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,
370 struct z3fold_buddy_slots *slots,
373 unsigned long h = (unsigned long)zhdr;
377 * For a headless page, its handle is its pointer with the extra
378 * PAGE_HEADLESS bit set
381 return h | (1 << PAGE_HEADLESS);
383 /* otherwise, return pointer to encoded handle */
384 idx = __idx(zhdr, bud);
387 h |= (zhdr->last_chunks << BUDDY_SHIFT);
389 slots->slot[idx] = h;
390 return (unsigned long)&slots->slot[idx];
393 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
395 return __encode_handle(zhdr, zhdr->slots, bud);
398 /* Returns the z3fold page where a given handle is stored */
399 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
401 unsigned long addr = h;
403 if (!(addr & (1 << PAGE_HEADLESS)))
404 addr = *(unsigned long *)h;
406 return (struct z3fold_header *)(addr & PAGE_MASK);
409 /* only for LAST bud, returns zero otherwise */
410 static unsigned short handle_to_chunks(unsigned long handle)
412 unsigned long addr = *(unsigned long *)handle;
414 return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
418 * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
419 * but that doesn't matter. because the masking will result in the
420 * correct buddy number.
422 static enum buddy handle_to_buddy(unsigned long handle)
424 struct z3fold_header *zhdr;
427 WARN_ON(handle & (1 << PAGE_HEADLESS));
428 addr = *(unsigned long *)handle;
429 zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
430 return (addr - zhdr->first_num) & BUDDY_MASK;
433 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
438 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
440 struct page *page = virt_to_page(zhdr);
441 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
443 WARN_ON(!list_empty(&zhdr->buddy));
444 set_bit(PAGE_STALE, &page->private);
445 clear_bit(NEEDS_COMPACTING, &page->private);
446 spin_lock(&pool->lock);
447 if (!list_empty(&page->lru))
448 list_del_init(&page->lru);
449 spin_unlock(&pool->lock);
451 z3fold_page_unlock(zhdr);
452 spin_lock(&pool->stale_lock);
453 list_add(&zhdr->buddy, &pool->stale);
454 queue_work(pool->release_wq, &pool->work);
455 spin_unlock(&pool->stale_lock);
458 static void __attribute__((__unused__))
459 release_z3fold_page(struct kref *ref)
461 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
463 __release_z3fold_page(zhdr, false);
466 static void release_z3fold_page_locked(struct kref *ref)
468 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
470 WARN_ON(z3fold_page_trylock(zhdr));
471 __release_z3fold_page(zhdr, true);
474 static void release_z3fold_page_locked_list(struct kref *ref)
476 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
478 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
479 spin_lock(&pool->lock);
480 list_del_init(&zhdr->buddy);
481 spin_unlock(&pool->lock);
483 WARN_ON(z3fold_page_trylock(zhdr));
484 __release_z3fold_page(zhdr, true);
487 static void free_pages_work(struct work_struct *w)
489 struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
491 spin_lock(&pool->stale_lock);
492 while (!list_empty(&pool->stale)) {
493 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
494 struct z3fold_header, buddy);
495 struct page *page = virt_to_page(zhdr);
497 list_del(&zhdr->buddy);
498 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
500 spin_unlock(&pool->stale_lock);
501 cancel_work_sync(&zhdr->work);
502 free_z3fold_page(page, false);
504 spin_lock(&pool->stale_lock);
506 spin_unlock(&pool->stale_lock);
510 * Returns the number of free chunks in a z3fold page.
511 * NB: can't be used with HEADLESS pages.
513 static int num_free_chunks(struct z3fold_header *zhdr)
517 * If there is a middle object, pick up the bigger free space
518 * either before or after it. Otherwise just subtract the number
519 * of chunks occupied by the first and the last objects.
521 if (zhdr->middle_chunks != 0) {
522 int nfree_before = zhdr->first_chunks ?
523 0 : zhdr->start_middle - ZHDR_CHUNKS;
524 int nfree_after = zhdr->last_chunks ?
