage *zspage;
1693 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1696 fg[0] = ZS_ALMOST_FULL;
1697 fg[1] = ZS_ALMOST_EMPTY;
1700 for (i = 0; i < 2; i++) {
1701 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1702 struct zspage, list);
1704 remove_zspage(class, zspage, fg[i]);
1713 * putback_zspage - add @zspage into right class's fullness list
1714 * @class: destination class
1715 * @zspage: target page
1717 * Return @zspage's fullness_group
1719 static enum fullness_group putback_zspage(struct size_class *class,
1720 struct zspage *zspage)
1722 enum fullness_group fullness;
1724 fullness = get_fullness_group(class, zspage);
1725 insert_zspage(class, zspage, fullness);
1726 set_zspage_mapping(zspage, class->index, fullness);
1731 #ifdef CONFIG_COMPACTION
1733 * To prevent zspage destroy during migration, zspage freeing should
1734 * hold locks of all pages in the zspage.
1736 static void lock_zspage(struct zspage *zspage)
1738 struct page *page = get_first_page(zspage);
1742 } while ((page = get_next_page(page)) != NULL);
1745 static int zs_init_fs_context(struct fs_context *fc)
1747 return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1750 static struct file_system_type zsmalloc_fs = {
1752 .init_fs_context = zs_init_fs_context,
1753 .kill_sb = kill_anon_super,
1756 static int zsmalloc_mount(void)
1760 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1761 if (IS_ERR(zsmalloc_mnt))
1762 ret = PTR_ERR(zsmalloc_mnt);
1767 static void zsmalloc_unmount(void)
1769 kern_unmount(zsmalloc_mnt);
1772 static void migrate_lock_init(struct zspage *zspage)
1774 rwlock_init(&zspage->lock);
1777 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1779 read_lock(&zspage->lock);
1782 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1784 read_unlock(&zspage->lock);
1787 static void migrate_write_lock(struct zspage *zspage)
1789 write_lock(&zspage->lock);
1792 static void migrate_write_unlock(struct zspage *zspage)
1794 write_unlock(&zspage->lock);
1797 /* Number of isolated subpage for *page migration* in this zspage */
1798 static void inc_zspage_isolation(struct zspage *zspage)
1803 static void dec_zspage_isolation(struct zspage *zspage)
1805 VM_BUG_ON(zspage->isolated == 0);
1809 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1810 struct page *newpage, struct page *oldpage)
1813 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1816 page = get_first_page(zspage);
1818 if (page == oldpage)
1819 pages[idx] = newpage;
1823 } while ((page = get_next_page(page)) != NULL);
1825 create_page_chain(class, zspage, pages);
1826 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1827 if (unlikely(ZsHugePage(zspage)))
1828 newpage->index = oldpage->index;
1829 __SetPageMovable(newpage, page_mapping(oldpage));
1832 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1834 struct zspage *zspage;
1837 * Page is locked so zspage couldn't be destroyed. For detail, look at
1838 * lock_zspage in free_zspage.
1840 VM_BUG_ON_PAGE(!PageMovable(page), page);
1841 VM_BUG_ON_PAGE(PageIsolated(page), page);
1843 zspage = get_zspage(page);
1844 migrate_write_lock(zspage);
1845 inc_zspage_isolation(zspage);
1846 migrate_write_unlock(zspage);
1851 static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1852 struct page *page, enum migrate_mode mode)
1854 struct zs_pool *pool;
1855 struct size_class *class;
1856 struct zspage *zspage;
1858 void *s_addr, *d_addr, *addr;
1860 unsigned long handle;
1861 unsigned long old_obj, new_obj;
1862 unsigned int obj_idx;
1866 * We cannot support the _NO_COPY case here, because copy needs to
1867 * happen under the zs lock, which does not work with
1868 * MIGRATE_SYNC_NO_COPY workflow.
