1 // SPDX-License-Identifier: GPL-2.0-only
5 * Author: Vitaly Wool <vitaly.wool@konsulko.com>
6 * Copyright (C) 2016, Sony Mobile Communications Inc.
8 * This implementation is based on zbud written by Seth Jennings.
10 * z3fold is an special purpose allocator for storing compressed pages. It
11 * can store up to three compressed pages per page which improves the
12 * compression ratio of zbud while retaining its main concepts (e. g. always
13 * storing an integral number of objects per page) and simplicity.
14 * It still has simple and deterministic reclaim properties that make it
15 * preferable to a higher density approach (with no requirement on integral
16 * number of object per page) when reclaim is used.
18 * As in zbud, pages are divided into "chunks". The size of the chunks is
19 * fixed at compile time and is determined by NCHUNKS_ORDER below.
21 * z3fold doesn't export any API and is meant to be used via zpool API.
24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
26 #include <linux/atomic.h>
27 #include <linux/sched.h>
28 #include <linux/cpumask.h>
29 #include <linux/list.h>
31 #include <linux/module.h>
32 #include <linux/page-flags.h>
33 #include <linux/migrate.h>
34 #include <linux/node.h>
35 #include <linux/compaction.h>
36 #include <linux/percpu.h>
37 #include <linux/mount.h>
38 #include <linux/pseudo_fs.h>
40 #include <linux/preempt.h>
41 #include <linux/workqueue.h>
42 #include <linux/slab.h>
43 #include <linux/spinlock.h>
44 #include <linux/zpool.h>
45 #include <linux/magic.h>
48 * NCHUNKS_ORDER determines the internal allocation granularity, effectively
49 * adjusting internal fragmentation. It also determines the number of
50 * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
51 * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
52 * in the beginning of an allocated page are occupied by z3fold header, so
53 * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
54 * which shows the max number of free chunks in z3fold page, also there will
55 * be 63, or 62, respectively, freelists per pool.
57 #define NCHUNKS_ORDER 6
59 #define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
60 #define CHUNK_SIZE (1 << CHUNK_SHIFT)
61 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
62 #define ZHDR_CHUNKS (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
63 #define TOTAL_CHUNKS (PAGE_SIZE >> CHUNK_SHIFT)
64 #define NCHUNKS ((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
66 #define BUDDY_MASK (0x3)
68 #define SLOTS_ALIGN (0x40)
75 int (*evict
)(struct z3fold_pool
*pool
, unsigned long handle
);
86 struct z3fold_buddy_slots
{
88 * we are using BUDDY_MASK in handle_to_buddy etc. so there should
89 * be enough slots to hold all possible variants
91 unsigned long slot
[BUDDY_MASK
+ 1];
92 unsigned long pool
; /* back link + flags */
94 #define HANDLE_FLAG_MASK (0x03)
97 * struct z3fold_header - z3fold page metadata occupying first chunks of each
98 * z3fold page, except for HEADLESS pages
99 * @buddy: links the z3fold page into the relevant list in the
101 * @page_lock: per-page lock
102 * @refcount: reference count for the z3fold page
103 * @work: work_struct for page layout optimization
104 * @slots: pointer to the structure holding buddy slots
105 * @pool: pointer to the containing pool
106 * @cpu: CPU which this page "belongs" to
107 * @first_chunks: the size of the first buddy in chunks, 0 if free
108 * @middle_chunks: the size of the middle buddy in chunks, 0 if free
109 * @last_chunks: the size of the last buddy in chunks, 0 if free
110 * @first_num: the starting number (for the first handle)
111 * @mapped_count: the number of objects currently mapped
113 struct z3fold_header
{
114 struct list_head buddy
;
115 spinlock_t page_lock
;
116 struct kref refcount
;
117 struct work_struct work
;
118 struct z3fold_buddy_slots
*slots
;
119 struct z3fold_pool
*pool
;
121 unsigned short first_chunks
;
122 unsigned short middle_chunks
;
123 unsigned short last_chunks
;
124 unsigned short start_middle
;
125 unsigned short first_num
:2;
126 unsigned short mapped_count
:2;
130 * struct z3fold_pool - stores metadata for each z3fold pool
132 * @lock: protects pool unbuddied/lru lists
133 * @stale_lock: protects pool stale page list
134 * @unbuddied: per-cpu array of lists tracking z3fold pages that contain 2-
135 * buddies; the list each z3fold page is added to depends on
136 * the size of its free region.
137 * @lru: list tracking the z3fold pages in LRU order by most recently
139 * @stale: list of pages marked for freeing
140 * @pages_nr: number of z3fold pages in the pool.
141 * @c_handle: cache for z3fold_buddy_slots allocation
142 * @ops: pointer to a structure of user defined operations specified at
143 * pool creation time.
144 * @compact_wq: workqueue for page layout background optimization
145 * @release_wq: workqueue for safe page release
146 * @work: work_struct for safe page release
147 * @inode: inode for z3fold pseudo filesystem
149 * This structure is allocated at pool creation time and maintains metadata
150 * pertaining to a particular z3fold pool.
