xref: /openbmc/linux/mm/zsmalloc.c (revision 4da722ca)
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/types.h>
50 #include <linux/debugfs.h>
51 #include <linux/zsmalloc.h>
52 #include <linux/zpool.h>
53 #include <linux/mount.h>
54 #include <linux/migrate.h>
55 #include <linux/pagemap.h>
56 
57 #define ZSPAGE_MAGIC	0x58
58 
59 /*
60  * This must be power of 2 and greater than of equal to sizeof(link_free).
61  * These two conditions ensure that any 'struct link_free' itself doesn't
62  * span more than 1 page which avoids complex case of mapping 2 pages simply
63  * to restore link_free pointer values.
64  */
65 #define ZS_ALIGN		8
66 
67 /*
68  * A single 'zspage' is composed of up to 2^N discontiguous 0-order (single)
69  * pages. ZS_MAX_ZSPAGE_ORDER defines upper limit on N.
70  */
71 #define ZS_MAX_ZSPAGE_ORDER 2
72 #define ZS_MAX_PAGES_PER_ZSPAGE (_AC(1, UL) << ZS_MAX_ZSPAGE_ORDER)
73 
74 #define ZS_HANDLE_SIZE (sizeof(unsigned long))
75 
76 /*
77  * Object location (<PFN>, <obj_idx>) is encoded as
78  * as single (unsigned long) handle value.
79  *
80  * Note that object index <obj_idx> starts from 0.
81  *
82  * This is made more complicated by various memory models and PAE.
83  */
84 
85 #ifndef MAX_PHYSMEM_BITS
86 #ifdef CONFIG_HIGHMEM64G
87 #define MAX_PHYSMEM_BITS 36
88 #else /* !CONFIG_HIGHMEM64G */
89 /*
90  * If this definition of MAX_PHYSMEM_BITS is used, OBJ_INDEX_BITS will just
91  * be PAGE_SHIFT
92  */
93 #define MAX_PHYSMEM_BITS BITS_PER_LONG
94 #endif
95 #endif
96 #define _PFN_BITS		(MAX_PHYSMEM_BITS - PAGE_SHIFT)
97 
98 /*
99  * Memory for allocating for handle keeps object position by
100  * encoding <page, obj_idx> and the encoded value has a room
101  * in least bit(ie, look at obj_to_location).
102  * We use the bit to synchronize between object access by
103  * user and migration.
104  */
105 #define HANDLE_PIN_BIT	0
106 
107 /*
108  * Head in allocated object should have OBJ_ALLOCATED_TAG
109  * to identify the object was allocated or not.
110  * It's okay to add the status bit in the least bit because
111  * header keeps handle which is 4byte-aligned address so we
112  * have room for two bit at least.
113  */
114 #define OBJ_ALLOCATED_TAG 1
115 #define OBJ_TAG_BITS 1
116 #define OBJ_INDEX_BITS	(BITS_PER_LONG - _PFN_BITS - OBJ_TAG_BITS)
117 #define OBJ_INDEX_MASK	((_AC(1, UL) << OBJ_INDEX_BITS) - 1)
118 
119 #define FULLNESS_BITS	2
120 #define CLASS_BITS	8
121 #define ISOLATED_BITS	3
122 #define MAGIC_VAL_BITS	8
123 
124 #define MAX(a, b) ((a) >= (b) ? (a) : (b))
125 /* ZS_MIN_ALLOC_SIZE must be multiple of ZS_ALIGN */
126 #define ZS_MIN_ALLOC_SIZE \
127 	MAX(32, (ZS_MAX_PAGES_PER_ZSPAGE << PAGE_SHIFT >> OBJ_INDEX_BITS))
128 /* each chunk includes extra space to keep handle */
129 #define ZS_MAX_ALLOC_SIZE	PAGE_SIZE
130 
131 /*
132  * On systems with 4K page size, this gives 255 size classes! There is a
133  * trader-off here:
134  *  - Large number of size classes is potentially wasteful as free page are
135  *    spread across these classes
136  *  - Small number of size classes causes large internal fragmentation
137  *  - Probably its better to use specific size classes (empirically
138  *    determined). NOTE: all those class sizes must be set as multiple of
139  *    ZS_ALIGN to make sure link_free itself never has to span 2 pages.
140  *
141  *  ZS_MIN_ALLOC_SIZE and ZS_SIZE_CLASS_DELTA must be multiple of ZS_ALIGN
142  *  (reason above)
143  */
144 #define ZS_SIZE_CLASS_DELTA	(PAGE_SIZE >> CLASS_BITS)
145 #define ZS_SIZE_CLASSES	(DIV_ROUND_UP(ZS_MAX_ALLOC_SIZE - ZS_MIN_ALLOC_SIZE, \
146 				      ZS_SIZE_CLASS_DELTA) + 1)
147 
148 enum fullness_group {
149 	ZS_EMPTY,
150 	ZS_ALMOST_EMPTY,
151 	ZS_ALMOST_FULL,
152 	ZS_FULL,
153 	NR_ZS_FULLNESS,
154 };
155 
156 enum zs_stat_type {
157 	CLASS_EMPTY,
158 	CLASS_ALMOST_EMPTY,
159 	CLASS_ALMOST_FULL,
160 	CLASS_FULL,
161 	OBJ_ALLOCATED,
162 	OBJ_USED,
163 	NR_ZS_STAT_TYPE,
164 };
165 
166 struct zs_size_stat {
167 	unsigned long objs[NR_ZS_STAT_TYPE];
168 };
169 
170 #ifdef CONFIG_ZSMALLOC_STAT
171 static struct dentry *zs_stat_root;
172 #endif
173 
174 #ifdef CONFIG_COMPACTION
175 static struct vfsmount *zsmalloc_mnt;
176 #endif
177 
178 /*
179  * We assign a page to ZS_ALMOST_EMPTY fullness group when:
180  *	n <= N / f, where
181  * n = number of allocated objects
182  * N = total number of objects zspage can store
183  * f = fullness_threshold_frac
184  *
185  * Similarly, we assign zspage to:
186  *	ZS_ALMOST_FULL	when n > N / f
187  *	ZS_EMPTY	when n == 0
188  *	ZS_FULL		when n == N
189  *
190  * (see: fix_fullness_group())
191  */
192 static const int fullness_threshold_frac = 4;
193 
194 struct size_class {
195 	spinlock_t lock;
196 	struct list_head fullness_list[NR_ZS_FULLNESS];
197 	/*
198 	 * Size of objects stored in this class. Must be multiple
199 	 * of ZS_ALIGN.
200 	 */
201 	int size;
202 	int objs_per_zspage;
203 	/* Number of PAGE_SIZE sized pages to combine to form a 'zspage' */
204 	int pages_per_zspage;
205 
206 	unsigned int index;
207 	struct zs_size_stat stats;
208 };
209 
210 /* huge object: pages_per_zspage == 1 && maxobj_per_zspage == 1 */
211 static void SetPageHugeObject(struct page *page)
212 {
213 	SetPageOwnerPriv1(page);
214 }
215 
216 static void ClearPageHugeObject(struct page *page)
217 {
218 	ClearPageOwnerPriv1(page);
219 }
220 
221 static int PageHugeObject(struct page *page)
222 {
223 	return PageOwnerPriv1(page);
224 }
225 
226 /*
227  * Placed within free objects to form a singly linked list.
228  * For every zspage, zspage->freeobj gives head of this list.
229  *
230  * This must be power of 2 and less than or equal to ZS_ALIGN
231  */
232 struct link_free {
233 	union {
234 		/*
235 		 * Free object index;
236 		 * It's valid for non-allocated object
237 		 */
238 		unsigned long next;
239 		/*
240 		 * Handle of allocated object.
241 		 */
242 		unsigned long handle;
243 	};
244 };
245 
246 struct zs_pool {
247 	const char *name;
248 
249 	struct size_class *size_class[ZS_SIZE_CLASSES];
250 	struct kmem_cache *handle_cachep;
251 	struct kmem_cache *zspage_cachep;
252 
253 	atomic_long_t pages_allocated;
254 
255 	struct zs_pool_stats stats;
256 
257 	/* Compact classes */
258 	struct shrinker shrinker;
259 	/*
260 	 * To signify that register_shrinker() was successful
261 	 * and unregister_shrinker() will not Oops.
