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
41 #include <linux/module.h>
42 #include <linux/kernel.h>
43 #include <linux/sched.h>
44 #include <linux/magic.h>
45 #include <linux/bitops.h>
46 #include <linux/errno.h>
47 #include <linux/highmem.h>
48 #include <linux/string.h>
49 #include <linux/slab.h>
50 #include <linux/pgtable.h>
51 #include <asm/tlbflush.h>
52 #include <linux/cpumask.h>
53 #include <linux/cpu.h>
54 #include <linux/vmalloc.h>
55 #include <linux/preempt.h>
56 #include <linux/spinlock.h>
57 #include <linux/shrinker.h>
58 #include <linux/types.h>
59 #include <linux/debugfs.h>
60 #include <linux/zsmalloc.h>
61 #include <linux/zpool.h>
62 #include <linux/mount.h>
63 #include <linux/pseudo_fs.h>
64 #include <linux/migrate.h>
65 #include <linux/wait.h>
66 #include <linux/pagemap.h>
69 #define ZSPAGE_MAGIC 0x58
72 * This must be power of 2 and greater than or equal to sizeof(link_free).
73 * These two conditions ensure that any 'struct link_free' itself doesn't
74 * span more than 1 page which avoids complex case of mapping 2 pages simply
75 * to restore link_free pointer values.
80 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
81 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
83 #define ZS_MAX_ZSPAGE_ORDER 2
84 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
86 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
89 * Object location (<PFN>, <obj_idx>) is encoded as
90 * a single (unsigned long) handle value.
92 * Note that object index <obj_idx> starts from 0.
94 * This is made more complicated by various memory models and PAE.
97 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
98 #ifdef MAX_PHYSMEM_BITS
99 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
102 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
105 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
109 #define _PFN_BITS (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
112 * Head in allocated object should have OBJ_ALLOCATED_TAG
113 * to identify the object was allocated or not.
114 * It's okay to add the status bit in the least bit because
115 * header keeps handle which is 4byte-aligned address so we
116 * have room for two bit at least.
118 #define OBJ_ALLOCATED_TAG 1
119 #define OBJ_TAG_BITS 1
120 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
121 #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;
217 * Placed within free objects to form a singly linked list.
218 * For every zspage, zspage->freeobj gives head of this list.
220 * This must be power of 2 and less than or equal to ZS_ALIGN
226 * It's valid for non-allocated object
230 * Handle of allocated object.
232 unsigned long handle;
239 struct size_class *size_class[ZS_SIZE_CLASSES];
240 struct kmem_cache *handle_cachep;
241 struct kmem_cache *zspage_cachep;
243 atomic_long_t pages_allocated;
245 struct zs_pool_stats stats;
247 /* Compact classes */
248 struct shrinker shrinker;
250 #ifdef CONFIG_ZSMALLOC_STAT
251 struct dentry *stat_dentry;
253 #ifdef CONFIG_COMPACTION
255 struct work_struct free_work;
257 /* protect page/zspage migration */
258 rwlock_t migrate_lock;
263 unsigned int huge:HUGE_BITS;
264 unsigned int fullness:FULLNESS_BITS;
265 unsigned int class:CLASS_BITS + 1;
266 unsigned int isolated:ISOLATED_BITS;
267 unsigned int magic:MAGIC_VAL_BITS;
270 unsigned int freeobj;
271 struct page *first_page;
272 struct list_head list; /* fullness list */
273 #ifdef CONFIG_COMPACTION
278 struct mapping_area {
279 char *vm_buf; /* copy buffer for objects that span pages */
280 char *vm_addr; /* address of kmap_atomic()'ed pages */
281 enum zs_mapmode vm_mm; /* mapping mode */
284 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
285 static void SetZsHugePage(struct zspage *zspage)
290 static bool ZsHugePage(struct zspage *zspage)
295 #ifdef CONFIG_COMPACTION
296 static int zs_register_migration(struct zs_pool *pool);
297 static void zs_unregister_migration(struct zs_pool *pool);
298 static void migrate_lock_init(struct zspage *zspage);
299 static void migrate_read_lock(struct zspage *zspage);
300 static void migrate_read_unlock(struct zspage *zspage);
301 static void migrate_write_lock(struct zspage *zspage);
302 static void migrate_write_lock_nested(struct zspage *zspage);
303 static void migrate_write_unlock(struct zspage *zspage);
304 static void kick_deferred_free(struct zs_pool *pool);
305 static void init_deferred_free(struct zs_pool *pool);
306 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
308 static int zsmalloc_mount(void) { return 0; }
309 static void zsmalloc_unmount(void) {}
310 static int zs_register_migration(struct zs_pool *pool) { return 0; }
311 static void zs_unregister_migration(struct zs_pool *pool) {}
312 static void migrate_lock_init(struct zspage *zspage) {}
313 static void migrate_read_lock(struct zspage *zspage) {}
314 static void migrate_read_unlock(struct zspage *zspage) {}
315 static void migrate_write_lock(struct zspage *zspage) {}
316 static void migrate_write_lock_nested(struct zspage *zspage) {}
317 static void migrate_write_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 /* class->lock(which owns the handle) synchronizes races */
370 static void record_obj(unsigned long handle, unsigned long obj)
372 *(unsigned long *)handle = obj;
379 static void *zs_zpool_create(const char *name, gfp_t gfp,
380 const struct zpool_ops *zpool_ops,
384 * Ignore global gfp flags: zs_malloc() may be invoked from
385 * different contexts and its caller must provide a valid
388 return zs_create_pool(name);
391 static void zs_zpool_destroy(void *pool)
393 zs_destroy_pool(pool);
396 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
397 unsigned long *handle)
399 *handle = zs_malloc(pool, size, gfp);
400 return *handle ? 0 : -1;
402 static void zs_zpool_free(void *pool, unsigned long handle)
404 zs_free(pool, handle);
407 static void *zs_zpool_map(void *pool, unsigned long handle,
408 enum zpool_mapmode mm)
410 enum zs_mapmode zs_mm;
425 return zs_map_object(pool, handle, zs_mm);
427 static void zs_zpool_unmap(void *pool, unsigned long handle)
429 zs_unmap_object(pool, handle);
432 static u64 zs_zpool_total_size(void *pool)
434 return zs_get_total_pages(pool) << PAGE_SHIFT;
437 static struct zpool_driver zs_zpool_driver = {
439 .owner = THIS_MODULE,
440 .create = zs_zpool_create,
441 .destroy = zs_zpool_destroy,
442 .malloc_support_movable = true,
443 .malloc = zs_zpool_malloc,
444 .free = zs_zpool_free,
446 .unmap = zs_zpool_unmap,
447 .total_size = zs_zpool_total_size,
450 MODULE_ALIAS("zpool-zsmalloc");
451 #endif /* CONFIG_ZPOOL */
453 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
454 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
456 static __maybe_unused int is_first_page(struct page *page)
458 return PagePrivate(page);
461 /* Protected by class->lock */
462 static inline int get_zspage_inuse(struct zspage *zspage)
464 return zspage->inuse;
468 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
470 zspage->inuse += val;
473 static inline struct page *get_first_page(struct zspage *zspage)
475 struct page *first_page = zspage->first_page;
477 VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
481 static inline int get_first_obj_offset(struct page *page)
483 return page->page_type;
486 static inline void set_first_obj_offset(struct page *page, int offset)
488 page->page_type = offset;
491 static inline unsigned int get_freeobj(struct zspage *zspage)
493 return zspage->freeobj;
496 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
498 zspage->freeobj = obj;
501 static void get_zspage_mapping(struct zspage *zspage,
502 unsigned int *class_idx,
503 enum fullness_group *fullness)
505 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
507 *fullness = zspage->fullness;
508 *class_idx = zspage->class;
511 static struct size_class *zspage_class(struct zs_pool *pool,
512 struct zspage *zspage)
514 return pool->size_class[zspage->class];
517 static void set_zspage_mapping(struct zspage *zspage,
518 unsigned int class_idx,
519 enum fullness_group fullness)
521 zspage->class = class_idx;
522 zspage->fullness = fullness;
526 * zsmalloc divides the pool into various size classes where each
527 * class maintains a list of zspages where each zspage is divided
528 * into equal sized chunks. Each allocation falls into one of these
529 * classes depending on its size. This function returns index of the
530 * size class which has chunk size big enough to hold the given size.
