ed19d98c9dcd1641c5e66d27d586e93f43fadce1
[muen/linux.git] / mm / z3fold.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * z3fold.c
4  *
5  * Author: Vitaly Wool <vitaly.wool@konsulko.com>
6  * Copyright (C) 2016, Sony Mobile Communications Inc.
7  *
8  * This implementation is based on zbud written by Seth Jennings.
9  *
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.
17  *
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.
20  *
21  * z3fold doesn't export any API and is meant to be used via zpool API.
22  */
23
24 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
25
26 #include <linux/atomic.h>
27 #include <linux/sched.h>
28 #include <linux/cpumask.h>
29 #include <linux/list.h>
30 #include <linux/mm.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>
39 #include <linux/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>
46
47 /*
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.
56  */
57 #define NCHUNKS_ORDER   6
58
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)
65
66 #define BUDDY_MASK      (0x3)
67 #define BUDDY_SHIFT     2
68 #define SLOTS_ALIGN     (0x40)
69
70 /*****************
71  * Structures
72 *****************/
73 struct z3fold_pool;
74 struct z3fold_ops {
75         int (*evict)(struct z3fold_pool *pool, unsigned long handle);
76 };
77
78 enum buddy {
79         HEADLESS = 0,
80         FIRST,
81         MIDDLE,
82         LAST,
83         BUDDIES_MAX = LAST
84 };
85
86 struct z3fold_buddy_slots {
87         /*
88          * we are using BUDDY_MASK in handle_to_buddy etc. so there should
89          * be enough slots to hold all possible variants
90          */
91         unsigned long slot[BUDDY_MASK + 1];
92         unsigned long pool; /* back link + flags */
93 };
94 #define HANDLE_FLAG_MASK        (0x03)
95
96 /*
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
100  *                      pool
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
112  */
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;
120         short cpu;
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;
127 };
128
129 /**
130  * struct z3fold_pool - stores metadata for each z3fold pool
131  * @name:       pool name
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
138  *              added buddy.
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
148  *
149  * This structure is allocated at pool creation time and maintains metadata
150  * pertaining to a particular z3fold pool.
151  */
152 struct z3fold_pool {
153         const char *name;
154         spinlock_t lock;
155         spinlock_t stale_lock;
156         struct list_head *unbuddied;
157         struct list_head lru;
158         struct list_head stale;
159         atomic64_t pages_nr;
160         struct kmem_cache *c_handle;
161         const struct z3fold_ops *ops;
162         struct zpool *zpool;
163         const struct zpool_ops *zpool_ops;
164         struct workqueue_struct *compact_wq;
165         struct workqueue_struct *release_wq;
166         struct work_struct work;
167         struct inode *inode;
168 };
169
170 /*
171  * Internal z3fold page flags
172  */
173 enum z3fold_page_flags {
174         PAGE_HEADLESS = 0,
175         MIDDLE_CHUNK_MAPPED,
176         NEEDS_COMPACTING,
177         PAGE_STALE,
178         PAGE_CLAIMED, /* by either reclaim or free */
179 };
180
181 /*****************
182  * Helpers
183 *****************/
184
185 /* Converts an allocation size in bytes to size in z3fold chunks */
186 static int size_to_chunks(size_t size)
187 {
188         return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
189 }
190
191 #define for_each_unbuddied_list(_iter, _begin) \
192         for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
193
194 static void compact_page_work(struct work_struct *w);
195
196 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
197                                                         gfp_t gfp)
198 {
199         struct z3fold_buddy_slots *slots;
200
201         slots = kmem_cache_alloc(pool->c_handle,
202                                  (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
203
204         if (slots) {
205                 memset(slots->slot, 0, sizeof(slots->slot));
206                 slots->pool = (unsigned long)pool;
207         }
208
209         return slots;
210 }
211
212 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
213 {
214         return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
215 }
216
217 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
218 {
219         return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
220 }
221
222 static inline void free_handle(unsigned long handle)
223 {
224         struct z3fold_buddy_slots *slots;
225         int i;
226         bool is_free;
227
228         if (handle & (1 << PAGE_HEADLESS))
229                 return;
230
231         WARN_ON(*(unsigned long *)handle == 0);
232         *(unsigned long *)handle = 0;
233         slots = handle_to_slots(handle);
234         is_free = true;
235         for (i = 0; i <= BUDDY_MASK; i++) {
236                 if (slots->slot[i]) {
237                         is_free = false;
238                         break;
239                 }
240         }
241
242         if (is_free) {
243                 struct z3fold_pool *pool = slots_to_pool(slots);
244
245                 kmem_cache_free(pool->c_handle, slots);
246         }
247 }
248
249 static int z3fold_init_fs_context(struct fs_context *fc)
250 {
251         return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
252 }
253
254 static struct file_system_type z3fold_fs = {
255         .name           = "z3fold",
256         .init_fs_context = z3fold_init_fs_context,
257         .kill_sb        = kill_anon_super,
258 };
259
260 static struct vfsmount *z3fold_mnt;
261 static int z3fold_mount(void)
262 {
263         int ret = 0;
264
265         z3fold_mnt = kern_mount(&z3fold_fs);
266         if (IS_ERR(z3fold_mnt))
267                 ret = PTR_ERR(z3fold_mnt);
268
269         return ret;
270 }
271
272 static void z3fold_unmount(void)
273 {
274         kern_unmount(z3fold_mnt);
275 }
276
277 static const struct address_space_operations z3fold_aops;
278 static int z3fold_register_migration(struct z3fold_pool *pool)
279 {
280         pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
