Merge tag 'batadv-next-for-davem-20200114' of git://git.open-mesh.org/linux-merge
[linux-2.6-microblaze.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/pseudo_fs.h>
56 #include <linux/migrate.h>
57 #include <linux/wait.h>
58 #include <linux/pagemap.h>
59 #include <linux/fs.h>
60
61 #define ZSPAGE_MAGIC    0x58
62
63 /*
64  * This must be power of 2 and greater than of equal to sizeof(link_free).
65  * These two conditions ensure that any 'struct link_free' itself doesn't
66  * span more than 1 page which avoids complex case of mapping 2 pages simply
67  * to restore link_free pointer values.
68  */
69 #define ZS_ALIGN                8
70
71 /*
72  * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
73  * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
74  */
75 #define ZS_MAX_ZSPAGE_ORDER 2
76 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
77
78 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
79
80 /*
81  * Object location (<PFN>, <obj_idx>) is encoded as
82  * as single (unsigned long) handle value.
83  *
84  * Note that object index <obj_idx> starts from 0.
85  *
86  * This is made more complicated by various memory models and PAE.
87  */
88
89 #ifndef MAX_POSSIBLE_PHYSMEM_BITS
90 #ifdef MAX_PHYSMEM_BITS
91 #define MAX_POSSIBLE_PHYSMEM_BITS MAX_PHYSMEM_BITS
92 #else
93 /*
94  * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
95  * be PAGE_SHIFT
96  */
97 #define MAX_POSSIBLE_PHYSMEM_BITS BITS_PER_LONG
98 #endif
99 #endif
100
101 #define _PFN_BITS               (MAX_POSSIBLE_PHYSMEM_BITS - PAGE_SHIFT)
102
103 /*
104  * Memory for allocating for handle keeps object position by
105  * encoding <page, obj_idx> and the encoded value has a room
106  * in least bit(ie, look at obj_to_location).
107  * We use the bit to synchronize between object access by
108  * user and migration.
109  */
110 #define HANDLE_PIN_BIT  0
111
112 /*
113  * Head in allocated object should have OBJ_ALLOCATED_TAG
114  * to identify the object was allocated or not.
115  * It's okay to add the status bit in the least bit because
116  * header keeps handle which is 4byte-aligned address so we
117  * have room for two bit at least.
118  */
119 #define OBJ_ALLOCATED_TAG 1
120 #define OBJ_TAG_BITS 1
121 #define OBJ_INDEX_BITS  (BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
122 #define OBJ_INDEX_MASK  ((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
123
124 #define FULLNESS_BITS   2
125 #define CLASS_BITS      8
126 #define ISOLATED_BITS   3
127 #define MAGIC_VAL_BITS  8
128
129 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
130 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
131 #define ZS_MIN_ALLOC_SIZE \
132         MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
133 /* each chunk includes extra space to keep handle */
134 #define ZS_MAX_ALLOC_SIZE       PAGE_SIZE
135
136 /*
137  * On systems with 4K page size, this gives 255 size classes! There is a
138  * trader-off here:
139  *  - Large number of size classes is potentially wasteful as free page are
140  *    spread across these classes
141  *  - Small number of size classes causes large internal fragmentation
142  *  - Probably its better to use specific size classes (empirically
143  *    determined). NOTE: all those class sizes must be set as multiple of
144  *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
145  *
146  *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
147  *  (reason above)
148  */
149 #define ZS_SIZE_CLASS_DELTA     (PAGE_SIZE >> CLASS_BITS)
150 #define ZS_SIZE_CLASSES (DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
151                                       ZS_SIZE_CLASS_DELTA) + 1)
152
153 enum fullness_group {
154         ZS_EMPTY,
155         ZS_ALMOST_EMPTY,
156         ZS_ALMOST_FULL,
157         ZS_FULL,
158         NR_ZS_FULLNESS,
159 };
160
161 enum zs_stat_type {
162         CLASS_EMPTY,
163         CLASS_ALMOST_EMPTY,
164         CLASS_ALMOST_FULL,
165         CLASS_FULL,
166         OBJ_ALLOCATED,
167         OBJ_USED,
168         NR_ZS_STAT_TYPE,
169 };
170
171 struct zs_size_stat {
172         unsigned long objs[NR_ZS_STAT_TYPE];
173 };
174
175 #ifdef CONFIG_ZSMALLOC_STAT
176 static struct dentry *zs_stat_root;
177 #endif
178
179 #ifdef CONFIG_COMPACTION
180 static struct vfsmount *zsmalloc_mnt;
181 #endif
182
183 /*
184  * We assign a page to ZS_ALMOST_EMPTY fullness group when:
185  *      n <= N / f, where
186  * n = number of allocated objects
187  * N = total number of objects zspage can store
188  * f = fullness_threshold_frac
189  *
190  * Similarly, we assign zspage to:
191  *      ZS_ALMOST_FULL  when n > N / f
192  *      ZS_EMPTY        when n == 0
193  *      ZS_FULL         when n == N
194  *
195  * (see: fix_fullness_group())
196  */
197 static const int fullness_threshold_frac = 4;
198 static size_t huge_class_size;
199
200 struct size_class {
201         spinlock_t lock;
202         struct list_head fullness_list[NR_ZS_FULLNESS];
203         /*
204          * Size of objects stored in this class. Must be multiple
205          * of ZS_ALIGN.
206          */
207         int size;
208         int objs_per_zspage;
209         /* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
210         int pages_per_zspage;
211
212         unsigned int index;
213         struct zs_size_stat stats;
214 };
215
216 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
217 static void SetPageHugeObject(struct page *page)
218 {
219         SetPageOwnerPriv1(page);
220 }
221
222 static void ClearPageHugeObject(struct page *page)
223 {
224         ClearPageOwnerPriv1(page);
225 }
226
227 static int PageHugeObject(struct page *page)
228 {
229         return PageOwnerPriv1(page);
230 }
231
232 /*
233  * Placed within free objects to form a singly linked list.
234  * For every zspage, zspage->freeobj gives head of this list.
235  *
236  * This must be power of 2 and less than or equal to ZS_ALIGN
237  */
238 struct link_free {
239         union {
240                 /*
241                  * Free object index;
242                  * It's valid for non-allocated object
243                  */
244                 unsigned long next;
245                 /*
246                  * Handle of allocated object.
247                  */
248                 unsigned long handle;
249         };
250 };
251
252 struct zs_pool {
253         const char *name;
254
255         struct size_class *size_class[ZS_SIZE_CLASSES];
256         struct kmem_cache *handle_cachep;
257         struct kmem_cache *zspage_cachep;
258
259         atomic_long_t pages_allocated;
260
261         struct zs_pool_stats stats;
262
263         /* Compact classes */
264         struct shrinker shrinker;
265
266 #ifdef CONFIG_ZSMALLOC_STAT
267         struct dentry *stat_dentry;
268 #endif
269 #ifdef CONFIG_COMPACTION
270         struct inode *inode;
271         struct work_struct free_work;
272         /* A wait queue for when migration races with async_free_zspage() */
273         struct wait_queue_head migration_wait;
274         atomic_long_t isolated_pages;
275         bool destroying;
276 #endif
277 };
278
279 struct zspage {
280         struct {
281                 unsigned int fullness:FULLNESS_BITS;
282                 unsigned int class:CLASS_BITS + 1;
283                 unsigned int isolated:ISOLATED_BITS;
284                 unsigned int magic:MAGIC_VAL_BITS;
285         };
286         unsigned int inuse;
287         unsigned int freeobj;
288         struct page *first_page;
289         struct list_head list; /* fullness list */
290 #ifdef CONFIG_COMPACTION
291         rwlock_t lock;
292 #endif
293 };
294
295 struct mapping_area {
296 #ifdef CONFIG_PGTABLE_MAPPING
297         struct vm_struct *vm; /* vm area for mapping object that span pages */
298 #else
299         char *vm_buf; /* copy buffer for objects that span pages */
300 #endif
301         char *vm_addr; /* address of kmap_atomic()'ed pages */
302         enum zs_mapmode vm_mm; /* mapping mode */
303 };
304
305 #ifdef CONFIG_COMPACTION
306 static int zs_register_migration(struct zs_pool *pool);
307 static void zs_unregister_migration(struct zs_pool *pool);
308 static void migrate_lock_init(struct zspage *zspage);
309 static void migrate_read_lock(struct zspage *zspage);
310 static void migrate_read_unlock(struct zspage *zspage);
311 static void kick_deferred_free(struct zs_pool *pool);
312 static void init_deferred_free(struct zs_pool *pool);
313 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
314 #else
315 static int zsmalloc_mount(void) { return 0; }
316 static void zsmalloc_unmount(void) {}
317 static int zs_register_migration(struct zs_pool *pool) { return 0; }
318 static void zs_unregister_migration(struct zs_pool *pool) {}
319 static void migrate_lock_init(struct zspage *zspage) {}
320 static void migrate_read_lock(struct zspage *zspage) {}
321 static void migrate_read_unlock(struct zspage *zspage) {}
322 static void kick_deferred_free(struct zs_pool *pool) {}
323 static void init_deferred_free(struct zs_pool *pool) {}
324 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
325 #endif
326
327 static int create_cache(struct zs_pool *pool)
328 {
329         pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
330                                         0, 0, NULL);
331         if (!pool->handle_cachep)
332                 return 1;
333
334         pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
335                                         0, 0, NULL);
336         if (!pool->zspage_cachep) {
337                 kmem_cache_destroy(pool->handle_cachep);
338                 pool->handle_cachep = NULL;
339                 return 1;
340         }
341
342         return 0;
343 }
344
345 static void destroy_cache(struct zs_pool *pool)
346 {
347         kmem_cache_destroy(pool->handle_cachep);
348         kmem_cache_destroy(pool->zspage_cachep);
349 }
350
351 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
352 {
353         return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
354                         gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
355 }
356
357 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
358 {
359         kmem_cache_free(pool->handle_cachep, (void *)handle);
360 }
361
362 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
363 {
364         return kmem_cache_alloc(pool->zspage_cachep,
365                         flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
366 }
367
368 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
369 {
370         kmem_cache_free(pool->zspage_cachep, zspage);
371 }
372
373 static void record_obj(unsigned long handle, unsigned long obj)
374 {
375         /*
376          * lsb of @obj represents handle lock while other bits
377          * represent object value the handle is pointing so
378          * updating shouldn't do store tearing.
