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->first_page: points to the first component (0-order) page
20 * page->index (union with page->freelist): offset of the first object
21 * starting in this page. For the first page, this is
22 * always 0, so we use this field (aka freelist) to point
23 * to the first free object in zspage.
24 * page->lru: links together all component pages (except the first page)
27 * For _first_ page only:
29 * page->private (union with page->first_page): refers to the
30 * component page after the first page
31 * If the page is first_page for huge object, it stores handle.
32 * Look at size_class->huge.
33 * page->freelist: points to the first free object in zspage.
34 * Free objects are linked together using in-place
36 * page->objects: maximum number of objects we can store in this
37 * zspage (class->zspage_order * PAGE_SIZE / class->size)
38 * page->lru: links together first pages of various zspages.
39 * Basically forming list of zspages in a fullness group.
40 * page->mapping: class index and fullness group of the zspage
42 * Usage of struct page flags:
43 * PG_private: identifies the first component page
44 * PG_private2: identifies the last component page
48 #ifdef CONFIG_ZSMALLOC_DEBUG
52 #include <linux/module.h>
53 #include <linux/kernel.h>
54 #include <linux/sched.h>
55 #include <linux/bitops.h>
56 #include <linux/errno.h>
57 #include <linux/highmem.h>
58 #include <linux/string.h>
59 #include <linux/slab.h>
60 #include <asm/tlbflush.h>
61 #include <asm/pgtable.h>
62 #include <linux/cpumask.h>
63 #include <linux/cpu.h>
64 #include <linux/vmalloc.h>
65 #include <linux/hardirq.h>
66 #include <linux/spinlock.h>
67 #include <linux/types.h>
68 #include <linux/debugfs.h>
69 #include <linux/zsmalloc.h>
70 #include <linux/zpool.h>
73 * This must be power of 2 and greater than of equal to sizeof(link_free).
74 * These two conditions ensure that any 'struct link_free' itself doesn't
75 * span more than 1 page which avoids complex case of mapping 2 pages simply
76 * to restore link_free pointer values.
81 * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
82 * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
84 #define ZS_MAX_ZSPAGE_ORDER 2
85 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
87 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
90 * Object location (<PFN>, <obj_idx>) is encoded as
91 * as single (unsigned long) handle value.
93 * Note that object index <obj_idx> is relative to system
94 * page <PFN> it is stored in, so for each sub-page belonging
95 * to a zspage, obj_idx starts with 0.
97 * This is made more complicated by various memory models and PAE.
100 #ifndef MAX_PHYSMEM_BITS
101 #ifdef CONFIG_HIGHMEM64G
102 #define MAX_PHYSMEM_BITS 36
103 #else /* !CONFIG_HIGHMEM64G */
105 * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
108 #define MAX_PHYSMEM_BITS BITS_PER_LONG
111 #define _PFN_BITS (MAX_PHYSMEM_BITS - PAGE_SHIFT)
114 * Memory for allocating for handle keeps object position by
115 * encoding <page, obj_idx> and the encoded value has a room
116 * in least bit(ie, look at obj_to_location).
117 * We use the bit to synchronize between object access by
118 * user and migration.
120 #define HANDLE_PIN_BIT 0
123 * Head in allocated object should have OBJ_ALLOCATED_TAG
124 * to identify the object was allocated or not.
125 * It's okay to add the status bit in the least bit because
126 * header keeps handle which is 4byte-aligned address so we
127 * have room for two bit at least.
129 #define OBJ_ALLOCATED_TAG 1
130 #define OBJ_TAG_BITS 1
131 #define OBJ_INDEX_BITS (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
132 #define OBJ_INDEX_MASK ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
134 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
135 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
136 #define ZS_MIN_ALLOC_SIZE \
137 MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
138 /* each chunk includes extra space to keep handle */
139 #define ZS_MAX_ALLOC_SIZE PAGE_SIZE
142 * On systems with 4K page size, this gives 255 size classes! There is a
144 * - Large number of size classes is potentially wasteful as free page are
145 * spread across these classes
146 * - Small number of size classes causes large internal fragmentation
147 * - Probably its better to use specific size classes (empirically
148 * determined). NOTE: all those class sizes must be set as multiple of
149 * ZS_ALIGN to make sure link_free itself never has to span 2 pages.
151 * ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
154 #define ZS_SIZE_CLASS_DELTA (PAGE_SIZE >> 8)
157 * We do not maintain any list for completely empty or full pages
159 enum fullness_group {
162 _ZS_NR_FULLNESS_GROUPS,
176 #ifdef CONFIG_ZSMALLOC_STAT
178 static struct dentry *zs_stat_root;
180 struct zs_size_stat {
181 unsigned long objs[NR_ZS_STAT_TYPE];
187 * number of size_classes
189 static int zs_size_classes;
192 * We assign a page to ZS_ALMOST_EMPTY fullness group when:
194 * n = number of allocated objects
195 * N = total number of objects zspage can store
196 * f = fullness_threshold_frac
198 * Similarly, we assign zspage to:
199 * ZS_ALMOST_FULL when n > N / f
200 * ZS_EMPTY when n == 0
201 * ZS_FULL when n == N
203 * (see: fix_fullness_group())
205 static const int fullness_threshold_frac = 4;
209 * Size of objects stored in this class. Must be multiple
215 /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
216 int pages_per_zspage;
217 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
220 #ifdef CONFIG_ZSMALLOC_STAT
221 struct zs_size_stat stats;
226 struct page *fullness_list[_ZS_NR_FULLNESS_GROUPS];
230 * Placed within free objects to form a singly linked list.
231 * For every zspage, first_page->freelist gives head of this list.
233 * This must be power of 2 and less than or equal to ZS_ALIGN
238 * Position of next free chunk (encodes <PFN, obj_idx>)
239 * It's valid for non-allocated object
243 * Handle of allocated object.
