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 * This allocator is designed for use with zram. Thus, the allocator is
16 * supposed to work well under low memory conditions. In particular, it
17 * never attempts higher order page allocation which is very likely to
18 * fail under memory pressure. On the other hand, if we just use single
19 * (0-order) pages, it would suffer from very high fragmentation --
20 * any object of size PAGE_SIZE/2 or larger would occupy an entire page.
21 * This was one of the major issues with its predecessor (xvmalloc).
23 * To overcome these issues, zsmalloc allocates a bunch of 0-order pages
24 * and links them together using various 'struct page' fields. These linked
25 * pages act as a single higher-order page i.e. an object can span 0-order
26 * page boundaries. The code refers to these linked pages as a single entity
29 * For simplicity, zsmalloc can only allocate objects of size up to PAGE_SIZE
30 * since this satisfies the requirements of all its current users (in the
31 * worst case, page is incompressible and is thus stored "as-is" i.e. in
32 * uncompressed form). For allocation requests larger than this size, failure
33 * is returned (see zs_malloc).
35 * Additionally, zs_malloc() does not return a dereferenceable pointer.
36 * Instead, it returns an opaque handle (unsigned long) which encodes actual
37 * location of the allocated object. The reason for this indirection is that
38 * zsmalloc does not keep zspages permanently mapped since that would cause
39 * issues on 32-bit systems where the VA region for kernel space mappings
40 * is very small. So, before using the allocating memory, the object has to
41 * be mapped using zs_map_object() to get a usable pointer and subsequently
42 * unmapped using zs_unmap_object().
44 * Following is how we use various fields and flags of underlying
45 * struct page(s) to form a zspage.
47 * Usage of struct page fields:
48 * page->first_page: points to the first component (0-order) page
49 * page->index (union with page->freelist): offset of the first object
50 * starting in this page. For the first page, this is
51 * always 0, so we use this field (aka freelist) to point
52 * to the first free object in zspage.
53 * page->lru: links together all component pages (except the first page)
56 * For _first_ page only:
58 * page->private (union with page->first_page): refers to the
59 * component page after the first page
60 * page->freelist: points to the first free object in zspage.
61 * Free objects are linked together using in-place
63 * page->objects: maximum number of objects we can store in this
64 * zspage (class->zspage_order * PAGE_SIZE / class->size)
65 * page->lru: links together first pages of various zspages.
66 * Basically forming list of zspages in a fullness group.
67 * page->mapping: class index and fullness group of the zspage
69 * Usage of struct page flags:
70 * PG_private: identifies the first component page
71 * PG_private2: identifies the last component page
75 #ifdef CONFIG_ZSMALLOC_DEBUG
79 #include <linux/module.h>
80 #include <linux/kernel.h>
81 #include <linux/bitops.h>
82 #include <linux/errno.h>
83 #include <linux/highmem.h>
84 #include <linux/string.h>
85 #include <linux/slab.h>
86 #include <asm/tlbflush.h>
87 #include <asm/pgtable.h>
88 #include <linux/cpumask.h>
89 #include <linux/cpu.h>
90 #include <linux/vmalloc.h>
91 #include <linux/hardirq.h>
92 #include <linux/spinlock.h>
93 #include <linux/types.h>
94 #include <linux/zsmalloc.h>
95 #include <linux/zpool.h>
98 * This must be power of 2 and greater than of equal to sizeof(link_free).
99 * These two conditions ensure that any 'struct link_free' itself doesn't
100 * span more than 1 page which avoids complex case of mapping 2 pages simply
101 * to restore link_free pointer values.
106 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
107 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
109 #define ZS_MAX_ZSPAGE_ORDER 2
110 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
113 * Object location (<PFN>, <obj_idx>) is encoded as
114 * as single (unsigned long) handle value.
116 * Note that object index <obj_idx> is relative to system
117 * page <PFN> it is stored in, so for each sub-page belonging
118 * to a zspage, obj_idx starts with 0.
120 * This is made more complicated by various memory models and PAE.