1870 if (mode == MIGRATE_SYNC_NO_COPY)
1873 VM_BUG_ON_PAGE(!PageMovable(page), page);
1874 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1876 zspage = get_zspage(page);
1878 /* Concurrent compactor cannot migrate any subpage in zspage */
1879 migrate_write_lock(zspage);
1880 pool = mapping->private_data;
1881 class = zspage_class(pool, zspage);
1882 offset = get_first_obj_offset(page);
1884 spin_lock(&class->lock);
1885 if (!get_zspage_inuse(zspage)) {
1887 * Set "offset" to end of the page so that every loops
1888 * skips unnecessary object scanning.
1894 s_addr = kmap_atomic(page);
1895 while (pos < PAGE_SIZE) {
1896 if (obj_allocated(page, s_addr + pos, &handle)) {
1897 if (!trypin_tag(handle))
1904 * Here, any user cannot access all objects in the zspage so let's move.
1906 d_addr = kmap_atomic(newpage);
1907 memcpy(d_addr, s_addr, PAGE_SIZE);
1908 kunmap_atomic(d_addr);
1910 for (addr = s_addr + offset; addr < s_addr + pos;
1911 addr += class->size) {
1912 if (obj_allocated(page, addr, &handle)) {
1913 BUG_ON(!testpin_tag(handle));
1915 old_obj = handle_to_obj(handle);
1916 obj_to_location(old_obj, &dummy, &obj_idx);
1917 new_obj = (unsigned long)location_to_obj(newpage,
1919 new_obj |= BIT(HANDLE_PIN_BIT);
1920 record_obj(handle, new_obj);
1924 replace_sub_page(class, zspage, newpage, page);
1927 dec_zspage_isolation(zspage);
1929 if (page_zone(newpage) != page_zone(page)) {
1930 dec_zone_page_state(page, NR_ZSPAGES);
1931 inc_zone_page_state(newpage, NR_ZSPAGES);
1938 ret = MIGRATEPAGE_SUCCESS;
1940 for (addr = s_addr + offset; addr < s_addr + pos;
1941 addr += class->size) {
1942 if (obj_allocated(page, addr, &handle)) {
1943 BUG_ON(!testpin_tag(handle));
1947 kunmap_atomic(s_addr);
1948 spin_unlock(&class->lock);
1949 migrate_write_unlock(zspage);
1954 static void zs_page_putback(struct page *page)
1956 struct zspage *zspage;
1958 VM_BUG_ON_PAGE(!PageMovable(page), page);
1959 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1961 zspage = get_zspage(page);
1962 migrate_write_lock(zspage);
1963 dec_zspage_isolation(zspage);
1964 migrate_write_unlock(zspage);
1967 static const struct address_space_operations zsmalloc_aops = {
1968 .isolate_page = zs_page_isolate,
1969 .migratepage = zs_page_migrate,
1970 .putback_page = zs_page_putback,
1973 static int zs_register_migration(struct zs_pool *pool)
1975 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
1976 if (IS_ERR(pool->inode)) {
1981 pool->inode->i_mapping->private_data = pool;
1982 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
1986 static void zs_unregister_migration(struct zs_pool *pool)
1988 flush_work(&pool->free_work);
1993 * Caller should hold page_lock of all pages in the zspage
1994 * In here, we cannot use zspage meta data.