532 static int get_size_class_index(int size)
536 if (likely(size > ZS_MIN_ALLOC_SIZE))
537 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
538 ZS_SIZE_CLASS_DELTA);
540 return min_t(int, ZS_SIZE_CLASSES - 1, idx);
543 /* type can be of enum type class_stat_type or fullness_group */
544 static inline void class_stat_inc(struct size_class *class,
545 int type, unsigned long cnt)
547 class->stats.objs[type] += cnt;
550 /* type can be of enum type class_stat_type or fullness_group */
551 static inline void class_stat_dec(struct size_class *class,
552 int type, unsigned long cnt)
554 class->stats.objs[type] -= cnt;
557 /* type can be of enum type class_stat_type or fullness_group */
558 static inline unsigned long zs_stat_get(struct size_class *class,
561 return class->stats.objs[type];
564 #ifdef CONFIG_ZSMALLOC_STAT
566 static void __init zs_stat_init(void)
568 if (!debugfs_initialized()) {
569 pr_warn("debugfs not available, stat dir not created\n");
573 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
576 static void __exit zs_stat_exit(void)
578 debugfs_remove_recursive(zs_stat_root);
581 static unsigned long zs_can_compact(struct size_class *class);
583 static int zs_stats_size_show(struct seq_file *s, void *v)
586 struct zs_pool *pool = s->private;
587 struct size_class *class;
589 unsigned long class_almost_full, class_almost_empty;
590 unsigned long obj_allocated, obj_used, pages_used, freeable;
591 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
592 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
593 unsigned long total_freeable = 0;
595 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
596 "class", "size", "almost_full", "almost_empty",
597 "obj_allocated", "obj_used", "pages_used",
598 "pages_per_zspage", "freeable");
600 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
601 class = pool->size_class[i];
603 if (class->index != i)
606 spin_lock(&class->lock);
607 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
608 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
609 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
610 obj_used = zs_stat_get(class, OBJ_USED);
611 freeable = zs_can_compact(class);
612 spin_unlock(&class->lock);
614 objs_per_zspage = class->objs_per_zspage;
615 pages_used = obj_allocated / objs_per_zspage *
616 class->pages_per_zspage;
618 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
619 " %10lu %10lu %16d %8lu\n",
620 i, class->size, class_almost_full, class_almost_empty,
621 obj_allocated, obj_used, pages_used,
622 class->pages_per_zspage, freeable);
624 total_class_almost_full += class_almost_full;
625 total_class_almost_empty += class_almost_empty;
626 total_objs += obj_allocated;
627 total_used_objs += obj_used;
628 total_pages += pages_used;
629 total_freeable += freeable;
633 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
634 "Total", "", total_class_almost_full,
635 total_class_almost_empty, total_objs,
636 total_used_objs, total_pages, "", total_freeable);
640 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
642 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
645 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
649 pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
651 debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
652 &zs_stats_size_fops);
655 static void zs_pool_stat_destroy(struct zs_pool *pool)
657 debugfs_remove_recursive(pool->stat_dentry);
660 #else /* CONFIG_ZSMALLOC_STAT */
661 static void __init zs_stat_init(void)
665 static void __exit zs_stat_exit(void)
669 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
673 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
680 * For each size class, zspages are divided into different groups
681 * depending on how "full" they are. This was done so that we could
682 * easily find empty or nearly empty zspages when we try to shrink
683 * the pool (not yet implemented). This function returns fullness
684 * status of the given page.
686 static enum fullness_group get_fullness_group(struct size_class *class,
687 struct zspage *zspage)
689 int inuse, objs_per_zspage;
690 enum fullness_group fg;
692 inuse = get_zspage_inuse(zspage);
693 objs_per_zspage = class->objs_per_zspage;
697 else if (inuse == objs_per_zspage)
699 else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
700 fg = ZS_ALMOST_EMPTY;
708 * Each size class maintains various freelists and zspages are assigned
709 * to one of these freelists based on the number of live objects they
710 * have. This functions inserts the given zspage into the freelist
711 * identified by <class, fullness_group>.
713 static void insert_zspage(struct size_class *class,
714 struct zspage *zspage,
715 enum fullness_group fullness)
719 class_stat_inc(class, fullness, 1);
720 head = list_first_entry_or_null(&class->fullness_list[fullness],
721 struct zspage, list);
723 * We want to see more ZS_FULL pages and less almost empty/full.
724 * Put pages with higher ->inuse first.
726 if (head && get_zspage_inuse(zspage) < get_zspage_inuse(head))
727 list_add(&zspage->list, &head->list);
729 list_add(&zspage->list, &class->fullness_list[fullness]);
733 * This function removes the given zspage from the freelist identified
734 * by <class, fullness_group>.