245 unsigned long handle;
252 struct size_class **size_class;
253 struct kmem_cache *handle_cachep;
255 gfp_t flags; /* allocation flags used when growing pool */
256 atomic_long_t pages_allocated;
258 #ifdef CONFIG_ZSMALLOC_STAT
259 struct dentry *stat_dentry;
264 * A zspage's class index and fullness group
265 * are encoded in its (first)page->mapping
267 #define CLASS_IDX_BITS 28
268 #define FULLNESS_BITS 4
269 #define CLASS_IDX_MASK ((1 << CLASS_IDX_BITS) - 1)
270 #define FULLNESS_MASK ((1 << FULLNESS_BITS) - 1)
272 struct mapping_area {
273 #ifdef CONFIG_PGTABLE_MAPPING
274 struct vm_struct *vm; /* vm area for mapping object that span pages */
276 char *vm_buf; /* copy buffer for objects that span pages */
278 char *vm_addr; /* address of kmap_atomic()'ed pages */
279 enum zs_mapmode vm_mm; /* mapping mode */
283 static int create_handle_cache(struct zs_pool *pool)
285 pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
287 return pool->handle_cachep ? 0 : 1;
290 static void destroy_handle_cache(struct zs_pool *pool)
292 kmem_cache_destroy(pool->handle_cachep);
295 static unsigned long alloc_handle(struct zs_pool *pool)
297 return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
298 pool->flags & ~__GFP_HIGHMEM);
301 static void free_handle(struct zs_pool *pool, unsigned long handle)
303 kmem_cache_free(pool->handle_cachep, (void *)handle);
306 static void record_obj(unsigned long handle, unsigned long obj)
308 *(unsigned long *)handle = obj;
315 static void *zs_zpool_create(char *name, gfp_t gfp, struct zpool_ops *zpool_ops)
317 return zs_create_pool(name, gfp);
320 static void zs_zpool_destroy(void *pool)
322 zs_destroy_pool(pool);
325 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
326 unsigned long *handle)
328 *handle = zs_malloc(pool, size);
329 return *handle ? 0 : -1;
331 static void zs_zpool_free(void *pool, unsigned long handle)
333 zs_free(pool, handle);
336 static int zs_zpool_shrink(void *pool, unsigned int pages,
337 unsigned int *reclaimed)
342 static void *zs_zpool_map(void *pool, unsigned long handle,
343 enum zpool_mapmode mm)
345 enum zs_mapmode zs_mm;
354 case ZPOOL_MM_RW: /* fallthru */
360 return zs_map_object(pool, handle, zs_mm);
362 static void zs_zpool_unmap(void *pool, unsigned long handle)
364 zs_unmap_object(pool, handle);
367 static u64 zs_zpool_total_size(void *pool)
369 return zs_get_total_pages(pool) << PAGE_SHIFT;
372 static struct zpool_driver zs_zpool_driver = {
374 .owner = THIS_MODULE,
375 .create = zs_zpool_create,
376 .destroy = zs_zpool_destroy,
377 .malloc = zs_zpool_malloc,
378 .free = zs_zpool_free,
379 .shrink = zs_zpool_shrink,
381 .unmap = zs_zpool_unmap,
382 .total_size = zs_zpool_total_size,
385 MODULE_ALIAS("zpool-zsmalloc");
386 #endif /* CONFIG_ZPOOL */
388 static unsigned int get_maxobj_per_zspage(int size, int pages_per_zspage)
390 return pages_per_zspage * PAGE_SIZE / size;
393 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
394 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
396 static int is_first_page(struct page *page)
398 return PagePrivate(page);
401 static int is_last_page(struct page *page)
403 return PagePrivate2(page);
406 static void get_zspage_mapping(struct page *page, unsigned int *class_idx,
407 enum fullness_group *fullness)
410 BUG_ON(!is_first_page(page));
412 m = (unsigned long)page->mapping;
413 *fullness = m & FULLNESS_MASK;
414 *class_idx = (m >> FULLNESS_BITS) & CLASS_IDX_MASK;
417 static void set_zspage_mapping(struct page *page, unsigned int class_idx,
418 enum fullness_group fullness)
421 BUG_ON(!is_first_page(page));
423 m = ((class_idx & CLASS_IDX_MASK) << FULLNESS_BITS) |
424 (fullness & FULLNESS_MASK);
425 page->mapping = (struct address_space *)m;
429 * zsmalloc divides the pool into various size classes where each
430 * class maintains a list of zspages where each zspage is divided
431 * into equal sized chunks. Each allocation falls into one of these
432 * classes depending on its size. This function returns index of the
433 * size class which has chunk size big enough to hold the give size.