262 	 */
263 	bool shrinker_enabled;
264 #ifdef CONFIG_ZSMALLOC_STAT
265 	struct dentry *stat_dentry;
266 #endif
267 #ifdef CONFIG_COMPACTION
268 	struct inode *inode;
269 	struct work_struct free_work;
270 #endif
271 };
272 
273 struct zspage {
274 	struct {
275 		unsigned int fullness:FULLNESS_BITS;
276 		unsigned int class:CLASS_BITS + 1;
277 		unsigned int isolated:ISOLATED_BITS;
278 		unsigned int magic:MAGIC_VAL_BITS;
279 	};
280 	unsigned int inuse;
281 	unsigned int freeobj;
282 	struct page *first_page;
283 	struct list_head list; /* fullness list */
284 #ifdef CONFIG_COMPACTION
285 	rwlock_t lock;
286 #endif
287 };
288 
289 struct mapping_area {
290 #ifdef CONFIG_PGTABLE_MAPPING
291 	struct vm_struct *vm; /* vm area for mapping object that span pages */
292 #else
293 	char *vm_buf; /* copy buffer for objects that span pages */
294 #endif
295 	char *vm_addr; /* address of kmap_atomic()'ed pages */
296 	enum zs_mapmode vm_mm; /* mapping mode */
297 };
298 
299 #ifdef CONFIG_COMPACTION
300 static int zs_register_migration(struct zs_pool *pool);
301 static void zs_unregister_migration(struct zs_pool *pool);
302 static void migrate_lock_init(struct zspage *zspage);
303 static void migrate_read_lock(struct zspage *zspage);
304 static void migrate_read_unlock(struct zspage *zspage);
305 static void kick_deferred_free(struct zs_pool *pool);
306 static void init_deferred_free(struct zs_pool *pool);
307 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage);
308 #else
309 static int zsmalloc_mount(void) { return 0; }
310 static void zsmalloc_unmount(void) {}
311 static int zs_register_migration(struct zs_pool *pool) { return 0; }
312 static void zs_unregister_migration(struct zs_pool *pool) {}
313 static void migrate_lock_init(struct zspage *zspage) {}
314 static void migrate_read_lock(struct zspage *zspage) {}
315 static void migrate_read_unlock(struct zspage *zspage) {}
316 static void kick_deferred_free(struct zs_pool *pool) {}
317 static void init_deferred_free(struct zs_pool *pool) {}
318 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage) {}
319 #endif
320 
321 static int create_cache(struct zs_pool *pool)
322 {
323 	pool->handle_cachep = kmem_cache_create("zs_handle", ZS_HANDLE_SIZE,
324 					0, 0, NULL);
325 	if (!pool->handle_cachep)
326 		return 1;
327 
328 	pool->zspage_cachep = kmem_cache_create("zspage", sizeof(struct zspage),
329 					0, 0, NULL);
330 	if (!pool->zspage_cachep) {
331 		kmem_cache_destroy(pool->handle_cachep);
332 		pool->handle_cachep = NULL;
333 		return 1;
334 	}
335 
336 	return 0;
337 }
338 
339 static void destroy_cache(struct zs_pool *pool)
340 {
341 	kmem_cache_destroy(pool->handle_cachep);
342 	kmem_cache_destroy(pool->zspage_cachep);
343 }
344 
345 static unsigned long cache_alloc_handle(struct zs_pool *pool, gfp_t gfp)
346 {
347 	return (unsigned long)kmem_cache_alloc(pool->handle_cachep,
348 			gfp & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
349 }
350 
351 static void cache_free_handle(struct zs_pool *pool, unsigned long handle)
352 {
353 	kmem_cache_free(pool->handle_cachep, (void *)handle);
354 }
355 
356 static struct zspage *cache_alloc_zspage(struct zs_pool *pool, gfp_t flags)
357 {
358 	return kmem_cache_alloc(pool->zspage_cachep,
359 			flags & ~(__GFP_HIGHMEM|__GFP_MOVABLE));
360 }
361 
362 static void cache_free_zspage(struct zs_pool *pool, struct zspage *zspage)
363 {
364 	kmem_cache_free(pool->zspage_cachep, zspage);
365 }
366 
367 static void record_obj(unsigned long handle, unsigned long obj)
368 {
369 	/*
370 	 * lsb of @obj represents handle lock while other bits
371 	 * represent object value the handle is pointing so
372 	 * updating shouldn't do store tearing.
373 	 */
374 	WRITE_ONCE(*(unsigned long *)handle, obj);
375 }
376 
377 /* zpool driver */
378 
379 #ifdef CONFIG_ZPOOL
380 
381 static void *zs_zpool_create(const char *name, gfp_t gfp,
382 			     const struct zpool_ops *zpool_ops,
383 			     struct zpool *zpool)
384 {
385 	/*
386 	 * Ignore global gfp flags: zs_malloc() may be invoked from
387 	 * different contexts and its caller must provide a valid
388 	 * gfp mask.
389 	 */
390 	return zs_create_pool(name);
391 }
392 
393 static void zs_zpool_destroy(void *pool)
394 {
395 	zs_destroy_pool(pool);
396 }
397 
398 static int zs_zpool_malloc(void *pool, size_t size, gfp_t gfp,
399 			unsigned long *handle)
400 {
401 	*handle = zs_malloc(pool, size, gfp);
402 	return *handle ? 0 : -1;
403 }
404 static void zs_zpool_free(void *pool, unsigned long handle)
405 {
406 	zs_free(pool, handle);
407 }
408 
409 static int zs_zpool_shrink(void *pool, unsigned int pages,
410 			unsigned int *reclaimed)
411 {
412 	return -EINVAL;
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: /* fallthru */
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 	.shrink =	zs_zpool_shrink,
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 static inline void set_zspage_inuse(struct zspage *zspage, int val)
481 {
482 	zspage->inuse = val;
483 }
484 
485 static inline void mod_zspage_inuse(struct zspage *zspage, int val)
486 {
487 	zspage->inuse += val;
488 }
489 
490 static inline struct page *get_first_page(struct zspage *zspage)
491 {
492 	struct page *first_page = zspage->first_page;
493 
494 	VM_BUG_ON_PAGE(!is_first_page(first_page), first_page);
495 	return first_page;
496 }
497 
498 static inline int get_first_obj_offset(struct page *page)
499 {
500 	return page->units;
501 }
502 
503 static inline void set_first_obj_offset(struct page *page, int offset)
504 {
505 	page->units = offset;
506 }
507 
508 static inline unsigned int get_freeobj(struct zspage *zspage)
509 {
510 	return zspage->freeobj;
511 }
512 
513 static inline void set_freeobj(struct zspage *zspage, unsigned int obj)
514 {
515 	zspage->freeobj = obj;
516 }
517 
518 static void get_zspage_mapping(struct zspage *zspage,
519 				unsigned int *class_idx,
520 				enum fullness_group *fullness)
521 {
522 	BUG_ON(zspage->magic != ZSPAGE_MAGIC);
523 
524 	*fullness = zspage->fullness;
525 	*class_idx = zspage->class;
526 }
527 
528 static void set_zspage_mapping(struct zspage *zspage,
529 				unsigned int class_idx,
530 				enum fullness_group fullness)
531 {
532 	zspage->class = class_idx;
533 	zspage->fullness = fullness;
534 }
535 
536 /*
537  * zsmalloc divides the pool into various size classes where each
538  * class maintains a list of zspages where each zspage is divided
539  * into equal sized chunks. Each allocation falls into one of these
540  * classes depending on its size. This function returns index of the
541  * size class which has chunk size big enough to hold the give size.
542  */
543 static int get_size_class_index(int size)
544 {
545 	int idx = 0;
546 
547 	if (likely(size > ZS_MIN_ALLOC_SIZE))
548 		idx = DIV_ROUND_UP(size - ZS_MIN_ALLOC_SIZE,
549 				ZS_SIZE_CLASS_DELTA);
550 
551 	return min_t(int, ZS_SIZE_CLASSES - 1, idx);
552 }
553 
554 static inline void zs_stat_inc(struct size_class *class,
555 				enum zs_stat_type type, unsigned long cnt)
556 {
557 	class->stats.objs[type] += cnt;
558 }
559 
560 static inline void zs_stat_dec(struct size_class *class,
561 				enum zs_stat_type type, unsigned long cnt)
562 {
563 	class->stats.objs[type] -= cnt;
564 }
565 
566 static inline unsigned long zs_stat_get(struct size_class *class,
567 				enum zs_stat_type type)
568 {
569 	return class->stats.objs[type];
570 }
571 
572 #ifdef CONFIG_ZSMALLOC_STAT
573 
574 static void __init zs_stat_init(void)
575 {
576 	if (!debugfs_initialized()) {
577 		pr_warn("debugfs not available, stat dir not created\n");
578 		return;
579 	}
580 
581 	zs_stat_root = debugfs_create_dir("zsmalloc", NULL);
582 	if (!zs_stat_root)
583 		pr_warn("debugfs 'zsmalloc' stat dir creation failed\n");
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 
651 static int zs_stats_size_open(struct inode *inode, struct file *file)
652 {
653 	return single_open(file, zs_stats_size_show, inode->i_private);
654 }
655 
656 static const struct file_operations zs_stat_size_ops = {
657 	.open           = zs_stats_size_open,
658 	.read           = seq_read,
659 	.llseek         = seq_lseek,
660 	.release        = single_release,
661 };
662 
663 static void zs_pool_stat_create(struct zs_pool *pool, const char *name)
664 {
665 	struct dentry *entry;
666 
667 	if (!zs_stat_root) {
668 		pr_warn("no root stat dir, not creating <%s> stat dir\n", name);
669 		return;
670 	}
671 
672 	entry = debugfs_create_dir(name, zs_stat_root);
673 	if (!entry) {
674 		pr_warn("debugfs dir <%s> creation failed\n", name);
675 		return;
676 	}
677 	pool->stat_dentry = entry;
678 
679 	entry = debugfs_create_file("classes", S_IFREG | S_IRUGO,
680 			pool->stat_dentry, pool, &zs_stat_size_ops);
681 	if (!entry) {
682 		pr_warn("%s: debugfs file entry <%s> creation failed\n",
683 				name, "classes");
684 		debugfs_remove_recursive(pool->stat_dentry);
685 		pool->stat_dentry = NULL;
686 	}
687 }
688 
689 static void zs_pool_stat_destroy(struct zs_pool *pool)
690 {
691 	debugfs_remove_recursive(pool->stat_dentry);
692 }
693 
694 #else /* CONFIG_ZSMALLOC_STAT */
695 static void __init zs_stat_init(void)
696 {
697 }
698 
699 static void __exit zs_stat_exit(void)
700 {
701 }
702 
703 static inline void zs_pool_stat_create(struct zs_pool *pool, const char *name)
704 {
705 }
706 
707 static inline void zs_pool_stat_destroy(struct zs_pool *pool)
708 {
709 }
710 #endif
711 
712 
713 /*
714  * For each size class, zspages are divided into different groups
715  * depending on how "full" they are. This was done so that we could
716  * easily find empty or nearly empty zspages when we try to shrink
717  * the pool (not yet implemented). This function returns fullness
718  * status of the given page.
719  */
720 static enum fullness_group get_fullness_group(struct size_class *class,
721 						struct zspage *zspage)
722 {
723 	int inuse, objs_per_zspage;
724 	enum fullness_group fg;
725 
726 	inuse = get_zspage_inuse(zspage);
727 	objs_per_zspage = class->objs_per_zspage;
728 
729 	if (inuse == 0)
730 		fg = ZS_EMPTY;
731 	else if (inuse == objs_per_zspage)
732 		fg = ZS_FULL;
733 	else if (inuse <= 3 * objs_per_zspage / fullness_threshold_frac)
734 		fg = ZS_ALMOST_EMPTY;
735 	else
736 		fg = ZS_ALMOST_FULL;
737 
738 	return fg;
739 }
740 
741 /*
742  * Each size class maintains various freelists and zspages are assigned
743  * to one of these freelists based on the number of live objects they
744  * have. This functions inserts the given zspage into the freelist
745  * identified by <class, fullness_group>.
746  */
747 static void insert_zspage(struct size_class *class,
748 				struct zspage *zspage,
749 				enum fullness_group fullness)
750 {
751 	struct zspage *head;
752 
753 	zs_stat_inc(class, fullness, 1);
754 	head = list_first_entry_or_null(&class->fullness_list[fullness],
755 					struct zspage, list);
756 	/*
757 	 * We want to see more ZS_FULL pages and less almost empty/full.