526 (zhdr->start_middle + zhdr->middle_chunks);
527 nfree = max(nfree_before, nfree_after);
529 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
533 /* Add to the appropriate unbuddied list */
534 static inline void add_to_unbuddied(struct z3fold_pool *pool,
535 struct z3fold_header *zhdr)
537 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
538 zhdr->middle_chunks == 0) {
539 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
541 int freechunks = num_free_chunks(zhdr);
542 spin_lock(&pool->lock);
543 list_add(&zhdr->buddy, &unbuddied[freechunks]);
544 spin_unlock(&pool->lock);
545 zhdr->cpu = smp_processor_id();
546 put_cpu_ptr(pool->unbuddied);
550 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
551 unsigned short dst_chunk)
554 return memmove(beg + (dst_chunk << CHUNK_SHIFT),
555 beg + (zhdr->start_middle << CHUNK_SHIFT),
556 zhdr->middle_chunks << CHUNK_SHIFT);
559 #define BIG_CHUNK_GAP 3
560 /* Has to be called with lock held */
561 static int z3fold_compact_page(struct z3fold_header *zhdr)
563 struct page *page = virt_to_page(zhdr);
565 if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
566 return 0; /* can't move middle chunk, it's used */
568 if (unlikely(PageIsolated(page)))
571 if (zhdr->middle_chunks == 0)
572 return 0; /* nothing to compact */
574 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
575 /* move to the beginning */
576 mchunk_memmove(zhdr, ZHDR_CHUNKS);
577 zhdr->first_chunks = zhdr->middle_chunks;
578 zhdr->middle_chunks = 0;
579 zhdr->start_middle = 0;
585 * moving data is expensive, so let's only do that if
586 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
588 if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
589 zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
591 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
592 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
594 } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
595 TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
596 + zhdr->middle_chunks) >=
598 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
600 mchunk_memmove(zhdr, new_start);
601 zhdr->start_middle = new_start;
608 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
610 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
613 page = virt_to_page(zhdr);
615 WARN_ON(z3fold_page_trylock(zhdr));
617 z3fold_page_lock(zhdr);
618 if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
619 z3fold_page_unlock(zhdr);
622 spin_lock(&pool->lock);
623 list_del_init(&zhdr->buddy);
624 spin_unlock(&pool->lock);
626 if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
627 atomic64_dec(&pool->pages_nr);
631 if (unlikely(PageIsolated(page) ||
632 test_bit(PAGE_CLAIMED, &page->private) ||
633 test_bit(PAGE_STALE, &page->private))) {
634 z3fold_page_unlock(zhdr);
638 z3fold_compact_page(zhdr);
639 add_to_unbuddied(pool, zhdr);
640 z3fold_page_unlock(zhdr);
643 static void compact_page_work(struct work_struct *w)
645 struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
648 do_compact_page(zhdr, false);
651 /* returns _locked_ z3fold page header or NULL */
652 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
653 size_t size, bool can_sleep)
655 struct z3fold_header *zhdr = NULL;
657 struct list_head *unbuddied;
658 int chunks = size_to_chunks(size), i;
661 /* First, try to find an unbuddied z3fold page. */
662 unbuddied = get_cpu_ptr(pool->unbuddied);
663 for_each_unbuddied_list(i, chunks) {
664 struct list_head *l = &unbuddied[i];
666 zhdr = list_first_entry_or_null(READ_ONCE(l),
667 struct z3fold_header, buddy);
672 /* Re-check under lock. */
673 spin_lock(&pool->lock);
675 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
676 struct z3fold_header, buddy)) ||
677 !z3fold_page_trylock(zhdr)) {
678 spin_unlock(&pool->lock);
680 put_cpu_ptr(pool->unbuddied);