379          */
380         WRITE_ONCE(*(unsigned long *)handle, obj);
381 }
382
383 /* zpool driver */
384
385 #ifdef CONFIG_ZPOOL
386
387 static void *zs_zpool_create(const char *name, gfp_t gfp,
388                              const struct zpool_ops *zpool_ops,
389                              struct zpool *zpool)
390 {
391         /*
392          * Ignore global gfp flags: zs_malloc() may be invoked from
393          * different contexts and its caller must provide a valid
394          * gfp mask.
395          */
396         return zs_create_pool(name);
397 }
398
399 static void zs_zpool_destroy(void *pool)
400 {
401         zs_destroy_pool(pool);
402 }
403
404 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
405                         unsigned long *handle)
406 {
407         *handle = zs_malloc(pool, size, gfp);
408         return *handle ? 0 : -1;
409 }
410 static void zs_zpool_free(void *pool, unsigned long handle)
411 {
412         zs_free(pool, handle);
413 }
414
415 static void *zs_zpool_map(void *pool, unsigned long handle,
416                         enum zpool_mapmode mm)
417 {
418         enum zs_mapmode zs_mm;
419
420         switch (mm) {
421         case ZPOOL_MM_RO:
422                 zs_mm = ZS_MM_RO;
423                 break;
424         case ZPOOL_MM_WO:
425                 zs_mm = ZS_MM_WO;
426                 break;
427         case ZPOOL_MM_RW: /* fall through */
428         default:
429                 zs_mm = ZS_MM_RW;
430                 break;
431         }
432
433         return zs_map_object(pool, handle, zs_mm);
434 }
435 static void zs_zpool_unmap(void *pool, unsigned long handle)
436 {
437         zs_unmap_object(pool, handle);
438 }
439
440 static u64 zs_zpool_total_size(void *pool)
441 {
442         return zs_get_total_pages(pool) << PAGE_SHIFT;
443 }
444
445 static struct zpool_driver zs_zpool_driver = {
446         .type =                   "zsmalloc",
447         .owner =                  THIS_MODULE,
448         .create =                 zs_zpool_create,
449         .destroy =                zs_zpool_destroy,
450         .malloc_support_movable = true,
451         .malloc =                 zs_zpool_malloc,
452         .free =                   zs_zpool_free,
453         .map =                    zs_zpool_map,
454         .unmap =                  zs_zpool_unmap,
455         .total_size =             zs_zpool_total_size,
456 };
457
458 MODULE_ALIAS("zpool-zsmalloc");
459 #endif /* CONFIG_ZPOOL */
460
461 /* per-cpu VM mapping areas for zspage accesses that cross page boundaries */
462 static DEFINE_PER_CPU(struct mapping_area, zs_map_area);
463
464 static bool is_zspage_isolated(struct zspage *zspage)
465 {
466         return zspage->isolated;
467 }
468
469 static __maybe_unused int is_first_page(struct page *page)
470 {
471         return PagePrivate(page);
472 }
473
474 /* Protected by class->lock */
475 static inline int get_zspage_inuse(struct zspage *zspage)
476 {
477         return zspage->inuse;
478 }
479
480
481 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
482 {
483         zspage->inuse += val;
484 }
485
486 static inline struct page *get_first_page(struct zspage *zspage)
487 {
488         struct page *first_page = zspage->first_page;
489
490         VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
491         return first_page;
492 }
493
494 static inline int get_first_obj_offset(struct page *page)
495 {
496         return page->units;
497 }
498
499 static inline void set_first_obj_offset(struct page *page, int offset)
500 {
501         page->units = offset;
502 }
503
504 static inline unsigned int get_freeobj(struct zspage *zspage)
505 {
506         return zspage->freeobj;
507 }
508
509 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
510 {
511         zspage->freeobj = obj;
512 }
513
514 static void get_zspage_mapping(struct zspage *zspage,
515                                 unsigned int *class_idx,
516                                 enum fullness_group *fullness)
517 {
518         BUG_ON(zspage->magic != ZSPAGE_MAGIC);
519
520         *fullness = zspage->fullness;
521         *class_idx = zspage->class;
522 }
523
524 static void set_zspage_mapping(struct zspage *zspage,
525                                 unsigned int class_idx,
526                                 enum fullness_group fullness)
527 {
528         zspage->class = class_idx;
529         zspage->fullness = fullness;
530 }
531
532 /*
533  * zsmalloc divides the pool into various size classes where each
534  * class maintains a list of zspages where each zspage is divided
535  * into equal sized chunks. Each allocation falls into one of these
536  * classes depending on its size. This function returns index of the
537  * size class which has chunk size big enough to hold the give size.
538  */
539 static int get_size_class_index(int size)
540 {
541         int idx = 0;
542
543         if (likely(size > ZS_MIN_ALLOC_SIZE))
544                 idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
545                                 ZS_SIZE_CLASS_DELTA);
546
547         return min_t(int, ZS_SIZE_CLASSES - 1, idx);
548 }
549
550 /* type can be of enum type zs_stat_type or fullness_group */
551 static inline void zs_stat_inc(struct size_class *class,
552                                 int type, unsigned long cnt)
553 {
554         class->stats.objs[type] += cnt;
555 }
556
557 /* type can be of enum type zs_stat_type or fullness_group */
558 static inline void zs_stat_dec(struct size_class *class,
559                                 int type, unsigned long cnt)
560 {
561         class->stats.objs[type] -= cnt;
562 }
563
564 /* type can be of enum type zs_stat_type or fullness_group */
565 static inline unsigned long zs_stat_get(struct size_class *class,
566                                 int type)
567 {
568         return class->stats.objs[type];
569 }
570
571 #ifdef CONFIG_ZSMALLOC_STAT
572
573 static void __init zs_stat_init(void)
574 {
575         if (!debugfs_initialized()) {
576                 pr_warn("debugfs not available, stat dir not created\n");
577                 return;
578         }
579
580         zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
581 }
582
583 static void __exit zs_stat_exit(void)
584 {
585         debugfs_remove_recursive(zs_stat_root);
586 }
587
588 static unsigned long zs_can_compact(struct size_class *class);
589
590 static int zs_stats_size_show(struct seq_file *s, void *v)
591 {
592         int i;
593         struct zs_pool *pool = s->private;
594         struct size_class *class;
595         int objs_per_zspage;
596         unsigned long class_almost_full, class_almost_empty;
597         unsigned long obj_allocated, obj_used, pages_used, freeable;
598         unsigned long total_class_almost_full = 0, total_class_almost_empty = 0;
599         unsigned long total_objs = 0, total_used_objs = 0, total_pages = 0;
600         unsigned long total_freeable = 0;
601
602         seq_printf(s, " %5s %5s %11s %12s %13s %10s %10s %16s %8s\n",
603                         "class", "size", "almost_full", "almost_empty",
604                         "obj_allocated", "obj_used", "pages_used",
605                         "pages_per_zspage", "freeable");
606
607         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
608                 class = pool->size_class[i];
609
610                 if (class->index != i)
611                         continue;
612
613                 spin_lock(&class->lock);
614                 class_almost_full = zs_stat_get(class, CLASS_ALMOST_FULL);
615                 class_almost_empty = zs_stat_get(class, CLASS_ALMOST_EMPTY);
616                 obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
617                 obj_used = zs_stat_get(class, OBJ_USED);
618                 freeable = zs_can_compact(class);
619                 spin_unlock(&class->lock);
620
621                 objs_per_zspage = class->objs_per_zspage;
622                 pages_used = obj_allocated / objs_per_zspage *
623                                 class->pages_per_zspage;
624
625                 seq_printf(s, " %5u %5u %11lu %12lu %13lu"
626                                 " %10lu %10lu %16d %8lu\n",
627                         i, class->size, class_almost_full, class_almost_empty,
628                         obj_allocated, obj_used, pages_used,
629                         class->pages_per_zspage, freeable);
630
631                 total_class_almost_full += class_almost_full;
632                 total_class_almost_empty += class_almost_empty;
633                 total_objs += obj_allocated;
634                 total_used_objs += obj_used;
635                 total_pages += pages_used;
636                 total_freeable += freeable;
637         }
638
639         seq_puts(s, "\n");
640         seq_printf(s, " %5s %5s %11lu %12lu %13lu %10lu %10lu %16s %8lu\n",
641                         "Total", "", total_class_almost_full,
642                         total_class_almost_empty, total_objs,
643                         total_used_objs, total_pages, "", total_freeable);
644
645         return 0;
646 }
647 DEFINE_SHOW_ATTRIBUTE(zs_stats_size);
648
649 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
650 {
651         if (!zs_stat_root) {
652                 pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
653                 return;
654         }
655
656         pool->stat_dentry = debugfs_create_dir(name, zs_stat_root);
657
658         debugfs_create_file("classes", S_IFREG | 0444, pool->stat_dentry, pool,
659                             &zs_stats_size_fops);
660 }
661
662 static void zs_pool_stat_destroy(struct zs_pool *pool)
663 {
664         debugfs_remove_recursive(pool->stat_dentry);
665 }
666
667 #else /* CONFIG_ZSMALLOC_STAT */
668 static void __init zs_stat_init(void)
669 {
670 }
671
672 static void __exit zs_stat_exit(void)
673 {
674 }
675
676 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
677 {
678 }
679
680 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
681 {
682 }
683 #endif
684
685
686 /*
687  * For each size class, zspages are divided into different groups
688  * depending on how "full" they are. This was done so that we could
689  * easily find empty or nearly empty zspages when we try to shrink
690  * the pool (not yet implemented). This function returns fullness
691  * status of the given page.
