Loading...
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*****************/
73struct z3fold_pool;
74struct z3fold_ops {
75 int (*evict)(struct z3fold_pool *pool, unsigned long handle);
76};
77
78enum buddy {
79 HEADLESS = 0,
80 FIRST,
81 MIDDLE,
82 LAST,
83 BUDDIES_MAX = LAST
84};
85
86struct 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 */
113struct 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 */
152struct 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 */
173enum 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 */
186static 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
194static void compact_page_work(struct work_struct *w);
195
196static 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
212static 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
217static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
218{
219 return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
220}
221
222static 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
249static int z3fold_init_fs_context(struct fs_context *fc)
250{
251 return init_pseudo(fc, Z3FOLD_MAGIC) ? 0 : -ENOMEM;
252}
253
254static 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
260static struct vfsmount *z3fold_mnt;
261static 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
272static void z3fold_unmount(void)
273{
274 kern_unmount(z3fold_mnt);
275}
276
277static const struct address_space_operations z3fold_aops;
278static 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
291static 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 */
298static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
299 struct z3fold_pool *pool, gfp_t gfp)
300{
301 struct z3fold_header *zhdr = page_address(page);
302 struct z3fold_buddy_slots *slots;
303
304 INIT_LIST_HEAD(&page->lru);
305 clear_bit(PAGE_HEADLESS, &page->private);
306 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
307 clear_bit(NEEDS_COMPACTING, &page->private);
308 clear_bit(PAGE_STALE, &page->private);
309 clear_bit(PAGE_CLAIMED, &page->private);
310 if (headless)
311 return zhdr;
312
313 slots = alloc_slots(pool, gfp);
314 if (!slots)
315 return NULL;
316
317 spin_lock_init(&zhdr->page_lock);
318 kref_init(&zhdr->refcount);
319 zhdr->first_chunks = 0;
320 zhdr->middle_chunks = 0;
321 zhdr->last_chunks = 0;
322 zhdr->first_num = 0;
323 zhdr->start_middle = 0;
324 zhdr->cpu = -1;
325 zhdr->slots = slots;
326 zhdr->pool = pool;
327 INIT_LIST_HEAD(&zhdr->buddy);
328 INIT_WORK(&zhdr->work, compact_page_work);
329 return zhdr;
330}
331
332/* Resets the struct page fields and frees the page */
333static void free_z3fold_page(struct page *page, bool headless)
334{
335 if (!headless) {
336 lock_page(page);
337 __ClearPageMovable(page);
338 unlock_page(page);
339 }
340 ClearPagePrivate(page);
341 __free_page(page);
342}
343
344/* Lock a z3fold page */
345static inline void z3fold_page_lock(struct z3fold_header *zhdr)
346{
347 spin_lock(&zhdr->page_lock);
348}
349
350/* Try to lock a z3fold page */
351static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
352{
353 return spin_trylock(&zhdr->page_lock);
354}
355
356/* Unlock a z3fold page */
357static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
358{
359 spin_unlock(&zhdr->page_lock);
360}
361
362/* Helper function to build the index */
363static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
364{
365 return (bud + zhdr->first_num) & BUDDY_MASK;
366}
367
368/*
369 * Encodes the handle of a particular buddy within a z3fold page
370 * Pool lock should be held as this function accesses first_num
371 */
372static unsigned long __encode_handle(struct z3fold_header *zhdr,
373 struct z3fold_buddy_slots *slots,
374 enum buddy bud)
375{
376 unsigned long h = (unsigned long)zhdr;
377 int idx = 0;
378
379 /*
380 * For a headless page, its handle is its pointer with the extra
381 * PAGE_HEADLESS bit set
382 */
383 if (bud == HEADLESS)
384 return h | (1 << PAGE_HEADLESS);
385
386 /* otherwise, return pointer to encoded handle */
387 idx = __idx(zhdr, bud);
388 h += idx;
389 if (bud == LAST)
390 h |= (zhdr->last_chunks << BUDDY_SHIFT);
391
392 slots->slot[idx] = h;
393 return (unsigned long)&slots->slot[idx];
394}
395
396static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
397{
398 return __encode_handle(zhdr, zhdr->slots, bud);
399}
400
401/* Returns the z3fold page where a given handle is stored */
402static inline struct z3fold_header *handle_to_z3fold_header(unsigned long h)
403{
404 unsigned long addr = h;
405
406 if (!(addr & (1 << PAGE_HEADLESS)))
407 addr = *(unsigned long *)h;
408
409 return (struct z3fold_header *)(addr & PAGE_MASK);
410}
411
412/* only for LAST bud, returns zero otherwise */
413static unsigned short handle_to_chunks(unsigned long handle)
414{
415 unsigned long addr = *(unsigned long *)handle;
416
417 return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
418}
419
420/*
421 * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
422 * but that doesn't matter. because the masking will result in the
423 * correct buddy number.
424 */
425static enum buddy handle_to_buddy(unsigned long handle)
426{
427 struct z3fold_header *zhdr;
428 unsigned long addr;
429
430 WARN_ON(handle & (1 << PAGE_HEADLESS));
431 addr = *(unsigned long *)handle;
432 zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
433 return (addr - zhdr->first_num) & BUDDY_MASK;
434}
435
436static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
437{
438 return zhdr->pool;
439}
440
441static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
442{
443 struct page *page = virt_to_page(zhdr);
444 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
445
446 WARN_ON(!list_empty(&zhdr->buddy));
447 set_bit(PAGE_STALE, &page->private);
448 clear_bit(NEEDS_COMPACTING, &page->private);
449 spin_lock(&pool->lock);
450 if (!list_empty(&page->lru))
451 list_del_init(&page->lru);
452 spin_unlock(&pool->lock);
453 if (locked)
454 z3fold_page_unlock(zhdr);
455 spin_lock(&pool->stale_lock);
456 list_add(&zhdr->buddy, &pool->stale);
457 queue_work(pool->release_wq, &pool->work);
458 spin_unlock(&pool->stale_lock);
459}
460
461static void __attribute__((__unused__))
462 release_z3fold_page(struct kref *ref)
463{
464 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
465 refcount);
466 __release_z3fold_page(zhdr, false);
467}
468
469static void release_z3fold_page_locked(struct kref *ref)
470{
471 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
472 refcount);
473 WARN_ON(z3fold_page_trylock(zhdr));
474 __release_z3fold_page(zhdr, true);
475}
476
477static void release_z3fold_page_locked_list(struct kref *ref)
478{
479 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
480 refcount);
481 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
482 spin_lock(&pool->lock);
483 list_del_init(&zhdr->buddy);
484 spin_unlock(&pool->lock);
485
486 WARN_ON(z3fold_page_trylock(zhdr));
487 __release_z3fold_page(zhdr, true);
488}
489
490static void free_pages_work(struct work_struct *w)
491{
492 struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
493
494 spin_lock(&pool->stale_lock);
495 while (!list_empty(&pool->stale)) {
496 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
497 struct z3fold_header, buddy);
498 struct page *page = virt_to_page(zhdr);
499
500 list_del(&zhdr->buddy);
501 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
502 continue;
503 spin_unlock(&pool->stale_lock);
504 cancel_work_sync(&zhdr->work);
505 free_z3fold_page(page, false);
506 cond_resched();
507 spin_lock(&pool->stale_lock);
508 }
509 spin_unlock(&pool->stale_lock);
510}
511
512/*
513 * Returns the number of free chunks in a z3fold page.
514 * NB: can't be used with HEADLESS pages.
515 */
516static int num_free_chunks(struct z3fold_header *zhdr)
517{
518 int nfree;
519 /*
520 * If there is a middle object, pick up the bigger free space
521 * either before or after it. Otherwise just subtract the number
522 * of chunks occupied by the first and the last objects.
523 */
524 if (zhdr->middle_chunks != 0) {
525 int nfree_before = zhdr->first_chunks ?
526 0 : zhdr->start_middle - ZHDR_CHUNKS;
527 int nfree_after = zhdr->last_chunks ?
528 0 : TOTAL_CHUNKS -
529 (zhdr->start_middle + zhdr->middle_chunks);
530 nfree = max(nfree_before, nfree_after);
531 } else
532 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
533 return nfree;
534}
535
536/* Add to the appropriate unbuddied list */
537static inline void add_to_unbuddied(struct z3fold_pool *pool,
538 struct z3fold_header *zhdr)
539{
540 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
541 zhdr->middle_chunks == 0) {
542 struct list_head *unbuddied = get_cpu_ptr(pool->unbuddied);
543
544 int freechunks = num_free_chunks(zhdr);
545 spin_lock(&pool->lock);
546 list_add(&zhdr->buddy, &unbuddied[freechunks]);
547 spin_unlock(&pool->lock);
548 zhdr->cpu = smp_processor_id();
549 put_cpu_ptr(pool->unbuddied);
550 }
551}
552
553static inline void *mchunk_memmove(struct z3fold_header *zhdr,
554 unsigned short dst_chunk)
555{
556 void *beg = zhdr;
557 return memmove(beg + (dst_chunk << CHUNK_SHIFT),
558 beg + (zhdr->start_middle << CHUNK_SHIFT),
559 zhdr->middle_chunks << CHUNK_SHIFT);
560}
561
562#define BIG_CHUNK_GAP 3
563/* Has to be called with lock held */
564static int z3fold_compact_page(struct z3fold_header *zhdr)
565{
566 struct page *page = virt_to_page(zhdr);
567
568 if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
569 return 0; /* can't move middle chunk, it's used */
570
571 if (unlikely(PageIsolated(page)))
572 return 0;
573
574 if (zhdr->middle_chunks == 0)
575 return 0; /* nothing to compact */
576
577 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
578 /* move to the beginning */
579 mchunk_memmove(zhdr, ZHDR_CHUNKS);
580 zhdr->first_chunks = zhdr->middle_chunks;
581 zhdr->middle_chunks = 0;
582 zhdr->start_middle = 0;
583 zhdr->first_num++;
584 return 1;
585 }
586
587 /*
588 * moving data is expensive, so let's only do that if
589 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
590 */
591 if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
592 zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
593 BIG_CHUNK_GAP) {
594 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
595 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
596 return 1;
597 } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
598 TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
599 + zhdr->middle_chunks) >=
600 BIG_CHUNK_GAP) {
601 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
602 zhdr->middle_chunks;
603 mchunk_memmove(zhdr, new_start);
604 zhdr->start_middle = new_start;
605 return 1;
606 }
607
608 return 0;
609}
610
611static void do_compact_page(struct z3fold_header *zhdr, bool locked)
612{
613 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
614 struct page *page;
615
616 page = virt_to_page(zhdr);
617 if (locked)
618 WARN_ON(z3fold_page_trylock(zhdr));
619 else
620 z3fold_page_lock(zhdr);
621 if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
622 z3fold_page_unlock(zhdr);
623 return;
624 }
625 spin_lock(&pool->lock);
626 list_del_init(&zhdr->buddy);
627 spin_unlock(&pool->lock);
628
629 if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
630 atomic64_dec(&pool->pages_nr);
631 return;
632 }
633
634 if (unlikely(PageIsolated(page) ||
635 test_bit(PAGE_CLAIMED, &page->private) ||
636 test_bit(PAGE_STALE, &page->private))) {
637 z3fold_page_unlock(zhdr);
638 return;
639 }
640
641 z3fold_compact_page(zhdr);
642 add_to_unbuddied(pool, zhdr);
643 z3fold_page_unlock(zhdr);
644}
645
646static void compact_page_work(struct work_struct *w)
647{
648 struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
649 work);
650
651 do_compact_page(zhdr, false);
652}
653
654/* returns _locked_ z3fold page header or NULL */
655static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
656 size_t size, bool can_sleep)
657{
658 struct z3fold_header *zhdr = NULL;
659 struct page *page;
660 struct list_head *unbuddied;
661 int chunks = size_to_chunks(size), i;
662
663lookup:
664 /* First, try to find an unbuddied z3fold page. */
665 unbuddied = get_cpu_ptr(pool->unbuddied);
666 for_each_unbuddied_list(i, chunks) {
667 struct list_head *l = &unbuddied[i];
668
669 zhdr = list_first_entry_or_null(READ_ONCE(l),
670 struct z3fold_header, buddy);
671
672 if (!zhdr)
673 continue;
674
675 /* Re-check under lock. */
676 spin_lock(&pool->lock);
677 l = &unbuddied[i];
678 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
679 struct z3fold_header, buddy)) ||
680 !z3fold_page_trylock(zhdr)) {
681 spin_unlock(&pool->lock);
682 zhdr = NULL;
683 put_cpu_ptr(pool->unbuddied);
684 if (can_sleep)
685 cond_resched();
686 goto lookup;
687 }
688 list_del_init(&zhdr->buddy);
689 zhdr->cpu = -1;
690 spin_unlock(&pool->lock);
691
692 page = virt_to_page(zhdr);
693 if (test_bit(NEEDS_COMPACTING, &page->private)) {
694 z3fold_page_unlock(zhdr);
695 zhdr = NULL;
696 put_cpu_ptr(pool->unbuddied);
697 if (can_sleep)
698 cond_resched();
699 goto lookup;
700 }
701
702 /*
703 * this page could not be removed from its unbuddied
704 * list while pool lock was held, and then we've taken
705 * page lock so kref_put could not be called before
706 * we got here, so it's safe to just call kref_get()
707 */
708 kref_get(&zhdr->refcount);
709 break;
710 }
711 put_cpu_ptr(pool->unbuddied);
712
713 if (!zhdr) {
714 int cpu;
715
716 /* look for _exact_ match on other cpus' lists */
717 for_each_online_cpu(cpu) {
718 struct list_head *l;
719
720 unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
721 spin_lock(&pool->lock);
722 l = &unbuddied[chunks];
723
724 zhdr = list_first_entry_or_null(READ_ONCE(l),
725 struct z3fold_header, buddy);
726
727 if (!zhdr || !z3fold_page_trylock(zhdr)) {
728 spin_unlock(&pool->lock);
729 zhdr = NULL;
730 continue;
731 }
732 list_del_init(&zhdr->buddy);
733 zhdr->cpu = -1;
734 spin_unlock(&pool->lock);
735
736 page = virt_to_page(zhdr);
737 if (test_bit(NEEDS_COMPACTING, &page->private)) {
738 z3fold_page_unlock(zhdr);
739 zhdr = NULL;
740 if (can_sleep)
741 cond_resched();
742 continue;
743 }
744 kref_get(&zhdr->refcount);
745 break;
746 }
747 }
748
749 return zhdr;
750}
751
752/*
753 * API Functions
754 */
755
756/**
757 * z3fold_create_pool() - create a new z3fold pool
758 * @name: pool name
759 * @gfp: gfp flags when allocating the z3fold pool structure
760 * @ops: user-defined operations for the z3fold pool
761 *
762 * Return: pointer to the new z3fold pool or NULL if the metadata allocation
763 * failed.
