xref: /openbmc/linux/mm/z3fold.c (revision 67e2fae3)
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 #include <linux/kmemleak.h>
47 
48 /*
49  * NCHUNKS_ORDER determines the internal allocation granularity, effectively
50  * adjusting internal fragmentation.  It also determines the number of
51  * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
52  * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
53  * in the beginning of an allocated page are occupied by z3fold header, so
54  * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
55  * which shows the max number of free chunks in z3fold page, also there will
56  * be 63, or 62, respectively, freelists per pool.
57  */
58 #define NCHUNKS_ORDER	6
59 
60 #define CHUNK_SHIFT	(PAGE_SHIFT - NCHUNKS_ORDER)
61 #define CHUNK_SIZE	(1 << CHUNK_SHIFT)
62 #define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
63 #define ZHDR_CHUNKS	(ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
64 #define TOTAL_CHUNKS	(PAGE_SIZE >> CHUNK_SHIFT)
65 #define NCHUNKS		((PAGE_SIZE - ZHDR_SIZE_ALIGNED) >> CHUNK_SHIFT)
66 
67 #define BUDDY_MASK	(0x3)
68 #define BUDDY_SHIFT	2
69 #define SLOTS_ALIGN	(0x40)
70 
71 /*****************
72  * Structures
73 *****************/
74 struct z3fold_pool;
75 struct z3fold_ops {
76 	int (*evict)(struct z3fold_pool *pool, unsigned long handle);
77 };
78 
79 enum buddy {
80 	HEADLESS = 0,
81 	FIRST,
82 	MIDDLE,
83 	LAST,
84 	BUDDIES_MAX = LAST
85 };
86 
87 struct z3fold_buddy_slots {
88 	/*
89 	 * we are using BUDDY_MASK in handle_to_buddy etc. so there should
90 	 * be enough slots to hold all possible variants
91 	 */
92 	unsigned long slot[BUDDY_MASK + 1];
93 	unsigned long pool; /* back link + flags */
94 	rwlock_t lock;
95 };
96 #define HANDLE_FLAG_MASK	(0x03)
97 
98 /*
99  * struct z3fold_header - z3fold page metadata occupying first chunks of each
100  *			z3fold page, except for HEADLESS pages
101  * @buddy:		links the z3fold page into the relevant list in the
102  *			pool
103  * @page_lock:		per-page lock
104  * @refcount:		reference count for the z3fold page
105  * @work:		work_struct for page layout optimization
106  * @slots:		pointer to the structure holding buddy slots
107  * @pool:		pointer to the containing pool
108  * @cpu:		CPU which this page "belongs" to
109  * @first_chunks:	the size of the first buddy in chunks, 0 if free
110  * @middle_chunks:	the size of the middle buddy in chunks, 0 if free
111  * @last_chunks:	the size of the last buddy in chunks, 0 if free
112  * @first_num:		the starting number (for the first handle)
113  * @mapped_count:	the number of objects currently mapped
114  */
115 struct z3fold_header {
116 	struct list_head buddy;
117 	spinlock_t page_lock;
118 	struct kref refcount;
119 	struct work_struct work;
120 	struct z3fold_buddy_slots *slots;
121 	struct z3fold_pool *pool;
122 	short cpu;
123 	unsigned short first_chunks;
124 	unsigned short middle_chunks;
125 	unsigned short last_chunks;
126 	unsigned short start_middle;
127 	unsigned short first_num:2;
128 	unsigned short mapped_count:2;
129 	unsigned short foreign_handles:2;
130 };
131 
132 /**
133  * struct z3fold_pool - stores metadata for each z3fold pool
134  * @name:	pool name
135  * @lock:	protects pool unbuddied/lru lists
136  * @stale_lock:	protects pool stale page list
137  * @unbuddied:	per-cpu array of lists tracking z3fold pages that contain 2-
138  *		buddies; the list each z3fold page is added to depends on
139  *		the size of its free region.
140  * @lru:	list tracking the z3fold pages in LRU order by most recently
141  *		added buddy.
142  * @stale:	list of pages marked for freeing
143  * @pages_nr:	number of z3fold pages in the pool.
144  * @c_handle:	cache for z3fold_buddy_slots allocation
145  * @ops:	pointer to a structure of user defined operations specified at
146  *		pool creation time.
147  * @compact_wq:	workqueue for page layout background optimization
148  * @release_wq:	workqueue for safe page release
149  * @work:	work_struct for safe page release
150  * @inode:	inode for z3fold pseudo filesystem
151  *
152  * This structure is allocated at pool creation time and maintains metadata
153  * pertaining to a particular z3fold pool.
154  */
155 struct z3fold_pool {
156 	const char *name;
157 	spinlock_t lock;
158 	spinlock_t stale_lock;
159 	struct list_head *unbuddied;
160 	struct list_head lru;
161 	struct list_head stale;
162 	atomic64_t pages_nr;
163 	struct kmem_cache *c_handle;
164 	const struct z3fold_ops *ops;
165 	struct zpool *zpool;
166 	const struct zpool_ops *zpool_ops;
167 	struct workqueue_struct *compact_wq;
168 	struct workqueue_struct *release_wq;
169 	struct work_struct work;
170 	struct inode *inode;
171 };
172 
173 /*
174  * Internal z3fold page flags
175  */
176 enum z3fold_page_flags {
177 	PAGE_HEADLESS = 0,
178 	MIDDLE_CHUNK_MAPPED,
179 	NEEDS_COMPACTING,
180 	PAGE_STALE,
181 	PAGE_CLAIMED, /* by either reclaim or free */
182 };
183 
184 /*
185  * handle flags, go under HANDLE_FLAG_MASK
186  */
187 enum z3fold_handle_flags {
188 	HANDLES_ORPHANED = 0,
189 };
190 
191 /*
192  * Forward declarations
193  */
194 static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
195 static void compact_page_work(struct work_struct *w);
196 
197 /*****************
198  * Helpers
199 *****************/
200 
201 /* Converts an allocation size in bytes to size in z3fold chunks */
202 static int size_to_chunks(size_t size)
203 {
204 	return (size + CHUNK_SIZE - 1) >> CHUNK_SHIFT;
205 }
206 
207 #define for_each_unbuddied_list(_iter, _begin) \
208 	for ((_iter) = (_begin); (_iter) < NCHUNKS; (_iter)++)
209 
210 static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
211 							gfp_t gfp)
212 {
213 	struct z3fold_buddy_slots *slots;
214 
215 	slots = kmem_cache_alloc(pool->c_handle,
216 				 (gfp & ~(__GFP_HIGHMEM | __GFP_MOVABLE)));
217 
218 	if (slots) {
219 		/* It will be freed separately in free_handle(). */
220 		kmemleak_not_leak(slots);
221 		memset(slots->slot, 0, sizeof(slots->slot));
222 		slots->pool = (unsigned long)pool;
223 		rwlock_init(&slots->lock);
224 	}
225 
226 	return slots;
227 }
228 
229 static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
230 {
231 	return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
232 }
233 
234 static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
235 {
236 	return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
237 }
238 
239 /* Lock a z3fold page */
240 static inline void z3fold_page_lock(struct z3fold_header *zhdr)
241 {
242 	spin_lock(&zhdr->page_lock);
243 }
244 
245 /* Try to lock a z3fold page */
246 static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
247 {
248 	return spin_trylock(&zhdr->page_lock);
249 }
250 
251 /* Unlock a z3fold page */
252 static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
253 {
254 	spin_unlock(&zhdr->page_lock);
