2 * zsmalloc memory allocator
4 * Copyright (C) 2011 Nitin Gupta
6 * This code is released using a dual license strategy: BSD/GPL
7 * You can choose the license that better fits your requirements.
9 * Released under the terms of 3-clause BSD License
10 * Released under the terms of GNU General Public License Version 2.0
13 #ifdef CONFIG_ZSMALLOC_DEBUG
17 #include <linux/module.h>
18 #include <linux/kernel.h>
19 #include <linux/bitops.h>
20 #include <linux/errno.h>
21 #include <linux/highmem.h>
22 #include <linux/init.h>
23 #include <linux/string.h>
24 #include <linux/slab.h>
25 #include <asm/tlbflush.h>
26 #include <asm/pgtable.h>
27 #include <linux/cpumask.h>
28 #include <linux/cpu.h>
29 #include <linux/vmalloc.h>
32 #include "zsmalloc_int.h"
35 * A zspage's class index and fullness group
36 * are encoded in its (first)page->mapping
38 #define CLASS_IDX_BITS 28
39 #define FULLNESS_BITS 4
40 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
41 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
43 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
44 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
46 static int is_first_page(struct page *page)
48 return test_bit(PG_private, &page->flags);
51 static int is_last_page(struct page *page)
53 return test_bit(PG_private_2, &page->flags);
56 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
57 enum fullness_group *fullness)
60 BUG_ON(!is_first_page(page));
62 m = (unsigned long)page->mapping;
63 *fullness = m & FULLNESS_MASK;
64 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
67 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
68 enum fullness_group fullness)
71 BUG_ON(!is_first_page(page));
73 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
74 (fullness & FULLNESS_MASK);
75 page->mapping = (struct address_space *)m;
78 static int get_size_class_index(int size)
82 if (likely(size > ZS_MIN_ALLOC_SIZE))
83 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
89 static enum fullness_group get_fullness_group(struct page *page)
91 int inuse, max_objects;
92 enum fullness_group fg;
93 BUG_ON(!is_first_page(page));
96 max_objects = page->objects;
100 else if (inuse == max_objects)
102 else if (inuse <= max_objects / fullness_threshold_frac)
103 fg = ZS_ALMOST_EMPTY;
110 static void insert_zspage(struct page *page, struct size_class *class,
111 enum fullness_group fullness)
115 BUG_ON(!is_first_page(page));
117 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
120 head = &class->fullness_list[fullness];
122 list_add_tail(&page->lru, &(*head)->lru);
127 static void remove_zspage(struct page *page, struct size_class *class,
128 enum fullness_group fullness)
132 BUG_ON(!is_first_page(page));
134 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
137 head = &class->fullness_list[fullness];
139 if (list_empty(&(*head)->lru))
141 else if (*head == page)
142 *head = (struct page *)list_entry((*head)->lru.next,
145 list_del_init(&page->lru);
148 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
152 struct size_class *class;
153 enum fullness_group currfg, newfg;
155 BUG_ON(!is_first_page(page));
157 get_zspage_mapping(page, &class_idx, &currfg);
158 newfg = get_fullness_group(page);
162 class = &pool->size_class[class_idx];
163 remove_zspage(page, class, currfg);
164 insert_zspage(page, class, newfg);
165 set_zspage_mapping(page, class_idx, newfg);
172 * We have to decide on how many pages to link together
173 * to form a zspage for each size class. This is important
174 * to reduce wastage due to unusable space left at end of
175 * each zspage which is given as:
176 * wastage = Zp - Zp % size_class
177 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
179 * For example, for size class of 3/8 * PAGE_SIZE, we should
180 * link together 3 PAGE_SIZE sized pages to form a zspage
181 * since then we can perfectly fit in 8 such objects.