764 */
765static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp,
766 const struct z3fold_ops *ops)
767{
768 struct z3fold_pool *pool = NULL;
769 int i, cpu;
770
771 pool = kzalloc(sizeof(struct z3fold_pool), gfp);
772 if (!pool)
773 goto out;
774 pool->c_handle = kmem_cache_create("z3fold_handle",
775 sizeof(struct z3fold_buddy_slots),
776 SLOTS_ALIGN, 0, NULL);
777 if (!pool->c_handle)
778 goto out_c;
779 spin_lock_init(&pool->lock);
780 spin_lock_init(&pool->stale_lock);
781 pool->unbuddied = __alloc_percpu(sizeof(struct list_head)*NCHUNKS, 2);
782 if (!pool->unbuddied)
783 goto out_pool;
784 for_each_possible_cpu(cpu) {
785 struct list_head *unbuddied =
786 per_cpu_ptr(pool->unbuddied, cpu);
787 for_each_unbuddied_list(i, 0)
788 INIT_LIST_HEAD(&unbuddied[i]);
789 }
790 INIT_LIST_HEAD(&pool->lru);
791 INIT_LIST_HEAD(&pool->stale);
792 atomic64_set(&pool->pages_nr, 0);
793 pool->name = name;
794 pool->compact_wq = create_singlethread_workqueue(pool->name);
795 if (!pool->compact_wq)
796 goto out_unbuddied;
797 pool->release_wq = create_singlethread_workqueue(pool->name);
798 if (!pool->release_wq)
799 goto out_wq;
800 if (z3fold_register_migration(pool))
801 goto out_rwq;
802 INIT_WORK(&pool->work, free_pages_work);
803 pool->ops = ops;
804 return pool;
805
806out_rwq:
807 destroy_workqueue(pool->release_wq);
808out_wq:
809 destroy_workqueue(pool->compact_wq);
810out_unbuddied:
811 free_percpu(pool->unbuddied);
812out_pool:
813 kmem_cache_destroy(pool->c_handle);
814out_c:
815 kfree(pool);
816out:
817 return NULL;
818}
819
820/**
821 * z3fold_destroy_pool() - destroys an existing z3fold pool
822 * @pool: the z3fold pool to be destroyed
823 *
824 * The pool should be emptied before this function is called.
825 */
826static void z3fold_destroy_pool(struct z3fold_pool *pool)
827{
828 kmem_cache_destroy(pool->c_handle);
829
830 /*
831 * We need to destroy pool->compact_wq before pool->release_wq,
832 * as any pending work on pool->compact_wq will call
833 * queue_work(pool->release_wq, &pool->work).
834 *
835 * There are still outstanding pages until both workqueues are drained,
836 * so we cannot unregister migration until then.
837 */
838
839 destroy_workqueue(pool->compact_wq);
840 destroy_workqueue(pool->release_wq);
841 z3fold_unregister_migration(pool);
842 kfree(pool);
843}
844
845/**
846 * z3fold_alloc() - allocates a region of a given size
847 * @pool: z3fold pool from which to allocate
848 * @size: size in bytes of the desired allocation
849 * @gfp: gfp flags used if the pool needs to grow
850 * @handle: handle of the new allocation
851 *
852 * This function will attempt to find a free region in the pool large enough to
853 * satisfy the allocation request. A search of the unbuddied lists is
854 * performed first. If no suitable free region is found, then a new page is
855 * allocated and added to the pool to satisfy the request.
856 *
857 * gfp should not set __GFP_HIGHMEM as highmem pages cannot be used
858 * as z3fold pool pages.
859 *
860 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
861 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
862 * a new page.
863 */
864static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
865 unsigned long *handle)
866{
867 int chunks = size_to_chunks(size);
868 struct z3fold_header *zhdr = NULL;
869 struct page *page = NULL;
870 enum buddy bud;
871 bool can_sleep = gfpflags_allow_blocking(gfp);
872
873 if (!size)
874 return -EINVAL;
875
876 if (size > PAGE_SIZE)
877 return -ENOSPC;
878
879 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
880 bud = HEADLESS;
881 else {
882retry:
883 zhdr = __z3fold_alloc(pool, size, can_sleep);
884 if (zhdr) {
885 if (zhdr->first_chunks == 0) {
886 if (zhdr->middle_chunks != 0 &&
887 chunks >= zhdr->start_middle)
888 bud = LAST;
889 else
890 bud = FIRST;
891 } else if (zhdr->last_chunks == 0)
892 bud = LAST;
893 else if (zhdr->middle_chunks == 0)
894 bud = MIDDLE;
895 else {
896 if (kref_put(&zhdr->refcount,
897 release_z3fold_page_locked))
898 atomic64_dec(&pool->pages_nr);
899 else
900 z3fold_page_unlock(zhdr);
901 pr_err("No free chunks in unbuddied\n");
902 WARN_ON(1);
903 goto retry;
904 }
905 page = virt_to_page(zhdr);
906 goto found;
907 }
908 bud = FIRST;
909 }
910
911 page = NULL;
912 if (can_sleep) {
913 spin_lock(&pool->stale_lock);
914 zhdr = list_first_entry_or_null(&pool->stale,
915 struct z3fold_header, buddy);
916 /*
917 * Before allocating a page, let's see if we can take one from
918 * the stale pages list. cancel_work_sync() can sleep so we
919 * limit this case to the contexts where we can sleep
920 */
921 if (zhdr) {
922 list_del(&zhdr->buddy);
923 spin_unlock(&pool->stale_lock);
924 cancel_work_sync(&zhdr->work);
925 page = virt_to_page(zhdr);
926 } else {
927 spin_unlock(&pool->stale_lock);
928 }
929 }
930 if (!page)
931 page = alloc_page(gfp);
932
933 if (!page)
934 return -ENOMEM;
935
936 zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
937 if (!zhdr) {
938 __free_page(page);
939 return -ENOMEM;
940 }
941 atomic64_inc(&pool->pages_nr);
942
943 if (bud == HEADLESS) {
944 set_bit(PAGE_HEADLESS, &page->private);
945 goto headless;
946 }
947 if (can_sleep) {
948 lock_page(page);
949 __SetPageMovable(page, pool->inode->i_mapping);
950 unlock_page(page);
951 } else {
952 if (trylock_page(page)) {
953 __SetPageMovable(page, pool->inode->i_mapping);
954 unlock_page(page);
955 }
956 }
957 z3fold_page_lock(zhdr);
958
959found:
960 if (bud == FIRST)
961 zhdr->first_chunks = chunks;
962 else if (bud == LAST)
963 zhdr->last_chunks = chunks;
964 else {
965 zhdr->middle_chunks = chunks;
966 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
967 }
968 add_to_unbuddied(pool, zhdr);
969
970headless:
971 spin_lock(&pool->lock);
972 /* Add/move z3fold page to beginning of LRU */
973 if (!list_empty(&page->lru))
974 list_del(&page->lru);
975
976 list_add(&page->lru, &pool->lru);
977
978 *handle = encode_handle(zhdr, bud);
979 spin_unlock(&pool->lock);
980 if (bud != HEADLESS)
981 z3fold_page_unlock(zhdr);
982
983 return 0;
984}
985
986/**
987 * z3fold_free() - frees the allocation associated with the given handle
988 * @pool: pool in which the allocation resided
989 * @handle: handle associated with the allocation returned by z3fold_alloc()
990 *
991 * In the case that the z3fold page in which the allocation resides is under
992 * reclaim, as indicated by the PG_reclaim flag being set, this function
993 * only sets the first|last_chunks to 0. The page is actually freed
994 * once both buddies are evicted (see z3fold_reclaim_page() below).
995 */
996static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
997{
998 struct z3fold_header *zhdr;
999 struct page *page;
1000 enum buddy bud;
1001 bool page_claimed;
1002
1003 zhdr = handle_to_z3fold_header(handle);
1004 page = virt_to_page(zhdr);
1005 page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1006
1007 if (test_bit(PAGE_HEADLESS, &page->private)) {
1008 /* if a headless page is under reclaim, just leave.
1009 * NB: we use test_and_set_bit for a reason: if the bit
1010 * has not been set before, we release this page
1011 * immediately so we don't care about its value any more.
