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, bool headless,
299 struct z3fold_pool *pool, gfp_t gfp)
301 struct z3fold_header *zhdr = page_address(page);
302 struct z3fold_buddy_slots *slots;
304 INIT_LIST_HEAD(&page->lru);
305 clear_bit(PAGE_HEADLESS, &page->private);
306 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
307 clear_bit(NEEDS_COMPACTING, &page->private);
308 clear_bit(PAGE_STALE, &page->private);
309 clear_bit(PAGE_CLAIMED, &page->private);
313 slots = alloc_slots(pool, gfp);
317 spin_lock_init(&zhdr->page_lock);
318 kref_init(&zhdr->refcount);
319 zhdr->first_chunks = 0;
320 zhdr->middle_chunks = 0;
321 zhdr->last_chunks = 0;
323 zhdr->start_middle = 0;
327 INIT_LIST_HEAD(&zhdr->buddy);
328 INIT_WORK(&zhdr->work, compact_page_work);
332 /* Resets the struct page fields and frees the page */
333 static void free_z3fold_page(struct page *page, bool headless)
337 __ClearPageMovable(page);
340 ClearPagePrivate(page);
344 /* Lock a z3fold page */
345 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
347 spin_lock(&zhdr->page_lock);
350 /* Try to lock a z3fold page */
351 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
353 return spin_trylock(&zhdr->page_lock);
356 /* Unlock a z3fold page */
357 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
359 spin_unlock(&zhdr->page_lock);
362 /* Helper function to build the index */
363 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
365 return (bud + zhdr->first_num) & BUDDY_MASK;
369 * Encodes the handle of a particular buddy within a z3fold page
370 * Pool lock should be held as this function accesses first_num
372 static unsigned long __encode_handle(struct z3fold_header *zhdr,
373 struct z3fold_buddy_slots *slots,
376 unsigned long h = (unsigned long)zhdr;
380 * For a headless page, its handle is its pointer with the extra
381 * PAGE_HEADLESS bit set
384 return h | (1 << PAGE_HEADLESS);
386 /* otherwise, return pointer to encoded handle */
387 idx = __idx(zhdr, bud);
390 h |= (zhdr->last_chunks << BUDDY_SHIFT);
392 slots->slot[idx] = h;
393 return (unsigned long)&slots->slot[idx];
396 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
398 return __encode_handle(zhdr, zhdr->slots, bud);
401 /* Returns the z3fold page where a given handle is stored */
402 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
404 unsigned long addr = h;
406 if (!(addr & (1 << PAGE_HEADLESS)))
407 addr = *(unsigned long *)h;
409 return (struct z3fold_header *)(addr & PAGE_MASK);
412 /* only for LAST bud, returns zero otherwise */
413 static unsigned short handle_to_chunks(unsigned long handle)
415 unsigned long addr = *(unsigned long *)handle;
417 return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
421 * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
422 * but that doesn't matter. because the masking will result in the
423 * correct buddy number.