123 #ifndef MAX_PHYSMEM_BITS
124 #ifdef CONFIG_HIGHMEM64G
125 #define MAX_PHYSMEM_BITS 36
126 #else /* !CONFIG_HIGHMEM64G */
128 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
131 #define MAX_PHYSMEM_BITS BITS_PER_LONG
134 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
135 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
136 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
138 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
139 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
140 #define ZS_MIN_ALLOC_SIZE \
141 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
142 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
145 * On systems with 4K page size, this gives 255 size classes! There is a
147 * - Large number of size classes is potentially wasteful as free page are
148 * spread across these classes
149 * - Small number of size classes causes large internal fragmentation
150 * - Probably its better to use specific size classes (empirically
151 * determined). NOTE: all those class sizes must be set as multiple of
152 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
154 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
157 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
158 #define ZS_SIZE_CLASSES ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / \
159 ZS_SIZE_CLASS_DELTA + 1)
162 * We do not maintain any list for completely empty or full pages
164 enum fullness_group {
167 _ZS_NR_FULLNESS_GROUPS,
174 * number of size_classes
176 static int zs_size_classes;
179 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
181 * n = number of allocated objects
182 * N = total number of objects zspage can store
183 * f = fullness_threshold_frac
185 * Similarly, we assign zspage to:
186 * ZS_ALMOST_FULL when n > N / f
187 * ZS_EMPTY when n == 0
188 * ZS_FULL when n == N
190 * (see: fix_fullness_group())
192 static const int fullness_threshold_frac = 4;
196 * Size of objects stored in this class. Must be multiple
202 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
203 int pages_per_zspage;
207 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
211 * Placed within free objects to form a singly linked list.
212 * For every zspage, first_page->freelist gives head of this list.
214 * This must be power of 2 and less than or equal to ZS_ALIGN
217 /* Handle of next free chunk (encodes <PFN, obj_idx>) */
222 struct size_class **size_class;
224 gfp_t flags; /* allocation flags used when growing pool */
225 atomic_long_t pages_allocated;
229 * A zspage's class index and fullness group
230 * are encoded in its (first)page->mapping
232 #define CLASS_IDX_BITS 28
233 #define FULLNESS_BITS 4
234 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
235 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
237 struct mapping_area {
238 #ifdef CONFIG_PGTABLE_MAPPING
239 struct vm_struct *vm; /* vm area for mapping object that span pages */
241 char *vm_buf; /* copy buffer for objects that span pages */
243 char *vm_addr; /* address of kmap_atomic()'ed pages */
244 enum zs_mapmode vm_mm; /* mapping mode */
251 static void *zs_zpool_create(gfp_t gfp, struct zpool_ops *zpool_ops)
253 return zs_create_pool(gfp);
256 static void zs_zpool_destroy(void *pool)
258 zs_destroy_pool(pool);
261 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
262 unsigned long *handle)
264 *handle = zs_malloc(pool, size);
265 return *handle ? 0 : -1;
267 static void zs_zpool_free(void *pool, unsigned long handle)
269 zs_free(pool, handle);
272 static int zs_zpool_shrink(void *pool, unsigned int pages,
273 unsigned int *reclaimed)
278 static void *zs_zpool_map(void *pool, unsigned long handle,
279 enum zpool_mapmode mm)
281 enum zs_mapmode zs_mm;
290 case ZPOOL_MM_RW: /* fallthru */
296 return zs_map_object(pool, handle, zs_mm);
298 static void zs_zpool_unmap(void *pool, unsigned long handle)
300 zs_unmap_object(pool, handle);
303 static u64 zs_zpool_total_size(void *pool)
305 return zs_get_total_pages(pool) << PAGE_SHIFT;
308 static struct zpool_driver zs_zpool_driver = {
310 .owner = THIS_MODULE,
311 .create = zs_zpool_create,
312 .destroy = zs_zpool_destroy,
313 .malloc = zs_zpool_malloc,
314 .free = zs_zpool_free,
315 .shrink = zs_zpool_shrink,
317 .unmap = zs_zpool_unmap,
318 .total_size = zs_zpool_total_size,
321 MODULE_ALIAS("zpool-zsmalloc");
322 #endif /* CONFIG_ZPOOL */
324 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
325 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
327 static int is_first_page(struct page *page)
329 return PagePrivate(page);
332 static int is_last_page(struct page *page)
334 return PagePrivate2(page);
337 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
338 enum fullness_group *fullness)
341 BUG_ON(!is_first_page(page));
343 m = (unsigned long)page->mapping;
344 *fullness = m & FULLNESS_MASK;
345 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
348 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
349 enum fullness_group fullness)
352 BUG_ON(!is_first_page(page));
354 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
355 (fullness & FULLNESS_MASK);
356 page->mapping = (struct address_space *)m;
360 * zsmalloc divides the pool into various size classes where each
361 * class maintains a list of zspages where each zspage is divided
362 * into equal sized chunks. Each allocation falls into one of these
363 * classes depending on its size. This function returns index of the
364 * size class which has chunk size big enough to hold the give size.