255 }
256 
257 
258 static inline struct z3fold_header *__get_z3fold_header(unsigned long handle,
259 							bool lock)
260 {
261 	struct z3fold_buddy_slots *slots;
262 	struct z3fold_header *zhdr;
263 	int locked = 0;
264 
265 	if (!(handle & (1 << PAGE_HEADLESS))) {
266 		slots = handle_to_slots(handle);
267 		do {
268 			unsigned long addr;
269 
270 			read_lock(&slots->lock);
271 			addr = *(unsigned long *)handle;
272 			zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
273 			if (lock)
274 				locked = z3fold_page_trylock(zhdr);
275 			read_unlock(&slots->lock);
276 			if (locked)
277 				break;
278 			cpu_relax();
279 		} while (lock);
280 	} else {
281 		zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
282 	}
283 
284 	return zhdr;
285 }
286 
287 /* Returns the z3fold page where a given handle is stored */
288 static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
289 {
290 	return __get_z3fold_header(h, false);
291 }
292 
293 /* return locked z3fold page if it's not headless */
294 static inline struct z3fold_header *get_z3fold_header(unsigned long h)
295 {
296 	return __get_z3fold_header(h, true);
297 }
298 
299 static inline void put_z3fold_header(struct z3fold_header *zhdr)
300 {
301 	struct page *page = virt_to_page(zhdr);
302 
303 	if (!test_bit(PAGE_HEADLESS, &page->private))
304 		z3fold_page_unlock(zhdr);
305 }
306 
307 static inline void free_handle(unsigned long handle)
308 {
309 	struct z3fold_buddy_slots *slots;
310 	struct z3fold_header *zhdr;
311 	int i;
312 	bool is_free;
313 
314 	if (handle & (1 << PAGE_HEADLESS))
315 		return;
316 
317 	if (WARN_ON(*(unsigned long *)handle == 0))
318 		return;
319 
320 	zhdr = handle_to_z3fold_header(handle);
321 	slots = handle_to_slots(handle);
322 	write_lock(&slots->lock);
323 	*(unsigned long *)handle = 0;
324 	if (zhdr->slots == slots) {
325 		write_unlock(&slots->lock);
326 		return; /* simple case, nothing else to do */
327 	}
328 
329 	/* we are freeing a foreign handle if we are here */
330 	zhdr->foreign_handles--;
331 	is_free = true;
332 	if (!test_bit(HANDLES_ORPHANED, &slots->pool)) {
333 		write_unlock(&slots->lock);
334 		return;
335 	}
336 	for (i = 0; i <= BUDDY_MASK; i++) {
337 		if (slots->slot[i]) {
338 			is_free = false;
339 			break;
340 		}
341 	}
342 	write_unlock(&slots->lock);
343 
344 	if (is_free) {
345 		struct z3fold_pool *pool = slots_to_pool(slots);
346 
347 		kmem_cache_free(pool->c_handle, slots);
348 	}
349 }
350 
351 static int z3fold_init_fs_context(struct fs_context *fc)
352 {
353 	return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
354 }
355 
356 static struct file_system_type z3fold_fs = {
357 	.name		= "z3fold",
358 	.init_fs_context = z3fold_init_fs_context,
359 	.kill_sb	= kill_anon_super,
360 };
361 
362 static struct vfsmount *z3fold_mnt;
363 static int z3fold_mount(void)
364 {
365 	int ret = 0;
366 
367 	z3fold_mnt = kern_mount(&z3fold_fs);
368 	if (IS_ERR(z3fold_mnt))
369 		ret = PTR_ERR(z3fold_mnt);
370 
371 	return ret;
372 }
373 
374 static void z3fold_unmount(void)
375 {
376 	kern_unmount(z3fold_mnt);
377 }
378 
379 static const struct address_space_operations z3fold_aops;
380 static int z3fold_register_migration(struct z3fold_pool *pool)
381 {
382 	pool->inode = alloc_anon_inode(z3fold_mnt->mnt_sb);
383 	if (IS_ERR(pool->inode)) {
384 		pool->inode = NULL;
385 		return 1;
386 	}
387 
388 	pool->inode->i_mapping->private_data = pool;
389 	pool->inode->i_mapping->a_ops = &z3fold_aops;
390 	return 0;
391 }
392 
393 static void z3fold_unregister_migration(struct z3fold_pool *pool)
394 {
395 	if (pool->inode)
396 		iput(pool->inode);
397  }
398 
399 /* Initializes the z3fold header of a newly allocated z3fold page */
400 static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
401 					struct z3fold_pool *pool, gfp_t gfp)
402 {
403 	struct z3fold_header *zhdr = page_address(page);
404 	struct z3fold_buddy_slots *slots;
405 
406 	INIT_LIST_HEAD(&page->lru);
407 	clear_bit(PAGE_HEADLESS, &page->private);
408 	clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
409 	clear_bit(NEEDS_COMPACTING, &page->private);
410 	clear_bit(PAGE_STALE, &page->private);
411 	clear_bit(PAGE_CLAIMED, &page->private);
412 	if (headless)
413 		return zhdr;
414 
415 	slots = alloc_slots(pool, gfp);
416 	if (!slots)
417 		return NULL;
418 
419 	spin_lock_init(&zhdr->page_lock);
420 	kref_init(&zhdr->refcount);
421 	zhdr->first_chunks = 0;
422 	zhdr->middle_chunks = 0;
423 	zhdr->last_chunks = 0;
424 	zhdr->first_num = 0;
425 	zhdr->start_middle = 0;
426 	zhdr->cpu = -1;
427 	zhdr->foreign_handles = 0;
428 	zhdr->mapped_count = 0;
429 	zhdr->slots = slots;
430 	zhdr->pool = pool;
431 	INIT_LIST_HEAD(&zhdr->buddy);
432 	INIT_WORK(&zhdr->work, compact_page_work);
433 	return zhdr;
434 }
435 
436 /* Resets the struct page fields and frees the page */
437 static void free_z3fold_page(struct page *page, bool headless)
438 {
439 	if (!headless) {
440 		lock_page(page);
441 		__ClearPageMovable(page);
442 		unlock_page(page);
443 	}
444 	ClearPagePrivate(page);
445 	__free_page(page);
446 }
447 
448 /* Helper function to build the index */
449 static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
450 {
451 	return (bud + zhdr->first_num) & BUDDY_MASK;
452 }
453 
454 /*
455  * Encodes the handle of a particular buddy within a z3fold page
456  * Pool lock should be held as this function accesses first_num
457  */
458 static unsigned long __encode_handle(struct z3fold_header *zhdr,
459 				struct z3fold_buddy_slots *slots,
460 				enum buddy bud)
461 {
462 	unsigned long h = (unsigned long)zhdr;
463 	int idx = 0;
464 
465 	/*
466 	 * For a headless page, its handle is its pointer with the extra
467 	 * PAGE_HEADLESS bit set
468 	 */
469 	if (bud == HEADLESS)
470 		return h | (1 << PAGE_HEADLESS);
471 
472 	/* otherwise, return pointer to encoded handle */
473 	idx = __idx(zhdr, bud);
474 	h += idx;
475 	if (bud == LAST)
476 		h |= (zhdr->last_chunks << BUDDY_SHIFT);
477 
478 	write_lock(&slots->lock);
479 	slots->slot[idx] = h;
480 	write_unlock(&slots->lock);
481 	return (unsigned long)&slots->slot[idx];
482 }
483 
484 static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
485 {
486 	return __encode_handle(zhdr, zhdr->slots, bud);
487 }
488 
489 /* only for LAST bud, returns zero otherwise */
490 static unsigned short handle_to_chunks(unsigned long handle)
491 {
492 	struct z3fold_buddy_slots *slots = handle_to_slots(handle);
493 	unsigned long addr;
494 
495 	read_lock(&slots->lock);
496 	addr = *(unsigned long *)handle;
497 	read_unlock(&slots->lock);
498 	return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
499 }
500 
501 /*
502  * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
503  *  but that doesn't matter. because the masking will result in the
504  *  correct buddy number.