281         if (IS_ERR(pool->inode)) {
282                 pool->inode = NULL;
283                 return 1;
284         }
285
286         pool->inode->i_mapping->private_data = pool;
287         pool->inode->i_mapping->a_ops = &z3fold_aops;
288         return 0;
289 }
290
291 static void z3fold_unregister_migration(struct z3fold_pool *pool)
292 {
293         if (pool->inode)
294                 iput(pool->inode);
295  }
296
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)
300 {
301         struct z3fold_header *zhdr = page_address(page);
302         struct z3fold_buddy_slots *slots = alloc_slots(pool, gfp);
303
304         if (!slots)
305                 return NULL;
306
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);
313
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;
319         zhdr->first_num = 0;
320         zhdr->start_middle = 0;
321         zhdr->cpu = -1;
322         zhdr->slots = slots;
323         zhdr->pool = pool;
324         INIT_LIST_HEAD(&zhdr->buddy);
325         INIT_WORK(&zhdr->work, compact_page_work);
326         return zhdr;
327 }
328
329 /* Resets the struct page fields and frees the page */
330 static void free_z3fold_page(struct page *page, bool headless)
331 {
332         if (!headless) {
333                 lock_page(page);
334                 __ClearPageMovable(page);
335                 unlock_page(page);
336         }
337         ClearPagePrivate(page);
338         __free_page(page);
339 }
340
341 /* Lock a z3fold page */
342 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
343 {
344         spin_lock(&zhdr->page_lock);
345 }
346
347 /* Try to lock a z3fold page */
348 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
349 {
350         return spin_trylock(&zhdr->page_lock);
351 }
352
353 /* Unlock a z3fold page */
354 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
355 {
356         spin_unlock(&zhdr->page_lock);
357 }
358
359 /* Helper function to build the index */
360 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
361 {
362         return (bud + zhdr->first_num) & BUDDY_MASK;
363 }
364
365 /*
366  * Encodes the handle of a particular buddy within a z3fold page
367  * Pool lock should be held as this function accesses first_num
368  */
369 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
370 {
371         struct z3fold_buddy_slots *slots;
372         unsigned long h = (unsigned long)zhdr;
373         int idx = 0;
374
375         /*
376          * For a headless page, its handle is its pointer with the extra
377          * PAGE_HEADLESS bit set
378          */
379         if (bud == HEADLESS)
380                 return h | (1 << PAGE_HEADLESS);
381
382         /* otherwise, return pointer to encoded handle */
383         idx = __idx(zhdr, bud);
384         h += idx;
385         if (bud == LAST)
386                 h |= (zhdr->last_chunks << BUDDY_SHIFT);
387
388         slots = zhdr->slots;
389         slots->slot[idx] = h;
390         return (unsigned long)&slots->slot[idx];
391 }
392
393 /* Returns the z3fold page where a given handle is stored */
394 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
395 {
396         unsigned long addr = h;
397
398         if (!(addr & (1 << PAGE_HEADLESS)))
399                 addr = *(unsigned long *)h;
400
401         return (struct z3fold_header *)(addr & PAGE_MASK);
402 }
403
404 /* only for LAST bud, returns zero otherwise */
405 static unsigned short handle_to_chunks(unsigned long handle)
406 {
407         unsigned long addr = *(unsigned long *)handle;
408
409         return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
410 }
411
412 /*
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.
416  */
417 static enum buddy handle_to_buddy(unsigned long handle)
418 {
419         struct z3fold_header *zhdr;
420         unsigned long addr;
421
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;
426 }
427
428 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
429 {
430         return zhdr->pool;
431 }
432
433 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
434 {
435         struct page *page = virt_to_page(zhdr);
436         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
437
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);
445         if (locked)
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);
451 }
452
453 static void __attribute__((__unused__))
454                         release_z3fold_page(struct kref *ref)
455 {
456         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
457                                                 refcount);
458         __release_z3fold_page(zhdr, false);
459 }
460
461 static void release_z3fold_page_locked(struct kref *ref)
462 {
463         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
464                                                 refcount);
465         WARN_ON(z3fold_page_trylock(zhdr));
466         __release_z3fold_page(zhdr, true);
467 }
468
469 static void release_z3fold_page_locked_list(struct kref *ref)
470 {
471         struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
472                                                refcount);
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);
477
478         WARN_ON(z3fold_page_trylock(zhdr));
479         __release_z3fold_page(zhdr, true);
480 }
481
482 static void free_pages_work(struct work_struct *w)
483 {
484         struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
485
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);
491
492                 list_del(&zhdr->buddy);
493                 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
494                         continue;
495                 spin_unlock(&pool->stale_lock);
496                 cancel_work_sync(&zhdr->work);
497                 free_z3fold_page(page, false);
498                 cond_resched();
499                 spin_lock(&pool->stale_lock);
500         }
501         spin_unlock(&pool->stale_lock);
502 }
503
504 /*
505  * Returns the number of free chunks in a z3fold page.
506  * NB: can't be used with HEADLESS pages.
507  */
508 static int num_free_chunks(struct z3fold_header *zhdr)
509 {
510         int nfree;
511         /*
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.
515          */
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 ?