758 	 * Put pages with higher ->inuse first.
759 	 */
760 	if (head) {
761 		if (get_zspage_inuse(zspage) < get_zspage_inuse(head)) {
762 			list_add(&zspage->list, &head->list);
763 			return;
764 		}
765 	}
766 	list_add(&zspage->list, &class->fullness_list[fullness]);
767 }
768 
769 /*
770  * This function removes the given zspage from the freelist identified
771  * by <class, fullness_group>.
772  */
773 static void remove_zspage(struct size_class *class,
774 				struct zspage *zspage,
775 				enum fullness_group fullness)
776 {
777 	VM_BUG_ON(list_empty(&class->fullness_list[fullness]));
778 	VM_BUG_ON(is_zspage_isolated(zspage));
779 
780 	list_del_init(&zspage->list);
781 	zs_stat_dec(class, fullness, 1);
782 }
783 
784 /*
785  * Each size class maintains zspages in different fullness groups depending
786  * on the number of live objects they contain. When allocating or freeing
787  * objects, the fullness status of the page can change, say, from ALMOST_FULL
788  * to ALMOST_EMPTY when freeing an object. This function checks if such
789  * a status change has occurred for the given page and accordingly moves the
790  * page from the freelist of the old fullness group to that of the new
791  * fullness group.
792  */
793 static enum fullness_group fix_fullness_group(struct size_class *class,
794 						struct zspage *zspage)
795 {
796 	int class_idx;
797 	enum fullness_group currfg, newfg;
798 
799 	get_zspage_mapping(zspage, &class_idx, &currfg);
800 	newfg = get_fullness_group(class, zspage);
801 	if (newfg == currfg)
802 		goto out;
803 
804 	if (!is_zspage_isolated(zspage)) {
805 		remove_zspage(class, zspage, currfg);
806 		insert_zspage(class, zspage, newfg);
807 	}
808 
809 	set_zspage_mapping(zspage, class_idx, newfg);
810 
811 out:
812 	return newfg;
813 }
814 
815 /*
816  * We have to decide on how many pages to link together
817  * to form a zspage for each size class. This is important
818  * to reduce wastage due to unusable space left at end of
819  * each zspage which is given as:
820  *     wastage = Zp % class_size
821  *     usage = Zp - wastage
822  * where Zp = zspage size = k * PAGE_SIZE where k = 1, 2, ...
823  *
824  * For example, for size class of 3/8 * PAGE_SIZE, we should
825  * link together 3 PAGE_SIZE sized pages to form a zspage
826  * since then we can perfectly fit in 8 such objects.
827  */
828 static int get_pages_per_zspage(int class_size)
829 {
830 	int i, max_usedpc = 0;
831 	/* zspage order which gives maximum used size per KB */
832 	int max_usedpc_order = 1;
833 
834 	for (i = 1; i <= ZS_MAX_PAGES_PER_ZSPAGE; i++) {
835 		int zspage_size;
836 		int waste, usedpc;
837 
838 		zspage_size = i * PAGE_SIZE;
839 		waste = zspage_size % class_size;
840 		usedpc = (zspage_size - waste) * 100 / zspage_size;
841 
842 		if (usedpc > max_usedpc) {
843 			max_usedpc = usedpc;
844 			max_usedpc_order = i;
845 		}
846 	}
847 
848 	return max_usedpc_order;
849 }
850 
851 static struct zspage *get_zspage(struct page *page)
852 {
853 	struct zspage *zspage = (struct zspage *)page->private;
854 
855 	BUG_ON(zspage->magic != ZSPAGE_MAGIC);
856 	return zspage;
857 }
858 
859 static struct page *get_next_page(struct page *page)
860 {
861 	if (unlikely(PageHugeObject(page)))
862 		return NULL;
863 
864 	return page->freelist;
865 }
866 
867 /**
868  * obj_to_location - get (<page>, <obj_idx>) from encoded object value
869  * @page: page object resides in zspage
870  * @obj_idx: object index
871  */
872 static void obj_to_location(unsigned long obj, struct page **page,
873 				unsigned int *obj_idx)
874 {
875 	obj >>= OBJ_TAG_BITS;
876 	*page = pfn_to_page(obj >> OBJ_INDEX_BITS);
877 	*obj_idx = (obj & OBJ_INDEX_MASK);
878 }
879 
880 /**
881  * location_to_obj - get obj value encoded from (<page>, <obj_idx>)
882  * @page: page object resides in zspage
883  * @obj_idx: object index
884  */
885 static unsigned long location_to_obj(struct page *page, unsigned int obj_idx)
886 {
887 	unsigned long obj;
888 
889 	obj = page_to_pfn(page) << OBJ_INDEX_BITS;
890 	obj |= obj_idx & OBJ_INDEX_MASK;
891 	obj <<= OBJ_TAG_BITS;
892 
893 	return obj;
894 }
895 
896 static unsigned long handle_to_obj(unsigned long handle)
897 {
898 	return *(unsigned long *)handle;
899 }
900 
901 static unsigned long obj_to_head(struct page *page, void *obj)
902 {
903 	if (unlikely(PageHugeObject(page))) {
904 		VM_BUG_ON_PAGE(!is_first_page(page), page);
905 		return page->index;
906 	} else
907 		return *(unsigned long *)obj;
908 }
909 
910 static inline int testpin_tag(unsigned long handle)
911 {
912 	return bit_spin_is_locked(HANDLE_PIN_BIT, (unsigned long *)handle);
913 }
914 
915 static inline int trypin_tag(unsigned long handle)
916 {
917 	return bit_spin_trylock(HANDLE_PIN_BIT, (unsigned long *)handle);
918 }
919 
920 static void pin_tag(unsigned long handle)
921 {
922 	bit_spin_lock(HANDLE_PIN_BIT, (unsigned long *)handle);
923 }
924 
925 static void unpin_tag(unsigned long handle)
926 {
927 	bit_spin_unlock(HANDLE_PIN_BIT, (unsigned long *)handle);
928 }
929 
930 static void reset_page(struct page *page)
931 {
932 	__ClearPageMovable(page);
933 	ClearPagePrivate(page);
934 	set_page_private(page, 0);
935 	page_mapcount_reset(page);
936 	ClearPageHugeObject(page);
937 	page->freelist = NULL;
938 }
939 
940 /*
941  * To prevent zspage destroy during migration, zspage freeing should
942  * hold locks of all pages in the zspage.
943  */
944 void lock_zspage(struct zspage *zspage)
945 {
946 	struct page *page = get_first_page(zspage);
947 
948 	do {
949 		lock_page(page);
950 	} while ((page = get_next_page(page)) != NULL);
951 }
952 
953 int trylock_zspage(struct zspage *zspage)
954 {
955 	struct page *cursor, *fail;
956 
957 	for (cursor = get_first_page(zspage); cursor != NULL; cursor =
958 					get_next_page(cursor)) {
959 		if (!trylock_page(cursor)) {
960 			fail = cursor;
961 			goto unlock;
962 		}
963 	}
964 
965 	return 1;
966 unlock:
967 	for (cursor = get_first_page(zspage); cursor != fail; cursor =
968 					get_next_page(cursor))
969 		unlock_page(cursor);
970 
971 	return 0;
972 }
973 
974 static void __free_zspage(struct zs_pool *pool, struct size_class *class,
975 				struct zspage *zspage)
976 {
977 	struct page *page, *next;
978 	enum fullness_group fg;
979 	unsigned int class_idx;
980 
981 	get_zspage_mapping(zspage, &class_idx, &fg);
982 
983 	assert_spin_locked(&class->lock);
984 
985 	VM_BUG_ON(get_zspage_inuse(zspage));
986 	VM_BUG_ON(fg != ZS_EMPTY);
987 
988 	next = page = get_first_page(zspage);
989 	do {
990 		VM_BUG_ON_PAGE(!PageLocked(page), page);
991 		next = get_next_page(page);
992 		reset_page(page);
993 		unlock_page(page);
994 		dec_zone_page_state(page, NR_ZSPAGES);
995 		put_page(page);
996 		page = next;
997 	} while (page != NULL);
998 
999 	cache_free_zspage(pool, zspage);
1000 
1001 	zs_stat_dec(class, OBJ_ALLOCATED, class->objs_per_zspage);
1002 	atomic_long_sub(class->pages_per_zspage,
1003 					&pool->pages_allocated);
1004 }
1005 
1006 static void free_zspage(struct zs_pool *pool, struct size_class *class,
1007 				struct zspage *zspage)
1008 {
1009 	VM_BUG_ON(get_zspage_inuse(zspage));
1010 	VM_BUG_ON(list_empty(&zspage->list));
1011 
1012 	if (!trylock_zspage(zspage)) {
1013 		kick_deferred_free(pool);
1014 		return;
1015 	}
1016 
1017 	remove_zspage(class, zspage, ZS_EMPTY);
1018 	__free_zspage(pool, class, zspage);
1019 }
1020 
1021 /* Initialize a newly allocated zspage */
1022 static void init_zspage(struct size_class *class, struct zspage *zspage)
1023 {
1024 	unsigned int freeobj = 1;
1025 	unsigned long off = 0;
1026 	struct page *page = get_first_page(zspage);
1027 
1028 	while (page) {
1029 		struct page *next_page;
1030 		struct link_free *link;
1031 		void *vaddr;
1032 
1033 		set_first_obj_offset(page, off);
1034 
1035 		vaddr = kmap_atomic(page);
1036 		link = (struct link_free *)vaddr + off / sizeof(*link);
1037 
1038 		while ((off += class->size) < PAGE_SIZE) {
1039 			link->next = freeobj++ << OBJ_TAG_BITS;
1040 			link += class->size / sizeof(*link);
1041 		}
1042 
1043 		/*
1044 		 * We now come to the last (full or partial) object on this
1045 		 * page, which must point to the first object on the next
1046 		 * page (if present)
1047 		 */
1048 		next_page = get_next_page(page);
1049 		if (next_page) {
1050 			link->next = freeobj++ << OBJ_TAG_BITS;
1051 		} else {
1052 			/*
1053 			 * Reset OBJ_TAG_BITS bit to last link to tell
1054 			 * whether it's allocated object or not.