692  */
693 static enum fullness_group get_fullness_group(struct size_class *class,
694                                                 struct zspage *zspage)
695 {
696         int inuse, objs_per_zspage;
697         enum fullness_group fg;
698
699         inuse = get_zspage_inuse(zspage);
700         objs_per_zspage = class->objs_per_zspage;
701
702         if (inuse == 0)
703                 fg = ZS_EMPTY;
704         else if (inuse == objs_per_zspage)
705                 fg = ZS_FULL;
706         else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
707                 fg = ZS_ALMOST_EMPTY;
708         else
709                 fg = ZS_ALMOST_FULL;
710
711         return fg;
712 }
713
714 /*
715  * Each size class maintains various freelists and zspages are assigned
716  * to one of these freelists based on the number of live objects they
717  * have. This functions inserts the given zspage into the freelist
718  * identified by <class, fullness_group>.
719  */
720 static void insert_zspage(struct size_class *class,
721                                 struct zspage *zspage,
722                                 enum fullness_group fullness)
723 {
724         struct zspage *head;
725
726         zs_stat_inc(class, fullness, 1);
727         head = list_first_entry_or_null(&class->fullness_list[fullness],
728                                         struct zspage, list);
729         /*
730          * We want to see more ZS_FULL pages and less almost empty/full.
731          * Put pages with higher ->inuse first.
732          */
733         if (head) {
734                 if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
735                         list_add(&zspage->list, &head->list);
736                         return;
737                 }
738         }
739         list_add(&zspage->list, &class->fullness_list[fullness]);
740 }
741
742 /*
743  * This function removes the given zspage from the freelist identified
744  * by <class, fullness_group>.
745  */
746 static void remove_zspage(struct size_class *class,
747                                 struct zspage *zspage,
748                                 enum fullness_group fullness)
749 {
750         VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
751         VM_BUG_ON(is_zspage_isolated(zspage));
752
753         list_del_init(&zspage->list);
754         zs_stat_dec(class, fullness, 1);
755 }
756
757 /*
758  * Each size class maintains zspages in different fullness groups depending
759  * on the number of live objects they contain. When allocating or freeing
760  * objects, the fullness status of the page can change, say, from ALMOST_FULL
761  * to ALMOST_EMPTY when freeing an object. This function checks if such
762  * a status change has occurred for the given page and accordingly moves the
763  * page from the freelist of the old fullness group to that of the new
764  * fullness group.
765  */
766 static enum fullness_group fix_fullness_group(struct size_class *class,
767                                                 struct zspage *zspage)
768 {
769         int class_idx;
770         enum fullness_group currfg, newfg;
771
772         get_zspage_mapping(zspage, &class_idx, &currfg);
773         newfg = get_fullness_group(class, zspage);
774         if (newfg == currfg)
775                 goto out;
776
777         if (!is_zspage_isolated(zspage)) {
778                 remove_zspage(class, zspage, currfg);
779                 insert_zspage(class, zspage, newfg);
780         }
781
782         set_zspage_mapping(zspage, class_idx, newfg);
783
784 out:
785         return newfg;
786 }
787
788 /*
789  * We have to decide on how many pages to link together
790  * to form a zspage for each size class. This is important
791  * to reduce wastage due to unusable space left at end of
792  * each zspage which is given as:
793  *     wastage = Zp % class_size
794  *     usage = Zp - wastage
795  * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
796  *
797  * For example, for size class of 3/8 * PAGE_SIZE, we should
798  * link together 3 PAGE_SIZE sized pages to form a zspage
799  * since then we can perfectly fit in 8 such objects.
800  */
801 static int get_pages_per_zspage(int class_size)
802 {
803         int i, max_usedpc = 0;
804         /* zspage order which gives maximum used size per KB */
805         int max_usedpc_order = 1;
806
807         for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
808                 int zspage_size;
809                 int waste, usedpc;
810
811                 zspage_size = i * PAGE_SIZE;
812                 waste = zspage_size % class_size;
813                 usedpc = (zspage_size - waste) * 100 / zspage_size;
814
815                 if (usedpc > max_usedpc) {
816                         max_usedpc = usedpc;
817                         max_usedpc_order = i;
818                 }
819         }
820
821         return max_usedpc_order;
822 }
823
824 static struct zspage *get_zspage(struct page *page)
825 {
826         struct zspage *zspage = (struct zspage *)page->private;
827
828         BUG_ON(zspage->magic != ZSPAGE_MAGIC);
829         return zspage;
830 }
831
832 static struct page *get_next_page(struct page *page)
833 {
834         if (unlikely(PageHugeObject(page)))
835                 return NULL;
836
837         return page->freelist;
838 }
839
840 /**
841  * obj_to_location - get (<page>, <obj_idx>) from encoded object value
842  * @obj: the encoded object value
843  * @page: page object resides in zspage
844  * @obj_idx: object index
845  */
846 static void obj_to_location(unsigned long obj, struct page **page,
847                                 unsigned int *obj_idx)
848 {
849         obj >>= OBJ_TAG_BITS;
850         *page = pfn_to_page(obj >> OBJ_INDEX_BITS);
851         *obj_idx = (obj & OBJ_INDEX_MASK);
852 }
853
854 /**
855  * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
856  * @page: page object resides in zspage
857  * @obj_idx: object index
858  */
859 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
860 {
861         unsigned long obj;
862
863         obj = page_to_pfn(page) << OBJ_INDEX_BITS;
864         obj |= obj_idx & OBJ_INDEX_MASK;
865         obj <<= OBJ_TAG_BITS;
866
867         return obj;
868 }
869
870 static unsigned long handle_to_obj(unsigned long handle)
871 {
872         return *(unsigned long *)handle;
873 }
874
875 static unsigned long obj_to_head(struct page *page, void *obj)
876 {
877         if (unlikely(PageHugeObject(page))) {
878                 VM_BUG_ON_PAGE(!is_first_page(page), page);
879                 return page->index;
880         } else
881                 return *(unsigned long *)obj;
882 }
883
884 static inline int testpin_tag(unsigned long handle)
885 {
886         return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
887 }
888
889 static inline int trypin_tag(unsigned long handle)
890 {
891         return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
892 }
893
894 static void pin_tag(unsigned long handle)
895 {
896         bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
897 }
898
899 static void unpin_tag(unsigned long handle)
900 {
901         bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
902 }
903
904 static void reset_page(struct page *page)
905 {
906         __ClearPageMovable(page);
907         ClearPagePrivate(page);
908         set_page_private(page, 0);
909         page_mapcount_reset(page);
910         ClearPageHugeObject(page);
911         page->freelist = NULL;
912 }
913
914 static int trylock_zspage(struct zspage *zspage)
915 {
916         struct page *cursor, *fail;
917
918         for (cursor = get_first_page(zspage); cursor != NULL; cursor =
919                                         get_next_page(cursor)) {
920                 if (!trylock_page(cursor)) {
921                         fail = cursor;
922                         goto unlock;
923                 }
924         }
925
926         return 1;
927 unlock:
928         for (cursor = get_first_page(zspage); cursor != fail; cursor =
929                                         get_next_page(cursor))
930                 unlock_page(cursor);
931
932         return 0;
933 }
934
935 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
936                                 struct zspage *zspage)
937 {
938         struct page *page, *next;
939         enum fullness_group fg;
940         unsigned int class_idx;
941
942         get_zspage_mapping(zspage, &class_idx, &fg);
943
944         assert_spin_locked(&class->lock);
945
946         VM_BUG_ON(get_zspage_inuse(zspage));
947         VM_BUG_ON(fg != ZS_EMPTY);
948
949         next = page = get_first_page(zspage);
950         do {
951                 VM_BUG_ON_PAGE(!PageLocked(page), page);
952                 next = get_next_page(page);
953                 reset_page(page);
954                 unlock_page(page);
955                 dec_zone_page_state(page, NR_ZSPAGES);
956                 put_page(page);
957                 page = next;
958         } while (page != NULL);
959
960         cache_free_zspage(pool, zspage);
961
962         zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
963         atomic_long_sub(class->pages_per_zspage,
964                                         &pool->pages_allocated);
965 }
966
967 static void free_zspage(struct zs_pool *pool, struct size_class *class,
968                                 struct zspage *zspage)
969 {
970         VM_BUG_ON(get_zspage_inuse(zspage));
971         VM_BUG_ON(list_empty(&zspage->list));
972
973         if (!trylock_zspage(zspage)) {
974                 kick_deferred_free(pool);
975                 return;
976         }
977
978         remove_zspage(class, zspage, ZS_EMPTY);
979         __free_zspage(pool, class, zspage);
980 }
981
982 /* Initialize a newly allocated zspage */
983 static void init_zspage(struct size_class *class, struct zspage *zspage)
984 {
985         unsigned int freeobj = 1;
986         unsigned long off = 0;
987         struct page *page = get_first_page(zspage);
988
989         while (page) {
990                 struct page *next_page;
991                 struct link_free *link;
992                 void *vaddr;
993
994                 set_first_obj_offset(page, off);
995
996                 vaddr = kmap_atomic(page);
997                 link = (struct link_free *)vaddr + off / sizeof(*link);
998
999                 while ((off += class->size) < PAGE_SIZE) {
1000                         link->next = freeobj++ << OBJ_TAG_BITS;
1001                         link += class->size / sizeof(*link);
1002                 }
1003
1004                 /*
1005                  * We now come to the last (full or partial) object on this
1006                  * page, which must point to the first object on the next
1007                  * page (if present)
1008                  */
1009                 next_page = get_next_page(page);
1010                 if (next_page) {
1011                         link->next = freeobj++ << OBJ_TAG_BITS;
1012                 } else {
1013                         /*
1014                          * Reset OBJ_TAG_BITS bit to last link to tell
1015                          * whether it's allocated object or not.
1016                          */
1017                         link->next = -1UL << OBJ_TAG_BITS;
1018                 }
1019                 kunmap_atomic(vaddr);
1020                 page = next_page;
1021                 off %= PAGE_SIZE;
1022         }
1023
1024         set_freeobj(zspage, 0);
1025 }
1026
1027 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1028                                 struct page *pages[])
1029 {
1030         int i;
1031         struct page *page;
1032         struct page *prev_page = NULL;
1033         int nr_pages = class->pages_per_zspage;
1034
1035         /*
1036          * Allocate individual pages and link them together as:
1037          * 1. all pages are linked together using page->freelist
1038          * 2. each sub-page point to zspage using page->private
1039          *
1040          * we set PG_private to identify the first page (i.e. no other sub-page
1041          * has this flag set).