505  */
506 static enum buddy handle_to_buddy(unsigned long handle)
507 {
508 	struct z3fold_header *zhdr;
509 	struct z3fold_buddy_slots *slots = handle_to_slots(handle);
510 	unsigned long addr;
511 
512 	read_lock(&slots->lock);
513 	WARN_ON(handle & (1 << PAGE_HEADLESS));
514 	addr = *(unsigned long *)handle;
515 	read_unlock(&slots->lock);
516 	zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
517 	return (addr - zhdr->first_num) & BUDDY_MASK;
518 }
519 
520 static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
521 {
522 	return zhdr->pool;
523 }
524 
525 static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
526 {
527 	struct page *page = virt_to_page(zhdr);
528 	struct z3fold_pool *pool = zhdr_to_pool(zhdr);
529 	bool is_free = true;
530 	int i;
531 
532 	WARN_ON(!list_empty(&zhdr->buddy));
533 	set_bit(PAGE_STALE, &page->private);
534 	clear_bit(NEEDS_COMPACTING, &page->private);
535 	spin_lock(&pool->lock);
536 	if (!list_empty(&page->lru))
537 		list_del_init(&page->lru);
538 	spin_unlock(&pool->lock);
539 
540 	/* If there are no foreign handles, free the handles array */
541 	read_lock(&zhdr->slots->lock);
542 	for (i = 0; i <= BUDDY_MASK; i++) {
543 		if (zhdr->slots->slot[i]) {
544 			is_free = false;
545 			break;
546 		}
547 	}
548 	if (!is_free)
549 		set_bit(HANDLES_ORPHANED, &zhdr->slots->pool);
550 	read_unlock(&zhdr->slots->lock);
551 
552 	if (is_free)
553 		kmem_cache_free(pool->c_handle, zhdr->slots);
554 
555 	if (locked)
556 		z3fold_page_unlock(zhdr);
557 
558 	spin_lock(&pool->stale_lock);
559 	list_add(&zhdr->buddy, &pool->stale);
560 	queue_work(pool->release_wq, &pool->work);
561 	spin_unlock(&pool->stale_lock);
562 }
563 
564 static void __attribute__((__unused__))
565 			release_z3fold_page(struct kref *ref)
566 {
567 	struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
568 						refcount);
569 	__release_z3fold_page(zhdr, false);
570 }
571 
572 static void release_z3fold_page_locked(struct kref *ref)
573 {
574 	struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
575 						refcount);
576 	WARN_ON(z3fold_page_trylock(zhdr));
577 	__release_z3fold_page(zhdr, true);
578 }
579 
580 static void release_z3fold_page_locked_list(struct kref *ref)
581 {
582 	struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
583 					       refcount);
584 	struct z3fold_pool *pool = zhdr_to_pool(zhdr);
585 
586 	spin_lock(&pool->lock);
587 	list_del_init(&zhdr->buddy);
588 	spin_unlock(&pool->lock);
589 
590 	WARN_ON(z3fold_page_trylock(zhdr));
591 	__release_z3fold_page(zhdr, true);
592 }
593 
594 static void free_pages_work(struct work_struct *w)
595 {
596 	struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
597 
598 	spin_lock(&pool->stale_lock);
599 	while (!list_empty(&pool->stale)) {
600 		struct z3fold_header *zhdr = list_first_entry(&pool->stale,
601 						struct z3fold_header, buddy);
602 		struct page *page = virt_to_page(zhdr);
603 
604 		list_del(&zhdr->buddy);
605 		if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
606 			continue;
607 		spin_unlock(&pool->stale_lock);
608 		cancel_work_sync(&zhdr->work);
609 		free_z3fold_page(page, false);
610 		cond_resched();
611 		spin_lock(&pool->stale_lock);
612 	}
613 	spin_unlock(&pool->stale_lock);
614 }
615 
616 /*
617  * Returns the number of free chunks in a z3fold page.
618  * NB: can't be used with HEADLESS pages.
619  */
620 static int num_free_chunks(struct z3fold_header *zhdr)
621 {
622 	int nfree;
623 	/*
624 	 * If there is a middle object, pick up the bigger free space
625 	 * either before or after it. Otherwise just subtract the number
626 	 * of chunks occupied by the first and the last objects.
627 	 */
628 	if (zhdr->middle_chunks != 0) {
629 		int nfree_before = zhdr->first_chunks ?
630 			0 : zhdr->start_middle - ZHDR_CHUNKS;
631 		int nfree_after = zhdr->last_chunks ?
632 			0 : TOTAL_CHUNKS -
633 				(zhdr->start_middle + zhdr->middle_chunks);
634 		nfree = max(nfree_before, nfree_after);
635 	} else
636 		nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
637 	return nfree;
638 }
639 
640 /* Add to the appropriate unbuddied list */
641 static inline void add_to_unbuddied(struct z3fold_pool *pool,
642 				struct z3fold_header *zhdr)
643 {
644 	if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
645 			zhdr->middle_chunks == 0) {
646 		struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
647 
648 		int freechunks = num_free_chunks(zhdr);
649 		spin_lock(&pool->lock);
650 		list_add(&zhdr->buddy, &unbuddied[freechunks]);
651 		spin_unlock(&pool->lock);
652 		zhdr->cpu = smp_processor_id();
653 		put_cpu_ptr(pool->unbuddied);
654 	}
655 }
656 
657 static inline void *mchunk_memmove(struct z3fold_header *zhdr,
658 				unsigned short dst_chunk)
659 {
660 	void *beg = zhdr;
661 	return memmove(beg + (dst_chunk << CHUNK_SHIFT),
662 		       beg + (zhdr->start_middle << CHUNK_SHIFT),
663 		       zhdr->middle_chunks << CHUNK_SHIFT);
664 }
665 
666 static inline bool buddy_single(struct z3fold_header *zhdr)
667 {
668 	return !((zhdr->first_chunks && zhdr->middle_chunks) ||
669 			(zhdr->first_chunks && zhdr->last_chunks) ||
670 			(zhdr->middle_chunks && zhdr->last_chunks));
671 }
672 
673 static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
674 {
675 	struct z3fold_pool *pool = zhdr_to_pool(zhdr);
676 	void *p = zhdr;
677 	unsigned long old_handle = 0;
678 	size_t sz = 0;
679 	struct z3fold_header *new_zhdr = NULL;
680 	int first_idx = __idx(zhdr, FIRST);
681 	int middle_idx = __idx(zhdr, MIDDLE);
682 	int last_idx = __idx(zhdr, LAST);
683 	unsigned short *moved_chunks = NULL;
684 
685 	/*
686 	 * No need to protect slots here -- all the slots are "local" and
687 	 * the page lock is already taken
688 	 */
689 	if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
690 		p += ZHDR_SIZE_ALIGNED;
691 		sz = zhdr->first_chunks << CHUNK_SHIFT;
692 		old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
693 		moved_chunks = &zhdr->first_chunks;
694 	} else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
695 		p += zhdr->start_middle << CHUNK_SHIFT;
696 		sz = zhdr->middle_chunks << CHUNK_SHIFT;
697 		old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
698 		moved_chunks = &zhdr->middle_chunks;
699 	} else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
700 		p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
701 		sz = zhdr->last_chunks << CHUNK_SHIFT;
702 		old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
703 		moved_chunks = &zhdr->last_chunks;
704 	}
705 
706 	if (sz > 0) {
707 		enum buddy new_bud = HEADLESS;
708 		short chunks = size_to_chunks(sz);
709 		void *q;
710 
711 		new_zhdr = __z3fold_alloc(pool, sz, false);
712 		if (!new_zhdr)
713 			return NULL;
714 
715 		if (WARN_ON(new_zhdr == zhdr))
716 			goto out_fail;
717 
718 		if (new_zhdr->first_chunks == 0) {
719 			if (new_zhdr->middle_chunks != 0 &&
720 					chunks >= new_zhdr->start_middle) {
721 				new_bud = LAST;
722 			} else {
723 				new_bud = FIRST;
724 			}
725 		} else if (new_zhdr->last_chunks == 0) {
726 			new_bud = LAST;
727 		} else if (new_zhdr->middle_chunks == 0) {
728 			new_bud = MIDDLE;
729 		}
730 		q = new_zhdr;
731 		switch (new_bud) {
732 		case FIRST:
733 			new_zhdr->first_chunks = chunks;
734 			q += ZHDR_SIZE_ALIGNED;