155 spinlock_t stale_lock
;
156 struct list_head
*unbuddied
;
157 struct list_head lru
;
158 struct list_head stale
;
160 struct kmem_cache
*c_handle
;
161 const struct z3fold_ops
*ops
;
163 const struct zpool_ops
*zpool_ops
;
164 struct workqueue_struct
*compact_wq
;
165 struct workqueue_struct
*release_wq
;
166 struct work_struct work
;
171 * Internal z3fold page flags
173 enum z3fold_page_flags
{
178 PAGE_CLAIMED
, /* by either reclaim or free */
185 /* Converts an allocation size in bytes to size in z3fold chunks */
186 static int size_to_chunks(size_t size
)
188 return (size
+ CHUNK_SIZE
- 1) >> CHUNK_SHIFT
;
191 #define for_each_unbuddied_list(_iter, _begin) \
192 for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
194 static void compact_page_work(struct work_struct
*w
);
196 static inline struct z3fold_buddy_slots
*alloc_slots(struct z3fold_pool
*pool
,
199 struct z3fold_buddy_slots
*slots
;
201 slots
= kmem_cache_alloc(pool
->c_handle
,
202 (gfp
& ~(__GFP_HIGHMEM
| __GFP_MOVABLE
)));
205 memset(slots
->slot
, 0, sizeof(slots
->slot
));
206 slots
->pool
= (unsigned long)pool
;
212 static inline struct z3fold_pool
*slots_to_pool(struct z3fold_buddy_slots
*s
)
214 return (struct z3fold_pool
*)(s
->pool
& ~HANDLE_FLAG_MASK
);
217 static inline struct z3fold_buddy_slots
*handle_to_slots(unsigned long handle
)
219 return (struct z3fold_buddy_slots
*)(handle
& ~(SLOTS_ALIGN
- 1));
222 static inline void free_handle(unsigned long handle
)
224 struct z3fold_buddy_slots
*slots
;
228 if (handle
& (1 << PAGE_HEADLESS
))
231 WARN_ON(*(unsigned long *)handle
== 0);
232 *(unsigned long *)handle
= 0;
233 slots
= handle_to_slots(handle
);
235 for (i
= 0; i
<= BUDDY_MASK
; i
++) {
236 if (slots
->slot
[i
]) {
243 struct z3fold_pool
*pool
= slots_to_pool(slots
);
245 kmem_cache_free(pool
->c_handle
, slots
);
249 static int z3fold_init_fs_context(struct fs_context
*fc
)
251 return init_pseudo(fc
, Z3FOLD_MAGIC
) ? 0 : -ENOMEM
;
254 static struct file_system_type z3fold_fs
= {
256 .init_fs_context
= z3fold_init_fs_context
,
257 .kill_sb
= kill_anon_super
,
260 static struct vfsmount
*z3fold_mnt
;
261 static int z3fold_mount(void)
265 z3fold_mnt
= kern_mount(&z3fold_fs
);
266 if (IS_ERR(z3fold_mnt
))
267 ret
= PTR_ERR(z3fold_mnt
);
272 static void z3fold_unmount(void)
274 kern_unmount(z3fold_mnt
);
277 static const struct address_space_operations z3fold_aops
;
278 static int z3fold_register_migration(struct z3fold_pool
*pool
)
280 pool
->inode
= alloc_anon_inode(z3fold_mnt
->mnt_sb
);
281 if (IS_ERR(pool
->inode
)) {
286 pool
->inode
->i_mapping
->private_data
= pool
;
287 pool
->inode
->i_mapping
->a_ops
= &z3fold_aops
;
291 static void z3fold_unregister_migration(struct z3fold_pool
*pool
)
297 /* Initializes the z3fold header of a newly allocated z3fold page */
298 static struct z3fold_header
*init_z3fold_page(struct page
*page
,
299 struct z3fold_pool
*pool
, gfp_t gfp
)
301 struct z3fold_header
*zhdr
= page_address(page
);
302 struct z3fold_buddy_slots
*slots
= alloc_slots(pool
, gfp
);
307 INIT_LIST_HEAD(&page
->lru
);
308 clear_bit(PAGE_HEADLESS
, &page
->private);
309 clear_bit(MIDDLE_CHUNK_MAPPED
, &page
->private);
310 clear_bit(NEEDS_COMPACTING
, &page
->private);
311 clear_bit(PAGE_STALE
, &page
->private);
312 clear_bit(PAGE_CLAIMED
, &page
->private);
314 spin_lock_init(&zhdr
->page_lock
);
315 kref_init(&zhdr
->refcount
);
316 zhdr
->first_chunks
= 0;
317 zhdr
->middle_chunks
= 0;
318 zhdr
->last_chunks
= 0;
320 zhdr
->start_middle
= 0;
324 INIT_LIST_HEAD(&zhdr
->buddy
);
325 INIT_WORK(&zhdr
->work
, compact_page_work
);
329 /* Resets the struct page fields and frees the page */
330 static void free_z3fold_page(struct page
*page
, bool headless
)
334 __ClearPageMovable(page
);
337 ClearPagePrivate(page
);
341 /* Lock a z3fold page */
342 static inline void z3fold_page_lock(struct z3fold_header
*zhdr
)
344 spin_lock(&zhdr
->page_lock
);
347 /* Try to lock a z3fold page */
348 static inline int z3fold_page_trylock(struct z3fold_header
*zhdr
)
350 return spin_trylock(&zhdr
->page_lock
);
353 /* Unlock a z3fold page */
354 static inline void z3fold_page_unlock(struct z3fold_header
*zhdr
)
356 spin_unlock(&zhdr
->page_lock
);
359 /* Helper function to build the index */
360 static inline int __idx(struct z3fold_header
*zhdr
, enum buddy bud
)
362 return (bud
+ zhdr
->first_num
) & BUDDY_MASK
;
366 * Encodes the handle of a particular buddy within a z3fold page
367 * Pool lock should be held as this function accesses first_num
369 static unsigned long encode_handle(struct z3fold_header
*zhdr
, enum buddy bud
)
371 struct z3fold_buddy_slots
*slots
;
372 unsigned long h
= (unsigned long)zhdr
;
376 * For a headless page, its handle is its pointer with the extra
377 * PAGE_HEADLESS bit set
380 return h
| (1 << PAGE_HEADLESS
);
382 /* otherwise, return pointer to encoded handle */
383 idx
= __idx(zhdr
, bud
);
386 h
|= (zhdr
->last_chunks
<< BUDDY_SHIFT
);
389 slots
->slot
[idx
] = h
;
390 return (unsigned long)&slots
->slot
[idx
];
393 /* Returns the z3fold page where a given handle is stored */
394 static inline struct z3fold_header
*handle_to_z3fold_header(unsigned long h
)
396 unsigned long addr
= h
;
398 if (!(addr
& (1 << PAGE_HEADLESS
)))
399 addr
= *(unsigned long *)h
;
401 return (struct z3fold_header
*)(addr
& PAGE_MASK
);
404 /* only for LAST bud, returns zero otherwise */
405 static unsigned short handle_to_chunks(unsigned long handle
)
407 unsigned long addr
= *(unsigned long *)handle
;
409 return (addr
& ~PAGE_MASK
) >> BUDDY_SHIFT
;
413 * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
414 * but that doesn't matter. because the masking will result in the
415 * correct buddy number.