1012 */
1013 if (!page_claimed) {
1014 spin_lock(&pool->lock);
1015 list_del(&page->lru);
1016 spin_unlock(&pool->lock);
1017 free_z3fold_page(page, true);
1018 atomic64_dec(&pool->pages_nr);
1019 }
1020 return;
1021 }
1022
1023 /* Non-headless case */
1024 z3fold_page_lock(zhdr);
1025 bud = handle_to_buddy(handle);
1026
1027 switch (bud) {
1028 case FIRST:
1029 zhdr->first_chunks = 0;
1030 break;
1031 case MIDDLE:
1032 zhdr->middle_chunks = 0;
1033 break;
1034 case LAST:
1035 zhdr->last_chunks = 0;
1036 break;
1037 default:
1038 pr_err("%s: unknown bud %d\n", __func__, bud);
1039 WARN_ON(1);
1040 z3fold_page_unlock(zhdr);
1041 return;
1042 }
1043
1044 free_handle(handle);
1045 if (kref_put(&zhdr->refcount, release_z3fold_page_locked_list)) {
1046 atomic64_dec(&pool->pages_nr);
1047 return;
1048 }
1049 if (page_claimed) {
1050 /* the page has not been claimed by us */
1051 z3fold_page_unlock(zhdr);
1052 return;
1053 }
1054 if (unlikely(PageIsolated(page)) ||
1055 test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1056 z3fold_page_unlock(zhdr);
1057 clear_bit(PAGE_CLAIMED, &page->private);
1058 return;
1059 }
1060 if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1061 spin_lock(&pool->lock);
1062 list_del_init(&zhdr->buddy);
1063 spin_unlock(&pool->lock);
1064 zhdr->cpu = -1;
1065 kref_get(&zhdr->refcount);
1066 do_compact_page(zhdr, true);
1067 clear_bit(PAGE_CLAIMED, &page->private);
1068 return;
1069 }
1070 kref_get(&zhdr->refcount);
1071 queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1072 clear_bit(PAGE_CLAIMED, &page->private);
1073 z3fold_page_unlock(zhdr);
1074}
1075
1076/**
1077 * z3fold_reclaim_page() - evicts allocations from a pool page and frees it
1078 * @pool: pool from which a page will attempt to be evicted
1079 * @retries: number of pages on the LRU list for which eviction will
1080 * be attempted before failing
1081 *
1082 * z3fold reclaim is different from normal system reclaim in that it is done
1083 * from the bottom, up. This is because only the bottom layer, z3fold, has
1084 * information on how the allocations are organized within each z3fold page.
1085 * This has the potential to create interesting locking situations between
1086 * z3fold and the user, however.
1087 *
1088 * To avoid these, this is how z3fold_reclaim_page() should be called:
1089 *
1090 * The user detects a page should be reclaimed and calls z3fold_reclaim_page().
1091 * z3fold_reclaim_page() will remove a z3fold page from the pool LRU list and
1092 * call the user-defined eviction handler with the pool and handle as
1093 * arguments.
1094 *
1095 * If the handle can not be evicted, the eviction handler should return
1096 * non-zero. z3fold_reclaim_page() will add the z3fold page back to the
1097 * appropriate list and try the next z3fold page on the LRU up to
1098 * a user defined number of retries.
1099 *
1100 * If the handle is successfully evicted, the eviction handler should
1101 * return 0 _and_ should have called z3fold_free() on the handle. z3fold_free()
1102 * contains logic to delay freeing the page if the page is under reclaim,
1103 * as indicated by the setting of the PG_reclaim flag on the underlying page.
1104 *
1105 * If all buddies in the z3fold page are successfully evicted, then the
1106 * z3fold page can be freed.
1107 *
1108 * Returns: 0 if page is successfully freed, otherwise -EINVAL if there are
1109 * no pages to evict or an eviction handler is not registered, -EAGAIN if
1110 * the retry limit was hit.
1111 */
1112static int z3fold_reclaim_page(struct z3fold_pool *pool, unsigned int retries)
1113{
1114 int i, ret = 0;
1115 struct z3fold_header *zhdr = NULL;
1116 struct page *page = NULL;
1117 struct list_head *pos;
1118 struct z3fold_buddy_slots slots;
1119 unsigned long first_handle = 0, middle_handle = 0, last_handle = 0;
1120
1121 spin_lock(&pool->lock);
1122 if (!pool->ops || !pool->ops->evict || retries == 0) {
1123 spin_unlock(&pool->lock);
1124 return -EINVAL;
1125 }
1126 for (i = 0; i < retries; i++) {
1127 if (list_empty(&pool->lru)) {
1128 spin_unlock(&pool->lock);
1129 return -EINVAL;
1130 }
1131 list_for_each_prev(pos, &pool->lru) {
1132 page = list_entry(pos, struct page, lru);
1133
1134 /* this bit could have been set by free, in which case
1135 * we pass over to the next page in the pool.
1136 */
1137 if (test_and_set_bit(PAGE_CLAIMED, &page->private)) {
1138 page = NULL;
1139 continue;
1140 }
1141
1142 if (unlikely(PageIsolated(page))) {
1143 clear_bit(PAGE_CLAIMED, &page->private);
1144 page = NULL;
1145 continue;
1146 }
1147 zhdr = page_address(page);
1148 if (test_bit(PAGE_HEADLESS, &page->private))
1149 break;
1150
1151 if (!z3fold_page_trylock(zhdr)) {
1152 clear_bit(PAGE_CLAIMED, &page->private);
1153 zhdr = NULL;
1154 continue; /* can't evict at this point */
1155 }
1156 kref_get(&zhdr->refcount);
1157 list_del_init(&zhdr->buddy);
1158 zhdr->cpu = -1;
1159 break;
1160 }
1161
1162 if (!zhdr)
1163 break;
1164
1165 list_del_init(&page->lru);
1166 spin_unlock(&pool->lock);
1167
1168 if (!test_bit(PAGE_HEADLESS, &page->private)) {
1169 /*
1170 * We need encode the handles before unlocking, and
1171 * use our local slots structure because z3fold_free
1172 * can zero out zhdr->slots and we can't do much
1173 * about that
1174 */
1175 first_handle = 0;
1176 last_handle = 0;
1177 middle_handle = 0;
1178 if (zhdr->first_chunks)
1179 first_handle = __encode_handle(zhdr, &slots,
1180 FIRST);
1181 if (zhdr->middle_chunks)
1182 middle_handle = __encode_handle(zhdr, &slots,
1183 MIDDLE);
1184 if (zhdr->last_chunks)
1185 last_handle = __encode_handle(zhdr, &slots,
1186 LAST);
1187 /*
1188 * it's safe to unlock here because we hold a
1189 * reference to this page
1190 */
1191 z3fold_page_unlock(zhdr);
1192 } else {
1193 first_handle = __encode_handle(zhdr, &slots, HEADLESS);
1194 last_handle = middle_handle = 0;
1195 }
1196
1197 /* Issue the eviction callback(s) */
1198 if (middle_handle) {
1199 ret = pool->ops->evict(pool, middle_handle);
1200 if (ret)
1201 goto next;
1202 }
1203 if (first_handle) {
1204 ret = pool->ops->evict(pool, first_handle);
1205 if (ret)
1206 goto next;
1207 }
1208 if (last_handle) {
1209 ret = pool->ops->evict(pool, last_handle);
1210 if (ret)
1211 goto next;
1212 }
1213next:
1214 if (test_bit(PAGE_HEADLESS, &page->private)) {
1215 if (ret == 0) {
1216 free_z3fold_page(page, true);
1217 atomic64_dec(&pool->pages_nr);
1218 return 0;
1219 }
1220 spin_lock(&pool->lock);
1221 list_add(&page->lru, &pool->lru);
1222 spin_unlock(&pool->lock);
1223 clear_bit(PAGE_CLAIMED, &page->private);
1224 } else {
1225 z3fold_page_lock(zhdr);
1226 if (kref_put(&zhdr->refcount,
1227 release_z3fold_page_locked)) {
1228 atomic64_dec(&pool->pages_nr);
1229 return 0;
1230 }
1231 /*
1232 * if we are here, the page is still not completely
1233 * free. Take the global pool lock then to be able
1234 * to add it back to the lru list
1235 */
1236 spin_lock(&pool->lock);
1237 list_add(&page->lru, &pool->lru);
1238 spin_unlock(&pool->lock);
1239 z3fold_page_unlock(zhdr);
1240 clear_bit(PAGE_CLAIMED, &page->private);
1241 }
1242
1243 /* We started off locked to we need to lock the pool back */
1244 spin_lock(&pool->lock);
1245 }
1246 spin_unlock(&pool->lock);
1247 return -EAGAIN;
1248}
1249
1250/**
1251 * z3fold_map() - maps the allocation associated with the given handle
1252 * @pool: pool in which the allocation resides
1253 * @handle: handle associated with the allocation to be mapped
1254 *
1255 * Extracts the buddy number from handle and constructs the pointer to the
1256 * correct starting chunk within the page.
1257 *
1258 * Returns: a pointer to the mapped allocation
1259 */
1260static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1261{
1262 struct z3fold_header *zhdr;
1263 struct page *page;
1264 void *addr;
1265 enum buddy buddy;
1266
1267 zhdr = handle_to_z3fold_header(handle);
1268 addr = zhdr;
1269 page = virt_to_page(zhdr);
1270
1271 if (test_bit(PAGE_HEADLESS, &page->private))
1272 goto out;
1273
1274 z3fold_page_lock(zhdr);
1275 buddy = handle_to_buddy(handle);
1276 switch (buddy) {
1277 case FIRST:
1278 addr += ZHDR_SIZE_ALIGNED;
1279 break;
1280 case MIDDLE:
1281 addr += zhdr->start_middle << CHUNK_SHIFT;
1282 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1283 break;
1284 case LAST:
1285 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1286 break;
1287 default:
1288 pr_err("unknown buddy id %d\n", buddy);
1289 WARN_ON(1);
1290 addr = NULL;
1291 break;
1292 }
1293
1294 if (addr)
1295 zhdr->mapped_count++;
1296 z3fold_page_unlock(zhdr);
1297out:
1298 return addr;
1299}
1300
1301/**
1302 * z3fold_unmap() - unmaps the allocation associated with the given handle
1303 * @pool: pool in which the allocation resides
1304 * @handle: handle associated with the allocation to be unmapped
1305 */
1306static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1307{
1308 struct z3fold_header *zhdr;
1309 struct page *page;
1310 enum buddy buddy;
1311
1312 zhdr = handle_to_z3fold_header(handle);
1313 page = virt_to_page(zhdr);
1314
1315 if (test_bit(PAGE_HEADLESS, &page->private))
1316 return;
1317
1318 z3fold_page_lock(zhdr);
1319 buddy = handle_to_buddy(handle);
1320 if (buddy == MIDDLE)
1321 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1322 zhdr->mapped_count--;
1323 z3fold_page_unlock(zhdr);
1324}
1325
1326/**
1327 * z3fold_get_pool_size() - gets the z3fold pool size in pages
1328 * @pool: pool whose size is being queried
1329 *
1330 * Returns: size in pages of the given pool.
