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>
31 #include "zsmalloc_int.h"
34 * A zspage's class index and fullness group
35 * are encoded in its (first)page->mapping
37 #define CLASS_IDX_BITS 28
38 #define FULLNESS_BITS 4
39 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
40 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
43 * Object location (<PFN>, <obj_idx>) is encoded as
44 * as single (void *) handle value.
46 * Note that object index <obj_idx> is relative to system
47 * page <PFN> it is stored in, so for each sub-page belonging
48 * to a zspage, obj_idx starts with 0.
50 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
51 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
52 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
54 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
55 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
57 static int is_first_page(struct page *page)
59 return test_bit(PG_private, &page->flags);
62 static int is_last_page(struct page *page)
64 return test_bit(PG_private_2, &page->flags);
67 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
68 enum fullness_group *fullness)
71 BUG_ON(!is_first_page(page));
73 m = (unsigned long)page->mapping;
74 *fullness = m & FULLNESS_MASK;
75 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
78 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
79 enum fullness_group fullness)
82 BUG_ON(!is_first_page(page));
84 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
85 (fullness & FULLNESS_MASK);
86 page->mapping = (struct address_space *)m;
89 static int get_size_class_index(int size)
93 if (likely(size > ZS_MIN_ALLOC_SIZE))
94 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
100 static enum fullness_group get_fullness_group(struct page *page)
102 int inuse, max_objects;
103 enum fullness_group fg;
104 BUG_ON(!is_first_page(page));
107 max_objects = page->objects;
111 else if (inuse == max_objects)
113 else if (inuse <= max_objects / fullness_threshold_frac)
114 fg = ZS_ALMOST_EMPTY;
121 static void insert_zspage(struct page *page, struct size_class *class,
122 enum fullness_group fullness)
126 BUG_ON(!is_first_page(page));
128 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
131 head = &class->fullness_list[fullness];
133 list_add_tail(&page->lru, &(*head)->lru);
138 static void remove_zspage(struct page *page, struct size_class *class,
139 enum fullness_group fullness)
143 BUG_ON(!is_first_page(page));
145 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
148 head = &class->fullness_list[fullness];
150 if (list_empty(&(*head)->lru))
152 else if (*head == page)
153 *head = (struct page *)list_entry((*head)->lru.next,
156 list_del_init(&page->lru);
159 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
163 struct size_class *class;
164 enum fullness_group currfg, newfg;
166 BUG_ON(!is_first_page(page));
168 get_zspage_mapping(page, &class_idx, &currfg);
169 newfg = get_fullness_group(page);
173 class = &pool->size_class[class_idx];
174 remove_zspage(page, class, currfg);
175 insert_zspage(page, class, newfg);
176 set_zspage_mapping(page, class_idx, newfg);
183 * We have to decide on how many pages to link together
184 * to form a zspage for each size class. This is important
185 * to reduce wastage due to unusable space left at end of
186 * each zspage which is given as:
187 * wastage = Zp - Zp % size_class
188 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
190 * For example, for size class of 3/8 * PAGE_SIZE, we should
191 * link together 3 PAGE_SIZE sized pages to form a zspage
192 * since then we can perfectly fit in 8 such objects.
194 static int get_zspage_order(int class_size)
196 int i, max_usedpc = 0;
197 /* zspage order which gives maximum used size per KB */
198 int max_usedpc_order = 1;
200 for (i = 1; i <= max_zspage_order; i++) {
204 zspage_size = i * PAGE_SIZE;
205 waste = zspage_size % class_size;
206 usedpc = (zspage_size - waste) * 100 / zspage_size;
208 if (usedpc > max_usedpc) {
210 max_usedpc_order = i;
214 return max_usedpc_order;
218 * A single 'zspage' is composed of many system pages which are
219 * linked together using fields in struct page. This function finds
220 * the first/head page, given any component page of a zspage.
222 static struct page *get_first_page(struct page *page)
224 if (is_first_page(page))
227 return page->first_page;
230 static struct page *get_next_page(struct page *page)
234 if (is_last_page(page))
236 else if (is_first_page(page))
237 next = (struct page *)page->private;
239 next = list_entry(page->lru.next, struct page, lru);
244 /* Encode <page, obj_idx> as a single handle value */
245 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
247 unsigned long handle;
254 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
255 handle |= (obj_idx & OBJ_INDEX_MASK);
257 return (void *)handle;
260 /* Decode <page, obj_idx> pair from the given object handle */
261 static void obj_handle_to_location(void *handle, struct page **page,
262 unsigned long *obj_idx)
264 unsigned long hval = (unsigned long)handle;
266 *page = pfn_to_page(hval >> OBJ_INDEX_BITS);
267 *obj_idx = hval & OBJ_INDEX_MASK;
270 static unsigned long obj_idx_to_offset(struct page *page,
271 unsigned long obj_idx, int class_size)
273 unsigned long off = 0;
275 if (!is_first_page(page))
278 return off + obj_idx * class_size;
281 static void free_zspage(struct page *first_page)
283 struct page *nextp, *tmp;
285 BUG_ON(!is_first_page(first_page));
286 BUG_ON(first_page->inuse);
288 nextp = (struct page *)page_private(first_page);
290 clear_bit(PG_private, &first_page->flags);
291 clear_bit(PG_private_2, &first_page->flags);
292 set_page_private(first_page, 0);
293 first_page->mapping = NULL;
294 first_page->freelist = NULL;
295 reset_page_mapcount(first_page);
296 __free_page(first_page);
298 /* zspage with only 1 system page */
302 list_for_each_entry_safe(nextp, tmp, &nextp->lru, lru) {
303 list_del(&nextp->lru);
304 clear_bit(PG_private_2, &nextp->flags);
310 /* Initialize a newly allocated zspage */
311 static void init_zspage(struct page *first_page, struct size_class *class)
313 unsigned long off = 0;
314 struct page *page = first_page;
316 BUG_ON(!is_first_page(first_page));
318 struct page *next_page;
319 struct link_free *link;
320 unsigned int i, objs_on_page;
323 * page->index stores offset of first object starting
324 * in the page. For the first page, this is always 0,
325 * so we use first_page->index (aka ->freelist) to store
326 * head of corresponding zspage's freelist.