417 static enum buddy
handle_to_buddy(unsigned long handle
)
419 struct z3fold_header
*zhdr
;
422 WARN_ON(handle
& (1 << PAGE_HEADLESS
));
423 addr
= *(unsigned long *)handle
;
424 zhdr
= (struct z3fold_header
*)(addr
& PAGE_MASK
);
425 return (addr
- zhdr
->first_num
) & BUDDY_MASK
;
428 static inline struct z3fold_pool
*zhdr_to_pool(struct z3fold_header
*zhdr
)
433 static void __release_z3fold_page(struct z3fold_header
*zhdr
, bool locked
)
435 struct page
*page
= virt_to_page(zhdr
);
436 struct z3fold_pool
*pool
= zhdr_to_pool(zhdr
);
438 WARN_ON(!list_empty(&zhdr
->buddy
));
439 set_bit(PAGE_STALE
, &page
->private);
440 clear_bit(NEEDS_COMPACTING
, &page
->private);
441 spin_lock(&pool
->lock
);
442 if (!list_empty(&page
->lru
))
443 list_del_init(&page
->lru
);
444 spin_unlock(&pool
->lock
);
446 z3fold_page_unlock(zhdr
);
447 spin_lock(&pool
->stale_lock
);
448 list_add(&zhdr
->buddy
, &pool
->stale
);
449 queue_work(pool
->release_wq
, &pool
->work
);
450 spin_unlock(&pool
->stale_lock
);
453 static void __attribute__((__unused__
))
454 release_z3fold_page(struct kref
*ref
)
456 struct z3fold_header
*zhdr
= container_of(ref
, struct z3fold_header
,
458 __release_z3fold_page(zhdr
, false);
461 static void release_z3fold_page_locked(struct kref
*ref
)
463 struct z3fold_header
*zhdr
= container_of(ref
, struct z3fold_header
,
465 WARN_ON(z3fold_page_trylock(zhdr
));
466 __release_z3fold_page(zhdr
, true);
469 static void release_z3fold_page_locked_list(struct kref
*ref
)
471 struct z3fold_header
*zhdr
= container_of(ref
, struct z3fold_header
,
473 struct z3fold_pool
*pool
= zhdr_to_pool(zhdr
);
474 spin_lock(&pool
->lock
);
475 list_del_init(&zhdr
->buddy
);
476 spin_unlock(&pool
->lock
);
478 WARN_ON(z3fold_page_trylock(zhdr
));
479 __release_z3fold_page(zhdr
, true);
482 static void free_pages_work(struct work_struct
*w
)
484 struct z3fold_pool
*pool
= container_of(w
, struct z3fold_pool
, work
);
486 spin_lock(&pool
->stale_lock
);
487 while (!list_empty(&pool
->stale
)) {
488 struct z3fold_header
*zhdr
= list_first_entry(&pool
->stale
,
489 struct z3fold_header
, buddy
);
490 struct page
*page
= virt_to_page(zhdr
);
492 list_del(&zhdr
->buddy
);
493 if (WARN_ON(!test_bit(PAGE_STALE
, &page
->private)))
495 spin_unlock(&pool
->stale_lock
);
496 cancel_work_sync(&zhdr
->work
);
497 free_z3fold_page(page
, false);
499 spin_lock(&pool
->stale_lock
);
501 spin_unlock(&pool
->stale_lock
);
505 * Returns the number of free chunks in a z3fold page.
506 * NB: can't be used with HEADLESS pages.
508 static int num_free_chunks(struct z3fold_header
*zhdr
)
512 * If there is a middle object, pick up the bigger free space
513 * either before or after it. Otherwise just subtract the number
514 * of chunks occupied by the first and the last objects.
516 if (zhdr
->middle_chunks
!= 0) {
517 int nfree_before
= zhdr
->first_chunks
?
518 0 : zhdr
->start_middle
- ZHDR_CHUNKS
;
519 int nfree_after
= zhdr
->last_chunks
?