183 static int get_zspage_order(int class_size)
185 int i, max_usedpc = 0;
186 /* zspage order which gives maximum used size per KB */
187 int max_usedpc_order = 1;
189 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
193 zspage_size = i * PAGE_SIZE;
194 waste = zspage_size % class_size;
195 usedpc = (zspage_size - waste) * 100 / zspage_size;
197 if (usedpc > max_usedpc) {
199 max_usedpc_order = i;
203 return max_usedpc_order;
207 * A single 'zspage' is composed of many system pages which are
208 * linked together using fields in struct page. This function finds
209 * the first/head page, given any component page of a zspage.
211 static struct page *get_first_page(struct page *page)
213 if (is_first_page(page))
216 return page->first_page;
219 static struct page *get_next_page(struct page *page)
223 if (is_last_page(page))
225 else if (is_first_page(page))
226 next = (struct page *)page->private;
228 next = list_entry(page->lru.next, struct page, lru);
233 /* Encode <page, obj_idx> as a single handle value */
234 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
236 unsigned long handle;
243 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
244 handle |= (obj_idx & OBJ_INDEX_MASK);
246 return (void *)handle;
249 /* Decode <page, obj_idx> pair from the given object handle */
250 static void obj_handle_to_location(void *handle, struct page **page,
251 unsigned long *obj_idx)
253 unsigned long hval = (unsigned long)handle;
255 *page = pfn_to_page(hval >> OBJ_INDEX_BITS);
256 *obj_idx = hval & OBJ_INDEX_MASK;
259 static unsigned long obj_idx_to_offset(struct page *page,
260 unsigned long obj_idx, int class_size)
262 unsigned long off = 0;
264 if (!is_first_page(page))
267 return off + obj_idx * class_size;
270 static void reset_page(struct page *page)
272 clear_bit(PG_private, &page->flags);
273 clear_bit(PG_private_2, &page->flags);
274 set_page_private(page, 0);
275 page->mapping = NULL;
276 page->freelist = NULL;
277 reset_page_mapcount(page);
280 static void free_zspage(struct page *first_page)
282 struct page *nextp, *tmp, *head_extra;
284 BUG_ON(!is_first_page(first_page));
285 BUG_ON(first_page->inuse);
287 head_extra = (struct page *)page_private(first_page);
289 reset_page(first_page);
290 __free_page(first_page);
292 /* zspage with only 1 system page */
296 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
297 list_del(&nextp->lru);
301 reset_page(head_extra);
302 __free_page(head_extra);
305 /* Initialize a newly allocated zspage */
306 static void init_zspage(struct page *first_page, struct size_class *class)
308 unsigned long off = 0;
309 struct page *page = first_page;
311 BUG_ON(!is_first_page(first_page));
313 struct page *next_page;
314 struct link_free *link;
315 unsigned int i, objs_on_page;
318 * page->index stores offset of first object starting
319 * in the page. For the first page, this is always 0,
320 * so we use first_page->index (aka ->freelist) to store
321 * head of corresponding zspage's freelist.
323 if (page != first_page)
326 link = (struct link_free *)kmap_atomic(page) +
328 objs_on_page = (PAGE_SIZE - off) / class->size;
330 for (i = 1; i <= objs_on_page; i++) {
332 if (off < PAGE_SIZE) {
333 link->next = obj_location_to_handle(page, i);
334 link += class->size / sizeof(*link);
339 * We now come to the last (full or partial) object on this
340 * page, which must point to the first object on the next
343 next_page = get_next_page(page);
344 link->next = obj_location_to_handle(next_page, 0);
347 off = (off + class->size) % PAGE_SIZE;
352 * Allocate a zspage for the given size class
354 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
357 struct page *first_page = NULL;
360 * Allocate individual pages and link them together as:
361 * 1. first page->private = first sub-page
362 * 2. all sub-pages are linked together using page->lru
363 * 3. each sub-page is linked to the first page using page->first_page
365 * For each size class, First/Head pages are linked together using
366 * page->lru. Also, we set PG_private to identify the first page
367 * (i.e. no other sub-page has this flag set) and PG_private_2 to
368 * identify the last page.