685 list_del_init(&zhdr->buddy);
687 spin_unlock(&pool->lock);
689 page = virt_to_page(zhdr);
690 if (test_bit(NEEDS_COMPACTING, &page->private)) {
691 z3fold_page_unlock(zhdr);
693 put_cpu_ptr(pool->unbuddied);
700 * this page could not be removed from its unbuddied
701 * list while pool lock was held, and then we've taken
702 * page lock so kref_put could not be called before
703 * we got here, so it's safe to just call kref_get()
705 kref_get(&zhdr->refcount);
708 put_cpu_ptr(pool->unbuddied);
713 /* look for _exact_ match on other cpus' lists */
714 for_each_online_cpu(cpu) {
717 unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
718 spin_lock(&pool->lock);
719 l = &unbuddied[chunks];
721 zhdr = list_first_entry_or_null(READ_ONCE(l),
722 struct z3fold_header, buddy);
724 if (!zhdr || !z3fold_page_trylock(zhdr)) {
725 spin_unlock(&pool->lock);
729 list_del_init(&zhdr->buddy);
731 spin_unlock(&pool->lock);
733 page = virt_to_page(zhdr);
734 if (test_bit(NEEDS_COMPACTING, &page->private)) {
735 z3fold_page_unlock(zhdr);
741 kref_get(&zhdr->refcount);
754 * z3fold_create_pool() - create a new z3fold pool
756 * @gfp: gfp flags when allocating the z3fold pool structure
757 * @ops: user-defined operations for the z3fold pool
759 * Return: pointer to the new z3fold pool or NULL if the metadata allocation
762 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
763 const struct z3fold_ops *ops)
765 struct z3fold_pool *pool = NULL;
768 pool = kzalloc(sizeof(struct z3fold_pool), gfp);
771 pool->c_handle = kmem_cache_create("z3fold_handle",
772 sizeof(struct z3fold_buddy_slots),
773 SLOTS_ALIGN, 0, NULL);
776 spin_lock_init(&pool->lock);
777 spin_lock_init(&pool->stale_lock);
778 pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
779 if (!pool->unbuddied)
781 for_each_possible_cpu(cpu) {
782 struct list_head *unbuddied =
783 per_cpu_ptr(pool->unbuddied, cpu);
784 for_each_unbuddied_list(i, 0)
785 INIT_LIST_HEAD(&unbuddied[i]);
787 INIT_LIST_HEAD(&pool->lru);
788 INIT_LIST_HEAD(&pool->stale);
789 atomic64_set(&pool->pages_nr, 0);
791 pool->compact_wq = create_singlethread_workqueue(pool->name);
792 if (!pool->compact_wq)
794 pool->release_wq = create_singlethread_workqueue(pool->name);
795 if (!pool->release_wq)
797 if (z3fold_register_migration(pool))
799 INIT_WORK(&pool->work, free_pages_work);
804 destroy_workqueue(pool->release_wq);
806 destroy_workqueue(pool->compact_wq);
808 free_percpu(pool->unbuddied);
810 kmem_cache_destroy(pool->c_handle);
818 * z3fold_destroy_pool() - destroys an existing z3fold pool
819 * @pool: the z3fold pool to be destroyed
821 * The pool should be emptied before this function is called.
823 static void z3fold_destroy_pool(struct z3fold_pool *pool)
825 kmem_cache_destroy(pool->c_handle);
828 * We need to destroy pool->compact_wq before pool->release_wq,
829 * as any pending work on pool->compact_wq will call
830 * queue_work(pool->release_wq, &pool->work).
832 * There are still outstanding pages until both workqueues are drained,
833 * so we cannot unregister migration until then.
836 destroy_workqueue(pool->compact_wq);
837 destroy_workqueue(pool->release_wq);
838 z3fold_unregister_migration(pool);
843 * z3fold_alloc() - allocates a region of a given size
844 * @pool: z3fold pool from which to allocate
845 * @size: size in bytes of the desired allocation
846 * @gfp: gfp flags used if the pool needs to grow
847 * @handle: handle of the new allocation
849 * This function will attempt to find a free region in the pool large enough to
850 * satisfy the allocation request. A search of the unbuddied lists is
851 * performed first. If no suitable free region is found, then a new page is
852 * allocated and added to the pool to satisfy the request.
854 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
855 * as z3fold pool pages.