520                         0 : TOTAL_CHUNKS -
521                                 (zhdr->start_middle + zhdr->middle_chunks);
522                 nfree = max(nfree_before, nfree_after);
523         } else
524                 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
525         return nfree;
526 }
527
528 /* Add to the appropriate unbuddied list */
529 static inline void add_to_unbuddied(struct z3fold_pool *pool,
530                                 struct z3fold_header *zhdr)
531 {
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);
535
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);
542         }
543 }
544
545 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
546                                 unsigned short dst_chunk)
547 {
548         void *beg = zhdr;
549         return memmove(beg + (dst_chunk << CHUNK_SHIFT),
550                        beg + (zhdr->start_middle << CHUNK_SHIFT),
551                        zhdr->middle_chunks << CHUNK_SHIFT);
552 }
553
554 #define BIG_CHUNK_GAP   3
555 /* Has to be called with lock held */
556 static int z3fold_compact_page(struct z3fold_header *zhdr)
557 {
558         struct page *page = virt_to_page(zhdr);
559
560         if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
561                 return 0; /* can't move middle chunk, it's used */
562
563         if (unlikely(PageIsolated(page)))
564                 return 0;
565
566         if (zhdr->middle_chunks == 0)
567                 return 0; /* nothing to compact */
568
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;
575                 zhdr->first_num++;
576                 return 1;
577         }
578
579         /*
580          * moving data is expensive, so let's only do that if
581          * there's substantial gain (at least BIG_CHUNK_GAP chunks)
582          */
583         if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
584             zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
585                         BIG_CHUNK_GAP) {
586                 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
587                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
588                 return 1;
589         } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
590                    TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
591                                         + zhdr->middle_chunks) >=
592                         BIG_CHUNK_GAP) {
593                 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
594                         zhdr->middle_chunks;
595                 mchunk_memmove(zhdr, new_start);
596                 zhdr->start_middle = new_start;
597                 return 1;
598         }
599
600         return 0;
601 }
602
603 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
604 {
605         struct z3fold_pool *pool = zhdr_to_pool(zhdr);
606         struct page *page;
607
608         page = virt_to_page(zhdr);
609         if (locked)
610                 WARN_ON(z3fold_page_trylock(zhdr));
611         else
612                 z3fold_page_lock(zhdr);
613         if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
614                 z3fold_page_unlock(zhdr);
615                 return;
616         }
617         spin_lock(&pool->lock);
618         list_del_init(&zhdr->buddy);
619         spin_unlock(&pool->lock);
620
621         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
622                 atomic64_dec(&pool->pages_nr);
623                 return;
624         }
625
626         if (unlikely(PageIsolated(page) ||
627                      test_bit(PAGE_STALE, &page->private))) {
628                 z3fold_page_unlock(zhdr);
629                 return;
630         }
631
632         z3fold_compact_page(zhdr);
633         add_to_unbuddied(pool, zhdr);
634         z3fold_page_unlock(zhdr);
635 }
636
637 static void compact_page_work(struct work_struct *w)
638 {
639         struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
640                                                 work);
641
642         do_compact_page(zhdr, false);
643 }
644
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)
648 {
649         struct z3fold_header *zhdr = NULL;
650         struct page *page;
651         struct list_head *unbuddied;
652         int chunks = size_to_chunks(size), i;
653
654 lookup:
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];
659
660                 zhdr = list_first_entry_or_null(READ_ONCE(l),
661                                         struct z3fold_header, buddy);
662
663                 if (!zhdr)
664                         continue;
665
666                 /* Re-check under lock. */
667                 spin_lock(&pool->lock);
668                 l = &unbuddied[i];
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);
673                         zhdr = NULL;
674                         put_cpu_ptr(pool->unbuddied);
675                         if (can_sleep)
676                                 cond_resched();
677                         goto lookup;
678                 }
679                 list_del_init(&zhdr->buddy);
680                 zhdr->cpu = -1;
681                 spin_unlock(&pool->lock);
682
683                 page = virt_to_page(zhdr);
684                 if (test_bit(NEEDS_COMPACTING, &page->private)) {
685                         z3fold_page_unlock(zhdr);
686                         zhdr = NULL;
687                         put_cpu_ptr(pool->unbuddied);
688                         if (can_sleep)
689                                 cond_resched();
690                         goto lookup;
691                 }
692
693                 /*
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()
698                  */
699                 kref_get(&zhdr->refcount);
700                 break;
701         }
702         put_cpu_ptr(pool->unbuddied);
703
704         if (!zhdr) {
705                 int cpu;
706
707                 /* look for _exact_ match on other cpus' lists */
708                 for_each_online_cpu(cpu) {
709                         struct list_head *l;
710
711                         unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
712                         spin_lock(&pool->lock);
713                         l = &unbuddied[chunks];
714
715                         zhdr = list_first_entry_or_null(READ_ONCE(l),
716                                                 struct z3fold_header, buddy);
717
718                         if (!zhdr || !z3fold_page_trylock(zhdr)) {
719                                 spin_unlock(&pool->lock);
720                                 zhdr = NULL;
721                                 continue;
722                         }
723                         list_del_init(&zhdr->buddy);
724                         zhdr->cpu = -1;
725                         spin_unlock(&pool->lock);
726
727                         page = virt_to_page(zhdr);
728                         if (test_bit(NEEDS_COMPACTING, &page->private)) {
729                                 z3fold_page_unlock(zhdr);
730                                 zhdr = NULL;
731                                 if (can_sleep)
732                                         cond_resched();
733                                 continue;
734                         }
735                         kref_get(&zhdr->refcount);
736                         break;
737                 }
738         }
739
740         return zhdr;
741 }
742
743 /*
744  * API Functions
745  */
746
747 /**
748  * z3fold_create_pool() - create a new z3fold pool
749  * @name:       pool name
750  * @gfp:        gfp flags when allocating the z3fold pool structure
751  * @ops:        user-defined operations for the z3fold pool
752  *
753  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
754  * failed.
