61cb05dc950caf394dbf375b5b7084a193cbaddb
[muen/linux.git] / mm / zsmalloc.c
1 /*
2  * zsmalloc memory allocator
3  *
4  * Copyright (C) 2011  Nitin Gupta
5  * Copyright (C) 2012, 2013 Minchan Kim
6  *
7  * This code is released using a dual license strategy: BSD/GPL
8  * You can choose the license that better fits your requirements.
9  *
10  * Released under the terms of 3-clause BSD License
11  * Released under the terms of GNU General Public License Version 2.0
12  */
13
14 /*
15  * Following is how we use various fields and flags of underlying
16  * struct page(s) to form a zspage.
17  *
18  * Usage of struct page fields:
19  *      page->private: points to zspage
20  *      page->freelist(index): links together all component pages of a zspage
21  *              For the huge page, this is always 0, so we use this field
22  *              to store handle.
23  *      page->units: first object offset in a subpage of zspage
24  *
25  * Usage of struct page flags:
26  *      PG_private: identifies the first component page
27  *      PG_owner_priv_1: identifies the huge component page
28  *
29  */
30
31 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32
33 #include <linux/module.h>
34 #include <linux/kernel.h>
35 #include <linux/sched.h>
36 #include <linux/magic.h>
37 #include <linux/bitops.h>
38 #include <linux/errno.h>
39 #include <linux/highmem.h>
40 #include <linux/string.h>
41 #include <linux/slab.h>
42 #include <asm/tlbflush.h>
43 #include <asm/pgtable.h>
44 #include <linux/cpumask.h>
45 #include <linux/cpu.h>
46 #include <linux/vmalloc.h>
47 #include <linux/preempt.h>
48 #include <linux/spinlock.h>
49 #include <linux/shrinker.h>
50 #include <linux/types.h>
51 #include <linux/debugfs.h>
52 #include <linux/zsmalloc.h>
53 #include <linux/zpool.h>
54 #include <linux/mount.h>
55 #include <linux/migrate.h>
56 #include <linux/pagemap.h>
57 #include <linux/fs.h>
58
59 #define ZSPAGE_MAGIC    0x58
60
61 /*
62  * This must be power of 2 and greater than of equal to sizeof(link_free).
63  * These two conditions ensure that any 'struct link_free' itself doesn't
64  * span more than 1 page which avoids complex case of mapping 2 pages simply
65  * to restore link_free pointer values.
66  */
67 #define ZS_ALIGN                8
68
69 /*
70  * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
71  * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
72  */
73 #define ZS_MAX_ZSPAGE_ORDER 2
74 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
75
76 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
77
78 /*
79  * Object location (<PFN>, <obj_idx>) is encoded as
80  * as single (unsigned long) handle value.
81  *
82  * Note that object index <obj_idx> starts from 0.
83  *
84  * This is made more complicated by various memory models and PAE.
85  */
86
87 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
88 #ifdef MAX_PHYSMEM_BITS
89 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
90 #else
91 /*
92  * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
93  * be PAGE_SHIFT
94  */
95 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
96 #endif
97 #endif
98
99 #define _PFN_BITS               (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
100
101 /*
102  * Memory for allocating for handle keeps object position by
103  * encoding <page, obj_idx> and the encoded value has a room
104  * in least bit(ie, look at obj_to_location).
105  * We use the bit to synchronize between object access by
106  * user and migration.
107  */
108 #define HANDLE_PIN_BIT  0
109
110 /*
111  * Head in allocated object should have OBJ_ALLOCATED_TAG
112  * to identify the object was allocated or not.
113  * It's okay to add the status bit in the least bit because
114  * header keeps handle which is 4byte-aligned address so we
115  * have room for two bit at least.
116  */
117 #define OBJ_ALLOCATED_TAG 1
118 #define OBJ_TAG_BITS 1
119 #define OBJ_INDEX_BITS  (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
120 #define OBJ_INDEX_MASK  ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
121
122 #define FULLNESS_BITS   2
123 #define CLASS_BITS      8
124 #define ISOLATED_BITS   3
125 #define MAGIC_VAL_BITS  8
126
127 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
128 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
129 #define ZS_MIN_ALLOC_SIZE \
130         MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
131 /* each chunk includes extra space to keep handle */
132 #define ZS_MAX_ALLOC_SIZE       PAGE_SIZE
133
134 /*
135  * On systems with 4K page size, this gives 255 size classes! There is a
136  * trader-off here:
137  *  - Large number of size classes is potentially wasteful as free page are
138  *    spread across these classes
139  *  - Small number of size classes causes large internal fragmentation
140  *  - Probably its better to use specific size classes (empirically
141  *    determined). NOTE: all those class sizes must be set as multiple of
142  *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
143  *
144  *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
145  *  (reason above)
146  */
147 #define ZS_SIZE_CLASS_DELTA     (PAGE_SIZE >> CLASS_BITS)
148 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
149                                       ZS_SIZE_CLASS_DELTA) + 1)
150
151 enum fullness_group {
152         ZS_EMPTY,
153         ZS_ALMOST_EMPTY,
154         ZS_ALMOST_FULL,
155         ZS_FULL,
156         NR_ZS_FULLNESS,
157 };
158
159 enum zs_stat_type {
160         CLASS_EMPTY,
161         CLASS_ALMOST_EMPTY,
162         CLASS_ALMOST_FULL,
163         CLASS_FULL,
164         OBJ_ALLOCATED,
165         OBJ_USED,
166         NR_ZS_STAT_TYPE,
167 };
168
169 struct zs_size_stat {
170         unsigned long objs[NR_ZS_STAT_TYPE];
171 };
172
173 #ifdef CONFIG_ZSMALLOC_STAT
174 static struct dentry *zs_stat_root;
175 #endif
176
177 #ifdef CONFIG_COMPACTION
178 static struct vfsmount *zsmalloc_mnt;
179 #endif
180
181 /*
182  * We assign a page to ZS_ALMOST_EMPTY fullness group when:
183  *      n <= N / f, where
184  * n = number of allocated objects
185  * N = total number of objects zspage can store
186  * f = fullness_threshold_frac
187  *
188  * Similarly, we assign zspage to:
189  *      ZS_ALMOST_FULL  when n > N / f
190  *      ZS_EMPTY        when n == 0
191  *      ZS_FULL         when n == N
192  *
193  * (see: fix_fullness_group())
194  */
195 static const int fullness_threshold_frac = 4;
196 static size_t huge_class_size;
197
198 struct size_class {
199         spinlock_t lock;
200         struct list_head fullness_list[NR_ZS_FULLNESS];
201         /*
202          * Size of objects stored in this class. Must be multiple
203          * of ZS_ALIGN.
204          */
205         int size;
206         int objs_per_zspage;
207         /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
208         int pages_per_zspage;
209
210         unsigned int index;
211         struct zs_size_stat stats;
212 };
213
214 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
215 static void SetPageHugeObject(struct page *page)
216 {
217         SetPageOwnerPriv1(page);
218 }
219
220 static void ClearPageHugeObject(struct page *page)
221 {
222         ClearPageOwnerPriv1(page);
223 }
224
225 static int PageHugeObject(struct page *page)
226 {
227         return PageOwnerPriv1(page);
228 }
229
230 /*
231  * Placed within free objects to form a singly linked list.
232  * For every zspage, zspage->freeobj gives head of this list.
233  *
234  * This must be power of 2 and less than or equal to ZS_ALIGN
235  */
236 struct link_free {
237         union {
238                 /*
239                  * Free object index;
240                  * It's valid for non-allocated object
241                  */
242                 unsigned long next;
243                 /*
244                  * Handle of allocated object.
245                  */
246                 unsigned long handle;
247         };
248 };
249
250 struct zs_pool {
251         const char *name;
252
253         struct size_class *size_class[ZS_SIZE_CLASSES];
254         struct kmem_cache *handle_cachep;
255         struct kmem_cache *zspage_cachep;
256
257         atomic_long_t pages_allocated;
258
259         struct zs_pool_stats stats;
260
261         /* Compact classes */
262         struct shrinker shrinker;
263
264 #ifdef CONFIG_ZSMALLOC_STAT
265         struct dentry *stat_dentry;
266 #endif
267 #ifdef CONFIG_COMPACTION
268         struct inode *inode;
269         struct work_struct free_work;
270 #endif
271 };
272
273 struct zspage {
274         struct {
275                 unsigned int fullness:FULLNESS_BITS;
276                 unsigned int class:CLASS_BITS + 1;
277                 unsigned int isolated:ISOLATED_BITS;
278                 unsigned int magic:MAGIC_VAL_BITS;
279         };
280         unsigned int inuse;
281         unsigned int freeobj;
282         struct page *first_page;
283         struct list_head list; /* fullness list */
284 #ifdef CONFIG_COMPACTION
285         rwlock_t lock;
286 #endif
287 };
288
289 struct mapping_area {
290 #ifdef CONFIG_PGTABLE_MAPPING
291         struct vm_struct *vm; /* vm area for mapping object that span pages */
292 #else
293         char *vm_buf; /* copy buffer for objects that span pages */
294 #endif
295         char *vm_addr; /* address of kmap_atomic()'ed pages */
296         enum zs_mapmode vm_mm; /* mapping mode */
297 };
298
299 #ifdef CONFIG_COMPACTION
300 static int zs_register_migration(struct zs_pool *pool);
301 static void zs_unregister_migration(struct zs_pool *pool);
302 static void migrate_lock_init(struct zspage *zspage);
303 static void migrate_read_lock(struct zspage *zspage);
304 static void migrate_read_unlock(struct zspage *zspage);
305 static void kick_deferred_free(struct zs_pool *pool);
306 static void init_deferred_free(struct zs_pool *pool);
307 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
308 #else
309 static int zsmalloc_mount(void) { return 0; }
310 static void zsmalloc_unmount(void) {}
311 static int zs_register_migration(struct zs_pool *pool) { return 0; }
312 static void zs_unregister_migration(struct zs_pool *pool) {}
313 static void migrate_lock_init(struct zspage *zspage) {}
314 static void migrate_read_lock(struct zspage *zspage) {}
315 static void migrate_read_unlock(struct zspage *zspage) {}
316 static void kick_deferred_free(struct zs_pool *pool) {}
317 static void init_deferred_free(struct zs_pool *pool) {}
318 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
319 #endif
320
321 static int create_cache(struct zs_pool *pool)
322 {
323         pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
324                                         0, 0, NULL);
325         if (!pool->handle_cachep)
326                 return 1;
327
328         pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
329                                         0, 0, NULL);
330         if (!pool->zspage_cachep) {
331                 kmem_cache_destroy(pool->handle_cachep);
332                 pool->handle_cachep = NULL;
333                 return 1;
334         }
335
336         return 0;
337 }
338
339 static void destroy_cache(struct zs_pool *pool)
340 {
341         kmem_cache_destroy(pool->handle_cachep);
342         kmem_cache_destroy(pool->zspage_cachep);
343 }
344
345 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
346 {
347         return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
348                         gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
349 }
350
351 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
352 {
353         kmem_cache_free(pool->handle_cachep, (void *)handle);
354 }
355
356 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
357 {
358         return kmem_cache_alloc(pool->zspage_cachep,
359                         flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
360 }
361
362 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
363 {
364         kmem_cache_free(pool->zspage_cachep, zspage);
365 }
366
367 static void record_obj(unsigned long handle, unsigned long obj)
368 {
369         /*
370          * lsb of @obj represents handle lock while other bits
371          * represent object value the handle is pointing so
372          * updating shouldn't do store tearing.