735 			break;
736 		case MIDDLE:
737 			new_zhdr->middle_chunks = chunks;
738 			new_zhdr->start_middle =
739 				new_zhdr->first_chunks + ZHDR_CHUNKS;
740 			q += new_zhdr->start_middle << CHUNK_SHIFT;
741 			break;
742 		case LAST:
743 			new_zhdr->last_chunks = chunks;
744 			q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
745 			break;
746 		default:
747 			goto out_fail;
748 		}
749 		new_zhdr->foreign_handles++;
750 		memcpy(q, p, sz);
751 		write_lock(&zhdr->slots->lock);
752 		*(unsigned long *)old_handle = (unsigned long)new_zhdr +
753 			__idx(new_zhdr, new_bud);
754 		if (new_bud == LAST)
755 			*(unsigned long *)old_handle |=
756 					(new_zhdr->last_chunks << BUDDY_SHIFT);
757 		write_unlock(&zhdr->slots->lock);
758 		add_to_unbuddied(pool, new_zhdr);
759 		z3fold_page_unlock(new_zhdr);
760 
761 		*moved_chunks = 0;
762 	}
763 
764 	return new_zhdr;
765 
766 out_fail:
767 	if (new_zhdr) {
768 		if (kref_put(&new_zhdr->refcount, release_z3fold_page_locked))
769 			atomic64_dec(&pool->pages_nr);
770 		else {
771 			add_to_unbuddied(pool, new_zhdr);
772 			z3fold_page_unlock(new_zhdr);
773 		}
774 	}
775 	return NULL;
776 
777 }
778 
779 #define BIG_CHUNK_GAP	3
780 /* Has to be called with lock held */
781 static int z3fold_compact_page(struct z3fold_header *zhdr)
782 {
783 	struct page *page = virt_to_page(zhdr);
784 
785 	if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
786 		return 0; /* can't move middle chunk, it's used */
787 
788 	if (unlikely(PageIsolated(page)))
789 		return 0;
790 
791 	if (zhdr->middle_chunks == 0)
792 		return 0; /* nothing to compact */
793 
794 	if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
795 		/* move to the beginning */
796 		mchunk_memmove(zhdr, ZHDR_CHUNKS);
797 		zhdr->first_chunks = zhdr->middle_chunks;
798 		zhdr->middle_chunks = 0;
799 		zhdr->start_middle = 0;
800 		zhdr->first_num++;
801 		return 1;
802 	}
803 
804 	/*
805 	 * moving data is expensive, so let's only do that if
806 	 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
807 	 */
808 	if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
809 	    zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
810 			BIG_CHUNK_GAP) {
811 		mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
812 		zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
813 		return 1;
814 	} else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
815 		   TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
816 					+ zhdr->middle_chunks) >=
817 			BIG_CHUNK_GAP) {
818 		unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
819 			zhdr->middle_chunks;
820 		mchunk_memmove(zhdr, new_start);
821 		zhdr->start_middle = new_start;
822 		return 1;
823 	}
824 
825 	return 0;
826 }
827 
828 static void do_compact_page(struct z3fold_header *zhdr, bool locked)
829 {
830 	struct z3fold_pool *pool = zhdr_to_pool(zhdr);
831 	struct page *page;
832 
833 	page = virt_to_page(zhdr);
834 	if (locked)
835 		WARN_ON(z3fold_page_trylock(zhdr));
836 	else
837 		z3fold_page_lock(zhdr);
838 	if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
839 		z3fold_page_unlock(zhdr);
840 		return;
841 	}
842 	spin_lock(&pool->lock);
843 	list_del_init(&zhdr->buddy);
844 	spin_unlock(&pool->lock);
845 
846 	if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
847 		atomic64_dec(&pool->pages_nr);
848 		return;
849 	}
850 
851 	if (unlikely(PageIsolated(page) ||
852 		     test_bit(PAGE_CLAIMED, &page->private) ||
853 		     test_bit(PAGE_STALE, &page->private))) {
854 		z3fold_page_unlock(zhdr);
855 		return;
856 	}
857 
858 	if (!zhdr->foreign_handles && buddy_single(zhdr) &&
859 	    zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
860 		if (kref_put(&zhdr->refcount, release_z3fold_page_locked))
861 			atomic64_dec(&pool->pages_nr);
862 		else
863 			z3fold_page_unlock(zhdr);
864 		return;
865 	}
866 
867 	z3fold_compact_page(zhdr);
868 	add_to_unbuddied(pool, zhdr);
869 	z3fold_page_unlock(zhdr);
870 }
871 
872 static void compact_page_work(struct work_struct *w)
873 {
874 	struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
875 						work);
876 
877 	do_compact_page(zhdr, false);
878 }
879 
880 /* returns _locked_ z3fold page header or NULL */
881 static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
882 						size_t size, bool can_sleep)
883 {
884 	struct z3fold_header *zhdr = NULL;
885 	struct page *page;
886 	struct list_head *unbuddied;
887 	int chunks = size_to_chunks(size), i;
888 
889 lookup:
890 	/* First, try to find an unbuddied z3fold page. */
891 	unbuddied = get_cpu_ptr(pool->unbuddied);
892 	for_each_unbuddied_list(i, chunks) {
893 		struct list_head *l = &unbuddied[i];
894 
895 		zhdr = list_first_entry_or_null(READ_ONCE(l),
896 					struct z3fold_header, buddy);
897 
898 		if (!zhdr)
899 			continue;
900 
901 		/* Re-check under lock. */
902 		spin_lock(&pool->lock);
903 		l = &unbuddied[i];
904 		if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
905 						struct z3fold_header, buddy)) ||
906 		    !z3fold_page_trylock(zhdr)) {
907 			spin_unlock(&pool->lock);
908 			zhdr = NULL;
909 			put_cpu_ptr(pool->unbuddied);
910 			if (can_sleep)
911 				cond_resched();
912 			goto lookup;
913 		}
914 		list_del_init(&zhdr->buddy);
915 		zhdr->cpu = -1;
916 		spin_unlock(&pool->lock);
917 
918 		page = virt_to_page(zhdr);
919 		if (test_bit(NEEDS_COMPACTING, &page->private) ||
920 		    test_bit(PAGE_CLAIMED, &page->private)) {
921 			z3fold_page_unlock(zhdr);
922 			zhdr = NULL;
923 			put_cpu_ptr(pool->unbuddied);
924 			if (can_sleep)
925 				cond_resched();
926 			goto lookup;
927 		}
928 
929 		/*
930 		 * this page could not be removed from its unbuddied
931 		 * list while pool lock was held, and then we've taken
932 		 * page lock so kref_put could not be called before
933 		 * we got here, so it's safe to just call kref_get()
934 		 */
935 		kref_get(&zhdr->refcount);
936 		break;
937 	}
938 	put_cpu_ptr(pool->unbuddied);
939 
940 	if (!zhdr) {
941 		int cpu;
942 
943 		/* look for _exact_ match on other cpus' lists */
944 		for_each_online_cpu(cpu) {
945 			struct list_head *l;
946 
947 			unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
948 			spin_lock(&pool->lock);
949 			l = &unbuddied[chunks];
950 
951 			zhdr = list_first_entry_or_null(READ_ONCE(l),
952 						struct z3fold_header, buddy);
953 
954 			if (!zhdr || !z3fold_page_trylock(zhdr)) {
955 				spin_unlock(&pool->lock);
956 				zhdr = NULL;
957 				continue;
958 			}
959 			list_del_init(&zhdr->buddy);
960 			zhdr->cpu = -1;
961 			spin_unlock(&pool->lock);
962 
963 			page = virt_to_page(zhdr);
964 			if (test_bit(NEEDS_COMPACTING, &page->private) ||
965 			    test_bit(PAGE_CLAIMED, &page->private)) {
966 				z3fold_page_unlock(zhdr);
967 				zhdr = NULL;
968 				if (can_sleep)
969 					cond_resched();
970 				continue;
971 			}
972 			kref_get(&zhdr->refcount);
973 			break;
974 		}
975 	}
976 
977 	return zhdr;
978 }
979 
980 /*
981  * API Functions
982  */
983 
984 /**
985  * z3fold_create_pool() - create a new z3fold pool
986  * @name:	pool name
987  * @gfp:	gfp flags when allocating the z3fold pool structure
988  * @ops:	user-defined operations for the z3fold pool
989  *
990  * Return: pointer to the new z3fold pool or NULL if the metadata allocation
991  * failed.