1331 */
1332static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1333{
1334 return atomic64_read(&pool->pages_nr);
1335}
1336
1337static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1338{
1339 struct z3fold_header *zhdr;
1340 struct z3fold_pool *pool;
1341
1342 VM_BUG_ON_PAGE(!PageMovable(page), page);
1343 VM_BUG_ON_PAGE(PageIsolated(page), page);
1344
1345 if (test_bit(PAGE_HEADLESS, &page->private) ||
1346 test_bit(PAGE_CLAIMED, &page->private))
1347 return false;
1348
1349 zhdr = page_address(page);
1350 z3fold_page_lock(zhdr);
1351 if (test_bit(NEEDS_COMPACTING, &page->private) ||
1352 test_bit(PAGE_STALE, &page->private))
1353 goto out;
1354
1355 pool = zhdr_to_pool(zhdr);
1356
1357 if (zhdr->mapped_count == 0) {
1358 kref_get(&zhdr->refcount);
1359 if (!list_empty(&zhdr->buddy))
1360 list_del_init(&zhdr->buddy);
1361 spin_lock(&pool->lock);
1362 if (!list_empty(&page->lru))
1363 list_del(&page->lru);
1364 spin_unlock(&pool->lock);
1365 z3fold_page_unlock(zhdr);
1366 return true;
1367 }
1368out:
1369 z3fold_page_unlock(zhdr);
1370 return false;
1371}
1372
1373static int z3fold_page_migrate(struct address_space *mapping, struct page *newpage,
1374 struct page *page, enum migrate_mode mode)
1375{
1376 struct z3fold_header *zhdr, *new_zhdr;
1377 struct z3fold_pool *pool;
1378 struct address_space *new_mapping;
1379
1380 VM_BUG_ON_PAGE(!PageMovable(page), page);
1381 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1382 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1383
1384 zhdr = page_address(page);
1385 pool = zhdr_to_pool(zhdr);
1386
1387 if (!z3fold_page_trylock(zhdr)) {
1388 return -EAGAIN;
1389 }
1390 if (zhdr->mapped_count != 0) {
1391 z3fold_page_unlock(zhdr);
1392 return -EBUSY;
1393 }
1394 if (work_pending(&zhdr->work)) {
1395 z3fold_page_unlock(zhdr);
1396 return -EAGAIN;
1397 }
1398 new_zhdr = page_address(newpage);
1399 memcpy(new_zhdr, zhdr, PAGE_SIZE);
1400 newpage->private = page->private;
1401 page->private = 0;
1402 z3fold_page_unlock(zhdr);
1403 spin_lock_init(&new_zhdr->page_lock);
1404 INIT_WORK(&new_zhdr->work, compact_page_work);
1405 /*
1406 * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1407 * so we only have to reinitialize it.
1408 */
1409 INIT_LIST_HEAD(&new_zhdr->buddy);
1410 new_mapping = page_mapping(page);
1411 __ClearPageMovable(page);
1412 ClearPagePrivate(page);
1413
1414 get_page(newpage);
1415 z3fold_page_lock(new_zhdr);
1416 if (new_zhdr->first_chunks)
1417 encode_handle(new_zhdr, FIRST);
1418 if (new_zhdr->last_chunks)
1419 encode_handle(new_zhdr, LAST);
1420 if (new_zhdr->middle_chunks)
1421 encode_handle(new_zhdr, MIDDLE);
1422 set_bit(NEEDS_COMPACTING, &newpage->private);
1423 new_zhdr->cpu = smp_processor_id();
1424 spin_lock(&pool->lock);
1425 list_add(&newpage->lru, &pool->lru);
1426 spin_unlock(&pool->lock);
1427 __SetPageMovable(newpage, new_mapping);
1428 z3fold_page_unlock(new_zhdr);
1429
1430 queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1431
1432 page_mapcount_reset(page);
1433 put_page(page);
1434 return 0;
1435}
1436
1437static void z3fold_page_putback(struct page *page)
1438{
1439 struct z3fold_header *zhdr;
1440 struct z3fold_pool *pool;
1441
1442 zhdr = page_address(page);
1443 pool = zhdr_to_pool(zhdr);
1444
1445 z3fold_page_lock(zhdr);
1446 if (!list_empty(&zhdr->buddy))
1447 list_del_init(&zhdr->buddy);
1448 INIT_LIST_HEAD(&page->lru);
1449 if (kref_put(&zhdr->refcount, release_z3fold_page_locked)) {
1450 atomic64_dec(&pool->pages_nr);
1451 return;
1452 }
1453 spin_lock(&pool->lock);
1454 list_add(&page->lru, &pool->lru);
1455 spin_unlock(&pool->lock);
1456 z3fold_page_unlock(zhdr);
1457}
1458
1459static const struct address_space_operations z3fold_aops = {
1460 .isolate_page = z3fold_page_isolate,
1461 .migratepage = z3fold_page_migrate,
1462 .putback_page = z3fold_page_putback,
1463};
1464
1465/*****************
1466 * zpool
1467 ****************/
1468
1469static int z3fold_zpool_evict(struct z3fold_pool *pool, unsigned long handle)
1470{
1471 if (pool->zpool && pool->zpool_ops && pool->zpool_ops->evict)
1472 return pool->zpool_ops->evict(pool->zpool, handle);
1473 else
1474 return -ENOENT;
1475}
1476
1477static const struct z3fold_ops z3fold_zpool_ops = {
1478 .evict = z3fold_zpool_evict
1479};
1480
1481static void *z3fold_zpool_create(const char *name, gfp_t gfp,
1482 const struct zpool_ops *zpool_ops,
1483 struct zpool *zpool)
1484{
1485 struct z3fold_pool *pool;
1486
1487 pool = z3fold_create_pool(name, gfp,
1488 zpool_ops ? &z3fold_zpool_ops : NULL);
1489 if (pool) {
1490 pool->zpool = zpool;
1491 pool->zpool_ops = zpool_ops;
1492 }
1493 return pool;
1494}
1495
1496static void z3fold_zpool_destroy(void *pool)
1497{
1498 z3fold_destroy_pool(pool);
1499}
1500
1501static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1502 unsigned long *handle)
1503{
1504 return z3fold_alloc(pool, size, gfp, handle);
1505}
1506static void z3fold_zpool_free(void *pool, unsigned long handle)
1507{
1508 z3fold_free(pool, handle);
1509}
1510
1511static int z3fold_zpool_shrink(void *pool, unsigned int pages,
1512 unsigned int *reclaimed)
1513{
1514 unsigned int total = 0;
1515 int ret = -EINVAL;
1516
1517 while (total < pages) {
1518 ret = z3fold_reclaim_page(pool, 8);
1519 if (ret < 0)
1520 break;
1521 total++;
1522 }
1523
1524 if (reclaimed)
1525 *reclaimed = total;
1526
1527 return ret;
1528}
1529
1530static void *z3fold_zpool_map(void *pool, unsigned long handle,
1531 enum zpool_mapmode mm)
1532{
1533 return z3fold_map(pool, handle);
1534}
1535static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1536{
1537 z3fold_unmap(pool, handle);
1538}
1539
1540static u64 z3fold_zpool_total_size(void *pool)
1541{
1542 return z3fold_get_pool_size(pool) * PAGE_SIZE;
1543}
1544
1545static struct zpool_driver z3fold_zpool_driver = {
1546 .type = "z3fold",
1547 .owner = THIS_MODULE,
1548 .create = z3fold_zpool_create,
1549 .destroy = z3fold_zpool_destroy,
1550 .malloc = z3fold_zpool_malloc,
1551 .free = z3fold_zpool_free,
1552 .shrink = z3fold_zpool_shrink,
1553 .map = z3fold_zpool_map,
1554 .unmap = z3fold_zpool_unmap,
1555 .total_size = z3fold_zpool_total_size,
1556};
1557
1558MODULE_ALIAS("zpool-z3fold");
1559
1560static int __init init_z3fold(void)
1561{
1562 int ret;
1563
1564 /* Make sure the z3fold header is not larger than the page size */
1565 BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE);
1566 ret = z3fold_mount();
1567 if (ret)
1568 return ret;
1569
1570 zpool_register_driver(&z3fold_zpool_driver);
1571
1572 return 0;
1573}
1574
1575static void __exit exit_z3fold(void)
1576{
1577 z3fold_unmount();
1578 zpool_unregister_driver(&z3fold_zpool_driver);
1579}
1580
1581module_init(init_z3fold);
1582module_exit(exit_z3fold);
1583
1584MODULE_LICENSE("GPL");
1585MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1586MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");
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/preempt.h>
38#include <linux/workqueue.h>
39#include <linux/slab.h>
40#include <linux/spinlock.h>
41#include <linux/zpool.h>
42#include <linux/kmemleak.h>
43
44/*
45 * NCHUNKS_ORDER determines the internal allocation granularity, effectively
46 * adjusting internal fragmentation. It also determines the number of
47 * freelists maintained in each pool. NCHUNKS_ORDER of 6 means that the
48 * allocation granularity will be in chunks of size PAGE_SIZE/64. Some chunks
49 * in the beginning of an allocated page are occupied by z3fold header, so
50 * NCHUNKS will be calculated to 63 (or 62 in case CONFIG_DEBUG_SPINLOCK=y),
51 * which shows the max number of free chunks in z3fold page, also there will
52 * be 63, or 62, respectively, freelists per pool.
53 */
54#define NCHUNKS_ORDER 6
55
56#define CHUNK_SHIFT (PAGE_SHIFT - NCHUNKS_ORDER)
57#define CHUNK_SIZE (1 << CHUNK_SHIFT)
58#define ZHDR_SIZE_ALIGNED round_up(sizeof(struct z3fold_header), CHUNK_SIZE)
59#define ZHDR_CHUNKS (ZHDR_SIZE_ALIGNED >> CHUNK_SHIFT)
60#define TOTAL_CHUNKS (PAGE_SIZE >> CHUNK_SHIFT)
61#define NCHUNKS (TOTAL_CHUNKS - ZHDR_CHUNKS)
62
63#define BUDDY_MASK (0x3)
64#define BUDDY_SHIFT 2
65#define SLOTS_ALIGN (0x40)
66
67/*****************
68 * Structures
69*****************/
70struct z3fold_pool;
71
72enum buddy {
73 HEADLESS = 0,
74 FIRST,
75 MIDDLE,
76 LAST,
77 BUDDIES_MAX = LAST
78};
79
80struct z3fold_buddy_slots {
81 /*
82 * we are using BUDDY_MASK in handle_to_buddy etc. so there should
83 * be enough slots to hold all possible variants
84 */
85 unsigned long slot[BUDDY_MASK + 1];
86 unsigned long pool; /* back link */
87 rwlock_t lock;
88};
89#define HANDLE_FLAG_MASK (0x03)
90
91/*
92 * struct z3fold_header - z3fold page metadata occupying first chunks of each
93 * z3fold page, except for HEADLESS pages
94 * @buddy: links the z3fold page into the relevant list in the
95 * pool
96 * @page_lock: per-page lock
97 * @refcount: reference count for the z3fold page
98 * @work: work_struct for page layout optimization
99 * @slots: pointer to the structure holding buddy slots
100 * @pool: pointer to the containing pool
101 * @cpu: CPU which this page "belongs" to
102 * @first_chunks: the size of the first buddy in chunks, 0 if free
103 * @middle_chunks: the size of the middle buddy in chunks, 0 if free
104 * @last_chunks: the size of the last buddy in chunks, 0 if free
105 * @first_num: the starting number (for the first handle)
106 * @mapped_count: the number of objects currently mapped
107 */
108struct z3fold_header {
109 struct list_head buddy;
110 spinlock_t page_lock;
111 struct kref refcount;
112 struct work_struct work;
113 struct z3fold_buddy_slots *slots;
114 struct z3fold_pool *pool;
115 short cpu;
116 unsigned short first_chunks;
117 unsigned short middle_chunks;
118 unsigned short last_chunks;
119 unsigned short start_middle;
120 unsigned short first_num:2;
121 unsigned short mapped_count:2;
122 unsigned short foreign_handles:2;
123};
124
125/**
126 * struct z3fold_pool - stores metadata for each z3fold pool
127 * @name: pool name
128 * @lock: protects pool unbuddied lists
129 * @stale_lock: protects pool stale page list
130 * @unbuddied: per-cpu array of lists tracking z3fold pages that contain 2-
131 * buddies; the list each z3fold page is added to depends on
132 * the size of its free region.