736 static void remove_zspage(struct size_class *class,
737 struct zspage *zspage,
738 enum fullness_group fullness)
740 VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
742 list_del_init(&zspage->list);
743 class_stat_dec(class, fullness, 1);
747 * Each size class maintains zspages in different fullness groups depending
748 * on the number of live objects they contain. When allocating or freeing
749 * objects, the fullness status of the page can change, say, from ALMOST_FULL
750 * to ALMOST_EMPTY when freeing an object. This function checks if such
751 * a status change has occurred for the given page and accordingly moves the
752 * page from the freelist of the old fullness group to that of the new
755 static enum fullness_group fix_fullness_group(struct size_class *class,
756 struct zspage *zspage)
759 enum fullness_group currfg, newfg;
761 get_zspage_mapping(zspage, &class_idx, &currfg);
762 newfg = get_fullness_group(class, zspage);
766 remove_zspage(class, zspage, currfg);
767 insert_zspage(class, zspage, newfg);
768 set_zspage_mapping(zspage, class_idx, newfg);
774 * We have to decide on how many pages to link together
775 * to form a zspage for each size class. This is important
776 * to reduce wastage due to unusable space left at end of
777 * each zspage which is given as:
778 * wastage = Zp % class_size
779 * usage = Zp - wastage
780 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
782 * For example, for size class of 3/8 * PAGE_SIZE, we should
783 * link together 3 PAGE_SIZE sized pages to form a zspage
784 * since then we can perfectly fit in 8 such objects.
786 static int get_pages_per_zspage(int class_size)
788 int i, max_usedpc = 0;
789 /* zspage order which gives maximum used size per KB */
790 int max_usedpc_order = 1;
792 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
796 zspage_size = i * PAGE_SIZE;
797 waste = zspage_size % class_size;
798 usedpc = (zspage_size - waste) * 100 / zspage_size;
800 if (usedpc > max_usedpc) {
802 max_usedpc_order = i;
806 return max_usedpc_order;
809 static struct zspage *get_zspage(struct page *page)
811 struct zspage *zspage = (struct zspage *)page_private(page);
813 BUG_ON(zspage->magic != ZSPAGE_MAGIC);
817 static struct page *get_next_page(struct page *page)
819 struct zspage *zspage = get_zspage(page);
821 if (unlikely(ZsHugePage(zspage)))
824 return (struct page *)page->index;
828 * obj_to_location - get (<page>, <obj_idx>) from encoded object value
829 * @obj: the encoded object value
830 * @page: page object resides in zspage
831 * @obj_idx: object index
833 static void obj_to_location(unsigned long obj, struct page **page,
834 unsigned int *obj_idx)
836 obj >>= OBJ_TAG_BITS;
837 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
838 *obj_idx = (obj & OBJ_INDEX_MASK);
841 static void obj_to_page(unsigned long obj, struct page **page)
843 obj >>= OBJ_TAG_BITS;
844 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
848 * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
849 * @page: page object resides in zspage
850 * @obj_idx: object index
852 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
856 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
857 obj |= obj_idx & OBJ_INDEX_MASK;
858 obj <<= OBJ_TAG_BITS;
863 static unsigned long handle_to_obj(unsigned long handle)
865 return *(unsigned long *)handle;
868 static bool obj_allocated(struct page *page, void *obj, unsigned long *phandle)
870 unsigned long handle;
871 struct zspage *zspage = get_zspage(page);
873 if (unlikely(ZsHugePage(zspage))) {
874 VM_BUG_ON_PAGE(!is_first_page(page), page);
875 handle = page->index;
877 handle = *(unsigned long *)obj;
879 if (!(handle & OBJ_ALLOCATED_TAG))
882 *phandle = handle & ~OBJ_ALLOCATED_TAG;
886 static void reset_page(struct page *page)
888 __ClearPageMovable(page);
889 ClearPagePrivate(page);
890 set_page_private(page, 0);
891 page_mapcount_reset(page);
895 static int trylock_zspage(struct zspage *zspage)
897 struct page *cursor, *fail;
899 for (cursor = get_first_page(zspage); cursor != NULL; cursor =
900 get_next_page(cursor)) {
901 if (!trylock_page(cursor)) {
909 for (cursor = get_first_page(zspage); cursor != fail; cursor =
910 get_next_page(cursor))
916 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
917 struct zspage *zspage)
919 struct page *page, *next;
920 enum fullness_group fg;
921 unsigned int class_idx;
923 get_zspage_mapping(zspage, &class_idx, &fg);
925 assert_spin_locked(&class->lock);
927 VM_BUG_ON(get_zspage_inuse(zspage));
928 VM_BUG_ON(fg != ZS_EMPTY);
930 next = page = get_first_page(zspage);
932 VM_BUG_ON_PAGE(!PageLocked(page), page);
933 next = get_next_page(page);
936 dec_zone_page_state(page, NR_ZSPAGES);
939 } while (page != NULL);
941 cache_free_zspage(pool, zspage);
943 class_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
944 atomic_long_sub(class->pages_per_zspage,
945 &pool->pages_allocated);
948 static void free_zspage(struct zs_pool *pool, struct size_class *class,
949 struct zspage *zspage)
951 VM_BUG_ON(get_zspage_inuse(zspage));
952 VM_BUG_ON(list_empty(&zspage->list));
955 * Since zs_free couldn't be sleepable, this function cannot call
956 * lock_page. The page locks trylock_zspage got will be released
959 if (!trylock_zspage(zspage)) {
960 kick_deferred_free(pool);
964 remove_zspage(class, zspage, ZS_EMPTY);
965 __free_zspage(pool, class, zspage);
968 /* Initialize a newly allocated zspage */
969 static void init_zspage(struct size_class *class, struct zspage *zspage)
971 unsigned int freeobj = 1;
972 unsigned long off = 0;
973 struct page *page = get_first_page(zspage);
976 struct page *next_page;
977 struct link_free *link;
980 set_first_obj_offset(page, off);
982 vaddr = kmap_atomic(page);
983 link = (struct link_free *)vaddr + off / sizeof(*link);
985 while ((off += class->size) < PAGE_SIZE) {
986 link->next = freeobj++ << OBJ_TAG_BITS;
987 link += class->size / sizeof(*link);
991 * We now come to the last (full or partial) object on this
992 * page, which must point to the first object on the next
995 next_page = get_next_page(page);
997 link->next = freeobj++ << OBJ_TAG_BITS;
1000 * Reset OBJ_TAG_BITS bit to last link to tell
1001 * whether it's allocated object or not.