435 static int get_size_class_index(int size)
439 if (likely(size > ZS_MIN_ALLOC_SIZE))
440 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
441 ZS_SIZE_CLASS_DELTA);
443 return min(zs_size_classes - 1, idx);
446 #ifdef CONFIG_ZSMALLOC_STAT
448 static inline void zs_stat_inc(struct size_class *class,
449 enum zs_stat_type type, unsigned long cnt)
451 class->stats.objs[type] += cnt;
454 static inline void zs_stat_dec(struct size_class *class,
455 enum zs_stat_type type, unsigned long cnt)
457 class->stats.objs[type] -= cnt;
460 static inline unsigned long zs_stat_get(struct size_class *class,
461 enum zs_stat_type type)
463 return class->stats.objs[type];
466 static int __init zs_stat_init(void)
468 if (!debugfs_initialized())
471 zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
478 static void __exit zs_stat_exit(void)
480 debugfs_remove_recursive(zs_stat_root);
483 static int zs_stats_size_show(struct seq_file *s, void *v)
486 struct zs_pool *pool = s->private;
487 struct size_class *class;
489 unsigned long class_almost_full, class_almost_empty;
490 unsigned long obj_allocated, obj_used, pages_used;
491 unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
492 unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
494 seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s\n",
495 "class", "size", "almost_full", "almost_empty",
496 "obj_allocated", "obj_used", "pages_used",
499 for (i = 0; i < zs_size_classes; i++) {
500 class = pool->size_class[i];
502 if (class->index != i)
505 spin_lock(&class->lock);
506 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
507 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
508 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
509 obj_used = zs_stat_get(class, OBJ_USED);
510 spin_unlock(&class->lock);
512 objs_per_zspage = get_maxobj_per_zspage(class->size,
513 class->pages_per_zspage);
514 pages_used = obj_allocated / objs_per_zspage *
515 class->pages_per_zspage;
517 seq_printf(s, " %5u %5u %11lu %12lu %13lu %10lu %10lu %16d\n",
518 i, class->size, class_almost_full, class_almost_empty,
519 obj_allocated, obj_used, pages_used,
520 class->pages_per_zspage);
522 total_class_almost_full += class_almost_full;
523 total_class_almost_empty += class_almost_empty;
524 total_objs += obj_allocated;
525 total_used_objs += obj_used;
526 total_pages += pages_used;
530 seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu\n",
531 "Total", "", total_class_almost_full,
532 total_class_almost_empty, total_objs,
533 total_used_objs, total_pages);
538 static int zs_stats_size_open(struct inode *inode, struct file *file)
540 return single_open(file, zs_stats_size_show, inode->i_private);
543 static const struct file_operations zs_stat_size_ops = {
544 .open = zs_stats_size_open,
547 .release = single_release,
550 static int zs_pool_stat_create(char *name, struct zs_pool *pool)
552 struct dentry *entry;
557 entry = debugfs_create_dir(name, zs_stat_root);
559 pr_warn("debugfs dir <%s> creation failed\n", name);
562 pool->stat_dentry = entry;
564 entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
565 pool->stat_dentry, pool, &zs_stat_size_ops);
567 pr_warn("%s: debugfs file entry <%s> creation failed\n",
575 static void zs_pool_stat_destroy(struct zs_pool *pool)
577 debugfs_remove_recursive(pool->stat_dentry);
580 #else /* CONFIG_ZSMALLOC_STAT */
582 static inline void zs_stat_inc(struct size_class *class,
583 enum zs_stat_type type, unsigned long cnt)
587 static inline void zs_stat_dec(struct size_class *class,
588 enum zs_stat_type type, unsigned long cnt)
592 static inline unsigned long zs_stat_get(struct size_class *class,
593 enum zs_stat_type type)
598 static int __init zs_stat_init(void)
603 static void __exit zs_stat_exit(void)
607 static inline int zs_pool_stat_create(char *name, struct zs_pool *pool)
612 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
620 * For each size class, zspages are divided into different groups
621 * depending on how "full" they are. This was done so that we could
622 * easily find empty or nearly empty zspages when we try to shrink
623 * the pool (not yet implemented). This function returns fullness
624 * status of the given page.
626 static enum fullness_group get_fullness_group(struct page *page)
628 int inuse, max_objects;
629 enum fullness_group fg;
630 BUG_ON(!is_first_page(page));
633 max_objects = page->objects;
637 else if (inuse == max_objects)
639 else if (inuse <= 3 * max_objects / fullness_threshold_frac)
640 fg = ZS_ALMOST_EMPTY;
648 * Each size class maintains various freelists and zspages are assigned
649 * to one of these freelists based on the number of live objects they
650 * have. This functions inserts the given zspage into the freelist
651 * identified by <class, fullness_group>.
653 static void insert_zspage(struct page *page, struct size_class *class,
654 enum fullness_group fullness)
658 BUG_ON(!is_first_page(page));
660 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
663 head = &class->fullness_list[fullness];
665 list_add_tail(&page->lru, &(*head)->lru);
668 zs_stat_inc(class, fullness == ZS_ALMOST_EMPTY ?
669 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
673 * This function removes the given zspage from the freelist identified
674 * by <class, fullness_group>.
676 static void remove_zspage(struct page *page, struct size_class *class,
677 enum fullness_group fullness)
681 BUG_ON(!is_first_page(page));
683 if (fullness >= _ZS_NR_FULLNESS_GROUPS)
686 head = &class->fullness_list[fullness];
688 if (list_empty(&(*head)->lru))
690 else if (*head == page)
691 *head = (struct page *)list_entry((*head)->lru.next,
694 list_del_init(&page->lru);
695 zs_stat_dec(class, fullness == ZS_ALMOST_EMPTY ?
696 CLASS_ALMOST_EMPTY : CLASS_ALMOST_FULL, 1);
700 * Each size class maintains zspages in different fullness groups depending
701 * on the number of live objects they contain. When allocating or freeing
702 * objects, the fullness status of the page can change, say, from ALMOST_FULL
703 * to ALMOST_EMPTY when freeing an object. This function checks if such
704 * a status change has occurred for the given page and accordingly moves the
705 * page from the freelist of the old fullness group to that of the new
708 static enum fullness_group fix_fullness_group(struct size_class *class,
712 enum fullness_group currfg, newfg;
714 BUG_ON(!is_first_page(page));
716 get_zspage_mapping(page, &class_idx, &currfg);
717 newfg = get_fullness_group(page);
721 remove_zspage(page, class, currfg);
722 insert_zspage(page, class, newfg);
723 set_zspage_mapping(page, class_idx, newfg);
730 * We have to decide on how many pages to link together
731 * to form a zspage for each size class. This is important
732 * to reduce wastage due to unusable space left at end of
733 * each zspage which is given as:
734 * wastage = Zp - Zp % size_class
735 * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
737 * For example, for size class of 3/8 * PAGE_SIZE, we should
738 * link together 3 PAGE_SIZE sized pages to form a zspage
739 * since then we can perfectly fit in 8 such objects.