755  */
756 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
757                 const struct z3fold_ops *ops)
758 {
759         struct z3fold_pool *pool = NULL;
760         int i, cpu;
761
762         pool = kzalloc(sizeof(struct z3fold_pool), gfp);
763         if (!pool)
764                 goto out;
765         pool->c_handle = kmem_cache_create("z3fold_handle",
766                                 sizeof(struct z3fold_buddy_slots),
767                                 SLOTS_ALIGN, 0, NULL);
768         if (!pool->c_handle)
769                 goto out_c;
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)
774                 goto out_pool;
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]);
780         }
781         INIT_LIST_HEAD(&pool->lru);
782         INIT_LIST_HEAD(&pool->stale);
783         atomic64_set(&pool->pages_nr, 0);
784         pool->name = name;
785         pool->compact_wq = create_singlethread_workqueue(pool->name);
786         if (!pool->compact_wq)
787                 goto out_unbuddied;
788         pool->release_wq = create_singlethread_workqueue(pool->name);
789         if (!pool->release_wq)
790                 goto out_wq;
791         if (z3fold_register_migration(pool))
792                 goto out_rwq;
793         INIT_WORK(&pool->work, free_pages_work);
794         pool->ops = ops;
795         return pool;
796
797 out_rwq:
798         destroy_workqueue(pool->release_wq);
799 out_wq:
800         destroy_workqueue(pool->compact_wq);
801 out_unbuddied:
802         free_percpu(pool->unbuddied);
803 out_pool:
804         kmem_cache_destroy(pool->c_handle);
805 out_c:
806         kfree(pool);
807 out:
808         return NULL;
809 }
810
811 /**
812  * z3fold_destroy_pool() - destroys an existing z3fold pool
813  * @pool:       the z3fold pool to be destroyed
814  *
815  * The pool should be emptied before this function is called.
816  */
817 static void z3fold_destroy_pool(struct z3fold_pool *pool)
818 {
819         kmem_cache_destroy(pool->c_handle);
820
821         /*
822          * We need to destroy pool->compact_wq before pool->release_wq,
823          * as any pending work on pool->compact_wq will call
824          * queue_work(pool->release_wq, &pool->work).
825          *
826          * There are still outstanding pages until both workqueues are drained,
827          * so we cannot unregister migration until then.
828          */
829
830         destroy_workqueue(pool->compact_wq);
831         destroy_workqueue(pool->release_wq);
832         z3fold_unregister_migration(pool);
833         kfree(pool);
834 }
835
836 /**
837  * z3fold_alloc() - allocates a region of a given size
838  * @pool:       z3fold pool from which to allocate
839  * @size:       size in bytes of the desired allocation
840  * @gfp:        gfp flags used if the pool needs to grow
841  * @handle:     handle of the new allocation
842  *
843  * This function will attempt to find a free region in the pool large enough to
844  * satisfy the allocation request.  A search of the unbuddied lists is
845  * performed first. If no suitable free region is found, then a new page is
846  * allocated and added to the pool to satisfy the request.
847  *
848  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
849  * as z3fold pool pages.
850  *
851  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
852  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
853  * a new page.
854  */
855 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
856                         unsigned long *handle)
857 {
858         int chunks = size_to_chunks(size);
859         struct z3fold_header *zhdr = NULL;
860         struct page *page = NULL;
861         enum buddy bud;
862         bool can_sleep = gfpflags_allow_blocking(gfp);
863
864         if (!size)
865                 return -EINVAL;
866
867         if (size > PAGE_SIZE)
868                 return -ENOSPC;
869
870         if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
871                 bud = HEADLESS;
872         else {
873 retry:
874                 zhdr = __z3fold_alloc(pool, size, can_sleep);
875                 if (zhdr) {
876                         if (zhdr->first_chunks == 0) {
877                                 if (zhdr->middle_chunks != 0 &&
878                                     chunks >= zhdr->start_middle)
879                                         bud = LAST;
880                                 else
881                                         bud = FIRST;
882                         } else if (zhdr->last_chunks == 0)
883                                 bud = LAST;
884                         else if (zhdr->middle_chunks == 0)
885                                 bud = MIDDLE;
886                         else {
887                                 if (kref_put(&zhdr->refcount,
888                                              release_z3fold_page_locked))
889                                         atomic64_dec(&pool->pages_nr);
890                                 else
891                                         z3fold_page_unlock(zhdr);
892                                 pr_err("No free chunks in unbuddied\n");
893                                 WARN_ON(1);
894                                 goto retry;
895                         }
896                         page = virt_to_page(zhdr);
897                         goto found;
898                 }
899                 bud = FIRST;
900         }
901
902         page = NULL;
903         if (can_sleep) {
904                 spin_lock(&pool->stale_lock);
905                 zhdr = list_first_entry_or_null(&pool->stale,
906                                                 struct z3fold_header, buddy);
907                 /*
908                  * Before allocating a page, let's see if we can take one from
909                  * the stale pages list. cancel_work_sync() can sleep so we
910                  * limit this case to the contexts where we can sleep
911                  */
912                 if (zhdr) {