1003 link->next = -1UL << OBJ_TAG_BITS;
1005 kunmap_atomic(vaddr);
1010 set_freeobj(zspage, 0);
1013 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1014 struct page *pages[])
1018 struct page *prev_page = NULL;
1019 int nr_pages = class->pages_per_zspage;
1022 * Allocate individual pages and link them together as:
1023 * 1. all pages are linked together using page->index
1024 * 2. each sub-page point to zspage using page->private
1026 * we set PG_private to identify the first page (i.e. no other sub-page
1027 * has this flag set).
1029 for (i = 0; i < nr_pages; i++) {
1031 set_page_private(page, (unsigned long)zspage);
1034 zspage->first_page = page;
1035 SetPagePrivate(page);
1036 if (unlikely(class->objs_per_zspage == 1 &&
1037 class->pages_per_zspage == 1))
1038 SetZsHugePage(zspage);
1040 prev_page->index = (unsigned long)page;
1047 * Allocate a zspage for the given size class
1049 static struct zspage *alloc_zspage(struct zs_pool *pool,
1050 struct size_class *class,
1054 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1055 struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1060 zspage->magic = ZSPAGE_MAGIC;
1061 migrate_lock_init(zspage);
1063 for (i = 0; i < class->pages_per_zspage; i++) {
1066 page = alloc_page(gfp);
1069 dec_zone_page_state(pages[i], NR_ZSPAGES);
1070 __free_page(pages[i]);
1072 cache_free_zspage(pool, zspage);
1076 inc_zone_page_state(page, NR_ZSPAGES);
1080 create_page_chain(class, zspage, pages);
1081 init_zspage(class, zspage);
1086 static struct zspage *find_get_zspage(struct size_class *class)
1089 struct zspage *zspage;
1091 for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1092 zspage = list_first_entry_or_null(&class->fullness_list[i],
1093 struct zspage, list);
1101 static inline int __zs_cpu_up(struct mapping_area *area)
1104 * Make sure we don't leak memory if a cpu UP notification
1105 * and zs_init() race and both call zs_cpu_up() on the same cpu
1109 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1115 static inline void __zs_cpu_down(struct mapping_area *area)
1117 kfree(area->vm_buf);
1118 area->vm_buf = NULL;
1121 static void *__zs_map_object(struct mapping_area *area,
1122 struct page *pages[2], int off, int size)
1126 char *buf = area->vm_buf;
1128 /* disable page faults to match kmap_atomic() return conditions */
1129 pagefault_disable();
1131 /* no read fastpath */
1132 if (area->vm_mm == ZS_MM_WO)
1135 sizes[0] = PAGE_SIZE - off;
1136 sizes[1] = size - sizes[0];
1138 /* copy object to per-cpu buffer */
1139 addr = kmap_atomic(pages[0]);
1140 memcpy(buf, addr + off, sizes[0]);
1141 kunmap_atomic(addr);
1142 addr = kmap_atomic(pages[1]);
1143 memcpy(buf + sizes[0], addr, sizes[1]);
1144 kunmap_atomic(addr);
1146 return area->vm_buf;
1149 static void __zs_unmap_object(struct mapping_area *area,
1150 struct page *pages[2], int off, int size)
1156 /* no write fastpath */
1157 if (area->vm_mm == ZS_MM_RO)
1161 buf = buf + ZS_HANDLE_SIZE;
1162 size -= ZS_HANDLE_SIZE;
1163 off += ZS_HANDLE_SIZE;
1165 sizes[0] = PAGE_SIZE - off;
1166 sizes[1] = size - sizes[0];
1168 /* copy per-cpu buffer to object */
1169 addr = kmap_atomic(pages[0]);
1170 memcpy(addr + off, buf, sizes[0]);
1171 kunmap_atomic(addr);
1172 addr = kmap_atomic(pages[1]);
1173 memcpy(addr, buf + sizes[0], sizes[1]);
1174 kunmap_atomic(addr);
1177 /* enable page faults to match kunmap_atomic() return conditions */
1181 static int zs_cpu_prepare(unsigned int cpu)
1183 struct mapping_area *area;
1185 area = &per_cpu(zs_map_area, cpu);
1186 return __zs_cpu_up(area);
1189 static int zs_cpu_dead(unsigned int cpu)
1191 struct mapping_area *area;
1193 area = &per_cpu(zs_map_area, cpu);
1194 __zs_cpu_down(area);
1198 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1199 int objs_per_zspage)
1201 if (prev->pages_per_zspage == pages_per_zspage &&
1202 prev->objs_per_zspage == objs_per_zspage)
1208 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1210 return get_zspage_inuse(zspage) == class->objs_per_zspage;
1213 unsigned long zs_get_total_pages(struct zs_pool *pool)
1215 return atomic_long_read(&pool->pages_allocated);
1217 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1220 * zs_map_object - get address of allocated object from handle.
1221 * @pool: pool from which the object was allocated
1222 * @handle: handle returned from zs_malloc
1223 * @mm: mapping mode to use
1225 * Before using an object allocated from zs_malloc, it must be mapped using
1226 * this function. When done with the object, it must be unmapped using
1229 * Only one object can be mapped per cpu at a time. There is no protection
1230 * against nested mappings.
1232 * This function returns with preemption and page faults disabled.
1234 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1237 struct zspage *zspage;
1239 unsigned long obj, off;
1240 unsigned int obj_idx;
1242 struct size_class *class;
1243 struct mapping_area *area;
1244 struct page *pages[2];
1248 * Because we use per-cpu mapping areas shared among the
1249 * pools/users, we can't allow mapping in interrupt context
1250 * because it can corrupt another users mappings.
1252 BUG_ON(in_interrupt());
1254 /* It guarantees it can get zspage from handle safely */
1255 read_lock(&pool->migrate_lock);
1256 obj = handle_to_obj(handle);
1257 obj_to_location(obj, &page, &obj_idx);
1258 zspage = get_zspage(page);
1261 * migration cannot move any zpages in this zspage. Here, class->lock
1262 * is too heavy since callers would take some time until they calls
1263 * zs_unmap_object API so delegate the locking from class to zspage
1264 * which is smaller granularity.