373          */
374         WRITE_ONCE(*(unsigned long *)handle, obj);
375 }
376
377 /* zpool driver */
378
379 #ifdef CONFIG_ZPOOL
380
381 static void *zs_zpool_create(const char *name, gfp_t gfp,
382                              const struct zpool_ops *zpool_ops,
383                              struct zpool *zpool)
384 {
385         /*
386          * Ignore global gfp flags: zs_malloc() may be invoked from
387          * different contexts and its caller must provide a valid
388          * gfp mask.
389          */
390         return zs_create_pool(name);
391 }
392
393 static void zs_zpool_destroy(void *pool)
394 {
395         zs_destroy_pool(pool);
396 }
397
398 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
399                         unsigned long *handle)
400 {
401         *handle = zs_malloc(pool, size, gfp);
402         return *handle ? 0 : -1;
403 }
404 static void zs_zpool_free(void *pool, unsigned long handle)
405 {
406         zs_free(pool, handle);
407 }
408
409 static void *zs_zpool_map(void *pool, unsigned long handle,
410                         enum zpool_mapmode mm)
411 {
412         enum zs_mapmode zs_mm;
413
414         switch (mm) {
415         case ZPOOL_MM_RO:
416                 zs_mm = ZS_MM_RO;
417                 break;
418         case ZPOOL_MM_WO:
419                 zs_mm = ZS_MM_WO;
420                 break;
421         case ZPOOL_MM_RW: /* fallthru */
422         default:
423                 zs_mm = ZS_MM_RW;
424                 break;
425         }
426
427         return zs_map_object(pool, handle, zs_mm);
428 }
429 static void zs_zpool_unmap(void *pool, unsigned long handle)
430 {
431         zs_unmap_object(pool, handle);
432 }
433
434 static u64 zs_zpool_total_size(void *pool)
435 {
436         return zs_get_total_pages(pool) << PAGE_SHIFT;
437 }
438
439 static struct zpool_driver zs_zpool_driver = {
440         .type =         "zsmalloc",
441         .owner =        THIS_MODULE,
442         .create =       zs_zpool_create,
443         .destroy =      zs_zpool_destroy,
444         .malloc =       zs_zpool_malloc,
445         .free =         zs_zpool_free,
446         .map =          zs_zpool_map,
447         .unmap =        zs_zpool_unmap,
448         .total_size =   zs_zpool_total_size,
449 };
450
451 MODULE_ALIAS("zpool-zsmalloc");
452 #endif /* CONFIG_ZPOOL */
453
454 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
455 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
456
457 static bool is_zspage_isolated(struct zspage *zspage)
458 {
459         return zspage->isolated;
460 }
461
462 static __maybe_unused int is_first_page(struct page *page)
463 {
464         return PagePrivate(page);
465 }
466
467 /* Protected by class->lock */
468 static inline int get_zspage_inuse(struct zspage *zspage)
469 {
470         return zspage->inuse;
471 }
472
473 static inline void set_zspage_inuse(struct zspage *zspage, int val)
474 {
475         zspage->inuse = val;
476 }
477
478 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
479 {
480         zspage->inuse += val;
481 }
482
483 static inline struct page *get_first_page(struct zspage *zspage)
484 {
485         struct page *first_page = zspage->first_page;
486
487         VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
488         return first_page;
489 }
490
491 static inline int get_first_obj_offset(struct page *page)
492 {
493         return page->units;
494 }
495
496 static inline void set_first_obj_offset(struct page *page, int offset)
497 {
498         page->units = offset;
499 }
500
501 static inline unsigned int get_freeobj(struct zspage *zspage)
502 {
503         return zspage->freeobj;
504 }
505
506 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
507 {
508         zspage->freeobj = obj;
509 }
510
511 static void get_zspage_mapping(struct zspage *zspage,
512                                 unsigned int *class_idx,
513                                 enum fullness_group *fullness)
514 {
515         BUG_ON(zspage->magic != ZSPAGE_MAGIC);
516
517         *fullness = zspage->fullness;
518         *class_idx = zspage->class;
519 }
520
521 static void set_zspage_mapping(struct zspage *zspage,
522                                 unsigned int class_idx,
523                                 enum fullness_group fullness)
524 {
525         zspage->class = class_idx;
526         zspage->fullness = fullness;
527 }
528
529 /*
530  * zsmalloc divides the pool into various size classes where each
531  * class maintains a list of zspages where each zspage is divided
532  * into equal sized chunks. Each allocation falls into one of these
533  * classes depending on its size. This function returns index of the
534  * size class which has chunk size big enough to hold the give size.
535  */
536 static int get_size_class_index(int size)
537 {
538         int idx = 0;
539
540         if (likely(size > ZS_MIN_ALLOC_SIZE))
541                 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
542                                 ZS_SIZE_CLASS_DELTA);
543
544         return min_t(int, ZS_SIZE_CLASSES - 1, idx);
545 }
546
547 /* type can be of enum type zs_stat_type or fullness_group */
548 static inline void zs_stat_inc(struct size_class *class,
549                                 int type, unsigned long cnt)
550 {
551         class->stats.objs[type] += cnt;
552 }
553
554 /* type can be of enum type zs_stat_type or fullness_group */
555 static inline void zs_stat_dec(struct size_class *class,
556                                 int type, unsigned long cnt)
557 {
558         class->stats.objs[type] -= cnt;
559 }
560
561 /* type can be of enum type zs_stat_type or fullness_group */
562 static inline unsigned long zs_stat_get(struct size_class *class,
563                                 int type)
564 {
565         return class->stats.objs[type];
566 }
567
568 #ifdef CONFIG_ZSMALLOC_STAT
569
570 static void __init zs_stat_init(void)
571 {
572         if (!debugfs_initialized()) {
573                 pr_warn("debugfs not available, stat dir not created\n");
574                 return;
575         }
576
577         zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
578         if (!zs_stat_root)
579                 pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
580 }
581
582 static void __exit zs_stat_exit(void)
583 {
584         debugfs_remove_recursive(zs_stat_root);
585 }
586
587 static unsigned long zs_can_compact(struct size_class *class);
588
589 static int zs_stats_size_show(struct seq_file *s, void *v)
590 {
591         int i;
592         struct zs_pool *pool = s->private;
593         struct size_class *class;
594         int objs_per_zspage;
595         unsigned long class_almost_full, class_almost_empty;
596         unsigned long obj_allocated, obj_used, pages_used, freeable;
597         unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
598         unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
599         unsigned long total_freeable = 0;
600
601         seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
602                         "class", "size", "almost_full", "almost_empty",
603                         "obj_allocated", "obj_used", "pages_used",
604                         "pages_per_zspage", "freeable");
605
606         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
607                 class = pool->size_class[i];
608
609                 if (class->index != i)
610                         continue;
611
612                 spin_lock(&class->lock);
613                 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
614                 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
615                 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
616                 obj_used = zs_stat_get(class, OBJ_USED);
617                 freeable = zs_can_compact(class);
618                 spin_unlock(&class->lock);
619
620                 objs_per_zspage = class->objs_per_zspage;
621                 pages_used = obj_allocated / objs_per_zspage *
622                                 class->pages_per_zspage;
623
624                 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
625                                 " %10lu %10lu %16d %8lu\n",
626                         i, class->size, class_almost_full, class_almost_empty,
627                         obj_allocated, obj_used, pages_used,
628                         class->pages_per_zspage, freeable);
629
630                 total_class_almost_full += class_almost_full;
631                 total_class_almost_empty += class_almost_empty;
632                 total_objs += obj_allocated;
633                 total_used_objs += obj_used;
634                 total_pages += pages_used;
635                 total_freeable += freeable;
636         }
637
638         seq_puts(s, "\n");
639         seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
640                         "Total", "", total_class_almost_full,
641                         total_class_almost_empty, total_objs,
642                         total_used_objs, total_pages, "", total_freeable);
643
644         return 0;
645 }
646 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
647
648 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
649 {
650         struct dentry *entry;
651
652         if (!zs_stat_root) {
653                 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
654                 return;
655         }
656
657         entry = debugfs_create_dir(name, zs_stat_root);
658         if (!entry) {
659                 pr_warn("debugfs dir <%s> creation failed\n", name);
660                 return;
661         }
662         pool->stat_dentry = entry;
663
664         entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
665                         pool->stat_dentry, pool, &zs_stats_size_fops);
666         if (!entry) {
667                 pr_warn("%s: debugfs file entry <%s> creation failed\n",
668                                 name, "classes");
669                 debugfs_remove_recursive(pool->stat_dentry);
670                 pool->stat_dentry = NULL;
671         }
672 }
673
674 static void zs_pool_stat_destroy(struct zs_pool *pool)
675 {
676         debugfs_remove_recursive(pool->stat_dentry);
677 }
678
679 #else /* CONFIG_ZSMALLOC_STAT */
680 static void __init zs_stat_init(void)
681 {
682 }
683
684 static void __exit zs_stat_exit(void)
685 {
686 }
687
688 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
689 {
690 }
691
692 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
693 {
694 }
695 #endif
696
697
698 /*
699  * For each size class, zspages are divided into different groups
700  * depending on how "full" they are. This was done so that we could
701  * easily find empty or nearly empty zspages when we try to shrink
702  * the pool (not yet implemented). This function returns fullness
703  * status of the given page.