913                         list_del(&zhdr->buddy);
914                         spin_unlock(&pool->stale_lock);
915                         cancel_work_sync(&zhdr->work);
916                         page = virt_to_page(zhdr);
917                 } else {
918                         spin_unlock(&pool->stale_lock);
919                 }
920         }
921         if (!page)
922                 page = alloc_page(gfp);
923
924         if (!page)
925                 return -ENOMEM;
926
927         zhdr = init_z3fold_page(page, pool, gfp);
928         if (!zhdr) {
929                 __free_page(page);
930                 return -ENOMEM;
931         }
932         atomic64_inc(&pool->pages_nr);
933
934         if (bud == HEADLESS) {
935                 set_bit(PAGE_HEADLESS, &page->private);
936                 goto headless;
937         }
938         if (can_sleep) {
939                 lock_page(page);
940                 __SetPageMovable(page, pool->inode->i_mapping);
941                 unlock_page(page);
942         } else {
943                 if (trylock_page(page)) {
944                         __SetPageMovable(page, pool->inode->i_mapping);
945                         unlock_page(page);
946                 }
947         }
948         z3fold_page_lock(zhdr);
949
950 found:
951         if (bud == FIRST)
952                 zhdr->first_chunks = chunks;
953         else if (bud == LAST)
954                 zhdr->last_chunks = chunks;
955         else {
956                 zhdr->middle_chunks = chunks;
957                 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
958         }
959         add_to_unbuddied(pool, zhdr);
960
961 headless:
962         spin_lock(&pool->lock);
963         /* Add/move z3fold page to beginning of LRU */
964         if (!list_empty(&page->lru))
965                 list_del(&page->lru);
966
967         list_add(&page->lru, &pool->lru);
968
969         *handle = encode_handle(zhdr, bud);
970         spin_unlock(&pool->lock);
971         if (bud != HEADLESS)
972                 z3fold_page_unlock(zhdr);
973
974         return 0;
975 }
976
977 /**
978  * z3fold_free() - frees the allocation associated with the given handle
979  * @pool:       pool in which the allocation resided
980  * @handle:     handle associated with the allocation returned by z3fold_alloc()
981  *
982  * In the case that the z3fold page in which the allocation resides is under
983  * reclaim, as indicated by the PG_reclaim flag being set, this function
984  * only sets the first|last_chunks to 0.  The page is actually freed
985  * once both buddies are evicted (see z3fold_reclaim_page() below).
986  */
987 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
988 {
989         struct z3fold_header *zhdr;
990         struct page *page;
991         enum buddy bud;
992
993         zhdr = handle_to_z3fold_header(handle);
994         page = virt_to_page(zhdr);
995
996         if (test_bit(PAGE_HEADLESS, &page->private)) {
997                 /* if a headless page is under reclaim, just leave.
998                  * NB: we use test_and_set_bit for a reason: if the bit
999                  * has not been set before, we release this page
1000                  * immediately so we don't care about its value any more.
1001                  */
1002                 if (!test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1003                         spin_lock(&pool->lock);
1004                         list_del(&page->lru);
1005                         spin_unlock(&pool->lock);
1006                         free_z3fold_page(page, true);
1007                         atomic64_dec(&pool->pages_nr);
1008                 }
1009                 return;
1010         }
1011
1012         /* Non-headless case */
1013         z3fold_page_lock(zhdr);
1014         bud = handle_to_buddy(handle);
1015
1016         switch (bud) {
1017         case FIRST:
1018                 zhdr->first_chunks = 0;
1019                 break;
1020         case MIDDLE:
1021                 zhdr->middle_chunks = 0;
1022                 break;
1023         case LAST:
1024                 zhdr->last_chunks = 0;
1025                 break;
1026         default:
1027                 pr_err("%s: unknown bud %d\n", __func__, bud);
1028                 WARN_ON(1);
1029                 z3fold_page_unlock(zhdr);
1030                 return;
1031         }
1032
1033         free_handle(handle);
1034         if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1035                 atomic64_dec(&pool->pages_nr);
1036                 return;
1037         }
1038         if (test_bit(PAGE_CLAIMED, &page->private)) {
1039                 z3fold_page_unlock(zhdr);
1040                 return;
1041         }
1042         if (unlikely(PageIsolated(page)) ||
1043             test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1044                 z3fold_page_unlock(zhdr);
1045                 return;
1046         }
1047         if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1048                 spin_lock(&pool->lock);
1049                 list_del_init(&zhdr->buddy);
1050                 spin_unlock(&pool->lock);
1051                 zhdr->cpu = -1;
1052                 kref_get(&zhdr->refcount);
1053                 do_compact_page(zhdr, true);
1054                 return;
1055         }
1056         kref_get(&zhdr->refcount);
1057         queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1058         z3fold_page_unlock(zhdr);
1059 }
1060
1061 /**
1062  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1063  * @pool:       pool from which a page will attempt to be evicted
1064  * @retries:    number of pages on the LRU list for which eviction will
1065  *              be attempted before failing
1066  *
1067  * z3fold reclaim is different from normal system reclaim in that it is done
1068  * from the bottom, up. This is because only the bottom layer, z3fold, has
1069  * information on how the allocations are organized within each z3fold page.