1266 migrate_read_lock(zspage);
1267 read_unlock(&pool->migrate_lock);
1269 class = zspage_class(pool, zspage);
1270 off = (class->size * obj_idx) & ~PAGE_MASK;
1272 area = &get_cpu_var(zs_map_area);
1274 if (off + class->size <= PAGE_SIZE) {
1275 /* this object is contained entirely within a page */
1276 area->vm_addr = kmap_atomic(page);
1277 ret = area->vm_addr + off;
1281 /* this object spans two pages */
1283 pages[1] = get_next_page(page);
1286 ret = __zs_map_object(area, pages, off, class->size);
1288 if (likely(!ZsHugePage(zspage)))
1289 ret += ZS_HANDLE_SIZE;
1293 EXPORT_SYMBOL_GPL(zs_map_object);
1295 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1297 struct zspage *zspage;
1299 unsigned long obj, off;
1300 unsigned int obj_idx;
1302 struct size_class *class;
1303 struct mapping_area *area;
1305 obj = handle_to_obj(handle);
1306 obj_to_location(obj, &page, &obj_idx);
1307 zspage = get_zspage(page);
1308 class = zspage_class(pool, zspage);
1309 off = (class->size * obj_idx) & ~PAGE_MASK;
1311 area = this_cpu_ptr(&zs_map_area);
1312 if (off + class->size <= PAGE_SIZE)
1313 kunmap_atomic(area->vm_addr);
1315 struct page *pages[2];
1318 pages[1] = get_next_page(page);
1321 __zs_unmap_object(area, pages, off, class->size);
1323 put_cpu_var(zs_map_area);
1325 migrate_read_unlock(zspage);
1327 EXPORT_SYMBOL_GPL(zs_unmap_object);
1330 * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1331 * zsmalloc &size_class.
1332 * @pool: zsmalloc pool to use
1334 * The function returns the size of the first huge class - any object of equal
1335 * or bigger size will be stored in zspage consisting of a single physical
1338 * Context: Any context.
1340 * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1342 size_t zs_huge_class_size(struct zs_pool *pool)
1344 return huge_class_size;
1346 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1348 static unsigned long obj_malloc(struct zs_pool *pool,
1349 struct zspage *zspage, unsigned long handle)
1351 int i, nr_page, offset;
1353 struct link_free *link;
1354 struct size_class *class;
1356 struct page *m_page;
1357 unsigned long m_offset;
1360 class = pool->size_class[zspage->class];
1361 handle |= OBJ_ALLOCATED_TAG;
1362 obj = get_freeobj(zspage);
1364 offset = obj * class->size;
1365 nr_page = offset >> PAGE_SHIFT;
1366 m_offset = offset & ~PAGE_MASK;
1367 m_page = get_first_page(zspage);
1369 for (i = 0; i < nr_page; i++)
1370 m_page = get_next_page(m_page);
1372 vaddr = kmap_atomic(m_page);
1373 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1374 set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1375 if (likely(!ZsHugePage(zspage)))
1376 /* record handle in the header of allocated chunk */
1377 link->handle = handle;
1379 /* record handle to page->index */
1380 zspage->first_page->index = handle;
1382 kunmap_atomic(vaddr);
1383 mod_zspage_inuse(zspage, 1);
1385 obj = location_to_obj(m_page, obj);
1392 * zs_malloc - Allocate block of given size from pool.
1393 * @pool: pool to allocate from
1394 * @size: size of block to allocate
1395 * @gfp: gfp flags when allocating object
1397 * On success, handle to the allocated object is returned,
1399 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1401 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1403 unsigned long handle, obj;
1404 struct size_class *class;
1405 enum fullness_group newfg;
1406 struct zspage *zspage;
1408 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1411 handle = cache_alloc_handle(pool, gfp);
1415 /* extra space in chunk to keep the handle */
1416 size += ZS_HANDLE_SIZE;
1417 class = pool->size_class[get_size_class_index(size)];
1419 /* class->lock effectively protects the zpage migration */
1420 spin_lock(&class->lock);
1421 zspage = find_get_zspage(class);
1422 if (likely(zspage)) {
1423 obj = obj_malloc(pool, zspage, handle);
1424 /* Now move the zspage to another fullness group, if required */
1425 fix_fullness_group(class, zspage);
1426 record_obj(handle, obj);
1427 class_stat_inc(class, OBJ_USED, 1);
1428 spin_unlock(&class->lock);
1433 spin_unlock(&class->lock);
1435 zspage = alloc_zspage(pool, class, gfp);
1437 cache_free_handle(pool, handle);
1441 spin_lock(&class->lock);
1442 obj = obj_malloc(pool, zspage, handle);
1443 newfg = get_fullness_group(class, zspage);
1444 insert_zspage(class, zspage, newfg);
1445 set_zspage_mapping(zspage, class->index, newfg);
1446 record_obj(handle, obj);
1447 atomic_long_add(class->pages_per_zspage,
1448 &pool->pages_allocated);
1449 class_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1450 class_stat_inc(class, OBJ_USED, 1);
1452 /* We completely set up zspage so mark them as movable */
1453 SetZsPageMovable(pool, zspage);
1454 spin_unlock(&class->lock);
1458 EXPORT_SYMBOL_GPL(zs_malloc);
1460 static void obj_free(int class_size, unsigned long obj)
1462 struct link_free *link;
1463 struct zspage *zspage;
1464 struct page *f_page;
1465 unsigned long f_offset;
1466 unsigned int f_objidx;
1469 obj_to_location(obj, &f_page, &f_objidx);
1470 f_offset = (class_size * f_objidx) & ~PAGE_MASK;
1471 zspage = get_zspage(f_page);
1473 vaddr = kmap_atomic(f_page);
1475 /* Insert this object in containing zspage's freelist */
1476 link = (struct link_free *)(vaddr + f_offset);
1477 if (likely(!ZsHugePage(zspage)))
1478 link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1481 kunmap_atomic(vaddr);
1482 set_freeobj(zspage, f_objidx);
1483 mod_zspage_inuse(zspage, -1);
1486 void zs_free(struct zs_pool *pool, unsigned long handle)
1488 struct zspage *zspage;
1489 struct page *f_page;
1491 struct size_class *class;
1492 enum fullness_group fullness;
1494 if (unlikely(!handle))
1498 * The pool->migrate_lock protects the race with zpage's migration
1499 * so it's safe to get the page from handle.