425 static enum buddy handle_to_buddy(unsigned long handle)
427 struct z3fold_header *zhdr;
430 WARN_ON(handle & (1 << PAGE_HEADLESS));
431 addr = *(unsigned long *)handle;
432 zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
433 return (addr - zhdr->first_num) & BUDDY_MASK;
436 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
441 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
443 struct page *page = virt_to_page(zhdr);
444 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
446 WARN_ON(!list_empty(&zhdr->buddy));
447 set_bit(PAGE_STALE, &page->private);
448 clear_bit(NEEDS_COMPACTING, &page->private);
449 spin_lock(&pool->lock);
450 if (!list_empty(&page->lru))
451 list_del_init(&page->lru);
452 spin_unlock(&pool->lock);
454 z3fold_page_unlock(zhdr);
455 spin_lock(&pool->stale_lock);
456 list_add(&zhdr->buddy, &pool->stale);
457 queue_work(pool->release_wq, &pool->work);
458 spin_unlock(&pool->stale_lock);
461 static void __attribute__((__unused__))
462 release_z3fold_page(struct kref *ref)
464 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
466 __release_z3fold_page(zhdr, false);
469 static void release_z3fold_page_locked(struct kref *ref)
471 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
473 WARN_ON(z3fold_page_trylock(zhdr));
474 __release_z3fold_page(zhdr, true);
477 static void release_z3fold_page_locked_list(struct kref *ref)
479 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
481 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
482 spin_lock(&pool->lock);
483 list_del_init(&zhdr->buddy);
484 spin_unlock(&pool->lock);
486 WARN_ON(z3fold_page_trylock(zhdr));
487 __release_z3fold_page(zhdr, true);
490 static void free_pages_work(struct work_struct *w)
492 struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
494 spin_lock(&pool->stale_lock);
495 while (!list_empty(&pool->stale)) {
496 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
497 struct z3fold_header, buddy);
498 struct page *page = virt_to_page(zhdr);
500 list_del(&zhdr->buddy);
501 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
503 spin_unlock(&pool->stale_lock);
504 cancel_work_sync(&zhdr->work);
505 free_z3fold_page(page, false);
507 spin_lock(&pool->stale_lock);
509 spin_unlock(&pool->stale_lock);
513 * Returns the number of free chunks in a z3fold page.
514 * NB: can't be used with HEADLESS pages.
516 static int num_free_chunks(struct z3fold_header *zhdr)
520 * If there is a middle object, pick up the bigger free space
521 * either before or after it. Otherwise just subtract the number
522 * of chunks occupied by the first and the last objects.
524 if (zhdr->middle_chunks != 0) {
525 int nfree_before = zhdr->first_chunks ?
526 0 : zhdr->start_middle - ZHDR_CHUNKS;
527 int nfree_after = zhdr->last_chunks ?
529 (zhdr->start_middle + zhdr->middle_chunks);
530 nfree = max(nfree_before, nfree_after);
532 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
536 /* Add to the appropriate unbuddied list */
537 static inline void add_to_unbuddied(struct z3fold_pool *pool,
538 struct z3fold_header *zhdr)
540 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
541 zhdr->middle_chunks == 0) {
542 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
544 int freechunks = num_free_chunks(zhdr);
545 spin_lock(&pool->lock);
546 list_add(&zhdr->buddy, &unbuddied[freechunks]);
547 spin_unlock(&pool->lock);
548 zhdr->cpu = smp_processor_id();
549 put_cpu_ptr(pool->unbuddied);
553 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
554 unsigned short dst_chunk)
557 return memmove(beg + (dst_chunk << CHUNK_SHIFT),
558 beg + (zhdr->start_middle << CHUNK_SHIFT),
559 zhdr->middle_chunks << CHUNK_SHIFT);
562 #define BIG_CHUNK_GAP 3
563 /* Has to be called with lock held */
564 static int z3fold_compact_page(struct z3fold_header *zhdr)
566 struct page *page = virt_to_page(zhdr);
568 if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
569 return 0; /* can't move middle chunk, it's used */
571 if (unlikely(PageIsolated(page)))
574 if (zhdr->middle_chunks == 0)
575 return 0; /* nothing to compact */
577 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
578 /* move to the beginning */
579 mchunk_memmove(zhdr, ZHDR_CHUNKS);
580 zhdr->first_chunks = zhdr->middle_chunks;
581 zhdr->middle_chunks = 0;
582 zhdr->start_middle = 0;
588 * moving data is expensive, so let's only do that if
589 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
591 if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
592 zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
594 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
595 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