1070  * This has the potential to create interesting locking situations between
1071  * z3fold and the user, however.
1072  *
1073  * To avoid these, this is how z3fold_reclaim_page() should be called:
1074  *
1075  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1076  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1077  * call the user-defined eviction handler with the pool and handle as
1078  * arguments.
1079  *
1080  * If the handle can not be evicted, the eviction handler should return
1081  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1082  * appropriate list and try the next z3fold page on the LRU up to
1083  * a user defined number of retries.
1084  *
1085  * If the handle is successfully evicted, the eviction handler should
1086  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1087  * contains logic to delay freeing the page if the page is under reclaim,
1088  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1089  *
1090  * If all buddies in the z3fold page are successfully evicted, then the
1091  * z3fold page can be freed.
1092  *
1093  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1094  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1095  * the retry limit was hit.
1096  */
1097 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1098 {
1099         int i, ret = 0;
1100         struct z3fold_header *zhdr = NULL;
1101         struct page *page = NULL;
1102         struct list_head *pos;
1103         unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1104
1105         spin_lock(&pool->lock);
1106         if (!pool->ops || !pool->ops->evict || retries == 0) {
1107                 spin_unlock(&pool->lock);
1108                 return -EINVAL;
1109         }
1110         for (i = 0; i < retries; i++) {
1111                 if (list_empty(&pool->lru)) {
1112                         spin_unlock(&pool->lock);
1113                         return -EINVAL;
1114                 }
1115                 list_for_each_prev(pos, &pool->lru) {
1116                         page = list_entry(pos, struct page, lru);
1117
1118                         /* this bit could have been set by free, in which case
1119                          * we pass over to the next page in the pool.
1120                          */
1121                         if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1122                                 continue;
1123
1124                         if (unlikely(PageIsolated(page)))
1125                                 continue;
1126                         if (test_bit(PAGE_HEADLESS, &page->private))
1127                                 break;
1128
1129                         zhdr = page_address(page);
1130                         if (!z3fold_page_trylock(zhdr)) {
1131                                 zhdr = NULL;
1132                                 continue; /* can't evict at this point */
1133                         }
1134                         kref_get(&zhdr->refcount);
1135                         list_del_init(&zhdr->buddy);
1136                         zhdr->cpu = -1;
1137                         break;
1138                 }
1139
1140                 if (!zhdr)
1141                         break;
1142
1143                 list_del_init(&page->lru);
1144                 spin_unlock(&pool->lock);
1145
1146                 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1147                         /*
1148                          * We need encode the handles before unlocking, since
1149                          * we can race with free that will set
1150                          * (first|last)_chunks to 0
1151                          */
1152                         first_handle = 0;
1153                         last_handle = 0;
1154                         middle_handle = 0;
1155                         if (zhdr->first_chunks)
1156                                 first_handle = encode_handle(zhdr, FIRST);
1157                         if (zhdr->middle_chunks)
1158                                 middle_handle = encode_handle(zhdr, MIDDLE);
1159                         if (zhdr->last_chunks)
1160                                 last_handle = encode_handle(zhdr, LAST);
1161                         /*
1162                          * it's safe to unlock here because we hold a
1163                          * reference to this page
1164                          */
1165                         z3fold_page_unlock(zhdr);
1166                 } else {
1167                         first_handle = encode_handle(zhdr, HEADLESS);
1168                         last_handle = middle_handle = 0;
1169                 }
1170
1171                 /* Issue the eviction callback(s) */
1172                 if (middle_handle) {
1173                         ret = pool->ops->evict(pool, middle_handle);
1174                         if (ret)
1175                                 goto next;
1176                 }
1177                 if (first_handle) {
1178                         ret = pool->ops->evict(pool, first_handle);
1179                         if (ret)
1180                                 goto next;
1181                 }
1182                 if (last_handle) {
1183                         ret = pool->ops->evict(pool, last_handle);
1184                         if (ret)
1185                                 goto next;
1186                 }
1187 next:
1188                 if (test_bit(PAGE_HEADLESS, &page->private)) {
1189                         if (ret == 0) {
1190                                 free_z3fold_page(page, true);
1191                                 atomic64_dec(&pool->pages_nr);
1192                                 return 0;
1193                         }
1194                         spin_lock(&pool->lock);
1195                         list_add(&page->lru, &pool->lru);
1196                         spin_unlock(&pool->lock);
1197                 } else {
1198                         z3fold_page_lock(zhdr);
1199                         clear_bit(PAGE_CLAIMED, &page->private);
1200                         if (kref_put(&zhdr->refcount,
1201                                         release_z3fold_page_locked)) {
1202                                 atomic64_dec(&pool->pages_nr);
1203                                 return 0;
1204                         }
1205                         /*
1206                          * if we are here, the page is still not completely
1207                          * free. Take the global pool lock then to be able
1208                          * to add it back to the lru list
1209                          */
1210                         spin_lock(&pool->lock);
1211                         list_add(&page->lru, &pool->lru);
1212                         spin_unlock(&pool->lock);
1213                         z3fold_page_unlock(zhdr);
1214                 }
1215
1216                 /* We started off locked to we need to lock the pool back */
1217                 spin_lock(&pool->lock);
1218         }
1219         spin_unlock(&pool->lock);
1220         return -EAGAIN;
1221 }
1222
1223 /**
1224  * z3fold_map() - maps the allocation associated with the given handle
1225  * @pool:       pool in which the allocation resides
1226  * @handle:     handle associated with the allocation to be mapped
1227  *
1228  * Extracts the buddy number from handle and constructs the pointer to the
1229  * correct starting chunk within the page.
