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)
44 * Object location (<PFN>, <obj_idx>) is encoded as
45 * as single (void *) handle value.
47 * Note that object index <obj_idx> is relative to system
48 * page <PFN> it is stored in, so for each sub-page belonging
49 * to a zspage, obj_idx starts with 0.
51 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
52 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS)
53 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
55 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
56 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
58 static int is_first_page(struct page *page)
60 return test_bit(PG_private, &page->flags);
63 static int is_last_page(struct page *page)
65 return test_bit(PG_private_2, &page->flags);
68 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
69 enum fullness_group *fullness)
72 BUG_ON(!is_first_page(page));
74 m = (unsigned long)page->mapping;
75 *fullness = m & FULLNESS_MASK;
76 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
79 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
80 enum fullness_group fullness)
83 BUG_ON(!is_first_page(page));
85 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
86 (fullness & FULLNESS_MASK);
87 page->mapping = (struct address_space *)m;
90 static int get_size_class_index(int size)
94 if (likely(size > ZS_MIN_ALLOC_SIZE))
95 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
101 static enum fullness_group get_fullness_group(struct page *page)
103 int inuse, max_objects;
104 enum fullness_group fg;
105 BUG_ON(!is_first_page(page));
108 max_objects = page->objects;
112 else if (inuse == max_objects)
114 else if (inuse <= max_objects / fullness_threshold_frac)
115 fg = ZS_ALMOST_EMPTY;
122 static void insert_zspage(struct page *page, struct size_class *class,
123 enum fullness_group fullness)
127 BUG_ON(!is_first_page(page));
129 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
132 head = &class->fullness_list[fullness];
134 list_add_tail(&page->lru, &(*head)->lru);
139 static void remove_zspage(struct page *page, struct size_class *class,
140 enum fullness_group fullness)
144 BUG_ON(!is_first_page(page));
146 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
149 head = &class->fullness_list[fullness];
151 if (list_empty(&(*head)->lru))
153 else if (*head == page)
154 *head = (struct page *)list_entry((*head)->lru.next,
157 list_del_init(&page->lru);
160 static enum fullness_group fix_fullness_group(struct zs_pool *pool,
164 struct size_class *class;
165 enum fullness_group currfg, newfg;
167 BUG_ON(!is_first_page(page));
169 get_zspage_mapping(page, &class_idx, &currfg);
170 newfg = get_fullness_group(page);
174 class = &pool->size_class[class_idx];
175 remove_zspage(page, class, currfg);
176 insert_zspage(page, class, newfg);
177 set_zspage_mapping(page, class_idx, newfg);
184 * We have to decide on how many pages to link together
185 * to form a zspage for each size class. This is important
186 * to reduce wastage due to unusable space left at end of
187 * each zspage which is given as:
188 * wastage = Zp - Zp % size_class
189 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
191 * For example, for size class of 3/8 * PAGE_SIZE, we should
192 * link together 3 PAGE_SIZE sized pages to form a zspage
193 * since then we can perfectly fit in 8 such objects.
195 static int get_zspage_order(int class_size)
197 int i, max_usedpc = 0;
198 /* zspage order which gives maximum used size per KB */
199 int max_usedpc_order = 1;
201 for (i = 1; i <= max_zspage_order; i++) {
205 zspage_size = i * PAGE_SIZE;
206 waste = zspage_size % class_size;
207 usedpc = (zspage_size - waste) * 100 / zspage_size;
209 if (usedpc > max_usedpc) {
211 max_usedpc_order = i;
215 return max_usedpc_order;
219 * A single 'zspage' is composed of many system pages which are
220 * linked together using fields in struct page. This function finds
221 * the first/head page, given any component page of a zspage.