992  */
993 static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
994 		const struct z3fold_ops *ops)
995 {
996 	struct z3fold_pool *pool = NULL;
997 	int i, cpu;
998 
999 	pool = kzalloc(sizeof(struct z3fold_pool), gfp);
1000 	if (!pool)
1001 		goto out;
1002 	pool->c_handle = kmem_cache_create("z3fold_handle",
1003 				sizeof(struct z3fold_buddy_slots),
1004 				SLOTS_ALIGN, 0, NULL);
1005 	if (!pool->c_handle)
1006 		goto out_c;
1007 	spin_lock_init(&pool->lock);
1008 	spin_lock_init(&pool->stale_lock);
1009 	pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
1010 	if (!pool->unbuddied)
1011 		goto out_pool;
1012 	for_each_possible_cpu(cpu) {
1013 		struct list_head *unbuddied =
1014 				per_cpu_ptr(pool->unbuddied, cpu);
1015 		for_each_unbuddied_list(i, 0)
1016 			INIT_LIST_HEAD(&unbuddied[i]);
1017 	}
1018 	INIT_LIST_HEAD(&pool->lru);
1019 	INIT_LIST_HEAD(&pool->stale);
1020 	atomic64_set(&pool->pages_nr, 0);
1021 	pool->name = name;
1022 	pool->compact_wq = create_singlethread_workqueue(pool->name);
1023 	if (!pool->compact_wq)
1024 		goto out_unbuddied;
1025 	pool->release_wq = create_singlethread_workqueue(pool->name);
1026 	if (!pool->release_wq)
1027 		goto out_wq;
1028 	if (z3fold_register_migration(pool))
1029 		goto out_rwq;
1030 	INIT_WORK(&pool->work, free_pages_work);
1031 	pool->ops = ops;
1032 	return pool;
1033 
1034 out_rwq:
1035 	destroy_workqueue(pool->release_wq);
1036 out_wq:
1037 	destroy_workqueue(pool->compact_wq);
1038 out_unbuddied:
1039 	free_percpu(pool->unbuddied);
1040 out_pool:
1041 	kmem_cache_destroy(pool->c_handle);
1042 out_c:
1043 	kfree(pool);
1044 out:
1045 	return NULL;
1046 }
1047 
1048 /**
1049  * z3fold_destroy_pool() - destroys an existing z3fold pool
1050  * @pool:	the z3fold pool to be destroyed
1051  *
1052  * The pool should be emptied before this function is called.
1053  */
1054 static void z3fold_destroy_pool(struct z3fold_pool *pool)
1055 {
1056 	kmem_cache_destroy(pool->c_handle);
1057 
1058 	/*
1059 	 * We need to destroy pool->compact_wq before pool->release_wq,
1060 	 * as any pending work on pool->compact_wq will call
1061 	 * queue_work(pool->release_wq, &pool->work).
1062 	 *
1063 	 * There are still outstanding pages until both workqueues are drained,
1064 	 * so we cannot unregister migration until then.
1065 	 */
1066 
1067 	destroy_workqueue(pool->compact_wq);
1068 	destroy_workqueue(pool->release_wq);
1069 	z3fold_unregister_migration(pool);
1070 	kfree(pool);
1071 }
1072 
1073 /**
1074  * z3fold_alloc() - allocates a region of a given size
1075  * @pool:	z3fold pool from which to allocate
1076  * @size:	size in bytes of the desired allocation
1077  * @gfp:	gfp flags used if the pool needs to grow
1078  * @handle:	handle of the new allocation
1079  *
1080  * This function will attempt to find a free region in the pool large enough to
1081  * satisfy the allocation request.  A search of the unbuddied lists is
1082  * performed first. If no suitable free region is found, then a new page is
1083  * allocated and added to the pool to satisfy the request.
1084  *
1085  * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
1086  * as z3fold pool pages.
1087  *
1088  * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
1089  * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
1090  * a new page.
1091  */
1092 static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1093 			unsigned long *handle)
1094 {
1095 	int chunks = size_to_chunks(size);
1096 	struct z3fold_header *zhdr = NULL;
1097 	struct page *page = NULL;
1098 	enum buddy bud;
1099 	bool can_sleep = gfpflags_allow_blocking(gfp);
1100 
1101 	if (!size)
1102 		return -EINVAL;
1103 
1104 	if (size > PAGE_SIZE)
1105 		return -ENOSPC;
1106 
1107 	if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1108 		bud = HEADLESS;
1109 	else {
1110 retry:
1111 		zhdr = __z3fold_alloc(pool, size, can_sleep);
1112 		if (zhdr) {
1113 			if (zhdr->first_chunks == 0) {
1114 				if (zhdr->middle_chunks != 0 &&
1115 				    chunks >= zhdr->start_middle)
1116 					bud = LAST;
1117 				else
1118 					bud = FIRST;
1119 			} else if (zhdr->last_chunks == 0)
1120 				bud = LAST;
1121 			else if (zhdr->middle_chunks == 0)
1122 				bud = MIDDLE;
1123 			else {
1124 				if (kref_put(&zhdr->refcount,
1125 					     release_z3fold_page_locked))
1126 					atomic64_dec(&pool->pages_nr);
1127 				else
1128 					z3fold_page_unlock(zhdr);
1129 				pr_err("No free chunks in unbuddied\n");
1130 				WARN_ON(1);
1131 				goto retry;
1132 			}
1133 			page = virt_to_page(zhdr);
1134 			goto found;
1135 		}
1136 		bud = FIRST;
1137 	}
1138 
1139 	page = NULL;
1140 	if (can_sleep) {
1141 		spin_lock(&pool->stale_lock);
1142 		zhdr = list_first_entry_or_null(&pool->stale,
1143 						struct z3fold_header, buddy);
1144 		/*
1145 		 * Before allocating a page, let's see if we can take one from
1146 		 * the stale pages list. cancel_work_sync() can sleep so we
1147 		 * limit this case to the contexts where we can sleep
1148 		 */
1149 		if (zhdr) {
1150 			list_del(&zhdr->buddy);
1151 			spin_unlock(&pool->stale_lock);
1152 			cancel_work_sync(&zhdr->work);
1153 			page = virt_to_page(zhdr);
1154 		} else {
1155 			spin_unlock(&pool->stale_lock);
1156 		}
1157 	}
1158 	if (!page)
1159 		page = alloc_page(gfp);
1160 
1161 	if (!page)
1162 		return -ENOMEM;
1163 
1164 	zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1165 	if (!zhdr) {
1166 		__free_page(page);
1167 		return -ENOMEM;
1168 	}
1169 	atomic64_inc(&pool->pages_nr);
1170 
1171 	if (bud == HEADLESS) {
1172 		set_bit(PAGE_HEADLESS, &page->private);
1173 		goto headless;
1174 	}
1175 	if (can_sleep) {
1176 		lock_page(page);
1177 		__SetPageMovable(page, pool->inode->i_mapping);
1178 		unlock_page(page);
1179 	} else {
1180 		if (trylock_page(page)) {
1181 			__SetPageMovable(page, pool->inode->i_mapping);
1182 			unlock_page(page);
1183 		}
1184 	}
1185 	z3fold_page_lock(zhdr);
1186 
1187 found:
1188 	if (bud == FIRST)
1189 		zhdr->first_chunks = chunks;
1190 	else if (bud == LAST)
1191 		zhdr->last_chunks = chunks;