741 static int get_pages_per_zspage(int class_size)
743 int i, max_usedpc = 0;
744 /* zspage order which gives maximum used size per KB */
745 int max_usedpc_order = 1;
747 for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
751 zspage_size = i * PAGE_SIZE;
752 waste = zspage_size % class_size;
753 usedpc = (zspage_size - waste) * 100 / zspage_size;
755 if (usedpc > max_usedpc) {
757 max_usedpc_order = i;
761 return max_usedpc_order;
765 * A single 'zspage' is composed of many system pages which are
766 * linked together using fields in struct page. This function finds
767 * the first/head page, given any component page of a zspage.
769 static struct page *get_first_page(struct page *page)
771 if (is_first_page(page))
774 return page->first_page;
777 static struct page *get_next_page(struct page *page)
781 if (is_last_page(page))
783 else if (is_first_page(page))
784 next = (struct page *)page_private(page);
786 next = list_entry(page->lru.next, struct page, lru);
792 * Encode <page, obj_idx> as a single handle value.
793 * We use the least bit of handle for tagging.
795 static void *location_to_obj(struct page *page, unsigned long obj_idx)
804 obj = page_to_pfn(page) << OBJ_INDEX_BITS;
805 obj |= ((obj_idx) & OBJ_INDEX_MASK);
806 obj <<= OBJ_TAG_BITS;
812 * Decode <page, obj_idx> pair from the given object handle. We adjust the
813 * decoded obj_idx back to its original value since it was adjusted in
816 static void obj_to_location(unsigned long obj, struct page **page,
817 unsigned long *obj_idx)
819 obj >>= OBJ_TAG_BITS;
820 *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
821 *obj_idx = (obj & OBJ_INDEX_MASK);
824 static unsigned long handle_to_obj(unsigned long handle)
826 return *(unsigned long *)handle;
829 static unsigned long obj_to_head(struct size_class *class, struct page *page,
833 VM_BUG_ON(!is_first_page(page));
834 return *(unsigned long *)page_private(page);
836 return *(unsigned long *)obj;
839 static unsigned long obj_idx_to_offset(struct page *page,
840 unsigned long obj_idx, int class_size)
842 unsigned long off = 0;
844 if (!is_first_page(page))
847 return off + obj_idx * class_size;
850 static inline int trypin_tag(unsigned long handle)
852 unsigned long *ptr = (unsigned long *)handle;
854 return !test_and_set_bit_lock(HANDLE_PIN_BIT, ptr);
857 static void pin_tag(unsigned long handle)
859 while (!trypin_tag(handle));
862 static void unpin_tag(unsigned long handle)
864 unsigned long *ptr = (unsigned long *)handle;
866 clear_bit_unlock(HANDLE_PIN_BIT, ptr);
869 static void reset_page(struct page *page)
871 clear_bit(PG_private, &page->flags);
872 clear_bit(PG_private_2, &page->flags);
873 set_page_private(page, 0);
874 page->mapping = NULL;
875 page->freelist = NULL;
876 page_mapcount_reset(page);
879 static void free_zspage(struct page *first_page)
881 struct page *nextp, *tmp, *head_extra;
883 BUG_ON(!is_first_page(first_page));
884 BUG_ON(first_page->inuse);
886 head_extra = (struct page *)page_private(first_page);
888 reset_page(first_page);
889 __free_page(first_page);
891 /* zspage with only 1 system page */
895 list_for_each_entry_safe(nextp, tmp, &head_extra->lru, lru) {
896 list_del(&nextp->lru);
900 reset_page(head_extra);
901 __free_page(head_extra);
904 /* Initialize a newly allocated zspage */
905 static void init_zspage(struct page *first_page, struct size_class *class)
907 unsigned long off = 0;
908 struct page *page = first_page;
910 BUG_ON(!is_first_page(first_page));
912 struct page *next_page;
913 struct link_free *link;
918 * page->index stores offset of first object starting
919 * in the page. For the first page, this is always 0,
920 * so we use first_page->index (aka ->freelist) to store
921 * head of corresponding zspage's freelist.
923 if (page != first_page)
926 vaddr = kmap_atomic(page);
927 link = (struct link_free *)vaddr + off / sizeof(*link);
929 while ((off += class->size) < PAGE_SIZE) {
930 link->next = location_to_obj(page, i++);
931 link += class->size / sizeof(*link);
935 * We now come to the last (full or partial) object on this
936 * page, which must point to the first object on the next
939 next_page = get_next_page(page);
940 link->next = location_to_obj(next_page, 0);
941 kunmap_atomic(vaddr);
948 * Allocate a zspage for the given size class
950 static struct page *alloc_zspage(struct size_class *class, gfp_t flags)
953 struct page *first_page = NULL, *uninitialized_var(prev_page);
956 * Allocate individual pages and link them together as:
957 * 1. first page->private = first sub-page
958 * 2. all sub-pages are linked together using page->lru
959 * 3. each sub-page is linked to the first page using page->first_page
961 * For each size class, First/Head pages are linked together using
962 * page->lru. Also, we set PG_private to identify the first page
963 * (i.e. no other sub-page has this flag set) and PG_private_2 to
964 * identify the last page.