857 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
858 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
861 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
862 unsigned long *handle)
864 int chunks = size_to_chunks(size);
865 struct z3fold_header *zhdr = NULL;
866 struct page *page = NULL;
868 bool can_sleep = gfpflags_allow_blocking(gfp);
873 if (size > PAGE_SIZE)
876 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
880 zhdr = __z3fold_alloc(pool, size, can_sleep);
882 if (zhdr->first_chunks == 0) {
883 if (zhdr->middle_chunks != 0 &&
884 chunks >= zhdr->start_middle)
888 } else if (zhdr->last_chunks == 0)
890 else if (zhdr->middle_chunks == 0)
893 if (kref_put(&zhdr->refcount,
894 release_z3fold_page_locked))
895 atomic64_dec(&pool->pages_nr);
897 z3fold_page_unlock(zhdr);
898 pr_err("No free chunks in unbuddied\n");
902 page = virt_to_page(zhdr);
910 spin_lock(&pool->stale_lock);
911 zhdr = list_first_entry_or_null(&pool->stale,
912 struct z3fold_header, buddy);
914 * Before allocating a page, let's see if we can take one from
915 * the stale pages list. cancel_work_sync() can sleep so we
916 * limit this case to the contexts where we can sleep
919 list_del(&zhdr->buddy);
920 spin_unlock(&pool->stale_lock);
921 cancel_work_sync(&zhdr->work);
922 page = virt_to_page(zhdr);
924 spin_unlock(&pool->stale_lock);
928 page = alloc_page(gfp);
933 zhdr = init_z3fold_page(page, pool, gfp);
938 atomic64_inc(&pool->pages_nr);
940 if (bud == HEADLESS) {
941 set_bit(PAGE_HEADLESS, &page->private);
946 __SetPageMovable(page, pool->inode->i_mapping);
949 if (trylock_page(page)) {
950 __SetPageMovable(page, pool->inode->i_mapping);
954 z3fold_page_lock(zhdr);
958 zhdr->first_chunks = chunks;
959 else if (bud == LAST)
960 zhdr->last_chunks = chunks;
962 zhdr->middle_chunks = chunks;
963 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
965 add_to_unbuddied(pool, zhdr);
968 spin_lock(&pool->lock);
969 /* Add/move z3fold page to beginning of LRU */
970 if (!list_empty(&page->lru))
971 list_del(&page->lru);
973 list_add(&page->lru, &pool->lru);
975 *handle = encode_handle(zhdr, bud);
976 spin_unlock(&pool->lock);
978 z3fold_page_unlock(zhdr);
984 * z3fold_free() - frees the allocation associated with the given handle
985 * @pool: pool in which the allocation resided
986 * @handle: handle associated with the allocation returned by z3fold_alloc()
988 * In the case that the z3fold page in which the allocation resides is under
989 * reclaim, as indicated by the PG_reclaim flag being set, this function
990 * only sets the first|last_chunks to 0. The page is actually freed
991 * once both buddies are evicted (see z3fold_reclaim_page() below).
993 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
995 struct z3fold_header *zhdr;
999 zhdr = handle_to_z3fold_header(handle);
1000 page = virt_to_page(zhdr);
1002 if (test_bit(PAGE_HEADLESS, &page->private)) {
1003 /* if a headless page is under reclaim, just leave.
1004 * NB: we use test_and_set_bit for a reason: if the bit
1005 * has not been set before, we release this page
1006 * immediately so we don't care about its value any more.
1008 if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1009 spin_lock(&pool->lock);
1010 list_del(&page->lru);
1011 spin_unlock(&pool->lock);
1012 free_z3fold_page(page, true);
1013 atomic64_dec(&pool->pages_nr);
1018 /* Non-headless case */
1019 z3fold_page_lock(zhdr);
1020 bud = handle_to_buddy(handle);
1024 zhdr->first_chunks = 0;
1027 zhdr->middle_chunks = 0;
1030 zhdr->last_chunks = 0;
1033 pr_err("%s: unknown bud %d\n", __func__, bud);
1035 z3fold_page_unlock(zhdr);
1039 free_handle(handle);
1040 if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1041 atomic64_dec(&pool->pages_nr);
1044 if (test_bit(PAGE_CLAIMED, &page->private)) {
1045 z3fold_page_unlock(zhdr);
1048 if (unlikely(PageIsolated(page)) ||
1049 test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1050 z3fold_page_unlock(zhdr);
1053 if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1054 spin_lock(&pool->lock);
1055 list_del_init(&zhdr->buddy);
1056 spin_unlock(&pool->lock);
1058 kref_get(&zhdr->refcount);
1059 do_compact_page(zhdr, true);
1062 kref_get(&zhdr->refcount);
1063 queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1064 z3fold_page_unlock(zhdr);
1068 * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1069 * @pool: pool from which a page will attempt to be evicted
1070 * @retries: number of pages on the LRU list for which eviction will
1071 * be attempted before failing
1073 * z3fold reclaim is different from normal system reclaim in that it is done
1074 * from the bottom, up. This is because only the bottom layer, z3fold, has
1075 * information on how the allocations are organized within each z3fold page.