133 * @stale: list of pages marked for freeing
134 * @pages_nr: number of z3fold pages in the pool.
135 * @c_handle: cache for z3fold_buddy_slots allocation
136 * @compact_wq: workqueue for page layout background optimization
137 * @release_wq: workqueue for safe page release
138 * @work: work_struct for safe page release
139 *
140 * This structure is allocated at pool creation time and maintains metadata
141 * pertaining to a particular z3fold pool.
142 */
143struct z3fold_pool {
144 const char *name;
145 spinlock_t lock;
146 spinlock_t stale_lock;
147 struct list_head *unbuddied;
148 struct list_head stale;
149 atomic64_t pages_nr;
150 struct kmem_cache *c_handle;
151 struct workqueue_struct *compact_wq;
152 struct workqueue_struct *release_wq;
153 struct work_struct work;
154};
155
156/*
157 * Internal z3fold page flags
158 */
159enum z3fold_page_flags {
160 PAGE_HEADLESS = 0,
161 MIDDLE_CHUNK_MAPPED,
162 NEEDS_COMPACTING,
163 PAGE_STALE,
164 PAGE_CLAIMED, /* by either reclaim or free */
165 PAGE_MIGRATED, /* page is migrated and soon to be released */
166};
167
168/*
169 * handle flags, go under HANDLE_FLAG_MASK
170 */
171enum z3fold_handle_flags {
172 HANDLES_NOFREE = 0,
173};
174
175/*
176 * Forward declarations
177 */
178static struct z3fold_header *__z3fold_alloc(struct z3fold_pool *, size_t, bool);
179static void compact_page_work(struct work_struct *w);
180
181/*****************
182 * Helpers
183*****************/
184
185/* Converts an allocation size in bytes to size in z3fold chunks */
186static 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
194static inline struct z3fold_buddy_slots *alloc_slots(struct z3fold_pool *pool,
195 gfp_t gfp)
196{
197 struct z3fold_buddy_slots *slots = kmem_cache_zalloc(pool->c_handle,
198 gfp);
199
200 if (slots) {
201 /* It will be freed separately in free_handle(). */
202 kmemleak_not_leak(slots);
203 slots->pool = (unsigned long)pool;
204 rwlock_init(&slots->lock);
205 }
206
207 return slots;
208}
209
210static inline struct z3fold_pool *slots_to_pool(struct z3fold_buddy_slots *s)
211{
212 return (struct z3fold_pool *)(s->pool & ~HANDLE_FLAG_MASK);
213}
214
215static inline struct z3fold_buddy_slots *handle_to_slots(unsigned long handle)
216{
217 return (struct z3fold_buddy_slots *)(handle & ~(SLOTS_ALIGN - 1));
218}
219
220/* Lock a z3fold page */
221static inline void z3fold_page_lock(struct z3fold_header *zhdr)
222{
223 spin_lock(&zhdr->page_lock);
224}
225
226/* Try to lock a z3fold page */
227static inline int z3fold_page_trylock(struct z3fold_header *zhdr)
228{
229 return spin_trylock(&zhdr->page_lock);
230}
231
232/* Unlock a z3fold page */
233static inline void z3fold_page_unlock(struct z3fold_header *zhdr)
234{
235 spin_unlock(&zhdr->page_lock);
236}
237
238/* return locked z3fold page if it's not headless */
239static inline struct z3fold_header *get_z3fold_header(unsigned long handle)
240{
241 struct z3fold_buddy_slots *slots;
242 struct z3fold_header *zhdr;
243 int locked = 0;
244
245 if (!(handle & (1 << PAGE_HEADLESS))) {
246 slots = handle_to_slots(handle);
247 do {
248 unsigned long addr;
249
250 read_lock(&slots->lock);
251 addr = *(unsigned long *)handle;
252 zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
253 locked = z3fold_page_trylock(zhdr);
254 read_unlock(&slots->lock);
255 if (locked) {
256 struct page *page = virt_to_page(zhdr);
257
258 if (!test_bit(PAGE_MIGRATED, &page->private))
259 break;
260 z3fold_page_unlock(zhdr);
261 }
262 cpu_relax();
263 } while (true);
264 } else {
265 zhdr = (struct z3fold_header *)(handle & PAGE_MASK);
266 }
267
268 return zhdr;
269}
270
271static inline void put_z3fold_header(struct z3fold_header *zhdr)
272{
273 struct page *page = virt_to_page(zhdr);
274
275 if (!test_bit(PAGE_HEADLESS, &page->private))
276 z3fold_page_unlock(zhdr);
277}
278
279static inline void free_handle(unsigned long handle, struct z3fold_header *zhdr)
280{
281 struct z3fold_buddy_slots *slots;
282 int i;
283 bool is_free;
284
285 if (WARN_ON(*(unsigned long *)handle == 0))
286 return;
287
288 slots = handle_to_slots(handle);
289 write_lock(&slots->lock);
290 *(unsigned long *)handle = 0;
291
292 if (test_bit(HANDLES_NOFREE, &slots->pool)) {
293 write_unlock(&slots->lock);
294 return; /* simple case, nothing else to do */
295 }
296
297 if (zhdr->slots != slots)
298 zhdr->foreign_handles--;
299
300 is_free = true;
301 for (i = 0; i <= BUDDY_MASK; i++) {
302 if (slots->slot[i]) {
303 is_free = false;
304 break;
305 }
306 }
307 write_unlock(&slots->lock);
308
309 if (is_free) {
310 struct z3fold_pool *pool = slots_to_pool(slots);
311
312 if (zhdr->slots == slots)
313 zhdr->slots = NULL;
314 kmem_cache_free(pool->c_handle, slots);
315 }
316}
317
318/* Initializes the z3fold header of a newly allocated z3fold page */
319static struct z3fold_header *init_z3fold_page(struct page *page, bool headless,
320 struct z3fold_pool *pool, gfp_t gfp)
321{
322 struct z3fold_header *zhdr = page_address(page);
323 struct z3fold_buddy_slots *slots;
324
325 clear_bit(PAGE_HEADLESS, &page->private);
326 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
327 clear_bit(NEEDS_COMPACTING, &page->private);
328 clear_bit(PAGE_STALE, &page->private);
329 clear_bit(PAGE_CLAIMED, &page->private);
330 clear_bit(PAGE_MIGRATED, &page->private);
331 if (headless)
332 return zhdr;
333
334 slots = alloc_slots(pool, gfp);
335 if (!slots)
336 return NULL;
337
338 memset(zhdr, 0, sizeof(*zhdr));
339 spin_lock_init(&zhdr->page_lock);
340 kref_init(&zhdr->refcount);
341 zhdr->cpu = -1;
342 zhdr->slots = slots;
343 zhdr->pool = pool;
344 INIT_LIST_HEAD(&zhdr->buddy);
345 INIT_WORK(&zhdr->work, compact_page_work);
346 return zhdr;
347}
348
349/* Resets the struct page fields and frees the page */
350static void free_z3fold_page(struct page *page, bool headless)
351{
352 if (!headless) {
353 lock_page(page);
354 __ClearPageMovable(page);
355 unlock_page(page);
356 }
357 __free_page(page);
358}
359
360/* Helper function to build the index */
361static inline int __idx(struct z3fold_header *zhdr, enum buddy bud)
362{
363 return (bud + zhdr->first_num) & BUDDY_MASK;
364}
365
366/*
367 * Encodes the handle of a particular buddy within a z3fold page
368 * Pool lock should be held as this function accesses first_num
369 */
370static unsigned long __encode_handle(struct z3fold_header *zhdr,
371 struct z3fold_buddy_slots *slots,
372 enum buddy bud)
373{
374 unsigned long h = (unsigned long)zhdr;
375 int idx = 0;
376
377 /*
378 * For a headless page, its handle is its pointer with the extra
379 * PAGE_HEADLESS bit set
380 */
381 if (bud == HEADLESS)
382 return h | (1 << PAGE_HEADLESS);
383
384 /* otherwise, return pointer to encoded handle */
385 idx = __idx(zhdr, bud);
386 h += idx;
387 if (bud == LAST)
388 h |= (zhdr->last_chunks << BUDDY_SHIFT);
389
390 write_lock(&slots->lock);
391 slots->slot[idx] = h;
392 write_unlock(&slots->lock);
393 return (unsigned long)&slots->slot[idx];
394}
395
396static unsigned long encode_handle(struct z3fold_header *zhdr, enum buddy bud)
397{
398 return __encode_handle(zhdr, zhdr->slots, bud);
399}
400
401/* only for LAST bud, returns zero otherwise */
402static unsigned short handle_to_chunks(unsigned long handle)
403{
404 struct z3fold_buddy_slots *slots = handle_to_slots(handle);
405 unsigned long addr;
406
407 read_lock(&slots->lock);
408 addr = *(unsigned long *)handle;
409 read_unlock(&slots->lock);
410 return (addr & ~PAGE_MASK) >> BUDDY_SHIFT;
411}
412
413/*
414 * (handle & BUDDY_MASK) < zhdr->first_num is possible in encode_handle
415 * but that doesn't matter. because the masking will result in the
416 * correct buddy number.
417 */
418static enum buddy handle_to_buddy(unsigned long handle)
419{
420 struct z3fold_header *zhdr;
421 struct z3fold_buddy_slots *slots = handle_to_slots(handle);
422 unsigned long addr;
423
424 read_lock(&slots->lock);
425 WARN_ON(handle & (1 << PAGE_HEADLESS));
426 addr = *(unsigned long *)handle;
427 read_unlock(&slots->lock);
428 zhdr = (struct z3fold_header *)(addr & PAGE_MASK);
429 return (addr - zhdr->first_num) & BUDDY_MASK;
430}
431
432static inline struct z3fold_pool *zhdr_to_pool(struct z3fold_header *zhdr)
433{
434 return zhdr->pool;
435}
436
437static void __release_z3fold_page(struct z3fold_header *zhdr, bool locked)
438{
439 struct page *page = virt_to_page(zhdr);
440 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
441
442 WARN_ON(!list_empty(&zhdr->buddy));
443 set_bit(PAGE_STALE, &page->private);
444 clear_bit(NEEDS_COMPACTING, &page->private);
445 spin_lock(&pool->lock);
446 spin_unlock(&pool->lock);
447
448 if (locked)
449 z3fold_page_unlock(zhdr);
450
451 spin_lock(&pool->stale_lock);
452 list_add(&zhdr->buddy, &pool->stale);
453 queue_work(pool->release_wq, &pool->work);
454 spin_unlock(&pool->stale_lock);
455
456 atomic64_dec(&pool->pages_nr);
457}
458
459static void release_z3fold_page_locked(struct kref *ref)
460{
461 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
462 refcount);
463 WARN_ON(z3fold_page_trylock(zhdr));
464 __release_z3fold_page(zhdr, true);
465}
466
467static void release_z3fold_page_locked_list(struct kref *ref)
468{
469 struct z3fold_header *zhdr = container_of(ref, struct z3fold_header,
470 refcount);
471 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
472
473 spin_lock(&pool->lock);
474 list_del_init(&zhdr->buddy);
475 spin_unlock(&pool->lock);
476
477 WARN_ON(z3fold_page_trylock(zhdr));
478 __release_z3fold_page(zhdr, true);
479}
480
481static inline int put_z3fold_locked(struct z3fold_header *zhdr)
482{
483 return kref_put(&zhdr->refcount, release_z3fold_page_locked);
484}
485
486static inline int put_z3fold_locked_list(struct z3fold_header *zhdr)
487{
488 return kref_put(&zhdr->refcount, release_z3fold_page_locked_list);
489}
490
491static void free_pages_work(struct work_struct *w)
492{
493 struct z3fold_pool *pool = container_of(w, struct z3fold_pool, work);
494
495 spin_lock(&pool->stale_lock);
496 while (!list_empty(&pool->stale)) {
497 struct z3fold_header *zhdr = list_first_entry(&pool->stale,
498 struct z3fold_header, buddy);
499 struct page *page = virt_to_page(zhdr);
500
501 list_del(&zhdr->buddy);
502 if (WARN_ON(!test_bit(PAGE_STALE, &page->private)))
503 continue;
504 spin_unlock(&pool->stale_lock);
505 cancel_work_sync(&zhdr->work);
506 free_z3fold_page(page, false);
507 cond_resched();
508 spin_lock(&pool->stale_lock);
509 }
510 spin_unlock(&pool->stale_lock);
511}
512
513/*
514 * Returns the number of free chunks in a z3fold page.