967 for (i = 0; i < class->pages_per_zspage; i++) {
970 page = alloc_page(flags);
974 INIT_LIST_HEAD(&page->lru);
975 if (i == 0) { /* first page */
976 SetPagePrivate(page);
977 set_page_private(page, 0);
979 first_page->inuse = 0;
982 set_page_private(first_page, (unsigned long)page);
984 page->first_page = first_page;
986 list_add(&page->lru, &prev_page->lru);
987 if (i == class->pages_per_zspage - 1) /* last page */
988 SetPagePrivate2(page);
992 init_zspage(first_page, class);
994 first_page->freelist = location_to_obj(first_page, 0);
995 /* Maximum number of objects we can store in this zspage */
996 first_page->objects = class->pages_per_zspage * PAGE_SIZE / class->size;
998 error = 0; /* Success */
1001 if (unlikely(error) && first_page) {
1002 free_zspage(first_page);
1009 static struct page *find_get_zspage(struct size_class *class)
1014 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1015 page = class->fullness_list[i];
1023 #ifdef CONFIG_PGTABLE_MAPPING
1024 static inline int __zs_cpu_up(struct mapping_area *area)
1027 * Make sure we don't leak memory if a cpu UP notification
1028 * and zs_init() race and both call zs_cpu_up() on the same cpu
1032 area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1038 static inline void __zs_cpu_down(struct mapping_area *area)
1041 free_vm_area(area->vm);
1045 static inline void *__zs_map_object(struct mapping_area *area,
1046 struct page *pages[2], int off, int size)
1048 BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1049 area->vm_addr = area->vm->addr;
1050 return area->vm_addr + off;
1053 static inline void __zs_unmap_object(struct mapping_area *area,
1054 struct page *pages[2], int off, int size)
1056 unsigned long addr = (unsigned long)area->vm_addr;
1058 unmap_kernel_range(addr, PAGE_SIZE * 2);
1061 #else /* CONFIG_PGTABLE_MAPPING */
1063 static inline int __zs_cpu_up(struct mapping_area *area)
1066 * Make sure we don't leak memory if a cpu UP notification
1067 * and zs_init() race and both call zs_cpu_up() on the same cpu
1071 area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1077 static inline void __zs_cpu_down(struct mapping_area *area)
1079 kfree(area->vm_buf);
1080 area->vm_buf = NULL;
1083 static void *__zs_map_object(struct mapping_area *area,
1084 struct page *pages[2], int off, int size)
1088 char *buf = area->vm_buf;
1090 /* disable page faults to match kmap_atomic() return conditions */
1091 pagefault_disable();
1093 /* no read fastpath */
1094 if (area->vm_mm == ZS_MM_WO)
1097 sizes[0] = PAGE_SIZE - off;
1098 sizes[1] = size - sizes[0];
1100 /* copy object to per-cpu buffer */
1101 addr = kmap_atomic(pages[0]);
1102 memcpy(buf, addr + off, sizes[0]);
1103 kunmap_atomic(addr);
1104 addr = kmap_atomic(pages[1]);
1105 memcpy(buf + sizes[0], addr, sizes[1]);
1106 kunmap_atomic(addr);
1108 return area->vm_buf;
1111 static void __zs_unmap_object(struct mapping_area *area,
1112 struct page *pages[2], int off, int size)
1118 /* no write fastpath */
1119 if (area->vm_mm == ZS_MM_RO)
1124 buf = buf + ZS_HANDLE_SIZE;
1125 size -= ZS_HANDLE_SIZE;
1126 off += ZS_HANDLE_SIZE;
1129 sizes[0] = PAGE_SIZE - off;
1130 sizes[1] = size - sizes[0];
1132 /* copy per-cpu buffer to object */
1133 addr = kmap_atomic(pages[0]);
1134 memcpy(addr + off, buf, sizes[0]);
1135 kunmap_atomic(addr);
1136 addr = kmap_atomic(pages[1]);
1137 memcpy(addr, buf + sizes[0], sizes[1]);
1138 kunmap_atomic(addr);
1141 /* enable page faults to match kunmap_atomic() return conditions */
1145 #endif /* CONFIG_PGTABLE_MAPPING */
1147 static int zs_cpu_notifier(struct notifier_block *nb, unsigned long action,
1150 int ret, cpu = (long)pcpu;
1151 struct mapping_area *area;
1154 case CPU_UP_PREPARE:
1155 area = &per_cpu(zs_map_area, cpu);
1156 ret = __zs_cpu_up(area);
1158 return notifier_from_errno(ret);
1161 case CPU_UP_CANCELED:
1162 area = &per_cpu(zs_map_area, cpu);
1163 __zs_cpu_down(area);
1170 static struct notifier_block zs_cpu_nb = {
1171 .notifier_call = zs_cpu_notifier
1174 static int zs_register_cpu_notifier(void)
1176 int cpu, uninitialized_var(ret);
1178 cpu_notifier_register_begin();
1180 __register_cpu_notifier(&zs_cpu_nb);
1181 for_each_online_cpu(cpu) {
1182 ret = zs_cpu_notifier(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
1183 if (notifier_to_errno(ret))
1187 cpu_notifier_register_done();
1188 return notifier_to_errno(ret);
1191 static void zs_unregister_cpu_notifier(void)
1195 cpu_notifier_register_begin();
1197 for_each_online_cpu(cpu)
1198 zs_cpu_notifier(NULL, CPU_DEAD, (void *)(long)cpu);
1199 __unregister_cpu_notifier(&zs_cpu_nb);
1201 cpu_notifier_register_done();
1204 static void init_zs_size_classes(void)
1208 nr = (ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) / ZS_SIZE_CLASS_DELTA + 1;
1209 if ((ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE) % ZS_SIZE_CLASS_DELTA)
1212 zs_size_classes = nr;
1215 static bool can_merge(struct size_class *prev, int size, int pages_per_zspage)
1217 if (prev->pages_per_zspage != pages_per_zspage)
1220 if (get_maxobj_per_zspage(prev->size, prev->pages_per_zspage)
1221 != get_maxobj_per_zspage(size, pages_per_zspage))
1227 static bool zspage_full(struct page *page)
1229 BUG_ON(!is_first_page(page));
1231 return page->inuse == page->objects;
1234 unsigned long zs_get_total_pages(struct zs_pool *pool)
1236 return atomic_long_read(&pool->pages_allocated);
1238 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1241 * zs_map_object - get address of allocated object from handle.