1230  *
1231  * Returns: a pointer to the mapped allocation
1232  */
1233 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1234 {
1235         struct z3fold_header *zhdr;
1236         struct page *page;
1237         void *addr;
1238         enum buddy buddy;
1239
1240         zhdr = handle_to_z3fold_header(handle);
1241         addr = zhdr;
1242         page = virt_to_page(zhdr);
1243
1244         if (test_bit(PAGE_HEADLESS, &page->private))
1245                 goto out;
1246
1247         z3fold_page_lock(zhdr);
1248         buddy = handle_to_buddy(handle);
1249         switch (buddy) {
1250         case FIRST:
1251                 addr += ZHDR_SIZE_ALIGNED;
1252                 break;
1253         case MIDDLE:
1254                 addr += zhdr->start_middle << CHUNK_SHIFT;
1255                 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1256                 break;
1257         case LAST:
1258                 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1259                 break;
1260         default:
1261                 pr_err("unknown buddy id %d\n", buddy);
1262                 WARN_ON(1);
1263                 addr = NULL;
1264                 break;
1265         }
1266
1267         if (addr)
1268                 zhdr->mapped_count++;
1269         z3fold_page_unlock(zhdr);
1270 out:
1271         return addr;
1272 }
1273
1274 /**
1275  * z3fold_unmap() - unmaps the allocation associated with the given handle
1276  * @pool:       pool in which the allocation resides
1277  * @handle:     handle associated with the allocation to be unmapped
1278  */
1279 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1280 {
1281         struct z3fold_header *zhdr;
1282         struct page *page;
1283         enum buddy buddy;
1284
1285         zhdr = handle_to_z3fold_header(handle);
1286         page = virt_to_page(zhdr);
1287
1288         if (test_bit(PAGE_HEADLESS, &page->private))
1289                 return;
1290
1291         z3fold_page_lock(zhdr);
1292         buddy = handle_to_buddy(handle);
1293         if (buddy == MIDDLE)
1294                 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1295         zhdr->mapped_count--;
1296         z3fold_page_unlock(zhdr);
1297 }
1298
1299 /**
1300  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1301  * @pool:       pool whose size is being queried
1302  *
1303  * Returns: size in pages of the given pool.
1304  */
1305 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1306 {
1307         return atomic64_read(&pool->pages_nr);
1308 }
1309
1310 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1311 {
1312         struct z3fold_header *zhdr;
1313         struct z3fold_pool *pool;
1314
1315         VM_BUG_ON_PAGE(!PageMovable(page), page);
1316         VM_BUG_ON_PAGE(PageIsolated(page), page);
1317
1318         if (test_bit(PAGE_HEADLESS, &page->private))
1319                 return false;
1320
1321         zhdr = page_address(page);
1322         z3fold_page_lock(zhdr);
1323         if (test_bit(NEEDS_COMPACTING, &page->private) ||
1324             test_bit(PAGE_STALE, &page->private))
1325                 goto out;
1326
1327         pool = zhdr_to_pool(zhdr);
1328
1329         if (zhdr->mapped_count == 0) {
1330                 kref_get(&zhdr->refcount);
1331                 if (!list_empty(&zhdr->buddy))
1332                         list_del_init(&zhdr->buddy);
1333                 spin_lock(&pool->lock);
1334                 if (!list_empty(&page->lru))
1335                         list_del(&page->lru);
1336                 spin_unlock(&pool->lock);
1337                 z3fold_page_unlock(zhdr);
1338                 return true;
1339         }
1340 out:
1341         z3fold_page_unlock(zhdr);
1342         return false;
1343 }
1344
1345 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1346                                struct page *page, enum migrate_mode mode)
1347 {
1348         struct z3fold_header *zhdr, *new_zhdr;
1349         struct z3fold_pool *pool;
1350         struct address_space *new_mapping;
1351
1352         VM_BUG_ON_PAGE(!PageMovable(page), page);
1353         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1354         VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1355
1356         zhdr = page_address(page);
1357         pool = zhdr_to_pool(zhdr);
1358
1359         if (!z3fold_page_trylock(zhdr)) {
1360                 return -EAGAIN;
1361         }
1362         if (zhdr->mapped_count != 0) {
1363                 z3fold_page_unlock(zhdr);
1364                 return -EBUSY;
1365         }
1366         if (work_pending(&zhdr->work)) {
1367                 z3fold_page_unlock(zhdr);
1368                 return -EAGAIN;
1369         }
1370         new_zhdr = page_address(newpage);
1371         memcpy(new_zhdr, zhdr, PAGE_SIZE);
1372         newpage->private = page->private;
1373         page->private = 0;
1374         z3fold_page_unlock(zhdr);
1375         spin_lock_init(&new_zhdr->page_lock);
1376         INIT_WORK(&new_zhdr->work, compact_page_work);
1377         /*
1378          * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1379          * so we only have to reinitialize it.