597 } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
598 TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
599 + zhdr->middle_chunks) >=
601 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
603 mchunk_memmove(zhdr, new_start);
604 zhdr->start_middle = new_start;
611 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
613 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
616 page = virt_to_page(zhdr);
618 WARN_ON(z3fold_page_trylock(zhdr));
620 z3fold_page_lock(zhdr);
621 if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
622 z3fold_page_unlock(zhdr);
625 spin_lock(&pool->lock);
626 list_del_init(&zhdr->buddy);
627 spin_unlock(&pool->lock);
629 if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
630 atomic64_dec(&pool->pages_nr);
634 if (unlikely(PageIsolated(page) ||
635 test_bit(PAGE_CLAIMED, &page->private) ||
636 test_bit(PAGE_STALE, &page->private))) {
637 z3fold_page_unlock(zhdr);
641 z3fold_compact_page(zhdr);
642 add_to_unbuddied(pool, zhdr);
643 z3fold_page_unlock(zhdr);
646 static void compact_page_work(struct work_struct *w)
648 struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
651 do_compact_page(zhdr, false);
654 /* returns _locked_ z3fold page header or NULL */
655 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
656 size_t size, bool can_sleep)
658 struct z3fold_header *zhdr = NULL;
660 struct list_head *unbuddied;
661 int chunks = size_to_chunks(size), i;
664 /* First, try to find an unbuddied z3fold page. */
665 unbuddied = get_cpu_ptr(pool->unbuddied);
666 for_each_unbuddied_list(i, chunks) {
667 struct list_head *l = &unbuddied[i];
669 zhdr = list_first_entry_or_null(READ_ONCE(l),
670 struct z3fold_header, buddy);
675 /* Re-check under lock. */
676 spin_lock(&pool->lock);
678 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
679 struct z3fold_header, buddy)) ||
680 !z3fold_page_trylock(zhdr)) {
681 spin_unlock(&pool->lock);
683 put_cpu_ptr(pool->unbuddied);
688 list_del_init(&zhdr->buddy);
690 spin_unlock(&pool->lock);
692 page = virt_to_page(zhdr);
693 if (test_bit(NEEDS_COMPACTING, &page->private)) {
694 z3fold_page_unlock(zhdr);
696 put_cpu_ptr(pool->unbuddied);
703 * this page could not be removed from its unbuddied
704 * list while pool lock was held, and then we've taken
705 * page lock so kref_put could not be called before
706 * we got here, so it's safe to just call kref_get()
708 kref_get(&zhdr->refcount);
711 put_cpu_ptr(pool->unbuddied);
716 /* look for _exact_ match on other cpus' lists */
717 for_each_online_cpu(cpu) {
720 unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
721 spin_lock(&pool->lock);
722 l = &unbuddied[chunks];
724 zhdr = list_first_entry_or_null(READ_ONCE(l),
725 struct z3fold_header, buddy);
727 if (!zhdr || !z3fold_page_trylock(zhdr)) {
728 spin_unlock(&pool->lock);
732 list_del_init(&zhdr->buddy);
734 spin_unlock(&pool->lock);
736 page = virt_to_page(zhdr);
737 if (test_bit(NEEDS_COMPACTING, &page->private)) {
738 z3fold_page_unlock(zhdr);
744 kref_get(&zhdr->refcount);
757 * z3fold_create_pool() - create a new z3fold pool
759 * @gfp: gfp flags when allocating the z3fold pool structure
760 * @ops: user-defined operations for the z3fold pool
762 * Return: pointer to the new z3fold pool or NULL if the metadata allocation
765 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
766 const struct z3fold_ops *ops)
768 struct z3fold_pool *pool = NULL;
771 pool = kzalloc(sizeof(struct z3fold_pool), gfp);
774 pool->c_handle = kmem_cache_create("z3fold_handle",
775 sizeof(struct z3fold_buddy_slots),
776 SLOTS_ALIGN, 0, NULL);
779 spin_lock_init(&pool->lock);
780 spin_lock_init(&pool->stale_lock);
781 pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
782 if (!pool->unbuddied)
784 for_each_possible_cpu(cpu) {
785 struct list_head *unbuddied =
786 per_cpu_ptr(pool->unbuddied, cpu);
787 for_each_unbuddied_list(i, 0)
788 INIT_LIST_HEAD(&unbuddied[i]);
790 INIT_LIST_HEAD(&pool->lru);
791 INIT_LIST_HEAD(&pool->stale);
792 atomic64_set(&pool->pages_nr, 0);
794 pool->compact_wq = create_singlethread_workqueue(pool->name);
795 if (!pool->compact_wq)
797 pool->release_wq = create_singlethread_workqueue(pool->name);
798 if (!pool->release_wq)
800 if (z3fold_register_migration(pool))
802 INIT_WORK(&pool->work, free_pages_work);
807 destroy_workqueue(pool->release_wq);
809 destroy_workqueue(pool->compact_wq);
811 free_percpu(pool->unbuddied);
813 kmem_cache_destroy(pool->c_handle);
821 * z3fold_destroy_pool() - destroys an existing z3fold pool
822 * @pool: the z3fold pool to be destroyed
824 * The pool should be emptied before this function is called.