704  */
705 static enum fullness_group get_fullness_group(struct size_class *class,
706                                                 struct zspage *zspage)
707 {
708         int inuse, objs_per_zspage;
709         enum fullness_group fg;
710
711         inuse = get_zspage_inuse(zspage);
712         objs_per_zspage = class->objs_per_zspage;
713
714         if (inuse == 0)
715                 fg = ZS_EMPTY;
716         else if (inuse == objs_per_zspage)
717                 fg = ZS_FULL;
718         else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
719                 fg = ZS_ALMOST_EMPTY;
720         else
721                 fg = ZS_ALMOST_FULL;
722
723         return fg;
724 }
725
726 /*
727  * Each size class maintains various freelists and zspages are assigned
728  * to one of these freelists based on the number of live objects they
729  * have. This functions inserts the given zspage into the freelist
730  * identified by <class, fullness_group>.
731  */
732 static void insert_zspage(struct size_class *class,
733                                 struct zspage *zspage,
734                                 enum fullness_group fullness)
735 {
736         struct zspage *head;
737
738         zs_stat_inc(class, fullness, 1);
739         head = list_first_entry_or_null(&class->fullness_list[fullness],
740                                         struct zspage, list);
741         /*
742          * We want to see more ZS_FULL pages and less almost empty/full.
743          * Put pages with higher ->inuse first.
744          */
745         if (head) {
746                 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
747                         list_add(&zspage->list, &head->list);
748                         return;
749                 }
750         }
751         list_add(&zspage->list, &class->fullness_list[fullness]);
752 }
753
754 /*
755  * This function removes the given zspage from the freelist identified
756  * by <class, fullness_group>.
757  */
758 static void remove_zspage(struct size_class *class,
759                                 struct zspage *zspage,
760                                 enum fullness_group fullness)
761 {
762         VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
763         VM_BUG_ON(is_zspage_isolated(zspage));
764
765         list_del_init(&zspage->list);
766         zs_stat_dec(class, fullness, 1);
767 }
768
769 /*
770  * Each size class maintains zspages in different fullness groups depending
771  * on the number of live objects they contain. When allocating or freeing
772  * objects, the fullness status of the page can change, say, from ALMOST_FULL
773  * to ALMOST_EMPTY when freeing an object. This function checks if such
774  * a status change has occurred for the given page and accordingly moves the
775  * page from the freelist of the old fullness group to that of the new
776  * fullness group.
777  */
778 static enum fullness_group fix_fullness_group(struct size_class *class,
779                                                 struct zspage *zspage)
780 {
781         int class_idx;
782         enum fullness_group currfg, newfg;
783
784         get_zspage_mapping(zspage, &class_idx, &currfg);
785         newfg = get_fullness_group(class, zspage);
786         if (newfg == currfg)
787                 goto out;
788
789         if (!is_zspage_isolated(zspage)) {
790                 remove_zspage(class, zspage, currfg);
791                 insert_zspage(class, zspage, newfg);
792         }
793
794         set_zspage_mapping(zspage, class_idx, newfg);
795
796 out:
797         return newfg;
798 }
799
800 /*
801  * We have to decide on how many pages to link together
802  * to form a zspage for each size class. This is important
803  * to reduce wastage due to unusable space left at end of
804  * each zspage which is given as:
805  *     wastage = Zp % class_size
806  *     usage = Zp - wastage
807  * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
808  *
809  * For example, for size class of 3/8 * PAGE_SIZE, we should
810  * link together 3 PAGE_SIZE sized pages to form a zspage
811  * since then we can perfectly fit in 8 such objects.
812  */
813 static int get_pages_per_zspage(int class_size)
814 {
815         int i, max_usedpc = 0;
816         /* zspage order which gives maximum used size per KB */
817         int max_usedpc_order = 1;
818
819         for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
820                 int zspage_size;
821                 int waste, usedpc;
822
823                 zspage_size = i * PAGE_SIZE;
824                 waste = zspage_size % class_size;
825                 usedpc = (zspage_size - waste) * 100 / zspage_size;
826
827                 if (usedpc > max_usedpc) {
828                         max_usedpc = usedpc;
829                         max_usedpc_order = i;
830                 }
831         }
832
833         return max_usedpc_order;
834 }
835
836 static struct zspage *get_zspage(struct page *page)
837 {
838         struct zspage *zspage = (struct zspage *)page->private;
839
840         BUG_ON(zspage->magic != ZSPAGE_MAGIC);
841         return zspage;
842 }
843
844 static struct page *get_next_page(struct page *page)
845 {
846         if (unlikely(PageHugeObject(page)))
847                 return NULL;
848
849         return page->freelist;
850 }
851
852 /**
853  * obj_to_location - get (<page>, <obj_idx>) from encoded object value
854  * @obj: the encoded object value
855  * @page: page object resides in zspage
856  * @obj_idx: object index
857  */
858 static void obj_to_location(unsigned long obj, struct page **page,
859                                 unsigned int *obj_idx)
860 {
861         obj >>= OBJ_TAG_BITS;
862         *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
863         *obj_idx = (obj & OBJ_INDEX_MASK);
864 }
865
866 /**
867  * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
868  * @page: page object resides in zspage
869  * @obj_idx: object index
870  */
871 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
872 {
873         unsigned long obj;
874
875         obj = page_to_pfn(page) << OBJ_INDEX_BITS;
876         obj |= obj_idx & OBJ_INDEX_MASK;
877         obj <<= OBJ_TAG_BITS;
878
879         return obj;
880 }
881
882 static unsigned long handle_to_obj(unsigned long handle)
883 {
884         return *(unsigned long *)handle;
885 }
886
887 static unsigned long obj_to_head(struct page *page, void *obj)
888 {
889         if (unlikely(PageHugeObject(page))) {
890                 VM_BUG_ON_PAGE(!is_first_page(page), page);
891                 return page->index;
892         } else
893                 return *(unsigned long *)obj;
894 }
895
896 static inline int testpin_tag(unsigned long handle)
897 {
898         return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
899 }
900
901 static inline int trypin_tag(unsigned long handle)
902 {
903         return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
904 }
905
906 static void pin_tag(unsigned long handle)
907 {
908         bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
909 }
910
911 static void unpin_tag(unsigned long handle)
912 {
913         bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
914 }
915
916 static void reset_page(struct page *page)
917 {
918         __ClearPageMovable(page);
919         ClearPagePrivate(page);
920         set_page_private(page, 0);
921         page_mapcount_reset(page);
922         ClearPageHugeObject(page);
923         page->freelist = NULL;
924 }
925
926 /*
927  * To prevent zspage destroy during migration, zspage freeing should
928  * hold locks of all pages in the zspage.
929  */
930 void lock_zspage(struct zspage *zspage)
931 {
932         struct page *page = get_first_page(zspage);
933
934         do {
935                 lock_page(page);
936         } while ((page = get_next_page(page)) != NULL);
937 }
938
939 int trylock_zspage(struct zspage *zspage)
940 {
941         struct page *cursor, *fail;
942
943         for (cursor = get_first_page(zspage); cursor != NULL; cursor =
944                                         get_next_page(cursor)) {
945                 if (!trylock_page(cursor)) {
946                         fail = cursor;
947                         goto unlock;
948                 }
949         }
950
951         return 1;
952 unlock:
953         for (cursor = get_first_page(zspage); cursor != fail; cursor =
954                                         get_next_page(cursor))
955                 unlock_page(cursor);
956
957         return 0;
958 }
959
960 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
961                                 struct zspage *zspage)
962 {
963         struct page *page, *next;
964         enum fullness_group fg;
965         unsigned int class_idx;
966
967         get_zspage_mapping(zspage, &class_idx, &fg);
968
969         assert_spin_locked(&class->lock);
970
971         VM_BUG_ON(get_zspage_inuse(zspage));
972         VM_BUG_ON(fg != ZS_EMPTY);
973
974         next = page = get_first_page(zspage);
975         do {
976                 VM_BUG_ON_PAGE(!PageLocked(page), page);
977                 next = get_next_page(page);
978                 reset_page(page);
979                 unlock_page(page);
980                 dec_zone_page_state(page, NR_ZSPAGES);
981                 put_page(page);
982                 page = next;
983         } while (page != NULL);
984
985         cache_free_zspage(pool, zspage);
986
987         zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
988         atomic_long_sub(class->pages_per_zspage,
989                                         &pool->pages_allocated);
990 }
991
992 static void free_zspage(struct zs_pool *pool, struct size_class *class,
993                                 struct zspage *zspage)
994 {
995         VM_BUG_ON(get_zspage_inuse(zspage));
996         VM_BUG_ON(list_empty(&zspage->list));
997
998         if (!trylock_zspage(zspage)) {
999                 kick_deferred_free(pool);
1000                 return;
1001         }
1002
1003         remove_zspage(class, zspage, ZS_EMPTY);
1004         __free_zspage(pool, class, zspage);
1005 }
1006
1007 /* Initialize a newly allocated zspage */
1008 static void init_zspage(struct size_class *class, struct zspage *zspage)
1009 {
1010         unsigned int freeobj = 1;
1011         unsigned long off = 0;
1012         struct page *page = get_first_page(zspage);
1013
1014         while (page) {
1015                 struct page *next_page;
1016                 struct link_free *link;
1017                 void *vaddr;
1018
1019                 set_first_obj_offset(page, off);
1020
1021                 vaddr = kmap_atomic(page);
1022                 link = (struct link_free *)vaddr + off / sizeof(*link);
1023
1024                 while ((off += class->size) < PAGE_SIZE) {
1025                         link->next = freeobj++ << OBJ_TAG_BITS;
1026                         link += class->size / sizeof(*link);
1027                 }
1028
1029                 /*
1030                  * We now come to the last (full or partial) object on this
1031                  * page, which must point to the first object on the next
1032                  * page (if present)
1033                  */
1034                 next_page = get_next_page(page);
1035                 if (next_page) {
1036                         link->next = freeobj++ << OBJ_TAG_BITS;
1037                 } else {
1038                         /*
1039                          * Reset OBJ_TAG_BITS bit to last link to tell
1040                          * whether it's allocated object or not.