223 static struct page *get_first_page(struct page *page)
225 if (is_first_page(page))
228 return page->first_page;
231 static struct page *get_next_page(struct page *page)
235 if (is_last_page(page))
237 else if (is_first_page(page))
238 next = (struct page *)page->private;
240 next = list_entry(page->lru.next, struct page, lru);
245 /* Encode <page, obj_idx> as a single handle value */
246 static void *obj_location_to_handle(struct page *page, unsigned long obj_idx)
248 unsigned long handle;
255 handle = page_to_pfn(page) << OBJ_INDEX_BITS;
256 handle |= (obj_idx & OBJ_INDEX_MASK);
258 return (void *)handle;
261 /* Decode <page, obj_idx> pair from the given object handle */
262 static void obj_handle_to_location(void *handle, struct page **page,
263 unsigned long *obj_idx)
265 unsigned long hval = (unsigned long)handle;
267 *page = pfn_to_page(hval >> OBJ_INDEX_BITS);
268 *obj_idx = hval & OBJ_INDEX_MASK;
271 static unsigned long obj_idx_to_offset(struct page *page,
272 unsigned long obj_idx, int class_size)
274 unsigned long off = 0;
276 if (!is_first_page(page))
279 return off + obj_idx * class_size;
282 static void free_zspage(struct page *first_page)
284 struct page *nextp, *tmp;
286 BUG_ON(!is_first_page(first_page));
287 BUG_ON(first_page->inuse);
289 nextp = (struct page *)page_private(first_page);
291 clear_bit(PG_private, &first_page->flags);
292 clear_bit(PG_private_2, &first_page->flags);
293 set_page_private(first_page, 0);
294 first_page->mapping = NULL;
295 first_page->freelist = NULL;
296 reset_page_mapcount(first_page);
297 __free_page(first_page);
299 /* zspage with only 1 system page */
303 list_for_each_entry_safe(nextp, tmp, &nextp->lru, lru) {
304 list_del(&nextp->lru);
305 clear_bit(PG_private_2, &nextp->flags);
311 /* Initialize a newly allocated zspage */
312 static void init_zspage(struct page *first_page, struct size_class *class)
314 unsigned long off = 0;
315 struct page *page = first_page;
317 BUG_ON(!is_first_page(first_page));
319 struct page *next_page;
320 struct link_free *link;
321 unsigned int i, objs_on_page;
324 * page->index stores offset of first object starting
325 * in the page. For the first page, this is always 0,
326 * so we use first_page->index (aka ->freelist) to store
327 * head of corresponding zspage's freelist.
329 if (page != first_page)
332 link = (struct link_free *)kmap_atomic(page) +
334 objs_on_page = (PAGE_SIZE - off) / class->size;
336 for (i = 1; i <= objs_on_page; i++) {
338 if (off < PAGE_SIZE) {
339 link->next = obj_location_to_handle(page, i);
340 link += class->size / sizeof(*link);
345 * We now come to the last (full or partial) object on this
346 * page, which must point to the first object on the next
349 next_page = get_next_page(page);
350 link->next = obj_location_to_handle(next_page, 0);
353 off = (off + class->size) % PAGE_SIZE;
358 * Allocate a zspage for the given size class
360 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
363 struct page *first_page = NULL;
366 * Allocate individual pages and link them together as:
367 * 1. first page->private = first sub-page
368 * 2. all sub-pages are linked together using page->lru
369 * 3. each sub-page is linked to the first page using page->first_page
371 * For each size class, First/Head pages are linked together using
372 * page->lru. Also, we set PG_private to identify the first page
373 * (i.e. no other sub-page has this flag set) and PG_private_2 to
374 * identify the last page.