1192 	else {
1193 		zhdr->middle_chunks = chunks;
1194 		zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1195 	}
1196 	add_to_unbuddied(pool, zhdr);
1197 
1198 headless:
1199 	spin_lock(&pool->lock);
1200 	/* Add/move z3fold page to beginning of LRU */
1201 	if (!list_empty(&page->lru))
1202 		list_del(&page->lru);
1203 
1204 	list_add(&page->lru, &pool->lru);
1205 
1206 	*handle = encode_handle(zhdr, bud);
1207 	spin_unlock(&pool->lock);
1208 	if (bud != HEADLESS)
1209 		z3fold_page_unlock(zhdr);
1210 
1211 	return 0;
1212 }
1213 
1214 /**
1215  * z3fold_free() - frees the allocation associated with the given handle
1216  * @pool:	pool in which the allocation resided
1217  * @handle:	handle associated with the allocation returned by z3fold_alloc()
1218  *
1219  * In the case that the z3fold page in which the allocation resides is under
1220  * reclaim, as indicated by the PG_reclaim flag being set, this function
1221  * only sets the first|last_chunks to 0.  The page is actually freed
1222  * once both buddies are evicted (see z3fold_reclaim_page() below).
1223  */
1224 static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1225 {
1226 	struct z3fold_header *zhdr;
1227 	struct page *page;
1228 	enum buddy bud;
1229 	bool page_claimed;
1230 
1231 	zhdr = get_z3fold_header(handle);
1232 	page = virt_to_page(zhdr);
1233 	page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1234 
1235 	if (test_bit(PAGE_HEADLESS, &page->private)) {
1236 		/* if a headless page is under reclaim, just leave.
1237 		 * NB: we use test_and_set_bit for a reason: if the bit
1238 		 * has not been set before, we release this page
1239 		 * immediately so we don't care about its value any more.
1240 		 */
1241 		if (!page_claimed) {
1242 			spin_lock(&pool->lock);
1243 			list_del(&page->lru);
1244 			spin_unlock(&pool->lock);
1245 			put_z3fold_header(zhdr);
1246 			free_z3fold_page(page, true);
1247 			atomic64_dec(&pool->pages_nr);
1248 		}
1249 		return;
1250 	}
1251 
1252 	/* Non-headless case */
1253 	bud = handle_to_buddy(handle);
1254 
1255 	switch (bud) {
1256 	case FIRST:
1257 		zhdr->first_chunks = 0;
1258 		break;
1259 	case MIDDLE:
1260 		zhdr->middle_chunks = 0;
1261 		break;
1262 	case LAST:
1263 		zhdr->last_chunks = 0;
1264 		break;
1265 	default:
1266 		pr_err("%s: unknown bud %d\n", __func__, bud);
1267 		WARN_ON(1);
1268 		put_z3fold_header(zhdr);
1269 		clear_bit(PAGE_CLAIMED, &page->private);
1270 		return;
1271 	}
1272 
1273 	if (!page_claimed)
1274 		free_handle(handle);
1275 	if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1276 		atomic64_dec(&pool->pages_nr);
1277 		return;
1278 	}
1279 	if (page_claimed) {
1280 		/* the page has not been claimed by us */
1281 		z3fold_page_unlock(zhdr);
1282 		return;
1283 	}
1284 	if (unlikely(PageIsolated(page)) ||
1285 	    test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1286 		put_z3fold_header(zhdr);
1287 		clear_bit(PAGE_CLAIMED, &page->private);
1288 		return;
1289 	}
1290 	if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1291 		spin_lock(&pool->lock);
1292 		list_del_init(&zhdr->buddy);
1293 		spin_unlock(&pool->lock);
1294 		zhdr->cpu = -1;
1295 		kref_get(&zhdr->refcount);
1296 		clear_bit(PAGE_CLAIMED, &page->private);
1297 		do_compact_page(zhdr, true);
1298 		return;
1299 	}
1300 	kref_get(&zhdr->refcount);
1301 	clear_bit(PAGE_CLAIMED, &page->private);
1302 	queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1303 	put_z3fold_header(zhdr);
1304 }
1305 
1306 /**
1307  * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1308  * @pool:	pool from which a page will attempt to be evicted
1309  * @retries:	number of pages on the LRU list for which eviction will
1310  *		be attempted before failing
1311  *
1312  * z3fold reclaim is different from normal system reclaim in that it is done
1313  * from the bottom, up. This is because only the bottom layer, z3fold, has
1314  * information on how the allocations are organized within each z3fold page.
1315  * This has the potential to create interesting locking situations between
1316  * z3fold and the user, however.
1317  *
1318  * To avoid these, this is how z3fold_reclaim_page() should be called:
1319  *
1320  * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1321  * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1322  * call the user-defined eviction handler with the pool and handle as
1323  * arguments.
1324  *
1325  * If the handle can not be evicted, the eviction handler should return
1326  * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1327  * appropriate list and try the next z3fold page on the LRU up to
1328  * a user defined number of retries.
1329  *
1330  * If the handle is successfully evicted, the eviction handler should
1331  * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1332  * contains logic to delay freeing the page if the page is under reclaim,
1333  * as indicated by the setting of the PG_reclaim flag on the underlying page.
1334  *
1335  * If all buddies in the z3fold page are successfully evicted, then the
1336  * z3fold page can be freed.
1337  *
1338  * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1339  * no pages to evict or an eviction handler is not registered, -EAGAIN if
1340  * the retry limit was hit.
1341  */
1342 static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1343 {
1344 	int i, ret = -1;
1345 	struct z3fold_header *zhdr = NULL;
1346 	struct page *page = NULL;
1347 	struct list_head *pos;
1348 	unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1349 
1350 	spin_lock(&pool->lock);
1351 	if (!pool->ops || !pool->ops->evict || retries == 0) {
1352 		spin_unlock(&pool->lock);
1353 		return -EINVAL;
1354 	}
1355 	for (i = 0; i < retries; i++) {
1356 		if (list_empty(&pool->lru)) {
1357 			spin_unlock(&pool->lock);
1358 			return -EINVAL;
1359 		}
1360 		list_for_each_prev(pos, &pool->lru) {
1361 			page = list_entry(pos, struct page, lru);
1362 
1363 			/* this bit could have been set by free, in which case
1364 			 * we pass over to the next page in the pool.