328 if (page != first_page)
331 link = (struct link_free *)kmap_atomic(page) +
333 objs_on_page = (PAGE_SIZE - off) / class->size;
335 for (i = 1; i <= objs_on_page; i++) {
337 if (off < PAGE_SIZE) {
338 link->next = obj_location_to_handle(page, i);
339 link += class->size / sizeof(*link);
344 * We now come to the last (full or partial) object on this
345 * page, which must point to the first object on the next
348 next_page = get_next_page(page);
349 link->next = obj_location_to_handle(next_page, 0);
352 off = (off + class->size) % PAGE_SIZE;
357 * Allocate a zspage for the given size class
359 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
362 struct page *first_page = NULL;
365 * Allocate individual pages and link them together as:
366 * 1. first page->private = first sub-page
367 * 2. all sub-pages are linked together using page->lru
368 * 3. each sub-page is linked to the first page using page->first_page
370 * For each size class, First/Head pages are linked together using
371 * page->lru. Also, we set PG_private to identify the first page
372 * (i.e. no other sub-page has this flag set) and PG_private_2 to
373 * identify the last page.
376 for (i = 0; i < class->zspage_order; i++) {
377 struct page *page, *prev_page;
379 page = alloc_page(flags);
383 INIT_LIST_HEAD(&page->lru);
384 if (i == 0) { /* first page */
385 set_bit(PG_private, &page->flags);
386 set_page_private(page, 0);
388 first_page->inuse = 0;
391 first_page->private = (unsigned long)page;
393 page->first_page = first_page;
395 list_add(&page->lru, &prev_page->lru);
396 if (i == class->zspage_order - 1) /* last page */
397 set_bit(PG_private_2, &page->flags);
402 init_zspage(first_page, class);
404 first_page->freelist = obj_location_to_handle(first_page, 0);
405 /* Maximum number of objects we can store in this zspage */
406 first_page->objects = class->zspage_order * PAGE_SIZE / class->size;
408 error = 0; /* Success */
411 if (unlikely(error) && first_page) {
412 free_zspage(first_page);
419 static struct page *find_get_zspage(struct size_class *class)
424 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
425 page = class->fullness_list[i];
435 * If this becomes a separate module, register zs_init() with
436 * module_init(), zs_exit with module_exit(), and remove zs_initialized
438 static int zs_initialized;
440 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
443 int cpu = (long)pcpu;
444 struct mapping_area *area;
448 area = &per_cpu(zs_map_area, cpu);
451 area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes);
453 return notifier_from_errno(-ENOMEM);
456 case CPU_UP_CANCELED:
457 area = &per_cpu(zs_map_area, cpu);
459 free_vm_area(area->vm);
467 static struct notifier_block zs_cpu_nb = {
468 .notifier_call = zs_cpu_notifier
471 static void zs_exit(void)
475 for_each_online_cpu(cpu)
476 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
477 unregister_cpu_notifier(&zs_cpu_nb);
480 static int zs_init(void)
484 register_cpu_notifier(&zs_cpu_nb);
485 for_each_online_cpu(cpu) {
486 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
487 if (notifier_to_errno(ret))
493 return notifier_to_errno(ret);
496 struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
498 int i, error, ovhd_size;
499 struct zs_pool *pool;
504 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
505 pool = kzalloc(ovhd_size, GFP_KERNEL);
509 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
511 struct size_class *class;
513 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
514 if (size > ZS_MAX_ALLOC_SIZE)
515 size = ZS_MAX_ALLOC_SIZE;
517 class = &pool->size_class[i];
520 spin_lock_init(&class->lock);
521 class->zspage_order = get_zspage_order(size);
526 * If this becomes a separate module, register zs_init with
527 * module_init, and remove this block
529 if (!zs_initialized) {
539 error = 0; /* Success */
543 zs_destroy_pool(pool);
549 EXPORT_SYMBOL_GPL(zs_create_pool);
551 void zs_destroy_pool(struct zs_pool *pool)
555 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
557 struct size_class *class = &pool->size_class[i];
559 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
560 if (class->fullness_list[fg]) {
561 pr_info("Freeing non-empty class with size "
562 "%db, fullness group %d\n",
569 EXPORT_SYMBOL_GPL(zs_destroy_pool);
572 * zs_malloc - Allocate block of given size from pool.