515 * NB: can't be used with HEADLESS pages.
516 */
517static int num_free_chunks(struct z3fold_header *zhdr)
518{
519 int nfree;
520 /*
521 * If there is a middle object, pick up the bigger free space
522 * either before or after it. Otherwise just subtract the number
523 * of chunks occupied by the first and the last objects.
524 */
525 if (zhdr->middle_chunks != 0) {
526 int nfree_before = zhdr->first_chunks ?
527 0 : zhdr->start_middle - ZHDR_CHUNKS;
528 int nfree_after = zhdr->last_chunks ?
529 0 : TOTAL_CHUNKS -
530 (zhdr->start_middle + zhdr->middle_chunks);
531 nfree = max(nfree_before, nfree_after);
532 } else
533 nfree = NCHUNKS - zhdr->first_chunks - zhdr->last_chunks;
534 return nfree;
535}
536
537/* Add to the appropriate unbuddied list */
538static inline void add_to_unbuddied(struct z3fold_pool *pool,
539 struct z3fold_header *zhdr)
540{
541 if (zhdr->first_chunks == 0 || zhdr->last_chunks == 0 ||
542 zhdr->middle_chunks == 0) {
543 struct list_head *unbuddied;
544 int freechunks = num_free_chunks(zhdr);
545
546 migrate_disable();
547 unbuddied = this_cpu_ptr(pool->unbuddied);
548 spin_lock(&pool->lock);
549 list_add(&zhdr->buddy, &unbuddied[freechunks]);
550 spin_unlock(&pool->lock);
551 zhdr->cpu = smp_processor_id();
552 migrate_enable();
553 }
554}
555
556static inline enum buddy get_free_buddy(struct z3fold_header *zhdr, int chunks)
557{
558 enum buddy bud = HEADLESS;
559
560 if (zhdr->middle_chunks) {
561 if (!zhdr->first_chunks &&
562 chunks <= zhdr->start_middle - ZHDR_CHUNKS)
563 bud = FIRST;
564 else if (!zhdr->last_chunks)
565 bud = LAST;
566 } else {
567 if (!zhdr->first_chunks)
568 bud = FIRST;
569 else if (!zhdr->last_chunks)
570 bud = LAST;
571 else
572 bud = MIDDLE;
573 }
574
575 return bud;
576}
577
578static inline void *mchunk_memmove(struct z3fold_header *zhdr,
579 unsigned short dst_chunk)
580{
581 void *beg = zhdr;
582 return memmove(beg + (dst_chunk << CHUNK_SHIFT),
583 beg + (zhdr->start_middle << CHUNK_SHIFT),
584 zhdr->middle_chunks << CHUNK_SHIFT);
585}
586
587static inline bool buddy_single(struct z3fold_header *zhdr)
588{
589 return !((zhdr->first_chunks && zhdr->middle_chunks) ||
590 (zhdr->first_chunks && zhdr->last_chunks) ||
591 (zhdr->middle_chunks && zhdr->last_chunks));
592}
593
594static struct z3fold_header *compact_single_buddy(struct z3fold_header *zhdr)
595{
596 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
597 void *p = zhdr;
598 unsigned long old_handle = 0;
599 size_t sz = 0;
600 struct z3fold_header *new_zhdr = NULL;
601 int first_idx = __idx(zhdr, FIRST);
602 int middle_idx = __idx(zhdr, MIDDLE);
603 int last_idx = __idx(zhdr, LAST);
604 unsigned short *moved_chunks = NULL;
605
606 /*
607 * No need to protect slots here -- all the slots are "local" and
608 * the page lock is already taken
609 */
610 if (zhdr->first_chunks && zhdr->slots->slot[first_idx]) {
611 p += ZHDR_SIZE_ALIGNED;
612 sz = zhdr->first_chunks << CHUNK_SHIFT;
613 old_handle = (unsigned long)&zhdr->slots->slot[first_idx];
614 moved_chunks = &zhdr->first_chunks;
615 } else if (zhdr->middle_chunks && zhdr->slots->slot[middle_idx]) {
616 p += zhdr->start_middle << CHUNK_SHIFT;
617 sz = zhdr->middle_chunks << CHUNK_SHIFT;
618 old_handle = (unsigned long)&zhdr->slots->slot[middle_idx];
619 moved_chunks = &zhdr->middle_chunks;
620 } else if (zhdr->last_chunks && zhdr->slots->slot[last_idx]) {
621 p += PAGE_SIZE - (zhdr->last_chunks << CHUNK_SHIFT);
622 sz = zhdr->last_chunks << CHUNK_SHIFT;
623 old_handle = (unsigned long)&zhdr->slots->slot[last_idx];
624 moved_chunks = &zhdr->last_chunks;
625 }
626
627 if (sz > 0) {
628 enum buddy new_bud = HEADLESS;
629 short chunks = size_to_chunks(sz);
630 void *q;
631
632 new_zhdr = __z3fold_alloc(pool, sz, false);
633 if (!new_zhdr)
634 return NULL;
635
636 if (WARN_ON(new_zhdr == zhdr))
637 goto out_fail;
638
639 new_bud = get_free_buddy(new_zhdr, chunks);
640 q = new_zhdr;
641 switch (new_bud) {
642 case FIRST:
643 new_zhdr->first_chunks = chunks;
644 q += ZHDR_SIZE_ALIGNED;
645 break;
646 case MIDDLE:
647 new_zhdr->middle_chunks = chunks;
648 new_zhdr->start_middle =
649 new_zhdr->first_chunks + ZHDR_CHUNKS;
650 q += new_zhdr->start_middle << CHUNK_SHIFT;
651 break;
652 case LAST:
653 new_zhdr->last_chunks = chunks;
654 q += PAGE_SIZE - (new_zhdr->last_chunks << CHUNK_SHIFT);
655 break;
656 default:
657 goto out_fail;
658 }
659 new_zhdr->foreign_handles++;
660 memcpy(q, p, sz);
661 write_lock(&zhdr->slots->lock);
662 *(unsigned long *)old_handle = (unsigned long)new_zhdr +
663 __idx(new_zhdr, new_bud);
664 if (new_bud == LAST)
665 *(unsigned long *)old_handle |=
666 (new_zhdr->last_chunks << BUDDY_SHIFT);
667 write_unlock(&zhdr->slots->lock);
668 add_to_unbuddied(pool, new_zhdr);
669 z3fold_page_unlock(new_zhdr);
670
671 *moved_chunks = 0;
672 }
673
674 return new_zhdr;
675
676out_fail:
677 if (new_zhdr && !put_z3fold_locked(new_zhdr)) {
678 add_to_unbuddied(pool, new_zhdr);
679 z3fold_page_unlock(new_zhdr);
680 }
681 return NULL;
682
683}
684
685#define BIG_CHUNK_GAP 3
686/* Has to be called with lock held */
687static int z3fold_compact_page(struct z3fold_header *zhdr)
688{
689 struct page *page = virt_to_page(zhdr);
690
691 if (test_bit(MIDDLE_CHUNK_MAPPED, &page->private))
692 return 0; /* can't move middle chunk, it's used */
693
694 if (unlikely(PageIsolated(page)))
695 return 0;
696
697 if (zhdr->middle_chunks == 0)
698 return 0; /* nothing to compact */
699
700 if (zhdr->first_chunks == 0 && zhdr->last_chunks == 0) {
701 /* move to the beginning */
702 mchunk_memmove(zhdr, ZHDR_CHUNKS);
703 zhdr->first_chunks = zhdr->middle_chunks;
704 zhdr->middle_chunks = 0;
705 zhdr->start_middle = 0;
706 zhdr->first_num++;
707 return 1;
708 }
709
710 /*
711 * moving data is expensive, so let's only do that if
712 * there's substantial gain (at least BIG_CHUNK_GAP chunks)
713 */
714 if (zhdr->first_chunks != 0 && zhdr->last_chunks == 0 &&
715 zhdr->start_middle - (zhdr->first_chunks + ZHDR_CHUNKS) >=
716 BIG_CHUNK_GAP) {
717 mchunk_memmove(zhdr, zhdr->first_chunks + ZHDR_CHUNKS);
718 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
719 return 1;
720 } else if (zhdr->last_chunks != 0 && zhdr->first_chunks == 0 &&
721 TOTAL_CHUNKS - (zhdr->last_chunks + zhdr->start_middle
722 + zhdr->middle_chunks) >=
723 BIG_CHUNK_GAP) {
724 unsigned short new_start = TOTAL_CHUNKS - zhdr->last_chunks -
725 zhdr->middle_chunks;
726 mchunk_memmove(zhdr, new_start);
727 zhdr->start_middle = new_start;
728 return 1;
729 }
730
731 return 0;
732}
733
734static void do_compact_page(struct z3fold_header *zhdr, bool locked)
735{
736 struct z3fold_pool *pool = zhdr_to_pool(zhdr);
737 struct page *page;
738
739 page = virt_to_page(zhdr);
740 if (locked)
741 WARN_ON(z3fold_page_trylock(zhdr));
742 else
743 z3fold_page_lock(zhdr);
744 if (WARN_ON(!test_and_clear_bit(NEEDS_COMPACTING, &page->private))) {
745 z3fold_page_unlock(zhdr);
746 return;
747 }
748 spin_lock(&pool->lock);
749 list_del_init(&zhdr->buddy);
750 spin_unlock(&pool->lock);
751
752 if (put_z3fold_locked(zhdr))
753 return;
754
755 if (test_bit(PAGE_STALE, &page->private) ||
756 test_and_set_bit(PAGE_CLAIMED, &page->private)) {
757 z3fold_page_unlock(zhdr);
758 return;
759 }
760
761 if (!zhdr->foreign_handles && buddy_single(zhdr) &&
762 zhdr->mapped_count == 0 && compact_single_buddy(zhdr)) {
763 if (!put_z3fold_locked(zhdr)) {
764 clear_bit(PAGE_CLAIMED, &page->private);
765 z3fold_page_unlock(zhdr);
766 }
767 return;
768 }
769
770 z3fold_compact_page(zhdr);
771 add_to_unbuddied(pool, zhdr);
772 clear_bit(PAGE_CLAIMED, &page->private);
773 z3fold_page_unlock(zhdr);
774}
775
776static void compact_page_work(struct work_struct *w)
777{
778 struct z3fold_header *zhdr = container_of(w, struct z3fold_header,
779 work);
780
781 do_compact_page(zhdr, false);
782}
783
784/* returns _locked_ z3fold page header or NULL */
785static inline struct z3fold_header *__z3fold_alloc(struct z3fold_pool *pool,
786 size_t size, bool can_sleep)
787{
788 struct z3fold_header *zhdr = NULL;
789 struct page *page;
790 struct list_head *unbuddied;
791 int chunks = size_to_chunks(size), i;
792
793lookup:
794 migrate_disable();
795 /* First, try to find an unbuddied z3fold page. */
796 unbuddied = this_cpu_ptr(pool->unbuddied);
797 for_each_unbuddied_list(i, chunks) {
798 struct list_head *l = &unbuddied[i];
799
800 zhdr = list_first_entry_or_null(READ_ONCE(l),
801 struct z3fold_header, buddy);
802
803 if (!zhdr)
804 continue;
805
806 /* Re-check under lock. */
807 spin_lock(&pool->lock);
808 if (unlikely(zhdr != list_first_entry(READ_ONCE(l),
809 struct z3fold_header, buddy)) ||
810 !z3fold_page_trylock(zhdr)) {
811 spin_unlock(&pool->lock);
812 zhdr = NULL;
813 migrate_enable();
814 if (can_sleep)
815 cond_resched();
816 goto lookup;
817 }
818 list_del_init(&zhdr->buddy);
819 zhdr->cpu = -1;
820 spin_unlock(&pool->lock);
821
822 page = virt_to_page(zhdr);
823 if (test_bit(NEEDS_COMPACTING, &page->private) ||
824 test_bit(PAGE_CLAIMED, &page->private)) {
825 z3fold_page_unlock(zhdr);
826 zhdr = NULL;
827 migrate_enable();
828 if (can_sleep)
829 cond_resched();
830 goto lookup;
831 }
832
833 /*
834 * this page could not be removed from its unbuddied
835 * list while pool lock was held, and then we've taken
836 * page lock so kref_put could not be called before
837 * we got here, so it's safe to just call kref_get()
838 */
839 kref_get(&zhdr->refcount);
840 break;
841 }
842 migrate_enable();
843
844 if (!zhdr) {
845 int cpu;
846
847 /* look for _exact_ match on other cpus' lists */
848 for_each_online_cpu(cpu) {
849 struct list_head *l;
850
851 unbuddied = per_cpu_ptr(pool->unbuddied, cpu);
852 spin_lock(&pool->lock);
853 l = &unbuddied[chunks];
854
855 zhdr = list_first_entry_or_null(READ_ONCE(l),
856 struct z3fold_header, buddy);
857
858 if (!zhdr || !z3fold_page_trylock(zhdr)) {
859 spin_unlock(&pool->lock);
860 zhdr = NULL;
861 continue;
862 }
863 list_del_init(&zhdr->buddy);
864 zhdr->cpu = -1;
865 spin_unlock(&pool->lock);
866
867 page = virt_to_page(zhdr);
868 if (test_bit(NEEDS_COMPACTING, &page->private) ||
869 test_bit(PAGE_CLAIMED, &page->private)) {
870 z3fold_page_unlock(zhdr);
871 zhdr = NULL;
872 if (can_sleep)
873 cond_resched();
874 continue;
875 }
876 kref_get(&zhdr->refcount);
877 break;
878 }
879 }
880
881 if (zhdr && !zhdr->slots) {
882 zhdr->slots = alloc_slots(pool, GFP_ATOMIC);
883 if (!zhdr->slots)
884 goto out_fail;
885 }
886 return zhdr;
887
888out_fail:
889 if (!put_z3fold_locked(zhdr)) {
890 add_to_unbuddied(pool, zhdr);
891 z3fold_page_unlock(zhdr);
892 }
893 return NULL;
894}
895
896/*
897 * API Functions
898 */
899
900/**
901 * z3fold_create_pool() - create a new z3fold pool
902 * @name: pool name
903 * @gfp: gfp flags when allocating the z3fold pool structure
904 *
905 * Return: pointer to the new z3fold pool or NULL if the metadata allocation
906 * failed.