1365 			 */
1366 			if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1367 				page = NULL;
1368 				continue;
1369 			}
1370 
1371 			if (unlikely(PageIsolated(page))) {
1372 				clear_bit(PAGE_CLAIMED, &page->private);
1373 				page = NULL;
1374 				continue;
1375 			}
1376 			zhdr = page_address(page);
1377 			if (test_bit(PAGE_HEADLESS, &page->private))
1378 				break;
1379 
1380 			if (!z3fold_page_trylock(zhdr)) {
1381 				clear_bit(PAGE_CLAIMED, &page->private);
1382 				zhdr = NULL;
1383 				continue; /* can't evict at this point */
1384 			}
1385 			if (zhdr->foreign_handles) {
1386 				clear_bit(PAGE_CLAIMED, &page->private);
1387 				z3fold_page_unlock(zhdr);
1388 				zhdr = NULL;
1389 				continue; /* can't evict such page */
1390 			}
1391 			kref_get(&zhdr->refcount);
1392 			list_del_init(&zhdr->buddy);
1393 			zhdr->cpu = -1;
1394 			break;
1395 		}
1396 
1397 		if (!zhdr)
1398 			break;
1399 
1400 		list_del_init(&page->lru);
1401 		spin_unlock(&pool->lock);
1402 
1403 		if (!test_bit(PAGE_HEADLESS, &page->private)) {
1404 			/*
1405 			 * We need encode the handles before unlocking, and
1406 			 * use our local slots structure because z3fold_free
1407 			 * can zero out zhdr->slots and we can't do much
1408 			 * about that
1409 			 */
1410 			first_handle = 0;
1411 			last_handle = 0;
1412 			middle_handle = 0;
1413 			if (zhdr->first_chunks)
1414 				first_handle = encode_handle(zhdr, FIRST);
1415 			if (zhdr->middle_chunks)
1416 				middle_handle = encode_handle(zhdr, MIDDLE);
1417 			if (zhdr->last_chunks)
1418 				last_handle = encode_handle(zhdr, LAST);
1419 			/*
1420 			 * it's safe to unlock here because we hold a
1421 			 * reference to this page
1422 			 */
1423 			z3fold_page_unlock(zhdr);
1424 		} else {
1425 			first_handle = encode_handle(zhdr, HEADLESS);
1426 			last_handle = middle_handle = 0;
1427 		}
1428 		/* Issue the eviction callback(s) */
1429 		if (middle_handle) {
1430 			ret = pool->ops->evict(pool, middle_handle);
1431 			if (ret)
1432 				goto next;
1433 			free_handle(middle_handle);
1434 		}
1435 		if (first_handle) {
1436 			ret = pool->ops->evict(pool, first_handle);
1437 			if (ret)
1438 				goto next;
1439 			free_handle(first_handle);
1440 		}
1441 		if (last_handle) {
1442 			ret = pool->ops->evict(pool, last_handle);
1443 			if (ret)
1444 				goto next;
1445 			free_handle(last_handle);
1446 		}
1447 next:
1448 		if (test_bit(PAGE_HEADLESS, &page->private)) {
1449 			if (ret == 0) {
1450 				free_z3fold_page(page, true);
1451 				atomic64_dec(&pool->pages_nr);
1452 				return 0;
1453 			}
1454 			spin_lock(&pool->lock);
1455 			list_add(&page->lru, &pool->lru);
1456 			spin_unlock(&pool->lock);
1457 			clear_bit(PAGE_CLAIMED, &page->private);
1458 		} else {
1459 			z3fold_page_lock(zhdr);
1460 			if (kref_put(&zhdr->refcount,
1461 					release_z3fold_page_locked)) {
1462 				atomic64_dec(&pool->pages_nr);
1463 				return 0;
1464 			}
1465 			/*
1466 			 * if we are here, the page is still not completely
1467 			 * free. Take the global pool lock then to be able
1468 			 * to add it back to the lru list
1469 			 */
1470 			spin_lock(&pool->lock);
1471 			list_add(&page->lru, &pool->lru);
1472 			spin_unlock(&pool->lock);
1473 			z3fold_page_unlock(zhdr);
1474 			clear_bit(PAGE_CLAIMED, &page->private);
1475 		}
1476 
1477 		/* We started off locked to we need to lock the pool back */
1478 		spin_lock(&pool->lock);
1479 	}
1480 	spin_unlock(&pool->lock);
1481 	return -EAGAIN;
1482 }
1483 
1484 /**
1485  * z3fold_map() - maps the allocation associated with the given handle
1486  * @pool:	pool in which the allocation resides
1487  * @handle:	handle associated with the allocation to be mapped
1488  *
1489  * Extracts the buddy number from handle and constructs the pointer to the
1490  * correct starting chunk within the page.
1491  *
1492  * Returns: a pointer to the mapped allocation
1493  */
1494 static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1495 {
1496 	struct z3fold_header *zhdr;
1497 	struct page *page;
1498 	void *addr;
1499 	enum buddy buddy;
1500 
1501 	zhdr = get_z3fold_header(handle);
1502 	addr = zhdr;
1503 	page = virt_to_page(zhdr);
1504 
1505 	if (test_bit(PAGE_HEADLESS, &page->private))
1506 		goto out;
1507 
1508 	buddy = handle_to_buddy(handle);
1509 	switch (buddy) {
1510 	case FIRST:
1511 		addr += ZHDR_SIZE_ALIGNED;
1512 		break;
1513 	case MIDDLE:
1514 		addr += zhdr->start_middle << CHUNK_SHIFT;
1515 		set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1516 		break;
1517 	case LAST:
1518 		addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1519 		break;
1520 	default:
1521 		pr_err("unknown buddy id %d\n", buddy);
1522 		WARN_ON(1);
1523 		addr = NULL;
1524 		break;
1525 	}
1526 
1527 	if (addr)
1528 		zhdr->mapped_count++;
1529 out:
1530 	put_z3fold_header(zhdr);
1531 	return addr;
1532 }
1533 
1534 /**
1535  * z3fold_unmap() - unmaps the allocation associated with the given handle
1536  * @pool:	pool in which the allocation resides
1537  * @handle:	handle associated with the allocation to be unmapped
1538  */
1539 static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1540 {
1541 	struct z3fold_header *zhdr;
1542 	struct page *page;
1543 	enum buddy buddy;
1544 
1545 	zhdr = get_z3fold_header(handle);
1546 	page = virt_to_page(zhdr);
1547 
1548 	if (test_bit(PAGE_HEADLESS, &page->private))
1549 		return;
1550 
1551 	buddy = handle_to_buddy(handle);
1552 	if (buddy == MIDDLE)
1553 		clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1554 	zhdr->mapped_count--;
1555 	put_z3fold_header(zhdr);
1556 }
1557 
1558 /**
1559  * z3fold_get_pool_size() - gets the z3fold pool size in pages
1560  * @pool:	pool whose size is being queried
1561  *
1562  * Returns: size in pages of the given pool.