1242 * @pool: pool from which the object was allocated
1243 * @handle: handle returned from zs_malloc
1245 * Before using an object allocated from zs_malloc, it must be mapped using
1246 * this function. When done with the object, it must be unmapped using
1249 * Only one object can be mapped per cpu at a time. There is no protection
1250 * against nested mappings.
1252 * This function returns with preemption and page faults disabled.
1254 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1258 unsigned long obj, obj_idx, off;
1260 unsigned int class_idx;
1261 enum fullness_group fg;
1262 struct size_class *class;
1263 struct mapping_area *area;
1264 struct page *pages[2];
1270 * Because we use per-cpu mapping areas shared among the
1271 * pools/users, we can't allow mapping in interrupt context
1272 * because it can corrupt another users mappings.
1274 BUG_ON(in_interrupt());
1276 /* From now on, migration cannot move the object */
1279 obj = handle_to_obj(handle);
1280 obj_to_location(obj, &page, &obj_idx);
1281 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1282 class = pool->size_class[class_idx];
1283 off = obj_idx_to_offset(page, obj_idx, class->size);
1285 area = &get_cpu_var(zs_map_area);
1287 if (off + class->size <= PAGE_SIZE) {
1288 /* this object is contained entirely within a page */
1289 area->vm_addr = kmap_atomic(page);
1290 ret = area->vm_addr + off;
1294 /* this object spans two pages */
1296 pages[1] = get_next_page(page);
1299 ret = __zs_map_object(area, pages, off, class->size);
1302 ret += ZS_HANDLE_SIZE;
1306 EXPORT_SYMBOL_GPL(zs_map_object);
1308 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1311 unsigned long obj, obj_idx, off;
1313 unsigned int class_idx;
1314 enum fullness_group fg;
1315 struct size_class *class;
1316 struct mapping_area *area;
1320 obj = handle_to_obj(handle);
1321 obj_to_location(obj, &page, &obj_idx);
1322 get_zspage_mapping(get_first_page(page), &class_idx, &fg);
1323 class = pool->size_class[class_idx];
1324 off = obj_idx_to_offset(page, obj_idx, class->size);
1326 area = this_cpu_ptr(&zs_map_area);
1327 if (off + class->size <= PAGE_SIZE)
1328 kunmap_atomic(area->vm_addr);
1330 struct page *pages[2];
1333 pages[1] = get_next_page(page);
1336 __zs_unmap_object(area, pages, off, class->size);
1338 put_cpu_var(zs_map_area);
1341 EXPORT_SYMBOL_GPL(zs_unmap_object);
1343 static unsigned long obj_malloc(struct page *first_page,
1344 struct size_class *class, unsigned long handle)
1347 struct link_free *link;
1349 struct page *m_page;
1350 unsigned long m_objidx, m_offset;
1353 handle |= OBJ_ALLOCATED_TAG;
1354 obj = (unsigned long)first_page->freelist;
1355 obj_to_location(obj, &m_page, &m_objidx);
1356 m_offset = obj_idx_to_offset(m_page, m_objidx, class->size);
1358 vaddr = kmap_atomic(m_page);
1359 link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1360 first_page->freelist = link->next;
1362 /* record handle in the header of allocated chunk */
1363 link->handle = handle;
1365 /* record handle in first_page->private */
1366 set_page_private(first_page, handle);
1367 kunmap_atomic(vaddr);
1368 first_page->inuse++;
1369 zs_stat_inc(class, OBJ_USED, 1);
1376 * zs_malloc - Allocate block of given size from pool.
1377 * @pool: pool to allocate from
1378 * @size: size of block to allocate
1380 * On success, handle to the allocated object is returned,
1382 * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1384 unsigned long zs_malloc(struct zs_pool *pool, size_t size)
1386 unsigned long handle, obj;
1387 struct size_class *class;
1388 struct page *first_page;
1390 if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1393 handle = alloc_handle(pool);
1397 /* extra space in chunk to keep the handle */
1398 size += ZS_HANDLE_SIZE;
1399 class = pool->size_class[get_size_class_index(size)];
1400 /* In huge class size, we store the handle into first_page->private */
1402 size -= ZS_HANDLE_SIZE;
1403 class = pool->size_class[get_size_class_index(size)];
1406 spin_lock(&class->lock);
1407 first_page = find_get_zspage(class);
1410 spin_unlock(&class->lock);
1411 first_page = alloc_zspage(class, pool->flags);
1412 if (unlikely(!first_page)) {
1413 free_handle(pool, handle);
1417 set_zspage_mapping(first_page, class->index, ZS_EMPTY);
1418 atomic_long_add(class->pages_per_zspage,
1419 &pool->pages_allocated);
1421 spin_lock(&class->lock);
1422 zs_stat_inc(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1423 class->size, class->pages_per_zspage));
1426 obj = obj_malloc(first_page, class, handle);
1427 /* Now move the zspage to another fullness group, if required */
1428 fix_fullness_group(class, first_page);
1429 record_obj(handle, obj);
1430 spin_unlock(&class->lock);
1434 EXPORT_SYMBOL_GPL(zs_malloc);
1436 static void obj_free(struct zs_pool *pool, struct size_class *class,
1439 struct link_free *link;
1440 struct page *first_page, *f_page;
1441 unsigned long f_objidx, f_offset;
1444 enum fullness_group fullness;
1448 obj &= ~OBJ_ALLOCATED_TAG;
1449 obj_to_location(obj, &f_page, &f_objidx);
1450 first_page = get_first_page(f_page);
1452 get_zspage_mapping(first_page, &class_idx, &fullness);
1453 f_offset = obj_idx_to_offset(f_page, f_objidx, class->size);
1455 vaddr = kmap_atomic(f_page);
1457 /* Insert this object in containing zspage's freelist */
1458 link = (struct link_free *)(vaddr + f_offset);
1459 link->next = first_page->freelist;
1461 set_page_private(first_page, 0);
1462 kunmap_atomic(vaddr);
1463 first_page->freelist = (void *)obj;
1464 first_page->inuse--;
1465 zs_stat_dec(class, OBJ_USED, 1);