371 for (i = 0; i < class->zspage_order; i++) {
372 struct page *page, *prev_page;
374 page = alloc_page(flags);
378 INIT_LIST_HEAD(&page->lru);
379 if (i == 0) { /* first page */
380 set_bit(PG_private, &page->flags);
381 set_page_private(page, 0);
383 first_page->inuse = 0;
386 first_page->private = (unsigned long)page;
388 page->first_page = first_page;
390 list_add(&page->lru, &prev_page->lru);
391 if (i == class->zspage_order - 1) /* last page */
392 set_bit(PG_private_2, &page->flags);
397 init_zspage(first_page, class);
399 first_page->freelist = obj_location_to_handle(first_page, 0);
400 /* Maximum number of objects we can store in this zspage */
401 first_page->objects = class->zspage_order * PAGE_SIZE / class->size;
403 error = 0; /* Success */
406 if (unlikely(error) && first_page) {
407 free_zspage(first_page);
414 static struct page *find_get_zspage(struct size_class *class)
419 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
420 page = class->fullness_list[i];
430 * If this becomes a separate module, register zs_init() with
431 * module_init(), zs_exit with module_exit(), and remove zs_initialized
433 static int zs_initialized;
435 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
438 int cpu = (long)pcpu;
439 struct mapping_area *area;
443 area = &per_cpu(zs_map_area, cpu);
446 area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes);
448 return notifier_from_errno(-ENOMEM);
451 case CPU_UP_CANCELED:
452 area = &per_cpu(zs_map_area, cpu);
454 free_vm_area(area->vm);
462 static struct notifier_block zs_cpu_nb = {
463 .notifier_call = zs_cpu_notifier
466 static void zs_exit(void)
470 for_each_online_cpu(cpu)
471 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
472 unregister_cpu_notifier(&zs_cpu_nb);
475 static int zs_init(void)
479 register_cpu_notifier(&zs_cpu_nb);
480 for_each_online_cpu(cpu) {
481 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
482 if (notifier_to_errno(ret))
488 return notifier_to_errno(ret);
491 struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
493 int i, error, ovhd_size;
494 struct zs_pool *pool;
499 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
500 pool = kzalloc(ovhd_size, GFP_KERNEL);
504 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
506 struct size_class *class;
508 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
509 if (size > ZS_MAX_ALLOC_SIZE)
510 size = ZS_MAX_ALLOC_SIZE;
512 class = &pool->size_class[i];
515 spin_lock_init(&class->lock);
516 class->zspage_order = get_zspage_order(size);
521 * If this becomes a separate module, register zs_init with
522 * module_init, and remove this block
524 if (!zs_initialized) {
534 error = 0; /* Success */
538 zs_destroy_pool(pool);
544 EXPORT_SYMBOL_GPL(zs_create_pool);
546 void zs_destroy_pool(struct zs_pool *pool)
550 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
552 struct size_class *class = &pool->size_class[i];
554 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
555 if (class->fullness_list[fg]) {
556 pr_info("Freeing non-empty class with size "
557 "%db, fullness group %d\n",
564 EXPORT_SYMBOL_GPL(zs_destroy_pool);
567 * zs_malloc - Allocate block of given size from pool.
568 * @pool: pool to allocate from
569 * @size: size of block to allocate
570 * @page: page no. that holds the object
571 * @offset: location of object within page
573 * On success, <page, offset> identifies block allocated
574 * and 0 is returned. On failure, <page, offset> is set to
575 * 0 and -ENOMEM is returned.