826 static void z3fold_destroy_pool(struct z3fold_pool *pool)
828 kmem_cache_destroy(pool->c_handle);
831 * We need to destroy pool->compact_wq before pool->release_wq,
832 * as any pending work on pool->compact_wq will call
833 * queue_work(pool->release_wq, &pool->work).
835 * There are still outstanding pages until both workqueues are drained,
836 * so we cannot unregister migration until then.
839 destroy_workqueue(pool->compact_wq);
840 destroy_workqueue(pool->release_wq);
841 z3fold_unregister_migration(pool);
846 * z3fold_alloc() - allocates a region of a given size
847 * @pool: z3fold pool from which to allocate
848 * @size: size in bytes of the desired allocation
849 * @gfp: gfp flags used if the pool needs to grow
850 * @handle: handle of the new allocation
852 * This function will attempt to find a free region in the pool large enough to
853 * satisfy the allocation request. A search of the unbuddied lists is
854 * performed first. If no suitable free region is found, then a new page is
855 * allocated and added to the pool to satisfy the request.
857 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
858 * as z3fold pool pages.
860 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
861 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
864 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
865 unsigned long *handle)
867 int chunks = size_to_chunks(size);
868 struct z3fold_header *zhdr = NULL;
869 struct page *page = NULL;
871 bool can_sleep = gfpflags_allow_blocking(gfp);
876 if (size > PAGE_SIZE)
879 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
883 zhdr = __z3fold_alloc(pool, size, can_sleep);
885 if (zhdr->first_chunks == 0) {
886 if (zhdr->middle_chunks != 0 &&
887 chunks >= zhdr->start_middle)
891 } else if (zhdr->last_chunks == 0)
893 else if (zhdr->middle_chunks == 0)
896 if (kref_put(&zhdr->refcount,
897 release_z3fold_page_locked))
898 atomic64_dec(&pool->pages_nr);
900 z3fold_page_unlock(zhdr);
901 pr_err("No free chunks in unbuddied\n");
905 page = virt_to_page(zhdr);
913 spin_lock(&pool->stale_lock);
914 zhdr = list_first_entry_or_null(&pool->stale,
915 struct z3fold_header, buddy);
917 * Before allocating a page, let's see if we can take one from
918 * the stale pages list. cancel_work_sync() can sleep so we
919 * limit this case to the contexts where we can sleep
922 list_del(&zhdr->buddy);
923 spin_unlock(&pool->stale_lock);
924 cancel_work_sync(&zhdr->work);
925 page = virt_to_page(zhdr);
927 spin_unlock(&pool->stale_lock);
931 page = alloc_page(gfp);
936 zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
941 atomic64_inc(&pool->pages_nr);
943 if (bud == HEADLESS) {
944 set_bit(PAGE_HEADLESS, &page->private);
949 __SetPageMovable(page, pool->inode->i_mapping);
952 if (trylock_page(page)) {
953 __SetPageMovable(page, pool->inode->i_mapping);
957 z3fold_page_lock(zhdr);
961 zhdr->first_chunks = chunks;
962 else if (bud == LAST)
963 zhdr->last_chunks = chunks;
965 zhdr->middle_chunks = chunks;
966 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
968 add_to_unbuddied(pool, zhdr);
971 spin_lock(&pool->lock);
972 /* Add/move z3fold page to beginning of LRU */
973 if (!list_empty(&page->lru))
974 list_del(&page->lru);
976 list_add(&page->lru, &pool->lru);
978 *handle = encode_handle(zhdr, bud);
979 spin_unlock(&pool->lock);
981 z3fold_page_unlock(zhdr);
987 * z3fold_free() - frees the allocation associated with the given handle
988 * @pool: pool in which the allocation resided
989 * @handle: handle associated with the allocation returned by z3fold_alloc()
991 * In the case that the z3fold page in which the allocation resides is under
992 * reclaim, as indicated by the PG_reclaim flag being set, this function
993 * only sets the first|last_chunks to 0. The page is actually freed
994 * once both buddies are evicted (see z3fold_reclaim_page() below).