1041                          */
1042                         link->next = -1UL << OBJ_TAG_BITS;
1043                 }
1044                 kunmap_atomic(vaddr);
1045                 page = next_page;
1046                 off %= PAGE_SIZE;
1047         }
1048
1049         set_freeobj(zspage, 0);
1050 }
1051
1052 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1053                                 struct page *pages[])
1054 {
1055         int i;
1056         struct page *page;
1057         struct page *prev_page = NULL;
1058         int nr_pages = class->pages_per_zspage;
1059
1060         /*
1061          * Allocate individual pages and link them together as:
1062          * 1. all pages are linked together using page->freelist
1063          * 2. each sub-page point to zspage using page->private
1064          *
1065          * we set PG_private to identify the first page (i.e. no other sub-page
1066          * has this flag set).
1067          */
1068         for (i = 0; i < nr_pages; i++) {
1069                 page = pages[i];
1070                 set_page_private(page, (unsigned long)zspage);
1071                 page->freelist = NULL;
1072                 if (i == 0) {
1073                         zspage->first_page = page;
1074                         SetPagePrivate(page);
1075                         if (unlikely(class->objs_per_zspage == 1 &&
1076                                         class->pages_per_zspage == 1))
1077                                 SetPageHugeObject(page);
1078                 } else {
1079                         prev_page->freelist = page;
1080                 }
1081                 prev_page = page;
1082         }
1083 }
1084
1085 /*
1086  * Allocate a zspage for the given size class
1087  */
1088 static struct zspage *alloc_zspage(struct zs_pool *pool,
1089                                         struct size_class *class,
1090                                         gfp_t gfp)
1091 {
1092         int i;
1093         struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1094         struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1095
1096         if (!zspage)
1097                 return NULL;
1098
1099         memset(zspage, 0, sizeof(struct zspage));
1100         zspage->magic = ZSPAGE_MAGIC;
1101         migrate_lock_init(zspage);
1102
1103         for (i = 0; i < class->pages_per_zspage; i++) {
1104                 struct page *page;
1105
1106                 page = alloc_page(gfp);
1107                 if (!page) {
1108                         while (--i >= 0) {
1109                                 dec_zone_page_state(pages[i], NR_ZSPAGES);
1110                                 __free_page(pages[i]);
1111                         }
1112                         cache_free_zspage(pool, zspage);
1113                         return NULL;
1114                 }
1115
1116                 inc_zone_page_state(page, NR_ZSPAGES);
1117                 pages[i] = page;
1118         }
1119
1120         create_page_chain(class, zspage, pages);
1121         init_zspage(class, zspage);
1122
1123         return zspage;
1124 }
1125
1126 static struct zspage *find_get_zspage(struct size_class *class)
1127 {
1128         int i;
1129         struct zspage *zspage;
1130
1131         for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1132                 zspage = list_first_entry_or_null(&class->fullness_list[i],
1133                                 struct zspage, list);
1134                 if (zspage)
1135                         break;
1136         }
1137
1138         return zspage;
1139 }
1140
1141 #ifdef CONFIG_PGTABLE_MAPPING
1142 static inline int __zs_cpu_up(struct mapping_area *area)
1143 {
1144         /*
1145          * Make sure we don't leak memory if a cpu UP notification
1146          * and zs_init() race and both call zs_cpu_up() on the same cpu
1147          */
1148         if (area->vm)
1149                 return 0;
1150         area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1151         if (!area->vm)
1152                 return -ENOMEM;
1153         return 0;
1154 }
1155
1156 static inline void __zs_cpu_down(struct mapping_area *area)
1157 {
1158         if (area->vm)
1159                 free_vm_area(area->vm);
1160         area->vm = NULL;
1161 }
1162
1163 static inline void *__zs_map_object(struct mapping_area *area,
1164                                 struct page *pages[2], int off, int size)
1165 {
1166         BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1167         area->vm_addr = area->vm->addr;
1168         return area->vm_addr + off;
1169 }
1170
1171 static inline void __zs_unmap_object(struct mapping_area *area,
1172                                 struct page *pages[2], int off, int size)
1173 {
1174         unsigned long addr = (unsigned long)area->vm_addr;
1175
1176         unmap_kernel_range(addr, PAGE_SIZE * 2);
1177 }
1178
1179 #else /* CONFIG_PGTABLE_MAPPING */
1180
1181 static inline int __zs_cpu_up(struct mapping_area *area)
1182 {
1183         /*
1184          * Make sure we don't leak memory if a cpu UP notification
1185          * and zs_init() race and both call zs_cpu_up() on the same cpu
1186          */
1187         if (area->vm_buf)
1188                 return 0;
1189         area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1190         if (!area->vm_buf)
1191                 return -ENOMEM;
1192         return 0;
1193 }
1194
1195 static inline void __zs_cpu_down(struct mapping_area *area)
1196 {
1197         kfree(area->vm_buf);
1198         area->vm_buf = NULL;
1199 }
1200
1201 static void *__zs_map_object(struct mapping_area *area,
1202                         struct page *pages[2], int off, int size)
1203 {
1204         int sizes[2];
1205         void *addr;
1206         char *buf = area->vm_buf;
1207
1208         /* disable page faults to match kmap_atomic() return conditions */
1209         pagefault_disable();
1210
1211         /* no read fastpath */
1212         if (area->vm_mm == ZS_MM_WO)
1213                 goto out;
1214
1215         sizes[0] = PAGE_SIZE - off;
1216         sizes[1] = size - sizes[0];
1217
1218         /* copy object to per-cpu buffer */
1219         addr = kmap_atomic(pages[0]);
1220         memcpy(buf, addr + off, sizes[0]);
1221         kunmap_atomic(addr);
1222         addr = kmap_atomic(pages[1]);
1223         memcpy(buf + sizes[0], addr, sizes[1]);
1224         kunmap_atomic(addr);
1225 out:
1226         return area->vm_buf;
1227 }
1228
1229 static void __zs_unmap_object(struct mapping_area *area,
1230                         struct page *pages[2], int off, int size)
1231 {
1232         int sizes[2];
1233         void *addr;
1234         char *buf;
1235
1236         /* no write fastpath */
1237         if (area->vm_mm == ZS_MM_RO)
1238                 goto out;
1239
1240         buf = area->vm_buf;
1241         buf = buf + ZS_HANDLE_SIZE;
1242         size -= ZS_HANDLE_SIZE;
1243         off += ZS_HANDLE_SIZE;
1244
1245         sizes[0] = PAGE_SIZE - off;
1246         sizes[1] = size - sizes[0];
1247
1248         /* copy per-cpu buffer to object */
1249         addr = kmap_atomic(pages[0]);
1250         memcpy(addr + off, buf, sizes[0]);
1251         kunmap_atomic(addr);
1252         addr = kmap_atomic(pages[1]);
1253         memcpy(addr, buf + sizes[0], sizes[1]);
1254         kunmap_atomic(addr);
1255
1256 out:
1257         /* enable page faults to match kunmap_atomic() return conditions */
1258         pagefault_enable();
1259 }
1260
1261 #endif /* CONFIG_PGTABLE_MAPPING */
1262
1263 static int zs_cpu_prepare(unsigned int cpu)
1264 {
1265         struct mapping_area *area;
1266
1267         area = &per_cpu(zs_map_area, cpu);
1268         return __zs_cpu_up(area);
1269 }
1270
1271 static int zs_cpu_dead(unsigned int cpu)
1272 {
1273         struct mapping_area *area;
1274
1275         area = &per_cpu(zs_map_area, cpu);
1276         __zs_cpu_down(area);
1277         return 0;
1278 }
1279
1280 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1281                                         int objs_per_zspage)
1282 {
1283         if (prev->pages_per_zspage == pages_per_zspage &&
1284                 prev->objs_per_zspage == objs_per_zspage)
1285                 return true;
1286
1287         return false;
1288 }
1289
1290 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1291 {
1292         return get_zspage_inuse(zspage) == class->objs_per_zspage;
1293 }
1294
1295 unsigned long zs_get_total_pages(struct zs_pool *pool)
1296 {
1297         return atomic_long_read(&pool->pages_allocated);
1298 }
1299 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1300
1301 /**
1302  * zs_map_object - get address of allocated object from handle.