577 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
579 void *zs_malloc(struct zs_pool *pool, size_t size)
582 struct link_free *link;
584 struct size_class *class;
586 struct page *first_page, *m_page;
587 unsigned long m_objidx, m_offset;
589 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
592 class_idx = get_size_class_index(size);
593 class = &pool->size_class[class_idx];
594 BUG_ON(class_idx != class->index);
596 spin_lock(&class->lock);
597 first_page = find_get_zspage(class);
600 spin_unlock(&class->lock);
601 first_page = alloc_zspage(class, pool->flags);
602 if (unlikely(!first_page))
605 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
606 spin_lock(&class->lock);
607 class->pages_allocated += class->zspage_order;
610 obj = first_page->freelist;
611 obj_handle_to_location(obj, &m_page, &m_objidx);
612 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
614 link = (struct link_free *)kmap_atomic(m_page) +
615 m_offset / sizeof(*link);
616 first_page->freelist = link->next;
617 memset(link, POISON_INUSE, sizeof(*link));
621 /* Now move the zspage to another fullness group, if required */
622 fix_fullness_group(pool, first_page);
623 spin_unlock(&class->lock);
627 EXPORT_SYMBOL_GPL(zs_malloc);
629 void zs_free(struct zs_pool *pool, void *obj)
631 struct link_free *link;
632 struct page *first_page, *f_page;
633 unsigned long f_objidx, f_offset;
636 struct size_class *class;
637 enum fullness_group fullness;
642 obj_handle_to_location(obj, &f_page, &f_objidx);
643 first_page = get_first_page(f_page);
645 get_zspage_mapping(first_page, &class_idx, &fullness);
646 class = &pool->size_class[class_idx];
647 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
649 spin_lock(&class->lock);
651 /* Insert this object in containing zspage's freelist */
652 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
654 link->next = first_page->freelist;
656 first_page->freelist = obj;
659 fullness = fix_fullness_group(pool, first_page);
661 if (fullness == ZS_EMPTY)
662 class->pages_allocated -= class->zspage_order;
664 spin_unlock(&class->lock);
666 if (fullness == ZS_EMPTY)
667 free_zspage(first_page);
669 EXPORT_SYMBOL_GPL(zs_free);
671 void *zs_map_object(struct zs_pool *pool, void *handle)
674 unsigned long obj_idx, off;
676 unsigned int class_idx;
677 enum fullness_group fg;
678 struct size_class *class;
679 struct mapping_area *area;
683 obj_handle_to_location(handle, &page, &obj_idx);
684 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
685 class = &pool->size_class[class_idx];
686 off = obj_idx_to_offset(page, obj_idx, class->size);
688 area = &get_cpu_var(zs_map_area);
689 if (off + class->size <= PAGE_SIZE) {
690 /* this object is contained entirely within a page */
691 area->vm_addr = kmap_atomic(page);
693 /* this object spans two pages */
696 nextp = get_next_page(page);
700 set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL));
701 set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL));
703 /* We pre-allocated VM area so mapping can never fail */
704 area->vm_addr = area->vm->addr;
707 return area->vm_addr + off;
709 EXPORT_SYMBOL_GPL(zs_map_object);
711 void zs_unmap_object(struct zs_pool *pool, void *handle)
714 unsigned long obj_idx, off;
716 unsigned int class_idx;
717 enum fullness_group fg;
718 struct size_class *class;
719 struct mapping_area *area;
723 obj_handle_to_location(handle, &page, &obj_idx);
724 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
725 class = &pool->size_class[class_idx];
726 off = obj_idx_to_offset(page, obj_idx, class->size);
728 area = &__get_cpu_var(zs_map_area);
729 if (off + class->size <= PAGE_SIZE) {
730 kunmap_atomic(area->vm_addr);
732 set_pte(area->vm_ptes[0], __pte(0));
733 set_pte(area->vm_ptes[1], __pte(0));
734 __flush_tlb_one((unsigned long)area->vm_addr);
735 __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE);
737 put_cpu_var(zs_map_area);
739 EXPORT_SYMBOL_GPL(zs_unmap_object);
741 u64 zs_get_total_size_bytes(struct zs_pool *pool)
746 for (i = 0; i < ZS_SIZE_CLASSES; i++)
747 npages += pool->size_class[i].pages_allocated;
749 return npages << PAGE_SHIFT;
751 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);