1468 void zs_free(struct zs_pool *pool, unsigned long handle)
1470 struct page *first_page, *f_page;
1471 unsigned long obj, f_objidx;
1473 struct size_class *class;
1474 enum fullness_group fullness;
1476 if (unlikely(!handle))
1480 obj = handle_to_obj(handle);
1481 obj_to_location(obj, &f_page, &f_objidx);
1482 first_page = get_first_page(f_page);
1484 get_zspage_mapping(first_page, &class_idx, &fullness);
1485 class = pool->size_class[class_idx];
1487 spin_lock(&class->lock);
1488 obj_free(pool, class, obj);
1489 fullness = fix_fullness_group(class, first_page);
1490 if (fullness == ZS_EMPTY) {
1491 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1492 class->size, class->pages_per_zspage));
1493 atomic_long_sub(class->pages_per_zspage,
1494 &pool->pages_allocated);
1495 free_zspage(first_page);
1497 spin_unlock(&class->lock);
1500 free_handle(pool, handle);
1502 EXPORT_SYMBOL_GPL(zs_free);
1504 static void zs_object_copy(unsigned long src, unsigned long dst,
1505 struct size_class *class)
1507 struct page *s_page, *d_page;
1508 unsigned long s_objidx, d_objidx;
1509 unsigned long s_off, d_off;
1510 void *s_addr, *d_addr;
1511 int s_size, d_size, size;
1514 s_size = d_size = class->size;
1516 obj_to_location(src, &s_page, &s_objidx);
1517 obj_to_location(dst, &d_page, &d_objidx);
1519 s_off = obj_idx_to_offset(s_page, s_objidx, class->size);
1520 d_off = obj_idx_to_offset(d_page, d_objidx, class->size);
1522 if (s_off + class->size > PAGE_SIZE)
1523 s_size = PAGE_SIZE - s_off;
1525 if (d_off + class->size > PAGE_SIZE)
1526 d_size = PAGE_SIZE - d_off;
1528 s_addr = kmap_atomic(s_page);
1529 d_addr = kmap_atomic(d_page);
1532 size = min(s_size, d_size);
1533 memcpy(d_addr + d_off, s_addr + s_off, size);
1536 if (written == class->size)
1539 if (s_off + size >= PAGE_SIZE) {
1540 kunmap_atomic(d_addr);
1541 kunmap_atomic(s_addr);
1542 s_page = get_next_page(s_page);
1544 s_addr = kmap_atomic(s_page);
1545 d_addr = kmap_atomic(d_page);
1546 s_size = class->size - written;
1553 if (d_off + size >= PAGE_SIZE) {
1554 kunmap_atomic(d_addr);
1555 d_page = get_next_page(d_page);
1557 d_addr = kmap_atomic(d_page);
1558 d_size = class->size - written;
1566 kunmap_atomic(d_addr);
1567 kunmap_atomic(s_addr);
1571 * Find alloced object in zspage from index object and
1574 static unsigned long find_alloced_obj(struct page *page, int index,
1575 struct size_class *class)
1579 unsigned long handle = 0;
1580 void *addr = kmap_atomic(page);
1582 if (!is_first_page(page))
1583 offset = page->index;
1584 offset += class->size * index;
1586 while (offset < PAGE_SIZE) {
1587 head = obj_to_head(class, page, addr + offset);
1588 if (head & OBJ_ALLOCATED_TAG) {
1589 handle = head & ~OBJ_ALLOCATED_TAG;
1590 if (trypin_tag(handle))
1595 offset += class->size;
1599 kunmap_atomic(addr);
1603 struct zs_compact_control {
1604 /* Source page for migration which could be a subpage of zspage. */
1605 struct page *s_page;
1606 /* Destination page for migration which should be a first page
1608 struct page *d_page;
1609 /* Starting object index within @s_page which used for live object
1610 * in the subpage. */
1612 /* how many of objects are migrated */
1616 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1617 struct zs_compact_control *cc)
1619 unsigned long used_obj, free_obj;
1620 unsigned long handle;
1621 struct page *s_page = cc->s_page;
1622 struct page *d_page = cc->d_page;
1623 unsigned long index = cc->index;
1624 int nr_migrated = 0;
1628 handle = find_alloced_obj(s_page, index, class);
1630 s_page = get_next_page(s_page);
1637 /* Stop if there is no more space */
1638 if (zspage_full(d_page)) {
1644 used_obj = handle_to_obj(handle);
1645 free_obj = obj_malloc(d_page, class, handle);
1646 zs_object_copy(used_obj, free_obj, class);
1648 record_obj(handle, free_obj);
1650 obj_free(pool, class, used_obj);
1654 /* Remember last position in this iteration */
1655 cc->s_page = s_page;
1657 cc->nr_migrated = nr_migrated;
1662 static struct page *alloc_target_page(struct size_class *class)
1667 for (i = 0; i < _ZS_NR_FULLNESS_GROUPS; i++) {
1668 page = class->fullness_list[i];
1670 remove_zspage(page, class, i);
1678 static void putback_zspage(struct zs_pool *pool, struct size_class *class,
1679 struct page *first_page)
1682 enum fullness_group fullness;
1684 BUG_ON(!is_first_page(first_page));
1686 get_zspage_mapping(first_page, &class_idx, &fullness);
1687 insert_zspage(first_page, class, fullness);
1688 fullness = fix_fullness_group(class, first_page);
1689 if (fullness == ZS_EMPTY) {
1690 zs_stat_dec(class, OBJ_ALLOCATED, get_maxobj_per_zspage(
1691 class->size, class->pages_per_zspage));
1692 atomic_long_sub(class->pages_per_zspage,
1693 &pool->pages_allocated);
1695 free_zspage(first_page);
1699 static struct page *isolate_source_page(struct size_class *class)
1703 page = class->fullness_list[ZS_ALMOST_EMPTY];
1705 remove_zspage(page, class, ZS_ALMOST_EMPTY);
1710 static unsigned long __zs_compact(struct zs_pool *pool,
1711 struct size_class *class)
1714 struct zs_compact_control cc;
1715 struct page *src_page;
1716 struct page *dst_page = NULL;
1717 unsigned long nr_total_migrated = 0;
1721 spin_lock(&class->lock);
1722 while ((src_page = isolate_source_page(class))) {
1724 BUG_ON(!is_first_page(src_page));
1726 /* The goal is to migrate all live objects in source page */
1727 nr_to_migrate = src_page->inuse;
1729 cc.s_page = src_page;
1731 while ((dst_page = alloc_target_page(class))) {
1732 cc.d_page = dst_page;
1734 * If there is no more space in dst_page, try to
1735 * allocate another zspage.