996 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
998 struct z3fold_header *zhdr;
1003 zhdr = handle_to_z3fold_header(handle);
1004 page = virt_to_page(zhdr);
1005 page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1007 if (test_bit(PAGE_HEADLESS, &page->private)) {
1008 /* if a headless page is under reclaim, just leave.
1009 * NB: we use test_and_set_bit for a reason: if the bit
1010 * has not been set before, we release this page
1011 * immediately so we don't care about its value any more.
1013 if (!page_claimed) {
1014 spin_lock(&pool->lock);
1015 list_del(&page->lru);
1016 spin_unlock(&pool->lock);
1017 free_z3fold_page(page, true);
1018 atomic64_dec(&pool->pages_nr);
1023 /* Non-headless case */
1024 z3fold_page_lock(zhdr);
1025 bud = handle_to_buddy(handle);
1029 zhdr->first_chunks = 0;
1032 zhdr->middle_chunks = 0;
1035 zhdr->last_chunks = 0;
1038 pr_err("%s: unknown bud %d\n", __func__, bud);
1040 z3fold_page_unlock(zhdr);
1044 free_handle(handle);
1045 if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1046 atomic64_dec(&pool->pages_nr);
1050 /* the page has not been claimed by us */
1051 z3fold_page_unlock(zhdr);
1054 if (unlikely(PageIsolated(page)) ||
1055 test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1056 z3fold_page_unlock(zhdr);
1057 clear_bit(PAGE_CLAIMED, &page->private);
1060 if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1061 spin_lock(&pool->lock);
1062 list_del_init(&zhdr->buddy);
1063 spin_unlock(&pool->lock);
1065 kref_get(&zhdr->refcount);
1066 do_compact_page(zhdr, true);
1067 clear_bit(PAGE_CLAIMED, &page->private);
1070 kref_get(&zhdr->refcount);
1071 queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1072 clear_bit(PAGE_CLAIMED, &page->private);
1073 z3fold_page_unlock(zhdr);
1077 * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1078 * @pool: pool from which a page will attempt to be evicted
1079 * @retries: number of pages on the LRU list for which eviction will
1080 * be attempted before failing
1082 * z3fold reclaim is different from normal system reclaim in that it is done
1083 * from the bottom, up. This is because only the bottom layer, z3fold, has
1084 * information on how the allocations are organized within each z3fold page.
1085 * This has the potential to create interesting locking situations between
1086 * z3fold and the user, however.
1088 * To avoid these, this is how z3fold_reclaim_page() should be called:
1090 * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1091 * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1092 * call the user-defined eviction handler with the pool and handle as
1095 * If the handle can not be evicted, the eviction handler should return
1096 * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1097 * appropriate list and try the next z3fold page on the LRU up to
1098 * a user defined number of retries.
1100 * If the handle is successfully evicted, the eviction handler should
1101 * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1102 * contains logic to delay freeing the page if the page is under reclaim,
1103 * as indicated by the setting of the PG_reclaim flag on the underlying page.
1105 * If all buddies in the z3fold page are successfully evicted, then the
1106 * z3fold page can be freed.
1108 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1109 * no pages to evict or an eviction handler is not registered, -EAGAIN if
1110 * the retry limit was hit.