521 (zhdr
->start_middle
+ zhdr
->middle_chunks
);
522 nfree
= max(nfree_before
, nfree_after
);
524 nfree
= NCHUNKS
- zhdr
->first_chunks
- zhdr
->last_chunks
;
528 /* Add to the appropriate unbuddied list */
529 static inline void add_to_unbuddied(struct z3fold_pool
*pool
,
530 struct z3fold_header
*zhdr
)
532 if (zhdr
->first_chunks
== 0 || zhdr
->last_chunks
== 0 ||
533 zhdr
->middle_chunks
== 0) {
534 struct list_head
*unbuddied
= get_cpu_ptr(pool
->unbuddied
);
536 int freechunks
= num_free_chunks(zhdr
);
537 spin_lock(&pool
->lock
);
538 list_add(&zhdr
->buddy
, &unbuddied
[freechunks
]);
539 spin_unlock(&pool
->lock
);
540 zhdr
->cpu
= smp_processor_id();
541 put_cpu_ptr(pool
->unbuddied
);
545 static inline void *mchunk_memmove(struct z3fold_header
*zhdr
,
546 unsigned short dst_chunk
)
549 return memmove(beg
+ (dst_chunk
<< CHUNK_SHIFT
),
550 beg
+ (zhdr
->start_middle
<< CHUNK_SHIFT
),
551 zhdr
->middle_chunks
<< CHUNK_SHIFT
);
554 #define BIG_CHUNK_GAP 3
555 /* Has to be called with lock held */
556 static int z3fold_compact_page(struct z3fold_header
*zhdr
)
558 struct page
*page
= virt_to_page(zhdr
);
560 if (test_bit(MIDDLE_CHUNK_MAPPED
, &page
->private))
561 return 0; /* can't move middle chunk, it's used */
563 if (unlikely(PageIsolated(page
)))
566 if (zhdr
->middle_chunks
== 0)
567 return 0; /* nothing to compact */
569 if (zhdr
->first_chunks
== 0 && zhdr
->last_chunks
== 0) {
570 /* move to the beginning */
571 mchunk_memmove(zhdr
, ZHDR_CHUNKS
);
572 zhdr
->first_chunks
= zhdr
->middle_chunks
;
573 zhdr
->middle_chunks
= 0;
574 zhdr
->start_middle
= 0;
580 * moving data is expensive, so let's only do that if
581 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
583 if (zhdr
->first_chunks
!= 0 && zhdr
->last_chunks
== 0 &&
584 zhdr
->start_middle
- (zhdr
->first_chunks
+ ZHDR_CHUNKS
) >=
586 mchunk_memmove(zhdr
, zhdr
->first_chunks
+ ZHDR_CHUNKS
);
587 zhdr
->start_middle
= zhdr
->first_chunks
+ ZHDR_CHUNKS
;
589 } else if (zhdr
->last_chunks
!= 0 && zhdr
->first_chunks
== 0 &&
590 TOTAL_CHUNKS
- (zhdr
->last_chunks
+ zhdr
->start_middle
591 + zhdr
->middle_chunks
) >=
593 unsigned short new_start
= TOTAL_CHUNKS
- zhdr
->last_chunks
-
595 mchunk_memmove(zhdr
, new_start
);
596 zhdr
->start_middle
= new_start
;
603 static void do_compact_page(struct z3fold_header
*zhdr
, bool locked
)
605 struct z3fold_pool
*pool
= zhdr_to_pool(zhdr
);
608 page
= virt_to_page(zhdr
);
610 WARN_ON(z3fold_page_trylock(zhdr
));
612 z3fold_page_lock(zhdr
);
613 if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING
, &page
->private))) {
614 z3fold_page_unlock(zhdr
);
617 spin_lock(&pool
->lock
);
618 list_del_init(&zhdr
->buddy
);
619 spin_unlock(&pool
->lock
);
621 if (kref_put(&zhdr
->refcount
, release_z3fold_page_locked
)) {
622 atomic64_dec(&pool
->pages_nr
);
626 if (unlikely(PageIsolated(page
) ||
627 test_bit(PAGE_STALE
, &page
->private))) {
628 z3fold_page_unlock(zhdr
);
632 z3fold_compact_page(zhdr
);
633 add_to_unbuddied(pool
, zhdr
);
634 z3fold_page_unlock(zhdr
);
637 static void compact_page_work(struct work_struct
*w
)
639 struct z3fold_header
*zhdr
= container_of(w
, struct z3fold_header
,
642 do_compact_page(zhdr
, false);
645 /* returns _locked_ z3fold page header or NULL */
646 static inline struct z3fold_header
*__z3fold_alloc(struct z3fold_pool
*pool
,
647 size_t size
, bool can_sleep
)
649 struct z3fold_header
*zhdr
= NULL
;
651 struct list_head
*unbuddied
;
652 int chunks
= size_to_chunks(size
), i
;
655 /* First, try to find an unbuddied z3fold page. */
656 unbuddied
= get_cpu_ptr(pool
->unbuddied
);
657 for_each_unbuddied_list(i
, chunks
) {
658 struct list_head
*l
= &unbuddied
[i
];
660 zhdr
= list_first_entry_or_null(READ_ONCE(l
),
661 struct z3fold_header
, buddy
);
666 /* Re-check under lock. */
667 spin_lock(&pool
->lock
);
669 if (unlikely(zhdr
!= list_first_entry(READ_ONCE(l
),
670 struct z3fold_header
, buddy
)) ||
671 !z3fold_page_trylock(zhdr
)) {
672 spin_unlock(&pool
->lock
);
674 put_cpu_ptr(pool
->unbuddied
);
679 list_del_init(&zhdr
->buddy
);
681 spin_unlock(&pool
->lock
);
683 page
= virt_to_page(zhdr
);
684 if (test_bit(NEEDS_COMPACTING
, &page
->private)) {
685 z3fold_page_unlock(zhdr
);
687 put_cpu_ptr(pool
->unbuddied
);
694 * this page could not be removed from its unbuddied
695 * list while pool lock was held, and then we've taken
696 * page lock so kref_put could not be called before
697 * we got here, so it's safe to just call kref_get()
699 kref_get(&zhdr
->refcount
);
702 put_cpu_ptr(pool
->unbuddied
);
707 /* look for _exact_ match on other cpus' lists */
708 for_each_online_cpu(cpu
) {
711 unbuddied
= per_cpu_ptr(pool
->unbuddied
, cpu
);
712 spin_lock(&pool
->lock
);
713 l
= &unbuddied
[chunks
];
715 zhdr
= list_first_entry_or_null(READ_ONCE(l
),
716 struct z3fold_header
, buddy
);
718 if (!zhdr
|| !z3fold_page_trylock(zhdr
)) {
719 spin_unlock(&pool
->lock
);
723 list_del_init(&zhdr
->buddy
);
725 spin_unlock(&pool
->lock
);
727 page
= virt_to_page(zhdr
);
728 if (test_bit(NEEDS_COMPACTING
, &page
->private)) {
729 z3fold_page_unlock(zhdr
);
735 kref_get(&zhdr
->refcount
);
748 * z3fold_create_pool() - create a new z3fold pool
750 * @gfp: gfp flags when allocating the z3fold pool structure
751 * @ops: user-defined operations for the z3fold pool
753 * Return: pointer to the new z3fold pool or NULL if the metadata allocation
756 static struct z3fold_pool
*z3fold_create_pool(const char *name
, gfp_t gfp
,
757 const struct z3fold_ops
*ops
)
759 struct z3fold_pool
*pool
= NULL
;
762 pool
= kzalloc(sizeof(struct z3fold_pool
), gfp
);
765 pool
->c_handle
= kmem_cache_create("z3fold_handle",
766 sizeof(struct z3fold_buddy_slots
),
767 SLOTS_ALIGN
, 0, NULL
);
770 spin_lock_init(&pool
->lock
);
771 spin_lock_init(&pool
->stale_lock
);
772 pool
->unbuddied
= __alloc_percpu(sizeof(struct list_head
)*NCHUNKS
, 2);
773 if (!pool
->unbuddied
)
775 for_each_possible_cpu(cpu
) {
776 struct list_head
*unbuddied
=
777 per_cpu_ptr(pool
->unbuddied
, cpu
);
778 for_each_unbuddied_list(i
, 0)
779 INIT_LIST_HEAD(&unbuddied
[i
]);
781 INIT_LIST_HEAD(&pool
->lru
);
782 INIT_LIST_HEAD(&pool
->stale
);
783 atomic64_set(&pool
->pages_nr
, 0);
785 pool
->compact_wq
= create_singlethread_workqueue(pool
->name
);
786 if (!pool
->compact_wq
)
788 pool
->release_wq
= create_singlethread_workqueue(pool
->name
);
789 if (!pool
->release_wq
)
791 if (z3fold_register_migration(pool
))
793 INIT_WORK(&pool
->work
, free_pages_work
);
798 destroy_workqueue(pool
->release_wq
);
800 destroy_workqueue(pool
->compact_wq
);
802 free_percpu(pool
->unbuddied
);
804 kmem_cache_destroy(pool
->c_handle
);
812 * z3fold_destroy_pool() - destroys an existing z3fold pool
813 * @pool: the z3fold pool to be destroyed
815 * The pool should be emptied before this function is called.