1501 read_lock(&pool->migrate_lock);
1502 obj = handle_to_obj(handle);
1503 obj_to_page(obj, &f_page);
1504 zspage = get_zspage(f_page);
1505 class = zspage_class(pool, zspage);
1506 spin_lock(&class->lock);
1507 read_unlock(&pool->migrate_lock);
1509 obj_free(class->size, obj);
1510 class_stat_dec(class, OBJ_USED, 1);
1511 fullness = fix_fullness_group(class, zspage);
1512 if (fullness != ZS_EMPTY)
1515 free_zspage(pool, class, zspage);
1517 spin_unlock(&class->lock);
1518 cache_free_handle(pool, handle);
1520 EXPORT_SYMBOL_GPL(zs_free);
1522 static void zs_object_copy(struct size_class *class, unsigned long dst,
1525 struct page *s_page, *d_page;
1526 unsigned int s_objidx, d_objidx;
1527 unsigned long s_off, d_off;
1528 void *s_addr, *d_addr;
1529 int s_size, d_size, size;
1532 s_size = d_size = class->size;
1534 obj_to_location(src, &s_page, &s_objidx);
1535 obj_to_location(dst, &d_page, &d_objidx);
1537 s_off = (class->size * s_objidx) & ~PAGE_MASK;
1538 d_off = (class->size * d_objidx) & ~PAGE_MASK;
1540 if (s_off + class->size > PAGE_SIZE)
1541 s_size = PAGE_SIZE - s_off;
1543 if (d_off + class->size > PAGE_SIZE)
1544 d_size = PAGE_SIZE - d_off;
1546 s_addr = kmap_atomic(s_page);
1547 d_addr = kmap_atomic(d_page);
1550 size = min(s_size, d_size);
1551 memcpy(d_addr + d_off, s_addr + s_off, size);
1554 if (written == class->size)
1562 if (s_off >= PAGE_SIZE) {
1563 kunmap_atomic(d_addr);
1564 kunmap_atomic(s_addr);
1565 s_page = get_next_page(s_page);
1566 s_addr = kmap_atomic(s_page);
1567 d_addr = kmap_atomic(d_page);
1568 s_size = class->size - written;
1572 if (d_off >= PAGE_SIZE) {
1573 kunmap_atomic(d_addr);
1574 d_page = get_next_page(d_page);
1575 d_addr = kmap_atomic(d_page);
1576 d_size = class->size - written;
1581 kunmap_atomic(d_addr);
1582 kunmap_atomic(s_addr);
1586 * Find alloced object in zspage from index object and
1589 static unsigned long find_alloced_obj(struct size_class *class,
1590 struct page *page, int *obj_idx)
1593 int index = *obj_idx;
1594 unsigned long handle = 0;
1595 void *addr = kmap_atomic(page);
1597 offset = get_first_obj_offset(page);
1598 offset += class->size * index;
1600 while (offset < PAGE_SIZE) {
1601 if (obj_allocated(page, addr + offset, &handle))
1604 offset += class->size;
1608 kunmap_atomic(addr);
1615 struct zs_compact_control {
1616 /* Source spage for migration which could be a subpage of zspage */
1617 struct page *s_page;
1618 /* Destination page for migration which should be a first page
1620 struct page *d_page;
1621 /* Starting object index within @s_page which used for live object
1622 * in the subpage. */
1626 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1627 struct zs_compact_control *cc)
1629 unsigned long used_obj, free_obj;
1630 unsigned long handle;
1631 struct page *s_page = cc->s_page;
1632 struct page *d_page = cc->d_page;
1633 int obj_idx = cc->obj_idx;
1637 handle = find_alloced_obj(class, s_page, &obj_idx);
1639 s_page = get_next_page(s_page);
1646 /* Stop if there is no more space */
1647 if (zspage_full(class, get_zspage(d_page))) {
1652 used_obj = handle_to_obj(handle);
1653 free_obj = obj_malloc(pool, get_zspage(d_page), handle);
1654 zs_object_copy(class, free_obj, used_obj);
1656 record_obj(handle, free_obj);
1657 obj_free(class->size, used_obj);
1660 /* Remember last position in this iteration */
1661 cc->s_page = s_page;
1662 cc->obj_idx = obj_idx;
1667 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1670 struct zspage *zspage;
1671 enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1674 fg[0] = ZS_ALMOST_FULL;
1675 fg[1] = ZS_ALMOST_EMPTY;
1678 for (i = 0; i < 2; i++) {
1679 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1680 struct zspage, list);
1682 remove_zspage(class, zspage, fg[i]);
1691 * putback_zspage - add @zspage into right class's fullness list
1692 * @class: destination class
1693 * @zspage: target page
1695 * Return @zspage's fullness_group
1697 static enum fullness_group putback_zspage(struct size_class *class,
1698 struct zspage *zspage)
1700 enum fullness_group fullness;
1702 fullness = get_fullness_group(class, zspage);
1703 insert_zspage(class, zspage, fullness);
1704 set_zspage_mapping(zspage, class->index, fullness);
1709 #ifdef CONFIG_COMPACTION
1711 * To prevent zspage destroy during migration, zspage freeing should
1712 * hold locks of all pages in the zspage.
1714 static void lock_zspage(struct zspage *zspage)
1716 struct page *page = get_first_page(zspage);
1720 } while ((page = get_next_page(page)) != NULL);
1723 static int zs_init_fs_context(struct fs_context *fc)
1725 return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1728 static struct file_system_type zsmalloc_fs = {
1730 .init_fs_context = zs_init_fs_context,
1731 .kill_sb = kill_anon_super,
1734 static int zsmalloc_mount(void)
1738 zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1739 if (IS_ERR(zsmalloc_mnt))
1740 ret = PTR_ERR(zsmalloc_mnt);
1745 static void zsmalloc_unmount(void)
1747 kern_unmount(zsmalloc_mnt);
1750 static void migrate_lock_init(struct zspage *zspage)
1752 rwlock_init(&zspage->lock);
1755 static void migrate_read_lock(struct zspage *zspage) __acquires(&zspage->lock)
1757 read_lock(&zspage->lock);
1760 static void migrate_read_unlock(struct zspage *zspage) __releases(&zspage->lock)
1762 read_unlock(&zspage->lock);
1765 static void migrate_write_lock(struct zspage *zspage)
1767 write_lock(&zspage->lock);
1770 static void migrate_write_lock_nested(struct zspage *zspage)
1772 write_lock_nested(&zspage->lock, SINGLE_DEPTH_NESTING);
1775 static void migrate_write_unlock(struct zspage *zspage)
1777 write_unlock(&zspage->lock);
1780 /* Number of isolated subpage for *page migration* in this zspage */
1781 static void inc_zspage_isolation(struct zspage *zspage)
1786 static void dec_zspage_isolation(struct zspage *zspage)
1788 VM_BUG_ON(zspage->isolated == 0);
1792 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1793 struct page *newpage, struct page *oldpage)
1796 struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1799 page = get_first_page(zspage);
1801 if (page == oldpage)
1802 pages[idx] = newpage;
1806 } while ((page = get_next_page(page)) != NULL);
1808 create_page_chain(class, zspage, pages);
1809 set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1810 if (unlikely(ZsHugePage(zspage)))
1811 newpage->index = oldpage->index;
1812 __SetPageMovable(newpage, page_mapping(oldpage));
1815 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1817 struct zspage *zspage;
1820 * Page is locked so zspage couldn't be destroyed. For detail, look at
1821 * lock_zspage in free_zspage.