1112 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1115 struct z3fold_header *zhdr = NULL;
1116 struct page *page = NULL;
1117 struct list_head *pos;
1118 struct z3fold_buddy_slots slots;
1119 unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1121 spin_lock(&pool->lock);
1122 if (!pool->ops || !pool->ops->evict || retries == 0) {
1123 spin_unlock(&pool->lock);
1126 for (i = 0; i < retries; i++) {
1127 if (list_empty(&pool->lru)) {
1128 spin_unlock(&pool->lock);
1131 list_for_each_prev(pos, &pool->lru) {
1132 page = list_entry(pos, struct page, lru);
1134 /* this bit could have been set by free, in which case
1135 * we pass over to the next page in the pool.
1137 if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1142 if (unlikely(PageIsolated(page))) {
1143 clear_bit(PAGE_CLAIMED, &page->private);
1147 zhdr = page_address(page);
1148 if (test_bit(PAGE_HEADLESS, &page->private))
1151 if (!z3fold_page_trylock(zhdr)) {
1152 clear_bit(PAGE_CLAIMED, &page->private);
1154 continue; /* can't evict at this point */
1156 kref_get(&zhdr->refcount);
1157 list_del_init(&zhdr->buddy);
1165 list_del_init(&page->lru);
1166 spin_unlock(&pool->lock);
1168 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1170 * We need encode the handles before unlocking, and
1171 * use our local slots structure because z3fold_free
1172 * can zero out zhdr->slots and we can't do much
1178 if (zhdr->first_chunks)
1179 first_handle = __encode_handle(zhdr, &slots,
1181 if (zhdr->middle_chunks)
1182 middle_handle = __encode_handle(zhdr, &slots,
1184 if (zhdr->last_chunks)
1185 last_handle = __encode_handle(zhdr, &slots,
1188 * it's safe to unlock here because we hold a
1189 * reference to this page
1191 z3fold_page_unlock(zhdr);
1193 first_handle = __encode_handle(zhdr, &slots, HEADLESS);
1194 last_handle = middle_handle = 0;
1197 /* Issue the eviction callback(s) */
1198 if (middle_handle) {
1199 ret = pool->ops->evict(pool, middle_handle);
1204 ret = pool->ops->evict(pool, first_handle);
1209 ret = pool->ops->evict(pool, last_handle);
1214 if (test_bit(PAGE_HEADLESS, &page->private)) {
1216 free_z3fold_page(page, true);
1217 atomic64_dec(&pool->pages_nr);
1220 spin_lock(&pool->lock);
1221 list_add(&page->lru, &pool->lru);
1222 spin_unlock(&pool->lock);
1223 clear_bit(PAGE_CLAIMED, &page->private);
1225 z3fold_page_lock(zhdr);
1226 if (kref_put(&zhdr->refcount,
1227 release_z3fold_page_locked)) {
1228 atomic64_dec(&pool->pages_nr);
1232 * if we are here, the page is still not completely
1233 * free. Take the global pool lock then to be able
1234 * to add it back to the lru list
1236 spin_lock(&pool->lock);
1237 list_add(&page->lru, &pool->lru);
1238 spin_unlock(&pool->lock);
1239 z3fold_page_unlock(zhdr);
1240 clear_bit(PAGE_CLAIMED, &page->private);
1243 /* We started off locked to we need to lock the pool back */
1244 spin_lock(&pool->lock);
1246 spin_unlock(&pool->lock);
1251 * z3fold_map() - maps the allocation associated with the given handle
1252 * @pool: pool in which the allocation resides
1253 * @handle: handle associated with the allocation to be mapped
1255 * Extracts the buddy number from handle and constructs the pointer to the
1256 * correct starting chunk within the page.