1042          */
1043         for (i = 0; i < nr_pages; i++) {
1044                 page = pages[i];
1045                 set_page_private(page, (unsigned long)zspage);
1046                 page->freelist = NULL;
1047                 if (i == 0) {
1048                         zspage->first_page = page;
1049                         SetPagePrivate(page);
1050                         if (unlikely(class->objs_per_zspage == 1 &&
1051                                         class->pages_per_zspage == 1))
1052                                 SetPageHugeObject(page);
1053                 } else {
1054                         prev_page->freelist = page;
1055                 }
1056                 prev_page = page;
1057         }
1058 }
1059
1060 /*
1061  * Allocate a zspage for the given size class
1062  */
1063 static struct zspage *alloc_zspage(struct zs_pool *pool,
1064                                         struct size_class *class,
1065                                         gfp_t gfp)
1066 {
1067         int i;
1068         struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1069         struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1070
1071         if (!zspage)
1072                 return NULL;
1073
1074         memset(zspage, 0, sizeof(struct zspage));
1075         zspage->magic = ZSPAGE_MAGIC;
1076         migrate_lock_init(zspage);
1077
1078         for (i = 0; i < class->pages_per_zspage; i++) {
1079                 struct page *page;
1080
1081                 page = alloc_page(gfp);
1082                 if (!page) {
1083                         while (--i >= 0) {
1084                                 dec_zone_page_state(pages[i], NR_ZSPAGES);
1085                                 __free_page(pages[i]);
1086                         }
1087                         cache_free_zspage(pool, zspage);
1088                         return NULL;
1089                 }
1090
1091                 inc_zone_page_state(page, NR_ZSPAGES);
1092                 pages[i] = page;
1093         }
1094
1095         create_page_chain(class, zspage, pages);
1096         init_zspage(class, zspage);
1097
1098         return zspage;
1099 }
1100
1101 static struct zspage *find_get_zspage(struct size_class *class)
1102 {
1103         int i;
1104         struct zspage *zspage;
1105
1106         for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1107                 zspage = list_first_entry_or_null(&class->fullness_list[i],
1108                                 struct zspage, list);
1109                 if (zspage)
1110                         break;
1111         }
1112
1113         return zspage;
1114 }
1115
1116 #ifdef CONFIG_PGTABLE_MAPPING
1117 static inline int __zs_cpu_up(struct mapping_area *area)
1118 {
1119         /*
1120          * Make sure we don't leak memory if a cpu UP notification
1121          * and zs_init() race and both call zs_cpu_up() on the same cpu
1122          */
1123         if (area->vm)
1124                 return 0;
1125         area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1126         if (!area->vm)
1127                 return -ENOMEM;
1128         return 0;
1129 }
1130
1131 static inline void __zs_cpu_down(struct mapping_area *area)
1132 {
1133         if (area->vm)
1134                 free_vm_area(area->vm);
1135         area->vm = NULL;
1136 }
1137
1138 static inline void *__zs_map_object(struct mapping_area *area,
1139                                 struct page *pages[2], int off, int size)
1140 {
1141         BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1142         area->vm_addr = area->vm->addr;
1143         return area->vm_addr + off;
1144 }
1145
1146 static inline void __zs_unmap_object(struct mapping_area *area,
1147                                 struct page *pages[2], int off, int size)
1148 {
1149         unsigned long addr = (unsigned long)area->vm_addr;
1150
1151         unmap_kernel_range(addr, PAGE_SIZE * 2);
1152 }
1153
1154 #else /* CONFIG_PGTABLE_MAPPING */
1155
1156 static inline int __zs_cpu_up(struct mapping_area *area)
1157 {
1158         /*
1159          * Make sure we don't leak memory if a cpu UP notification
1160          * and zs_init() race and both call zs_cpu_up() on the same cpu
1161          */
1162         if (area->vm_buf)
1163                 return 0;
1164         area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1165         if (!area->vm_buf)
1166                 return -ENOMEM;
1167         return 0;
1168 }
1169
1170 static inline void __zs_cpu_down(struct mapping_area *area)
1171 {
1172         kfree(area->vm_buf);
1173         area->vm_buf = NULL;
1174 }
1175
1176 static void *__zs_map_object(struct mapping_area *area,
1177                         struct page *pages[2], int off, int size)
1178 {
1179         int sizes[2];
1180         void *addr;
1181         char *buf = area->vm_buf;
1182
1183         /* disable page faults to match kmap_atomic() return conditions */
1184         pagefault_disable();
1185
1186         /* no read fastpath */
1187         if (area->vm_mm == ZS_MM_WO)
1188                 goto out;
1189
1190         sizes[0] = PAGE_SIZE - off;
1191         sizes[1] = size - sizes[0];
1192
1193         /* copy object to per-cpu buffer */
1194         addr = kmap_atomic(pages[0]);
1195         memcpy(buf, addr + off, sizes[0]);
1196         kunmap_atomic(addr);
1197         addr = kmap_atomic(pages[1]);
1198         memcpy(buf + sizes[0], addr, sizes[1]);
1199         kunmap_atomic(addr);
1200 out:
1201         return area->vm_buf;
1202 }
1203
1204 static void __zs_unmap_object(struct mapping_area *area,
1205                         struct page *pages[2], int off, int size)
1206 {
1207         int sizes[2];
1208         void *addr;
1209         char *buf;
1210
1211         /* no write fastpath */
1212         if (area->vm_mm == ZS_MM_RO)
1213                 goto out;
1214
1215         buf = area->vm_buf;
1216         buf = buf + ZS_HANDLE_SIZE;
1217         size -= ZS_HANDLE_SIZE;
1218         off += ZS_HANDLE_SIZE;
1219
1220         sizes[0] = PAGE_SIZE - off;
1221         sizes[1] = size - sizes[0];
1222
1223         /* copy per-cpu buffer to object */
1224         addr = kmap_atomic(pages[0]);
1225         memcpy(addr + off, buf, sizes[0]);
1226         kunmap_atomic(addr);
1227         addr = kmap_atomic(pages[1]);
1228         memcpy(addr, buf + sizes[0], sizes[1]);
1229         kunmap_atomic(addr);
1230
1231 out:
1232         /* enable page faults to match kunmap_atomic() return conditions */
1233         pagefault_enable();
1234 }
1235
1236 #endif /* CONFIG_PGTABLE_MAPPING */
1237
1238 static int zs_cpu_prepare(unsigned int cpu)
1239 {
1240         struct mapping_area *area;
1241
1242         area = &per_cpu(zs_map_area, cpu);
1243         return __zs_cpu_up(area);
1244 }
1245
1246 static int zs_cpu_dead(unsigned int cpu)
1247 {
1248         struct mapping_area *area;
1249
1250         area = &per_cpu(zs_map_area, cpu);
1251         __zs_cpu_down(area);
1252         return 0;
1253 }
1254
1255 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1256                                         int objs_per_zspage)
1257 {
1258         if (prev->pages_per_zspage == pages_per_zspage &&
1259                 prev->objs_per_zspage == objs_per_zspage)
1260                 return true;
1261
1262         return false;
1263 }
1264
1265 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1266 {
1267         return get_zspage_inuse(zspage) == class->objs_per_zspage;
1268 }
1269
1270 unsigned long zs_get_total_pages(struct zs_pool *pool)
1271 {
1272         return atomic_long_read(&pool->pages_allocated);
1273 }
1274 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1275
1276 /**
1277  * zs_map_object - get address of allocated object from handle.
1278  * @pool: pool from which the object was allocated
1279  * @handle: handle returned from zs_malloc
1280  * @mm: maping mode to use
1281  *
1282  * Before using an object allocated from zs_malloc, it must be mapped using
1283  * this function. When done with the object, it must be unmapped using
1284  * zs_unmap_object.
1285  *
1286  * Only one object can be mapped per cpu at a time. There is no protection
1287  * against nested mappings.
1288  *
1289  * This function returns with preemption and page faults disabled.
1290  */
1291 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1292                         enum zs_mapmode mm)
1293 {
1294         struct zspage *zspage;
1295         struct page *page;
1296         unsigned long obj, off;
1297         unsigned int obj_idx;
1298
1299         unsigned int class_idx;
1300         enum fullness_group fg;
1301         struct size_class *class;
1302         struct mapping_area *area;
1303         struct page *pages[2];
1304         void *ret;
1305
1306         /*
1307          * Because we use per-cpu mapping areas shared among the
1308          * pools/users, we can't allow mapping in interrupt context
1309          * because it can corrupt another users mappings.