377 for (i = 0; i < class->zspage_order; i++) {
378 struct page *page, *prev_page;
380 page = alloc_page(flags);
384 INIT_LIST_HEAD(&page->lru);
385 if (i == 0) { /* first page */
386 set_bit(PG_private, &page->flags);
387 set_page_private(page, 0);
389 first_page->inuse = 0;
392 first_page->private = (unsigned long)page;
394 page->first_page = first_page;
396 list_add(&page->lru, &prev_page->lru);
397 if (i == class->zspage_order - 1) /* last page */
398 set_bit(PG_private_2, &page->flags);
403 init_zspage(first_page, class);
405 first_page->freelist = obj_location_to_handle(first_page, 0);
406 /* Maximum number of objects we can store in this zspage */
407 first_page->objects = class->zspage_order * PAGE_SIZE / class->size;
409 error = 0; /* Success */
412 if (unlikely(error) && first_page) {
413 free_zspage(first_page);
420 static struct page *find_get_zspage(struct size_class *class)
425 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
426 page = class->fullness_list[i];
436 * If this becomes a separate module, register zs_init() with
437 * module_init(), zs_exit with module_exit(), and remove zs_initialized
439 static int zs_initialized;
441 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
444 int cpu = (long)pcpu;
445 struct mapping_area *area;
449 area = &per_cpu(zs_map_area, cpu);
452 area->vm = alloc_vm_area(2 * PAGE_SIZE, area->vm_ptes);
454 return notifier_from_errno(-ENOMEM);
457 case CPU_UP_CANCELED:
458 area = &per_cpu(zs_map_area, cpu);
460 free_vm_area(area->vm);
468 static struct notifier_block zs_cpu_nb = {
469 .notifier_call = zs_cpu_notifier
472 static void zs_exit(void)
476 for_each_online_cpu(cpu)
477 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
478 unregister_cpu_notifier(&zs_cpu_nb);
481 static int zs_init(void)
485 register_cpu_notifier(&zs_cpu_nb);
486 for_each_online_cpu(cpu) {
487 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
488 if (notifier_to_errno(ret))
494 return notifier_to_errno(ret);
497 struct zs_pool *zs_create_pool(const char *name, gfp_t flags)
499 int i, error, ovhd_size;
500 struct zs_pool *pool;
505 ovhd_size = roundup(sizeof(*pool), PAGE_SIZE);
506 pool = kzalloc(ovhd_size, GFP_KERNEL);
510 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
512 struct size_class *class;
514 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
515 if (size > ZS_MAX_ALLOC_SIZE)
516 size = ZS_MAX_ALLOC_SIZE;
518 class = &pool->size_class[i];
521 spin_lock_init(&class->lock);
522 class->zspage_order = get_zspage_order(size);
527 * If this becomes a separate module, register zs_init with
528 * module_init, and remove this block
530 if (!zs_initialized) {
540 error = 0; /* Success */
544 zs_destroy_pool(pool);
550 EXPORT_SYMBOL_GPL(zs_create_pool);
552 void zs_destroy_pool(struct zs_pool *pool)
556 for (i = 0; i < ZS_SIZE_CLASSES; i++) {
558 struct size_class *class = &pool->size_class[i];
560 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
561 if (class->fullness_list[fg]) {
562 pr_info("Freeing non-empty class with size "
563 "%db, fullness group %d\n",
570 EXPORT_SYMBOL_GPL(zs_destroy_pool);
573 * zs_malloc - Allocate block of given size from pool.
574 * @pool: pool to allocate from
575 * @size: size of block to allocate
576 * @page: page no. that holds the object
577 * @offset: location of object within page
579 * On success, <page, offset> identifies block allocated
580 * and 0 is returned. On failure, <page, offset> is set to
581 * 0 and -ENOMEM is returned.