366 static int get_size_class_index(int size)
370 if (likely(size > ZS_MIN_ALLOC_SIZE))
371 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
372 ZS_SIZE_CLASS_DELTA);
378 * For each size class, zspages are divided into different groups
379 * depending on how "full" they are. This was done so that we could
380 * easily find empty or nearly empty zspages when we try to shrink
381 * the pool (not yet implemented). This function returns fullness
382 * status of the given page.
384 static enum fullness_group get_fullness_group(struct page *page)
386 int inuse, max_objects;
387 enum fullness_group fg;
388 BUG_ON(!is_first_page(page));
391 max_objects = page->objects;
395 else if (inuse == max_objects)
397 else if (inuse <= max_objects / fullness_threshold_frac)
398 fg = ZS_ALMOST_EMPTY;
406 * Each size class maintains various freelists and zspages are assigned
407 * to one of these freelists based on the number of live objects they
408 * have. This functions inserts the given zspage into the freelist
409 * identified by <class, fullness_group>.
411 static void insert_zspage(struct page *page, struct size_class *class,
412 enum fullness_group fullness)
416 BUG_ON(!is_first_page(page));
418 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
421 head = &class->fullness_list[fullness];
423 list_add_tail(&page->lru, &(*head)->lru);
429 * This function removes the given zspage from the freelist identified
430 * by <class, fullness_group>.
432 static void remove_zspage(struct page *page, struct size_class *class,
433 enum fullness_group fullness)
437 BUG_ON(!is_first_page(page));
439 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
442 head = &class->fullness_list[fullness];
444 if (list_empty(&(*head)->lru))
446 else if (*head == page)
447 *head = (struct page *)list_entry((*head)->lru.next,
450 list_del_init(&page->lru);
454 * Each size class maintains zspages in different fullness groups depending
455 * on the number of live objects they contain. When allocating or freeing
456 * objects, the fullness status of the page can change, say, from ALMOST_FULL
457 * to ALMOST_EMPTY when freeing an object. This function checks if such
458 * a status change has occurred for the given page and accordingly moves the
459 * page from the freelist of the old fullness group to that of the new
462 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
466 struct size_class *class;
467 enum fullness_group currfg, newfg;
469 BUG_ON(!is_first_page(page));
471 get_zspage_mapping(page, &class_idx, &currfg);
472 newfg = get_fullness_group(page);
476 class = pool->size_class[class_idx];
477 remove_zspage(page, class, currfg);
478 insert_zspage(page, class, newfg);
479 set_zspage_mapping(page, class_idx, newfg);
486 * We have to decide on how many pages to link together
487 * to form a zspage for each size class. This is important
488 * to reduce wastage due to unusable space left at end of
489 * each zspage which is given as:
490 * wastage = Zp - Zp % size_class
491 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
493 * For example, for size class of 3/8 * PAGE_SIZE, we should
494 * link together 3 PAGE_SIZE sized pages to form a zspage
495 * since then we can perfectly fit in 8 such objects.
497 static int get_pages_per_zspage(int class_size)
499 int i, max_usedpc = 0;
500 /* zspage order which gives maximum used size per KB */
501 int max_usedpc_order = 1;
503 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
507 zspage_size = i * PAGE_SIZE;
508 waste = zspage_size % class_size;
509 usedpc = (zspage_size - waste) * 100 / zspage_size;
511 if (usedpc > max_usedpc) {
513 max_usedpc_order = i;
517 return max_usedpc_order;
521 * A single 'zspage' is composed of many system pages which are
522 * linked together using fields in struct page. This function finds
523 * the first/head page, given any component page of a zspage.