1823 VM_BUG_ON_PAGE(!PageMovable(page), page);
1824 VM_BUG_ON_PAGE(PageIsolated(page), page);
1826 zspage = get_zspage(page);
1827 migrate_write_lock(zspage);
1828 inc_zspage_isolation(zspage);
1829 migrate_write_unlock(zspage);
1834 static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1835 struct page *page, enum migrate_mode mode)
1837 struct zs_pool *pool;
1838 struct size_class *class;
1839 struct zspage *zspage;
1841 void *s_addr, *d_addr, *addr;
1843 unsigned long handle;
1844 unsigned long old_obj, new_obj;
1845 unsigned int obj_idx;
1848 * We cannot support the _NO_COPY case here, because copy needs to
1849 * happen under the zs lock, which does not work with
1850 * MIGRATE_SYNC_NO_COPY workflow.
1852 if (mode == MIGRATE_SYNC_NO_COPY)
1855 VM_BUG_ON_PAGE(!PageMovable(page), page);
1856 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1858 pool = mapping->private_data;
1861 * The pool migrate_lock protects the race between zpage migration
1864 write_lock(&pool->migrate_lock);
1865 zspage = get_zspage(page);
1866 class = zspage_class(pool, zspage);
1869 * the class lock protects zpage alloc/free in the zspage.
1871 spin_lock(&class->lock);
1872 /* the migrate_write_lock protects zpage access via zs_map_object */
1873 migrate_write_lock(zspage);
1875 offset = get_first_obj_offset(page);
1876 s_addr = kmap_atomic(page);
1879 * Here, any user cannot access all objects in the zspage so let's move.
1881 d_addr = kmap_atomic(newpage);
1882 memcpy(d_addr, s_addr, PAGE_SIZE);
1883 kunmap_atomic(d_addr);
1885 for (addr = s_addr + offset; addr < s_addr + PAGE_SIZE;
1886 addr += class->size) {
1887 if (obj_allocated(page, addr, &handle)) {
1889 old_obj = handle_to_obj(handle);
1890 obj_to_location(old_obj, &dummy, &obj_idx);
1891 new_obj = (unsigned long)location_to_obj(newpage,
1893 record_obj(handle, new_obj);
1896 kunmap_atomic(s_addr);
1898 replace_sub_page(class, zspage, newpage, page);
1900 * Since we complete the data copy and set up new zspage structure,
1901 * it's okay to release migration_lock.
1903 write_unlock(&pool->migrate_lock);
1904 spin_unlock(&class->lock);
1905 dec_zspage_isolation(zspage);
1906 migrate_write_unlock(zspage);
1909 if (page_zone(newpage) != page_zone(page)) {
1910 dec_zone_page_state(page, NR_ZSPAGES);
1911 inc_zone_page_state(newpage, NR_ZSPAGES);
1917 return MIGRATEPAGE_SUCCESS;
1920 static void zs_page_putback(struct page *page)
1922 struct zspage *zspage;
1924 VM_BUG_ON_PAGE(!PageMovable(page), page);
1925 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1927 zspage = get_zspage(page);
1928 migrate_write_lock(zspage);
1929 dec_zspage_isolation(zspage);
1930 migrate_write_unlock(zspage);
1933 static const struct address_space_operations zsmalloc_aops = {
1934 .isolate_page = zs_page_isolate,
1935 .migratepage = zs_page_migrate,
1936 .putback_page = zs_page_putback,
1939 static int zs_register_migration(struct zs_pool *pool)
1941 pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
1942 if (IS_ERR(pool->inode)) {
1947 pool->inode->i_mapping->private_data = pool;
1948 pool->inode->i_mapping->a_ops = &zsmalloc_aops;
1952 static void zs_unregister_migration(struct zs_pool *pool)
1954 flush_work(&pool->free_work);
1959 * Caller should hold page_lock of all pages in the zspage
1960 * In here, we cannot use zspage meta data.
1962 static void async_free_zspage(struct work_struct *work)
1965 struct size_class *class;
1966 unsigned int class_idx;
1967 enum fullness_group fullness;
1968 struct zspage *zspage, *tmp;
1969 LIST_HEAD(free_pages);
1970 struct zs_pool *pool = container_of(work, struct zs_pool,
1973 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
1974 class = pool->size_class[i];
1975 if (class->index != i)
1978 spin_lock(&class->lock);
1979 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
1980 spin_unlock(&class->lock);
1983 list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
1984 list_del(&zspage->list);
1985 lock_zspage(zspage);
1987 get_zspage_mapping(zspage, &class_idx, &fullness);
1988 VM_BUG_ON(fullness != ZS_EMPTY);
1989 class = pool->size_class[class_idx];
1990 spin_lock(&class->lock);
1991 __free_zspage(pool, class, zspage);
1992 spin_unlock(&class->lock);
1996 static void kick_deferred_free(struct zs_pool *pool)
1998 schedule_work(&pool->free_work);
2001 static void init_deferred_free(struct zs_pool *pool)
2003 INIT_WORK(&pool->free_work, async_free_zspage);
2006 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2008 struct page *page = get_first_page(zspage);
2011 WARN_ON(!trylock_page(page));
2012 __SetPageMovable(page, pool->inode->i_mapping);
2014 } while ((page = get_next_page(page)) != NULL);
2020 * Based on the number of unused allocated objects calculate
2021 * and return the number of pages that we can free.
2023 static unsigned long zs_can_compact(struct size_class *class)
2025 unsigned long obj_wasted;
2026 unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2027 unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2029 if (obj_allocated <= obj_used)
2032 obj_wasted = obj_allocated - obj_used;
2033 obj_wasted /= class->objs_per_zspage;
2035 return obj_wasted * class->pages_per_zspage;
2038 static unsigned long __zs_compact(struct zs_pool *pool,
2039 struct size_class *class)
2041 struct zs_compact_control cc;
2042 struct zspage *src_zspage;
2043 struct zspage *dst_zspage = NULL;
2044 unsigned long pages_freed = 0;
2046 /* protect the race between zpage migration and zs_free */
2047 write_lock(&pool->migrate_lock);
2048 /* protect zpage allocation/free */
2049 spin_lock(&class->lock);
2050 while ((src_zspage = isolate_zspage(class, true))) {
2051 /* protect someone accessing the zspage(i.e., zs_map_object) */
2052 migrate_write_lock(src_zspage);
2054 if (!zs_can_compact(class))
2058 cc.s_page = get_first_page(src_zspage);
2060 while ((dst_zspage = isolate_zspage(class, false))) {
2061 migrate_write_lock_nested(dst_zspage);
2063 cc.d_page = get_first_page(dst_zspage);
2065 * If there is no more space in dst_page, resched
2066 * and see if anyone had allocated another zspage.