1310          */
1311         BUG_ON(in_interrupt());
1312
1313         /* From now on, migration cannot move the object */
1314         pin_tag(handle);
1315
1316         obj = handle_to_obj(handle);
1317         obj_to_location(obj, &page, &obj_idx);
1318         zspage = get_zspage(page);
1319
1320         /* migration cannot move any subpage in this zspage */
1321         migrate_read_lock(zspage);
1322
1323         get_zspage_mapping(zspage, &class_idx, &fg);
1324         class = pool->size_class[class_idx];
1325         off = (class->size * obj_idx) & ~PAGE_MASK;
1326
1327         area = &get_cpu_var(zs_map_area);
1328         area->vm_mm = mm;
1329         if (off + class->size <= PAGE_SIZE) {
1330                 /* this object is contained entirely within a page */
1331                 area->vm_addr = kmap_atomic(page);
1332                 ret = area->vm_addr + off;
1333                 goto out;
1334         }
1335
1336         /* this object spans two pages */
1337         pages[0] = page;
1338         pages[1] = get_next_page(page);
1339         BUG_ON(!pages[1]);
1340
1341         ret = __zs_map_object(area, pages, off, class->size);
1342 out:
1343         if (likely(!PageHugeObject(page)))
1344                 ret += ZS_HANDLE_SIZE;
1345
1346         return ret;
1347 }
1348 EXPORT_SYMBOL_GPL(zs_map_object);
1349
1350 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1351 {
1352         struct zspage *zspage;
1353         struct page *page;
1354         unsigned long obj, off;
1355         unsigned int obj_idx;
1356
1357         unsigned int class_idx;
1358         enum fullness_group fg;
1359         struct size_class *class;
1360         struct mapping_area *area;
1361
1362         obj = handle_to_obj(handle);
1363         obj_to_location(obj, &page, &obj_idx);
1364         zspage = get_zspage(page);
1365         get_zspage_mapping(zspage, &class_idx, &fg);
1366         class = pool->size_class[class_idx];
1367         off = (class->size * obj_idx) & ~PAGE_MASK;
1368
1369         area = this_cpu_ptr(&zs_map_area);
1370         if (off + class->size <= PAGE_SIZE)
1371                 kunmap_atomic(area->vm_addr);
1372         else {
1373                 struct page *pages[2];
1374
1375                 pages[0] = page;
1376                 pages[1] = get_next_page(page);
1377                 BUG_ON(!pages[1]);
1378
1379                 __zs_unmap_object(area, pages, off, class->size);
1380         }
1381         put_cpu_var(zs_map_area);
1382
1383         migrate_read_unlock(zspage);
1384         unpin_tag(handle);
1385 }
1386 EXPORT_SYMBOL_GPL(zs_unmap_object);
1387
1388 /**
1389  * zs_huge_class_size() - Returns the size (in bytes) of the first huge
1390  *                        zsmalloc &size_class.
1391  * @pool: zsmalloc pool to use
1392  *
1393  * The function returns the size of the first huge class - any object of equal
1394  * or bigger size will be stored in zspage consisting of a single physical
1395  * page.
1396  *
1397  * Context: Any context.
1398  *
1399  * Return: the size (in bytes) of the first huge zsmalloc &size_class.
1400  */
1401 size_t zs_huge_class_size(struct zs_pool *pool)
1402 {
1403         return huge_class_size;
1404 }
1405 EXPORT_SYMBOL_GPL(zs_huge_class_size);
1406
1407 static unsigned long obj_malloc(struct size_class *class,
1408                                 struct zspage *zspage, unsigned long handle)
1409 {
1410         int i, nr_page, offset;
1411         unsigned long obj;
1412         struct link_free *link;
1413
1414         struct page *m_page;
1415         unsigned long m_offset;
1416         void *vaddr;
1417
1418         handle |= OBJ_ALLOCATED_TAG;
1419         obj = get_freeobj(zspage);
1420
1421         offset = obj * class->size;
1422         nr_page = offset >> PAGE_SHIFT;
1423         m_offset = offset & ~PAGE_MASK;
1424         m_page = get_first_page(zspage);
1425
1426         for (i = 0; i < nr_page; i++)
1427                 m_page = get_next_page(m_page);
1428
1429         vaddr = kmap_atomic(m_page);
1430         link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1431         set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1432         if (likely(!PageHugeObject(m_page)))
1433                 /* record handle in the header of allocated chunk */
1434                 link->handle = handle;
1435         else
1436                 /* record handle to page->index */
1437                 zspage->first_page->index = handle;
1438
1439         kunmap_atomic(vaddr);
1440         mod_zspage_inuse(zspage, 1);
1441         zs_stat_inc(class, OBJ_USED, 1);
1442
1443         obj = location_to_obj(m_page, obj);
1444
1445         return obj;
1446 }
1447
1448
1449 /**
1450  * zs_malloc - Allocate block of given size from pool.
1451  * @pool: pool to allocate from
1452  * @size: size of block to allocate
1453  * @gfp: gfp flags when allocating object
1454  *
1455  * On success, handle to the allocated object is returned,
1456  * otherwise 0.
1457  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1458  */
1459 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1460 {
1461         unsigned long handle, obj;
1462         struct size_class *class;
1463         enum fullness_group newfg;
1464         struct zspage *zspage;
1465
1466         if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1467                 return 0;
1468
1469         handle = cache_alloc_handle(pool, gfp);
1470         if (!handle)
1471                 return 0;
1472
1473         /* extra space in chunk to keep the handle */
1474         size += ZS_HANDLE_SIZE;
1475         class = pool->size_class[get_size_class_index(size)];
1476
1477         spin_lock(&class->lock);
1478         zspage = find_get_zspage(class);
1479         if (likely(zspage)) {
1480                 obj = obj_malloc(class, zspage, handle);
1481                 /* Now move the zspage to another fullness group, if required */
1482                 fix_fullness_group(class, zspage);
1483                 record_obj(handle, obj);
1484                 spin_unlock(&class->lock);
1485
1486                 return handle;
1487         }
1488
1489         spin_unlock(&class->lock);
1490
1491         zspage = alloc_zspage(pool, class, gfp);
1492         if (!zspage) {
1493                 cache_free_handle(pool, handle);
1494                 return 0;
1495         }
1496
1497         spin_lock(&class->lock);
1498         obj = obj_malloc(class, zspage, handle);
1499         newfg = get_fullness_group(class, zspage);
1500         insert_zspage(class, zspage, newfg);
1501         set_zspage_mapping(zspage, class->index, newfg);
1502         record_obj(handle, obj);
1503         atomic_long_add(class->pages_per_zspage,
1504                                 &pool->pages_allocated);
1505         zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1506
1507         /* We completely set up zspage so mark them as movable */
1508         SetZsPageMovable(pool, zspage);
1509         spin_unlock(&class->lock);
1510
1511         return handle;
1512 }
1513 EXPORT_SYMBOL_GPL(zs_malloc);
1514
1515 static void obj_free(struct size_class *class, unsigned long obj)
1516 {
1517         struct link_free *link;
1518         struct zspage *zspage;
1519         struct page *f_page;
1520         unsigned long f_offset;
1521         unsigned int f_objidx;
1522         void *vaddr;
1523
1524         obj &= ~OBJ_ALLOCATED_TAG;
1525         obj_to_location(obj, &f_page, &f_objidx);
1526         f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1527         zspage = get_zspage(f_page);
1528
1529         vaddr = kmap_atomic(f_page);
1530
1531         /* Insert this object in containing zspage's freelist */
1532         link = (struct link_free *)(vaddr + f_offset);
1533         link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1534         kunmap_atomic(vaddr);
1535         set_freeobj(zspage, f_objidx);
1536         mod_zspage_inuse(zspage, -1);
1537         zs_stat_dec(class, OBJ_USED, 1);
1538 }
1539
1540 void zs_free(struct zs_pool *pool, unsigned long handle)
1541 {
1542         struct zspage *zspage;
1543         struct page *f_page;
1544         unsigned long obj;
1545         unsigned int f_objidx;
1546         int class_idx;
1547         struct size_class *class;
1548         enum fullness_group fullness;
1549         bool isolated;
1550
1551         if (unlikely(!handle))
1552                 return;
1553
1554         pin_tag(handle);
1555         obj = handle_to_obj(handle);
1556         obj_to_location(obj, &f_page, &f_objidx);
1557         zspage = get_zspage(f_page);
1558
1559         migrate_read_lock(zspage);
1560
1561         get_zspage_mapping(zspage, &class_idx, &fullness);
1562         class = pool->size_class[class_idx];
1563
1564         spin_lock(&class->lock);
1565         obj_free(class, obj);
1566         fullness = fix_fullness_group(class, zspage);
1567         if (fullness != ZS_EMPTY) {
1568                 migrate_read_unlock(zspage);
1569                 goto out;
1570         }
1571
1572         isolated = is_zspage_isolated(zspage);
1573         migrate_read_unlock(zspage);
1574         /* If zspage is isolated, zs_page_putback will free the zspage */
1575         if (likely(!isolated))
1576                 free_zspage(pool, class, zspage);
1577 out:
1578
1579         spin_unlock(&class->lock);
1580         unpin_tag(handle);
1581         cache_free_handle(pool, handle);
1582 }
1583 EXPORT_SYMBOL_GPL(zs_free);
1584
1585 static void zs_object_copy(struct size_class *class, unsigned long dst,
1586                                 unsigned long src)
1587 {
1588         struct page *s_page, *d_page;
1589         unsigned int s_objidx, d_objidx;
1590         unsigned long s_off, d_off;
1591         void *s_addr, *d_addr;
1592         int s_size, d_size, size;
1593         int written = 0;
1594
1595         s_size = d_size = class->size;
1596
1597         obj_to_location(src, &s_page, &s_objidx);
1598         obj_to_location(dst, &d_page, &d_objidx);
1599
1600         s_off = (class->size * s_objidx) & ~PAGE_MASK;
1601         d_off = (class->size * d_objidx) & ~PAGE_MASK;
1602
1603         if (s_off + class->size > PAGE_SIZE)
1604                 s_size = PAGE_SIZE - s_off;
1605
1606         if (d_off + class->size > PAGE_SIZE)
1607                 d_size = PAGE_SIZE - d_off;
1608
1609         s_addr = kmap_atomic(s_page);
1610         d_addr = kmap_atomic(d_page);
1611
1612         while (1) {
1613                 size = min(s_size, d_size);
1614                 memcpy(d_addr + d_off, s_addr + s_off, size);
1615                 written += size;
1616
1617                 if (written == class->size)
1618                         break;
1619
1620                 s_off += size;
1621                 s_size -= size;
1622                 d_off += size;
1623                 d_size -= size;
1624
1625                 if (s_off >= PAGE_SIZE) {
1626                         kunmap_atomic(d_addr);
1627                         kunmap_atomic(s_addr);
1628                         s_page = get_next_page(s_page);
1629                         s_addr = kmap_atomic(s_page);
1630                         d_addr = kmap_atomic(d_page);
1631                         s_size = class->size - written;
1632                         s_off = 0;
1633                 }
1634
1635                 if (d_off >= PAGE_SIZE) {
1636                         kunmap_atomic(d_addr);
1637                         d_page = get_next_page(d_page);
1638                         d_addr = kmap_atomic(d_page);
1639                         d_size = class->size - written;
1640                         d_off = 0;
1641                 }
1642         }
1643
1644         kunmap_atomic(d_addr);
1645         kunmap_atomic(s_addr);
1646 }
1647
1648 /*
1649  * Find alloced object in zspage from index object and
1650  * return handle.