525 static struct page *get_first_page(struct page *page)
527 if (is_first_page(page))
530 return page->first_page;
533 static struct page *get_next_page(struct page *page)
537 if (is_last_page(page))
539 else if (is_first_page(page))
540 next = (struct page *)page_private(page);
542 next = list_entry(page->lru.next, struct page, lru);
548 * Encode <page, obj_idx> as a single handle value.
549 * On hardware platforms with physical memory starting at 0x0 the pfn
550 * could be 0 so we ensure that the handle will never be 0 by adjusting the
551 * encoded obj_idx value before encoding.
553 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
555 unsigned long handle;
562 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
563 handle |= ((obj_idx + 1) & OBJ_INDEX_MASK);
565 return (void *)handle;
569 * Decode <page, obj_idx> pair from the given object handle. We adjust the
570 * decoded obj_idx back to its original value since it was adjusted in
571 * obj_location_to_handle().
573 static void obj_handle_to_location(unsigned long handle, struct page **page,
574 unsigned long *obj_idx)
576 *page = pfn_to_page(handle >> OBJ_INDEX_BITS);
577 *obj_idx = (handle & OBJ_INDEX_MASK) - 1;
580 static unsigned long obj_idx_to_offset(struct page *page,
581 unsigned long obj_idx, int class_size)
583 unsigned long off = 0;
585 if (!is_first_page(page))
588 return off + obj_idx * class_size;
591 static void reset_page(struct page *page)
593 clear_bit(PG_private, &page->flags);
594 clear_bit(PG_private_2, &page->flags);
595 set_page_private(page, 0);
596 page->mapping = NULL;
597 page->freelist = NULL;
598 page_mapcount_reset(page);
601 static void free_zspage(struct page *first_page)
603 struct page *nextp, *tmp, *head_extra;
605 BUG_ON(!is_first_page(first_page));
606 BUG_ON(first_page->inuse);
608 head_extra = (struct page *)page_private(first_page);
610 reset_page(first_page);
611 __free_page(first_page);
613 /* zspage with only 1 system page */
617 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
618 list_del(&nextp->lru);
622 reset_page(head_extra);
623 __free_page(head_extra);
626 /* Initialize a newly allocated zspage */
627 static void init_zspage(struct page *first_page, struct size_class *class)
629 unsigned long off = 0;
630 struct page *page = first_page;
632 BUG_ON(!is_first_page(first_page));
634 struct page *next_page;
635 struct link_free *link;
640 * page->index stores offset of first object starting
641 * in the page. For the first page, this is always 0,
642 * so we use first_page->index (aka ->freelist) to store
643 * head of corresponding zspage's freelist.
645 if (page != first_page)
648 vaddr = kmap_atomic(page);
649 link = (struct link_free *)vaddr + off / sizeof(*link);
651 while ((off += class->size) < PAGE_SIZE) {
652 link->next = obj_location_to_handle(page, i++);
653 link += class->size / sizeof(*link);
657 * We now come to the last (full or partial) object on this
658 * page, which must point to the first object on the next
661 next_page = get_next_page(page);
662 link->next = obj_location_to_handle(next_page, 0);
663 kunmap_atomic(vaddr);
670 * Allocate a zspage for the given size class
672 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
675 struct page *first_page = NULL, *uninitialized_var(prev_page);
678 * Allocate individual pages and link them together as:
679 * 1. first page->private = first sub-page
680 * 2. all sub-pages are linked together using page->lru
681 * 3. each sub-page is linked to the first page using page->first_page
683 * For each size class, First/Head pages are linked together using
684 * page->lru. Also, we set PG_private to identify the first page
685 * (i.e. no other sub-page has this flag set) and PG_private_2 to
686 * identify the last page.