1055 			 */
1056 			link->next = -1 << OBJ_TAG_BITS;
1057 		}
1058 		kunmap_atomic(vaddr);
1059 		page = next_page;
1060 		off %= PAGE_SIZE;
1061 	}
1062 
1063 	set_freeobj(zspage, 0);
1064 }
1065 
1066 static void create_page_chain(struct size_class *class, struct zspage *zspage,
1067 				struct page *pages[])
1068 {
1069 	int i;
1070 	struct page *page;
1071 	struct page *prev_page = NULL;
1072 	int nr_pages = class->pages_per_zspage;
1073 
1074 	/*
1075 	 * Allocate individual pages and link them together as:
1076 	 * 1. all pages are linked together using page->freelist
1077 	 * 2. each sub-page point to zspage using page->private
1078 	 *
1079 	 * we set PG_private to identify the first page (i.e. no other sub-page
1080 	 * has this flag set).
1081 	 */
1082 	for (i = 0; i < nr_pages; i++) {
1083 		page = pages[i];
1084 		set_page_private(page, (unsigned long)zspage);
1085 		page->freelist = NULL;
1086 		if (i == 0) {
1087 			zspage->first_page = page;
1088 			SetPagePrivate(page);
1089 			if (unlikely(class->objs_per_zspage == 1 &&
1090 					class->pages_per_zspage == 1))
1091 				SetPageHugeObject(page);
1092 		} else {
1093 			prev_page->freelist = page;
1094 		}
1095 		prev_page = page;
1096 	}
1097 }
1098 
1099 /*
1100  * Allocate a zspage for the given size class
1101  */
1102 static struct zspage *alloc_zspage(struct zs_pool *pool,
1103 					struct size_class *class,
1104 					gfp_t gfp)
1105 {
1106 	int i;
1107 	struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE];
1108 	struct zspage *zspage = cache_alloc_zspage(pool, gfp);
1109 
1110 	if (!zspage)
1111 		return NULL;
1112 
1113 	memset(zspage, 0, sizeof(struct zspage));
1114 	zspage->magic = ZSPAGE_MAGIC;
1115 	migrate_lock_init(zspage);
1116 
1117 	for (i = 0; i < class->pages_per_zspage; i++) {
1118 		struct page *page;
1119 
1120 		page = alloc_page(gfp);
1121 		if (!page) {
1122 			while (--i >= 0) {
1123 				dec_zone_page_state(pages[i], NR_ZSPAGES);
1124 				__free_page(pages[i]);
1125 			}
1126 			cache_free_zspage(pool, zspage);
1127 			return NULL;
1128 		}
1129 
1130 		inc_zone_page_state(page, NR_ZSPAGES);
1131 		pages[i] = page;
1132 	}
1133 
1134 	create_page_chain(class, zspage, pages);
1135 	init_zspage(class, zspage);
1136 
1137 	return zspage;
1138 }
1139 
1140 static struct zspage *find_get_zspage(struct size_class *class)
1141 {
1142 	int i;
1143 	struct zspage *zspage;
1144 
1145 	for (i = ZS_ALMOST_FULL; i >= ZS_EMPTY; i--) {
1146 		zspage = list_first_entry_or_null(&class->fullness_list[i],
1147 				struct zspage, list);
1148 		if (zspage)
1149 			break;
1150 	}
1151 
1152 	return zspage;
1153 }
1154 
1155 #ifdef CONFIG_PGTABLE_MAPPING
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)
1163 		return 0;
1164 	area->vm = alloc_vm_area(PAGE_SIZE * 2, NULL);
1165 	if (!area->vm)
1166 		return -ENOMEM;
1167 	return 0;
1168 }
1169 
1170 static inline void __zs_cpu_down(struct mapping_area *area)
1171 {
1172 	if (area->vm)
1173 		free_vm_area(area->vm);
1174 	area->vm = NULL;
1175 }
1176 
1177 static inline void *__zs_map_object(struct mapping_area *area,
1178 				struct page *pages[2], int off, int size)
1179 {
1180 	BUG_ON(map_vm_area(area->vm, PAGE_KERNEL, pages));
1181 	area->vm_addr = area->vm->addr;
1182 	return area->vm_addr + off;
1183 }
1184 
1185 static inline void __zs_unmap_object(struct mapping_area *area,
1186 				struct page *pages[2], int off, int size)
1187 {
1188 	unsigned long addr = (unsigned long)area->vm_addr;
1189 
1190 	unmap_kernel_range(addr, PAGE_SIZE * 2);
1191 }
1192 
1193 #else /* CONFIG_PGTABLE_MAPPING */
1194 
1195 static inline int __zs_cpu_up(struct mapping_area *area)
1196 {
1197 	/*
1198 	 * Make sure we don't leak memory if a cpu UP notification
1199 	 * and zs_init() race and both call zs_cpu_up() on the same cpu
1200 	 */
1201 	if (area->vm_buf)
1202 		return 0;
1203 	area->vm_buf = kmalloc(ZS_MAX_ALLOC_SIZE, GFP_KERNEL);
1204 	if (!area->vm_buf)
1205 		return -ENOMEM;
1206 	return 0;
1207 }
1208 
1209 static inline void __zs_cpu_down(struct mapping_area *area)
1210 {
1211 	kfree(area->vm_buf);
1212 	area->vm_buf = NULL;
1213 }
1214 
1215 static void *__zs_map_object(struct mapping_area *area,
1216 			struct page *pages[2], int off, int size)
1217 {
1218 	int sizes[2];
1219 	void *addr;
1220 	char *buf = area->vm_buf;
1221 
1222 	/* disable page faults to match kmap_atomic() return conditions */
1223 	pagefault_disable();
1224 
1225 	/* no read fastpath */
1226 	if (area->vm_mm == ZS_MM_WO)
1227 		goto out;
1228 
1229 	sizes[0] = PAGE_SIZE - off;
1230 	sizes[1] = size - sizes[0];
1231 
1232 	/* copy object to per-cpu buffer */
1233 	addr = kmap_atomic(pages[0]);
1234 	memcpy(buf, addr + off, sizes[0]);
1235 	kunmap_atomic(addr);
1236 	addr = kmap_atomic(pages[1]);
1237 	memcpy(buf + sizes[0], addr, sizes[1]);
1238 	kunmap_atomic(addr);
1239 out:
1240 	return area->vm_buf;
1241 }
1242 
1243 static void __zs_unmap_object(struct mapping_area *area,
1244 			struct page *pages[2], int off, int size)
1245 {
1246 	int sizes[2];
1247 	void *addr;
1248 	char *buf;
1249 
1250 	/* no write fastpath */
1251 	if (area->vm_mm == ZS_MM_RO)
1252 		goto out;
1253 
1254 	buf = area->vm_buf;
1255 	buf = buf + ZS_HANDLE_SIZE;
1256 	size -= ZS_HANDLE_SIZE;
1257 	off += ZS_HANDLE_SIZE;
1258 
1259 	sizes[0] = PAGE_SIZE - off;
1260 	sizes[1] = size - sizes[0];
1261 
1262 	/* copy per-cpu buffer to object */
1263 	addr = kmap_atomic(pages[0]);
1264 	memcpy(addr + off, buf, sizes[0]);
1265 	kunmap_atomic(addr);
1266 	addr = kmap_atomic(pages[1]);
1267 	memcpy(addr, buf + sizes[0], sizes[1]);
1268 	kunmap_atomic(addr);
1269 
1270 out:
1271 	/* enable page faults to match kunmap_atomic() return conditions */
1272 	pagefault_enable();
1273 }
1274 
1275 #endif /* CONFIG_PGTABLE_MAPPING */
1276 
1277 static int zs_cpu_prepare(unsigned int cpu)
1278 {
1279 	struct mapping_area *area;
1280 
1281 	area = &per_cpu(zs_map_area, cpu);
1282 	return __zs_cpu_up(area);
1283 }
1284 
1285 static int zs_cpu_dead(unsigned int cpu)
1286 {
1287 	struct mapping_area *area;
1288 
1289 	area = &per_cpu(zs_map_area, cpu);
1290 	__zs_cpu_down(area);
1291 	return 0;
1292 }
1293 
1294 static bool can_merge(struct size_class *prev, int pages_per_zspage,
1295 					int objs_per_zspage)
1296 {
1297 	if (prev->pages_per_zspage == pages_per_zspage &&
1298 		prev->objs_per_zspage == objs_per_zspage)
1299 		return true;
1300 
1301 	return false;
1302 }
1303 
1304 static bool zspage_full(struct size_class *class, struct zspage *zspage)
1305 {
1306 	return get_zspage_inuse(zspage) == class->objs_per_zspage;
1307 }
1308 
1309 unsigned long zs_get_total_pages(struct zs_pool *pool)
1310 {
1311 	return atomic_long_read(&pool->pages_allocated);
1312 }
1313 EXPORT_SYMBOL_GPL(zs_get_total_pages);
1314 
1315 /**
1316  * zs_map_object - get address of allocated object from handle.
1317  * @pool: pool from which the object was allocated
1318  * @handle: handle returned from zs_malloc
1319  *
1320  * Before using an object allocated from zs_malloc, it must be mapped using
1321  * this function. When done with the object, it must be unmapped using
1322  * zs_unmap_object.
1323  *
1324  * Only one object can be mapped per cpu at a time. There is no protection
1325  * against nested mappings.
1326  *
1327  * This function returns with preemption and page faults disabled.
1328  */
1329 void *zs_map_object(struct zs_pool *pool, unsigned long handle,
1330 			enum zs_mapmode mm)
1331 {
1332 	struct zspage *zspage;
1333 	struct page *page;
1334 	unsigned long obj, off;
1335 	unsigned int obj_idx;
1336 
1337 	unsigned int class_idx;
1338 	enum fullness_group fg;
1339 	struct size_class *class;
1340 	struct mapping_area *area;
1341 	struct page *pages[2];
1342 	void *ret;
1343 
1344 	/*
1345 	 * Because we use per-cpu mapping areas shared among the
1346 	 * pools/users, we can't allow mapping in interrupt context
1347 	 * because it can corrupt another users mappings.