907 */
908static struct z3fold_pool *z3fold_create_pool(const char *name, gfp_t gfp)
909{
910 struct z3fold_pool *pool = NULL;
911 int i, cpu;
912
913 pool = kzalloc(sizeof(struct z3fold_pool), gfp);
914 if (!pool)
915 goto out;
916 pool->c_handle = kmem_cache_create("z3fold_handle",
917 sizeof(struct z3fold_buddy_slots),
918 SLOTS_ALIGN, 0, NULL);
919 if (!pool->c_handle)
920 goto out_c;
921 spin_lock_init(&pool->lock);
922 spin_lock_init(&pool->stale_lock);
923 pool->unbuddied = __alloc_percpu(sizeof(struct list_head) * NCHUNKS,
924 __alignof__(struct list_head));
925 if (!pool->unbuddied)
926 goto out_pool;
927 for_each_possible_cpu(cpu) {
928 struct list_head *unbuddied =
929 per_cpu_ptr(pool->unbuddied, cpu);
930 for_each_unbuddied_list(i, 0)
931 INIT_LIST_HEAD(&unbuddied[i]);
932 }
933 INIT_LIST_HEAD(&pool->stale);
934 atomic64_set(&pool->pages_nr, 0);
935 pool->name = name;
936 pool->compact_wq = create_singlethread_workqueue(pool->name);
937 if (!pool->compact_wq)
938 goto out_unbuddied;
939 pool->release_wq = create_singlethread_workqueue(pool->name);
940 if (!pool->release_wq)
941 goto out_wq;
942 INIT_WORK(&pool->work, free_pages_work);
943 return pool;
944
945out_wq:
946 destroy_workqueue(pool->compact_wq);
947out_unbuddied:
948 free_percpu(pool->unbuddied);
949out_pool:
950 kmem_cache_destroy(pool->c_handle);
951out_c:
952 kfree(pool);
953out:
954 return NULL;
955}
956
957/**
958 * z3fold_destroy_pool() - destroys an existing z3fold pool
959 * @pool: the z3fold pool to be destroyed
960 *
961 * The pool should be emptied before this function is called.
962 */
963static void z3fold_destroy_pool(struct z3fold_pool *pool)
964{
965 kmem_cache_destroy(pool->c_handle);
966
967 /*
968 * We need to destroy pool->compact_wq before pool->release_wq,
969 * as any pending work on pool->compact_wq will call
970 * queue_work(pool->release_wq, &pool->work).
971 *
972 * There are still outstanding pages until both workqueues are drained,
973 * so we cannot unregister migration until then.
974 */
975
976 destroy_workqueue(pool->compact_wq);
977 destroy_workqueue(pool->release_wq);
978 free_percpu(pool->unbuddied);
979 kfree(pool);
980}
981
982static const struct movable_operations z3fold_mops;
983
984/**
985 * z3fold_alloc() - allocates a region of a given size
986 * @pool: z3fold pool from which to allocate
987 * @size: size in bytes of the desired allocation
988 * @gfp: gfp flags used if the pool needs to grow
989 * @handle: handle of the new allocation
990 *
991 * This function will attempt to find a free region in the pool large enough to
992 * satisfy the allocation request. A search of the unbuddied lists is
993 * performed first. If no suitable free region is found, then a new page is
994 * allocated and added to the pool to satisfy the request.
995 *
996 * Return: 0 if success and handle is set, otherwise -EINVAL if the size or
997 * gfp arguments are invalid or -ENOMEM if the pool was unable to allocate
998 * a new page.
999 */
1000static int z3fold_alloc(struct z3fold_pool *pool, size_t size, gfp_t gfp,
1001 unsigned long *handle)
1002{
1003 int chunks = size_to_chunks(size);
1004 struct z3fold_header *zhdr = NULL;
1005 struct page *page = NULL;
1006 enum buddy bud;
1007 bool can_sleep = gfpflags_allow_blocking(gfp);
1008
1009 if (!size || (gfp & __GFP_HIGHMEM))
1010 return -EINVAL;
1011
1012 if (size > PAGE_SIZE)
1013 return -ENOSPC;
1014
1015 if (size > PAGE_SIZE - ZHDR_SIZE_ALIGNED - CHUNK_SIZE)
1016 bud = HEADLESS;
1017 else {
1018retry:
1019 zhdr = __z3fold_alloc(pool, size, can_sleep);
1020 if (zhdr) {
1021 bud = get_free_buddy(zhdr, chunks);
1022 if (bud == HEADLESS) {
1023 if (!put_z3fold_locked(zhdr))
1024 z3fold_page_unlock(zhdr);
1025 pr_err("No free chunks in unbuddied\n");
1026 WARN_ON(1);
1027 goto retry;
1028 }
1029 page = virt_to_page(zhdr);
1030 goto found;
1031 }
1032 bud = FIRST;
1033 }
1034
1035 page = alloc_page(gfp);
1036 if (!page)
1037 return -ENOMEM;
1038
1039 zhdr = init_z3fold_page(page, bud == HEADLESS, pool, gfp);
1040 if (!zhdr) {
1041 __free_page(page);
1042 return -ENOMEM;
1043 }
1044 atomic64_inc(&pool->pages_nr);
1045
1046 if (bud == HEADLESS) {
1047 set_bit(PAGE_HEADLESS, &page->private);
1048 goto headless;
1049 }
1050 if (can_sleep) {
1051 lock_page(page);
1052 __SetPageMovable(page, &z3fold_mops);
1053 unlock_page(page);
1054 } else {
1055 WARN_ON(!trylock_page(page));
1056 __SetPageMovable(page, &z3fold_mops);
1057 unlock_page(page);
1058 }
1059 z3fold_page_lock(zhdr);
1060
1061found:
1062 if (bud == FIRST)
1063 zhdr->first_chunks = chunks;
1064 else if (bud == LAST)
1065 zhdr->last_chunks = chunks;
1066 else {
1067 zhdr->middle_chunks = chunks;
1068 zhdr->start_middle = zhdr->first_chunks + ZHDR_CHUNKS;
1069 }
1070 add_to_unbuddied(pool, zhdr);
1071
1072headless:
1073 spin_lock(&pool->lock);
1074 *handle = encode_handle(zhdr, bud);
1075 spin_unlock(&pool->lock);
1076 if (bud != HEADLESS)
1077 z3fold_page_unlock(zhdr);
1078
1079 return 0;
1080}
1081
1082/**
1083 * z3fold_free() - frees the allocation associated with the given handle
1084 * @pool: pool in which the allocation resided
1085 * @handle: handle associated with the allocation returned by z3fold_alloc()
1086 *
1087 * In the case that the z3fold page in which the allocation resides is under
1088 * reclaim, as indicated by the PAGE_CLAIMED flag being set, this function
1089 * only sets the first|middle|last_chunks to 0. The page is actually freed
1090 * once all buddies are evicted (see z3fold_reclaim_page() below).
1091 */
1092static void z3fold_free(struct z3fold_pool *pool, unsigned long handle)
1093{
1094 struct z3fold_header *zhdr;
1095 struct page *page;
1096 enum buddy bud;
1097 bool page_claimed;
1098
1099 zhdr = get_z3fold_header(handle);
1100 page = virt_to_page(zhdr);
1101 page_claimed = test_and_set_bit(PAGE_CLAIMED, &page->private);
1102
1103 if (test_bit(PAGE_HEADLESS, &page->private)) {
1104 /* if a headless page is under reclaim, just leave.
1105 * NB: we use test_and_set_bit for a reason: if the bit
1106 * has not been set before, we release this page
1107 * immediately so we don't care about its value any more.