1258 * Returns: a pointer to the mapped allocation
1260 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1262 struct z3fold_header *zhdr;
1267 zhdr = handle_to_z3fold_header(handle);
1269 page = virt_to_page(zhdr);
1271 if (test_bit(PAGE_HEADLESS, &page->private))
1274 z3fold_page_lock(zhdr);
1275 buddy = handle_to_buddy(handle);
1278 addr += ZHDR_SIZE_ALIGNED;
1281 addr += zhdr->start_middle << CHUNK_SHIFT;
1282 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1285 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1288 pr_err("unknown buddy id %d\n", buddy);
1295 zhdr->mapped_count++;
1296 z3fold_page_unlock(zhdr);
1302 * z3fold_unmap() - unmaps the allocation associated with the given handle
1303 * @pool: pool in which the allocation resides
1304 * @handle: handle associated with the allocation to be unmapped
1306 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1308 struct z3fold_header *zhdr;
1312 zhdr = handle_to_z3fold_header(handle);
1313 page = virt_to_page(zhdr);
1315 if (test_bit(PAGE_HEADLESS, &page->private))
1318 z3fold_page_lock(zhdr);
1319 buddy = handle_to_buddy(handle);
1320 if (buddy == MIDDLE)
1321 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1322 zhdr->mapped_count--;
1323 z3fold_page_unlock(zhdr);
1327 * z3fold_get_pool_size() - gets the z3fold pool size in pages
1328 * @pool: pool whose size is being queried
1330 * Returns: size in pages of the given pool.
1332 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1334 return atomic64_read(&pool->pages_nr);
1337 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1339 struct z3fold_header *zhdr;
1340 struct z3fold_pool *pool;
1342 VM_BUG_ON_PAGE(!PageMovable(page), page);
1343 VM_BUG_ON_PAGE(PageIsolated(page), page);
1345 if (test_bit(PAGE_HEADLESS, &page->private) ||
1346 test_bit(PAGE_CLAIMED, &page->private))
1349 zhdr = page_address(page);
1350 z3fold_page_lock(zhdr);
1351 if (test_bit(NEEDS_COMPACTING, &page->private) ||
1352 test_bit(PAGE_STALE, &page->private))
1355 pool = zhdr_to_pool(zhdr);
1357 if (zhdr->mapped_count == 0) {
1358 kref_get(&zhdr->refcount);
1359 if (!list_empty(&zhdr->buddy))
1360 list_del_init(&zhdr->buddy);
1361 spin_lock(&pool->lock);
1362 if (!list_empty(&page->lru))
1363 list_del(&page->lru);
1364 spin_unlock(&pool->lock);
1365 z3fold_page_unlock(zhdr);
1369 z3fold_page_unlock(zhdr);
1373 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1374 struct page *page, enum migrate_mode mode)
1376 struct z3fold_header *zhdr, *new_zhdr;
1377 struct z3fold_pool *pool;
1378 struct address_space *new_mapping;
1380 VM_BUG_ON_PAGE(!PageMovable(page), page);
1381 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1382 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1384 zhdr = page_address(page);
1385 pool = zhdr_to_pool(zhdr);
1387 if (!z3fold_page_trylock(zhdr)) {
1390 if (zhdr->mapped_count != 0) {
1391 z3fold_page_unlock(zhdr);
1394 if (work_pending(&zhdr->work)) {
1395 z3fold_page_unlock(zhdr);
1398 new_zhdr = page_address(newpage);
1399 memcpy(new_zhdr, zhdr, PAGE_SIZE);
1400 newpage->private = page->private;
1402 z3fold_page_unlock(zhdr);
1403 spin_lock_init(&new_zhdr->page_lock);
1404 INIT_WORK(&new_zhdr->work, compact_page_work);
1406 * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1407 * so we only have to reinitialize it.