1996 static void async_free_zspage(struct work_struct *work)
1999 struct size_class *class;
2000 unsigned int class_idx;
2001 enum fullness_group fullness;
2002 struct zspage *zspage, *tmp;
2003 LIST_HEAD(free_pages);
2004 struct zs_pool *pool = container_of(work, struct zs_pool,
2007 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2008 class = pool->size_class[i];
2009 if (class->index != i)
2012 spin_lock(&class->lock);
2013 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2014 spin_unlock(&class->lock);
2017 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2018 list_del(&zspage->list);
2019 lock_zspage(zspage);
2021 get_zspage_mapping(zspage, &class_idx, &fullness);
2022 VM_BUG_ON(fullness != ZS_EMPTY);
2023 class = pool->size_class[class_idx];
2024 spin_lock(&class->lock);
2025 __free_zspage(pool, class, zspage);
2026 spin_unlock(&class->lock);
2030 static void kick_deferred_free(struct zs_pool *pool)
2032 schedule_work(&pool->free_work);
2035 static void init_deferred_free(struct zs_pool *pool)
2037 INIT_WORK(&pool->free_work, async_free_zspage);
2040 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2042 struct page *page = get_first_page(zspage);
2045 WARN_ON(!trylock_page(page));
2046 __SetPageMovable(page, pool->inode->i_mapping);
2048 } while ((page = get_next_page(page)) != NULL);
2054 * Based on the number of unused allocated objects calculate
2055 * and return the number of pages that we can free.
2057 static unsigned long zs_can_compact(struct size_class *class)
2059 unsigned long obj_wasted;
2060 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2061 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2063 if (obj_allocated <= obj_used)
2066 obj_wasted = obj_allocated - obj_used;
2067 obj_wasted /= class->objs_per_zspage;
2069 return obj_wasted * class->pages_per_zspage;
2072 static unsigned long __zs_compact(struct zs_pool *pool,
2073 struct size_class *class)
2075 struct zs_compact_control cc;
2076 struct zspage *src_zspage;
2077 struct zspage *dst_zspage = NULL;
2078 unsigned long pages_freed = 0;
2080 spin_lock(&class->lock);
2081 while ((src_zspage = isolate_zspage(class, true))) {
2083 if (!zs_can_compact(class))
2087 cc.s_page = get_first_page(src_zspage);
2089 while ((dst_zspage = isolate_zspage(class, false))) {
2090 cc.d_page = get_first_page(dst_zspage);
2092 * If there is no more space in dst_page, resched
2093 * and see if anyone had allocated another zspage.
2095 if (!migrate_zspage(pool, class, &cc))
2098 putback_zspage(class, dst_zspage);
2101 /* Stop if we couldn't find slot */
2102 if (dst_zspage == NULL)
2105 putback_zspage(class, dst_zspage);
2106 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2107 free_zspage(pool, class, src_zspage);
2108 pages_freed += class->pages_per_zspage;
2110 spin_unlock(&class->lock);
2112 spin_lock(&class->lock);
2116 putback_zspage(class, src_zspage);
2118 spin_unlock(&class->lock);
2123 unsigned long zs_compact(struct zs_pool *pool)
2126 struct size_class *class;
2127 unsigned long pages_freed = 0;
2129 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2130 class = pool->size_class[i];
2133 if (class->index != i)
2135 pages_freed += __zs_compact(pool, class);
2137 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2141 EXPORT_SYMBOL_GPL(zs_compact);
2143 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2145 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2147 EXPORT_SYMBOL_GPL(zs_pool_stats);
2149 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2150 struct shrink_control *sc)
2152 unsigned long pages_freed;
2153 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2157 * Compact classes and calculate compaction delta.
2158 * Can run concurrently with a manually triggered
2159 * (by user) compaction.
2161 pages_freed = zs_compact(pool);
2163 return pages_freed ? pages_freed : SHRINK_STOP;
2166 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2167 struct shrink_control *sc)
2170 struct size_class *class;
2171 unsigned long pages_to_free = 0;
2172 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2175 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2176 class = pool->size_class[i];
2179 if (class->index != i)
2182 pages_to_free += zs_can_compact(class);
2185 return pages_to_free;
2188 static void zs_unregister_shrinker(struct zs_pool *pool)
2190 unregister_shrinker(&pool->shrinker);
2193 static int zs_register_shrinker(struct zs_pool *pool)
2195 pool->shrinker.scan_objects = zs_shrinker_scan;
2196 pool->shrinker.count_objects = zs_shrinker_count;
2197 pool->shrinker.batch = 0;
2198 pool->shrinker.seeks = DEFAULT_SEEKS;
2200 return register_shrinker(&pool->shrinker);
2204 * zs_create_pool - Creates an allocation pool to work from.