1348 	 */
1349 	WARN_ON_ONCE(in_interrupt());
1350 
1351 	/* From now on, migration cannot move the object */
1352 	pin_tag(handle);
1353 
1354 	obj = handle_to_obj(handle);
1355 	obj_to_location(obj, &page, &obj_idx);
1356 	zspage = get_zspage(page);
1357 
1358 	/* migration cannot move any subpage in this zspage */
1359 	migrate_read_lock(zspage);
1360 
1361 	get_zspage_mapping(zspage, &class_idx, &fg);
1362 	class = pool->size_class[class_idx];
1363 	off = (class->size * obj_idx) & ~PAGE_MASK;
1364 
1365 	area = &get_cpu_var(zs_map_area);
1366 	area->vm_mm = mm;
1367 	if (off + class->size <= PAGE_SIZE) {
1368 		/* this object is contained entirely within a page */
1369 		area->vm_addr = kmap_atomic(page);
1370 		ret = area->vm_addr + off;
1371 		goto out;
1372 	}
1373 
1374 	/* this object spans two pages */
1375 	pages[0] = page;
1376 	pages[1] = get_next_page(page);
1377 	BUG_ON(!pages[1]);
1378 
1379 	ret = __zs_map_object(area, pages, off, class->size);
1380 out:
1381 	if (likely(!PageHugeObject(page)))
1382 		ret += ZS_HANDLE_SIZE;
1383 
1384 	return ret;
1385 }
1386 EXPORT_SYMBOL_GPL(zs_map_object);
1387 
1388 void zs_unmap_object(struct zs_pool *pool, unsigned long handle)
1389 {
1390 	struct zspage *zspage;
1391 	struct page *page;
1392 	unsigned long obj, off;
1393 	unsigned int obj_idx;
1394 
1395 	unsigned int class_idx;
1396 	enum fullness_group fg;
1397 	struct size_class *class;
1398 	struct mapping_area *area;
1399 
1400 	obj = handle_to_obj(handle);
1401 	obj_to_location(obj, &page, &obj_idx);
1402 	zspage = get_zspage(page);
1403 	get_zspage_mapping(zspage, &class_idx, &fg);
1404 	class = pool->size_class[class_idx];
1405 	off = (class->size * obj_idx) & ~PAGE_MASK;
1406 
1407 	area = this_cpu_ptr(&zs_map_area);
1408 	if (off + class->size <= PAGE_SIZE)
1409 		kunmap_atomic(area->vm_addr);
1410 	else {
1411 		struct page *pages[2];
1412 
1413 		pages[0] = page;
1414 		pages[1] = get_next_page(page);
1415 		BUG_ON(!pages[1]);
1416 
1417 		__zs_unmap_object(area, pages, off, class->size);
1418 	}
1419 	put_cpu_var(zs_map_area);
1420 
1421 	migrate_read_unlock(zspage);
1422 	unpin_tag(handle);
1423 }
1424 EXPORT_SYMBOL_GPL(zs_unmap_object);
1425 
1426 static unsigned long obj_malloc(struct size_class *class,
1427 				struct zspage *zspage, unsigned long handle)
1428 {
1429 	int i, nr_page, offset;
1430 	unsigned long obj;
1431 	struct link_free *link;
1432 
1433 	struct page *m_page;
1434 	unsigned long m_offset;
1435 	void *vaddr;
1436 
1437 	handle |= OBJ_ALLOCATED_TAG;
1438 	obj = get_freeobj(zspage);
1439 
1440 	offset = obj * class->size;
1441 	nr_page = offset >> PAGE_SHIFT;
1442 	m_offset = offset & ~PAGE_MASK;
1443 	m_page = get_first_page(zspage);
1444 
1445 	for (i = 0; i < nr_page; i++)
1446 		m_page = get_next_page(m_page);
1447 
1448 	vaddr = kmap_atomic(m_page);
1449 	link = (struct link_free *)vaddr + m_offset / sizeof(*link);
1450 	set_freeobj(zspage, link->next >> OBJ_TAG_BITS);
1451 	if (likely(!PageHugeObject(m_page)))
1452 		/* record handle in the header of allocated chunk */
1453 		link->handle = handle;
1454 	else
1455 		/* record handle to page->index */
1456 		zspage->first_page->index = handle;
1457 
1458 	kunmap_atomic(vaddr);
1459 	mod_zspage_inuse(zspage, 1);
1460 	zs_stat_inc(class, OBJ_USED, 1);
1461 
1462 	obj = location_to_obj(m_page, obj);
1463 
1464 	return obj;
1465 }
1466 
1467 
1468 /**
1469  * zs_malloc - Allocate block of given size from pool.
1470  * @pool: pool to allocate from
1471  * @size: size of block to allocate
1472  * @gfp: gfp flags when allocating object
1473  *
1474  * On success, handle to the allocated object is returned,
1475  * otherwise 0.
1476  * Allocation requests with size > ZS_MAX_ALLOC_SIZE will fail.
1477  */
1478 unsigned long zs_malloc(struct zs_pool *pool, size_t size, gfp_t gfp)
1479 {
1480 	unsigned long handle, obj;
1481 	struct size_class *class;
1482 	enum fullness_group newfg;
1483 	struct zspage *zspage;
1484 
1485 	if (unlikely(!size || size > ZS_MAX_ALLOC_SIZE))
1486 		return 0;
1487 
1488 	handle = cache_alloc_handle(pool, gfp);
1489 	if (!handle)
1490 		return 0;
1491 
1492 	/* extra space in chunk to keep the handle */
1493 	size += ZS_HANDLE_SIZE;
1494 	class = pool->size_class[get_size_class_index(size)];
1495 
1496 	spin_lock(&class->lock);
1497 	zspage = find_get_zspage(class);
1498 	if (likely(zspage)) {
1499 		obj = obj_malloc(class, zspage, handle);
1500 		/* Now move the zspage to another fullness group, if required */
1501 		fix_fullness_group(class, zspage);
1502 		record_obj(handle, obj);
1503 		spin_unlock(&class->lock);
1504 
1505 		return handle;
1506 	}
1507 
1508 	spin_unlock(&class->lock);
1509 
1510 	zspage = alloc_zspage(pool, class, gfp);
1511 	if (!zspage) {
1512 		cache_free_handle(pool, handle);
1513 		return 0;
1514 	}
1515 
1516 	spin_lock(&class->lock);
1517 	obj = obj_malloc(class, zspage, handle);
1518 	newfg = get_fullness_group(class, zspage);
1519 	insert_zspage(class, zspage, newfg);
1520 	set_zspage_mapping(zspage, class->index, newfg);
1521 	record_obj(handle, obj);
1522 	atomic_long_add(class->pages_per_zspage,
1523 				&pool->pages_allocated);
1524 	zs_stat_inc(class, OBJ_ALLOCATED, class->objs_per_zspage);
1525 
1526 	/* We completely set up zspage so mark them as movable */
1527 	SetZsPageMovable(pool, zspage);
1528 	spin_unlock(&class->lock);
1529 
1530 	return handle;
1531 }
1532 EXPORT_SYMBOL_GPL(zs_malloc);
1533 
1534 static void obj_free(struct size_class *class, unsigned long obj)
1535 {
1536 	struct link_free *link;
1537 	struct zspage *zspage;
1538 	struct page *f_page;
1539 	unsigned long f_offset;
1540 	unsigned int f_objidx;
1541 	void *vaddr;
1542 
1543 	obj &= ~OBJ_ALLOCATED_TAG;
1544 	obj_to_location(obj, &f_page, &f_objidx);
1545 	f_offset = (class->size * f_objidx) & ~PAGE_MASK;
1546 	zspage = get_zspage(f_page);
1547 
1548 	vaddr = kmap_atomic(f_page);
1549 
1550 	/* Insert this object in containing zspage's freelist */
1551 	link = (struct link_free *)(vaddr + f_offset);
1552 	link->next = get_freeobj(zspage) << OBJ_TAG_BITS;
1553 	kunmap_atomic(vaddr);
1554 	set_freeobj(zspage, f_objidx);
1555 	mod_zspage_inuse(zspage, -1);
1556 	zs_stat_dec(class, OBJ_USED, 1);
1557 }
1558 
1559 void zs_free(struct zs_pool *pool, unsigned long handle)
1560 {
1561 	struct zspage *zspage;
1562 	struct page *f_page;
1563 	unsigned long obj;
1564 	unsigned int f_objidx;
1565 	int class_idx;
1566 	struct size_class *class;
1567 	enum fullness_group fullness;
1568 	bool isolated;
1569 
1570 	if (unlikely(!handle))
1571 		return;
1572 
1573 	pin_tag(handle);
1574 	obj = handle_to_obj(handle);
1575 	obj_to_location(obj, &f_page, &f_objidx);
1576 	zspage = get_zspage(f_page);
1577 
1578 	migrate_read_lock(zspage);
1579 
1580 	get_zspage_mapping(zspage, &class_idx, &fullness);
1581 	class = pool->size_class[class_idx];
1582 
1583 	spin_lock(&class->lock);
1584 	obj_free(class, obj);
1585 	fullness = fix_fullness_group(class, zspage);
1586 	if (fullness != ZS_EMPTY) {
1587 		migrate_read_unlock(zspage);
1588 		goto out;
1589 	}
1590 
1591 	isolated = is_zspage_isolated(zspage);
1592 	migrate_read_unlock(zspage);
1593 	/* If zspage is isolated, zs_page_putback will free the zspage */
1594 	if (likely(!isolated))
1595 		free_zspage(pool, class, zspage);
1596 out:
1597 
1598 	spin_unlock(&class->lock);
1599 	unpin_tag(handle);
1600 	cache_free_handle(pool, handle);
1601 }
1602 EXPORT_SYMBOL_GPL(zs_free);
1603 
1604 static void zs_object_copy(struct size_class *class, unsigned long dst,
1605 				unsigned long src)
1606 {
1607 	struct page *s_page, *d_page;
1608 	unsigned int s_objidx, d_objidx;
1609 	unsigned long s_off, d_off;
1610 	void *s_addr, *d_addr;
1611 	int s_size, d_size, size;
1612 	int written = 0;
1613 
1614 	s_size = d_size = class->size;
1615 
1616 	obj_to_location(src, &s_page, &s_objidx);
1617 	obj_to_location(dst, &d_page, &d_objidx);
1618 
1619 	s_off = (class->size * s_objidx) & ~PAGE_MASK;
1620 	d_off = (class->size * d_objidx) & ~PAGE_MASK;
1621 
1622 	if (s_off + class->size > PAGE_SIZE)
1623 		s_size = PAGE_SIZE - s_off;
1624 
1625 	if (d_off + class->size > PAGE_SIZE)
1626 		d_size = PAGE_SIZE - d_off;
1627 
1628 	s_addr = kmap_atomic(s_page);
1629 	d_addr = kmap_atomic(d_page);
1630 
1631 	while (1) {
1632 		size = min(s_size, d_size);
1633 		memcpy(d_addr + d_off, s_addr + s_off, size);
1634 		written += size;
1635 
1636 		if (written == class->size)
1637 			break;
1638 
1639 		s_off += size;
1640 		s_size -= size;
1641 		d_off += size;
1642 		d_size -= size;
1643 
1644 		if (s_off >= PAGE_SIZE) {
1645 			kunmap_atomic(d_addr);
1646 			kunmap_atomic(s_addr);
1647 			s_page = get_next_page(s_page);
1648 			s_addr = kmap_atomic(s_page);
1649 			d_addr = kmap_atomic(d_page);
1650 			s_size = class->size - written;
1651 			s_off = 0;
1652 		}
1653 
1654 		if (d_off >= PAGE_SIZE) {
1655 			kunmap_atomic(d_addr);
1656 			d_page = get_next_page(d_page);
1657 			d_addr = kmap_atomic(d_page);
1658 			d_size = class->size - written;
1659 			d_off = 0;
1660 		}
1661 	}
1662 
1663 	kunmap_atomic(d_addr);
1664 	kunmap_atomic(s_addr);
1665 }
1666 
1667 /*
1668  * Find alloced object in zspage from index object and
1669  * return handle.