1108 */
1109 if (!page_claimed) {
1110 put_z3fold_header(zhdr);
1111 free_z3fold_page(page, true);
1112 atomic64_dec(&pool->pages_nr);
1113 }
1114 return;
1115 }
1116
1117 /* Non-headless case */
1118 bud = handle_to_buddy(handle);
1119
1120 switch (bud) {
1121 case FIRST:
1122 zhdr->first_chunks = 0;
1123 break;
1124 case MIDDLE:
1125 zhdr->middle_chunks = 0;
1126 break;
1127 case LAST:
1128 zhdr->last_chunks = 0;
1129 break;
1130 default:
1131 pr_err("%s: unknown bud %d\n", __func__, bud);
1132 WARN_ON(1);
1133 put_z3fold_header(zhdr);
1134 return;
1135 }
1136
1137 if (!page_claimed)
1138 free_handle(handle, zhdr);
1139 if (put_z3fold_locked_list(zhdr))
1140 return;
1141 if (page_claimed) {
1142 /* the page has not been claimed by us */
1143 put_z3fold_header(zhdr);
1144 return;
1145 }
1146 if (test_and_set_bit(NEEDS_COMPACTING, &page->private)) {
1147 clear_bit(PAGE_CLAIMED, &page->private);
1148 put_z3fold_header(zhdr);
1149 return;
1150 }
1151 if (zhdr->cpu < 0 || !cpu_online(zhdr->cpu)) {
1152 zhdr->cpu = -1;
1153 kref_get(&zhdr->refcount);
1154 clear_bit(PAGE_CLAIMED, &page->private);
1155 do_compact_page(zhdr, true);
1156 return;
1157 }
1158 kref_get(&zhdr->refcount);
1159 clear_bit(PAGE_CLAIMED, &page->private);
1160 queue_work_on(zhdr->cpu, pool->compact_wq, &zhdr->work);
1161 put_z3fold_header(zhdr);
1162}
1163
1164/**
1165 * z3fold_map() - maps the allocation associated with the given handle
1166 * @pool: pool in which the allocation resides
1167 * @handle: handle associated with the allocation to be mapped
1168 *
1169 * Extracts the buddy number from handle and constructs the pointer to the
1170 * correct starting chunk within the page.
1171 *
1172 * Returns: a pointer to the mapped allocation
1173 */
1174static void *z3fold_map(struct z3fold_pool *pool, unsigned long handle)
1175{
1176 struct z3fold_header *zhdr;
1177 struct page *page;
1178 void *addr;
1179 enum buddy buddy;
1180
1181 zhdr = get_z3fold_header(handle);
1182 addr = zhdr;
1183 page = virt_to_page(zhdr);
1184
1185 if (test_bit(PAGE_HEADLESS, &page->private))
1186 goto out;
1187
1188 buddy = handle_to_buddy(handle);
1189 switch (buddy) {
1190 case FIRST:
1191 addr += ZHDR_SIZE_ALIGNED;
1192 break;
1193 case MIDDLE:
1194 addr += zhdr->start_middle << CHUNK_SHIFT;
1195 set_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1196 break;
1197 case LAST:
1198 addr += PAGE_SIZE - (handle_to_chunks(handle) << CHUNK_SHIFT);
1199 break;
1200 default:
1201 pr_err("unknown buddy id %d\n", buddy);
1202 WARN_ON(1);
1203 addr = NULL;
1204 break;
1205 }
1206
1207 if (addr)
1208 zhdr->mapped_count++;
1209out:
1210 put_z3fold_header(zhdr);
1211 return addr;
1212}
1213
1214/**
1215 * z3fold_unmap() - unmaps the allocation associated with the given handle
1216 * @pool: pool in which the allocation resides
1217 * @handle: handle associated with the allocation to be unmapped
1218 */
1219static void z3fold_unmap(struct z3fold_pool *pool, unsigned long handle)
1220{
1221 struct z3fold_header *zhdr;
1222 struct page *page;
1223 enum buddy buddy;
1224
1225 zhdr = get_z3fold_header(handle);
1226 page = virt_to_page(zhdr);
1227
1228 if (test_bit(PAGE_HEADLESS, &page->private))
1229 return;
1230
1231 buddy = handle_to_buddy(handle);
1232 if (buddy == MIDDLE)
1233 clear_bit(MIDDLE_CHUNK_MAPPED, &page->private);
1234 zhdr->mapped_count--;
1235 put_z3fold_header(zhdr);
1236}
1237
1238/**
1239 * z3fold_get_pool_size() - gets the z3fold pool size in pages
1240 * @pool: pool whose size is being queried
1241 *
1242 * Returns: size in pages of the given pool.
1243 */
1244static u64 z3fold_get_pool_size(struct z3fold_pool *pool)
1245{
1246 return atomic64_read(&pool->pages_nr);
1247}
1248
1249static bool z3fold_page_isolate(struct page *page, isolate_mode_t mode)
1250{
1251 struct z3fold_header *zhdr;
1252 struct z3fold_pool *pool;
1253
1254 VM_BUG_ON_PAGE(PageIsolated(page), page);
1255
1256 if (test_bit(PAGE_HEADLESS, &page->private))
1257 return false;
1258
1259 zhdr = page_address(page);
1260 z3fold_page_lock(zhdr);
1261 if (test_bit(NEEDS_COMPACTING, &page->private) ||
1262 test_bit(PAGE_STALE, &page->private))
1263 goto out;
1264
1265 if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0)
1266 goto out;
1267
1268 if (test_and_set_bit(PAGE_CLAIMED, &page->private))
1269 goto out;
1270 pool = zhdr_to_pool(zhdr);
1271 spin_lock(&pool->lock);
1272 if (!list_empty(&zhdr->buddy))
1273 list_del_init(&zhdr->buddy);
1274 spin_unlock(&pool->lock);
1275
1276 kref_get(&zhdr->refcount);
1277 z3fold_page_unlock(zhdr);
1278 return true;
1279
1280out:
1281 z3fold_page_unlock(zhdr);
1282 return false;
1283}
1284
1285static int z3fold_page_migrate(struct page *newpage, struct page *page,
1286 enum migrate_mode mode)
1287{
1288 struct z3fold_header *zhdr, *new_zhdr;
1289 struct z3fold_pool *pool;
1290
1291 VM_BUG_ON_PAGE(!PageIsolated(page), page);
1292 VM_BUG_ON_PAGE(!test_bit(PAGE_CLAIMED, &page->private), page);
1293 VM_BUG_ON_PAGE(!PageLocked(newpage), newpage);
1294
1295 zhdr = page_address(page);
1296 pool = zhdr_to_pool(zhdr);
1297
1298 if (!z3fold_page_trylock(zhdr))
1299 return -EAGAIN;
1300 if (zhdr->mapped_count != 0 || zhdr->foreign_handles != 0) {
1301 clear_bit(PAGE_CLAIMED, &page->private);
1302 z3fold_page_unlock(zhdr);
1303 return -EBUSY;
1304 }
1305 if (work_pending(&zhdr->work)) {
1306 z3fold_page_unlock(zhdr);
1307 return -EAGAIN;
1308 }
1309 new_zhdr = page_address(newpage);
1310 memcpy(new_zhdr, zhdr, PAGE_SIZE);
1311 newpage->private = page->private;
1312 set_bit(PAGE_MIGRATED, &page->private);
1313 z3fold_page_unlock(zhdr);
1314 spin_lock_init(&new_zhdr->page_lock);
1315 INIT_WORK(&new_zhdr->work, compact_page_work);
1316 /*
1317 * z3fold_page_isolate() ensures that new_zhdr->buddy is empty,
1318 * so we only have to reinitialize it.
1319 */
1320 INIT_LIST_HEAD(&new_zhdr->buddy);
1321 __ClearPageMovable(page);
1322
1323 get_page(newpage);
1324 z3fold_page_lock(new_zhdr);
1325 if (new_zhdr->first_chunks)
1326 encode_handle(new_zhdr, FIRST);
1327 if (new_zhdr->last_chunks)
1328 encode_handle(new_zhdr, LAST);
1329 if (new_zhdr->middle_chunks)
1330 encode_handle(new_zhdr, MIDDLE);
1331 set_bit(NEEDS_COMPACTING, &newpage->private);
1332 new_zhdr->cpu = smp_processor_id();
1333 __SetPageMovable(newpage, &z3fold_mops);
1334 z3fold_page_unlock(new_zhdr);
1335
1336 queue_work_on(new_zhdr->cpu, pool->compact_wq, &new_zhdr->work);
1337
1338 /* PAGE_CLAIMED and PAGE_MIGRATED are cleared now. */
1339 page->private = 0;
1340 put_page(page);
1341 return 0;
1342}
1343
1344static void z3fold_page_putback(struct page *page)
1345{
1346 struct z3fold_header *zhdr;
1347 struct z3fold_pool *pool;
1348
1349 zhdr = page_address(page);
1350 pool = zhdr_to_pool(zhdr);
1351
1352 z3fold_page_lock(zhdr);
1353 if (!list_empty(&zhdr->buddy))
1354 list_del_init(&zhdr->buddy);
1355 INIT_LIST_HEAD(&page->lru);
1356 if (put_z3fold_locked(zhdr))
1357 return;
1358 if (list_empty(&zhdr->buddy))
1359 add_to_unbuddied(pool, zhdr);
1360 clear_bit(PAGE_CLAIMED, &page->private);
1361 z3fold_page_unlock(zhdr);
1362}
1363
1364static const struct movable_operations z3fold_mops = {
1365 .isolate_page = z3fold_page_isolate,
1366 .migrate_page = z3fold_page_migrate,
1367 .putback_page = z3fold_page_putback,
1368};
1369
1370/*****************
1371 * zpool
1372 ****************/
1373
1374static void *z3fold_zpool_create(const char *name, gfp_t gfp)
1375{
1376 return z3fold_create_pool(name, gfp);
1377}
1378
1379static void z3fold_zpool_destroy(void *pool)
1380{
1381 z3fold_destroy_pool(pool);
1382}
1383
1384static int z3fold_zpool_malloc(void *pool, size_t size, gfp_t gfp,
1385 unsigned long *handle)
1386{
1387 return z3fold_alloc(pool, size, gfp, handle);
1388}
1389static void z3fold_zpool_free(void *pool, unsigned long handle)
1390{
1391 z3fold_free(pool, handle);
1392}
1393
1394static void *z3fold_zpool_map(void *pool, unsigned long handle,
1395 enum zpool_mapmode mm)
1396{
1397 return z3fold_map(pool, handle);
1398}
1399static void z3fold_zpool_unmap(void *pool, unsigned long handle)
1400{
1401 z3fold_unmap(pool, handle);
1402}
1403
1404static u64 z3fold_zpool_total_size(void *pool)
1405{
1406 return z3fold_get_pool_size(pool) * PAGE_SIZE;
1407}
1408
1409static struct zpool_driver z3fold_zpool_driver = {
1410 .type = "z3fold",
1411 .sleep_mapped = true,
1412 .owner = THIS_MODULE,
1413 .create = z3fold_zpool_create,
1414 .destroy = z3fold_zpool_destroy,
1415 .malloc = z3fold_zpool_malloc,
1416 .free = z3fold_zpool_free,
1417 .map = z3fold_zpool_map,
1418 .unmap = z3fold_zpool_unmap,
1419 .total_size = z3fold_zpool_total_size,
1420};
1421
1422MODULE_ALIAS("zpool-z3fold");
1423
1424static int __init init_z3fold(void)
1425{
1426 /*
1427 * Make sure the z3fold header is not larger than the page size and
1428 * there has remaining spaces for its buddy.
1429 */
1430 BUILD_BUG_ON(ZHDR_SIZE_ALIGNED > PAGE_SIZE - CHUNK_SIZE);
1431 zpool_register_driver(&z3fold_zpool_driver);
1432
1433 return 0;
1434}
1435
1436static void __exit exit_z3fold(void)
1437{
1438 zpool_unregister_driver(&z3fold_zpool_driver);
1439}
1440
1441module_init(init_z3fold);
1442module_exit(exit_z3fold);
1443
1444MODULE_LICENSE("GPL");
1445MODULE_AUTHOR("Vitaly Wool <vitalywool@gmail.com>");
1446MODULE_DESCRIPTION("3-Fold Allocator for Compressed Pages");