1651  */
1652 static unsigned long find_alloced_obj(struct size_class *class,
1653                                         struct page *page, int *obj_idx)
1654 {
1655         unsigned long head;
1656         int offset = 0;
1657         int index = *obj_idx;
1658         unsigned long handle = 0;
1659         void *addr = kmap_atomic(page);
1660
1661         offset = get_first_obj_offset(page);
1662         offset += class->size * index;
1663
1664         while (offset < PAGE_SIZE) {
1665                 head = obj_to_head(page, addr + offset);
1666                 if (head & OBJ_ALLOCATED_TAG) {
1667                         handle = head & ~OBJ_ALLOCATED_TAG;
1668                         if (trypin_tag(handle))
1669                                 break;
1670                         handle = 0;
1671                 }
1672
1673                 offset += class->size;
1674                 index++;
1675         }
1676
1677         kunmap_atomic(addr);
1678
1679         *obj_idx = index;
1680
1681         return handle;
1682 }
1683
1684 struct zs_compact_control {
1685         /* Source spage for migration which could be a subpage of zspage */
1686         struct page *s_page;
1687         /* Destination page for migration which should be a first page
1688          * of zspage. */
1689         struct page *d_page;
1690          /* Starting object index within @s_page which used for live object
1691           * in the subpage. */
1692         int obj_idx;
1693 };
1694
1695 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1696                                 struct zs_compact_control *cc)
1697 {
1698         unsigned long used_obj, free_obj;
1699         unsigned long handle;
1700         struct page *s_page = cc->s_page;
1701         struct page *d_page = cc->d_page;
1702         int obj_idx = cc->obj_idx;
1703         int ret = 0;
1704
1705         while (1) {
1706                 handle = find_alloced_obj(class, s_page, &obj_idx);
1707                 if (!handle) {
1708                         s_page = get_next_page(s_page);
1709                         if (!s_page)
1710                                 break;
1711                         obj_idx = 0;
1712                         continue;
1713                 }
1714
1715                 /* Stop if there is no more space */
1716                 if (zspage_full(class, get_zspage(d_page))) {
1717                         unpin_tag(handle);
1718                         ret = -ENOMEM;
1719                         break;
1720                 }
1721
1722                 used_obj = handle_to_obj(handle);
1723                 free_obj = obj_malloc(class, get_zspage(d_page), handle);
1724                 zs_object_copy(class, free_obj, used_obj);
1725                 obj_idx++;
1726                 /*
1727                  * record_obj updates handle's value to free_obj and it will
1728                  * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1729                  * breaks synchronization using pin_tag(e,g, zs_free) so
1730                  * let's keep the lock bit.
1731                  */
1732                 free_obj |= BIT(HANDLE_PIN_BIT);
1733                 record_obj(handle, free_obj);
1734                 unpin_tag(handle);
1735                 obj_free(class, used_obj);
1736         }
1737
1738         /* Remember last position in this iteration */
1739         cc->s_page = s_page;
1740         cc->obj_idx = obj_idx;
1741
1742         return ret;
1743 }
1744
1745 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1746 {
1747         int i;
1748         struct zspage *zspage;
1749         enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1750
1751         if (!source) {
1752                 fg[0] = ZS_ALMOST_FULL;
1753                 fg[1] = ZS_ALMOST_EMPTY;
1754         }
1755
1756         for (i = 0; i < 2; i++) {
1757                 zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1758                                                         struct zspage, list);
1759                 if (zspage) {
1760                         VM_BUG_ON(is_zspage_isolated(zspage));
1761                         remove_zspage(class, zspage, fg[i]);
1762                         return zspage;
1763                 }
1764         }
1765
1766         return zspage;
1767 }
1768
1769 /*
1770  * putback_zspage - add @zspage into right class's fullness list
1771  * @class: destination class
1772  * @zspage: target page
1773  *
1774  * Return @zspage's fullness_group
1775  */
1776 static enum fullness_group putback_zspage(struct size_class *class,
1777                         struct zspage *zspage)
1778 {
1779         enum fullness_group fullness;
1780
1781         VM_BUG_ON(is_zspage_isolated(zspage));
1782
1783         fullness = get_fullness_group(class, zspage);
1784         insert_zspage(class, zspage, fullness);
1785         set_zspage_mapping(zspage, class->index, fullness);
1786
1787         return fullness;
1788 }
1789
1790 #ifdef CONFIG_COMPACTION
1791 /*
1792  * To prevent zspage destroy during migration, zspage freeing should
1793  * hold locks of all pages in the zspage.
1794  */
1795 static void lock_zspage(struct zspage *zspage)
1796 {
1797         struct page *page = get_first_page(zspage);
1798
1799         do {
1800                 lock_page(page);
1801         } while ((page = get_next_page(page)) != NULL);
1802 }
1803
1804 static int zs_init_fs_context(struct fs_context *fc)
1805 {
1806         return init_pseudo(fc, ZSMALLOC_MAGIC) ? 0 : -ENOMEM;
1807 }
1808
1809 static struct file_system_type zsmalloc_fs = {
1810         .name           = "zsmalloc",
1811         .init_fs_context = zs_init_fs_context,
1812         .kill_sb        = kill_anon_super,
1813 };
1814
1815 static int zsmalloc_mount(void)
1816 {
1817         int ret = 0;
1818
1819         zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1820         if (IS_ERR(zsmalloc_mnt))
1821                 ret = PTR_ERR(zsmalloc_mnt);
1822
1823         return ret;
1824 }
1825
1826 static void zsmalloc_unmount(void)
1827 {
1828         kern_unmount(zsmalloc_mnt);
1829 }
1830
1831 static void migrate_lock_init(struct zspage *zspage)
1832 {
1833         rwlock_init(&zspage->lock);
1834 }
1835
1836 static void migrate_read_lock(struct zspage *zspage)
1837 {
1838         read_lock(&zspage->lock);
1839 }
1840
1841 static void migrate_read_unlock(struct zspage *zspage)
1842 {
1843         read_unlock(&zspage->lock);
1844 }
1845
1846 static void migrate_write_lock(struct zspage *zspage)
1847 {
1848         write_lock(&zspage->lock);
1849 }
1850
1851 static void migrate_write_unlock(struct zspage *zspage)
1852 {
1853         write_unlock(&zspage->lock);
1854 }
1855
1856 /* Number of isolated subpage for *page migration* in this zspage */
1857 static void inc_zspage_isolation(struct zspage *zspage)
1858 {
1859         zspage->isolated++;
1860 }
1861
1862 static void dec_zspage_isolation(struct zspage *zspage)
1863 {
1864         zspage->isolated--;
1865 }
1866
1867 static void putback_zspage_deferred(struct zs_pool *pool,
1868                                     struct size_class *class,
1869                                     struct zspage *zspage)
1870 {
1871         enum fullness_group fg;
1872
1873         fg = putback_zspage(class, zspage);
1874         if (fg == ZS_EMPTY)
1875                 schedule_work(&pool->free_work);
1876
1877 }
1878
1879 static inline void zs_pool_dec_isolated(struct zs_pool *pool)
1880 {
1881         VM_BUG_ON(atomic_long_read(&pool->isolated_pages) <= 0);
1882         atomic_long_dec(&pool->isolated_pages);
1883         /*
1884          * There's no possibility of racing, since wait_for_isolated_drain()
1885          * checks the isolated count under &class->lock after enqueuing
1886          * on migration_wait.
1887          */
1888         if (atomic_long_read(&pool->isolated_pages) == 0 && pool->destroying)
1889                 wake_up_all(&pool->migration_wait);
1890 }
1891
1892 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1893                                 struct page *newpage, struct page *oldpage)
1894 {
1895         struct page *page;
1896         struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1897         int idx = 0;
1898
1899         page = get_first_page(zspage);
1900         do {
1901                 if (page == oldpage)
1902                         pages[idx] = newpage;
1903                 else
1904                         pages[idx] = page;
1905                 idx++;
1906         } while ((page = get_next_page(page)) != NULL);
1907
1908         create_page_chain(class, zspage, pages);
1909         set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1910         if (unlikely(PageHugeObject(oldpage)))
1911                 newpage->index = oldpage->index;
1912         __SetPageMovable(newpage, page_mapping(oldpage));
1913 }
1914
1915 static bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1916 {
1917         struct zs_pool *pool;
1918         struct size_class *class;
1919         int class_idx;
1920         enum fullness_group fullness;
1921         struct zspage *zspage;
1922         struct address_space *mapping;
1923
1924         /*
1925          * Page is locked so zspage couldn't be destroyed. For detail, look at
1926          * lock_zspage in free_zspage.
1927          */
1928         VM_BUG_ON_PAGE(!PageMovable(page), page);
1929         VM_BUG_ON_PAGE(PageIsolated(page), page);
1930
1931         zspage = get_zspage(page);
1932
1933         /*
1934          * Without class lock, fullness could be stale while class_idx is okay
1935          * because class_idx is constant unless page is freed so we should get
1936          * fullness again under class lock.
1937          */
1938         get_zspage_mapping(zspage, &class_idx, &fullness);
1939         mapping = page_mapping(page);
1940         pool = mapping->private_data;
1941         class = pool->size_class[class_idx];
1942
1943         spin_lock(&class->lock);
1944         if (get_zspage_inuse(zspage) == 0) {
1945                 spin_unlock(&class->lock);
1946                 return false;
1947         }
1948
1949         /* zspage is isolated for object migration */
1950         if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1951                 spin_unlock(&class->lock);
1952                 return false;
1953         }
1954
1955         /*
1956          * If this is first time isolation for the zspage, isolate zspage from
1957          * size_class to prevent further object allocation from the zspage.