1076 * This has the potential to create interesting locking situations between
1077 * z3fold and the user, however.
1079 * To avoid these, this is how z3fold_reclaim_page() should be called:
1081 * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1082 * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1083 * call the user-defined eviction handler with the pool and handle as
1086 * If the handle can not be evicted, the eviction handler should return
1087 * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1088 * appropriate list and try the next z3fold page on the LRU up to
1089 * a user defined number of retries.
1091 * If the handle is successfully evicted, the eviction handler should
1092 * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1093 * contains logic to delay freeing the page if the page is under reclaim,
1094 * as indicated by the setting of the PG_reclaim flag on the underlying page.
1096 * If all buddies in the z3fold page are successfully evicted, then the
1097 * z3fold page can be freed.
1099 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1100 * no pages to evict or an eviction handler is not registered, -EAGAIN if
1101 * the retry limit was hit.
1103 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1106 struct z3fold_header *zhdr = NULL;
1107 struct page *page = NULL;
1108 struct list_head *pos;
1109 struct z3fold_buddy_slots slots;
1110 unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1112 spin_lock(&pool->lock);
1113 if (!pool->ops || !pool->ops->evict || retries == 0) {
1114 spin_unlock(&pool->lock);
1117 for (i = 0; i < retries; i++) {
1118 if (list_empty(&pool->lru)) {
1119 spin_unlock(&pool->lock);
1122 list_for_each_prev(pos, &pool->lru) {
1123 page = list_entry(pos, struct page, lru);
1125 /* this bit could have been set by free, in which case
1126 * we pass over to the next page in the pool.
1128 if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1133 if (unlikely(PageIsolated(page))) {
1134 clear_bit(PAGE_CLAIMED, &page->private);
1138 zhdr = page_address(page);
1139 if (test_bit(PAGE_HEADLESS, &page->private))
1142 if (!z3fold_page_trylock(zhdr)) {
1143 clear_bit(PAGE_CLAIMED, &page->private);
1145 continue; /* can't evict at this point */
1147 kref_get(&zhdr->refcount);
1148 list_del_init(&zhdr->buddy);
1156 list_del_init(&page->lru);
1157 spin_unlock(&pool->lock);
1159 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1161 * We need encode the handles before unlocking, and
1162 * use our local slots structure because z3fold_free
1163 * can zero out zhdr->slots and we can't do much
1169 if (zhdr->first_chunks)
1170 first_handle = __encode_handle(zhdr, &slots,
1172 if (zhdr->middle_chunks)
1173 middle_handle = __encode_handle(zhdr, &slots,
1175 if (zhdr->last_chunks)
1176 last_handle = __encode_handle(zhdr, &slots,
1179 * it's safe to unlock here because we hold a
1180 * reference to this page
1182 z3fold_page_unlock(zhdr);
1184 first_handle = __encode_handle(zhdr, &slots, HEADLESS);
1185 last_handle = middle_handle = 0;
1188 /* Issue the eviction callback(s) */
1189 if (middle_handle) {
1190 ret = pool->ops->evict(pool, middle_handle);
1195 ret = pool->ops->evict(pool, first_handle);
1200 ret = pool->ops->evict(pool, last_handle);
1205 if (test_bit(PAGE_HEADLESS, &page->private)) {
1207 free_z3fold_page(page, true);
1208 atomic64_dec(&pool->pages_nr);
1211 spin_lock(&pool->lock);
1212 list_add(&page->lru, &pool->lru);
1213 spin_unlock(&pool->lock);
1214 clear_bit(PAGE_CLAIMED, &page->private);
1216 z3fold_page_lock(zhdr);
1217 if (kref_put(&zhdr->refcount,
1218 release_z3fold_page_locked)) {
1219 atomic64_dec(&pool->pages_nr);
1223 * if we are here, the page is still not completely
1224 * free. Take the global pool lock then to be able
1225 * to add it back to the lru list
1227 spin_lock(&pool->lock);
1228 list_add(&page->lru, &pool->lru);
1229 spin_unlock(&pool->lock);
1230 z3fold_page_unlock(zhdr);
1231 clear_bit(PAGE_CLAIMED, &page->private);
1234 /* We started off locked to we need to lock the pool back */
1235 spin_lock(&pool->lock);
1237 spin_unlock(&pool->lock);
1242 * z3fold_map() - maps the allocation associated with the given handle
1243 * @pool: pool in which the allocation resides
1244 * @handle: handle associated with the allocation to be mapped
1246 * Extracts the buddy number from handle and constructs the pointer to the
1247 * correct starting chunk within the page.