817 static void z3fold_destroy_pool(struct z3fold_pool
*pool
)
819 kmem_cache_destroy(pool
->c_handle
);
820 z3fold_unregister_migration(pool
);
821 destroy_workqueue(pool
->release_wq
);
822 destroy_workqueue(pool
->compact_wq
);
827 * z3fold_alloc() - allocates a region of a given size
828 * @pool: z3fold pool from which to allocate
829 * @size: size in bytes of the desired allocation
830 * @gfp: gfp flags used if the pool needs to grow
831 * @handle: handle of the new allocation
833 * This function will attempt to find a free region in the pool large enough to
834 * satisfy the allocation request. A search of the unbuddied lists is
835 * performed first. If no suitable free region is found, then a new page is
836 * allocated and added to the pool to satisfy the request.
838 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
839 * as z3fold pool pages.
841 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
842 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
845 static int z3fold_alloc(struct z3fold_pool
*pool
, size_t size
, gfp_t gfp
,
846 unsigned long *handle
)
848 int chunks
= size_to_chunks(size
);
849 struct z3fold_header
*zhdr
= NULL
;
850 struct page
*page
= NULL
;
852 bool can_sleep
= gfpflags_allow_blocking(gfp
);
857 if (size
> PAGE_SIZE
)
860 if (size
> PAGE_SIZE
- ZHDR_SIZE_ALIGNED
- CHUNK_SIZE
)
864 zhdr
= __z3fold_alloc(pool
, size
, can_sleep
);
866 if (zhdr
->first_chunks
== 0) {
867 if (zhdr
->middle_chunks
!= 0 &&
868 chunks
>= zhdr
->start_middle
)
872 } else if (zhdr
->last_chunks
== 0)
874 else if (zhdr
->middle_chunks
== 0)
877 if (kref_put(&zhdr
->refcount
,
878 release_z3fold_page_locked
))
879 atomic64_dec(&pool
->pages_nr
);
881 z3fold_page_unlock(zhdr
);
882 pr_err("No free chunks in unbuddied\n");
886 page
= virt_to_page(zhdr
);
894 spin_lock(&pool
->stale_lock
);
895 zhdr
= list_first_entry_or_null(&pool
->stale
,
896 struct z3fold_header
, buddy
);
898 * Before allocating a page, let's see if we can take one from
899 * the stale pages list. cancel_work_sync() can sleep so we
900 * limit this case to the contexts where we can sleep
903 list_del(&zhdr
->buddy
);
904 spin_unlock(&pool
->stale_lock
);
905 cancel_work_sync(&zhdr
->work
);
906 page
= virt_to_page(zhdr
);
908 spin_unlock(&pool
->stale_lock
);
912 page
= alloc_page(gfp
);
917 zhdr
= init_z3fold_page(page
, pool
, gfp
);
922 atomic64_inc(&pool
->pages_nr
);
924 if (bud
== HEADLESS
) {
925 set_bit(PAGE_HEADLESS
, &page
->private);
930 __SetPageMovable(page
, pool
->inode
->i_mapping
);
933 if (trylock_page(page
)) {
934 __SetPageMovable(page
, pool
->inode
->i_mapping
);
938 z3fold_page_lock(zhdr
);
942 zhdr
->first_chunks
= chunks
;
943 else if (bud
== LAST
)
944 zhdr
->last_chunks
= chunks
;
946 zhdr
->middle_chunks
= chunks
;
947 zhdr
->start_middle
= zhdr
->first_chunks
+ ZHDR_CHUNKS
;
949 add_to_unbuddied(pool
, zhdr
);
952 spin_lock(&pool
->lock
);
953 /* Add/move z3fold page to beginning of LRU */
954 if (!list_empty(&page
->lru
))
955 list_del(&page
->lru
);
957 list_add(&page
->lru
, &pool
->lru
);
959 *handle
= encode_handle(zhdr
, bud
);
960 spin_unlock(&pool
->lock
);
962 z3fold_page_unlock(zhdr
);
968 * z3fold_free() - frees the allocation associated with the given handle
969 * @pool: pool in which the allocation resided
970 * @handle: handle associated with the allocation returned by z3fold_alloc()
972 * In the case that the z3fold page in which the allocation resides is under
973 * reclaim, as indicated by the PG_reclaim flag being set, this function
974 * only sets the first|last_chunks to 0. The page is actually freed
975 * once both buddies are evicted (see z3fold_reclaim_page() below).