1409 INIT_LIST_HEAD(&new_zhdr->buddy);
1410 new_mapping = page_mapping(page);
1411 __ClearPageMovable(page);
1412 ClearPagePrivate(page);
1415 z3fold_page_lock(new_zhdr);
1416 if (new_zhdr->first_chunks)
1417 encode_handle(new_zhdr, FIRST);
1418 if (new_zhdr->last_chunks)
1419 encode_handle(new_zhdr, LAST);
1420 if (new_zhdr->middle_chunks)
1421 encode_handle(new_zhdr, MIDDLE);
1422 set_bit(NEEDS_COMPACTING, &newpage->private);
1423 new_zhdr->cpu = smp_processor_id();
1424 spin_lock(&pool->lock);
1425 list_add(&newpage->lru, &pool->lru);
1426 spin_unlock(&pool->lock);
1427 __SetPageMovable(newpage, new_mapping);
1428 z3fold_page_unlock(new_zhdr);
1430 queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1432 page_mapcount_reset(page);
1437 static void z3fold_page_putback(struct page *page)
1439 struct z3fold_header *zhdr;
1440 struct z3fold_pool *pool;
1442 zhdr = page_address(page);
1443 pool = zhdr_to_pool(zhdr);
1445 z3fold_page_lock(zhdr);
1446 if (!list_empty(&zhdr->buddy))
1447 list_del_init(&zhdr->buddy);
1448 INIT_LIST_HEAD(&page->lru);
1449 if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1450 atomic64_dec(&pool->pages_nr);
1453 spin_lock(&pool->lock);
1454 list_add(&page->lru, &pool->lru);
1455 spin_unlock(&pool->lock);
1456 z3fold_page_unlock(zhdr);
1459 static const struct address_space_operations z3fold_aops = {
1460 .isolate_page = z3fold_page_isolate,
1461 .migratepage = z3fold_page_migrate,
1462 .putback_page = z3fold_page_putback,
1469 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1471 if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1472 return pool->zpool_ops->evict(pool->zpool, handle);
1477 static const struct z3fold_ops z3fold_zpool_ops = {
1478 .evict = z3fold_zpool_evict
1481 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1482 const struct zpool_ops *zpool_ops,
1483 struct zpool *zpool)
1485 struct z3fold_pool *pool;
1487 pool = z3fold_create_pool(name, gfp,
1488 zpool_ops ? &z3fold_zpool_ops : NULL);
1490 pool->zpool = zpool;
1491 pool->zpool_ops = zpool_ops;
1496 static void z3fold_zpool_destroy(void *pool)
1498 z3fold_destroy_pool(pool);
1501 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1502 unsigned long *handle)
1504 return z3fold_alloc(pool, size, gfp, handle);
1506 static void z3fold_zpool_free(void *pool, unsigned long handle)
1508 z3fold_free(pool, handle);
1511 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1512 unsigned int *reclaimed)
1514 unsigned int total = 0;
1517 while (total < pages) {
1518 ret = z3fold_reclaim_page(pool, 8);
1530 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1531 enum zpool_mapmode mm)
1533 return z3fold_map(pool, handle);
1535 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1537 z3fold_unmap(pool, handle);
1540 static u64 z3fold_zpool_total_size(void *pool)
1542 return z3fold_get_pool_size(pool) * PAGE_SIZE;
1545 static struct zpool_driver z3fold_zpool_driver = {
1547 .owner = THIS_MODULE,
1548 .create = z3fold_zpool_create,
1549 .destroy = z3fold_zpool_destroy,
1550 .malloc = z3fold_zpool_malloc,
1551 .free = z3fold_zpool_free,
1552 .shrink = z3fold_zpool_shrink,
1553 .map = z3fold_zpool_map,
1554 .unmap = z3fold_zpool_unmap,
1555 .total_size = z3fold_zpool_total_size,
1558 MODULE_ALIAS("zpool-z3fold");
1560 static int __init init_z3fold(void)
1564 /* Make sure the z3fold header is not larger than the page size */
1565 BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1566 ret = z3fold_mount();
1570 zpool_register_driver(&z3fold_zpool_driver);
1575 static void __exit exit_z3fold(void)
1578 zpool_unregister_driver(&z3fold_zpool_driver);
1581 module_init(init_z3fold);
1582 module_exit(exit_z3fold);
1584 MODULE_LICENSE("GPL");
1585 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1586 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");