2205 * @name: pool name to be created
2207 * This function must be called before anything when using
2208 * the zsmalloc allocator.
2210 * On success, a pointer to the newly created pool is returned,
2213 struct zs_pool *zs_create_pool(const char *name)
2216 struct zs_pool *pool;
2217 struct size_class *prev_class = NULL;
2219 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2223 init_deferred_free(pool);
2225 pool->name = kstrdup(name, GFP_KERNEL);
2229 if (create_cache(pool))
2233 * Iterate reversely, because, size of size_class that we want to use
2234 * for merging should be larger or equal to current size.
2236 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2238 int pages_per_zspage;
2239 int objs_per_zspage;
2240 struct size_class *class;
2243 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2244 if (size > ZS_MAX_ALLOC_SIZE)
2245 size = ZS_MAX_ALLOC_SIZE;
2246 pages_per_zspage = get_pages_per_zspage(size);
2247 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2250 * We iterate from biggest down to smallest classes,
2251 * so huge_class_size holds the size of the first huge
2252 * class. Any object bigger than or equal to that will
2253 * endup in the huge class.
2255 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2257 huge_class_size = size;
2259 * The object uses ZS_HANDLE_SIZE bytes to store the
2260 * handle. We need to subtract it, because zs_malloc()
2261 * unconditionally adds handle size before it performs
2262 * size class search - so object may be smaller than
2263 * huge class size, yet it still can end up in the huge
2264 * class because it grows by ZS_HANDLE_SIZE extra bytes
2265 * right before class lookup.
2267 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2271 * size_class is used for normal zsmalloc operation such
2272 * as alloc/free for that size. Although it is natural that we
2273 * have one size_class for each size, there is a chance that we
2274 * can get more memory utilization if we use one size_class for
2275 * many different sizes whose size_class have same
2276 * characteristics. So, we makes size_class point to
2277 * previous size_class if possible.
2280 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2281 pool->size_class[i] = prev_class;
2286 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2292 class->pages_per_zspage = pages_per_zspage;
2293 class->objs_per_zspage = objs_per_zspage;
2294 spin_lock_init(&class->lock);
2295 pool->size_class[i] = class;
2296 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2298 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2303 /* debug only, don't abort if it fails */
2304 zs_pool_stat_create(pool, name);
2306 if (zs_register_migration(pool))
2310 * Not critical since shrinker is only used to trigger internal
2311 * defragmentation of the pool which is pretty optional thing. If
2312 * registration fails we still can use the pool normally and user can
2313 * trigger compaction manually. Thus, ignore return code.
2315 zs_register_shrinker(pool);
2320 zs_destroy_pool(pool);
2323 EXPORT_SYMBOL_GPL(zs_create_pool);
2325 void zs_destroy_pool(struct zs_pool *pool)
2329 zs_unregister_shrinker(pool);
2330 zs_unregister_migration(pool);
2331 zs_pool_stat_destroy(pool);
2333 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2335 struct size_class *class = pool->size_class[i];
2340 if (class->index != i)
2343 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2344 if (!list_empty(&class->fullness_list[fg])) {
2345 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2352 destroy_cache(pool);
2356 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2358 static int __init zs_init(void)
2362 ret = zsmalloc_mount();
2366 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2367 zs_cpu_prepare, zs_cpu_dead);
2372 zpool_register_driver(&zs_zpool_driver);
2385 static void __exit zs_exit(void)
2388 zpool_unregister_driver(&zs_zpool_driver);
2391 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2396 module_init(zs_init);
2397 module_exit(zs_exit);
2399 MODULE_LICENSE("Dual BSD/GPL");
2400 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");