573 * @pool: pool to allocate from
574 * @size: size of block to allocate
575 * @page: page no. that holds the object
576 * @offset: location of object within page
578 * On success, <page, offset> identifies block allocated
579 * and 0 is returned. On failure, <page, offset> is set to
580 * 0 and -ENOMEM is returned.
582 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
584 void *zs_malloc(struct zs_pool *pool, size_t size)
587 struct link_free *link;
589 struct size_class *class;
591 struct page *first_page, *m_page;
592 unsigned long m_objidx, m_offset;
594 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
597 class_idx = get_size_class_index(size);
598 class = &pool->size_class[class_idx];
599 BUG_ON(class_idx != class->index);
601 spin_lock(&class->lock);
602 first_page = find_get_zspage(class);
605 spin_unlock(&class->lock);
606 first_page = alloc_zspage(class, pool->flags);
607 if (unlikely(!first_page))
610 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
611 spin_lock(&class->lock);
612 class->pages_allocated += class->zspage_order;
615 obj = first_page->freelist;
616 obj_handle_to_location(obj, &m_page, &m_objidx);
617 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
619 link = (struct link_free *)kmap_atomic(m_page) +
620 m_offset / sizeof(*link);
621 first_page->freelist = link->next;
622 memset(link, POISON_INUSE, sizeof(*link));
626 /* Now move the zspage to another fullness group, if required */
627 fix_fullness_group(pool, first_page);
628 spin_unlock(&class->lock);
632 EXPORT_SYMBOL_GPL(zs_malloc);
634 void zs_free(struct zs_pool *pool, void *obj)
636 struct link_free *link;
637 struct page *first_page, *f_page;
638 unsigned long f_objidx, f_offset;
641 struct size_class *class;
642 enum fullness_group fullness;
647 obj_handle_to_location(obj, &f_page, &f_objidx);
648 first_page = get_first_page(f_page);
650 get_zspage_mapping(first_page, &class_idx, &fullness);
651 class = &pool->size_class[class_idx];
652 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
654 spin_lock(&class->lock);
656 /* Insert this object in containing zspage's freelist */
657 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
659 link->next = first_page->freelist;
661 first_page->freelist = obj;
664 fullness = fix_fullness_group(pool, first_page);
666 if (fullness == ZS_EMPTY)
667 class->pages_allocated -= class->zspage_order;
669 spin_unlock(&class->lock);
671 if (fullness == ZS_EMPTY)
672 free_zspage(first_page);
674 EXPORT_SYMBOL_GPL(zs_free);
676 void *zs_map_object(struct zs_pool *pool, void *handle)
679 unsigned long obj_idx, off;
681 unsigned int class_idx;
682 enum fullness_group fg;
683 struct size_class *class;
684 struct mapping_area *area;
688 obj_handle_to_location(handle, &page, &obj_idx);
689 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
690 class = &pool->size_class[class_idx];
691 off = obj_idx_to_offset(page, obj_idx, class->size);
693 area = &get_cpu_var(zs_map_area);
694 if (off + class->size <= PAGE_SIZE) {
695 /* this object is contained entirely within a page */
696 area->vm_addr = kmap_atomic(page);
698 /* this object spans two pages */
701 nextp = get_next_page(page);
705 set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL));
706 set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL));
708 /* We pre-allocated VM area so mapping can never fail */
709 area->vm_addr = area->vm->addr;
712 return area->vm_addr + off;
714 EXPORT_SYMBOL_GPL(zs_map_object);
716 void zs_unmap_object(struct zs_pool *pool, void *handle)
719 unsigned long obj_idx, off;
721 unsigned int class_idx;
722 enum fullness_group fg;
723 struct size_class *class;
724 struct mapping_area *area;
728 obj_handle_to_location(handle, &page, &obj_idx);
729 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
730 class = &pool->size_class[class_idx];
731 off = obj_idx_to_offset(page, obj_idx, class->size);
733 area = &__get_cpu_var(zs_map_area);
734 if (off + class->size <= PAGE_SIZE) {
735 kunmap_atomic(area->vm_addr);
737 set_pte(area->vm_ptes[0], __pte(0));
738 set_pte(area->vm_ptes[1], __pte(0));
739 __flush_tlb_one((unsigned long)area->vm_addr);
740 __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE);
742 put_cpu_var(zs_map_area);
744 EXPORT_SYMBOL_GPL(zs_unmap_object);
746 u64 zs_get_total_size_bytes(struct zs_pool *pool)
751 for (i = 0; i < ZS_SIZE_CLASSES; i++)
752 npages += pool->size_class[i].pages_allocated;
754 return npages << PAGE_SHIFT;
756 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);