689 for (i = 0; i < class->pages_per_zspage; i++) {
692 page = alloc_page(flags);
696 INIT_LIST_HEAD(&page->lru);
697 if (i == 0) { /* first page */
698 SetPagePrivate(page);
699 set_page_private(page, 0);
701 first_page->inuse = 0;
704 set_page_private(first_page, (unsigned long)page);
706 page->first_page = first_page;
708 list_add(&page->lru, &prev_page->lru);
709 if (i == class->pages_per_zspage - 1) /* last page */
710 SetPagePrivate2(page);
714 init_zspage(first_page, class);
716 first_page->freelist = obj_location_to_handle(first_page, 0);
717 /* Maximum number of objects we can store in this zspage */
718 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
720 error = 0; /* Success */
723 if (unlikely(error) && first_page) {
724 free_zspage(first_page);
731 static struct page *find_get_zspage(struct size_class *class)
736 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
737 page = class->fullness_list[i];
745 #ifdef CONFIG_PGTABLE_MAPPING
746 static inline int __zs_cpu_up(struct mapping_area *area)
749 * Make sure we don't leak memory if a cpu UP notification
750 * and zs_init() race and both call zs_cpu_up() on the same cpu
754 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
760 static inline void __zs_cpu_down(struct mapping_area *area)
763 free_vm_area(area->vm);
767 static inline void *__zs_map_object(struct mapping_area *area,
768 struct page *pages[2], int off, int size)
770 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
771 area->vm_addr = area->vm->addr;
772 return area->vm_addr + off;
775 static inline void __zs_unmap_object(struct mapping_area *area,
776 struct page *pages[2], int off, int size)
778 unsigned long addr = (unsigned long)area->vm_addr;
780 unmap_kernel_range(addr, PAGE_SIZE * 2);
783 #else /* CONFIG_PGTABLE_MAPPING */
785 static inline int __zs_cpu_up(struct mapping_area *area)
788 * Make sure we don't leak memory if a cpu UP notification
789 * and zs_init() race and both call zs_cpu_up() on the same cpu
793 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
799 static inline void __zs_cpu_down(struct mapping_area *area)
805 static void *__zs_map_object(struct mapping_area *area,
806 struct page *pages[2], int off, int size)
810 char *buf = area->vm_buf;
812 /* disable page faults to match kmap_atomic() return conditions */
815 /* no read fastpath */
816 if (area->vm_mm == ZS_MM_WO)
819 sizes[0] = PAGE_SIZE - off;
820 sizes[1] = size - sizes[0];
822 /* copy object to per-cpu buffer */
823 addr = kmap_atomic(pages[0]);
824 memcpy(buf, addr + off, sizes[0]);
826 addr = kmap_atomic(pages[1]);
827 memcpy(buf + sizes[0], addr, sizes[1]);
833 static void __zs_unmap_object(struct mapping_area *area,
834 struct page *pages[2], int off, int size)
838 char *buf = area->vm_buf;
840 /* no write fastpath */
841 if (area->vm_mm == ZS_MM_RO)
844 sizes[0] = PAGE_SIZE - off;
845 sizes[1] = size - sizes[0];
847 /* copy per-cpu buffer to object */
848 addr = kmap_atomic(pages[0]);
849 memcpy(addr + off, buf, sizes[0]);
851 addr = kmap_atomic(pages[1]);
852 memcpy(addr, buf + sizes[0], sizes[1]);
856 /* enable page faults to match kunmap_atomic() return conditions */
860 #endif /* CONFIG_PGTABLE_MAPPING */
862 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
865 int ret, cpu = (long)pcpu;
866 struct mapping_area *area;
870 area = &per_cpu(zs_map_area, cpu);
871 ret = __zs_cpu_up(area);
873 return notifier_from_errno(ret);
876 case CPU_UP_CANCELED:
877 area = &per_cpu(zs_map_area, cpu);
885 static struct notifier_block zs_cpu_nb = {
886 .notifier_call = zs_cpu_notifier
889 static void zs_unregister_cpu_notifier(void)
893 cpu_notifier_register_begin();
895 for_each_online_cpu(cpu)
896 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
897 __unregister_cpu_notifier(&zs_cpu_nb);
899 cpu_notifier_register_done();
902 static int zs_register_cpu_notifier(void)
904 int cpu, uninitialized_var(ret);
906 cpu_notifier_register_begin();
908 __register_cpu_notifier(&zs_cpu_nb);
909 for_each_online_cpu(cpu) {
910 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
911 if (notifier_to_errno(ret))
915 cpu_notifier_register_done();
916 return notifier_to_errno(ret);
919 static void init_zs_size_classes(void)
923 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
924 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
927 zs_size_classes = nr;
930 static void __exit zs_exit(void)
933 zpool_unregister_driver(&zs_zpool_driver);
935 zs_unregister_cpu_notifier();
938 static int __init zs_init(void)
940 int ret = zs_register_cpu_notifier();
943 zs_unregister_cpu_notifier();
947 init_zs_size_classes();
950 zpool_register_driver(&zs_zpool_driver);
955 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
957 return pages_per_zspage * PAGE_SIZE / size;
960 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
962 if (prev->pages_per_zspage != pages_per_zspage)
965 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
966 != get_maxobj_per_zspage(size, pages_per_zspage))
973 * zs_create_pool - Creates an allocation pool to work from.