977 static void z3fold_free(struct z3fold_pool
*pool
, unsigned long handle
)
979 struct z3fold_header
*zhdr
;
983 zhdr
= handle_to_z3fold_header(handle
);
984 page
= virt_to_page(zhdr
);
986 if (test_bit(PAGE_HEADLESS
, &page
->private)) {
987 /* if a headless page is under reclaim, just leave.
988 * NB: we use test_and_set_bit for a reason: if the bit
989 * has not been set before, we release this page
990 * immediately so we don't care about its value any more.
992 if (!test_and_set_bit(PAGE_CLAIMED
, &page
->private)) {
993 spin_lock(&pool
->lock
);
994 list_del(&page
->lru
);
995 spin_unlock(&pool
->lock
);
996 free_z3fold_page(page
, true);
997 atomic64_dec(&pool
->pages_nr
);
1002 /* Non-headless case */
1003 z3fold_page_lock(zhdr
);
1004 bud
= handle_to_buddy(handle
);
1008 zhdr
->first_chunks
= 0;
1011 zhdr
->middle_chunks
= 0;
1014 zhdr
->last_chunks
= 0;
1017 pr_err("%s: unknown bud %d\n", __func__
, bud
);
1019 z3fold_page_unlock(zhdr
);
1023 free_handle(handle
);
1024 if (kref_put(&zhdr
->refcount
, release_z3fold_page_locked_list
)) {
1025 atomic64_dec(&pool
->pages_nr
);
1028 if (test_bit(PAGE_CLAIMED
, &page
->private)) {
1029 z3fold_page_unlock(zhdr
);
1032 if (unlikely(PageIsolated(page
)) ||
1033 test_and_set_bit(NEEDS_COMPACTING
, &page
->private)) {
1034 z3fold_page_unlock(zhdr
);
1037 if (zhdr
->cpu
< 0 || !cpu_online(zhdr
->cpu
)) {
1038 spin_lock(&pool
->lock
);
1039 list_del_init(&zhdr
->buddy
);
1040 spin_unlock(&pool
->lock
);
1042 kref_get(&zhdr
->refcount
);
1043 do_compact_page(zhdr
, true);
1046 kref_get(&zhdr
->refcount
);
1047 queue_work_on(zhdr
->cpu
, pool
->compact_wq
, &zhdr
->work
);
1048 z3fold_page_unlock(zhdr
);
1052 * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1053 * @pool: pool from which a page will attempt to be evicted
1054 * @retries: number of pages on the LRU list for which eviction will
1055 * be attempted before failing
1057 * z3fold reclaim is different from normal system reclaim in that it is done
1058 * from the bottom, up. This is because only the bottom layer, z3fold, has
1059 * information on how the allocations are organized within each z3fold page.
1060 * This has the potential to create interesting locking situations between
1061 * z3fold and the user, however.
1063 * To avoid these, this is how z3fold_reclaim_page() should be called:
1065 * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1066 * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1067 * call the user-defined eviction handler with the pool and handle as
1070 * If the handle can not be evicted, the eviction handler should return
1071 * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1072 * appropriate list and try the next z3fold page on the LRU up to
1073 * a user defined number of retries.
1075 * If the handle is successfully evicted, the eviction handler should
1076 * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1077 * contains logic to delay freeing the page if the page is under reclaim,
1078 * as indicated by the setting of the PG_reclaim flag on the underlying page.
1080 * If all buddies in the z3fold page are successfully evicted, then the
1081 * z3fold page can be freed.
1083 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1084 * no pages to evict or an eviction handler is not registered, -EAGAIN if
1085 * the retry limit was hit.
1087 static int z3fold_reclaim_page(struct z3fold_pool
*pool
, unsigned int retries
)
1090 struct z3fold_header
*zhdr
= NULL
;
1091 struct page
*page
= NULL
;
1092 struct list_head
*pos
;
1093 unsigned long first_handle
= 0, middle_handle
= 0, last_handle
= 0;
1095 spin_lock(&pool
->lock
);
1096 if (!pool
->ops
|| !pool
->ops
->evict
|| retries
== 0) {
1097 spin_unlock(&pool
->lock
);
1100 for (i
= 0; i
< retries
; i
++) {
1101 if (list_empty(&pool
->lru
)) {
1102 spin_unlock(&pool
->lock
);
1105 list_for_each_prev(pos
, &pool
->lru
) {
1106 page
= list_entry(pos
, struct page
, lru
);
1108 /* this bit could have been set by free, in which case
1109 * we pass over to the next page in the pool.