1303  * @pool: pool from which the object was allocated
1304  * @handle: handle returned from zs_malloc
1305  * @mm: maping mode to use
1306  *
1307  * Before using an object allocated from zs_malloc, it must be mapped using
1308  * this function. When done with the object, it must be unmapped using
1309  * zs_unmap_object.
1310  *
1311  * Only one object can be mapped per cpu at a time. There is no protection
1312  * against nested mappings.
1313  *
1314  * This function returns with preemption and page faults disabled.
1315  */
1316 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1317                         enum zs_mapmode mm)
1318 {
1319         struct zspage *zspage;
1320         struct page *page;
1321         unsigned long obj, off;
1322         unsigned int obj_idx;
1323
1324         unsigned int class_idx;
1325         enum fullness_group fg;
1326         struct size_class *class;
1327         struct mapping_area *area;
1328         struct page *pages[2];
1329         void *ret;
1330
1331         /*
1332          * Because we use per-cpu mapping areas shared among the
1333          * pools/users, we can't allow mapping in interrupt context
1334          * because it can corrupt another users mappings.
1335          */
1336         BUG_ON(in_interrupt());
1337
1338         /* From now on, migration cannot move the object */
1339         pin_tag(handle);
1340
1341         obj = handle_to_obj(handle);
1342         obj_to_location(obj, &page, &obj_idx);
1343         zspage = get_zspage(page);
1344
1345         /* migration cannot move any subpage in this zspage */
1346         migrate_read_lock(zspage);
1347
1348         get_zspage_mapping(zspage, &class_idx, &fg);
1349         class = pool->size_class[class_idx];
1350         off = (class->size * obj_idx) & ~PAGE_MASK;
1351
1352         area = &get_cpu_var(zs_map_area);
1353         area->vm_mm = mm;
1354         if (off + class->size <= PAGE_SIZE) {
1355                 /* this object is contained entirely within a page */
1356                 area->vm_addr = kmap_atomic(page);
1357                 ret = area->vm_addr + off;
1358                 goto out;
1359         }
1360
1361         /* this object spans two pages */
1362         pages[0] = page;
1363         pages[1] = get_next_page(page);
1364         BUG_ON(!pages[1]);
1365
1366         ret = __zs_map_object(area, pages, off, class->size);
1367 out:
1368         if (likely(!PageHugeObject(page)))
1369                 ret += ZS_HANDLE_SIZE;
1370
1371         return ret;
1372 }
1373 EXPORT_SYMBOL_GPL(zs_map_object);
1374
1375 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1376 {
1377         struct zspage *zspage;
1378         struct page *page;
1379         unsigned long obj, off;
1380         unsigned int obj_idx;
1381
1382         unsigned int class_idx;
1383         enum fullness_group fg;
1384         struct size_class *class;
1385         struct mapping_area *area;
1386
1387         obj = handle_to_obj(handle);
1388         obj_to_location(obj, &page, &obj_idx);
1389         zspage = get_zspage(page);
1390         get_zspage_mapping(zspage, &class_idx, &fg);
1391         class = pool->size_class[class_idx];
1392         off = (class->size * obj_idx) & ~PAGE_MASK;
1393
1394         area = this_cpu_ptr(&zs_map_area);
1395         if (off + class->size <= PAGE_SIZE)
1396                 kunmap_atomic(area->vm_addr);
1397         else {
1398                 struct page *pages[2];
1399
1400                 pages[0] = page;
1401                 pages[1] = get_next_page(page);
1402                 BUG_ON(!pages[1]);
1403
1404                 __zs_unmap_object(area, pages, off, class->size);
1405         }
1406         put_cpu_var(zs_map_area);
1407
1408         migrate_read_unlock(zspage);
1409         unpin_tag(handle);
1410 }
1411 EXPORT_SYMBOL_GPL(zs_unmap_object);
1412
1413 /**
1414  * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1415  *                        zsmalloc &size_class.
1416  * @pool: zsmalloc pool to use
1417  *
1418  * The function returns the size of the first huge class - any object of equal
1419  * or bigger size will be stored in zspage consisting of a single physical
1420  * page.
1421  *
1422  * Context: Any context.
1423  *
1424  * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1425  */
1426 size_t zs_huge_class_size(struct zs_pool *pool)
1427 {
1428         return huge_class_size;
1429 }
1430 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1431
1432 static unsigned long obj_malloc(struct size_class *class,
1433                                 struct zspage *zspage, unsigned long handle)
1434 {
1435         int i, nr_page, offset;
1436         unsigned long obj;
1437         struct link_free *link;
1438
1439         struct page *m_page;
1440         unsigned long m_offset;
1441         void *vaddr;
1442
1443         handle |= OBJ_ALLOCATED_TAG;
1444         obj = get_freeobj(zspage);
1445
1446         offset = obj * class->size;
1447         nr_page = offset >> PAGE_SHIFT;
1448         m_offset = offset & ~PAGE_MASK;
1449         m_page = get_first_page(zspage);
1450
1451         for (i = 0; i < nr_page; i++)
1452                 m_page = get_next_page(m_page);
1453
1454         vaddr = kmap_atomic(m_page);
1455         link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1456         set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1457         if (likely(!PageHugeObject(m_page)))
1458                 /* record handle in the header of allocated chunk */
1459                 link->handle = handle;
1460         else
1461                 /* record handle to page->index */
1462                 zspage->first_page->index = handle;
1463
1464         kunmap_atomic(vaddr);
1465         mod_zspage_inuse(zspage, 1);
1466         zs_stat_inc(class, OBJ_USED, 1);
1467
1468         obj = location_to_obj(m_page, obj);
1469
1470         return obj;
1471 }
1472
1473
1474 /**
1475  * zs_malloc - Allocate block of given size from pool.
1476  * @pool: pool to allocate from
1477  * @size: size of block to allocate
1478  * @gfp: gfp flags when allocating object
1479  *
1480  * On success, handle to the allocated object is returned,
1481  * otherwise 0.
1482  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1483  */
1484 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1485 {
1486         unsigned long handle, obj;
1487         struct size_class *class;
1488         enum fullness_group newfg;
1489         struct zspage *zspage;
1490
1491         if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1492                 return 0;
1493
1494         handle = cache_alloc_handle(pool, gfp);
1495         if (!handle)
1496                 return 0;
1497
1498         /* extra space in chunk to keep the handle */
1499         size += ZS_HANDLE_SIZE;
1500         class = pool->size_class[get_size_class_index(size)];
1501
1502         spin_lock(&class->lock);
1503         zspage = find_get_zspage(class);
1504         if (likely(zspage)) {
1505                 obj = obj_malloc(class, zspage, handle);
1506                 /* Now move the zspage to another fullness group, if required */
1507                 fix_fullness_group(class, zspage);
1508                 record_obj(handle, obj);
1509                 spin_unlock(&class->lock);
1510
1511                 return handle;
1512         }
1513
1514         spin_unlock(&class->lock);
1515
1516         zspage = alloc_zspage(pool, class, gfp);
1517         if (!zspage) {
1518                 cache_free_handle(pool, handle);
1519                 return 0;
1520         }
1521
1522         spin_lock(&class->lock);
1523         obj = obj_malloc(class, zspage, handle);
1524         newfg = get_fullness_group(class, zspage);
1525         insert_zspage(class, zspage, newfg);
1526         set_zspage_mapping(zspage, class->index, newfg);
1527         record_obj(handle, obj);
1528         atomic_long_add(class->pages_per_zspage,
1529                                 &pool->pages_allocated);
1530         zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1531
1532         /* We completely set up zspage so mark them as movable */
1533         SetZsPageMovable(pool, zspage);
1534         spin_unlock(&class->lock);
1535
1536         return handle;
1537 }
1538 EXPORT_SYMBOL_GPL(zs_malloc);
1539
1540 static void obj_free(struct size_class *class, unsigned long obj)
1541 {
1542         struct link_free *link;
1543         struct zspage *zspage;
1544         struct page *f_page;
1545         unsigned long f_offset;
1546         unsigned int f_objidx;
1547         void *vaddr;
1548
1549         obj &= ~OBJ_ALLOCATED_TAG;
1550         obj_to_location(obj, &f_page, &f_objidx);
1551         f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1552         zspage = get_zspage(f_page);
1553
1554         vaddr = kmap_atomic(f_page);
1555
1556         /* Insert this object in containing zspage's freelist */
1557         link = (struct link_free *)(vaddr + f_offset);
1558         link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1559         kunmap_atomic(vaddr);
1560         set_freeobj(zspage, f_objidx);
1561         mod_zspage_inuse(zspage, -1);
1562         zs_stat_dec(class, OBJ_USED, 1);
1563 }
1564
1565 void zs_free(struct zs_pool *pool, unsigned long handle)
1566 {
1567         struct zspage *zspage;
1568         struct page *f_page;
1569         unsigned long obj;
1570         unsigned int f_objidx;
1571         int class_idx;
1572         struct size_class *class;
1573         enum fullness_group fullness;
1574         bool isolated;
1575
1576         if (unlikely(!handle))
1577                 return;
1578
1579         pin_tag(handle);
1580         obj = handle_to_obj(handle);
1581         obj_to_location(obj, &f_page, &f_objidx);
1582         zspage = get_zspage(f_page);
1583
1584         migrate_read_lock(zspage);
1585
1586         get_zspage_mapping(zspage, &class_idx, &fullness);
1587         class = pool->size_class[class_idx];
1588
1589         spin_lock(&class->lock);
1590         obj_free(class, obj);
1591         fullness = fix_fullness_group(class, zspage);
1592         if (fullness != ZS_EMPTY) {
1593                 migrate_read_unlock(zspage);
1594                 goto out;
1595         }
1596
1597         isolated = is_zspage_isolated(zspage);
1598         migrate_read_unlock(zspage);
1599         /* If zspage is isolated, zs_page_putback will free the zspage */
1600         if (likely(!isolated))
1601                 free_zspage(pool, class, zspage);
1602 out:
1603
1604         spin_unlock(&class->lock);
1605         unpin_tag(handle);
1606         cache_free_handle(pool, handle);
1607 }
1608 EXPORT_SYMBOL_GPL(zs_free);
1609
1610 static void zs_object_copy(struct size_class *class, unsigned long dst,
1611                                 unsigned long src)
1612 {
1613         struct page *s_page, *d_page;
1614         unsigned int s_objidx, d_objidx;
1615         unsigned long s_off, d_off;
1616         void *s_addr, *d_addr;
1617         int s_size, d_size, size;
1618         int written = 0;
1619
1620         s_size = d_size = class->size;
1621
1622         obj_to_location(src, &s_page, &s_objidx);
1623         obj_to_location(dst, &d_page, &d_objidx);
1624
1625         s_off = (class->size * s_objidx) & ~PAGE_MASK;
1626         d_off = (class->size * d_objidx) & ~PAGE_MASK;
1627
1628         if (s_off + class->size > PAGE_SIZE)
1629                 s_size = PAGE_SIZE - s_off;
1630
1631         if (d_off + class->size > PAGE_SIZE)
1632                 d_size = PAGE_SIZE - d_off;
1633
1634         s_addr = kmap_atomic(s_page);
1635         d_addr = kmap_atomic(d_page);
1636
1637         while (1) {
1638                 size = min(s_size, d_size);
1639                 memcpy(d_addr + d_off, s_addr + s_off, size);
1640                 written += size;
1641
1642                 if (written == class->size)
1643                         break;
1644
1645                 s_off += size;
1646                 s_size -= size;
1647                 d_off += size;
1648                 d_size -= size;
1649
1650                 if (s_off >= PAGE_SIZE) {
1651                         kunmap_atomic(d_addr);
1652                         kunmap_atomic(s_addr);
1653                         s_page = get_next_page(s_page);
1654                         s_addr = kmap_atomic(s_page);
1655                         d_addr = kmap_atomic(d_page);
1656                         s_size = class->size - written;
1657                         s_off = 0;
1658                 }
1659
1660                 if (d_off >= PAGE_SIZE) {
1661                         kunmap_atomic(d_addr);
1662                         d_page = get_next_page(d_page);
1663                         d_addr = kmap_atomic(d_page);
1664                         d_size = class->size - written;
1665                         d_off = 0;
1666                 }
1667         }
1668
1669         kunmap_atomic(d_addr);
1670         kunmap_atomic(s_addr);
1671 }
1672
1673 /*
1674  * Find alloced object in zspage from index object and
1675  * return handle.