583 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
585 void *zs_malloc(struct zs_pool *pool, size_t size)
588 struct link_free *link;
590 struct size_class *class;
592 struct page *first_page, *m_page;
593 unsigned long m_objidx, m_offset;
595 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
598 class_idx = get_size_class_index(size);
599 class = &pool->size_class[class_idx];
600 BUG_ON(class_idx != class->index);
602 spin_lock(&class->lock);
603 first_page = find_get_zspage(class);
606 spin_unlock(&class->lock);
607 first_page = alloc_zspage(class, pool->flags);
608 if (unlikely(!first_page))
611 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
612 spin_lock(&class->lock);
613 class->pages_allocated += class->zspage_order;
616 obj = first_page->freelist;
617 obj_handle_to_location(obj, &m_page, &m_objidx);
618 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
620 link = (struct link_free *)kmap_atomic(m_page) +
621 m_offset / sizeof(*link);
622 first_page->freelist = link->next;
623 memset(link, POISON_INUSE, sizeof(*link));
627 /* Now move the zspage to another fullness group, if required */
628 fix_fullness_group(pool, first_page);
629 spin_unlock(&class->lock);
633 EXPORT_SYMBOL_GPL(zs_malloc);
635 void zs_free(struct zs_pool *pool, void *obj)
637 struct link_free *link;
638 struct page *first_page, *f_page;
639 unsigned long f_objidx, f_offset;
642 struct size_class *class;
643 enum fullness_group fullness;
648 obj_handle_to_location(obj, &f_page, &f_objidx);
649 first_page = get_first_page(f_page);
651 get_zspage_mapping(first_page, &class_idx, &fullness);
652 class = &pool->size_class[class_idx];
653 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
655 spin_lock(&class->lock);
657 /* Insert this object in containing zspage's freelist */
658 link = (struct link_free *)((unsigned char *)kmap_atomic(f_page)
660 link->next = first_page->freelist;
662 first_page->freelist = obj;
665 fullness = fix_fullness_group(pool, first_page);
667 if (fullness == ZS_EMPTY)
668 class->pages_allocated -= class->zspage_order;
670 spin_unlock(&class->lock);
672 if (fullness == ZS_EMPTY)
673 free_zspage(first_page);
675 EXPORT_SYMBOL_GPL(zs_free);
677 void *zs_map_object(struct zs_pool *pool, void *handle)
680 unsigned long obj_idx, off;
682 unsigned int class_idx;
683 enum fullness_group fg;
684 struct size_class *class;
685 struct mapping_area *area;
689 obj_handle_to_location(handle, &page, &obj_idx);
690 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
691 class = &pool->size_class[class_idx];
692 off = obj_idx_to_offset(page, obj_idx, class->size);
694 area = &get_cpu_var(zs_map_area);
695 if (off + class->size <= PAGE_SIZE) {
696 /* this object is contained entirely within a page */
697 area->vm_addr = kmap_atomic(page);
699 /* this object spans two pages */
702 nextp = get_next_page(page);
706 set_pte(area->vm_ptes[0], mk_pte(page, PAGE_KERNEL));
707 set_pte(area->vm_ptes[1], mk_pte(nextp, PAGE_KERNEL));
709 /* We pre-allocated VM area so mapping can never fail */
710 area->vm_addr = area->vm->addr;
713 return area->vm_addr + off;
715 EXPORT_SYMBOL_GPL(zs_map_object);
717 void zs_unmap_object(struct zs_pool *pool, void *handle)
720 unsigned long obj_idx, off;
722 unsigned int class_idx;
723 enum fullness_group fg;
724 struct size_class *class;
725 struct mapping_area *area;
729 obj_handle_to_location(handle, &page, &obj_idx);
730 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
731 class = &pool->size_class[class_idx];
732 off = obj_idx_to_offset(page, obj_idx, class->size);
734 area = &__get_cpu_var(zs_map_area);
735 if (off + class->size <= PAGE_SIZE) {
736 kunmap_atomic(area->vm_addr);
738 set_pte(area->vm_ptes[0], __pte(0));
739 set_pte(area->vm_ptes[1], __pte(0));
740 __flush_tlb_one((unsigned long)area->vm_addr);
741 __flush_tlb_one((unsigned long)area->vm_addr + PAGE_SIZE);
743 put_cpu_var(zs_map_area);
745 EXPORT_SYMBOL_GPL(zs_unmap_object);
747 u64 zs_get_total_size_bytes(struct zs_pool *pool)
752 for (i = 0; i < ZS_SIZE_CLASSES; i++)
753 npages += pool->size_class[i].pages_allocated;
755 return npages << PAGE_SHIFT;
757 EXPORT_SYMBOL_GPL(zs_get_total_size_bytes);