1563  */
1564 static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1565 {
1566 	return atomic64_read(&pool->pages_nr);
1567 }
1568 
1569 static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1570 {
1571 	struct z3fold_header *zhdr;
1572 	struct z3fold_pool *pool;
1573 
1574 	VM_BUG_ON_PAGE(!PageMovable(page), page);
1575 	VM_BUG_ON_PAGE(PageIsolated(page), page);
1576 
1577 	if (test_bit(PAGE_HEADLESS, &page->private) ||
1578 	    test_bit(PAGE_CLAIMED, &page->private))
1579 		return false;
1580 
1581 	zhdr = page_address(page);
1582 	z3fold_page_lock(zhdr);
1583 	if (test_bit(NEEDS_COMPACTING, &page->private) ||
1584 	    test_bit(PAGE_STALE, &page->private))
1585 		goto out;
1586 
1587 	if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1588 		goto out;
1589 
1590 	pool = zhdr_to_pool(zhdr);
1591 	spin_lock(&pool->lock);
1592 	if (!list_empty(&zhdr->buddy))
1593 		list_del_init(&zhdr->buddy);
1594 	if (!list_empty(&page->lru))
1595 		list_del_init(&page->lru);
1596 	spin_unlock(&pool->lock);
1597 
1598 	kref_get(&zhdr->refcount);
1599 	z3fold_page_unlock(zhdr);
1600 	return true;
1601 
1602 out:
1603 	z3fold_page_unlock(zhdr);
1604 	return false;
1605 }
1606 
1607 static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1608 			       struct page *page, enum migrate_mode mode)
1609 {
1610 	struct z3fold_header *zhdr, *new_zhdr;
1611 	struct z3fold_pool *pool;
1612 	struct address_space *new_mapping;
1613 
1614 	VM_BUG_ON_PAGE(!PageMovable(page), page);
1615 	VM_BUG_ON_PAGE(!PageIsolated(page), page);
1616 	VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1617 
1618 	zhdr = page_address(page);
1619 	pool = zhdr_to_pool(zhdr);
1620 
1621 	if (!z3fold_page_trylock(zhdr)) {
1622 		return -EAGAIN;
1623 	}
1624 	if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1625 		z3fold_page_unlock(zhdr);
1626 		return -EBUSY;
1627 	}
1628 	if (work_pending(&zhdr->work)) {
1629 		z3fold_page_unlock(zhdr);
1630 		return -EAGAIN;
1631 	}
1632 	new_zhdr = page_address(newpage);
1633 	memcpy(new_zhdr, zhdr, PAGE_SIZE);
1634 	newpage->private = page->private;
1635 	page->private = 0;
1636 	z3fold_page_unlock(zhdr);
1637 	spin_lock_init(&new_zhdr->page_lock);
1638 	INIT_WORK(&new_zhdr->work, compact_page_work);
1639 	/*
1640 	 * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1641 	 * so we only have to reinitialize it.
1642 	 */
1643 	INIT_LIST_HEAD(&new_zhdr->buddy);
1644 	new_mapping = page_mapping(page);
1645 	__ClearPageMovable(page);
1646 	ClearPagePrivate(page);
1647 
1648 	get_page(newpage);
1649 	z3fold_page_lock(new_zhdr);
1650 	if (new_zhdr->first_chunks)
1651 		encode_handle(new_zhdr, FIRST);
1652 	if (new_zhdr->last_chunks)
1653 		encode_handle(new_zhdr, LAST);
1654 	if (new_zhdr->middle_chunks)
1655 		encode_handle(new_zhdr, MIDDLE);
1656 	set_bit(NEEDS_COMPACTING, &newpage->private);
1657 	new_zhdr->cpu = smp_processor_id();
1658 	spin_lock(&pool->lock);
1659 	list_add(&newpage->lru, &pool->lru);
1660 	spin_unlock(&pool->lock);
1661 	__SetPageMovable(newpage, new_mapping);
1662 	z3fold_page_unlock(new_zhdr);
1663 
1664 	queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1665 
1666 	page_mapcount_reset(page);
1667 	put_page(page);
1668 	return 0;
1669 }
1670 
1671 static void z3fold_page_putback(struct page *page)
1672 {
1673 	struct z3fold_header *zhdr;
1674 	struct z3fold_pool *pool;
1675 
1676 	zhdr = page_address(page);
1677 	pool = zhdr_to_pool(zhdr);
1678 
1679 	z3fold_page_lock(zhdr);
1680 	if (!list_empty(&zhdr->buddy))
1681 		list_del_init(&zhdr->buddy);
1682 	INIT_LIST_HEAD(&page->lru);
1683 	if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1684 		atomic64_dec(&pool->pages_nr);
1685 		return;
1686 	}
1687 	spin_lock(&pool->lock);
1688 	list_add(&page->lru, &pool->lru);
1689 	spin_unlock(&pool->lock);
1690 	z3fold_page_unlock(zhdr);
1691 }
1692 
1693 static const struct address_space_operations z3fold_aops = {
1694 	.isolate_page = z3fold_page_isolate,
1695 	.migratepage = z3fold_page_migrate,
1696 	.putback_page = z3fold_page_putback,
1697 };
1698 
1699 /*****************
1700  * zpool
1701  ****************/
1702 
1703 static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1704 {
1705 	if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1706 		return pool->zpool_ops->evict(pool->zpool, handle);
1707 	else
1708 		return -ENOENT;
1709 }
1710 
1711 static const struct z3fold_ops z3fold_zpool_ops = {
1712 	.evict =	z3fold_zpool_evict
1713 };
1714 
1715 static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1716 			       const struct zpool_ops *zpool_ops,
1717 			       struct zpool *zpool)
1718 {
1719 	struct z3fold_pool *pool;
1720 
1721 	pool = z3fold_create_pool(name, gfp,
1722 				zpool_ops ? &z3fold_zpool_ops : NULL);
1723 	if (pool) {
1724 		pool->zpool = zpool;
1725 		pool->zpool_ops = zpool_ops;
1726 	}
1727 	return pool;
1728 }
1729 
1730 static void z3fold_zpool_destroy(void *pool)
1731 {
1732 	z3fold_destroy_pool(pool);
1733 }
1734 
1735 static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1736 			unsigned long *handle)
1737 {
1738 	return z3fold_alloc(pool, size, gfp, handle);
1739 }
1740 static void z3fold_zpool_free(void *pool, unsigned long handle)
1741 {
1742 	z3fold_free(pool, handle);
1743 }
1744 
1745 static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1746 			unsigned int *reclaimed)
1747 {
1748 	unsigned int total = 0;
1749 	int ret = -EINVAL;
1750 
1751 	while (total < pages) {
1752 		ret = z3fold_reclaim_page(pool, 8);
1753 		if (ret < 0)
1754 			break;
1755 		total++;
1756 	}
1757 
1758 	if (reclaimed)
1759 		*reclaimed = total;
1760 
1761 	return ret;
1762 }
1763 
1764 static void *z3fold_zpool_map(void *pool, unsigned long handle,
1765 			enum zpool_mapmode mm)
1766 {
1767 	return z3fold_map(pool, handle);
1768 }
1769 static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1770 {
1771 	z3fold_unmap(pool, handle);
1772 }
1773 
1774 static u64 z3fold_zpool_total_size(void *pool)
1775 {
1776 	return z3fold_get_pool_size(pool) * PAGE_SIZE;
1777 }
1778 
1779 static struct zpool_driver z3fold_zpool_driver = {
1780 	.type =		"z3fold",
1781 	.owner =	THIS_MODULE,
1782 	.create =	z3fold_zpool_create,
1783 	.destroy =	z3fold_zpool_destroy,
1784 	.malloc =	z3fold_zpool_malloc,
1785 	.free =		z3fold_zpool_free,
1786 	.shrink =	z3fold_zpool_shrink,
1787 	.map =		z3fold_zpool_map,
1788 	.unmap =	z3fold_zpool_unmap,
1789 	.total_size =	z3fold_zpool_total_size,
1790 };
1791 
1792 MODULE_ALIAS("zpool-z3fold");
1793 
1794 static int __init init_z3fold(void)
1795 {
1796 	int ret;
1797 
1798 	/* Make sure the z3fold header is not larger than the page size */
1799 	BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1800 	ret = z3fold_mount();
1801 	if (ret)
1802 		return ret;
1803 
1804 	zpool_register_driver(&z3fold_zpool_driver);
1805 
1806 	return 0;
1807 }
1808 
1809 static void __exit exit_z3fold(void)
1810 {
1811 	z3fold_unmount();
1812 	zpool_unregister_driver(&z3fold_zpool_driver);
1813 }
1814 
1815 module_init(init_z3fold);
1816 module_exit(exit_z3fold);
1817 
1818 MODULE_LICENSE("GPL");
1819 MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1820 MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");
1821