1670  */
1671 static unsigned long find_alloced_obj(struct size_class *class,
1672 					struct page *page, int *obj_idx)
1673 {
1674 	unsigned long head;
1675 	int offset = 0;
1676 	int index = *obj_idx;
1677 	unsigned long handle = 0;
1678 	void *addr = kmap_atomic(page);
1679 
1680 	offset = get_first_obj_offset(page);
1681 	offset += class->size * index;
1682 
1683 	while (offset < PAGE_SIZE) {
1684 		head = obj_to_head(page, addr + offset);
1685 		if (head & OBJ_ALLOCATED_TAG) {
1686 			handle = head & ~OBJ_ALLOCATED_TAG;
1687 			if (trypin_tag(handle))
1688 				break;
1689 			handle = 0;
1690 		}
1691 
1692 		offset += class->size;
1693 		index++;
1694 	}
1695 
1696 	kunmap_atomic(addr);
1697 
1698 	*obj_idx = index;
1699 
1700 	return handle;
1701 }
1702 
1703 struct zs_compact_control {
1704 	/* Source spage for migration which could be a subpage of zspage */
1705 	struct page *s_page;
1706 	/* Destination page for migration which should be a first page
1707 	 * of zspage. */
1708 	struct page *d_page;
1709 	 /* Starting object index within @s_page which used for live object
1710 	  * in the subpage. */
1711 	int obj_idx;
1712 };
1713 
1714 static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
1715 				struct zs_compact_control *cc)
1716 {
1717 	unsigned long used_obj, free_obj;
1718 	unsigned long handle;
1719 	struct page *s_page = cc->s_page;
1720 	struct page *d_page = cc->d_page;
1721 	int obj_idx = cc->obj_idx;
1722 	int ret = 0;
1723 
1724 	while (1) {
1725 		handle = find_alloced_obj(class, s_page, &obj_idx);
1726 		if (!handle) {
1727 			s_page = get_next_page(s_page);
1728 			if (!s_page)
1729 				break;
1730 			obj_idx = 0;
1731 			continue;
1732 		}
1733 
1734 		/* Stop if there is no more space */
1735 		if (zspage_full(class, get_zspage(d_page))) {
1736 			unpin_tag(handle);
1737 			ret = -ENOMEM;
1738 			break;
1739 		}
1740 
1741 		used_obj = handle_to_obj(handle);
1742 		free_obj = obj_malloc(class, get_zspage(d_page), handle);
1743 		zs_object_copy(class, free_obj, used_obj);
1744 		obj_idx++;
1745 		/*
1746 		 * record_obj updates handle's value to free_obj and it will
1747 		 * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
1748 		 * breaks synchronization using pin_tag(e,g, zs_free) so
1749 		 * let's keep the lock bit.
1750 		 */
1751 		free_obj |= BIT(HANDLE_PIN_BIT);
1752 		record_obj(handle, free_obj);
1753 		unpin_tag(handle);
1754 		obj_free(class, used_obj);
1755 	}
1756 
1757 	/* Remember last position in this iteration */
1758 	cc->s_page = s_page;
1759 	cc->obj_idx = obj_idx;
1760 
1761 	return ret;
1762 }
1763 
1764 static struct zspage *isolate_zspage(struct size_class *class, bool source)
1765 {
1766 	int i;
1767 	struct zspage *zspage;
1768 	enum fullness_group fg[2] = {ZS_ALMOST_EMPTY, ZS_ALMOST_FULL};
1769 
1770 	if (!source) {
1771 		fg[0] = ZS_ALMOST_FULL;
1772 		fg[1] = ZS_ALMOST_EMPTY;
1773 	}
1774 
1775 	for (i = 0; i < 2; i++) {
1776 		zspage = list_first_entry_or_null(&class->fullness_list[fg[i]],
1777 							struct zspage, list);
1778 		if (zspage) {
1779 			VM_BUG_ON(is_zspage_isolated(zspage));
1780 			remove_zspage(class, zspage, fg[i]);
1781 			return zspage;
1782 		}
1783 	}
1784 
1785 	return zspage;
1786 }
1787 
1788 /*
1789  * putback_zspage - add @zspage into right class's fullness list
1790  * @class: destination class
1791  * @zspage: target page
1792  *
1793  * Return @zspage's fullness_group
1794  */
1795 static enum fullness_group putback_zspage(struct size_class *class,
1796 			struct zspage *zspage)
1797 {
1798 	enum fullness_group fullness;
1799 
1800 	VM_BUG_ON(is_zspage_isolated(zspage));
1801 
1802 	fullness = get_fullness_group(class, zspage);
1803 	insert_zspage(class, zspage, fullness);
1804 	set_zspage_mapping(zspage, class->index, fullness);
1805 
1806 	return fullness;
1807 }
1808 
1809 #ifdef CONFIG_COMPACTION
1810 static struct dentry *zs_mount(struct file_system_type *fs_type,
1811 				int flags, const char *dev_name, void *data)
1812 {
1813 	static const struct dentry_operations ops = {
1814 		.d_dname = simple_dname,
1815 	};
1816 
1817 	return mount_pseudo(fs_type, "zsmalloc:", NULL, &ops, ZSMALLOC_MAGIC);
1818 }
1819 
1820 static struct file_system_type zsmalloc_fs = {
1821 	.name		= "zsmalloc",
1822 	.mount		= zs_mount,
1823 	.kill_sb	= kill_anon_super,
1824 };
1825 
1826 static int zsmalloc_mount(void)
1827 {
1828 	int ret = 0;
1829 
1830 	zsmalloc_mnt = kern_mount(&zsmalloc_fs);
1831 	if (IS_ERR(zsmalloc_mnt))
1832 		ret = PTR_ERR(zsmalloc_mnt);
1833 
1834 	return ret;
1835 }
1836 
1837 static void zsmalloc_unmount(void)
1838 {
1839 	kern_unmount(zsmalloc_mnt);
1840 }
1841 
1842 static void migrate_lock_init(struct zspage *zspage)
1843 {
1844 	rwlock_init(&zspage->lock);
1845 }
1846 
1847 static void migrate_read_lock(struct zspage *zspage)
1848 {
1849 	read_lock(&zspage->lock);
1850 }
1851 
1852 static void migrate_read_unlock(struct zspage *zspage)
1853 {
1854 	read_unlock(&zspage->lock);
1855 }
1856 
1857 static void migrate_write_lock(struct zspage *zspage)
1858 {
1859 	write_lock(&zspage->lock);
1860 }
1861 
1862 static void migrate_write_unlock(struct zspage *zspage)
1863 {
1864 	write_unlock(&zspage->lock);
1865 }
1866 
1867 /* Number of isolated subpage for *page migration* in this zspage */
1868 static void inc_zspage_isolation(struct zspage *zspage)
1869 {
1870 	zspage->isolated++;
1871 }
1872 
1873 static void dec_zspage_isolation(struct zspage *zspage)
1874 {
1875 	zspage->isolated--;
1876 }
1877 
1878 static void replace_sub_page(struct size_class *class, struct zspage *zspage,
1879 				struct page *newpage, struct page *oldpage)
1880 {
1881 	struct page *page;
1882 	struct page *pages[ZS_MAX_PAGES_PER_ZSPAGE] = {NULL, };
1883 	int idx = 0;
1884 
1885 	page = get_first_page(zspage);
1886 	do {
1887 		if (page == oldpage)
1888 			pages[idx] = newpage;
1889 		else
1890 			pages[idx] = page;
1891 		idx++;
1892 	} while ((page = get_next_page(page)) != NULL);
1893 
1894 	create_page_chain(class, zspage, pages);
1895 	set_first_obj_offset(newpage, get_first_obj_offset(oldpage));
1896 	if (unlikely(PageHugeObject(oldpage)))
1897 		newpage->index = oldpage->index;
1898 	__SetPageMovable(newpage, page_mapping(oldpage));
1899 }
1900 
1901 bool zs_page_isolate(struct page *page, isolate_mode_t mode)
1902 {
1903 	struct zs_pool *pool;
1904 	struct size_class *class;
1905 	int class_idx;
1906 	enum fullness_group fullness;
1907 	struct zspage *zspage;
1908 	struct address_space *mapping;
1909 
1910 	/*
1911 	 * Page is locked so zspage couldn't be destroyed. For detail, look at
1912 	 * lock_zspage in free_zspage.
1913 	 */
1914 	VM_BUG_ON_PAGE(!PageMovable(page), page);
1915 	VM_BUG_ON_PAGE(PageIsolated(page), page);
1916 
1917 	zspage = get_zspage(page);
1918 
1919 	/*
1920 	 * Without class lock, fullness could be stale while class_idx is okay
1921 	 * because class_idx is constant unless page is freed so we should get
1922 	 * fullness again under class lock.