1958          */
1959         if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1960                 get_zspage_mapping(zspage, &class_idx, &fullness);
1961                 atomic_long_inc(&pool->isolated_pages);
1962                 remove_zspage(class, zspage, fullness);
1963         }
1964
1965         inc_zspage_isolation(zspage);
1966         spin_unlock(&class->lock);
1967
1968         return true;
1969 }
1970
1971 static int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1972                 struct page *page, enum migrate_mode mode)
1973 {
1974         struct zs_pool *pool;
1975         struct size_class *class;
1976         int class_idx;
1977         enum fullness_group fullness;
1978         struct zspage *zspage;
1979         struct page *dummy;
1980         void *s_addr, *d_addr, *addr;
1981         int offset, pos;
1982         unsigned long handle, head;
1983         unsigned long old_obj, new_obj;
1984         unsigned int obj_idx;
1985         int ret = -EAGAIN;
1986
1987         /*
1988          * We cannot support the _NO_COPY case here, because copy needs to
1989          * happen under the zs lock, which does not work with
1990          * MIGRATE_SYNC_NO_COPY workflow.
1991          */
1992         if (mode == MIGRATE_SYNC_NO_COPY)
1993                 return -EINVAL;
1994
1995         VM_BUG_ON_PAGE(!PageMovable(page), page);
1996         VM_BUG_ON_PAGE(!PageIsolated(page), page);
1997
1998         zspage = get_zspage(page);
1999
2000         /* Concurrent compactor cannot migrate any subpage in zspage */
2001         migrate_write_lock(zspage);
2002         get_zspage_mapping(zspage, &class_idx, &fullness);
2003         pool = mapping->private_data;
2004         class = pool->size_class[class_idx];
2005         offset = get_first_obj_offset(page);
2006
2007         spin_lock(&class->lock);
2008         if (!get_zspage_inuse(zspage)) {
2009                 /*
2010                  * Set "offset" to end of the page so that every loops
2011                  * skips unnecessary object scanning.
2012                  */
2013                 offset = PAGE_SIZE;
2014         }
2015
2016         pos = offset;
2017         s_addr = kmap_atomic(page);
2018         while (pos < PAGE_SIZE) {
2019                 head = obj_to_head(page, s_addr + pos);
2020                 if (head & OBJ_ALLOCATED_TAG) {
2021                         handle = head & ~OBJ_ALLOCATED_TAG;
2022                         if (!trypin_tag(handle))
2023                                 goto unpin_objects;
2024                 }
2025                 pos += class->size;
2026         }
2027
2028         /*
2029          * Here, any user cannot access all objects in the zspage so let's move.
2030          */
2031         d_addr = kmap_atomic(newpage);
2032         memcpy(d_addr, s_addr, PAGE_SIZE);
2033         kunmap_atomic(d_addr);
2034
2035         for (addr = s_addr + offset; addr < s_addr + pos;
2036                                         addr += class->size) {
2037                 head = obj_to_head(page, addr);
2038                 if (head & OBJ_ALLOCATED_TAG) {
2039                         handle = head & ~OBJ_ALLOCATED_TAG;
2040                         if (!testpin_tag(handle))
2041                                 BUG();
2042
2043                         old_obj = handle_to_obj(handle);
2044                         obj_to_location(old_obj, &dummy, &obj_idx);
2045                         new_obj = (unsigned long)location_to_obj(newpage,
2046                                                                 obj_idx);
2047                         new_obj |= BIT(HANDLE_PIN_BIT);
2048                         record_obj(handle, new_obj);
2049                 }
2050         }
2051
2052         replace_sub_page(class, zspage, newpage, page);
2053         get_page(newpage);
2054
2055         dec_zspage_isolation(zspage);
2056
2057         /*
2058          * Page migration is done so let's putback isolated zspage to
2059          * the list if @page is final isolated subpage in the zspage.
2060          */
2061         if (!is_zspage_isolated(zspage)) {
2062                 /*
2063                  * We cannot race with zs_destroy_pool() here because we wait
2064                  * for isolation to hit zero before we start destroying.
2065                  * Also, we ensure that everyone can see pool->destroying before
2066                  * we start waiting.
2067                  */
2068                 putback_zspage_deferred(pool, class, zspage);
2069                 zs_pool_dec_isolated(pool);
2070         }
2071
2072         if (page_zone(newpage) != page_zone(page)) {
2073                 dec_zone_page_state(page, NR_ZSPAGES);
2074                 inc_zone_page_state(newpage, NR_ZSPAGES);
2075         }
2076
2077         reset_page(page);
2078         put_page(page);
2079         page = newpage;
2080
2081         ret = MIGRATEPAGE_SUCCESS;
2082 unpin_objects:
2083         for (addr = s_addr + offset; addr < s_addr + pos;
2084                                                 addr += class->size) {
2085                 head = obj_to_head(page, addr);
2086                 if (head & OBJ_ALLOCATED_TAG) {
2087                         handle = head & ~OBJ_ALLOCATED_TAG;
2088                         if (!testpin_tag(handle))
2089                                 BUG();
2090                         unpin_tag(handle);
2091                 }
2092         }
2093         kunmap_atomic(s_addr);
2094         spin_unlock(&class->lock);
2095         migrate_write_unlock(zspage);
2096
2097         return ret;
2098 }
2099
2100 static void zs_page_putback(struct page *page)
2101 {
2102         struct zs_pool *pool;
2103         struct size_class *class;
2104         int class_idx;
2105         enum fullness_group fg;
2106         struct address_space *mapping;
2107         struct zspage *zspage;
2108
2109         VM_BUG_ON_PAGE(!PageMovable(page), page);
2110         VM_BUG_ON_PAGE(!PageIsolated(page), page);
2111
2112         zspage = get_zspage(page);
2113         get_zspage_mapping(zspage, &class_idx, &fg);
2114         mapping = page_mapping(page);
2115         pool = mapping->private_data;
2116         class = pool->size_class[class_idx];
2117
2118         spin_lock(&class->lock);
2119         dec_zspage_isolation(zspage);
2120         if (!is_zspage_isolated(zspage)) {
2121                 /*
2122                  * Due to page_lock, we cannot free zspage immediately
2123                  * so let's defer.
2124                  */
2125                 putback_zspage_deferred(pool, class, zspage);
2126                 zs_pool_dec_isolated(pool);
2127         }
2128         spin_unlock(&class->lock);
2129 }
2130
2131 static const struct address_space_operations zsmalloc_aops = {
2132         .isolate_page = zs_page_isolate,
2133         .migratepage = zs_page_migrate,
2134         .putback_page = zs_page_putback,
2135 };
2136
2137 static int zs_register_migration(struct zs_pool *pool)
2138 {
2139         pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2140         if (IS_ERR(pool->inode)) {
2141                 pool->inode = NULL;
2142                 return 1;
2143         }
2144
2145         pool->inode->i_mapping->private_data = pool;
2146         pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2147         return 0;
2148 }
2149
2150 static bool pool_isolated_are_drained(struct zs_pool *pool)
2151 {
2152         return atomic_long_read(&pool->isolated_pages) == 0;
2153 }
2154
2155 /* Function for resolving migration */
2156 static void wait_for_isolated_drain(struct zs_pool *pool)
2157 {
2158
2159         /*
2160          * We're in the process of destroying the pool, so there are no
2161          * active allocations. zs_page_isolate() fails for completely free
2162          * zspages, so we need only wait for the zs_pool's isolated
2163          * count to hit zero.
2164          */
2165         wait_event(pool->migration_wait,
2166                    pool_isolated_are_drained(pool));
2167 }
2168
2169 static void zs_unregister_migration(struct zs_pool *pool)
2170 {
2171         pool->destroying = true;
2172         /*
2173          * We need a memory barrier here to ensure global visibility of
2174          * pool->destroying. Thus pool->isolated pages will either be 0 in which
2175          * case we don't care, or it will be > 0 and pool->destroying will
2176          * ensure that we wake up once isolation hits 0.
2177          */
2178         smp_mb();
2179         wait_for_isolated_drain(pool); /* This can block */
2180         flush_work(&pool->free_work);
2181         iput(pool->inode);
2182 }
2183
2184 /*
2185  * Caller should hold page_lock of all pages in the zspage
2186  * In here, we cannot use zspage meta data.
2187  */
2188 static void async_free_zspage(struct work_struct *work)
2189 {
2190         int i;
2191         struct size_class *class;
2192         unsigned int class_idx;
2193         enum fullness_group fullness;
2194         struct zspage *zspage, *tmp;
2195         LIST_HEAD(free_pages);
2196         struct zs_pool *pool = container_of(work, struct zs_pool,
2197                                         free_work);
2198
2199         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2200                 class = pool->size_class[i];
2201                 if (class->index != i)
2202                         continue;
2203
2204                 spin_lock(&class->lock);
2205                 list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2206                 spin_unlock(&class->lock);
2207         }
2208
2209
2210         list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2211                 list_del(&zspage->list);
2212                 lock_zspage(zspage);
2213
2214                 get_zspage_mapping(zspage, &class_idx, &fullness);
2215                 VM_BUG_ON(fullness != ZS_EMPTY);
2216                 class = pool->size_class[class_idx];
2217                 spin_lock(&class->lock);
2218                 __free_zspage(pool, pool->size_class[class_idx], zspage);
2219                 spin_unlock(&class->lock);
2220         }
2221 };
2222
2223 static void kick_deferred_free(struct zs_pool *pool)
2224 {
2225         schedule_work(&pool->free_work);
2226 }
2227
2228 static void init_deferred_free(struct zs_pool *pool)
2229 {
2230         INIT_WORK(&pool->free_work, async_free_zspage);
2231 }
2232
2233 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2234 {
2235         struct page *page = get_first_page(zspage);
2236
2237         do {
2238                 WARN_ON(!trylock_page(page));
2239                 __SetPageMovable(page, pool->inode->i_mapping);
2240                 unlock_page(page);
2241         } while ((page = get_next_page(page)) != NULL);
2242 }
2243 #endif
2244
2245 /*
2246  *
2247  * Based on the number of unused allocated objects calculate
2248  * and return the number of pages that we can free.