1676  */
1677 static unsigned long find_alloced_obj(struct size_class *class,
1678                                         struct page *page, int *obj_idx)
1679 {
1680         unsigned long head;
1681         int offset = 0;
1682         int index = *obj_idx;
1683         unsigned long handle = 0;
1684         void *addr = kmap_atomic(page);
1685
1686         offset = get_first_obj_offset(page);
1687         offset += class->size * index;
1688
1689         while (offset < PAGE_SIZE) {
1690                 head = obj_to_head(page, addr + offset);
1691                 if (head & OBJ_ALLOCATED_TAG) {
1692                         handle = head & ~OBJ_ALLOCATED_TAG;
1693                         if (trypin_tag(handle))
1694                                 break;
1695                         handle = 0;
1696                 }
1697
1698                 offset += class->size;
1699                 index++;
1700         }
1701
1702         kunmap_atomic(addr);
1703
1704         *obj_idx = index;
1705
1706         return handle;
1707 }
1708
1709 struct zs_compact_control {
1710         /* Source spage for migration which could be a subpage of zspage */
1711         struct page *s_page;
1712         /* Destination page for migration which should be a first page
1713          * of zspage. */
1714         struct page *d_page;
1715          /* Starting object index within @s_page which used for live object
1716           * in the subpage. */
1717         int obj_idx;
1718 };
1719
1720 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1721                                 struct zs_compact_control *cc)
1722 {
1723         unsigned long used_obj, free_obj;
1724         unsigned long handle;
1725         struct page *s_page = cc->s_page;
1726         struct page *d_page = cc->d_page;
1727         int obj_idx = cc->obj_idx;
1728         int ret = 0;
1729
1730         while (1) {
1731                 handle = find_alloced_obj(class, s_page, &obj_idx);
1732                 if (!handle) {
1733                         s_page = get_next_page(s_page);
1734                         if (!s_page)
1735                                 break;
1736                         obj_idx = 0;
1737                         continue;
1738                 }
1739
1740                 /* Stop if there is no more space */
1741                 if (zspage_full(class, get_zspage(d_page))) {
1742                         unpin_tag(handle);
1743                         ret = -ENOMEM;
1744                         break;
1745                 }
1746
1747                 used_obj = handle_to_obj(handle);
1748                 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1749                 zs_object_copy(class, free_obj, used_obj);
1750                 obj_idx++;
1751                 /*
1752                  * record_obj updates handle's value to free_obj and it will
1753                  * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1754                  * breaks synchronization using pin_tag(e,g, zs_free) so
1755                  * let's keep the lock bit.
1756                  */
1757                 free_obj |= BIT(HANDLE_PIN_BIT);
1758                 record_obj(handle, free_obj);
1759                 unpin_tag(handle);
1760                 obj_free(class, used_obj);
1761         }
1762
1763         /* Remember last position in this iteration */
1764         cc->s_page = s_page;
1765         cc->obj_idx = obj_idx;
1766
1767         return ret;
1768 }
1769
1770 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1771 {
1772         int i;
1773         struct zspage *zspage;
1774         enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1775
1776         if (!source) {
1777                 fg[0] = ZS_ALMOST_FULL;
1778                 fg[1] = ZS_ALMOST_EMPTY;
1779         }
1780
1781         for (i = 0; i < 2; i++) {
1782                 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1783                                                         struct zspage, list);
1784                 if (zspage) {
1785                         VM_BUG_ON(is_zspage_isolated(zspage));
1786                         remove_zspage(class, zspage, fg[i]);
1787                         return zspage;
1788                 }
1789         }
1790
1791         return zspage;
1792 }
1793
1794 /*
1795  * putback_zspage - add @zspage into right class's fullness list
1796  * @class: destination class
1797  * @zspage: target page
1798  *
1799  * Return @zspage's fullness_group
1800  */
1801 static enum fullness_group putback_zspage(struct size_class *class,
1802                         struct zspage *zspage)
1803 {
1804         enum fullness_group fullness;
1805
1806         VM_BUG_ON(is_zspage_isolated(zspage));
1807
1808         fullness = get_fullness_group(class, zspage);
1809         insert_zspage(class, zspage, fullness);
1810         set_zspage_mapping(zspage, class->index, fullness);
1811
1812         return fullness;
1813 }
1814
1815 #ifdef CONFIG_COMPACTION
1816 static struct dentry *zs_mount(struct file_system_type *fs_type,
1817                                 int flags, const char *dev_name, void *data)
1818 {
1819         static const struct dentry_operations ops = {
1820                 .d_dname = simple_dname,
1821         };
1822
1823         return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1824 }
1825
1826 static struct file_system_type zsmalloc_fs = {
1827         .name           = "zsmalloc",
1828         .mount          = zs_mount,
1829         .kill_sb        = kill_anon_super,
1830 };
1831
1832 static int zsmalloc_mount(void)
1833 {
1834         int ret = 0;
1835
1836         zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1837         if (IS_ERR(zsmalloc_mnt))
1838                 ret = PTR_ERR(zsmalloc_mnt);
1839
1840         return ret;
1841 }
1842
1843 static void zsmalloc_unmount(void)
1844 {
1845         kern_unmount(zsmalloc_mnt);
1846 }
1847
1848 static void migrate_lock_init(struct zspage *zspage)
1849 {
1850         rwlock_init(&zspage->lock);
1851 }
1852
1853 static void migrate_read_lock(struct zspage *zspage)
1854 {
1855         read_lock(&zspage->lock);
1856 }
1857
1858 static void migrate_read_unlock(struct zspage *zspage)
1859 {
1860         read_unlock(&zspage->lock);
1861 }
1862
1863 static void migrate_write_lock(struct zspage *zspage)
1864 {
1865         write_lock(&zspage->lock);
1866 }
1867
1868 static void migrate_write_unlock(struct zspage *zspage)
1869 {
1870         write_unlock(&zspage->lock);
1871 }
1872
1873 /* Number of isolated subpage for *page migration* in this zspage */
1874 static void inc_zspage_isolation(struct zspage *zspage)
1875 {
1876         zspage->isolated++;
1877 }
1878
1879 static void dec_zspage_isolation(struct zspage *zspage)
1880 {
1881         zspage->isolated--;
1882 }
1883
1884 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1885                                 struct page *newpage, struct page *oldpage)
1886 {
1887         struct page *page;
1888         struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1889         int idx = 0;
1890
1891         page = get_first_page(zspage);
1892         do {
1893                 if (page == oldpage)
1894                         pages[idx] = newpage;
1895                 else
1896                         pages[idx] = page;
1897                 idx++;
1898         } while ((page = get_next_page(page)) != NULL);
1899
1900         create_page_chain(class, zspage, pages);
1901         set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1902         if (unlikely(PageHugeObject(oldpage)))
1903                 newpage->index = oldpage->index;
1904         __SetPageMovable(newpage, page_mapping(oldpage));
1905 }
1906
1907 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1908 {
1909         struct zs_pool *pool;
1910         struct size_class *class;
1911         int class_idx;
1912         enum fullness_group fullness;
1913         struct zspage *zspage;
1914         struct address_space *mapping;
1915
1916         /*
1917          * Page is locked so zspage couldn't be destroyed. For detail, look at
1918          * lock_zspage in free_zspage.
