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