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