1919          */
1920         VM_BUG_ON_PAGE(!PageMovable(page), page);
1921         VM_BUG_ON_PAGE(PageIsolated(page), page);
1922
1923         zspage = get_zspage(page);
1924
1925         /*
1926          * Without class lock, fullness could be stale while class_idx is okay
1927          * because class_idx is constant unless page is freed so we should get
1928          * fullness again under class lock.
1929          */
1930         get_zspage_mapping(zspage, &class_idx, &fullness);
1931         mapping = page_mapping(page);
1932         pool = mapping->private_data;
1933         class = pool->size_class[class_idx];
1934
1935         spin_lock(&class->lock);
1936         if (get_zspage_inuse(zspage) == 0) {
1937                 spin_unlock(&class->lock);
1938                 return false;
1939         }
1940
1941         /* zspage is isolated for object migration */
1942         if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1943                 spin_unlock(&class->lock);
1944                 return false;
1945         }
1946
1947         /*
1948          * If this is first time isolation for the zspage, isolate zspage from
1949          * size_class to prevent further object allocation from the zspage.
1950          */
1951         if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1952                 get_zspage_mapping(zspage, &class_idx, &fullness);
1953                 remove_zspage(class, zspage, fullness);
1954         }
1955
1956         inc_zspage_isolation(zspage);
1957         spin_unlock(&class->lock);
1958
1959         return true;
1960 }
1961
1962 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1963                 struct page *page, enum migrate_mode mode)
1964 {
1965         struct zs_pool *pool;
1966         struct size_class *class;
1967         int class_idx;
1968         enum fullness_group fullness;
1969         struct zspage *zspage;
1970         struct page *dummy;
1971         void *s_addr, *d_addr, *addr;
1972         int offset, pos;
1973         unsigned long handle, head;
1974         unsigned long old_obj, new_obj;
1975         unsigned int obj_idx;
1976         int ret = -EAGAIN;
1977
1978         /*
1979          * We cannot support the _NO_COPY case here, because copy needs to
1980          * happen under the zs lock, which does not work with
1981          * MIGRATE_SYNC_NO_COPY workflow.
1982          */
1983         if (mode == MIGRATE_SYNC_NO_COPY)
1984                 return -EINVAL;
1985
1986         VM_BUG_ON_PAGE(!PageMovable(page), page);
1987         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1988
1989         zspage = get_zspage(page);
1990
1991         /* Concurrent compactor cannot migrate any subpage in zspage */
1992         migrate_write_lock(zspage);
1993         get_zspage_mapping(zspage, &class_idx, &fullness);
1994         pool = mapping->private_data;
1995         class = pool->size_class[class_idx];
1996         offset = get_first_obj_offset(page);
1997
1998         spin_lock(&class->lock);
1999         if (!get_zspage_inuse(zspage)) {
2000                 /*
2001                  * Set "offset" to end of the page so that every loops
2002                  * skips unnecessary object scanning.
2003                  */
2004                 offset = PAGE_SIZE;
2005         }
2006
2007         pos = offset;
2008         s_addr = kmap_atomic(page);
2009         while (pos < PAGE_SIZE) {
2010                 head = obj_to_head(page, s_addr + pos);
2011                 if (head & OBJ_ALLOCATED_TAG) {
2012                         handle = head & ~OBJ_ALLOCATED_TAG;
2013                         if (!trypin_tag(handle))
2014                                 goto unpin_objects;
2015                 }
2016                 pos += class->size;
2017         }
2018
2019         /*
2020          * Here, any user cannot access all objects in the zspage so let's move.
2021          */
2022         d_addr = kmap_atomic(newpage);
2023         memcpy(d_addr, s_addr, PAGE_SIZE);
2024         kunmap_atomic(d_addr);
2025
2026         for (addr = s_addr + offset; addr < s_addr + pos;
2027                                         addr += class->size) {
2028                 head = obj_to_head(page, addr);
2029                 if (head & OBJ_ALLOCATED_TAG) {
2030                         handle = head & ~OBJ_ALLOCATED_TAG;
2031                         if (!testpin_tag(handle))
2032                                 BUG();
2033
2034                         old_obj = handle_to_obj(handle);
2035                         obj_to_location(old_obj, &dummy, &obj_idx);
2036                         new_obj = (unsigned long)location_to_obj(newpage,
2037                                                                 obj_idx);
2038                         new_obj |= BIT(HANDLE_PIN_BIT);
2039                         record_obj(handle, new_obj);
2040                 }
2041         }
2042
2043         replace_sub_page(class, zspage, newpage, page);
2044         get_page(newpage);
2045
2046         dec_zspage_isolation(zspage);
2047
2048         /*
2049          * Page migration is done so let's putback isolated zspage to
2050          * the list if @page is final isolated subpage in the zspage.
2051          */
2052         if (!is_zspage_isolated(zspage))
2053                 putback_zspage(class, zspage);
2054
2055         reset_page(page);
2056         put_page(page);
2057         page = newpage;
2058
2059         ret = MIGRATEPAGE_SUCCESS;
2060 unpin_objects:
2061         for (addr = s_addr + offset; addr < s_addr + pos;
2062                                                 addr += class->size) {
2063                 head = obj_to_head(page, addr);
2064                 if (head & OBJ_ALLOCATED_TAG) {
2065                         handle = head & ~OBJ_ALLOCATED_TAG;
2066                         if (!testpin_tag(handle))
2067                                 BUG();
2068                         unpin_tag(handle);
2069                 }
2070         }
2071         kunmap_atomic(s_addr);
2072         spin_unlock(&class->lock);
2073         migrate_write_unlock(zspage);
2074
2075         return ret;
2076 }
2077
2078 void zs_page_putback(struct page *page)
2079 {
2080         struct zs_pool *pool;
2081         struct size_class *class;
2082         int class_idx;
2083         enum fullness_group fg;
2084         struct address_space *mapping;
2085         struct zspage *zspage;
2086
2087         VM_BUG_ON_PAGE(!PageMovable(page), page);
2088         VM_BUG_ON_PAGE(!PageIsolated(page), page);
2089
2090         zspage = get_zspage(page);
2091         get_zspage_mapping(zspage, &class_idx, &fg);
2092         mapping = page_mapping(page);
2093         pool = mapping->private_data;
2094         class = pool->size_class[class_idx];
2095
2096         spin_lock(&class->lock);
2097         dec_zspage_isolation(zspage);
2098         if (!is_zspage_isolated(zspage)) {
2099                 fg = putback_zspage(class, zspage);
2100                 /*
2101                  * Due to page_lock, we cannot free zspage immediately
2102                  * so let's defer.
2103                  */
2104                 if (fg == ZS_EMPTY)
2105                         schedule_work(&pool->free_work);
2106         }
2107         spin_unlock(&class->lock);
2108 }
2109
2110 const struct address_space_operations zsmalloc_aops = {
2111         .isolate_page = zs_page_isolate,
2112         .migratepage = zs_page_migrate,
2113         .putback_page = zs_page_putback,
2114 };
2115
2116 static int zs_register_migration(struct zs_pool *pool)
2117 {
2118         pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2119         if (IS_ERR(pool->inode)) {
2120                 pool->inode = NULL;
2121                 return 1;
2122         }
2123
2124         pool->inode->i_mapping->private_data = pool;
2125         pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2126         return 0;
2127 }
2128
2129 static void zs_unregister_migration(struct zs_pool *pool)
2130 {
2131         flush_work(&pool->free_work);
2132         iput(pool->inode);
2133 }
2134
2135 /*
2136  * Caller should hold page_lock of all pages in the zspage
2137  * In here, we cannot use zspage meta data.
2138  */
2139 static void async_free_zspage(struct work_struct *work)
2140 {
2141         int i;
2142         struct size_class *class;
2143         unsigned int class_idx;
2144         enum fullness_group fullness;
2145         struct zspage *zspage, *tmp;
2146         LIST_HEAD(free_pages);
2147         struct zs_pool *pool = container_of(work, struct zs_pool,
2148                                         free_work);
2149
2150         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2151                 class = pool->size_class[i];
2152                 if (class->index != i)
2153                         continue;
2154
2155                 spin_lock(&class->lock);
2156                 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2157                 spin_unlock(&class->lock);
2158         }
2159
2160
2161         list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2162                 list_del(&zspage->list);
2163                 lock_zspage(zspage);
2164
2165                 get_zspage_mapping(zspage, &class_idx, &fullness);
2166                 VM_BUG_ON(fullness != ZS_EMPTY);
2167                 class = pool->size_class[class_idx];
2168                 spin_lock(&class->lock);
2169                 __free_zspage(pool, pool->size_class[class_idx], zspage);
2170                 spin_unlock(&class->lock);
2171         }
2172 };
2173
2174 static void kick_deferred_free(struct zs_pool *pool)
2175 {
2176         schedule_work(&pool->free_work);
2177 }
2178
2179 static void init_deferred_free(struct zs_pool *pool)
2180 {
2181         INIT_WORK(&pool->free_work, async_free_zspage);
2182 }
2183
2184 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2185 {
2186         struct page *page = get_first_page(zspage);
2187
2188         do {
2189                 WARN_ON(!trylock_page(page));
2190                 __SetPageMovable(page, pool->inode->i_mapping);
2191                 unlock_page(page);
2192         } while ((page = get_next_page(page)) != NULL);
2193 }
2194 #endif
2195
2196 /*
2197  *
2198  * Based on the number of unused allocated objects calculate
2199  * and return the number of pages that we can free.