1923 	 */
1924 	get_zspage_mapping(zspage, &class_idx, &fullness);
1925 	mapping = page_mapping(page);
1926 	pool = mapping->private_data;
1927 	class = pool->size_class[class_idx];
1928 
1929 	spin_lock(&class->lock);
1930 	if (get_zspage_inuse(zspage) == 0) {
1931 		spin_unlock(&class->lock);
1932 		return false;
1933 	}
1934 
1935 	/* zspage is isolated for object migration */
1936 	if (list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1937 		spin_unlock(&class->lock);
1938 		return false;
1939 	}
1940 
1941 	/*
1942 	 * If this is first time isolation for the zspage, isolate zspage from
1943 	 * size_class to prevent further object allocation from the zspage.
1944 	 */
1945 	if (!list_empty(&zspage->list) && !is_zspage_isolated(zspage)) {
1946 		get_zspage_mapping(zspage, &class_idx, &fullness);
1947 		remove_zspage(class, zspage, fullness);
1948 	}
1949 
1950 	inc_zspage_isolation(zspage);
1951 	spin_unlock(&class->lock);
1952 
1953 	return true;
1954 }
1955 
1956 int zs_page_migrate(struct address_space *mapping, struct page *newpage,
1957 		struct page *page, enum migrate_mode mode)
1958 {
1959 	struct zs_pool *pool;
1960 	struct size_class *class;
1961 	int class_idx;
1962 	enum fullness_group fullness;
1963 	struct zspage *zspage;
1964 	struct page *dummy;
1965 	void *s_addr, *d_addr, *addr;
1966 	int offset, pos;
1967 	unsigned long handle, head;
1968 	unsigned long old_obj, new_obj;
1969 	unsigned int obj_idx;
1970 	int ret = -EAGAIN;
1971 
1972 	VM_BUG_ON_PAGE(!PageMovable(page), page);
1973 	VM_BUG_ON_PAGE(!PageIsolated(page), page);
1974 
1975 	zspage = get_zspage(page);
1976 
1977 	/* Concurrent compactor cannot migrate any subpage in zspage */
1978 	migrate_write_lock(zspage);
1979 	get_zspage_mapping(zspage, &class_idx, &fullness);
1980 	pool = mapping->private_data;
1981 	class = pool->size_class[class_idx];
1982 	offset = get_first_obj_offset(page);
1983 
1984 	spin_lock(&class->lock);
1985 	if (!get_zspage_inuse(zspage)) {
1986 		ret = -EBUSY;
1987 		goto unlock_class;
1988 	}
1989 
1990 	pos = offset;
1991 	s_addr = kmap_atomic(page);
1992 	while (pos < PAGE_SIZE) {
1993 		head = obj_to_head(page, s_addr + pos);
1994 		if (head & OBJ_ALLOCATED_TAG) {
1995 			handle = head & ~OBJ_ALLOCATED_TAG;
1996 			if (!trypin_tag(handle))
1997 				goto unpin_objects;
1998 		}
1999 		pos += class->size;
2000 	}
2001 
2002 	/*
2003 	 * Here, any user cannot access all objects in the zspage so let's move.
2004 	 */
2005 	d_addr = kmap_atomic(newpage);
2006 	memcpy(d_addr, s_addr, PAGE_SIZE);
2007 	kunmap_atomic(d_addr);
2008 
2009 	for (addr = s_addr + offset; addr < s_addr + pos;
2010 					addr += class->size) {
2011 		head = obj_to_head(page, addr);
2012 		if (head & OBJ_ALLOCATED_TAG) {
2013 			handle = head & ~OBJ_ALLOCATED_TAG;
2014 			if (!testpin_tag(handle))
2015 				BUG();
2016 
2017 			old_obj = handle_to_obj(handle);
2018 			obj_to_location(old_obj, &dummy, &obj_idx);
2019 			new_obj = (unsigned long)location_to_obj(newpage,
2020 								obj_idx);
2021 			new_obj |= BIT(HANDLE_PIN_BIT);
2022 			record_obj(handle, new_obj);
2023 		}
2024 	}
2025 
2026 	replace_sub_page(class, zspage, newpage, page);
2027 	get_page(newpage);
2028 
2029 	dec_zspage_isolation(zspage);
2030 
2031 	/*
2032 	 * Page migration is done so let's putback isolated zspage to
2033 	 * the list if @page is final isolated subpage in the zspage.
2034 	 */
2035 	if (!is_zspage_isolated(zspage))
2036 		putback_zspage(class, zspage);
2037 
2038 	reset_page(page);
2039 	put_page(page);
2040 	page = newpage;
2041 
2042 	ret = MIGRATEPAGE_SUCCESS;
2043 unpin_objects:
2044 	for (addr = s_addr + offset; addr < s_addr + pos;
2045 						addr += class->size) {
2046 		head = obj_to_head(page, addr);
2047 		if (head & OBJ_ALLOCATED_TAG) {
2048 			handle = head & ~OBJ_ALLOCATED_TAG;
2049 			if (!testpin_tag(handle))
2050 				BUG();
2051 			unpin_tag(handle);
2052 		}
2053 	}
2054 	kunmap_atomic(s_addr);
2055 unlock_class:
2056 	spin_unlock(&class->lock);
2057 	migrate_write_unlock(zspage);
2058 
2059 	return ret;
2060 }
2061 
2062 void zs_page_putback(struct page *page)
2063 {
2064 	struct zs_pool *pool;
2065 	struct size_class *class;
2066 	int class_idx;
2067 	enum fullness_group fg;
2068 	struct address_space *mapping;
2069 	struct zspage *zspage;
2070 
2071 	VM_BUG_ON_PAGE(!PageMovable(page), page);
2072 	VM_BUG_ON_PAGE(!PageIsolated(page), page);
2073 
2074 	zspage = get_zspage(page);
2075 	get_zspage_mapping(zspage, &class_idx, &fg);
2076 	mapping = page_mapping(page);
2077 	pool = mapping->private_data;
2078 	class = pool->size_class[class_idx];
2079 
2080 	spin_lock(&class->lock);
2081 	dec_zspage_isolation(zspage);
2082 	if (!is_zspage_isolated(zspage)) {
2083 		fg = putback_zspage(class, zspage);
2084 		/*
2085 		 * Due to page_lock, we cannot free zspage immediately
2086 		 * so let's defer.
2087 		 */
2088 		if (fg == ZS_EMPTY)
2089 			schedule_work(&pool->free_work);
2090 	}
2091 	spin_unlock(&class->lock);
2092 }
2093 
2094 const struct address_space_operations zsmalloc_aops = {
2095 	.isolate_page = zs_page_isolate,
2096 	.migratepage = zs_page_migrate,
2097 	.putback_page = zs_page_putback,
2098 };
2099 
2100 static int zs_register_migration(struct zs_pool *pool)
2101 {
2102 	pool->inode = alloc_anon_inode(zsmalloc_mnt->mnt_sb);
2103 	if (IS_ERR(pool->inode)) {
2104 		pool->inode = NULL;
2105 		return 1;
2106 	}
2107 
2108 	pool->inode->i_mapping->private_data = pool;
2109 	pool->inode->i_mapping->a_ops = &zsmalloc_aops;
2110 	return 0;
2111 }
2112 
2113 static void zs_unregister_migration(struct zs_pool *pool)
2114 {
2115 	flush_work(&pool->free_work);
2116 	iput(pool->inode);
2117 }
2118 
2119 /*
2120  * Caller should hold page_lock of all pages in the zspage
2121  * In here, we cannot use zspage meta data.
2122  */
2123 static void async_free_zspage(struct work_struct *work)
2124 {
2125 	int i;
2126 	struct size_class *class;
2127 	unsigned int class_idx;
2128 	enum fullness_group fullness;
2129 	struct zspage *zspage, *tmp;
2130 	LIST_HEAD(free_pages);
2131 	struct zs_pool *pool = container_of(work, struct zs_pool,
2132 					free_work);
2133 
2134 	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2135 		class = pool->size_class[i];
2136 		if (class->index != i)
2137 			continue;
2138 
2139 		spin_lock(&class->lock);
2140 		list_splice_init(&class->fullness_list[ZS_EMPTY], &free_pages);
2141 		spin_unlock(&class->lock);
2142 	}
2143 
2144 
2145 	list_for_each_entry_safe(zspage, tmp, &free_pages, list) {
2146 		list_del(&zspage->list);
2147 		lock_zspage(zspage);
2148 
2149 		get_zspage_mapping(zspage, &class_idx, &fullness);
2150 		VM_BUG_ON(fullness != ZS_EMPTY);
2151 		class = pool->size_class[class_idx];
2152 		spin_lock(&class->lock);
2153 		__free_zspage(pool, pool->size_class[class_idx], zspage);
2154 		spin_unlock(&class->lock);
2155 	}
2156 };
2157 
2158 static void kick_deferred_free(struct zs_pool *pool)
2159 {
2160 	schedule_work(&pool->free_work);
2161 }
2162 
2163 static void init_deferred_free(struct zs_pool *pool)
2164 {
2165 	INIT_WORK(&pool->free_work, async_free_zspage);
2166 }
2167 
2168 static void SetZsPageMovable(struct zs_pool *pool, struct zspage *zspage)
2169 {
2170 	struct page *page = get_first_page(zspage);
2171 
2172 	do {
2173 		WARN_ON(!trylock_page(page));
2174 		__SetPageMovable(page, pool->inode->i_mapping);
2175 		unlock_page(page);
2176 	} while ((page = get_next_page(page)) != NULL);
2177 }
2178 #endif
2179 
2180 /*
2181  *
2182  * Based on the number of unused allocated objects calculate
2183  * and return the number of pages that we can free.
2184  */
2185 static unsigned long zs_can_compact(struct size_class *class)
2186 {
2187 	unsigned long obj_wasted;
2188 	unsigned long obj_allocated = zs_stat_get(class, OBJ_ALLOCATED);
2189 	unsigned long obj_used = zs_stat_get(class, OBJ_USED);
2190 
2191 	if (obj_allocated <= obj_used)
2192 		return 0;
2193 
2194 	obj_wasted = obj_allocated - obj_used;
2195 	obj_wasted /= class->objs_per_zspage;
2196 
2197 	return obj_wasted * class->pages_per_zspage;
2198 }
2199 
2200 static void __zs_compact(struct zs_pool *pool, struct size_class *class)
2201 {
2202 	struct zs_compact_control cc;
2203 	struct zspage *src_zspage;
2204 	struct zspage *dst_zspage = NULL;
2205 
2206 	spin_lock(&class->lock);
2207 	while ((src_zspage = isolate_zspage(class, true))) {
2208 
2209 		if (!zs_can_compact(class))
2210 			break;
2211 
2212 		cc.obj_idx = 0;
2213 		cc.s_page = get_first_page(src_zspage);
2214 
2215 		while ((dst_zspage = isolate_zspage(class, false))) {
2216 			cc.d_page = get_first_page(dst_zspage);
2217 			/*
2218 			 * If there is no more space in dst_page, resched
2219 			 * and see if anyone had allocated another zspage.