1111 if (test_and_set_bit(PAGE_CLAIMED
, &page
->private))
1114 if (unlikely(PageIsolated(page
)))
1116 if (test_bit(PAGE_HEADLESS
, &page
->private))
1119 zhdr
= page_address(page
);
1120 if (!z3fold_page_trylock(zhdr
)) {
1122 continue; /* can't evict at this point */
1124 kref_get(&zhdr
->refcount
);
1125 list_del_init(&zhdr
->buddy
);
1133 list_del_init(&page
->lru
);
1134 spin_unlock(&pool
->lock
);
1136 if (!test_bit(PAGE_HEADLESS
, &page
->private)) {
1138 * We need encode the handles before unlocking, since
1139 * we can race with free that will set
1140 * (first|last)_chunks to 0
1145 if (zhdr
->first_chunks
)
1146 first_handle
= encode_handle(zhdr
, FIRST
);
1147 if (zhdr
->middle_chunks
)
1148 middle_handle
= encode_handle(zhdr
, MIDDLE
);
1149 if (zhdr
->last_chunks
)
1150 last_handle
= encode_handle(zhdr
, LAST
);
1152 * it's safe to unlock here because we hold a
1153 * reference to this page
1155 z3fold_page_unlock(zhdr
);
1157 first_handle
= encode_handle(zhdr
, HEADLESS
);
1158 last_handle
= middle_handle
= 0;
1161 /* Issue the eviction callback(s) */
1162 if (middle_handle
) {
1163 ret
= pool
->ops
->evict(pool
, middle_handle
);
1168 ret
= pool
->ops
->evict(pool
, first_handle
);
1173 ret
= pool
->ops
->evict(pool
, last_handle
);
1178 if (test_bit(PAGE_HEADLESS
, &page
->private)) {
1180 free_z3fold_page(page
, true);
1181 atomic64_dec(&pool
->pages_nr
);
1184 spin_lock(&pool
->lock
);
1185 list_add(&page
->lru
, &pool
->lru
);
1186 spin_unlock(&pool
->lock
);
1188 z3fold_page_lock(zhdr
);
1189 clear_bit(PAGE_CLAIMED
, &page
->private);
1190 if (kref_put(&zhdr
->refcount
,
1191 release_z3fold_page_locked
)) {
1192 atomic64_dec(&pool
->pages_nr
);
1196 * if we are here, the page is still not completely
1197 * free. Take the global pool lock then to be able
1198 * to add it back to the lru list
1200 spin_lock(&pool
->lock
);
1201 list_add(&page
->lru
, &pool
->lru
);
1202 spin_unlock(&pool
->lock
);
1203 z3fold_page_unlock(zhdr
);
1206 /* We started off locked to we need to lock the pool back */
1207 spin_lock(&pool
->lock
);
1209 spin_unlock(&pool
->lock
);
1214 * z3fold_map() - maps the allocation associated with the given handle
1215 * @pool: pool in which the allocation resides
1216 * @handle: handle associated with the allocation to be mapped
1218 * Extracts the buddy number from handle and constructs the pointer to the
1219 * correct starting chunk within the page.
1221 * Returns: a pointer to the mapped allocation
1223 static void *z3fold_map(struct z3fold_pool
*pool
, unsigned long handle
)
1225 struct z3fold_header
*zhdr
;
1230 zhdr
= handle_to_z3fold_header(handle
);
1232 page
= virt_to_page(zhdr
);
1234 if (test_bit(PAGE_HEADLESS
, &page
->private))
1237 z3fold_page_lock(zhdr
);
1238 buddy
= handle_to_buddy(handle
);
1241 addr
+= ZHDR_SIZE_ALIGNED
;
1244 addr
+= zhdr
->start_middle
<< CHUNK_SHIFT
;
1245 set_bit(MIDDLE_CHUNK_MAPPED
, &page
->private);
1248 addr
+= PAGE_SIZE
- (handle_to_chunks(handle
) << CHUNK_SHIFT
);
1251 pr_err("unknown buddy id %d\n", buddy
);
1258 zhdr
->mapped_count
++;
1259 z3fold_page_unlock(zhdr
);
1265 * z3fold_unmap() - unmaps the allocation associated with the given handle
1266 * @pool: pool in which the allocation resides
1267 * @handle: handle associated with the allocation to be unmapped
1269 static void z3fold_unmap(struct z3fold_pool
*pool
, unsigned long handle
)
1271 struct z3fold_header
*zhdr
;
1275 zhdr
= handle_to_z3fold_header(handle
);
1276 page
= virt_to_page(zhdr
);
1278 if (test_bit(PAGE_HEADLESS
, &page
->private))
1281 z3fold_page_lock(zhdr
);
1282 buddy
= handle_to_buddy(handle
);
1283 if (buddy
== MIDDLE
)
1284 clear_bit(MIDDLE_CHUNK_MAPPED
, &page
->private);
1285 zhdr
->mapped_count
--;
1286 z3fold_page_unlock(zhdr
);
1290 * z3fold_get_pool_size() - gets the z3fold pool size in pages
1291 * @pool: pool whose size is being queried
1293 * Returns: size in pages of the given pool.
1295 static u64
z3fold_get_pool_size(struct z3fold_pool
*pool
)
1297 return atomic64_read(&pool
->pages_nr
);
1300 static bool z3fold_page_isolate(struct page
*page
, isolate_mode_t mode
)
1302 struct z3fold_header
*zhdr
;
1303 struct z3fold_pool
*pool
;
1305 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1306 VM_BUG_ON_PAGE(PageIsolated(page
), page
);
1308 if (test_bit(PAGE_HEADLESS
, &page
->private))
1311 zhdr
= page_address(page
);
1312 z3fold_page_lock(zhdr
);
1313 if (test_bit(NEEDS_COMPACTING
, &page
->private) ||
1314 test_bit(PAGE_STALE
, &page
->private))
1317 pool
= zhdr_to_pool(zhdr
);
1319 if (zhdr
->mapped_count
== 0) {
1320 kref_get(&zhdr
->refcount
);
1321 if (!list_empty(&zhdr
->buddy
))
1322 list_del_init(&zhdr
->buddy
);
1323 spin_lock(&pool
->lock
);
1324 if (!list_empty(&page
->lru
))
1325 list_del(&page
->lru
);
1326 spin_unlock(&pool
->lock
);
1327 z3fold_page_unlock(zhdr
);
1331 z3fold_page_unlock(zhdr
);
1335 static int z3fold_page_migrate(struct address_space
*mapping
, struct page
*newpage
,
1336 struct page
*page
, enum migrate_mode mode
)
1338 struct z3fold_header
*zhdr
, *new_zhdr
;
1339 struct z3fold_pool
*pool
;
1340 struct address_space
*new_mapping
;
1342 VM_BUG_ON_PAGE(!PageMovable(page
), page
);
1343 VM_BUG_ON_PAGE(!PageIsolated(page
), page
);
1344 VM_BUG_ON_PAGE(!PageLocked(newpage
), newpage
);
1346 zhdr
= page_address(page
);
1347 pool
= zhdr_to_pool(zhdr
);
1349 if (!z3fold_page_trylock(zhdr
)) {
1352 if (zhdr
->mapped_count
!= 0) {
1353 z3fold_page_unlock(zhdr
);
1356 if (work_pending(&zhdr
->work
)) {
1357 z3fold_page_unlock(zhdr
);
1360 new_zhdr
= page_address(newpage
);
1361 memcpy(new_zhdr
, zhdr
, PAGE_SIZE
);
1362 newpage
->private = page
->private;
1364 z3fold_page_unlock(zhdr
);
1365 spin_lock_init(&new_zhdr
->page_lock
);
1366 INIT_WORK(&new_zhdr
->work
, compact_page_work
);
1368 * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1369 * so we only have to reinitialize it.