2200  */
2201 static unsigned long zs_can_compact(struct size_class *class)
2202 {
2203         unsigned long obj_wasted;
2204         unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2205         unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2206
2207         if (obj_allocated <= obj_used)
2208                 return 0;
2209
2210         obj_wasted = obj_allocated - obj_used;
2211         obj_wasted /= class->objs_per_zspage;
2212
2213         return obj_wasted * class->pages_per_zspage;
2214 }
2215
2216 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2217 {
2218         struct zs_compact_control cc;
2219         struct zspage *src_zspage;
2220         struct zspage *dst_zspage = NULL;
2221
2222         spin_lock(&class->lock);
2223         while ((src_zspage = isolate_zspage(class, true))) {
2224
2225                 if (!zs_can_compact(class))
2226                         break;
2227
2228                 cc.obj_idx = 0;
2229                 cc.s_page = get_first_page(src_zspage);
2230
2231                 while ((dst_zspage = isolate_zspage(class, false))) {
2232                         cc.d_page = get_first_page(dst_zspage);
2233                         /*
2234                          * If there is no more space in dst_page, resched
2235                          * and see if anyone had allocated another zspage.
2236                          */
2237                         if (!migrate_zspage(pool, class, &cc))
2238                                 break;
2239
2240                         putback_zspage(class, dst_zspage);
2241                 }
2242
2243                 /* Stop if we couldn't find slot */
2244                 if (dst_zspage == NULL)
2245                         break;
2246
2247                 putback_zspage(class, dst_zspage);
2248                 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2249                         free_zspage(pool, class, src_zspage);
2250                         pool->stats.pages_compacted += class->pages_per_zspage;
2251                 }
2252                 spin_unlock(&class->lock);
2253                 cond_resched();
2254                 spin_lock(&class->lock);
2255         }
2256
2257         if (src_zspage)
2258                 putback_zspage(class, src_zspage);
2259
2260         spin_unlock(&class->lock);
2261 }
2262
2263 unsigned long zs_compact(struct zs_pool *pool)
2264 {
2265         int i;
2266         struct size_class *class;
2267
2268         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2269                 class = pool->size_class[i];
2270                 if (!class)
2271                         continue;
2272                 if (class->index != i)
2273                         continue;
2274                 __zs_compact(pool, class);
2275         }
2276
2277         return pool->stats.pages_compacted;
2278 }
2279 EXPORT_SYMBOL_GPL(zs_compact);
2280
2281 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2282 {
2283         memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2284 }
2285 EXPORT_SYMBOL_GPL(zs_pool_stats);
2286
2287 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2288                 struct shrink_control *sc)
2289 {
2290         unsigned long pages_freed;
2291         struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2292                         shrinker);
2293
2294         pages_freed = pool->stats.pages_compacted;
2295         /*
2296          * Compact classes and calculate compaction delta.
2297          * Can run concurrently with a manually triggered
2298          * (by user) compaction.
2299          */
2300         pages_freed = zs_compact(pool) - pages_freed;
2301
2302         return pages_freed ? pages_freed : SHRINK_STOP;
2303 }
2304
2305 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2306                 struct shrink_control *sc)
2307 {
2308         int i;
2309         struct size_class *class;
2310         unsigned long pages_to_free = 0;
2311         struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2312                         shrinker);
2313
2314         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2315                 class = pool->size_class[i];
2316                 if (!class)
2317                         continue;
2318                 if (class->index != i)
2319                         continue;
2320
2321                 pages_to_free += zs_can_compact(class);
2322         }
2323
2324         return pages_to_free;
2325 }
2326
2327 static void zs_unregister_shrinker(struct zs_pool *pool)
2328 {
2329         unregister_shrinker(&pool->shrinker);
2330 }
2331
2332 static int zs_register_shrinker(struct zs_pool *pool)
2333 {
2334         pool->shrinker.scan_objects = zs_shrinker_scan;
2335         pool->shrinker.count_objects = zs_shrinker_count;
2336         pool->shrinker.batch = 0;
2337         pool->shrinker.seeks = DEFAULT_SEEKS;
2338
2339         return register_shrinker(&pool->shrinker);
2340 }
2341
2342 /**
2343  * zs_create_pool - Creates an allocation pool to work from.
2344  * @name: pool name to be created
2345  *
2346  * This function must be called before anything when using
2347  * the zsmalloc allocator.
2348  *
2349  * On success, a pointer to the newly created pool is returned,
2350  * otherwise NULL.
2351  */
2352 struct zs_pool *zs_create_pool(const char *name)
2353 {
2354         int i;
2355         struct zs_pool *pool;
2356         struct size_class *prev_class = NULL;
2357
2358         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2359         if (!pool)
2360                 return NULL;
2361
2362         init_deferred_free(pool);
2363
2364         pool->name = kstrdup(name, GFP_KERNEL);
2365         if (!pool->name)
2366                 goto err;
2367
2368         if (create_cache(pool))
2369                 goto err;
2370
2371         /*
2372          * Iterate reversely, because, size of size_class that we want to use
2373          * for merging should be larger or equal to current size.
2374          */
2375         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2376                 int size;
2377                 int pages_per_zspage;
2378                 int objs_per_zspage;
2379                 struct size_class *class;
2380                 int fullness = 0;
2381
2382                 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2383                 if (size > ZS_MAX_ALLOC_SIZE)
2384                         size = ZS_MAX_ALLOC_SIZE;
2385                 pages_per_zspage = get_pages_per_zspage(size);
2386                 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2387
2388                 /*
2389                  * We iterate from biggest down to smallest classes,
2390                  * so huge_class_size holds the size of the first huge
2391                  * class. Any object bigger than or equal to that will
2392                  * endup in the huge class.
2393                  */
2394                 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2395                                 !huge_class_size) {
2396                         huge_class_size = size;
2397                         /*
2398                          * The object uses ZS_HANDLE_SIZE bytes to store the
2399                          * handle. We need to subtract it, because zs_malloc()
2400                          * unconditionally adds handle size before it performs
2401                          * size class search - so object may be smaller than
2402                          * huge class size, yet it still can end up in the huge
2403                          * class because it grows by ZS_HANDLE_SIZE extra bytes
2404                          * right before class lookup.
2405                          */
2406                         huge_class_size -= (ZS_HANDLE_SIZE - 1);
2407                 }
2408
2409                 /*
2410                  * size_class is used for normal zsmalloc operation such
2411                  * as alloc/free for that size. Although it is natural that we
2412                  * have one size_class for each size, there is a chance that we
2413                  * can get more memory utilization if we use one size_class for
2414                  * many different sizes whose size_class have same
2415                  * characteristics. So, we makes size_class point to
2416                  * previous size_class if possible.
2417                  */
2418                 if (prev_class) {
2419                         if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2420                                 pool->size_class[i] = prev_class;
2421                                 continue;
2422                         }
2423                 }
2424
2425                 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2426                 if (!class)
2427                         goto err;
2428
2429                 class->size = size;
2430                 class->index = i;
2431                 class->pages_per_zspage = pages_per_zspage;
2432                 class->objs_per_zspage = objs_per_zspage;
2433                 spin_lock_init(&class->lock);
2434                 pool->size_class[i] = class;
2435                 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2436                                                         fullness++)
2437                         INIT_LIST_HEAD(&class->fullness_list[fullness]);
2438
2439                 prev_class = class;
2440         }
2441
2442         /* debug only, don't abort if it fails */
2443         zs_pool_stat_create(pool, name);
2444
2445         if (zs_register_migration(pool))
2446                 goto err;
2447
2448         /*
2449          * Not critical since shrinker is only used to trigger internal
2450          * defragmentation of the pool which is pretty optional thing.  If
2451          * registration fails we still can use the pool normally and user can
2452          * trigger compaction manually. Thus, ignore return code.
2453          */
2454         zs_register_shrinker(pool);
2455
2456         return pool;
2457
2458 err:
2459         zs_destroy_pool(pool);
2460         return NULL;
2461 }
2462 EXPORT_SYMBOL_GPL(zs_create_pool);
2463
2464 void zs_destroy_pool(struct zs_pool *pool)
2465 {
2466         int i;
2467
2468         zs_unregister_shrinker(pool);
2469         zs_unregister_migration(pool);
2470         zs_pool_stat_destroy(pool);
2471
2472         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2473                 int fg;
2474                 struct size_class *class = pool->size_class[i];
2475
2476                 if (!class)
2477                         continue;
2478
2479                 if (class->index != i)
2480                         continue;
2481
2482                 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2483                         if (!list_empty(&class->fullness_list[fg])) {
2484                                 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2485                                         class->size, fg);
2486                         }
2487                 }
2488                 kfree(class);
2489         }
2490
2491         destroy_cache(pool);
2492         kfree(pool->name);
2493         kfree(pool);
2494 }
2495 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2496
2497 static int __init zs_init(void)
2498 {
2499         int ret;
2500
2501         ret = zsmalloc_mount();
2502         if (ret)
2503                 goto out;
2504
2505         ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2506                                 zs_cpu_prepare, zs_cpu_dead);
2507         if (ret)
2508                 goto hp_setup_fail;
2509
2510 #ifdef CONFIG_ZPOOL
2511         zpool_register_driver(&zs_zpool_driver);
2512 #endif
2513
2514         zs_stat_init();
2515
2516         return 0;
2517
2518 hp_setup_fail:
2519         zsmalloc_unmount();
2520 out:
2521         return ret;
2522 }
2523
2524 static void __exit zs_exit(void)
2525 {
2526 #ifdef CONFIG_ZPOOL
2527         zpool_unregister_driver(&zs_zpool_driver);
2528 #endif
2529         zsmalloc_unmount();
2530         cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2531
2532         zs_stat_exit();
2533 }
2534
2535 module_init(zs_init);
2536 module_exit(zs_exit);
2537
2538 MODULE_LICENSE("Dual BSD/GPL");
2539 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");