974 * @flags: allocation flags used to allocate pool metadata
976 * This function must be called before anything when using
977 * the zsmalloc allocator.
979 * On success, a pointer to the newly created pool is returned,
982 struct zs_pool *zs_create_pool(gfp_t flags)
985 struct zs_pool *pool;
987 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
988 pool = kzalloc(ovhd_size, GFP_KERNEL);
992 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
994 if (!pool->size_class) {
1000 * Iterate reversly, because, size of size_class that we want to use
1001 * for merging should be larger or equal to current size.
1003 for (i = zs_size_classes - 1; i >= 0; i--) {
1005 int pages_per_zspage;
1006 struct size_class *class;
1007 struct size_class *prev_class;
1009 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1010 if (size > ZS_MAX_ALLOC_SIZE)
1011 size = ZS_MAX_ALLOC_SIZE;
1012 pages_per_zspage = get_pages_per_zspage(size);
1015 * size_class is used for normal zsmalloc operation such
1016 * as alloc/free for that size. Although it is natural that we
1017 * have one size_class for each size, there is a chance that we
1018 * can get more memory utilization if we use one size_class for
1019 * many different sizes whose size_class have same
1020 * characteristics. So, we makes size_class point to
1021 * previous size_class if possible.
1023 if (i < ZS_SIZE_CLASSES - 1) {
1024 prev_class = pool->size_class[i + 1];
1025 if (can_merge(prev_class, size, pages_per_zspage)) {
1026 pool->size_class[i] = prev_class;
1031 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1037 class->pages_per_zspage = pages_per_zspage;
1038 spin_lock_init(&class->lock);
1039 pool->size_class[i] = class;
1042 pool->flags = flags;
1047 zs_destroy_pool(pool);
1050 EXPORT_SYMBOL_GPL(zs_create_pool);
1052 void zs_destroy_pool(struct zs_pool *pool)
1056 for (i = 0; i < zs_size_classes; i++) {
1058 struct size_class *class = pool->size_class[i];
1063 if (class->index != i)
1066 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1067 if (class->fullness_list[fg]) {
1068 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1075 kfree(pool->size_class);
1078 EXPORT_SYMBOL_GPL(zs_destroy_pool);
1081 * zs_malloc - Allocate block of given size from pool.