2068 if (!migrate_zspage(pool, class, &cc))
2071 putback_zspage(class, dst_zspage);
2072 migrate_write_unlock(dst_zspage);
2074 if (rwlock_is_contended(&pool->migrate_lock))
2078 /* Stop if we couldn't find slot */
2079 if (dst_zspage == NULL)
2082 putback_zspage(class, dst_zspage);
2083 migrate_write_unlock(dst_zspage);
2085 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2086 migrate_write_unlock(src_zspage);
2087 free_zspage(pool, class, src_zspage);
2088 pages_freed += class->pages_per_zspage;
2090 migrate_write_unlock(src_zspage);
2091 spin_unlock(&class->lock);
2092 write_unlock(&pool->migrate_lock);
2094 write_lock(&pool->migrate_lock);
2095 spin_lock(&class->lock);
2099 putback_zspage(class, src_zspage);
2100 migrate_write_unlock(src_zspage);
2103 spin_unlock(&class->lock);
2104 write_unlock(&pool->migrate_lock);
2109 unsigned long zs_compact(struct zs_pool *pool)
2112 struct size_class *class;
2113 unsigned long pages_freed = 0;
2115 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2116 class = pool->size_class[i];
2119 if (class->index != i)
2121 pages_freed += __zs_compact(pool, class);
2123 atomic_long_add(pages_freed, &pool->stats.pages_compacted);
2127 EXPORT_SYMBOL_GPL(zs_compact);
2129 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2131 memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2133 EXPORT_SYMBOL_GPL(zs_pool_stats);
2135 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2136 struct shrink_control *sc)
2138 unsigned long pages_freed;
2139 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2143 * Compact classes and calculate compaction delta.
2144 * Can run concurrently with a manually triggered
2145 * (by user) compaction.
2147 pages_freed = zs_compact(pool);
2149 return pages_freed ? pages_freed : SHRINK_STOP;
2152 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2153 struct shrink_control *sc)
2156 struct size_class *class;
2157 unsigned long pages_to_free = 0;
2158 struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2161 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2162 class = pool->size_class[i];
2165 if (class->index != i)
2168 pages_to_free += zs_can_compact(class);
2171 return pages_to_free;
2174 static void zs_unregister_shrinker(struct zs_pool *pool)
2176 unregister_shrinker(&pool->shrinker);
2179 static int zs_register_shrinker(struct zs_pool *pool)
2181 pool->shrinker.scan_objects = zs_shrinker_scan;
2182 pool->shrinker.count_objects = zs_shrinker_count;
2183 pool->shrinker.batch = 0;
2184 pool->shrinker.seeks = DEFAULT_SEEKS;
2186 return register_shrinker(&pool->shrinker);
2190 * zs_create_pool - Creates an allocation pool to work from.
2191 * @name: pool name to be created
2193 * This function must be called before anything when using
2194 * the zsmalloc allocator.
2196 * On success, a pointer to the newly created pool is returned,
2199 struct zs_pool *zs_create_pool(const char *name)
2202 struct zs_pool *pool;
2203 struct size_class *prev_class = NULL;
2205 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2209 init_deferred_free(pool);
2210 rwlock_init(&pool->migrate_lock);
2212 pool->name = kstrdup(name, GFP_KERNEL);
2216 if (create_cache(pool))
2220 * Iterate reversely, because, size of size_class that we want to use
2221 * for merging should be larger or equal to current size.
2223 for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2225 int pages_per_zspage;
2226 int objs_per_zspage;
2227 struct size_class *class;
2230 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2231 if (size > ZS_MAX_ALLOC_SIZE)
2232 size = ZS_MAX_ALLOC_SIZE;
2233 pages_per_zspage = get_pages_per_zspage(size);
2234 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2237 * We iterate from biggest down to smallest classes,
2238 * so huge_class_size holds the size of the first huge
2239 * class. Any object bigger than or equal to that will
2240 * endup in the huge class.
2242 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2244 huge_class_size = size;
2246 * The object uses ZS_HANDLE_SIZE bytes to store the
2247 * handle. We need to subtract it, because zs_malloc()
2248 * unconditionally adds handle size before it performs
2249 * size class search - so object may be smaller than
2250 * huge class size, yet it still can end up in the huge
2251 * class because it grows by ZS_HANDLE_SIZE extra bytes
2252 * right before class lookup.
2254 huge_class_size -= (ZS_HANDLE_SIZE - 1);
2258 * size_class is used for normal zsmalloc operation such
2259 * as alloc/free for that size. Although it is natural that we
2260 * have one size_class for each size, there is a chance that we
2261 * can get more memory utilization if we use one size_class for
2262 * many different sizes whose size_class have same
2263 * characteristics. So, we makes size_class point to
2264 * previous size_class if possible.
2267 if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2268 pool->size_class[i] = prev_class;
2273 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2279 class->pages_per_zspage = pages_per_zspage;
2280 class->objs_per_zspage = objs_per_zspage;
2281 spin_lock_init(&class->lock);
2282 pool->size_class[i] = class;
2283 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2285 INIT_LIST_HEAD(&class->fullness_list[fullness]);
2290 /* debug only, don't abort if it fails */
2291 zs_pool_stat_create(pool, name);
2293 if (zs_register_migration(pool))
2297 * Not critical since shrinker is only used to trigger internal
2298 * defragmentation of the pool which is pretty optional thing. If
2299 * registration fails we still can use the pool normally and user can
2300 * trigger compaction manually. Thus, ignore return code.
2302 zs_register_shrinker(pool);
2307 zs_destroy_pool(pool);
2310 EXPORT_SYMBOL_GPL(zs_create_pool);
2312 void zs_destroy_pool(struct zs_pool *pool)
2316 zs_unregister_shrinker(pool);
2317 zs_unregister_migration(pool);
2318 zs_pool_stat_destroy(pool);
2320 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2322 struct size_class *class = pool->size_class[i];
2327 if (class->index != i)
2330 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2331 if (!list_empty(&class->fullness_list[fg])) {
2332 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2339 destroy_cache(pool);
2343 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2345 static int __init zs_init(void)
2349 ret = zsmalloc_mount();
2353 ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2354 zs_cpu_prepare, zs_cpu_dead);
2359 zpool_register_driver(&zs_zpool_driver);
2372 static void __exit zs_exit(void)
2375 zpool_unregister_driver(&zs_zpool_driver);
2378 cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2383 module_init(zs_init);
2384 module_exit(zs_exit);
2386 MODULE_LICENSE("Dual BSD/GPL");
2387 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");