1737 if (!migrate_zspage(pool, class, &cc))
1740 putback_zspage(pool, class, dst_page);
1741 nr_total_migrated += cc.nr_migrated;
1742 nr_to_migrate -= cc.nr_migrated;
1745 /* Stop if we couldn't find slot */
1746 if (dst_page == NULL)
1749 putback_zspage(pool, class, dst_page);
1750 putback_zspage(pool, class, src_page);
1751 spin_unlock(&class->lock);
1752 nr_total_migrated += cc.nr_migrated;
1754 spin_lock(&class->lock);
1758 putback_zspage(pool, class, src_page);
1760 spin_unlock(&class->lock);
1762 return nr_total_migrated;
1765 unsigned long zs_compact(struct zs_pool *pool)
1768 unsigned long nr_migrated = 0;
1769 struct size_class *class;
1771 for (i = zs_size_classes - 1; i >= 0; i--) {
1772 class = pool->size_class[i];
1775 if (class->index != i)
1777 nr_migrated += __zs_compact(pool, class);
1784 EXPORT_SYMBOL_GPL(zs_compact);
1787 * zs_create_pool - Creates an allocation pool to work from.
1788 * @flags: allocation flags used to allocate pool metadata
1790 * This function must be called before anything when using
1791 * the zsmalloc allocator.
1793 * On success, a pointer to the newly created pool is returned,
1796 struct zs_pool *zs_create_pool(char *name, gfp_t flags)
1799 struct zs_pool *pool;
1800 struct size_class *prev_class = NULL;
1802 pool = kzalloc(sizeof(*pool), GFP_KERNEL);
1806 pool->size_class = kcalloc(zs_size_classes, sizeof(struct size_class *),
1808 if (!pool->size_class) {
1813 pool->name = kstrdup(name, GFP_KERNEL);
1817 if (create_handle_cache(pool))
1821 * Iterate reversly, because, size of size_class that we want to use
1822 * for merging should be larger or equal to current size.
1824 for (i = zs_size_classes - 1; i >= 0; i--) {
1826 int pages_per_zspage;
1827 struct size_class *class;
1829 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
1830 if (size > ZS_MAX_ALLOC_SIZE)
1831 size = ZS_MAX_ALLOC_SIZE;
1832 pages_per_zspage = get_pages_per_zspage(size);
1835 * size_class is used for normal zsmalloc operation such
1836 * as alloc/free for that size. Although it is natural that we
1837 * have one size_class for each size, there is a chance that we
1838 * can get more memory utilization if we use one size_class for
1839 * many different sizes whose size_class have same
1840 * characteristics. So, we makes size_class point to
1841 * previous size_class if possible.
1844 if (can_merge(prev_class, size, pages_per_zspage)) {
1845 pool->size_class[i] = prev_class;
1850 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
1856 class->pages_per_zspage = pages_per_zspage;
1857 if (pages_per_zspage == 1 &&
1858 get_maxobj_per_zspage(size, pages_per_zspage) == 1)
1860 spin_lock_init(&class->lock);
1861 pool->size_class[i] = class;
1866 pool->flags = flags;
1868 if (zs_pool_stat_create(name, pool))
1874 zs_destroy_pool(pool);
1877 EXPORT_SYMBOL_GPL(zs_create_pool);
1879 void zs_destroy_pool(struct zs_pool *pool)
1883 zs_pool_stat_destroy(pool);
1885 for (i = 0; i < zs_size_classes; i++) {
1887 struct size_class *class = pool->size_class[i];
1892 if (class->index != i)
1895 for (fg = 0; fg < _ZS_NR_FULLNESS_GROUPS; fg++) {
1896 if (class->fullness_list[fg]) {
1897 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
1904 destroy_handle_cache(pool);
1905 kfree(pool->size_class);
1909 EXPORT_SYMBOL_GPL(zs_destroy_pool);
1911 static int __init zs_init(void)
1913 int ret = zs_register_cpu_notifier();
1918 init_zs_size_classes();
1921 zpool_register_driver(&zs_zpool_driver);
1924 ret = zs_stat_init();
1926 pr_err("zs stat initialization failed\n");
1933 zpool_unregister_driver(&zs_zpool_driver);
1936 zs_unregister_cpu_notifier();
1941 static void __exit zs_exit(void)
1944 zpool_unregister_driver(&zs_zpool_driver);
1946 zs_unregister_cpu_notifier();
1951 module_init(zs_init);
1952 module_exit(zs_exit);
1954 MODULE_LICENSE("Dual BSD/GPL");
1955 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");