1082 * @pool: pool to allocate from
1083 * @size: size of block to allocate
1085 * On success, handle to the allocated object is returned,
1087 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1089 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1092 struct link_free *link;
1093 struct size_class *class;
1096 struct page *first_page, *m_page;
1097 unsigned long m_objidx, m_offset;
1099 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1102 class = pool->size_class[get_size_class_index(size)];
1104 spin_lock(&class->lock);
1105 first_page = find_get_zspage(class);
1108 spin_unlock(&class->lock);
1109 first_page = alloc_zspage(class, pool->flags);
1110 if (unlikely(!first_page))
1113 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1114 atomic_long_add(class->pages_per_zspage,
1115 &pool->pages_allocated);
1116 spin_lock(&class->lock);
1119 obj = (unsigned long)first_page->freelist;
1120 obj_handle_to_location(obj, &m_page, &m_objidx);
1121 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1123 vaddr = kmap_atomic(m_page);
1124 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1125 first_page->freelist = link->next;
1126 memset(link, POISON_INUSE, sizeof(*link));
1127 kunmap_atomic(vaddr);
1129 first_page->inuse++;
1130 /* Now move the zspage to another fullness group, if required */
1131 fix_fullness_group(pool, first_page);
1132 spin_unlock(&class->lock);
1136 EXPORT_SYMBOL_GPL(zs_malloc);
1138 void zs_free(struct zs_pool *pool, unsigned long obj)
1140 struct link_free *link;
1141 struct page *first_page, *f_page;
1142 unsigned long f_objidx, f_offset;
1146 struct size_class *class;
1147 enum fullness_group fullness;
1152 obj_handle_to_location(obj, &f_page, &f_objidx);
1153 first_page = get_first_page(f_page);
1155 get_zspage_mapping(first_page, &class_idx, &fullness);
1156 class = pool->size_class[class_idx];
1157 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1159 spin_lock(&class->lock);
1161 /* Insert this object in containing zspage's freelist */
1162 vaddr = kmap_atomic(f_page);
1163 link = (struct link_free *)(vaddr + f_offset);
1164 link->next = first_page->freelist;
1165 kunmap_atomic(vaddr);
1166 first_page->freelist = (void *)obj;
1168 first_page->inuse--;
1169 fullness = fix_fullness_group(pool, first_page);
1170 spin_unlock(&class->lock);
1172 if (fullness == ZS_EMPTY) {
1173 atomic_long_sub(class->pages_per_zspage,
1174 &pool->pages_allocated);
1175 free_zspage(first_page);
1178 EXPORT_SYMBOL_GPL(zs_free);
1181 * zs_map_object - get address of allocated object from handle.
1182 * @pool: pool from which the object was allocated
1183 * @handle: handle returned from zs_malloc
1185 * Before using an object allocated from zs_malloc, it must be mapped using
1186 * this function. When done with the object, it must be unmapped using
1189 * Only one object can be mapped per cpu at a time. There is no protection
1190 * against nested mappings.
1192 * This function returns with preemption and page faults disabled.
1194 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1198 unsigned long obj_idx, off;
1200 unsigned int class_idx;
1201 enum fullness_group fg;
1202 struct size_class *class;
1203 struct mapping_area *area;
1204 struct page *pages[2];
1209 * Because we use per-cpu mapping areas shared among the
1210 * pools/users, we can't allow mapping in interrupt context
1211 * because it can corrupt another users mappings.
1213 BUG_ON(in_interrupt());
1215 obj_handle_to_location(handle, &page, &obj_idx);
1216 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1217 class = pool->size_class[class_idx];
1218 off = obj_idx_to_offset(page, obj_idx, class->size);
1220 area = &get_cpu_var(zs_map_area);
1222 if (off + class->size <= PAGE_SIZE) {
1223 /* this object is contained entirely within a page */
1224 area->vm_addr = kmap_atomic(page);
1225 return area->vm_addr + off;
1228 /* this object spans two pages */
1230 pages[1] = get_next_page(page);
1233 return __zs_map_object(area, pages, off, class->size);
1235 EXPORT_SYMBOL_GPL(zs_map_object);
1237 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1240 unsigned long obj_idx, off;
1242 unsigned int class_idx;
1243 enum fullness_group fg;
1244 struct size_class *class;
1245 struct mapping_area *area;
1249 obj_handle_to_location(handle, &page, &obj_idx);
1250 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1251 class = pool->size_class[class_idx];
1252 off = obj_idx_to_offset(page, obj_idx, class->size);
1254 area = this_cpu_ptr(&zs_map_area);
1255 if (off + class->size <= PAGE_SIZE)
1256 kunmap_atomic(area->vm_addr);
1258 struct page *pages[2];
1261 pages[1] = get_next_page(page);
1264 __zs_unmap_object(area, pages, off, class->size);
1266 put_cpu_var(zs_map_area);
1268 EXPORT_SYMBOL_GPL(zs_unmap_object);
1270 unsigned long zs_get_total_pages(struct zs_pool *pool)
1272 return atomic_long_read(&pool->pages_allocated);
1274 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1276 module_init(zs_init);
1277 module_exit(zs_exit);
1279 MODULE_LICENSE("Dual BSD/GPL");
1280 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");