1249 * Returns: a pointer to the mapped allocation
1251 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1253 struct z3fold_header *zhdr;
1258 zhdr = handle_to_z3fold_header(handle);
1260 page = virt_to_page(zhdr);
1262 if (test_bit(PAGE_HEADLESS, &page->private))
1265 z3fold_page_lock(zhdr);
1266 buddy = handle_to_buddy(handle);
1269 addr += ZHDR_SIZE_ALIGNED;
1272 addr += zhdr->start_middle << CHUNK_SHIFT;
1273 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1276 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1279 pr_err("unknown buddy id %d\n", buddy);
1286 zhdr->mapped_count++;
1287 z3fold_page_unlock(zhdr);
1293 * z3fold_unmap() - unmaps the allocation associated with the given handle
1294 * @pool: pool in which the allocation resides
1295 * @handle: handle associated with the allocation to be unmapped
1297 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1299 struct z3fold_header *zhdr;
1303 zhdr = handle_to_z3fold_header(handle);
1304 page = virt_to_page(zhdr);
1306 if (test_bit(PAGE_HEADLESS, &page->private))
1309 z3fold_page_lock(zhdr);
1310 buddy = handle_to_buddy(handle);
1311 if (buddy == MIDDLE)
1312 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1313 zhdr->mapped_count--;
1314 z3fold_page_unlock(zhdr);
1318 * z3fold_get_pool_size() - gets the z3fold pool size in pages
1319 * @pool: pool whose size is being queried
1321 * Returns: size in pages of the given pool.
1323 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1325 return atomic64_read(&pool->pages_nr);
1328 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1330 struct z3fold_header *zhdr;
1331 struct z3fold_pool *pool;
1333 VM_BUG_ON_PAGE(!PageMovable(page), page);
1334 VM_BUG_ON_PAGE(PageIsolated(page), page);
1336 if (test_bit(PAGE_HEADLESS, &page->private) ||
1337 test_bit(PAGE_CLAIMED, &page->private))
1340 zhdr = page_address(page);
1341 z3fold_page_lock(zhdr);
1342 if (test_bit(NEEDS_COMPACTING, &page->private) ||
1343 test_bit(PAGE_STALE, &page->private))
1346 pool = zhdr_to_pool(zhdr);
1348 if (zhdr->mapped_count == 0) {
1349 kref_get(&zhdr->refcount);
1350 if (!list_empty(&zhdr->buddy))
1351 list_del_init(&zhdr->buddy);
1352 spin_lock(&pool->lock);
1353 if (!list_empty(&page->lru))
1354 list_del(&page->lru);
1355 spin_unlock(&pool->lock);
1356 z3fold_page_unlock(zhdr);
1360 z3fold_page_unlock(zhdr);
1364 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1365 struct page *page, enum migrate_mode mode)
1367 struct z3fold_header *zhdr, *new_zhdr;
1368 struct z3fold_pool *pool;
1369 struct address_space *new_mapping;
1371 VM_BUG_ON_PAGE(!PageMovable(page), page);
1372 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1373 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1375 zhdr = page_address(page);
1376 pool = zhdr_to_pool(zhdr);
1378 if (!z3fold_page_trylock(zhdr)) {
1381 if (zhdr->mapped_count != 0) {
1382 z3fold_page_unlock(zhdr);
1385 if (work_pending(&zhdr->work)) {
1386 z3fold_page_unlock(zhdr);
1389 new_zhdr = page_address(newpage);
1390 memcpy(new_zhdr, zhdr, PAGE_SIZE);
1391 newpage->private = page->private;
1393 z3fold_page_unlock(zhdr);
1394 spin_lock_init(&new_zhdr->page_lock);
1395 INIT_WORK(&new_zhdr->work, compact_page_work);
1397 * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1398 * so we only have to reinitialize it.