2249  */
2250 static unsigned long zs_can_compact(struct size_class *class)
2251 {
2252         unsigned long obj_wasted;
2253         unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2254         unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2255
2256         if (obj_allocated <= obj_used)
2257                 return 0;
2258
2259         obj_wasted = obj_allocated - obj_used;
2260         obj_wasted /= class->objs_per_zspage;
2261
2262         return obj_wasted * class->pages_per_zspage;
2263 }
2264
2265 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2266 {
2267         struct zs_compact_control cc;
2268         struct zspage *src_zspage;
2269         struct zspage *dst_zspage = NULL;
2270
2271         spin_lock(&class->lock);
2272         while ((src_zspage = isolate_zspage(class, true))) {
2273
2274                 if (!zs_can_compact(class))
2275                         break;
2276
2277                 cc.obj_idx = 0;
2278                 cc.s_page = get_first_page(src_zspage);
2279
2280                 while ((dst_zspage = isolate_zspage(class, false))) {
2281                         cc.d_page = get_first_page(dst_zspage);
2282                         /*
2283                          * If there is no more space in dst_page, resched
2284                          * and see if anyone had allocated another zspage.
2285                          */
2286                         if (!migrate_zspage(pool, class, &cc))
2287                                 break;
2288
2289                         putback_zspage(class, dst_zspage);
2290                 }
2291
2292                 /* Stop if we couldn't find slot */
2293                 if (dst_zspage == NULL)
2294                         break;
2295
2296                 putback_zspage(class, dst_zspage);
2297                 if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2298                         free_zspage(pool, class, src_zspage);
2299                         pool->stats.pages_compacted += class->pages_per_zspage;
2300                 }
2301                 spin_unlock(&class->lock);
2302                 cond_resched();
2303                 spin_lock(&class->lock);
2304         }
2305
2306         if (src_zspage)
2307                 putback_zspage(class, src_zspage);
2308
2309         spin_unlock(&class->lock);
2310 }
2311
2312 unsigned long zs_compact(struct zs_pool *pool)
2313 {
2314         int i;
2315         struct size_class *class;
2316
2317         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2318                 class = pool->size_class[i];
2319                 if (!class)
2320                         continue;
2321                 if (class->index != i)
2322                         continue;
2323                 __zs_compact(pool, class);
2324         }
2325
2326         return pool->stats.pages_compacted;
2327 }
2328 EXPORT_SYMBOL_GPL(zs_compact);
2329
2330 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2331 {
2332         memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2333 }
2334 EXPORT_SYMBOL_GPL(zs_pool_stats);
2335
2336 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2337                 struct shrink_control *sc)
2338 {
2339         unsigned long pages_freed;
2340         struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2341                         shrinker);
2342
2343         pages_freed = pool->stats.pages_compacted;
2344         /*
2345          * Compact classes and calculate compaction delta.
2346          * Can run concurrently with a manually triggered
2347          * (by user) compaction.
2348          */
2349         pages_freed = zs_compact(pool) - pages_freed;
2350
2351         return pages_freed ? pages_freed : SHRINK_STOP;
2352 }
2353
2354 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2355                 struct shrink_control *sc)
2356 {
2357         int i;
2358         struct size_class *class;
2359         unsigned long pages_to_free = 0;
2360         struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2361                         shrinker);
2362
2363         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2364                 class = pool->size_class[i];
2365                 if (!class)
2366                         continue;
2367                 if (class->index != i)
2368                         continue;
2369
2370                 pages_to_free += zs_can_compact(class);
2371         }
2372
2373         return pages_to_free;
2374 }
2375
2376 static void zs_unregister_shrinker(struct zs_pool *pool)
2377 {
2378         unregister_shrinker(&pool->shrinker);
2379 }
2380
2381 static int zs_register_shrinker(struct zs_pool *pool)
2382 {
2383         pool->shrinker.scan_objects = zs_shrinker_scan;
2384         pool->shrinker.count_objects = zs_shrinker_count;
2385         pool->shrinker.batch = 0;
2386         pool->shrinker.seeks = DEFAULT_SEEKS;
2387
2388         return register_shrinker(&pool->shrinker);
2389 }
2390
2391 /**
2392  * zs_create_pool - Creates an allocation pool to work from.
2393  * @name: pool name to be created
2394  *
2395  * This function must be called before anything when using
2396  * the zsmalloc allocator.
2397  *
2398  * On success, a pointer to the newly created pool is returned,
2399  * otherwise NULL.
2400  */
2401 struct zs_pool *zs_create_pool(const char *name)
2402 {
2403         int i;
2404         struct zs_pool *pool;
2405         struct size_class *prev_class = NULL;
2406
2407         pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2408         if (!pool)
2409                 return NULL;
2410
2411         init_deferred_free(pool);
2412
2413         pool->name = kstrdup(name, GFP_KERNEL);
2414         if (!pool->name)
2415                 goto err;
2416
2417 #ifdef CONFIG_COMPACTION
2418         init_waitqueue_head(&pool->migration_wait);
2419 #endif
2420
2421         if (create_cache(pool))
2422                 goto err;
2423
2424         /*
2425          * Iterate reversely, because, size of size_class that we want to use
2426          * for merging should be larger or equal to current size.
2427          */
2428         for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2429                 int size;
2430                 int pages_per_zspage;
2431                 int objs_per_zspage;
2432                 struct size_class *class;
2433                 int fullness = 0;
2434
2435                 size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2436                 if (size > ZS_MAX_ALLOC_SIZE)
2437                         size = ZS_MAX_ALLOC_SIZE;
2438                 pages_per_zspage = get_pages_per_zspage(size);
2439                 objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2440
2441                 /*
2442                  * We iterate from biggest down to smallest classes,
2443                  * so huge_class_size holds the size of the first huge
2444                  * class. Any object bigger than or equal to that will
2445                  * endup in the huge class.
2446                  */
2447                 if (pages_per_zspage != 1 && objs_per_zspage != 1 &&
2448                                 !huge_class_size) {
2449                         huge_class_size = size;
2450                         /*
2451                          * The object uses ZS_HANDLE_SIZE bytes to store the
2452                          * handle. We need to subtract it, because zs_malloc()
2453                          * unconditionally adds handle size before it performs
2454                          * size class search - so object may be smaller than
2455                          * huge class size, yet it still can end up in the huge
2456                          * class because it grows by ZS_HANDLE_SIZE extra bytes
2457                          * right before class lookup.
2458                          */
2459                         huge_class_size -= (ZS_HANDLE_SIZE - 1);
2460                 }
2461
2462                 /*
2463                  * size_class is used for normal zsmalloc operation such
2464                  * as alloc/free for that size. Although it is natural that we
2465                  * have one size_class for each size, there is a chance that we
2466                  * can get more memory utilization if we use one size_class for
2467                  * many different sizes whose size_class have same
2468                  * characteristics. So, we makes size_class point to
2469                  * previous size_class if possible.
2470                  */
2471                 if (prev_class) {
2472                         if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2473                                 pool->size_class[i] = prev_class;
2474                                 continue;
2475                         }
2476                 }
2477
2478                 class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2479                 if (!class)
2480                         goto err;
2481
2482                 class->size = size;
2483                 class->index = i;
2484                 class->pages_per_zspage = pages_per_zspage;
2485                 class->objs_per_zspage = objs_per_zspage;
2486                 spin_lock_init(&class->lock);
2487                 pool->size_class[i] = class;
2488                 for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2489                                                         fullness++)
2490                         INIT_LIST_HEAD(&class->fullness_list[fullness]);
2491
2492                 prev_class = class;
2493         }
2494
2495         /* debug only, don't abort if it fails */
2496         zs_pool_stat_create(pool, name);
2497
2498         if (zs_register_migration(pool))
2499                 goto err;
2500
2501         /*
2502          * Not critical since shrinker is only used to trigger internal
2503          * defragmentation of the pool which is pretty optional thing.  If
2504          * registration fails we still can use the pool normally and user can
2505          * trigger compaction manually. Thus, ignore return code.
2506          */
2507         zs_register_shrinker(pool);
2508
2509         return pool;
2510
2511 err:
2512         zs_destroy_pool(pool);
2513         return NULL;
2514 }
2515 EXPORT_SYMBOL_GPL(zs_create_pool);
2516
2517 void zs_destroy_pool(struct zs_pool *pool)
2518 {
2519         int i;
2520
2521         zs_unregister_shrinker(pool);
2522         zs_unregister_migration(pool);
2523         zs_pool_stat_destroy(pool);
2524
2525         for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2526                 int fg;
2527                 struct size_class *class = pool->size_class[i];
2528
2529                 if (!class)
2530                         continue;
2531
2532                 if (class->index != i)
2533                         continue;
2534
2535                 for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2536                         if (!list_empty(&class->fullness_list[fg])) {
2537                                 pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2538                                         class->size, fg);
2539                         }
2540                 }
2541                 kfree(class);
2542         }
2543
2544         destroy_cache(pool);
2545         kfree(pool->name);
2546         kfree(pool);
2547 }
2548 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2549
2550 static int __init zs_init(void)
2551 {
2552         int ret;
2553
2554         ret = zsmalloc_mount();
2555         if (ret)
2556                 goto out;
2557
2558         ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2559                                 zs_cpu_prepare, zs_cpu_dead);
2560         if (ret)
2561                 goto hp_setup_fail;
2562
2563 #ifdef CONFIG_ZPOOL
2564         zpool_register_driver(&zs_zpool_driver);
2565 #endif
2566
2567         zs_stat_init();
2568
2569         return 0;
2570
2571 hp_setup_fail:
2572         zsmalloc_unmount();
2573 out:
2574         return ret;
2575 }
2576
2577 static void __exit zs_exit(void)
2578 {
2579 #ifdef CONFIG_ZPOOL
2580         zpool_unregister_driver(&zs_zpool_driver);
2581 #endif
2582         zsmalloc_unmount();
2583         cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2584
2585         zs_stat_exit();
2586 }
2587
2588 module_init(zs_init);
2589 module_exit(zs_exit);
2590
2591 MODULE_LICENSE("Dual BSD/GPL");
2592 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");