1371 INIT_LIST_HEAD(&new_zhdr
->buddy
);
1372 new_mapping
= page_mapping(page
);
1373 __ClearPageMovable(page
);
1374 ClearPagePrivate(page
);
1377 z3fold_page_lock(new_zhdr
);
1378 if (new_zhdr
->first_chunks
)
1379 encode_handle(new_zhdr
, FIRST
);
1380 if (new_zhdr
->last_chunks
)
1381 encode_handle(new_zhdr
, LAST
);
1382 if (new_zhdr
->middle_chunks
)
1383 encode_handle(new_zhdr
, MIDDLE
);
1384 set_bit(NEEDS_COMPACTING
, &newpage
->private);
1385 new_zhdr
->cpu
= smp_processor_id();
1386 spin_lock(&pool
->lock
);
1387 list_add(&newpage
->lru
, &pool
->lru
);
1388 spin_unlock(&pool
->lock
);
1389 __SetPageMovable(newpage
, new_mapping
);
1390 z3fold_page_unlock(new_zhdr
);
1392 queue_work_on(new_zhdr
->cpu
, pool
->compact_wq
, &new_zhdr
->work
);
1394 page_mapcount_reset(page
);
1399 static void z3fold_page_putback(struct page
*page
)
1401 struct z3fold_header
*zhdr
;
1402 struct z3fold_pool
*pool
;
1404 zhdr
= page_address(page
);
1405 pool
= zhdr_to_pool(zhdr
);
1407 z3fold_page_lock(zhdr
);
1408 if (!list_empty(&zhdr
->buddy
))
1409 list_del_init(&zhdr
->buddy
);
1410 INIT_LIST_HEAD(&page
->lru
);
1411 if (kref_put(&zhdr
->refcount
, release_z3fold_page_locked
)) {
1412 atomic64_dec(&pool
->pages_nr
);
1415 spin_lock(&pool
->lock
);
1416 list_add(&page
->lru
, &pool
->lru
);
1417 spin_unlock(&pool
->lock
);
1418 z3fold_page_unlock(zhdr
);
1421 static const struct address_space_operations z3fold_aops
= {
1422 .isolate_page
= z3fold_page_isolate
,
1423 .migratepage
= z3fold_page_migrate
,
1424 .putback_page
= z3fold_page_putback
,
1431 static int z3fold_zpool_evict(struct z3fold_pool
*pool
, unsigned long handle
)
1433 if (pool
->zpool
&& pool
->zpool_ops
&& pool
->zpool_ops
->evict
)
1434 return pool
->zpool_ops
->evict(pool
->zpool
, handle
);
1439 static const struct z3fold_ops z3fold_zpool_ops
= {
1440 .evict
= z3fold_zpool_evict
1443 static void *z3fold_zpool_create(const char *name
, gfp_t gfp
,
1444 const struct zpool_ops
*zpool_ops
,
1445 struct zpool
*zpool
)
1447 struct z3fold_pool
*pool
;
1449 pool
= z3fold_create_pool(name
, gfp
,
1450 zpool_ops
? &z3fold_zpool_ops
: NULL
);
1452 pool
->zpool
= zpool
;
1453 pool
->zpool_ops
= zpool_ops
;
1458 static void z3fold_zpool_destroy(void *pool
)
1460 z3fold_destroy_pool(pool
);
1463 static int z3fold_zpool_malloc(void *pool
, size_t size
, gfp_t gfp
,
1464 unsigned long *handle
)
1466 return z3fold_alloc(pool
, size
, gfp
, handle
);
1468 static void z3fold_zpool_free(void *pool
, unsigned long handle
)
1470 z3fold_free(pool
, handle
);
1473 static int z3fold_zpool_shrink(void *pool
, unsigned int pages
,
1474 unsigned int *reclaimed
)
1476 unsigned int total
= 0;
1479 while (total
< pages
) {
1480 ret
= z3fold_reclaim_page(pool
, 8);
1492 static void *z3fold_zpool_map(void *pool
, unsigned long handle
,
1493 enum zpool_mapmode mm
)
1495 return z3fold_map(pool
, handle
);
1497 static void z3fold_zpool_unmap(void *pool
, unsigned long handle
)
1499 z3fold_unmap(pool
, handle
);
1502 static u64
z3fold_zpool_total_size(void *pool
)
1504 return z3fold_get_pool_size(pool
) * PAGE_SIZE
;
1507 static struct zpool_driver z3fold_zpool_driver
= {
1509 .owner
= THIS_MODULE
,
1510 .create
= z3fold_zpool_create
,
1511 .destroy
= z3fold_zpool_destroy
,
1512 .malloc
= z3fold_zpool_malloc
,
1513 .free
= z3fold_zpool_free
,
1514 .shrink
= z3fold_zpool_shrink
,
1515 .map
= z3fold_zpool_map
,
1516 .unmap
= z3fold_zpool_unmap
,
1517 .total_size
= z3fold_zpool_total_size
,
1520 MODULE_ALIAS("zpool-z3fold");
1522 static int __init
init_z3fold(void)
1526 /* Make sure the z3fold header is not larger than the page size */
1527 BUILD_BUG_ON(ZHDR_SIZE_ALIGNED
> PAGE_SIZE
);
1528 ret
= z3fold_mount();
1532 zpool_register_driver(&z3fold_zpool_driver
);
1537 static void __exit
exit_z3fold(void)
1540 zpool_unregister_driver(&z3fold_zpool_driver
);
1543 module_init(init_z3fold
);
1544 module_exit(exit_z3fold
);
1546 MODULE_LICENSE("GPL");
1547 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1548 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");