1400 INIT_LIST_HEAD(&new_zhdr->buddy);
1401 new_mapping = page_mapping(page);
1402 __ClearPageMovable(page);
1403 ClearPagePrivate(page);
1406 z3fold_page_lock(new_zhdr);
1407 if (new_zhdr->first_chunks)
1408 encode_handle(new_zhdr, FIRST);
1409 if (new_zhdr->last_chunks)
1410 encode_handle(new_zhdr, LAST);
1411 if (new_zhdr->middle_chunks)
1412 encode_handle(new_zhdr, MIDDLE);
1413 set_bit(NEEDS_COMPACTING, &newpage->private);
1414 new_zhdr->cpu = smp_processor_id();
1415 spin_lock(&pool->lock);
1416 list_add(&newpage->lru, &pool->lru);
1417 spin_unlock(&pool->lock);
1418 __SetPageMovable(newpage, new_mapping);
1419 z3fold_page_unlock(new_zhdr);
1421 queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1423 page_mapcount_reset(page);
1428 static void z3fold_page_putback(struct page *page)
1430 struct z3fold_header *zhdr;
1431 struct z3fold_pool *pool;
1433 zhdr = page_address(page);
1434 pool = zhdr_to_pool(zhdr);
1436 z3fold_page_lock(zhdr);
1437 if (!list_empty(&zhdr->buddy))
1438 list_del_init(&zhdr->buddy);
1439 INIT_LIST_HEAD(&page->lru);
1440 if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1441 atomic64_dec(&pool->pages_nr);
1444 spin_lock(&pool->lock);
1445 list_add(&page->lru, &pool->lru);
1446 spin_unlock(&pool->lock);
1447 z3fold_page_unlock(zhdr);
1450 static const struct address_space_operations z3fold_aops = {
1451 .isolate_page = z3fold_page_isolate,
1452 .migratepage = z3fold_page_migrate,
1453 .putback_page = z3fold_page_putback,
1460 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1462 if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1463 return pool->zpool_ops->evict(pool->zpool, handle);
1468 static const struct z3fold_ops z3fold_zpool_ops = {
1469 .evict = z3fold_zpool_evict
1472 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1473 const struct zpool_ops *zpool_ops,
1474 struct zpool *zpool)
1476 struct z3fold_pool *pool;
1478 pool = z3fold_create_pool(name, gfp,
1479 zpool_ops ? &z3fold_zpool_ops : NULL);
1481 pool->zpool = zpool;
1482 pool->zpool_ops = zpool_ops;
1487 static void z3fold_zpool_destroy(void *pool)
1489 z3fold_destroy_pool(pool);
1492 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1493 unsigned long *handle)
1495 return z3fold_alloc(pool, size, gfp, handle);
1497 static void z3fold_zpool_free(void *pool, unsigned long handle)
1499 z3fold_free(pool, handle);
1502 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1503 unsigned int *reclaimed)
1505 unsigned int total = 0;
1508 while (total < pages) {
1509 ret = z3fold_reclaim_page(pool, 8);
1521 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1522 enum zpool_mapmode mm)
1524 return z3fold_map(pool, handle);
1526 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1528 z3fold_unmap(pool, handle);
1531 static u64 z3fold_zpool_total_size(void *pool)
1533 return z3fold_get_pool_size(pool) * PAGE_SIZE;
1536 static struct zpool_driver z3fold_zpool_driver = {
1538 .owner = THIS_MODULE,
1539 .create = z3fold_zpool_create,
1540 .destroy = z3fold_zpool_destroy,
1541 .malloc = z3fold_zpool_malloc,
1542 .free = z3fold_zpool_free,
1543 .shrink = z3fold_zpool_shrink,
1544 .map = z3fold_zpool_map,
1545 .unmap = z3fold_zpool_unmap,
1546 .total_size = z3fold_zpool_total_size,
1549 MODULE_ALIAS("zpool-z3fold");
1551 static int __init init_z3fold(void)
1555 /* Make sure the z3fold header is not larger than the page size */
1556 BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1557 ret = z3fold_mount();
1561 zpool_register_driver(&z3fold_zpool_driver);
1566 static void __exit exit_z3fold(void)
1569 zpool_unregister_driver(&z3fold_zpool_driver);
1572 module_init(init_z3fold);
1573 module_exit(exit_z3fold);
1575 MODULE_LICENSE("GPL");
1576 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1577 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");