1380          */
1381         INIT_LIST_HEAD(&new_zhdr->buddy);
1382         new_mapping = page_mapping(page);
1383         __ClearPageMovable(page);
1384         ClearPagePrivate(page);
1385
1386         get_page(newpage);
1387         z3fold_page_lock(new_zhdr);
1388         if (new_zhdr->first_chunks)
1389                 encode_handle(new_zhdr, FIRST);
1390         if (new_zhdr->last_chunks)
1391                 encode_handle(new_zhdr, LAST);
1392         if (new_zhdr->middle_chunks)
1393                 encode_handle(new_zhdr, MIDDLE);
1394         set_bit(NEEDS_COMPACTING, &newpage->private);
1395         new_zhdr->cpu = smp_processor_id();
1396         spin_lock(&pool->lock);
1397         list_add(&newpage->lru, &pool->lru);
1398         spin_unlock(&pool->lock);
1399         __SetPageMovable(newpage, new_mapping);
1400         z3fold_page_unlock(new_zhdr);
1401
1402         queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1403
1404         page_mapcount_reset(page);
1405         put_page(page);
1406         return 0;
1407 }
1408
1409 static void z3fold_page_putback(struct page *page)
1410 {
1411         struct z3fold_header *zhdr;
1412         struct z3fold_pool *pool;
1413
1414         zhdr = page_address(page);
1415         pool = zhdr_to_pool(zhdr);
1416
1417         z3fold_page_lock(zhdr);
1418         if (!list_empty(&zhdr->buddy))
1419                 list_del_init(&zhdr->buddy);
1420         INIT_LIST_HEAD(&page->lru);
1421         if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1422                 atomic64_dec(&pool->pages_nr);
1423                 return;
1424         }
1425         spin_lock(&pool->lock);
1426         list_add(&page->lru, &pool->lru);
1427         spin_unlock(&pool->lock);
1428         z3fold_page_unlock(zhdr);
1429 }
1430
1431 static const struct address_space_operations z3fold_aops = {
1432         .isolate_page = z3fold_page_isolate,
1433         .migratepage = z3fold_page_migrate,
1434         .putback_page = z3fold_page_putback,
1435 };
1436
1437 /*****************
1438  * zpool
1439  ****************/
1440
1441 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1442 {
1443         if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1444                 return pool->zpool_ops->evict(pool->zpool, handle);
1445         else
1446                 return -ENOENT;
1447 }
1448
1449 static const struct z3fold_ops z3fold_zpool_ops = {
1450         .evict =        z3fold_zpool_evict
1451 };
1452
1453 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1454                                const struct zpool_ops *zpool_ops,
1455                                struct zpool *zpool)
1456 {
1457         struct z3fold_pool *pool;
1458
1459         pool = z3fold_create_pool(name, gfp,
1460                                 zpool_ops ? &z3fold_zpool_ops : NULL);
1461         if (pool) {
1462                 pool->zpool = zpool;
1463                 pool->zpool_ops = zpool_ops;
1464         }
1465         return pool;
1466 }
1467
1468 static void z3fold_zpool_destroy(void *pool)
1469 {
1470         z3fold_destroy_pool(pool);
1471 }
1472
1473 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1474                         unsigned long *handle)
1475 {
1476         return z3fold_alloc(pool, size, gfp, handle);
1477 }
1478 static void z3fold_zpool_free(void *pool, unsigned long handle)
1479 {
1480         z3fold_free(pool, handle);
1481 }
1482
1483 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1484                         unsigned int *reclaimed)
1485 {
1486         unsigned int total = 0;
1487         int ret = -EINVAL;
1488
1489         while (total < pages) {
1490                 ret = z3fold_reclaim_page(pool, 8);
1491                 if (ret < 0)
1492                         break;
1493                 total++;
1494         }
1495
1496         if (reclaimed)
1497                 *reclaimed = total;
1498
1499         return ret;
1500 }
1501
1502 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1503                         enum zpool_mapmode mm)
1504 {
1505         return z3fold_map(pool, handle);
1506 }
1507 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1508 {
1509         z3fold_unmap(pool, handle);
1510 }
1511
1512 static u64 z3fold_zpool_total_size(void *pool)
1513 {
1514         return z3fold_get_pool_size(pool) * PAGE_SIZE;
1515 }
1516
1517 static struct zpool_driver z3fold_zpool_driver = {
1518         .type =         "z3fold",
1519         .owner =        THIS_MODULE,
1520         .create =       z3fold_zpool_create,
1521         .destroy =      z3fold_zpool_destroy,
1522         .malloc =       z3fold_zpool_malloc,
1523         .free =         z3fold_zpool_free,
1524         .shrink =       z3fold_zpool_shrink,
1525         .map =          z3fold_zpool_map,
1526         .unmap =        z3fold_zpool_unmap,
1527         .total_size =   z3fold_zpool_total_size,
1528 };
1529
1530 MODULE_ALIAS("zpool-z3fold");
1531
1532 static int __init init_z3fold(void)
1533 {
1534         int ret;
1535
1536         /* Make sure the z3fold header is not larger than the page size */
1537         BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1538         ret = z3fold_mount();
1539         if (ret)
1540                 return ret;
1541
1542         zpool_register_driver(&z3fold_zpool_driver);
1543
1544         return 0;
1545 }
1546
1547 static void __exit exit_z3fold(void)
1548 {
1549         z3fold_unmount();
1550         zpool_unregister_driver(&z3fold_zpool_driver);
1551 }
1552
1553 module_init(init_z3fold);
1554 module_exit(exit_z3fold);
1555
1556 MODULE_LICENSE("GPL");
1557 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1558 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");