2220 			 */
2221 			if (!migrate_zspage(pool, class, &cc))
2222 				break;
2223 
2224 			putback_zspage(class, dst_zspage);
2225 		}
2226 
2227 		/* Stop if we couldn't find slot */
2228 		if (dst_zspage == NULL)
2229 			break;
2230 
2231 		putback_zspage(class, dst_zspage);
2232 		if (putback_zspage(class, src_zspage) == ZS_EMPTY) {
2233 			free_zspage(pool, class, src_zspage);
2234 			pool->stats.pages_compacted += class->pages_per_zspage;
2235 		}
2236 		spin_unlock(&class->lock);
2237 		cond_resched();
2238 		spin_lock(&class->lock);
2239 	}
2240 
2241 	if (src_zspage)
2242 		putback_zspage(class, src_zspage);
2243 
2244 	spin_unlock(&class->lock);
2245 }
2246 
2247 unsigned long zs_compact(struct zs_pool *pool)
2248 {
2249 	int i;
2250 	struct size_class *class;
2251 
2252 	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2253 		class = pool->size_class[i];
2254 		if (!class)
2255 			continue;
2256 		if (class->index != i)
2257 			continue;
2258 		__zs_compact(pool, class);
2259 	}
2260 
2261 	return pool->stats.pages_compacted;
2262 }
2263 EXPORT_SYMBOL_GPL(zs_compact);
2264 
2265 void zs_pool_stats(struct zs_pool *pool, struct zs_pool_stats *stats)
2266 {
2267 	memcpy(stats, &pool->stats, sizeof(struct zs_pool_stats));
2268 }
2269 EXPORT_SYMBOL_GPL(zs_pool_stats);
2270 
2271 static unsigned long zs_shrinker_scan(struct shrinker *shrinker,
2272 		struct shrink_control *sc)
2273 {
2274 	unsigned long pages_freed;
2275 	struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2276 			shrinker);
2277 
2278 	pages_freed = pool->stats.pages_compacted;
2279 	/*
2280 	 * Compact classes and calculate compaction delta.
2281 	 * Can run concurrently with a manually triggered
2282 	 * (by user) compaction.
2283 	 */
2284 	pages_freed = zs_compact(pool) - pages_freed;
2285 
2286 	return pages_freed ? pages_freed : SHRINK_STOP;
2287 }
2288 
2289 static unsigned long zs_shrinker_count(struct shrinker *shrinker,
2290 		struct shrink_control *sc)
2291 {
2292 	int i;
2293 	struct size_class *class;
2294 	unsigned long pages_to_free = 0;
2295 	struct zs_pool *pool = container_of(shrinker, struct zs_pool,
2296 			shrinker);
2297 
2298 	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2299 		class = pool->size_class[i];
2300 		if (!class)
2301 			continue;
2302 		if (class->index != i)
2303 			continue;
2304 
2305 		pages_to_free += zs_can_compact(class);
2306 	}
2307 
2308 	return pages_to_free;
2309 }
2310 
2311 static void zs_unregister_shrinker(struct zs_pool *pool)
2312 {
2313 	if (pool->shrinker_enabled) {
2314 		unregister_shrinker(&pool->shrinker);
2315 		pool->shrinker_enabled = false;
2316 	}
2317 }
2318 
2319 static int zs_register_shrinker(struct zs_pool *pool)
2320 {
2321 	pool->shrinker.scan_objects = zs_shrinker_scan;
2322 	pool->shrinker.count_objects = zs_shrinker_count;
2323 	pool->shrinker.batch = 0;
2324 	pool->shrinker.seeks = DEFAULT_SEEKS;
2325 
2326 	return register_shrinker(&pool->shrinker);
2327 }
2328 
2329 /**
2330  * zs_create_pool - Creates an allocation pool to work from.
2331  * @name: pool name to be created
2332  *
2333  * This function must be called before anything when using
2334  * the zsmalloc allocator.
2335  *
2336  * On success, a pointer to the newly created pool is returned,
2337  * otherwise NULL.
2338  */
2339 struct zs_pool *zs_create_pool(const char *name)
2340 {
2341 	int i;
2342 	struct zs_pool *pool;
2343 	struct size_class *prev_class = NULL;
2344 
2345 	pool = kzalloc(sizeof(*pool), GFP_KERNEL);
2346 	if (!pool)
2347 		return NULL;
2348 
2349 	init_deferred_free(pool);
2350 
2351 	pool->name = kstrdup(name, GFP_KERNEL);
2352 	if (!pool->name)
2353 		goto err;
2354 
2355 	if (create_cache(pool))
2356 		goto err;
2357 
2358 	/*
2359 	 * Iterate reversely, because, size of size_class that we want to use
2360 	 * for merging should be larger or equal to current size.
2361 	 */
2362 	for (i = ZS_SIZE_CLASSES - 1; i >= 0; i--) {
2363 		int size;
2364 		int pages_per_zspage;
2365 		int objs_per_zspage;
2366 		struct size_class *class;
2367 		int fullness = 0;
2368 
2369 		size = ZS_MIN_ALLOC_SIZE + i * ZS_SIZE_CLASS_DELTA;
2370 		if (size > ZS_MAX_ALLOC_SIZE)
2371 			size = ZS_MAX_ALLOC_SIZE;
2372 		pages_per_zspage = get_pages_per_zspage(size);
2373 		objs_per_zspage = pages_per_zspage * PAGE_SIZE / size;
2374 
2375 		/*
2376 		 * size_class is used for normal zsmalloc operation such
2377 		 * as alloc/free for that size. Although it is natural that we
2378 		 * have one size_class for each size, there is a chance that we
2379 		 * can get more memory utilization if we use one size_class for
2380 		 * many different sizes whose size_class have same
2381 		 * characteristics. So, we makes size_class point to
2382 		 * previous size_class if possible.
2383 		 */
2384 		if (prev_class) {
2385 			if (can_merge(prev_class, pages_per_zspage, objs_per_zspage)) {
2386 				pool->size_class[i] = prev_class;
2387 				continue;
2388 			}
2389 		}
2390 
2391 		class = kzalloc(sizeof(struct size_class), GFP_KERNEL);
2392 		if (!class)
2393 			goto err;
2394 
2395 		class->size = size;
2396 		class->index = i;
2397 		class->pages_per_zspage = pages_per_zspage;
2398 		class->objs_per_zspage = objs_per_zspage;
2399 		spin_lock_init(&class->lock);
2400 		pool->size_class[i] = class;
2401 		for (fullness = ZS_EMPTY; fullness < NR_ZS_FULLNESS;
2402 							fullness++)
2403 			INIT_LIST_HEAD(&class->fullness_list[fullness]);
2404 
2405 		prev_class = class;
2406 	}
2407 
2408 	/* debug only, don't abort if it fails */
2409 	zs_pool_stat_create(pool, name);
2410 
2411 	if (zs_register_migration(pool))
2412 		goto err;
2413 
2414 	/*
2415 	 * Not critical, we still can use the pool
2416 	 * and user can trigger compaction manually.
2417 	 */
2418 	if (zs_register_shrinker(pool) == 0)
2419 		pool->shrinker_enabled = true;
2420 	return pool;
2421 
2422 err:
2423 	zs_destroy_pool(pool);
2424 	return NULL;
2425 }
2426 EXPORT_SYMBOL_GPL(zs_create_pool);
2427 
2428 void zs_destroy_pool(struct zs_pool *pool)
2429 {
2430 	int i;
2431 
2432 	zs_unregister_shrinker(pool);
2433 	zs_unregister_migration(pool);
2434 	zs_pool_stat_destroy(pool);
2435 
2436 	for (i = 0; i < ZS_SIZE_CLASSES; i++) {
2437 		int fg;
2438 		struct size_class *class = pool->size_class[i];
2439 
2440 		if (!class)
2441 			continue;
2442 
2443 		if (class->index != i)
2444 			continue;
2445 
2446 		for (fg = ZS_EMPTY; fg < NR_ZS_FULLNESS; fg++) {
2447 			if (!list_empty(&class->fullness_list[fg])) {
2448 				pr_info("Freeing non-empty class with size %db, fullness group %d\n",
2449 					class->size, fg);
2450 			}
2451 		}
2452 		kfree(class);
2453 	}
2454 
2455 	destroy_cache(pool);
2456 	kfree(pool->size_class);
2457 	kfree(pool->name);
2458 	kfree(pool);
2459 }
2460 EXPORT_SYMBOL_GPL(zs_destroy_pool);
2461 
2462 static int __init zs_init(void)
2463 {
2464 	int ret;
2465 
2466 	ret = zsmalloc_mount();
2467 	if (ret)
2468 		goto out;
2469 
2470 	ret = cpuhp_setup_state(CPUHP_MM_ZS_PREPARE, "mm/zsmalloc:prepare",
2471 				zs_cpu_prepare, zs_cpu_dead);
2472 	if (ret)
2473 		goto hp_setup_fail;
2474 
2475 #ifdef CONFIG_ZPOOL
2476 	zpool_register_driver(&zs_zpool_driver);
2477 #endif
2478 
2479 	zs_stat_init();
2480 
2481 	return 0;
2482 
2483 hp_setup_fail:
2484 	zsmalloc_unmount();
2485 out:
2486 	return ret;
2487 }
2488 
2489 static void __exit zs_exit(void)
2490 {
2491 #ifdef CONFIG_ZPOOL
2492 	zpool_unregister_driver(&zs_zpool_driver);
2493 #endif
2494 	zsmalloc_unmount();
2495 	cpuhp_remove_state(CPUHP_MM_ZS_PREPARE);
2496 
2497 	zs_stat_exit();
2498 }
2499 
2500 module_init(zs_init);
2501 module_exit(zs_exit);
2502 
2503 MODULE_LICENSE("Dual BSD/GPL");
2504 MODULE_AUTHOR("Nitin Gupta <ngupta@vflare.org>");
2505